doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.ppat.1003626 | Small RNA sX13: A Multifaceted Regulator of Virulence in the Plant Pathogen Xanthomonas | Small noncoding RNAs (sRNAs) are ubiquitous posttranscriptional regulators of gene expression. Using the model plant-pathogenic bacterium Xanthomonas campestris pv. vesicatoria (Xcv), we investigated the highly expressed and conserved sRNA sX13 in detail. Deletion of sX13 impinged on Xcv virulence and the expression of genes encoding components and substrates of the Hrp type III secretion (T3S) system. qRT-PCR analyses revealed that sX13 promotes mRNA accumulation of HrpX, a key regulator of the T3S system, whereas the mRNA level of the master regulator HrpG was unaffected. Complementation studies suggest that sX13 acts upstream of HrpG. Microarray analyses identified 63 sX13-regulated genes, which are involved in signal transduction, motility, transcriptional and posttranscriptional regulation and virulence. Structure analyses of in vitro transcribed sX13 revealed a structure with three stable stems and three apical C-rich loops. A computational search for putative regulatory motifs revealed that sX13-repressed mRNAs predominantly harbor G-rich motifs in proximity of translation start sites. Mutation of sX13 loops differentially affected Xcv virulence and the mRNA abundance of putative targets. Using a GFP-based reporter system, we demonstrated that sX13-mediated repression of protein synthesis requires both the C-rich motifs in sX13 and G-rich motifs in potential target mRNAs. Although the RNA-binding protein Hfq was dispensable for sX13 activity, the hfq mRNA and Hfq::GFP abundance were negatively regulated by sX13. In addition, we found that G-rich motifs in sX13-repressed mRNAs can serve as translational enhancers and are located at the ribosome-binding site in 5% of all protein-coding Xcv genes. Our study revealed that sX13 represents a novel class of virulence regulators and provides insights into sRNA-mediated modulation of adaptive processes in the plant pathogen Xanthomonas.
| Since the discovery of the first regulatory RNA in 1981, hundreds of small RNAs (sRNAs) have been identified in bacteria. Although sRNA-mediated control of virulence was demonstrated for numerous animal- and human-pathogenic bacteria, sRNAs and their functions in plant-pathogenic bacteria have been enigmatic. We discovered that the sRNA sX13 is a novel virulence regulator of Xanthomonas campestris pv. vesicatoria (Xcv), which causes bacterial spot disease on pepper and tomato. sX13 contributes to the Xcv-plant interaction by promoting the synthesis of an essential pathogenicity factor of Xcv, i. e., the type III secretion system. Thus, in addition to transcriptional regulation, sRNA-mediated posttranscriptional regulation contributes to virulence of plant-pathogenic xanthomonads. To repress target mRNAs carrying G-rich motifs, sX13 employs C-rich loops. Hence, sX13 exhibits striking structural similarity to sRNAs in distantly related human pathogens, e. g., Staphylococcus aureus and Helicobacter pylori, suggesting that structure-driven target regulation via C-rich motifs represents a conserved feature of sRNA-mediated posttranscriptional regulation. Furthermore, sX13 is the first sRNA shown to control the mRNA level of hfq, which encodes a conserved RNA-binding protein required for sRNA activity and virulence in many enteric bacteria.
| The survival and prosperity of bacteria depends on their ability to adapt to a variety of environmental cues such as nutrient availability, osmolarity and temperature. Besides the adaptation to the environment by transcriptional regulation of gene expression bacteria express regulatory RNAs that modulate expression on the posttranscriptional level [1], [2]. Small regulatory RNAs (sRNAs; ∼50–300 nt) have been intensively studied in the enterobacteria Escherichia coli and Salmonella spp. and, in most cases, regulate translation and/or stability of target mRNAs through short and imperfect base-pairing (10 to 25 nucleotides) [1], [3], [4], [5]. The majority of characterized sRNAs inhibits translation of target mRNAs by pairing near or at the ribosome-binding site (RBS) [1], [6]. In addition, sRNAs can promote target mRNA translation, e. g., the sRNAs ArcZ, DsrA and RprA activate translation of sigma factor RpoS [7], [8], [9]. Regulation of multiple rather than single genes has emerged as a major feature of sRNAs affecting processes like iron homeostasis, carbon metabolism, stress responses and quorum sensing (QS) [1], [2], [6]. In numerous cases, sRNAs are under transcriptional control of two-component systems (TCS), which themselves are often controlled by sRNAs [10]. The activity and stability of most enterobacterial sRNAs requires the hexameric RNA-binding protein Hfq, which facilitates the formation of sRNA-mRNA duplexes and their subsequent degradation by the RNA degradosome [1], [11]. Hfq is present in approximately 50% of all bacterial species and acts in concert with sRNAs to regulate stress responses and virulence of a number of animal- and human-pathogenic bacteria [5], [12].
To date, little is known about sRNAs in plant-pathogenic bacteria. Only recently, high throughput RNA-sequencing approaches uncovered potential sRNAs in the plant-pathogenic α-proteobacterium Agrobacterium tumefaciens [13], the γ-proteobacteria Pseudomonas syringae pv. tomato [14] and Xanthomonas campestris pv. vesicatoria (Xcv) [15], [16]. Additional studies identified four and eight sRNAs in X. campestris pv. campestris (Xcc) and X. oryzae pv. oryzae (Xoo), respectively [17], [18]. So far, only few sRNAs of plant-pathogenic bacteria were characterized with respect to potential targets. Examples include the A. tumefaciens antisense RNA RepE and the sRNA AbcR1, which regulate Ti-plasmid replication and the expression of ABC transporters, respectively [19], [20]. RNAs involved in virulence of plant-pathogenic bacteria were so far only reported for Erwinia spp. and Xcv. In Erwinia, the protein-binding RNA RsmB modulates the activity of the translational repressor protein RsmA, which impacts on QS, the production of extracellular enzymes and virulence [21], [22], [23]. In Xcv, sX12 was reported to be required for full virulence [16].
Xanthomonads are ubiquitous plant-pathogenic bacteria that infect approximately 120 monocotyledonous and 270 dicotyledonous plant species, many of which are economically important crops [24], [25]. These pathogens, usually only found in association with plants and plant parts, differ from most other Gram-negative bacteria in their high G+C content (∼65%), and high numbers of TonB-dependent transporters and signaling proteins [26]. Pathogenicity of most Xanthomonas spp. and other Gram-negative plant- and animal-pathogenic bacteria relies on a type III secretion (T3S) system which translocates bacterial effector proteins into the eukaryotic host cell [27], [28]. In addition, other protein secretion systems, degradative enzymes and QS regulation contribute to virulence of Xanthomonas spp. [29], [30].
One of the models to study plant-pathogen interactions is Xcv, the causal agent of bacterial spot disease on pepper and tomato [31], [32]. The T3S system of Xcv is encoded by the hrp− [hypersensitive response (HR) and pathogenicity] gene cluster and translocates effector proteins into the plant cell where they interfere with host cellular processes to the benefit of the pathogen [29], [33], [34]. The mutation of hrp-genes abolishes bacterial growth in the plant tissue and the induction of the HR in resistant plants. The HR is a local and rapid programmed plant cell death at the infection site and coincides with the arrest of bacterial multiplication [33], [35], [36]. The expression of the T3S system is transcriptionally induced in the plant and in the synthetic medium XVM2, and is controlled by the key regulators HrpG and HrpX [37], [38], [39], [40]. The OmpR-type regulator HrpG induces transcription of hrpX which encodes an AraC-type activator [39], [41]. HrpG and HrpX control the expression of hrp, type III effector and other virulence genes [16], [29], [40], [42], [43]. Recently, dRNA-seq identified 24 sRNAs in Xcv strain 85-10, expression of eight of which is controlled by HrpG and HrpX, including the aforementioned sX12 sRNA [15], [16].
In this study, we aimed at a detailed characterization of sX13 from Xcv strain 85-10, which is an abundant and HrpG-/HrpX-independently expressed sRNA [16]. Using mutant and complementation studies, we discovered that sX13 promotes the expression of the T3S system and contributes to virulence of Xcv. Microarray and quantitative reverse transcription PCR (qRT-PCR) analyses identified a large sX13 regulon and G-rich motifs in presumed sX13-target mRNAs. Selected putative targets were analyzed by site-directed mutagenesis of sX13 and mRNA::gfp fusions. Furthermore, we provide evidence that sX13 acts Hfq-independently. Our study provides the first comprehensive characterization of a trans-encoded sRNA that contributes to virulence of a plant-pathogenic bacterium.
The sRNA sX13 (115 nt; [16]) is encoded in a 437-bp intergenic region of the Xcv 85-10 chromosome, i. e., 175 bp downstream of the stop codon of the DNA polymerase I-encoding gene polA and 148 bp upstream of the translation start site (TLS) of XCV4199, which encodes a hypothetical protein. Sequence comparisons revealed that the sX13 gene is exclusively found in members of the Xanthomonadaceae family, i. e., in the genomes of plant-pathogenic Xanthomonas and Xylella species, the human pathogen Stenotrophomonas maltophilia and non-pathogenic bacteria of the genus Pseudoxanthomonas. Interestingly, sX13 homologs are highly conserved [16] and always located downstream of polA. By contrast, sX13-flanking sequences are highly variable.
To characterize sX13 in Xcv strain 85-10, we introduced an unmarked sX13 deletion into the chromosome (see ‘Materials and Methods’). Analysis of bacterial growth of the sX13 mutant strain (ΔsX13) revealed a significantly reduced stationary-phase density compared to the Xcv wild-type strain 85-10 in complex medium (NYG; Figure 1A) and in minimal medium A (MMA; Figure 1B). The mutant phenotype of XcvΔsX13 was complemented by chromosomal re-integration of sX13 into the sX13 locus, termed ΔsX13+sX13ch (Figure 1A, B; see ‘Materials and Methods’).
To address a potential role of sX13 in virulence, we performed plant infection assays. As shown in Figure 1C, the Xcv strains 85-10 and ΔsX13 grew similarly in leaves of susceptible pepper plants (ECW). Strikingly, infection with the sX13 mutant resulted in strongly delayed disease symptoms in susceptible and a delayed HR in resistant pepper plants (ECW-10R) (Figure 1D). Ectopic expression of sX13 under control of the lac promoter (psX13), which is constitutive in Xcv [38], and re-integration of sX13 into the ΔsX13 locus fully complemented the mutant phenotype of XcvΔsX13 (Figure 1D; data not shown). Strain Xcv 85-10 carrying psX13 induced an accelerated HR in comparison to the wild type (data not shown).
As the HR induction in ECW-10R plants depends on the recognition of the bacterial type III effector protein AvrBs1 by the plant disease resistance gene Bs1 [44], [45], the in planta phenotype of XcvΔsX13 suggested a reduced activity of the T3S system. To address this question, we investigated protein accumulation of selected components of the T3S system. Given that T3S apparatus proteins are not detectable in NYG-grown bacteria, we incubated the bacteria for 3.5 hours in the hrp-gene inducing XVM2 medium [38], [40]. Western blot analysis revealed reduced amounts of the translocon protein HrpF, the T3S-ATPase HrcN and the T3S-apparatus component HrcJ in XcvΔsX13 compared to the wild type, ΔsX13(psX13) (Figure 2A) and strain ΔsX13+sX13ch (selectively tested for HrcJ; Figure 2B). Thus, sX13 positively affects the synthesis of T3S components.
As HrpG controls the expression of the hrp-regulon [39], we analyzed whether the reduced virulence of strain ΔsX13 is due to a reduced activity of HrpG. Therefore, we ectopically expressed a constitutively active version of HrpG (HrpG*; phrpG*; [41]) in XcvΔsX13 and performed plant-infection assays. The disease symptoms induced by XcvΔsX13 and the wild type were comparable in presence of phrpG*, whereas with low inoculum of Xcv 85-10ΔsX13 the HR was slightly delayed (Figure 1D). This suggests that HrpG* suppresses the 85-10ΔsX13 phenotype. HrpF, HrcN and HrcJ protein accumulation in strain ΔsX13(phrpG*) was identical to the wild type suggesting full complementation (Figure 2A, B).
To investigate whether the reduced protein amounts of T3S-system components in XcvΔsX13 are due to altered mRNA levels, we performed qRT-PCR analyses. mRNA accumulation of hrpF, hrcJ and the type III effector genes avrBs1 and xopJ was two-fold lower in XcvΔsX13 than in the wild type and the complemented strain ΔsX13+sX13ch (Figure 2C). In addition, the mRNA amount of hrpX, but not of hrpG, was reduced in the sX13 mutant (Figure 2C). In presence of phrpG*, comparable mRNA amounts of hrpG, hrpX, hrpF, hrcJ and xopJ were detected in Xcv 85-10, ΔsX13 and ΔsX13+sX13ch, whereas the avrBs1 mRNA accumulation was significantly reduced in strain 85-10ΔsX13 (Figure 2C). Taken together, our data suggest that the reduced virulence of the 85-10ΔsX13 mutant is caused by a lower expression of components and substrates of the T3S system (Figure 1D; Figure 2A–C).
The deletion and chromosomal re-insertion of sX13 in XcvΔsX13 and ΔsX13+sX13ch, respectively, were verified by Northern blot using an sX13-specific probe (Figure S1). The sX13 abundance was not affected by expression of HrpG*, which confirms our previous findings [16] and suggests that expression of sX13 is independent of HrpG and HrpX (Figure S1).
The expression of known bacterial sRNAs depends on a variety of environmental stimuli, which often reflect the physiological functions of sRNAs [2], [46], e. g., the E. coli sRNA Spot42 is repressed in the absence of glucose and regulates carbon metabolism [47], [48]. Northern blots revealed similar sX13 amounts in bacteria incubated in NYG medium at 30°C (standard condition), in presence of H2O2, at 4°C and in NYG medium lacking nitrogen (Figure 3A). By contrast, sX13 accumulation was increased in presence of high salt (NaCl), 37°C and in MMA (Figure 3A). Hence, sX13 is differentially expressed in different growth conditions and might contribute to environmental adaptation of Xcv.
To gain an insight into the sX13 regulon we performed microarray analyses. For this, cDNA derived from Xcv strains 85-10 and ΔsX13 grown in NYG and MMA, respectively, was used as a probe. The hybridization data were evaluated using EMMA 2.8.2 [49] (see ‘Materials and Methods’). In XcvΔsX13 grown in NYG, 23 mRNAs were upregulated and 21 mRNAs were downregulated by a factor of at least 1.5 compared to the wild type (Table S2). In the MMA-grown sX13 mutant, 23 upregulated mRNAs were detected, four of which were also upregulated in NYG-grown bacteria, whereas no downregulated genes were identified (Table S2). With respect to both growth conditions, 42 and 21 genes were upregulated and downregulated, respectively, in the sX13 mutant. qRT-PCR analyses of 11 selected upregulated and four downregulated genes confirmed the microarray data (Table 1; Figure 4).
Based on the annotated genome sequence of Xcv 85-10 [32], genes upregulated in XcvΔsX13 can be grouped (Table S2): 18 genes encode proteins with unknown function, e. g., the putative LysM-domain protein XCV3927. 14 genes encode proteins involved in type IV pilus (Tfp) biogenesis, e. g., the putative Tfp assembly protein XCV2821, the pilus component PilE and the TCS response regulator PilG. Tfp enable twitching motility, i. e., adhesion to and movement on solid surfaces [50], [51]. Three genes encode proteins assigned to signal transduction, i. e., the TCS regulator AlgR, the GGDEF-domain protein XCV2041 and the chemotaxis regulator XCV2186. Moreover, hfq mRNA accumulation was two-fold increased in XcvΔsX13.
The microarray data suggested that most upregulated genes in XcvΔsX13 were only expressed in NYG- or MMA-grown bacteria (Table S2), which might be explained by the P-value and signal-strength thresholds applied for data evaluation. qRT-PCR analyses showed that the mRNA accumulation of hfq, XCV2186, pilG and XCV3927 was increased in both the NYG- and MMA-grown sX13 mutant compared to the wild type (Figure 4; Table 1). qRT-PCR analyses also revealed an upregulation of pilH in the NYG- and MMA-grown XcvΔsX13 compared to the wild type (Figure 4; Table 1). Because pilH is the second gene in the pilG operon and was not detected as expressed in the microarray data, the number of mRNAs affected by sX13 deletion might be higher than suggested by the microarray data.
Five of 21 genes downregulated in XcvΔsX13 presumably encode proteins involved in flagellum-mediated chemotaxis, e. g., the sensor kinase CheA1, the corresponding response regulator CheY and the flagellum components FliD and FliC (Table S2). qRT-PCR analyses revealed 17-fold lower fliC mRNA abundance in XcvΔsX13 grown in NYG compared to the wild type, whereas the accumulation in MMA-grown cells was identical (Figure 4; Table 1). Similarly, XCV3572, which encodes a TonB-dependent receptor, was downregulated in NYG- but not in MMA-grown XcvΔsX13 (Figure 4; Table 1). Gene XCV3573, which is encoded adjacent to XCV3572 and encodes an AraC-type regulator, was also downregulated (Figure 4; Table 1). As mentioned above, sX13 positively affected the mRNA accumulation of hrpX in XVM2 medium (see Figure 2C), which was also true for bacteria grown in NYG and MMA (Figure 4; Table 1). Since HrpX controls the expression of many type III effector genes, we analyzed xopS [52] by qRT-PCR and detected similarly decreased levels in NYG-grown XcvΔsX13 as for hrpX (Figure 4; Table 1). Taken together, our data suggest that the sX13 regulon comprises genes involved in signal transduction, motility, transcriptional and posttranscriptional regulation and virulence.
To address whether differential expression of sX13 under different conditions (see Figure 3A) affects the mRNA abundance of sX13-regulated genes, we performed qRT-PCR. We detected elevated sX13 levels in Xcv strain 85-10 cultivated in high salt conditions, at 37°C and in MMA compared to standard conditions and an increased hrpX and decreased XCV3927 mRNA accumulation (Figure 3B). In addition, low amounts of the hfq mRNA were detected in presence of high sX13 levels, whereas the abundance of the sX13-independent XCV0612 mRNA (see Table 1) was not altered (Figure 3B).
The hfq mRNA accumulation in XcvΔsX13 (Figure 3B; Figure 4; Table 1) prompted us to test whether sX13 activity depends on the RNA-binding protein Hfq. For this, we introduced a frameshift mutation into the hfq gene of Xcv strains 85-10 and 85-10ΔsX13. Northern blot analyses revealed comparable sX13 accumulation in both strains and the complemented hfq mutant, which ectopically expressed Hfq (phfq) (Figure 5A). By contrast, the accumulation of the sRNA sX14 [16] was strongly reduced in the hfq mutant; this was restored by phfq (Figure 5A). Unexpectedly, the hfq mutant strain was not altered in the induction of in planta phenotypes, i. e., in virulence (Figure 5B).
To investigate whether sX13 affects translation of putative target mRNAs, we established a GFP-based in vivo reporter system for Xcv similar to the one described for E. coli [53]. The promoterless broad-host range plasmid pFX-P permits generation of translational gfp fusions in a one-step restriction-ligation reaction (Golden Gate cloning [54]; see ‘Materials and Methods’). We cloned the promoter, 5′-UTRs, and 10 and 25 codons of XCV3927 and hfq, respectively, into pFX-P resulting in pFX3927 and pFXhfq. XCV3927 was selected because of a strongly increased mRNA accumulation in XcvΔsX13 compared to the wild type (see Table 1). In presence of pFX3927 or pFXhfq, fluorescence of XCV3927::GFP or Hfq::GFP fusion proteins was comparable in the Xcv wild type and hfq mutant (Figure 5C). The XCV3927::GFP and Hfq::GFP fluorescence was about 4-fold and 2-fold increased, respectively, in XcvΔsX13 compared to strain 85-10 (Figure 5C), suggesting that the synthesis of the fusion proteins is repressed by sX13. Interestingly, the XCV3927::GFP and Hfq::GFP fluorescence was similarly increased in XcvΔsX13 and the sX13hfq double mutant (Figure 5C). As abundance and activity of sX13 were not affected by the hfq mutation, we assume that sX13 acts Hfq-independently.
The predicted secondary structure of sX13 obtained by mfold [55] displays an unstructured 5′-region and three stable stem-loops, termed stem 1 to 3, and loop 1 to 3 (Figure 6A). Interestingly, loop 1 and loop 2 contain a ‘CCCC’ (4C) motif, whereas loop 3 harbors a ‘CCCCC’ (5C) motif (Figure 6A). To experimentally verify the predicted structure, we performed structure analyses of in vitro transcribed and radioactively-labeled sX13 treated with RNase V1 or RNase T1. While RNase T1 cleaves single-stranded RNA with a preference for G residues, RNase V1 randomly cleaves double-stranded RNA. We detected RNase T1-cleavage products for the 5′-region and RNase V1-cleavage products for stem 1 and 2, which is in good agreement with the predicted structure (Figure 6A; Figure S2). Moreover, RNase V1-cleavage products were less abundant for the 4C-motif of loop 1 and loop 2, suggesting single-stranded sequences (Figure 6A; Figure S2). The results did not allow conclusions about stem 3 and loop 3 structures.
To assess the contribution of the 4C-/5C-motifs to sX13 function, we mutated psX13 in loop 1 and 2, respectively, to ‘GCGC’, and the 5C-motif in loop 3 to ‘GCGCC’ resulting in pL1, pL2 and pL3 (Figure 6A). In addition, loop mutations were combined (pL1/2, pL1/3, pL2/3) and analyzed for their ability to complement the in planta phenotype of strain ΔsX13. As shown above, XcvΔsX13 induced a delayed HR, which was complemented by psX13 (Figure 1D). Similar phenotypes were observed with sX13 mutants carrying pL1 or psX13Δ5′, which encodes a 5′-truncated sX13 derivative lacking the terminal 14 nucleotides (Figure 6B). The HR induced by the sX13 mutant containing pL2 or pL1/2 was intermediate, whereas pL3, pL1/3 and pL2/3 failed to complement XcvΔsX13 (Figure 6B). Northern blot analyses revealed expression of all sX13-loop mutant derivatives (Figure S3). The different RNA species derived from ectopically expressed sX13 and derivatives compared to chromosomally encoded sX13 might be due to alternative transcription termination of plasmid-derived sX13 and derivatives.
As mutation of sX13 loops impinged on Xcv virulence (Figure 6B), we addressed by qRT-PCR whether loop mutations affect the mRNA abundance of XCV2821, XCV3927, hfq and pilH, which were upregulated in XcvΔsX13 (see Figure 4; Table 1). In addition, we analyzed a downregulated gene, XCV3572, and XCV0612, which was not affected by sX13 deletion. As shown in Figure 7A–E, sX13 negatively affected the mRNA abundance of XCV2821, XCV3927, hfq and pilH, whereas sX13 promoted mRNA accumulation of XCV3572. Mutation of sX13 loops differentially affected the mRNA abundance of the tested genes: pL2 and pL1/2 failed to complement XcvΔsX13 with respect to the mRNA abundance of XCV2821, XCV3927 and hfq (Figure 7A–C). Intermediate mRNA amounts of XCV3927 and hfq were detected in XcvΔsX13 carrying pL1/3 or pL2/3 compared to pB and psX13 (Figure 7B, C). Taken together, the mRNA abundance of XCV2821, XCV3927 and hfq appears to be mainly controlled by sX13-loop 2. In contrast, pilH mRNA accumulation appears to depend on multiple sX13 loops as only psX13 and pL1 complemented XcvΔsX13 (Figure 7D). The reduced mRNA amount of XCV3572 in XcvΔsX13 was complemented by pL1 and pL3 but not by pL1/3 (Figure 7E), which suggests redundant roles of sX13-loops. In presence of pL2, pL1/2 or pL2/3 in XcvΔsX13, the XCV3572 mRNA levels were intermediate compared to XcvΔsX13 carrying pB or psX13 (Figure 7E). As expected, the mRNA abundance of XCV0612 was identical in the different strains (Figure 7F).
To identify potential regulatory motifs in sX13-regulated mRNAs, a discriminative motif search was performed using DREME [56]. For this, sequences surrounding the TLSs of the 42 up- and 21 downregulated genes identified by microarray analyses (Table S2) were compared. More precisely, sequences spanning from known transcription start sites (TSSs) to 100 bp downstream of TLSs or, in case of unknown TSSs, 100 bp up- and 100 bp downstream of the TLS were inspected.
We found that up- and downregulated genes differ in the presence of ‘GGGG’ (4G) motifs. In the NYG-grown sX13 mutant, 15 out of 23 (65%) upregulated genes contain up to three 4G-motifs which are predominantly located upstream of the TLS (Figure S4A; Table S2). 70% of the genes upregulated in MMA (16 out of 23), but only 14% of the genes downregulated in NYG medium (3 out of 21) contain 4G-motifs (Figure S4A; Table S2). Thus, 4G-motifs appear to be enriched in sX13-repressed mRNAs. However, the position of the motifs and flanking nucleotides are not conserved among sX13-regulated genes. Note that the term ‘4G-motif’ also refers to motifs containing more than four G-residues in a row. The complementarity of C-rich sX13-loop sequences and G-rich mRNA motifs suggests sX13-mRNA interactions via antisense-base pairing (Figure 6A; Table 1; Table S2).
Compared to the occurrence of 4G-motifs in approximately 70% of sX13-repressed genes, only 30.71% of all chromosomally encoded Xcv genes (1,378 out of 4,487) carry 4G-motifs in proximity of their TLS (Figure S4A). Interestingly, 4G-motifs in 241 of the chromosomally encoded genes (5.37%) are located between nucleotide position 8 and 15 upstream of the TLS (Figure S4B). This position corresponds to the presumed location of the RBS and suggests a role of 4G-motifs in translation control.
To study the effect of sX13 on translation of selected putative targets, i. e., XCV3927 and hfq, we used the above-mentioned GFP-reporter plasmids pFX3927 and pFXhfq. In addition, we generated pilH::gfp (pFXpilH) and XCV0612::gfp (pFX0612) fusions (see ‘Materials and Methods’). All mRNA::gfp fusions contain a G-rich motif in the proximity of their TLS which is complementary to C-rich sX13-loop regions (see ‘Materials and Methods’). The fluorescence of the sX13 deletion mutant carrying pFX3927, pFXhfq and pFXpilH was about 3.5-, 1.6- and 2.5-fold higher, respectively, compared to the Xcv wild type (Figure 8A–C). In presence of psX13, pL1, pL3 or pL1/3 in XcvΔsX13, the XCV3927::GFP and Hfq::GFP fluorescence levels were comparable to the Xcv wild type (Figure 8A, B). By contrast, the XCV3927::GFP and Hfq::GFP fluorescence of strain ΔsX13 carrying pL2, pL1/2 or pL2/3 was similarly increased as compared to XcvΔsX13 carrying pB (Figure 8A, B). This suggests that the 4C-motif in sX13-loop 2 is required to repress XCV3927::GFP and Hfq::GFP synthesis. The increased PilH::GFP fluorescence of XcvΔsX13 was complemented by psX13 and pL1, in contrast to other sX13-loop mutant derivatives (Figure 8C). Fluorescence values of all analyzed Xcv strains carrying pFX0612 were comparable confirming sX13-independency (Figure 8D).
To address whether the G-rich motif in presumed target mRNAs is required for sX13 dependency of mRNA::gfp translation, we introduced compensatory mutations. i. e., mutated the motif to ‘GCGC’. Xcv strains carrying the resulting plasmids, pFX3927MUT, pFXhfqMUT or pFXpilHMUT, exhibited a similar fluorescence in absence and presence of sX13 and mutated sX13 derivatives (Figure 8A–C). This suggests that the G-rich motif is required for sX13-dependency of target::GFP synthesis. However, mutation of the C-rich motifs in sX13 and the G-rich motifs in mRNA::gfp fusions did not restore sX13 dependency (Figure 8A–C). Unexpectedly, the fluorescence detected for Xcv strains containing pFX3927MUT or pFXhfqMUT was comparable to the fluorescence of Xcv 85-10 carrying the non-mutated plasmids pFX3927 and pFXhfq, respectively (Figure 8A, B). The mutation of the 5G-motif in pilH abolished the fluorescence of strains containing pFXpilHMUT indicating an essential role of the motif in pilH translation (Figure 8C).
Because sX13 was more abundant in MMA- than NYG-grown bacteria (Figure 3), we also analyzed the fluorescence of MMA-grown Xcv strains containing pFX-derivatives. XCV3927::GFP and PilH::GFP synthesis in MMA was sX13-dependently repressed to a greater extent than in NYG (Figure S5; see Figure 8).
Because sX13 negatively affected both the mRNA accumulation of chromosomally encoded XCV3927, hfq and pilH genes and accumulation of the corresponding GFP-fusion proteins, we exemplarily analyzed whether this is due to an altered mRNA abundance. However, qRT-PCR analyses revealed that the XCV3927::gfp mRNA accumulation was sX13-independent suggesting that sX13 posttranscriptionally affects the synthesis of XCV3927::GFP (Figure S6).
To discriminate between transcriptional and posttranscriptional effects of sX13 on target::GFP synthesis we generated reporter fusions controlled by plac (see ‘Materials and Methods’). Note that the activity of the lac promoter is not affected by deletion of sX13 (data not shown). As shown in Figure S7, the fluorescence of XcvΔsX13 carrying pFXpl-3927 (XCV3927) and pFXpl-pilH (pilH) was 2.5- and 4-fold higher, respectively, compared to the Xcv wild type and the complemented sX13 mutant strain. Interestingly, mutation of the 4G-motif in the XCV3927 5′-UTR did not only abolish sX13-dependency but also led to a significantly reduced fluorescence compared to the Xcv wild type which carried the non-mutated reporter plasmid (Figure S7). This suggests that the 4G-motif in the XCV3927 5′-UTR promotes translation, i. e., acts as translational enhancer element. In presence of pFXpl-pilH, the fluorescence of the fusion protein was only detectable in the sX13 mutant but not in the wild type or complemented strain, confirming that PilH::GFP synthesis is repressed by sX13 (Figure S7). Overall, the data confirm that sX13 represses the synthesis of XCV3927 and PilH on the posttranscriptional level.
This study provides a first insight into the posttranscriptional modulation of clade-specific adaptive processes in a plant-pathogenic γ-proteobacterium. We identified sX13 as a major regulator of Xcv virulence in that it promotes expression of genes in the hrp-regulon, i. e., components and substrates of the T3S system (Figure 2A–C). This finding is remarkable because the hrp-regulon is only expressed under certain conditions, whereas sX13 is constitutively expressed (Figure S1) [16]. The sX13 gene is exclusively found and highly conserved in members of the Xanthomonadaceae family of Gram-negative bacteria [16]. Intriguingly, several species with an sX13 homolog lack a T3S system, e. g., the plant pathogen X. fastidiosa and the opportunistic human pathogen S. maltophilia. This suggests a role of sX13 apart from regulation of the hrp-regulon in these organisms.
The expression of the hrp-regulon depends on HrpG and HrpX [39], [40]. HrpG is presumably posttranslationally activated in the plant and in XVM2 medium and induces the expression of hrpX [38], [39], [40], [41]. As the XVM2-grown sX13 mutant displayed decreased mRNA amounts of hrpX but not of hrpG (Figure 2C), we suppose that sX13 acts upstream of HrpG. This idea is supported by the finding that constitutively active HrpG (HrpG* [41]) suppressed the sX13 mutation with respect to virulence and the expression of hrpX and downstream genes (Figure 1D; Figure 2A–C). In addition, sX13 affected the basal expression level and, hence, the activity of HrpX under non-inducing conditions, which might impact on the efficiency of hrp-gene induction during infection. Based on the fact that HrpG::GFP and HrpX::GFP synthesis was sX13-independent (Figure S8) we assume that sX13 indirectly controls the expression of the hrp-regulon.
Deletion of sX13 affected the mRNA abundance of more than 60 genes involved in signaling, motility, transcriptional and posttranscriptional regulation (Table S2). sX13 negatively regulated mRNAs involved in Tfp biogenesis but promoted the accumulation of mRNAs involved in flagellum-mediated chemotaxis in a growth-phase dependent manner (Table 1; Table S2). This, together with the fact that sX13 is differentially expressed under certain stress conditions (Figure 3), implies a central role of sX13 in the transduction of environmental signals into comprehensive cellular responses affecting virulence gene expression, motility and QS regulation. The latter is corroborated by the reduced stationary-phase cell density of the sX13 mutant compared to the Xcv wild type (Figure 1A, B) and the sX13-dependency of the XCV2041 mRNA (Table S2), which encodes a GGDEF-/EAL-domain protein. Such domains play a role in the control of cyclic-di-GMP levels and QS regulation [57]. Interestingly, XCV2041 shares 94% identity with the Xcc protein XC2226 which is a repressor of Tfp-mediated motility [58].
Another remarkable finding of our study was the sX13-dependent accumulation of the hfq mRNA. To the best of our knowledge, sX13 is the first sRNA which affects expression of this conserved RNA-binding protein (Table 1). The Xcv hfq mutant was unaltered in virulence on its host plant (Figure 5B), which is in good agreement with recent findings for Xoo [18]. By contrast, Hfq contributes to virulence in a number of other bacteria, including the plant pathogen A. tumefaciens, and is also involved in symbiotic plant interactions of Sinorhizobium meliloti [5], [12], [59], [60]. In Vibrio cholerae, four redundantly acting and Hfq-dependent sRNAs (Qrr) destabilize hapR mRNA, which encodes the master regulator of QS, the T3S system and other virulence genes [61], [62]. In the Gram-positive human pathogen Staphylococcus aureus, the Hfq-independent RNAIII is induced by the agr QS system and mediates the switch between the expression of surface proteins and excreted toxins through translational repression of Rot (repressor of toxins) [63], [64], [65].
Xcv sRNAs are strongly structured and lack extended single-stranded regions [15], [16], [32]. In contrast, enterobacterial sRNAs commonly exhibit a modular structure consisting of a single-stranded mRNA-targeting domain, often located at the 5′-end, an A/U-rich Hfq-binding site and a Rho-independent terminator [1]. The sX13 structure suggests that direct sRNA-mRNA interactions are energetically confined to the unstructured 5′-region and its three C-rich loops (Figure 6A). However, the 5′-region of sX13 was dispensable for full virulence of Xcv and sX13 activity appears to be exerted via loops 2 and 3 (Figure 6B). Although loops 1 and 2 just differ in the 3′-adjacent nucleotide (U/A) (Figure 6A), only loop 2 was required to repress the synthesis of XCV3927::GFP and Hfq::GFP, which might depend on the position of stem-loops in the sRNA and, thus, accessibility. By contrast, repression of PilH::GFP appears to depend on multiple sX13 regions (Figure 8C).
An important question is whether sX13 controls target gene expression on the level of mRNA stability or translation. On one hand, sX13-loop mutant derivatives affected the mRNA levels of presumed targets (Figure 7). On the other hand, protein levels, but not the mRNA level of an XCV3927::gfp fusion, harboring only the 5′-UTR and 10 codons of XCV3927, were sX13-dependent (Figure 8A; Figure S6). This suggests that the impact of sX13 on XCV3927 mRNA abundance and translation are separate events and hints at the presence of additional regulatory sites in the XCV3927 mRNA. It should be noted that the assessment of RNA stability by rifampicin treatment is hampered by the fact that our Xcv strains are rifampicin resistant.
The sX13 loops are reminiscent of regulatory RNAs in S. aureus, many of which contain ‘UCCC’-motifs in loops [66]. For example, RNAIII contains C-rich stem-loops, which interact with the RBS of target mRNAs [63], [65], [67]. RNAIII represses Rot synthesis through formation of kissing complexes between two stem-loops of each RNAIII and rot mRNA [64], [65]. Such multiple loop interactions are also employed by the E. coli sRNA OxyS to target fhlA [68]. In Helicobacter pylori, the sRNA HPnc5490 represses the synthesis of the chemotaxis regulator TlpB [69]. Interestingly, the central part of the HPnc5490-loop sequence is identical to the ‘UCCCCCU’-motif of loop 3 in sX13 [69].
Similarly to RNAIII targets in S. aureus and the tlpB mRNA in H. pylori [65], [69], mRNAs repressed by sX13 are enriched for G-rich motifs in proximity of the TLS (Figure S4; Table S2). The complementarity between these motifs and the 4C-/5C-motif in the sX13 loops suggests sX13-mRNA interactions through antisense base pairing. Our data emphasize that sX13 acts posttranscriptionally on target genes that contain G-rich motifs, as shown for XCV3927 and pilH (Figure S7). This idea is supported by the fact that mutation of the G-rich motifs, located in the mRNA of XCV3927, pilH and hfq, abolished sX13-dependency of protein synthesis (Figure 8; Figure S5; Figure S7). However, the presence of a G-rich motif does not necessarily confer regulation by sX13 (see XCV0612; Figure 7F; Figure 8D). Given that eight of 28 repressed and 4G-motif-containing mRNAs contain at least two 4G-motifs close to the TLSs (Figure S4), we assume that sX13 loops can interact with multiple 4G-motifs in certain mRNAs. As positively regulated mRNAs lack G-rich motifs, sX13 presumably acts indirectly on the corresponding genes (Figure S4; Table S2).
Direct sRNA-mRNA interactions are commonly validated by compensatory mutant studies [1]. However, in case of the E. coli sRNA RyhB, mutations were suggested to interfere with Hfq-binding and rendered compensatory mutants non-functional [70]. Here, mutation and deletion of sX13 increased the synthesis of XCV3927::GFP and Hfq::GFP fusions, whereas mutation of corresponding 4G-motifs resulted in similar fluorescence values as non-mutated mRNA::gfp fusions in Xcv wild type. In addition, the reduced fluorescence of mutated target::GFP fusions was unaffected by compensatory sX13-mutant derivatives (Figure 8). This suggests that G-rich motifs in sX13-repressed mRNAs play a role besides mediation of sRNA interactions. While Xanthomonas spp., like other G+C-rich bacteria, lack a consensus RBS [16], [71], 5% of the chromosomal Xcv coding sequences (241 of 4,487) contain a G-rich motif 8–15 nucleotides upstream of their TLS (Figure S4). As anticipated, mutation of the 5G-motif at the RBS position of pilH abolished translation (Figure 8; Figure S5; Figure S7). By contrast, the 4G-motifs in XCV3927 and hfq mRNAs, located 21 nucleotides upstream and nine nucleotides downstream of the AUG, respectively, confer sX13-dependency but were not essential for translation (Figure 8). Thus, G-rich motifs confer sX13-dependency and mRNA translation in a position-dependent manner. As mutation of the 4G-motif in XCV3927 reduced protein synthesis, the motif appears to function as translational enhancer (Figure S7). We suggest that sequestration of a G-rich motif by sX13 as well as mutation of the motif precludes the binding of an unknown factor, which promotes mRNA translation. Such a factor could be RNA, protein or the ribosome.
The presumed sX13 mode of action is reminiscent of the Salmonella sRNA GcvB, which inhibits translation of mRNAs by targeting C/A-rich enhancer elements [72], [73]. By increasing the ribosome-binding affinity, C/A-rich motifs enhance mRNA translation, irrespective of their localization upstream or downstream of the TLS [72], [74].
Homologs of Xcv sRNAs are predominantly found in members of the Xanthomonadaceae family but not in other bacteria [15], [16]. The sX13 gene is located adjacent to the DNA polymerase I-encoding polA gene, which resembles a locus encoding the Spot42 sRNA in E. coli and members of the αr7 sRNA family in α-proteobacteria [75], [76], [77], [78]. In contrast to sX13, Spot42 requires Hfq and regulates targets involved in carbon metabolism [48], [79], e. g., the discoordinated expression of genes within the gal galactose utilization operon [47], which is absent in Xcv. Although sX13 lacks sequence similarity to Spot42 and αr7 sRNAs, the latter also contain three stem-loops and apical C-rich motifs [80] suggesting that sRNAs in distantly related bacteria evolved divergently but retained structural conservation. Thus, it will be interesting to see whether the polA locus in other bacteria also encodes sRNAs, and whether sX13 and structurally related sRNAs act in a similar manner.
For bacterial strains, plasmids and oligonucleotides used in this study see Table S1. E. coli strains were grown at 37°C in lysogeny broth (LB), Xcv strains at 30°C in nutrient-yeast-glycerol (NYG) [81], XVM2 [40] or minimal medium A (MMA) [82], which was supplemented with casamino acids (0.3%) and sucrose (10 mM). Plasmids were introduced into E. coli by chemical transformation and into Xcv by tri-parental conjugation, using pRK2013 as helper plasmid [83]. Antibiotics were added to the final concentrations: ampicillin, 100 µg/ml; gentamycin, 15 µg/ml; kanamycin, 25 µg/ml; rifampicin, 100 µg/ml; spectinomycin, 100 µg/ml.
To generate the sRNA-expression vector pBRS, a 28-bp fragment between the lac promoter and the EcoRI cloning site in pBBR1mod1 [84] was replaced by a truncated fragment, amplified by PCR from pBBR1mod1 using primers pBRS-EcoRI-fw and pBRS-NcoI-rev. For generation of constructs expressing sX13 (psX13) and sX13Δ5′ (psX13Δ5′; lacking 14 nt at the 5′-end), respective fragments were PCR-amplified from Xcv 85-10 using primers sX13-fw/-rev or sX13d5-fw/-rev. PCR products were digested with EcoRI/HindIII and cloned into pBRS. The sX13-mutant plasmids pL1, pL2, pL3 and pL2/3 were generated by PCR amplification of plasmid psX13 using primers L1-fw/-rev, L2-fw/-rev, L3-fw/-rev and L3-fw/L2/3-rev, respectively; plasmid pL1/3 was generated with primers L3-fw/-rev and pL1 as template; pL1/2 was generated with primers L2-fw/L1/2-rev and pL2 as template. For ectopic expression of hfq under control of its own promoter, a PCR product obtained from Xcv 85-10 using primers pMphfq-fw/-rev was cloned into the promoterless vector pBRM-P [84].
The GFP-based promoter-less reporter plasmid pFX-P permits BsaI-mediated cloning of PCR amplicons (Golden Gate cloning) in a one-step restriction-ligation reaction [54] and was generated as follows: pDSK602 [85] was digested with PstI/BamHI to remove the lac promoter and multiple-cloning site. To allow blue-white selection, a dummy module containing 5′- and 3′-BsaI recognition sites, plac and lacZ was PCR-amplified from pBRM-P [84] using primers pFX-lz-fw/-rev. A fragment containing both the gfp coding sequence without translation start codon and the rrnB terminator was PCR-amplified from pXG-1 [53] using primers pFXgfp-fw/-rev. After blunt-end ligation of dummy- and gfp-module, the fragment was digested with Mph1103I/BglII and ligated into the PstI/BamHI sites of the pDSK602 backbone, resulting in pFX-P. For generation of the GFP-control plasmid pFX0, a promoterless fragment (138 bp) of the sRNA gene sX6 [16] was PCR-amplified from Xcv 85-10 using primers pFX0-fw/-rev and cloned into pFX-P.
To generate mRNA::gfp expression constructs, fragments containing the promoter, 5′-UTR and 10 to 25 codons of the respective genes were PCR-amplified from Xcv 85-10 using corresponding pFX-fw/-rev primers (Table S1) and cloned into pFX-P. Plasmids pFX3927, pFXhfq and pFX0612 were generated by cloning of nucleotide sequences −98 to +30, −160 to +75 and −116 to +33 relative to the translation start codon of XCV3927, hfq and XCV0612, respectively. pFXpilH was constructed by cloning a fragment spanning nucleotides −99 upstream of the pilG translation start codon to nucleotide +60 within the coding sequence of pilH.
The 4G-motif in XCV3927::gfp and hfq::gfp is located 21–24 bp upstream and 9–12 bp downstream of the ATG, respectively. pilH::gfp and XCV0612::gfp contain a 5G-motif at nucleotide positions 10–14 and 8–12 upstream of the ATG, respectively. Plasmids pFXMUT were constructed as follows: to mutate the ‘GGGG’ motif to ‘GCGC’, sequences upstream and downstream of the motif were PCR-amplified from Xcv 85-10 using primers pFX-fw/pFX-mut-L-rev and pFX-mut-R-fw/pFX-rev, respectively. Primers pFX-mut-L-rev and pFX-mut-R-fw contain the mutation flanked by a BsaI-recognition site. pFX and corresponding pFXMUT derivatives only differ in the sequence of the G-rich motif at nucleotide positions relative to the translation start codons: −24/−22 in pFX3927MUT, +10/+12 in pFXhfqMUT and −12/−10 in pFXpilHMUT.
Plasmids pFXpl, which express plac-driven mRNA::gfp fusions, were constructed by cloning respective fragments into pFX-P: plac was PCR-amplified from pBRM-P [84] using primers plac-fw/rev; sequences −59 to +54 and −147 to +54 relative to the ATG of hrpX and hrpG, respectively, were PCR-amplified from Xcv 85-10 using primers pFXpl-hrpX-fw/-rev and pFXpl-hrpG-fw/-rev; fragments of XCV3927 and pilH were PCR-amplified from respective pFX and pFXMUT plasmids using primers pFXpl3927-fw/pFXpl3927mut-fw/pFX3927-rev and pFXplpilH-fw/pFXpilH-rev.
To generate a chromosomal sX13 deletion mutant, flanking sequences of ∼650 bp up- and downstream of sX13 [16] were amplified by PCR from Xcv 85-10 using primers d13L-fw/-rev and d13R-fw/-rev. PCR products were digested with BamHI/HindIII and HindIII/XbaI, respectively, and ligated into the suicide vector pOK1 [86]. XcvΔsX13+sX13ch, which carries an sX13 copy at the ΔsX13 locus, was created as follows: two PCR fragments amplified from Xcv 85-10 using primers int13L-fw/rev and int13R-fw/rev were digested with Psp1406I, ligated and cloned into the BamHI/XbaI site of pOK1. Conjugation of pOKΔsX13 into Xcv 85-10 and pOKint13 into XcvΔsX13 and selection of the correct double crossing-over events were performed as described [86]. XcvΔsX13+sX13ch was identified by PCR amplification of the sX13 locus and Psp1406I restriction.
To introduce a frameshift mutation into chromosomal hfq, PCR products obtained from Xcv 85-10 using primers hfqL-fw/-rev and hfqR-fw/-rev were digested with BamHI/BsaI and BsaI/XbaI, respectively, and cloned into pOK1. Conjugation of pOK-fshfq into Xcv and selection of double crossing-over events were performed as described [86]. The resulting hfq mutant strain carries a 4 bp deletion in an MnlI site (nucleotides 33–36 in hfq) and was identified by PCR using primers seqhfq-fw/-rev followed by digestion with MnlI.
Pepper (Capsicum annuum) plants of the near-isogenic cultivars ECW and ECW-10R [87] were grown at 25°C with 60–70% relative humidity and 16 hours light. For infection assays, Xcv bacteria were resuspended in 10 mM MgCl2 and inoculated with a needleless syringe into the intercellular spaces of leaves using concentrations of 1–4×108 colony-forming units (CFU) per ml for scoring plant reactions and 104 CFU/ml for in planta growth curves. For better visualization of the HR, leaves were bleached in 70% ethanol. In planta growth curves were performed as described [33]. All experiments were repeated at least two times.
Xcv cells grown overnight in NYG medium were washed, incoculated at OD600 = 0.2 into XVM2 medium and incubated for 3.5 hours at 30°C. Total cell extracts were analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis and immunoblotting using specific polyclonal antibodies directed against HrpF [88], HrcN [89], HrcJ [89] and GroEL (Stressgen). A horseradish peroxidase-labeled anti-rabbit antibody (Amersham Pharmacia Biotech) was used as secondary antibody. Antibody reactions were visualized by enhanced chemiluminescence (Amersham Pharmacia Biotech).
To determine GFP fluorescence, bacteria were adjusted to OD600 = 1.0 in 10 mM MgCl2. Fluorescence was measured at 485-nm excitation and 535-nm emission using a microplate reader (SpectraFluor Plus; Tecan Trading AG).
sX13 [16] was PCR-amplified from Xcv 85-10 using primers sX13T7-fw, containing the T7-promoter, and sX13T7-rev and cloned into pUC57 (Thermo Fisher Scientific), resulting in pUC-13T7. Template DNA for in vitro transcription was amplified from pUC-13T7 using primers sX13-ITC-fw/-rev. sX13 transcription and DNase treatment were performed according to manufacturer's instructions (MEGAscript®Kit; Invitrogen). RNA labeling using [γ-32P]-ATP, treatment with RNase T1 (1 Pharmacia unit; Ambion) or RNase V1 (0.01 to 0.0002 Pharmacia units; Ambion) and generation of nucleotide ladders were performed as described [90]. Samples were analyzed on 12% polyacrylamide gels containing 7 M urea. Signals were visualized with a phosphoimager (FLA-3000 Series; Fuji).
Bacteria were grown overnight in NYG and inoculated at OD600 = 0.2 into NYG or XVM2 medium. XVM2 cultures were incubated for 3.5 hours at 30°C. NYG-grown cells were harvested at exponential growth phase (OD600 = 0.5–0.7) or used to inoculate the following media at OD600 = 0.5: NYG containing 0.3 M NaCl, 0.2 M H2O2 or NYG lacking a nitrogen source, MMA or MMA lacking a carbon source followed by incubation for 3 hours.
RNA isolation and Northern blot hybridization was performed as described [16], [91]. Oligonucleotide probes for detection of sX13 and 5S rRNA are described in [16].
For qRT-PCR analyses, cDNA was synthesized using RevertAid H Minus First Strand cDNA-Synthesis Kit according to manufacturer's instructions (Fermentas). qRT-PCR was performed using 2 ng cDNA and ABsolute BlueSYBR Green Fluorescein (Thermo Scientific) and analyzed on MyiQ2 and CFX Connect systems (Bio-Rad). The efficiency and specificity of PCR amplifications was determined by standard curves derived from a dilution series of template cDNA and melting curve analysis, respectively. Mean transcript levels were calculated based on values obtained from technical duplicates of at least three independent biological replicates and the levels of 16S rRNA (reference gene) as described (ABI user bulletin 2; Applied Biosystems).
For isolation of total RNA, NYG-grown cells were harvested at exponential growth phase (OD600 = 0.5–0.7) or used to inoculate MMA at OD600 = 0.5 followed by incubation for 3 hours. Fluorescently labeled cDNA was prepared as described [92]. Starting from 10–15 µg total RNA, aminoallyl-modified first strand cDNA was synthesized by reverse transcription using random hexamer primers, Bioscript reverse transcriptase (Bioline) and 0.5 mM dNTP, dTTP∶aminoallyl-dUTP (1∶4). After hydrolysis and clean up using Nucleotide removal kit (Qiagen), Cy3- and Cy5-N-Hydroxysuccinimidyl ester dyes (GE Healthcare) were coupled to the aminoallyl-labeled first strand cDNA. Uncoupled dye was removed using the Nucleotide removal kit (Qiagen). For RNA from NYG- and MMA-grown bacteria, four and three microarray hybridizations were performed, respectively, using labeled cDNA obtained from independent bacterial cultures.
The genome-wide microarray for Xcv strain 85-10 (Xcv5KOLI) carried 50–70 nt unique oligonucleotides representing CDSs, with each oligonucleotide spotted in three technical replicates per microarray [93]. Preprocessing of microarrays was performed as described [94]. Hybridization was performed in EasyHyb hybridization solution (Roche) supplemented with sonicated salmon sperm DNA at 50 µg/ml in a final volume of 130 µl for 90 min at 45°C using the HS400 Pro hybridization station (Tecan Trading AG). Labeled cDNA samples were denatured for 5 min at 65°C prior hybridization. After hybridization microarrays were washed as described [94].
Mean signal and mean local background intensities were obtained for each spot on the microarray images using ImaGene 8.0 software for spot detection, image segmentation and signal quantification (Biodiscovery Inc.). Spots were flagged as empty if R≤0.5 in both channels, where R = (signal mean−background mean)/background standard deviation. Remaining spots were analyzed further. The log2 value of the ratio of intensities was calculated for each spot using the formula Mi = log2(Ri/Gi). Ri = Ich1(i)-Bgch1(i) and Gi = Ich2(i)-Bgch2(i), where Ich1(i) or Ich2(i) is the intensity of a spot in channel 1 or channel 2, and Bgch1(i) or Bgch2(i) is the background intensity of a spot in channel 1 or channel 2. The mean intensity was calculated for each spot, Ai = log2(RiGi)0.5 [95]. For data normalization (Median), significance test (Holm) and t-statistics analysis, the EMMA 2.8.2 open source platform was used [49]. Genes were accounted as differentially expressed if P adjusted ≤0.05, A≥8, and if the ratio showed at least a 1.5-fold difference between the two experimental conditions.
Homology searches were performed using BLASTn and the NCBI genome database (http://blast.ncbi.nlm.nih.gov; http://www.ncbi.nlm.nih.gov/genome; date: 11/22/2012).
The secondary structure of sX13 [16] was predicted using Mfold version 3.5 (http://mfold.rna.albany.edu/?q=mfold/RNA-Folding-Form) and default folding parameters [55]. To identify putative regulatory motifs in the 5′-regions of sX13-regulated mRNAs, a discriminative motif search was performed using DREME version 4.9.0 (http://meme.nbcr.net/meme/cgi-bin/dreme.cgi) [56]. Sequences of regulated genes comprising nucleotide positions −100 to +100 relative to translation start codons or in case of known TSSs [16] (see Table S2), sequences comprising the 5′-UTR to position +100 downstream of translation start codons, were extracted from the genome of Xcv strain 85-10 (NC_007508 and NC_007507) [32]. DREME motif search was performed with negatively regulated genes as input and positively regulated genes as comparative sequences and an E-value of ≤5.
YP_363045.1; YP_363046.1; YP_362142.1; YP_362163.1; YP_362160.1; YP_361663.1; YP_363499.1; YP_365931.1; YP_363887.1; YP_365930.1; YP_365658.1; YP_364552.1; YP_363772.1; YP_363917.1; YP_364964.1; YP_364963.1; YP_364545.1; YP_365303.1; YP_365304.1; YP_362343.1; YP_362302.1; YP_362409.1; YP_364550.1; YP_364798.1; YP_364827.1; YP_365231.1; YP_363688.1; YP_363753.1; YP_361904.1; YP_363264.1; YP_362055.1; YP_363957.1
|
10.1371/journal.pntd.0007406 | Provision of deworming intervention to pregnant women by antenatal services in countries endemic for soil-transmitted helminthiasis | The World Health Organization has recently reemphasized the importance of providing preventive chemotherapy to women of reproductive age in countries endemic for soil-transmitted helminthiasis as they are at heightened risk of associated morbidity. The Demographic and Health Surveys (DHS) Program is responsible for collecting and disseminating accurate, nationally representative data on health and population in developing countries. Our study aims to estimate the number of pregnant women at risk of soil-transmitted helminthiasis that self-reported deworming by antenatal services in endemic countries that conducted Demographic and Health Surveys.
The number of pregnant women living in endemic countries was extrapolated from the United Nations World Population Prospects 2015. National deworming coverage among pregnant women were extracted from Demographic and Health Surveys and applied to total numbers of pregnant women in the country.
Sub-national DHS with data on self-reported deworming were available from 49 of the 102 endemic countries. In some regions more than 73% of STH endemic countries had a DHS. The DHS report an average deworming coverage of 23% (CI 19–28), ranging from 2% (CI 1–3) to 35% (CI 29–40) in the different regions, meaning more than 16 million pregnant women were dewormed in countries surveyed by DHS. The deworming rates amongst the 43 million pregnant women in STH endemic countries not surveyed by DHS remains unknown.
These estimates will serve to establish baseline numbers of deworming coverage among pregnant women, monitor progress, and urge endemic countries to continue working toward reducing the burden of soil-transmitted helminthiasis. The DHS program should be extended to STH-endemic countries currently not covering the topic of deworming during pregnancy.
| Soil-transmitted helminths are intestinal worms that cause significant suffering among the poorest communities in the world. They are transmitted via contaminated water, food or soil, all of which result from poor sanitation. Children and women of reproductive age are at heightened risk of related morbidities such as malnutrition, cognitive impairment and anaemia. Pregnant women are particularly susceptible to severe maternal and neonatal complications. Deworming drugs are cheap, safe, and effective in reducing morbidity related to soil-transmitted helminthiasis. Large scale drug administration campaigns have distributed donated medicines to children in endemic countries, but women of reproductive age are currently not well covered. Yet, demographic surveys show that they are being treated for soil-transmitted helminthiasis through health care services. This study provides estimates for the number of pregnant women at risk of soil-transmitted helminthiasis being dewormed by antenatal services in endemic countries conducting Demographic Health Surveys. These estimates mark the preliminary reference point for deworming coverage among pregnant women in endemic countries, and will thus prove useful for tracking overall progress in the ongoing effort to eliminate neglected tropical diseases.
| Soil-transmitted helminthiasis (STH) are caused by the intestinal worms Ascaris lumbricoides (roundworm,) Trichuris trichiura (whipworm), Necator americanus and Ancylostoma duodenale (hookworm). STH are transmitted when individuals come in contact with environment contaminated by faeces containing nematode eggs. The parasites’ eggs or larvae enter the human body via ingestion of infested food, the larvae by skin penetration[1,2].
More than two billion people [1] in 102 endemic countries [3] were considered to be infected with STH in 2015, causing a loss of 39 million of disability-adjusted life years. Most of the disease burden occurs in the tropical and sub-tropical areas of sub-Saharan Africa and South East Asia [4]. STH endemicity is a result of poor sanitary infrastructure and lack of understanding regarding the importance of safe disposal of faeces characteristic of the most impoverished communities of the world [5].
The World Health Organization (WHO) recommends administration of albendazole (400 mg) or mebendazole (500 mg) coupled with hygiene education to reduce STH-related morbidity among high risk populations. Annual PC is recommended in areas where STH prevalence is between 20% and 50%, and biannual PC in areas where STH prevalence exceeds 50%[1]. Completely eliminating STH requires long-term commitment and the allocation of extensive resources toward improving water and sanitation[6].
WHO recognizes three population groups at highest risk of STH-related morbidity: preschool-age children (pre-SAC), school-age children (SAC), and women of reproductive age (WRA)[1]. WRA are especially vulnerable as they are at elevated risk of certain comorbidities such as anaemia (exacerbated particularly by hookworm and whipworm infections)[7]. Drug donations are presently available for the regular PC of pre-SAC and SAC in STH-endemic countries. These countries consistently report the use of donated drugs among risk groups to WHO[8]. In 2015, the implementation of PC averted over 44% of lost disability-adjusted life years due to STH in pre-SAC and SAC [9].
WHO has recently reiterated the importance of WRA as a target population for the scaling up of PC in the Bellagio Declaration[7] as STH among pregnant women can produce or seriously aggravate maternal and neonatal complications[4]. Yet, anthelminthic drug donations are currently not available for WRA and countries are not reporting coverage, making the estimation of PC provided to this group difficult to assess.
In this study, we quantified the number of pregnant women being dewormed by health services in countries considered endemic for STH (>20% prevalence) [3,8] that conducted recent Demographic and Health Surveys (DHS). The estimates offered by this study will provide useful information for the scaling up of PC programmes among WRA, however it is important that control programme managers collect further coverage and context specific epidemiological data to improve these estimates.
The United Nations World Population Prospects 2015 consists of population estimates and projections executed by the Population Division of the Department of Economic and Social Affairs of the United Nations Secretariat. It includes results from national population censuses and demographic and health surveys, outlining key indicators by region, sub-region, country, and development group. The United Nations World Population Prospects 2015 was used to estimate the total number of pregnant women in STH-endemic countries[10].
DHS are standardized nationally-representative household surveys implemented over a span of 18–20 months about every five years. The surveys are implemented by ICF International and funded by the United States Agency for International Development (USAID) with contributions from other donors such as UNICEF, UNFPA, WHO, and UNAIDS. The surveys consist of four types of questionnaires (Household, Woman’s, Man’s, and Biomarker) that collect data on demographic, environmental, socio-economic, and health-related characteristics. One of the aim of DHS is to evaluate the performances of health services; in this context DHS asked to women in the sample with a live birth in the past five years, if the “took intestinal parasite drug” during the pregnancy of their last birth [11]. All DHS survey data were extracted and analysed using R statistical program and the rDHS and survey packages. Permission for sub-national DHS data between 2000 and 2017 in STH-endemic countries was first sought, downloaded through the DHS API, and national means and confidence intervals of coverages of deworming among pregnant women by antenatal services summarised using sample weights.
The number of live births obtained from the United Nations World Population Prospects 2015 was assumed to be equal to the total number of pregnant women in each country. The 2015 report was used, as opposed to a more recent version, to maintain consistency with previously estimated deworming coverage among not pregnant WRA.
Deworming coverage measured by DHS were assumed to be representative of the national situation. For DHS reports reporting deworming, the proportion of women dewormed during their last pregnancy was applied to number of pregnant women in each country. Total pregnant women and pregnant women dewormed in each of the endemic countries were aggregated by WHO region and globally.
Out of 102 STH-endemic countries, DHS reports for 49 countries were available, as depicted in Fig 1. 53 countries did not have DHS reports or were excluded from the analysis as they were based on surveys conducted before 2000. Overall, DHS was available for 48% of all endemic countries (Table 1). Out of the 1.24 million household interviews conducted, 85% were in the South East Asian (636,551) and African (425,650) regions.
45 out of 49 DHS Reports were based on surveys conducted after 2010, with the majority of surveys being from 2011–2015. Only 6 DHS reports were based on surveys conducted before 2011. The proportion of DHS reports which include the percentage of pregnant WRA dewormed by antenatal services increased from 0% between 2000 and 2005 to 92% after 2016 (see S1 Fig).
Among the 49 countries investigated with DHS, it was estimated that out of 68 million pregnant women at risk of STH, over 16 million received deworming medication during their last pregnancy (Table 2). The kind of anthelminthic drug given was not specified in the DHS reports.
The average coverage of deworming was estimated at 23% (CI 19–28) among the countries surveyed, and varied between 0.6% (CI 0.3–1.2) and 83% (CI 75–89).
Overall, the 49 countries with DHS reports accounted for 62% of total pregnant women in all STH-endemic countries. DHS represented deworming coverage during pregnancy in the African region the best, with 31 out of 42 STH endemic countries surveyed. In contrast, the Western Pacific Region was least represented by DHS, with only 2 out 15 STH endemic countries surveyed for deworming coverage during pregnancy. There was no DHS report available for China, which accounts for 77.8% of pregnant women at risk of STH in the Western Pacific region and 14.6% of pregnant WRA in all endemic countries. Consequently, the Western Pacific region had the lowest percentage of pregnant women represented by available DHS (Table 2).
Table 2 describes the regional distribution of total pregnant women at risk of STH as well as the estimated number of pregnant women dewormed in all STH-endemic countries.
In this study we estimate that in 49 STH endemic countries where DHS was recently conducted over 16 million have received deworming. More than 1.2 million households were interviewed and we consider the estimated deworming coverage representative of the performances of health services in those countries.
At the moment countries are not requested to report coverage data on WRA because this risk group is not among the ones targeted by the NTD road map. Our data shows an overall positive trend in the number of STH-endemic countries reporting deworming of pregnant women in the absence of global monitoring, the initial spike in this trend coinciding with the first Global Partner’s Meeting on Neglected Tropical Diseases organized in 2007 [12]. Deworming initiatives for pregnant women are likely to continue growing as countries scale up their STH control programs. These estimates constitute a valuable step toward the coverage of this group at risk as they represent baseline figures of deworming coverage among pregnant women in multiple endemic countries.
Although we do not have access to more recent DHS information on countries that conducted surveys between 2000 and 2005, it is probable that those countries have since initiated deworming activity for pregnant women. Additionally, we recognize that countries that do not report deworming as a component of antenatal services may be providing anthelminthic drugs to pregnant women through other platforms, or women are procuring deworming tablets in local pharmacies or markets. The coverage of deworming among pregnant women in this study could therefore be underestimated.
It is important to note that we found DHS for 73% of the STH-endemic countries in Africa, and for 50% of the STH-endemic countries in South East Asia, both accounting for the vast majority of pregnant women at risk of STH worldwide. As these two WHO regions comprise the highest concentration of STH, we consider that the data collected is particularly representative of those countries most in need of targeted PC. There was, however, no data on deworming during pregnancy from 53 STH endemic countries, meaning an evidence gap still remains in these countries and should be addressed, particularly in the American and Western Pacific regions, to better understand global deworming coverage in pregnant women. In addition to expanding the geographic scope of DHS, the questionnaire could be adapted in future surveys to elicit better quality data on deworming in WRA by, for example, by limiting the question on use of deworming to the latest pregnancy.
A limitation of this study is the assumption that the number of pregnant women in a country is equal to the number of live births, as it omits all pregnancies that resulted in abortions, miscarriages, and stillbirths. On the other hand, multiple pregnancies (i.e. those giving birth to twins) may lead to overestimation of total number of pregnant women, granted an overall underestimation is more likely.
In our opinion, antenatal services represent a relevant potential channel for systemic PC delivery considering that countries such as Congo have achieved 85.8% deworming coverage of pregnant women solely through antenatal services.
Targeted PC can significantly reduce the burden of STH among WRA, improving maternal and child health outcomes and spurring productivity[13]. Considering the prospective impact of STH elimination and the relatively low implementation costs, PC among WRA should be at the forefront of public health operations in all endemic countries. We urge countries that have already integrated deworming into health programming to continue scaling up the provision of anthelminthic treatment among WRA. The estimates provided by this study will aid strategic efforts by informing the expansion of PC among WRA in endemic countries and promoting the definitive goal of eliminating STH globally.
|
10.1371/journal.pgen.1007679 | Mechanisms of acquired resistance to rapalogs in metastatic renal cell carcinoma | The mechanistic target of rapamycin (mTOR) is an established therapeutic target in renal cell carcinoma (RCC). Mechanisms of secondary resistance to rapalog therapy in RCC have not been studied previously. We identified six patients with metastatic RCC who initially responded to mTOR inhibitor therapy and then progressed, and had pre-treatment and post-treatment tumor samples available for analysis. We performed deep whole exome sequencing on the paired tumor samples and a blood sample. Sequence data was analyzed using Mutect, CapSeg, Absolute, and Phylogic to identify mutations, copy number changes, and their changes over time. We also performed in vitro functional assays on PBRM1 in RCC cell lines. Five patients had clear cell and one had chromophobe RCC. 434 somatic mutations in 416 genes were identified in the 12 tumor samples. 201 (46%) of mutations were clonal in both samples while 129 (30%) were acquired in the post-treatment samples. Tumor heterogeneity or sampling issues are likely to account for some mutations that were acquired in the post-treatment samples. Three samples had mutations in TSC1; one in PTEN; and none in MTOR. PBRM1 was the only gene in which mutations were acquired in more than one post-treatment sample. We examined the effect of PBRM1 loss in multiple RCC cell lines, and could not identify any effect on rapalog sensitivity in in vitro culture assays. We conclude that mTOR pathway gene mutations did not contribute to rapalog resistance development in these six patients with advanced RCC. Furthermore, mechanisms of resistance to rapalogs in RCC remain unclear and our results suggest that PBRM1 loss may contribute to sensitivity through complex transcriptional effects.
| Mammalian target of rapamycin (mTOR) inhibitors, everolimus and temsirolimus, are FDA-approved for treatment of metastatic renal cell carcinoma (mRCC), but molecular mechanisms of acquired or secondary resistance to these agents are unknown. We evaluated six mRCC patients with available pre-treatment specimens who were treated with mTOR inhibitors and had a good clinical outcome, and then had a second biopsy at the time of resistance. We found that mutations in PBRM1 appeared to be enriched in post-treatment samples. However, modulation of PBRM1 levels in vitro in cell lines had no apparent effect on rapalog sensitivity. We conclude that mechanisms of resistance to rapalog therapy in RCC are not easily explained by gene mutations in most cases, and may depend on more subtle transcriptional and/or epigenetic changes.
| Both everolimus and temsirolimus, analogs of rapamycin termed rapalogs, are FDA-approved and in common used for treatment of metastatic RCC based on seminal randomized clinical trials [1–3]. However, these drugs are known to cause disease stabilization in most cases, with a 5% objective response rate by standard RECIST criteria.
The Phosphatidylinositol 3-kinase (PI3K)/AKT/mechanistic target of rapamycin (mTOR] pathway plays a critical role in cell growth, differentiation, survival and metabolism. It is frequently activated in a variety of cancer types [4], and new uses of rapalogs in combination with other therapies continue to be discovered [5,6].
mTOR is a serine threonine kinase which occurs in cells in two large multi-component complexes termed mTORC1 and mTORC2 [7,8]. mTORC1 is negatively regulated by the TSC protein complex which consists of TSC1, TSC2, and TBC1D7, which converts the small GTPase RHEB into its inactive GDP-bound form. When both alleles of either TSC1 or TSC2 are mutated or lost, as is the rule in tumors occurring in individuals with the genetic disorder tuberous sclerosis complex, RHEB-GTP levels are high, leading to activation of mTORC1. mTORC1 activity is also regulated by PI3K, AKT, MAPK, AMPK, growth factors, nutrient availability, stress levels and oxygen levels. Activation of mTORC1 leads to protein synthesis, lipid synthesis, nucleotide synthesis, autophagy inhibition, leading to cell enlargement and preparation for cell division [9]. Somatic mutations in MTOR which deregulate and activate its kinase [10,11] are known to occur in several cancer types, predominantly RCC in which mutation is seen in about 5% [12]. Activating RHEB mutations which activate mTORC1 are quite rare but also known to occur in cancer [13].
Rapalogs are allosteric inhibitors of mTORC1 through binding to FKBP12, which binds to a specific domain of mTORC1 to inhibit its kinase activity. Previously we have reported that response to rapalog therapy in RCC is associated with mutation in the mTOR pathway genes: TSC1, TSC2, and MTOR [14]. Another recent study reported that mutations in PBRM1 were associated with response to rapalog therapy in the RECORD-3 trial [15]. To our knowledge no previous study has examined molecular mechanisms of acquired or secondary resistance of rapalog therapy in responding patients with RCC.
We identified six mRCC patients who developed resistance to rapalog therapy after initial clinical benefit, and who had available pre-treatment and post-treatment tumor samples. Five of these 6 patients had been studied in our earlier analysis of the association between mTOR pathway mutations and response to rapalogs in RCC [14]. (However, note that the earlier study did not include analysis of post-treatment as well as pre-treatment samples, and was only gene panel sequencing, not whole exome sequencing.) Five patients had clear cell RCC (ccRCC) and one had chromophobe RCC (Table 1). Most patients had received prior treatment with vascular endothelial growth factor targeted therapies (n = 5) and received a rapalog in the second (n = 3) or third (n = 2) line setting. Five patients received treatment with everolimus and one with temsirolimus. The patients received rapalogs for a median of 9.5 months (range: 5.5–46 months), after which they progressed. One (the chromophobe RCC) had a complete response, four had a partial response, and one had 10% tumor shrinkage (Stable Disease).
434 somatic variants were identified in these six patients’ biopsies, including both pre-treatment and post-treatment samples (S1 Table). The mutation profiles of the six individual tumors matched well with the expected genes and mutations from past studies in RCC. Three patients’ tumors showed mutations in TSC1, while one had a mutation in PTEN, and some of these patients had been included in our previous study showing that there is enrichment for mutations in mTOR pathway genes in RCC patients who respond to rapalog therapy [14]. Recurrent mutations (seen in more than one patients’ samples) were seen in 10 genes (Table 2). VHL mutations were seen in all 5 ccRCCs, as expected. PBRM1 mutations were also seen in all 5 ccRCC samples. Three samples had inactivating mutations in TSC1, as noted; three including the chromophobe RCC had TP53 mutations; while two had BAP1 mutations. Several genes with recurrent mutations were likely chance events, enhanced by their large size: DNAH11 (4516aa), TTN (34350aa), PIEZO1 (2521aa), TRPM6 (2022aa); none are thought to be involved in the pathogenesis of any form of cancer. Furthermore, several of these mutations were silent, also suggesting that they were random events. Recurrent mutations in PTPRN2 may have also been due to random chance, and one of those was also silent.
ABSOLUTE was used to calculate the tumor purity of the samples, and relative allelic frequency of each mutation in the cancer cells, termed clonal cell fraction (CCF) [16–18]. Phylogic was used to construct phylogenetic trees for the pre- and post-treatment pairs of samples, and to generate diagrams showing the major mutational events and clonal evolution (S1 Fig).
46% of the somatic variants identified were clonal in both the pre-treatment and post-treatment samples, while 25% of variants were not seen in the pre-treatment sample but were clonal in the post-treatment sample (S2 Table). VHL mutations were clonal or high subclonal in all 5 ccRCC samples, as expected, and showed no significant change in clonal representation in the paired samples. A TSC1 mutation went from subclonal to zero in one sample (MT_002) after treatment, consistent with a model in which TSC1-mutation bearing cells might have been sensitive to rapalog therapy, and were selectively killed or inhibited by such therapy. However, in another sample (MT_006), an inactivating TSC1 mutation went from zero cancer cell fraction in the pre-treatment sample to clonal in the post-treatment sample, in direct contrast to this model. Mutations in each of TP53 and PTPRN2 also were enriched in one patient’s post-treatment sample but lost in another patient’s post-treatment sample.
Multiple copy number (CN) events were also identified in these tumor samples (S3 Table). All ccRCC samples showed loss of one copy of 3p, as expected, and the chromophobe sample showed loss of multiple chromosomes (1, 2, 6, 8, 10, 11p, 13, 17, 21). There were no focal amplifications identified in the post-treatment samples, but rather a variety of chromosome and arm level gains and losses of uncertain significance.
PBRM1 mutations were clonal in both pre-treatment and post-treatment samples from two patients, but went from zero to clonal in 3 patients samples’ following treatment. Two of these acquired mutations were nonsense mutations in PBRM1 while the third was a synonymous change. PBRM1 mutations are seen in about 30% of ccRCC samples [19], and it is possible that the appearance of PBRM1 mutations in the post-treatment samples was due solely to tumor heterogeneity and/or sampling issues. However, as noted above, VHL mutations were present at or near clonal frequency in all 5 ccRCC samples both pre- and post-treatment. Hence, the finding that PBRM1 mutations were present in all 5 ccRCC samples at the time of resistance, including acquisition of mutations in 3 ccRCC samples led us to explore the hypothesis that PBRM1 loss is a mechanism of resistance to rapalog therapy in ccRCC.
We studied several ccRCC cell lines with native wild type (786-O, SNU-349) or native mutant PBRM1 (A704, RCC4) [20–22]. In 786-O cells, 4 different shRNA clones (sh889, sh890, sh994, sh326) were used to generate stable lines with reduced PBRM1 expression (Fig 1A). The 3 lines with greatest knock-down showed somewhat variable growth rates in comparison to a similarly derived line expressing a control shRNA, and all lines showed moderate growth inhibition in response to rapamycin treatment at 20 nM for up to 3 days with no difference between the PBRM1 knock-down cells and those expressing control shRNAs (Fig 1B). Similar results were obtained from SNU349 cells with stable downregulation of PBRM1 (S2 Fig).
We also studied RCC4 cells that are known to have biallelic mutation in PBRM1 that leads to a complete loss of expression of the protein [21]. We used derivative RCC4 cell lines expressing either control vector or wild type VHL [20]. We observed significant growth inhibition of both RCC4-vector and RCC4-VHL cells in response to each of rapamycin and Torin1 in 96well plate assays (S3 Fig). This also suggested that PBRM1 loss did not lead to rapamycin resistance.
We also performed the converse experiment, in that we examined rapalog sensitivity of derivatives of a native PBRM1 null cell line (A704) expressing either empty vector (A704_EV), wild type PBRM1 (A704_WT), or a mutant Q1298* PBRM1 (A704_Q1298*) under regulation of doxycycline [23] (Fig 2A). There were minor differences in the growth rate of these various A704 derivative lines, but there was no appreciable difference among them in response to rapamycin treatment for up to 6 days with and without doxycycline induction (Fig 2B and 2C). We also examined the growth of these various A704 sublines in a clonogenic assay under rapamycin treatment for 30 days (Fig 2D and 2E). Different numbers of colonies were seen for the different A704 derivative lines that varied to a small extent with and without doxycycline. All three lines showed a significant reduction in colony growth in response to rapamycin, and for the A704_WT line, a similar reduction in colony number was seen with and without doxycycline induction of wild type PBRM1.
One hundred twenty-nine genes showed a significant increase in the cancer cell fraction for a single mutation in a single post-treatment tumor sample in comparison to the pre-treatment sample (S4 Table). To assess whether these genes were enriched in a pathway that might be consistent with resistance to rapalogs, we performed gene set enrichment analysis using hallmark gene sets [24]. The 129 genes showed modest overlap with two hallmark gene sets (E2F targets, and mitotic spindle), 4 genes each with FDR q = 0.044, and no enrichment for any other hallmark gene set. None of these genes were obvious members of the PI3K-AKT-mTOR signaling cascade. Some were known cancer genes, including CDKN2A, KEAP1, MYCN, PLK4, SETD2, TP53. It is possible that any of these singleton genetic changes contributed to resistance to rapalog therapy in an individual patient.
In this study, we sought to identify molecular mechanisms of secondary resistance to rapalog therapy in patients with RCC who had demonstrated initial clinical benefit (5 of 6 had PR/CR). Our analyses were limited by both the relatively small number of samples available to us (n = 6), and that only one of the six patients had experienced a durable CR, whereas four had PRs only, and one had prolonged stable disease.
Furthermore, tumor heterogeneity is well-known in RCC [12,25–27], and complicates interpretation of genetic differences seen in the two paired tumor samples. 63 to 69% of all somatic mutations were not detectable across every tumor region of RCC when multiple samples were analyzed [25], similar to our findings that 46% of somatic mutations were not seen in both of our paired samples for these 6 patients (S2 Table). Consequently the 129 genes with somatic mutations that were enriched (CCF increased by > 0.5) in a single post-treatment sample likely reflect that heterogeneity, and are each unlikely to contribute to rapalog resistance. However, it is possible that some of those ‘acquired’ mutations seen only in the post-treatment samples may have contributed to resistance.
PBRM1 was the only gene to show gain of mutation in more than one patient in the post-treatment sample. Two ccRCC patients showed gain of a nonsense mutation in PBRM1, while a third showed gain of a synonymous mutation, and all three were clonal in the post-treatment sample. However, through analysis of RCC cell lines with both native PBRM1 expression, and those with native loss of PBRM1, we could find no evidence of differential sensitivity to rapamycin therapy in standard and clonal growth assays at the standard dose of rapamycin 20nM, which is similar to serum trough levels of this compound achieved in patients with standard dosing, 10-15nM. It remains possible that the effects of PBRM1 loss with respect to rapalog sensitivity are not modeled well in tissue culture systems, and that this genetic event still contributed to rapalog resistance in patients. Alternatively, tumor heterogeneity or sampling issues may account for these PBRM1 mutations seen only in the post-treatment samples.
Interestingly, Hsieh et al. reported that PBRM1 mutations were associated with longer progression free survival (PFS) in metastatic RCC patients treated with first-line everolimus in the RECORD-3 trial [15]. Hence it is possible that the finding of PBRM1 mutations in our 5 ccRCC patients correlates with response to rapalog therapy, though not seen in the pre-treatment tumor specimen, rather than representing a mechanism of resistance. On the other hand, if PBRM1 mutations cause response, then loss of PBRM1 mutation might be expected in the post-treatment resistance sample, given tumor heterogeneity in RCC, and we did not observe this.
We were somewhat surprised that we did not identify any secondary mutations in MTOR in these samples. Previous studies have identified MTOR mutations capable of preventing each of rapalog and ATP-competitive kinase inhibition of mTOR kinase activity in vivo and in vitro [28,29]. Furthermore, a variety of activating mutations in MTOR are well-known in both RCC and other cancer types [10–12,27,30], and in some cases are associated with exceptional response to rapalog therapy. Nonetheless, no MTOR mutations were seen in any of these six patients, nor were MTOR mutations associated with resistance development in even a single case.
Hence, we conclude that mechanisms of resistance to rapalog therapy in RCC are not easily explained by mutations in most cases, and likely depend on more subtle transcriptional and/or epigenetic changes. Transcriptional effects of PBRM1 mutation have recently been identified in analysis of the association of response of ccRCC to immune checkpoint therapies [31], and may have a similar effect in enhancing response to rapalogs.
This research study was approved by the Dana Farber/Harvard Cancer Center Office for Human Research Studies, protocol 07–336, and all subjects provided written informed consent.
We identified patients with metastatic RCC who initially responded to treatment with rapalog for at least 5 months, and then showed progressive disease, with available pre- and post-treatment (at time of progression) biopsies through a search of our own medical facilities and national and international collaborators. The six patients were treated with temsirolimus or everolimus at one of three medical centers: Dana-Farber Cancer Institute, Beth Israel Deaconess Medical Center, and the University of Utah Hospital.
From each patient, we collected 1) a pre-treatment nephrectomy specimen, 2) post-treatment metastatic tumor specimen, and 3) a venous blood specimen. An expert genitourinary pathologist (SS) reviewed hematoxylin and eosin stained slides. For each case regions containing at least an estimated 50% tumor cells were selected for DNA extraction. Selected tumor areas were scraped off unstained slides and DNA was extracted using the QIAamp DNA FFPE Tissue Kit (QIAGEN, Valencia, CA), according to the manufacturer guidelines.
Whole exome sequencing (WES) was performed at the Broad Institute following standard protocols. Sequencing data was analyzed using standard analytic pipelines deployed in the Firehose environment. Mutect and Indelocator were used to identify somatic mutations in tumor-normal pairs. Every single mutation that was called by these pipelines was scrutinized using IGV to assess the reliability of the variant call, and to confirm allele frequencies seen in the various samples. Many variants were discarded due to misaligned reads, repetitive sequence tracts, low quality base or read scores, or reads seen in only a single direction.
Recapseg and AllelicCapseg were used to determine copy number profiles. ABSOLUTE was used to estimate sample purity and ploidy, absolute copy number for each chromosome and segment, and clonal cell fraction (CCF) values for each mutation [16,17]. Phylogic was used to perform Bayesian clustering of mutation CCFs, and to construct phylogenetic trees for the pre- and post-treatment samples, as described previously [17,18].
We studied RCC cell lines SNU349, 786-O, RCC4, and several versions of the A704 cell line (A704+BAF180_WT, A704+BAF180_Q1298*, A704+BAF180_EV) previously generated by one of the co-authors (WG) [19]. Stable PBRM1 knock down was performed using four different shRNAs (Sigma) in lentiviruses following standard methods; reduced PBRM1 expression was confirmed by SDS-PAGE and immunoblotting of cell lysates. Cell growth assays were performed for the indicated time points in clear 96-well plates for Crystal Violet staining or white opaque 96-well plates (Corning) for cellular ATP measurement using Cell Titer Glo (Promega). Rapamycin was used at 20nM and compared with vehicle (DMSO) treatment. Clonogenic cell proliferation assays were performed by plating 200 cells in 10 cm dishes (n = 3 for each cell line/condition), treating them every 3 days with Rapamycin (20nM) or corresponding DMSO control, with and without doxycycline (1ug/ml) for 30 days, and then counting visible colonies without magnification following Crystal Violet staining.
|
10.1371/journal.pntd.0007404 | Pregnancy outcomes and mother-to-child transmission rate in HTLV-1/2 infected women attending two public hospitals in the metropolitan area of Rio de Janeiro | HTLV-1/2 are transmitted sexually, by whole cell blood products and from mother-to-child (MTC), mainly through breastfeeding. HTLV-1/2 prevalence in pregnant women is high in Rio de Janeiro, however there were no local studies addressing the rate of adverse pregnancy outcomes (APO) and MTC transmission. The aim was to study sociodemographic characteristics which may be associated to HTLV-1/2 infection and describe pregnancy outcomes and MTC transmission in HTLV-1/2-positive women. The cross-sectional study screened 1,628 pregnant women in of Rio de Janeiro (2012–2014) and found 12 asymptomatic carrier mothers (prevalence = 0.74%). Pregnancy outcome information was retrieved from medical records. Sociodemographic characteristics were similar between the positive and negative groups except for maternal age, which was higher in carrier mothers. The incidence of adverse pregnancy outcomes was similar in infected and non-infected patients (p = 0.33), however there was a high rate of premature rupture of membranes (PROM) amid infected mothers (3/12). Multilevel logistic regression found that for each additional year of age, the chance of being HTLV-1/2-positive increased 11% and that having another sexually transmitted infection (STI) increased 9 times the chance of being infected. Carrier mothers had more antenatal visits (OR = 5.26). Among the children of HTLV-1/2-positive mothers there was one fetal death, one infant death and one loss of follow-up. After two years of follow-up there was one case of MTC transmission (1/9). The mother reported breastfeeding for one month only. Knowledge about factors associated to HTLV-1/2 infection, its impact on pregnancy outcomes and the MTC transmission rate is important to guide public health policies on antenatal screening and management.
| HTLV-1/2 are retroviruses transmitted by sex, blood products and from mother to child (MTC), mainly through breastfeeding. There is currently no vaccine, treatment or cure. Although it’s mostly asymptomatic it can cause disabling and even lethal diseases in carriers. The prevalence of HTLV-1/2 in pregnant women at the metropolitan area of Rio de Janeiro is high (0.74%). Our aim was to study the sociodemographic characteristics which may be associated to HTLV-1/2 infection and describe pregnancy outcomes and MTC transmission in the infected population. Apart from being slightly older, there were no differences in the carrier mothers’ sociodemographic profile. Pregnant women with sexually transmitted infections had a 9-fold chance of being HTLV-1/2 positive. Although adverse pregnancy outcomes were not increased, infected mothers had a high rate of ruptured membranes. Among the children of HTLV-1/2-positive mothers there was one fetal death, one infant death and one loss of follow-up. There was one case of MTC transmission (1/9), after one month of breastfeeding. Knowledge about factors associated to HTLV-1/2 infection, its impact on pregnancy, and the MTC transmission rate is important to guide further research and public health policies.
| Human T-lymphotropic virus types 1 and 2 (HTLV-1/2) are human oncogenic retroviruses first identified in the early 1980’s [1]. There are six subtypes of HTLV-1 (A to F), which have no impact on the clinical expression of the disease [2]. There are two other types of HTLV (3 and 4); but there is no evidence of their pathogenicity in humans [3].
HTLV-1/2 viruses are globally distributed and there may be up to 10 million infected worldwide [4]. Prevalence is characterized by endemic clusters occurring next to low prevalence areas. It also varies considerably according to the ethnical and social background of the population. Since transmission occurs through infected body fluids, intravenous (IV) drug users and sex workers have been reported as high-risk groups [4]. The association between low social and economic level and lower education is not homogeneous among studies and most likely represents a bias. Endemic HTLV-1 clusters are found in Sub-Saharan Africa, South-western Japan, Central and South America as well as the Middle East and Melanesia [4]. Regardless of the area, seroprevalence increases with age, particularly in women due the excess efficiency of the male-female sexual transmission. HTLV-2 is endemic in Pygmy tribes of Central Africa and in several Native American populations, particularly in the Amazon area [5,6]. It is also frequent in IV drug users, often in co-infection with HIV [5,7].
Brazil may be the country with the highest absolute number of HTLV-1/2 carriers in the world. Estimates range from 800,000 to 2.5 million people [5,8,9]. Such variation in numbers can be explained both by the epidemiological characteristic of the infection and by the lack of data, with large areas of the country unmapped.
Infection is perennial and most of the patients are asymptomatic reservoirs, sustaining the chain of transmission. In contrast, up to 8% of HTLV-1 carriers develop severe diseases, mainly the highly aggressive adult T-cell leukaemia/lymphoma (ATLL) and the painful and disabling HTLV-1-associated myelopathy/Tropical Spastic Paraparesis. Type 1 virus also causes a spectrum of inflammatory conditions, such as dermatitis and uveitis [10]. In turn, HTLV-2 has been associated to erythrodermatitis, neurologic disorders and opportunistic infections [7].
Literature about the effect of HTLV-1/2 infection on pregnancy outcomes is scarce. Only one study on the subject was published. It was conducted in Africa, between 1986 and 1988, involving 45 HTLV-1/2 positive pregnant women and 90 negative ones. No statistically significant differences were found between the groups regarding neither sociodemographic profile, pregnancy and neonatal outcomes [11].
Only four Brazilian researches on HTLV-1/2 in pregnant women and puerperae assessed previous obstetric history and pregnancy outcome, and they reported only on miscarriage [12–15]. Large regional studies which altogether included over 130,000 pregnant women report miscarriage rates between 22% and 30%, notably higher than that observed in non-infected patients [13–15]. On the other hand, the only research conducted on an endemic area found a pregnancy loss rate of 10%, similar to the general population [12]. Dal Fabbro’s study was the only to report the frequency of two or more previous miscarriages, which was 0.8% in HTLV-1/2 infected women. This number is equivalent to the incidence of recurrent pregnancy losses in the general population [13].
HTLV-1/2 is transmitted via whole cell containing body fluids, mainly through sexual contact, exposure to blood products or viscera and from mother-to-child (MTC). The relative importance of each mode of transmission is still largely unknown and most likely it varies with the population involved. In endemic areas such as Japan MTC transmission has been described as the main source of transmission, mainly through breastfeeding [12,16]. Only 2.5–5.0% of children are seroconverted in the absence of breastfeeding while up to 25% are infected if breastfed for over 12 months [16–19]. In fact, a recent Brazilian study found a vertical transmission rate of 50% in children who were breastfed for over 24 months [17]. This research also detected an increased risk of infection in siblings, confirming the trend for familial clustering of the disease [17]. Higher proviral load and antibody titers in maternal blood and breastmilk are also associated with increased MTC transmission rate [17,20–22]. On the other hand, peripartum transmission has been shown to have little impact on the burden of disease [19,23].
The aim of the research was to describe the epidemiological profile of pregnant women diagnosed with HTLV-1/2 in the metropolitan area of Rio de Janeiro, the occurrence of adverse pregnancy outcomes (APO) and the rate of mother-to-child transmission.
The study population consisted of 1,628 pregnant women. The first 1,204 were enrolled at admission for delivery as part of a research on HTLV-1/2 prevalence conducted at two public hospitals in the metropolitan area of Rio de Janeiro: the ‘Pedro Ernesto’ University Hospital of the Rio de Janeiro State University (Universidade do Estado do Rio de Janeiro–HUPE/UERJ) and the ‘Hospital Estadual da Mãe’ (HEM). HUPE is a referral centre for high-risk patients while HEM, situated at the adjacent city of Mesquita, assists low and medium-complexity cases [24]. As a result of the relevant prevalence found in the local population (0.66%) at the first part of the study, routine HTLV-1/2 screening was instituted in HUPE’s perinatal unit in July 2013. The other 424 women were recruited at their first antenatal visit. During the first part of the study, roughly 24% of the women attending the HEM were recruited while the uptake at UERJ was 40%. During the antenatal screening period at UERJ only four approached women refused to take part in the study. The study was conducted between November 2012 and December 2017. Subjects were recruited until 2014 and the children of HTLV-1/2 positive women were followed for three years. Women who were mentally unable to give consent or who declined to take part in the research were excluded. A structured questionnaire with socio-epidemiological, clinical and reproductive data was applied at recruitment. Women found to be HTLV-1/2 positive were counseled by a multidisciplinary team which provided health information and psychosocial support. They were advised about the risk of vertical transmission though breastfeeding and formula milk was provided to safeguard the infants’ nutrition. Children of carrier mothers were monitored for at least three years after the birth on the paediatric infectology department of HUPE. There were two exceptions: one who was followed up at her local health care center and another who was lost to follow up.
Blood for HTLV-1/2 screening was collected either during the routine antenatal care or at the admission for delivery by chemiluminescent microparticle immunoassay (CMIA—Architect rHTLV-I/II, Abbott). Children’s samples were also screened by CMIA within the month after birth, at six months, one and two years of age. Infection was confirmed if the child remained seropositive after 24 months, and additional yearly exams were performed for follow-up. Reactive samples were confirmed by Western blot (WB, Inno Lia HTLV-I/II score Biomerieux). Two pregnant women with positive screening tests and negative WB results were considered false positive and allocated to the negative group in the statistical analysis. Routine antenatal screening for sexually transmitted infections (STI) in Brazil consists of VDRL test for syphilis and ELISA tests for HIV, hepatitis B and C. In case of positive screening the confirmation tests are FTA-Abs for syphilis and western blot for HIV. All tests were done at HUPE/UERJ’s Clinical Analysis Laboratory.
This research complies with the Declaration of Helsinki and the Resolution 466 of December 12, 2012 of the Brazilian Ministry of Health. The project was approved by the Rio de Janeiro State University Research and Ethics Committee (COEP-UERJ, process 034.3.2012) and sponsored by the Fundação de Amparo à Pesquisa do Estado do Rio de Janeiro (FAPERJ, E-26/110.351/2012). Written informed consent was obtained from all the subjects and from the legally authorized representatives of the minors who agreed to take part in the research. Anonymity and data confidentiality were guaranteed.
Means, medians, standard deviations and percentages were used to describe the results. Medians were used to define the cutoff point used to convert numerical variables to categorical ones (age, family income, number of partners and number of pregnancies). The following variables were considered as adverse pregnancy outcomes: fetal demise, perinatal death, maternal hypertensive syndrome, preterm birth and admission to the neonatal intensive care unit. A composite variable named ‘adverse outcome’ was created in view of the low frequency of each adverse obstetric outcome studied. Missing data were excluded from the statistical analyses. Mann-Whitney and Fisher’s exact tests were used to compare categorical variables between the HTLV-1/2 positive (G1) and negative (G2) groups. Multilevel logistic regression was performed to assess the factors associated to HTLV-1/2 infection. Age, number of pregnancies, coinfection with other STI, condom use in pregnancy and number of antenatal appointments were included in the model as fixed effects and "Hospital" was included as random effects (random intercept). The Epi-info software version 3.5.2 and the R-Project version 3.3.1 were used for building the database and performing the statistical analyses.
Prevalence of HTLV-1/2 infection in this study population was 0.74% (12/1628), with no significant difference between the hospitals (0.82% in HUPE, 0.67% in Mesquita; p = 0.78).
Among the sociodemographic characteristics (Table 1) only maternal age was significantly different between the HTLV positive and negative groups. The age was over 24 years in 83% (n = 10) of the HTLV+ group (G1) and in 53% of G2 (p = 0.03). Most women in both groups reported being non-white (p = 0.99); having at least 10 years of formal education (p = 0.99) and being in a stable marital relationship (p = 0.74). About half the patients in both groups had monthly household income higher than two minimum wages (p = 0.99). No subjects reported behavioural risk factors like the use of IV drugs, having multiple sexual partners and being a sex worker.
Sexual and reproductive characteristics are summarized in Table 2. HTLV-1/2 positive women (G1) were more frequently coinfected with another STI than patients in G2 (41.7% vs. 8.9%, p< 0.01). Three of the HTLV-1/2 positive women had syphilis, one had HIV and another had histopathologically confirmed HPV condilomata. Most women in both groups had their first antenatal visit before 12 weeks of pregnancy (66.6% and 59.2%, p = 0.46). The number of antenatal appointments was higher in the HTLV-1/2 positive group, even after the multivariate analysis (81.8% vs. 52%, p = 0.03). Regarding the obstetric history, around 2/3 of infected women had more than two previous pregnancies (66.7%), while the proportion was inverse in G2 (34.9%). Number of sexual partners, previous reproductive history and frequency of condom use were similar in both groups.
The prevalence of comorbidities was high in the study population (37.6%), being significantly more frequent in HUPE (78.1%) than in HEM (4.8%, p<0.001). However, the rate was similar between the infected and non-infected groups (p = 0.16).
There were five adverse pregnancy outcomes in the HTLV positive group. The difference was not significant when compared to the negative group (p = 0.33) (Table 3). There were three cases of coinfection with syphilis (3/12) and three of premature rupture of membranes (PROM -3/12) at term. One patient with PROM also had syphilis, but the other two cases had no other risk factors (2/12; 16.6%). The prevalence of PROM in the HTLV-1/2 negative group was 4.7% (71/1,518). Although the difference seems significant, it was not confirmed by Fisher’s exact test (p = 0.26), probably due to the small number of cases. One HTLV-1/2-infected newborn was admitted to the neonatal intensive care unit (NICU) due to prolonged PROM. Among non-infected babies, 137 were admitted to the NICU (8.3% vs. 10.0%, p = 0.66). There were missing data on 210 seronegative pregnancy outcomes.
Differences in maternal age, STI coinfection and the number of antenatal visits remained significant after multivariate logistic regression (Table 4).
Multilevel Logistic regression model–dependent variable: HLTV; fixed effects: age, number of pregnancies, coinfection with other STI, condom use in pregnancy, number of antenatal appointments; random effects: hospital.
Each additional year of maternal age increased the chance of being HTLV-1/2 positive in 11% (OR = 1.11). Having another STI increased 9 times the chance of being infected (OR = 9.27). In G1 there was a higher frequency of antenatal visits (OR = 5.31) (Table 4).
Among the children of the 12 HTLV-1/2 infected mothers, there was one fetal demise, one infant death and one loss of follow-up. The fetal death occurred at 24 weeks of pregnancy and the infant died at two months of age due to pneumonia. One child had its seroconversion (1/9) confirmed after two years of age. She was born at January 2013 and remains asymptomatic under medical surveillance at the HUPE to date. Her mother referred breastfeeding for less than one month. The other eight children were periodically monitored, seven at the pediatric infectology department of UERJ and one at its local health care center. Despite their economic difficulty, their mothers reported avoiding breastfeeding since they were aware of their carrier status.
The HTLV-1/2 prevalence found was consistent with the study previously published by our group which recruited only women admitted for delivery [24].
The sociodemographic profile of both groups was similar except for the older age found in the infected group. This is in accordance with the international literature on the disease’s epidemiology [4] and with the studies performed in Brazilian areas with high prevalence of HTLV-1/2 [17, 25,26]. On the other hand, in a large research performed in Gabon, where HTLV-1/2 prevalence is over 10%, there was no sociodemographic difference between the groups [11]. The multivariate analysis of sexual and reproductive characteristics found two significant differences between the infected and non-infected groups: STI coinfection and number of antenatal visits. The increased frequency of STI coinfection in HTLV-1/2 carriers, particularly syphilis, was consistent with studies from endemic areas [12,14,27,28], although it’s not a universal finding across publications [11,13,15,26,29]. On the Gabon research there were twice as many cases coinfected with syphilis than controls, however statistical significance was not reached, probably because insufficient sample size [11]. The latest Salvador study found that 21.5% of HTLV-1/2 infected subjects also had syphilis (OR = 36.7) [28]. As for the higher number of antenatal visits in the HTLV-1/2 positive group, it cannot be explained by the knowledge of the carrier status itself since only one patient was aware of the infection at the beginning of the pregnancy. There was also no significant correlation with the presence of comorbidities (p = 0.38). The hypothesis of the more frequent antenatal visits being due to these women having more previous adverse pregnancy outcomes seemed significant (p = 0.04), but its confidence interval was too wide (0.91–10.2). It is true, though, that the lack of significance could be caused by the small number of cases. A previous case-control study which addressed this variable also failed to find any difference between infected and non-infected patients [11].
Regarding the reproductive history, it’s striking that almost half of the carrier mothers (5/12) had previous adverse obstetric outcomes (three first trimester miscarriages, one fetal demise and one FGR with neonatal death). Three of these women had no comorbidities. It must be said however, that no causal link to HTLV-1/2 can be inferred since their infectious status was unknown during the previous pregnancies. The prevalence of early miscarriages was similar (circa 21%) among HTLV-1/2-positive and negative patients (p = 0.86). This was in accordance with data from other endemic areas [11,12].
Regarding the adverse pregnancy outcomes studied, no difference was observed between infected and non-infected patients. It’s important to stress that this finding cannot be generalized in view of the small sample size and the low incidence of the outcomes. The two studies reporting on obstetric results of HTLV-1/2 infected women did not find association between the infection and adverse pregnancy outcomes as well [11,12]. The Gabon research found a trend for preterm delivery and complicated pregnancies in HTLV-1/2 positive women [11]. Unfortunately, even this study, which followed 45 infected patients and 90 controls, was underpowered for this statistical analysis.
There were no cases of preterm delivery or low birth weight among HTLV-1/2 positive patients, in accordance with the findings of Bittencourt et al [12]. There was one case of fetal growth restriction and intrauterine demise in an otherwise healthy HTLV-1 infected mother. This woman reported two previous adverse pregnancy outcomes, but it’s unknown whether she was already infected at the time. At the Salvador case-control study[12], there were also fetal deaths on the HTLV-1/2 carrier mothers’ group, but those happened in patients with additional comorbidities such as hypertension and falciform anaemia. Additionally, that study reported three cases of hypertension in pregnancy and one admission to the NICU due to neonatal sepsis after PROM.
Among the infected mothers there were three cases of term PROM (25%). In a recent Brazilian study using data from the Ministry of Health [30], PROM was found to complicate approximately 4.2% of all livebirths in the country. This number is in line with the prevalence found in our HTLV-1/2 negative group. Uterine inflammation and sexually transmitted infections have been shown to be associated with obstetric complications such as PROM [31–34]. HTLV-1/2 infection is also known to be linked to different inflammatory and infectious manifestations. Thus, it seems reasonable to interrogate whether HTLV-1/2 infection may increase the risk of PROM. Unfortunately, the largest study on HTLV-1/2 pregnancy outcomes, performed in Gabon, didn’t assess the incidence of PROM [11]. This finding prompts the need for further research, adequately powered to elucidate the matter.
In our study, the mother of the only infected child reported breastfeeding for less than one month. Mother-to-child transmission of HTLV-1/2 occurs mainly through breastfeeding, ranging from 3.9% to 22% in endemic areas [35]. The policy of universal HTLV-1/2 antenatal screening and contraindication of breastfeeding for infected mothers at the Nagasaki province reduced the local MTC transmission rate from 20.3% to 2.5% [16], which are the known seroconversion rates of prolonged breastfeeding (> 6 months) and exclusive bottle-feeding, respectively. However, even with shorter periods of breastfeeding the MTC transmission rate is greater than using only infant formula (7.4% vs. 2.5%) [22]. Other possible reasons for the MTC transmission in this case are peripartum infection or additional risk factors such as: high maternal antigenemia, concentration of gp46 HTLV-1/2 antibodies, the presence of anti-Tax antibodies, or the human leucocyte antigen system (HLA) type concordance between mother and child [17,18,35,36,37,38,39]. The hypothesis of peripartum infection seems unlikely since the child was delivered by caesarean section due to hypertensive syndrome without PROM or labour. A limitation of this study is that it could not assess the other variables mentioned, such as proviral load and HLA type. Other two Brazilian studies report MTC transmission after less than a month of breastfeeding. In both cases the mothers’ proviral loads were extremely high [17, 37]. Another limitation of the study was the small sample of infected patients and the low frequency of adverse pregnancy outcomes, which were grouped for the statistical analysis. A study which is properly powered for statistical analysis on this matter would require a much greater sample size, and that may prove impeditive in areas of intermediate prevalence such as ours.
On the other hand, a strength of the study is that the children of infected mothers were followed up for three years, a gold standard set by the Nagazaki study group [16]. This was proposed since some cases of seropositivity in children are caused by maternal antibodies, which generally disappear after 12 months of life. Our MTC transmission rate was 1/9 (11%), similar to the ones found in Haiti and Guyana [11,35].
The study confirmed the high prevalence of HTLV-1/2 in pregnant women at the metropolitan area of Rio de Janeiro and found no sociodemographic difference between infected and non-infected patients. Carrier mothers frequently reported previous adverse pregnancy outcomes (5/12), but at the current pregnancy there was only one unexplained fetal demise (growth restricted) and one admission to the NICU due to sepsis. There was a significant association to other STI, but the intriguing point was the number of PROM cases among infected women (3/12).
Since there is no treatment or immunization for HTLV-1/2, preventive measures are currently the only effective way to break the chain of transmission. Thus, it is vital to increase awareness about the infection among health providers and the population. Safe sex campaigns are already extensive, therefore tackling MTC transmission is likely to have the most significant effect on the longitudinal perpetuation of the virus and the reduction of HTLV-1/2 associated diseases, particularly in endemic areas. There are no studies on antiretroviral therapy or mode of delivery to address the potential for reducing MTC transmission. Thus, avoidance of breastfeeding remains the only effective way to block the MTC transmission of the virus. The use of this strategy as a public policy in low income areas is still not a consensus, since breastfeeding also plays a role in reducing infant mortality and morbidity through immunity boosting and protecting against infections.
Since there is no clear difference in the sociodemographic profile of HTLV-1/2 carriers, routine prenatal screening at endemic areas is supported by several research groups [9,10,13,15–17,21,25,27,28,30,31,40–42]. Considering the number of livebirths at the metropolitan area of Rio de Janeiro (236,960 LB in 2016) and the HTLV-1/2 seroprevalence in pregnant women found in this study, the introduction of local routine antenatal screening could avoid over one thousand cases of MTC transmission in a year[41]. A similar estimate was found by a recent epidemiological study [42].
Further studies are needed on the cost-effectiveness of such strategy across different prevalence areas and socioeconomic resources. Moreover, epidemiological mapping of prevalence and mother-to-child transmission is needed to guide public health policies on antenatal screening and management. Additionally, better understanding of the burden of the infection in pregnancy may help to improve HTLV-1/2 infected mothers’ antenatal care and their children’s outcome.
|
10.1371/journal.ppat.1005339 | EBNA2 Drives Formation of New Chromosome Binding Sites and Target Genes for B-Cell Master Regulatory Transcription Factors RBP-jκ and EBF1 | Epstein-Barr Virus (EBV) transforms resting B-lymphocytes into proliferating lymphoblasts to establish latent infections that can give rise to malignancies. We show here that EBV-encoded transcriptional regulator EBNA2 drives the cooperative and combinatorial genome-wide binding of two master regulators of B-cell fate, namely EBF1 and RBP-jκ. Previous studies suggest that these B-cell factors are statically bound to target gene promoters. In contrast, we found that EBNA2 induces the formation of new binding for both RBP-jκ and EBF1, many of which are in close physical proximity in the cellular and viral genome. These newly induced binding sites co-occupied by EBNA2-EBF1-RBP-jκ correlate strongly with transcriptional activation of linked genes that are important for B-lymphoblast function. Conditional expression or repression of EBNA2 leads to a rapid alteration in RBP-jκ and EBF1 binding. Biochemical and shRNA depletion studies provide evidence for cooperative assembly at co-occupied sites. These findings reveal that EBNA2 facilitate combinatorial interactions to induce new patterns of transcription factor occupancy and gene programming necessary to drive B-lymphoblast growth and survival.
| Epstein-Barr Virus (EBV) reprograms host cell transcription through multiple mechanisms. Here, we show that EBV-encoded transcriptional co-activator EBNA2 drives the formation of new chromosome binding sites for host cell factors RBP-jκ and EBF1. The formation of these new sites is EBNA2-dependent. These newly formed sites have overlapping or neighboring consensus binding sites for these factors, but are only co-occupied in the presence of EBNA2. Newly formed, co-occupied binding sites are highly enriched at promoter and enhancer regulatory elements of genes activated by EBV and required for B-cell proliferation and survival. These findings indicate that EBNA2 drives cooperative and combinatorial transcription factor interactions on chromosomal DNA. We suggest that models depicting the static binding of master regulatory transcription factors to consensus binding sites be revised, and that co-activators, like EBNA2, induce dynamic and combinatorial selection of genome-wide binding sites to alter gene regulation.
| Tumor viruses encode many factors that mimic and alter host cell processes. Epstein-Barr Virus (EBV) is a human tumor virus associated with various lymphoid and epithelial cell malignancies, including Burkitt’s lymphoma, nasopharyngeal carcinoma, and lymphoproliferative disorders in the immunosuppressed [1, 2]. EBV can efficiently immortalize naïve B-lymphocytes and establish long-term latent infection in these immortalized cells [3]. EBV expresses different viral gene products depending on the cell or tumor type where latent infection is established [4]. During natural infection, EBV drives naïve resting B-cells into a proliferative program resembling antigen driven B-cell germinal center reaction [5]. EBV drives B-cell proliferation and differentiation through a complex combination of viral proteins and non-coding RNAs [4]. EBV gene expression programs also change coordinately with the differentiating host-cell, and these changes have been referred to as latency types [5]. Thus, EBV infection mimics B-cell developmental programs in the absence of normal exogenous antigenic signal [5].
EBV encodes several nuclear proteins that modulate host and viral gene transcription [6]. EBNA2 is potent transcriptional co-activator that is expressed in early stages of EBV-induced proliferation of naïve B-cells, but its expression is extinguished at later stages of B-cell development. EBNA2 is best characterized for its physical interaction with the sequence-specific transcription factors RBP-jκ (also called RBPJ, CBF1, and CSL) at promoters and enhancers of EBNA2-regulated genes ([7, 8](reviewed in [9, 10]). RBP-jκ is thought to bind constitutively to many of transcriptional regulatory elements, and function as a scaffold for co-activators, like intracellular Notch, or various co-repressors in the absence of Notch [11]. While EBNA2 and Notch are not interchangeable for either viral or cellular functions, EBNA2 may be considered a viral mimic of Notch at some transcriptional regulatory elements. Recent genome-wide studies have revealed that EBNA2 frequently colocalizes with RBP-jκ, as well as other cellular factor binding sites, including Early B-cell Factor 1 (EBF1) [12]. EBF1 is a sequence specific DNA binding protein that functions as a B-cell identity factor necessary for the establishment of pro-B cells and maintenance of B-cell specific transcription programs [13]. EBNA2 colocalizes with EBF1 and RBP-jκ at many enhancer and super-enhancer regions in EBV positive lymphoblastoid cell lines [14]. Both EBF1 and RBP-jκ are master regulatory transcription factors that play fate determining roles in lymphoid cell development, but their functional cooperativity has not been experimentally established. Moreover, EBNA2 is not thought to alter the chromosome binding site selection of these sequence specific DNA binding factors.
Patterns of transcription factor binding to chromosomal regulatory elements is a primary determinant of cell fate and identity [13, 15, 16]. Transcription programs depend on many intrinsic and extrinsic factors that converge ultimately on the selective binding of transcription factors to recognition elements that may be embedded in non-permissive chromatin environments and subject to repressive epigenetic modifications. In addition to overcoming chromatin and epigenetic barriers, transcription factors engage in complex combinatorial interactions at promoter and enhancer regulatory elements. How combinatorial interactions between DNA-binding and non-binding factors contribute to site selection and transcription programming in a genome-wide context remains largely unexplored. Here we investigate how a viral-encoded non-DNA binding coactivator, EBNA2, can reprogram the target binding sites and functional cooperativity of two host cell master regulatory transcription factors, EBF1 and RBP-jκ.
Transcription factors with sequence-specific DNA binding capability are known to recognize a distinct primary DNA sequence element. However, not all consensus sequences are occupied by these transcription factors. Epigenetic factors, including DNA methylation, nucleosome positioning, heterochromatin, and competitive transcription factor binding may influence the occupancy at different chromosomal locations in cell-type dependent manners. It is also known that some chromosome occupancy may depend on cooperative binding with additional factors, including sequence specific factors and non-DNA binding co-regulatory factors. Here, we present evidence that EBNA2 is one such co-regulatory factor that alters the genome wide distribution of sequence-specific B-cell master regulatory factors EBF1 and RBP-jκ.
To investigate the landscape of transcription factor binding on identical genomes in two closely related cell types, we first compared the binding patterns of EBF1 and RBP-jκ on the EBV genomes in B-cells with distinct latency types. We compared an EBV positive lymphoblastoid cell line (LCL) with a type III latency expression pattern (EBNA2 positive) to an EBV positive Burkitt lymphoma (BL) cell line (Mutu I) with a type I latency expression pattern (EBNA2 negative) (Figs 1A, 1B, 1C and S1). We found that EBF1 and RBP-jκ bound with very similar occupancy in LCLs, but with different binding patterns in Mutu I. In LCL, EBF1 and RBP-jκ bound in close proximity to each other at the promoter regulatory elements for LMP1 and LMP2A. In contrast, neither EBF1 nor RBP-jκ binding was observed at the LMP1 and LMP2A locus in Mutu I. Both LMP1 and LMP2A are expressed at high levels in LCLs, but are repressed in Mutu I cells (Fig 1D and 1E). Mutu I-selective enrichment of EBF1 was observed at the EBER and OriP control region, which has important transcription regulatory functions in type I cells [17]. The LCL-specific enrichment of EBF1 and RBP-jκ binding at LMP1 and LMP2A was confirmed by ChIP-qPCR (Fig 1B). We also show that this general pattern of EBF1 and RBP-jκ occupancy is observed in an isogenic BL with type I (Kem I) or type III (Kem III) pattern of viral latency gene expression (Fig 1C). Protein expression of EBF1 and RBP-jκ, as well as EBV proteins specific for type III latency, were monitored by Western blot (Fig 1D). Although some differences in protein expression levels were observed, these differences alone are unlikely to account for the global changes in chromosome occupancies. We also observed some changes in EBF1 and RBP-jκ protein mobility, suggesting that post-translational modifications or alternative isoforms may contribute to the different chromosome occupancy. As expected, EBV type III latency genes were expressed exclusively in type III cells (Fig 1E). Taken together, these findings suggest that EBF1 and RBP-jκ can have cell type-dependent DNA binding patterns.
To determine if the cell-type specific binding of EBF1 and RBP-jκ extended to the cellular genome, we compared ChIP-Seq data for cellular genome binding patterns in LCL and Mutu I cells (Fig 2). We found that both EBF1 and RBP-jκ had large numbers of cell-type dependent binding sites, as well as a subset of binding sites common to both cells (Fig 2A). Total and cell-type specific EBF1 binding sites were lower in LCL relative to Mutu I, while RBP-jκ binding sites showed the inverse trend. The peak intensities of RBP-jκ and EBF1 binding sites also correlated with cell-type specific distribution as shown by scatter plot (Fig 2B). A heat map of binding sites centered on EBF1 site (Fig 2C left panels), or RBP-jκ sites (Fig 2C right panels) shows the relative distribution and co-occupancies of EBF1 and RBP-jκ in the different cell types. We observed that RBP-jκ sites trended to colocalize with EBF1. While the overlap of EBF1 and RBP-jκ binding was not restricted to LCLs, a distinct cluster of colocalized sites could be assigned as specific to each cell type. ChIP-Seq data sets were analyzed for localization to nearest gene transcription start sites, and these sites were then selected and retested for their cell-type specific enrichment by ChIP-qPCR (Fig 2D). Several of the co-occupied LCL-specific sites were found at well-established RBP-jκ regulated genes, including IL7, HES1, FCER2, and ICA1. These showed strong and selective enrichment for both EBF1 and RBP-jκ in LCL relative to Mutu I. In contrast, ZNF595, miR4325, and RNF144B showed Mutu I-specific co-occupancy (Fig 2F). We also examined the binding patterns of RBP-jκ and EBF1 in the isogenic BL-cells Κem I and Kem III, that have different EBV latency type gene expression programs (Fig 2E and 2G). EBF1 and RBP-jκ were highly enriched at the IL7, FCER2, and ICA1 gene loci in Kem III, but not Kem I (Fig 2E). In contrast, ZNF595, miR4325, and RNF144B were enriched for EBF1 and RBP-jκ in Kem I, but not Kem III (Fig 2G). HES1 was enriched for RBP-jκ in Kem III, but not for EBF1, indicating there are some cell type specific differences between Kem III and LCLs. Nevertheless, these findings suggest that master regulatory transcription factors, like RBP-jκ and EBF1, have different genomic binding patterns in different cell types and this appears to correlate with latency type gene programming.
EBV-encodes several transcriptional co-regulators that binds RBP-jκ and are expressed in type III, but not type I latency. We therefore aligned our ChIP-Seq data sets for RBP-jκ and EBF1 with several published ChIP-Seq data sets for EBNA2, EBNA3C, and EBNALP from both LCL and Mutu III cells [12, 18–20] (Fig 3A). The de novo motif analysis on EBNA2 peaks from LCL identified RBP-jκ, as well as EBF1 as significant motifs (S2 Fig). Aligning these data sets with EBNA2 from LCL revealed strong correlations with EBNA2 from Mutu III, as well as EBNA3C and EBNALP, as previously reported. Analysis of EBF1 and RBP-jκ ChIP-Seq revealed significant colocalization with EBNA2 in LCLs, but not in Mutu I (Figs 3A and S3). To further investigate the correlation of EBNA2 peaks with the LCL-specific EBF1 and RBP-jκ binding sites, we investigated the ratio of the peak intensity for EBF1 (x-axis) and RBP-jκ (y-axis) between LCL and Mutu I (Fig 3B). The ratio of EBF1 and RBP-jκ peak intensity (L/M) was highly correlated with EBNA2 co-occupancy (correlation coefficient = 0.8), suggesting that EBNA2 trends towards EBF1 and RBP-jκ sites that are of greater peak intensity in LCLs relative to Mutu I. These findings indicate that EBNA2 binds frequently at EBF1 and RBP-jκ co-occupied sites, and that these sites are enriched in LCL relative to a BL cell line with type I latency.
The convergent binding of EBNA2, EBF1, and RBP-jκ was further analyzed for colocalizations with other well-established histone modifications and transcription factors (Fig 3C–3F). We generated new ChIP-Seq data sets for histone modifications H3K4me3, H3K4me1, and H3K27me3, and combined it with existing published data sets for transcription factors BATF and JunD [21], because the motif for AP1 factor (TGAnTCA) was significantly observed in the de novo motif finding results (S2 Fig). We also oriented the analysis centered over the EBF1 (Fig 3C and 3D) or RBP-jκ (Fig 3E and 3F) binding sites in LCLs. Several conclusions could be drawn from these analyses. First, the LCL-specific binding sites (cluster i) for EBF1 colocalize with RBP-jκ and EBNA2, and are highly enriched for H3K4me1, as well as BATF and JunD (Fig 3C and 3D). This suggests that these LCL-specific sites have enhancer-like properties since they are highly enriched in H3K4me1, a well characterized mark of transcriptional enhancers [22]. EBF1 sites that are highly enriched in Mutu I (cluster ii) also colocalize with Mutu I RBP-jκ, but are less enriched for EBNA2, BATF, and JunD, and have reduced H3K4me1 in LCLs, while H3K4me3 remains enriched in Mutu I co-occupied sites. This suggests that cluster ii sites colocalize with transcriptional start sites, and less commonly with enhancer elements. When the analysis was centered on RBP-jκ binding sites (Fig 3E and 3F), we observed strong co-localization of RBP-jκ with EBF1 in both LCL (cluster i) and Mutu I (cluster ii) enriched sites. LCL-specific sites (cluster i) were highly enriched for H3K4me1, BATF, and JunD. Mutu I specific sites (cluster ii) were highly enriched for H3K4me3, suggesting that these sites colocalize at transcription start sites. Taken together, these findings indicate that RBP-jκ and EBF1 have functionally distinct interactions at different locations in a cell-type specific manner, and that EBNA2 strongly colocalizes with enhancer-like histone modification H3K4me1 in LCL.
To investigate the functional significance of the co-occupancy of EBNA2 with EBF1 and RBP-jκ, we analyzed the mRNA expression of LCL and Mutu I cells using Illumina microarrays (Figs 4 and S4). We focused on genes that had transcription start sites within 10 kb of an EBNA2-EBF1-RBP-jκ co-occupied binding site in LCLs (Fig 4A). We identified 165 such genes, and observed that 84 of those genes were significantly up-regulated in LCL compared to Mutu I, an enrichment of 3.2 fold more than expected by random chance (p<10−12 by Fisher exact test, Fig 4B). Ingenuity Pathway analysis of these upregulated genes revealed a network of important B-lymphocyte functions, including activation of lymphocytes, mobilization of calcium, cell viability, and chemotaxis (Fig 4C).
Correlations studies suggest that EBNA2 may facilitate binding and co-occupancy of RBP-jκ and EBF1 in LCLs. To experimentally determine if EBNA2 can alter EBF1 and RBP-jκ chromosome occupancy, we utilized an EBNA2-inducible cell line, EREB2.5, that contains an EBNA2 protein fused to the estrogen receptor (ER) [23]. The ER-EBNA2 fusion in EREB2.5 cells regulates the nuclear-cytoplasmic transport of EBNA2 and therefore has a rapid post-translational effect on EBNA2 function in the nucleus. Upon withdrawal of estrogen and inactivation of EBNA2, we found a rapid loss of binding for EBF1 and RBP-jκ at sites that colocalize with EBNA2 (Fig 5A–5D). As expected, EBNA2 binding was rapidly eliminated from all sites tested (Fig 5E and 5F). EBF1 binding was not disrupted at most sites that were not overlapping with EBNA2 binding sites, while RBP-jκ binding was modestly affected at some of sites that were not overlapping with EBNA2 sites (S5A, S5B and S5C Fig). In the EBV genome, we observed a strong EBNA2-dependence of EBF1 and RBP-jκ binding to LMP2A, LMP1, and C promoter regulatory elements (Fig 5B and 5D). Estradiol withdrawal did not have similar inhibitory effects on binding of CTCF, PU.1, or PAX5 (Fig 5G), suggesting that these effects are selective for the EBNA2 interacting proteins RBP-jκ and EBF1. Western blot analysis indicated that protein levels for EBF1 and RBP-jκ were not significantly altered upon estradiol withdrawal, indicating that loss of protein expression is not the mechanism for the altered binding site selectivity (Fig 5H). Although estradiol withdrawal induced a G1 cell cycle arrest at 48 hrs (S6 Fig), the re-addition of estradiol after 72 hrs withdrawal rescued EBF1 and RBB-jκ binding to their co-occupied cellular target sites (S7 Fig). These findings indicate that EBNA2 is necessary and sufficient to maintain the stable binding of EBF1 and RBP-jκ at co-occupied sites in LCLs.
To explore the biochemical basis for the altered DNA binding properties of EBF1 and RBP-jκ in the presence of EBNA2, we first performed co-IP experiments (S8 Fig). We confirmed that EBNA2 can co-IP with RBP-jκ, as expected, but did not find any evidence for a stable interaction between EBF1 with either RBP-jκ or with EBNA2 (S8 Fig). We next tested whether EBNA2 may facilitate EBF1 and RBP-jκ DNA binding using in vitro DNA affinity assays followed by Western blotting (Fig 6A). We compared extracts derived from EREB2.5 cells treated with or without estradiol for their ability to bind to several different DNA elements, including those with closely overlapping EBF1 and RBP-jκ sites that colocalize with EBNA2 (IL7, HES1, and LMP1) (S9 Fig), as well as negative controls lacking EBNA2 binding sites (Qp and TOM1). We observed that nuclear extracts derived from E2 treated EREB2.5 cells were enriched for EBNA2 protein and bound EBNA2 efficiently in DNA affinity assays. We found that EBF1 and RBP-jκ binding were enhanced in the presence of EBNA2 at the IL7, HES1, and LMP1 DNA elements, but not at Qp or TOM1 DNA controls. The failure to bind TOM1 was surprising, since EBF1 and RBP-jκ can bind in EREB cells in the absence of EBNA2. Experimental conditions were optimized for EBNA2 binding, and these conditions may limit detection of weaker interactions that occur at sites that lack EBNA2 colocalization. The non-specific DNA binding protein PARP1 bound independently of ER treatment, while the non-DNA binding protein Actin did not bind DNA templates, providing some control for specificity. These findings suggest that EBNA2 protein can partially facilitate EBF1 and RBP-jκ binding at some DNA elements, but that other factors, perhaps chromatin structure and epigenetic modifications, must contribute substantially to the cell-type selective binding observed in living cells.
To determine if EBF1 and RBP-jκ bind simultaneously to the same DNA elements, we performed ChIP-reChIP assays (Fig 6B and 6C). We compared LMP1 and IL7 promoter regions that have strong overlapping co-occupied sites for EBF1, RBP-jκ and EBNA2. We found that ChIP-reChIP of EBF1 primary ChIPs with either EBF1, RBP-jκ, or EBNA2 resulted in significant enrichment at LMP1 and IL7, suggesting that these proteins form complexes on the same DNA elements. The BDH2 gene has an EBF1 site that is not co-occupied by RBP-jκ or EBNA2 in ChIP-Seq data, and showed no significant enrichment when subject to reChIP (Fig 6B, right panel). Similarly, when RBP-jκ was used as the primary ChIP, reChIP with either EBF1, RBP-jκ, or EBNA2 showed significant enrichment at the LMP1 and IL7 promoters, but not at the LZTFL1 gene which has an RBP-jκ site that lacks EBF1 or EBNA2 co-occupancy in ChIP-Seq data sets (Fig 6C). These findings suggest that EBNA2, RBP-jκ, and EBF1 can form a stable complex simultaneously on the same DNA elements that where they colocalize in ChIP-Seq experiments.
To determine whether EBF1 or RBP-jκ may contribute to transcription regulation at some of these co-occupied sites, we used shRNA depletion of either EBF1 or RBP-jκ in LCL cells (Fig 7). Both EBF1 and RBP-jκ shRNA led to an efficient depletion of their respective targets, as determined by Western blot (Fig 7A). Interestingly, EBF1 depletion led to a loss of LMP1 protein expression from LCLs, but no loss of other type III proteins, including EBNA2 or EBNA3C. In contrast, RBP-jκ depletion led to a loss of EBNA3C, but no loss EBNA2 or LMP1. This suggests that EBF1 and RBP-jκ have non-redundant functions in the transcription regulation of EBV type III gene expression. Both LMP1 and EBNA3C are essential for LCL viability, and as expected, both EBF1 and RBP-jκ depletion decreased LCL cell viability (Fig 7B). RT-qPCR analysis confirmed that EBF1 shRNA depletion led to a loss of LMP1 transcription, while RBP-jκ led to loss of EBNA3C (Fig 7C). RT-PCR also revealed several other changes in EBV gene expression in response to EBF1 or RBP-jκ depletion, including a modest decrease in EBNA2 expression. Examination of cellular gene mRNA revealed some selective sensitivity to either EBF1 or RBP-jκ depletion. IL7, like LMP1, was more sensitive to EBF1 depletion, while HES1 was more sensitive to RBP-jκ depletion. FCER2, like EBNA2, was similarly affected by either EBF1 or RBP-jκ (Fig 7C). ChIP assays revealed a strong loss of EBF1 from shEBF1 depleted cells (Fig 7D and 7E), and RBP-jκ from RBP-jκ depleted cells (Fig 7F and 7G) at all viral and cellular binding sites tested. Interestingly, in shEBF1 depleted cells, RBP-jκ binding was reduced at the LMP1 promoter (Fig 7H) and at the IL7 and FCER2 cellular promoters (Fig 7I). On the other hand, in shRBP-jκ depleted cells, EBF1 binding was reduced at the LMP2A, LMP1, and Cp viral promoters (Fig 7J) and HES1 cellular promoter (Fig 7K). These findings suggest that some, but not all EBF1 and RBP-jκ binding sites are interdependent. Finally, we examined EBNA2 binding in shEBF1 or shRBP-jκ depleted cells (Fig 7L and 7M). We found that EBNA2 binding was reduced at LMP2A, LMP1, and Cp viral promoters, and IL7 and FCER2 cellular promoters in shEBF1 depleted cells. On the other hand, in shRBP-jκ depleted cells, we found that EBNA2 binding was reduced at LMP2A, Cp, and cellular HES1. These findings suggest that EBNA2 may be recruited to some sites by either RBP-jκ or EBF1, or both. Taken together, these findings indicate that EBF1 and RBP-jκ have non-redundant, cooperative and context-dependent functions at viral and cellular promoters where they colocalize with EBNA2.
A fundamental unresolved question in eukaryotic gene regulation is how transcription factors find their appropriate cell-type specific binding sites in the context of chromatin and other epigenetic parameters. Pioneering factors, like EBF1, are thought to have the intrinsic capacity to access their chromatin-embedded binding sites to form new promoter and enhancers elements necessary for cell-specific gene transcription [15]. While this is well-established for many genomic locations, our findings provide evidence that a significant number of transcription factor binding sites require cooperative interactions with neighboring factors and accessory co-factors. Our findings indicate that transcriptional cofactors, like EBNA2, can function as pioneer factor ‘scouts’ that reorient chromosome occupancy for client pioneer proteins, like RBP-jκ and EBF1.
EBV genomes adopt different gene expression programs that depend on host cell or tumor type, as well as on epigenetic factors [24, 25]. EBV encodes several transcription cofactors, in addition to EBNA2, that remodel viral and host cell gene programs to confer the EBV-induced proliferation and immortalization to resting B-cells. EBNA2 binds to several sequence specific transcription factors, like RBP-jκ and PU.1, and also interacts with epigenetic co-activators, like histone acetyltransferases CBP/p300 and chromatin remodeling factors containing SNF2 [6]. The prevailing model posits that EBNA2 binds to pre-established sequence-specific binding factors bound to their cognate binding sites. EBNA2 is thought to displace co-repressors bound to these sequence specific factors and then form a stable co-activator complex at selected promoters and enhancers to stimulate target gene transcription. While constitutive DNA binding occurs at some chromosomal sites, our findings indicate that EBNA2 can also drive the formation of new chromosome occupancies for factors like RBP-jκ and EBF1. These newly formed sites are commonly co-occupied by RBP-jκ, EBF1 and EBNA2, and correlate with activated chromatin and transcription function (Fig 8).
A recent study concluded that EBNA2 does not fundamentally alter the binding patterns of RBP-jκ or other sequence specific binding factors [18, 26]. These studies compared LCL to resting B-lymphocytes (RBLs) which do not express EBNA2. Upon EBV infection, EBNA2 was found to colocalize only to pre-existing RBP-jκ binding sites, and only increased RBP-jκ binding peak intensity. In this study, EBNA2 was found to function largely through the formation of enhancer-promoter DNA-loop interactions [18]. Our study differs from these previous reports in several important respects. We compare established LCLs to type I BL cells, as well as established LCLs with a conditional EBNA2-ER fusion protein. Our findings indicate that a large percentage of RBP-jκ and EBF1 binding sites are detected only in the presence of EBNA2. It is likely that some of the RBP-jκ and EBF1 sites exist at low levels in the absence of EBNA2 expression in resting B-lymphocytes, and that threshold cutoffs for ChIP may explain some of these differences in interpretation of data. We also found that some non-colocalized EBF1 and RBP-jκ sites are stable in the absence of EBNA2. However, we find numerous EBF1 and RBP-jκ colocalized sites that form selectively and exclusively in the presence of EBNA2. This suggests that the EBNA2 increases the stability and probability of RBP-jκ and EBF1 binding to a specific group of genomic sites and regulatory elements.
Recent genome-wide studies reveal that these EBV-encoded transcriptional modulators have complex and disparate functions on the cellular genome [12, 18–20, 27]. ChIP-Seq analysis of EBNA2 in a type III BL cells (Mutu III) revealed a high-colocalization with the predicted binding sites of RBP-jκ and EBF1 [19]. Our experimental data confirms this computational prediction in type III LCLs. Additionally, EBNA2 has been implicated in both cooperative and competitive interactions with other viral proteins. For example, EBNA3A can compete with EBNA2 at RBP-jκ sites that regulate CXCL10 and CXCL9 [28]. EBNA-LP colocalizes with ~ 29% of EBNA2 sites, many of which were enriched for EBF1 and RBP-jκ [12]. EBNALP is proposed to remove the co-repressor NCoR from RBP-jκ to induce transcription activation at RBP-jκ bound sites. EBNA3A and 3C were shown to colocalize preferentially with BATF and IRF4, and to a lesser extent with RBP-jκ [20, 27]. Additional levels of complexity may also arise from amino acid variations in EBNA2 proteins derived from different EBV strains. Variants of EBNA2 that more efficiently immortalize B-lymphocytes were found to selectively regulate genes with composite motifs for ETS-interferon regulatory factor (IRF) binding motifs [29]. More recently, EBNA3 proteins have been shown to associate with ubiquitin modifying enzymes that could further alter the stability and functionality of interaction partners [30]. These findings highlight the complex and significant role of non-DNA binding co-factors in altering transcription factor interactions, chromatin occupancy, and function.
Context-dependent binding of RBP-jκ in response to factors like EBNA2 and Notch have been predicted based on meta-data analysis of several reports of various EBNA2 and Notch binding patterns [31]. Other studies have found that RBP-jκ binds dynamically in response to Notch signaling [32–34]. In T-lymphoblastic leukemia cells, RBP-jκ binding was enhanced by the conditional expression of Notch 1 [33]. These Notch 1 induced RBP-jκ sites colocalized with histone H3K4me1 and H3K27ac, suggesting these sites function as enhancer regulatory elements. RBP-jκ chromosome occupancy was also enhanced by Notch1 expression in mouse myogenic cells [34]. The modulatory effects of Notch 1 on RBP-jκ binding site selection are similar to what we observe with EBNA2 induced binding of RBP-jκ. EBNA2 does not have identical activity to Notch, and studies suggest that Notch 2 expression can antagonize EBNA2 growth proliferative activity in EBV immortalized B-lymphocytes [35]. Thus, while EBNA2 and Notch share the ability to bind and modulate RBP-jκ, they are likely to have different patterns of chromosome binding sites and gene activation profiles.
EBF1 has been referred to as a ‘pioneer’ factor due to its ability to initiate an early B-cell gene program and activate genes that are otherwise repressed by chromatin [36]. EBF1 binds and regulates both constitutively active and inducible gene targets during stage-specific B-cell differentiation [37]. However, it is also reported that EBF1 requires priming factors to bind to its chromosomal sites, as it is unable to bind B-cell specific sites in non-hematopoietic cells [13, 36]. EBF1 has been shown to interact with several proteins that could facilitate its pioneering activity, including epigenetic modifiers like TET2, which can promote DNA demethylation [38] and BRG1, which can facilitate nucleosome remodeling [39]. Thus, cellular pioneering factors, like EBF1, are likely to require co-factors to facilitate access to epigenetically repressed sites.
Pioneering co-factors, like EBNA2, may function through several mechanisms Recruiting epigenetic modifiers and chromatin remodelers have been demonstrated at several promoter regulatory elements. But perhaps equally significant, are the combinatorial interactions of transcription factors and co-factors that determine the lineage- and cell-specific enhancer priming [40]. EBNA2 may facilitate combinatorial and cooperative interactions to initiate formation of new enhancer and promoter bound transcription factor complexes. EBNA2 may function as a scaffold that stabilizes multiple protein interactions, including cooperative binding between RBP-jκ and EBF1 at some genomic locations. Our DNA affinity binding assays provide partial support for this model (Fig 6). However, we did not observe a stable interaction of EBF1 with either EBNA2 or RBP-jκ in co-IPs, suggesting that the cooperativity occurs at a functional level distinct form physical association. We also observed gene-specific sensitivities to the depletion of either RBP-jκ or EBF1, suggesting that their functions in transcription regulation are non-redundant (Fig 7). Furthermore, we observed some interdependent chromosome occupancy of these factors, which also appeared gene specific (Fig 7). EBNA2 is also known to recruit histone acetyltransferases p300/CBP and chromatin remodeling factors that can further facilitate transcription factor binding in the context of chromatin. Taken together, these findings suggest that factors like EBNA2, RBP-jκ, and EBF1 can functionally interact in a context-dependent manner to stabilize their coordinated chromatin occupancy and transcription enhancement.
An important biological outcome of EBNA2 function is the overall skewing of binding patterns for the overlapping targets of EBF1 and RBP-jκ. In particular, EBNA2 expression correlated with an increase in the total number and peak intensity of RBP-jκ bound sites (Fig 2). This observation from ChIP-Seq analysis is also consistent with the increase steady state protein expression levels for RBP-jκ in EBNA2 positive cells (Fig 1D). These trends may reflect the role of EBNA2 in driving resting B-cells into proliferating and activated B-lymphoblasts. RBP-jκ is thought to have a major role in T-cell development, while having a more nuanced role in B-cell development, like promoting marginal zone B-lymphocyte differentiation [41]. Our findings suggest that EBV type III latency promotes a B-cell marginal zone developmental program, as the number of RBP-jκ binding sites is largely increased.
In conclusion, our findings demonstrate that the EBV-encoded transcriptional co-activator EBNA2 reprograms the chromosome binding patterns of host master regulatory transcription factors RBP-jκ and EBF1. We propose that EBNA2 drives new combinatorial interactions on DNA at promoter and enhancer regulatory elements of genes essential for EBV lymphoblast proliferation and survival.
The LCL cell line was generated from human B-cells (provided by the Wistar Institute Phlebotomy Core) transformed with Mutu I virus. EBV positive BL cell lines Mutu I, Kem I, and Kem III were obtained from Dr. Jeffrey Sample, Penn State University Hershey Medical School, PA. LCLs, Mutu I, Kem I and Kem III cells, were maintained in RPMI containing 12% FBS and antibiotics (penicillin and streptomycin). EREB 2.5, a lymphoblastoid cell line expressing the estrogen-inducible EBNA2-estrogen receptor (ER) fusion protein [23], was maintained in RPMI containing 12% FBS, antibiotics (penicillin and streptomycin), and 1μM estradiol.
ChIP-Seq experiments were performed with 5 x 107 cells per assay with 10 μg of either rabbit anti- EBF1 (EMD Millipore AB10523), RBP-jκ Abcam AB25949), histone H3K4me3 (EMD Millipore 07–473), H3K4me1 (Active Motif 39297), H3K27me3 (Active Motif 39155) antibodies or control rabbit IgG (Santa Cruz Biotechnology sc-2027). ChIP was performed by as described previously with some modifications [42]. Briefly, crosslinked Mutu I or LCL lysates were sonicated to achieve a DNA fragment length of ~100–500 bp, incubated overnight with antibody-coated Dynabeads protein A/G, then washed with ChIP-seq wash buffer (50 mM HEPES, pH 7.5, 500 mM LiCl, 1 mM EDTA, 1% NP-40, 0.7% Na-Deoxycholate, 1x protease inhibitors) for 5 times, then washed once with 50 mM NaCl in TE buffer. Immunoprecipitated DNA was eluted with ChIP-seq elution buffer (50 mM Tris-HCl, pH 8, 10 mM EDTA, 1% SDS), reverse-crosslinked at 65°C, treated with RNase A (0.2 mg/ml) and proteinase K (0.2 mg/ml), purified with phenol and chloroform, then subjected to qPCR validation. Validated ChIP DNA was isolated by agarose gel purification, ligated to primers, and then subject to Illumina-based sequencing using manufacturers recommendations (Illumina).
ChIP-qPCR assays were performed as described previously [43]. Quantification of precipitated DNA was determined using real-time PCR and the delta Ct method for relative quantitation (ABI 7900HT Fast Real-Time PCR System). Rabbit IgG, EBF1 and RBP-jκ antibodies used for conventional ChIP assays were same as antibodies used for ChIP-seq. Rabbit CTCF (EMD Millipore 07–729), PU.1 (EMD Millipore 04–1072), PAX5 (Santa Cruz Biotechnology sc-1974x), mouse IgG (Santa Cruz Biotechnology sc-2025), mouse EBNA2 (gift from Paul Farrell, UK) antibodies were also used in ChIP assays. Primers for ChIP assays are listed in Supplement Tables S1 and S2.
ChIP-re-ChIP assays were performed as described [44]. Mouse EBNA2 (gift from Paul Farrell, UK), rabbit EBF1 or RBP-jκ antibodies used in ChIP-reChIP were same as antibodies used in conventional ChIP assays.
RNA was isolated from 2 x 106 cells using RNeasy Kit (Qiagen) and then further treated with DNase I by using DNase treatment and removal kit (Ambion). RT-PCR was performed as previously described [45]. Real-time PCR was performed with SYBR green probe in an ABI Prism 7900 and the delta Ct method for relative quantitation. Primer sequences for RT-PCR are available upon request.
DNA-affinity purification was performed as previously described [46]. Soluble nuclear extract fractions were obtained from EREB cells cultured with or without 1μM estradiol for 24 hours. The ~400 bp sequence from the center of EBF1/RPB-jκ peaks at IL7, HES1, or LMP1 loci were cloned into pCR-blunt II-TOPO vector (Invitrogene). The ~500 bp sequence from EBV Qp region (EBV 49712–50250) was cloned into pBSKII. The ~ 400 bp sequence from the center of EBF1/RBPjκ peaks at TOM1 was cloned into pUCIDT vector (IDT). DNA fragments were PCR amplified from their respective plasmid templates using a biotinylated forward primer. The biotinylated PCR fragments were coupled to M-280 streptavidin Dynabeads. The coupled beads were washed and incubated with soluble nuclear extract for 1 hour. The bound proteins and beads were washed three times with D100 buffer (20 mM HEPES, pH 7.9, 20% glycerol, 0.2 mM EDTA, 0.05% NP-40, 100 mM KCl, 1 mM phenylmethylsulfonyl fluoride, 1 mM DTT, and 1x protease inhibitors). The bound proteins were eluted from the beads using 2 x Laemmli buffer and heated at 95°C for 10 minutes, loaded on an 8–16% SDS-PAGE, then analyzed by Western blotting.
Rabbit polyclonal anti-EBF1 (EMD Millipore AB10523), RBP- jκ (Abcam AB25949), CTCF (EMD Millipore 07–729), PARP-1 (Alexis 210-302-R100), GAPDH (Cell signaling 2118); Goat polyclonal anti-EBF1 (R&D System AF5165); Sheep polyclonal anti-EBNA3C (Exalpha F125P); Rat monoclonal anti-EBNA2 (EMD Millipore MABE8); Mouse monoclonal anti-LMP1 (DAKO M0897); Actin-Peroxidase antibody (Sigma A3854).
pLKO.1 vector-based shRNA construct for EBF1 (TRCN0000013831) or RBPjκ (TRCN0000016203) was obtained from Open Biosystems. shControl were generated in pLKO.1 vector with target sequence 5′-TTATCGCGCATATCACGCG-3′. Lentiviruses were produced by the use of envelope and packaging vectors pMD2.G and pSPAX2 as described previously [47]. Mutu I or LCL cells were infected with lentiviruses carrying pLKO.1-puro vectors by spin-infection at 450 g for 90 minutes at room temperature. The cell pellets were resuspended and incubated in fresh RPMI medium, then treated with 2.5 μg/ml puromycin at 48 hr after the infection. The RPMI medium with 2.5 μg/ml puromycin was replaced every 2 to 3 days. The cells were collected after 7 days of puromycin selection, then subject to Western blotting, RT-PCR, ChIP and cell viability assays.
48 hours after lentivirus infection, 96-well assay plate was set up with 104 cells in 100 μL complete RPMI medium with 2.5 μg/ml puromycin in each well. The transduced cells were cultured for 5 days then 10 μl of 0.5 mM resazurin (Sigma) solution were added to each well. The plate was incubated in cell culture incubator for 3–4 hours then read for fluorescence at 560/590 nm. The relative cell viability of shEBF1 transduced cells was presented as a percentage relative to shCtrl.
All ChIP-seq tags for EBNA2, EBF1, RBP-jκ, histone H3K4me3, H3K4me1, H3K27me3, BATF, and JunD were aligned to the human genome hg19 using Bowtie [48] with options ‘-v 2 -m 1 –best–strata’ and all of the redundant tags were removed before downstream analysis. All ChIP-seq data to the human genome were normalized to 10 reads per million mapped reads (RPM). From ENCODE [49], we used ChIP-seq data for BATF (GSM803538) and JunD (GSM754331) in GM12878. For comparison we used ChIP-seq data for EBNA2 in LCL (GSE29498) and Mutu III (GSE47629, sample ID GSM1153765). We also used EBNA3C (GSE52632), and EBNA-LP (GSE49338) in LCL and EBNA3C in Mutu III (GSE47629, sample ID GSM1153766). Peak calling was performed using the findPeaks command in Homer [40]. After initial calling, all of the peaks were resized to 200 bp; the 1 reads per million (RPM) cutoff was applied for EBNA2, EBF1, RBP-jκ to select strong peaks. Specific peaks were defined as having at least a four-fold difference in enrichment within a 200 bp region between the two cell populations. The remaining peaks were defined as common peaks. To compare the ChIP occupancy at cell-type specific peaks, paired t-test was applied. De novo motif finding on the EBNA2 peaks was performed using the findMotifsGenome command in Homer [40]. In the visualization of heatmaps, normalized reads were obtained around the peaks (+/- 4kbp) with a 20bp resolution.
RNA samples extracted from three independently cultured Mutu I or LCL cells then further treated with DNase I by using DNase treatment and removal kit (Ambion). RNA quality was determined using the Bioanalyzer (Agilent). Only samples with RIN numbers >7.5 were used for further studies. Equal amounts (400ng) of total RNA was amplified as recommended by Illumina and hybridized to the HumanHT-12 v4 human whole genome bead arrays. Illumina BeadStudio v.3.0 software was used to export expression levels and detection p-values for each probe of each sample. Arrays were quantile-normalized and filtered to remove probes with a detection p-value>0.05 in all samples. Expression level comparisons between the two cell lines were done using two sample t-test and correction for multiple testing to estimate False Discovery Rate (FDR) was performed [50]. FDR<5% genes were considered significant unless stated otherwise. Gene set enrichment analyses were performed using QIAGEN’s Ingenuity Pathway Analysis software (IPA, QIAGEN Redwood City,www.qiagen.com/ingenuity) and only enrichments from “Diseases & Functions” results that passed p<0.001 and Z-score>2 for predicted function activation were reported. Additionally, genes uniquely occupied by EBF1/RBP-jκ/EBNA2 in LCL within 10kb from TSS and significantly over-expressed vs Mutu I at least 1.2 fold were analyzed using IPA’s “Regulator Effect” analysis option to generate a network of predicted affected cellular functions. Differences in proportions of up vs down-regulated or up-regulated vs unchanged genes between classes of occupied genes were tested using Fisher Exact Test. Enrichment values for heatmap with different fold change/distance to TSS windows were calculated using genes significantly different between LCL and Mutu I at nominal p<0.05. Enrichment p-values were estimated by Fisher Exact Test and adjusted for multiple testing with Bonferroni correction. Results with adjusted p<0.05 were considered significant. Windows for any TSS distance X were log10-scaled by using formula [10^(log10X-0.5) to 10^(log10X+0.5)], e.g for 10kb from TSS, a window of [3.16kb to 31.6kb] was generated. TSS distance X for the plot varied from log10X = 2 to log10X = 5.2 using step of 0.1.
|
10.1371/journal.pbio.2002909 | The rostromedial tegmental nucleus is essential for non-rapid eye movement sleep | The rostromedial tegmental nucleus (RMTg), also called the GABAergic tail of the ventral tegmental area, projects to the midbrain dopaminergic system, dorsal raphe nucleus, locus coeruleus, and other regions. Whether the RMTg is involved in sleep–wake regulation is unknown. In the present study, pharmacogenetic activation of rat RMTg neurons promoted non-rapid eye movement (NREM) sleep with increased slow-wave activity (SWA). Conversely, rats after neurotoxic lesions of 8 or 16 days showed decreased NREM sleep with reduced SWA at lights on. The reduced SWA persisted at least 25 days after lesions. Similarly, pharmacological and pharmacogenetic inactivation of rat RMTg neurons decreased NREM sleep. Electrophysiological experiments combined with optogenetics showed a direct inhibitory connection between the terminals of RMTg neurons and midbrain dopaminergic neurons. The bidirectional effects of the RMTg on the sleep–wake cycle were mimicked by the modulation of ventral tegmental area (VTA)/substantia nigra compacta (SNc) dopaminergic neuronal activity using a pharmacogenetic approach. Furthermore, during the 2-hour recovery period following 6-hour sleep deprivation, the amount of NREM sleep in both the lesion and control rats was significantly increased compared with baseline levels; however, only the control rats showed a significant increase in SWA compared with baseline levels. Collectively, our findings reveal an essential role of the RMTg in the promotion of NREM sleep and homeostatic regulation.
| Sleep–wake behavior is controlled by networks of neurons and neurotransmitters in the brain. There are multiple populations of wake-promoting neurons, but few sleep-promoting neurons have been identified. In this study, we revealed that the rostromedial tegmental nucleus, the GABAergic tail of the ventral tegmental area, regulates non-rapid eye movement sleep. We show that neurons in the rat rostromedial tegmental nucleus, when activated by pharmacogenetics, increase and deepen non-rapid eye movement sleep. Inhibition of these neurons exhibits the opposite effects. Furthermore, rats with lesion in the rostromedial tegmental nucleus have a reduced response of sleep homeostasis following sleep deprivation. We show that stimulation of the terminals of the neurons in the rostromedial tegmental nucleus inhibits dopaminergic neurons in the midbrain. Interestingly, inhibition of these dopaminergic neurons also has sleep-promoting effects. The current results provide a potential target for prolonging non-rapid eye movement sleep, improving sleep quality, and treating sleep disorders in dopamine-implicated mental illness.
| Dopamine (DA) produced by neurons in the midbrain plays a key role in processing reward, aversive, and cognitive signals [1]. Abnormal DA is closely associated with neuropsychiatric disorders such as Parkinson disease, schizophrenia, and substance abuse. Severe sleep disturbances have been observed in nearly all of these types of patients [2–5]. Growing evidence suggests that DA-containing neurons are important for arousal maintenance in both humans [6, 7] and animals [8–11]. Moreover, it has been found that activation of ventral tegmental area (VTA) γ-amino-butyric acid (GABA) neurons, which indirectly inhibits the firing of VTA DAergic neurons, is sufficient to elicit a place aversion [12, 13]. This suggests that the ability of midbrain DAergic neurons to regulate sleep–wake behavior and sleep problems in DA-associated mental illnesses may be affected by upstream inhibitory neuronal systems.
The rostromedial tegmental nucleus (RMTg) is a newly identified structure in the brainstem that is rich in μ-opioid receptors. It primarily comprises GABAergic neurons that are distributed dorsolateral to the interpeduncular nucleus (IPN). The RMTg is strikingly innervated by the afferent input from the lateral habenula and additional inputs from the extended amygdala and other closely connected regions, such as the lateral septum and periaqueductal gray matter. The GABAergic axons from the RMTg densely project to midbrain DAergic neurons [14–18]. The RMTg acts as a hub converging and integrating widespread signals toward DAergic systems [19]. Neuroanatomical and electron microscopic studies have found that most RMTg axons form symmetric synapses with tyrosine hydroxylase (TH)-containing dendrites in the substantia nigra compacta (SNc) and VTA [20, 21]. An in vivo electrophysiological study showed that the RMTg exerted greater inhibition of the SNc DAergic neurons than the inhibitory afferents arising from the striatum, globus pallidus, or substantia nigra parsreticulata [20, 22, 23]. Likewise, GABAergic RMTg neurons inhibited the activity of VTA DAergic cells by inhibiting synaptic transmission more effectively than intermediate GABAergic neurons in the VTA [21, 24, 25]. DAergic cells are controlled by excitatory and inhibitory inputs whose balance finely tunes cell activity [26]. The RMTg is now recognized as a GABA brake for midbrain DAergic systems [19].
Aside from the heavy output to the midbrain DAergic neurons, the RMTg sends projections to the dorsal raphe nucleus (DRN), pedunculopontine tegmental and laterodorsal tegmental nuclei (PPT, LDT), and locus coeruleus (LC) and has relatively meager output to the forebrain, including the lateral hypothalamus and lateral preoptic area [14, 15]. Generally, neurons in these cell groups fire most actively during wakefulness [7].
Although the RMTg has been confirmed to inhibit the electrical activity of midbrain DAergic neurons associated with wakefulness activation, whether it is implicated in sleep–wake behavior is unknown. Considering that RMTg neurons are prominently GABAergic and are thus speculated to inhibit rather than facilitate the activity of targeted neurons, we propose that the RMTg is involved in promoting sleep. To test this hypothesis, we employed pharmacogenetics using designer receptors exclusively activated by designer drugs (DREADDs) [27] and pharmacological approaches to manipulate neuronal activity, neurochemistry, electrophysiology, and immunohistochemistry along with optogenetics and transgenic mice to investigate whether the RMTg plays a role in the regulation of sleep and homeostasis. We then explored whether the RMTg nucleus controls sleep through the modulation of DAergic neuron activity.
To test the effects of RMTg neuron activation on sleep and waking regulation, adeno-associated virus (AAV) vectors containing excitatory modified muscarinic G protein-coupled receptors (hM3Dq) (Fig 1A) were bilaterally microinjected into the rat RMTg (Fig 1B). Cell-surface expression of hM3Dq receptors was observed via red fluorescent mCherry protein. Confocal images of double labeling with mCherry and GABA immunofluorescence showed that 87% of mCherry-positive neurons in the RMTg region coexpressed GABA (245 of 280), which indicated GABAergic neurons in the RMTg were targeted. Moreover, we found 56% of GABA-positive neurons colabeled hM3Dq/mCherry (245 of 437) (Fig 1C and 1D), indicating that the virus was efficient for RMTg neurons.
Immunohistochemistry showed that clozapine-N-oxide (CNO, 0.3 mg/kg), a specific hM3Dq agonist (Fig 1E), but not saline (Fig 1F), could drive c-Fos expression in hM3Dq-expressing neurons in the RMTg. Confocal image of double labeling with mCherry and c-Fos immunofluorescence showed that the 56% of mCherry-positive neurons in the RMTg region expressed c-Fos in the CNO group compared to 2% in the saline group (Fig 1G). In addition, bath application of CNO (500 nM) depolarized the RMTg hM3Dq-expressing neurons and significantly increased the firing of action potentials in hM3Dq/mCherry-positive neurons, as indicated by whole-cell current clamp recordings (Fig 1H–1K). Thus, the DREADD system used in this study stimulates the activity of rat RMTg neurons both in vivo and in vitro.
On average, the CNO-injected rats showed a 32.2% increase in non-rapid eye movement (NREM) sleep and reductions of 84.6% and 34.8% in rapid eye movement (REM) sleep and wakefulness, respectively, during the 7-hour post-CNO injection period (Fig 2A and 2B). The number of stage conversions from wakefulness to NREM sleep (Fig 2C) and the total NREM sleep episodes (Fig 2D) did not change, even though CNO induced fewer NREM sleep bouts of 1–2 minutes. However, the number of prolonged NREM sleep episodes with durations between 4–16 minutes was significantly increased (Fig 2E), resulting in a longer mean duration of NREM sleep (Fig 2F). CNO administration promoted NREM sleep at the expense of REM sleep, significantly reducing REM sleep bouts of all durations (Fig 2G). Electroencephalogram (EEG) power spectrum analysis revealed that during the 7 hours after CNO injection, the average slow-wave activity (SWA)—a commonly used quantitative measure of sleep intensity, indicated by an EEG power between 0.5 and 4 Hz within NREM sleep [28–30]—in CNO-injected rats increased by 11.0% compared with the saline-injected controls. In contrast, the CNO-injected rats displayed a 9.6% decrease in the power density of REM sleep during the theta band range of 6–10 Hz (Fig 2H and 2I).
The above results indicate that activation of RMTg neurons played an important role in maintaining NREM sleep in rats.
In the saline group, a total of 8 c-Fos+ neurons were found in the RMTg of rats infected with viral vectors encoding hM3Dq, indicating that RMTg neurons were not remarkably activated during the spontaneous sleep–wake cycle. When we analyzed distribution of the c-Fos+ cells in the CNO group, we found that 78% of them expressed hM3Dq/mCherry (151 of 194) and that the other 22% were mCherry− (43 of 194). These c-Fos+/mCherry− neurons may possibly be activated by CNO-induced sleep. Since the regulation of cell activity is very complicated, the increased RMTg c-Fos expression induced by sleep promotion must be a comprehensive result, in which many neural circuits are involved.
To explore the functional nature of the RMTg-to-midbrain DAergic connections, we employed an optogenetic-assisted circuit mapping approach. Channelrhodopsin-2 (ChR2), a blue light-gated cation channel, was expressed in RMTg neurons by injecting AAV-ChR2-mCherry into the RMTg of Sprague–Dawley rats. After 3 weeks, acute coronal brain slices containing the RMTg or VTA/SNc were prepared for in vitro patch-clamp recording (Fig 3A). The expression of ChR2 allowed the activation of cell bodies within the RMTg and the selective stimulation of terminals from the RMTg that projected to the VTA and SNc.
First, we found that there were dense mCherry+ terminals of RMTg neurons in VTA and SNc, which showed anatomical connections between RMTg neurons and midbrain DAergic neurons (Fig 3B). We then tested the responses of ChR2-expressing neurons within the RMTg to optogenetic stimulation. Blue light pulses at 20 Hz evoked action potentials with high fidelity and elicited robust photocurrents under voltage mode (Fig 3C). Next, cells in the VTA and the SNc were patch clamped while blue light flashes were used to stimulate the axon terminals of RMTg neurons. To identify the cell types of recorded midbrain neurons, we added biocytin to the pipette solution and performed immunostaining using TH as a marker for DAergic neurons after recording. We found that light-evoked inhibition could be recorded in TH-positive neurons within the VTA and the SNc (Fig 3D and 3E), and the connected neurons were distributed throughout the rostrocaudal extent of the VTA/SNc (Fig 3F). In the cell-attached patch mode, photostimulation (5-millisecond pulses, 20 Hz) of RMTg terminals in the VTA or the SNc was sufficient to decrease the firing rate; in some cases, light application totally inhibited the spikes of midbrain neurons, and the firing rate recovered immediately upon the termination of photostimulation (Fig 3G and 3J). In the whole-cell voltage-clamp mode, light evoked fast inhibitory postsynaptic currents (IPSCs) in VTA and SNc TH-positive neurons with latencies less than 5 milliseconds in both cases (Fig 3H and 3K), indicating a direct inhibitory connection between the terminals of RMTg neurons and midbrain TH-positive neurons. Moreover, the light-evoked IPSCs were completely abolished by 100 μM picrotoxin (PTX, a GABAAR antagonist; Fig 3I and 3L), indicating that these responses were mediated by GABA released from axon terminals of RMTg neurons.
The present study confirmed that midbrain DAergic neurons receive direct inhibitory innervation from RMTg neurons, which is now known to be the predominant GABAergic control for midbrain DAergic neuron activity. Furthermore, functional studies have revealed that the roles of the RMTg are mediated by the modulation of DAergic neurons. Thus, we wondered whether sleep control by the RMTg also occurs through the modulation of DAergic neuron activity.
To specifically manipulate DAergic neuron activity, Cre-dependent AAVs (Fig 4A) were microinjected into the VTA or SNc areas (Fig 4E) of TH-Cre mice to express the sequence of hM3Dq or hM4Di, which was activated by CNO treatment; as a result, the DAergic neurons were reversibly excited or inhibited, respectively. To detect whether AAVs could be expressed specifically in DAergic neurons rather than in other kinds of neurons, hM4Di-expressing AAVs combined with red fluorescent protein mCherry were microinjected into the VTA of TH-Cre mice. Immunofluorescent staining of colocalization (yellow) of mCherry (red), DAPI (blue), and TH (green) showed that 56% of TH-expressing neurons in the VTA coexpressed hM4Di/mCherry (170 of 302) and that 79% of hM4Di/mCherry-positive neurons were colabeled with TH (170 of 214) (Fig 4B–4D), which validated the efficiency and specificity of this targeting strategy for midbrain DAergic neurons.
To test the brain states in which the DAergic neurons were inhibited, we microinjected Cre-inducible AAVs expressing hM4Di fused with red fluorescent protein into the target of TH-Cre mice. The microinjection sites and AAV-hM4Di-infected areas were confirmed by mCherry expression in the VTA (Fig 4F) and SNc (Fig 4H). Intraperitoneal injection of both CNO and saline did not induce c-Fos expression in hM4Di-expressing mice (Fig 4G and 4I). Next, we performed in vitro electrophysiological experiments to confirm the inhibitory effects of CNO on hM4Di-expressing neuron activity. The recorded mCherry-expressing neuron displayed firing properties with hyperpolarization-activated cation current (Ih) (Fig 4J, top) that were similar to the properties previously reported for DAergic neurons [31]. Whole-cell current clamp recordings demonstrated that bath application of CNO (500 nM) inhibited the firing rate of a VTA hM4Di-expressing DAergic neuron (Fig 4J, bottom). These results indicated that after AAV-hM4Di microinjection, DAergic neurons were inhibited by CNO in vivo and in vitro.
When CNO was administered at 09:00 hours, the TH-Cre mice microinjected with AAV-hM4Di into the VTA or SNc showed an increase in NREM sleep, a decrease in wakefulness, and no change in REM sleep. Compared with the saline control, when VTA DAergic neurons were inhibited, total NREM sleep at 4 hours following CNO treatment was increased by 25.6%, while wakefulness decreased by 36.9% (Fig 5A and 5B). Similarly, the inhibition of SNc DAergic neurons by CNO administration induced a 3-hour increase of 45.5% in NREM sleep, with a 37.8% decrease in wakefulness (Fig 5C and 5D). The number of prolonged NREM sleep episodes with durations between 8 and 64 minutes showed a tendency of increase, which may have caused the increase of NREM sleep when VTA DAergic neurons were inhibited using the pharmacogenetic approach (Fig 5E). Similarly, inhibition of SNc DAergic neurons produced an increase in the number of NREM sleep episodes lasting between 4 and 16 minutes (Fig 5G). Although the amount of NREM sleep induced by pharmacogenetic activation of the RMTg was mimicked by inhibiting midbrain DAergic neurons, the enhanced quality of NREM sleep with higher SWA levels induced by activation of the RMTg was not demonstrated (Fig 5F and 5H).
In order to investigate whether the RMTg plays a role in physiological sleep promotion, bilateral lesions were formed in the RMTg neurons by microinjection of ibotenic acid, a cell-specific neurochemical toxin, using a glass microelectrode technique. After a recovery period, continuous EEG was performed for 48 hours. We compared lesioned and saline-injected control animals for sleep–wake parameters, including time spent in each sleep–wake state, bout numbers, average bout durations, and SWA. On completion of EEG recordings, the lesion position and extent was immunohistologically confirmed by neuron-specific nuclear-binding protein (NeuN) staining. Compared with the intact neurons in the control rats, the animals with ibotenic acid-induced lesions showed extensive cell loss within the RMTg. The lesion extent of each rat was outlined by drawing the boundary of the NeuN-positive neuron population when the staining images had been scaled by the same proportion as the referenced atlas in the same canvas (Fig 6A and 6B).
At 8 days after lesions, rats showed a 25.2% decrease in NREM sleep and a 47.8% decrease in REM sleep along with a 35.9% increase in wakefulness during the 12-hour period from 04:00 to 16:00 hours (Fig 6C and 6D). Analysis of sleep architecture showed that the average duration of NREM sleep was not shortened, and the REM sleep duration was prolonged in the lesioned rats from 04:00 to 16:00 hours. However, the total number of episodes of NREM and REM sleep was markedly decreased in the lesioned rats (Fig 6E and 6F). The rats with RMTg lesions tended to have fewer NREM sleep bouts in the range of 10–50 seconds than controls, although the difference was not statistically significant, and had significantly fewer NREM sleep bouts between 1–2 minutes and REM sleep bouts between 10–50 seconds than controls (Fig 6G and 6H). The lesioned rats also had fewer conversions from wakefulness to NREM sleep, NREM sleep to REM sleep, and REM sleep to wakefulness (Fig 6I). Therefore, the decreased NREM and REM sleep amount was mainly the result of a reduction in the number of short fragments of NREM sleep (<2 minutes) and REM sleep (<1 minute) in the lesioned rats. Across the 24-hour period, SWA was at the highest level—48.0% ± 2.5% and 38.6% ± 1.6% in control and lesioned animals, respectively—during 07:00–08:00, immediately after light onset. The highest level of SWA in the NREM sleep was significantly lower in the lesioned rats than in the control rats (p < 0.01). Then, SWA gradually decreased until immediately before and after lights off, when it was at the lowest level (Fig 6J).
Similarly, at 16 days after RMTg lesions, the amount of sleep and wakefulness and sleep–wake architecture differed significantly between the lesion and control groups. NREM sleep decreased by 9.4%, and REM sleep decreased by 14.3%, with corresponding increases in wakefulness by 10.7% was observed in the lesioned rats in comparison with the controls over a 24-hour period (see S1A and S1B Fig). The lesioned rats had fewer episodes of NREM and REM sleep and fewer conversions from wakefulness to NREM sleep, NREM sleep to REM sleep, and REM sleep to wakefulness than the control rats (see S1C–S1E Fig). SWA in NREM sleep during the hour immediately after lights turned on was also significantly lower in the lesioned rats (42.1% ± 2.3%) than in the control rats (50.6% ± 2.9%) (see S1F Fig).
The loss of NREM and REM sleep in the RMTg lesion rats was not observed at 25 days after neuron damage (see S2A and S2B Fig). There were no differences in episode bouts, durations, or conversions among NREM and REM sleep and wakefulness between the lesioned and control rats (see S2C–S2E Fig). However, at 25 days after RMTg lesions, the rats still had lower SWA levels in NREM sleep, similar to those at 8 and 16 days. SWA in NREM sleep during the first hour of the light phase (34.9% ± 6.4%) was continually and markedly lower than that of the control rats (48.0% ± 2.0%) (see S2F Fig).
The results suggest that the RMTg is involved in the initiation of NREM sleep, which is compensated by some brain structures at 25 days after damage to RMTg neurons. Furthermore, the RMTg maintains NREM sleep depth, and this effect lasts longer than that on sleep initiation.
Sleep homeostasis is primarily studied through sleep deprivation (SD) experiments [32]. Thus, we performed 6-hour SD from 13:00 to 19:00 hours in control and lesioned rats and compared their rebound sleep to determine the role of the RMTg in homeostatic regulation of sleep. The control rats and rats with 8 days of RMTg lesions exhibited similar responses in their NREM sleep during the subsequent recovery period following SD (Fig 7A and 7B). We calculated the total time spent in NREM sleep for 2 hours after SD because the increase in NREM sleep was maintained for 2 hours during the recovery period after 6-hour SD. From 19:00 to 21:00 hours, there was no difference between the control and lesioned rats in the baseline level of NREM sleep, which increased by 57.6% and 128.2%, respectively, after 6-hour SD (Fig 7C). In both groups, the increase in NREM sleep was mainly due to prolonged bout duration (Fig 7D).
To compare the rebound of SWA within NREM sleep after 6-hour SD in the control and lesioned groups, the SWA data were analyzed using 2-way ANOVA followed by paired t tests. The analysis revealed a significant increase in SWA in the control rats during the first 7 hours (from 19:00 to 02:00 hours) following 6-hour SD (p = 0.041) compared with the baseline level. In contrast, 6-hour SD induced no increase in SWA in the lesioned rats even during the first 2 hours (from 19:00 to 21:00 hours) immediately after SD (p = 0.394) compared with the baseline level. There was no change from baseline in the mean SWA in control (33.1% ± 2.2%) or lesioned (27.5% ± 2.7%) rats during the first 2-hour interval from 19:00 to 21:00 hours. After 6-hour SD, the average SWA increased by 38.2% ± 4.3% and 21.5% ± 4.2% from the baseline levels to 46.1% ± 4.0% and 33.5% ± 3.7% in the control and lesioned rats, respectively (Fig 7E and 7F). The magnitude of the increase in SWA after 6-hour SD was significantly lower in lesioned rats than in control rats (p < 0.05, unpaired t test).
The above results showed that in the recovery period following 6-hour SD in RMTg lesioned rats, the amount of NREM sleep rebounded normally, but the rebound of NREM sleep depth was impaired. Therefore, the RMTg plays an important role in the homeostatic regulation of NREM sleep.
The RMTg is rich in μ receptors, and in vivo electrophysiological experiments have found that morphine, a μ-receptor agonist, inactivates RMTg neurons [25, 33]. Therefore, to rapidly and reversibly inhibit RMTg neurons, we bilaterally microinjected morphine through guide cannulas. After EEG recording, Nissl staining was performed to confirm the microinjection site. Only samples in which the needle tips were located in the RMTg were included in data analysis (Fig 8A). Under current-clamp conditions, morphine perfusion at 10 μmol/L (μM) was found to decrease the firing rate and induce significant hyperpolarization of RMTg neurons from −49.3 ± 1.3 to −56.3 ± 2.2 mV (p < 0.05). In some cases, morphine completely inhibited the firing of RMTg neurons within 1–2 minutes (Fig 8B and 8C). These results were consistent with previous reports that opioids could inactivate RMTg neurons [33].
During the 3-hour post-microinjection of morphine at 2 nmol/side, the rats showed an immediate 48.5% decrease in NREM sleep and a 93.1% decrease in REM sleep, with an increase in wakefulness by 107.6% compared with the artificial cerebrospinal fluid (ACSF) controls (Fig 8D and 8E). EEG architecture analysis showed that during the 3 hours after microinjection of morphine at 2 nmol/side, although the episode numbers of NREM and REM sleep did not change (Fig 8F), the longer fragments of NREM sleep with durations between 1 and 4 minutes and longer fragments of REM sleep between 1 and 2 minutes were significantly decreased (Fig 8G and 8H). As a result, the mean duration of NREM and REM sleep was shortened compared with ACSF microinjection controls (Fig 8I). In addition, the morphine-induced inhibition of RMTg neurons did not affect the number of stage shifts between wakefulness and NREM sleep (Fig 8J). EEG power spectrum analysis revealed that the power density of NREM sleep was significantly decreased in the rats microinjected in the RMTg with 2 nmol/side morphine compared with the ACSF controls over the frequency range of delta activity from 0.5 to 4 Hz during the 3 hours from 10:00 to 13:00 hours following microinjection (p < 0.05) (Fig 8K). The above results suggest that the RMTg plays roles in the maintenance of the duration and quality of NREM sleep.
Similar results were observed after pharmacogenetic inhibition of cells in the RMTg with Gi-coupled DREADDs (see S3 Fig). When RMTg neurons were inactivated (see S4A and S4B Fig), the rats showed a decrease in time spent in NREM sleep and a corresponding increase in time awake, but no decrease in REM sleep (see S4C Fig). The decreased NREM sleep resulted mainly from the reduction of NREM bout occurrence, particularly the number of shorter episodes with durations between 10–110 seconds (see S4D and S4E Fig) and fewer conversions from wakefulness to NREM sleep (see S4F Fig) following CNO injections. These results suggest that the RMTg plays a role in the initiation of NREM sleep. However, the changes in the mean duration and SWA level of NREM sleep did not differ significantly between saline-treated rats and CNO-treated rats (see S4G and S4H Fig). The decrease in NREM sleep without any change in SWA induced by pharmacogenetic inhibition of the RMTg using hM4Di was mimicked by selective pharmacogenetic activation of VTA/SNc DAergic neurons using hM3Dq in TH-Cre mice (see S5 Fig).
Using pharmacogenetic activation and neurotoxic lesions, as well as microinfusion and pharmacogenetic inactivation approaches, the present study found that RMTg neurons were essential for physiological NREM sleep. After RMTg neurons were damaged, the rats had reduced NREM sleep, which was mainly due to the decreased number of short-duration bouts and fewer transitions from wakefulness to NREM sleep. The results suggest that the RMTg is involved in the initiation of NREM sleep. Moreover, the RMTg mediates the maintenance of NREM sleep quality. SWA in the NREM sleep of the control and lesioned rats was prominent at lights on, when sleep pressure is highest, and gradually decreased as sleep progressed, which was in line with previous reports [30, 32]. However, the highest level of SWA in the RMTg-lesioned rats was significantly and continuously weaker than that in the control animals. NREM sleep quality was impaired by RMTg lesions, and the impairment lasted longer than the reduction in the amount of NREM sleep, which returned to baseline at 25 days after lesions. Therefore, the RMTg is essential to maintaining SWA at the peak level in NREM sleep; in contrast, its role in maintaining NREM sleep quantity can be fulfilled by other brain regions.
Although lesion formation is a widely used approach in neuroscience research, it has some limitations, such as causing irreversible damage to neurons and fibers of passage, which is not a natural physiological situation, and as a result, it is not possible to regulate neuronal activity reversibly [34]. Moreover, functional recovery may be seen, and the effects of lesions are confounded by secondary compensatory changes in response to the damage [35]. Complementary to the research on permanent neurotoxic lesions, temporary inactivation studies will help to clarify the engagement of the RMTg in the physiological regulation of sleep. We observed that the inactivation of RMTg neurons by both intra-RMTg injections of morphine and the Gi-coupled DREADD approach markedly reduced NREM sleep duration, which confirms the necessity of RMTg for the control of NREM sleep. The present study also shows that the normal neuronal activities of RMTg are necessary for the maintenance of NREM sleep, as demonstrated by the reduction in the numbers of longer bouts of NREM sleep with morphine-RMTg injection, as well as for the initiation of NREM sleep, as shown by the decreased numbers of short-duration bouts of NREM sleep and fewer transitions from wakefulness to NREM sleep with Gi-coupled DREADDs and neurotoxic lesions. The role of the RMTg in the promotion of NREM sleep is suggested again by the pronounced increase in NREM sleep duration with higher SWA levels after pharmacogenetic activation of RMTg neurons.
Most likely, the different effects of the RMTg on sleep architecture and sleep depth arise from differences in pharmacological and pharmacogenetic effects. The lesion method is an excellent approach, but it does not exclude a neuromodulatory role. Another conventional pharmacological intervention of microinfusion as well as the pharmacogenetic method with Gi-coupled AAV vectors that has developed rapidly over the past decade can be used to inactivate neurons. Intra-RMTg injection of morphine inhibits neurons by stimulating μ-opioid receptors and is likely to inhibit many more cells within the nucleus and diffuse to adjacent nontargeted areas, whereas the DREADD method is mainly used in the local nucleus to inhibit modified muscarinic receptors (M-Rs), especially with lower doses of AAVs, and generally produces mild effects. Moreover, the downstream signaling pathways of morphine and CNO (ligand for M-Rs) are also different.
It is worth noting that the vast majority of mCherry-positive cells in our study were confined to the RMTg (see S6 Fig). However, some labeled cells were found ventrally in the IPN. It has been reported that lesions of the bilateral fasciculus retroflexus, a major input to the IPN, result in reduced REM sleep [36]. Another study found that the Gscl (a homeobox transcription factor, Goosecoid-like) knockout mice exhibited a decrease in REM sleep and an increase in NREM sleep [37]. These findings indicate that the normal function of IPN is required for wakefulness and REM sleep. The neighboring dorsal and median raphe also had a small number of infected neurons. The main neurons in dorsal and median raphe are serotonergic neurons, which are sufficient for inducing wakefulness. Drugs that increase serotonin tone, such as selective serotonin reuptake inhibitors, generally increase wakefulness in both humans and rodents [38]. In all, the effects of the expression of hM3Dq/hM4Di spread outward somewhat are not linked to the NREM sleep promotion of RMTg neurons.
Sleep is an inactive state, characterized by the cessation of locomotion, reduced responsiveness, and easy reversibility [28]. From this point of view, the present results were in agreement with Lavezzi et al., who found that the activation of RMTg neurons through bilateral infusions of the GABAA receptor antagonist bicuculline suppressed locomotor activity in rats [39].
REM sleep was also inhibited during CNO-induced NREM sleep in rats with DREADD activation of RMTg neurons. Why REM sleep was so strongly inhibited is unclear, but one explanation may be that the effect of NREM sleep-promotion is too pronounced to be completely compensated by wakefulness alone. In a previous study, similar REM sleep reduction was observed when activation of the parafacial zone markedly induced NREM sleep [40]. It is also possible that RMTg GABAergic neurons project to and inhibit REM sleep-promoting PPT and LDT neurons [38]. However, considering that lesions or inactivation of the rat RMTg did not markedly affect REM sleep quantity, we feel that the RMTg does not play an important role in the physiological control of REM sleep.
Homeostatic control of sleep, 1 of 2 sleep regulation processes, is the increased propensity for sleep during prolonged wakefulness [41]. Animals may compensate for sleep loss by an increased amount of NREM sleep and SWA in NREM sleep [28]. Our study showed that 6-hour SD induced a 2-hour increase in NREM sleep, which was similar to previous studies, which found that NREM sleep was significantly above the corresponding baseline in the first 2-hour interval after 6-hour SD [42, 43]. However, in the present study, NREM sleep intensity significantly increased in control rats, whereas the RMTg-lesioned rats showed a reduced response of sleep homeostasis, as indicated by the lack of enhanced delta EEG power in NREM sleep after 6-hour SD. Several previous studies showed the same results. Rats with nucleus accumbens core lesions showed reductions in SWA rebound, but not in sleep duration response, after 6-hour SD [43], and the A1-R (a kind of adenosine receptor) knockout mice responded similarly following 3-hour SD [44].
SWA in NREM sleep is considered to be functionally involved in sleep–wake homeostasis. However, the neuronal mechanisms and neurochemical factors that drive these homeostatic responses are largely unknown [7, 32]. Studies have suggested that SWA is generated in the cerebral cortex and thalamus [38], and the basal forebrain is another important structure for sleep homeostasis [45]. The DRN and PPT nucleus are important target nuclei of RMTg [15]. Inhibition of DRN neurons by inhibition of calmodulin-dependent kinase II or damage with 3,4-methylenedioxymethamphetamine was found to significantly enhance NREM sleep and delta power density during NREM sleep [46, 47]. In addition, activation of cholinergic LDT neurons suppressed slow EEG activity during NREM sleep [48]. Moreover, the VTA and SNc, which are major strong outputs of RMTg, have efferent and afferent connections with the DRN, LDT, locus coeruleus, lateral and posterior hypothalamus, basal forebrain, and thalamus [49]. Collectively, the downstream nuclei may be involved in the circuit through which the RMTg regulates sleep homeostasis. Further studies are needed to clarify the inconsistency of RMTg’s effects on the 2 aspects of homeostasis regulation.
RMTg is widely recognized as a brake for the DAergic system. Early electrophysiological findings, which showed that VTA and SNc DAergic neurons do not change their firing rates across the sleep–wake cycle [50, 51] and that there is no change in the time spent in electrocortical waking in animals with electrolytic lesions of the catecholamine-containing neurons of the VTA and SNc [52], suggest that DAergic neurons are not involved in the regulation of sleep–wake states [7]. Meanwhile, some evidence suggests that DA signals are associated with sleep promotion [49, 53]. Since DAergic neurons are mainly distributed in the VTA and SNc, we directly manipulated the activity of midbrain DAergic neurons using pharmacogenetic methods to observe their effects on sleep–wake behavior. We found that selective inhibition of VTA/SNc DAergic neurons promoted NREM sleep, which was also observed with activation of RMTg neurons. In contrast, selective activation of VTA/SNc DAergic neurons produced an increase in wakefulness. The current results are in agreement with a recent finding that optogenetic stimulation or inhibition of VTA DAergic neurons initiates and maintains wakefulness or increases NREM sleep, respectively [10].
We find that the RMTg is essential for NREM sleep and homeostatic regulation. Since the RMTg exerts major inhibitory control over the midbrain DAergic system, we hypothesize that the RMTg regulates sleep–wake behavior through the modulation of DAergic neuron activity. However, we cannot exclude the possibility that other RMTg GABAergic projected targets also contribute to the roles of the RMTg in sleep–wake behavior.
For further study, we will use transgenic VGAT (vesicular GABA transporter)-Cre mice to specifically target the GABAergic neurons in order to expand on our findings in rats that RMTg neurons are necessary for sleep and use pharmacogenetic and optogenetic manipulations together with polysomnographic recordings to elucidate the role of the downstream elements of GABAergic RMTg neurons in the control of sleep.
There is evidence that sleep, especially NREM sleep with high levels of EEG delta power, is important not only for neurobehavioral functions such as memory consolidation [54] but also for peripheral physiological functions such as the maintenance of normal glucose homeostasis [55, 56]. Thus, the current results suggest that the RMTg may be considered as a potential intervention for prolonging NREM sleep duration and improving sleep quality to maintain human health. Moreover, the data obtained in the present study enrich our understanding of how the brain regulates sleep–wake behavior and provide a potential target for understanding the roles of DA in the physiological regulation of sleep–wake states as well as in the pathologic process of sleep disturbances. Our results will inform the development of potential therapeutic targets against sleep disorders in DA-implicated mental illness.
All experimental procedures involving animals were approved by the Committee on the Ethics of Animal Experiments of School of Basic Medical Sciences, Fudan University, with license identification number 20150119–067. The animals were anesthetized with an IP injection of chloral hydrate before surgery or killing. The animals were put on the heating pad until they woke up from anesthesia after surgery. During the postoperative recovery period, the animals were observed every day, and the sawdust was kept clean. Every effort was made to minimize animal suffering or discomfort and to reduce the number of animals used.
Male Sprague–Dawley rats (280–370 g) were obtained from the Laboratory Animal Center, Chinese Academy of Sciences (Shanghai, China). Adult male TH-Cre mice and non-Cre-expressing littermate mice (8–16 weeks old, 25–30 g) were also used. The animals were housed in individual cages at an ambient temperature (22 ± 0.5°C) with relative humidity of 60% ± 2% in an automatically controlled 12:12-hour light/dark cycle (lights on at 07:00 hours, illumination intensity approximately 100 lux), with free access to food and water.
The AAVs of serotype rh10 for AAV-hSyn-DIO-hM3Dq-mCherry, AAV-hSyn-DIO-hM4Di-mCherry, and AAV-hSyn-DIO-ChR2-mCherry were generated by tripartite transfection (AAV-rep2/caprh10 expression plasmid, adenovirus helper plasmid, and pAAV plasmid) into 293A cells. After 3 days, the 293A cells were resuspended in ACSF, freeze-thawed 4 times, and treated with benzonase nuclease (Millipore) to degrade all forms of DNA and RNA. Subsequently, the cell debris was removed by centrifugation, and the virus titer in the supernatant was determined using an AAVpro Titration Kit for Real Time PCR (Takara). The final viral concentrations of the transgenes were 1 × 1012–2 × 1012 genome copies/mL. Aliquots of viral vectors were stored at −80°C before stereotaxic injection.
To induce lesions in the RMTg, rats were anesthetized with chloral hydrate (10% in saline, 360 mg/kg) and immobilized in a Stereotaxic Alignment System (RWD Life Science, Shenzhen, China). Ibotenic acid (1% in saline, 200 nL/side; Sigma, St. Louis, Missouri, United States) was bilaterally injected through a glass pipette (glass stock: 1 mm in diameter; tip: 10–20 μm) with nitrogen gas pulses of 20–40 psi using an air compression system [57] into the RMTg according to the atlas [58] (coordinate relative to bregma: AP −6.8 mm; ML ± 0.3 mm; DV −8.4 mm) for 5–10 minutes. After leaving the pipette in the brain for an additional 5 minutes, the pipette was slowly retracted. Control rats received 200 nL/side saline.
For microinjection of morphine in the RMTg, rats were implanted with 2 guide cannula (30 gauge). The cannulas were inserted at stereotaxic coordinates based on the rat brain atlas [58] (coordinate relative to bregma: AP −6.8 mm; ML ± 0.3 mm; DV −7.4 mm). The 2 cannulas were fixed to the skull with dental cement and 3 stainless steel screws for anchorage [59].
In order to manipulate neuronal activity, we used hM3Dq or hM4Di that was selectively activated or inhibited by the pharmacologically inert agent CNO [60]. Under anesthesia with chloral hydrate (10% in saline, 360 mg/kg), a burr hole was made, and a fine glass pipette containing AAVs carrying Cre-independent hM3Dq, hM4Di, or ChR2 was lowered bilaterally into the rat RMTg (coordinate relative to bregma: AP −6.8 mm; ML ± 0.3 mm; DV −8.4 mm). The AAV vectors were delivered with 300 nL/side.
Mice were anesthetized with chloral hydrate (5% in saline, 720 mg/kg) and then placed in a stereotaxic frame so that the head was fixed. After opening a burr hole, a fine glass pipette containing AAV carrying Cre-dependent hM3Dq/hM4Di (Taiting, Shanghai, China) was bilaterally lowered into the VTA (coordinate relative to bregma: AP −3.4 mm; ML ± 0.3 mm; DV −4.0 mm) or SNc (coordinate relative to bregma: AP −3.4 mm; ML ± 1.2 mm; DV −4.0 mm) according to the atlas[5]. The AAV vectors (50 nL/side) were delivered over a 5-minute period per hemisphere. After an additional 10 minutes, the pipette was slowly withdrawn.
Following ibotenic acid or saline injection (or guide cannula implantation) or after 2 weeks of recovery from virus injection, the rats were implanted with EEG and electromyography (EMG) electrodes for polysomnographic recordings. The implant consisted of 2 stainless steel screws (1-mm diameter) inserted through the frontal (AP +2 mm; ML +3 mm) and parietal (AP −4 mm; ML +3 mm) bones, and a stainless steel screw (1.5-mm diameter) inserted in the left frontal bone (AP +3 mm; ML −3 mm) as a reference electrode. All of the above positions were coordinately relative to the bregma [58]. The mice were recovered for 2 weeks after virus injection before electrodes were implanted. Two stainless steel screws (1-mm diameter) were inserted through the skull into the cortex (AP +1 mm to the bregma; AP +1 mm to the lambda; ML +1.5 mm to the midline) [61] and served as EEG electrodes.
Two Teflon-coated, stainless steel wires were bilaterally placed into both trapezius muscles for EMG recordings in rats or mice. All electrodes for the rats or mice were attached to a connector and fixed to the skull with dental cement. The animals were then allowed to recover on a heating pad until awakening from anesthesia [62].
After a 1-week recovery period from EEG electrode implantation, the animals were transferred to the recording room and habituated to the recording cables and conditions for 2–3 days. Following this habituation period, 48 hours of EEG/EMG recordings were performed on all the animals. The data collected during the first 24 hours served as baseline, and the second 24 hours served as experimental data.
Morphine hydrochloride, a nonselective μ-opioid receptor agonist, diluted in ACSF, was injected at 09:00 hours through a syringe connected with a lengthened flexible pipe under red illumination; thus, the rats could freely move or have rest without irritant body touch, disturbing noises, and natural light. A syringe that was designed with an oblique tip for easy aligning was gently and smoothly inserted into the hollow and straight guide cannula (O.D. = 0.6 mm). A volume of 0.5 μL morphine was injected in each side by microinjector at a slow and constant speed for 0.5 minutes. Injection was made unilaterally in sequence. After the end of each injection, the syringe was left for an additional 3 minutes for complete local absorption. The control group was injected with ACSF.
In the DREADD experiment, saline was administered IP at 09:00 hours on day 1 of the EEG recording. On the next day, CNO (LKT Laboratories, Minneapolis, Minnesota, US) was dissolved in saline before use and injected at 09:00 hours at 1 mg/kg for mice (0.1 mL/10 g) or 0.3 mg/kg for rats (1 mL/100 g).
Cortical EEG and EMG signals were amplified, filtered (EEG, 0.5–30 Hz; EMG, 20–200 Hz), digitized at a sampling rate of 128 Hz, and recorded using VitalRecorder (Kissei Comtec, Nagano, Japan). When complete, polygraphic recordings were automatically scored offline by 10-second epochs as waking, NREM sleep, and REM sleep using SleepSign according to standard criteria. Defined sleep–wake stages were examined visually and corrected if necessary [57].
SD was achieved by gentle handling that included tapping the cage, introducing novel objects into the cage, or removing the rat from the cage when behavioral signs of sleep were observed. Rats were deprived of sleep during the light phase for 6 hours from 13:00 to 19:00 hours. Undisturbed rats that served as a control group were never disturbed when they were spontaneously awake, feeding, or drinking in the same time period as the corresponding 6-hour SD group [30, 42].
On completion of EEG recordings of the rats with lesions or morphine microinjection, the rats were anesthetized with chloral hydrate (10% in saline, 360 mg/kg), with 150–200 mL saline followed by 500 mL 4% paraformaldehyde (PFA) in PBS through the heart. The brains were removed, post-fixed for 4–5 hours at 4°C in 4% PFA, and then equilibrated in phosphate buffer containing 10%, 20%, and 30% sucrose solution at 4°C. The brain sections were serially cut in the coronal plane at 30 μm on a freezing microtome (CM1950, Leica, Wetzlar, Germany), protected in cryoprotectant solution, and stored at −20°C until further processing for immunostaining.
For verification of ibotenic-acid-induced brain lesions, one series of tissues was processed for NeuN staining. Brain sections were incubated in primary antibody in PBS containing Tween-20 (PBST) (mouse anti-NeuN, 1:50,000; Millipore, Bedford, Maryland, US) overnight, and then staining was revealed using the avidin–biotin complex method (ABC kit SC-2017; Santa Cruz Biotechnology, Santa Cruz, California, US). The sections were incubated for 2 hours in biotinylated secondary antibody in PBST (1:500), followed by incubation with avidin–biotin–horseradish peroxidase (HRP) conjugate and staining with 3,3-diaminobenzidine tetrahydrochloride (DAB). Sections were then mounted, dried, dehydrated, and cover slipped. Only samples in which the lesion sites were confined to the RMTg were included in data analysis.
For confirmation of the microinjection site of morphine, one series of sections was subjected to Nissl staining. Brain sections were mounted on adhesive slides, dried naturally for 2 consecutive days, and stained by cresyl violet method. Sections were washed in water and PBS successively, incubated in 0.1% cresyl violet for 15 minutes, differentiated in graded ethanol, and cleared in xylene before being cover slipped. All sections containing a cannula-insertion site were compared with a rat brain atlas to confirm the 3-dimensional coordinates of the site relative to bregma. Only experiments in which the tip of the microinjection cannula was located above the RMTg were included in data analysis [59].
The AAV vectors were linked with a red fluorescent protein mCherry; therefore, the viral injection site was determined by mCherry expression. Induction of c-Fos, the protein of an immediate early gene, is supposed to be an indicator of neuronal activity [63]. Thus, whether the virus-infected neurons were activated or inhibited could be determined through c-Fos expression. After EEG/EMG recording, the animals were injected with saline or CNO (0.3 mg/kg for rats; 1 mg/kg for mice). Ninety minutes later, the animals were anesthetized and perfused, and the brain coronal sections were prepared. The brain tissue sections were rinsed in 0.1 M PBS (3 times, 5 minutes/each wash) and then incubated in rabbit anti-c-Fos primary antibody (1:5,000; Millipore) diluted in PBST for 48 hours. The tissues were washed 3 times in PBS and incubated with Alexa Fluor 488-conjugated donkey anti-rabbit secondary antibody in PBST (1:1,000; Invitrogen, Carlsbad, California, US) for 2 hours in the dark. After being washed in PBS, the sections were mounted on glass slides, cover slipped using FluoroGuard Mounting Medium, and kept at 4°C before imaging [40]. Colocalization of mCherry and c-Fos expression was observed by confocal microscopy. Localization of the rat RMTg and mouse VTA and SNc were confirmed by staining and reference to the brain atlas.
For double immunofluorescence staining of TH/mCherry or GABA/mCherry, brain sections were incubated with a rabbit antibody against TH (1:3,000; Millipore) or GABA (1:1,000; Invitrogen) in PBST over night at 4°C. The sections were then rinsed and incubated in a donkey anti-rabbit Alexa Fluor 488-conjugated secondary antibody (1:1,000; Invitrogen) at room temperature for 2 hours in the dark. After 3 washes in PBS, sections were incubated in 4,6-diamidino-2-phenylindole (DAPI; 1:3,000; Invitrogen) for 10 min at room temperature. Finally, sections were washed in PBS and mounted on glass slides using FluoroGuard Mounting Medium.
For quantification of the colocalization of mCherry-expressing RMTg neurons in rats and VTA/SNc DAergic neurons in mice microinjected with hM3Dq/hM4Di-containing vectors and other histological markers (GABA, c-Fos, TH), the area used for counting was demarcated by cells that were labeled by mCherry. All cell counting was conducted blindly on 3 × 2 tiled confocal images of the target area. For each rat or mouse, brain sections were analyzed bilaterally [64].
To investigate whether CNO manipulates the activity of AAV-infected cells or RMTg has an inhibitory projection to midbrain DAergic neurons, male Sprague–Dawley rats, 20–30 days old weighing 50–60 g, were microinjected with AAV vectors carrying Cre-independent hsyn–hM3Dq (hM4Di, ChR2)-mCherry under anesthesia with chloral hydrate into the RMTg, and TH-Cre mice were microinjected with AAV vectors carrying Cre-dependent hsyn–hM3Dq/ hM4Di–mCherry into the VTA or SNc. Slices containing the RMTg of the rats or VTA/SNc of TH-Cre mice were prepared from the animals 3 weeks after AAV microinjection or from male Sprague–Dawley rats, 20–30 days old weighing 50–60 g, without AAV injection to investigate whether morphine inhibited RMTg neurons in vitro. The RMTg or VTA/SNc was identified according to stereotaxic coordinates [58, 61].
Coronal slices of the rats (280-μm thick) were cut using a vibratome (VT-1200S; Leica) in ice-cold sucrose-based ACSF, bubbled with 95% O2 and 5% CO2, containing 230 mM sucrose, 2.5 mM KCl, 3 mM MgSO4, 1.25 mM NaH2PO4, 26 mM NaHCO3, 0.5 mM CaCl2, and 10 mM d-glucose. The slices were allowed to recover for at least 1 hour in a holding chamber with ACSF without sucrose in a water bath (32°C) before recording. For preparation of the mouse brain slices, the coronal slices were cut at 300-μm thickness in ice-cold glycerol-based ACSF, which was different from the rats, containing 260 mM glycerol, 5 mM KCl, 1.25 mM KH2PO4, 1.3 mM MgSO4, 0.5 mM CaCl2, 20 mM NaHCO3, and 10 mM glucose.
Coronal slices were transferred to the recording chamber, where they were held down with a platinum ring. Carbonated ACSF with 95% O2 and 5% CO2 flowed through the bath (2 mL/minute). Patch pipettes were pulled from thick-walled borosilicate glass capillaries (1.5-mm outer diameter, 0.84-mm internal diameter, Sutter Instruments, San Rafael, California, US) using a 2-step vertical puller (PC-10; Narishige, Japan). Pipette resistance was typically 4–7 MΩ when filled with internal solution containing 120 mM potassium gluconate, 20 mM KCl, 1 mM MgCl2, 0.16 mM CaCl2, 10 mM HEPES, 0.5 mM EGTA, 2 mM Mg-ATP, and 0.5 mM NaGTP. RMTg or VTA/SNc neurons were identified under visual guidance using a fixed-stage upright microscope (BX-51; Olympus, Tokyo, Japan) fitted with a 40 × water immersion objective lens. The image was detected with an infrared sensitive charge coupled device camera (U-TV1X-2; Olympus) and displayed on a screen in real time. The output signals were amplified (Molecular Devices, Eugene, Oregon, US), filtered at 5 kHz, and digitized at 20 kHz using a National Instruments digitization board (NI-DAQmx, PCI-6052E; National Instruments, Austin, Texas, US). Neurons were current clamped to record spontaneous action potentials and/or membrane potentials. The series resistance and input resistance were monitored throughout the cell recording, and data were discarded when either of the 2 resistances changed by >20% [57, 65]. To visualize the recorded cells in the RMTg, biocytin (0.2%) was included in the pipette solution to confirm the position of patched cells. Slices were fixed immediately after recording in 4% formaldehyde for 2 hours and then immersed in 0.3% PBST. Slices were incubated in Fluor-488-conjugated streptavidin (Invitrogen, 1:2,000, 12 hours at 4°C). Sections were mounted on slides using FluoroGuard Antifade Reagent (Bio-Rad, Hercules, California, US) and visualized under an Olympus microscope.
For the optogenetic experiment, whole-cell and cell-attached recordings were made from RMTg or VTA/SNc DAergic neurons. ChR2 was stimulated by 473-nm light delivered via an optical fiber coupled to a laser source (Guang Teng, Shanghai, China). For recording light evoked inhibitory synaptic currents, the internal solution contained 105 mM potassium gluconate, 30 mM KCl, 4 mM ATP-Mg, 10 mM phosphocreatine, 0.3 mM EGTA, 0.3 mM GTP-Na, and 10 mM HEPES. In the voltage-clamp mode, cells were held at −70 mV. When needed, 25 μM d-(-)-2-amino-5-phosphonopentanoic acid (d-APV), 5 μM NBQX, and 100 μM PTX were added to block NMDA, AMPA, and GABAA receptors, respectively. The internal solution also contained 0.2% biocytin. To test for expression of TH, brain slices were incubated in rabbit anti-TH antibody (1:2,000, Millipore) containing 3% normal donkey serum (v/v), 0.5% Triton X-100 (v/v) for 24 hours at 4°C. This was followed by incubation with Alexa Fluor 488-conjugated donkey anti-rabbit (1:800; Invitrogen) and Alexa Fluor 405 streptavidin (1:1,000, Invitrogen) for 12 hours at RT.
The sum of sleep and wakefulness and other sleep architecture parameters in the RMTg-lesioned and control rats were compared using unpaired t tests. The sleep–wake profiles among groups with different doses of morphine microinjection and vehicles were assessed using 1-way ANOVA followed by least significant difference tests. The hourly durations of each stage and SWA analyses were compared using 2-way ANOVA (repeated measures) followed by unpaired t tests. Two-way ANOVA (repeated measures) followed by paired t tests were conducted to analyze the sleep rebound induced by 6-hour SD in rats.
For viral microinjection, the sum of sleep and wakefulness and other sleep architecture parameters after CNO or saline injection as well as membrane potentials of rat RMTg neurons before and after bath application of CNO were compared using paired t tests. The hourly duration of each stage of sleep and wake profiles was compared using 2-way ANOVA (repeated measures) followed by paired t tests.
For GABA, c-Fos, or TH immunohistochemistry analysis, each group consisted of data obtained from 3 rats or TH-Cre mice; as a result, a total of 3 bilateral sections containing target area were analyzed for each animal. The quantification results were compared using unpaired t tests between CNO and saline control groups.
All results were expressed as the mean ± SEM. We analyzed the data using Prism 5.0 (GraphPad software, San Diego, California, US). In all cases, p < 0.05 was taken as the level of significance.
|
10.1371/journal.pcbi.1004628 | The Essential Complexity of Auditory Receptive Fields | Encoding properties of sensory neurons are commonly modeled using linear finite impulse response (FIR) filters. For the auditory system, the FIR filter is instantiated in the spectro-temporal receptive field (STRF), often in the framework of the generalized linear model. Despite widespread use of the FIR STRF, numerous formulations for linear filters are possible that require many fewer parameters, potentially permitting more efficient and accurate model estimates. To explore these alternative STRF architectures, we recorded single-unit neural activity from auditory cortex of awake ferrets during presentation of natural sound stimuli. We compared performance of > 1000 linear STRF architectures, evaluating their ability to predict neural responses to a novel natural stimulus. Many were able to outperform the FIR filter. Two basic constraints on the architecture lead to the improved performance: (1) factorization of the STRF matrix into a small number of spectral and temporal filters and (2) low-dimensional parameterization of the factorized filters. The best parameterized model was able to outperform the full FIR filter in both primary and secondary auditory cortex, despite requiring fewer than 30 parameters, about 10% of the number required by the FIR filter. After accounting for noise from finite data sampling, these STRFs were able to explain an average of 40% of A1 response variance. The simpler models permitted more straightforward interpretation of sensory tuning properties. They also showed greater benefit from incorporating nonlinear terms, such as short term plasticity, that provide theoretical advances over the linear model. Architectures that minimize parameter count while maintaining maximum predictive power provide insight into the essential degrees of freedom governing auditory cortical function. They also maximize statistical power available for characterizing additional nonlinear properties that limit current auditory models.
| Understanding how the brain solves sensory problems can provide useful insight for the development of automated systems such as speech recognizers and image classifiers. Recent developments in nonlinear regression and machine learning have produced powerful algorithms for characterizing the input-output relationship of complex systems. However, the complexity of sensory neural systems, combined with practical limitations on experimental data, make it difficult to apply arbitrarily complex analyses to neural data. In this study we pushed analysis in the opposite direction, toward simpler models. We asked how simple a model can be while still capturing the essential sensory properties of neurons in auditory cortex. We found that substantially simpler formulations of the widely-used spectro-temporal receptive field are able to perform as well as the best current models. These simpler formulations define new basis sets that can be incorporated into state-of-the-art machine learning algorithms for a more exhaustive exploration of sensory processing.
| Encoding models provide a powerful, objective means to evaluate our understanding of how sensory neural systems represent complex natural stimuli [1, 2]. An encoding model describes any time-varying neural signal (single- or multiunit activity [3, 4], local field potential [5], hemodynamic activity [6], or behavior [7]) as a function of the input stimulus, and it can predict the neural response to an arbitrary novel stimulus, including complex natural sounds. Prediction accuracy provides a quantitative measure of how well a model describes sensory-evoked activity; a completely accurate model should predict neural responses to any stimulus without error. More accurate models of sensory neural activity provide insight into algorithms that can be integrated into automated systems, such as speech recognizers and image classifiers.
In the auditory system, the linear spectro-temporal receptive field (STRF), implemented as a finite impulse response (FIR) filter, is the established “standard model” for neural representation [2, 4, 8–13]. This filter forms the core of generalized linear models (GLMs) applied to the auditory system [14, 15], and models sharing the same analytical form as the FIR STRF have been developed for studying visual [16–18], somatosensory [19, 20], and olfactory systems [21]. Despite its widespread use, careful assessments of how well the linear STRF actually describes auditory neural activity are limited [22]. A few studies have shown that the linear STRF can explain only a limited portion of sound-evoked activity in cortex, especially for complex natural stimuli [9, 23]. Others have argued that nonlinear variants of the classical linear STRF can improve predictive power [3, 24–33]. These nonlinear variants of the STRF show improved predictive power under specific experimental conditions. However, the more complex models are difficult to estimate reliably when experimental data are limited [1, 18, 22], especially for natural stimuli [12, 23, 34]. Difficulties associated with fitting and testing have prevented any single alternative from replacing the linear STRF as a new standard.
The challenges encountered when evaluating alternatives to the FIR STRF highlight the trade-off between model performance, how accurately it predicts neural activity, and complexity, the degrees of freedom governing the stimulus-response relationship [35, 36]. In order to completely describe a system’s function, an encoding model must account for all the degrees of freedom of the actual system. If the system is not well understood, some degrees of freedom in a model are likely to be mismatched to the system’s function. Any mismatched complexity does not provide additional explanatory power, but it does introduce noise into model parameter estimates. Because this complexity does not improve performance, there should exist a model with fewer degrees of freedom that can perform as well as the more complex model.
In this study we focus on the problem of complexity. Rather than simply seeking the model that performs best, we identify the simplest possible model that attains a minimum level of performance. Specifically, we ask, can we produce a low-dimensional approximation of the linear STRF that performs as well as the full FIR STRF? The idea of improving STRF performance by dimensionality reduction has been proposed previously. Isolated studies have shown benefits of low-rank approximations of the STRF [28, 31, 37, 38]. In the visual system, several studies have also proposed low-dimensional, system-specific parameterizations [18, 29, 39–43]. Despite the many parameterizations that have been proposed, however, direct comparisons between them have been limited, especially for natural stimuli. Thus it remains difficult to identify the important features of these different models.
We approached the complexity problem directly by systematic comparison of a large set of alternative parameterizations. We generated a collection of models that instantiate a variety of low-dimensional approximations to the FIR STRF. We then compared their performance on single-unit data collected from primary auditory cortex during presentation of natural vocalizations. By exploring the performance of this family of models, we were able to identify the minimal essential components required by linear STRFs that best described the data and to study the relationship between the amount of data available and optimal model complexity.
We found that the standard FIR STRF is suboptimal according to the complexity criterion. Instead, a much simpler model, which defines the STRF as a product of three Gaussian-tuned spectral filters and biphasic temporal filters, outperformed the FIR STRF, while requiring only about 10% of the parameters (29 vs. 276 free parameters). These results indicate that, for the average A1 neuron, a model with about 30 free parameters is able to capture its linear filter properties. The total degrees of freedom of a comprehensive nonlinear model is likely to be higher, but our minimally complex linear STRF provides a starting point for developing better-performing nonlinear models.
We recorded single-unit neural activity from the auditory cortex (A1) of awake, passively listening ferrets during presentation of natural ferret vocalizations. The same set of 42 3-second vocalizations was presented during recordings from all neurons (N = 176). We then fit a large number of encoding models with different architectures to data from each neuron and compared their performance. Data for each neuron were grouped into an estimation data set (40 vocalizations), which was used for fitting, and a validation data set (2 vocalizations), which was used only to test how well each fit predicted responses to a novel stimulus (Fig 1A). Our primary performance metric was prediction correlation, i.e., the correlation coefficient (Pearson’s R) between the actual peri-stimulus time histogram (PSTH), r(t), and the PSTH predicted by the model, p(t) (Fig 1C). Other commonly used performance metrics showed the same pattern of results (e.g., log-likelihood and mutual information, see below).
Models were structured as a sequence of signal transformations, or functional modules, corresponding to the block diagram in Fig 1B,
x 0 ( t ) → f 1 ( · ) x 1 ( t ) → f 2 ( · ) ⋯ → f n ( · ) y ( t ) (1)
where the output, xi(t), of each module, fi(⋅), provides the input into the subsequent module. The final module produced the predicted time-varying spike rate, y(t). In most models tested, this sequence consisted of three modules, a cochlear filterbank [26, 44], followed by a linear spectro-temporal filter [8, 9, 11, 12], and finally an output nonlinearity to account for spike generation thresholds [13, 17].
Alternative model architectures were compared by replacing one or more modules in Eq 1, while keeping the others the same. Thus the impact of the choice for each module on model performance could be tested individually (see Fig 2C). Using this empirical approach, we selected optimal modules for the cochlear filterbank (Eqs 11–13) and output nonlinearity (Eq 14) for the same linear filter module (FIR filter, see below, Eq 3). These modules were then held constant while we compared performance for the different formulations of the linear filter module that follow.
Models were fit using an iterated coordinate descent (a.k.a. boosting) algorithm [34]. On each iteration, the algorithm cycled through each module sequentially and performed a few steps of coordinate descent within that module before moving on to the next one (see Methods). We have previously demonstrated that this coordinate descent algorithm is able to accurately recover linear STRFs in simulation [30, 34].
Because datasets are finite, the performance of any model will be limited by sampling noise. This noise impacts the analysis at two stages: producing error in the estimation of model parameters and in validation of prediction accuracy [18, 22, 45]. Accounting for the first problem is a nuanced issue: more complex models that require a large number of parameters are more susceptible to noise than simpler models. We address the issue of finite estimation data in a later section (see Parameterized models perform similarly to FIR models in the limit of infinite data, below). To account for the latter problem, measures of prediction correlation were normalized by a factor reflecting response reliability in the validation stimulus (Eq 23, [45]). This factor was fixed for an individual neuron’s validation data. Thus it does not affect the performance of one model relative to another. Numerically, this correction increased prediction correlations in A1 by a mean of 20% (ranging from 3% to 39% for individual neurons).
Model complexity is often factored into cost functions for model fitting, in order to positively weigh simpler models [35, 46]. Our goal was to study in depth the relationship between model complexity and performance. Thus, rather than combining them into a single cost function, we studied the trade-off between these criteria in detail, exploring the family of solutions that are optimal with respect to both. This optimal set of solutions is known as the Pareto front [36, 47]. Formally, all items belonging to this front are non-dominated in the Pareto sense [47] which means that for all pairs of models on the front, one is less complex while the other fits more closely to the data. All models below the Pareto front are non-optimal: there is at least one model on the front that is both less complex and more accurate.
We generated Pareto plots for the 1061 different linear STRF architectures tested, comparing model parameter count against average prediction correlation for estimation data (Fig 2A) and validation data (Fig 2B). Most models lie under the Pareto front (red line) and are suboptimal relative to models that are less complex, better performing, or both. More complex models tend to perform better for estimation data, but they do not necessarily predict novel validation data more accurately. The differences between estimation and validation plots illustrate the problem of overfitting when available estimation data are finite. Among the more complex models, the FIR STRF falls below the Pareto front for the validation data (black point, Fig 2B). Instead, best performance in the current dataset is achieved by a model requiring just 29 parameters (orange point).
In the following sections, we discuss in detail the subset of 260 architectures in which only the linear filtering module was varied, while all other modules (cochlear filterbank, input nonlinearity, output nonlinearity) and the fitting algorithm were held constant (dark gray points, Fig 2A and 2B). Our focus is on identifying model architectures that fall on or near the Pareto front, making them optimal for a given level of complexity. The remaining models were generated by manipulating one or more modules other than the linear filter (Fig 2C). Varying the other modules had less dramatic effects on model complexity and performance, but they provide a dense sampling of the complexity-performance space. A complete list of architectures evaluated is included in the supplementary materials (S1 Table).
Parameterized STRFs are approximations of the FIR STRF. Thus, in theory, the FIR STRF should perform as well as or better than any parameterized STRF. In practice, however, data available for estimation are finite, and simpler models can be estimated more accurately than the full FIR STRF. Thus simpler models are able to perform better than the FIR STRF in our analysis (Fig 6). The results so far demonstrate a clear practical advantage of the factorized and parameterized models, but they do not answer the question of whether any simpler model fully accounts for the linear STRF. Such a question can only be answered by comparing the relative performance of these models in the limit of infinite estimation data [18, 22].
Extrapolating performance to infinite estimation data is challenging because there is no widely agreed upon model of variability in sensory-evoked neural activity. We made a simplifying assumption that prediction error from estimation noise is additive and inversely proportional to the square root of the number of samples used to estimate the STRF, T (see Methods, Eq 31, [18, 45]). When these assumptions hold, then the effect of noise on model variance explained (square of prediction correlation, R2) also decreases proportionally to T. We varied T by subsampling the available estimation data (10%–75%) and measured the average RT across neurons for models fit with the different data subsets. We then fit the free parameters in Eq 31 to determine the theoretical limit on performance for each model, Rinf.
We measured the asymptotic performance limit of four model architectures, ranging from high to low complexity: the full FIR model (FIR, 276 parameters), D = 3 factorized model (Factorized x3, 109 parameters), D = 3 Gaussian spectral/P3Z1 temporal parameterization (P3Z1x3, 29 parameters), and D = 1 Gaussian spectral/P3Z1 temporal parameterization (P3Z1x1, 13 parameters). We removed very noisy data and focused on the subset of 124 neurons that produced reliable auditory-evoked responses (SNR > 0.005, see Methods, Eq 21).
For all models, performance improved as more estimation data became available (Fig 8A). As expected, the lower-dimensional models performed better for small data sets and neared asymptotic performance sooner than higher-dimensional models. Consistent with this observation, performance of the FIR STRF showed the greatest improvement in the asymptote (Rinf = 0.63, Fig 8B). However, performance of the Factorized x3 (Rinf = 0.63) and P3Z1x3 models (Rinf = 0.62) was not significantly different from the FIR STRF (jackknifed t-test). Thus within the precision we could achieve with this analysis, both models captured the essential features of the FIR STRF. Error bars on asymptotic performance are relatively large, especially for the FIR STRF, so a strong conclusion about relative performance of these models is difficult. However, asymptotic performance of the P3Z1x1 model was significantly worse than the other models (Rinf = 0.56, p < 0.001), indicating a failure of this very simple model to capture the full linear model.
For comparison with a previous analysis [22], we also measured asymptotic performance for the FIR STRF with no output nonlinearity. This model performed better than the standard FIR STRF for smaller estimation sets, presumably due to its reduced complexity, but its advantage diminished for larger datasets. Asymptotic performance was slightly lower than the standard FIR STRF that included an output nonlinearity (Rinf = 0.61, p < 0.05, Fig 8B).
In addition to outperforming the FIR model in finite data conditions, reduced-dimensionality factorized and parameterized STRFs demonstrated several other benefits over the FIR STRF, which we detail below. For brevity in this section, factorized model refers specifically to the D = 2 factorized model, and parameterized model refers to the D = 3 Gaussian spectral parameterization with P3Z1 temporal parameterization. These models were chosen because they represent the best-performing models, respectively, among the factorized and parameterized models tested (Fig 6C).
The finite impulse response (FIR) STRF represents the current standard model for stimulus-response filtering in the auditory system [2, 4, 8–13]. Our results agree with previous findings that, as a general architecture, the linear STRF accounts only partially for the neural response to natural sounds in A1 [9, 23]. However, we find that the same level of performance can be achieved by much simpler models. A model requiring fewer than 30 parameters not only matches performance of the FIR STRF (> 250 parameters) but actually outperforms it for large but finite datasets. The simplest parameterization that works optimally for a neural population provides insight into the neural circuitry underlying system function [36]. According to this logic, the average linear STRF of an A1 neuron can be captured by the sum of three channels with Gaussian spectral tuning and an IIR temporal filter.
When data are finite, a critical issue is that a simpler model with fewer free parameters will be less susceptible to estimation noise than a more complex model. Thus the simpler model may perform better, even if it fails to account for important degrees of freedom in the more complex one. Accounting for the impact of estimation noise on model performance is difficult, as it requires extrapolation to the condition where data are infinite [18, 22]. By assuming that estimation noise is additive, we found that a simple inverse relationship between estimation set size and prediction error accurately described performance for several different architectures (Fig 8A). In the limit of infinite data and under these assumptions, the FIR STRF did not perform significantly better than the simple parameterized model. These results should be confirmed with a larger dataset, but the current analysis suggests that the essential degrees of freedom for the linear STRF are much closer to 29 than to the 276 specified by the FIR STRF.
The average linear STRF in A1 may be described by about 30 parameters, but STRFs for individual neurons do vary substantially in their complexity. Some neurons require only one spectral channel for optimal performance while others require four or more channels (Fig 12B). The fact that only a minority of neurons were best described by a single dimension argues that most linear STRFs are not frequency-time separable [37, 51]. At the other extreme, even STRFs with four or five spectral channels required substantially fewer parameters than the standard FIR STRF.
This low dimensionality generalizes across other natural and synthentic stimuli in A1, but our analysis of data from the belt area dPEG indicates that more complex models are required for non-primary cortex (Fig 11). Moreover, even in A1, the full dimensionality of encoding models is likely to be greater than what is required to specify the linear STRF. As demonstrated by the enhanced performance of the nonlinear STP STRFs (Fig 11), introducing additional dimensionality that extends outside of the linear STRF architecture can improve model performance.
How well can the linear STRF actually describe sensory responses in A1? Issues surrounding finite sampling of experimental data again make it difficult to answer this question definitively [18, 45]. After implementing our estimation noise model, we found that the FIR STRF is able to account for 40% of A1 response variance on average (i.e., variance explained is 100R2 for R = 0.63, Fig 8). Factorized and parameterized STRFs very nearly matched performance of the FIR model (39% of response variance), indicating that these approximations capture the essential features of the more complex model, despite requiring only about 50% and 10% of the parameters, respectively. These measurements establish baseline performance by the linear STRF that must be surpassed by any more accurate model. At the Pareto frontier, a better model must either produce more accurate predictions or require fewer parameters and perform as well.
Only one previous study has attempted to answer this question rigorously, using activity driven by random chord stimuli in anesthetized mice [22]. Although we focused primarily on models that included an output nonlinearity [13, 14], we also computed asymptotic performance of STRFs without this nonlinear term in order to make a more direct comparison to the previous analysis of asymptotic performance. Without a spiking nonlinearity, the average FIR STRF was able to account for about 37% of response variance. This result falls in the range of 18–40% reported previously [22], although several factors make a direct comparison difficult. In the current study, recordings were performed in awake ferrets and used natural vocalizations rather than anesthetized mice and noise stimuli. Anesthesia can impact auditory neural activity [58, 59], and natural sounds evoke nonlinear response properties in a different functional domain than noise stimuli [9, 60].
The number of models tested here was relatively large, but they are still likely to be suboptimal compared to as-yet-untested parameterizations. The current study explored only two spectral parameterizations (Gaussian and Morlet functions) and the pole-zero family of IIR temporal filters. Numerous other basis functions could be considered, including Gabor wavelets [42, 61] or empirically-derived basis functions [29, 31, 33]. There is a clear trade-off between basis function complexity and the number of spectral dimensions needed. Better-performing temporal kernels like the P3Z1 filter reach their peak performance when D = 3, while simpler kernels like P1Z0 need D ≥ 4 to reach the same performance. Thus the interaction between channel count and basis function complexity will be relevant to identifying optimal parameterizations.
The efficiency of estimating parameterized STRFs allows the introduction of new, nonlinear terms that can account for encoding properties that are not captured by the linear model [30, 31]. When nonlinear short-term plasticity was introduced to the FIR STRF, it did not change model performance, but when it was introduced to the parameterized model, it improved predictive power. Thus the benefits of nonlinear terms may only become apparent when sufficient statistical power is available in the current dataset.
The family of models used in this study incorporate static nonlinearities that are commonly part of STRFs. This include log-compression of the input spectrogram to account for basilar membrane mechanics [25, 26] and an output nonlinearity to account for spike threshold and saturation [13, 14]. Other studies have incorporated nonlinear terms into the core computation of the filter. Some use general Volterra series expansions to account for second- and higher-order nonlinearities [3, 27, 57, 62]. Others incorporate more specific terms aimed at capturing contextual influences [28, 29] or mimicking biological circuit elements [26, 31]. These additional nonlinear terms can be incorporated into the parameterized framework, potentially providing substantial improvements in predictive power.
Neurons also undergo plasticity at multiple timescales due to stimulus context [12, 30, 63, 64], changes in behavioral state [50, 65, 66], and learning [49, 67]. In many experimental settings, the quantity of data available in a single behavioral state may provide a critical limitation on statistical power. Low-dimensional parameterized models may be particularly beneficial for exploring changes in spectro-temporal response properties in these experimental settings.
From a general analytical perspective, parameterization is similar to regularization during model estimation [1, 12, 46, 68]. In both cases, pre-existing knowledge or a hypothesis about the system’s function is used to constrain model fits. The idea that sensory receptive fields should vary smoothly in space and time has motivated the use of priors for smoothly varying STRFs [46, 68]. Similarly, the idea that receptive fields should have a relatively small number of non-zero parameters has motivated a sparse prior on model fits [14, 46]. Constraining the STRF to have analytical form of the factorized or parameterized models serves the same purpose of imposing a prior on the fit [38]. In the current study, the spectral and temporal parameterizations constrain both sparseness (limiting the model’s degrees of freedom) and smoothness (Gaussian spectral tuning and exponential temporal tuning).
To simplify model comparisons in this study, we used a single fit algorithm across all models. Thus it was not optimized specifically for the FIR STRF. Incorporating stricter stop criteria and sparseness constraints improve FIR STRF performance, but even after tuning the cost function, it did not match the performance of the parameterized model. The factorized and parameterized models were less sensitive to details of the fit algorithm such as the stop criterion, emphasizing the benefits of regularization effectively built into parameterization.
Most real world optimization problems involve the simultaneous minimization of several objectives [69]. Thus when comparing different model architectures, it may be helpful to consider trade-offs separately along different dimensions [36, 70, 71]. The current study focused in particular on the trade-off between model prediction accuracy and parameter count. In general, however, such an approach can be used to define an N-dimensional Pareto front containing the best models according to numerous other measures, including alternative performance metrics (Fig 13, see also [72]), alternative model complexity metrics [73, 74], data required to fit (Fig 8), computational cost [75], or model plausibility [76].
Pareto fronts are extensively used in the context of multiobjective optimization for the formulation of heuristics [69]. Given the complexity of performing a search on the space of model architectures, we relied here on inspection of the Pareto front to guide model design. While developing new analytic models to test, we found it most helpful to generate new models by adding to or discarding from a model on the current Pareto front. Variants of non-Pareto-optimal models rarely improved performance or provided insight into the relevancy of new parameters. Of particular note, the FIR implementation falls far from the Pareto front (Fig 2B), making it difficult to test variants based on the FIR STRF.
Single-unit neural activity was recorded from five awake, passively listening ferrets. For the main analysis of responses to vocalizations, a total of 176 single units were recorded in primary auditory cortex (A1) and 130 units in belt auditory cortex (dPEG). For one analysis (Fig 11C and 11D), responses were analyzed for 808 A1 units recorded during the presentation of continuous speech (reanalyzed from a previous publication [9]) and for 139 A1 units recorded during the presentation of 1/f noise [56].
Data used in this study will be made publicly available online via the Neural Prediction Challenge (http://neuralprediction.berkeley.edu/).
The relationship between the time-varying input auditory stimulus, x(t), and simultaneously recorded single-unit firing rate response, y(t), is described by the spectro-temporal receptive field (STRF [8, 9, 11, 12]) or, more generally, any function that maps x to y. In the current study, this mapping was cast as a sequence of functional modules, in which each function was applied to the output of the previous one (Eq 1, Fig 1C). The series of functions maps roughly to the physical elements that transmit auditory information to cortex. A detailed list of all models tested in this framework is included in S1 Table.
For most models, stimulus and response data were binned at 10 ms (100 Hz) and averaged across repetitions. Stimulus binning was applied after transformation to the spectrogram.
Data recorded from each neuron were divided into two subsets, one used only for model estimation (4–6 repetitions of 40 3-sec vocalization sequences) and the other for validation (20 repetitions of 2 3-sec sequences). Model parameters were fit using an iterated, greedy version of boosting that minimized mean-squared error prediction of the neural PSTH in the estimation dataset (details below). Each model was then evaluated based on its ability to predict the time-varying PSTH response in the reserved validation data set. Prediction accuracy was measured as the correlation coefficient (Pearson’s R) between the predicted and observed PSTH [12, 34]. The correlation coefficient provides a useful metric because it scales performance between 0 (completely random) and 1 (perfect correlation). Model performance can be variable across single neurons. Thus to compare models we focused on average performance across the entire set of neurons studied, using the nonparametric Wilcoxon signed rank test (sign test) to assess significant differences in performance. Error bars for average prediction correlation plots were computed on the difference between prediction correlation for each model and the FIR STRF fit to the same neuron. Computing error bars based on the difference per neuron removed variability in overall neural response SNR (e.g., Figs 4A and 10B) and revealed model differences commensurate with the sign test.
Our goal was to compare the ability of different analytical model structures to describe the neural data. Ideally, the details of the fitting algorithm used to fit the different models should not be relevant to this comparison, but in practice, there is no single algorithm that can be applied to different models without some bias [1]. Thus, the best fitting algorithm and model analytical structures are not separable in practice. We tested a variety of fit algorithms (Fig 2, S1 Table), but we focused on a single algorithm that performed best, on average, across all the models tested.
The fit algorithm consisted of nested iterations through each STRF module, initially optimizing each module with a conservative stop criterion. Once all modules had converged for the current stop criterion, its value was reduced and procedure was repeated for the smaller criterion. When fitting each module, two different coordinate descent algorithms were used. For non-parameterized modules (FIR filter, factorized spectral filter, and factorized temporal filter), a standard coordinate descent algorithm was used. For the remaining, parameterized modules (including the input filterbank and spike nonlinearities), greedy coordinate descent was used. The details of the fit algorithm are as follows:
In general, we found that fitting parameters separately within modules and iterating through modules with progressively smaller stop criteria helped avoid local minima. Fitting first without the spike nonlinearity also helped avoid local minima. The greedy algorithm increased the risk of overfitting complex models, but on average greatly improved predictions for models with nonlinear and parameterized modules. The non-greedy algorithm worked best for non-parameterized modules where all parameters are of similar scale.
Experimental procedures were approved by the Oregon Health and Science University Institutional Animal Care and Use Committee and conformed to standards of the National Institutes of Health.
|
10.1371/journal.pgen.1006969 | A mutation in Nischarin causes otitis media via LIMK1 and NF-κB pathways | Otitis media (OM), inflammation of the middle ear (ME), is a common cause of conductive hearing impairment. Despite the importance of the disease, the aetiology of chronic and recurrent forms of middle ear inflammatory disease remains poorly understood. Studies of the human population suggest that there is a significant genetic component predisposing to the development of chronic OM, although the underlying genes are largely unknown. Using N-ethyl-N-nitrosourea mutagenesis we identified a recessive mouse mutant, edison, that spontaneously develops a conductive hearing loss due to chronic OM. The causal mutation was identified as a missense change, L972P, in the Nischarin (NISCH) gene. edison mice develop a serous or granulocytic effusion, increasingly macrophage and neutrophil rich with age, along with a thickened, inflamed mucoperiosteum. We also identified a second hypomorphic allele, V33A, with only modest increases in auditory thresholds and reduced incidence of OM. NISCH interacts with several proteins, including ITGA5 that is thought to have a role in modulating VEGF-induced angiogenesis and vascularization. We identified a significant genetic interaction between Nisch and Itga5; mice heterozygous for Itga5-null and homozygous for edison mutations display a significantly increased penetrance and severity of chronic OM. In order to understand the pathological mechanisms underlying the OM phenotype, we studied interacting partners to NISCH along with downstream signalling molecules in the middle ear epithelia of edison mouse. Our analysis implicates PAK1 and RAC1, and downstream signalling in LIMK1 and NF-κB pathways in the development of chronic OM.
| Otitis media (OM) is the most common cause of deafness in children and is primarily characterised by inflammation of the middle ear. It is the most common cause of surgery in children in the developed world, with many children developing recurrent and chronic forms of OM undergoing tympanostomy tube insertion. There is evidence that a significant genetic component contributes towards the development of recurrent and chronic forms of OM. The mouse has been a powerful tool for identifying the genes involved in chronic OM. In this study we identified and characterised edison, a novel mouse model of chronic OM that shares important features with the chronic disease in humans. A mutation in the Nisch gene causes edison mice to spontaneously develop OM following birth and subsequently develop chronic OM, with an associated hearing loss. Our molecular analysis of the mutation reveals the underlying pathological mechanisms and pathways involved in OM in the edison mouse, involving PAK1, RAC1 and downstream signalling in LIMK1 and NF-κB pathways. Identification of the edison mutant provides an important genetic disease model of chronic OM and implicates a new gene and genetic pathways involved in predisposition to OM.
| Otitis media (OM) is characterised by inflammation of the middle ear (ME), often associated with a conductive hearing impairment, and is the commonest cause of hearing loss in children. It is perceived by many to be a transient affliction that in reality places a substantial social, medical and economic burden on healthcare systems globally [1]. Evidence from studies of the human population suggests that there is a significant genetic component predisposing to the development of recurrent and chronic forms of OM [2,3]. Despite the importance of the disease, many of the genes involved in OM susceptibility have still yet to be identified. At present, the use of mouse models is the most promising method to identify candidate loci underlying susceptibility to OM. Mouse models have highlighted the role of Toll-like receptors (TLRs) in acute OM, in particular the protection against commensal and pathogenic bacteria, and that persistent NF-κB or TGF-β signalling could be two mechanisms leading to the overactive pro-inflammatory response seen in chronic OM [4,5].
The large-scale phenotype-driven mouse ENU (N-ethyl-N-nitrosourea) mutagenesis program at MRC Harwell [6,7] has previously identified two novel mouse mutants, Jeff and Junbo, that develop a conductive hearing loss characterised by ME fluid and mucosal inflammation [8,9]. The Jeff mouse has a mutation in the Fbxo11 gene [10] and the Junbo mouse has a mutation in Evi1 [9]. Studies have revealed that these genes, are involved in signalling of the TGF-β superfamily, via SMAD proteins [11,12]; negatively regulate NF-κB–dependent inflammation [13]; and highlight the role of HIF–VEGF pathways in the underlying genetic and pathophysiological mechanisms that predispose to chronic OM [14].
OM mouse models with single gene mutations have identified a number of genes as candidate susceptibility genes for human OM, including Tlr2, Tlr4, p73, E2f4, Plg, Tgif1, Evi1 and Fbxo11 [4]. These genes identified from mouse models of OM are beginning to be studied in the human population; with significant associations between OM and polymorphisms in FBXO11 [15,16], TLR2 [17] and TLR4 [17–19].
We have identified and characterised a novel OM mouse mutant, edison, from the ENU mutagenesis program at MRC Harwell. Homozygous edison mice spontaneously develop a conductive hearing loss associated with chronic inflammation of the ME, sharing many features with chronic OM in humans. The underlying mutation of this phenotype has been identified as a mutation in the Nisch gene. We have explored the role of Nisch in chronic OM, relating the edison phenotype to the underlying mechanisms of Nisch function. We have utilised double mutants to assess genetic interactions and pathways involved, implicating PAK1 and RAC1, and downstream signalling events in LIMK1 and NF-κB signalling pathways in the development of chronic OM. Overall, the edison mouse highlights a new candidate gene for susceptibility to chronic OM and has provided further insight into the genetic pathways and pathogenic processes involved.
A phenotype-driven ENU mutagenesis screen [20] identified a new recessive mutant, edison (edsn), with hearing loss. Preliminary phenotyping using a click-box test (20 kHz, 90 dB sound pressure level (SPL) tone burst) of an age-matched cohort derived from the founder mouse indicated that 6-week-old (wk) mice demonstrated a reduced startle response. SNP-based linkage analysis and mapping identified an approximately 9 Mb interval on chromosome 14 delineated by marker rs30778552 and rs46823676 containing 119 genes (Fig 1A). Whole-genome sequencing identified 60 ENU-induced, de novo variants within the critical 9Mb interval and importantly only one missense variant was identified. This missense variant was a c.3079T>C substitution in Nischarin (Nisch) (open reading frame of NCBI RefSeq transcript NM_022656.2) that results in a Leu972Pro substitution (Fig 1B). The change occurs in a highly conserved region that has been maintained through evolution (Fig 1C). PROVEAN analysis predicts that this change is ‘‘deleterious” and SIFT predicts that it is ‘‘not tolerated”. No other non-synonymous sequence changes were identified within the minimal interval. The Nisch locus encodes a protein of 1,593 amino acids, coded for by 22 exons. The protein consists of an N-terminal phox homology (PX) domain, six putative leucine-rich repeats (LRRs), a predicted coiled-coil (CC) domain, an alanine/proline-rich region and a long C-terminal region (Fig 1D).
DNA and sperm archives derived from ENU mutagenesis programmes [21] were utilised to identify an additional allele at the Nisch locus. We screened ten exons of Nisch employing high resolution melting analysis of ~10,000 mutant mice and identified a c.98T>C substitution resulting in a Val33Ala substitution within a conserved region of the NISCH PX domain. We rederived this second allele, NischV33A, and examined the phenotypes.
NISCH binds to the cytoplasmic domain of ITGA5 [23], and is thought to regulate its expression [24]. Cross-talk between VEGF and integrins has been shown to be a critical factor in the regulation of angiogenesis and vascularization [25]. Given these findings we examined the genetic interaction between Nisch and Itga5. Nischedsn/+ mice and Itga5tm1Hyn/+; Nischedsn/+ double heterozygotes were intercrossed to produce Itga5tm1Hyn/+; Nischedsn/edsn littermates for the study, with Itga5tm1Hyn/+; Nisch+/+ and Itga5+/+; Nischedsn/edsn progeny as littermate controls (Fig 5).
Itga5+/+; Nisch+/+, Itga5+/+; Nischedsn/+ and Itga5tm1Hyn/+; Nisch+/+ mice showed normal click ABR thresholds across the study (Fig 5A). Both Itga5+/+; Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice displayed a progressive hearing loss with onset at 4 wk. Interestingly, Itga5tm1Hyn/+; Nischedsn/edsn mice exhibited significantly elevated auditory thresholds compared to Itga5+/+; Nischedsn/edsn mice, throughout the time course (Kruskall-Wallis: p < 0.01). At 20 wk, Itga5tm1Hyn/+; Nischedsn/edsn animals were recorded with a mean click ABR threshold of 61 ± 3 dB SPL, compared to 40 ± 4 dB SPL for Itga5+/+; Nischedsn/edsn mice. There was a consistently high prevalence of OM in Itga5tm1Hyn/+; Nischedsn/edsn mice compared to Itga5+/+; Nischedsn/edsn animals (Fig 5C and 5D). Itga5tm1Hyn/+; Nischedsn/edsn mice exhibited significantly increased OM prevalence at 4 wk compared to Itga5+/+; Nischedsn/edsn mice (Fisher Exact: p = 0.026). At 4 wk, 67% of Itga5tm1Hyn/+; Nischedsn/edsn mutants had bilateral OM and 33% had unilateral OM (n = 12), whereas in Itga5+/+; Nischedsn/edsn mice at 4 wk, 13% had bilateral OM, 63% unilateral OM and 25% showed no OM phenotype (n = 8). By 20 wk, the difference in OM prevalence observed between Itga5tm1Hyn/+; Nischedsn/edsn and Itga5+/+; Nischedsn/edsn mice was still increased (Fisher Exact: p = 0.051). In Itga5tm1Hyn/+; Nischedsn/edsn mutants at 20 wk, 83% had bilateral OM and 17% had unilateral OM (n = 12). While, in Itga5+/+; Nischedsn/edsn mice at 20 wk, 38% had bilateral OM, 25% unilateral OM and 37% showed no OM phenotype (n = 8). One Itga5tm1Hyn/+; Nisch+/+ mouse (n = 15) displayed unilateral OM at 4 wk, with no other recordings at later time points (S4A Fig).
Histological examination confirmed that Itga5+/+; Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice develop chronic OM (Fig 5E–5J). Itga5tm1Hyn/+; Nischedsn/edsn mice displayed a more severe mucosal inflammation, with increased polypoid exophytic growths and a thick cellular effusion. Blinded assessment of the mucoperiosteum thickness (Fig 5K) indicated that both Itga5+/+; Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice had significant mucosal thickening compared to wild-type littermates (Kruskall-Wallis: p = 0.003 and p < 0.001 respectively). Additionally, when only OM ears from Itga5+/+; Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice were compared (Fig 5L), there was a significant increase in mucosal thickness observed in Itga5tm1Hyn/+; Nischedsn/edsn mice (Itga5+/+; Nischedsn/edsn, 97.8 ± 12.12 μm, n = 8; Itga5tm1Hyn/+; Nischedsn/edsn, 137.9 ± 10.85 μm, n = 9; Kruskall-Wallis: p = 0.026).
Additionally, double heterozygotes (Itga5tm1Hyn/+; Nischedsn/+) exhibited a mild late-onset hearing loss, with onset at 12 wk (Fig 5A). Visualisation of the tympanic membrane showed this mild hearing loss was associated with a late-onset in the prevalence of OM from 12 wk in Itga5tm1Hyn/+; Nischedsn/+ mice (S4B Fig). By 20 wk, Itga5tm1Hyn/+; Nischedsn/+ mice exhibited significantly increased prevalence of OM compared to wild-type littermates (Fisher Exact: p = 0.041). In Itga5tm1Hyn/+; Nischedsn/+ mutants at 20 wk, 31% had unilateral OM and 69% had no OM phenotype (n = 13). Histological examination confirmed that Itga5tm1Hyn/+; Nischedsn/+ mice develop chronic OM (S4D and S4F Fig). The mucosal inflammation was diffuse and of mild severity, with the presence of a cellular effusion. Blinded assessment of the mucoperiosteum thickness indicated that Itga5tm1Hyn/+; Nischedsn/+ mice had significant mucosal thickening compared to wild-type littermates (Fig 5K).
We proceeded to investigate the expression of interacting partners to NISCH, including ITGA5, as well as key downstream effectors in order to relate the underlying mutation to the edison phenotype. In addition to binding ITGA5, NISCH has also been shown to interact directly with PAK1 [23,26], and RAC1 [27]. As well as being a downstream effector of PAK1, LIMK1 [28] also interacts directly with NISCH. In addition PAK controls NF-κB activation [29] and RAC1 leads to activation of NF-κB [30,31]. We thus performed IHC expression analysis of NISCH, ITGA5, phosphorylated-PAK1 (p-PAK1), phosphorylated-LIMK1/2 (p-LIMK1/2), RAC1 and NF-κB p65 on ME epithelia from wild-type, Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice. In addition, we assessed protein expression in ME epithelia by western blot analysis for each of these proteins.
IHC labelling for NISCH, ITGA5 and p-PAK1 was observed in ME epithelial cells with similar patterns of expression in Nisch+/+, Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice. No obvious differences in localisation were observed for these three proteins, although staining was stronger in Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice (Fig 6A–6C). Western analysis of ME epithelial cell lysates showed raised levels of ITGA5 in Nischedsn/edsn mice but not significantly different compared to wild-type (t-test: p = 0.071) and not surprisingly a significant decline in levels between Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice (t-test: p = 0.020) (Fig 7A). PAK1 protein levels were significantly raised in both Nischedsn/edsn (t-test: p = 0.002), and Itga5tm1Hyn/+; Nischedsn/edsn mice (t-test: p = 0.006) when compared to wild-type, but not between Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice (t-test: p = 0.282) (Fig 7B).
Available antibodies for p-LIMK1 were ineffective. However, we carried out expression analyses using a p-LIMK1/2 and LIMK1 antibody. IHC analysis of p-LIMK1/2 expression in ME epithelial cells revealed nuclear localisation and increased expression from both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice (one-way ANOVA: p < 0.001) compared to wild-type. However, no significant difference was observed between ME epithelial cells in Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice (one-way ANOVA: p = 0.704) (Fig 6D). Similarly, protein levels of p-LIMK1/2 in ME epithelial cells were significantly raised in Nischedsn/edsn (t-test: p = 0.039) and Itga5tm1Hyn/+; Nischedsn/edsn (t-test: p = 0.038) compared to wild-type. No difference was detected between Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn samples (t-test: p = 0.598), consistent with the observations from immunohistochemistry (Fig 7C). In addition, levels of LIMK1 in ME epithelial cells were also significantly raised in Itga5tm1Hyn/+; Nischedsn/edsn compared to wild-type (t-test: p = 0.030) and Nischedsn/edsn (t-test: p = 0.041) mice (Fig 7D).
IHC analysis of RAC1 demonstrated nuclear localisation and increased expression from Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type (one-way ANOVA: p < 0.001) (Fig 6E). Moreover, in agreement with these observations, protein levels of RAC1 were significantly raised in Itga5tm1Hyn/+; Nischedsn/edsn mice (t-test: p = 0.033), compared to wild-type (Fig 7E).
IHC analysis of NF-κB p65 expression showed nuclear localisation and raised levels of protein in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type (one-way ANOVA: p = 0.017 and p <0.001 respectively). Moreover, there was a significant enhancement of NF-κB labelling in the double mutant compared to Nischedsn/edsn (one-way ANOVA: p = 0.026) (Fig 6F). In agreement with IHC, we found significantly higher levels of NF-κB p65 by western blot in both mutants compared to wild-type (t-test: p = 0.039 and p = 0.041) (Fig 7F). We also examined levels of activated NF-κB p65 by western blot employing an antibody that recognises phosphorylated-NF-κB [Ser 276] p65 (p-NF-κB p65) and detected significantly higher levels of the activated protein in both mutants compared to wild-type (t-test: p = 0.030 and p = 0.035) (Fig 7G).
In addition, we investigated the expression of two pathways that are implicated in the development of chronic OM (see Introduction), or are regulated by NISCH interacting partners. First, we evaluated levels of focal adhesion kinase (FAK) and SRC. Cross-talk between integrins and VEGF is a critical factor in the regulation of angiogenesis, vascularisation and vascular permeability [25]. Integrins regulate VE-cadherin via the activation of SRC, and FAK catalytic activity is required when α5β1 integrin stimulates SRC activation through FAK phosphorylation. FAK inhibition prevents VEGF-stimulated vascular permeability underlining the importance of FAK activity in the regulation of adherens junctions [32]. Tyr 397 in human FAK becomes phosphorylated upon integrin engagement and creates a binding site for SRC. This results in release of the inactive conformation of SRC (Tyr 527) and leads to autophosphorylation of SRC on Tyr 416. Activated SRC further phosphorylates FAK on additional residues, one of which is Tyr 576. The activated FAK-SRC complex then initiates multiple downstream signalling pathways [33,34]. IHC of ME epithelial cells showed there was increased expression of total FAK in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn compared to wild-type mice (one-way ANOVA: p = 0.015 and p < 0.001 respectively). A significant difference was also observed between Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn ears (one-way ANOVA: p = 0.005) (Fig 6G).
To study protein levels of FAK we used an antibody raised against the last 50 amino acids at the C-terminal of the human protein. There are nine known mouse isoforms of FAK (http://www.uniprot.org/uniprot/P34152) produced by alternative promoter usage and alternative splicing and the antibody is potentially able to detect six of them. We detected three main bands in ME epithelial cell lysates at 124, 115 and 100kDa. The full-length canonical isoform is 124kDa. We detected significantly increased levels of the full length 124kDa FAK1 protein in Nischedsn/edsn mutants compared to wild-type (t-test: p = 0.014). The level of the 115kDa form was significantly raised in Nischedsn/edsn (t-test: p = 0.014) and Itga5tm1Hyn/+; Nischedsn/edsn (t-test: p = 0.001) mice compared to wild-type, while the 100kDa form was the main isoform detected in wild-type samples. A significant difference was detected in the levels of the 100kDa between wild-type and Itga5tm1Hyn/+; Nischedsn/edsn (t-test: p = 0.0005) (Fig 7H). Furthermore we used phosphorylated-FAK (p-FAK) [Y576] and phosphorylated-SRC (p-SRC) [Y527] antibodies to test the activity of the FAK-SRC complex in middle ear epithelia. Using a p-FAK [Y576] antibody, which recognises activated FAK, we detected an increase in the activated FAK in the Nischedsn/edsn tissue compared to the wild-type ME epithelial cell lysate (t-test: p = 0.007) (Fig 7I). Not surprisingly, in the double mutant Itga5tm1Hyn/+; Nischedsn/edsn, levels of activated FAK were not significantly different to wild-type. Using a p-SRC [Y527] antibody, which recognises inactive SRC, we detected complementary results to that seen with activated FAK. We found reduced levels of protein in Nischedsn/edsn ME lysates compared to the wild types (t-test: p = 0.002) but no differences between wild-type and Itga5tm1Hyn/+; Nischedsn/edsn (Fig 7J).
Finally, we proceeded to evaluate activation of the TGF-β pathway during chronic ME disease by assessment of phosphorylated-SMAD2 (p-SMAD2). IHC analysis of p-SMAD2 revealed significantly raised levels in Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type (one-way ANOVA: p < 0.001). However, there was no significant difference between Nischedsn/edsn and wild-type mice (one-way ANOVA: p = 0.680) (Fig 6H). In addition, protein levels of p-SMAD2 in ME epithelial cells were significantly higher in Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type (t-test: p = 0.029) or Nischedsn/edsn (t-test: p = 0.024) mice. Again no difference was detected between wild-type and Nischedsn/edsn mice (t-test: p = 0.864) (Fig 7K).
IHC analysis of this suite of proteins in airway epithelia revealed many similarities to that seen in ME epithelia (S5 Fig). Levels of NISCH appeared to be raised in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type. Furthermore, increased epithelia expression of p-LIMK1/2 in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice mirrored the findings in ME epithelia. Similarly, RAC1 epithelial levels were significantly raised in Itga5tm1Hyn/+; Nischedsn/edsn mice, while NF-κB and FAK expression was raised in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice. However, we observed no significant changes in p-SMAD2 levels in mutant airways. In contrast, in lung tissue we observed very few significant changes in protein levels by western blot analysis (S5 Fig). These results may reflect the complexity of tissues and cell types isolated from dissected material and the visible mesenchymal expression for many of these proteins in airway tissue.
In a large-scale ENU mutagenesis screen we recovered a new recessive mouse model of chronic OM, edison. The edison mutant carries a Leu972Pro change in the Nischarin gene. Nischedsn/edsn homozygotes display a progressive middle ear disease with 56% of mice displaying bilateral OM by 20 weeks and elevated ABR thresholds of 20–30 dB SPL indicative of a conductive hearing loss. We derived an additional ENU allele (NischV33A) in the Nischarin gene which also presents with progressive chronic OM, but where ABR thresholds were only very moderately increased and at 12 weeks only 10% of the mice had bilateral OM. It appears that the NischV33A allele is severely hypomorphic. Compound heterozygotes of the edison and NischV33A alleles showed an intermediate non-complementing OM phenotype. We did not identify any sensorineural element to the hearing loss in the Nischedsn/edsn mutant.
The chronic OM in Nischedsn/edsn mice is exemplified by exudate within the ME cavity and a thickened mucoperiosteum and polypoid exophytic growths, sometimes associated with an inflamed tympanic membrane. Serous or granulocyte-rich effusions were observed, but as the mice aged a thick effusion was predominately observed rich in macrophages and PMNs. Examination of ME exudates for upregulation of both inflammatory genes and hypoxia genes found that both Il-1b and Tnfa were both upregulated, as were Hif1a and the HIF responsive gene Vegfa. This is similar to the findings reported for other mouse models of chronic OM, such as the Jeff [8], Junbo [9] and Tgif1 [35] mutants. Both the middle ear and the lungs have substantial similarities in structure and function [36] and, intriguingly, we identified a lung defect in Nischedsn/edsn mice. During embryonic development, we found a significant reduction in airspace width, while in the adults, we observed an emphysema-like phenotype with enlarged airspace width and a reduced number of airspaces. In utero, the network of airways is generated first followed by formation of the gas-exchanging units (alveoli) that develop from the distal ends of the small airways. As a consequence disruption to lung development often results in narrower airspaces in embryonic lungs but enlarged airspaces post-natally because of insufficient generation of alveoli. These lung abnormalities likely account for the deficit of Nischedsn/edsn mice recovered from the various crosses that we report.
The discovery of the involvement of Nischarin in the development of inflammatory middle ear disease identifies a novel gene and associated pathways that are involved in OM. This led us to explore the intersection with known pathways of OM [11,13,14] and the downstream signalling mechanisms that lead to the OM phenotype. Nischarin is a highly conserved protein across mammalian species, consisting of an N-terminal phox homology (PX) domain, 6 putative leucine-rich repeats, a coiled-coil domain, an alanine/proline-rich region and a long C-terminal region. NISCH has a multitude of interacting partners, including ITGA5 [37], PAK1 [26], Rac1 [27,38], LIMK1 [28], Rab14, PI3P [38], and LKB1 [39]. Association of NISCH with these interacting partners underlines its broad impact on the regulation of cell motility, cell invasion, vesicle maturation, as well as its role as a tumour suppressor [28,37,38,40]. Most notably, the binding of NISCH to ITGA5 [37] is thought to mediate the translocation of ITGA5 from the cell membrane to endosomes [24] thus regulating ITGA5 levels. As we discuss above, integrins have been shown to play a critical role in modulating VEGF-induced angiogenesis and vascularization [25]. These pathways thus intersect with the hypoxia-response pathways mediated by HIF-1a which have been demonstrated to be involved with the development of chronic OM in the Junbo and Jeff models [14]. Hypoxia leads to upregulation of VEGFA and downstream pathway genes resulting in VEGF-induced angiogenesis and vascular leak. VEGFR inhibitors moderated angiogenesis and lymphangiogenesis in the Junbo mouse. For these reasons we sought to explore the role of ITGA5 in the edison mutant, and also to characterise the responses of downstream pathways which may illuminate the mechanisms of OM development in edison.
We found a strong genetic interaction between edison and Itga5 mutants. Itga5tm1Hyn/+; Nischedsn/edsn double mutants compared to Nischedsn/edsn mice showed significantly elevated ABR thresholds as well as a very significant raised frequency of bilateral OM in mice from 4 weeks onwards. In summary, the OM was more highly penetrant from an earlier age, and commensurately less progressive. Given the interactions of ITGA5 and NISCH, along with the interactions of NISCH with diverse molecules involved in LIMK1 and NF-κB signalling, we sought to interpret this genetic interaction in the context of known signalling pathways and interactions downstream of NISCH. Moreover, in developing a mechanistic model, we took into account the reports that interaction of ITGA5 with NISCH affects some of the downstream interactions of the NISCH molecule itself.
RAC1 signalling regulates disparate cellular functions mediated through a variety of effector proteins [30,31]. PAK1 is a key downstream effector of RAC1 [41,42], with binding of RAC1 leading to activation of PAK1 [43]. NISCH represses this pathway, and NISCH has been shown to block RAC1 induced cell migration through binding to PAK1; interaction with NISCH strongly inhibits the kinase activity of PAK1 [23]. RAC1 activation of PAK1 enhances the interaction between NISCH and PAK1, while notably expression of ITGA5 also increases the association between NISCH and PAK1 [23]. LIMK1 is a downstream effector of PAK1 [44] and has been shown to be involved in vascular permeability [45]. NISCH is also an interacting partner with LIMK1, regulating cell invasion through repression of the LIMK1-cofilin pathway [28]. NISCH also regulates PAK1-independent RAC1 signalling through direct interaction with RAC1 [27]. RAC1 stimulates the phosphorylation and degradation of IkB and up-regulates NF-κB [46,47]. Overexpression of NISCH has been shown to suppress the ability of RAC1 to stimulate NF-κB activation [27]. The Junbo OM mutant carries a mutation in the Evi1 gene, and it is noteworthy that EVI1 is a negative-feedback regulator of NF-κB [13]. The mutation in Junbo leads to activation of NF-κB and inappropriate regulation of the inflammatory response [13].
We have assessed the expression of the critical genes within these pathways in the ME epithelia of wild-type, Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice. First, we observed that levels of ITGA5 were raised in Nischedsn/edsn mice though not significantly, reflecting the role of NISCH in regulating ITGA5 levels. Not surprisingly, ITGA5 was significantly reduced compared to edison mice in the double mutant, Itga5tm1Hyn/+; Nischedsn/edsn. We found significantly raised levels of activated p-PAK1 in both mutant lines compared to wild-type, which was mirrored by downstream raised levels of p-LIMK1/2 in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice. Raised levels of p-LIMK1/2 were also reflected in the IHC assessment. LIMK1 levels were also raised in Nischedsn/edsn Itga5tm1Hyn/+ mice. We were unable to assess directly levels of p-LIMK1 due to the ineffectiveness of available antibodies. While RAC1 protein levels were not affected in Nischedsn/edsn mice, they were significantly raised in the double mutant, Itga5tm1Hyn/+; Nischedsn/edsn mice which was mirrored in the IHC assessment. We also found by both IHC and western analysis that NF-κB levels were raised in both Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice compared to wild-type. Moreover, there is evidence from IHC that the effect of the two mutations are additive and that the double mutant shows a significantly higher level of NF-κB expression compared to Nischedsn/edsn mice. While the protein analysis shows a similar trend the differences between Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice are not significant.
Overall, our analysis indicates that the edison mutation leads to activation of PAK1 and PAK1-independent RAC1 pathways with increased levels of NF-κB and p-LIMK1/2 each of which may lead to inflammatory and vascular permeability effects. In addition, a combination of mutations in NISCH and ITGA5 can lead to an exacerbation of raised protein levels which may underlie the more severe phenotype seen in the double mutant, Itga5tm1Hyn/+; Nischedsn/edsn. This may reflect the role of ITGA5 in enhancing binding of NISCH to PAK1, and provides us with a model of the mechanism underlying the Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn phenotypes (Fig 8). We surmise that impairing function of NISCH leads to derepression of RAC1 pathways, while reducing levels of ITGA5 in combination with impaired NISCH leads to further derepression and activation of downstream pathways manifested in the more severe phenotype. However, we did not observe significantly raised levels of RAC1 in the single mutant. Nevertheless, we found raised levels of NF-κB that may reflect activation of PAK1 independent pathways.
ITGA5 is known to be involved with the activation of SRC protein tyrosine kinase, as well as FAK, both of which are involved with mediating VEGF induced vascular leak [32]. ITGA5 can mediate its effects by phosphorylation of FAK and binding of activated FAK to SRC leading to conformational SRC activation [33]. Alternatively, for example in the context of neuroblastoma cell motility, FAK is required for integrin α5β1-mediated SRC phosphorylation [34]. We thus investigated FAK and SRC levels in the mutant mice, focusing on activated FAK and inactive SRC. We found total levels of the full length isoform of the FAK protein raised in Nischedsn/edsn mice. However, in Itga5tm1Hyn/+; Nischedsn/edsn mice, full length FAK protein was not significantly different to wild-type. Importantly, we detected raised levels of activated FAK along with reduced levels of inactive SRC in Nischedsn/edsn mice. These changes may contribute to angiogenesis and vascular permeability through upregulation of VEGF signalling pathways.
The genes mutated in previously characterised OM models (Junbo, Jeff and Tgif1), have been reported to regulate the TGF-β signalling pathway [11,12,35]. There is considerable cross-talk between TGF-β signalling and hypoxia pathways and mutations that perturb the TGF-β pathway might be expected to perturb hypoxia responses with downstream consequences on VEGFA and VEGF signalling, as is observed in the mutants studied [14]. We discuss above the raised levels of HIF-1a and VEGFA found in the edison mutant. We thus assessed TGF-β signalling (in edison), and found that p-SMAD2 levels are significantly raised in Itga5tm1Hyn/+; Nischedsn/edsn mice, but not in Nischedsn/edsn. The lack of raised p-SMAD2 levels in the Nischedsn/edsn mouse suggests that activation of TGF-β signalling is not a primary event underlying the development of chronic OM in the edison mouse. Rather, the upregulation of p-SMAD2 in the double mutant may reflect the very severe inflammatory state of the middle ear leading to activation of TGF-β pathways.
In summary (see Fig 8), we conclude that mutations in Nischarin impact upon PAK-dependent and PAK-independent RAC pathways with downstream signalling effects on LIMK1 and NF-κB. Moreover, the interplay between Nischarin and ITGA5 likely underlies impacts on FAK and SRC signalling which may lead to VEGF mediated vascular leak. The combined effects on LIMK1, NF-κB and FAK signalling can account for the observed inflammatory changes in the edison middle ear, which along with vascular leak and middle ear exudate, produce a chronic OM. Together, the pathways changes we describe provide a mechanism underlying the observed phenotypic changes in edison. Moreover, these studies further enhance our knowledge of the relevant genetic pathways that contribute to middle ear inflammatory disease, as well as a panel of new genes that are candidates for genetic susceptibility to chronic OM in the human population.
Mice were bred and maintained by Mary Lyon Centre, MRC Harwell and were housed in specific-pathogen free conditions. All animal experimentation was approved by the Animal Welfare and Ethical Review Body at MRC Harwell (License Numbers: 30/3015 and 30/3280). The humane care and use of mice in this study was under the authority of the appropriate UK Home Office Project License.
The founder mouse carrying the edison mutation was generated in a large-scale phenotype-driven ENU mutagenesis program at MRC Harwell [20]. Briefly, Male C57BL/6J mice were mutagenized and mated to C3H.Pde6b+ females (a C3H stock that does not carry the retinal degeneration allele Pde6brd). G3 offspring were screened for a variety of abnormalities, including deafness and vestibular dysfunction. The edison founder was identified due to lack of a Preyer reflex when presented with a calibrated 20 kHz, 90 dB SPL tone burst via a click-box test.
The founder edison mouse was maintained by repeated outcrossing to C3H/HeH and intercrossing to produce homozygous mutant progeny, identified by the lack of a Preyer reflex. For linkage analysis genomic DNA from 13 affected mice were screened with 63 strain specific SNP markers spaced equidistantly across the genome using the Pyrosequencing SNP genotyping system (QIAGEN). Additional SNP markers were used within linked regions to further fine map the causal mutation. Markers and primer sequences are available on request from the authors.
Genomic DNA from a single affected edison mouse was sent for next-generation sequencing (High-Throughput Genomics, WTCHG). Identification, analysis and dissemination of sequence variant information identified through whole-genome sequencing were achieved with the tools provided by the MRC Harwell Biocomputing custom sequence analysis pipeline. Sequence reads were mapped to the NCBI37/mm9 assembly of the reference mouse genome. Within the critical interval, the mean read depth was 8.17 and the read coverage was 99.64%. Sequence variants were identified and were subsequently categorised into those which occurred within exons and splice donor/acceptor sites, and were not known or strain-specific variants. This led to the identification of a T to C base substitution within exon 14 of Nisch, which is predicted to cause a Leu972Pro missense change in NISCH protein. This sequence variant was validated using Sanger sequencing (Source BioScience, UK). Affected and unaffected mice were genotyped using the LightScanner SNP genotyping system (Idaho Technology Inc., USA) to confirm the association of genotype to phenotype. In all cases the genotype correlated with the phenotype.
DNA from the MRC Harwell ENU-DNA sperm archive (http://www.har.mrc.ac.uk/services/archiving-distribution/enu-dna-archive) was screened with the LightScanner platform (Idaho Technology Inc., USA). Briefly, male C57BL/6J mice were treated with ENU and crossed to C3H/HeH females. F1 progeny (C3H/HeH.C57BL/6J) were rederived and male F1 animals had sperm and DNA samples taken for archiving. Ten exons of Nisch were screened in DNA from ~10,000 F1 ENU mutagenised animals and potential mutations confirmed with Sanger sequencing (Source BioScience, UK).
All edison mice used for phenotyping were congenic on a C3H/HeH background. They were backcrossed for at least ten generations. The NischV33A strain was rederived by in vitro fertilisation of C57BL/6J oocytes with F1 sperm from the MRC Harwell ENU-DNA sperm archive and maintained on a mixed C3H/HeH and C57BL/6J genetic background. NischV33A/+ mice were intercrossed for phenotypic analysis and crossed to congenic Nischedsn/+ mice for complementation testing. Cryopreserved Itga5tm1Hyn mutant sperm was imported from the Jackson Laboratory (Stock No. 002274) [48] and the colony was rederived using in vitro fertilisation of C57BL/6J oocytes. Itga5tm1Hyn mice had been backcrossed 4–5 generations onto a C3H/HeH background, before crossing to congenic Nischedsn mice for phenotypic analysis.
Genotyping for edison mice was performed using an allelic discrimination assay with primers 5’-GGC AGC ACA AAG ATG GCG GTA AC-3’ and 5’-AAC TGC CGC AAC CGC AAC A-3’ and labelled probes 5’-[6-FAM]_AGC AGC TCG AGC ACA T-3’ (edsn) and 5’-[TET]_CAG CTC GGG CAC ATG-3’ (Wild-type). The Applied Biosystems 7900HT Fast System (Applied Biosystems, USA) was used for amplification and analysis. To genotype NischV33A mice, PCR amplification was performed with primers 5’-GAC TGA GTA CCT TGC AGC TA-3’ and 5’-CTG TAA CGG TGT TTG ATC GTC-3’ and an unlabelled probe 5’-CCC TTT AGG CTT ATG TCA TCC AGG TTA C_[SpC3]-3’. The LightScanner System (Idaho Technology Inc., USA) was used for subsequent unlabelled probe genotyping analysis. Itga5tm1Hyn mice were genotyped with mutant and control specific primers, as described on the Jackson Laboratory Mice Database (http://jaxmice.jax.org/strain/002274).
Analysis of skulls from 20 wk mice was performed using a Faxitron Mx-20 DC-4 specimen X-ray System. ImageJ software was used to measure the skull length, nasal bone length, frontal bone length, parietal bone length and skull width. Allometric comparisons were performed against skull length with at least 12 mice of each genotype.
Mice were anesthetized and hearing thresholds determined using ABR, as previously described [7]. The ABR threshold was measured for each ear. Click-evoked hearing assessments for edison mice were conducted at 3, 4, 6, 8, 12, 16 and 20 wk with cohorts containing at least 14 mice of each genotype at each time point. Frequency-specific (8, 16, and 32 kHz) analysis of auditory function for edison mice was conducted over a longitudinal time course at 4, 6, 8, 12 and 20 wk with 5 mice of each genotype. At least 10 NischV33A mice were used for click-evoked ABR analysis across a longitudinal time course at 4, 6, 8 and 12 wk. Finally, click-evoked analysis of Itga5tm1Hyn Nischedsn hearing thresholds were measured across a longitudinal time course at 4, 6, 8, 12, 16 and 20 wk with at least 8 mice of each genotype.
Mouse 3, 4, 6, 8, 12, 16 and 20 wk heads from Nisch+/+, Nischedsn/+ and Nischedsn/edsn mice were fixed for 48 hours in 10% neutral buffered formaldehyde, decalcified in D.F.B decalcifying agent (Kristensen; Pioneer Research Chemicals) for 72 hours and embedded in paraffin following routine procedures. Perinatal heads were processed in the same manner without the decalcification steps. Four-micrometre-thick sections were obtained, de-paraffinized in xylene substitute and rehydrated via a graded ethanol. For morphological observations, sections were stained with haemotoxylin and eosin (H&E). The histological sections were used to investigate the middle ear inflammation of the mice. Evaluation of mean mucosal thickness was by blinded assessment of a standard 1000 μm length of ME mucosa (avoiding the cochlea and the region close to the Eustachian tube), the mucosal thickness was averaged from five measurements.
To study the lung morphology of edison mice, H&E stained sections from adult and perinatal lungs were viewed using a Zeiss Axiostar Plus bright-field microscope and analysed using cellB imaging software (Olympus). The data was analysed as previously described [49].
To study the ultra-structure of the organ of Corti we dissected the inner ears from five 20 wk Nisch+/+ and Nischedsn/edsn mice and prepared the samples as previously described [50]. Inner ears were imaged using a JEOL 6010 LV scanning electron microscope under high vacuum conditions.
Blood and bulla fluids were collected, as previously described [14], for analysis using Real-time quantitative PCR.
Real-time quantitative PCR was performed as previously described [14]. For ear fluid analysis, each sample pool comprised the fluid from both ears of four individual samples. For blood analysis each sample pool comprised four individual samples. Murine TaqMan gene expression assays used for analysis were Hif1a (Mm01283756_m1), Il1b (Mm01336189_m1), Tnfa (Mm00443258_m1), Vegfa (Mm00437304_m1), Src (Mm00436785_m1), Evi1 (Mm00491303_m1) and Fbxo11 (Mm01227499_m1). Ppia (Mm02342429_g1) was used as the endogenous control.
For immunohistochemical analysis, the avidin–biotin complex (ABC) method was used to look for the localization of NISCH, ITGA5, p-PAK, p-LIMK1/2, RAC1, NF-κB p65, FAK and p-SMAD2 in wild-type and mutant mouse ME and lungs. The sections through the ears of mice were de-parafinized, and endogenous peroxidase activity was quenched with 3% hydrogen peroxide in isopropanol for 30 min. Vectastain Elite ABC kit (Vector Laboratories, PK 6101) was used to perform the immunohistochemistry. The antibodies were as follows: rabbit polyclonal anti-NISCH (sc-98980, Santa Cruz Biotechnology), rabbit polyclonal anti-ITGA5 (sc-10729, Santa Cruz Biotechnology), rabbit polyclonal anti-p-αPAK (Thr212) (sc-101772, Santa Cruz Biotechnology), rabbit polyclonal anti-p-LIMK1/2 (Thr508/505) (sc-28409-R, Santa Cruz Biotechnology), rabbit polyclonal anti-RAC1 (sc-95, Santa Cruz Biotechnology), rabbit polyclonal anti-p-SMAD2 (Ser465/467) (AB3849, Chemicon International), rabbit polyclonal anti-FAK (sc-558, Santa Cruz Biotechnology) rabbit polyclonal anti-NF-κB p65 (ab131485, Abcam). The sections were incubated with the antibodies overnight at the following dilutions: p-PAK, 1:50; NISCH, ITGA5, p-LIMK1/2, NF-κB p65, FAK and p-SMAD2, 1:200; Rac1, 1:400. For F4/80 visualisation, sections were treated with 0.05% trypsin in calcium chloride for 20 min at 37°C, blocked with 10% rabbit serum (X0902, DAKO), incubated with rat anti mouse F4/80 (MCA497GA, Serotec) antibody overnight at 1:100 dilution and the next day after the washes were incubated with biotinylated rabbit anti-rat secondary antibody at 1:400 dilution (E0468, DAKO). The serum and the secondary antibody for all the other antibodies were from the Vectastain Elite ABC kit and they were used according to the manufacturer's instructions. DAB+ chromogen system (DAKO K3468) was used to develop the specific signals. The slides were counterstained with haematoxylin.
Total protein extracted from the ME epithelial cells and lungs of two-months-old wild-type, Nischedsn/edsn and Itga5tm1Hyn/+; Nischedsn/edsn mice were used for the western blot analysis. Each middle ear sample consisted of combined epithelial cells scooped out of both ears of one mouse. Each lung sample consisted of whole lung tissue from one mouse. Either three or four biological replicates were performed for each antibody. The tissues were homogenised in CelLytic MT Cell Lysis Reagent (Sigma C3228), protease inhibitors, phosphatase inhibitors and vanadate and centrifuged at 4°C. Protein concentration was determined using the DC Protein Assay kit (Bio-Rad). Samples (30 μg from the lung samples and 10 μg from the middle ear samples) were loaded into 12% NuPAGE Bis-Tris gel, 7% NuPAGE Tris Acetate gel or 3–8% NuPAGE Tris acetate gels (Invitrogen), blotted onto nitrocellulose membrane (Invitrogen) and immunostained. 5% non-fat milk in TBST was used as blocking solution and antibody diluent. The antibodies and the dilutions they were used at for the western blot analysis were as follows: rabbit polyclonal anti-ITGA5 (sc-10729, Santa Cruz Biotechnology) 1:500, rabbit polyclonal anti-PAK1 (2602, Cell Signaling) 1:1000, rabbit polyclonal anti-p-LIMK1/2 (Thr508/505) (sc-28409-R, Santa Cruz Biotechnology) 1:500, goat polyclonal anti LIMK1 (sc-8387, Santa Cruz Biotechnology) 1:500, rabbit polyclonal anti-RAC1 (sc-95, Santa Cruz Biotechnology) 1:500, rabbit polyclonal anti-FAK (sc-558, Santa Cruz Biotechnology) 1:500, rabbit polyclonal anti-NF-κB p65 (ab131485, Abcam) 1:1000, rabbit polyclonal p-NFκB p65 (Ser 276) (sc-101749, Santa Cruz Biotechnology) 1:500, rabbit polyclonal anti-phospho-FAK (Tyr576) (44-652G, Invitrogen), rabbit polyclonal anti-phospho-SRC (Tyr527) (2105, Cell Signaling), rabbit polyclonal anti-phospho-SMAD2 (Ser465 /467) (3101, Cell Signaling) 1:500 and actin (A 2066, Sigma). Goat anti-rabbit IgG (H+L)-HRP conjugate (1706515, Bio-Rad), 1:3000, was used as a secondary antibody for all the primary antibodies except for LIMK1 for which Rabbit anti-goat IgG (H+L) secondary antibody, HRP, 1:5000 (81–1620, Invitrogen) was used. ECL or ECL 2 (GE Healthcare) were used as detection system.
All data are given as unadjusted mean +/- SEM (standard error of the mean) unless stated otherwise. Data were analysed to establish normal distribution. Where data was normally distributed an ANOVA or Student’s t-test were conducted. If data was not normally distributed the non-parametric equivalents of these tests were used (Kruskal-Wallis One Way Analysis of Variance on Ranks or Mann-Whitney Rank Sum Test) to establish if data were significant. The Holm-Sidak method (ANOVA) or Dunn’s method (Ranks) was used for multiple comparisons versus a control group. Results with values of P < 0.05 were considered statistically significant. SigmaPlot 11.0 software was used to perform all statistical analysis.
|
10.1371/journal.pgen.1004020 | TATN-1 Mutations Reveal a Novel Role for Tyrosine as a Metabolic Signal That Influences Developmental Decisions and Longevity in Caenorhabditis elegans | Recent work has identified changes in the metabolism of the aromatic amino acid tyrosine as a risk factor for diabetes and a contributor to the development of liver cancer. While these findings could suggest a role for tyrosine as a direct regulator of the behavior of cells and tissues, evidence for this model is currently lacking. Through the use of RNAi and genetic mutants, we identify tatn-1, which is the worm ortholog of tyrosine aminotransferase and catalyzes the first step of the conserved tyrosine degradation pathway, as a novel regulator of the dauer decision and modulator of the daf-2 insulin/IGF-1-like (IGFR) signaling pathway in Caenorhabditis elegans. Mutations affecting tatn-1 elevate tyrosine levels in the animal, and enhance the effects of mutations in genes that lie within the daf-2/insulin signaling pathway or are otherwise upstream of daf-16/FOXO on both dauer formation and worm longevity. These effects are mediated by elevated tyrosine levels as supplemental dietary tyrosine mimics the phenotypes produced by a tatn-1 mutation, and the effects still occur when the enzymes needed to convert tyrosine into catecholamine neurotransmitters are missing. The effects on dauer formation and lifespan require the aak-2/AMPK gene, and tatn-1 mutations increase phospho-AAK-2 levels. In contrast, the daf-16/FOXO transcription factor is only partially required for the effects on dauer formation and not required for increased longevity. We also find that the controlled metabolism of tyrosine by tatn-1 may function normally in dauer formation because the expression of the TATN-1 protein is regulated both by daf-2/IGFR signaling and also by the same dietary and environmental cues which influence dauer formation. Our findings point to a novel role for tyrosine as a developmental regulator and modulator of longevity, and support a model where elevated tyrosine levels play a causal role in the development of diabetes and cancer in people.
| In people, elevated blood levels of the amino acid tyrosine are seen in obese individuals, and these elevations represent a novel risk factor for the development of diabetes. The enzyme tyrosine aminotransferase, which removes tyrosine from the body, has also been identified as a tumor suppressor gene, and this enzyme normally acts to prevent the development of liver cancer. In our work, we identify tyrosine aminotransferase as a regulator of larval development and adult longevity in the non-parasitic worm Caenorhabditis elegans. Worms with mutations impairing tyrosine aminotransferase activity show elevated levels of tyrosine, are prone to arresting development in a larval stage called a dauer, and show increased longevity. Part of the effect of tyrosine aminotransferase is due to inhibitory effects on an insulin-like signaling pathway in the worms. Our work suggests that levels of the amino acid tyrosine are sensed and can lead to changes in cell signaling. These results may provide insights into how tyrosine could be involved in obesity, diabetes, and cancer in people.
| The aromatic amino acid tyrosine serves many metabolic roles including being a building block for protein synthesis, a source of energy, and a precursor for the synthesis of melanin and several neurotransmitters including dopamine and other catecholamines. Beyond these currently known functions for tyrosine, recent work has suggested that tyrosine could also play regulatory roles in both metabolism and the control of cell proliferation. Specifically, in people elevated serum tyrosine levels occur with obesity and represent a risk factor for the development of diabetes [1]–[6]. Additionally, the enzyme tyrosine aminotransferase (TAT), which acts to normally convert tyrosine to energy, has been identified as a tumor suppressor gene which acts to promote apoptosis and prevent the development of hepatocellular carcinoma [7]. How changes in tyrosine metabolism could contribute to these disease processes is currently unknown, but it is possible that levels of this amino acid could play a direct regulatory role for the behavior of specific cells and tissues. While consistent with the available data, direct evidence for this model is currently lacking.
The nematode Caenorhabditis elegans normally progresses through four larval stages before developing into a reproductive adult animal. However specific cues, such as crowding, low food availability, or elevated temperature, can be sensed by the developing worm and lead to developmental arrest in a diapause state called a dauer larva [8]–[11]. Entry into dauer permits worms to delay the completion of development and the initiation of reproduction in environments which are not favorable, and instead the animals can survive as a dauer for up to several months before resuming normal development when conditions become favorable for reproductive success. This developmental decision requires a complicated interplay of sensory neurons with specific cGMP, TGF-β, and insulin-like signaling cascades controlling the choice of reproductive versus dauer development [9], [10].
In the worm, the daf-2 insulin/IGF-1 receptor (IGFR) signaling pathway is involved in both dauer development and adult longevity [12]–[14]. Active signaling through the pathway during development enables animals to reach reproductive adulthood whereas reductions in daf-2/IGFR signaling due to either environmental triggers or genetic mutations lead to arrest as a dauer [9], [14]. In adult worms, daf-2/IGFR signaling is a major modulator of longevity and mutations impairing the pathway can result in 100% increases in lifespan [13].
At a molecular level, the daf-2/IGFR pathway consists of daf-28 and other insulin-like peptides, which are thought to act as ligands for the DAF-2 insulin/IGF-1 receptor [15]–[18]. Downstream of daf-2/IGFR, is the age-1 PI3 kinase and a kinase cascade consisting of the phosphoinositide-dependent kinase pdk-1 and the protein kinase B genes akt-1 and akt-2 (Figure 1A) [19]–[22]. Both akt-1 and akt-2 normally act to phosphorylate the DAF-16 FOXO transcription factor which leads to its retention in the cytoplasm [23]–[25]. Reductions in either daf-2/IGFR or combined akt-1 and akt-2 activity result in the entry of DAF-16/FOXO into the nucleus and strong activation of DAF-16/FOXO target genes [23]–[25]. In contrast, loss of only akt-1 activity leads to the translocation of DAF-16/FOXO into the nucleus but a lesser increase in the expression of DAF-16/FOXO target genes (Figure 1A and Table S1) [26], [27]. This finding suggested that additional pathways could be involved in controlling the transcriptional activity of DAF-16/FOXO (Figure 1A). One group of potential regulators is the eak (enhancer of akt-1 null) genes, which were identified in a forward genetic screen and act in a non-cell autonomous manner to control the transcriptional activity of nuclear localized daf-16/FOXO (Figure 1A) [26]–[29]. The identified eak genes lack structural homology to one another and appear to lie in one or more poorly characterized pathways that act in parallel to akt-1. The identification of the eak genes suggests that additional novel pathways either downstream or parallel to insulin signaling may await discovery.
In both vertebrates and worms, there is data suggesting a link between tyrosine metabolism and insulin signaling. TAT, which catalyzes the first step in a conserved degradation pathway that converts tyrosine to fumarate and acetoacetate (Figure 1B), has been well studied as a target of regulation by insulin signaling in vertebrates with insulin effects seen at both the transcriptional and translational level [30]–[38]. Further, in C. elegans the hpd-1 gene, which encodes the enzyme 4-hydroxyphenylpyruvate dioxygenase and lies immediately downstream of TAT in the tyrosine degradation pathway, is a target gene for the daf-16/FOXO transcription factor and positively regulated by daf-2/IGFR signaling [39]. The down-regulation of hpd-1 in daf-2/IGFR mutants could lead to a reduction in tyrosine clearance and could, at least in part, account for the increases tyrosine levels observed in these animals [40]. Furthermore, the inhibition of hpd-1 by RNAi was shown to both extend lifespan and delay dauer exit through unknown mechanisms [39]. Hence, tyrosine metabolism appears to be actively controlled by insulin signaling though the consequences of this regulation are currently unclear.
In our work, we identify tatn-1, which is the worm ortholog of TAT, to be a novel dauer formation regulator that is under the control of several dauer-inducing stimuli, including daf-2/IGFR signaling, and ultimately regulates tyrosine levels in the worm. We further find that tatn-1 mutations enhance the dauer-formation and lifespan phenotypes of both daf-2/IGFR and eak mutants suggesting that elevated tyrosine levels have inhibitory effects on insulin signaling. These effects require the aak-2/AMPK gene, and tatn-1 mutants have elevated levels of the activated phospho-AAK-2 protein consistent with activation of AAK-2 signaling in response to elevated tyrosine. The activation of AAK-2 may lead to effects on the downstream transcription factors daf-16/FOXO and crh-1/CREB. We see a partial dependence on daf-16/FOXO for some tatn-1 phenotypes and activation of daf-16/FOXO target genes, and the loss of crh-1/CREB, which is inhibited by activated aak-2, mimics some tatn-1 phenotypes. Together our findings establish a novel role for tyrosine as a metabolic signal that influences insulin signaling, development, and lifespan through effects on aak-2/AMPK signaling. While further study is necessary, our results also suggest that the recently observed associations between tyrosine metabolism and both diabetes and cancer are due to elevated tyrosine levels playing a direct causal role in disease pathogenesis.
The eak genes were identified in a genetic screen as enhancers of the weak dauer formation phenotype shown by akt-1 mutants, and these genes normally act to suppress the transcriptional activity of nuclear localized daf-16/FOXO [26]–[29]. To identify new genes that act in parallel to eak genes to control dauer formation, we performed a genome-wide RNAi screen for gene inactivations that enhance the weak dauer-constitutive phenotype of the eak-4(mg348) mutant. Since RNAi of dauer-constitutive genes typically yields a weaker phenotype than the corresponding mutants, we constructed an eri-3(mg408); eak-4(mg348) double mutant to enhance the sensitivity of the eak-4 mutant to RNAi [41], [42]. Control experiments demonstrated that RNAi inhibition of daf-2/IGFR, akt-1, or the 14-3-3 gene ftt-2 enhance dauer arrest by the eri-3; eak-4 mutants whereas RNAi inhibition of daf-7, daf-9, or daf-11, which encode components of TGF-β, dafachronic acid, and cGMP pathways, respectively, do not (Figure 2A). These data suggested that the screen could be enriched for genes that act in the daf-2/IGFR pathway to control dauer arrest.
Among the RNAi clones identified from the screen was tatn-1, which encodes the worm ortholog of tyrosine aminotransferase. Inhibition of tatn-1 by RNAi enhanced dauer arrest by eri-3; eak-4 mutants to the same extent as akt-1 RNAi (Figure 2A) [43]. Tyrosine aminotransferase is the first enzyme in the conserved five step tyrosine degradation pathway present in worms and other eukaryotes, and this enzyme catalyzes the deamination of tyrosine to produce 4-hydroxyphenylpyruvate (Figure 1B). The subsequent steps in the degradation pathway convert 4-hydroxyphenylpyruvate into fumarate and acetoacetate which can be ultimately metabolized by the Krebs' cycle (Figure 1B).
Similarly to the effects of tatn-1 RNAi, we also found that the tatn-1(baf1) mutation, which has a P224S mutation in a conserved region of the protein and is likely a partial loss of function allele [43], [44], also promoted dauer arrest by the eak-4(mg348) mutants (Figure 2B and Table S2). Interestingly, the interaction between tatn-1 and eak-4 was strongly influenced by worm diet with diets consisting of the E. coli K12-derived HT115 or K12-B hybrid HB101 bacterial strains showing the strongest interaction (Figure S1). Additionally, we observed that a second tatn-1 allele, tatn-1(qd182), both enhanced dauer arrest by the eak-4 mutant and also had a weak effect on dauer formation in isolation (Figure S2). The tatn-1(qd182) allele encodes a protein with a G171E mutation affecting a highly conserved glycine residue (Figure S2). Together these findings demonstrate a novel role for tyrosine aminotransferase as a regulator of the dauer development decision.
We then tested whether tatn-1 interacted with other eak genes via the construction of tatn-1; eak mutants. We found that tatn-1(baf1) enhanced the dauer-constitutive phenotype of all eak mutants tested, including eak-4(mg348), eak-3(mg344),eak-5/sdf-9(mg337), eak-2/hsd-1(mg433), and eak-7(tm3188) (Figure 2C and 2D). eak-7 showed a particularly strong interaction with tatn-1, with 92.4% of the population forming dauers at 25°C (Figure 2D), and we observed dauers in eak-7; tatn-1 in cultures growing at lower temperatures or on other worm diets. These data suggest that tatn-1 is a general enhancer of dauer formation by eak mutants.
In addition to enhancing dauer formation, eak-7 mutations, but not other eak mutations, extend the lifespan of wild-type and akt-1 mutant worms [27]. We tested whether a tatn-1 mutation also enhanced the lifespan of eak-7 mutants by conducting survival assays using wild type N2, tatn-1(baf1), eak-7(tm3188), and eak-7(tm3188); tatn-1(baf1) worm populations. We found that tatn-1, eak-7, and eak-7; tatn-1 mutants all showed increased longevity relative to N2 (mean survivals of 21.0, 23.0, 30.0, and 33.5 days, respectively) (Figure 2E and Table S3). Specifically, tatn-1 extends mean survival 10.4%, eak-7 extends mean survival 43.1%, and eak-7; tatn-1 increases survival 59.2% over wild type (Figure 2E). These findings demonstrate a novel role for tatn-1 in modulating lifespan and also demonstrate that the effects of the eak-7 – tatn-1 genetic interaction also influence adult longevity.
Mutations in the eak genes enhance the dauer formation phenotype of loss-of-function mutations affecting genes in the daf-2/IGFR signaling pathway, so we tested whether tatn-1 mutations also enhanced daf-2/IGFR mutant phenotypes [28]. We used the daf-2(e1368) allele which has a strong dauer-constitutive phenotype when grown at 25°C but a weaker phenotype when grown at lower temperatures. Compared to tatn-1(baf1) and daf-2(e1368) alone, we found that daf-2(e1368); tatn-1(baf1) mutants showed increased levels of dauer formation when grown at 23°C (Figure 3A). Further, we found that the tatn-1 mutation extends the adult lifespan of worms treated with daf-2 RNAi starting at day 1 of adulthood (Figure 3B). We used RNAi treatment due to the high levels of dauer arrest that we observed with the daf-2(e1368); tatn-1(baf1) animals. In these RNAi experiments, the mean survival of worms were 22.0 days for N2 on control RNAi, 24.0 days for tatn-1 on control RNAi, 30.5 days for N2 on daf-2 RNAi, and 37.0 days for tatn-1 on daf-2 RNAi (Figure 3B and Table S3). Hence, while daf-2 RNAi treatment extends the survival of wild type N2 worms by 37.4%, the inclusion of the tatn-1 allele further extends the lifespan of daf-2(RNAi) treated worms by an additional 21%. These findings show a genetic interaction between tatn-1 and daf-2/IGFR with impaired tyrosine degradation enhancing the daf-2/IGFR dauer formation and lifespan phenotypes.
Since the eak mutations interact with akt-1 mutations to enhance dauer formation, the effects of tatn-1 mutations could be due to inhibition of the PI3 kinase signaling pathway (Figure 1A) [20]–[22], [28], [45]. To test this possibility, we looked for genetic interactions between loss-of-function and gain-of-function mutations affecting the PI3 kinase pathway with tatn-1. First, we constructed tatn-1 mutants containing the loss-of-function mutations affecting the PKB genes akt-1 and akt-2, and the related kinase sgk-1. We found that none of these genes interacts strongly with a tatn-1 mutation to enhance dauer arrest (Figure 4A). This finding suggests that tatn-1 is not an upstream regulator of either akt-1 or akt-2. To further test whether tyrosine affected the PI3 kinase signaling cascade, we tested whether the eak-4(mg348); tatn-1(baf1) interaction was blocked by gain-of-function mutations in either pdk-1 or akt-1. These mutations were identified in genetic screens as suppressors of the dauer-constitutive phenotype of an age-1/PI3K null mutant [20], [21]. We constructed both eak-4(mg348); pdk-1(mg142), tatn-1(baf1) and eak-4(mg348); akt-1(mg144); tatn-1(baf1) mutants and examined the effects of these mutations on dauer formation. We found that pdk-1(mg142), which is a dominant gain-of-function allele of 3-phosphoinositide-dependent kinase and lies upstream of akt-1, akt-2, and sgk-1, had little effect on dauer formation by the eak-4(mg348); tatn-1(baf1) mutants (70% with pdk-1(mg142) versus 75% without) (Figure 4B). We further found that akt-1(mg144) had at most a modest effect on dauer formation by eak-4(mg348); tatn-1(baf1) mutants (51% with akt-1(mg144) versus 77.6% without) (Figure 4C). Together these results suggest that the interaction between eak-4 and tatn-1 likely does not involve changes in the PI3 kinase signaling cascade.
In addition to the daf-2/IGFR signaling pathway, dauer formation in worms is also regulated by a TGF-β signaling pathway [9]–[11]. While traditionally these two pathways are viewed as independent, more recent work has indicated that cross-talk between the pathways may occur. Specifically, genes in one pathway, such as sdf-9/eak-5 or pdp-1, have been shown to augment the dauer formation phenotypes of genes in the other pathway [46], [47]. Notably, the pdp-1 phosphatase was identified from an RNAi screen using daf-2/IGFR pathway mutants as was tatn-1. To test whether tatn-1 could act in the TGF-β signaling pathway, we blocked this pathway with either the daf-3/SMAD or daf-5/Sno mutations [41], [48], [49]. We found that neither mutant reduced dauer formation by the eak-4; tatn-1 mutants (Figure S3), which is consistent with tatn-1 acting independently of the TGF-β pathway.
Since tatn-1 did not interact with akt-1, akt-2, or sgk-1, we looked for alternate signaling pathways that could be activated by a tatn-1 mutation, and then interact with the daf-2/IGFR signaling pathway. The AMP-activated protein kinase (AMPK) ortholog AAK-2 was considered as a candidate because aak-2 interacts with daf-2 signaling, plays roles in dauer development, modulates worm longevity, and acts in part through the daf-16/FOXO transcription factor which is part of the daf-2/IGFR signaling pathway [50]–[53]. To test the involvement of aak-2/AMPK, we compared dauer formation between eak-4(mg348); tatn-1(baf1) and eak-4(mg348); tatn-1(baf1), aak-2(gt33) mutants. We found that loss of aak-2 strongly reduced dauer arrest from 84.7% to 10.7% (Figure 5A and Table S2). The aak-2 mutation also reduced dauer formation by an eak-4(mg348); tatn-1(qd182) mutant though to a lesser degree than with the tatn-1(baf1) allele (Figure S4). This may be due to the tatn-1(qd182) allele being a stronger loss-of-function allele than tatn-1(baf1) as suggested by the developmental delay phenotype shown by tatn-1(qd182) and the higher tyrosine levels found in this mutant (Figure S4 and below). Together these findings support a necessary role for aak-2/AMPK for worms to respond to reductions in tatn-1 activity, especially with weaker alleles.
To further explore the role of aak-2/AMPK in dauer formation, we treated worms with the AMPK agonist AICAR. Treatment of wild-type N2 worms with 0.125 mM AICAR had no effect on development or dauer formation, but eak-4 mutants treated with AICAR showed a significant increase in dauer formation compared to an untreated control (Figure 5B). This effect required aak-2/AMPK as an eak-4(mg348); aak-2(gt33) mutant failed to respond to AICAR (Figure 5B). These findings demonstrate that aak-2 activity is both necessary and sufficient to promote dauer formation by eak-4 mutants.
Our findings could suggest that AAK-2 is activated in the tatn-1 mutants. To test this possibility we used western blotting to measure levels of AAK-2 phosphorylated on the activating Thr243 residue, which is analogous to Thr172 in the vertebrate orthologs. We found that treatment of N2 wild-type animals with either sodium azide which inhibits mitochondrial function and depletes ATP levels by approximately 50% at this dose or with 1 mM AICAR leads to increases in phosphorylated AAK-2 levels compared with untreated L2 larval N2 worms grown in parallel (Figure 5C) [54]. Furthermore levels of phospho-AAK-2 were also increased in eak-4; tatn1 mutant L2 larvae compared to the N2 control larvae (Figure 5C). Importantly, the phopho-AAK-2 signal was lost in aak-2(gt33) mutants confirming the identity of this band as AAK-2 (Figure 5C). These findings demonstrate that both tatn-1 mutations and AICAR treatment serve to activate AMPK in C. elegans, and that this activation of AAK-2 is required for the effects of tatn-1 mutations on development.
The requirement for aak-2/AMPK for the effects of tatn-1 could reflect impaired tyrosine degradation leading to reduced production of the TCA cycle precursors fumarate and acetoacetate and a consequent decrease in energy production. To test this possibility we analyzed worm lysates for levels of AMP and ATP. We found that eak-4(mg348); tatn-1(baf1) mutants had a lower AMP/ATP ratio than N2 worms grown in parallel (Figure S5). This suggests that the role of aak-2/AMPK does not reflect reduced energy production in the tatn-1 mutants.
Finally, we tested whether the effect of tatn-1(baf1) on adult lifespan requires aak-2 activity, and we found that a tatn-1 mutation increased wild-type lifespan by 17.8%, but decreased aak-2(gt33) lifespan by 8.9% compared to aak-2(gt33) alone (N2 mean survival 18.0 days, tatn-1(baf1) 21.0 days, aak-2(gt33) 14.0 days, and tatn-1(baf1), aak-2(gt33) 13.0 days) (Figure 5C and Table S3). Hence aak-2 is both required for the tatn-1 lifespan increase and may even play a protective role in animals with reduced tatn-1 activity.
Since the effects of eak genes on dauer formation are dependent on both the FOXO transcription factor DAF-16 and the nuclear hormone receptor DAF-12, we tested whether these genes are required for dauer formation by eak-4(mg348); tatn-1(baf1) mutants through the construction of daf-16(mgDf47); eak-4(mg348); tatn-1(baf1) and eak-4(mg348); tatn-1(baf1), daf-12(rh61rh411) mutants [26]–[28]. We found that the genetic interaction is completely dependent on daf-12 (Figure 6A). However, we identified both daf-16 dependent and independent effects of tatn-1. The tatn-1 effects on dauer formation are largely blocked by the daf-16(mgDf47) mutation, but a small percentage of worms still form dauers (Figure 6A). This daf-16 independent pathway was also seen in experiments using the stronger tatn-1(qd182) allele (Figure S6). These findings suggest that daf-16/FOXO does play a vital role in the interaction between eak genes and tatn-1, but that daf-16/FOXO independent pathways are also involved.
The daf-16/FOXO gene encodes multiple isoforms which have been recently shown to have differential modes of regulation and have distinct effects on development and longevity [24], [55], [56]. For dauer development, three isoforms appear to be involved with DAF-16A being the predominant isoform, DAF-16DF playing a somewhat lesser role, and DAF-16B playing a modest role, at best [55]. As a result, we tested whether the developmental effects of eak-4 and tatn-1 are dependent on particular daf-16 isoforms through the use of isoform-specific transgenes to rescue the dauer-constitutive phenotypes lost in daf-16/FOXO mutants (Figure 6B). We found that a transgene encoding a DAF-16A:mRFP fusion protein was able to strongly rescue the formation of dauers by a daf-16(mgDf50); eak-4(mg348); tatn-1(baf1) mutant, whereas a transgene encoding a DAF-16DF:GFP fusion protein only weakly rescued dauer formation (Figure 6B). These data suggest that the daf-16a isoform is most involved in the developmental effects produced by tatn-1 mutants.
To further explore the role of daf-16/FOXO in tatn-1 phenotypes, we examined the effects of eak-4 and tatn-1 mutations on both daf-16/FOXO target gene expression and DAF-16 subcellular localization. We used a sod-3:GFP transgene to examine the effects of eak-4 and tatn-1 mutations on expression of the daf-16 target gene sod-3, which encodes a manganese superoxide dismutase enzyme [57], [58]. We generated combinations of eak-4(mg348) and tatn-1(baf1) with the transgene, and examined GFP expression in L2 larvae. We found that the combination of eak-4 and tatn-1 mutations resulted in the highest expression of sod-3:GFP even prior to dauer formation (Figure 6C). Furthermore, aak-2/AMPK was required for this enhanced activation of the sod-3:GFP reporter (Figure 6C). Similar effects of tatn-1 and eak-4 on sod-3 expression were also seen when mRNA levels for the endogenous gene were measured by Q-PCR in L2 larvae, and this effect on sod-3 expression also required aak-2/AMPK (Figure 6D). Together these findings demonstrate that tatn-1 and eak-4 act in an aak-2/AMPK dependent manner to promote at least some aspects of daf-16/FOXO transcriptional activity.
Since the eak genes act to inhibit DAF-16/FOXO within the nucleus, a possible explanation for the enhancement of dauer formation in eak; tatn-1 double mutants may be explained by the tatn-1 allele causing increased nuclear localization of DAF-16 as do mutations affecting akt-1 [26]. We tested for changes in DAF-16 localization via the use of transgenic animals expressing a DAF-16A:GFP fusion protein [59]. We found that eak-4; tatn-1 mutants failed to show clearly visible nuclear localization of DAF-16:GFP in synchronized L2 worms grown at 25°C (Figure 6E). In contrast, exposure of animals to a 1 hour heat-shock at 35°C led to strong nuclear localization of DAF-16A:GFP (Figure 6E). Together these findings suggest that the tatn-1 enhancement of the eak dauer formation phenotype could be due to an increase in daf-16 transcriptional activity without an accompanying significant change in DAF-16 subcellular localization.
We then used daf-16 RNAi treatment to test whether daf-16/FOXO is required for the lifespan extension of tatn-1 mutants. We found that silencing of daf-16 through RNAi results in a shortened lifespan for both tatn-1 and wild-type N2 worms compared to control RNAi treatment (Figure 6F). However in daf-16 RNAi treated worms, tatn-1 still produces lifespan extension over wild type (mean survival for N2 control RNAi 21.0 days, tatn-1(baf1) control RNAi 24.0 days, N2 daf-16(RNAi) 14.5 days, and tatn-1(baf1) daf-16(RNAi) 17.0 days). Specifically, tatn-1(baf1) produced a 13.7% increase in lifespan in control RNAi treated worms, but a 17.8% increase in lifespan in daf-16 RNAi treated worms. These data suggest that the increased lifespan resulting from tatn-1 is either independent of daf-16 or occurs in a site resistant to the effects of daf-16 RNAi.
To explore the effects of impaired tyrosine metabolism in C. elegans, we performed whole transcriptome RNA sequencing (RNA-seq) to identify genes that are differentially regulated in the tatn-1 mutants. To maximize the gene expression changes seen, we used the stronger tatn-1(qd182) allele and compared its transcriptome to that of wild-type N2 animals. Using ANOVA testing with a false-discovery rate of 5%, we identified 890 up-regulated and 3732 down-regulated genes in the tatn-1(qd182) mutant relative to N2 (Table S4).
To understand how the tatn-1 mutants might affect the daf-2/IGFR pathway, we used this data set to examine whether changes in the expression of genes in pathway are seen. We found that there was no change in the expression of daf-2/IGFR, daf-16/FOXO, aak-2/AMPK, eak-4, or akt-2 (Table S4). However we did find that levels of the age-1 PI3-kinase are reduced almost 76% and levels of akt-1 are reduced almost 70% compared to N2 while levels of the daf-18/PTEN tumor suppressor, which normally inhibits signaling through the PI3-kinase signaling pathway, is reduced almost 95% compared to wild-type animals (Table S4). Despite the observed changes in the expression of genes in the PI3-kinase signaling pathway, there is likely little net effect on the regulation of downstream targets by the pathway as we failed to observe differences in DAF-16:GFP localization, which would translocate to the nucleus if the pathway was inhibited (Figure 6E).
We then used both the DAVID program and the Panther database both to identify biologic themes within the up-regulated and down-regulated genes by testing for over-represented gene classes based on structural and functional annotations and to visualize the gene classes seen in both groups of genes (Figure S7 and Table S5) [60], [61]. Within the up-regulated set, we found that genes involved in tyrosine metabolism and neuropeptide signaling were strongly over-represented (7–10 fold) (Table S5). Specifically, we found that every gene in the tyrosine degradation pathway is up-regulated in the tatn-1(qd182) mutant suggesting that the altered metabolism is detected and leads to a compensatory change in expression of the pathway (Table S4). The significance of the changes in neuropeptide signaling gene expression is currently unclear, but could suggest that impaired tyrosine metabolism or the resulting increased in AAK-2/AMPK activity produces direct changes in neuronal activity or that these changes could be the direct downstream effectors responsible for the tatn-1 phenotypes. Graphically, we saw a greater percentage of genes which were expressed in the extracellular compartment and had catalytic or receptor activity among the up-regulated genes compared to those that were down-regulated in the tatn-1 mutants (Figure S7). In contrast, within the down-regulated genes, we identified over-representation of a broad range of genes involved in germline development, cell cycle, DNA replication, and larval development (Table S5). Further, these genes were more likely to be expressed in the intracellular compartment and to have regulatory effects on translation or enzyme activity (Figure S7). The expression changes in genes involved in cell cycle regulation and development could perhaps account for the developmental delay observed in the tatn-1(qd182) mutant compared to N2 (Figure S4).
Given the requirement we found for daf-16/FOXO for aspects of the tatn-1 phenotypes, we used Gene Set Association Analysis (GSAA) to test for enrichment of genes known to be regulated by daf-16/FOXO in the context of daf-2/IGFR signaling [62], [63]. Via this approach, we found both the up-regulated and down-regulated daf-16/FOXO target genes identified by Murphy et. al. to be enriched within the tatn-1(qd182) transcriptome (Figure 7A). This suggests that the expression of a subset of daf-16/FOXO target genes is altered by changes in tyrosine metabolism.
Since our genetic studies suggested the involvement of a daf-16-independent pathway, we also used GSAA to test whether target genes recently identified for the CREB transcription factor crh-1 are enriched in the tatn-1(qd182) mutant [53]. We chose to focus on crh-1 because recent work has demonstrated that crh-1/CREB lies downstream of aak-2/AMPK [53]. crh-1/CREB and aak-2/AMPK are mechanistically linked because AAK-2 directly phosphorylates and inactivates the crh-1/CREB coactivator crtc-1, and as a result both aak-2/AMPK over-expressing and crh-1/CREB mutant animals are long-lived and share gene expression profiles [53]. We found that in the tatn-1(qd182) mutants, there is differential expression of both genes up-regulated and genes down-regulated in crh-1/CREB mutants (Figure 7B). This suggests that altered tyrosine metabolism could lead to changes in crh-1 target gene expression and could suggest a role for crh-1/CREB in the tatn-1 phenotypes. To test for crh-1/CREB involvement, we combined the crh-1(tz2) null allele with tatn-1 and eak-4 and examined the effects on dauer formation. We found that crh-1 showed a similar interaction as tatn-1 with eak-4, but did not promote dauer formation by the tatn-1 mutant (Figure 7C). Together these findings suggest that tatn-1 mutants share phenotypes and gene expression profiles with crh-1/CREB mutants and could be consistent with crh-1/CREB acting as an additional downstream effector of the response to impaired tyrosine metabolism.
In vertebrates, tyrosine aminotransferase has been reported to be an insulin target gene with insulin treatment leading to reduced expression [32]–[35], [38]. As a result, we asked whether tatn-1 could also be regulated by daf-2/IGFR signaling in worms. To test the effects of daf-2 signaling on tatn-1 expression, we generated transgenic worms with an integrated transgene expressing a TATN-1:GFP fusion protein under the control of the tatn-1 promoter. This transgene rescues the tatn-1(baf1) mutation and blocks dauer formation by eak-4(mg348); tatn-1(baf-1) mutants, which demonstrates that the fusion protein is both functional and expressed in the correct anatomical locations (Figure 8A).
GFP fluorescence representing the TATN-1:GFP fusion protein is observed in the intestine and hypodermis of worms (Figure 8E). When we crossed the transgene into the daf-2(e1368) mutant, and we found that the presence of the daf-2/IGFR mutation led to a 20% decline in GFP expression in adult worms grown on OP50-1 at 20° (Figure 8B and Figure 8E). This is consistent with daf-2/IGFR signaling acting positively to promote TATN-1 expression in adult worms. Interestingly, the effect of daf-2/IGFR signaling on TATN-1 expression likely occurs either at the translational or protein stability level because Q-PCR experiments demonstrated almost a 50% increase in tatn-1 mRNA expression in the daf-2 mutant animals (Figure S8). As a result of the divergent regulation of tatn-1 mRNA and protein levels, we focused on TATN-1:GFP expression in our subsequent experiments.
Beyond daf-2/IGFR signaling, we found that both diet and environmental temperature affected TATN-1 levels to a similar or even greater degree. Specifically, adult worms grown on OP50-1, HB101, or HT115 show decreases in TATN-1:GFP expression when shifted from 20°C to 25°C for 24 hours with worms grown on OP50-1 showing a 26.2% decrease, on HB101 a 44.7% decrease, and on HT115 a 24.2% decrease (Figure 8C and Figure 8E). Further, we found that the OP50-1 fed worms showed greater TATN-1:GFP expression compared to HB101 and HT115 fed worms. This difference was especially apparent in worms grown at 25°C, due to the variability of GFP intensity seen at 20°C, with HB101 and HT115 fed animals showing a 25.3% and 17.3% decrease, respectively, compared to OP50-1 fed worms (Figure 8C and Figure 8E).
The effects of diet on TATN-1:GFP expression suggests that the E. coli bacterial strains vary in nutrient composition in a way that can be detected by the worms. Prior work has demonstrated that protein is the primary component of these bacteria but that the overall protein levels are not significantly different between strains [64]. However, this work also suggested that specific amino acids could vary between the strains and account for differences in fat content in worms fed each strain. Specifically, pept-1 mutants, which lack an intestinal peptide transporter, fail to show the expected differences in fat content when fed different bacterial strains [64]. In vertebrates, tyrosine aminotransferase expression is controlled by dietary amino acid intake, most notably for tryptophan [65]–[67]. To test whether dietary amino acid intake could affect TATN-1:GFP expression, we supplemented HB101 spotted NGA plates spotted with either tyrosine or tryptophan at a final concentration of 1 mg/mL, and compared TATN-1:GFP expression to worms fed HB101 alone or OP50-1. This concentration is 8 times the level found in standard NGA media (0.125 mg/mL). We found that the addition of tyrosine or tryptophan increases the GFP expression level in HB101 fed worms up to that seen in worms grown on OP50-1 (Figure 8D and Figure 8E). Together these data demonstrate that TATN-1 levels are dynamic and under the control of both daf-2/IGFR signaling as well as dietary and environmental cues. Importantly many of these signals that control TATN-1 expression also influence dauer formation suggesting that tatn-1 could be a regulated modulator of daf-2/IGFR signaling and developmental decisions.
Since TATN-1 is the first enzyme in the tyrosine degradation pathway, decreased activity should both increase the levels of tyrosine and also decrease the levels of the downstream metabolites (Figure 1B). The tatn-1 mutant phenotypes could be a direct result of changes in the level of a particular metabolite. For example, elevated fumarate levels are known promote hypoxia inducible factor (HIF) activity in certain renal cancers [68]. Alternately, tatn-1 could have a novel function that is independent of its metabolic activity. For example, subunits of the phenylalanine hydroxylase enzyme are known to have an additional, non-enzymatic role as a transcriptional co-activator [69]. To explore whether either of these models accounts for the tatn-1 phenotypes, we tested for interactions between eak-4 and mutant alleles of other enzymes in the tyrosine degradation pathway (Figure 1B) by constructing pah-1; eak-4 and hpd-1; eak-4 mutants. These mutants lack pah-1, which encodes the enzyme phenylalanine-4-hydroxylase that converts phenylalanine into tyrosine, or hpd-1, which encodes 4-hydroxyphenylpyruvate dioxygenase and catalyzes the step immediately downstream of tatn-1 [39], [43], [70]. We found that pah-1 did not enhance dauer arrest by the eak-4 mutants (Figure 9A) whereas both tatn-1 and hpd-1 increased dauer formation. Lower numbers of eak-4; tatn-1 dauers were seen because the experiment was scored after 3 days due to the slower development of the hpd-1 mutant. Additionally, when we used the pah-1 mutant to block the synthesis of tyrosine in a tatn-1; eak-4 mutant, we found that this reduced dauer formation (Figure 9B). These results suggested that the effects of tatn-1 on dauer formation were directly linked to the metabolic effects of tatn-1, and that the accumulation of tyrosine instead of deficiency of a downstream metabolite could be responsible for the tatn-1 phenotype.
As a result, we measured the levels of amino acids in wild-type N2, tatn-1(baf1), and tatn-1(qd182) larval animals grown at 25°C on HB101 plates by liquid chromatography mass spectrometry (LC-MS/MS). We found that wild-type N2 worms contained an average of 78.1 pmol of tyrosine per 100 worms whereas tatn-1(baf1) worms contained 237.4 pmol per 100 worms (Figure 9C). Hence the tatn-1(baf1) mutation produced a roughly three fold increase in tyrosine levels in the mutant animals. Further to compare the effects of tatn-1(qd182) on tyrosine levels compared to tatn-1(baf1), we measured tyrosine levels in additional samples grown and prepared in parallel. We found in these samples that wild-type N2 worms contained an average of 99.6 pmol of tyrosine per 100 worms, tatn-1(baf1) contained an average of 327.2 pmol per 100 worms, and tatn-1(qd182) contained an average of 470.2 pmol per 100 worms, which is an 43.7% increase over tatn-1(baf1) (Figure 9D). However, we noted that the tatn-1(qd182) worms were smaller than N2 or tatn-1(baf1) and the levels of many other amino acids measured in parallel were lower in tatn-1(qd182) compared to tatn-1(baf1). This suggested that the tatn-1(qd182) samples may have contained less overall biomass, and hence our normalization to worm counts alone may underestimate the effect of the tatn-1(qd182) mutation on tyrosine levels. To correct for this difference, we normalized tyrosine levels to the levels of all non-aromatic amino acids in the samples with the assumption that the net effect of these mutations on the levels of these amino acids is neutral. After normalization, we found that tatn-1(baf1) produced a 3.4 fold increase in tyrosine levels compared to N2 whereas tatn-1(qd182) produced a 12.6 fold increase compared to N2, which is also a 3.7 fold increase over tatn-1(baf1) levels (Figure 9E). These findings demonstrate that both tatn-1 alleles increase tyrosine levels compared to wild-type animals, and that the stronger phenotypes of the tatn-1(qd182) allele are likely due to the further increases in tyrosine levels observed.
To directly test whether elevated tyrosine levels are responsible for the tatn-1 phenotype, we treated worms with exogenous tyrosine cast into the NGA plates at 1 mg/mL. This treatment results in tyrosine levels in the worms that are elevated compared to untreated animals, but lower than those seen in the tatn-1 mutants (Figure S9). We found that supplementation had no effect on the development of wild-type worms but lead to dauer arrest by the eak-4 mutants (Figure 9F). These results directly demonstrate changes in tyrosine levels alter the development of eak-4 mutant worms and are responsible for the tatn-1 phenotype.
Amino acids are known to antagonize insulin actions in vertebrates, so our results could represent either the non-specific effects of any amino acid or a tyrosine-specific effect [71], [72]. To test these possibilities we directly compared the ability of a variety of amino acids to enhance dauer formation by eak-4 mutants. We grew eak-4 mutants on HB101 spotted NGA supplemented with tyrosine, glycine, leucine, isoleucine, glutamate, glutamine, asparagine, or aspartate, each at the concentration of 1 mg/mL. Since tyrosine is the largest of these amino acids, this resulted in worms being treated with higher molar equivalents of the other amino acids compared to tyrosine. We found that while other amino acids do increase the formation of dauers by eak-4 mutants, none was as potent as tyrosine (Figure 9G). This suggests that the effect on dauer formation shows selectivity for the presence of tyrosine. The effects of tyrosine and to a lesser extent the other amino acids is not due to a toxic effect of the amino acid supplementation as treated worms showed a similar lifespan to untreated worms (Figure S9).
Besides being a building block for proteins, tyrosine serves as a precursor for the synthesis of catecholamine neurotransmitters. In vertebrates, there is evidence that the levels of tyrosine as a precursor influences the synthesis of these neurotransmitters [73]. Hence, one possibility is that tatn-1 mutations raise tyrosine levels and facilitate its conversion into the neurotransmitters dopamine, octopamine, or tyramine which could produce the observed phenotypes. To test this possibility, we blocked dopamine synthesis with the cat-2 mutation, which affects the worm tyrosine hydroxylase gene, and we blocked octopamine and tyramine synthesis with the tdc-1 mutation, which removes the enzyme tyrosine decarboxylase [74], [75]. We found that both cat-2; eak-4; tatn-1 and tdc-1; eak-4; tatn-1 mutants are similar to eak-4; tatn-1 mutants with regards to the formation of dauers (Figure 9H and Figure 9I). These data demonstrate that excessive synthesis of dopamine, octopamine, or tyramine is not responsible for the tatn-1 phenotype. Instead tyrosine is directly sensed by the worms and acts as a developmental regulator.
Together our results identify tyrosine and tyrosine aminotransferase activity as a modifier of daf-2/IGFR effects in C. elegans (Figure 10). While the control of tyrosine aminotransferase expression and activity has been extensively studied as a target of insulin signaling in vertebrates [30]–[34], [36], [38], a connection between tyrosine aminotransferase or tyrosine metabolism and insulin action has not been demonstrated. Prior work in C. elegans has suggested that the hpd-1 gene, which encodes 4-hydroxyphenylpyruvate dioxygenase, is repressed in daf-2 mutants and that knock-down of hpd-1 by RNAi delayed dauer exit and extended lifespan [39]. However, the mechanism involved has been unclear. Our work shows that both tatn-1, and likely hpd-1, impact on daf-2/IGFR signaling through increasing tyrosine levels in the animal.
We find the effects of tyrosine on daf-2/IGFR signaling to be complex with roles for both the daf-16/FOXO transcription factor and the aak-2/AMPK seen (Figure 10). One way that high tyrosine levels could interact with daf-2/IGFR signaling would be for tyrosine to somehow activate aak-2/AMPK. AMP kinases are a known regulator of both daf-16 and the vertebrate homolog FOXO3 [51], [76]. AMPK regulates FOXO transcriptional activity through the phosphorylation of up to six sites on these proteins. In our work, aak-2/AMPK mutations suppress the dauer promoting effects of tatn-1 mutations, treatment of worms with the AMPK agonist AICAR is able to mimic the effects of tyrosine, and increases in the active phosphorylated form of AAK-2 are seen in the tatn-1 mutant. These findings demonstrate that elevated tyrosine levels activate aak-2/AMPK which could then phosphorylate daf-16/FOXO. This phosphorylation event could then interfere with the inhibitory effects of an intact daf-2/IGFR pathway on daf-16/FOXO activity. Further work would be needed to test this model, and the ability of mutants lacking aak-2/AMPK or daf-16/FOXO to still respond to elevated tyrosine levels also supports the presence of alternate, currently unknown downstream pathways.
These alternate pathways could either lie in parallel to daf-16/FOXO or could be the dominant response pathway with daf-16/FOXO only playing a permissive role, especially at lower tyrosine levels. One possible alternate pathway involves the CREB transcription factor crh-1 (Figure 10). The crtc-1 co-activator for the CREB transcription factors has been shown to be a target of regulation by aak-2/AMPK, and in vertebrates, CRTC co-activators are known to interact with insulin signaling in mediating the hepatic metabolic adaptation to the fed versus fasting state [53], [77], [78]. We find that tatn-1(qd182) mutants show evidence of crh-1/CREB-regulated gene expression, and a crh-1/CREB mutant mimics the interaction of eak-4 and tatn-1. Perhaps elevated tyrosine levels lead to the activation of aak-2/AMPK which then results in the activation of daf-16/FOXO and inhibition of crtc-1 and crh-1/CREB (Figure 10). The presence of paired downstream pathways could explain the partial requirement for daf-16/FOXO, especially at higher tyrosine levels.
Beyond effects on daf-16/FOXO and crh-1/CREB, elevated tyrosine, especially at high levels, could also have hormetic effects via changing cellular redox status, producing ER stress, or perturbing the protein folding environment [79]. These effects could account for the positive effects of increased tyrosine on longevity, and some of the genes involved in sensing hormetic stresses, such as the HSF-1 ortholog hsf-1, also interact with daf-16/FOXO and play roles in dauer formation [80]–[82]. Alternately, tyrosine could act via a novel pathway that operates independently of daf-16/FOXO or crh-1/CREB, especially at higher levels. For example, the vertebrate calcium-sensing receptor, PPARγ nuclear receptor, and aryl hydrocarbon receptor (AHR) have all been shown to respond to aromatic amino acids, though their connection to insulin action is currently unclear [83]–[85]. The recent finding that Akt and Foxo1 are largely dispensable for the control of hepatic metabolism by insulin in vivo has suggested that FOXO- independent pathways exist and play important roles in metabolic control [86].
Our work also provides insights into the eak genes which are known to act via unclear mechanisms to reduce daf-16/FOXO transcriptional activity while not significantly affecting the subcellular localization of DAF-16/FOXO [26], [27], [29]. We find that beyond enhancing the inhibitory effects of akt-1 on daf-16/FOXO activity, the eak genes also suppress the effects of amino acids and AMPK activity on daf-16/FOXO activity. Additional work will be needed to understand if the eak genes normally represent a control point where the effects of these metabolic signals on insulin signaling can be enhanced or suppressed.
We find that the regulation of tatn-1 expression in worms is complex with daf-2 activity, diet, and environmental conditions each contributing to the expression level (Figure 10). In vertebrates, tyrosine aminotransferase has also been shown to undergo regulation at the transcriptional, translation, and degradation levels in response to hormonal and nutritional cues [66]. We find that daf-2/IGFR activity inhibits tatn-1 gene transcription but raises TATN-1 protein levels. This is consistent with work in vertebrates showing that insulin shows complex effects on tyrosine aminotransferase expression with actions at both the transcriptional and translational level [30]–[34], [36], [38]. Nutritional cues appear to also be an important regulator because we find that the E. coli strain used as food has a powerful effect on the expression of tatn-1 and these effects parallel the effects of the weaker tatn-1(baf1) allele on dauer formation. In rats, the activity of hepatic tyrosine aminotransferase varies several-fold during the day with a peak during the evening and nadir in the early morning [87]. Studies of the cyclic variation have demonstrated that dietary protein intake is a prime inducer of tyrosine aminotransferase levels [65]. Feeding animals a protein-free diet results in a constant low level of tyrosine aminotransferase, whereas feeding animals protein meals at differing times produces corresponding shifts in enzyme production. Among amino acids, some such as tryptophan are potent inducers of tyrosine aminotransferase expression [66]. The mechanisms accounting for the dietary effects of amino acids on tyrosine aminotransferase are currently unclear. This could suggest a role for additional nutrient sensitive pathways which may well be conserved as we find that both tryptophan and tyrosine act as tatn-1 inducers in worms. Finally, we find a novel role for environmental conditions on tatn-1 expression as lower temperatures promote expression and higher temperatures inhibit it. How changes in temperature translate into the observed effects is unclear but perhaps hormonal changes mediated by the cytochrome P450 daf-9 and the nuclear hormone receptor daf-12 or changes mediated by thermosensory neurons are involved [88]. Together this suggests that the control of tyrosine aminotransferase activity, which modulates tyrosine levels, could be controlled via a complex network of internal and external cues. Our finding that changes in tyrosine levels alter both signaling pathways and gene expression patterns could suggest that carefully controlling tyrosine metabolism and ultimately tyrosine levels plays an important role in overall homeostasis (Figure 10).
Recent work has suggested that levels of specific amino acids, particularly branched chain and aromatic amino acids, could influence insulin sensitivity in people and mice [1]–[6]. While the exact role of aromatic amino acids in metabolic disease is unknown, our results suggest that these could play a causal role in either insulin-resistance or the development of diabetes [89]. Given the complex nature of tyrosine aminotransferase regulation, subtle changes in hormone levels, diet, and perhaps other factors could lead to changes in hepatic tyrosine metabolism and contribute to changes in serum aromatic amino acid levels. There may also be significant changes during the day due to dietary intake or release from internal stores such as muscle. As tyrosine levels increase, it is possible that, as in worms, the increases modify responses to insulin signaling and augment pre-existing insulin resistance in a harmful way (Figure 10). The connection between insulin signaling and tyrosine metabolism could potentially even lead to a vicious cycle of reduced insulin signaling producing elevated tyrosine levels which then lead to a further reduction in insulin signaling
Tyrosine aminotransferase has also been found to be a tumor suppressor gene in human hepatocellular carcinoma (HCC) [7]. The human tyrosine aminotransferase gene is located on 16q, which is frequently deleted in HCC, and analysis of tumors reveals that gene deletion or silencing via hypermethylation is common [7]. Consistent with an inhibitory role in the pathogenesis of liver cancer, transfection of HCC cancer cell lines with a tyrosine aminotransferase transgene suppressed malignant behavior such as growth in soft agar and the formation of tumors in nude mice. In these cells, tyrosine aminotransferase expression also acted to inhibit tumor formation via the stimulation of apoptosis, but the exact molecular events are still unclear [7]. Our data would suggest that the activation of AMPK or the downstream effects of AMPK on FOXO transcription factors, such as FOXO3, or CREB would be attractive targets for future study. Alternately, we also saw down-regulation of genes involved in DNA repair so the elevated tyrosine levels could also promote the accumulation of additional cancer promoting mutations (Table S5). Together these findings suggest that extracellular or intracellular tyrosine levels could act as signaling molecules involved in the control of cell growth, differentiation, and physiology.
All C. elegans strains were propagated on standard nematode growth agar (NGA) plates containing streptomycin (200 µg/mL) and spotted with OP50-1, as previously described [90]. For specific experiments, worms were fed HB101, OP50, or HT115 E. coli strains using NGA containing streptomycin (HB101) or no antibiotics (OP50 and HT115).
The following C. elegans mutants were obtained from the C. elegans Genetics Center, which is supported in part by NIH Office of Research Infrastructure Programs (P40 OD010440): daf-16(mgDf50) I, cat-2(e1112) II, pah-1(ok687) II, tdc-1(ok914) II, crh-1(tz2) III, daf-2(e1368) III, hpd-1(ok1955) III, unc-119(ed3) III, akt-1(mg144) V, akt-1(mg306) V, aak-2(gt33) X, akt-2(ok393) X, daf-12(rh61rh411) X, pdk-1(mg142) X, sgk-1(ok538) X, muIs84[pAD76(sod-3::GFP)], lpIs14 [daf-16f::GFP+unc-119(+)], and lpIs12 [daf-16a::RFP+unc-119(+)]. tatn-1(baf1) X has been described previously, and tatn-1(qd182) was identified in an unrelated mutagenesis screen and is a gift from Daniel Pagano and Dennis Kim [43], [44]. muIs109[daf-16::GFP] X has been described previously and is a gift from Malene Hansen [59]. daf-16(mgDf47), hsd-1(mg433) I, eak-3(mg344) III, eak-4(mg348) IV, eak-7(tm3188) IV, sdf-9(mg337) V have been described previously [26]–[28], [91]. Double and triple mutants were generated by standard genetic crosses, and the genotypes of strains were confirmed by PCR using oligos which detect gene deletions or RFLP's associated with the mutation (Table S6). Throughout this work tatn-1 is implied to refer to the tatn-1(baf1) allele except specifically as noted otherwise.
Worm embryos were isolated by sodium hypochlorite treatment, and eggs were transferred to plates, and grown at the indicated temperature in a designated incubator. The plates were scored two or three days later under a dissecting microscope for the presence of L2, dauer, and L3/L4 and older worms. We conducted control experiments to determine the robustness and reproducibility of visual scoring by having several scorers evaluate a series of still images and corresponding movies of larvae of different developmental stages. We then compared the correlation between the scorers for the entire series via the use of a kappa statistic [92]. These experiments indicated that scoring was consistent between raters within the lab with all comparisons showing “substantial” to “almost perfect” agreement (Table S7) [92].
For each assay, approximately 100 worms were scored from each of two to three plates set up in parallel for each genotype used in an experiment. This resulted in 200 to, more typically, 300 animals being scored for each genotype within an experiment. Each experiment was repeated at least once with comparable results, which resulted in between 400–600 worms being scored per genotype in total. The percentages of each stage were graphed using Prism5 software, and the graphs show pooled data from a single trial. To perform pairwise comparisons between mutant strains, a contingency table was set up using the counts for L2, dauer, and L3/L4 categories, and p-values were calculated using Fisher's exact contingency test within SAS version 9.3. To determine SDS resistance, worms were washed from plates with 1% SDS, and then incubated for 20 minutes with gentle rocking. Worms were then pelleted and washed with water. Aliquots were scored for survival as demonstrated by movement, and each experiment was repeated at least once. The percentages of living worms were graphed with Prism5.
Lifespan assays were conducted as previously described at 20°C using either NGA or RNAi plates containing 50 µM FUDR [93]. Lifespan assays for N2, tatn-1(baf1), eak-7(tm3188), and eak-7(tm3188); tatn-1(baf1) used NGA plates spotted with HB101, and worms were grown from eggs at 16°C to minimize larval arrest. Lifespan assays for N2, tatn-1(baf1), aak-2(gt33), and tatn-1(baf1), aak-2(gt33) used NGA plates spotted with HB101. Lifespan assays for amino acid treated N2 worms used either NGA plates or NGA plates supplemented with 1 mg./mL tyrosine, glycine, or isoleucine, and then spotted with HB101. Lifespan assays using daf-2 and daf-16 RNAi treatment used NGA media supplemented with carbenicillin (50 µg/mL) and isopropyl β-d-thiogalactopyranoside (IPTG, 1 mM). RNAi treatment for daf-16 was started at egg hatching while daf-2 RNAi treatment started on day 1 of adulthood with larval development occurring on NGA plates spotted with HB101 at 20°C.
For all lifespan assays three plates containing 40 worms each, for each genotype were set up, as well as an extra plate, with worms to replace worms that had crawled off the plate, bagged, or exploded, to reduce the number of censored events. Prism5 (Graphpad Software) was used to generate graphs and perform log-rank testing for curve comparisons. SAS was used to create lifetables and calculate mean survival.
Amino acids (Sigma-Aldrich) were dissolved as 45 mg/mL stock solutions in water, and then added to molten NGM to obtain a 1 mg/mL final concentration. These plates were dried and spotted with HB101 before use.
Worms were grown on HB101 spotted NGM plates for two days at 25°C before being washed from the plates and then being rinsed twice with miliQ water. For Figure S9A, NGM plates either with or without 1 mg./mL tyrosine cast into the agar were used to grow the worm culture. To normalize the samples for worm number, a 5 µL aliquot was removed and scored for worm number. Amino acids were then extracted using aqueous methanol and crushing with a mortar and pestle as previously described [94]. The methanol solution was removed by evaporation and the residue stored frozen at −80°C. Amino acid analysis was performed via liquid chromatography tandem mass spectrometry following reconstitution of the residue in 0.1 mL water as described previously for urine [95]. Amino acid content was then either normalized to total worm number in the sample (Figure 9B, Figure 9C, and Figure S9A) or normalized to the levels of individual non-aromatic amino acids and then divided by the average normalized level observed in the wild-type N2 samples (Figure 9D).
Aliquots from a 250 mM AICAR solution dissolved in water (Cell Signaling Technology) were spotted onto NGA plates spotted with HB101 to give a final concentration of 0.125 mM 1 hour before eggs were added to the plates. A comparable volume of water alone was used as a negative control.
Worm embryos were isolated by sodium hypochlorite treatment from N2 and eak-4(mg348); tatn-1(baf1) adults, and the eggs were transferred to NGA plates spotted with HB101. The plates were incubated at 25°C for 24 hours so most of the population was L2 larvae. The worms were washed from plates with water and washed with water to remove bacteria. Nucleotides were then extracted from the worms as previously described [50]. The resulting extract was stored at −80°C until analysis. ATP, ADP, and AMP levels were measured by HPLC with UV detection of individual nucleotides.
Images of worms were obtained with a BX51 fluorescence microscope and quantified using ImageJ software as previously described [44].
For sod-3 expression, worm embryos were isolated by sodium hypochlorite treatment, and eggs were transferred to NGA plates spotted with HB101 and incubated at 25°C for 24 hours. The worms were washed from plates in water, pelleted by centrifugation, washed with water, and frozen for storage. RNA extraction, reverse transcription, and quantitative PCR were performed as previously described [27], [44]. The geometric mean level of the control genes pmp-3, cdc-42, and Y45F10D.4 were used to normalize the samples, and the relative levels of sod-3 expression were determined using the 2−ΔΔCt approach [96], [97].
For tatn-1 expression, N2 and daf-2(e1368) embryos were isolated by sodium hypochlorite treatment, and eggs were transferred to S-basal to arrest the worms at the L1 stage. L1 larvae were added to NGA plates spotted with OP50-1 and the plates were incubated at 20°C for 3 days. Adult worms were washed from plates and RNA was isolated as described above. The geometric mean level of the control genes pmp-3, cdc-42, and Y45F10D.4 were used to normalize the samples, and the relative levels of tatn-1 expression were determined using the 2−ΔΔCt approach [96], [97].
The oligos used to detect pmp-3, cdc-42, and Y45F10D.4 have been previously described [98]. The expression of sod-3 was detected using the oligos 5′-CCAACCAGCGCTGAAATTCAATGG-3′ and 5′-GGAACCGAAGTCGCGCTTAATAGT-3′ [99]. The expression of tatn-1 was detected using 5′-CTTGATCAGAGAAGAATCAGTG-3′ and 5′-GAGTGTTGATTGAAGTTGCG-3′. These oligos were designed to cross intron-exon boundaries using the PerlPrimer program [100].
N2 wild-type control and tatn-1(qd182) mutant worms were synchronized via the use of hypochlorite treatment and grown on HB101 spotted NGM plates at 25°C for 2 days. These conditions and time point correspond to the conditions used for the amino acid analysis separately performed using these strains. The worms were washed from the plates and were then washed twice with miliQ water. The worm pellet was then suspended in QIAzol lysis reagent (Qiagen, Valencia, CA) and frozen at −80°C. Total RNA was isolated using the Qiagen miRNeasy mini kit and the RNA yield was measured by spectrophotometry. Total RNA was sent to Expression Analysis (Durham, NC) for analysis including bioanalyzer electrophoresis to ensure RNA quality followed by library preparation using the Illumina TruSeq RNA sample prep kit. The resulting library was subjected to high-throughput 50 nucleotide paired end sequencing using an Illumina sequencer at a depth of 17 million reads per sample.
The resulting sequence data was clipped using internally developed software from Expression Analysis and matched to the C. elegans genome using RSEM [101]. The resulting transcript counts were then normalized using the upper quartile normalization approach [102]. Differentially expressed genes were then identified through the use of ANOVA testing and genes with a FDR score of 5% or lower were considered to be differentially expressed. This lead to the identification of 4622 genes as being differentially expressed (890 up-regulated and 3732 down-regulated) between tatn-1(qd182) and wild-type N2.
Over-represented gene classes were identified in the up-regulated and down-regulated genes through the use of DAVID [60]. Analysis of the transcriptome data for daf-16/FOXO and crh-1/CREB regulated genes was performed using the Gene Set Association Analysis for RNA-seq program (available at http://gsaa.unc.edu/login/index.html) [63]. GSAA calculates a differential expression score for each gene in the entire RNA-seq dataset, 20408 genes in all, and then uses a running weighted Kolmogorov-Smirnov test to examine association of an entire gene set with each phenotypic class. The strength of the association is measured by the association score (AS) where positive scores indicate association of the gene set with the phenotype, and statistical significance is measured by a false discovery rate (FDR) that is adjusted for multiple testing. The daf-16/FOXO regulated genes were from previously published microarray data from Murphy et. al., and the crh-1/CREB regulated genes were from recently published microarray data from Mair et. al. [53], [62].
Clones for the tatn-1 promoter and cDNA where purchased from Open Biosystems and verified by sequencing [103], [104]. A tatn-1p:tatn-1 cDNA:GFP transgene was generated using Gateway cloning and the vector pDEST-MB14 [104]. The resulting transgene or punc-119cbr, which contains the unc-119 gene from Caenorhabditis briggsae, was used to bombard HT1593 (unc-119(ed3)) as previously described [105], [106]. From bombardment we obtained bafIs130 from punc-119cbr, which carries the unc-119 gene from Caenorhabditis briggsae, and bafIs131 from the tatn-1p:tatn-1 cDNA:GFP transgene [106]. Both transgenes were outcrossed with N2 and then mated with eak-4(mg348); tatn-1(baf1) to test for rescue of the dauer formation phenotype. The bafIs131 transgene was also crossed into a daf-2(e1368) mutant to generate daf-2(e1368); bafIs131.
Worm embryos were isolated by sodium hypochlorite treatment, and eggs were transferred to NGA plates spotted with HB101 and incubated at 25°C for 24 hours. For treatment with AICAR, the plate was spotted 1 hour before adding the eggs with aliquots from a 250 mM AICAR stock that was dissolved in water. As a positive control, N2 wild-type worms were grown as above, washed off in S-basal, and then exposed to 10 mM sodium azide in S-basal. This dose of S-basal has been previously shown to result in a 50% decrease in ATP concentrations in treated worms [54]. The worms were washed from plates in water, pelleted by centrifugation, washed with water, and suspended in 1× LDS loading buffer (Invitrogen) before being heated to 70°C in a Bransonic sonicator waterbath (Branson) for 30–40 minutes [107]. After heating the samples were centrifuged and the supernatant retained for analysis. The protein levels were measured using the CB-X protein assay kit (G-Biosciences), and 30 µg of protein was run on a 10% Nupage SDS-PAGE gel (Invitrogen) and blotted to a nitrocellulose membrane. Phospho-AAK-2 was detected using a rabbit monoclonal antibody (Cell Signaling Technology #2535), and actin was detected using a rabbit anti-actin antibody (Cell Signaling Technology #4967) followed by detection by an anti-rabbit HRP secondary antibody and visualization using chemiluminescence (Bio-Rad). The resulting X-ray film was scanned and quantified using gel analysis tools in ImageJ [108].
|
10.1371/journal.pntd.0007191 | The importance of extracellular vesicle purification for downstream analysis: A comparison of differential centrifugation and size exclusion chromatography for helminth pathogens | Robust protocols for the isolation of extracellular vesicles (EVs) from the rest of their excretory-secretory products are necessary for downstream studies and application development. The most widely used purification method of EVs for helminth pathogens is currently differential centrifugation (DC). In contrast, size exclusion chromatography (SEC) has been included in the purification pipeline for EVs from other pathogens, highlighting there is not an agreed research community ‘gold standard’ for EV isolation. In this case study, Fasciola hepatica from natural populations were cultured in order to collect EVs from culture media and evaluate a SEC or DC approach to pathogen helminth EV purification.
Transmission electron and atomic force microscopy demonstrated that EVs prepared by SEC were both smaller in size and less diverse than EV resolved by DC. Protein quantification and Western blotting further demonstrated that SEC purification realised a higher EV purity to free excretory-secretory protein (ESP) yield ratio compared to DC approaches as evident by the reduction of soluble free cathepsin L proteases in SEC EV preparations. Proteomic analysis further highlighted DC contamination from ESP as shown by an increased diversity of protein identifications and unique peptide hits in DC EVs as compared to SEC EVs. In addition, SEC purified EVs contained less tegumental based proteins than DC purified EVs.
The data suggests that DC and SEC purification methods do not isolate equivalent EV population profiles and caution should be taken in the choice of EV purification utilised, with certain protocols for DC preparations including more free ES proteins and tegumental artefacts. We propose that SEC methods should be used for EV purification prior to downstream studies.
| Recent pathogen research has identified extracellular vesicle (EV) release from many organisms. EVs are small membrane bound organelles, which have different origins, sizes and composition. It is important that the optimal EV purification method is realised in order to obtain high quality EVs to have confidence in understanding EV biology and function. In this study, the zoonotic parasite, Fasciola hepatica, was cultured as a case study to investigate the importance of EV purification from helminth culture media. Investigating two purification methods, it was found that size exclusion chromatography EV isolation led to a reduction of contaminating excretory-secretory and tegumental proteins. This research highlighted that purification methods do not isolate equivalent EV population profiles with similar EV purities and that size exclusion chromatography methods are likely better suited for downstream helminth EV studies and application development, compared to a differential centrifugation method.
| Extracellular vesicle (EV) purification is challenging to standardise due to the diversity of sample composition producing EVs (cell cultures and body fluids), the need for high recovery of functional EVs, the quality of EV preparation and the simplicity of isolation [1–5]. Therefore, there is no current gold standard for EV isolation [6].
The most widely accepted method to isolate EVs involves differential centrifugation (DC). This method encompasses sequential centrifugations, increasing in speed and time, to pellet particles decreasing in size [1,2,4–11]. DC is highly reliable, yet it is a labour-intensive procedure, requiring large sample volumes to obtain low EV yields [2,4,12]. Furthermore, EV yield and purity is dependent upon DC parameters such as the rotor type, centrifugal force, centrifugation period and temperature [4,11,13]. Consequently, protocol standardisation is difficult, leading to incomparable EV characterisation and functional investigations [4,5]. To overcome EV purification challenges, many downstream EV investigations require further isolation procedures to accommodate DC, to improve sample purity and validate research analysis.
Recently, size exclusion chromatography (SEC) has been used to successfully purify EVs. SEC has been observed to separate protein aggregates and lipoproteins from EV samples as well as preclude EV or protein aggregation [1,3,4,9,14–16]. However, elution fractions containing a high yield of protein, also contain small EVs (<75 nm), meaning SEC can be selective upon EV size collection. In addition, the purest vesicle elution fraction post SEC produces low yields, so the quantity of fraction collection is dependent upon the EV purity needed for experimental investigation [14]. However, research has observed that SEC EV isolation produces greater sample purity, compared to DC and precipitating agents (Polyethylene glycol and PRotein Organic Solvent PRecipitation, PROSPR) as evidenced by EV markers (Alix and CD9) in greater concentration intensity in EV samples [2,15,17]. Potentially, SEC could be reproducible in all research settings, as it is easy to use, removes most contaminating proteins and maintains EV morphology.
Many studies in independent laboratories confirm both pathogenic flatworm and nematode helminths produce EVs, albeit during in vitro ex-host experimentation. Of note is the recent expansion of information on the EVs of the liver fluke Fasciola hepatica. Fascioliasis, infection with either F. hepatica or F. gigantica, is a major neglected zoonotic disease that infects humans and ruminant species worldwide [18,19]. At least 2.4 million people are currently infected in over seventy countries, with millions more at risk of this food borne disease [20]. Furthermore, the disease is a significant animal health and food security issue, costing the global livestock industry an excess of $3 billion per annum [21]. In the absence of protective vaccines, control is usually via anthelmintics, with triclabendazole (TCBZ) being the drug of choice, especially for acute disease caused by pathogenic juvenile F. hepatica. F. hepatica resistance towards TCBZ has spread widely, threatening future chemotherapeutic based control [22]. Therefore, the development of novel approaches and options for F. hepatica control must be considered a research and government priority.
The recent discovery of F. hepatica EVs identified in excretory-secretory products (ESP) has led to us re-evaluating the host-pathogen interface [23,24]. EVs of pathogen origin are enriched with pathogen molecules, thus they could potentially be utilised for improved control. In pathogens, EVs have been found to function to either promote or inhibit host immunity with recognised EV immunogenic properties highlighting EVs as vaccine preparations [25]. Furthermore, EVs have the ability to transport molecules to recipient cells, which could be utilised for drug delivery [26]. Therefore, EVs have a major role in pathogen infection and could be exploited to develop novel therapeutic approaches. Thus, it is important that the optimal EV purification method is utilised to obtain high purity EVs, which represent the whole biological EV population that would likely interact with the host environment.
Furthermore, given that all parasitic helminth EV studies completed to date have utilised DC without further EV enrichment, it is vital for the field to look at alternative EV isolation strategies [23,24,27–35]. This is particularly pertinent given the need for more functional studies into EV biology thus requiring EVs of high purity. Literature suggests that SEC is a suitable EV purification method to compare against DC purification methodology, which has been primarily used in helminthology, for the aim of investigating the most advantageous method for purifying helminth EVs from culture media for further downstream analysis, where EVs isolated are a biological representation of the helminth EV population. Therefore, in the current study, the EVs from adult F. hepatica, as a model pathogenic helminth, have been purified allowing the comparison of SEC EV purification, to the widely accepted method of DC EV preparation.
This study has discovered that SEC EV isolation leads to smaller sized and lower diversity EV populations, with importantly a higher EV purity to free ESP yield, a less diverse EV proteome and different EV gene enrichment profiles compared to DC purified EVs. Therefore, the data suggests that DC and SEC purification methods do not isolate equivalent EV population profiles and caution should be taken in the choice of EV purification utilised with functional assays incorporated into the isolation pipeline. This research has highlighted SEC methods with functional assays as the methodology of choice for helminth pathogen EV studies and application development in the absence of a gold standard purification method.
A local collaborating abattoir (Wales, U.K.), where animals were processed as part of the normal work of the abattoir, gave consent for the collection of live adult F. hepatica from naturally infected ovine livers of animals immediately post-slaughter, where F. hepatica were maintained as previously described [36]. Briefly, F. hepatica were thoroughly washed in phosphate buffered saline (PBS) at 37°C to remove host material. F. hepatica were then transported to the laboratory in Fasciola saline (Dulbecco’s modified Eagle’s medium (DMEM) (w/o NaHPO3 and PO4) plus 2.2 mM Ca (C2H3O2), 2.7 mM MgSO4, 61.1 mM glucose, 1 μM serotonin, 5 μg/mL gentamycin, 15 mM N-2-hydroxyethylpiperazine-N′-2-ethanesulfonic acid (HEPES), pH7.4) at 37°C for 1 hour. Following incubation, F. hepatica were maintained in fresh Fasciola saline (1 ml/ F. hepatica) at 37°C for 5 hours. All F. hepatica remained alive after in vitro culture incubations. Both F. hepatica and the culture media was immediately frozen and stored at -80°C, until further experimentation.
EVs were purified from the culture media based on the method according to Thery et al. [37] with slight modifications. In brief, culture media was centrifuged at 300 x g for 10 minutes at 4°C and then at 700 x g for 30 minutes at 4°C, removing any large particulates. The supernatant was ultracentrifuged at 100,000 x g for 80 minutes at 4°C (Optima L-100 XP ultracentrifuge (Beckman Coulter, High Wycombe, UK) using a Type 50.2 Ti rotor) and the EV depleted culture media was removed and stored at -80°C. Of note, the 10,000 x g centrifugation step for 30 minutes carried out by Thery et al. [37] was not undertaken within the current methodology to ensure isolation of the whole biological EV population. Therefore, it should be acknowledged that debris may be still present in the EV enriched sample. The pellet was washed in PBS and vortexed until the pellet had been suspended in solution. The ultracentrifugation step was repeated and the supernatant was discarded. The pellet was then re-suspended in 200 μl PBS by vortexing and stored at -80°C until further experimentation.
Culture media was centrifuged at 300 x g for 10 minutes at 4°C and then at 700 x g for 30 minutes at 4°C, removing any large particulates. The sample was then concentrated using 10 KDa MWCO Amicon Ultra-15 Centrifugal Filter Units (Merck Millipore), following the manufacturer’s guidelines. Briefly, the sample was centrifuged at 4000 x g for 20 minutes at 4°C, until approximately 500 μl of sample was retained in the filter. Filtration flow through was stored at -80°C. The sample was passed through qEVoriginal SEC columns (IZON science), utilising the manufacturer’s optimised protocol. Briefly, the column was rinsed with 10 ml of filtered (0.2 μm, Life Sciences) PBS. The sample was then added to the SEC column and the first 2.5 ml of flow through was discarded. The next 2.5 ml of flow through, containing EVs, was collected and stored at -80°C. The column was then washed with 10 ml PBS, which was combined with the filtration flow through to create EV depleted SEC ESP.
EV samples and culture media, containing F. hepatica ESP depleted of EVs, were fixed onto formvar/carbon coated copper grids (agar scientific) by adding 10 μl sample to the grid for 45 minutes on ice. Grids were then placed on the viscous of 4% v/v uranyl acetate for five minutes on ice. Grids were stored at room temperature for at least 24 hours before being imaged using the transmission electron microscope (Jeol JEM1010 microscope at 60 kV). Images were developed and the size of 200 EVs per purification method were measured (nm) using imageJ (https://imagej.nih.gov/ij/) and statistically analysed by Mann-Whitney U test in R studio (https://www.rstudio.com/).
EV samples were diluted 1:10 with deionised water and adsorbed onto freshly cleaved mica sheets (Agar Scientific AGG250-1) for two minutes and dried under a nitrogen stream. Samples were then scanned with a Park Systems E100 AFM, using silicon probes (NT-MTD: NSG-01 & NSG-03PT) in non-contact mode. Topographic height and phase images were scanned at 512 x 512 pixels at a rate of 0.2 Hz. Images were analysed using Gwyddion software (www.gwyddion.net/).
F. hepatica somatic sample was prepared by homogenising F. hepatica in homogenisation buffer (20 mM potassium phosphate buffer (pH 7.4), 0.1% (v/v) triton X-100 and protease inhibitor tablets, EDTA-free) (0.5 ml/ F. hepatica) on ice, before centrifuging at 21,000 x g at 4°C for 30 minutes. The supernatant was termed the cytosolic fraction, and was stored at -20°C until further experimentation.
Culture media, containing F. hepatica ESP depleted from EVs from both purification methods, was precipitated with an equal volume of ice-cold 20% (w/v) trichloroacetic acid (TCA) in acetone for one hour at -20°C. The sample was centrifuged at 21,000 x g, 4°C for 20 minutes. The pellet was washed twice in ice-cold acetone with centrifugation 21,000 x g, 4°C for 20 minutes between washes. The pellet was left to air dry at -20°C for 15 minutes, before being resuspended in PBS and stored at -20°C until further experimentation.
EV samples were centrifuged at 100,000 x g at 4°C for 30 minutes and the supernatant was discarded. Lysis buffer (PBS, 0.1% (v/v) triton X-100 and protease inhibitor tablets, EDTA-free) was added to the pellet and the sample was sonicated for 30 seconds and then rested on ice for 30 seconds and repeated three times. Following lysis, the sample was centrifuged at 100,000 x g at 4°C for 30 minutes and the soluble fraction in the supernatant was TCA precipitated as above and stored at -20°C until further use. The insoluble fraction was washed following the protocol of Hart et al. [38]. Briefly, the pellet was re-suspended in sodium carbonate buffer (0.1 M Na2CO3 (pH 11), 10 mM EDTA, 20 mM DTT and protease inhibitor tablets, EDTA-free), vortexed and left for one hour at 4°C. The sample was then centrifuged at 100,000 x g at 4°C for 30 minutes and the supernatant was discarded. The pellet was washed in sodium carbonate buffer, vortexed, left for 30 minutes at 4°C and centrifuged as previously described. This wash step was repeated before the pellet was re-suspended in solubilisation buffer (20 mM potassium phosphate (pH 7.4), 4% SDS (w/v) and protease inhibitor tablets, EDTA-free). The sample was heated to 95°C for five minutes and then stored at -20°C until further experimentation.
Samples were quantified using either the bicinchoninic acid assay (Thermo scientific) or Bradford assay (Sigma), following the manufacturer’s protocol. Comparisons between sample protein concentrations were statistically analysed using Kruskal-Wallis test with the Dunn’s Post-hoc test using Sidak correction in R package (https://www.r-project.org/), where significance was considered p<0.05. Laemmli protein 4 x sample buffer (Bio-rad) was added to the sample (3:1 ratio) and heated to 95°C for 10 minutes. The samples were then loaded into 7 cm 12.5% Tris/glycine polyacrylamide gels and run using the Protean III system (Bio-Rad). Gels were run at 70 V, until the dye front passed through the stacking gel, and then the voltage was increased to 150 V until completion. Gels were fixed (40% ethanol (v/v), 10% acetic acid (v/v)) and stained with colloidal Coomassie Brilliant blue (Sigma). Gels were imaged using the GS-800 calibrated densitometer (Bio-rad).
Samples were prepared and run upon SDS-PAGE gels as previously mentioned. Gels were electrophoretically transferred onto Hybond-C extra nitrocellulose paper (GE Healthcare) using a Trans-Blot Cell at 40 V for two hours in transfer buffer (192 mM Glycine, 25 mM Tris-HCL (pH 8.3), 20% (v/v) methanol) according to the method of Towbin et al. [39]. Transfer efficiency was determined by staining the membrane with amido black (0.1% amido black (w/v), 10% acetic acid (v/v), 25% isopropanol (v/v)) for one minute. The membrane was then destained (10% acetic acid (v/v) and 25% isopropanol (v/v)), and washed three times in Tris buffered saline (100 mM Tris-HCL (pH 7.5), 0.9% sodium chloride (w/v)) with 1% Tween 20 (v/v) (TTBS). The membrane was then blocked overnight in TTBS and 5% skimmed milk powder (v/v) on a rocker at 4°C.
Primary antibody, either anti-glutathione transferase sigma class (Anti-FhGST-S1) at 1:20,000 [40], Anti-Fasciola cathepsin L1 at 1:6,000 (Anti-FhCat-L1) which was commercially made (Lampire) from polyclonal antibodies to a purified recombinant cathepsin L1 from F. hepatica expressed in yeast and raised in rabbits or Anti-fatty acid binding protein V (Anti-FhFABP-V) at 1:2,000 which was commercially made (Lampire) from polyclonal antibodies to a purified recombinant fatty acid binding protein V from F. hepatica expressed in yeast and raised in rabbits, in TTBS and 1% skimmed milk (v/v) was added to blocked membranes. Anti-FhGST-S1 and Anti-FhCat-L1 primary antibodies are known to have high specificity however, Anti-FhFABP-V is known to be reactive to FABP I, FABP II, FABP III and FABP V. The membrane was rocked for one hour at room temperature. The membrane was then washed three times for five minutes in TBS. The secondary antibody (IgG, anti-rabbit IgG (whole molecule) conjugated to alkaline phosphatase (AKP, Sigma) dilution 1: 30,000) in TTBS was then added to the membrane and rocked for one hour at room temperature. Membrane washing was then repeated as previously described. Recognised proteins were visualised using 5-bromo-4-chloro-3-indoyl phosphate/nitro blue tetrazolium liquid substrate system. In brief, 33 μl 5-bromo-4-chloro-3-indoyl phosphate (50 mg/ml) and 330 μl nitro blue tetrazolium (10 mg/ml) was added to 10 ml substrate buffer (0.1 M Tris, 100 mM sodium chloride, 5 mM magnesium chloride, (pH 9.5)). The detection solution was added to the membrane on the rocker at room temperature until either visualisation of banding occurred or up to five minutes. The reaction was stopped by water washes. Membranes were scanned using the GS-800 calibrated densitometer (Bio-rad).
All protein bands were excised from one dimensional SDS-PAGE electrophoresis and digested as described in Morphew et al. [41]. Briefly, protein bands were washed in 50% (v/v) acetonitrile and 50% (v/v) 50 mM ammonium bicarbonate at 37°C until destained. Destained bands were dehydrated in 100% acetonitrile at 37°C for 15 minutes, before being removed and dried at 50°C. Protein bands were then incubated with 10 mM dithiothreitol in 50 mM ammonium bicarbonate for 30 minutes at 80°C. The supernatant was discarded before bands were incubated with 55 mM iodoacetamide in 50 mM ammonium bicarbonate for 20 minutes at room temperature in the dark. The supernatant was discarded and the bands were washed twice for 15 minutes at room temperature with 50% (v/v) acetonitrile and 50% (v/v) 50 mM ammonium bicarbonate. Excess was removed before 100% acetonitrile was added to the bands at room temperature for 15 minutes. The supernatant was removed and bands were dried at 50°C. Bands were rehydrated and digested using 50 mM ammonium bicarbonate containing trypsin (modified trypsin sequencing grade, Roche, UK) at 10 ng/μl at 37°C overnight. The supernatant was stored, before 100% acetonitrile was added to the bands at room temperature for 15 minutes, followed by adding 50 mM ammonium bicarbonate to the bands at room temperature for 15 minutes. This step was repeated and from each step, the supernatant was removed and pooled for each band. The 100% acetonitrile step was repeated, and supernatant pooled, before the samples were vacuum dried (Maxi dry plus, Heto) and re-suspended in 20 μl of 0.1% (v/v) formic acid for tandem mass spectrometry.
Liquid Chromatography tandem mass spectrometry (Agilent 6550 iFunnel Q-TOF) coupled to a HPLC-Chip (1200 series, Agilent Technologies, Cheshire, UK) was used for peptide separations. The HPLC-Chip/Q-TOF system was equipped with a capillary loading pump (1200 series, Agilent Technologies) and a nano pump (1200 series, Agilent Technologies). Sample injection was conducted with a micro auto sampler (1100 series, Agilent Technologies), where 1 μl of sample in 0.1% formic acid was loaded on to the enrichment column at a flow of 2.5 μL/min followed by separation at a flow of 300 nL/min. A Polaris Chip was used (G4240-62030, Agilent Technologies), comprising a C18 enrichment/trap column (360 nl) and a C18 separation column (150 mm x 75 Âμm), where ions were generated at a capillary voltage of 1950 V. The solvent system was: solvent A (ultra-pure water with 0.1% formic acid), and solvent B (90% acetonitrile with 0.1% formic acid). The liquid chromatography was performed with a piece-linear gradient using 3–8% of solvent B over 0.1 minutes, 8–35% solvent B over 14.9 minutes, 35–90% solvent B over five minutes and hold at 90% solvent B for two minutes. Tandem mass spectrometry was performed in AutoMS2 mode in the 300–1700 Da range, at a rate of 5 spectra per second, performing MS2 on the 5 most intense ions in the precursor scan. Masses were excluded for 0.1 minutes after MS/MS was performed. Reference mass locking was used for internal calibration using the mass of 391.2843 Da.
Peak lists were generated with Mass Hunter Qualitative Analysis software (V B.06, Agilent Technologies) and exported as Mascot Generic Files. Samples were processed following Morphew et al. [42]. Briefly, samples were submitted to MASCOT daemon (Matrix Science, v2.4.1) MS/MS ions search against F. hepatica gene sequences, accessed through WormBase ParaSite (http://parasite.wormbase.org/, accession PRJEB6687, version WBPS9). Search parameters included setting the enzyme to trypsin with one missed cleavage allowed, setting fixed modifications to carbamidomethylation with variable modifications set for oxidation of methionine, fixing monoisotopic masses with unrestricted protein masses with peptide and fragment mass tolerances at ±1.2 Da and ±0.6 Da respectively (project accession PXD008737). Protein identifications were reported at a false discovery rate of 1%. For the overall list of proteins identified (data in S1 File), only proteins with at least 2 unique peptides and present in both biological replicates (n = 2) were selected. Protein sequences were searched using BLAST2GO (https://www.blast2go.com/) obtaining BLAST descriptions and gene ontology terms (data in S2 File). Gene ontology enrichment analysis was completed using GOATOOLS python package (https://github.com/tanghaibao/goatools) (data in S4 File) where the GO terms were not propagated up the hierarchy and p<0.05 identified significance.
In order to assess EV morphology post DC and SEC purification, both atomic force microscopy (AFM) and transmission electron microscopy (TEM) imaging were utilised. AFM and TEM identified that EV structures were characteristically diverse in size and morphology using both DC and SEC purification methods, representing a biological population of EVs (Fig 1). When comparing EV dimensions following TEM analysis, SEC EVs were significantly smaller (76 nm ± 44 SD) than DC EVs (95 nm ± 58 SD) (n = 200) (W = 14,726, p < 0.001) and DC purified EVs displayed a greater range of EV sizes than SEC (DC range = 505 nm, SEC range = 285 nm). All observed EVs were intact in TEM images for both purification methods. Aggregation of EVs and contaminants in the image background was observed using both purification methods in AFM and TEM images.
TEM micrographs produced for the culture media containing F. hepatica ESP depleted of EVs identified few EVs (data not shown). In addition, EVs were not found within the ultrafiltration flow through during the SEC purification method (data not shown). In all non-EV preparations debris and additional non-EV contaminants were identified.
EV preparations obtained from both DC and SEC purification methods were quantified for protein content for ESP (protein content for residual ESP after EV purification), whole lysed EV samples and soluble and insoluble EV fractions (Fig 2). All samples quantified were unrelated biological replicates. Analysis of the EV preparation via the DC purification method revealed comparable protein quantities between ESP (average = 30 μg, n = 3) and whole lysed EV samples (average = 40 μg, n = 3), even though the whole lysed EV samples had greater quantity variation (ESP range = 20 μg, whole lysed EV range = 140 μg). However, this protein profile was not replicated for SEC derived EV preparations. Protein quantity of SEC ESP was nearly threefold that of the lysed EV samples (ESP average = 240 μg, n = 3, whole lysed EV average = 80 μg, n = 3), even though whole lysed EV samples had a greater quantity variation (ESP range = 90 μg, whole lysed EV range = 180 μg). Furthermore, SEC purified samples produced a higher protein yield than DC in both ESP and whole lysed EV samples, with ESP showing a significant increase (D = -3.13, p = 0.02, n = 3). Similarly, SEC produced a greater protein yield of soluble and insoluble EV fraction protein (soluble protein average = 170 μg, n = 9, insoluble protein average = 210 μg, n = 9) than DC (soluble protein average = 30 μg, n = 9, insoluble protein average = 60 μg, n = 9). Both EV purification methods showed that there was a greater protein quantity of insoluble EV protein compared to soluble EV protein, with the SEC soluble fraction containing significantly less protein than insoluble protein fraction (D = -0.43, p = 0, n = 9).
Whole lysed EV and ESP samples displayed different protein profiles as observed by one dimensional sodium dodecyl sulfate-polyacrylamide gel (SDS-PAGE) electrophoresis, suggesting that both EV purification methods selectively isolate EVs away from F. hepatica secretions or ESP (Fig 3). In addition, biological replication of EV preparations indicated that protein quantification and preparation was reproducible as all sample replicates produced comparable protein banding patterns. Comparing lysed EV preparation protein profiles using DC and SEC purification methods, banding differences were identifiable, especially between 50–100 kDa and 25–37 kDa markers. Differences were less noticeable within the protein profiles of the soluble and insoluble EV fraction proteins and between the method by which the EVs were purified.
Following SDS-PAGE, mass spectrometry was employed to identify key differences between the profiles from a combined analysis of the soluble and insoluble EV protein fractions from SEC and DC isolated EVs. Following analysis, SEC purification revealed 68% protein matching between the soluble and insoluble SEC EV fractions with a 63% similarity of proteins for DC purified soluble and insoluble fractions. Following the combination of the soluble and insoluble protein identifications the whole SEC and DC purified EV proteomes were further analysed to identify key protein differences between the purification methods. Interestingly, 77% of proteins were comparable between SEC biological replicates and 77% between DC biological replicates.
In DC purified EVs, 392 proteins were identified, while 321 proteins were observed using SEC purification across both replicates. Of these proteins, 276 were comparable between purification methods demonstrating differences and similarities between the two methodologies. When looking at unique proteins to each method, 116 proteins were found exclusively in DC purified EVs and 45 proteins were found exclusively in SEC purified EVs (data in S5 File). Gene ontology analysis identified that EV proteins mostly showed biologically function in organic substance and primary metabolic processes as well as single-organism cellular processes irrespective of EV purification method (S6 File). These categories encompass proteins such as cathepsin L and GAPDH. However, there were significant differences in gene ontology terms such as catabolic process and microtubule-based process which were only found in DC purified EVs, while many more ‘regulatory-cell invasion’ terms (i.e. cellular localisation and response to stress) were only found in SEC purified EVs. Catabolic process and microtubule-based process only found in DC purified EVs refers to proteins such as Mov34/MPN/PAD-1 family protein, ubiquitin—protein ligase and alpha-tubulin. While, ‘regulatory-cell invasion’ terms only found in SEC purified EVs includes ADP-ribosylation factor family protein, T-complex protein 1 zeta subunit and glutathione peroxidase. Considering molecular function, transmembrane transporter activity and substrate-specific transporter activity gene ontology terms were only identified in SEC purified EVs including proteins such as IC domain protein HAD ATPase P-type family and hypothetical proteins. For cellular components, catalytic complex and intrinsic component of the membrane were the only categories found uniquely in EVs purified from DC including multicatalytic endopeptidase, ATPase family protein and proteasome subunit alpha domain protein and cell periphery, cell projection, plasma membrane, plasma membrane part and plasma membrane region were the only unique components found in EVs purified from SEC including IC domain protein, HAD ATPase, P-type family, ADP-ribosylation factor family protein and hypothetical proteins.
When specifically looking at protein identifications, DC EVs consistently demonstrated a greater amount of unique peptides hit per protein than SEC. The top 50 proteins found in EVs purified using SEC were also present in the DC EV purified proteome. However, proteins relating to gene scaffolds BN1106_s90B000599 (ATPase family protein), BN1106_s1277B000102 (HSP90 protein), BN1106_s63B000395 (hypothetical protein D915_01544) and BN1106_s285B000846 (unnamed protein product) were found in the top 50 proteins of DC purified EVs, but were not present in the SEC isolated EV proteome (highlighted within Table 1). Common EV markers from ExoCarta database (http://www.exocarta.org/) present in the top 50 EV proteins using both purification methods included heat shock proteins, glyceraldehyde-3-phosphate dehydrogenase, actin, 14-3-3 protein, annexin and tubulin. Unique EV markers, gelosin and phosphoglycerate kinase were found in the top 50 EV protein hits using DC purification, while a tetraspanin EV marker was found in the top 50 protein hits using SEC purified EVs.
EV proteins unique to SEC or DC purification were further investigated to assess the likelihood of non-EV contamination in each preparation likely from F. hepatica ESP or the tegument using data from recent proteomic studies (data in S5 File) [23,24,43–46]. Of the 45 proteins identified in SEC purified EVs only, 7 proteins were observed previously in ESP with 11 proteins identified previously in tegumental preparations. Looking at the 116 proteins identified in DC purified EVs only, 11 proteins were previously observed in ESP proteomic studies with an additional 27 proteins located in tegumental proteomic studies. Of note is the reduced abundance of the potential tegumental proteins in the SEC preparation in comparison to DC EV preparations. Based on the number of unique or total peptides for each protein as an relative assessment of abundance, potential tegument identifications from DC prepared EVs were in greater quantities (Average of 14 unique peptides and 47 total peptides per protein) compared to EVs from SEC preparations (Average of 9 unique peptides and 15 total peptides per protein).
Of interest in the DC prepared EVs was the identification of 5 proteins that have previously been well documented in the proteome of F. hepatica eggs. Specifically, two ferritin-like proteins (BN1106_s709B000642 & BN1106_s709B000627), heat shock protein-35a (BN1106_s7572B000046), oxidoreductase (BN1106_s3715B000086) and an omega class glutathione S-transferase (BN1106_s1029B000154) which were all identified within DC EVs only.
Gene enrichment was undertaken to determine significant over-representation of biological characteristics within the DC and SEC purified proteomes, compared to the F. hepatica genome set. Gene enrichment analysis revealed enriched gene ontology terms for both purification methods in biological process, molecular function and cellular component categories against the F. hepatica genome (S7 File). Several gene ontology terms were enriched, which were not represented in EVs from both purification methods. For example, in biological processes, ATP synthesis coupled proton transport, carbohydrate metabolic process, cellular iron ion homeostasis, DNA replication, gluconeogenesis, iron ion transport, protein transport, proton transport, regulation of protein phosphorylation, DNA templated transcription and transmembrane transport were enriched in DC purified EVs only, where categories included proteins such as ATP synthase F1 beta subunit and peptidase T1 family. While nucleocytoplasmic transport, nucleosome assembly, oxidation-reduction process, phosphatidylinositol metabolic process, protein dephosphorylation, proteolysis, regulation of actin filament polymerization, small GTPase mediated signal transduction and translation including proteins such as Ras family protein and aldehyde dehydrogenase family protein were not enriched in DC purified EVs. Pairwise comparisons of DC purified EVs compared to SEC purified EVs demonstrated that, when considering biological processes, there were significantly less translation gene ontology terms (p = 0.02) and significantly more DNA-templated regulation of transcription gene ontology terms (p = 0.04), when considering cellular components there were significantly less ribosome gene ontology terms (p = 0.04) and when considering molecular function, structural constituents of the ribosome gene ontology terms (p = 0.02) were significantly less enriched.
The relative abundances of three proteins, which are well recognised to be located in F. hepatica ESP, EVs and somatic fractions, were compared utilising Western blotting. Somatic, ESP, soluble and insoluble EV fraction samples were assessed using both DC and SEC EV purification approaches (Fig 4). Equal concentrations of each protein sample was assayed for each EV purification Western blot. Anti-fatty acid binding protein V (Anti-FhFABP-V) antibody recognition was observed within somatic, ESP and soluble EV fractions using both purification methods but not within the insoluble EV fraction yet, DC EV purification demonstrated a marginally higher anti-FABP V recognition in EV soluble protein concentrations than SEC methods. Recognition with anti-glutathione transferase sigma class (Anti-FhGST-S1), identified in all fractions (somatic, ESP, insoluble and soluble EV), was notably observed to a greater extent within the insoluble EV fraction using SEC purification rather than DC purification.
Anti-Fasciola cathepsin L1 (Anti-FhCat-L1) antibody recognition was identified in all fractions (somatic, ESP and soluble EV) with the exception of insoluble EV preparations. Of note is the increased recognition of Fasciola cathepsin L1 observed in the soluble EV fraction of DC purified EVs when compared to SEC EVs. Furthermore, there was greater recognition by Anti-FhCat-L1 in the ESP following SEC purification rather than DC. To further investigate the abundance of cathepsin L proteases in DC and SEC EVs the proteomic data sets generated in the current work were examined to reveal which cathepsin L proteases were identified in the respective EV preparations (Table 2). Cathepsin L protease identifications revealed a greater number of CL1 clade (CL1A and CL1D) identified in EVs purified by SEC, rather than EVs purified by DC. In addition, DC purified EVs contained a greater number of the CL5 clade members.
Given recent discoveries that EVs function in host-pathogen communication promoting or inhibiting host immunity, it is vital that reliable protocols for EV isolation, away from ESP, are conducted for downstream studies and application development. Therefore, the current work aimed to assess the current standard procedure for helminth pathogen EV analysis, DC, with a SEC approach well-established outside of the pathogen field when purifying EVs from culture media.
EV morphological characteristics assessed after both DC and SEC purification methods, identified that SEC purified EVs on average were smaller (76 nm ± 44 SD) and less diverse in size than DC purified EVs (95 nm ± 58 SD), indicating different EV populations were likely isolated by each method. DC purified EV size and diversity (27 nm and above) were similar to EV size predictions made by the centrifugation calculator [49]. Centrifugation calculators further predict that 38% of free protein will be purified in solution using DC purification methods. Furthermore, DC of EVs is more successful when sedimentation coefficients of the particles to be distinguished differ significantly [49]. For example, different centrifugation forces are required to obtain optimal EV to protein yield for different cell lines [50]. Thus, the difficulty of DC based standardisation across laboratory settings and sample types is becoming clearer.
During the current analysis, EVs were observed to aggregate together and contaminants co-purified were visible utilising both purification methods. However, Nordin et al. [2] demonstrated that DC purified EVs were fused, disrupted (62%) and aggregated when compared to SEC purified EVs yet, in the current investigation, EVs from either purification method were not damaged. It has been noted that forced filtration of samples can cause EV rupture and deformity, so sample preparation and SEC methods are advised to be performed using gravity only methods [1,6,11]. Aggregates are more likely to occur when high concentrations of EVs are present in a small volume with non-vesicular material; circumstances which are more likely within DC purification strategies rather than SEC purification of EVs [8]. However, EV aggregates may be overestimated by TEM and AFM methodology whereby surfaces capture EVs. Furthermore, vesicle morphology can be influenced by TEM and AFM methodology processes [6,51,52]. Both methods exhibited EV loss as demonstrated by the presence of EVs in the flow through (ESP), confirming that both purification methods are selective upon EV isolation.
When analysing the protein composition of purified EVs, it was apparent that SEC methods purified a greater protein yield within the EV samples (whole lysed, soluble protein fraction and insoluble protein fraction) when compared to DC preparations. These findings compare well with research identifying a greater number of particles to protein yield via SEC methods compared to DC methods and other commercial EV purification systems [2,14,15,53,54]. SEC has further been found to remove contaminating proteins from difficult sample types compared to DC such as albumin, cholesterol and apolipoprotein AI from blood plasma EVs and lipopolysaccharide binding protein from prokaryote EVs [14,55,56]. Conversely, Mol et al., [57] observed no significant differences in EV protein or particle yield between both purification methods. There is an inverse relationship between EV purity and the protein yield, where the lower the ratio the more impure the sample [1,10]. Therefore, sample purity may have consequent effects on experimental investigations [10]. Similar protein levels were observed between whole lysed EV and ESP samples purified by DC, yet the protein concentration of ESP was nearly threefold that of the whole lysed EV samples during SEC purification. This suggests again that SEC purification methods have an increased EV purity to protein yield ratio than DC based methods. This significant variation in ESP protein yield is most likely a direct result from sample wash steps involved with DC purification of EVs. Therefore, multiple wash steps are likely to increase EV loss and, as Webber and Clayton [10] acknowledged, are ineffective at removing protein contaminants, where there was only a 2-fold increase in particle to protein yield, compared to the crude pellet. Furthermore, there is the possibility that protein will sediment along with EVs during DC rather than be separated from EVs. Thus, in the current work SEC EV purification delivered a higher yield EV preparation in addition to higher yields of EV depleted ESP.
Whole lysed EV and ESP samples displayed different protein profiles as observed by one dimensional SDS-PAGE electrophoresis, suggesting that both EV purification methods isolate EVs from remaining F. hepatica secretions or ESP to varying degrees of purity. Following mass spectrometry analysis, we identified a greater total number of proteins associated with DC purified EVs than SEC purified EVs. However, SEC purified EV samples showed a greater diversity of EV functions in biological process, molecular function and cellular component gene ontology categories than DC purified EVs. The most common gene ontology terms identified in biological process, molecular function and cellular component were present in EVs from both purification methods. A comparable result was found investigating cell culture EVs, where fewer proteins (388 proteins) were identified in SEC purified EVs than in DC purified EVs (421 proteins), although fewer proteins (147 proteins) overlapped between SEC and DC EVs compared to the current study [53]. This result further suggests that DC purification alters EV protein composition to a greater extent than SEC purification, possibly isolating different EV subpopulations, as well as purifying greater amounts of contaminant. Thus, supporting SEC purification methods for an increased EV purity to protein yield ratio than DC based methods.
The most abundant proteins found in F. hepatica EVs, per rank, showed DC consistently had a greater amount of unique peptides hit per protein than SEC. Cwiklinski et al. [24] similarly analysed F. hepatica EV proteome where all 180 proteins identified were also found in the current investigation. However, the top unique peptide hits (where soluble and insoluble fraction data was combined) from Cwiklinski et al. [24] were not comparable to the current experimental results. Cwiklinski et al. [24] did not investigate a biological representation of the whole F. hepatica EV proteome, but instead EVs isolated from 120,000 x g ultracentrifugation spin only. In addition, a different mass spectrometry methodology and analysis was performed likely explaining DC EV differences. The top 50 proteins observed in EVs purified using SEC were also present in the DC EV purified proteome. However, three proteins were identified in the top 50 proteins of DC purified EVs, but were not present in the full SEC isolated EV proteome. Common EV markers were present in the top 50 EV proteins using both purification methods although tetraspanin was identified in the top 50 EV protein hits using SEC purification only and gelosin and phosphoglycerate kinase EV markers were observed in the top 50 protein hit using DC purification only. This may indicate that SEC and DC purification methods also purify different EV sub-populations, or that EV sub-populations are purified in differing quantities within each method.
Further investigation discovered that there was a 2.6 fold greater number of protein identifications in DC purified EVs only, compared to SEC purified proteins. Of these DC EV unique proteins many have been identified previously in ESP and tegument proteomic studies. With pathogens, the separation of EVs from additional pathogen components such as the platyhelminth tegument is vital to fully assess the role of the pathogen EV. The recognised increase of tegumental proteins found within the DC purified EVs further supports that SEC purified EVs have a greater EV purity to protein yield than DC purified EVs. Furthermore, current international EV purification methods report that chromatographic methods, including SEC, produce less contamination by non-vesicular proteins and macromolecule structures than ultracentrifugation, thus is likely to account for the reduction of tegumental proteins in SEC EVs [58].
Akin to the work of Benedikter et al. [53] gene enrichment analysis demonstrated comparable enriched gene ontology terms for both purification methods in biological process, molecular function and cellular component categories against the F. hepatica genome. Interestingly, only GO terms for translation biological process ribosome cellular component and structural constituent of the ribosome were significantly less enriched in DC purified EVs compared to SEC purified EVs, while DNA-templated regulation of transcription was enriched significantly in DC purified EVs compared to SEC purified EVs. This further suggests that DC and SEC purification methods isolate differing EV populations with altered functions. Cwiklinski et al. [24] previously discovered that EV sub-populations secreted from F. hepatica, contained different relative peptidase activities. A range of other studies have also identified that nucleic acids [59–64] and proteins [60,64–67] are selectively packed into EV subtypes. It has further been found that purification methodologies have differential affinity for protein EV markers and by extension for different EV sub-population using human urine samples [68]. This research further supports the current work in that DC and SEC purification methods isolated EVs of different function and different levels of purity.
An increased EV purity to protein yield ratio demonstrated in SEC purification over DC based methods was further reinforced by Western blotting revealing higher recognition by Anti-FhFABP V, Anti-FhGST-S1 and Anti-FhCat-L1 in the EV soluble fraction using DC purification rather than SEC. Only protein fragments from FABP III were identified within the mass spectrometry peptide analysis from the FABP protein family. Importantly, comparable levels of recognition by Anti-FhFABP V and Anti-FhGST-S1 were seen in DC and SEC ESP depleted of EVs, yet a greater recognition from Anti-FhCat-L1 was observed in SEC EV depleted ESP demonstrating increased separation of ESP from EVs in SEC methods. Baranyai et al. [55], also using Western blotting, demonstrated that SEC produced a greater EV purity to protein yield ratio, as higher albumin concentrations were seen in DC methods than SEC methods in mouse blood plasma. Correspondingly, vesicle markers have been found at greater concentrations in EV samples using SEC methods than DC [2]. Interestingly, there were a greater variety of cathepsin L protease isoforms identified in EVs purified by SEC, rather than EVs purified by DC. Of these proteins, there was a greater number of cathepsin L1A and cathepsin L1D, but a lower abundance of cathepsin L5 in SEC purified EVs, compared to DC purified EVs. Cathepsin L proteases have functional roles within immune evasion, nutrition and migration. In particular, cathepsin L1 and L2 proteases degrade host haemoglobin, immunoglobulin and interstitial matrix proteins such as fibronectin, laminin, and native collagen [69,70]. Different cathepsin L protease clades, have distinct substrate abilities, indicating that they have different roles in parasite biology. In the case of cathepsin L5, these roles are yet to still be determined [47,71]. In previous proteomic studies investigating the surface and membrane fractions of F. hepatica EVs [24], members of the CL1 clade (1A,1B and 1D) have all been identified to be membrane associated, yet members of CL5 have not. Therefore, the SEC approach identified more membrane associated cathepsin L in comparison to DC, identifying increased soluble cathepsin L derived from the EV contents or from contamination from ESP.
Interestingly, Anti-FhGST-S1 recognition was strongly observed in the insoluble EV fraction and a higher protein abundance of FhGST-S1 was suggested using SEC purification. Therefore, FhGST-S1 could potentially act as a novel marker for F. hepatica EVs given its identification in the insoluble EV fraction. The function of FhGST-S1 in F. hepatica includes prostaglandin synthesis which establishes host infection, egg development and embryogenesis, host immune modulation and potential triclabendazole drug response in phase III sequestration based detoxification [40]. Recent studies have found that parasite EVs can communicate with host cells therefore, the function FhGST-S1 in F. hepatica EVs could involve host infection establishment and host immune modulation [23,28,29,72–82]. However, further investigation upon this proposal is required to determine the role and location of FhGST-S1 in EVs. This finding indicates the importance of EV purification methodology upon accurately identifying the abundance of pathogen markers and vaccine candidates in EVs.
To claim that SEC should be used as a gold standard method for isolating EVs from parasite culture media, additional isolation methodologies (density gradient, sucrose cushion, precipitation methods and immunoaffinity isolation) will need to be compared with SEC methodology. It is acknowledged that within any purification method, many variables, such as the number of wash steps, will affect the purity and functionality of purified EVs. Given that all EV helminth studies to date have used DC purification for downstream analysis, the present study findings suggest a change in approach may be required with SEC providing a promising purification method for isolating EVs from in vitro helminth cultures for downstream functional analysis in terms of minimising non-EV contaminants.
There is a strong demand for an established uniform protocol for EV purification. However, instead of finding the ultimate ‘gold standard’ for EV purification, from this investigation it appears to be of more importance that EV purity is standardised, rather than the purification protocol. Possibly, rather than the helminth EV field aiming to isolate EVs using a ‘gold standard’ method, there should be a ‘gold standard’ of purity assessment of EV samples used for experimentation, dependent upon the sample type (e.g. plasma, urine, cell culture media and parasite culture media). EV specific markers are also likely to be important to identify EV sample purity. Specific EV markers would add value to the assessment of EV sample purity and thus could be species specific. This is especially pertinent given that current EV markers are based on mammalian work [83]. In the current study, differential abundance of protein families (CAT L clades and egg based proteins) were noted during proteomic analysis that could be utilised for Fasciola specific EV purity markers. This would improve the standardisation of protocols and the comparability of results from scientific research.
Thorough proteomic investigation on EV protein composition from EVs purified by DC and SEC methodology identified that SEC purified EVs contained proteins with more functional properties than DC purified EVs. However, in order to confirm increased EV functionality and the improved functional benefit of using SEC for EV purification compared to other methodologies, additional functional studies comparing the isolation methodologies by independent host-parasite interaction groups would be required. For example, SEC and DC purified EVs could be cultured with host cells in vitro and host cell transcriptome data could be compared between the two isolation conditions, following the methodology undertaken using DC purified O. viverrini EVs [76]. However, as the findings from the current study support that SEC purified EVs contained proteins with more functional properties than DC purified EVs, speculation upon the validity of other EV functional studies which have purified EVs from parasite culture media using DC methodology is necessary. Therefore, more comparative research is required to understand the influence of EV purification methodologies upon functional studies.
In summary, the current study has challenged whether the most accepted EV purification technique in helminth research is optimal for functional studies, in comparison to SEC methods. Our discoveries using F. hepatica as a pathogen case study, suggest that SEC purification has a higher EV purity to protein yield ratio than DC purified EVs evidenced by reduced contamination from ESP and tegumental components, whilst still maintaining EV morphological characteristics. Furthermore, gene ontology terminology proposed that DC and SEC purification methods isolate differential EV sub-populations. Given the demonstrated variation in purification methodologies and the importance of understanding the function of EVs for potential downstream studies and application development, the authors suggest that for EV functional assays the purification methodology used should be of importance when designing experiments. This research has highlighted SEC EV isolation as a potential key methodology for functional EV research.
|
10.1371/journal.pgen.1004986 | Sequence Features and Transcriptional Stalling within Centromere DNA Promote Establishment of CENP-A Chromatin | Centromere sequences are not conserved between species, and there is compelling evidence for epigenetic regulation of centromere identity, with location being dictated by the presence of chromatin containing the histone H3 variant CENP-A. Paradoxically, in most organisms CENP-A chromatin generally occurs on particular sequences. To investigate the contribution of primary DNA sequence to establishment of CENP-A chromatin in vivo, we utilised the fission yeast Schizosaccharomyces pombe. CENP-ACnp1 chromatin is normally assembled on ∼10 kb of central domain DNA within these regional centromeres. We demonstrate that overproduction of S. pombe CENP-ACnp1 bypasses the usual requirement for adjacent heterochromatin in establishing CENP-ACnp1 chromatin, and show that central domain DNA is a preferred substrate for de novo establishment of CENP-ACnp1 chromatin. When multimerised, a 2 kb sub-region can establish CENP-ACnp1 chromatin and form functional centromeres. Randomization of the 2 kb sequence to generate a sequence that maintains AT content and predicted nucleosome positioning is unable to establish CENP-ACnp1 chromatin. These analyses indicate that central domain DNA from fission yeast centromeres contains specific information that promotes CENP-ACnp1 incorporation into chromatin. Numerous transcriptional start sites were detected on the forward and reverse strands within the functional 2 kb sub-region and active promoters were identified. RNAPII is enriched on central domain DNA in wild-type cells, but only low levels of transcripts are detected, consistent with RNAPII stalling during transcription of centromeric DNA. Cells lacking factors involved in restarting transcription—TFIIS and Ubp3—assemble CENP-ACnp1 on central domain DNA when CENP-ACnp1 is at wild-type levels, suggesting that persistent stalling of RNAPII on centromere DNA triggers chromatin remodelling events that deposit CENP-ACnp1. Thus, sequence-encoded features of centromeric DNA create an environment of pervasive low quality RNAPII transcription that is an important determinant of CENP-ACnp1 assembly. These observations emphasise roles for both genetic and epigenetic processes in centromere establishment.
| The kinetochore directs the separation of chromosomes and is assembled at a special region of the chromosome—the centromere. DNA is wrapped around particles called nucleosomes, which contain histone proteins. The nucleosomes at centromeres are specialized, and contain the centromere-specific histone CENP-A. CENP-A nucleosomes form the platform upon which the kinetochore is built. Thus, CENP-A and centromere function go hand-in-hand. How the cell ensures that CENP-A is deposited at centromeres and not elsewhere is not well understood. We investigated the role that DNA sequence plays in defining centromere function in fission yeast. Our observations suggest that it is not the DNA sequence per se that is important for attracting CENP-A, but rather, the particular environment that the sequence creates. During transcription of centromeric DNA, RNA polymerase (RNAPII) appears to get stuck or stalled. Particular proteins—such as TFIIS and Ubp3—are known to help restart RNAPII so it can continue transcribing. We found that when cells lack Ubp3 or TFIIS, CENP-A becomes deposited on centromere sequences. We propose that persistent stalling of RNAPII on centromere DNA attracts factors that help deposit CENP-A. This study highlights the influence of DNA sequence in creating an attractive environment for CENP-A assembly.
| Centromeres are the chromosomal sites upon which kinetochores are assembled to ensure accurate segregation of sister chromatids into daughter cells. Most kinetochores are built upon a specialized type of chromatin in which canonical histone H3 is replaced by the histone variant CENP-A. Although the centromere-kinetochore complex performs conserved essential functions, and kinetochore proteins are generally conserved [1], centromeric DNA is not conserved, even between related species, and a huge variety of centromere sequences and structures exist [2–5]. The point centromeres of budding yeast consist of 125 bp of DNA and utilize an essential centromere-specific DNA binding protein [6]. At the other extreme, the nematode, Caenorhabditis elegans, has holocentric centromeres, in which kinetochore proteins assemble at multiple loci along each chromosome arm [7,8]. The majority of centromeres studied to date are regional. Centromeres in various plant and animal species are composed of arrays of different types of satellite, repetitive sequences and transposable elements, for instance, human centromeres encompass several megabases of tandem repetitive arrays of alpha-satellite sequence [9–11]. Fission yeast centromeres represent another type of regional centromere, in which a unique central core of 4–7 kb is flanked by inverted repeat elements and blocks of relatively large repeat units, resulting in centromeres of 40–120 kb [12]. Even the centromeres of different chromosomes in individual species are not necessarily homologous; each Candida albicans centromere has a unique central core, whilst chicken and potato each utilize both repeat-rich and unique sequence centromeres [13–15]. Thus, functional centromeres are assembled on diverse types of sequences in different organisms and it remains unknown if there is a universal fundamental property that defines centromeric sequences.
Abundant evidence indicates that centromeres are epigenetically regulated [16]. Although rare, neocentromeres have been observed in many species, forming on DNA sequences that do not normally possess centromere function and share no sequence homology with normal centromeres [17]. The best-characterized example in human is the neocentromere in 10q25 on the long arm of chromosome 10 that arose upon deletion of the centromere and loss of the entire alpha satellite array [18]. In S. pombe, neocentromeres form in close proximity to telomeres following the engineered deletion of a centromere [19]. Conversely, centromeres can be inactivated on dicentric human chromosomes despite the continued presence of alpha-satellite sequence at both centromeric loci [20]. In S. pombe one centromere on a dicentric chromosome can be inactivated by mechanisms such as heterochromatinisation or formation of a domain of histone hypoacetylation [21]. These and numerous other examples demonstrate that centromeric sequences are neither necessary nor sufficient for kinetochore assembly.
The histone H3 variant, CENP-A acts as the epigenetic mark that specifies centromere identity [22–24]. CENP-A is found only at active centromeres, including neocentromeres, and is absent at inactivated centromeres. The forced recruitment of CENP-A either by directly tethering CENP-A or its chaperone (HJURP) to a non-centromeric locus leads to the accumulation of CENP-A and kinetochore proteins at that location [24,25]. It is thought that continued deposition of CENP-A at centromere regions through cell and organism generations involves a self-propagation mechanism in which CENP-A chromatin, or features of the kinetochore which is assembled upon it, are recognized and attract additional CENP-A [26,27].
In most organisms there is no obligate coupling of sequence and CENP-A assembly, yet kinetochores are normally assembled upon particular centromeric sequences in any given species [4]. This suggests that centromeric sequences possess underlying properties that promote CENP-A incorporation. Alternatively, the preponderance of particular sequences at centromeres could be driven by properties of CENP-A chromatin or kinetochores themselves [28]. However, centromeric DNA allows the de novo assembly of functional centromeres following its introduction into cells in many organisms. Alpha satellite arrays are able to direct the de novo assembly of centromeres when introduced into certain cell lines as naked DNA [29,30]. De novo assembly of centromeres also occurs when centromeric DNA from S. pombe is introduced into cells. However, de novo establishment does not seem to be a universal property: despite promiscuous neocentromere formation in C. albicans, transformation with bone fide centromeric sequences does not result in kinetochore assembly[13]. At the other extreme, many sequences introduced into the holocentric organism C. elegans appear able to assemble CENP-A chromatin [31,32]. Thus, the relationship between centromeric sequence and the establishment and maintenance of CENP-A chromatin is enigmatic.
Transcription has received a lot of attention as a possible contributor to assembly of CENP-A chromatin. Transcripts emanating from centromeric regions have been detected in many organisms, including maize, human, rice, budding yeast, fission yeast and tammar wallaby [33–38]. Interfering with the chromatin status or transcriptional properties of centromeric repeats affects maintenance of CENP-A chromatin and segregation function on human artificial chromosomes (HACs) [39,40]. RNA Polymerase II (RNAPII) has been detected at mitotic mammalian centromeres where it may influence centromere function [38]. In fission yeast, transient H2B ubiquitylation may loosen centromeric chromatin to promote transcription and CENP-ACnp1 incorporation and defective reassembly of H3 chromatin behind elongating RNAPII aids CENP-ACnp1 incorporation [41,42]. However, although there are numerous tantalising hints that transcriptional activity contributes to centromere function or identity, much remains to be understood [33,38,43,44].
Here we investigate the contribution of DNA sequence to the establishment of CENP-A chromatin in fission yeast, an organism in which epigenetic mechanisms clearly influence centromere identity. Normally proximal heterochromatin is required to facilitate establishment of CENP-ACnp1 chromatin on centromere central domain sequences [45,46]. We show that this requirement can be bypassed by overexpression of CENP-ACnp1 and that central domain DNA is a preferred substrate for establishment of CENP-ACnp1 chromatin. We find that there is functional redundancy within the central domain but that sub-regions are non-equivalent in their ability to establish CENP-ACnp1 chromatin. Analysis of a 2 kb region capable of directing CENP-ACnp1 assembly indicates that it contains numerous transcriptional start sites, along with promoter elements, and that relatively high levels of RNAPII are recruited, despite low levels of transcripts produced, consistent with the presence of stalled RNAPII. Our observations suggest that redundant sequence features in the centromere central domain create a unique transcriptional environment that is permissive for CENP-ACnp1 establishment. Consistent with this, defective transcriptional elongation where stalled RNAPII is increased promotes the establishment of CENP-ACnp1 chromatin.
In wild-type fission yeast cells, de novo CENP-ACnp1 chromatin establishment on circular plasmid-based minichromosomes requires an outer repeat or tethered Clr4 histone H3K9 methyltransferase to form a block of heterochromatin in close proximity to central domain DNA from centromeres [45,46]. CENP-ACnp1 can also be deposited at other non-centromeric locations in the genome when it is overexpressed, however the level incorporated at these ectopic sites is much lower than that detected at natural centromeres [41,47]. To determine whether central domain DNA is a preferential substrate for the establishment of CENP-ACnp1 chromatin, plasmid pMcc2 bearing 8.5 kb of central domain from cen2 (imr2-cc2-imr2) sequence, but no heterochromatic outer repeat sequences, was transformed into cells expressing additional GFP-CENP-ACnp1 (Fig. 1A).
All strains used have 6 kb of cen2 central domain DNA replaced with 5.5 kb of cen1 central domain DNA (cc2Δ::cc1—Fig. 1A, S1 Fig.) so that only 2.5 kb of normal cen2 central domain DNA remains at this modified cen2 (imr2L, regions J, K, R; Fig. 1A). The resulting deletion of fragments L-Q from the cen2 central domain allows detailed and specific analysis of 6 kb of central domain DNA when borne by plasmid-based minichromosomes. Quantitative chromatin immunoprecipitation assays (qChIP) shows that CENP-ACnp1 chromatin does not assemble on regions L, M N, O or P when a plasmid (pMcc2) containing the 8.5 kb cc2 sequence, but lacking heterochromatin, was transformed into wild-type cells [45]. However, when pMcc2 was transformed into cells over-expressing CENP-ACnp1 (hi-CENP-ACnp1; ∼15 fold more than wild-type cells [41]), CENP-ACnp1 and the kinetochore proteins CENP-CCnp3 and CENP-KSim4 were easily detected over the central domain of pMcc2 by qChIP (Fig. 1B-E). Importantly, these centromeric proteins were enriched on centromeric DNA but not on the plasmid backbone, indicating that CENP-ACnp1 chromatin assembles specifically on central domain DNA from centromeres (Fig. 1). The relative level of enrichment of CENP-ACnp1 and the other kinetochore proteins on different parts of pMcc2 suggests all proteins are distributed uniformly across this plasmid-borne central domain (Fig. 1B-D). Furthermore, the levels of histone H3 associated with the L-P regions of pMcc2 were reduced in cells expressing additional CENP-ACnp1 compared to control cells (Fig. 1E). We conclude that H3 chromatin is normally assembled on central domain DNA on pMcc2 in wild-type cells but CENP-ACnp1 chromatin assembles instead when pMcc2 is placed in hi-CENP-ACnp1 cells.
To determine whether CENP-ACnp1 can become established on plasmids that are already assembled in chromatin, the pMcc2 plasmid was transformed into cells expressing wt-CENP-ACnp1 levels and subsequently crossed with hi-CENP-ACnp1 cells. qChIP analyses indicate that CENP-ACnp1 is initially absent from pMcc2 in the wt-CENP-ACnp1 parental strain and then becomes assembled in CENP-ACnp1 chromatin when transferred into the hi-CENP-ACnp1 environment, indicating that plasmid-borne cc2 initially assembled in normal (H3) chromatin can be converted to CENP-ACnp1 chromatin (Fig. 2A).
In addition, a copy of cc2 (8.5 kb) was inserted on the arm of the 530 kb Ch16 linear minichromosome which carries a complete cen3 [48] (Ch16-cc2; Fig. 2B). When the expression of additional GFP-CENP-ACnp1 was repressed (0h+T), no CENP-ACnp1 was detected on cc2. However, when GFP-CENP-ACnp1 was induced (48h-T) both CENP-ACnp1 and CENP-CCnp3 were detected on cc2 (Fig. 2B). Thus, cc2 borne on a linear minichromosome can be converted from a pre-chromatinised state to a CENP-ACnp1 state. Moreover, colony colour assays indicate that hi-CENP-ACnp1 expression induces increased loss of Ch16-cc2, which is consistent with a second functional kinetochore being formed at cc2 on Ch16 (Fig. 2C). Thus, Ch16-cc2 behaves as an inducible dicentric chromosome controlled by CENP-ACnp1 levels.
It is possible that high levels of CENP-ACnp1 are continuously required to maintain CENP-ACnp1 on pMcc2, or alternatively, once established, CENP-ACnp1 and kinetochore proteins may persist even when CENP-ACnp1 is returned to wild-type levels (wt-CENP-ACnp1). To investigate the maintenance of CENP-ACnp1 chromatin, pMcc2 was first transformed into hi-CENP-ACnp1 cells to allow the assembly of centromeric chromatin and subsequently these pMcc2-containing cells were crossed with wt-CENP-ACnp1 cells to transfer the pMcc2 plasmid into cells expressing wild-type CENP-ACnp1 levels. ChIP analyses show that CENP-ACnp1 persisted on the pMcc2 in this wild-type background (Fig. 2D). Western analysis of extracts from Parental, F1 and F2 cells confirmed that GFP-CENP-ACnp1 was lost in F1 and F2 cells (Fig. 2D). Thus, CENP-ACnp1 chromatin behaves as a true epigenetic entity in that once established it carries its own efficient propagation mechanism, persisting even though the original stimulus has been removed. More remarkably, this CENP-ACnp1 chromatin is maintained through 2 rounds of meiosis and at least 50 mitotic divisions. Thus central domain sequences are particularly receptive to the establishment and maintenance of CENP-ACnp1 chromatin.
To determine if specific regions from the central domain of cen2 are required to establish CENP-ACnp1 chromatin, plasmids bearing different sub-fragments from cc2 were transformed into wt-CENP-ACnp1 or hi-CENP-ACnp1 cells (Fig. 3). We used an unbiased approach to divide the 8.5 kb cc2 into 1 kb regions (J-R). Deletion of 1 kb from the centre of cc2 (N) does not affect CENP-ACnp1 establishment (pΔN; Fig. 3A, compare with pMcc2, Fig. 1B). Notably, CENP-ACnp1 incorporation on a plasmid carrying identical centromeric DNA as pΔN but with the right half (O-R) inverted relative to the native sequence, was less efficient (pΔN-rev; Fig. 3B). Thus, the relative orientation of central domain sequences within cc2 influences the degree of CENP-ACnp1 deposition; directionality or the juxtaposition of certain sequences may be important for promoting CENP-ACnp1 incorporation. The central CENP-ACnp1 domain at endogenous fission yeast centromeres is composed of inverted imr repeats that flank the central core. ChIP analyses demonstrated that, in a plasmid-based establishment assay, the imr repeats are dispensable for de novo CENP-ACnp1 incorporation on the remaining central domain sequences (pΔimr; Fig. 3C).
Deletion of additional regions (LMN) of cc2 markedly decreased the efficiency of CENP-ACnp1 incorporation relative to pMcc2 and pΔN (Fig. 3D; compare Fig. 1B, Fig. 3A), suggesting that either the LM region is critical for promoting CENP-ACnp1 incorporation or that the overall reduced centromeric DNA length diminishes CENP-ACnp1 deposition (Fig. 3D). However, further investigation using plasmids bearing smaller cc2 fragments suggests that the specific sequences present have a more significant influence on CENP-ACnp1 deposition than the overall length of cc2 DNA present (Fig. 3E,F). For example, pΔJM and pΔNR differ by only 500 bp, however, pΔJM incorporated substantially more CENP-ACnp1 than pΔNR (Fig. 3F). We conclude that specific sequences from the central domain of fission yeast centromeres, combined with their overall length, promote the efficient de novo assembly of CENP-ACnp1 chromatin.
It is possible that shorter fragments of centromere DNA from within the central domain can actively promote CENP-ACnp1 assembly but that because longer total lengths are required to stabilise incorporated CENP-ACnp1 the activity of shorter fragments cannot be detected. To address this possibility we selected two distinct sequences from the central domain of cen2 for analyses. The 2 kb OP region was present on all the pMcc2 derivatives with which we detected significant CENP-ACnp1 incorporation following transformation into hi-CENP-ACnp1 cells (Fig. 3). In addition, ChIP-seq analysis indicates particularly high CENP-ACnp1 nucleosome occupancy within OP at endogenous cen2 [49 and Fig. 4A]. In contrast, the 2 kb LM region appears to be dispensable for de novo CENP-ACnp1 assembly on pMcc2 derived plasmids and exhibits low CENP-ACnp1 nucleosome occupancy (Fig. 3, Fig. 4A).
Initial tests showed that neither OP (p1xOP) nor LM (p1xLM) sequences alone were capable of inducing significant de novo CENP-ACnp1 incorporation when introduced into hi-CENP-ACnp1 cells (Fig. 4B). This finding is consistent with a minimal length of central domain DNA being required for stable CENP-ACnp1 chromatin assembly and retention. To satisfy this apparent length requirement, the OP and the LM fragments were multimerised as tandem repeats to create 3xOP and 3xLM (p3xOP, p3xLM; Fig. 4C). Remarkably, when transformed into hi-CENP-ACnp1 cells no CENP-ACnp1 was detectable on p3xOP whereas p3xLM allowed a reasonable level of CENP-ACnp1 incorporation (Fig. 4C). This suggests that in isolation the OP region is unable to promote CENP-ACnp1 deposition even though in the context of an entire central domain it normally accepts CENP-ACnp1 and ends up with high CENP-ACnp1 nucleosome occupancy (Fig. 4A). We note that the removal of the LM region from the central domain of pMcc2 derived plasmids greatly reduced the level of CENP-ACnp1 incorporated on OP (compare pΔN Fig. 3A with pΔLMN Fig. 3D). Thus, in contrast to OP, the LM region appears to have the ability to induce CENP-ACnp1 deposition. To directly test this possibility, a single copy of LM was placed adjacent to two tandem copies of OP (pLM-2xOP) and transformed into hi-CENP-ACnp1 cells. High levels of CENP-ACnp1 were detected on both the LM and OP regions of pLM-2xOP (Fig. 4D), thus the LM region has an innate ability to stimulate CENP-ACnp1 deposition on the OP region. A different arrangement of the same sequences (pOPLMOP) also attracted CENP-ACnp1 in hi-CENP-ACnp1 cells (S2 Fig.). These analyses indicate that the 2 kb LM sequence contains all the features that are required to promote and accept CENP-ACnp1 assembly, and thus LM defines a 2 kb region of S. pombe centromeric sequence that allows the de novo assembly of CENP-ACnp1 chromatin.
Plasmids bearing an entire central core domain flanked by outer heterochromatin repeats assemble functional centromeres when transformed into wild-type cells [45,50]. To determine if the 2 kb LM region imparts centromere function, a plasmid carrying the 3xLM tandem repeat adjacent to a 5 kb outer repeat heterochromatin forming element (pH-3xLM) was transformed into wild-type cells expressing CENP-ACnp1 at normal levels (Fig. 5A). The establishment of functional centromeres in the resulting transformants was monitored by an ade6-based colony colour sectoring assay [51]. Minichromosomes carrying full-length cc2 and 5 kb of outer repeat heterochromatin were able to establish functional centromeres upon transformation (Fig. 5B and S2 Fig.). pH-3xLM and pH-LM-2xOP transformants also established functional centromeres, but at lower frequency than pH-cc2 (Fig. 5B and S2 Fig.). Differences in the ability of various constructs to form functional centromeres may reflect the particular configuration of sequences in individual minichromosomes. In contrast, pH-3xOP (3xOP flanked by heterochromatin) was unable to establish functional centromeres. Thus the LM sequence in a 3x tandem array, flanked by heterochromatin, is sufficient to form functional centromeres. ChIP analyses confirmed that kinetochores were assembled on pH-3xLM since CENP-ACnp1 and the kinetochore proteins CENP-CCnp3 and CENP-KSim4 were enriched over the LM sequences at levels comparable to endogenous centromeres (Fig. 5C). We conclude that the LM sequence within pH-3xLM not only promotes incorporation of CENP-ACnp1 into chromatin but also supports the assembly of a functional centromere.
Nucleosome occupancy is known to be influenced by a combination of DNA sequence and the action of chromatin remodelers [52]. Primary DNA sequence itself influences nucleosome occupancy since DNA sequences with a high GC content and periodic dinucleotide patterns, that are devoid of poly(dA:dT) sequences, are strongly favored for nucleosome occupancy because of biophysical constraints that allow such sequences to wrap more easily around nucleosomes. These constraints have led to the development of algorithms that predict the probability of nucleosome occupancy [53,54]. In common with centromeres of many organisms, fission yeast centromeric DNA is AT-rich with a higher frequency of poly(dA:dT) tracts. It is therefore possible that H3 nucleosomes have a lower affinity for such sequences whereas CENP-A nucleosomes may be unperturbed by such AT rich DNA. To examine the underlying sequence specificity within centromeric DNA that favours the deposition of CENP-ACnp1 nucleosomes, the sequence of LM DNA was altered by randomisation using a 5 bp sliding window throughout the entire 2 kb element. This generated a synthetic LM sequence (SynR-LM) that is 62.6% identical to the wild-type LM sequence, retaining the same AT content and dinucleotide periodicity, and thus the same predicted nucleosome occupancy as the wild-type LM element (Fig. 6A) [55].
Synthesised SynR-LM assembled as a 3xSynR-LM tandem array was placed in the same plasmid backbone as p3xLM to generate p3xSynR-LM. p3xSynR-LM was transformed into wt-CENP-ACnp1 and hi-CENP-ACnp1 cells. In contrast to p3xLM, CENP-ACnp1 was not detectable on the shuffled LM sequence of pSynR-LM (Fig. 6B, compare with Fig. 4B). These analyses demonstrate that preservation of nucleotide composition (AT-content, dinucleotide periodicity) and predicted nucleosome occupancy within an altered centromeric DNA is not sufficient to allow CENP-ACnp1 deposition. The fact that the natural 2 kb LM sequence is active whereas the artificial SynR-LM is inactive reveals that the primary sequence of wild-type centromeric LM DNA encodes properties that somehow allow its recognition in vivo and consequent de novo assembly of CENP-ACnp1 chromatin.
Upon transformation into cells innate features within 3xLM sequence must allow it to be either immediately assembled in CENP-ACnp1 chromatin, or, initially assembled in H3 chromatin with subsequent remodelling that exchanges canonical H3 for CENP-ACnp1. The process of transcription is obviously accompanied by chromatin remodelling and non-coding transcripts synthesised from within the central CENP-ACnp1 domains of fission yeast centromeres have been detected [37,42]. The transcription of central domain DNA might influence the assembly of CENP-ACnp1 chromatin. In cells expressing CENP-ACnp1 at wild-type levels, plasmid-borne central domain sequences are assembled in H3 rather than CENP-ACnp1 chromatin (Fig. 1E). Higher levels of RNAPII are detected on plasmid-borne central domain sequences (pMcc2) introduced into wild-type cells than when cc2 is assembled in CENP-ACnp1 chromatin on pMcc2 or at endogenous centromeres (Fig. 7A, S3 Fig.). Although relatively high levels of RNAPII associate with the pMcc2 central domain when assembled as H3 chromatin in wild-type cells (10–30% of levels at act1+) (Fig. 1E, Fig. 7A), the level of transcripts emanating from the central domain is very low (<0.1% of act1+), even when analysed in exosome defective cells (dis3–54; Fig. 7B). Thus, although ample RNAPII is recruited to the central domain of pMcc2 few transcripts are generated, suggesting that transcriptional stalling occurs.
To map transcriptional start sites (TSSs) within the LM and OP regions, 5’ RACE was performed on RNA extracted from dis3–54 exosome mutant cells harbouring p3xLM or p3xOP (Fig. 7C, S4 Fig.). Many TSSs were identified within LM and OP, suggesting that these regions contain several promoters (Fig. 7C). 200 bp regions from both LM and OP were tested for their ability to drive production of β-galactosidase when placed upstream of a lacZ reporter in fission yeast and as shown in Fig. 7D, the regions displayed promoter activity. Mutated or inverted versions of promoter region M2 did not promote transcription of LacZ (S4 Fig.). Whilst most regions of LM and OP exhibit promoter activity that is lower than that of nmt81 control promoter, it is notable that region-O1 and region-P2 from OP have equivalent and 10-fold higher activity, respectively (Fig. 7D). It is possible that the higher promoter activity possessed by some regions of OP may affect its ability to establish CENP-ACnp1. We surmise that the central domain from cen2 is peppered with promoters that can drive the production of transcripts on both strands. Their relative arrangement along with the strength and pattern of transcription may affect CENP-ACnp1 incorporation.
The progression of RNAPII is impeded by obstacles such as nucleosomes, DNA damage, bound proteins and by sequences that are intrinsically difficult to transcribe, causing transcriptional pausing, stalling or arrest [56]. RNAPII-associated proteins ease the passage of RNAPII through such impediments, contributing to the processivity of the polymerase [57]. TFIIS facilitates transcriptional elongation of stalled/backtracked RNAPII by stimulating cleavage of nascent transcripts [58–60]. Upon stalling an elongating RNAPII becomes mono- then poly-ubiquitylated on the largest Rpb1 subunit. A rescue pathway involving de-ubiquitylation by the ubiquitin hydrolase Ubp3 is deployed to restart stalled RNAPII [56,61].
Our analyses suggest that the central domain chromatin landscape contains numerous promoters on both strands and multiple TSSs. In addition, long poly(dA:dT) tracts are likely to be an intrinsically problematic sequence for RNAPII transcription and present a barrier to RNAPII elongation [62,63]. We reasoned that mutants that are defective in the response to transcriptional stalling might influence the ability of the central domain to become assembled in CENP-ACnp1 chromatin. To test this possibility, wild-type and TFIIS (tfs1Δ) mutant cells expressing hi-CENP-ACnp1 were transformed with pMcc2. Surprisingly, slightly increased levels of CENP-ACnp1 were detected on pMcc2 in the tfs1Δ mutant compared to wild-type cells, suggesting that loss of TFSII promotes CENP-ACnp1 deposition (S5 Fig.). Consistent with this, even when pMcc2 was transformed into tfs1Δ cells expressing wt-CENP-ACnp1 levels, CENP-ACnp1 was detected on the pMcc2 central domain (Fig. 8A). In order to determine whether the effect on CENP-ACnp1 establishment was specific to tfs1Δ or a general consequence of increased RNAPII stalling, we also investigated if loss of the ubiquitin hydrolase Ubp3, which normally rescues arrested RNAPII, affects CENP-ACnp1 deposition. Strikingly, CENP-ACnp1 was detected at high levels on central domain sequences in ubp3Δ cells transformed with pMcc2. CENP-ACnp1 was also detected on p3xLM, but not p3xOP in ubp3Δ (Fig. 8B, S6 Fig.). CENP-CCnp3 and CENP-KSim4 centromere proteins were also significantly enriched on pMcc2 in ubp3Δ cells (S8 Fig.). These effects were not due to increased abundance of CENP-ACnp1 in tfs1Δ or ubp3Δ cells as protein levels were similar to wild-type cells (S7 Fig.). In fact, a reduction in CENP-ACnp1 and CENP-CCnp3 levels was detected at endogenous centromeres in ubp3Δ, but not tfs1Δ cells (S8 Fig.). Tfs1 and Ubp3 were previously reported to modulate RNAi-independent heterochromatin assembly [64]. To test whether the effect on CENP-ACnp1 establishment in tfs1Δ or ubp3Δ cells could be due to spurious assembly of heterochromatin on pMcc2, H3K9me2 ChIP was performed. The level of H3K9me2 on pMcc2 in tfs1Δand ubp3Δ was similar to that on a negative control locus, act1+, and assembly of CENP-ACnp1 on pMcc2 in these mutants was not dependent on the H3K9-methyltransferase Clr4 (S9 Fig.). Thus, CENP-ACnp1 assembly on pMcc2 in the absence of TFIIS or Ubp3 does not result from induction by ectopic heterochromatin.
If lack of TFIIS or Ubp3 hinders transcriptional elongation, an increased level of RNAPII would be expected on affected chromatin templates. Elevated levels of Rpb1/RNAPII were detected on the central domain of pMcc2 in tfs1Δ (TFIIS) and ubp3Δ cells (Fig. 8C). In addition, increased levels of the elongation-specific Phospho-Ser2 form of RNAPII were observed on the central domain of pMcc2 in ubp3Δ cells, suggestive of failure to efficiently clear stalled RNAPII (Fig. 8D). Thus, two mutants, which perturb the progress of RNAPII elongation complexes in different ways, lead to deposition of CENP-ACnp1. These observations suggest that altering the transcriptional properties of the central domain chromatin through increased RNAPII stalling creates an environment that is permissive for establishment of CENP-ACnp1 chromatin in place of H3 chromatin.
It is thought that once established, CENP-A chromatin has the ability to be ‘self-propagating’, and through the recruitment of factors that are themselves involved in deposition of CENP-A, it ensures its own maintenance [16,17,23,24,26,65]. Epigenetic inheritance can be defined as the propagation of a state in the absence of the initial inducer of that state. In this study, the inducer—overexpression of CENP-ACnp1—causes an event that would not normally occur, the assembly of CENP-ACnp1 chromatin on episomal centromeric DNA (pMcc2). When CENP-ACnp1-assembled pMcc2 is crossed from hi-CENP-ACnp1 cells into wt-CENP-ACnp1 cells, CENP-ACnp1 is propagated in the absence of the initial inducer through many generations and through meiosis. These observations further strengthen the evidence that CENP-A behaves as a bona fide epigenetic entity [24].
It is clear that both epigenetic and genetic factors influence CENP-A assembly. We have investigated the role of DNA sequence in establishment of CENP-A chromatin in fission yeast, an organism where analysis is not confounded by repetitive arrays of short satellite sequences. CENP-ACnp1 is normally restricted to the central domain of centromeres where it forms the basis for the kinetochore. Central domain DNA is a preferred substrate for establishment of CENP-ACnp1 chromatin upon overexpression, whilst other genomic loci do not support accumulation of high levels of CENP-ACnp1 [47], and even vector DNA adjacent to the central domain is not a good substrate. Conditions and mechanisms that influence assembly of CENP-ACnp1 on naïve plasmid DNA are also able to convert pre-chromatinised cc2 present on episomal plasmids or linear minichromosomes. What makes central domain DNA a preferred site for CENP-ACnp1 assembly? The lack of homology between cc2 and cc1/cc3 sequences suggests that it is not a simple case of specific sequence that is critical [66–68]. Our analyses indicate that there is functional redundancy within the central domain and no one particular sequence is either necessary or sufficient for CENP-ACnp1 establishment, consistent with previous findings [50]. Despite this redundancy it appears that there are inherent distinctions between different regions of cc2. The 2 kb sub-regions, LM and OP, are functionally non-equivalent and consistently behaved differently when challenged to assemble CENP-ACnp1 chromatin. LM is competent to establish centromeric chromatin upon CENP-ACnp1 overexpression, contains sufficient information to make a functional centromere when placed next to heterochromatin (pH-3xLM), and assembles CENP-ACnp1 chromatin in cells lacking Ubp3. On the other hand, the OP region fails to become assembled in CENP-ACnp1 chromatin in all these situations, yet can accept CENP-ACnp1 when adjacent to one copy of LM, which apparently acts as an initiator. The ability of LM, but not OP, to substitute for full-length cc2 sequence indicates that not all sequences are equivalent and LM must contain all information necessary to make this region permissive for CENP-ACnp1 establishment. It is possible that the observed higher promoter activity observed in the OP region (Fig. 7D) prevents stabilisation of CENP-ACnp1 nucleosomes on this sequence.
In common with many organisms, the central domain of S. pombe centromeres is AT rich and this property might contribute to the propensity of centromeric DNA to attract CENP-A [5,68]. S. pombe central domain DNA has an AT content of 72% (genome average of 64%), as does the establishment competent LM sequence. However, other regions that alone fail to support CENP-ACnp1 establishment have a similar AT content, such as OP (71% AT) and intergenic regions (72% AT). Moreover, randomisation of the LM sequence resulted in SynR-LM that, even with identical nucleotide composition (72% AT), was incompetent for CENP-ACnp1 establishment. Thus, high AT content alone, even when it mimics natural nucleosome positioning predictions, is not a defining factor in CENP-ACnp1 assembly. Together our observations indicate that rather than there being a specific critical sequence, central domain sequences encode unique properties capable of triggering or promoting the establishment of CENP-ACnp1 chromatin.
Transcription-coupled remodelling is associated with the deposition of histone variants and could potentially contribute to the assembly of CENP-A chromatin [69,70]. However, the simple act of transcription cannot be sufficient to provide specificity to the deposition of CENP-A. Our observations suggest that the transcriptional landscape of the centromeric central domain is unusual: scattered promoters of various strengths resulting in pervasive low quality transcription and numerous TSSs on both strands, in conjunction with poly(dA:dT) tracts that are inherently difficult to transcribe are likely to cause collision between convergently transcribing RNAPIIs and pile-ups at difficult sequences [63,71]. The relatively high density of RNAPII on pMcc2 contrasts with very low levels of transcripts (Fig. 7), consistent with inefficient progress of transcription by RNAPII on cc2, and many stalled elongation complexes. In addition, long tracts of poly(dA:dT) are known to disfavour nucleosome assembly, consistent with the apparently wide spacing of nucleosomes at endogenous centromeres [49,72]. These regions may be de facto nucleosome free regions, similar to those at promoters, allowing cryptic initiation of transcription to occur [72,73]. The randomized synthetic sequence SynR-LM that is a poor substrate for CENP-ACnp1 deposition has similar long A tracts, but transcription-related sequence-sensitive elements—such as promoters and transcription factor binding sites—would be destroyed. Thus, the central domain, due to its sequence-encoded properties, may produce a distinctive chromatin and transcriptional environment.
CENP-ACnp1 chromatin does not assemble de novo on cc2 sequence alone in wild-type cells expressing normal CENP-ACnp1 levels [45]. Instead, we envisage that the unique transcriptional chromatin environment created by the cc2 sequence renders it permissive for CENP-ACnp1 establishment, but that establishment occurs only if other favourable conditions exist. CENP-ACnp1 is preferentially incorporated on these central domain sequences upon overexpression, when adjacent to heterochromatin, and in the absence of factors that usually enhance transcriptional elongation. Any explanation of CENP-ACnp1 chromatin establishment on central domain DNA must also account for how CENP-ACnp1 is incorporated instead of H3. Serine 2 in the CTD heptad repeat of Rpb1 is phosphorylated in elongating RNAPII, and this Ser2P-Rbp1/RNAPII becomes ubiquitylated upon stalling [74–76]. The ubiquitin hydrolase Ubp3 normally acts as a proof-reading activity to prevent degradation of stalled but rescuable RNAPII [56,61]. Absence of Ubp3 compromises the processing of stalled RNAPII, resulting in the accumulation of ubiquitylated Ser2P-Rbp1/RNAPII complexes. We propose that such modifications contribute to the distinctive status of central domain chromatin, leading to recruitment of factors that promote CENP-ACnp1 deposition (Fig. 8E). Alternatively, it may create an environment in which H3 nucleosomes are efficiently turned over/evicted, whereas CENP-ACnp1 nucleosomes are poorly evicted specifically in the context of stalled RNAPII. In cells lacking Ubp3, severe or prolonged stalling, even with normal levels of CENP-ACnp1, would provide extended opportunities for CENP-ACnp1 recruitment, or poor eviction of CENP-ACnp1 during prolonged stalling. TFIIS promotes transcriptional elongation by cleaving nascent transcripts in the context of stalled/backtracked RNAPII [57,58,77]. Although the effects of TFIIS deletion are more subtle than lack of Ubp3, the accumulation of RNAPII correlates with assembly CENP-ACnp1 chromatin, supporting a mechanism where persistent RNAPII stalling within central domain triggers remodelling that results in CENP-ACnp1 deposition.
In this model, when naïve central domain DNA (pMcc2) is introduced into wild-type cells, transient stalling occurs but it is efficiently cleared with the aid of factors such as TFIIS and Ubp3 (Fig. 8E). Because in wild-type cells CENP-ACnp1 levels are extremely low compared to histone H3 there would be little opportunity for CENP-ACnp1 to gain access to cc2, and with efficient clearing of stalled RNAPII, CENP-ACnp1 would fail to accumulate in cc2 [78]. CENP-ACnp1 overexpression would increase the probability of interaction with the transiently stalled RNAPII in central domain chromatin, increasing the likelihood of recruitment. Alternatively, increased access coupled with poor eviction would lead to CENP-ACnp1 accumulation. In addition, CENP-ACnp1 nucleosomes themselves, which have distinct N-terminal tails that lack the conserved lysine residues of H3 whose modification aids transcription, are likely to present a greater barrier to transcription than H3 nucleosomes [79]. Thus, once incorporated, CENP-ACnp1 nucleosomes might exacerbate the poor transcriptional elongation, creating conditions permissive for recruitment of more CENP-ACnp1 in a self-perpetuating system. Longer regions of central domain DNA would have greater probability of triggering stalling events and thus be more likely to initiate the incorporation of CENP-ACnp1. In the context of this model, heterochromatin could promote establishment of CENP-ACnp1 chromatin on adjacent cc2 sequence by drawing plasmids to sites of endogenous heterochromatin such as the spindle pole body where they would encounter a higher concentration of CENP-ACnp1 than non-heterochromatinized plasmids located in the nuclear interior [80]. Alternatively, heterochromatin-associated chromatin modifying activities may influence transcriptional elongation by RNAPII within cc2, causing enhanced stalling and deposition of CENP-ACnp1[41].
Following establishment of CENP-A chromatin and kinetochore assembly, transcription could play a proof-reading role that evicts H3 deposited at centromeres during S phase [81]. Indeed, transcription and RNAPII have been detected at centromeres in mammalian cells and transcription/RNAPII may play a role in centromere integrity [33,34,38]. Transcription of human α-satellite arrays introduced as HACs is known to occur. Although CENP-A assembly is compatible with targeting of mild transcriptional activators, targeting of a strong transcriptional activator is deleterious [30,38,82]. Thus transcription and/or the transcription-coupled histone modifications detected at centromeres may promote CENP-A deposition at mammalian centromeres.
In conclusion, we show that the sequence of fission yeast centromere central domain DNA is important only in so far as it encodes for certain properties that contribute to the region’s unusual chromatin and transcriptional landscape. Establishment of CENP-ACnp1 chromatin is driven by these sequence-encoded properties that when combined with the presence of nearby heterochromatin, overexpressed CENP-ACnp1 or increased RNAPII stalling, tips the balance in favour CENP-ACnp1 chromatin assembly. It seems likely that a similar combination of factors, which together favour CENP-A incorporation, must also contribute to the formation of neocentromeres at novel chromosomal locations.
Standard genetic and molecular techniques were followed. Fission yeast methods were as described [83]. Fission yeast strains are listed in Table 1. Minichromosomes used in this study were transformed by electroporation. Transformants were selected by growth on PMG—ura—ade at 32°C. As circular minichromosomes lack heterochromatin and therefore centromeric cohesion, plasmids were maintained in cells by selection in medium lacking adenine and uracil. 3 independent colonies from each transformation were analysed for the presence of kinetochore proteins by chromatin immunoprecipitation (ChIP).
Plasmids bearing centromere fragments contained a minimal ars1 element to ensure efficient replication in S. pombe, in addition to selectable markers sup3-5 (complements ade6-704), ura4+ and KANR. 8.5 kb of central domain DNA (cc2 plus inner part of imr2L and imr2R) was cloned into the multiple cloning site as a SalI-NcoI fragment to create pMcc2. Various sub-fragments of cc2 (J-Q) were amplified by PCR and cloned into the multiple cloning site as BamHI/BglII fragments. 5.6 kb of heterochromatin-forming outer repeat sequence was inserted adjacent to central domain sequences to test ability to form functional centromeres.
A plasmid, pMC28, bearing cc2, a KAN resistance marker and an inverted ura4 sequence was constructed from pMcc2. Linearisation of the plasmid at NotI within the inverted ura4 sequence allowed integration at ura4+ located on the arm of Ch16-m23:ura4+. Ch16-m23: ura4+ is a derivative of Ch16, a 530 kb minichromosome, itself derived from Chromosome III [48]. It also bears the ade6-216 allele which complements the ade6-210 allele present on endogenous Chromosome III by interallelic complementation. Integration of linearised pMC28 on Ch16-m23:ura4 allowed selection on the counter-selective drug 5-fluoro-orotic acid and G418 (KAN). Cells that lost the Ch16-m23:ura4::cc2-KAN (abbreviated as Ch16-cc2) became red on limiting adenine and were sensitive to G418. For growth in liquid, cells containing Ch16-cc2 were grown in media lacking adenine.
ChIP was performed as previously described [84] using anti-CENP-Acnp1 antibody, anti CENP-CCnp3 antibody, anti-CENP-KSim4 antibody, anti-H3 antibody (ab1791; Abcam,), anti-H3K9me2 antibody (T. Urano) and anti-total RNA polymerase II (4F8; 61081, Active Motif), anti-Rpb1-Ser2P (3E10; 61083, Active Motif) and analysed by qPCR. Primers are listed in Table 2. P-values were calculated by standard t-test on 3 replicates between wild-type and mutant; p<0.05 was considered significant.
For the establishment assay, cells were transformed with minichromosomes (containing 5.6 kb of outer repeat sequence in addition to cc2 sequences), by electroporation with ∼200 ng of DNA and plated on selective medium. Resultant colonies were replicated onto rich medium containing limiting adenine. The presence of pale pink/white colonies indicates establishment of a functional centromere on the minichromosome. Establishment efficiency is calculated as percentage of these colonies divided by the total number of transformants. Colonies were streaked on limiting adenine plates to confirm the presence of sectoring that is indicative of centromere function.
Quantitative PCR reactions were carried out in 10 μl volume, with 5μl Light Cycler 480 SybrGreen Master Mix (Roche), 0.5μl each primer (10 μM) and 3μl ChIP or total template. The data were analysed using Light Cycler 480 Software 1.5 (Roche).
5’RACE-PCR was performed as previously described [37]. In brief, RNA was isolated with RNeasy mini/midi kit (Qiagen) according to the manufacturer’s protocol. Poly(A) containing RNA was purified from 500 μg of total RNA by affinity purification with biotinylated oligo-dT using PolyATtract mRNA Isolation Systems (Promega). 5’RACE PCR was performed using SMARTer 5’/3’ RACE (Clontech) according to the manufacturer’s protocol. PCR products were then run on 1% agarose gel, purified and cloned into pGEM-T Easy vector (Promega) and subsequently sequenced. Reverse transcription reaction for 5’RACE and qRT-PCR was performed using Superscript III Reverse Transcriptase (Invitrogen) using RNA extracted from 3 independent colonies. For qRT-PCR, transcript levels were normalized over gDNA to take into account differences in copy number between plasmids and normalized relative to act1+.
LacZ assay was performed as described [85]. pREP81X-LacZ was digested with XhoI and PstI and the nmt81 promoter upstream of LacZ replaced with sequences from centromere 2. Plasmids were transformed into wild-type and grown on minimal medium (n = 3).
DNA was extracted as previously described [83]. The DNA was digested with BglII/SpeI or SphI/SpeI, run on a 1% agarose gel, blotted on nylon membrane (Hybond N, Amersham) and UV-crosslinked. The membrane was hybridized with DNA probes specific for central domain 1 or central domain 2. To make the probes, PCR products were used as template in the labelling reaction using High Prime (Roche). Primers sequences are listed in Table 2.
Western analysis was performed as described previously using anti-GFP antibody (Roche) and anti-H3 antibody (ab1794-abcam) [86]. The intensities of GFP and H3 signals were acquired using LICOR Odyssey Infrared Imaging System software (Li-COR Bioscience).
|
10.1371/journal.pcbi.1004362 | Prediction of Functionally Important Phospho-Regulatory Events in Xenopus laevis Oocytes | The African clawed frog Xenopus laevis is an important model organism for studies in developmental and cell biology, including cell-signaling. However, our knowledge of X. laevis protein post-translational modifications remains scarce. Here, we used a mass spectrometry-based approach to survey the phosphoproteome of this species, compiling a list of 2636 phosphosites. We used structural information and phosphoproteomic data for 13 other species in order to predict functionally important phospho-regulatory events. We found that the degree of conservation of phosphosites across species is predictive of sites with known molecular function. In addition, we predicted kinase-protein interactions for a set of cell-cycle kinases across all species. The degree of conservation of kinase-protein interactions was found to be predictive of functionally relevant regulatory interactions. Finally, using comparative protein structure models, we find that phosphosites within structured domains tend to be located at positions with high conformational flexibility. Our analysis suggests that a small class of phosphosites occurs in positions that have the potential to regulate protein conformation.
| Proteins can be modified during their life-cycle in order to regulate their function. The addition of a phosphate group is one of the most abundant and well understood protein modifications. Recent technological developments are now allowing us to uncover thousands of phosphorylation sites within proteins in a single experiment. Here we applied these approaches to identify phosphorylation sites in the African clawed frog Xenopus laevis. It has been suggested that many such modifications might serve no biological function inside the cell. To address this we have used a comparative approach to identify highly conserved phosphorylation sites and regulatory interactions. In addition we have used 3D structural models to identify a set of phosphorylation sites that might regulate protein conformations.
| Protein function can be regulated by post-translational modifications (PTMs) by altering diverse protein properties such as their localization, activity or interactions. Protein phosphorylation is one of the most well studied PTMs with over 50 years of research since the pioneering work of Krebs and Fischer on glycogen phosphorylase [1]. It is estimated that approximately 30% of the human proteome can be phosphorylated and this modification has been shown to play a role in a very broad set of cellular and developmental functions as well as dysregulation in disease [2]. Recent advances in phosphoenrichment procedures and mass spectrometry (MS) technologies have resulted in a tremendous increase in the capacity to identify phosphorylation sites on a large scale [3] and over the past few years over 200.000 phosphorylation sites have been identified across a varied number of species (ptmfunc.com). These studies have highlighted the true extent and complexity of PTM regulation and underscored the need to develop large-scale approaches to study PTM function. The availability of such data for different species has allowed for comparative studies. While a significant level of constraint on phosphorylated residues has been detected [4] a large fraction of phosphosites are not conserved across species [5–12]. Given the high evolutionary turn-over of these modification sites it is plausible that a fraction of these serve no biological purpose in extant species [5,13]. Therefore, it has become important to develop methods to discern the functional relevance of PTMs [14]. For example, the conservation of phosphosites has been used to highlight sites that are more likely to be important. Different kinases have specific preferences for the amino acids in the vicinity of the target phosphorylated residue [15–17]. This local sequence context is often referred to as the kinase target consensus sequence or motif and the conservation of these kinase motifs across orthologous proteins can be used to improve the predictions of kinase regulated sites [18,19]. In parallel to conservation based approaches, computational and experimental methods have been developed to identify phosphosites that are more likely to be functionally important by regulating protein interactions [13,20], protein activities [13], metabolic enzymes [21], or cross-regulate other types of modifications [22–24].
Phosphoproteomic approaches have been applied extensively to several species [7,8,25–28]. However, although X. laevis is a well-established model organism there has been little previous knowledge of the extent and conservation of its phosphoproteome. To address this we have used a MS approach identify phosphorylation sites in X. laevis egg extracts. This approach resulted in the identification of 1738 phosphorylation sites. For the subsequent analysis, we combined these sites with sites identified in a previous study [29] resulting in a total of 2636 phosphosites for analysis. Using a compilation of phosphorylation information for 13 other species we identified a small number of highly conserved phosphosites, which were found to be enriched in sites with known molecular functions. In addition we used kinase specificity predictions for kinases involved in cell cycle regulation to predict conserved kinase-protein associations. The degree of conservation of predicted cell cycle-related kinase interactions was correlated with known kinase-protein regulatory interactions. Conserved putative interactions were also enriched in proteins that are phosphoregulated during the cell cycle and in genes that when knocked down cause mitotic phenotypes. In order to study the structural properties of these sites we obtained comparative models for 518 phosphosites. Structural analysis revealed a number of solvent inaccessible phosphosites that likely indicate protein regions that can exist in more accessible conformations. The analysis of these sites suggests that a significant fraction may regulate protein conformation.
Xenopus laevis egg extracts were prepared in two cell cycle stages (interphase and mitosis) in order to increase the coverage of phospho-regulatory events. After protein extraction the samples were trypsin digested, subjected to a phosphopeptide enrichment protocol and LC-MS analysis (see Materials and Methods). The spectra were matched to a reference proteome for X. laevis and phosphosites were identified with a false discovery rate below 1% as estimated by a decoy library search. The localization of the phosphosite acceptor residue with the peptides was scored using the SLIP score [30]. We identified a total of 1738 non-redundant phosphosites in both samples. Using the SLIP scores we estimate the 77% of these phosphosites are well localized within the phosphopeptide (see S2 Table). In addition we have also compiled 1076 sites from a previous study [29] that we were able to map to the same reference proteome. The two sets obtained in this study have a total of 2636 non-redundant sites (S1 Table). The distribution of modified residues is similar to previous phosphoproteomics studies with 2072 (78%) phospho-serines, 453 (17%) phospho-threonine and 111 (4%) phospho-tyrosines (Fig 1A). For most of the analysis we do not require unambiguous localization of the phospho-acceptor residue within the phosphopeptide and used the total set of 2636 phosphosites. To facilitate the re-use of this data we also selected a higher quality subset of 941 phosphosites with high expectation and localization scores (see Methods and S1 Table).
In order to study the structural properties of these phosphosites we obtained comparative models for X. laevis phosphoproteins. The models were created with ModPipe [31] using templates with at least 25% sequence identity and using established model quality criteria (see Materials and Methods). When different models were available for the same phosphosite-containing region we selected the largest model available. We were able to obtain models for a total of 518 phosphosites of which 267 are contained within known PFAM domain (Fig 1A). An additional 280 are within PFAM domain boundaries for which we could not build a model. Of the total number of sites that are likely to be within globular protein domains, as defined by PFAM, we could model approximately 49%. In addition to structural information we also determined the level of the conservation of the identified phosphosites using a compilation of phosphorylation information for a set of 13 other species obtained from ptmfunc.com (Saccharomyces cerevisiae, Schizosaccharomyces pombe, Plasmodium falciparum, Toxoplasma gondii, Trypanosoma brucei, Trypanosoma cruzi, Oryza sativa, Arabidopsis thaliana, Drosophila melanogaster, Caenorhabditis elegans, Homo sapiens, Rattus norvegicus and Mus musculus). The number of predicted orthologs and known phosphosites used for each species is provided in S3 Table. X. laevis proteins were aligned with putative orthologs in these species and a phosphosite was considered to be conserved in a target species when the aligned peptide region was known to be phosphorylated in that species (see Materials and Methods). Previous studies have noted that regulation by protein phosphorylation can diverge quickly during evolution [5–7,9,10]. In line with these studies we find that only 1050 (39.8%) sites were found to be conserved in one or more species analyzed (Fig 1A). We note that the conservation values of the phosphorylation status are under-estimated due to lack of complete coverage for most phosphoproteomes. We next combined the structural information and the conservation information to study the surface accessibility of the phosphosites. As expected [32,33], phosphosites are on average more likely to have higher all-atom relative surface accessibility than non-modified residues. This is apparent for serines and threonines (Fig 1B, phospho-serines vs. serines p-value = 2.328x10-9, phospho-threonines vs. threonine p-value = 3.518x10-5 with a two sample Kolmogorov-Smirnov test) but not for tyrosines. Phosphosites conserved in at least one species do not appear to be more surface-exposed than average phosphosites (Fig 1B).
A small fraction of phosphosites was found to be conserved across several species. We observed for example that 82 sites were conserved in 4 or more species (Fig 1A). In order to test the usefulness of this comparative approach we used a list of human sites known to have a molecular function from small-scale studies (from phosphosite.org). We first restricted the analysis to 596 X. laevis phosphosites conserved in H. sapiens. We then tested if the level of conservation in additional species beyond human was predictive of a known function in human. While 62 of these 596 (10.4%) X. laevis sites have a known human function we observed that the fraction of sites with known function increased with the level of conservation of the phosphorylation status (Fig 1C). Of the 66 Xenopus sites that are conserved in human and in at least 3 other species 22 have a known human function (33%). We next predicted protein disorder for X. laevis proteins using DISOPRED (version 3.1) and repeated this analysis separately for protein regions predicted to be ordered and disordered. In both cases the trend is the same with a stronger enrichment observed for disordered regions (S1 Fig). This large and significant increase suggests that phosphosites conserved across many distantly related species are more likely to be functionally relevant. Increasing coverage of experimentally determined phosphorylation sites across a varied number of species will further facilitate the identification of such highly conserved sites. Examples of highly conserved sites with available homology models are shown in Fig 1D. For example, the phosphorylation of the activation loop region of the GSK3b protein kinase (Fig 1D, left) is one of the most conserved phosphorylation sites across all species. Phosphorylation of the activation loop region of protein kinases is a very well established mechanism to regulate kinase activity. HSP90 proteins are one of the most conserved molecular chaperones involved in the folding of a varied set of client proteins [34]. This chaperone is a highly flexible protein typical forming a homo-dimmer via a c-terminal region [34]. One of the most conserved phosphorylation sites we identified here is in a protein homologous to the human HSP90AB1 (Fig 1D, center). The phosphorylation occurs in the flexible n-terminal region that opens and closes during the ATPase cycle [34]. This position is equivalent to S226 in the human HSP90AB1 that has been previously shown to regulate chaperone activity [35]. Another interesting phosphosite, conserved in 4 other species, is the threonine near the active site of NDP kinase A (Fig 1D, right) which due to the proximity to the substrate is very likely to influence enzyme activity. Given the broad phylogenetic distribution of the species used in this study, these highly conserved sites are expected to be enriched in ancient and functionally important phospho-regulatory modifications. A list of the X. laevis sites that are phosphorylated in the predicted human ortholog and in at least 4 other species are listed in Table 1, along with annotations on the molecular role from PhosphositePlus (www.phosphosite.org).
Although phosphosites tend to have high all-atom relative surface accessibility (RSA) around 20% of the sites appear to be poorly accessible—here defined as having below 20% RSA. Given that surface accessibility should be a requirement for kinase regulation we explored potential explanations for these low accessibility sites. We hypothesized that this observation could be due to three potential factors: incorrect homology models, false positive phosphosites, or changes in protein conformation. We reasoned that if model quality was a determinant factor in explaining the inaccessible sites then the fraction of such sites should decrease with increasing quality of the models. However, homology models obtained from templates of higher sequence identity had a similar distribution of phosphosite RSA (Fig 2A). In order to test the impact of false positive sites we relied on the idea that conserved phosphosites are very unlikely to be experimental false positives. We noted that conserved sites have a similar fraction of poorly accessible phosphosites than non-conserved sites and no clear trend of accessibility relative to conservation is apparent. Overall these results suggest that false positive phosphosites are unlikely to be a major determinant for low surface accessibility sites.
If inaccessible phosphosites are not mostly due to incorrect homology models or false positive sites then conformation flexibility might explain why some sites appear to have low accessibility. Phosphosites are well known to regulate protein function by controlling protein conformation [36–39]. Xin and colleagues have compared structures of the same proteins in their modified and un-modified form and noted that PTMs tend to be associated with changes in protein conformation [37]. In addition, a structural analysis of 7 protein structures using normal mode analysis suggested that low accessibility sites could, in some cases, become accessible by conformational changes [40]. These studies suggest that the phosphosites that appear to be poorly accessible may occur in regions of proteins that can become more accessible in a conformation that is not captured in the structural template used to create these models. In order to study this we analyzed phosphoproteins for which we had more than one model structure obtained from different templates. For each pair of comparative models we correlated the accessibility for all serine, threonine and tyrosine residues. Overall, there is a high correlation of all-atom RSA value for different templates (Fig 2B, r = 0.67). However, phosphorylated residues showed a significantly higher change in accessibility in different models when compared to non-modified residues (Fig 2C, p-value = 6x10-6, two sample Kolmogorov-Smirnov test). The median absolute change in accessibility is 7.1 for the phospho-acceptor residues, 11.55 for phosphosites and 12.3 for conserved phosphosites. This result suggests that phosphosites are more likely to be in regions that show high conformation flexibility across different structural models. Phosphorylation sites are known to preferentially occur in disordered regions [41,42]. It is possible that the high conformational variability of phosphosites could be due to higher flexibility and/or lower modeling quality of disordered loop. We used DISOPRED to predict protein disorder across all phosphosites. As expected, we observed that phosphosites predicted to be in disordered regions had a higher median change in accessibility when compared to ordered sites (14.6 versus 10.7, Fig 2C). However, phosphosites predicted to be within ordered regions still have a significantly higher change in accessibility when compared to all acceptor residues (Fig 2C, 10.7 versus 7.1, p-value 6x10-4, two-sample Kolmogorov-Smirnov test). Phosphorylation can, in some cases, regulate protein conformation [43] and this analysis suggests that it is possible to use comparative models to define a class of functional phosphosites that can play a role in conformation regulation.
We analyzed in more detail sites that were at positions with large changes in accessibility across different templates and where the site had RSA below 20% in at least one model. Three of such cases are shown in Fig 2D were we superimposed the two models highlighting the differences in accessibility. YBX1 and YBX2 are RNA binding proteins and are phosphorylated in the cold-shock protein (CDP) domain (PFAM:PF00313) in a loop region (Fig 2D, left) that is known to be highly flexible and does not appear to play a direct role in RNA binding [44]. We cannot discount the possibility that large changes in accessibility in such large flexible loops are due to difficulties in modeling such protein regions. The Ran GTPase was also observed to be phosphorylated in a position with different accessibility in different templates (Fig 2D, center). The phosphosite is in a position equivalent to S135 in human RAN that has been previously shown to be regulated by p21-activated kinase 4 (PAK4) during the cell-cycle in human and X. laevis [45]. In addition, Ran S135 alanine and phosphomimetic mutants had an impact on microtubule nucleation in X. laevis extracts and in binding the exchange factor RCC1 in human cell lines [45]. It is possible that PAK4 regulation of S135 could promote a conformation that does not favor the interaction with RCC1. The third example is a phosphosite on the Dynamin 1-Like pleckstrin homology (PH) domain that comprises the “foot” region of dynamin like proteins (Fig 2D, right). The PH domain can adopt different conformations relative to the rest of the protein and the phosphorylated position changes drastically in accessibility depending on the conformation. The phosphorylated position is equivalent to the S635 in rat Drp1 and S616 in human Drp1 that has been previously shown to be regulated during cell-cycle and have functional roles in mitochondrial fission and microtubule targeting [46,47]. We hypothesize that the phosphosite may regulate Drp1 function by restricting the possible conformation variability of the PH domain relative to the rest of the protein. Although these examples would require experimental validation they illustrate how the structural analysis of phosphosites using different structural templates might suggest mechanist explanations for the function of MS identified sites.
Previous studies have shown that the conservation of kinase sequence motifs in alignments of orthologous proteins could be used to improve the predictions of kinase target sites [18,19]. However, some kinases are known to regulate proteins in clusters of sites [48–50] and in some cases it is plausible that the exact position of the target site might change within the protein during evolution yet maintaining the kinase-protein regulatory interaction [7,12,51,52]. In line with this reasoning, predicted kinase-protein interactions were often found to be conserved across species even when the phosphosite positions were not conserved [7]. We hypothesize that the conservation of predicted kinase-protein interactions across the large number of species analyzed here could be used to predict functionally important interactions. We decided to test this hypothesis focusing on cell-cycle kinases which allowed us to take advantage of previous large scale studies of cell cycle phosphoregulation [53] and phenotypes [54] for benchmarking purposes. To predict the targets of cell cycle-related kinases, we selected 16 kinases with many known target sites (Akt, Atr, AurA, AurB, Cdk1, Cdk2, Cdk3, Cdk7, Chk1, Nek2, Nek6, Nek9, Plk1, Plk2, Plk3, Ttk). Position specific scoring matrices (PSSMs) were derived for each kinase and benchmarked on a set of known target sites using cross validation (Methods and S2 Fig). From the 16 kinases we then selected 10 that had a cross validation area under the ROC curve (AROC) >0.7 (Akt, Atr, AurB, Cdk1, Cdk2, Cdk3, Chk1, Nek6, Plk1, Plk3, S3 Fig). For each species we predicted which phosphosites match the kinase preference of the 10 kinases studied. A phosphoprotein was predicted to be regulated by a kinase if at least one phosphosite was a predicted target of that kinase. The number of X. laevis sites and proteins predicted to be regulated by each kinase is shown in Fig 3A.
We next tested if the degree of conservation of kinase interactions is a useful predictor for known and functionally important kinase-target interactions. A putative X. laevis kinase-protein interaction was considered to be conserved in another species if an ortholog in that species had at least one phosphosite that was predicted to be a target of the same kinase model. We note that the sites do not have to be in the same region of the orthologous proteins. We ranked X. laevis kinase interactions according to the degree of conservation in the other 13 species. We then measured the enrichment of known or functionally important kinase-interactions relative to all predicted X. laevis kinase-protein interactions as a function of the degree of conservation. We observed that the degree of conservation was a significant predictor of previously described human kinase-protein interactions (Fig 3B AROC = 0.76 and 3C, “known interactions”). Highly conserved kinase-protein interactions were also enriched in target proteins that were previously known to be phosphoregulated during the cell cycle in human [53] (Fig 3B AROC = 0.78 and 3C, phospho-regulated) and enriched in genes that when knocked-down cause cell cycle-related phenotypes [54] (Fig 3B AROC = 0.66 and 3C, mitocheck phenotype). To exclude the possibility that the enrichment is due to the high scoring sequences that may be orthologous to the known human target sites of these kinases we repeated the analysis excluding all phosphosites that are 100% identical to known target sites of each kinase. No significant differences in the enrichment were observed as measured by the AROC curves (S4 Fig). To facilitate the re-use of these predictions we annotated 68 putative kinase-interactions conserved in 7 or more species (Fig 3D, S4 Table). When compared to all putative kinase-interactions this network is 6.8-fold enriched in known kinase-interactions (Fig 3C and 3D green arrows), 1.9-fold enriched in cell cycle phospho-regulated proteins (Fig 3C and 3D orange circles) and 1.3-fold enriched in genes associated with cell cycle-related phenotypes (Fig 3C and 3D red outline).
To facilitate the study of X. laevis PTM signaling, we obtained here a current survey of this species’ phosphoproteome. Previous studies have indicated that phosphoregulation can diverge quickly during evolution and that a fraction of phosphosites might serve no biological function in extant species. We used a compilation of phosphorylation data for 13 other species to identify highly conserved phosphosites and potential kinase-protein interactions. While others have shown that conservation of kinase sequence motifs across orthologous positions is a useful filter to predict kinase interactions [18,19] we show here that the degree of conservation of experimentally determined phosphorylation states is a strong predictor of sites with known function, of kinase-protein interactions and of function specific annotations (i.e. cell-cycle regulated and phenotypes). A small fraction of sites appear to be conserved over a large number of species. Given the divergence times separating the species studied here, these phosphosites likely have very ancient origins despite the fast evolutionary turn-over of phospho-regulation. As the experimental data on protein phosphorylation is incomplete for most species the conservation values presented here are under-estimated. Additional data will allow for further identification of such “ultra-conserved” and functionally important sites. Given that the collection of phosphorylation data across most species does not take into account different environmental or developmental conditions, these highly conserved sites are potentially biased for phosphosites that are constitutively on. Also, due to MS bias for higher protein abundance, highly conserved phosphosites described here are also potentially biased for proteins of higher abundance.
Although conservation is useful predictor of functional phosphosites, there are functionally important sites that are not highly conserved. For these reasons it is important to develop approaches that do not rely on conservation to rank PTMs according to biological importance. In this context, we and others have previously made use of structural information to predict PTMs that have the potential to regulate protein-protein interactions [13,20]. Using comparative models for X. laevis phosphoproteins we observed that some phosphosites appear to be in inaccessible regions and that phosphosites tend to be in positions of higher variability in surface accessibility across different structural templates. Our analysis suggests that structural models can therefore be used to predict, in an unbiased way, PTMs with the potential to regulate protein conformation. It will be important to verify this finding across different species and other PTM types. This approach could be further extended by including other structure based approaches such as normal mode analysis [40] and molecular dynamics [55] as well as sequence based approaches such as statistical coupling analysis [56].
The majority of PTMs identified to date for human and other species has no known function. Given the large throughput of MS approaches and the low fraction of PTMs with currently known functions much additional effort needs to be committed to the development of computational and experimental methods to elucidate PTM function. The evolutionary and structural observations presented here can be used to facilitate the prioritization of PTM functional studies in any species.
All animal work was conducted according to relevant national and international guidelines. Animal protocols were approved by the Stanford University Administrative Panel on Laboratory Animal Care.
Frog eggs were obtained from female Xenopus laevis as described in [57,58]. Briefly, frogs were primed by injecting 50U of pregnant mare serum gonadotropin [PMSG) into the dorsal sacs 72h before egg collection. Egg laying was induced by injecting 500U of human chorionic gonadotropin (HCG) 18h before egg collection. 50ml of laid eggs were used for preparing interphase and mitotic extracts. A total of 500ul (at 20mg/ml) of interphase egg extracts supplemented with an ATP regenerating system and cycloheximide (100mg/ml) was prepared in the presence of protease inhibitors (leupeptin, pepstatin, cytochalasin and chymostatin). Half of the interphase extract was quick freeze for MS analysis. In order to make mitotic egg extracts, 250ul of interphase egg extracts were treated with 100nM non-degradable Xenopus D65-cyclin B1 and reactions were incubated for 1 hour and 30min at 22C. De-membranated sperm chromatin was added at 500/μl and samples were collected and stained with DAPI in order to monitor nuclear morphology and mitotic progression by fluorescence and phase microscopy. The criteria for M phase entry were condensed chromatin and a lack of a discernable nuclear envelope in at least 90% of the nuclei.
Xenopus extracts were denatured in a buffer containing 8M urea, 0.1M Tris pH 8.0, and 150 mM NaCl. Disulfide bonds were reduced by incubation with 4 mM TCEP for 30 minutes at room temperature, and free sulfhydryl groups were alkylated by incubation with 10 mM iodoacetamide for 30 minutes at room temperature in the dark. Samples were diluted back to 2 M urea by addition of 0.1 M Tris pH 8.0, and trypsin was added at an enzyme:substrate ratios of 1:100. Lysates were digested overnight at 37 degrees Celsius. Following digestion the samples were concentrated using SepPak C18 cartridges (Waters). The C18 cartridge was washed once with 1 mL of 80% ACN, 0.1% TFA, followed by a 3 mL wash with 0.1% TFA. 10% TFA was added to each samples to a final concentration of 0.1% after which the samples were applied to the cartridge. The cartridge was washed with 3 mL of 0.1% TFA after binding, and the peptides were eluted with 40% ACN, 0.1% TFA. Following elution the peptides were lyophilized to dryness. Phosphopeptides were fractionated using hydrophilic interaction chromatography (HILIC) adapted from a method published by McNulty and Annan [59]. Buffers used for HILIC separation were HILIC buffer A (2% ACN, 0.1% TFA) and HILIC buffer B (98% ACN, 0.1% TFA). Peptides were resuspended in 90% HILIC buffer B and loaded onto a TSKgel amide-80 column (Tosoh Biosciences, 4.6 mm I.D. x 25 cm packed with 5 um particles). Peptides were separated at a flow rate of 0.5 mL / min using a gradient from 90% to 85% HILIC buffer B for 5 minutes, 85% to 55% HILIC buffer B for 80 minutes, then 55% to 0% HILIC buffer B for 5 minutes. Fractions were collected every 2 minutes and the 22 fractions previously determined to contain the majority of phosphopeptides were evaporated to dryness. Following HILIC fractionation, fractions were further enriched for phosphopeptides using titanium dioxide magnetic beads (Pierce) using the manufacturer’s protocol. Following titanium dioxide enrichment, sample were evaporated to dryness and resuspended in 0.1% formic acid for mass spectrometry analysis.
Each fraction was analyzed separately by a Thermo Scientific LTQ Orbitrap Elite mass spectrometry system equipped with an Easy-nLC 1000 HPLC and autosampler. Samples were injected directly onto a reverse phase column (25 cm x 75 um I.D. packed with ReproSil-Pur C18-AQ 1.9 um particles) in buffer A [0.1% formic acid) at a maximum pressure of 800 bar. Peptides were separated with a gradient from 0% to 5% buffer B (100% ACN, 0.1% formic acid) over 5 minutes, then 5% to 30% buffer B over 52 minutes, then 30% to 95% buffer B over 1 minute, then held at 95% buffer B for 6 minutes. The separation was performed at a flow rate of 400 nl/min. The mass spectrometer continuously collected spectra in a data-dependent manner, acquiring a full scan in the Orbitrap (at 120,000 resolution with an automatic gain control target of 1,000,000 and a maximum injection time of 100 ms) followed by collision-induced dissociation spectra for the 20 most abundant ions in the ion trap (with an automatic gain control target of 10,000, a maximum injection time of 10 ms, a normalized collision energy of 35.0, activation Q of 0.250, isolation width of 2.0 m/z, and an activation time of 10.0). Singly and unassigned charge states were rejected for data-dependent selection. Dynamic exclusion was enabled to data-dependent selection of ions with a repeat count of 1, a repeat duration of 20.0 s, an exclusion duration of 20.0 s, an exclusion list size of 500, and exclusion mass width of + or—10.00 ppm. Raw mass spectrometry data was converted to peaklists using the PAVA algorithm. Data were searched using the Protein Prospector suite of algorithms (prospector.ucsf.edu). The data were searched against a X. laevis proteome obtained from the genome sequencing project. Specifically, we used a version containing 24,762 gene models obtained from sequencing of tissue samples and containing the longest gene model for each putative orthologous group. ("OrthoGeneOne" model of Taira201203_XENLA_tissue data, available at http://www.marcottelab.org/index.php/XENLA_GeneModel2012). Searches were run with a concatenated decoy database comprised of all sequences with their amino acids randomized. The algorithm searched for fully tryptic peptides with up to 2 missed cleavages using a parent mass tolerance of 20 ppm and a fragment mass tolerance of 0.8 Da. The algorithm indicated a static modification for carboxyamidomethyl of cysteine residues, and for variable modifications acetylation of protein N-termini, glutamine to pyroglutamate conversion, methionine oxidation, and phosphorylation of serine, threonine, or tyrosine residues. Data were filtered using a Protein Prospector expectation value that was resulted in a false discovery rate of 1% as determined by the number of matches to the randomized decoy database. The number of phosphosites identified in each HILIC fraction are provided in S5 Table. The phosphosite localization within the peptide was scored using the SLIP score [30]. For most of the analysis we made use of all phosphosites, included those that are ambiguously localized within the peptide sequence. A higher confidence list of well localized sites was generated by selected phosphosites with an Evalue<0.001 and a SLIP score > = 3. Benchmarks for the SLIP score suggest that at this cut-off over 90% of the sites are well localized. The list of identified phosphopetides and corresponding quality scores is provided in S1 Table.
Putative orthologs of X. laevis phosphoproteins were predicted using the reciprocal best-BLAST hits method [60] against a set of 13 proteomes with currently available phosphorylation information. Putative orthologs were aligned using MUSCLE [61]. The phosphorylation information for the 13 species was retrieved from the ptmfunc database (http://ptmfunc.com) and includes phosphosite information for Saccharomyces cerevisiae, Schizosaccharomyces pombe, Plasmodium falciparum, Toxoplasma gondii, Trypanosoma brucei, Trypanosoma cruzi, Oryza sativa, Arabidopsis thaliana, Drosophila melanogaster, Caenorhabditis elegans, Homo sapiens, Rattus norvegicus and Mus musculus. An X. laevis phosphosite was considered to be conserved in a target species if the predicted orthologous protein was known to be phosphorylated in the target species in a window of +/-2 residues around the aligned position. A window was used to take into account the alignment uncertainty and the ambiguity in identifying the exact position of the phosphorylated residue within a phosphopeptide. To predict the targets of cell cycle-related kinases, we selected 16 kinases (Akt, Atr, AurA, AurB, Cdk1, Cdk2, Cdk3, Cdk7, Chk1, Nek2, Nek6, Nek9, Plk1, Plk2, Plk3, Ttk) to train specificity models based on known target site data. Kinase substrate data for 357 kinases was obtained from public databases (PhosphoSitePlus [62], PhosphoELM [63] and HPRD [64]). In total we collected 9,595 kinase substrate relationships (KSR), based on 6,747 kinase-associated phosphosites. The positive set of phosphosites for a given kinase model was defined as the set of phosphosites annotated to that kinase, whereas the negative set is defined as sites annotated to any other kinase. Position specific scoring matrices (PSSMs) were derived for each of the 16 kinases and benchmarked on a set of known target sites using a cross-fold validation and area under the ROC curve (S2 Fig). From the 16 kinases we then selected 10 that had a cross-fold validation AUC>0.7 (Akt, Atr, AurB, Cdk1, Cdk2, Cdk3, Chk1, Nek6, Plk1, Plk3). Each model was used to score the corresponding positive and negative sets of peptides using the Matrix Similarity Score (MSS) as described in [65] and the MSS threshold that maximised the accuracy was used. Positive and negative sample sizes are required to be similar if not equal; therefore, the positive set was up sampled with replacement to the size of the negative sample, prior to computing the accuracy. MSS cutoffs for each kinase is as follows: Plk1 (0.436), Plk3 (0.386), Chk1 (0.236), Cdk1 (0.929), Cdk2 (0.126), Cdk3 (0.455), AurB (0.14), Atr (0.932), Akt (0.662), Nek6 (0.402). Structural models of X. laevis phosphoproteins were built automatically using ModPipe [31] relying on Modeller 9.10 [66]. A model was considered acceptable if the template sequence identify was at least 25% and met one additional criterion: TSVMod NO35 > = 40%, GA341 > = 0.7, E-value <0.0001 or zDOPE <0. All-atom residue relative surface accessibility was computed using NACCESS [67]. A list of models created with selected PDB codes and model quality values are available in S6 Table.
|
10.1371/journal.pgen.1003993 | Quantifying Missing Heritability at Known GWAS Loci | Recent work has shown that much of the missing heritability of complex traits can be resolved by estimates of heritability explained by all genotyped SNPs. However, it is currently unknown how much heritability is missing due to poor tagging or additional causal variants at known GWAS loci. Here, we use variance components to quantify the heritability explained by all SNPs at known GWAS loci in nine diseases from WTCCC1 and WTCCC2. After accounting for expectation, we observed all SNPs at known GWAS loci to explain more heritability than GWAS-associated SNPs on average (). For some diseases, this increase was individually significant: for Multiple Sclerosis (MS) () and for Crohn's Disease (CD) (); all analyses of autoimmune diseases excluded the well-studied MHC region. Additionally, we found that GWAS loci from other related traits also explained significant heritability. The union of all autoimmune disease loci explained more MS heritability than known MS SNPs () and more CD heritability than known CD SNPs (), with an analogous increase for all autoimmune diseases analyzed. We also observed significant increases in an analysis of Rheumatoid Arthritis (RA) samples typed on ImmunoChip, with more heritability from all SNPs at GWAS loci () and more heritability from all autoimmune disease loci () compared to known RA SNPs (including those identified in this cohort). Our methods adjust for LD between SNPs, which can bias standard estimates of heritability from SNPs even if all causal variants are typed. By comparing adjusted estimates, we hypothesize that the genome-wide distribution of causal variants is enriched for low-frequency alleles, but that causal variants at known GWAS loci are skewed towards common alleles. These findings have important ramifications for fine-mapping study design and our understanding of complex disease architecture.
| Heritable diseases have an unknown underlying “genetic architecture” that defines the distribution of effect-sizes for disease-causing mutations. Understanding this genetic architecture is an important first step in designing disease-mapping studies, and many theories have been developed on the nature of this distribution. Here, we evaluate the hypothesis that additional heritable variation lies at previously known associated loci but is not fully explained by the single most associated marker. We develop methods based on variance-components analysis to quantify this type of “local” heritability, demonstrating that standard strategies can be falsely inflated or deflated due to correlation between neighboring markers and propose a robust adjustment. In analysis of nine common diseases we find a significant average increase of local heritability, consistent with multiple common causal variants at an average locus. Intriguingly, for autoimmune diseases we also observe significant local heritability in loci not associated with the specific disease but with other autoimmune diseases, implying a highly correlated underlying disease architecture. These findings have important implications to the design of future studies and our general understanding of common disease.
| While association studies have been successful in finding a large number of significant variants for many complex traits, they have individually explained relatively little of the total heritability, motivating analyses that seek to identify this so-called “missing” heritability [1]–[3]. One hypothesis is that additional causal variation is present at the known GWAS loci but not fully quantified by individual GWAS markers [1], [2], [4]–[7]. This scenario may arise if the true causal variant is poorly tagged by any single GWAS marker [8] or if multiple independent causal variants exist at the locus [9]. In this case, the variance explained by the most-significant marker would only provide a lower bound on the local contribution, and some of the “missing” heritability would in fact be hidden at the previously discovered loci. If we consider “local” heritability to be the measure of aggregate variance from all causal variants at a locus, its quantification is an important step towards fully understanding the contributions made by association studies. Moreover, estimating components of local heritability indirectly from the vast amount of GWAS-level data already available would enrich our current understanding of complex disease architecture and provide insights into further study-design for post-GWAS fine-mapping studies. Here, we investigate methods for inferring components of local heritability at previously identified GWAS loci.
As study sample sizes continue to grow, researchers have focused on quantifying the amount of heritability explained by individually significant single-marker associations [4], [10]–[14]. In well-powered GWAS, one can also look for secondary variants that are conditionally independent of the leading SNP and estimate the joint contribution to phenotype. This conditional analysis has recently proven effective in GWAS for height [4], [15], [16] and multiple case-control traits [17], where a handful of loci were found to contain independent secondary associations. This strategy inherently focuses on a small number of independent markers and the outcome strongly depends on power to detect the primary association as well as any secondary variants. Such complexities make it difficult to compare this estimate across different studies and disease architectures. With additional resources, one can fine-map implicated loci using denser genotyping or sequencing platforms and look for more strongly significant markers. Recent studies involving re-sequencing around known GWAS-associated regions have identified additional variants explaining significant heritability in several complex traits [5], [18]–[20]. Looking beyond individual traits, a fine-mapping study of Celiac disease examined loci associated with other autoimmune diseases and nearly doubled the number of significant associations [21]. This approach can leverage the shared genetic architecture observed in some groups of related traits [22]–[24]. Still, such studies have not always yielded significant associations; a targeted re-sequencing analysis of Type 2 Diabetes did not yield any additional variants beyond what was known from GWAS [25] and recent work with dense genotyping did not uncover significant additional heritability at known loci for Type 2 Diabetes and Coronary Artery Disease [20]. Overall, these findings motivate methods that can infer components of additional local heritability using available GWAS data to guide fine-mapping analysis for identifying additional risk variants.
We propose to address this challenge by making use of all observed markers in a variance-component analysis, which optimizes a single measure of effect-size over a sample relatedness matrix. When sample relatedness is computed directly from the observed markers - referred to as the genetic relatedness matrix (GRM) - this variance-component can be used to infer the narrow-sense heritability explained by these markers. This measurement of narrow-sense heritability represents the aggregate effect of all causal variants observed or tagged in the data assuming additive, normally-distributed effect sizes. Recent work in variance-components analysis has shown that the contribution of all genotyped SNPs and any markers in LD with them, denoted , can be estimated directly from large-sample GWAS data in this way [26]–[29]. Similarly, our aim is to apply the variance-component model locally, by constructing the GRM from all typed SNPs at known GWAS regions and estimating the corresponding local . The excess of this quantity over the variance explained by known associations provides a lower bound on additional heritability at the locus. Uniquely, this method allows the analysis of loci that have no known association in the focal trait but have been associated with other related traits, quantifying sources of missing heritability implicated by shared disease architecture.
In this study, we apply these methods to both simulated and real phenotypes. Using simulations involving real genotypes, we find that LD between typed markers can significantly bias the estimate and propose a correction to the GRM calculation, which we compare to a recently proposed approach [30]. In local analysis, we observe higher estimates of heritability with the adjusted variance-component strategy compared to traditional association and conditional analysis, particularly when the locus harbors multiple causal variants. Importantly, our LD residual correction ensures these statistics are not inflated under the range of disease architectures considered (unlike the correction of [30]). We estimate local at known loci for nine common diseases finding a significant average increase vs. the variance explained by known associations, with individually significant increases for three of the traits. We also estimate local heritability at loci identified only in other related traits, showing significant enrichment in autoimmune disease for within-trait heritability at cross-trait loci. For RA, we analyze dense genotypes from samples typed on the ImmunoChip data as part of the Rheumatoid Arthritis Consortium International (RACI). This significantly larger sample-size and deep genotyping empowers us to provide precise estimates on the significant increases in local heritability within RA and across non-RA autoimmune traits. Our results have important implications for fine-mapping study-design as well as the broader understanding of disease architecture and allelic heterogeneity.
Our fundamental goal is to explain as much of the local heritability as possible without upward bias. We consider four different estimators with unique individual properties: , the variance explained by the single most associated SNP at a locus, computed directly from the effect-size of a univariate regression; , the variance explained by a conditional linear model of significant SNPs constructed by step-wise regression over all SNPs in the locus as described by [15]–[17]; , the heritability inferred with a standard variance-component constructed from all SNPs in the locus; and , the heritability inferred with an LD-residual adjusted variance-component constructed from all SNPs in the locus. The LD adjustment is crucial in scenarios where LD patterns that are systematically different at causal variants can distort the observed sample relatedness and bias traditional estimates of , as previously demonstrated by [30]. Our proposed correction uses linear regression to transform each SNP into an “LD residual” of any correlated preceding markers and construct the GRM from these residuals. We compare this correction to LDAK, the re-weighing solution of [30], as well as other strategies (see Methods).
We first analyze the genome-wide heritability explained by genotyped SNPs in the nine WTCCC1 and WTCCC2 traits (Table S1). Figure 1 shows the results of this analysis for unadjusted and LD-adjusted estimates performed over genotyped and genotyped+imputed SNPs (2.1 million 1000 Genomes [31] SNPs on average; see Table S1) separately. Results are shown on the observed scale. (Results on the liability scale are provided in Figure S4; all numerical values are provided in Table S14,S15.) We note that stringent quality-control is imperative for heritability analysis, where many small artifacts can compound into significant inflation of the genome-wide estimate [23]; this effect can be exacerbated by LD-adjustment methods, which will tend to promote low frequency variants that may be especially prone to QC issues. As in other studies [23], [30], we use a series of highly conservative QC filters to stem this problem, at the cost of filtering out many potentially informative markers (see Methods). The absence of any significant false heritability between the two control cohorts, particularly after LD-adjustment, indicates that genotyping artifacts are unlikely to be substantial (Figure 1). We note that in the presence of strong artifacts [30], propose an elegant solution of estimating SNP weighing scores from an independent population, and a similar strategy can be applied to the LD-residual adjustment.
For all traits we see that the LD-adjusted estimate from typed SNPs is higher than the corresponding unadjusted estimate, with an average of for genotyped SNPs. Previous work has shown the standard estimate to be robust when the trait is infinitesimal, i.e. where all SNPs are causal with normally distributed effect-sizes [32], [33]. However, as demonstrated in our simulations and in [30], non-infinitesimal traits with systematically less LD between rare and low-frequency variants will underrepresent those variants in the un-adjusted kinship, resulting in deflated estimates when a majority of the causal variants are low-frequency (Figure S1). The increase in adjusted estimates on real data therefore implies a genome-wide genetic architecture for these traits that is generally shifted towards low-frequency variants. As in our simulations, the effect of LD-adjustment is even stronger when imputed SNPs are included ( more on average, comparing dark-green to light-green bars), demonstrating the downwards bias introduced by an abundance of imputed markers without LD adjustment. Indeed, without adjustment, all of the traits exhibit lower after imputation. Interestingly, even though imputation increases the total number of markers by , the adjusted estimate from imputed SNPs is, on average, only higher than the corresponding estimate from genotyped SNPs. Because the LD adjustment effectively removes any new SNP that is a linear combination of nearby SNPs, this would be consistent with imputation providing information similar to such linear combinations [34]. This is further supported by the fact that the sum of LD-adjusted SNP variances (roughly corresponding to the independent number of SNPs) for imputed SNPs was only higher than that of typed SNPs. These findings do not minimize the utility of imputation for mapping, where individual effect sizes are important, but does imply that imputed variants are not explaining dramatically more missing heritability. Based on these findings and our previous simulations with imputed variants, we restrict our subsequent variance-components analysis to the genotyped data only.
Next, we infer the amount of local around the GWAS loci for the nine traits and compare to the corresponding and values (Figure 2, Table S16). When computing the increase in (and its statistical significance), we always account for and the local expectation, i.e. the increase that would be expected by chance based on the total genome-wide and the fraction of genome covered by the variance-component (see Methods). Across all the nine traits we find a consistent excess of local heritability, with an average increase of over the local expectation (combined ). These results were consistent with the LDAK-based adjustment, which had a mean increase of (Table S18). P-values were computed using a z-test and consistent to different definitions of (see Methods), but an analysis involving a comparison to random regions of the genome also produced similar results (see Methods, Table S17, Figure S5). Three of these traits (CD, UC, and MS) show individually significant increases (; ; and respectively). The regression-based analysis of jointly significant markers () yields an average of more heritability than . In instances where there are multiple known associations at a locus, only the leading SNP is included in but all of the known associated SNPs are automatically included in , demonstrating that previously known locus heterogeneity still does not explain as much heritability as the estimate. On average, these loci are explaining 11% of the genome-wide with 1.1% of the genome. Interestingly, the estimate with no LD-adjustment also yields increased local heritability for all phenotypes with an even higher average increase (Table S18). Given that our simulations show an increase in unadjusted estimates only when the underlying causal variant is common (Table 1), this increase in real data suggests that most causal variation in these GWAS loci originates from common causal variants (in contrast to the rest of the genome; see above).
The presence of significant additional heritability in individual traits raises the question of whether it is coming from a single poorly-tagged causal variant or multiple independent causal variants. In our previous simulations, an increase in local heritability is not expected under the single causal-variant model and the ratio of to has a direct relationship to the number of causal variants. For the WTCCC2 data, a single rare or common untyped causal variant is expected to yield an of and , respectively (Table 1 C,D). Both are lower than our observed average of in real data, and much lower than significant increases of and in UC and MS (Table S16). These results are therefore unlikely to arise simply due to all loci harboring a single poorly-tagged causal variant, with the point estimate of 1.29 indicating a likely architecture of 2–3 causal variants at the average locus. However, we caution that the variance of this ratio observed in simulations is very high (for example, 18% of the single common causal simulations have a local increase greater than 1.29), making it difficult to reject the single-causal variant hypothesis at this sample-size. From our previous power estimates (Table S13), we observe that at a sample-size of 15,000 power to detect multiple causal variants approaches 100%, allowing us to distinguish between these two scenarios.
We note that some of the GWAS loci we analyzed were genome-wide significant in the WTCCC data and could potentially exhibit inflated effect-sizes due to winner's curse if discovered in this cohort. However, because the heritability from variance-components and GWAS SNPs are inferred in the same data, we expect any effect-size inflation to impact both estimates equally, making our relative comparisons robust even in the presence of biases. In light of this and the small fraction of such loci actually present (8% averaged over the 7 WTCCC1 traits) we do not believe winner's curse to have had an impact on these results.
Recent analyses of multiple phenotypes have demonstrated significant correlations in genetic architecture for certain groups of related traits [23], [24], [35], [36]. Unique to the local variance-components approach, we can also compute components of heritability at known GWAS loci from multiple related traits without having genotypes for those traits. This measure provides an estimate of the additional variation that would be explained by fine-mapping loci associated with one trait within the affected samples of another; for example, analyzing known Ulcerative Colitis loci in a study of Crohn's Disease. We expect this to be informative when the traits have correlated genetic architectures, with causal variants that only reached statistical significance in one trait potentially explaining heritability in the other. One example of such related traits is the class of autoimmune disorders, which are known to have a shared disease architecture as well as many instances of overlapping GWAS loci [22], [37]–[40]. For each of the nine traits, we consider the amount of heritability explained by loci that were previously associated to one or more other autoimmune diseases but not to the focal trait. By definition, the for these loci is zero, and so we compare to the local expectation, i.e. what would be expected by chance from the genome-wide and locus size (see Methods). As with all other analyses, we specifically exclude the MHC for all autoimmune diseases so as to investigate the patterns of shared heritability outside of this well-studied region.
Figure 3 (numerical results in Table S19,S20) shows the results of this analysis, as well as the increase in heritability explained compared to the local expectation. The five autoimmune traits have the highest relative increases and are unique in being statistically significant. On average, the loci in the autoimmune traits explain more heritability than the local expectation (combined ), compared to more for the non-autoimmune traits (combined ). Both results were consistent with the LDAK-adjusted estimate of and respectively (Table S20). We again confirmed all significant z-test results using an empirical expectation by sampling random regions of the genome (see Methods, Table S17, Figure S5). Importantly, these results were not substantially different after accounting for increased heritability in coding regions, with the average increase after correction still significant at (see Methods, Table S21). We stress that these estimates specifically exclude any known loci for the respective disease; for example, the results from RA represent analysis of known autoimmune disease loci not identified in RA, and likewise for all of the other traits. As such, the additional heritability we identify would not have been found in a traditional targeted fine-mapping study that focuses only on trait-specific loci. Combining these results with the trait-specific analysis, we observe an average of more than at the union of autoimmune and disease-specific loci, individually significant across all the autoimmune traits (Table S22). On average, these loci are explaining 27% of the genome-wide . Most significant are the increases for MS and CD, with () and () more local , respectively.
Overall, we find that the class of autoimmune traits has a shared genetic architecture at known GWAS loci that can be leveraged to explain significant additional heritability. Loci found in one autoimmune trait are expected to harbor significantly more for other traits (beyond what is expected from lying near coding regions) and can therefore be important targets for fine-mapping analysis.
We estimate components of local heritability for Rheumatoid Arthritis in 23,092 samples of European origin typed on the ImmunoChip platform, recently analyzed for association by Eyre et al. [41]. The increased SNP density of this data is expected to provide higher power for local heritability analyses, and we again compare , , , and using simulated phenotypes from ImmunoChip genotypes (see Methods). We again observe an inflated and un-inflated , though the latter is more conservative than in previous simulations (Table S23). Overall, the higher density ImmunoChip results in a greater expected increase when considering all SNPs, particularly when variants are low-frequency.
We now consider real RA phenotypes. Of the 13 RA GWAS loci analyzed in the WTCCC1 data, 10 are also present on the ImmunoChip and we re-estimate local at this subset of 10 loci in both studies for comparison (Table 2A). The ImmunoChip data exhibits an increase in additional heritability explained over local expectation of (), compared to (non-significant at ) in the corresponding WTCCC1 loci. The ImmunoChip also exhibits a significant increase in heritability explained compared to and local expectation, with an increase of (). The ImmunoChip also contains 17 of the 24 non-RA autoimmune disease loci, also allowing us to perform the analysis of non-RA autoimmune loci. Again, we observe the local heritability to increase between the WTCCC1 and ImmunoChip data from 0.012 to 0.018, with the latter resulting in an increase of compared to local expectation (, Table 2B). Examining all relevant loci on the ImmunoChip, which are more likely to come from studies performed after the WTCCC, both local increases were lower but more significant due to the additional data analyzed.
For consistency, we have assumed the same total of 0.14 in both of the data-sets when computing the local heritability expected by chance, though this is likely an underestimate for the dense typing on the ImmunoChip. Likewise, the densely typed ImmunoChip sites also tag some markers outside of the variance-component region, effectively increasing the local expectation. Using 1,000 Genomes data, we find that a sequenced variant within 500 kbp of the studied regions is tagged with an average of 0.33 by the ImmunoChip sites in these loci, so we also consider a local expectation where each region is increased by of “flanking” length. However, irrespective of whether we use a total of 0.40 (the total estimated in previous studies excluding MHC [42]) and/or include the flanking regions, the local heritability identified at these loci remains strongly significant (Table S24). Overall, the ImmunoChip data shows local for RA at 27 known (RA+other) autoimmune loci to be 0.032, higher than that explained by the individual RA GWAS SNPs (0.006) and higher than the joint GWAS model (0.009).
The variance-component method allows us to estimate local at regions that are suggestive of harboring a secondary signal in this data. Specifically, Eyre et al. [41] analyzed these samples for conditional association and identified six loci that had a significant secondary signal. Predictably, when we restrict our analysis to these loci we confirm that the joint model increases heritability by over the associated SNP, but we also find the local to be even higher with a increase over the associated SNP and highly significant compared to local expectation (Table S25). Though the joint analysis has high power in this large cohort, the variance-components model still reveals additional hidden heritability. Similarly, Diogo et al. [43] fine-mapped 25 known RA loci and searched for the presence of secondary associations driven by variants in the protein-coding sequence of biological candidate genes, identifying strong enrichment of association at 10 coding variants (9 loci) but no individually significant variant. We examine these 9 loci in the ImmunoChip data and again observe an increase in heritability from the joint analysis of compared to the leading SNPs, but an even higher increase in local of which is more significant at than the permutation-based reported by Diogo et al. (Table S25).
Overall, the higher density and sample-size of the ImmunoChip data empowers us to identify the presence of significant additional at known RA loci as well as known non-RA autoimmune loci, beyond the heritability explained by standard mapping approaches analyzing the same data.
In this work we have sought to explain additional heritability at known GWAS loci by using large-sample SNP data. Specifically, we have utilized variance-components models that estimate the total contribution of all typed markers in the sample and do not require individual markers to be genome-wide significant. In applying these methods we have quantified biases in the standard estimate when the underlying disease architecture is non-infinitesimal and LD is systematically different at causal variants (as recently identified by [30]). To address this, we have proposed and compared several methods that seek to adjust the covariance matrix such that this correlation between markers is accounted for. In particular, we find the method of using LD residuals in computing the kinship to provide accurate estimates with no observed upward bias, in contrast to the proposed LDAK strategy [30] which yielded upward bias in our genome-wide simulations (though it exhibited lower mean error in imputed data). We thus recommend that the LD-residual approach be used in preference to LDAK when one is seeking lower bounds on the estimate of , as we are here.
Applying the LD-residual to known GWAS loci for nine WTCCC1 and WTCCC2 traits, we see that LD-adjusted estimates are nearly always higher than the unadjusted estimates, suggesting that the disease architecture is indeed shifted towards low-frequency variants for most traits. Understanding this phenomenon and applying and LD-adjustment method is therefore important for accurate estimation of in future studies. An alternative framework is the Bayesian sparse linear mixed model, which attempts to infer the underlying genetic architecture jointly with the and can provide more accurate estimates under certain disease architectures but requires significant computational resources (e.g. running time of 77 hours for a data set with 3,925 samples) [44].
Looking at previously known GWAS loci, we showed by simulation that the LD-residual adjusted variance-components approach is not inflated and can uncover additional heritability beyond that observed by the leading tag SNP, particularly when there are multiple underlying causal variants or tags. In analysis of nine dichotomous traits, we find a significant average increase in heritability explained of (combined ), with three traits exhibiting individually significant increases consistent with the presence of multiple causal variants on average. The latter finding is supported by previous work showing that loci with a single causal variant are unlikely to explain substantially more heritability then the GWAS SNP and hypothesizing multiple underlying causal variants [8]. However, though our simulations show that increased heritability is an indicator of multiple causal variants on average, the current sample size is not sufficient to reject the possibility that this local increase is caused by a single causal variant being poorly tagged by the leading GWAS SNP. We extrapolate that as sample sizes reach the tens of thousands our method can conclusively draw distinctions between these two scenarios.
Because the LD-unadjusted method tends to be deflated when the underlying causal variant is low-frequency (Table 1), we can use the unadjusted estimate as an indicator of the causal allele frequency. The fact that all but one of these traits exhibit an unadjusted local that is higher than the strongly suggests that the bulk of causal variation at these known loci does not lie in low-frequency variants. This is consistent with the recent findings of Hunt et al. [45] in a large-scale sequencing study that demonstrated minimal rare-variant heritability for 25 known auto-immune disease risk genes. This is in contrast to our genome-wide analysis that yielded additional heritability after LD-adjustment, indicative of a shift toward low-frequency markers. Taken together, we hypothesize that the causal frequency spectrum at these known loci is substantially different from that of the rest of the genome. In light of this finding, we caution against extrapolating the genome-wide disease architecture from known GWAS loci, as done in Hunt et al. and other studies [45]–[48].
We also applied this technique to loci that have been discovered in related traits but not in the focal trait. Additional variation would be found in instances where causal loci are shared across multiple traits but have only been mapped in one trait, allowing us to estimate the efficacy of a fine-mapping study design incorporating these loci. For autoimmune diseases we see a significant amount of excess heritability at such related-trait loci with an average of more than expected by chance. Relative to the known , the greatest increase from the union of trait-specific and related-trait loci is observed in MS () and CD (). This finding is substantiated by the fact that non-autoimmune traits exhibit no such significant increase and serve as negative controls. Where previous studies have documented overlap between causal variants from autoimmune disease [22], [40], we show that this is a wide-spread phenomenon expected to account for an average of 27% of total over five auto-immune traits. Our analysis is complementary to recent methods that construct multivariate variance-components models which directly estimate the genetic correlation between multiple traits [23], [24]. In contrast to those studies, our approach requires only the genetic information from a single trait of interest, allowing us to analyze components of heritability between many autoimmune traits without having their genetic data. Looking forward, this strategy can be used to analyze other classes of related phenotypes such as metabolic traits [24] and psychiatric disorders [36]. Given that we observe GWAS loci to have fundamentally different disease architectures from the rest of the genome, our method will still not capture the genome-wide correlation between the two traits. A potential future application is local heritability analysis with the multivariate variance-components model, merging these two strategies.
For RA, we repeated our analysis in a much larger cohort typed on the ImmunoChip and found significant additional heritability. Where the GWAS analysis of this data by Eyre et al. [41] found 6/45 loci containing a secondary marker, we quantify the overall amount of additional heritability to be than . While Eyre et al. identified a significant correlation between their associated loci and genes with auto-immune function, we additionally observe more heritability than expected by chance in non-RA auto-immune loci (Table 2), a highly significant increase. These findings demonstrate the effectiveness of our method in quantifying components of heritability from high-density data. Loci from the other traits we examined have also recently been analyzed large fine-mapping studies. Jostins et al. [40] found that 30/163 loci associated with Crohn's Disease or Ulcerative Colitis exhibit significant secondary effects, and all loci have an higher chance of being associated with immune-function genes. Likewise, we observe significant local and related-trait heritability for Crohn's Disease. On the other hand, Shea et al. [25] re-sequenced one locus for T2D and Maller et al. [20] densely genotyped 11 loci for CAD and T2D, with neither study identifying significantly more heritability. This too is consistent with our failure to observe significant increases in heritability for these traits, though both sets of negative results may be due to the small number of loci and samples examined.
Two recent publications by Ehret et al. and Ke [16], [17] propose methods to quantify the amount of recoverable heritability at known loci by selecting a conditional linear model. The conceptual distinction between these methods and our approach is that they explicitly focus on a pruned and p-value restricted set of markers and are therefore limited by power to detect association within the analyzed sample. The Ke strategy differs from that of Ehret et al. in the specific threshold values and that it does not depend on an external set of samples for estimating unbiased effects; as such, it is likely to be the less conservative estimate of local heritability and the one we selected for comparison. Because these strategies only focus on loci where conditionally nominal SNPs are present, they do not provide a complete analysis of all known loci together. While it is possible to incorporate many more SNPs into a complex multiple regression and estimate the total fraction of phenotypic variance explained, this estimate will be highly biased proportional to the effective number of SNPs divided by the effective number of samples, a difficult ratio to quantify in the presence of LD between SNPs and sample structure. On the other hand, the local variance-components model provides an approximately unbiased estimate of the total heritability explained by all SNPs, allowing us estimate components from putative loci without significant associations, as we do here with related traits. Both in simulations and in real data, we find that our strategy identifies more additional variation than the standard linear model.
One limitation of the current variance-components strategy is that analysis of ascertained case-control traits can lead to underestimates of when the ratio of SNPs to samples is low (A.L.P., unpublished data), as can be the case when analyzing a small number of loci. This would lower the power to detect significant additional heritability and yield local estimates that are a conservative lower bound. Quantifying and correcting for this phenomena in case-control traits is an important area of future study. Other future directions for this work include the estimation of local heritability over more complex annotations of putative regions [49] as well as the use of local heritability for mapping previously unknown loci akin to group-wise tests [50], [51].
The torrent of large-scale sequencing studies will do much to inform our understanding of the genetic architecture of common diseases, but the design of such studies also motivates the inference of disease architecture from currently available data. The strategies outlined here demonstrate a great diversity of allelic heterogeneity within and between traits, informing our assumptions for future GWAS and fine-mapping analysis.
We examined data from the Wellcome Trust Case Control Consortium (WTCCC) versions 1 and 2. These datasets have been outlined in [13] and [12], [40], and we provide summary details in Table S1. Unlike GWAS studies, heritability estimates can be particularly sensitive to individually small artifacts/batch-effects [52], which can add up over many SNPs to exhibit false heritability [29]. To account for this, we apply several additional layers of quality control.
We also examined 23,092 samples of European origin typed on the ImmunoChip platform (32% cases for Rhematoid Arthritis), recently analyzed for association by Eyre et al. [41]. For this data, we followed the QC protocol of Eyre et al. [41] and also excluded any SNPs below 1% allele frequency.
The variance-components model assumes an idealized infinitesimal genetic architecture where every marker is causal and effect-sizes are normally distributed over the normalized variants. [33] showed that the model remains unbiased when causal variants are randomly sampled from the typed SNPs (though the analytical standard error on the estimate does exhibit bias as the number of causal variants becomes very low [30]). However, as demonstrated in [33], when causal variants are not randomly drawn from the typed SNPs, LD between markers can lead to over-representation of certain SNPs in the sample GRM and distort the estimated relationships between individuals, thereby distorting the final estimate of SNP-heritability. We describe and evaluate several methods that account for correlations between markers when constructing a GRM. In all cases, the goal is to reweigh or transform each SNP so that it is equally represented in a new adjusted genotype matrix. We caution that our simulations do not explore the robustness of this model in the presence of very rare variants (e.g. whole-genome sequence) where assumptions of normality may be strongly violated.
Open-source software implementing the LD residual adjustment we have described is implemented in EIGENSOFT 5.0 at http://www.hsph.harvard.edu/alkes-price/software. HAPI-UR software is available at https://code.google.com/p/hapi-ur/ GCTA software is available at http://www.complextraitgenomics.com/software/gcta/
|
10.1371/journal.pbio.0050232 | ISWI Regulates Higher-Order Chromatin Structure and Histone H1 Assembly In Vivo | Imitation SWI (ISWI) and other ATP-dependent chromatin-remodeling factors play key roles in transcription and other processes by altering the structure and positioning of nucleosomes. Recent studies have also implicated ISWI in the regulation of higher-order chromatin structure, but its role in this process remains poorly understood. To clarify the role of ISWI in vivo, we examined defects in chromosome structure and gene expression resulting from the loss of Iswi function in Drosophila. Consistent with a broad role in transcriptional regulation, the expression of a large number of genes is altered in Iswi mutant larvae. The expression of a dominant-negative form of ISWI leads to dramatic alterations in higher-order chromatin structure, including the apparent decondensation of both mitotic and polytene chromosomes. The loss of ISWI function does not cause obvious defects in nucleosome assembly, but results in a significant reduction in the level of histone H1 associated with chromatin in vivo. These findings suggest that ISWI plays a global role in chromatin compaction in vivo by promoting the association of the linker histone H1 with chromatin.
| Chromatin-remodeling factors such as ISWI play a role in transcription and other nuclear processes by altering the structure and positioning of nucleosomes (the protein–DNA complexes that organize chromatin). Recent studies have suggested that chromatin-remodeling factors can also influence higher-order chromatin structure, but how they do this is not well understood. Using Drosophila melanogaster as a model organism, we investigated the role of ISWI in gene expression and the regulation of chromosome structure in higher eukaryotes. Loss of ISWI alters the expression of a large number of genes. The loss of ISWI function also causes dramatic alterations in higher-order chromatin structure—including the decondensation of mitotic and polytene chromosomes—accompanied by a striking reduction in the amount of the linker histone H1 associated with chromatin. Based on these findings, we propose that ISWI plays a global role in chromosome compaction by promoting the association of a linker histone with chromatin.
| The packaging of DNA into chromatin is critical for the organization and expression of eukaryotic genomes. The basic unit of chromatin structure—the nucleosome—can repress transcription by blocking the access of transcription factors and other regulatory proteins to DNA [1]. Interactions between nucleosomes lead to the formation of 30-nm fibers, which can be further packaged into increasingly compact structures [2–4]. The regulation of higher-order chromatin structure is critical for chromosome condensation and segregation during mitosis and meiosis [5,6]. A growing body of evidence suggests that chromatin folding or looping is also important for the regulation of enhancer–promoter interactions and the subdivision of chromosomes into discrete functional domains [7]. The molecular mechanisms used to regulate chromatin structure have therefore been the topic of extensive study.
The repressive effect of chromatin on transcription is regulated via two general mechanisms: ATP-dependent chromatin remodeling and the covalent modification of histones. Chromatin remodeling reactions—including changes in the structure and spacing of nucleosomes—are catalyzed by ATPases that often function as subunits of large complexes, including the SWI/SNF, Imitation SWI (ISWI), and CHD complexes [8]. Histone-modifying enzymes alter the acetylation, methylation, phosphorylation, or ubiquitinylation of N-terminal histone tails and other regions on the surface of the nucleosome; these modifications modulate interactions between nucleosomes and a wide variety of structural and regulatory proteins [9]. Both histone-modifying enzymes and chromatin-remodeling complexes can be targeted to specific promoters by gene-specific or general transcription factors. By locally altering the structure or positioning of nucleosomes, histone-modifying enzymes and chromatin-remodeling complexes can activate or repress transcription of specific genes [1]. The coordinated activities of histone-modifying and chromatin-remodeling enzymes are therefore critical for transcription in a chromatin environment.
Interactions between histone-modifying and chromatin-remodeling enzymes can have profound effects on higher-order chromatin structure [10], as illustrated by recent studies of the Drosophila chromatin-remodeling factor ISWI. ISWI functions as the ATPase subunit of at least three distinct chromatin-remodeling complexes: ACF, CHRAC, and NURF [11,12]. These complexes use the energy of ATP hydrolysis to slide nucleosomes and alter the spacing of nucleosome arrays [12]. The loss of Iswi function in the larval salivary gland leads to the dramatic decondensation of a specific chromosome: the male X [13]. Similar defects in the structure of the male X chromosome are caused by loss-of-function mutations in E(bx), the gene encoding the largest subunit of NURF (NURF301) [14]. These findings suggest that ISWI plays a relatively global role in chromatin compaction in vivo.
The unusual sensitivity of the male X chromosome to the loss of ISWI function suggests that changes in chromatin structure that accompany dosage compensation might regulate the ability of ISWI to remodel chromatin in vivo. In Drosophila, dosage compensation is dependent on an RNA–protein complex that contains the MOF histone acetyltransferase [15,16]. This complex is specifically targeted to the male X chromosome, leading to the widespread acetylation of lysine 16 of the histone H4 tail (H4K16) by MOF [17]. This covalent modification is restricted to the male X chromosome and is thought to “open” chromatin structure by disrupting interactions between adjacent nucleosomes, thereby leading to the increased transcription of X-linked genes in males [18,19]. Dosage compensation is necessary and sufficient for the chromosome defects observed in Iswi mutant larvae, and genetic interactions between ISWI and MOF are consistent with functional antagonism between the two proteins [20]. Furthermore, biochemical studies have shown that the acetylation of H4K16 inhibits interactions between ISWI and its nucleosomal substrate in vitro [19–22]. Based on these observations, it has been proposed that the acetylation of H4K16 regulates chromatin compaction mediated by ISWI [19,20].
Although tremendous progress has been made toward understanding how ISWI alters the structure and positioning of nucleosomes, relatively little is known about how it alters higher-order chromatin structure and whether this activity is used to regulate transcription in a chromatin environment. To address these issues, we characterized defects in chromatin structure and gene expression resulting from the loss of Iswi function in Drosophila. Here we report that ISWI plays a surprisingly global role in the regulation of higher-order chromatin structure and transcription in vivo. Loss of ISWI function leads to widespread changes in gene expression, including the derepression of numerous genes. Defects in chromosome structure resulting from the loss of ISWI function are accompanied by a dramatic reduction in the level of histone H1 associated with chromatin. These findings suggest that ISWI plays a global role in chromatin compaction in vivo by promoting the association of histone H1 with chromatin.
Although genetic studies have shown that ISWI regulates the structure of the male X chromosome, its role in this process has remained unclear. ISWI might specifically regulate the structure of the male X chromosome, perhaps by dampening the effect of H4K16 acetylation on higher-order chromatin structure. In support of this view, the acetylation of H4K16 by the dosage compensation machinery is both necessary and sufficient for the chromosome defects observed in Iswi mutant larvae [20]. Alternatively, ISWI might regulate the structure of all chromosomes, with the male X chromosome being exquisitely sensitive to the loss of ISWI function. This sensitivity could be due to negative regulation of ISWI function by H4K16 acetylation, an idea that is supported by in vitro studies demonstrating that H4K16 acetylation blocks interactions between ISWI and its nucleosomal substrate [19,20]. To distinguish between the two models, we further investigated chromosome defects resulting from the loss of ISWI function in vivo.
Like many Drosophila genes, Iswi is expressed at high levels both maternally and zygotically; we therefore suspected that the high maternal contribution of Iswi gene products to the unfertilized egg might mask phenotypes resulting from the loss of zygotic Iswi function [13]. Our previous genetic studies of ISWI relied on two null alleles, Iswi1 and Iswi2 [13,20]. Individuals that are homozygous or trans-heterozygous for these alleles survive until late larval development, presumably due to the high maternal contribution of ISWI [13]. Consistent with this possibility, we detected low levels of ISWI in the salivary glands of Iswi1/Iswi2 larvae by protein blotting (data not shown). Thus, the striking defects in the structure of the male X chromosome observed in these individuals [13] (Figure 1) did not reflect the true null phenotype of Iswi mutations. We therefore examined the consequences of further reducing Iswi function by producing a dominant-negative form of the ISWI protein, ISWIK159R.
The IswiK159R mutation eliminates the ATPase activity of ISWI without affecting its ability to interact with other proteins; as a result, the expression of ISWIK159R has very strong dominant-negative effects on ISWI function in vivo [13]. The expression of a GAL4-responsive IswiK159R transgene (UAS-IswiK159R) in salivary gland nuclei under the control of an ey-GAL4 driver led to the dramatic decondensation of the X chromosome and the autosomes in both sexes (Figure 1F and data not shown). By contrast, the production of a variety of other proteins, including ISWI, GFP, and a dominant-negative form of another chromatin-remodeling factor (BRMK804R) did not cause similar defects in chromosome structure in either sex (Figure 1E and data not shown). These findings suggested that ISWI plays a global role in the regulation of polytene chromosome structure, as opposed to acting specifically on the male X chromosome.
Although salivary gland polytene chromosomes provide a useful model for studying interphase chromosomes, their structure is unusual in many respects. Unlike chromosomes in diploid cells, the structure of polytene chromosomes is dependent on physical interactions between hundreds of sister chromatids formed by DNA replication in the absence of cell division [23]. This led us to wonder whether ISWI affects an aspect of chromatin organization unique to polytene chromosomes, as opposed to a more general aspect of higher-order chromatin structure. To investigate this issue, we examined whether the loss of ISWI function alters the structure of chromosomes in larval neuroblasts, a diploid cell type that is particularly well suited for cytological studies [24]. To allow observation of fully condensed chromosomes, neuroblasts were arrested in metaphase with colchicine. Mitotic chromosomes prepared from neuroblasts of wild-type and Iswi1/Iswi2 third-instar larvae were indistinguishable (Figure 2A and 2B). Again, we suspected that the lack of a discernable phenotype might be due to the high maternal contribution of ISWI, since low levels of ISWI can be detected in extracts of neuroblasts by protein blotting (data not shown). We therefore expressed ISWIK159R in neuroblasts of third-instar larvae using a da-GAL4 transgene.
The expression of ISWIK159R in larval neuroblasts dramatically altered the structure of metaphase chromosomes (Figure 2C–2E). Chromatin appeared highly decondensed and disorganized, and sister chromatids were often indistinguishable from each other. Loss of Iswi function most strongly affected euchromatic regions, which appeared hazy and diffuse. The effect on heterochromatic regions was much less striking, as evidenced by the relatively normal condensation of the largely heterochromatic fourth and Y chromosomes (Figure 2C–2E). Similar phenotypes were not observed in larvae expressing a variety of control proteins, including GFP, ISWI, and BRMK804R (Figure 2F and unpublished data). The chromosome defects resulting from the expression of ISWIK159R were highly penetrant. Thus, ISWI is required for the generation or maintenance of higher-order chromatin structure in both polytene and diploid cells.
Based on the known biochemical activities of ISWI, it is tempting to speculate that it regulates higher-order chromatin structure by altering the spacing or fluidity of nucleosome arrays. However, the chromosome defects observed in Iswi mutants might be a secondary consequence of changes in gene expression resulting from the loss of ISWI function. This is not a trivial concern, since salivary gland and neuroblast chromosomes cannot be examined until relatively late stages of larval development, approximately 5 d after fertilization. To help exclude this possibility, we examined whether the expression of ISWIK159R alters chromosome structure in the early embryo. To minimize the maternal contribution of ISWI, we expressed ISWIK159R in embryos produced by heterozygous Iswi2 females at 25 °C. The GAL4 system used to drive the expression of ISWIK159R is inherently temperature sensitive; much stronger phenotypes are observed at 25 °C than at 18 °C [25]. At 18 °C, individuals expressing ISWIK159R under the control of a da-GAL4 driver survive until late larval stages, but at 25 °C they fail to complete embryogenesis (data not shown). Defects in chromosome structure resulting from the loss of Iswi function were examined by staining fixed embryos with DAPI and antibodies against α-tubulin. Defects in chromosome condensation and the organization of the mitotic spindle were observed as early as nuclear cycle 12, shortly after the onset of zygotic transcription (Figure 2). Similar defects were not observed in control embryos expressing a variety of control proteins, including BRMK804R (data not shown). These findings suggest that ISWI directly regulates higher-order chromatin structure.
ISWI is associated with hundreds of euchromatic sites of Drosophila polytene chromosomes in a pattern that is largely complementary to that of RNA polymerase II (Pol II) [13] (Figure 3A). This bias is even more pronounced when the distribution of ISWI is compared to that of the elongating form of Pol II (Pol IIoser2) (Figure 3B). To determine if ISWI plays a global role in transcriptional repression, we monitored changes in gene expression resulting from the loss of Iswi function in male and female third-instar larvae using whole-genome microarrays. Of nearly 15,000 genes analyzed, the expression of approximately 500 changed at least 2-fold in mutant male or female larvae relative to wild-type (Tables 1 and S1). Consistent with a predominant role for ISWI in transcriptional repression, nearly 75% of the genes are expressed at higher levels in Iswi mutants. The genes affected in Iswi mutants appear to be randomly distributed between the X chromosome and autosomes in both sexes and represent a functionally diverse group of genes with no obvious common properties or involvement in the regulation of higher-order chromatin structure (Tables 1 and S1).
To address whether ISWI directly regulates the genes identified in our microarray studies, we examined whether it is physically associated with 16 of its potential target genes in salivary gland chromosomes. After staining polytene chromosomes from wild-type larvae with antibodies against ISWI and the elongating form of Pol II (Pol IIoser2), the positions of potential target genes were identified by fluorescent in situ hybridization (FISH; Figure 3C and 3D). For these experiments, we selected genes that exhibited at least a 2.5-fold change in gene expression in Iswi mutant versus wild-type male larvae. ISWI was not associated with five of seven genes that are expressed at reduced levels in the salivary glands of Iswi mutant larvae (Figure 3C and unpublished data), suggesting that ISWI may indirectly activate their expression. By contrast, for eight of nine potential targets of repression examined, ISWI co-localized with the FISH signal, suggesting that it may directly repress the transcription of these genes (Figure 3D and unpublished data). These findings—together with the results of our microarray experiments and the preferential association of ISWI with weakly transcribed or silent regions of chromatin—is consistent with a predominant role for ISWI in transcriptional repression.
The changes in gene expression observed in Iswi mutants could be a consequence of global changes in chromatin compaction that increase the access of transcription factors or RNA polymerase to the DNA template. If this hypothesis is correct, one would expect that the loss of Iswi function in the salivary gland of male larvae would have a greater effect on the expression of X-linked genes than autosomal genes. We therefore examined the effect of Iswi mutations on gene expression in the salivary glands of third-instar larvae. In the salivary glands of Iswi1/Iswi2 males, the expression of 339 genes changed at least 2-fold relative to wild-type; 76% of these genes were expressed at higher levels because of the loss of ISWI function (Tables 1 and S2). The potential targets of ISWI regulation in the salivary gland appear to be randomly distributed between the X chromosome and autosomes (Tables 1 and S2; Figure 4). Furthermore, the magnitude of the changes in the expression of potential ISWI target genes does not vary significantly as a function of their chromosomal locations (Figure 4; Table S2). Thus, the dramatic alterations in the structure of the male X chromosome in Iswi mutants are not accompanied by similarly dramatic changes in the expression of X-linked genes.
We next investigated the molecular basis of the chromosome defects resulting from the loss of ISWI function. ISWI and MOF (the histone acetyltransferase that acetylates H4K16) have opposite effects on chromatin structure, and genetic studies have revealed a strong functional antagonism between the two proteins [20]. Based on these observations, we suspected that ISWI might promote chromatin compaction by blocking the acetylation of H4K16. To investigate this possibility, we stained polytene chromosomes of wild-type and Iswi mutant larvae with antibodies that specifically recognize acetylated H4K16. The levels of H4K16 acetylation on the male X chromosomes of wild-type and Iswi mutant larvae appear similar (Figure 5A and 5B) [13]. Furthermore, the chromosome defects observed in larvae expressing ISWIK159R are not due to the spread of H4K16 acetylation to the autosomes (Figure 5C). These findings suggest that ISWI can promote the formation of higher-order chromatin structure independently of H4K16 acetylation.
Biochemical studies have suggested several other mechanisms by which ISWI might regulate higher-order chromatin structure. For example, ISWI has been implicated in nucleosome assembly—a prerequisite for the formation of higher-order chromatin structure [26–28]. ISWI could also influence the packaging of nucleosome arrays by altering their spacing or fluidity. To determine if the loss of ISWI function causes gross defects in chromatin assembly, we stained wild-type and Iswi mutant larvae with a monoclonal antibody, MAB052, that has been reported to recognize the core histones H2A, H2B, H3, and H4 and the linker histone H1. As expected for a “pan-histone” antibody, MAB052 stained the X chromosome and autosomes of wild-type larvae in a pattern similar to that of DAPI-stained DNA (Figure 6A). In Iswi1/Iswi2 larvae, MAB052 staining of the male X chromosome, but not the female X chromosome or autosomes of either sex, was dramatically reduced (Figure 6A). Furthermore, the chromosome defects resulting from the expression of ISWIK159R were accompanied by loss of MAB052 staining from the autosomes as well as the X chromosome (Figure 6A). Similar results were obtained using a variety of fixation techniques, including treatment with either formaldehyde or citric acid (data not shown). The strong correlation between the loss of MAB052 staining and the severity of the chromosome defects observed in Iswi mutant larvae suggested that ISWI might indirectly regulate higher-order chromatin structure by promoting nucleosome assembly.
As an alternative approach for examining potential defects in nucleosome assembly resulting from the loss of ISWI function, we analyzed chromatin isolated from salivary glands of wild-type larvae and larvae expressing ISWIK159R by partial digestion with micrococcal nuclease. As described above, the expression of ISWIK159R in salivary gland nuclei leads to the decondensation of all chromosomes in both sexes. However, these defects are not accompanied by the appearance of subnucleosomal fragments characteristic of nucleosomes lacking histone H2A/B dimers or other core histones (Figure 6B). Furthermore, the loss of ISWI function did not lead to obvious changes in either the sensitivity of chromatin to micrococcal nuclease digestion or nucleosome spacing (Figure 6B and data not shown). Consistent with these observations, the decondensation of the male X chromosome in Iswi1/Iswi2 larvae is not accompanied by decreased chromosomal levels of several core histones, including histone H4 acetylated on K16 (Figure 5), histone H3 trimethylated on K4 or K27 (Figure S1), and histone H2AvD (data not shown). These observations suggest that the chromosome defects observed in Iswi mutants do not result from gross defects in nucleosome assembly.
To reconcile the seemingly contradictory results obtained using the above assays, we investigated the specificity of the supposedly pan-histone MAB052 antibody by protein blotting. Interestingly, we found that MAB052 has a strong preference for purified histone H1 (Figure 6C). The antibody only weakly detects histone H3 on a protein blot and does not recognize other recombinant core histones (Figure 6C). This strong preference for histone H1 was even more pronounced when extracts of Drosophila embryos and salivary glands were assayed by protein blotting using MAB052 (Figure 6C). We therefore repeated the above staining experiments using a polyclonal antibody specific for histone H1 (Figure 7). The loss of ISWI function does not significantly alter the total level of histone H1 in salivary glands as assayed by protein blotting (data not shown). However, as expected based on the specificity of MAB052, we found that histone H1 levels are dramatically reduced on the X chromosome of Iswi1/Iswi2 males, but not the autosomes (Figure 7B). Furthermore, the expression of dominant-negative ISWIK159R protein in salivary gland nuclei led to a dramatic reduction of histone H1 levels on all chromosomes of both sexes (Figure 7D and data not shown). In all cases, we observed an excellent correlation between the severity of the chromosome defects observed and the loss of histone H1 staining.
To confirm the above findings, we used a biochemical assay to determine whether the loss of ISWI function reduces the level of histone H1 associated with chromatin. Salivary glands were dissected from wild-type larvae and larvae expressing the ISWIK159R protein and fixed with formaldehyde. Following chromatin extraction and the reversal of cross-linking, the levels of histone H3 and histone H1 associated with chromatin were assayed by protein blotting. As expected, the loss of ISWI function led to a significant and reproducible reduction in the ratio of histone H1 to histone H3 associated with chromatin (Figure 7E). These findings strongly suggest that defects in higher-order chromatin structure resulting from the loss of ISWI function are due to a failure to efficiently incorporate histone H1 into chromatin in vivo.
Most studies of ISWI complexes in Drosophila and other organisms have focused on their ability to alter the structure or spacing of nucleosomes, the fundamental unit of chromatin structure. Our findings reveal that ISWI also plays a global role in the regulation of higher-order chromatin structure. The Iswi mutations used in this study eliminate the function of multiple chromatin-remodeling complexes, including ACF, NURF, and CHRAC [11]. Which of these complexes are required for the formation of higher-order chromatin structure? Loss of function mutations in Acf1—which encodes a subunit protein shared by ACF and CHRAC—do not cause obvious defects in higher-order chromatin structure [27]. By contrast, loss of function mutations in E(bx)—which encodes a subunit specific to NURF—cause male X chromosome defects similar to those observed in Iswi mutants [14]. These findings suggest that ISWI modulates higher-order chromatin structure within the context of NURF, as opposed to ACF or CHRAC.
We observed a striking correlation between the severity of the chromosome defects resulting from the loss of ISWI function and the loss of the linker histone H1. This correlation suggests that ISWI regulates higher-order chromatin structure by promoting the association of histone H1 with chromatin. Histone H1 and other linker histones influence higher-order chromatin structure in vitro by stabilizing interactions between nucleosomes and chromatin fibers [29]. Although the ability of histone H1 to promote chromatin compaction in vitro is well established, its function in vivo has been a topic of considerable debate [30,31]. A protein with biochemical properties reminiscent of linker histones—HHO1—is present in budding yeast; surprisingly, HHO1 is not essential for viability in yeast, and hho1 mutations have little effect on either gene expression or chromatin structure [32,33]. Genetic studies in Tetrahymena have suggested roles for linker histones in chromatin condensation and gene expression [34,35], but the relevance of these studies to histone H1 function in higher eukaryotes remains unclear. Studies of histone H1 function in higher eukaryotes have been complicated by the presence of redundant genes encoding histone H1 or histone H1 subtypes [36]. In spite of these difficulties, recent studies have revealed important roles for histone H1 in chromosome compaction in Xenopus and mice [37–40]. Thus, the chromosome defects observed in Iswi mutants could easily result from inefficient incorporation of histone H1 into chromatin.
How might ISWI promote the association of histone H1 with chromatin? Since ISWI is not required for histone H1 synthesis, ISWI may directly promote the assembly of chromatin containing histone H1 following DNA replication. Recent biochemical studies provide support for this possibility: ACF promotes the ATP-dependent assembly of H1-containing chromatin in vitro [26]. Loss of ACF1 function does not cause obvious changes in chromosome structure, however, suggesting that ACF either does not regulate higher-order chromatin structure in vivo or plays a redundant role in this process [27]. It remains possible that ISWI promotes the assembly of histone-H1-containing chromatin within the context of NURF or another chromatin-remodeling complex.
The ability to promote histone H1 assembly is not a common property of all chromatin-remodeling factors, as illustrated by recent biochemical studies of CHD1 [26]. Like ACF and other ISWI complexes, the CHD1 ATPase promotes the assembly of regularly spaced nucleosomes in vitro [26]. By contrast, CHD1 does not promote the incorporation of histone H1 during chromatin assembly in vitro [26]. These biochemical studies provide a plausible explanation for why the loss of ISWI function leads to the loss of histone H1 without causing dramatic changes in nucleosome assembly in vivo.
In other organisms, depletion of histone H1 leads to a significant decrease in the nucleosome repeat length [29], presumably because of the failure to efficiently incorporate histone H1 during replication-coupled chromatin assembly. By contrast, the loss of ISWI function in salivary gland nuclei leads to a decrease in the amount of histone H1 associated with chromatin without causing dramatic changes in nucleosome repeat length (Figure 6B). It is therefore tempting to speculate that ISWI promotes histone H1 incorporation via a replication-independent process. The association of histone H1 with chromatin is far less stable than that of core histones; histone H1 undergoes dynamic, global exchange throughout the cell cycle [41]. Photobleaching experiments in Tetrahymena and vertebrates have suggested that the majority of histone H1 molecules associated with chromatin are exchanged every few minutes [37,42–44], but little is known about the factors that regulate this process. Based on our findings, ISWI is an excellent candidate for a factor that regulates the dynamic exchange of histone H1 in vivo. Further work will be necessary to determine whether ISWI promotes histone H1 incorporation via replication-dependent or -independent mechanisms.
Our findings, together with previous studies, suggest that acetylation of H4K16 may regulate the association of linker histones with chromatin in vivo. The histone H4 tail is required for the nucleosome-stimulated ATPase activity of ISWI, and for its ability to slide nucleosomes and alter their spacing in vitro [21,22,45–47]. The region of the H4 tail that is critical for ISWI function in vitro is a DNA-bound basic patch (R17H18R19) adjacent to H4K16, the residue that is acetylated by the MOF histone acetyltransferase [21,22,47]. The acetylation of H4K16 interferes with the ability of ISWI to interact with the histone H4 tail and alter the spacing of nucleosome arrays in vitro [19,20,47]. Consistent with these findings, dosage compensation is necessary and sufficient for the decondensation of the X chromosome in Iswi mutant larvae, and genetic studies have revealed a strong functional antagonism between ISWI and MOF [20]. Thus, H4K16 acetylation may function as a switch that regulates the histone H1 assembly mediated by ISWI.
Our microarray studies revealed that ISWI is required for the proper expression of a large number of genes. These findings are consistent with numerous studies implicating ISWI in transcriptional regulation in vitro and in vivo [11,48,49]. Does ISWI modulate transcription by altering higher-order chromatin structure? We suspect that ISWI regulates transcription and higher-order chromatin structure via distinct mechanisms, since we observed no obvious correlation between the magnitude of the changes in gene expression and chromosome structure observed in Iswi mutant larvae. This is consistent with genetic studies in other organisms that have revealed that the loss of histone H1 does not cause dramatic changes in gene expression [33,34,37,40]. We also failed to observe a correlation between the magnitude of transcriptional derepression and gene size in Iswi mutant larvae (data not shown), as would be expected if ISWI relieved a general block to transcriptional elongation by Pol II. It should be noted, however, that relatively subtle, but biologically important, changes in gene expression may have escaped detection in our microarray studies. Further work will be necessary to clarify this issue and to determine whether ISWI regulates transcription and higher-order chromatin structure via distinct or related mechanisms.
Flies were raised on cornmeal, agar, yeast, and molasses medium, supplemented with methyl paraben and propionic acid. Oregon R was used as the wild-type strain in all experiments unless otherwise noted. Drosophila strains were obtained from the Bloomington Drosophila Stock Center (http://flystocks.bio.indiana.edu/) and are described in FlyBase (http://www.flybase.org/), unless otherwise noted. The Iswi mutations used in this study are described elsewhere [13].
To generate larvae lacking zygotic Iswi function, y w; Iswi2 sp; +/T(2;3) B3 CyO, TM6B, Tb females were crossed to Iswi1 Bc/SM5, Cy sp males. The transheterozygous Iswi progeny resulting from this cross are referred to as Iswi1/Iswi2 elsewhere in the text.
The GAL4 system [25] was used to drive the expression of ISWIK159R and other proteins in the salivary gland and other tissues. Two GAL4 driver lines were used in this study: da-GAL4 [50] and ey-GAL4 [51]. ey-GAL4 drives expression in the larval salivary glands and eye-antennal disc; da-GAL4 is widely expressed at all stages of development. The UAS lines used in this study include (1) Df(1)w67c2 y; P[w+, UAS-IswiK159R-HA-6His] 11–4/TM6B, P[w+ Ubi-GFP], Tb (referred to as UAS-IswiK159R); (2) w1118; P[w+ UAS-GFP.nls] (referred to as UAS-GFP); (3) w; P[w+mC UAS-LacZ.B] 4-1-2 (referred to as UAS-LacZ); (4) Df(1)w67c2 y; P[w+ UAS-brmK804R] 2–2 (referred to as UAS-brmK804R); (5) Df(1)w67c2 y P[w+ UASG-Iswi-HA-6His] 18 (referred to as UAS-Iswi); (6) w; al b cn Iswi2 sp; P[w+, UAS-IswiK159R-HA-6HIS] 11–4/T(2;3) B3 CyO, TM6B, Tb; and (7) w; P[w+, eyGAL4], P[w, UAS-IswiK159R-HA-6His] 11–4/TM3, Sb.
Polytene chromosomes were isolated from salivary glands of third-instar larvae maintained at 18 °C. To examine the effect of Iswi mutations on polytene chromosome structure, salivary glands were isolated from Iswi1/Iswi2 larvae. The effect of ISWIK159R expression on polytene chromosome structure was examined by crossing w; P[w+, UAS-IswiK159R-HA-6His] 11–4/TM6B, P[w+Ubi-GFP], Tb flies to da-GAL4 or w; P[w+, eyGAL4] flies at 18 °C. Salivary glands were isolated from larvae bearing UAS-IswiK159R in trans to ey-GAL4 or da-GAL4. Salivary glands bearing UAS-LacZ, UAS-GFP, and UAS-brmK804R in trans to da-GAL4 or ey-GAL4 were analyzed for control experiments.
To analyze polytene chromosome structure, salivary glands were dissected in 0.7% NaCl and squashed in 1.85% formaldehyde/45% acetic acid. Slides were frozen in liquid nitrogen, air dried, and counterstained and mounted in Vectashield containing DAPI (Vector Laboratories, http://www.vectorlabs.com/).
Antibodies used in this study include affinity-purified rabbit anti-ISWI [52]; mouse anti-RPII140 [53]; rabbit anti-H4AcK16 (catalog number ab1762, Abcam, http://www.abcam.com/); rabbit anti-H3K27Me(3×) (catalog number 07–449, Upstate, http://www.upstate.com/); rabbit anti-H3K4Me(3×) (Upstate, catalog number 07–473); mouse anti-Pol IIoser2 (Covance, http://www.covance.com/); mouse pan-histone monoclonal antibody MAB052 (catalog number MAB052, Chemicon, http://www.chemicon.com/); and rabbit anti-histone H1 [54]. Secondary antibodies were obtained from Jackson ImmunoResearch Laboratories (http://www.jacksonimmuno.com/). Slides were counterstained and mounted in Vectashield containing DAPI (Vector Laboratories). Immunostaining of wild-type polytene chromosomes with antibodies against ISWI and Pol II subunits was carried out as described previously [55]. For the staining of polytene chromosomes with monoclonal antibody MAB052 and antibodies against acetylated H4K16, trimethyl H3K27, trimethyl H3K4, and Drosophila histone H1, glands were dissected in 0.7% NaCl and fixed in 6 mM MgCl2/1% citric acid/1% Triton X-100 for 2 min. Chromosome preparations were analyzed using a Zeiss Axioskop 2 plus fluorescent microscope equipped with an Axioplan HRm CCD camera and Axiovision 4.2 software (Zeiss, http://www.zeiss.com/). To directly compare the chromosomal levels of acetylated H4K16 and histone H1 in wild-type and mutant larvae, chromosomes were squashed and processed in parallel, and images were captured using identical exposure times.
The association of ISWI with nine potential targets of ISWI repression—CG7951 (sima); CG18166; CG7997; CG2162; CG1787 (Hexo2); CG7696; CG3765 (Pde9); CG1214 (ru); and CG12002 (Pxn)—and seven potential targets of ISWI activation—CG4690 (Tsp5D); CG18350 (Sxl); CG10934; CG8094 (Hex-C); CG1146; CG4281; and CG9531—was examined. cDNAs corresponding to each gene were obtained from the Drosophila Genomics Resource Center (https://dgrc.cgb.indiana.edu). Wild-type polytene chromosomes were stained with antibodies against ISWI and Pol IIoser2 as described previously [55]. The chromosome squashes were then washed in PBS for 5 min, fixed in 3.7% formaldehyde/PBS for 10 min, and washed again in PBS for 5 min. FISH was subsequently carried out as described in Pimpinelli et al. [56], except that denaturation in 70% formamide/2× SSC was carried out for 8 min and probes were labeled using a biotin nick translation mix (Roche, http://www.roche.com/). Slides were counterstained and mounted in Vectashield containing DAPI (Vector Laboratories).
To examine the effect of loss of zygotic ISWI function on mitotic chromosome structure, neuroblasts were isolated from wild-type and Iswi1/Iswi2 larvae. The effect of ISWIK159R expression on chromosome structure was examined by crossing w; P[w+, UAS-IswiK159R-HA-6His] 11–4/TM6B, P[w+Ubi-GFP], Tb flies to da-GAL4 flies at 18 °C. Neuroblasts were isolated from larvae bearing UAS-IswiK159R in trans to da-GAL4. Neuroblasts bearing UAS-LacZ, UAS-GFP, and UAS-brmK804R in trans to da-GAL4 were analyzed for control experiments. Metaphase chromosomes were prepared from neuroblasts of third-instar larvae as described by Cenci et al. [57]. Slides were counterstained and mounted in Vectashield containing DAPI (Vector Laboratories). Chromosome preparations were analyzed using a Zeiss Axioskop 2 plus fluorescent microscope equipped with an Axioplan HRm CCD camera, and Axiovision 4.2 software (Zeiss).
To express IswiK159R in heterozygous Iswi2 embryos, w; al b cn Iswi2 sp; P[w+, UAS-IswiK159R-HA-6HIS] 11–4/T(2;3) B3 CyO, TM6B, Tb females were crossed to da-GAL4 males at 25 °C. Embryos were collected which were either +; da-GAL4/T(2;3) B3 CyO, TM6B, Tb or al b cn Iswi2 sp/+; P[w+, UAS-IswiK159R-HA-6HIS] 11–4/da-GAL4. Embryos bearing UAS-LacZ, UAS-GFP, or UAS-brmK804R in trans to da-GAL4 were analyzed for control experiments. Embryos were collected and fixed with methanol as described by Rothwell and Sullivan [58], stained with a mouse anti-α-tubulin antibody (catalog number T9026, Sigma-Aldrich, http://www.sigmaaldrich.com/), counterstained, and mounted in Vectashield containing propidium iodide (Vector Laboratories). Embryo preparations were analyzed using an inverted microscope (DM IRB, Leica Microsystems, http://www.leica-microsystems.com/) equipped with a laser confocal imaging system (TCS SP2, Leica Microsystems), and images were acquired and analyzed using Leica Microsystems confocal software version 2.61.
Proteins were extracted from Drosophila embryos as described in Srinivasan et al. [59]. To prepare salivary gland protein extracts, salivary glands were dissected from third-instar larvae in 0.7% NaCl; transferred to a microfuge tube containing 100 μl PBS, 0.8% NP-40, 1 mM DTT, and 1 mM PMSF; pelleted by centrifugation at 2,100g for 5 min at 4 °C; resuspended in boiling SDS–polyacrylamide gel electrophoresis (SDS-PAGE) loading buffer; homogenized with a pestle; and frozen in liquid nitrogen. Proteins were fractionated by SDS-PAGE and analyzed by protein blotting as described in Srinivasan et al. [59]. Primary antibodies were detected using horseradish-peroxidase-coupled secondary antibody (Bio-Rad, http://www.bio-rad.com/) and Super Signal chemiluminescent reagent (Pierce, http://www.piercenet.com/).
To prepare salivary gland chromatin extracts, salivary glands were dissected from third-instar larvae in 0.7% NaCl; transferred to a microfuge containing 2% formaldehyde in 100 μl of 0.5× M buffer (10 mM Hepes-KOH [pH 7.6], 25 mM KCl, 5 mM MgCl2, 5% glycerol), 1 mM DTT, and 1 mM PMSF; and incubated for 15 min at room temperature. The cross-linking reaction was stopped with 125 mM glycine. Salivary glands were transferred to a microfuge tube containing 100 μl of M buffer, 0.8% NP-40, 1 mM DTT, and 1 mM PMSF; incubated on ice for 15 min; homogenized with a pestle; pelleted by centrifugation at 2,100g for 5 min at 4 °C; and resuspended in boiling SDS-PAGE loading buffer. Proteins were fractionated by SDS-PAGE and analyzed by protein blotting as described above using rabbit antibodies against rabbit Drosophila histone H1 [54] and histone H3 (Abcam, catalog number ab1791). Chemiluminescent signals were quantified using a Bio-Rad Molecular Imager.
Partial micrococcal nuclease digestion of salivary gland chromatin was conducted using a modification of a protocol described in Cartwright et al. [60]. Chromatin was isolated from wild-type or w; P[w+, ey-GAL4], P[w+, UAS-IswiK159R-HA-6-His] 11–4/TM3, Sb third-instar larvae. Salivary glands were dissected in M buffer (10 mM Hepes-KOH [pH 7.6], 25 mM KCl, 5 mM MgCl2, 5% glycerol) and transferred to 100 μl of M buffer, 0.5 mM PMSF, and 0.5 mM DTT. After addition of 4 μl of 20% NP-40, the samples were incubated on ice for 15 min and homogenized with a pestle. The samples were centrifuged at 2,100g for 5 min at 4 °C; washed with MNase buffer (M buffer, 2 mM CaCl2, 0.5 mM PMSF); centrifuged at 2,100g for 5 min at 4 °C; and resuspended in 200 μl of MNase buffer containing 10,180 units of micrococcal nuclease (USB, http://www.usbweb.com/). After digestion for 4 min at room temperature, the reaction was stopped by the addition of 200 μl of S buffer (20 mM Tris-HCl [pH 7.4], 200 mM NaCl, 2 mM EDTA, 2% SDS, 30 mM EGTA). Proteinase K was added to a final concentration of 1 mg/ml, and the samples were incubated for 1 h at 45 °C. Following phenol-chloroform extraction and the addition of 1 μl of glycogen (Roche), the DNA was ethanol precipitated and resuspended in 120 μl of RNase buffer (50 mM Tris-HCl [pH 7.5], 100 mM NaCl, 10 mM EDTA). RNase A (Sigma) was added to a final concentration of 250 μg/ml, and the samples were incubated for 1 h at 37 °C. Following another round of phenol-chloroform extraction and ethanol precipitation, DNA was resuspended in 10 μl of TE and analyzed by agarose gel electrophoresis.
Microarrays containing ∼14,400 cDNA fragments representing over 14,000 different Drosophila genes were used to characterize changes in gene expression resulting from the loss of ISWI function. Control larvae were generated by crossing Oregon R males to Df(1)w67c2 y virgin females. Iswi1/Iswi2 larvae were generated by crossing w; al b cn Iswi2 sp; +/T(2;3) B3 CyO, TM6B, Tb virgin females to Iswi1 Bc/SM5, Cy sp males. Male and female third-instar larvae of the appropriate sex and genotype were identified using the markers y, Bc, and Tb. RNA was isolated from third-instar larvae or dissected salivary glands by homogenization in Trizol reagent (Invitrogen, http://www.invitrogen.com/) followed by chloroform extraction and precipitation with isopropanol. Detailed protocols for the RNA isolation, cDNA synthesis, construction of microarrays, and hybridization are provided in Protocol S1.
Microarrays were scanned and analyzed using an Axon scanner and GenePix software (MDS Analytical Technologies, http://www.moleculardevices.com/). The data acquired with GenePix were uploaded to the Stanford MicroArray Database (http://genome-www5.stanford.edu/) and analyzed using Stanford MicroArray Database data analysis software. Data were normalized by ribosomal RNA inputs, by spot intensity/background readings, and by data taken from mitochondrial genes (i.e., mt:ND3 and mt:ND6). Data obtained from triplicate hybridizations per experiment were filtered for measurement values with a regression correlation coefficient equal or greater than 0.6. Data that passed the spot criteria were filtered based on gene data values whose log2R/G normalized ratio (mean) was equal to or greater than four. Finally, data that passed all the above criteria were clustered using a non-centered Pearson correlation algorithm.
|
10.1371/journal.ppat.1005025 | A Unique Human Norovirus Lineage with a Distinct HBGA Binding Interface | Norovirus (NoV) causes epidemic acute gastroenteritis in humans, whereby histo-blood group antigens (HBGAs) play an important role in host susceptibility. Each of the two major genogroups (GI and GII) of human NoVs recognizes a unique set of HBGAs through a distinct binding interface that is conserved within a genogroup, indicating a distinct evolutionary path for each genogroup. Here, we characterize a Lewis a (Lea) antigen binding strain (OIF virus) in the GII.21 genotype that does not share the conserved GII binding interface, revealing a new evolution lineage with a distinct HBGA binding interface. Sequence alignment showed that the major residues contributing to the new HBGA binding interface are conserved among most members of the GII.21, as well as a closely related GII.13 genotype. In addition, we found that glycerol inhibits OIF binding to HBGAs, potentially allowing production of cheap antivirals against human NoVs. Taken together, our results reveal a new evolutionary lineage of NoVs selected by HBGAs, a finding that is important for understanding the diversity and widespread nature of NoVs.
| Human norovirus (huNoV) has diverged into two major lineages (GI and GII) selected by the host histo-blood group antigens (HBGAs). Both lineages further diverge into various sub-lineages (genotypes) that recognize different ABH and Lewis antigens through a common HBGA binding interface shared among strains within each genogroup. In this study, through X-ray crystallography of the P domain of a GII.21 huNoV (OIF) we identified a unique lineage in GII consisting of GII.13 and GII.21 genotypes that recognize HBGAs through a binding interface distinct from the GII conventional binding interface. While the mechanism remains unknown, our finding raises an alert on future emergence of new lineages by the same way via developing new receptor binding interfaces, as well as further divergence of this new lineage into more sub-lineages recognizing different HBGAs, which may impact future epidemiology and strategies for disease control and prevention against huNoVs.
| Noroviruses (NoVs) are a group of non-enveloped, single stranded, positive-sense RNA viruses that constitute the Norovirus genus in the family Caliciviridae. NoVs are genetically diverse, containing six genogroups (GI to GVI) with over 35 genetic genotypes. NoVs exhibit wide host tropisms causing diseases in various mammalians including human. Human NoVs (huNoVs), consisting of mainly GI and GII NoVs, are the most important viral cause of epidemic acute gastroenteritis in humans [1], claiming over 200,000 lives each year [2]. NoVs are encapsulated by a protein capsid that is assembled by a single major structural protein, the capsid protein VP1. Each NoV capsid contains 180 copies of VP1, which are organized as a T = 3 icosahedron [3]. VP1 is divided into the N-terminal shell (S) and the C-terminal protruding (P) domain, forming the interior shell and the multiple protrusions of the capsid, respectively [3]. The P domain can be further divided into P1 and P2 subdomains, corresponding to the head as the outermost portion of NoVs, and the leg of the protrusion, respectively. Compared with S and the P1 domain, the P2 domain exhibits the most variable sequences, which are responsible for strain-specific virus-host interactions and immune responses of NoVs.
HuNoVs interact with histo-blood group antigens (HBGAs) in a strain-specific manner [4, 5]. HBGAs are fucose-containing glycans in specific sequences as determinants of human and animal blood types, including A/B/O, secretor (H), and Lewis (Le) or non-secretor (H negative) types. They are often present as parts of the carbohydrate moiety of cell surface glycoproteins and glycolipids with N- or O-linkage [6]. The biological roles of HBGAs in huNoV infection have been revealed by human volunteer challenge studies [7–9] and outbreak investigations [10, 11] of huNoVs, in which an association between the host susceptibility and the HBGA binding patterns of huNoVs has been established.
Despite the recent breakthroughs in culturing huNoVs in BJAB cells [12], the use of huNoV reverse genetics system [13], and the development of an immunocompromised mouse model [14] for huNoV propagation, an effective cell culture system or an animal model for huNoVs remains lacking. As a result, our understanding on huNoV-HBGA interactions relies mainly on data from in vitro studies using various recombinant subviral particles as models of huNoVs. Virus-like particles (VLPs), which are produced by expression full-length VP1 via a eukaryotic expression system, share similar structures and functions with the capsid of a native virion, and they have been used extensively as a huNoV surrogate. In addition, smaller P domain complexes that are self-assembled by expression of the huNoV P domain, including the P dimer [15], small P particle [16] and P particles [17, 18], are also used for the study of huNoV-HBGA interactions. Knowledge on the structures of the HBGA binding interfaces has been derived mainly from crystal structures of P domain dimers in complex with HBGA oligosaccharides [19–26], by taking advantage of its small size (~69 kDa) and easy production.
Structural analysis of known HBGA binding interfaces of huNoVs showed that GI and GII huNoVs recognize HBGAs through a conserved, genogroup-specific binding interface (reviewed in [27–29]), suggesting a strong selection of huNoV evolution by human HBGAs. On the other hand, the GI and GII HBGA-binding interfaces are distinct in the locations, structures, residue compositions, and HBGA binding modes [27–29], indicating a long separation of the two genetic lineages. In this study we report a new evolutionary lineage consisting of GII.13 and GII.21 genotypes within GII, which does not share the conventional GII HBGA binding interface (Fig 1), but remains binding ability to HBGAs [5, 30]. X-ray crystallography of the GII.21 OIF virus P domain complexed with a Lea antigen revealed a new HBGA binding interface that is distinct from the GII conventional binding interface. Sequence alignment further showed that the amino acid composition of the new binding interface of OIF is highly conserved among all members of both GII.21 and a closely related GII.13 genotype. These results indicate that the genetic branch consisting of GII.21 and GII.13 developed along a novel evolutionary path that split from the mainstream lineage of GII NoVs selected by HBGAs. While many questions on the cause and mechanisms behind the emergence of this new lineage remain unanswered, our data point towards a continual occurrence of new lineages, which may significantly impact future epidemiology and prevention strategies against huNoVs.
The crystal of the native P domain protein belongs to the P21 space group, with two monomers forming a dimer in an asymmetric unit. The final refined structure of native P domain includes residues 223 to 527, with the exception of a loop region comprising residues 340~342 due to the lack of recognizable electron density. While the P domain of OIF virus shares only ~50% sequence homology with other GII P domains of known crystal structures, it shows an arrangement of overall and secondary structures similar to those huNoV P proteins, including GII.4 VA387 [20], GII.10 Vietnam 026 [23], GII.9 VA207 [21], and GII.12 Hiro [23] with Cα atoms r.m.s.d. = 0.85 Å, 0.83 Å, 0.82 Å, and 0.62 Å, respectively. Like other huNoVs, the OIF P domain has two moieties (Fig 2), with the inner portion or P1 subdomain (residues 223 to 272, and 416 to 527) constituting the leg, and the outer moiety or P2 subdomain (residues 273 to 415) forming the protruding head of the P dimer.
The two monomers of the OIF P dimer are related by a non-crystallographic two-fold axis, which forms the biologically active protrusion of the NoV capsid. The P dimer has a dimension of 55 Å×64 Å×70 Å with an extensively buried interface of 3,500 Å2 between two protomers, including hydrophobic and hydrophilic interacting residues from both P1 and P2 subdomains (Fig 2A). These extensive inter-molecular interactions contribute to the stability of the P dimer. Although the OIF P dimer shares similar global structures with the previously reported P dimers of other huNoVs, significant differences on the top surface are clearly seen, mainly due to the differences in the sequences, lengths, and conformations of several surface loops (Figs 3 and 4) (see below).
The P-loop of the OIF P dimer exhibits a unique conformation, taking a ~90° flip (Fig 3), which results in its extruding out of the flat surface and approaching the opposite protomer. This leads to greater exposure of the P-loop on the top surface of the P dimer compared to that in other GII huNoVs (Fig 4). As a result, the majority of the S-loop is covered in the OIF virus. Such structural rearrangements destroy the structural integrity of the conventional GII HBGA binding interfaces that are formed mainly by P- and S-loops (Figs 2 and 4). On the other hand, the B- and the T-loops also exhibit unique conformations (Fig 3), approaching to each other in closer proximity with the N-loop (Fig 4A). This special structural feature allows the formation of new HBGA binding interface (see below), formed by the B-, T- and N-loops, a scenario that is absent in all other known GII huNoV structures (Fig 4). The protrusions of the P- and A-loops directed towards the opposite protomer restructure the boundary between two P monomers into a saw tooth-like structure, in contrast to the straight line-like pattern in other GII NoVs (Fig 4, compare G and H). These different surface structures and conformations of OIF P dimer might also contribute to the difference in antigenic and immunological properties of GII.21 compared to other huNoV genotypes.
As shown in previous studies [5, 30], the OIF virus that does not share the conventional GII HBGA binding interface only recognizes Lea antigen. To understand their binding mechanism, we co-crystallized the OIF P domain with Lea trisaccharide (Lea-tri). The P domain-Lea-tri complex was crystallized in P212121 space group, with one homodimer in each asymmetric unit. The Lea-tri is clearly visible in the 2Fo-Fc difference electron density map, with all three rings of the Lea-tri well fitted into the map (Fig 5A). Two symmetric Lea binding interfaces are identified on the top surface of the P dimer (Fig 5B), each of which is formed by eleven residues from the P2 domain of a single protomer. Specifically, residues W296 from the B-loop, S354 and S357 from the N-loop, N392 and T395 from the T-loop form the depressed region at the bottom of the binding pocket (Fig 5C–5E). We noticed that, although S357 does not interact directly with the Lea-tri, it forms a strong hydrogen bond (2.6Å) with the D297 to stabilize the side chain of the latter, ensuring the structural integrity of the binding pocket (Fig 5E and 5F). The surrounding wall of the binding pocket is built by D294 and Y295 from B-loop, T356 and E358 from N-loop, and N394 from T-loop (Fig 5C–5E). None of these residue compositions are conserved with those of the known GII HBGA binding interfaces that is contributed by residues from P-, S-, and A-loops (Figs 2 and 4). The OIF HBGA binding interface also differs completely from those of the GI huNoVs (Fig 4F). Thus, this OIF binding interface represents a previously unrecognized HBGA binding interface.
The HBGA binding pocket interacts with Lea antigen through the β-galactose (β-Gal), one of the two saccharides of the precursor disaccharide, as the major binding saccharide (MaBS). The β-Gal is firmly held inside the binding pocket by nine direct hydrogen bonds formed by the side chains of S354, T356, N392, N394 and T395, in which the side chains of S354 from the bottom and N394 from the wall of the pocket form three strong hydrogen bonds (2.7 Å or 2.6 Å) with β-Gal (Fig 5E and 5F). It is noteworthy that the side chain of W296 is oriented nearly in parallel with the β-Gal ring, which leads to a hydrophobic interaction/stacking with the latter to further support the interaction between β-Gal and the binding pocket. In addition, the side chains of D294, D297 and E358 form water-bridged hydrogen bonds with β-Gal via a common water molecule, contributing to the binding outcomes between the OIF P domain and the Lea antigen. Together, these amino acids form the β-Gal binding site that is the major component of the HBGA binding interface of OIF virus.
As a result of the strong interactions between the β-Gal and the binding pocket, the Lewis fucose (Le-Fuc, α-1,4-fucose) is orientated between the walls of N394 and E358. Only a single direct hydrogen bond (2.7 Å) forms between E358 and the Le-Fuc (Fig 5E and 5F) and therefore, the Le-Fuc is a minor binding saccharide (MiBS). As a result of the β-Gal/Le-Fuc-binding pocket interactions, the N-acetyl-β-glucosamine (β-GlcNAc), the other saccharide of the precursor disaccharide, is pointing away from the surface of the binding interface. No hydrogen bond was observed between β-GlcNAc and the binding interface (Fig 5C–5F).
We further investigated the roles of the eleven residues that form the HBGA binding interface of the OIF virus using single-point mutagenesis (Figs 5 and 6). Compared with the wild type P particle of OIF virus that binds strongly to Lea oligosaccharide, but not other oligosaccharides representing different HBGA types, including precursors (Fig 6A), mutant P particles with a single mutation at one of those amino acids to an alanine lost the binding to Lea antigen completely (Fig 6C–6L) or nearly completely (Fig 6B). These results confirmed the importance of these residues for the structural and functional integrity of the HBGA binding interface. In addition, except for a slight increase in binding to the A antigen of the E358A mutant, no changes in binding to other HBGA types were observed for these interface mutants.
Analysis of the native OIF P domain structure revealed that a glycerol molecule occupies the HBGA binding pocket, which was clearly visible from the 2Fo-Fc omit map (Fig 7A). Originally, glycerol was part of the protease solution acting as a stabilizer during the cleavage of the GST-P domain fusion protein. The glycerol molecule is held firmly inside the HBGA binding pocket by eight direct hydrogen bonds between the hydroxyl groups of the glycerol and the side chains of N392, N394 and T395 from the T-loop, T356 and the main chain of S357 from the N-loop (Fig 7B and 7C). Interestingly, the glycerol molecule resembles partial structures (C2, C3, and C4 with their hydroxyl groups) of the β-Gal that contributes the vast majority of the interactions with the binding pocket (Compare Fig 5F with Fig 7C), thus explaining the observed interactions.
We found that inclusion of glycerol in any process of purification or crystallization would prevent the binding of the Lea antigen to the same binding pocket. The later success in identification of HBGA binding interface (Fig 5) of OIF virus using the P protein without glycerol proved our assumption that glycerol occupies the HBGA binding pocket and blocks further interaction of the binding pocket with the Lea antigen. Co-crystallization followed by structural determination using the P protein in the presence or absence of glycerol confirmed our observation. We then performed ELISA binding assays to test if the monomeric glycerol (1–10%) can block binding of polyvalent Lea-tri-PAA (polyacrylamide) conjugates (2 μg/mL) to OIF P particles (4 μg/mL). We did not observe any detectable blocking effects (Fig 7D), indicating that the free glycerol can block binding of free, monomeric Lea-tri, but cannot inhibit the binding of polyvalent Lea-tri molecules to the HBGA binding pocket.
The unique HBGA binding interface of the GII.21 OIF virus prompted us to examine whether this is a common feature of this specific huNoV lineage. Representative P domain sequences of GII.13/21 huNoVs were aligned and compared with those of GII.17, which is genetically closest to the GII.13/21 lineage (Fig 1A), as well as with that of GII.4, which is the most prevalent genotype. We focused on the surface loops that form the OIF HBGA binding pocket (B-, T, and N-loops) and the conventional GII binding interface (P-, S- and A-loops) (Fig 8). We found that most members of the GII.13/21 lineage share the major residues that form the novel HBGA binding pocket, indicating a new evolutionary selection has occurred among members of this genetic lineage. In contrast, GII.17 retained the conventional GII HBGA binding interfaces, indicating that the occurrence of the OIF-like binding pocket in the GII.13/21 lineage was after the evolutionary divergence with the GII.17 genotype.
GI and GII NoVs constitute the vast majority of huNoVs, and they were shown previously to interact with HBGAs through two distinct, genogroup-specific HBGA binding interfaces [31]. The two HBGA binding interfaces differ in their locations on the top of the protruding dimer, their residue composition as well as their interacting modes with HBGAs, but are highly conserved among genotypes within each of the two genogroups (reviewed in [27–29, 31]). It is therefore suggested that GI and GII NoVs must split from their common ancestor a long time ago during evolution into two independent evolutionary lineages, from which the two genogroup specific HBGA binding interfaces evolved individually. It is noteworthy that both binding interfaces recognize most, if not all, of the same repertoire of the polymorphic human HBGAs despite their distinct features, indicating that HBGA is an important selection factor in the evolution of huNoVs.
The presence of the two different HBGA binding interfaces in huNoVs raises the question as to whether such an event may occur again during evolution, especially within the same genogroup. We noted earlier on that a genetic branch that consists of GII.13 and GII.21 genotypes does not share the residue composition of the HBGA binding interface of the mainstream GII NoVs, but maintains HBGA binding function [5, 30], suggesting that this GII.13/21 lineage may interact with HBGAs through an unknown binding interface. By performing X-ray crystallography of the native P domain of the GII.21 OIF virus and its complex with Lea trisaccharide (Lea-tri), we found that GII.21 huNoVs recognize HBGAs via a previously unrecognized binding interface. The newly identified HBGA binding interface is spatially, structurally, and compositionally distinct from the well-studied conventional GII and GI HBGA binding interface.
The conventional GII HBGA binding interface is formed by three surface loops, i.e. the P- and A-loops from the P2 subdomain and the S-loop from the P1 subdomain. The binding interface is located directly on the boundary between the two P protomers, which makes each binding interface bivalent contributed by residues from both P protomers. In the GII.21 OIF virus, however, the conventional HBGA binding interface has been clearly abolished by the rearrangement of the three loops. Through a 90 degree flip, the P-loop extends to the opposite protomer, covering the vast majority of the S-loop and moving away from the A-loop. Nevertheless, this dramatic structural rearrangement of the OIF P dimer also helps to reconstruct a new HBGA binding interface, formed by other three surface loops, i.e. the B-, N- and T loops from the P2 subdomain of a single P protomer. We also noted that the binding interface corresponding to that of OIF virus does not exist in the mainstream GII NoVs, because the B- and T-loops are far away from each other in those GII NoVs.
In addition to the location, residue composition, and structure, the two GII HBGA binding interfaces are also different in their binding modes to HBGAs. The general binding modes of huNoVs have recently been summarized [28, 29]. GI NoVs bind HBGAs via a Gal (theα- or the β-Gal), while GII NoVs bind via a Fuc (the α-1, 2/3/4 Fuc) as the major binding saccharide (MaBS). The OIF virus interacts with the Lea antigen via the β-Gal as the MaBS, distinct from the convention GII, but similar to the GI binding interfaces. In addition, the OIF-Lea interaction relies mainly on the β-Gal that contributes 11 out of total 12 hydrogen bonds, including 9 direct ones. The α-Fuc (Le-Fuc) contributes only one hydrogen bond, while the β-GlcNAc does not participate directly in binding. This “upright” binding mode is also different from the “flat” mode observed in the conventional GII binding interfaces, which often includes two saccharides as the minor binding saccharides (MiBS), which usually contribute three or more interacting bonds. Again, the binding mode of OIF virus to HBGA appears to be more similar to that of GI NoV than to GII NoV [28, 29].
Our crystal structures of the HBGA binding interface of OIF virus complexed with Lea-tri reveal why OIF virus does not bind H, A, and B antigens. The H epitope (α-1,2 Fuc) and the A/B epitope (α-GalNAc/Gal) are linked to the 2’ and 3’ hydroxyl groups of β-Gal, respectively [29]. However, both the 2’ and 3’ hydroxyl groups of the β-Gal are located at the bottom of the binding pocket (Fig 5). Any addition of an extra saccharide at one of these positions would cause serious steric hindrance, and thus prevent the binding of H, A, and B antigens to the binding pocket of the OIF virus. This binding mode may also explain the absence of changes of the binding specificity through the eleven single residue mutations at the amino acids forming the OIF binding interfaces (Fig 6). This HBGA binding interface and its binding mode appear not to provide much flexibility to different HBGAs. Thus, this scenario of OIF virus differs clearly from observed flexibility of the conventional GII binding interface that is capable to bind more than one HBGA type and that can change binding specificity through a single residue mutation in or around the binding interface [32–36].
The formation of the new HBGA binding interface in the GII 13/21 lineage also raises questions as to how the conventional GII HBGA binding interface has been abandoned, and how a new interface was formed. The huge structural variation of the conventional GII binding interface in OIF virus may represent the accumulation of multiple gradual changes over time, following the acquisition of the novel binding interface and loss of binding function of the old one and HBGA selection. The movements of the B-, T- and N-loops to form the new HBGA binding interface can be monitored through the huNoVs with known crystal structures (Fig 4). The three loops are far away in GI.1 (Norwalk virus), GII.4 (VA387), GII.9 (VA207), GII.12 (Hiro), but are closer sterically in GII.10 (Vietnam026). Thus, movements of these loops are the structural basis for the formation of the functional HBGA binding interface found in OIF virus, although how such a functional binding interface was created in detail remains to be elucidated.
How the GII.13/21 lineage with the new HBGA binding interface split from the GII.17 genotype with the conventional one remains unclear. The capsid proteins (VP1s) of the GII.13/21 lineage share generally ~75% sequence identify with that of the GII.17 NoVs and these differences concentrate in their P domains that shows only ~62% sequence identity. This great sequence difference indicates that the two lineages have split from each other for a long time, which makes an attempt to reconstruct their common ancestor difficult. We speculate that the GII.17 P domain must have similar surface topology and loop organization to those of the mainstream GII NoVs due to the fact that the GII.17 NoV retains the conventional binding interface. In contrast, the GII.21 OIF shows very different surface topology and loop arrangements compared with the other GII NoVs, leading to the formation of the new HBGA binding interface. Unfortunately, the lack of the crystal structures of a GII.17 P domain prevents our further understanding on the evolutionary scenarios of the separation between the GII.13/GII.21 lineage and the GII.17 genotype. Thus, future study to solve the structures of a GII.17 P domain is of significance.
While the conservation of the conventional GI and GII HBGA binding interfaces indicates that there was selection pressure imposed by HBGAs functioning as attachment factors or receptors, the conservation of the HBGA binding interface of the GII.13/21 lineage (Fig 8) also indicates that new selection pressure of HBGA for the new binding interface has been established. Thus, it is plausible to expect that the conserved binding interface of the GII.13/21 lineage will continue to be maintained, just like the GI and GII binding interfaces. It would be important to show whether different strains of the GII.13/21 lineage can also recognize different HBGAs, just as GI and GII NoVs do. It was noted that Y295 in the binding pocket of GII.21 OIF is replaced by an asparagine (N) in GII.13 (Fig 8) that is highly conserved among most known GII.13 isolates [37]. It would be of significance to assess how this mutation could affect the HBGA binding profiles of GII.13 NoVs. In addition, N392 is another important residue for the structural and functional integrity of the GII.21 OIF binding interface, which was confirmed by the loss of the binding function through a N392A mutation (Fig 6J). This residue is also conserved among GII.13 NoVs (N395, Fig 8). However, we noted an exception, a GII.13 isolate called Kashiwa 47 (BAC05515) has an N395D mutation [37] and the recombinant VLPs of this isolate did not bind human saliva samples and tested synthetic HBGAs [38], consistent with our result of N392A mutation in GII.21 OIF.
Members of the GII.13/21 lineage revealed increased prevalence in the recent years. Surveillance indicates that the GII.13 and GII.21 huNoVs continue to cause outbreaks in the USA (CaliciNet http://www.cdc.gov/norovirus/reporting/calicinet/data.html), Europe (NoroNet, http://www.rivm.nl/dsresource?objectid=rivmp:248062&type=org&disposition=inline), and other countries [37, 39, 40], indicating that the GII.13/21 lineage is of clinical importance. It was also noted that, after being silent for the past decades, GII.17 NoVs became dominant in southern China in the past winter [41], overwhelmed GII.4 NoVs. Thus, continual surveillance of the epidemiological trends of the GII.13/17/21 lineage will be necessary in the future.
The fact that a new functional carbohydrate binding interface can be created through evolution raises the possibility of zoonotic transmission of NoVs. A new binding interface could recognize different host factors or receptors among human populations or animal species. The currently known host tropisms of different human and animal NoVs may be examples of such possibility. The genetic branch of porcine NoVs, consisting of the GII.11, GII.18 and GII.19 genotypes may be another example. Since these porcine NoVs share the conventional GII binding interface, even though they do not infect humans, it would be important to determine what host factor that porcine NoVs may recognize and understand how such molecule(s) interact with the conserved GII binding interface of porcine NoVs.
Another important finding of this study is the strong interaction of the HBGA binding pocket with a glycerol molecule, a fact mirrored by its inhibition of OIF P domain interaction with Lea antigen. Thus, glycerol, a small, low cost compound, may be a promising antiviral candidate against infection of the GII.13/21 NoVs. The glycerol molecule shares the structure of the β-Gal that is involved in the interaction to the binding pocket. A similar scenario was also observed in a previous study showing that a citrate molecule occupies the binding pocket of a GII.10 NoV (Vietnam026) [42] and importantly, does so with similar binding affinity observed between HBGAs and GII.10 P domains. However, our results showed that the free, monomeric glycerol cannot inhibit the interaction of the OIF P particle with multivalent Lea-tri-PAA conjugates, most likely due to avidity effects. Therefore, further investigations will be necessary to develop polymers that contain multivalent structures of glycerol molecules as antiviral candidates. Most importantly, our study should provide a solid structural basis and a model for future studies.
In summary, we have identified and proved the genetic branch of GII.13/21 as new evolutionary lineage with a novel binding interface for HBGAs. We have elucidated the structural basis for the abolishment of the conventional GII HBGA binding interface and the development of the new binding site. We also showed that the new HBGA binding interface is conserved among members of the GII.13/21 lineage, indicating that a new selection pressure has been asserted through the interaction with HBGAs. Finally, we have identified glycerol as a potential low-cost compound to become an antiviral candidate for future development. In conclusion, our data provide new insights into the complex interactions between the diverse huNoV and the polymorphic HBGAs.
OIF (Operation Iraqi Freedom) virus is a GII.21 huNoV that was isolated from a huNoV outbreak in a US military deployed to Iraq in 2003. The cDNA sequences encoding the P domain of the OIF virus (accession number: AY675554; residues 220 to 527) was cloned into pGEX-6P-1 expression vector between the EcoR I and Xho I sites of the MCS. The OIF P domain was expressed as a GST fusion protein in Escherichia Coli BL 21 (DE 3) upon induction with 0.5 mM Isopropyl β-D-1-Thiogalactopyranoside (IPTG) at 16°C for 18 hours, at an OD600 nm of 0.6–0.8. The recombinant protein was purified using glutathione-sepharose 4B (GE Healthcare Life Sciences) following the manufacturer’s guidelines. The GST-P domain fusion protein was cleaved on the resin by Prescission Protease (GE Healthcare Life Sciences) at 4 °C overnight, and the P proteins were further purified by Resource Q anion ion exchange (GE), using a buffer containing 20 mM HEPES (pH 7.5), with P domain protein eluted at approx. 100 mM NaCl.
The purified native P domain protein was dialyzed against buffer containing 20 mM HEPES (pH 7.5), 150mM NaCl and 5% (v/v) glycerol, before it was concentrated to 13 mg/mL for crystallization. Native P domain crystals were grown by hanging-drop vapor diffusion method, with the crystallization droplet containing 1 μL protein and 1 μL reservoir solution containing 0.1 M MES (pH 6.5), 0.25 M (NH4)2SO4, 18% (w/v) polyethylene glycol (PEG) 3350. The crystals were grown at 16°C and harvested after approx. one week.
While performing the crystallization experiments, we observed that the presence of glycerol inhibited the co-crystallization of the OIF P domain-Lea complex. The P protein used for co-crystallization with Lea-trisaccharide was purified with identical solutions (see above), but in the absence of glycerol. The Lea trisaccharide [β-Gal-(1,3)-(α-Fuc-(1,4))- GlcNAc] (J&K, China) was dissolved in double distilled water and prepared as 20 mM solution, and then mixed with an equal volume of native P domain protein (26 mg/mL), and incubated at 4°C for 1 hour before crystallization. The final reservoir for the growth of complex crystals contained 0.25 M (NH4)2SO4, 18% (w/v) PEG 3350. Micro-seeding technique was used to aid the growth of complex crystal with native crystal seeds 16 hours after the setting of crystallization droplets. OIF P domain-Lea complex crystals were harvested in three days.
The cryo-protectants of crystals of unliganded P domain and P domain-Lea complex were the corresponding reservoir solutions complemented with 15% (v/v) PEG 400. Crystals were briefly soaked in the cryo-protectant for 5 seconds before being mounted for diffraction test. The diffraction data for native crystal were collected at the beamline 41U of SPRING8 (Japan) at a wavelength of 1.0000 Å, while the Lea complex data were collected at rotating-anode X-ray source MicroMax-007/Satun 944 HG/Varimax HF (Institute of Biophysics, CAS, Beijing) at the wavelength of 1.5418 Å. Diffraction data were processed, scaled, and merged using the HKL-2000 program package [43]. Data collection statistics are summarized in Table 1.
The native crystal structures were solved by molecular replacement method using Phaser of CCP4 suite [44] and the GII.9 NoV VA207 P domain (PDB ID:3PUN) structure as the initial search model. Automatic structure building and refinement were carried out using Phenix program [45] and manual adjustment was done using the program COOT [46] with guidance of (2Fo-Fc) and (Fo-Fc) electron density maps. Water molecules were added at the final round of structure optimization at (Fo-Fc) electron density map peaks (>2.5 σ) where they can form stable hydrogen bonding with nearby amino acid residues. The phases and structures of the P domain-Lea complex were solved using the final refined structure of native P domain protein as model. Structure refinement statistics are summarized in Table 2. The final structure validation was done with the PROCHECK [47], with no residue found at disallowed region of the Ramachandran plot. Structural analysis was performed using EdPDB [48] and Pymol [49].
Single residue mutations were introduced to the HBGA binding site of the OIF P domain by site-directed mutagenesis using the expression plasmid of the wild type OIF P domain as template. Mutagenesis was carried out using the QuickChange Site-Directed Mutagenesis Kit (Agilent Technology, CA) and corresponding primer pairs containing the mutation sites. After confirmation of mutations, the P proteins were expressed as P particles and purified using an E. coli system (BL21) as described elsewhere [31, 50, 51]. The GST-P domain fusion proteins were cleaved by thrombin (GE Healthcare Life Sciences) to allow the P proteins to self-assemble into P particles. P particle formation was monitored by gel filtration through a size-exclusion column (Superdex 200, GE Healthcare Life Sciences) powered by an AKTA-FPLC system (model 920, GE Healthcare Life Sciences, Piscataway, NJ) followed by SDS-PAGE electrophoresis. The P particles showed a molecular weight of 830 kDa. None of the mutations in this study affected the formation of P particles, as confirmed by FPLC (data not shown). All mutant P particles can be well detected by the hyperimmune serum against norovirus VLPs that was used in the HBGA binding and blocking assays (below).
HBGA assays were carried out as described previously [5]. Briefly, a panel of synthetic oligosaccharides (GlycoTech, Gaithersburg, MD) representing types A, B, H1, H2, H3, Lea, Leb, Lex, Ley, sialyl Lea, and sialyl Lex as well as the disaccharides representing the type 1 and type 2 precursors of HBGAs at 2 μg/mL were coated on microtiter plates at 4°C overnight. After blocking with 5% (w/v) non-fat milk, the coated oligosaccharides were incubated with the affinity-column purified OIF P particles (10 ng/μL) for 60 minutes at 37°C. The bound OIF P particles were detected as described previously, using an in-house hyperimmune rabbit serum against various huNoVs [5, 30]. For glycerol blocking assays, the OIF P particles (4 μg/mL) were mixed with the monomeric glycerol (1–10%) and incubated with the coated polyvalent Lea-tri-PAA (polyacrylamide) conjugates (2 μg/mL). The bound OIF P particles were detected as described above.
The cDNA sequences encoding the P domain of the OIF virus has been submitted to GenBank previously with an accession number of AY675554. Coordinates and structure factors of the native OIF P protein and the complex with Lea trisaccharide have been deposited in the Protein Data Bank (www.pdb.org) with the pdb accession codes of 4RLZ and 4RM0, respectively. The GenBank accession numbers of the sequences that were used in Figs 1 and 8 were shown in their figure legends.
|
10.1371/journal.pgen.1002761 | The Histone Demethylase Jhdm1a Regulates Hepatic Gluconeogenesis | Hepatic gluconeogenesis is required for maintaining blood glucose homeostasis; yet, in diabetes mellitus, this process is unrestrained and is a major contributor to fasting hyperglycemia. To date, the impacts of chromatin modifying enzymes and chromatin landscape on gluconeogenesis are poorly understood. Through catalyzing the removal of methyl groups from specific lysine residues in the histone tail, histone demethylases modulate chromatin structure and, hence, gene expression. Here we perform an RNA interference screen against the known histone demethylases and identify a histone H3 lysine 36 (H3K36) demethylase, Jhdm1a, as a key negative regulator of gluconeogenic gene expression. In vivo, silencing of Jhdm1a promotes liver glucose synthesis, while its exogenous expression reduces blood glucose level. Importantly, the regulation of gluconeogenesis by Jhdm1a requires its demethylation activity. Mechanistically, we find that Jhdm1a regulates the expression of a major gluconeogenic regulator, C/EBPα. This is achieved, at least in part, by its USF1-dependent association with the C/EBPα promoter and its subsequent demethylation of dimethylated H3K36 on the C/EBPα locus. Our work provides compelling evidence that links histone demethylation to transcriptional regulation of gluconeogenesis and has important implications for the treatment of diabetes.
| Histones are small proteins that are essential for packaging and ordering genetic information (DNA) into high-order chromatin structures. Methylation of specific lysine residues of histones alters chromatin structure, serving as an important epigenetic mechanism for regulation of gene expression. The dynamic nature of histone methylation is controlled by a balance of methyltransferases and demethylases. We have discovered here that the demethylase Jhdm1a negatively regulates gluconeogenesis (de novo glucose synthesis) through suppressing the expression of two rate-limiting gluconeogenic enzymes. Gluconeogenesis is required for maintaining blood glucose homeostasis; yet, in diabetes mellitus, this process is unrestrained and is a major contributor to hyperglycemia. Indeed, we have found that manipulation of Jhdm1a level in liver affects glucose production in normal mice and hyperglycemia in diabetic mice. Mechanistically, Jhdm1a actively removes dimethyl groups from histone H3K36 along the locus of a key gluconeogenic regulator, C/EBPα, which in turn results in decreased C/EBPα expression. Our findings thus identify histone demethylation as a novel regulatory mechanism for gluconeogenesis and have important implications for the treatment of diabetes.
| Hepatic glucose production is critical for the maintenance of normal blood levels to meet whole-body fuel requirements. In the early phase of postabsorptive state, circulating glucose is supplied from breakdown of liver glycogen stores. When fasting progresses, gluconeogenesis, which utilizes non-carbohydrate precursors to de novo synthesize glucose, becomes the major form of hepatic glucose production [1], [2]. In both type 1 and type 2 diabetes, gluconeogenesis is exaggerated and contributes to hyperglycemia [3]–[5].
The rate of gluconeogenesis is largely determined by three rate-limiting enzymes, Phosphoenolpyruvate carboxykinase (PEPCK), fructose-1,6-bisphosphatase (FBP-1) and glucose 6-phosphatase (G6Pase). The levels of these gluconeogenic enzymes are controlled by hormonal signals, notably glucagon and glucocorticoids, and the opposing hormone insulin, at the transcription level. Key DNA elements responsible for the hormonal regulation have been well characterized on the promoters of PEPCK and G6Pase gene [6]–[9]. These elements serve as platforms for setting up a complex transcriptional machinery that includes transcription factors (e.g., CREB, FOXO1, FOXA2, C/EBPs, HNF4α, GR, Nur77) and co-factors (e.g., PGC-1α, CRTC2, SIRT1, p300/CBP, SRC-1), thereby driving gluconeogenic gene expression [10], [11]. Despite these tremendous progresses, the regulatory mechanisms upstream of this transcriptional network are incompletely understood. Furthermore, it is unclear how the chromatin landscape affects gluconeogenesis, what chromatin modifying enzymes (in addition to p300/CBP) are involved, and how these enzymes coordinate with the aforementioned transcriptional regulators.
One determinant for chromatin structure and functional state is histone methylation that occurs on specific lysine residues in histones [12], [13]. Five lysine residues within the N-terminal tail of histone H3 (K4, K9, K27, and K36) and H4 (K20) have been shown to be the sites for methylation. These lysine residues can be mono-, di-, or trimethylated. Depending on the specific lysine residues and the degree of methylation, histone methylation can have distinct effects on gene expression. In general, histone H3K4 and K36 di-and trimethylation, and H3K27 monomethylation are associated with actively transcribed genes, whereas H3K9 and K27 di- and trimethylation are considered repressive markers for gene expression. The distribution pattern of histone methylation on gene loci can also be quite different. For example, H3K4 and K9 methylation are enriched in the promoter regions, whereas K36 di- and trimethylation are mainly located in the coding regions and their levels peak toward the 3′end of the gene [14]–[16]. By altering chromatin structure, histone methylation fine-tunes transcriptional outputs.
Histone methylation is reversible and its dynamic nature is controlled by a balance between histone methyltransferases and histone demethylases. A number of histone demethylases have been identified in recent years and they are classified into two groups [17]–[20]. The first group contains two genes, LSD1 and LSD2, in human genome. These enzymes catalyze demethylation via an FAD-dependent oxidative reaction that requires protonated nitrogen in the substrate [19]. The second group are genes that contain a JmjC domain. Nineteen members of the JmjC domain-containing proteins in the human genome have been shown to be demethylases. The JmjC domain is the catalytic domain that possesses demethylation activity. These enzymes use Fe(II) and the intermediate metabolite α-ketoglutarate as co-factors to catalyze a hydroxylation-based demethylation [20]. Because of their enzymatic requirement for either FAD or α-ketoglutarate, it has been postulated that histone demethylases might be important for energy homeostasis by linking metabolic signals to chromatin status and transcriptional regulation [21]. Here, through both in vitro and in vivo studies, we reveal an important regulatory function of histone demethylase Jhdm1a in gluconeogenesis that is mediated by its active demethylation on the C/EBPα locus.
To assess whether JmjC domain-containing histone demethylase(s) is involved in the regulation of gluconeogenesis, we treated human hepatoma HepG2 cells with N-oxalylglycine (NOG) or its derivative, dimethyloxalylglycine (DMOG), and examined the expression of the gluconeogenic enzymes. NOG and DMOG are analogues of α-ketoglutarate and are general enzymatic inhibitors of the JmjC domain-containing histone demethylases [22], [23]. Treatment with either compound led to an increase of PEPCK and G6Pase expression (Figure 1A), indicating a potential requirement of histone demethylation activity in the regulation of gluconeogenesis.
We next decided to use shRNA knockdown to identify the involved histone demethylase(s). As both the hormonal and molecular pathways that regulate PEPCK and G6Pase transcription are retained in HepG2 cells, we performed our screening experiments in these cells. We obtained a collection of human lentiviral shRNA constructs against the known JmjC domain-containing demethylases and a few JmjC domain-containing proteins where an enzymatic function has not been ascribed. We also included shRNA constructs against the FAD-dependent histone demethylases (LSD1 and LSD2). We stably expressed individual knockdown constructs in HepG2 cells and screened by quantitative RT-PCR for an increase of PEPCK expression compared with scramble controls. We found that knockdown of the JmjC-domain-containing protein Jhdm1a had the strongest effect (Figure 1B). Jhdm1a is a histone demethylase that specifically demethylates dimethylated H3K36 [20]. Knockdown of this demethylase also robustly promoted G6Pase expression, but not FBP-1 expression (Figure 1C). Not surprisingly, this led to an increase of PEPCK and G6Pase protein levels (Figure 1C). It is likely that NOG-induced PEPCK and G6Pase expression is mediated through inhibition of Jhdm1a, as the induction was lost in Jhdm1a knockdown cells (Figure 1D). As expected, treatment of HepG2 cells with dibutyryl cyclic-AMP and dexamethasone stimulated PEPCK and G6Pase expression; knockdown of Jhdm1a further led to an additive/synergistic increase, indicating a possibility that the effect of Jhdm1a knockdown is independent of the pathway activated by the hormones (Figure 1E). Similar results were obtained in HepG2 cells with an independent Jhdm1a silencing construct (Figure S1). In addition, we generated a lentiviral knockdown construct that targeted mouse Jhdm1a and expressed it in mouse hepatoma HepA1-6 cells. These cells express low level of PEPCK and undetectable level of G6Pase. Silencing of Jhdm1a in these cells elevated PEPCK expression (Figure 1F), while expression of key gluconeogenic transcriptional regulator Foxo1 and PGC-1α was not increased (Figure S2). We next studied gluconeogenic gene expression in a more physiological setting. We knocked down Jhdm1a in primary mouse hepatocytes using adenovirus and found that PEPCK and G6Pase expression was increased as well (Figure 1G). These results collectively demonstrated a negative role of Jhdm1a in gluconeogenic gene expression. Finally, we determined whether Jhdm1a regulates other metabolic pathways in HepG2 cells. We found that knockdown of Jhdm1a did not affect expression of any examined genes involved in lipogenesis, fatty acid oxidation, glycolysis, or glycogenolysis (Figure 1C), suggesting a quite specific metabolic function of Jhdm1a.
Given that knockdown of Jhdm1a elevates gluconeogenic gene expression, we examined whether an opposite effect could be observed in cells expressing Jhdm1a. We stably expressed Jhdm1a via lentivirus in liver cells and found that this expression decreased both basal and hormonal-stimulated levels of PEPCK and G6Pase mRNA (Figure 2A and Figure S3). Interestingly, and in agreement with the knockdown data (Figure 1B), ectopic expression of demethylase Jhdm1b, which is closely related to Jhdm1a, did not inhibit gluconeogenic gene expression (Figure 2B and Figure S4). To determine the domains in Jhdm1a that are required for its suppressive function, we generated a series of Jhdm1a mutants. We first confirmed that these mutants were capable of producing stable proteins at a similar level, as judged by plasmid transfection in Hela cells (Figure 2A). We then expressed the mutants in HepG2 cells through lentivirus with a similar, low infection efficiency. Deletion of the JmjC domain or the CXXC Zinc finger domain abolished the suppression on PEPCK and G6Pase expression, whereas mutant lacking either the PHD domain or the F-box and Leucine-rich repeats remained fully functional (Figure 2A and Figure S3). Note that these Jhdm1a mutants were expressed at similar mRNA levels as their wild-type counterpart. The JmjC domain harbors the histone demethylation activity. Consistent with the effect of the JmjC deletion mutant, a demethylation-dead point mutant (H212A) [20] of Jhdm1a was no longer able to suppress PEPCK and G6Pase expression (Figure 2A and Figure S3). We next determined the effect of Jhdm1a on glucose production in vitro. We found that ectopic expression of wild type Jhdm1a, but not the demethylation defective mutants, inhibited glucose production in rat hepatoma FAO cells (Figure 2C). Taken together, these results demonstrate that both the demethylation activity and the CXXC Zinc finger domain of Jhdm1a are required for its negative modulation of gluconeogenic gene expression.
Based on our discovery of the regulatory role of Jhdm1a in vitro, we tested whether Jhdm1a regulates gluconeogenesis in live animals. We obtained five lentiviral Jhdm1a knockdown constructs from Open Biosystems and tested their knockdown efficiency by RT-QPCR in mouse cell culture. We transferred two best ones into an adenoviral vector, generated adenoviruses, and further confirmed that they were able to reduce ectopically expressed Jhdm1a protein level in vitro (Figure S5). The viruses were infused into the liver of wild-type C57BL/6J mice via tail vein injection and endogenous Jhdm1a expression was decreased, which led to a significantly increase in hepatic expression of PEPCK and G6Pase in both fed and fasting states, compared with the scramble control (Figure 3A, Figure S6 and S7). A corresponding enhanced PEPCK and G6Pase protein production was observed (Figure 3A). Blood insulin levels examined at fed state were not significantly different (Figure 3B). Although the Jhdm1a knockdown mice were still able to maintain normal glycemia, they displayed higher glucose production upon injection of the gluconeogenic substrate pyruvate (Figure 3C). We next ectopically expressed either wild-type Jhdm1a or the H212A point mutant in the liver of diabetic ob/ob mice. Expression of the wild-type Jhdm1a, but not the H212A point mutant, decreased the expression of PEPCK and G6Pase (Figure 3D). Accordingly, we observed a statistically significant reduction of blood glucose level in ob/ob mice expressing wild-type Jhdm1a (Figure 3E). Thus, Jhdm1a indeed has a physiological role in hepatic gluconeogenesis in vivo, and this role is mediated by its histone demethylation activity.
Gluconeogenesis is activated during fasting and suppressed by a meal. Interestingly, the hepatic expression Jhdm1a was not changed during either a short-fasting (5 hr) or a long-fasting (20 hr) (Figure 3F). Furthermore, administration of either glucagon or insulin in vivo revealed no difference in Jhdm1a expression (Figure 3G). Likewise, treatment of HepG2 cells with dibutyryl cyclic-AMP and dexamethasone or insulin had no effect on Jhdm1a expression (Figure S8). Although we cannot rule out the possibility of post-transcriptional regulation of Jhdm1a by hormonal signaling, these data, together with the observed effects of Jhdm1a on PEPCK and G6Pase expression in both non-stimulatory and stimulatory conditions (Figure 1C and 1E, Figure 3A, and Figures S6 and S7), indicate that Jhdm1a acts as a negative regulatory mechanism to fine-tune baseline gluconeogenesis. In diabetic ob/ob mice, Jhdm1a expression was elevated (Figure 3H), possibly reflecting a feedback response.
We explored how Jhdm1a regulates gluconeogenesis. We initially speculated that Jhdm1a might associate with the transcriptional regulator complex on the promoters of PEPCK and G6Pase and directly regulate their expression. To test this idea, we performed chromatin immunoprecipitation experiments in HepG2 cells ectopically expressing HA-tagged Jhdm1a. Unexpectedly, Jhdm1a did not associate with either PEPCK promoter or G6Pase promoter (Figure S9). The promoter regions we examined have been well characterized previously and are subjected to extensive regulation by an array of transcription regulators [6]–[9]. The lack of association of Jhdm1a with PEPCK and G6Pase promoters indicates to us that Jhdm1a might not directly regulate the expression of these two genes. We thus considered a possibility that Jhdm1a instead regulates the expression of any of the involved transcription factors or co-factors [10], [11]. We knocked down Jhdm1a in HepG2 cells and examined their expression. We found that the transcription factor C/EBPα was the only one whose expression level was significantly increased (Figure 4A). Similarly, knockdown of Jhdm1a promoted C/EBPα expression in primary mouse hepatocytes (Figure 4B). As a result of increased C/EBPα level, the association of C/EBPα with its binding sites within the PEPCK and G6Pase promoters was strongly enhanced in Jhdm1a knockdown HepG2 cells (Figure 4C). Members of C/EBPs were shown to activate the expression of PEPCK and G6Pase in vitro [24], [25]. We confirmed these previous results and also observed a remarkably similar target gene expression pattern between Jhdm1a silencing and C/EBPα ectopic expression (comparing Figure 1C and Figure 4D), supporting a functional connection between Jhdm1a and C/EBPα.
Previous work by others has also demonstrated an essential in vivo role of C/EBPα in hepatic PEPCK and G6Pase expression [26]–[29]. Importantly, we found that in vivo knockdown of Jhdm1a in the mouse liver increased the level of C/EBPα (Figure 4E and Figure S10). Conversely, exogenous expression of Jhdm1a in the liver suppressed C/EBPα expression, whereas the H212A point mutant had no effect (Figure 4F). To further examine whether the action of Jhdm1a is C/EBPα-dependent, we knocked down both Jhdm1a and C/EBPα in hepatic cells. The increase of PEPCK and G6Pase expression caused by Jhdm1a knockdown was greatly diminished in the double knockdown cells (Figure 4G). The results together suggest that Jhdm1a regulates gluconeogenesis, at least in part, through its control of C/EBPα expression. As previously noted [11], C/EBPα expression remained unchanged during both short and long fasting (data not shown), in agreement with our observation that Jhdm1a expression was not affected by these conditions.
To identify the molecular mechanism by which Jhdm1a regulates C/EBPα expression, we first examined whether Jhdm1a associates with the C/EBPα locus. The C/EBPα locus contains a single exon. We expressed HA-tagged Jhdm1a in hepatic cells and performed chromatin immunoprecipitation experiments using antibody against the HA tag. We found that Jhdm1a was associated with the C/EBPα promoter region but not with the intragenic region (Figure 5A). Interestingly, this promoter region contains four separate transcription factor USF1 binding sites that have been implicated in C/EBPα expression [30], [31] and Jhdm1a was present on three of them. An interaction between Jhdm1a and USF1 was readily detected in cells expressing both of them (Figure 5B). Moreover, reduction of USF1 level by shRNA-mediated silencing diminished the association of exogenous Jhdm1a with these sites (Figure 5C). Despite the high background of the Jhdm1a antibody, we were also able to show that endogenous Jhdm1a associated with the USF1 binding sites, since knockdown of Jhdm1a decreased its association with these sites (Figure S11). Functionally, knockdown of USF1 led to an increase of C/EBPα expression and accordingly, an increase of PEPCK expression (Figure 5D). These data suggest a model in which USF1 recruits Jhdm1a to the C/EBPα promoter to negatively regulate its expression.
Given the association of Jhdm1a with the C/EBPα promoter, we examined whether Jhdm1a modulates the H3K36 methylation pattern on the C/EBPα locus. Knockdown of Jhdm1a increased H3K36 dimethylation in the 3′ exon region and 3′ UTR that is close to the exon, but had little effect on H3K36 dimethylation on the promoter, 5′ exon region, and 3′UTR that is located far away from the exon (Figure 6A). This pattern of modulation is in concord with the previously shown genome-wide distribution of H3K36 dimethylation where it is mostly found in the intragenic region and usually peaks toward 3′ exon [15]. The demethylation by Jhdm1a is gene-specific, as knockdown of Jhdm1a did not increase H3K36 dimethylation at the C/EBPβ locus (Figure 6B). Moreover, knockdown of Jhdm1a did not affect the H3K36 trimethylation pattern at the C/EBPα locus (Figure 6C), consistent with the enzymatic property of Jhdm1a to specifically demethylate dimethylated H3K36 [20]. Next, we examined the effect of ectopically expressed Jhdm1a on H3K36 dimethylation at the C/EBPα locus. We found that expression of wild type Jhdm1a, but not of the H212A mutant, led to a significant decrease of K36 dimethylation (Figure 6D). These results suggest that Jhdm1a demethylates dimethylated H3K36 at the C/EBPα locus, hence directly regulating its expression.
We determined whether the H3K36 dimethylation status at the C/EBPα locus is modulated by hormonal signaling or metabolic states. In agreement with Jhdm1a expression (Figure 3F and 3G, and Figure S8), we found that levels of H3K36 dimethylation remained unchanged in HepG2 cells treated with hormones (Figure S12) or in livers of fasted mice (Figure 6E), supporting the idea that Jhdm1a and H3K36 dimethylation at the C/EBPα locus are primarily involved in basal control of gluconeogenesis. Interestingly, H3K36 dimethylation was significantly decreased at the exon region of C/EBPα locus in diabetic ob/ob mice (Figure 6F), likely due to increased Jhdm1a expression (Figure 3H). These data suggest a possible physiological, compensatory attempt to suppress hyperglycemia in ob/ob mice.
In recent years, a number of histone demethylases have been identified [17]–[20]. While these exciting discoveries dramatically reversed our previous view that histone methylation was a stable, non-erasable marker, our knowledge regarding the functions of these demethylases in biological processes and diseases is very limited. Here, through an shRNA screen against the known histone demethylases, we identify Jhdm1a negatively regulates gluconeogenic gene PEPCK and G6Pase expression both in vitro and in vivo. Phenotypically, silencing of Jhdm1a elevates glucose production, whereas its ectopic expression lowers blood glucose levels in diabetes. Interestingly, our studies suggest that Jhdm1a does not appear to control PEPCK and G6Pase expression directly. Rather, Jhdm1a exerts its function through C/EBPα. The role of C/EBPα in gluconeogenesis has been well established [25]–[29]. We found that Jhdm1a negatively modulates the expression of C/EBPα through active demethylation on the C/EBPα locus. Therefore, our work potentially uncovers a novel molecular mechanism in gluconeogenesis, where histone demethylation regulates a key gluconeogenic transcription factor. However, it is important to note that our in vivo studies were performed using adenoviral infusion to acutely manipulate hepatic Jhdm1a level, therefore, chronic and more physiological and pathophysiological roles of Jhdm1a in gluconeogenesis remain to be addressed in detail with liver-specific Jhdm1a knockout and transgenic models. In addition, as genetic variations at the Jhdm1a locus are present in human population, it will be interesting to analyze whether these variations are associated with type 2 diabetes.
It was hypothesized that histone demethylases might be important for metabolic homeostasis [21]. This is supported by the obese phenotype of mice deficient for H3K9 histone demethylase, Jhdm2a [32], [33]. Our demonstration of Jhdm1a functioning in gluconeogenesis provides another example. It is anticipated that future studies will reveal additional histone demethylases as important regulators of energy metabolism. Histone demethylases are considered as global modifiers of chromatin structure, however, it is clear that a particular demethylase only regulates a small subset of genes and therefore, a specific metabolic pathway. This specificity is likely to be determined by the target tissue, the repertoire of transcriptional regulators in that tissue, and whether this particular demethylation on individual gene locus is sufficient to translate into a gene expression readout.
Histone H3K36 di- and trimethylation have been shown to be associated with actively transcribed genes and their levels peak near the 3′ end of the gene [14]–[16], [34]. In yeast, K36 di- and trimethylation have been implicated in transcriptional elongation by preventing cryptic, intragenic transcription [35]–[37]. In higher eukaryotes, the exact function of K36 methylation is poorly understood. We show here that Jhdm1a demethylates dimethylated H3K36 on the C/EBPα locus and negatively regulates its expression. Although we cannot rule out the possibility that changes of dimethylated H3K36 level are secondary due to C/EBPα expression, the requirement for the demethylase activity of Jhdm1a and the unaffected H3K36 trimethylation on the C/EBPα locus strongly argue that this is unlikely. Jhdm1b is another demethylase that targets dimethylated H3K36. Jhdm1b-mediated demethylation was recently shown to negatively regulate the expression of the p15Ink4b tumor suppressor [38]. These studies suggest a positive role of H3K36 dimethylation for gene expression.
Our data suggest that Jhdm1a is recruited by USF1 to the USF1-binding sites within the C/EBPα promoter. A recent study shows that Jhdm1a, through its CXXC Zinc finger domain, associates with unmethylated CpG islands on gene promoters [39]. Indeed, the C/EBPα promoter is considerably CpG-rich, and we find that the CXXC Zinc finger domain is required for the suppressive function of Jhdm1a. Therefore, it is possible that the CpG-rich sequences and USF1 cooperatively mediate the recruitment of Jhdm1a. One interesting observation in our study is that Jhdm1a demethylates C/EBPα intragenic region that lacks detectable association. It is possible that the initial recruitment by USF1 to the promoter is a relatively stable state, but following recruitment, Jhdm1a moves along the gene body to demethylate dimethylated H3K36. Thus the association of Jhdm1a with the gene body might be transient and difficult to capture. There are precedents of similar observations. For example, ChIP-seq studies reveal that, for actively transcribed genes, Pol II is predominantly detected at transcription start sites, not transcribed regions [16]. PHF8, a H4K20/H3K9 demethylase, was found to demethylate regions that it does not associate with [40]. Clearly, how epigenetic enzymes are recruited and are able to modify chromatin structure in a widespread fashion is a fascinating question to be fully understood.
While Jhdm1a-catalyzed histone demethylation regulates gluconeogenesis through an indirect mechanism by targeting C/EBPα, a previous report has postulated that dimethylation of histone H3 arginine 17 has a direct impact on gluconeogenic gene expression, as the level of this modification on the PEPCK promoter increases with dexamethasone treatment and decreases upon subsequent addition of insulin [41]. However, the molecular events responsible for and the functional outcome of this change were unknown. Nevertheless, their studies, along with ours, indicate that histone methylation/demethylation could be more commonly employed than we appreciated to regulate gluconeogenesis at multiple layers. To our surprise, Jhdm1a expression, hence the H3K36 dimethylation status at the C/EBPα locus, are not influenced by fed and fasted states and hormonal signaling. Our data indicate that, under normal conditions, Jhdm1a-mediated demethylation primarily function in maintaining basal-state gluconeogenesis irrespective of nutritional and hormonal cues. In support of this model, we found that knockdown of Jhdm1a in mice elevates the expression of C/EBPα, PEPCK and G6Pase in both fed and fasted states. Mechanisms controlling hormonal-regulated gluconeogenesis have been extensively studies [11], less was understood for basal-state gluconeogenesis. Our work provides insights into this key process. Interestingly, in diabetic state, Jhdm1a expression is increased and H3K36 dimethylation at the C/EBPα locus is decreased, indicating a possible involvement of Jhdm1a in counteracting hyperglycemia. Thus, under pathophysiological conditions such as obesity and insulin resistance, the expression and/or activity of Jhdm1a can be modulated by currently unknown mechanisms. In addition, we find that in fetal liver, Jhdm1a is highly expressed and C/EBPα level is very low; in neonatal stage, hepatic Jhdm1a level decreases and C/EBPα level increases (our unpublished data). As gluconeogenesis occurs in neonatal stage but not in embryonic stage, whether Jhdm1a is involved in this metabolic transition during development remains to be determined. In summary, our results illustrate how the dynamics of H3K36 dimethylation regulates basal gluconeogenesis and indicate that increasing the demethylation activity of Jhdm1a could potentially offer therapeutic benefits to curb hyperglycemia.
Lentiviral shRNA constructs (pGIPZ-based; Open Biosystems) against the known human demethylases were obtained through the RNAi Core Facility at University of Massachusetts Medical School. All other lentiviral shRNA constructs were obtained directly from Open Biosystems. All relevant constructs were verified and their targeting sequences are provided in Table S1. Lentiviruses were packaged as described [42]. After virus infection, cells were re-plated next day and selected with puromysin for three days. Cells were then trypsinized and plated at a similar confluency. Cells were cultured in the presence of puromycin for two more days and total RNA was isolated.
Mouse wild-type and mutant Jhdm1a expression plasmids were generated by standard procedure and were fully sequenced. They were then transferred to pENTR-1A vector and recombined with pLenti-CMV/neo to generate lentiviral constructs essentially as described [43]. The titers of packaged lentiviruses were determined in liver cells. Cells were infected with similar number of viral particles, selected with G418, and cultured as above.
Total RNA was extracted with Trizol reagent. Gene expression was measured by quantitative RT-PCR and normalized to internal control genes (β-actin for cells, U36b4 or cyclophilin for liver tissue). Primer sequences are provided in Table S1.
Rat hepatic FAO cells expressing lentiviral Jhdm1a constructs were washed 3 times with PBS and then incubated in glucose free DMEM medium containing 2 mM sodium pyruvate and 20 mM sodium lactate for 6 hr. Glucose levels in the medium were measured with a Amplex red glucose assay kit (Invitrogen, #A22189).
Cells were prepared and cultured as described [24]. Cells were infected with adenoviruses at a multiplicity of infection of 50. Two days after infection, cells were starved for 6 hr in DMEM supplemented with 0.2% BSA and 2 mM sodium pyruvate before RNA isolation.
Adenoviral Jhdm1a expression and knockdown constructs and their respective control constructs were generated, and adenoviruses were produced and purified as described [42], [44]. Viral titers were determined in HEK293 cells by scoring GFP positive cells. Male wild-type C57BL/6J and ob/ob (on C57BL/6J background) mice were obtained from The Jackson Laboratory. Adenoviruses (4×109 and 9×109 viral particles for expression and knockdown, respectively) suspended in 0.2 ml PBS were injected through tail vein when animals were 10-week-old. Blood glucose levels were measured at indicated time and animals were sacrificed at Day 5. For pyruvate tolerance test, mice were fasted for 16 hr and sodium pyruvate dissolved in PBS was i.p. injected (2 g/kg body weight).
To determine the levels of liver PEPCK, G6Pase and C/EBPα protein, 50 mg liver sample were homogenized in 1 ml lysis buffer [100 mM NaCl, 50 mM Tris (pH 7.5), 0.5% Triton X-100, 5% (w/v) glycerol]. 26 µg protein extracts were separated by SDS-PAGE and probed with antibody against C/EBPα (Santa Cruz, sc-61), PEPCK (ABcam, ab28455) or G6Pase (Santa Cruz, sc-25840).
HA-Jhdm1a and Flag-USF1 plasmids were co-transfected into Hela cells. Cells were lysed in buffer [100 mM NaCl, 50 mM Tris (pH 7.5), 0.5% Triton X-100, 5% (w/v) glycerol]. Cell extracts were incubated with anti-HA beads (Santa Cruz, sc-7392AC) for overnight and the beads were washed 4 times with buffer [100 mM NaCl, 50 mM Tris (pH 7.5), 0.1% Triton X-100, 5% glycerol]. Immunoprecipitates were probed with an anti-Flag antibody (Sigma, F7425).
Assays were performed as described [42] using antibodies against HA (Sigma, #H6908), C/EBPα (Santa Cruz, sc-61), dimethyl-H3K36 (Millipore, #07274), trimethyl-H3K36 (Abcam, #9050). Immunoprecipitate signal was normalized with input signal; both were measured by real-time QPCR. Primer sequences are provided in the Table S1.
For ChIP assays performed with liver samples, samples were generated as described with minor modifications [45]. Briefly, parts of liver from same locations were excised, cut into small pieces with a razor blade, cross-linked with 1% formaldehyde for 15 minutes at room temperature. The samples were then ground and filtered through a 40 µm cell strainer to produce a single liver cell suspension. Nuclear extracts were prepared, chromatin was sonicated using a ultrasonic processor, and immunoprecipitation was performed as described [42]. An equivalent of 40 mg of liver tissue was used for each immunoprecipitaiton. After normalized with inputs, ChIP signals were calculated as folds relative to background signal (IgG) generated from the same animal.
Student's t test (two-tailed) was used for statistical analysis. P<0.05 was considered significant. Data are presented as mean ± s.e.m.
|
10.1371/journal.pbio.1000248 | Patterning of the Dorsal-Ventral Axis in Echinoderms: Insights into the Evolution of the BMP-Chordin Signaling Network | Formation of the dorsal-ventral axis of the sea urchin embryo relies on cell interactions initiated by the TGFβ Nodal. Intriguingly, although nodal expression is restricted to the ventral side of the embryo, Nodal function is required for specification of both the ventral and the dorsal territories and is able to restore both ventral and dorsal regions in nodal morpholino injected embryos. The molecular basis for the long-range organizing activity of Nodal is not understood. In this paper, we provide evidence that the long-range organizing activity of Nodal is assured by a relay molecule synthesized in the ventral ectoderm, then translocated to the opposite side of the embryo. We identified this relay molecule as BMP2/4 based on the following arguments. First, blocking BMP2/4 function eliminated the long-range organizing activity of an activated Nodal receptor in an axis rescue assay. Second, we demonstrate that BMP2/4 and the corresponding type I receptor Alk3/6 functions are both essential for specification of the dorsal region of the embryo. Third, using anti-phospho-Smad1/5/8 immunostaining, we show that, despite its ventral transcription, the BMP2/4 ligand triggers receptor mediated signaling exclusively on the dorsal side of the embryo, one of the most extreme cases of BMP translocation described so far. We further report that the pattern of pSmad1/5/8 is graded along the dorsal-ventral axis and that two BMP2/4 target genes are expressed in nested patterns centered on the region with highest levels of pSmad1/5/8, strongly suggesting that BMP2/4 is acting as a morphogen. We also describe the very unusual ventral co-expression of chordin and bmp2/4 downstream of Nodal and demonstrate that Chordin is largely responsible for the spatial restriction of BMP2/4 signaling to the dorsal side. Thus, unlike in most organisms, in the sea urchin, a single ventral signaling centre is responsible for induction of ventral and dorsal cell fates. Finally, we show that Chordin may not be required for long-range diffusion of BMP2/4, describe a striking dorsal-ventral asymmetry in the expression of Glypican 5, a heparin sulphated proteoglycan that regulates BMP mobility, and show that this asymmetry depends on BMP2/4 signaling. Our study provides new insights into the mechanisms by which positional information is established along the dorsal-ventral axis of the sea urchin embryo, and more generally on how a BMP morphogen gradient is established in a multicellular embryo. From an evolutionary point of view, it highlights that although the genes used for dorsal-ventral patterning are highly conserved in bilateria, there are considerable variations, even among deuterostomes, in the manner these genes are used to shape a BMP morphogen gradient.
| During early development of many organisms, patterning along the dorsal-ventral axis is regulated by the activities of two signaling centers located on the ventral and dorsal sides of the embryo. One of these centers produces growth factors of the BMP family that act as morphogens, whereas the other center secretes BMP antagonists such as Chordin that regulate the flow of BMPs along the dorsal-ventral axis. Expression from these two signaling centers results in roughly complementary distributions of BMP and BMP antagonist. We have analyzed BMP-mediated dorsal-ventral axis patterning in embryos of sea urchins, which are phylogenetically close to vertebrates and extensively rely on cell-cell interactions for their development. We found that in sea urchins, unlike in most organisms, the activity of a single signaling center located on the ventral side is responsible for generating both the ventral and the dorsal sides of the embryo. In addition, we discovered that the BMP2/4 gene is co-expressed with Chordin in this ventral center but that the BMP2/4 protein is translocated to the opposite side of the embryo where it activates the genetic program responsible for dorsal differentiation. Our study reveals an unusual example of signaling at a distance by a BMP growth factor. It also highlights that although the proteins used for dorsal-ventral patterning are evolutionarily conserved, there are considerable variations in the manner in which these proteins can be used in different species to generate a gradient of BMP morphogen.
| Genetic and molecular studies carried out in vertebrates and invertebrates have shown that dorsal-ventral (D/V) patterning in bilaterians is regulated by a remarkably conserved patterning system which relies on production of secreted BMP inhibitors such as Chordin (Sog in Drosophila) which antagonize the activity of BMP signals (Dpp in Drosophila) resulting in a gradient of BMP activity along the D/V axis [1]–[4]. Accordingly, in all the bilaterians where expression of chordin and BMP2/4 has been analyzed, these genes show complementary expression in opposite territories along the D/V axis leading to the important concept that dorsal and ventral cells communicate by signals emanating from dorsal and ventral signaling centers [5]. Intriguingly, although the function of these genes has been conserved between vertebrates and invertebrates, their expression pattern along the D/V axis are inverted, such that dpp is expressed dorsally in invertebrates while BMP2/4 is expressed ventrally in vertebrates, suggesting that an inversion of the D/V axis has occurred in the course of evolution [6],[7].
Echinoderms offer interesting models for the comparative analysis of the mechanisms of axis specification. Their position at the basis of the deuterostome lineage makes them a valuable model to reconstruct the evolution of deuterostomes from an ancestral metazoan and to understand how the chordate body plan emerged (Figure 1A). One peculiar feature of echinoderms is that most of them develop indirectly, i.e. the adult, which emerges through a metamorphosis and displays very little similarity with the larva, is built from groups of cells that are set aside during embryogenesis [8]. Although the adult body plan of echinoderms is typically radial, this feature is a recent modification of a bilateral body plan and echinoderm larvae are indeed bilaterally symmetrical as are larvae of their closest relatives, the hemichordates, which undergo no axial remodeling. Indeed, it has been recognized for a long time that larvae of indirectly developing echinoderms and larvae from directly developing hemichordates share many morphological features [9]. It is therefore most likely that despite considerable differences in their adult body plans and mode of development, these homologous larval features of echinoderms and hemichordates reflect utilization of similar genes and gene regulatory mechanisms during embryogenesis. Since the ambulacraria form a sister group of the chordates, it is also reasonable to postulate that the genes, transcription factors, and signaling pathways that orchestrate embryonic axial patterning in chordates may have a conserved role and echinoderms. This idea is largely supported by recent findings showing that echinoderms use many of the same regulatory genes and signaling pathways employed by more complex bilateria. The conserved role of the Wnt/beta catenin in animal vegetal patterning [10],[11] as well as the conserved role of the Nodal/Univin/vg1 in D/V patterning [12],[13] and in left right axis determination [14] even suggests a conservation at the wider gene regulatory network level.
The D/V polarity of the sea urchin larva first becomes morphologically apparent at the early gastrula stage when bilateral clusters of skeletogenic mesenchyme cells form on the presumptive ventral side (Figure 1B). Later during gastrulation, the larva flattens on the presumptive ventral side and the archenteron bends towards the presumptive ventral ectoderm where the mouth will open. As morphogenesis continues, D/V polarity becomes even more prominent with the larva elongating along the D/V axis and acquiring its characteristic easel-like shape. At the pluteus stage, the ectoderm is divided into two main territories: the ventral (or oral) ectoderm and the dorsal (or aboral) ectoderm separated by a belt of cuboidal ciliated cells and neurons, the ciliary band (Figure 1B).
Classical studies revealed that D/V axis formation in sea urchin embryos extensively relies on cell interactions [15]. In one of the most famous experiments of embryology, Hans Driesch demonstrated that after isolation, each of the four blastomeres of a four-cell stage embryo can give rise to a normal pluteus larva with a perfectly normal D/V axis. Not only did this experiment demonstrate the astonishing regulative capacity of sea urchin embryos, but it also showed that the D/V axis is not fixed in the early embryo and that formation of this axis must therefore rely on cellular communication.
Recent studies have shown that the TGF-β ligands Nodal, Vg1/Univin, and BMP2/4 all play crucial roles in these cell interactions [12],[13],[16],[17]. Expression of nodal is initiated around the 32–60-cell stage in most cells of the presumptive ectoderm and is rapidly restricted to the presumptive ventral ectoderm. nodal is the earliest zygotic gene found to display a restricted expression along the D/V axis during sea urchin development, and Nodal function appears to be required for all aspects of D/V polarity of the embryos [16]. When translation of the Nodal transcript is prevented, specification of both the ventral and the dorsal ectoderm fails and most of the ectoderm (except the ectoderm derived from the animal pole and vegetal pole regions) differentiates into a thickened ciliated ectoderm that shares features with the neurogenic territory of the ciliary band. The ciliary band may thus represent the default fate of most of the ectoderm in the absence of Nodal [16]. Functional studies have shown that overexpression of Nodal causes most ectodermal cells to adopt a ventral fate. Even more strikingly, injection of synthetic nodal mRNA into one blastomere at the eight-cell stage is sufficient to fully rescue axis formation in nodal morpholino injected embryos. It is thus clear that Nodal expressing cells have a long-range organizing activity and are capable of restoring D/V polarity over the whole embryo [14],[16]. Consistent with these observations, sea urchin Nodal activates a regulatory network of genes encoding key transcription factors such as FoxA, Goosecoid, and Brachyury as well as signaling molecules such as BMP2/4 and the Nodal antagonist Lefty, which restricts Nodal activity [12],[14],[18].
Although the organizing activity of Nodal expressing ventral cells is the key to D/V patterning in the sea urchin, the nature of the long-range signal that specifies the dorsal territory is not understood. Theoretically, the non-autonomous effect of Nodal could either be direct, with Nodal working as a long-range morphogen, or indirect, via induction of a second relay molecule. A direct role for Nodal is supported by previous studies in zebrafish and Xenopus demonstrating that secreted Nodal and Activin ligands can act as morphogens and diffuse over long distances to specify different cell fates [19],[20]. It is thus possible that in the sea urchin embryo, Nodal could diffuse from its site of production and specify the ventral and dorsal ectoderm territories in a dose-dependent manner. Various observations indicate, however, that Nodal may not act directly to specify dorsal cell fates but may induce production of a diffusible relay molecule. First, it has been shown that the range of action of Nodal ligands is limited by the activity of Lefty antagonists [18],[21]–[25]. Second, studies on the inhibitory effects of Nodal on specification of neural fates at the animal pole of the embryo have indicated that the effects of Nodal are short range and restricted to only a few cell rows away from the Nodal expressing territory [26],[27]. Finally, an attractive candidate for a relay molecule exists in BMP2/4 [16],[17]. BMP2/4 is expressed zygotically downstream of Nodal signaling from the 128-cell stage to the prism stage, with its transcripts detected exclusively on the presumptive ventral side of the embryo. Despite being ventrally expressed, BMP2/4 appears to be required on the opposite side of the embryo for expression of dorsal marker genes such as tbx2/3 and the novel transmembrane protein 29D [16]. Although BMP2/4 is thus an appealing candidate for a relay molecule downstream of Nodal, this hypothesis raises important issues concerning the evolution of the D/V patterning system in bilateria, since both Dpp in Drosophila and BMP2/4 in vertebrates act on the side of the embryo where they are expressed and not on the opposite side as predicted in the sea urchin.
In this study, we have analyzed the mechanism by which Nodal and BMP2/4 ligands pattern the D/V axis of the sea urchin embryo. We provide evidence that Nodal does not work as a long-range morphogen along the D/V axis but requires a relay molecule that we confirmed as BMP2/4. We first show that microinjection of an activated form of the Nodal receptor non-autonomously rescues dorsal structures in nodal morpholino injected embryo by inducing BMP signaling in cells located on the opposite side of the embryo. Second, by using a phospho-Smad immunostaining, we showed that, although BMP2/4 is expressed on the ventral side, BMP2/4 signaling is active exclusively on the dorsal side of the embryo, far away from its site of production. Third, we showed that the functions of BMP2/4 and of its putative receptor Alk3/6 are not required for specification of ventral cell fates but are essential for specification of dorsal cell fates. In addition, we have examined the expression and function of sea urchin Chordin and found it to be necessary to prevent BMP signaling on the ventral side. In contrast with the situation in other bilaterian models, however, the territories expressing chordin and BMP2/4 are not complementary but are congruent. Thus, although sea urchin embryos have a conserved BMP-Chordin axis, it is manifest by the activity of these molecules but not their site of production. Finally, we report that Chordin may not be required for diffusibility of BMP2/4 and that BMP2/4 signaling induces a positive feedback mechanism in responsive cells by inducing the expression of glypican 5, a positive regulator of BMP signaling and mobility. Based on our detailed dissection of D/V patterning in the sea urchin embryo, we propose a new model and discuss the evolutionary implications of these findings.
In order to test whether Nodal works as a morphogen along the D/V axis or through a relay molecule, we used an axis rescue assay [16]. We showed previously that microinjection of synthetic mRNA encoding the secreted Nodal ligand into one animal blastomere at the eight-cell stage fully rescues the D/V axis of embryos pre-injected with an antisense nodal morpholino oligonucleotide at the egg stage. To determine whether a long-range non-autonomous activity of Nodal is required for the rescue to be effective, we made use of the sea urchin Nodal receptor Alk4/5/7 [13]. To activate Nodal signaling cell autonomously, we constructed an activated form of this receptor by following a strategy similar to that used to make an activated version of the zebrafish Nodal receptor TARAMA, mutating the glutamine residue found in position 265 into an aspartic acid (see Figure S1–S3) [28] and tested if injection of mRNA encoding this activated Nodal receptor into an animal blastomere at the eight-cell stage was able to rescue the radialized phenotype of nodal morphants (Figure 2A). Strikingly, all embryos (n>50) that received the alk4/5/7QD mRNA into an animal blastomere developed into normal pluteus larvae with a harmoniously patterned D/V axis indicating that the activated Nodal receptor was just as efficient as the Nodal ligand in this rescue experiment (Figure 2Bi, ii, vi, vii). Even embryos injected into a vegetal blastomere that, according to the fate map, will contribute to the endomesoderm as well as to the most vegetal part of the ectoderm were rescued to a considerable extent (Figure S1). As observed in the case of the rescue with the Nodal ligand, the progeny of the blastomere injected with alk4/5/7QD mRNA was always found on the ventral face of the rescued pluteus larvae, consistent with the idea that Nodal signaling promotes ventral fates (Figure 2Bii, vii). More importantly, this result indicates that Nodal signaling is not required outside the ventral ectoderm to specify the dorsal ectodermal fates, implying that Nodal induces a relay molecule that is in turn responsible for specification of the dorsal ectoderm.
We then tested if BMP2/4 function is required for the rescue of dorsal structures by ectopic activation of the Nodal pathway in nodal morphants (Figure 2Biii–v, viii–x). Embryos were injected with both nodal and BMP2/4 morpholinos at the one-cell stage and alk4/5/7QD was injected into one blastomere at the eight-cell stage. Control embryos injected with a mixture of the BMP2/4 morpholino and the nodal morpholino were radialized and undistinguishable from embryos injected with the nodal morpholino alone (Figure 2Bxii–xv) consistent with BMP2/4 being a downstream target of Nodal. Injection of alk4/5/7QD into these double-morpholino embryos failed to rescue a full D/V axis, producing embryos that were polarized but that failed to elongate on the dorsal side (Figure 2Biii, iv, viii, ix). The arms of the pluteus larva failed to form, however the ventral ectoderm was specified in these embryos as shown by the expression of goosecoid in the clone of injected cells and by the asymmetrical positioning of the gut close to the alk4/5/7 expressing cells (Figure 2Bv, x) (n = 17). On the presumptive dorsal side, the epithelium remained thick, and typically, ectopic spicules were present (Figure 2Biii). These embryos resemble the BMP2/4 morpholino injected embryos (Figure 2Bxiv). We conclude that in the double nodal and BMP2/4 morpholino injected embryos, ectopic expression of alk4/5/7QD was able to restore ventral cell fates but failed to restore the dorsal cell fates—in other words, that BMP2/4 expression is required in the ventral ectoderm downstream of Alk4/5/7 to induce dorsal cell fates.
We had previously reported that inhibition of BMP2/4 mRNA translation by morpholino injection causes embryos to develop with a severe radialized phenotype [16]. A careful re-examination of this phenotype revealed several interesting features. First, injection of low concentrations (0.25 mM) of this morpholino frequently caused embryos to develop with a truncated dorsal region (Figure 3B, 3H) leaving the ventral region relatively unaffected, while injection of higher doses (0.4 mM) resulted in severe radialization (Figure 3A, 3G) as reported previously. This suggested that specification of the dorsal most region of the embryo is highly sensitive to BMP2/4 inhibition and that the lateral region of the embryo, from which the arms emerge, is less sensitive to reductions of BMP2/4 function. Second, we noted that although all the embryos injected at 0.4 mM appeared radialized, they retained a D/V polarity as shown by the opening of the mouth and by the bending of the gut towards one side of the embryo (Figure 3A, 3G). Thus, even high doses of this morpholino do not block differentiation of the ventral ectoderm. Intriguingly, most of the ectoderm covering these embryos was made of a thick ectoderm that resembled the ciliary band of pluteus larvae and multiple ectopic spicules frequently developed in association with this thickened ectoderm (Figure 3A, 3G).
To further investigate BMP2/4 function and determine if it fulfills the requirements for a relay molecule that acts downstream of Nodal to induce dorsal fates, we cloned and characterized one of its putative receptors. The genome of Strongylocentrotus purpuratus contains two genes encoding type I BMP receptors, alk3/6 and alk1/2, homologous to the vertebrate BMP type I receptors. Previous phylogenetic analyses indicated that Sp-Alk3/6 is mostly related to the Drosophila Thickveins receptor, the primary receptor for Dpp, while Sp-Alk1/2 is mostly related to Drosophila Saxophone, which preferentially binds Screw (Scw), a distantly related BMP ligand [29]–[35]. We isolated a cDNA encoding the Paracentrotus lividus Alk3/6 (Figures S2 and S3) and found that alk3/6 transcripts are expressed maternally and ubiquitously at all stages (unpublished data). This suggests that all cells of the sea urchin embryo are likely competent to respond to BMP signaling via this receptor and that spatial and temporal restriction of these signals is most likely due to the availability of a restricted source of ligands. Embryos injected with an antisense oligonucleotide morpholino directed against the translation initiation site of the alk3/6 transcript developed apparently normally up to the prism stage but then displayed a characteristic and highly penetrant phenotype. At 72 h, while the control embryos reached the pluteus stage, the injected embryos failed to elongate along the D/V axis and showed ectopic spicules associated with a thickened epithelium reminiscent of the ciliary band (Figure 3C, 3I). Importantly, the squamous epithelium that is normally found on the dorsal side never formed in alk3/6 morpholino embryos suggesting that specification of the dorsal ectoderm had failed. As in the case of BMP2/4, high concentration (0.8 mM) of this morpholino caused a complete radialization while lower concentrations (0.6 mM) prevented development of the dorsal side but allowed formation of a pair of ventral arms (Figure 3D, 3J). Also, as observed in the case of BMP2/4 morphants, these embryos retained some polarity along the D/V axis as indicated by the opening of the mouth, the asymmetric positioning of the gut on one side of the embryo, and the asymmetric morphogenesis of the ectoderm. This set of phenotypes, i.e. failure of the embryo to elongate, presence of ectopic spicules underneath a thickened ectoderm, is strikingly similar to the set of phenotypes observed following inhibition of BMP2/4 translation (see Duboc et al. 2004 [16] and, in this paper, Figure 2Bxiv and Figure 3A, 3G). To test the specificity of the alk3/6 morpholino, we performed a rescue experiment. Eggs were first injected with the alk3/6 morpholino, then later in the first cell cycle re-injected with a synthetic alk3/6 mRNA containing nine mismatches (alk3/6 mm) in the sequence recognized by the morpholino (Figure 3M–O). While injection of the alk3/6 morpholino caused embryos to develop with the radialized phenotype described above (Figure 3P, 3Q), half of the embryos that subsequently received injection of the alk3/6 mm mRNA developed with well developed dorsal arms and dorsal sides (Figure 3M, 3O).
We next constructed an activated form of the Alk3/6 receptor, using the same strategy as for the Alk4/5/7 receptor, mutating Glutamine in 230 into an Aspartic acid (Figure S2B). Embryos injected with mRNA encoding this mutant receptor (Alk3/6QD) developed with a completely radialized phenotype and lacked all D/V polarity at 48 h (Figure 3E). Except for the animal pole, where a thick proboscis formed, all the ectoderm of these embryos differentiated into a thin, squamous, and pigmented epithelium typical of the dorsal ectoderm. This phenotype is largely similar to the phenotype obtained by overexpressing BMP2/4 (Figure 3F), which was shown to reflect re-specification of most of the ectoderm into dorsal ectoderm [16]. Strikingly, starting at 72 h, long unbranched spicules strongly resembling the spicules normally found on the dorsal side of wild type embryos started to form in both the BMP2/4 and Alk3/6QD overexpressing embryos (Figure 3K, 3L). Taken together these results strongly suggest that signals transduced by the Alk3/6 receptor are required to induce dorsal ectoderm fates probably downstream of BMP2/4 binding.
To compare the phenotypes caused by inhibition of BMP2/4 and Alk3/6 at the molecular level, we examined the expression of tbx2/3, goosecoid, and onecut/hnf6 transcripts, normally expressed in ectodermal domains corresponding, respectively, to the dorsal, ventral, and ciliary band territories (Figure 4B) [16],[36]–[40]. Blocking BMP2/4 function or Alk3/6 function abolished the expression of tbx2/3 (Figure 4A–C) but did not affect the ventral marker gene goosecoid (Figure 4D–I). In contrast, expression of onecut/hnf6, which is expressed in the neurogenic territory of the ciliary band, was dramatically affected following inhibition of BMP signaling. Both the alk3/6 and BMP2/4 morpholinos caused an expansion of onecut/hnf6 in the territory where tbx2/3 failed to be expressed (Figure 4J–R). This prompted us to examine the expression of Delta, a marker gene expressed in neurons of the larva. Expression of Delta, which is normally restricted to individual neurons of the ciliary band and facial ectoderm, was dramatically expanded to the thickened ectoderm region facing the ventral side in the BMP2/4 and alk3/6 morphants (Figure 4S–U). This strongly suggested that the dorsal ectoderm adopted the fate of the ciliary band, or in other words that the dorsal ectoderm failed to form and was replaced by a more lateral ectoderm. To confirm this hypothesis, we examined the expression of the sm30 gene, which is normally expressed in the bilateral clusters of skeletogenic mesenchymal cells that form at the level of the presumptive ciliary band (Figure 1) [41]. Indeed, sm30 was expressed ectopically in clusters of skeletogenic cells underlying the thickened ectoderm of BMP2/4 and alk3/6 morphants (Figure 4V–X). This supports our previous suggestion that in the absence of Nodal, most of the ectoderm adopts a default fate and differentiates into a neurogenic ectoderm expressing gene markers of the ciliary band [16],[27]. Importantly, the phenotypes obtained by blocking BMP2/4 or Alk3/6 are indistinguishable both at the morphological and molecular level, indicating that Alk3/6 function is essential to transduce BMP2/4 signaling.
The experiments described above strongly suggest that BMP2/4 acts via Alk3/6 to specify the dorsal ectoderm of the embryo. To address whether ventrally expressed BMP2/4 diffuses from the ventral side where it is synthesized and activates Alk3/6 on the opposite side of the embryo or if it induces a second relay molecule in the ventral ectoderm, we sought to determine where in the embryo the BMP pathways are active by performing an anti-phospho-Smad immunostaining. In the sea urchin, there are only two genes encoding R-Smads: Smad1/5/8 and Smad2/3 [29]. To visualize the phosphorylated form of these Smad factors, we used an antibody directed against the phosphorylated form of human Smad5. On Western blot of P. lividus protein extracts, this antibody recognized predominantly one protein, the abundance of which increased abruptly at the mesenchyme blastula stage (Figure 5A). By immunostaining, this antibody detected asymmetrical nuclear phospho-Smad staining in roughly one half of the embryo starting at the onset of ingression of the primary mesenchyme cells (Figure 5Bi, ii, v). Interestingly, this pSmad staining appeared graded with the highest levels being detected in the dorsal midline (Figure 5Bviii–x). This staining intensified at the mesenchyme blastula and gastrula stages with strong signaling being detected in one half of the embryo in all three germ layers (Figure 5Biii, iv, vi, vii) and peak levels in the dorsal midline. In embryos in which BMP2/4 function was blocked, this pSmad staining was totally lost (Figure 5Ciii, vi), while in the alk3/6 morpholino injected embryos, the pSmad staining was strongly reduced, with only a residual staining being detected in patches of cells (Figure 5Cii, v). Since at the mesenchyme blastula stage BMP2/4 morphants do express nodal [16], the absence of pSmad staining in the BMP2/4 morphants together with the absence of pSmad staining at the early blastula stages in wild type embryos indicate that, in P. lividus, this antibody recognizes specifically the phosphorylated form of Smad1/5/8.
To test directly if Nodal activates BMP signaling within the ventral ectoderm or if it induces a BMP signal that acts as a relay to specify the dorsal side, we performed pSmad staining on nodal morpholino injected embryos rescued by ectopic expression of alk4/5/7QD into one blastomere at the eight-cell stage. As expected, the pSmad staining was abolished in nodal morpholino injected embryos (Figure 6Aiii), consistent with the absence of BMP2/4 transcripts in the nodal morphants [16]. In contrast, pSmad staining was restored in all the embryos injected with the nodal morpholino and later rescued with the activated alk4/5/7 mRNA (Figure 6Aiv–vi). Most strikingly, this pSmad staining was detected in a territory located not on the same but on the opposite side of the Alk4/5/7 expressing clone, the progeny of which was later restricted to the ventral side (Figure 6Aiv). Combined immunostaining with this antibody together with in situ hybridization with dorsal and ventral marker genes confirmed that in all cases, the pSmad nuclear staining was found on the opposite side of the territory expressing the ventral marker goosecoid (Figure 6Bi, ii) and on the same side as the territory expressing the dorsal marker tbx2-3 (Figure 6Biii, iv). Interestingly, a number of dorsally expressed marker genes including tbx2/3 and msx [42],[43] are expressed in nested partially overlapping patterns along the D/V axis (Figure 6C Lepage et al. unpublished data). The finding that pSmad1/5/8 staining is graded along the D/V axis and the nested expression of BMP2/4 target genes such as tbx2/3 and msx both suggest that BMP2/4 may act as morphogen and may regulate different target genes as a function of its concentration along the D/V axis (Figure 6D).
Taken together, these results indicate that cells that receive a high level of BMP signal in the embryo are not located on the ventral side but on the dorsal side of the embryo. They also show that activation and translocation of pSmad1/5/8 into the nucleus in these dorsal cells is dependent on the Alk3/6 receptor and on the BMP2/4 ligand. They reveal that Nodal induces a BMP signal in the ventral ectoderm that activates BMP signaling in dorsal cells and therefore that in the sea urchin embryo, the territories where the signal is expressed and where the signal is received are on the two opposite sides of the embryo.
The finding that BMP2/4 is produced on the ventral side but that BMP signaling is active on the dorsal side raised two intriguing questions: what prevents BMP2/4 from signaling in ventral cells and what promotes diffusion of BMP2/4 towards the dorsal side? In other systems, secreted BMP inhibitors such as Chordin in vertebrates or Sog in Drosophila are key regulators of the activity of BMP signals that sequester BMP ligands and prevent them from binding to and activating their receptors while allowing them to diffuse over long distances [1]. We thus isolated a P.lividus cDNA encoding Chordin and examined the expression of this gene during early development of the sea urchin embryo. Like the vertebrate Chordin and Drosophila Sog proteins, the sea urchin Chordin protein contains a hydrophobic leader sequence and four cysteine rich regions that are present in the same relative positions (Figure 7A and Figure S4). Northern blot analysis indicated that chordin is a strictly zygotic message that starts to accumulate in the embryo after hatching. Its expression peaks at the mesenchyme blastula stage, then decreases progressively during gastrulation and pluteus stages (Figure 7B). Chordin transcripts are downregulated in embryos treated with the vegetalizing agent lithium or in cultures of dissociated blastomeres but are overexpressed in embryos treated with the ventralizing agent nickel chloride (Figure 7B). Consistent with these observations, in situ hybridization revealed that during blastula and gastrula stages, the chordin (Figure 7Cii–x) and BMP2/4 genes (Figure 7Cxi–xiv) are transcribed in highly similar patterns within the ventral ectoderm, the BMP2/4 territory being slightly larger than the chordin expression domain at mesenchyme blastula and gastrula stages. Starting at the pluteus stage however, BMP2/4 expression shifted to the dorsal skeletogenic mesenchymal cells while chordin remained expressed in a subdomain of the ciliary band (Figure 7Cx, xv). Like all the genes expressed in the ventral ectoderm of the sea urchin embryo identified so far, chordin is under the control of Nodal signaling: chordin transcripts are absent from embryos injected with nodal morpholinos and are ectopically expressed throughout the ectoderm following activation of the Nodal pathway (Figure 7D). We conclude that chordin and BMP2/4 are expressed downstream of Nodal signaling. Thus, unlike in most organisms in which Chordin and BMPs are expressed in mutually exclusive regions, in the sea urchin embryo, their expression territories are congruent.
To test if Chordin is required to restrict BMP signaling to the dorsal side, we attempted to block its function with antisense morpholino oligonucleotides (Figure 8). Injection of Morpholinos oligonucleotides directed against the translation initiation codon of chordin (Mo1 chordin) produced a range of phenotypes of various severity. The most severely affected embryos were either completely radialized as shown by the presence of multiple ectopic spicules (Figure 8Ai, vi) or displayed profound defects in the establishment of D/V polarity such as the absence of ventral arms and dorsal apex but still retained a D/V polarity (Figure 8Aii, vii). The milder phenotypes were characterized by a poorly differentiated animal region that conserved a round shape and the absence or underdevelopment of the oral arms. A very peculiar and characteristic skeletal defect was observed in these embryos: the body rods spicules grew parallel instead of forming a triangle in the dorsal region, giving the embryos a luge-like shape (Figure 8Aiii, viii). These phenotypes (abnormal patterning of the ectoderm in the animal hemisphere and parallel growth of the spicules) were consistently observed following injection of a second non-overlapping morpholino directed against the 5′ UTR of the chordin transcript (Mo2 chordin) (Figure 8Aiv, ix). Interestingly this phenotype is also identical to the phenotype obtained following partial inhibition of Alk4/5/7 function using low doses of the SB431542 inhibitor, which strongly reduce chordin expression (Figure 8Av, x and unpublished data).
To test if Chordin is required to prevent BMP2/4 signaling on the ventral side, we monitored the distribution of pSmad1/5/8 in wild type embryos and in embryos injected with morpholino oligonucleotides against the chordin transcript (Figure 8B). In wild type embryos at mesenchyme blastula stage, the pSmad1/5/8 pattern was highly reproducible and staining was consistently detected in half of the embryo corresponding to the dorsal region. In embryos injected with either the morpholino directed against the translation start site or the morpholino directed against the 5′ UTR, however, a dramatic expansion of the pSmad1/5/8 pattern was observed such that staining was now detected both in dorsal and in ventral cells (Figure 8Bii, iii). Consistent with the expansion of the pSmad1/5/8 staining, these embryos showed an expansion of tbx2/3 expression at mesenchyme blastula stages (Figure 8Cii, iii, vi). This expansion of tbx2/3 expression was also visible at gastrula stages, although it appeared less dramatic at this stage, suggesting the existence of compensatory regulatory mechanisms (Figure 8Bvii). This result indicates that one function of Chordin is to prevent BMP2/4 from binding to and activating its receptor in the ventral ectoderm. Consistent with this idea, we found that the sea urchin Chordin protein has a very potent dorsalizing activity when overexpressed in zebrafish embryos mimicking the swirl mutant phenotype caused by inactivation of bmp2 (Figure S5) [44]. To further test if Chordin contributes to the establishment of the pSmad1/5/8 pattern, we overexpressed it by mRNA injection into the egg. Overexpression of chordin strongly affected D/V polarity (Figure 8Axi–xv). In 30% of the embryos it inhibited development of the dorsal region and caused a strong radialization as indicated by the formation of ectopic spicules throughout the circumference of the embryo (Figure 8Axii, xiii). The remaining embryos developed with a bilateral symmetry, but development of the ventral arms and dorsal apex of the larva were reduced (Figure 8Axiv, xv). pSmad1/5/8 staining (Figure 8Biv) and tbx2/3 expression (Figure 8Civ) was eliminated in these embryos.
Taken together these results demonstrate that in the sea urchin embryo as in other bilaterian models, Chordin plays a key role in secondary axis specification through the spatial restriction of BMP2/4 signaling.
That the cells expressing the BMP2/4 ligand are located on the ventral side while the cells receiving the signal are located on the dorsal side strongly implied that secreted BMP2/4 ligand could act over a long distance.
To test this possibility and to investigate the role of Chordin in BMP2/4 diffusion, we injected mRNA encoding either Alk3/6QD, the activated BMP receptor, or the secreted BMP2/4 ligand into one blastomere at the two-cell stage. Then at the mesenchyme blastula stage, embryos were fixed and immunostained with the anti-phospho-Smad1/5/8 antibody (Figure 9A). To prevent synthesis of Chordin and distinguish the ectopic from the endogenous BMP2/4 signal, the experiment was performed in the presence of the Alk4/5/7 inhibitor SB431542 added after fertilization. Embryos treated with SB431542 fail to express Nodal target genes such as chordin (Figure 7Diii) and BMP2/4 (unpublished data) and therefore the endogenous nuclear pSmad1/5/8 staining should be eliminated [14],[16]. As predicted, the dorsal phospho-Smad1/5/8 staining present in control embryos at blastula stages was absent in SB431542 treated embryos (Figure 9Bi–iv). Embryos injected with mRNA encoding the activated form of the receptor Alk3/6QD at the two-cell stage and treated with the Alk4/5/7 inhibitor showed a strong nuclearization of pSmad1/5/8, but this signal was restricted to the clone of cells that received the mRNA (Figure 9Bv–xii). In contrast, injection of BMP2/4 mRNA at the two-cell stage combined with SB431542 treatment resulted in a dramatic increase of nuclear pSmad1/5/8 staining, all cells of the ectoderm displaying strong nuclear staining (Figure 9Bxiii–xx). This indicates that in a SB431542 context and therefore even in the absence of the carrier protein Chordin, BMP2/4 is widely diffusible and can trigger long-range signaling within the ectoderm.
Proteoglycans such as Glypicans have emerged as major regulators of morphogen stability and mobility across tissues [45]–[47]. We therefore examined the expression of the genes encoding proteoglycans of the Glypican family, which have been shown to play crucial roles in movement of BMPs across a field of cells. The sea urchin genome contains two glypican genes, glypican 5 and glypican 6 [29],[48]. glypican 6 is abundantly expressed maternally, and its transcripts are uniformly distributed during cleavage, blastula, and gastrula stages (unpublished data). In contrast, transcription of glypican 5 is first activated at blastula stages in a belt of cell that includes the whole presumptive ectoderm except the animal pole region (Figure 10Aiv). Interestingly, starting at mesenchyme blastula stage, glypican 5 expression becomes restricted to the dorsal ectoderm (Figure 10Av–viii). This suggested that glypican 5 expression at this stage may be regulated by BMP2/4 signaling. Indeed, blocking BMP2/4 or Alk3/6 function eliminated expression of glypican 5 at gastrula stages (Figure 10Bii, iii, vi, vii). Conversely, ectopic expression of BMP2/4 or Alk3/6QD (unpublished data) induced strong expression of glypican 5 throughout the ectoderm (Figure 10Biv, viii). This indicates that transcription of this regulator of BMP signaling is itself under the control of BMP signaling and may thus be involved in a positive feedback loop.
Sea urchin Nodal has a remarkable ability to organize the embryonic D/V axis, and understanding the underlying molecular mechanism should provide important insights into the evolution of metazoan axial patterning. Based on the ability of ectopic nodal mRNA to rescue patterning over the whole D/V axis in rescue experiments, we had speculated that Nodal most likely required a relay molecule to induce the dorsal ectoderm [16]. However, direct evidence for the existence of a relay was lacking. The identity of the relay molecule(s) acting as the dorsal inducer(s) had not been demonstrated and direct evidence that BMP signaling was active on the dorsal side was missing. Similarly, the function of conserved regulators of BMP signaling such as Chordin in patterning the D/V axis was not known. In this study we have addressed these issues. We performed a functional analysis of several key genes involved in D/V patterning in the sea urchin embryo: the BMP ligand BMP2/4, the BMP receptor alk3/6, the BMP antagonist chordin, and the extracellular regulator glypican5. By using the rescue experiment as a functional assay, we provided a conclusive demonstration that Nodal does not work as a morphogen but requires a relay identified as BMP2/4. By using anti-phospho-Smad immunostaining as a direct read-out of BMP signaling, we discovered that BMP2/4 produced in ventral cells activates BMP signaling on the dorsal side of the embryo, and therefore that in the sea urchin embryo, the dorsal ectoderm is induced by ventrally produced signals. We also found that in the sea urchin, chordin is co-expressed with BMP2/4 but that despite this very unusual expression pattern, it has a conserved role in antagonizing BMP2/4. These results allow us to provide a new model for patterning along the D/V axis of the sea urchin embryo (Figure 11). They also show that although genes used for D/V patterning such as BMP2/4 and chordin are conserved in this organism, the relationships between their expression patterns are dramatically different from what was described so far in other models. These findings provide valuable information to understand the evolution of the pathways used for D/V axis specification in the ancestor of deuterostomes. They also raise important and intriguing new questions regarding the regulation of BMP activity and the regulation of BMP diffusion in embryos.
In zebrafish, ligands of the Activin/Nodal family have been proposed to control antero-posterior patterning and to work as long-range morphogens [49]. In particular, Squint has been demonstrated to work as a morphogen capable of diffusing over several cell diameters to induce target genes [19],[20]. In sea urchins, however, the ability of nodal mRNA to efficiently rescue normal development and to generate a morphologically normal pluteus larva is not very sensitive to the concentration of nodal mRNA injected, arguing against the morphogen hypothesis. Furthermore, it has been shown recently in vertebrates that Nodal antagonists such as Lefty spatially restrict the activity of Nodal ligands. Lefty factors are induced by Nodal signaling but are thought to diffuse faster than Nodal outside their expression territory, inhibiting the ability of Nodal to auto-activate and to induce downstream target genes. In the sea urchin embryo, Nodal and Lefty most likely obey the same rules [18] and constitute a typical reaction diffusion system. Further evidence for a lack of long-range action of Nodal is the demonstration of its short range of action in the animal pole region [26].
Nodal does not need to diffuse over a long range to rescue the dorsal structures, and therefore it must work by inducing a relay. We demonstrated that by comparing the effects of ectopically expressing the secreted Nodal ligand to the effects of ectopically activating the receptor complex that transduces Nodal signals in early embryos previously injected with a nodal morpholino. Furthermore we showed that in this rescue experiment, the Nodal receptor complex not only specifies ventral cell fates in a cell autonomous manner, but that it causes the BMP pathway to be activated on the opposite side demonstrating that Nodal does not work as a morphogen but that it induces a long range relay molecule that regulates morphogenesis of the dorsal region. It therefore follows that specification and patterning of the D/V axis of the sea urchin embryo does not result from the interactions between a BMP expressing ventral centre and a Chordin expressing dorsal centre, as in vertebrates, but that the activities of these two centres, which are both induced by Nodal, are concentrated within the ventral ectoderm.
In all the organisms where D/V patterning by BMP signaling has been studied, BMPs always act within the territory in which they are expressed. In this study, however, we provided several lines of evidence that, in the sea urchin embryo, specification of dorsal cell fates stringently relies on expression of the BMP2/4 ligand on the ventral side, but with the BMP molecule being translocated to the dorsal side where it activates BMP signaling. First, we showed that BMP2/4 and the type I receptor Alk3/6 are not required for specification of ventral cell fates but are essential for specification of the dorsal region of the embryos. In the absence of BMP2/4 or Alk3/6 function, the expression of ventral marker genes is unaffected, the mouth opens, but the expression of dorsal marker genes is abolished and the dorsal region of the embryo is not specified. Second, by directly visualizing the activation of the BMP pathway at blastula stages using pSmad1/5/8 immunostaining, we showed that cells that receive a high level of BMP signals are located on the dorsal side of the embryo, opposite to the side where the Nodal pathway is active and where BMP2/4 is expressed. Third, we showed that BMP signaling monitored by this anti p-Smad antibody is critically dependent on BMP2/4 and Alk3/6 function. Finally, we showed that BMP2/4 can signal over a long range from one side of the embryo to the opposite side. This does not rule out the possibility that other BMP ligands may be important for D/V patterning in the sea urchin embryo as shown in the case of Dpp and Scw in Drosophila. However, the abolition of pSmad1/5/8 staining in dorsal cells following inhibition of BMP2/4 function indicates that BMP2/4 is critically required for activation of Smad1/5/8 during gastrulation and that in its absence, other potential BMP ligands are not sufficient to activate BMP signaling in these dorsal cells. Intriguingly, down regulation of BMP2/4 abolished the pSmad1/5/8 staining while down regulation of Alk3/6 drastically reduced but did not eliminate pSmad1/5/8 staining. The residual pSmad1/5/8 staining observed in the alk3/6 morpholino injected embryos may be due to maternal Alk3/6 protein or it may reflect the activity of Alk1/2, the second BMP type I receptor present in the sea urchin genome, which could partially compensate for the loss of Alk3/6. However, the phenotype of alk3/6 morphants and the absence of expression of dorsal ectodermal markers in these embryos indicate that this residual nuclear phospho-Smad1/5/8 is not sufficient to trigger the genetic networks leading to specification of dorsal cell fates.
The mechanisms of BMP gradient formation have been best studied in Drosophila (reviewed in [1]). The prevalent view is that Drosophila Sog binds to Dpp and inhibits Dpp signaling within and near its territory of expression but that it also prevents receptor mediated internalization and turnover and thereby allows translocation of Dpp towards the dorsal midline where it is released by the metalloprotease Tolloid. Both genetic and biochemical studies in Drosophila have provided evidence for this translocation mechanism as a way to concentrate Dpp in a subdomain of a territory in which it is expressed uniformly. A similar situation is found in vertebrates, with BMP2/4 and chordin being expressed in mutually exclusive territories in the ventral and dorsal regions. Furthermore, experimental manipulations in Xenopus indicate that a shuttling mechanism of BMP ligands is responsible for translocation of BMP ligands ventrally through association with Chordin [50].
In the sea urchin embryo, BMP2/4 signaling is activated in a domain located outside its domain of expression, on the opposite side of the embryo. In this respect, it provides one of the most extreme examples known so far of translocation of a BMP ligand. It will be interesting in the future to dissect the mechanism and to identify the proteins involved in this process. Additional extracellular proteins are likely involved in the regulation of BMP2/4 translocation and activity, including the metalloprotease Tolloid, and the adaptator protein Twisted gastrulation, both of which have been shown to participate in the shuttling of BMP ligands in vertebrates and Drosophila [1]. The sea urchin genome contains a family of more than 10 Tolloid/BMP1 related genes, the functions of which are poorly understood [51]–[53].
Our finding that there is a gradient of nuclear pSmad1/5/8 along the D/V axis together with the nested patterns of expression of tbx2/3 and msx, two BMP target genes, strongly suggests that BMP2/4 is acting as a morphogen along the D/V axis. One would therefore predict that msx, which is expressed in the domain of highest BMP signaling, may be activated by high levels of BMP signaling while tbx2/3, which is expressed in the whole dorsal ectoderm, may require lower levels of BMP signals. The next question to answer will be to determine whether the graded pSmad1/5/8 pattern reflects a gradient of BMP2/4 protein or a gradient of BMP2/4 activity. One hypothesis is that BMP2/4 may be transported toward the dorsal side to form a D/V gradient of protein. An alternative hypothesis may be that BMP2/4 is present at homogenous levels throughout the dorsal side but that one of its antagonist is present in an opposite ventral-dorsal gradient. It will therefore be important to visualize the distribution of BMP2/4 protein in the extracellular space and to monitor BMP2/4 shuttling and accumulation on the dorsal side as described in Drosophila and vertebrates [54],[55].
Our finding that pSmad1/5/8 signaling is restricted to dorsal cells while the source of BMP2/4 ligand is located on the ventral side raised an intriguing question: What prevents BMP2/4 from activating its receptor in the ventral ectoderm? The distribution of the BMP2/4 receptors cannot explain this spatial restriction since Alk3/6 is expressed ubiquitously. Our finding that the chordin gene is expressed abundantly in the ventral ectoderm and that its product acts as an inhibitor of BMP2/4 signaling resolves this issue. The finding that chordin expression is regulated by Nodal signaling also helps to understand how a genetic circuit leading to specification of the dorsal ectoderm can be embedded into the genetic program leading to specification of the ventral ectoderm: Nodal induces BMP2/4 in the ventral ectoderm, and at the same time, it induces a potent antagonist of BMP signaling that prevents signaling in the ventral ectoderm and promotes diffusion of BMP2/4 outside this territory.
In Drosophila sog mutants, extracellular Dpp binds to all cells in the dorsal domain leading to broad pSmad staining. This is similar to the situation in the sea urchin embryo. In chordin morphants, BMP signaling invades the ventral side of the embryo leading to ectopic expression of dorsal marker genes such as tbx2/3. Furthermore, overexpression of chordin eliminated pSmad1/5/8 signaling and suppressed expression of dorsal markers genes. Our results therefore establish that in the sea urchin embryo, like in all the other organisms where the function of this gene has been studied, chordin plays a key role in D/V patterning through the spatial regulation of BMP2/4 signaling. These results contradict the conclusions of a recent study performed in a different sea urchin species on the role of Chordin in sea urchin development [56]. Bradham et al. reported that both overexpression of chordin and injection of the chordin morpholino strongly perturbed D/V polarity and suppressed formation of the dorsal side and oral arms, but surprisingly these perturbations did not affect tbx2/3 expression at late gastrula stage. The reasons for this discrepancy are not clear but we feel that explanations may reside in the different doses of RNA used, the different methods used for phenotypic characterization, and the different stages analyzed. In our study, we used doses of synthetic chordin mRNA up to 1 mg/ml while Bradham et al. used RNA concentrations one order of magnitude lower. Also, we performed both anti-phospho Smad1/5/8 staining and in situ hybridization at blastula stages while Bradham analyzed late gastrula embryos by in situ hybridization. At this late gastrula stage, the effects of the chordin morpholino on tbx2/3 expression are much less apparent probably due to compensatory mechanisms (see Figure 8Cvii).
The ability of BMP ligands to diffuse over long distances in the absence of carrier proteins such as Sog/Chordin is still a controversial issue. Early studies of Drosophila Sog mutants revealed that Dpp is widely diffusible in the presence of Sog but tightly localized in its absence, suggesting that diffusion of Dpp is spatially restricted [57]. The restricted diffusion of BMP ligands in the absence of carrier proteins such as Sog/Chordin has been proposed to be a key parameter and a condition for robustness in one of the mathematical models proposed to explain formation of the Dpp gradient. Limited ability of BMPs to diffuse in the absence of carrier proteins such as Chordin has also been documented in Xenopus [50]. In contrast, by examining embryos in which Dpp was ectopically expressed, Mizutani et al. found that the range of Dpp action was reduced but still substantial, consistent with free diffusion [58].
In this study, we examined the range of BMP2/4 signaling in embryos treated with the Nodal receptor inhibitor SB431542 that prevents chordin expression. We found that BMP2/4 is highly diffusible in these conditions. Therefore, our data argue in favor of free diffusion of BMP ligands in early embryos.
An interesting finding of our study that may explain the high diffusibility of BMP2/4 in the sea urchin embryo is that glypican 5, a cell-surface BMP binding protein, is expressed early in the dorsal and ventral ectoderm. Glypicans have been implicated as major regulators of BMP stability and/or BMP movement across fields of cells [45],[46],[59],[60]. In Drosophila, analysis of clones of cells deficient for or overexpressing Dally revealed that Dally regulates Dpp mobility, leading to the proposal that it is required to transmit the Dpp protein from cell to cell [45]. Interestingly, it has been shown that Dally promotes Dpp signaling and Dpp mobility in a non-autonomous manner [46]. For example, overexpressing Dally in the posterior compartment of the haltere, where it is normally absent, leads to a shift in Dpp signaling from the anterior to the posterior compartment [46],[61],[62]. On the basis of this and other observations, it was suggested that Dally biases Dpp mobility towards cells with higher level of Dally or, in other words, that cells with a high level of Dally attract Dpp [46]. Although asymmetries in glypican expression have been reported in the wing and haltere discs, differential glypican expression has never been reported to our knowledge in early embryos. Our finding of a strong asymmetry of glypican 5 expression in the early sea urchin embryo, with high level in the dorsal region, distant from the source of BMP2/4 signals, strongly suggests that a similar mechanism may also be used in early embryos to shape the BMP morphogen gradient that is responsible for patterning the D/V axis.
In the Drosophila embryo, localized injection of mRNA encoding an activated form of Thickveins promotes accumulation of extracellular Dpp, implying the existence of a positive feedback loop [54]. This observation led to the hypothesis that BMP signaling induces synthesis of a hypothetical cell-surface BMP binding protein that could either reduce the interaction of ligands with an inhibitory component or enhance future ligand receptor interactions. Mathematical modeling of BMP signaling during D/V patterning as well as experimental studies in other signaling pathways predict that positive feedback circuits can convert graded inputs into all or no outputs leading to production of bistable signaling states in which there is a sharp transition between cells displaying a high level of signaling and cells with very low signaling [63],[64]. We speculate that the BMP2/4-glypican 5 positive regulatory input we have characterized likely plays a role in intensifying BMP2/4-Alk3/6 interactions in the dorsal ectoderm.
Our results allow us to propose a new model of D/V patterning by BMP signals in the sea urchin embryo (Figure 11). Nodal signaling induces BMP2/4 and chordin expression in the ventral ectoderm. High levels of Chordin on the ventral side prevent BMP2/4 from binding to its receptor. Therefore, BMP2/4-Chordin complexes diffuse or are transported toward the dorsal side resulting in formation of a shallow D/V gradient of BMP2/4 signaling. Glypican 5, which is expressed in the whole ectoderm at this stage, may also participate in this initial phase of BMP translocation. Starting at the mesenchyme blastula stage, BMP signaling then feeds back onto glypican 5 expression. Based on the established activities of Glypicans as positive modulators of BMP signaling and BMP mobility in other systems, we speculate that this preferential expression of glypican 5 may bias BMP2/4 mobility towards the dorsal side and may increase the ability of BMP2/4 to bind to its receptor. This may further reinforce the accumulation of BMP2/4 in dorsal cells and augment BMP signaling in dorsal versus ventral cells.
A conserved feature between Drosophila and vertebrates is that both Chordin and Sog display a proneural activity, via inhibition of BMP2/4 or Dpp function in the neurectoderm. Surprisingly, in embryos of the hemichordate Saccoglossus kowalevkii, which have a diffuse neural system, the BMP signaling pathway does not repress neural gene expression [3]. Echinoderms, like hemichordates, develop a rather diffuse nervous system. However, a notable difference between the nervous system of echinoderms and hemichordates embryos is that echinoderms have retained a strong polarization of their nervous system along the D/V axis, most neurons arising in the animal pole and in the ciliary band territory, which lies at the interface between the ventral and the dorsal ectoderm. In this study, we showed that inhibiting the function of BMP2/4 or its receptor Alk3/6 dramatically expanded the neurogenic territory of the ciliary band towards the dorsal side. This strongly suggests that, unlike in hemichordates, neural tissue formation in echinoderms requires inhibition of BMP signaling. Taken together, our results are consistent with a model for specification of the ectoderm in the sea urchin embryo that relies on the hypothesis that the default state of most of the ectoderm in the absence of Nodal signaling is a proneural, ciliary band-like fate and that Nodal and BMP2/4 act in cascade to specify, respectively, the ventral and the dorsal ectoderm, restricting the ciliary band to the territory located in between these two specified territories.
Based on the conservation of the molecular pathway used both to develop a secondary axis and to form a nervous system in Drosophila and vertebrates, it was proposed that this molecular pathway was already used for specification of the D/V axis and nervous system in their common bilaterian ancestor but that an inversion of the D/V axis had occurred during evolution [65]–[67]. According to this hypothesis, the common ancestor would have possessed a centralized nervous system and a plausible mechanism to explain the D/V axis inversion is that the mouth would have moved from one side of the embryo to the other or, that the trunk would have rotated 180° relative to the head [6]. A recent study showed that in Cephalochordates, the most basal chordates according to recent phylogenetic analyses, the expression domain of chordin and BMP2/4 retained the same orientation as in protostomes (i.e., BMP2/4 being dorsal and chordin ventral) [4], suggesting that any inversion occurred after the emergence of chordates or, alternatively, that a second inversion would have occurred independently in this lineage. In the hemichordate Saccoglossus kowalevkii, an “invertebrate like” BMP-chordin axis is also present, with chordin being expressed on the ventral side and BMP on the dorsal side [3].
One novel finding of this study is that BMP2/4 and chordin are co-expressed in the ventral ectoderm of the embryo. The co-expression of BMP2/4 and chordin in the ventral ectoderm of the sea urchin embryo contrasts with the situation present in embryos of all the other bilaterians organisms (Figure 12). Thus, it indicates that although a BMP Chordin axis is present in the sea urchin embryo, this axis is defined by the activities of these factors and not by their expression territory.
Recent studies have shown that the dpp and chordin genes are present in the genomes of cnidarians, which are positioned phylogenetically as an outgroup to the bilaterians. Unexpectedly, in Nematostella vectensis gastrulae, the homologs of both genes are expressed asymmetrically along a secondary axis perpendicular to the oral-aboral axis. Even more intriguing is the fact that, in Nematostella like in sea urchin, BMP2/4-dpp and chordin are expressed on the same side of the embryo [68]–[71]. It has been suggested that animals with a radial symmetry could have developed a secondary axis that was lost later during evolution [72]. The asymmetrical expression of dpp and chordin in the Nematostella embryo could therefore reflect the presence of a cryptic D/V axis or a convergent use of BMP signaling for secondary axis specification. That chordin and BMP2/4-dpp are expressed in the same territory both in embryos of a basal deuterostome and in embryos of N. Vectensis raises the intriguing possibility that co-expression of BMP ligands and BMP antagonists such as Chordin on the same side of the embryo could have been the ancestral way to develop a secondary axis and the switch to opposite expression territories for the ligand and the antagonist a more recent evolutionary innovation. Another hypothesis is that in the two radially symmetrical organisms, the sea urchin and the sea anemone, BMP signaling may have evolved independently on the basis of co-expression. Functional analysis of the cnidarian chordin and dpp genes and investigation of the role of the BMP chordin signaling network in additional protostomes organisms are required to address these issues.
Adult sea urchins (Paracentrotus lividus) were collected in the bay of Villefranche-sur-Mer. Embryos were cultured at 18°C as described previously [73],[74]. Fertilization envelopes were removed by adding 2 mM 3-amino-1,2,4 triazole 1 min before insemination to prevent hardening of this envelope followed by filtration through a 75 µm nylon net [16]. Treatments with SB431542 were performed by adding the chemical at 1, 3, or 5 µM diluted from stocks in DMSO, in 24 well plates protected from light. Treatments with NiCl2 were performed by exposing embryos to 0.5 mM of chemical. All treatments were carried out from 30 min to 48 h post-fertilization.
Pl-alk3/6: Degenerated primers were derived from an alignment of the kinase domain from Activin Like Kinase (ALK) Type I receptors. Pl-chordin: Degenerated primers were derived from an alignment of the conserved first von Willebrand factor type C of vertebrate Chordin sequences. These primers were used to amplify partial cDNA fragments for both genes using RT PCR and a cDNA mix from mixed embryonic stages. The Primers used to amplify alk3/6 were as follows: Alk3/6 F1, 5′-GCNTTYATHGCNGCNGAYATHCC-3′, which is derived from the peptide sequence AFIAADIP; Alk3/6 R1, 5′-CATRTCNGCDATCATRCANGTNCC-3′, which is derived from the peptide sequence GTCMIADM. The Primers used to amplify chordin were: Chordin F1, 5′-TGYACNTTYGGNGCNGAYTTYTA-3′, which is derived from the peptide sequence CTFGADFY; Chordin R1, 5′-GGRCANGTYTTRCARCARAANCC-3′, which is derived from the peptide sequence GFCCKTCP. PCR fragments for alk3/6 and chordin were subsequently used as probes to screen a pluteus stage cDNA library, and the largest inserts were entirely sequenced on both strands.
The accession numbers of alk3/6, chordin, and Smad1/5/8 are, respectively, FJ976181, FJ976182, and FJ976183.
To make pCS2 Alk4/5/7-Q265D, the CAG codon encoding Glutamine in position 265 of pBS Alk4/5/7 was mutated to GAT. To make pCS2 Alk3/6-Q230D the CAA codon encoding Glutamine in position 230 of pBS Alk3/6 was mutated to GAT. Both constructions were obtained by PCR splicing using two successive rounds of PCR.
Oligonucleotides used for the pCS2-Alk4/5/7-Q265D are:
Alk4/5/7-Cla-ATG, 5′-ACCATCGATACCATGGCATTGGAACGTGCT-3′;
Alk4/5/7-TAG-Xho, 5′-AGGCTCGAGCTACATCTGTAGTTGAGGACG-3′;
Alk4/5/7-Q265D-Fw, 5′-ACCATCGCTCGAGATATCGTGATTCAG-3′;
Alk4/5/7-Q265D-Rev, 5′-CTGAATCACGATATCTCGAGCGATGGT-3′.
Oligonucleotides used for the pCS2-Alk3/6-Q230D are:
Alk3/6-Cla-ATG, 5′-CGGGATCCACCATGGCGACAGATGTAACACTAACCG-3′;
Alk3/6-TAG-Xho, 5′-CCGCTCGAGCTAAACTTTGAATTCCGTTTCTTG-3′;
Alk3/6-Q230D-Fw, 5′-CCGCTCGAGCTAAACTTTGAATTCCGTTTCTTG-3′;
Alk3/6-Q230D-Rev, 5′-CCTAATCAGCTGAACATCCTTCGCGATAGTGCG-3′.
To make the Alk3/6 mm construct, an oligonucleotide containing nine mismatches in the sequence recognized by the morpholino was used to amplify the coding sequence. The sequence of this oligo is:
5′-AGGGGATCCACCATGGCTACGGACGTCACGTTGACCGGACGAAAA-3′,
(mismatches underlined).
Oligonucleotides for making the pCS2 chordin construct are:
Chd-Xho-ATG, 5′ AGGCTCGAGACCATGTACCGTGTCGTGATTTATAC-3′;
Chd-Xba-TGA, 5′ TGCTCTAGACTATGAAAGCTTCTCTTTCCTTC-3′.
All PCR reactions were made using the Pfx DNA polymerase and the constructs were verified by sequencing.
For overexpression experiments, capped mRNA were synthesized from the pCS2 templates constructs [75] linearized with Not1 using the SP6 mMessage mMachine kit (Ambion). After synthesis, the capped RNA were purified on Sephadex G50 columns and quantitated by spectrophotometry. Synthesis of capped mRNA coding for Nodal and BMP2/4 is described in Duboc et al. (2004) [16]. RNAs were mixed with Tetramethyl Rhodamine Dextran (10,000 MW) or Texas Red Dextran (70,000 MW) or Fluoresceinated Dextran (70,000 MW) at 5 mg/ml and injected in the concentration range 500–800 µg/ml for nodal and alk3/6QD and in the range 1–1.5 mg/ml for chordin. Morpholino antisense oligonucleotides were obtained from GeneTools LLC (Eugene, OR). The sequence of the morpholino against alk3/6 is 5′-TAGTGTTACATCTGTCGCCATATTC-3′, in which the three underlined bases are complementary to the initiation ATG codon of the transcript. The sequences of the chordin morpholinos used are:
Mo1 chordin: 5′-GGTATAAATCACGACACGGTACATG-3′; Mo2 chordin: CGAAGATAAAAACTTCCAAGGTGTC. Injection of Mo1 chordin morpholino caused absolutely no toxicity even when high doses (up to 2 mM) were used while the Mo2 chordin started to cause non-specific defects when injected at concentrations above 0.8 mM. As a control, we used a morpholino directed against the hatching enzyme gene: 5′-GCAATATCAAGCCAGAATTCGCCAT-3′. Embryos injected with this morpholino at 1 mM developed into normal pluteus larvae. The nodal, BMP2/4, and Alk4/5/7 morpholinos are described in Duboc et al. (2004) [16] and Range et al. (2007) [13]. Morpholinos oligonucleotides were dissolved in sterile water and injected at the one-cell stage together with Tetramethyl Rhodamine Dextran (10,000 MW) at 5 mg/ml. For each morpholino a dose-response curve was obtained and a concentration at which it did not elicit non-specific defect was chosen. Approximately 2–4 pl of oligonucleotide solution at 0.5 mM for Mo-nodal1, 0.8–0.6 mM for Mo-alk3/6, 1–1.5 mM for Mo-1 chordin, and 0.4–0.6 mM for Mo2-chordin were used in the experiments described here. For each experiment about 150–200 eggs were injected for morphological observations, and 200–300 for in situ hybridization. All the experiments were repeated at least two to three times and only representative phenotypes observed in a majority of embryos are presented here.
In situ hybridization was performed following a protocol adapted from Harland (1991) [76] with antisense RNA probes and staged embryos. The goosecoid, tbx2/3, Delta, and sm30 probes have been described previously [77],[78]. The hnf6/onecut probe was derived from a pBluescript clone obtained by screening a lZAP cDNA library and the glypican 5 probe from a plasmid library constructed in pSport. Double in situ hybridizations were performed following the procedure of Thisse [79].
Total RNA from staged embryos was extracted by the method of Chomczynski and Sacchi (1987) [80]. Samples of total RNA (20 µg per lane) were fractioned on 1% agarose gel containing 0.66 M formaldehyde and transferred to membrane by standard methods (Sambrook et al. 1989) [81].
The antibody we used is an anti-phospho-Smad1/5/8 from Cell Signaling (Ref 9511) raised against a synthetic phosphopeptide corresponding to residues around Ser463/465 contained in the motif SSVS of human Smad5. In S.purpuratus, both Smad1/5/8 and Smad2/3 share the motif SSVS, but in P.lividus, only Smad1/5/8 contains the motif (SSVS), the corresponding motif in P. lividus Smad2/3 being SSMS. This antibody specifically recognizes Smad1/5/8 in P. lividus (see Figures 5, 6), but it is predicted to recognize both Smad1/5/8 and Smad 2/3 in S.purpuratus. Embryos were fixed in paraformaldehyde 4% in MFSW for 15 min, then briefly permeabilized with methanol. Embryos were rinsed with PBST, then PBST-BSA 2% four times and incubated overnight a +4°C with the primary antibody (anti-phospho-Smad1/5/8 Cell Signaling Ref 9511) diluted in PBST supplemented with 2% BSA. Embryos were then washed six times with PBST-BSA 2%, then the secondary antibody diluted in PBST-BSA 2% was added to the embryos. In all cases the antibody was incubated overnight at +4°C. For immunofluorescence, the secondary antibody was washed six times with PBST. Two last rinses were made with PBST-Glycerol 25% and 50%, respectively. Embryos were mounted in a drop of the Citifluor anti-bleaching mounting medium, then observed under a conventional fluorescence microscope or with a confocal microscope. For Alkaline phosphatase revelation, two rinses were made with PBST following the secondary antibody incubation, then two with TBST. Embryos were then washed twice with the alkaline phosphatase buffer supplemented with Tween 0.1%, then staining was performed either with NBT and BCIP as substrates at the final concentration of 50 mM each, or embryos were washed twice with Tris 100 mM pH 8.2 and stained using a FastRed as substrate in Tris100 mM pH 8.2. In both cases staining was stopped by four rinses with PBST+EDTA 5 mM, then two rinses with PBST 25% Glycérol and 50% Glycerol. Embryos were then mounted and observed with a DIC microscope. In the case of the combined immunostaining and in situ hybridization, the immunostaining was performed first as described above. Staining was stopped with four rinses with PBST, then embryos were fixed 1 h in paraformaldehyde 4% in MFSW before proceeding with in situ hybridization.
|
10.1371/journal.pbio.0050223 | Conserving Biodiversity Efficiently: What to Do, Where, and When | Conservation priority-setting schemes have not yet combined geographic priorities with a framework that can guide the allocation of funds among alternate conservation actions that address specific threats. We develop such a framework, and apply it to 17 of the world's 39 Mediterranean ecoregions. This framework offers an improvement over approaches that only focus on land purchase or species richness and do not account for threats. We discover that one could protect many more plant and vertebrate species by investing in a sequence of conservation actions targeted towards specific threats, such as invasive species control, land acquisition, and off-reserve management, than by relying solely on acquiring land for protected areas. Applying this new framework will ensure investment in actions that provide the most cost-effective outcomes for biodiversity conservation. This will help to minimise the misallocation of scarce conservation resources.
| Given limited funds for biodiversity conservation, we need to carefully prioritise where funds are spent. Various schemes have been developed to set priorities for conservation spending among different countries and regions. However, there is no framework for guiding the allocation of funds among alternative conservation actions that address specific threats. Here, we develop such a framework, and apply it to 17 of the world's 39 Mediterranean-climate ecoregions. We discover that one could protect many more plant and vertebrate species by investing in a sequence of conservation actions targeted towards specific threats, such as invasive species control and fire management, rather than by relying solely on acquiring land for protected areas. Applying this new framework will ensure investment in actions that provide the most cost-effective outcomes for biodiversity conservation.
| Many sophisticated approaches exist for identifying priority areas for conservation at a global scale. These “biodiversity hotspots” or “crisis ecoregions” are typically identified using data on endemic species richness, total biodiversity, and past habitat conversion [1–3]. With few exceptions, these approaches neglect economic costs and provide a static assessment of conservation priorities. They therefore cannot provide guidance on how funds should be distributed between regions, nor can they inform when the funds should be spent. Recent theoretical advances incorporate economic considerations and landscape dynamics into priority-setting, and provide an analytical framework for deciding where, when, and how much money should be invested for biodiversity conservation [4–8].
While these theoretical advances incorporate economic considerations, they treat land acquisition, or the creation of protected areas, as a surrogate for the broader suite of actions available to protect biodiversity. Conservation practitioners routinely invest in a diverse array of activities such as fire management, invasive species control, and revegetation, with the aim of enhancing or sustaining biodiversity. In many places land acquisition is not feasible, and neither appropriate nor affordable. In addition, the spatial extent of many threats is usually greater than the area of land that can be acquired. A framework is urgently needed that can support the more sophisticated funding allocation decisions required from conservation practitioners. Such a framework could help to allocate limited conservation funds to threat-specific conservation actions in areas where they are likely to achieve the greatest potential biodiversity benefit.
Here, we develop an action- and area-specific framework for conservation investment and illustrate its application using Mediterranean-type habitats (Figure 1). Mediterranean ecoregions boast exceptional species diversity but are poorly protected, highly degraded, and exposed to multiple persistent threats [9–13]. Consequently, they have been ranked among the world's highest conservation priorities [3,14,15]. How might funds be allocated to conserve Mediterranean ecoregions in the most cost-effective way?
To apply our framework (Figure 2 and see Materials and Methods) we require an explicit statement of the overall conservation objective and the budget (steps 1 and 2 of Figure 2), and an understanding of the threats operating in each ecoregion and the potential conservation actions to abate them (steps 3 and 4). Our objective is to maximise the total number of species (vascular plants and vertebrates combined) conserved across these ecoregions, through strategic investment in a suite of conservation actions, given a fixed annual budget. The amount of money allocated annually to each conservation action in each ecoregion depends on the area of land currently receiving and requiring the action, the cost of the action per unit area, and the biodiversity benefited by the investment (the number of plant and vertebrate species predicted to persist in an ecoregion after investment in a conservation action; see Materials and Methods).
Our aim is to develop investment schedules for Mediterranean ecoregions that reflect the relative returns from investing in different conservation actions in order to maximise our objective (step 5 of Figure 2). We deliver investment priorities that change through time depending on the cumulative impacts of investments (step 6). While we address a global-scale problem, our framework and analytical approach is also applicable at national and regional scales.
We apply our framework to the 17 (of 39) terrestrial Mediterranean ecoregions for which data are most readily accessible. This subset of Mediterranean ecoregions covers parts of Australia (ten ecoregions; Table 1), Chile (one ecoregion; Table 2), South Africa (three ecoregions; Table 3), and California and Baja California (three ecoregions; Table 4) (Figure 1). Although we recognise that alternative delineations of Mediterranean habitats are available, we employ the delineations provided by the World Wildlife Fund, given their utility for global-scale analyses.
Through consultation with regional experts, we identify the key threats in each ecoregion to achieving our objective of maximising the total number of species (in our case, vascular plants and vertebrates combined) conserved given an annual budget of US$100 million (steps 1–3 of Figure 2). We also identified the actions undertaken to abate these threats (step 4). Hereafter we term each ecoregion–conservation action combination an “ecoaction”. We assume that the impact of each ecoaction is independent. Through a combination of expert input, literature review, and analysis of regional datasets in geographic information systems we determine the areas requiring, and already receiving, each ecoaction and estimate the cost associated with its implementation (steps 4a–4c of Figure 2; Text S1).
While the costs incurred for some conservation actions, such as land acquisition or revegetation, are one-time costs, the costs of other actions, such as invasive predator control, are incurred annually. To convert the latter to one-time costs, we endow the annual cost over 20 y (unless otherwise stated), after which further funds are required for these conservation actions to continue. We determine the endowed value by calculating the net present value over the timeframe of interest, assuming an inflation rate of 3.2% and discount rate of 6.04%. This discount rate is equivalent to a 10-y US government bond rate, and the inflation rate represents that of the US dollar in 2005. We account for the costs of ongoing management for ecoactions that involve land acquisition and for the costs of establishing agreements with private landholders (if such investments are considered necessary for an ecoaction to proceed or to be long-lasting; Text S1). The cost of each ecoaction is based on the perceived expenditure required for successful interventions, and we therefore assume that investment in each ecoaction will prevent the local extinction of species at risk from the relevant threat.
In this paper, the number of species benefited by each ecoaction—its “biodiversity benefit”—is the number of plant and vertebrate species predicted to persist in an ecoregion after investment in the ecoaction (step 4d of Figure 2). If appropriate data were available for each ecoaction, we could modify the predicted biodiversity benefit by the likelihood that the ecoaction will succeed in abating the relevant threat. To operationalise our approach we need a functional form for the relationship between investment in an ecoaction and its biodiversity benefit. Every investment shows diminishing returns and therefore we assume that the marginal benefit of investment in a particular ecoaction decreases as the size of the investment increases. We represent diminishing returns using the functional form of the species–area relationship, where the total number of species (S) present in area A is a power-law function of that area [16]:
We calculate the constant α by dividing the total number of species in the ecoregion by the estimated area of original habitat, after raising this area to the power of z (see Text S2). In the baseline scenario we assign z a value of 0.2, a typical value for terrestrial, non-island regions [16].
We therefore assume that the incremental number of species protected with a given increase in area protected follows the form of a standard species–area curve. When we account for the cost of each ecoaction, we simply replace the area protected by the cost of protecting the equivalent area (to generate a species–investment curve; step 5a of Figure 2). This relationship is straightforward for habitat protection or restoration, but requires further thought for the diverse array of conservation actions considered here.
The adaptation of species–area curves to conservation actions other than reserving or restoring land is based on the premise that investment in these actions will also exhibit diminishing returns. The major refinements required are that the area of “protected” habitat is the area of investment in each ecoaction (each with a pre-specified cost) and the number of species protected is threat-specific. Since we currently do not have an ecological basis for an alternative parameterisation of this relationship for the range of ecoactions considered here, we evaluate the sensitivity of the allocation schedules to the value of z. We choose z randomly from a uniform distribution (between 0.1 and 0.4, n = 30) to reflect the uncertainty about the relationship between the number of species protected and the amount of money invested in each ecoaction, specifically, the rate at which the returns from investment diminish.
To determine the biodiversity benefit of an ecoaction that abates a specific threat we consider only those species impacted by that threat. We calculate the number of “at risk” species by determining the proportion of plant and vertebrate species regarded as threatened by each type of threat (using the World Conservation Union [IUCN] Red List for each country [17] and excluding those species that are of least concern or data deficient), and multiply this proportion by the total number of plant and vertebrate species occurring in each ecoregion [18,19]. Thus we assume that the proportion of species that would be protected by investment in a particular ecoaction is the same as the proportion of IUCN-listed species identified nationally as being at risk from the relevant threat, and that the species in this subset benefit equally from an investment. For invasive predator control in Australia, we limit the biodiversity benefit calculation to just vertebrates by restricting the IUCN search to vertebrate species and multiplying this proportion by the number of vertebrate species occurring in each ecoregion.
Obtaining an optimal allocation schedule through time amongst such a large number of ecoactions is computationally intractable [4,5], so we adopt a “rule of thumb”, or heuristic, to approximate the optimal investment schedule. This heuristic, which we term “maximise short-term gain”, directs funds each year to ecoactions that provide the greatest short-term increase in biodiversity benefit per dollar invested (steps 5b and 5c of Figure 2). Using this heuristic we generate an investment schedule over 20 y, given a fixed annual budget of US$100 million (step 6 of Figure 2). We assess the sensitivity of the investment schedule to the budget size by repeating the analysis with the annual budget reduced to US$10 million.
We use the Spearman coefficient of rank correlation to compare the priority rankings based on the ecoaction-specific framework to those based on a ranking of vertebrate species richness [20]. Because of lack of independence, we test significance against the distributions of Spearman values derived from 100,000 random pairings of X and Y variables. The null hypothesis is that the observed coefficient is zero, or the distribution of Y is the same for all values of X [21].
We also compare the ecoaction-specific framework to a simplified model of conservation that focuses only on land acquisition for the creation of protected areas. In this analysis, we estimate the cost of land acquisition using a statistical model (Table S1) and the area requiring acquisition as the area of natural habitat that is currently unprotected (IUCN status I–IV). We estimate the biodiversity benefit of this action as the total number of plant and vertebrate species. Therefore, while the biodiversity benefit under the ecoaction-specific framework is a proportion of the total species richness (those threatened by each threat type), the biodiversity benefit under the acquisition-only framework is the total richness of vertebrate and plant species. As with the ecoaction approach, we assume diminishing returns with cumulative investment and model this relationship using species–investment curves. For the acquisition-only approach we rescale the annual budget to US$148 million dollars (from US$100 million), since the overall cost of achieving our objective under this scenario is 48% greater.
Different ecoregions have different mixtures of threats and candidate conservation actions. In total, we evaluated 51 ecoregion–conservation action combinations (denoted “ecoactions”; Tables 1–4) across 17 Mediterranean ecoregions (Figure 1). Using species–investment curves for each of the 51 ecoactions and the “maximise short-term gain” heuristic (see Materials and Methods), we obtained investment schedules based on an annual budget of US$100 million. These schedules reflect shifts in the allocation of funds as the return from investing in each ecoaction diminishes. The investment schedules are determined by the interplay of three main factors: (1) the relationship between the additional area invested in each ecoaction and the biodiversity benefit; (2) the cost of this investment; and (3) the existing level of investment.
We illustrate how these three factors interact within our conservation investment framework (steps 4–6 of Figure 2) using a regional-scale case study from the Swan Coastal Plain scrub and woodlands ecoregion of Australia (Figure 3). The curve to the left of the circles indicates the total area of conservation interest (indicated by the squares), that is, the area already receiving the actions (step 4a), and the area requiring them (step 4b). For example, the area of interest for invasive predator control in this ecoregion is about 5,832 km2 (see Text S1). The curve to the right of the circles acknowledges that the persistence of species depends on past changes to habitat and that knowledge of the distribution of species is uncertain (see Text S2).
We estimated the original extent of habitat in this ecoregion to be approximately 15,210 km2 and that this area supported a total of 565 vertebrate and plant species now at risk due to habitat fragmentation, 256 plant and vertebrate species now at risk from a soil-borne pseudo-fungus, Phytophthora cinnamomi, and 143 vertebrate species now at risk due to invasive predators (step 4d of Figure 2). In each case, the total number of species estimated to be at risk is represented by the right-hand endpoint of the species–area curves (Figure 3A).
Assuming the costs of undertaking the different actions is the same, we determined that conducting invasive predator control or revegetation over an additional 200 km2 in the Swan Coastal Plain ecoregion will potentially protect three and four species, respectively (panels II and III of Figure 3B). Conducting Phytophthora management over the same area has the potential to protect 108 species because the area of conservation interest lies in the steepest part of the species–area curve (Text S1; Table 1; panel I of Figure 3B).
When we modelled the relationship between the biodiversity benefit and dollars invested using species–investment curves (step 5a of Figure 2; Figure 3C), we found that the cost-effectiveness of each action varies widely (step 5b). Revegetation in the Swan Coastal Plain ecoregion costs US$301,118 per square kilometre, Phytophthora management costs US$514,626 per square kilometre, and invasive predator control costs US$7,125 per square kilometre (Table 1). Phytophthora management was still the most cost-effective action: US$2 million spent on this action will potentially protect 49 species, although the potential benefit reduces rapidly with cumulative investment (panel I of Figure 3D). An initial expenditure of US$2 million on invasive predator control in the Swan Coastal Plain ecoregion has the potential to protect four species, whereas there is negligible benefit from spending US$2 million on revegetation (panels II and III of Figure 3D). The comparatively greater marginal returns from investing in invasive predator control are due to its low cost, despite the fact that the direct biodiversity benefit for this action is restricted to vertebrates.
Based on this analysis, initial investment within the Swan Coastal Plain scrub and woodlands ecoregion is prioritised to Phythopthora management (step 5c of Figure 2). The species–investment curves are then updated given changes in the area receiving the conservation action and the area requiring the conservation action (step 6). In the next time step, the budget is allocated to the conservation action that now maximises the biodiversity benefit per dollar invested. This regional case study therefore illustrates how the species–investment curves are constructed, and how the actions are prioritised for investment at each time step based on their cost and biodiversity benefits, and the current level of investment in each conservation action.
In applying the conservation investment framework at the global scale we encompass a greater mix of threats and candidate conservation actions. Across all 17 ecoregions, only 24 ecoactions (of the 51 possible) received investment in the model during the first 5 y. During this time, most funds were allocated to land protection and management (through land acquisition, off-reserve management, and on-going management) in the three South African ecoregions (66% of the total budget, six ecoactions in total; Table 5). Much of the remaining funds were allocated to invasive plant control in the Chilean ecoregion, the three South African ecoregions, and one of the Californian/Baja Californian ecoregions (24% of the total budget; five ecoactions in total; Table 5). These conservation actions yielded the greatest marginal return on investment over 5 y because the potential biodiversity benefit is high and the costs are comparatively low. Over 5 y the greatest amount of money (21% of the total budget) is allocated to land protection and management (through land acquisition, off-reserve management, and on-going management) to abate agricultural conversion in the montane fynbos and renosterveld ecoregion of South Africa. This broad ecoregion contains a large area of arable land that is unconverted but largely unprotected. Furthermore, the potential biodiversity benefit of abating agricultural conversion in this region is high, while the cost of this ecoaction is comparatively low (Table 3).
Beyond the first 5 y, we see additional ecoactions prioritised for investment because further investments in initially selected ecoactions exhibit diminishing returns. Consequently, as one moves from a 5-y to a 20-y timeframe, the number of ecoactions identified for investment increases from 24 to 30, despite a 4-fold increase in the funds available. The greatest investment over 20 y is directed, in equal proportions to the montane and lowland fynbos and renosterveld ecoregions of South Africa, to the conservation action of land protection and management to abate agricultural conversion (both ecoregions are allocated approximately 14% of the total budget for investment in this conservation action; Table 5). Over 20 y all 17 ecoregions are allocated some funds (Table 5).
Overall, the investment schedule was insensitive to the annual available budget, though some of the lower-priority ecoactions did not receive funding when the annual budget was reduced to US$10 million. For example, under a reduced budget, the Coolgardie woodlands was the only ecoregion in Australia allocated investment in invasive predator control over 20 y, since the current level of investment in this ecoaction is small. Likewise, under a reduced budget, funding was not allocated to invasive plant control in the Californian/Baja Californian ecoregions.
It is informative to compare the outcome of this analysis to that of a simpler analysis that ignores costs and benefits, and instead prioritises the ecoregions for investment on the basis of a single ecological criterion—in this case, vertebrate species richness per unit area. There was a lack of concordance (rs = 0.39, p = 0.12) between the priorities based on the two approaches, indicating that they would recommend profoundly different ecoregions as investment priorities (Figure 4). For example, the Chilean matorral ecoregion has the fewest vertebrate species per unit area but received the fourth greatest allocation under the ecoaction approach because of the potentially high biodiversity benefit per dollar invested. Conversely, the Jarrah-Karri forest and shrublands ecoregion in Australia has the greatest vertebrate species richness per unit area but was not a priority using the ecoaction approach (Figure 4).
When we compared the ecoaction-specific framework to a simplified model of conservation that focuses only on land acquisition (see Materials and Methods), we found that greater biodiversity benefits are accrued by investing in actions targeted towards specific threats. The decision steps in the resource allocation process are identical regardless of the investment approach (Figure 2), with the exception that the land-acquisition-only approach considers only a single conservation action—land acquisition. Based on the data available for this analysis, we estimate that over 5 y many more species could be protected using an ecoaction approach (2,780 versus 703 species). After 20 y slightly more than twice as many species could be protected. The difference is reduced through time because of diminishing returns regardless of the investment approach. Therefore, after accounting for the existing level of investment in each ecoaction (which in some cases includes land acquisition with the costs of management added; Text S1) or in land acquisition alone (where the costs of managing specific threats are not accounted for; Table S1), greater returns can be achieved using the ecoaction-specific framework. These results were relatively insensitive to the parameterisation of the species–investment relationship, specifically, the rate at which the returns from investment diminish (as determined by the z exponent). The average ratio of species protected using the ecoaction approach and the land-acquisition-only approach over 5 y was approximately 3.49 (this ratio varied from 3.46 to 3.51 with the value of z for each ecoaction randomly chosen from a uniform distribution between 0.1 and 0.4, n = 30; see Materials and Methods).
These results illustrate the advantages of an ecoaction-specific framework over priority-setting approaches that ignore economic costs, or that focus only on the acquisition of land for protected areas. The Mediterranean example shows that an ecoaction-specific framework provides better outcomes for biodiversity conservation than the simpler approaches that have dominated the scientific literature [22]. In practice, very few conservation practitioners adopt species richness priorities identified by simple numerical ranking. Instead, they routinely consider the costs of investments, and more complex measures of biodiversity benefits. Our framework provides a standard, transparent, and quantitative template in which to solve complex resource allocation problems.
By specifying costs and benefits and a total budget, we produced an investment schedule that reveals shifting priorities through time as the returns from investment change. Because conservation budgets are often reallocated every year, it could be practical to follow flexible and time-varying investment schedules, as opposed to being tied to specific actions simply because they were previously regarded a high priority. Furthermore, if an equitable distribution of a base level of funds is important, then a pre-specified amount of funds could be allocated to each ecoregion, with funds directed to particular actions according to their relative return on investment.
Various refinements to our approach would be valuable. The calculation of biodiversity benefits could be improved by incorporating more detailed information from conservation practitioners, either in the form of empirical or expert data. We could also extend our analysis to consider other types of benefits, including the potential returns from the protection of ecosystem services [23]. Under such circumstances, it would be possible to assess the potential collateral benefits of conservation investments beyond the protection of biodiversity and to evaluate the trade-offs involved, as it is likely that different areas will be prioritised to achieve the alternative objectives. Other improvements might entail identifying the individual species that are most at risk due to the different threats, the impact of investment in each ecoaction on the persistence of these species, the likelihood of success of each ecoaction, and the potential for leverage.
As with typical conservation planning exercises that focus on protected area establishment, we have assumed that each ecoaction will be totally effective in abating the relevant threat. A plethora of factors (ranging from natural community succession to climate change) render this assumption unreliable [24,25]. It would be ideal if we had estimates of the likelihood of long-term success of each conservation action in conserving biodiversity, both for the duration of the action, and after investment ceases. These data are unlikely to be available at any time in the near future. Instead, we base the cost of each ecoaction on the assumption that enough funds are invested to have a high likelihood of success. Presently we assume that alleviating the most important threat will protect the species at risk, but the number of protected species will likely be overestimated if some species need to be protected from multiple threats that require different ecoactions. With knowledge of the individual species at risk due to each threat we could identify which species are affected by more than one threat. With this knowledge, the complementarity of each ecoaction in improving species persistence could be incorporated, and this would help to minimise the degree to which benefits are overestimated, assuming of course that all important threats have been identified. In addition, the assumption of diminishing returns with cumulative investment could, in some instances, be replaced with threshold relationships for those conservation actions that yield no benefit until some minimum level of investment is reached.
These, however, are straightforward technical modifications of the approach; obtaining the relevant data represents the greatest challenge. Within our dynamic framework, the investment schedules can be updated as our knowledge improves. Application of our framework can also provide insights into research priorities. For example, through our Mediterranean application it has become apparent that information on the likelihood of success and patterns in threat co-variation among species are important subjects of future research. We hope that our framework for conservation investment will encourage conservation practitioners to track and report action-specific data to allow a refined framework to be parameterised.
To examine the sensitivity of our results to the budget, we reduced the amount of money available per annum from US$100 million to US$10 million. In this example, varying the annual budget simply altered what was able to be achieved over the timeframe of interest: an investment of US$10 million over 10 y will achieve approximately the same outcomes as an investment of US$100 million in 1 y. This is because the investment schedules are determined only by the area requiring investment and the relative returns of the investment. While the “maximise short-term gain” heuristic closely approximates the optimal solution, especially with funding uncertainty, the urgency of investment could also be incorporated if information were available on the rate of species loss in each ecoregion due to each threat [4]. Under such circumstances, the investment schedules would change over different timeframes because of the rates of species loss influencing both the area requiring each ecoaction and the relative return from investment. Explicitly accounting for ongoing species losses would change our objective to “minimising losses” rather than “maximising gains” [4,6,26]. Incorporating information on the rates of species loss would further improve the ability to determine when conservation actions should be implemented in order to achieve the greatest outcomes for biodiversity. Presently, data on the rates of loss of species due to particular threats are scarce, and there is limited understanding of how species loss varies with changes in available habitat.
We have applied the approach at a global scale, but it will be more effectively applied at local or regional scales, if only because in many cases the required data are more likely to be available and to be more accurately estimated. Application at a global scale is nevertheless important, despite the sparseness of the data. First, global non-governmental organisations and international agencies are interested in decision-making at a global scale, and will make investment decisions at such a scale. Second, there is now a large academic literature on setting conservation priorities at a global scale; these studies are equally beset by sparse data and poorly tested assumptions, they have mainly ignored costs, and they have focussed on protected area establishment.
Analysis at a finer spatial scale would further increase the efficiency of the investment schedules by accounting for the heterogeneity in the costs and benefits of conservation actions. Such an analysis will likely require assessment of empirical, modelled, and expert data. With a more detailed and refined analysis we could also account for the actual costs and relative success of conservation actions undertaken in the past. Such an analysis could also allow finer-scale socio-economic and policy data to be incorporated. For example, the area of land predicted to be vulnerable to agricultural conversion in the montane fynbos and renosterveld ecoregion of South Africa is likely to be overestimated by the biophysical models employed as these ignore socio-economic and political factors. The collapse of subsidies in this region may mean that only small areas are currently experiencing conversion pressures [19]. While analysis at a finer scale would allow a refined assessment of investment priorities, it would be at the expense of the global-scale evaluation of investment priorities presented here. By being transferable across scales, our framework can help to bridge the current gap between global-scale analyses and investment decisions that are implemented within regions, as it can provide an understanding of the relative importance of each ecoaction for conserving biodiversity within a global context.
Regardless of scale, stakeholders and experts are integral to the success of the ecoaction approach, through identifying threats and actions, determining the relative costs and benefits of each action, and identifying local constraints for their implementation (see Materials and Methods). The results of any assessment must also be interpreted in the context of the value systems of stakeholders, as well as other factors such as the implementation capacity of the relevant management agencies. These factors reflect the fine-tuning of quantitative analyses that is required to account for real-world constraints and opportunities, although the aim is to avoid such post hoc refinements and integrate all important considerations into the analysis. The results of a systematic and transparent assessment make explicit any trade-offs, compromises, and opportunities. When we qualitatively compare the results of our analyses to planning approaches within South Africa [27], we find a high degree of concordance, as we do when we compare the relative levels of investment in Phytophthora management and predator control in Australia [28–30]. Nevertheless, the identification of investment priorities through a systematic and transparent process will be extremely useful when local experts are not available or there is a need to remove individual biases. At a global or even national scale, there is likely to be a deficit of experts with knowledge of multiple regions and conservation actions and an ability to identify investment schedules across these in an integrated manner.
Our conservation investment framework offers substantial dividends for biodiversity conservation by prioritising the most appropriate and feasible conservation actions to abate the threats that operate in a region. Already, there has been a call for conservation organisations to audit their investments and measure their returns [31–34]. For conservation practitioners, this framework represents a much needed tool for incorporating their insights and experience regarding the costs, benefits, and dynamics of a suite of conservation actions to maximise conservation outcomes.
|
10.1371/journal.pbio.1002449 | Structural and Molecular Mechanism of CdpR Involved in Quorum-Sensing and Bacterial Virulence in Pseudomonas aeruginosa | Although quorum-sensing (QS) systems are important regulators of virulence gene expression in the opportunistic human pathogen Pseudomonas aeruginosa, their detailed regulatory mechanisms have not been fully characterized. Here, we show that deletion of PA2588 resulted in increased production of pyocyanin and biofilm, as well as enhanced pathogenicity in a mouse model. To gain insights into the function of PA2588, we performed a ChIP-seq assay and identified 28 targets of PA2588, including the intergenic region between PA2588 and pqsH, which encodes the key synthase of Pseudomonas quinolone signal (PQS). Though the C-terminal domain was similar to DNA-binding regions of other AraC family members, structural studies revealed that PA2588 has a novel fold at the N-terminal region (NTR), and its C-terminal HTH (helix-turn-helix) domain is also unique in DNA recognition. We also demonstrated that the adaptor protein ClpS, an essential regulator of ATP-dependent protease ClpAP, directly interacted with PA2588 before delivering CdpR to ClpAP for degradation. We named PA2588 as CdpR (ClpAP-degradation and pathogenicity Regulator). Moreover, deletion of clpP or clpS/clpA promotes bacterial survival in a mouse model of acute pneumonia infection. Taken together, this study uncovered that CdpR is an important QS regulator, which can interact with the ClpAS-P system to regulate the expression of virulence factors and pathogenicity.
| Although many transcriptional regulators tune P. aeruginosa virulence factor expression and secretion, the molecular mechanisms of the underlying regulatory network are still elusive. Quorum sensing, the ability of bacteria to communicate and detect cell density to determine the most advantageous time to orchestrate collective events, is known to govern P. aeruginosa virulence. In this study, we present a novel AraC-family transcription factor, CdpR (PA2588), that controls numerous virulence factors via directly regulating the quorum sensing Pseudomonas quinolone signal (PQS) system. We solved the crystal structure of CdpR, which showed that its N-terminal domain contains a unique fold that is different from other AraC-family proteins. In addition, we found that CdpR is regulated by the ClpAS-ClpP protease. CdpR interacts with the adaptor protein ClpS and then is degraded by the ATP-dependent ClpAP protease. This is the first example of a quorum-sensing regulator as a substrate of the ClpAS-ClpP protease in P. aeruginosa. These findings significantly extend our understandings of quorum sensing and virulence regulation and provide insights into the function of proteases in pathogenic bacteria.
| P. aeruginosa is one of the most common nosocomial pathogens associated with fatal lung disease in cystic fibrosis patients [1]. This bacterium synthesizes a group of virulence factors consisting of pyocyanin, rhamnolipids, proteases, and biofilms that are regulated by a cell density-dependent quorum-sensing (QS) system [2,3] and second messenger c-di-GMP [4].
QS systems enable bacteria to communicate and regulate a large number of genes. P. aeruginosa has one of the most sophisticated QS systems of all bacterial species, which includes two systems (las and rhl) based on N-acyl-homoserine lactones (AHLs) and one based on 2-alkyl-4-quinolone (AQ) [2]. The LasI synthase produces the signal of the las system, N-3-oxo-dodecanoyl homoserine lactone (3-oxo-C12-HSL), which is recognized by the transcriptional regulator LasR [5]. The RhlI synthase catalyzes the synthesis of N-butanoyl homoserine lactone (C4-HSL), which is sensed by the transcriptional regulator RhlR [6]. These two systems control 10% of the P. aeruginosa genome [7,8]. Besides the AHLs systems, P. aeruginosa also produces the AQ signal molecule [9]. Pseudomonas quinolone signal (PQS) is synthesized from the precursor molecule 2-heptyo-4(1H)-quinolone (HHQ) and finally converted into PQS by PqsH [10]. Moreover, these three systems and a group of transcriptional regulators (such as VqsR, QscR, VqsM, Vfr, and RpoN) form a complex regulatory network [11].
Intracellular protein degradation plays essential roles in many physiological processes such as cell growth and division. In Escherichia coli, proteolysis is mainly mediated by ATP-dependent proteolytic machineries (termed AAA+ proteases) [12], which consist of a proteolytic barrel (e.g., heptameric ClpP) and a hexameric ATPase (e.g., ClpA, ClpX). The ATP-binding proteins ClpX and ClpA recognize a specific peptide sequence of unfolded proteins [13,14], and translocate the tags into the peptidase barrel of ClpP, thus degrading the unfolded protein into small peptides [15]. In addition, adaptor proteins (i.e., ClpS) are responsible for binding and transporting their target proteins to the protease complexes for degradation [16]. In P. aeruginosa, the Lon and ClpP proteases share 84% and 86% identity with their E. coli counterparts, which suggest that these proteases also play important roles in the degradation of unstable or misfolded proteins in P. aeruginosa. Moreover, the ATP-dependent ClpP protease has been shown to control virulence, antibiotic resistance, and a group of metabolic pathways [17]. However, the substrates and molecular mechanism of ClpP in regulating bacterial pathogenesis are still unclear.
In this study, we showed that the new QS regulator CdpR (we named PA2588 in this study) directly activated PqsH, the PQS synthase. CdpR regulated the expression of virulence factors (i.e., pyocyanin and motility) and bacterial pathogenicity of P. aeruginosa. Unlike other AraC-family proteins, the C-terminal HTH (helix-turn-helix) motif of CdpR has a unique conformation, which may play important roles in DNA recognition. Structural comparison also revealed that the N-terminal region (NTR) of CdpR has a novel fold whose function remains elusive. We also found that CdpR interacts with adaptor protein ClpS and is degraded by the ClpAP protease in vivo. Moreover, the deletion of clpP or clpS/clpA-promoted bacterial survival in a mouse model of acute pneumonia. To the best of our knowledge, this is the first report showing the degradation of a regulator of the alginate system by ClpAS-P proteases in P. aeruginosa. Overall, these findings establish a link between proteases ClpAS-P and virulence, which provides us with new insights into investigate the roles of proteases in the future.
Previously, we showed that the QS regulator VqsM directly binds to the promoter region of the PA2588(cdpR) gene [18]. Its N-terminus (amino acids 59–198) is homologous to the arabinose-binding domain of the AraC transcription regulator (N-term; pfam12625), suggesting that it may be an AraC-family transcriptional regulator. Given that lasI and cdpR are directly controlled by VqsM, we hypothesized that CdpR also regulates P. aeruginosa QS-associated virulence factors, such as pyocyanin production. To this end, we made a cdpR null mutant in the PAO1 strain by following procedures from our previous study [18]. The resulting ΔcdpR mutant exhibited higher pyocyanin production than wild-type PAO1 after 24 h of growth in Luria-Bertani (LB) medium. Overexpression of cdpR in a pAK1900 plasmid (T7 promoter) decreased the pyocyanin production in the cdpR mutant (Fig 1A).
Two homologous operons are involved in the synthesis of phenazine compounds (i.e, pyocyanin) in P. aeruginosa, phzA1B1C1D1G1 (phzA1) and phzA2B2C2D2G2(phzA2) [19]. The increased pyocyanin production in the ΔcdpR strain led us to test if cdpR regulates the promoter activity of phzA1. To this end, a phzA1-lux reporter was introduced into the wild-type, the ΔcdpR mutant, and its complemented strain. The expression of phzA1-lux in the ΔcdpR mutant was 3-fold higher than in the wild-type PAO1 or its complemented strain (Fig 1B). These results suggest that CdpR is a negative regulator of pyocyanin synthesis.
The enhanced pyocyanin production in the ΔcdpR mutant led us to test if cdpR regulates other QS-related pathways, such as biofilm formation and bacterial motility. We performed microtiter dish and borosilicate tube binding assays to evaluate the biofilm formation of these strains. The ΔcdpR mutant exhibited better adherence than either the wild-type or the complemented strain (Fig 1C). In addition, deletion of cdpR resulted in reduced swarming motility (Fig 1D). Collectively, these results demonstrated that CdpR is a new and important regulator of P. aeruginosa virulence factors.
In order to dissect the regulatory mechanism of CdpR, we performed a ChIP-seq assay to identify its direct targets in the Pseudomonas genome. Like untagged CdpR, CdpR-VSV (overexpressed in the plasmid pAK1900) was also able to decrease the expression of phzA1 (Fig 1B).Using MACS software for identifying conserved DNA motif bound by CdpR [20], 28 enriched loci (p-value = e−5) were identified to carry CdpR-binding peaks (S1 Table), which were enriched > 1.5 folds compared to the control samples using the wild-type without VSV tags. Eighty-nine percent of these 28 loci are located within coding regions (Fig 2A). For example, CdpR binds to the coding region of PA0440 gene in the immunoprecipitated sample, while the enrichment in PA0440 was not present in the input sample (the wild-type PAO1 strain with empty pAK1900 vector) (Fig 2B). We further categorized the biological processes of CdpR targets based on gene ontology [21], including metabolism (28%), secretion (7%), regulation (4%), nucleic acid metabolism (3%), transportation (11%), virulence (11%), and unknown functions (36%) (Fig 2C). The ChIP-seq data also allowed us to define a 15-bp consensus-binding motif of CdpR (Fig 2D) using multiple EM for motif elicitation (MEME) analysis.
We next sought to perform electrophoretic mobility shift assays (EMSA) to verify the CdpR targets in vitro. We first expressed and purified the CdpR protein (S1A Fig). CdpR bound to two promoter regions of cerN and sphR, but not to the lasR promoter as a negative control (S1B–S1F Fig). We further tested 18 binding sites, and 13 showed specific binding (S2A Fig). Overall, these EMSA results are consistent with the ChIP-seq data and confirm the binding motif of CdpR in the P. aeruginosa genome.
The highest enriched region by CdpR was the shared promoter region between pqsH and cdpR (Fig 3A). EMSA showed that CdpR bound to the pqsH-cdpR intergenic region (pqsH-p) but not to the negative control (pqsR-p) (Fig 3B). Using a dye-based DNaseI footprinting assay, we further found a specific CdpR-bound sequence containing a 15-bp motif (5'-CTGCGCCTGGATGAT-3') (Fig 3C). The protected region extended from nucleotides -377 to -361 ahead of the pqsH translational start codon, which is consistent with the identified sequences by MEME analysis (Fig 2D). We have blasted the intergenic sequences between cdpR and pqsH in other Pseudomonas species. This region is largely conserved in P. aeruginosa, but not in other Pseudomonas species, suggesting that the shared promoter between cdpR and pqsH is universal in P. aeruginosa. We also noted that the motif (CTGCGCCTGGATGAT) is also present in the promoter regions of cerN (5'-CTTCGCCTGGTTGCC-3', -87 to -72 from the cerN translational start codon) and sphR (5'-CTGCGCCTGGCCGCC-3', -266 to -252 from the sphR translational start codon) (S2B Fig).
In order to determine if the motif (CTGCGCCTGGATGAT) was important for binding to CdpR, we performed EMSAs again using a pqsH-p1 DNA fragment (−365 to +88 relative to the start codon of pqsH, without the binding motif) and a pqsH-M-p DNA fragment (the motif CTGCGCCTGGATGAT was mutated to TGACTTCTGGATGAT). Neither the pqsH-p1 nor pqsH-M-p fragments were able to bind to CdpR (S3A Fig). As expected, CdpR could not bind to the sphR-p1 or cerN-p1 fragment (without the binding motif) (S1D and S1E Fig).
We next tested if CdpR controls the expression of pqsH or itself in vivo. As shown in Fig 3D, the promoter activity of pqsH (pqsH-lux, −500 to +88 relative to the start codon) in the ΔcdpR mutant strain was drastically weaker than that in the parental strain. Overexpression of cdpR gene in the ΔcdpR strain restored the expression of pqsH to the wild-type level. This result was further echoed by western blotting of the strains carrying an integrated mini-CTX-pqsH-flag plasmid (Fig 3E). We further measured QS molecules including PQS and C4-HSL, which showed that PQS production was lower in the cdpR mutant than in the wild-type PAO1 strain (Fig 3F); but no difference in C4-HSL production was observed between them (S3B Fig). Moreover, mutation in the consensus motif abolished the promoter activity of pqsH-M-lux in the wild-type strain (S3C Fig), suggesting that the binding sequence is required for full activity of the pqsH promoter. In addition, the expression of cdpR in the ΔcdpR mutant was 4-fold higher than that in the wild-type strain (S3D Fig), suggesting that the cdpR is negatively autoregulated. Taken together, these results indicate that CdpR regulates both pqsH and itself.
As aforementioned, the deletion of cdpR increased pyocyanin production and compromised the swarming motility and improved biofilm formation (Fig 1). We next sought to test whether the deletion of cdpR changes P. aeruginosa virulence in a mouse model of acute pneumonia. Therefore, 6-wk-old C57BL/6 mice were intranasally inoculated with 1×107 wild-type, ΔcdpR mutant, or complemented strain (single copy integration of the cdpR gene). Kaplan-Meier survival analysis showed that loss of cdpR significantly decreased mouse survival compared to the wild-type PAO1. The ΔcdpR strain caused 70% and 100% of infected mice to die by 12 h and 48 h, respectively. In contrast, 60% of mice infected with wild-type PAO1 or the complemented strain were alive at 48 h (Fig 4A). We also observed more lung injury and inflammation in the ΔcdpR strain-infected mice than the wild-type PAO1 or complemented strain-infected mice by histological analysis with H&E staining (Fig 4B). The colony-forming units (CFUs) in alveolar macrophages (AM) infected with ΔcdpR were higher than those with the wild-type or the complemented strain (Fig 4C). In addition, the ΔcdpR-infected mice produced more superoxide and more polymorphonuclear neutrophil (PMN) cells than wild-type-infected mice (Fig 4D and 4E). In summary, these results indicate that CdpR is a negative regulator of P. aeruginosa pathogenicity in a mouse model.
To gain more insights into the structural basis underlying the functions of PaCdpR, we carried out structural studies. The structure belongs to P41212 space group, and it contains one PaCdpR protein molecule in the asymmetric unit. As depicted in Fig 5A and 5B, the overall structure of PaCdpR can be divided into three regions, i.e., NTR (aa 5–203), connector region (CR, aa 204–228), and C-terminal HTH domain (aa 229–329). The NTR (Fig 5C) is of PaCdpR that is of α/β fold in nature; it is composed of ten α-helices (α1 to α10) and five β-strands (β1 to β5). The β strands form a single-layer β-sheet. The last strand (β5) sits in the center, whereas β2 and β1 locate on one side, and β3 and β4 locate on the other side. Except β4, which is parallel to β3, all the other strands are antiparallel to each other. The α-helices can be divided in two groups: group A (α1 to α5) and group B (α6 to α10). Group B α-helices were clamped between the A group helices and the β-sheet. Four helices, α9, α10, α6, and α7 were packed against the β-sheet at the edge in an anti-clockwise direction, while α8 (the longest helices) crossed the β-sheet in the middle and formed extensive hydrophobic interactions with the three central strands (β2, β5, and β3).
E. coli AraC (EcAraC) is composed of two domains, i.e., N-terminal domain (NTD) and C-terminal HTH domain, and is the most well studied member of AraC family. Upon the binding of arabinose, EcAraC NTD can change its dimerization mode, which will in turn alter the HTH–DNA interaction and affect its gene regulation activity [22]. Though both CdpR NTR and EcAraC NTD are of α/β fold in nature, their overall folds are different from each other. EcAraC NTD has a jelly-roll fold β-barrel, which holds the arabinose in between the double-layer beta-sheet (S4A Fig). Moreover, no sequence similarity was observed between CdpR NTR and EcAraC NTD. These differences suggested that CdpR NTR and EcAraC NTD may have different functions. To this end, we carried out an in vitro DNA binding assay. As shown in S4B Fig, the CdpR–DNA interaction was almost identical in the presence or absence of arabinose, suggesting that arabinose is not a cofactor of CdpR. Structural comparison using DALI server [23] did not reveal significant similarity between CdpR NTR and other known AraC family structures. Altogether, these data indicate that CdpR NTR has a novel fold among AraC family, and its functions need to be further investigated.
HTH domains are conserved in AraC family. Besides CdpR, the HTH domain structures of EcAraC (PDB ID: 2K9S) [24], EcMarA (multiple antibiotic resistant regulon A, PDB ID: 1BL0) [25], and Streptomyces griseus AdpA (PDB ID: 3W6V) [26] have also been reported. Although the overall folds are similar, structural comparison revealed that the detailed conformation of CdpR HTH is different from other AraC proteins (Fig 6A). CdpR HTH contains seven α helices (α13–α19), which can be further divided into two subdomains (S5A Fig), i.e., HTH1 (aa 229–271) and HTH2 (aa 291–329). The largest difference occurs at the HTH1 motif, especially the helix α13. Consistently, the sequence similarity at this region is also very low (S5B Fig).
The interaction modes between HTH1 and HTH2 motifs are also different within these structures. In EcMarA and sgAdpA structures, loop A and loop B interact with each other via backbone van der Waals contacts; however, as depicted in Fig 6B, such interaction was not observed in CdpR, indicated by the shortest distance (6.7 Å) between the loop backbones. Instead, CdpR HTH1 and HTH2 mainly interact with each other through the backbone of loop B (aa 303–304) and the side chain of Tyr245 from loop A. This interaction was further enhanced by Arg281 and Arg284 of helice α16, which connects HTH1 and HTH2 and Tyr305 of loop B, via H-bonding and stacking. Both Tyr245 and Arg284 of CdpR are not conserved in other AraC-family proteins.
HTH domains are responsible for target DNA recognition, and the complex structures have been solved for EcMarA and sgAdpA. To better understand the mechanism of CdpR-DNA recognition, we also carried out molecular modeling study. CdpR HTH may recognize the DNA base pairs using two positive charged residues (Arg260 and Arg264), corresponding to Arg262 and Arg266 of sgAdpA. In sgAdpA HTH–DNA structure, DNA backbones form several H-bonds with the side chains of positive charged residues (Fig 6C and 6D). Arg312 and Arg317 of CdpR HTH are corresponding to Arg315 and Arg320 of sgAdpA. The other four residues (Arg261, Arg269, Arg309, and His317) of sgAdpA are not conserved, but Arg263 of CdpR may replace Arg261 of sgAdpA during DNA recognition, supported by the similar orientations of their guanine groups. Structural analysis did not reveal any residues of CdpR, which could form H-bonds with DNA similar to Arg269, Arg309, and His317 of sgAdpA; lack of extensive interactions with DNA backbones may be the main cause of the weak CdpR-DNA binding affinity (Fig 6D).
In order to experimentally confirm the proposed structural model for CdpR-DNA complex, conserved residues predicted to contact DNA (Arg263, Arg274, Arg312, and Arg317) were mutated to Ala. We next performed EMSAs for comparing the ability of purified wild-type CdpR to bind a DNA probe encompassing the pqsH-cdpR intergenic region with those of its mutated derivatives. As shown in Fig 6E, CdpRwt was able to shift about 50% of the probe at a concentration of 0.5 μM, while CdpRR263A, CdpRR274A, and CdpRR312A/R317A were unable to shift the same DNA probe even at a concentration of 2 μM. To future confirm these residues are important for CdpR activity, we tested the expression of pqsH-lux in the cdpR mutant containing the expressing plasmid p-cdpRR263A, p-cdpRR274A, or p-cdpRR312A/317A. As shown in Fig 6F, mutation of these residues did not restore the pqsH activity of cdpR mutant to wild-type levels. These results clearly suggest that these residues are important for CdpR activity and confirm the CdpR-DNA model is accurate.
In order to identify regulators of cdpR, its promoter was used as a reporter (CTX-cdpR-lux) and subjected to transposon mutagenesis. After three rounds of screening of 8,000 mutants, 4 positive and 10 negative mutants were selected. Genes inserted by transposon were determined via arbitrary primed PCR and subsequent DNA sequencing (S2 Table). Interestingly, mutation of lasI compromised the expression of cdpR compared to wild-type PAO1, suggesting that LasI is a positive regulator of cdpR. The expression of cdpR was increased in the M6 (cdpR::Tn) strain compared to the wild-type strain, which is consistent with the observation in the cdpR deletion mutant (S6A Fig). As expected, pyocyanin production was also elevated in the M6 strain (S6B Fig). We noted that vqsM was missing in our screen, which might result from limited number (8,000) of mutants.
Of these genes, we are most interested in the ATP-dependent protease ClpP, which associates with ClpS (ATP-dependent Clp protease adaptor protein) and ClpA (ATP-binding protease component). These mutants showed very strong phenotypes, revealing that they are involved in CdpR-controlled pathway. Moreover, previous study has shown that ClpP is related to bacterial virulence [27]. clpS and clpA share the same operon [28]. To further verify the effects of the clpS/clpP transposon insertion, we made a clpS-clpA double mutant (ΔclpSΔclpA) and a clpP single mutant (ΔclpP). As expected, the activity of cdpR-lux in these two deletion strains was higher than that in the wild-type strain (S7A Fig).
We next wanted to verify whether the protein level of CdpR was affected by these proteases. To this end, we constructed the CTX-cdpR-flag fusion and the activity of CdpR-flag was shown the same as the wild-type CdpR (S7B Fig). Next, we tested the expression of CdpR-Flag in the wild-type PAO1, the ΔclpSΔclpA mutant, the ΔclpP mutant, and their complemented strains (ΔclpSΔclpA/p-clpSclpA, and ΔclpP/p-clpP) using western-blot assay. Higher levels of CdpR-Flag were detected in the ΔclpSΔclpA or ΔclpP mutant than in the wild-type strain. The complemented strains restored CdpR-Flag to the wild-type levels (S7C Fig). As expected, the expression of pqsH in ΔclpSΔclpA or ΔclpP mutants was also higher than the wild-type PAO1 (S8 Fig). Taken together, these results indicate that ClpS/ClpA and ClpP negatively regulate the expression of CdpR.
Given that ClpS is a bacterial adaptor that recognizes and delivers N-degron (N-terminal amino acids) substrates to the ClpAP or ClpCP AAA+ proteases [29,30], we hypothesized that ClpS interacts with CdpR by a bacterial two-hybrid system. Like strain 1 (the positive control), strain 4 (bearing pBT-CdpR and pTRG-ClpS) also grew on dual-selective medium (Fig 7A), indicating a direct interaction between CdpR and ClpS.
The direct interaction between ClpS and CdpR led us to examine if the CdpR protein can be degraded by the ATP-dependent ClpAP protease in vivo. We transformed mini-CTX-cdpR-flag into the wild-type PAO1, the ΔclpP, the ΔclpSΔclpA, and their complemented strains. Western-blotting analyses of cell lysates were performed over the course of 90 min. Degradation of CdpR-Flag was observed in the wild-type PAO1 and the complemented strains with a half-life of less than 30 min (Fig 7B). However, CdpR levels remained steady in the ΔclpP and the ΔclpSΔclpA strains (the second and fourth panels in Fig 7B), demonstrating that the ClpSA-ClpP protease system is responsible for CdpR degradation.
A group of proteases has been shown to modulate P. aeruginosa virulence. For example, the deletion of Lon protease exhibits a defect in cell division and virulence-related properties, such as swarming, twitching, and biofilm formation [31]. It was also demonstrated that the lon mutant was less virulent in a mouse acute lung infection model as well as in a rat model of chronic infection [32]. The protease ClpP is also involved in antibiotic resistance and virulence-related phenotypes, such as motility and biofilm formation [17]. In the present study, ClpS/ClpA and ClpP controlled the expression of CdpR, which negatively regulated pyocyanin production and bacterial virulence (Fig 1). We showed that the ΔclpSΔclpA and ΔclpP mutants produced more pigments than the wild-type PAO1 (Fig 8A and 8B). The deletion of cdpR in the ΔclpSΔclpA or ΔclpP mutants restored pigment production and phzA1-lux activity to the wild-type level (Fig 8C). We also observed that the swarming motility of the ΔclpAΔclpS and the ΔclpP strains could be partially reversed by CdpR (S9 Fig). Taken together, these results indicate that ClpSA-ClpP plays important roles in regulating pigment and motility. CdpR interacts with these intracellular proteases with important functions.
These altered virulence-associated phenotypes in the ΔclpP and ΔclpSΔclpA mutants led us to evaluate the pathogenicity of these mutants in a mouse model. We found that deletion of clpS/clpA or clpP improved mouse survival compared to the wild-type PAO1. The ΔclpSΔclpA and ΔclpP strains exhibited 100% and 85% of mice survival after 96 h, whereas the wild-type bacteria caused 70% death at 80 h (Fig 8D). These results demonstrated that ClpS/ClpA and ClpP proteases play important roles in P. aeruginosa pathogenicity in a mouse model.
Here, we present the first characterization of CdpR as an important regulator of QS system that directly binds to the promoter region of pqsH. We also solved the crystal structure of CdpR and revealed that CdpR NTD has a unique fold that is different from other AraC-family proteins. In addition, we found that CdpR interacts with ClpS before degradation by ClpAP proteases, thus regulates the P. aeruginosa virulence.
Although the QS systems in P. aeruginosa have been widely characterized in the last several decades, the complete signaling network and interactions with other systems still remain elusive. Previously, we showed that the QS regulator VqsM interacts with the promoter of cdpR [18]. Like VqsM, CdpR belongs to the AraC-family transcriptional regulators, which control a variety of cellular processes in bacteria, including carbon metabolism, stress responses, and virulence [21]. We performed a ChIP-seq assay that revealed 28 CdpR targets in the P. aeruginosa genome. More importantly, CdpR directly regulates the PQS system by binding to the promoter of pqsH (Fig 3 and S3 Fig). The ΔcdpR mutant exhibited higher pyocyanin and biofilm production as well as reduced bacterial motility (Fig 1), which are controlled by the PQS system [9,33]. Moreover, CdpR plays important roles in regulating bacterial pathogenicity in a mouse experiment. Mice inoculated with the ΔcdpR mutant showed more bacterial loads than mice infected with the wild-type strain (Fig 4A). The increased pathogenicity corresponds to the higher production of pyocyanin in the ΔcdpR mutant (Fig 1A).
Like the other AraC-family proteins, CdpR has an NTD whose function is deemed as binding to small molecules and a HTH region, which binds to a special substrate. From our ChIP-seq results, CdpR can target many genes, but its DNA binding affinity is relatively lower than other identified AraC proteins. This is due to the lack of positively charged residues in HTH domain that can interact with the DNA backbones (Fig 6 and S5 Fig). Unlike other AraC-family proteins that can bind arabinose using NTD domain and alter the gene regulation activity, CdpR has a novel NTR region, and its interaction with target DNAs is not affected by arabinose (S4 Fig). CdpR NTR may also play a role in the gene regulation via the interaction with other proteins or small molecules. Besides, analysis of the CdpR structure revealed that the conformation of CdpR HTH is different from other AraC proteins (Fig 6A). The proposed structural model for CdpR-DNA complex demonstrated that Arg263, Arg274, Arg312, and Arg317 are important for DNA-binding ability of CdpR (Fig 6D), which has been verified by in vitro EMSA and in vivo complemented assays where these residues were mutated (Fig 6E and 6F). Overall, our results provide insights into the function of AraC-family protein in different bacteria.
To further understand the regulatory mechanism upstream of CdpR, we performed a transposon mutagenesis assay that revealed several mutants showing altered expression of cdpR (S2 Table). The expression of cdpR was highly compromised in a mutant containing a transposon insertion in clpAP, which encodes ATP-dependent proteases [34]. Proteases ClpAP and Lon are involved in bacterial virulence expression, antibiotic resistance, and metabolism [27]. Lon protease negatively modulates QS by degrading autoinducer synthase LasI [35]. Our present study showed that CdpR interacted with the adaptor ClpS, and CdpR protein was degraded by the ClpAP protease in vivo (Fig 7). To the best of our knowledge, this is the first evidence showing the substrates of ClpS-ClpAP in P. aeruginosa. Further work needs to focus on the precise binding sites between ClpS and CdpR.
In conclusion, we identified an AraC-family regulator, CdpR, which negatively modulates bacterial virulence (Fig 9). CdpR is a substrate of ATP-dependent ClpAP protease. CdpR also regulates the expression of pqsH by directly binding to its promoter regions. CdpR is positively modulated by LasI (S2 Table) and VqsM [18]. CdpR may work with other negative QS regulators, such as QscR and RsaL, to negatively regulate QS. In addition, ClpS interacts with CdpR and then associates with ClpAP to degrade CdpR. CdpR might indirectly regulate the expression of virulence factors, thus negatively controls P. aeruginosa pathogenicity. This work provides insight into roles of proteases in QS and bacterial virulence, which provides more cues to design effective antimicrobial drugs for the prevention of P. aeruginosa infections in the future.
All the experiments involving animals were approved by the University of North Dakota Institutional Animal Care and Use Committee and performed in accordance with the animal care and institutional guidelines (IACUC approval #: 1204–5) (Assurance Number: A3917-01). Animal experimental procedures, including treatment, care, and end point choice, followed Animal Research: Reporting in Vivo Experiment guidelines. When moribund animals manifested distress and/or mortal signs, such as lethargy, weight loss (15%–25%), lack of eating or drinking, tissue rupture, or edema, we considered these as endpoints and humanely euthanized them immediately by CO2. Recovery of the animals was achieved without complications, but we monitored hourly until they were awake and active after procedures and then every 6 h. Animals were monitored daily throughout the experimental process. Once sedation was diminished, the mice acted normally and appeared playful. There generally was no apparent pain from the intranasal instilling procedure. If any pain sign occurred, a 1% solution of lidocaine was applied to lessen pain.
The bacterial strains and plasmids used in this study are listed in S3 Table. P. aeruginosa PAO1 and derivatives were grown at 37°C on LB agar dishes or in broth with shaking at 220 rpm. Antibiotics were used at the following concentrations: for E. coli, gentamicin (Gm) at 15 μg/ml, ampicillin at 100 μg/ml, and tetracycline 10 μg/ml; for P. aeruginosa, gentamicin (Gm) at 50 μg/ml in LB or 150 μg/ml in PIA (Pseudomonas Isolate Agar); Tetracycline at 150 μg/ml in LB or 300 μg/ml in PIA and Carbenicillin at 500 μg/ml in LB.
The procedures of chromatin immunoprecipitation (ChIP) were modified from previously described studies [18,36]. Wild-type PAO1 strain carrying pAK1900 or pAK1900-CdpR-VSV was cultured in LB medium until OD = 0.6 and then crosslinked with 1% formaldehyde (final concentration) for 10 min at 37°C with shaking. 125 mM glycine (final concentration) was added into the culture to stop the crosslinking. Bacteria were centrifuged and washed three times with a Tris buffer (150 mM NaCl, 20 mM Tris-HCl, pH 7.5). The pellets were resuspended in 500 μl IP buffer (150 mM NaCl, 50 mM HEPES-KOH pH 7.5, 1 mM EDTA, 0.1% SDS, 1% Triton X-100, 0.1% sodium deoxycholate and mini-protease inhibitor cocktail (Roche), and the DNA was sonicated to 100–300 bp. The chromatins were centrifuged (12,000 rpm, 4°C), and the supernatant was saved. All samples were treated by protein A beads (General Electric) before adding 50 μl agarose-conjugated anti-VSV antibodies. Washing, crosslink reversal, and purification of the ChIP DNA were performed by following previous procedures [36]. We used agarose gel to cut DNA fragments between 150 and 250 bp, which were then used for library construction with NEXTflex ChIP-Seq Kit (Bioo Scientific). After sequencing the libraries on HiSeq 2000 system (Illumina), the reads were mapped to the genome of P. aeruginosa PAO1 with TopHat (Version 2.0.0) [37]. MACS software (version 2.0.0) were then used to call peaks [20], which was subjected to MEME analyses to calculate the specific CdpR-binding motif [38].
Plasmids p-cdpR, p-clpS/clpA, and p-clpP were constructed respectively by amplifying fragments with the corresponding primer pairs (S4 Table) pAK-cdpR-A/pAK-cdpR-S, pAK-clpS-A/pAK-clpA-S, and pAK-clpP-A/pAK-clpP-S by PCR. The PCR products were digested with the indicated enzymes and cloned into PAK1900 [39]. For construction of CTX-clpP and CTX-clpA/clpS, the sequence including its promoter and coding region was PCR-amplified with primers: CTX-clpP-F/CTX-clpP-R and CTX-clpA/clpS-F/CTX-clpA/clpS-R. These fragments were respectively ligated with Mini-CTX-lux [40].
The mini-cdpR-flag-A/mini-cdpR-flag-S primers were used to amplify the cdpR gene that intended to fuse with a C-terminal Flag-tag. The indicated enzyme-digested PCR products were cloned into the corresponding enzyme sites of Mini-CTX-lacZ to generate either Mini-CTX-cdpR-flag.
The plasmid pMS402 carrying a promoterless luxCDABE reporter gene cluster was used to construct promoter-luxCDABE reporter fusions of the cdpR as previously described [18]. The cdpR promoter region was amplified by PCR using the primers cdpR-lux-F (with XhoI site) and cdpR-lux-R (with BamHI site) (S4 Table). The PCR products were cloned into the pMS402, yielding p-cdpR-lux. Besides the plasmid-based reporter system, an integration plasmid CTX6.1 originating from plasmid mini-CTX-lux was used to construct chromosomal fusion reporter. The pMS402 fragment containing the kanamycin-resistance marker, the MCS, and the promoter-luxCDABE reporter cassette was then isolated and ligated to CTX6.1, yielding CTX-cdpR-lux. The plasmid generated was first transferred into E. coli SM10-λ pir, and the P. aeruginosa reporter integration strain was obtained using biparental mating as previously reported [41]. All constructs were sequenced to verify that no mutations occurred for these constructs.
For construction of gene knockout mutants, a SacB-based strategy was employed as described in previous and our studies [42]. To construct the cdpR null mutant (ΔcdpR), PCRs were performed to amplify sequences upstream (1,975 bp) and downstream (1,242 bp) of the intended deletion. The upstream fragment was amplified from PAO1 genomic DNA using primer pair, pEX-cdpR-up-S, and pEX-cdpR-up-A, while the downstream fragment was amplified with primer pair, pEX-cdpR-down-S, and pEX-cdpR-down-A (S4 Table). The two PCR products were digested and then cloned into BamHI/HindIII-digested gene replacement vector pEX18Ap, yielding pEX18Ap-cdpR. A 0.9 kb gentamicin resistance cassette cut from pPS858 with XbaI was cloned into pEX18Ap-cdpR, yielding pEX18Ap-cdpR-Gm. The resultant plasmids were electroporated into PAO1 with selection for gentamicin resistance. Colonies showing both gentamicin resistance and loss of sucrose (5%) susceptibility were selected on LB agar plates containing 50 μg/ml of gentamicin and 5% sucrose, which typically indicates a double-crossover event and thus of gene replacement occurring. The pEX18Ap-cdpR-Tc was constructed by a similar strategy as described above. A 2.3 kb tetracycline resistance cassette was amplified from integration vector mini-CTX-lacZ with primer pair Tc-S/Tc-A (with XbaI site) (S4 Table) for replacing the cdpR gene in PAO1. The ΔcdpR mutant was further confirmed by PCR. The pqsH, clpS/clpA, and clpP mutant was generated by a similar strategy with deletion of cdpR gene in PAO1.
For generating cdpR/pqsH, cdpR/clpSclpA, and cdpR/clpP deletion strains (ΔcdpRΔpqsH, ΔcdpRΔclpSΔclpA, ΔcdpRΔclpP), the cdpR gene in ΔpqsH, ΔclpSΔclpA, and ΔclpP mutant was deleted by a similar strategy with plasmid pEX18Ap-cdpR-Tc. These resultant mutants were verified by PCR.
The plasmid encoding the His-Sumo-PaCdpR was constructed, through PCR reaction and DNA ligation, and transferred into E. coli strain BL21 DE3. The recombinant strains were cultured in 1L LB medium supplemented with 50 ug/ml kanamycin at 37°C. Protein expression was induced at OD600 ≈ 0.6 by addition of isopropyl β-D-1-thiogalacto-pyranoside (IPTG) with a final concentration of 0.2 mM. The induced cultures were then grown at 18°C for additional 18 h. The cells were harvested by centrifugation and resuspended in lysis buffer (20 mM Mes pH 6.5, 200 mM (NH4)2SO4, 200 mM NaCl, 25 mM Imidazole pH8.0) and then were lysed twice under high pressure via JNBIO homogenizer. The homogenate was clarified by centrifugation, and the supernatant was loaded onto Ni-NTA column (GE healthcare). The protein was eluted by elution buffer (20 mM Mes pH 6.5, 300 mM (NH4)2SO4, 150 mM NaCl, 500 mM Imidazole pH 8.0) using a stage-wise gradient. The fractions containing the recombinant His-Sumo-PaCdpR protein were pooled and dialyzed against Buffer S (20 mM Mes pH6.5, 200 mM (NH4)2SO4, 200 mM NaCl, 1‰ β-Me) for three hours at room temperature with Ulp1 protease added. The cleaved sample was not stable during dialysis and additional 100 mM (NH4)2SO4, 100 mM NaCl, and 2‰ β-Me were added to redissolve the protein precipitation. The sample was loaded onto Ni-NTA column again to remove the cleaved His-SUMO tag. Target PaCdPR protein was contained in the flow-through, and its purity was analyzed by SDS-PAGE. Protein was concentrated via Amicon-Ultra centrifugal device from Millipore and stored at -80°C freezer until use. Selenomethionine-substituted PaCdPR were expressed in M9 medium supplemented with 60 mg/L Se-Met (J&K), its purification procedure is similar as the full-length native protein. The purity was verified by SDS-PAGE gel (S1A Fig).
Crystallization was performed at 16°C using the Gryphon robot system from Arts Robbin Instrument Company, which identified the initial crystallization condition for full-length PaCdpR. The condition is composed of 1.0 M (NH4)2HPO4 and 0.1 M Acetate (pH4.5), which gives very tiny rod-shaped crystals. Relative larger crystals of Se-PaCdpR (0.1×0.1× 0.2 mm) were obtained after several runs of optimization. The protein concentration is 3 mg/ml, and the well solution is similar as that of native proteins. 0.001 M TCEP was added to the crystallization condition as additive. Crystals of Se-PaCdpR were cryoprotected by soaking in the mother liquid supplemented with 20% glycerol for 30 sec and then were flash-frozen by liquid nitrogen. The X-ray diffraction data was collected on beamline BL17U at Shanghai Synchrotron Radiation Facility (SSRF) at cryogenic temperature, maintained with Cryogenic system. One single crystal was used, and data processing was carried out with HKL2000 [43]. The data collection and processing statistics were summarized in S5 Table.
The structure of Se-PaCdpR was solved using the SAD method with the Autosol program [44] embedded in the Phenix suite, which built an initial model that covered about 75% of the residues. The refinement was done using the Refmac5 program of ccp4 [45]; and during refinement, 5% of the data was randomly selected and set aside for free R-factor cross validation calculations. The 2Fobs-Fcalc and Fobs-Fcalc electron density maps were regularly calculated and used as a guide for the building of the missing amino acids and solvent molecules using Coot [46]. The Rwork and Rfree of the final model are 18.8% and 23.4%, respectively. The rmsd of bond and angle is 0.005 Å and 0.963°, respectively. Other refinement parameters were also summarized in S5 Table. The structure factors and atomic coordinates can be found in the Protein Data Bank with the access code 5CHH.
The complex structures have been solved for two AraC family proteins, EcMarA (PDB code: 1BL0) and sgAdpA (PDB code: 3W6V). The overall structure of PaCdpR HTH is more similar to sgAdpA HTH than EcMarA HTH, with rmsds of 1.7 Å and 2.1 Å, respectively. In addition, the target DNAs of sgAdpA and PaCdpR are GC rich, whereas it is AT rich for EcMarA targeting DNA; therefore, the sgAdpA HTH–DNA structure was used as reference during the model building. The PaCdpR HTH–DNA complex was modeled in three steps. In the sgAdpA HTH–DNA complex structure, two symmetry-related DNA duplexes interact with the two DNA recognition helices (which are α15and α18 in PaCdpR) from one protein molecule; and the two DNA duplexes are not broken, due to the lacking of 5'-phosphate groups. Therefore, the first step was the generation of 5'-phosphate groups and bridging it with the 3'-OH of its neighboring nucleotide, resulting continuous DNA duplex. Second, the modified DNA was docked on the PaCdpR HTH structure, via the superimposing of the modified sgAdpA HTH–DNA structure with PaCdpR HTH. The rmsd between the overall structure of PaCdpR HTH and sgAdpA HTH is 1.7 Å, based on 111 pairs of Catoms. Finally, the side chains of some residues, including Arg260, Arg263, Arg264, Arg265, Arg274, and Arg317, were adjusted to avoid the clashes between protein and DNA. No further conformational change or energy minimization was involved in the modeling.
CdpR proteins were mixed with DNA probes (S4 Table) in 20 μl of the gel shift-loading buffer (20 mM HEPES, pH 8.0, 100 mM NaCl, 0.5 mM dithiothreitol, 10% Glycerol, and 3 μg/ml sheared salmon sperm DNA). After incubation at room temperature for 20 min, the samples were analyzed by 6% polyacrylamide gel electrophoresis in 0.5×TBE (Tris/Boric Acid/EDTA) buffer at 90 V for 90 min. The gels were stained by SYBR GOLD dye and subjected to screen on a phosphor screen (Tanon 5500).
The DNA footprint assay was performed by following previous procedures [18,47]. A 588-bp DNA containing the pqsH promoter region (–500 to +88) was amplified with primers pqsH-gf (with a 6-FAM modification at the 5’) and pqsH-gr (S4 Table). Forty nM of 6-FAM-labeled pqsH promoter probe was incubated with 2 μM of CdpR in gel-shift loading buffer. The protein-DNA mixtures were then partially digested with 0.05 units of DNase I (NEB) for 5 min at 25°C. The reaction was quenched by 0.25 M EDTA and purified with phenol-chloroform-isoamylalcohol (25:24:1) and then QIAquick PCR Purification kit (Qiagen). Control samples were done without CdpR protein. The genotype samples were run with the 3730 DNA Analyzer, and viewed with Peak Scanner (Applied Biosystems).
The PAO1 containing the CTX-cdpR-lux reporter fusion was subjected to transposon mutagenesis using the mariner transposon vector pBT20 [48]. Briefly, the donor strain (E. coli SM10-λpir) containing pBT20 and recipient PAO1-containing CTX-cdpR-lux were scraped from overnight plates and resuspended in 1 ml of M9 minimal medium. The bacterial suspensions were adjusted to an OD600 of 40 for the donor and an OD600 of 20 for the recipient. Next, 25 μl of each donor and recipient were mixed together and spotted on a dry LB agar plate and incubated at 37°C for overnight. Mating mixtures were scraped and resuspended in 1 ml of M9 minimal medium. Transposon-mutagenized bacteria were selected by plating on PIA plates containing gentamicin at 150 μg/ml. A transposon mutant library was constructed by picking 20,000 colonies grown on these selective plates. The mutants with altered expression of CTX-cdpR-lux were selected. The transposon insertion sites were determined by arbitrary primed PCR and subsequent sequencing of the PCR product [49].
Bacterial two-hybrid experiments were accomplished using the BacterioMatch II Two-Hybrid System Vector Kit (Agilent Technologies). The fragments of CdpR were cloned into the bait vector pBT in order to create a fusion protein with λ repressor protein (λcI). The DNA fragment of ClpS was inserted into the vector pTRG in frame with the α-subunit of RNA polymerase. The pBT-derived plasmids and pTRG-derived plasmid were then cotransformed into the validation reporter strain. Then, the resulting strains, grown at 37°C in SOC medium for 90 min, were harvested by centrifugation (5,000 g, 2 min) and washed twice with M9 + His-dropout broth. The cells were diluted to different gradients, spotted on the nonselective screening medium and selective screening medium, and finally incubated at 30°C for 2–4 d. The colonies grown on the selective screening medium were selected and streaked on the dual screening medium for further verification. The cotransformant containing pBT-LGF2 and pTRG-Gal 11P plasmids was used as the positive control.
Expression of lux-based reporters from cells grown in liquid culture was measured as counts per second (cps) of light production in a Synergy 2 (Biotek) as previously described [18]. Overnight cultures were diluted to OD600 = 0.2 with fresh LB medium, and then cultured for another 2 h. The cultures were added into a black 96-well plate with a transparent bottom. Sixty μl of sterilized mineral oil was added to culture. Promoter activities bacterial growth (OD = 595) were continuously measured every 30 min for 24 h in a Synergy 2 Plate Reader (BioTek).
The in vivo degradation assay was carried out to assess the stability of CdpR-Flag protein in P. aeruginosa as previously described [50,51]. Briefly, overnight LB cultures of tested strains were diluted 100-fold in fresh 20 ml LB medium in an Erlenmeyer flask with a flask volume-to-medium volume ratio of 5:1, and were aerated by shaking at 220 rpm. When OD600 value of the culture reached to approximately 0.4, chloramphenicol (a final concentration at 75 μg/ml) was added to the culture in order to block the translation, and samples for western blot analysis were removed at the indicated times. For the detection of CdpR-Flag, overnight LB cultures of the indicated strains were 1:100 diluted with LB medium recultured until A600 = 0.6. One hundred μl cultures were washed and resuspended in 15 μl PBS buffer.
The samples were solubilized in the SDS-PAGE loading buffer (50 mM Tris-HCl, pH 6.8; 2% SDS; 0.1% bromophenol blue; 1% mercaptoethanol; 10% glycerol) and then heated at 100°C for 15 min. SDS polyacrylamide gel electrophoresis was carried out according to the method of Laemmli [52] using a 10% slab gel with a 5% stacking gel and transferred onto PVDF (Bio-Rad) membranes. And then incubated with a mouse anti-Flag antibody (AOGMA, AGM12165) or anti-RNAP (Neoclone, WP003), followed by a sheep anti-mouse IgG antibody conjugated to horseradish peroxidase (HRP) (Code#: NA931, GE Healthcare), respectively. The relative abundance was determined by densitometric analysis using ImageQuant software.
Biofilm formation was measured in a static system as previously described [53], with minor modifications. Visualization of biofilm formation was carried out in 15-mL borosilicate tubes. Briefly, cells from overnight cultures were inoculated at 1:100 dilutions into LB medium supplemented with appropriate antibiotics and grown at 30°C for 10 h. Biofilms were stained with 0.1% crystal violet (CV) and tubes were washed with water to remove unbound dye. Quantification of biofim formation was performed in 24-well polystyrene microtiter plates. LB and appropriate antibiotics was inoculated to a final OD600nm of 0.01. The plates were incubated for 8 h or 14 h at 30°C. Crystal violet was added to each tube and stained for 15 min prior to removal by aspiration. Wells were rinsed three times by submerging the tubes in distilled water, and the remaining crystal violet was dissolved in 1 ml of 95% ethanol. A 1 ml portion of this solution was transferred to a new polystyrene tube, and the absorbance was measured at 600 nm.
Pyocyanin was extracted from culture supernatants and measured using previously reported methods [54]. Briefly, 3 ml chloroform was added to 5 ml culture supernatant. After extraction, the chloroform layer was transferred to a fresh tube and mixed with 1 ml 0.2 M HCI. After centrifugation, the top layer was removed and its A520 was measured. The amount of pyocyanin, in μg/ml, was calculated using the following formula: A520/A600 × 17.072 = μg of pyocyanin per ml.
Quantification of PQS production as described previously [55]. Briefly, singles colonies were inoculated in 10 ml LB for 16 h at 37°C. Cultures were diluted 1:100 into fresh media and grown for 24 h as above. A 500 μl aliquot of each culture was mixed with 1 ml of acidified ethyl acetate, vortexed vigorously for 2 min, and then centrifuged for 10 min at 16,000 × g. The organic phase was transferred to a fresh tube and dried to completion. The solute was dissolved in 50 μl of a 1:1 mix of acidified ethyl acetate: acetonitrile for analysis.
The autoinducer of the rhl system, C4-HSL, was measured using an rhlA promoter-based P. aeruginosa, pDO100 (pKD-rhlA) as previously described [56]. Two microlitres of test bacterial culture (OD600 = 1.0) were inoculated onto the seeded bioassay plates, and the plates were incubated at 37°C for 18 h. The dark halo zone around bacterial colonies indicates AHL activity.
Overnight culture of bacteria were 1:100 diluted with fresh LB medium and recultured until OD600 = 0.6. Six-week-old female C57BL6 mice were bought from the Harlan Laboratory (Indianapolis, IN). The animal procedures have been approved by the University of North Dakota institutional animal care and use committee (UND IACUC). Mice were anesthetized with 40 mg/kg ketamine plus 5 mg/kg diazepam. 1 × 107 CFUs of P. aeruginosa were intranasally instilled into mice and survival rates were calculated for every bacterial strain.
AM cells were cultured in 96-well plates at 37°C with 5% CO2 overnight. NBT dye (Sigma) was added to AM cells following the manufacturer’s procedures. The yellow-colored NBT changes to blue when oxidized by superoxide from AM cells [57]. The plate was kept at room temperature overnight for complete formation of dye product, which is monitored by a plate reader at 560 nm. Each experiment was conducted in three repeats.
After BAL procedures and serum collection, lung tissues were fixed in 10% formalin using a routine histologic procedure. Ten μl of BAL and serum were applied on microscope slide. The PMN numbers were counted under a light microscope using a Hema staining kit (Thermofisher). Homogenizations of lung tissues were done using liquid nitrogen then dissolved in RIPA buffer and sonicated (10 s sonication with 10 s intervals, 3 times) for next analysis. Tissue damage was determined by H&E staining in formalin-fixed tissues [57].
|
10.1371/journal.ppat.1000386 | High Throughput Functional Assays of the Variant Antigen PfEMP1 Reveal a Single Domain in the 3D7 Plasmodium falciparum Genome that Binds ICAM1 with High Affinity and Is Targeted by Naturally Acquired Neutralizing Antibodies | Plasmodium falciparum–infected erythrocytes bind endothelial receptors to sequester in vascular beds, and binding to ICAM1 has been implicated in cerebral malaria. Binding to ICAM1 may be mediated by the variant surface antigen family PfEMP1: for example, 6 of 21 DBLβC2 domains from the IT4 strain PfEMP1 repertoire were shown to bind ICAM1, and the PfEMP1 containing these 6 domains are all classified as Group B or C type. In this study, we surveyed binding of ICAM1 to 16 DBLβC2 domains of the 3D7 strain PfEMP1 repertoire, using a high throughput Bioplex assay format. Only one DBL2βC2 domain from the Group A PfEMP1 PF11_0521 showed strong specific binding. Among these 16 domains, DBL2βC2PF11_0521 best preserved the residues previously identified as conserved in ICAM1-binding versus non-binding domains. Our analyses further highlighted the potential role of conserved residues within predominantly non-conserved flexible loops in adhesion, and, therefore, as targets for intervention. Our studies also suggest that the structural/functional DBLβC2 domain involved in ICAM1 binding includes about 80 amino acid residues upstream of the previously suggested DBLβC2 domain. DBL2βC2PF11_0521 binding to ICAM1 was inhibited by immune sera from east Africa but not by control US sera. Neutralizing antibodies were uncommon in children but common in immune adults from east Africa. Inhibition of binding was much more efficient than reversal of binding, indicating a strong interaction between DBL2βC2PF11_0521 and ICAM1. Our high throughput approach will significantly accelerate studies of PfEMP1 binding domains and protective antibody responses.
| Plasmodium falciparum exports the protein PfEMP1 to the surface of parasitized erythrocytes for roles in immunoevasion and adhesion. The size and structural complexity of this diverse protein family have limited earlier studies of PfEMP1 biology to low throughput and semi-quantitative approaches. We developed a high throughput quantitative assay of PfEMP1 adhesion and used it to analyze structural features of domains that bind the putative cerebral receptor ICAM1, as well as to survey the acquisition of functional antibodies in malaria-exposed children and adults. In studies of the PfEMP1 repertoire of clone 3D7 parasites, a single specific domain bound ICAM1 strongly. PfEMP1 domains that bind ICAM1 strongly have conserved features, including conserved amino acids within otherwise highly variant flexible loops of the protein. While neutralizing antibodies against the PfEMP1–ICAM1 interaction were uncommon in Tanzanian children, such antibodies were common in east African adults, possibly explaining why immune adults rarely carry ICAM1-binding parasites. This high throughput platform will significantly accelerate studies of PfEMP1 binding domains and the corresponding antibody responses involved in protective immunity.
| The variant surface antigen Plasmodium falciparum erythrocyte membrane protein 1 (PfEMP1) is a virulence factor of the human malaria parasite P. falciparum. PfEMP1 variants are encoded by about 60 var genes per parasite, and have been implicated in the cytoadhesion of P. falciparum-infected erythrocytes (PE) to vascular endothelium [1]. PE bind numerous receptors (reviewed in [2] and [3]), including thrombospondin [4], CD36 [5], ICAM1 [6], E-selectin and VCAM-1 [7], chondroitin sulfate A (CSA) [8],[9], complement receptor 1 [10], PECAM-1 [11], heparan sulfate [12],[13], bloodgroup sugars A and B [14], and the serum proteins IgG/IgM and fibrinogen [15]. Cytoadhesion allows sequestration of PE in deep vascular beds, prevents clearance of PE in spleen, causes vascular occlusion and inflammation of different organs, and is related to cerebral malaria [16] and placental malaria [17]. PE sequestration may lead to occlusion of the microvasculature and thereby contributes to the acute pathology of severe forms of malaria [18]–[22].
Distinct domains of different var genes have been shown to bind specific ligands in vitro. For example, the CIDR1-α domain was implicated in binding CD36 [23],[24]. Different DBL1-α domains from various PfEMP1 were shown to bind the CR1 receptor on RBC [10], glycosaminoglycans on RBC, and heparan sulfate on the endothelial surface [23],[25]. The DBLβC2 domain combination binds ICAM1 [26], and DBLβ alone binds PECAM1 [23].
Specific host receptors have been implicated in specific malaria syndromes (reviewed in [27]). PE sequestration in cerebral capillaries and venules is a hallmark of cerebral malaria. ICAM1 is expressed at high levels in brains of patients with cerebral malaria, and has been implicated in this syndrome [28]. Some DBLβC2 domains of different PfEMP1 proteins bind ICAM1 [26]. In a survey of the entire repertoire of DBLβ-C2 domains (n = 25) from the IT4 line genome [29], 6 domains bound ICAM1. These studies employed a complex assay based on adhesion of ICAM1-coated beads to COS-7 cells that express PfEMP1 domains, and manual counting by microscopy. This approach is semiquantitative, time-consuming, and low throughput.
We have now developed a high throughput DBL domain-receptor binding assay and used it to study ICAM1 binding to the DBLβC2 domain repertoire of 3D7 clone parasite. Of the 16 DBLβC2 domains tested, we find that a single domain from a Group A PfEMP1 protein (PF11_0521) binds ICAM1 strongly. Our structural analyses suggest that DBL2βC2PF11_0521 binding to ICAM1 may be due to key conserved residues previously identified in IT4 line domains. Also, our truncation and binding analyses suggest that DBLβC2 domain extends about 80 amino acid residues upstream of its previously suggested boundary [26]. Binding-inhibition studies using our high throughput platform suggest that neutralizing antibodies may be infrequent in African children, but are common in immune African adults.
In our expression system, recombinant protein levels can be monitored during immobilization/purification on immobilized anti-GFP antibodies by fluorescence, as we previously described using the multi-well plate format [30]. To immobilize DBLβC2 constructs on BioRad BioPlex beads, we used a similar scheme here (Figure S1A): anti-GFP antibody was cross-linked to beads of different fluorescence intensity (i.e., different bead regions), then bead regions of distinct intensity were incubated with lysates of COS cells expressing individual domains as GFP-fusion proteins, and washed extensively for fast immobilization/purification [30]. Domain immobilization was confirmed by reactivity of beads with biotinylated anti-GFP. Signal intensity was similar for all constructs (other than mock-transfected cells, data not shown) indicating saturation of beads with recombinant domains.
In binding assays that used this bead array (Figure S1B), only the DBL2βC2PF11_0521 domain among the 16 domains tested bound to ICAM1 at high levels (Figure 1). This result was reproducible in numerous assays using different preparations of recombinant domains. We also tested the DBL2βC2PF11_0521 domain and several other non-binding domains using the traditional multi-well plate format [30], and obtained identical results (data not shown). In addition, ICAM1-Fc binding to DBLβC2 domains was detected by anti-human IgG at a nearly identical level to detection by monoclonal antibody (mAb) RR1 (data not shown), independently confirming the above results.
To confirm the specificity of ICAM1 binding to DBL2βC2PF11_0521, we tested the well-characterized mAb My13 for its ability to inhibit the interaction. According to earlier studies, My13 strongly inhibits binding of infected erythrocytes to ICAM1, but does not block binding of non-inhibitory mAb RR1 to the ICAM1 molecule [31]. Our results (Figure 2) demonstrate complete inhibition of ICAM1 binding to DBL2βC2PF11_0521 by an excess of My13, confirming the specificity of DBL2βC2PF11_0521-ICAM1 binding in our assay.
Using CLUSTALW 2.0.5 and subsequent manual curation, we aligned and analyzed sequences of DBLβC2 domains that bind and do not bind ICAM1. Figure S2 shows the alignment of DBLβC2 domains from both FCR3/IT and 3D7 strains, with four loops predicted to participate in ICAM1 contacts [32] indicated in boxes. We find that residues previously shown to be conserved in the ICAM1-binding DBL2βC2 domains of FCR3/IT are conserved in ICAM1-binding domain of 3D7 as well. The level of conservation among the residues in or directly adjacent to the ICAM1-binding structural loops is much higher in the binding versus non-binding DBLβC2 domains. This may indicate that these residues have an important role in structure or a direct interaction with the ligand.
Detailed analysis of the four ICAM1-binding loops revealed additional conserved residues. In loop 1, a Thr residue in the middle of the loop is conserved in every ICAM1-binding domain (and absent in 50% of non-binding domains). In loop 3, a 3-amino acid motif containing a hydrophilic residue-hydrophobic residue-hydrophilic residue, is completely conserved in binding domains. The hydrophobic residue in this motif is Ile with a single conservative exception (Val in var16), and appears to be in close contact with the ICAM1 molecule in the model [32]. This residue is absent in 37% of non-binding sequences. We speculated that preference in usage of Ile over Val may be explained by slightly larger surface of Ile in contact with the residues of ICAM1 (Figure S3A and S3B).
In previous analyses, an Ala or Leu residue was observed in position 3 of loop 4, in all but one ICAM1 binding domain (the exception being the FCR3/IT var1 DBL2βC2 domain where His is present) (Figure S2). In the 3D7 repertoire, 3 other non-binding domains carry Ala or Leu at this position, in addition to the ICAM1 binding DBL2βC2PF11_0521 domain. Similarly, in FCR3 strain parasites, 3 non-binding domains carry Ala or Leu at this position. However, the non-binding domains with the Ala or Leu residue in loop 4 contain multiple substitutions in other conserved regions and positions, which may explain their non-binding status.
Generally, ligand interactions involving substantial surfaces of amino acid residues are not significantly altered by substituting a single residue that participates in binding, so long as the substitution fits into the structure without clashes and does not affect structural integrity (e.g., substitutions in flexible loops) [33]. This was elegantly demonstrated by the Smith group [29], which examined amino acid substitutions and their combinations on ICAM1 binding by DBL2βC2 domains from two genes, var16 and var31 from FCR3/IT parasite. Using the model of DBL2βC2 complexed with ICAM1 [32] and the Deep View/Swiss-pdb viewer program (v.3.7), we examined the effect of replacing conserved Ala286 with Tyr in loop 4. The substitution does not introduce any amino acid residue clashes, and may provide an additional intra-domain hydrogen bond to R113 (data not shown). We infer that the loss of binding to ICAM1 is not due solely to substitutions of Ala or Leu in loop 4, but results from combinations of substitutions that involve this and other residues in the protein. We are currently testing this hypothesis using our quantitative BioPlex approach and site-specific mutagenesis. Other residues that we predict may have an effect on the domain structure or ligand binding are indicated in red in Figure S2.
Our sequence and binding analyses indicate that structural DBLβC2 domain is larger than previously suggested [26]. We propose that the domain starts about 80 aa residues upstream of the first Cys residue of A4tres DBL2βC2-ICAM1 domain. With a single exception (var6 from FCR3/IT parasite that preserves only Cys) this N-terminal region starts with a conserved Asn-Pro-Cys sequence and contains multiple conserved residues (Figure S2), independent of the type of domain located upstream of DBLβC2. With regard to downstream C2 region, our analysis demonstrates that previously described isolated DBLβ domains (e.g. DBL6β in 3D7 PFE1640, DBL5β in FCR3/IT var14, DBL3β in MC var1) are, in fact, DBLβC2 domains with degenerate C2 as well as upstream N-terminal sequences. These domains contain easily recognizable C2 features including the Y-motif and other conserved residues, as well as the upstream N-terminal fragment described above (Figure S4). A degenerate Y-motif was previously recognized in DBL6β of FCR3 VAR1CSA protein [34]. N-terminal and C2 sequences appear to diverge from the consensus sequence to similar degrees, suggesting a possible interaction between these fragments in 3-dimensional structure.
We inferred that the additional N-terminal sequence contributes to the complete DBLβC2 domains, and tested its effect on ICAM1-binding activity. We re-cloned DBL2βC2PF11_0521 domain (amino acid residues 1–522 in Figure S2) into pHisAdEx vector as well as two truncated constructs: one (named N-term) lacked 32 amino acid residues at the C-terminus (construct ends with conserved ACNC sequence plus two residues at the C-terminus), and another one (named C-term) lacked 68 amino acid residues at the N-terminus (construct starts with conserved Asn-69 at the N-terminus and includes complete ICAM1 minimal binding domain [34]) (Figure 3). The amount of all proteins immobilized on beads was similar by reactivity with anti-GFP antibody, and all proteins had His-tag at their N-termini confirmed by reactivity with anti-His antibody (data not shown). Binding of ICAM1 (Figure 3) clearly indicate that removal of the N-terminal fragment profoundly reduces ICAM1 binding activity. Since full-length and truncated variants all demonstrated similar and strong GFP fluorescence, which is a good indicator of correct folding of the entire membrane protein [35], our results suggest an important role for the N-terminal sequence in ICAM1 binding. A similar effect on adhesion was previously observed earlier in a semi-quantitative assay of the A4tres DBL2βC2 domain with an N-terminal truncation down to the first conserved Trp (var31 Trp-106 in Figure S2). This N-terminal truncation combined with C-terminal truncation up to the end of Y-motif (var31 His-449 in Figure S2) completely abolished ICAM1 binding [34].
We tested inhibition of ICAM1 binding to DBL2βC2PF11_0521 domain using pooled human plasma from immune adult males living in East Africa and from non-immune US adults (Figure 4). Pooled immune plasma from Africa blocked binding of ICAM1 to the DBL2βC2PF11_0521 domain by 78%, compared to binding in NI plasma that did not reduce binding compared to media alone. However, immune plasma was not efficient (∼15% reduced binding) in assays that measured reversal of adhesion (Figure 4), indicating a strong association between ICAM1 and the DBL2βC2PF11_0521 domain with low OFF rate. This result complements previous data with ICAM1-binding parasites that demonstrated less than 30% reversal of parasitized erythrocyte (PE) adhesion with immune sera [36].
To study the acquisition of neutralizing antibodies against the ICAM1 binding interaction, we assayed plasma samples collected from infants and toddlers participating in longitudinal birth cohort studies in Tanzania. Plasma from 7 children that were collected at several time points between 24 and 148 weeks of age, were tested for inhibition of ICAM1 binding to DBL2βC2PF11_0521 domain (Figure 5A). In parallel, we tested reactivity of IgG from the same plasma to DBL2βC2PF11_0521 domain (Figure 5B). Inhibition of ICAM1 binding activity was uncommon, and appeared to be short-lived in at least one child. The IgG reactivity curves appear almost as mirror images of the ICAM1 binding-inhibition curves (with the exception of the 148 week time point for the brown line child, discussed below), suggesting that naturally acquired anti-DBL2βC2PF11_0521 domain antibodies include at least a fraction of functional antibodies as they develop in individual children.
We tested levels of neutralizing antibody in additional plasma samples collected from children during the first 2 years of life, compared to plasma from adult males, all living in malaria endemic areas in East Africa (Figure 6). Plasma from most adults but only a few children contained neutralizing antibodies against the ICAM1—DBL2βC2PF11_0521 domain interaction, and neutralizing activity was significantly higher in plasma from adults versus children (P<0.01 when adults were compared to children 24 to 76 weeks old, and P<0.05 when compared to children 100 weeks or older, Kruskal-Wallis test). Neutralizing activity did not increase significantly in the first 2 years of life, although a trend to increasing activity was observed after 100 weeks of life. The same trend appeared in our longitudinal cohort study of 7 patients described above (Figure 5A), with a statistically significant difference in neutralizing activity between 48 and 148 weeks (P<0.05 by Kruskal Wallis test with Dunn's multiple comparison post-test).
PE bind endothelial receptors to sequester in vascular beds, and PE binding to ICAM1 has been implicated in cerebral malaria. In this study, we developed a functional BioPlex micro-bead protein array, and applied it to study ICAM1-binding P. falciparum ligands and the acquisition of neutralizing antibodies in naturally exposed individuals. Our results indicate that a single DBL2βC2PF11_0521 domain in the 3D7 genome binds at high levels to ICAM1, and the corresponding PfEMP1 protein is classified as Group A. Binding involves an N-terminal region that has not previously been recognized as an integral part of the DBLβC2 domain. While immune adults in East Africa commonly display neutralizing antibodies against this interaction, such antibodies are uncommon in infants and toddlers in the same region.
Immunological profiling of sera for reactivity against different antigens is a common method for assessing acquired immunity and identifying potential vaccine candidates [37]. However, relating immune responses to malaria resistance is not straightforward since exposed individuals are typically infected repeatedly throughout life, and develop diversified immune responses against multiple antigens, in many cases without comprehensible relevance to disease severity. Multiple studies have sought to relate seroreactivity with disease susceptibility in young African children [38]–[44], but no candidate antigens for a severe malaria vaccine have been identified. While seroreactivity studies are useful for defining the immunoepidemiology of existing vaccine candidate antigens [30],[45], functional assays may be essential for the discovery of novel vaccine candidates. Functional antibody responses are likely to be less diverse, to target fewer antigens, and to have a stronger association to protection from severe forms of malaria.
In this paper, we describe a high throughput approach to measure the presence and relative amount of functional antibodies in patient sera. An earlier approach, though elegant, is semi-quantitative and does not allow for high throughput studies [29]. The earlier approach was based on expression of recombinant PfEMP1 domains on the surface of mammalian cells; incubation of these mammalian cells with small resin beads chemically cross-linked to the host cell receptors; removal of unbound beads from mammalian cells attached to microscope glass by inversion and gravity sedimentation of unbound beads; and manual counting of the beads bound to the surface of mammalian cells. The approach exploited in our work is based on expression of functional antigens in mammalian cells, and rapid antigen immobilization in a directed manner on the surface of BioPlex fluorescence-coded beads. This approach allows multiplexed analyses of protein features including receptor binding activity (Figure S1) as well as seroreactivity studies in a high throughput manner.
Our studies focused on the construction of a 3D7 genome-wide array of the DBLβC2 domain, which was previously shown to bind ICAM1 in studies of other parasite lines [26],[28]. Analysis of ICAM1-binding activity in this array revealed that only the DBL2βC2PF11_0521 variant out of 17 domain variants, binds the receptor at high levels. Alignment of 3D7 ICAM1 binding and non-binding domains with previously identified ICAM1-binding domains from other parasite strains revealed new structural features related to the ICAM1 interaction (Figure S2), in particular, a conserved Thr residue in loop 1 and a conserved 3-amino acid motif in loop 3. This analysis further highlighted the potential role of conserved residues within predominantly non-conserved flexible loops in adhesion, and, therefore, as targets for intervention.
All DBLβC2 domains that we tested share highly conserved structural features, like helices and loops, and therefore, their general architecture should be similar. Constructs of these domains used the same boundaries and yielded recombinant protein at similar levels (according to GFP fluorescence) in a system that is well-suited for folding of transmembrane disulfide-rich proteins. Therefore, all recombinant domains have a high probability of folding similarly well. Because one of DBLβC2 domain variants clearly demonstrated binding activity, we assume that other variants that did not bind were properly folded but do not function as ICAM1 ligands. Nevertheless, since every protein is unique, false negatives can not be excluded completely without direct proof of correct folding by methods like X-ray or NMR, which are outside the scope of this study. Our findings also suggest that the structural/functional DBLβC2 domain involved in ICAM1 binding includes about 80 amino acid residues upstream of the previously suggested DBLβC2 domain [26]. This N-terminal sequence contains an alpha-helix (shown in Figure S4) predicted by several algorithms [46] in each DBLβC2 domain described here. Two other short segments associated with conserved and semi-conserved residues downstream of the alpha-helix were variously predicted to be alpha-helical or extended strands (not shown).
This high throughput assay platform can be used to profile functional antibody levels among naturally exposed children and adults. We find that antibodies that inhibit ICAM1 binding to DBL2βC2PF11_0521 appear sporadically in the first 2 years of life (Figure 5 and 6). Conversely, many immune adults have these antibodies in their sera. Neutralizing activity in adult plasma did not correlate with age in this group of adults (18 to 54 years old), consistent with the solid and stable protective immunity enjoyed by all adults in these communities. The slow acquisition of functional antibody may reflect that this domain variant is rare in the community, that the immature immune system of young children responds poorly to some PfEMP1, or that some other host-parasite interaction thwarts the development of functional immunity. In another study in a malaria endemic area [47], serum anti-rosetting activity against a particular lab strain (FCR3) appeared in only about 10% of children 2–5 years old, but in up to 60% of 15–16 year old adolescents, demonstrating a similar slow accumulation of functional responses. We do not know at present whether the functional response (inhibition of ICAM1 binding) is variant-specific. Future studies will clarify this question.
With regard to longevity of the immune response, an earlier study [48] found that anti-PfEMP1-like responses are short lived and variant-specific, at least in a low malaria endemicity area. We observed functional antibody to ICAM1 binding was short-lived in one child (green line in Figure 5). We are preparing to test whether this phenomenon is common using a larger set of children's plasma collected in longitudinal cohort studies. We will also examine other features of the natural immune response to malaria, such as the apparent discordance of seroreactivity and functional activity observed in some children (brown line in Figure 5). This may indicate that the amount of non-functional antibodies may increase without increasing the amount of functional antibodies, or that non-functional antibodies that increase with time may successfully compete with functional antibodies and block their activity. However, this dataset is limited and it would be premature to make definitive conclusions at present.
Our high throughput approach will now allow us to test numerous additional ICAM1 binding domains, and to determine which of these is targeted by neutralizing antibodies that also block parasite binding. With an expanded dataset, we can correlate functional antibody responses with clinical outcomes in these vulnerable populations. These future studies will also examine the concordance between this assay and the traditional binding-inhibition studies using parasitized red blood cells, including the ability to detect variant-specific versus broadly reactive functional antibodies.
Human plasma samples used in these studies were collected from East African donors under protocols approved by relevant ethical review committees. Study participants provided written informed consent before donating samples. Ethical clearance was obtained from Institutional Review Boards of SBRI and the National Medical Research Coordinating Committee in Tanzania.
All constructs were cloned into the pAdEx vector described earlier [30]. Expression of constructs in COS-7 cells and lysate preparation were also described in [30]. All expressed constructs are GFP-fusion proteins that contain an extracellular DBLβC2 domain, a short trans-membrane region, and a cytoplasmic domain fused to green fluorescent protein (GFP). In addition, full-length and truncated forms of DBL2βC2PF11_0521 domain were cloned into modified vector pHisAdEx. This vector was constructed as follows: pAdEx plasmid was digested with SfiI and BamHI restriction enzymes, then the large fragment was isolated by agarose gel electrophoresis and ligated with double-stranded oligonucleotide adaptor prepared by annealing of two oligonucleotides 5′-GAT CCC TGC GTG GTG GTG GTG GTG GTG CT-3′ and 5′-ACC ACC ACC ACC ACC ACG CAG G-3′. The resulting construct was verified by sequencing. Proteins expressed from this vector are similar to proteins expressed from pAdEx vector but contain His6-tag at their N-termini. Various PfEMP1 domains that supported binding of ICAM1 (see above) and CD36 (data not shown) in the pAdEx expression system also supported binding in the pHisAdEx expression system. Primers used for cDNA amplification using 3D7 genomic DNA are shown in supplementary Table S1. Alignment of all DBLβC2 domains is shown in supplementary Figure S2. Five DBLβC2 domains, that were shorter at their N-termini than other 11 domains after amplification with primers indicated in Table S1, were also obtained within the same boundaries as for other domains and cloned into pAdEx and pHisAdEx for expression. New forward primers for their PCR amplifications are shown in Table S2.
25 µg of anti-GFP antibody (Rockland, Gilbertsville, PA) was coupled to 200 µl of each different BioPlex bead region (Bio-Rad) (19 regions total) as described by the manufacturer, then resuspended in PBS containing 1 mg/ml BSA, 0.05% Tween-20, and 0.02% sodium azide (PBS-TBN buffer). Anti-GFP-coupled beads were incubated for 2 hours at 4°C with COS-cell lysates containing expressed domains, washed in PBS-TBN, and used in ligand binding experiments. These beads are designated as DBLβC2-coupled beads.
All DBLβC2-coupled beads were mixed together in quantities of ∼40–60 beads of each bead region (beads with distinct fluorescence intensity) per µl. 50 µl of bead mixtures were transferred into individual wells of HTS 96-well plates (Whatmann) that were pre-incubated with PBS-TBN for 30 minutes. Beads were washed in wells 3 times with PBS-TBN and incubated with different concentrations (20 – 0.1 µM) of ICAM1-human Fc receptor (R&D Systems, Minneapolis, MN). After 1 hour incubation at room temperature (RT) at constant rotation at 600 rpm, beads were washed in PBS-TBN and incubated in similar fashion with 1∶10 diluted biotinylated anti-ICAM1 monoclonal antibody (mAb) RR-1 (Axxora, San Diego, CA) followed by 1 hour incubation with 1∶250 diluted streptavidin-phycoerythrin (SA-PE) fluorescent molecules (Jackson ImmunoResearch, West Grove, PA). Also, in some experiments (binding only, not binding-inhibition by human serum) we used anti-human IgG coupled to phycoerythrin (1∶250 dilution, Jackson ImmunoResearch) to detect bound ICAM1-human Fc and obtained almost identical results. After a final wash, 96-well plates were transferred into the BioPlex apparatus (Bio-Rad) to quantify ICAM1 binding (measured in phycoerythrin channel) to the individual DBLβC2 domains. For negative controls, lysates prepared from mock-transfected cells, and from pAdEx vector transfected cells, were also coupled to beads. These control beads were mixed with the DBLβC2-coupled beads and assayed simultaneously. The pAdEx vector produces GFP-fusion proteins that contain an irrelevant peptide of 37 amino acids in the extracellular domain.
To confirm the specificity of ICAM1 binding to DBL2C2PF11_0521, ICAM1-Fc (1 µg/ml) was incubated with various concentrations of mAb My13 (Axxora) for 1 hour at room temperature and then used to bind to the mixture of His-DBL2C2PF11_0521 and HisAdEx (negative control) coupled beads as described above using RR1 mAb for detection.
For binding inhibition assays DBLβC2-coupled beads were pre-incubated for 1 hour at RT with various plasma samples diluted (1∶5) in PBS-TBN. The beads were then assayed in the same fashion as for the binding assay described above using RR1 mAb for detection. For binding-reversal assays, the binding assay was performed as described above, except that the beads were incubated with various plasma samples diluted (1∶5) in PBS-TBN for 1 hour at RT after the reaction with SA-PE, and then washed just prior to quantification in the BioPlex apparatus.
Human plasma samples used in these studies were collected from East African donors under protocols approved by relevant ethical review committees. Study participants provided written informed consent before donating samples, and included adult males from Kenya [49],[50] and children of different ages from Tanzania [51]. Malaria is endemic in both these regions. Plasma from 5 randomly selected non-immune donors in the US were separated from whole blood obtained from commercial sources (Valley Biomedical) and used in a pool as a negative control.
DBL2βC2PF11_0521 domain-coupled beads were washed in wells 3 times with PBS-TBN and incubated with children plasma samples at 1∶100 dilution in PBS-TBN. After 1 hour incubation at room temperature at constant rotation at 600 rpm, beads were washed in PBS-TBN and incubated in similar fashion with 1∶250 diluted anti-human IgG coupled to phycoerythrin (Jackson ImmunoResearch) for 1 hour. Signal was measured in the BioPlex apparatus to quantify bound IgG. Signals obtained for beads coated with protein expressed by pAdEx vector were used as negative controls and were subtracted from signal obtained with DBL2βC2PF11_0521 domain-coupled beads. Also, pooled samples of non-immune US plasma (n = 5) and immune plasma from adults living in malaria endemic region (n = 5) were used as additional negative and positive controls, respectively.
|
10.1371/journal.pgen.1007691 | Amn1 governs post-mitotic cell separation in Saccharomyces cerevisiae | Post-mitotic cell separation is one of the most prominent events in the life cycle of eukaryotic cells, but the molecular underpinning of this fundamental biological process is far from being concluded and fully characterized. We use budding yeast Saccharomyces cerevisiae as a model and demonstrate AMN1 as a major gene underlying post-mitotic cell separation in a natural yeast strain, YL1C. Specifically, we define a novel 11-residue domain by which Amn1 binds to Ace2. Moreover, we demonstrate that Amn1 induces proteolysis of Ace2 through the ubiquitin proteasome system and in turn, down-regulates Ace2’s downstream target genes involved in hydrolysis of the primary septum, thus leading to inhibition of cell separation and clumping of haploid yeast cells. Using ChIP assays and site-specific mutation experiments, we show that Ste12 and the a1-α12 heterodimer are two direct regulators of AMN1. Specifically, a1-α2, a diploid-specific heterodimer, prevents Ste12 from inactivating AMN1 through binding to its promoter. This demonstrates how the Amn1-governed cell separation is highly cell type dependent. Finally, we show that AMN1368D from YL1C is a dominant allele in most strains of S. cerevisiae and evolutionarily conserved in both genic structure and phenotypic effect in two closely related yeast species, K. lactis and C. glabrata.
| Separation of mother and daughter cells after mitosis in eukaryotes enacts various functional and/or developmental needs and has significant medical and industrial implications. How this cellular behaviour is regulated is far from being concluded. We report here a novel Amn1 mediated post-mitotic cell separation in a budding yeast strain, YL1C and demonstrate that the post-mitotic cell separation can be regulated through a ubiquitin-conjugated protein degradation of Ace2 by Amn1. The Amn1-governed switch of cell separation is evolutionarily conserved and highly cell type dependent. These findings provide insights into the molecular mechanism of how post-mitotic cell separation is regulated in budding yeast, and data for translating into medical and industrial applications.
| The switch between effective and inhibited separation of mother and daughter cells in eukaryotic mitosis represents a fundamental process for understanding the evolution of organizational and functional complexity of organisms, and also has significant medical and industrial value [1–3]. It has been well documented that division of the cytoplasm in Saccharomyces cerevisiae is comprised of a series of coordinated events including assembly and contraction of the contractile actomyosin ring in mitosis, formation of the primary and secondary septa and finally separation of mother and daughter cells [1]. The molecular machinery and regulatory networks that underlie this process has been significantly advanced in recent studies in the simple eukaryotic model yeast S. cerevisae. In particular, the RAM (regulation of Ace2 and morphogenesis) network, a pathway regulated by the MEN (Mitotic Exit Network), has been proposed to be responsible for nuclear importation of the transcription factor Ace2, which is needed for septum cleavage and post-mitotic cell separation [4–7]. Nuclear importation of Ace2 drives a sharp increase in the transcription of genes involved in septum cleavage, including CTS1 and SCW11, and ultimately leads to separation of mother and daughter cells [4, 5]. So far, six proteins have been identified as key nodes in the RAM network: Sog2, Tao3, Hym1, Kic1, Mob2 and Cbk1. Cells lacking Ace2 or any of these six core components show a phenotypic defect of indistinguishable post-mitotic cell separation and cell clumping [6, 7]. Failure of daughter cells to separate from their mothers after mitosis, showing a snowflake phenotype under the micropscope, is recognized as the early evolution from unicellular to multicellular populations and the transition can evolve quickly over the course of multiple rounds of selection for the snowflake phenotype [8–10].
In our earlier work, we dissected phenotypic variation in cell clumping in a segregating population created by crossing two phenotypically divergent strains (YL1C with a strong clumpy phenotype and YH1A with effective cell separation), into four major cell clumping Quantitative Trait Loci (QTLs) [11]. These major QTLs together explained 45% of the trait phenotypic variation. We resolved the major QTL explaining 25% of the clumping phenotypic variation into the QTL gene AMN1. We further identified the V368D substitution in Amn1 as the causative variation of the QTL gene through site specific mutation and allele replacement experiments [11]. Amn1 was previously found to be required to turn off the mitotic exit network (MEN), a pathway that promotes spindle breakdown, degradation of mitotic cyclins, cytokinesis, and post-mitotic cell separation, through obstructing the binding of Tem1 to Cdc15 [12, 13].
This paper presents a novel mechanism of Amn1-mediated cell separation inhibition after mitosis in YL1C. We show here for the first time that Amn1 can post-translationally control the degradation of Ace2 through the ubiquitin proteasome system (UPS). This establishes that Amn1 modulates post-mitotic cell separation through down-regulating Ace2 and its downstream genes. The data not only advances Amn1 as an antagonist of MEN [12], but also provides a new insight into how the RAM mediates post-mitotic cell separation [4–7]. Moreoever, we demonstrate that the Amn1-governed post-mitotic cell separation is cell-type dependent. The clumping phenotype governed by Amn1 is highly dependent on the ploidy level in natural S. cerevisiae cells, while the functional of Amn1368D from YL1C in controlling post-mitotic cell separation, is evolutionarily conserved in both genic structure and phenotypic effect.
Firstly, the clumping cells of the S. cerevisiae strain YL1C became separated when AMN1 was deleted (Fig 1A) as we previously observed [11]. To explore the underlying mechanism by which AMN1 causes cell clumping, we conducted an RNA-seq assay and identified 43 significantly differentially expressed genes between YL1C cells showing a strong clumpy phenotype and YL1C with AMN1 deleted (Fig 1B). Of these 43 genes, 18 were up-regulated when AMN1 was deleted, including DSE1, DSE2, DSE3, DSE4, EGT2, SCW11 and CTS1 with known roles in post-mitotic cell separation, acting directly to degrade the primary septum at the bud neck [14–16]. From these 7 known genes, we chose the 4 most up-regulated genes, DSE1, DSE2, SCW11 and CTS1, and confirmed the results of the RNA-seq assay by using RT-qPCR (S1 Fig). We then stained YL1C cells with calcofluor white (CFW), a fluorescent dye specifically staining chitin, the major component of the septum, as previously suggested [17], and confirmed that the YL1C cells remained attached with the undegraded primary septum at the bud neck. In contrast, when AMN1 was deleted, the bud scars were deeply stained by CFW, indicating complete mother-daughter cell separation (Fig 1C). These results indicate that AMN1 inhibits cell separation after mitosis and induces cell clumping as seen in the YL1C strain.
Yeast cell clumps can also be formed through an aggregate cellular behavior genetically controlled by the FLO family of genes that regulate interactions between cell wall glycoproteins [18–20]. We invesigated the influence of FLO genes on the clumping of YL1C cells. FLO1 and FLO8 deletion showed a comparable clumping phenotype to that of YL1C (S2 Fig), suggesting that clumping phenotype of YL1C cells was largely attributable rather to defective post-mitotic cell separation governed by Amn1 than to the cell wall glycoproteins encoded by the FLO gene family.
Ace2 is the major transcription factor of DSE1, DSE2, DSE3, DSE4, EGT2, SCW11 and CTS1, its mutation or deletion may sharply decrease the RNA levels of these target genes [4–6, 14–16]. Based on our observation that deletion of AMN1 occurred in parallel with down-regulation of these Ace2 target genes (Fig 1B), and the fact that deletion of ACE2 restored cell clumping phenotype in Δamn1 mutant cells (Fig 1A), we hypothesized that AMN1 inhibited post-mitotic cell separation in the YL1C strain through inactivating Ace2. To test the hypothesis, we firstly profiled the RNA level of ACE2, and did not find any significant change in the RNA level (Fig 2A upper), but the protein level of the gene measured by the western blotting assay was substantially up-regulated in the YL1C strain with AMN1 deleted when compared to that in the YL1C strain (Fig 2A lower). The protein level of Ace2 was also found to vary over the cell cycle and to be negatively correlated with the protein level of Amn1 in the YL1C strain in a cell synchronization analysis by using nocodazole. However, the protein level of Ace2 did not show a marked change across the cell cycle when the cells carried an Amn1368V variant (Fig 2B and 2C).
We induced expression of AMN1 under the PGAL10 drive in YL1C using galactose, and profiled the protein levels of Ace2 and Amn1. The results showed that the protein level of Ace2 decreased as the protein level of Amn1 increased. At the 6-hour time point, Amn1 protein expression reached its highest level and the Ace2 protein was not detectable (Fig 2D). However, the RNA expression of ACE2 did not show any change, while the RNA level of AMN1 progressively increased (S3 Fig). Moreover, when the 368D was replaced by the Val residue, the protein level of Ace2 was no longer dependent on the Amn1368V level (Fig 2D). The negative correlation in protein level between Ace2 and of Amn1 in both cell synchronization assay and galactose pulse-chase assay strongly support the down-regulation of Ace2 by Amn1.
We also tested the nuclear accumulation of Ace2, which was necessary to induce transcription of its target genes [4, 5]. We marked Ace2 with GFP and found that Ace2 was absent in daughter cell nuclei of synchronized YL1C cells. In contrast, when AMN1 was deleted, Ace2 efficiently accumulated in daughter cell nuclei (Fig 2E). Using site-directed mutagenesis, we constructed two YL1C strains with a continuously activated Ace2, carrying either of two sets of multiple substitutions, either S122D, S137D, T575A, S701A and S714A (referred to as Ace2-AAA-2D), or F127V, T575A, S701A and S714A (referred to as Ace2-AAA-F127V). Both were previously reported to locate in the nucleus in all cells regardless of cell cycle position and to continuously transcribe Ace2’s target genes [5, 21]. Nevertheless, post-mitotic cell separation was inhibited in both genetically modified strains, while the RNA levels of Ace2 target genes did not significantly change in the two strains compared to YL1C (S4A Fig). Ace2 protein levels remained extremely low in the wild-type YL1C cells and its engineered strains, but were boosted when AMN1 was deleted (S4B Fig). The data supports that Amn1 mediated degradation of Ace2 overrides the well-established view that regulation of Ace2 function is through phosphoration and protein localization during the process of cell separation after mitosis [4,5].
To further explore downregulation of Ace2 by Amn1, we firstly performed ribosome profiling to test whether Amn1 affects the translational efficiency of Ace2. It shows no marked change in the level of ACE2 mRNAs occupied by the ribosomes in YL1C compared with the corresponding Δamn1 strain (S5 Fig). The downregulation of Ace2 by Amn1 is therefore unlikely due to altered translational efficiency.
To examine the apparent impact of Amn1 on the stability of Ace2, we performed a GAL promoter shut-off chase experiment in which PGAL10-ACE2 was transformed into YL1C so as to drive ACE2 expression. The engineered cells were initially grown in galactose medium to induce ACE2 expression, and then transferred to glucose medium to switch off ACE2 transcription. The protein level of Ace2 was then measured every 15 minutes using western blotting. Ace2 was extremely unstable and vulnerable to degradation in the YL1C strain (Fig 3A left panel), while no marked decrease in the protein level was observed in the AMN1 deleted strain (Fig 3A right panel). When the proteolytic activity of the 26S proteasome was blocked using MG132, then Ace2 levels no longer markedly decreased (Fig 3B). Additionally, YL1C cells treated with MG132 had a stable endogenous Ace2 protein level (Fig 3C), while the endogenous Ace2 protein was markedly boosted when the ubiquitin coding gene UBI4 was deleted (Fig 3D). These results indicated that the down-regulation of Ace2 by Amn1 can be explained by the ubiquitin-conjugated protein degradation machinery.
Amn1 was previously predicted to be a member of the F-box protein family in S. cerevisiae, and proposed to be a potential ubiquitin ligase E3 in a bioinformatic analysis [22]. However, the putative F-box in Amn1 is atypical since the motif is separated by a 56-amino acid insertion. However, there is no experimental evidence so far for whether Amn1 catalyzes ubiquitination of Ace2 directly and facilitates its degradation. To tackle this open question, we firstly constructed two genetic modified Amn1 proteins without a normal function as an E3 ligase, one with deletion of only the F-box (referred to as AMN1-Δ(496–552)&(721–789)) and the other with deletion of both the F-box and the 56 amino acid insertion (referred to as AMN1-Δ(496–789)). YL1C cells with either modified Amn1 protein showed effective cell separation and substantial up-regulation of the Ace2 protein in vivo (Fig 3E). This suggested that the atypical F-box in Amn1 was functional and controlled the stability of Ace2.
Furthermore, we carried out co-immunoprecipitation assays and showed that Amn1 physically binds to Skp1 and Cdc53, two key components of SCF complex (Fig 3F left panel)[23]. We also performed a size-exclusion chromatography experiment, and found that the three proteins, Amn1-Flag, Skp1-Myc and Cdc53-Ha, could be co-eluted at #16-#21, corresponding to the apparent size of ~440kDa. Additionally, we found that the three proteins overlapped again at fraction #27-#29, corresponding to the apparent size of ~ 158kDa. We thus speculated that the fraction #27-#29 contained the complexes of Amn1-Cdc53, Cdc53- Skp1, and Amn1-Skp1 (Fig 3F right panel). We then examined the stability of Ace2 in cdc53 and skp1 mutants using the GAL promoter shut off assay. And found Ace2 was greatly stabilized by mutations in both CDC53 and SKP1 (Fig 3G). Moreover, we enriched the Ace2 protein complex using immunoprecipitation and profiled Amn1 and Ace2. Amn1 in YL1C was clearly seen to physically interact with Ace2 (Fig 3H). However, when the 368D in Amn1 was substituted by Val, the Amn1368V level was detected but its physical interaction with Ace2 was substantially weakened. A higher molecular ladder of ubiquitinated Ace2 was also detected in YL1C (Fig 3H and 3I), while ubiquitinated Ace2 was almost completely eliminated (Fig 3H and 3I) when AMN1 was deleted or substituted with AMN1368V. An in vitro protein ubiquitination assay also showed that Ace2 could be labelled with exogenous human ubiquitin in the presence of Amn1368D, but not in the absence of Amn1, and the ubiquitination was substantially weakened when Amn1368D was replaced by Amn1368V (Fig 3J). These observations support Amn1’s role in mediating Ace2 proteolysis through the UPS, and the essentiality of the 368D residue to enable Amn1 to adequately bind and hence to ubiquitinate Ace2.
ACE2 has a paralog SWI5 which arose from whole genome duplication [24]. Swi5 has an almost identical DNA-binding domain to Ace2, but regulates a different set of genes in vivo [16]. However, we did not see any significant change in RNA or protein levels of SWI5 in YL1C with or without AMN1, clearly indicating that Swi5, unlike Ace2, is not regulated by AMN1 (Fig 4A). We aligned protein sequences of Ace2 and Swi5, compared their functional domains and identified 3 aligned regions (A, B and C), with region A subdivided into A1, A2, A3 and A4 regions. We then constructed a series of Ace2-Swi5 chimeras, and identified a novel 11-residue domain ‘ELRDLDIPLVP’ as the site for Amn1 to bind to and directly degrade Ace2 (see Supplementary Materials and Methods, S6 Fig). We replaced this domain with the corresponding Swi5 derived ‘EINDLNLPLGP’ in YL1C, creating a modified Ace2* (Fig 4B). YL1C cells carrying the Ace2* separated effectively and a western blotting assay indicates that the modified Ace2* was expressed at markedly higher levels than the wild-typed Ace2 in vivo (Fig 4C and 4D). In addition, the RNA levels of four Ace2 target genes were clearly up-regulated (Fig 4D). An Ace2 protein with the 11 residue domain mutated may therefore escape from the negative control by Amn1. Moreover, we did not observe any detectable signal of interaction of Ace2* with Amn1, nor its ubiquitination, further supporting the role of the identified 11-residue domain in the post-translational regulation of Ace2 (Fig 4E). We then constructed a series of strains bearing single amino acid substitutions (Ace2R71N, Ace2D74N and Ace2V78G), but none of these led to effective cell separation, suggesting that the whole 11-residue domain be essential for its phenotypic effect (S6C Fig).
The above analyses were established in haploid cells. We found that inhibition of post-mitotic cell separation was released and the cells separated effectively in diploids (Fig 5A). Additionally, overexpression of AMN1 led to strong inhibition of post-mitotic cell separation in the diploid cells (Fig 5A). We then checked the protein level of Amn1 and Ace2 in diploid cells using western blotting and found Amn1 protein level was much lower, agreeing with its RNA level (Fig 5A and 5B). Conversely, the protein level of Ace2 was markedly boosted in diploids compared to haploid cells (Fig 5B). Overexpression of AMN1 in the diploid cells depressed the Ace2 protein level, as expected (Fig 5B). Using RNA-seq, we surveyed the RNA levels of the 18 genes that were up-regulated when AMN1 was deleted in haploid cells. Of the 18, 15 showed an inflated expression in MATa/α diploid cells, but their RNA levels were repressed when the cells had AMN1 over-expressed (Fig 5C). These results were confirmed by RT-qPCR assays (Fig 5A). The data supports Amn1 as the key inhibitor of post-mitotic cell separation in diploid as well as in haploid cells, but it must be noted that its endogenous expression is not activated in diploid cells in nature as to be explained below.
Diploid S. cerevisiae cells differ from the corresponding haploids in two ways. First, haploid cells and diploid cells created from merging of the haploid cells may be identical at every gene in the genome except at the MAT locus. Second, they differ in gene dosage [25]. To compare the diploids and haploids in exactly the same genetic background, we created diploids with a homozygous genotype MATa/a or MATα/α at the MAT locus. We observed that the cell clumping phenotype as well as the expression patterns of AMN1 and its downstream regulated genes in these diploids were comparable to those in the haploids (Fig 5A). These observations effectively excluded the possibility that differentiation in the post-mitotic cell separation between natural haploid and diploid cells was due to the difference in gene dosage.
We next explored the difference in post-mitotic cell separation behavior between diploid cells with either homozygous or heterozygous genotypes at the MAT locus. First, we noted that the MATa/α diploid cells of S. cerevisiae encoded a heterodimer, a1-α2, which played a role as a transcriptional repressor of haploid-specific genes [26] and might contribute to ploidy specific phenotypes observed between natural haploid and diploid yeasts [26–28]. To establish association of the a1-α2 heterodimer with the Amn1-regulated post mitotic cell separation, we notified that the binding sites of the heterodimer in the upstream promoter region of AMN1 predicted from publicly available large scale ChIP data [26,29]. The two predicted binding sites were also shared by another regulator, Ste12 (Fig 5D). At the two a1-α2 binding sites, we constructed single-site and double-site deletion mutants of the MATa/α diploid cells, and observed that, the MATa/α diploid cells in which both binding sites were deleted displayed inhibited mother-daughter separation after mitosis, similar to that observed in the corresponding a/a or α/α diploid cells (Fig 5A). At the same time, the expression of AMN1 was markedly boosted, while the expression of Ace2’s downstream genes, DSE1, DSE2, SCW11 and CTS1, was reduced (Fig 5A). We thus hypothesized that the post-mitotic cell separation of the natural MATa/α diploid cells was caused by the a1-α2 heterodimer blocking the binding of Ste12 to the AMN1 promoter.
We carried out a ChIP assay to test if Ste12 binds to the AMN1 promoter in YL1C haploids and diploids. The assay showed that the fold change of enrichment was highly significant in the haploid cells, with the enrichment peak observed in a region from -600 to -700 bp spanning the upstream region of AMN1, while no enrichment was observed in the diploid cells (Fig 5D). Consistently, the RNA level of AMN1 was reduced when STE12 was deleted in haploid cells, and in turn, the depressed expression of Amn1 has led to a remarkable boost in expression of DSE1, DSE2, SCW11 and CTS1 (Fig 5E). These results show that Ste12 and the a1-α2 heterodimer share common binding sites in the promoter of AMN1. The a1-α2 heterodimer down-regulates Amn1 and make the diploid cells effectively separate after mitosis.
Although Ace2 was a well-documented transcription factor of AMN1 in S288C or W303 yeasts, Ace2 was absent in haploid YL1C [16]. Thus, the question rises how AMN1 is regulated in YL1C? We showed that the RNA level of AMN1 in haploid cells with ACE2 deleted was not reduced to the background level (i.e. the level when AMN1’s promoter was completely removed) until STE12 was additionally deleted. In contrast, a single deletion of ACE2 was sufficient to reduce the RNA level of AMN1 to the background in the diploid cells (Fig 6A). All these together show that transcription regulation of AMN1 in YL1C cells differs between the two types (haploid and diploid) of cells.
In summary, the heterodimer is absent in haploids and Ste12 activates AMN1. Expression of AMN1 silences the protein expression of Ace2 through its N-terminal 11-residue domain as illustrated in Fig 4 and thus inhibits expression of Ace2 target genes. Down-regulation of the genes involved in septum cleavage consequently leads to post-mitotic cell separation inhibition, causing cells to be clumped in haploids. On the other hand, in MATa/α diploid strains, the presence of the a1-α2 heterodimer prevented Ste12 from efficiently binding to the AMN1 promoter, which in turn, inactivates the transcription of AMN1. This leads to effective post-mitotic cell separation because Ace2 is then released from negative regulation by Amn1. These observations explain the cell type dependency of the Amn1 regulated cell separation after mitosis (Fig 6B).
All the conclusions above were drawn from experiments with natural, cell clumped strain YL1C, which carrying an AMN1368D. To explore the conservation of Amn1368D among yeast strains of Saccharomyces cerevisiae, we collected and aligned the Amn1 protein sequences for all 46 sequenced yeast strains from the SGD database (www.yeastgenome.org). Of the 46 strains, 34 shared the 368D in Amn1, while the remaining 12, most of which have a common S288C background, (e.g. S288C, W303, BY4741, BY4742, JK9-3d and X2180-1A), carry the 368V (S2 Table). These data prompt us that AMN1368D is a dominant allele, while AMN1368V probably sourced from the human-selection for a lab-used strain which was less clumpy and thereby greatly easily handled.
Amn1 was reported as an inhibitor of the mitotic exit network (MEN), through binding to Tem1, and in doing so, inhibiting Tem1 binding to Cdc15 [12,13]. However, these findings were established from the haploid strain W303 carrying Amn1368V. To investigate whether MEN would also be inhibited in the YL1C haploid strain carrying Amn1368D, we performed a co-immunoprecipitation assay and observed effective binding of Amn1368D to Tem1 (Fig 7A). When AMN1 was deleted in YL1C, the interaction between Tem1 and Cdc15 was strengthened; however, once either AMN1368D or AMN1368V was overexpressed in the YL1C Δamn1 cells, the interaction between Cdc15 and Tem1 was markedly weakened (Fig 7B). These observations show that Amn1368D has the same function as Amn1368V in suppressing the interaction of Tem1 with Cdc15, suggesting that both proteins can inhibit MEN. We also compared Amn1368D and Amn1368V for their role in regulating post-mitotic cell separation in both S288C and W303 strains. The comparison shows that overexpressed AMN1368D may still result in post-mitotic cell separation inhibition, whereas overexpression of AMN1368V does not lead to the same phenotype in these two strains (Fig 7C).
AMN1 was predicted as an orthologous gene among at least 10 yeast species from the orthologous groups database (OrthoDB, http://orthodb.org/orthodb7), including S. cerevisiae, C. glabrata and K. lactis [30]. Notably, the conserved regions of the Amn1 orthologous proteins cover the functional 368D residue (S7A Fig). We tested functional conservation of the orthologous proteins in two yeast species related to S. cerevisiae, diverging ~100 million years ago [31]. First, C. glabrata, a highly opportunistic pathogen is usually found in the urogenital tract or bloodstream and especially prevalent in HIV positive or elderly populations [32]. Second, K. lactis, a Kluyveromyces yeast is commonly used in basic research and industry [33]. We aligned Amn1Kl (from K. lactis) and Amn1Cg (from C. glabrata) with Amn1Sc (from S. cerevisiae) using clustal W and found 24.8% and 37.6% amino acid sequence identity with Amn1Sc respectively (S7D Fig). When we replaced the coding region of the AMN1 allele in situ in YL1C with the coding region of either AMN1Cg or AMN1Kl, the clumping phenotype was restored in Δamn1 mutant cells (S7B Fig). The RNA levels of the four major downstream genes responsible for septum degradation, DSE1, DSE2, SCW11 and CTS1, were consistently repressed in the two gene replacement strains (S7C Fig). These results indicate that AMN1368D, but not AMN1368V is more likely the conserved allele.
The evolution of multicellularity from unicellularity in eukaryotes has attracted wide interest in both basic research and its translational value in medicine and industry. However, the underlying molecular mechanism that governs this important transition in cellular behavior is far from well established. In model organism budding yeast, the defective post-mitotic mother and daughter cell separation has been recognized as a key step in the evolution of multicellularity in the recent literature. In these works, clumped yeast cells from inhibited mother-daughter cell separation during mitosis were attributed to various mutants of ace2, particularly a truncated ace2 mutant [8–10], or the inhibition of Ace2’s nuclear importation in budding yeast [4, 5]. Here, we report a novel Amn1 governed post-mitotic cell separation and demonstrate that Amn1 induces proteolysis of Ace2 through the ubiquitin proteasome system, and in turn, down-regulates Ace2’s downstream target genes involved in hydrolysis of the primary septum. This leads to inhibition of cell separation and clumping of haploid yeast cells.
Moreover, the present study reveals that separation of mother and daughter cells after mitosis regulated by Amn1 is also highly dependent on cell type in the budding yeast S. cerevisiae, with effective separation after mitosis observed when haploid cells merge with other haploid cells of an opposite mating type to form diploids. In light of the Amn1 governed post-mitotic cell separation and our observation that AMN1 has an unusually lengthy upstream region, we demonstrate that Ste12 and the a1-α2 heterodimer share the same binding sites in the promoter of AMN1, which leads to the transcription regulation of AMN1 in YL1C cells differs between the two types (haploid and diploid) of cells. The present study clarified that divergence in the regulation of post-mitotic cell separation between haploids and diploids is not attributable to gene dosage, but actually to the presence or absence of the a1-α2 heterodimer. Furthermore, up to 65 other regulators that conditionally activate AMN1 were proposed to target the upstream region of AMN1, which may change the transcription regulation of AMN1 [34]. Among them, Phd1 in diploid yeast cells may bind to the AMN1 promoter when the cells are starved of nitrogen [35]. These observations indicate that AMN1 could be activated conditionally by common regulators, leading to inhibition of post-mitotic cell separation.
The functionality of AMN1 in regulating post-mitotic cell separation is highly conserved among S. cerevisiae and its two closely related yeast species, K. lactis and C. glabrata. Amn1 has not only preserved a high level of amino acid sequence similarity among yeast species, but also conservation of the functional domain of the C terminus can be extended to at least 13 other species, including Homo sapiens, Xenopus tropicalis and Danio rerio [36]. therefore, Amn1 may also be able to mediate turnover of substrates through the UPS in higher species, and regulate other biological processes in addition to post-mitotic cell separation.
All yeast strains and plasmids used in this research are listed in S3 and S4 Tables, and the methods for genetic modification of yeast strains and plasmid construction are detailed in supplementary material and methods.
Yeast cells were cultured in 50 ml YPD liquid medium until OD600 = 1.0 was reached. Cells were then harvested by centrifugation and the cell pellet was washed with ice-cold water. Total RNA was extracted with the Qiagen RNeasy mini kit and then RNA-seq libraries were constructed using the Illumina TruSeq kit. RNA was quantified using the NanoDrop and RNA integrity was assessed using the Agilent BioAnalyzer 2100. Using the Illumina HiSeq 2000, about 5 million 125 bp paired end reads were collected per sample. Reads were mapped to the S288C genome for expression analysis using tophat [37]. Gene expression was measured as reads mapped per kilobase of exon per million reads mapped (RPKM). A list of differentially expressed genes was obtained by using cufflinks and cuffdiff [37]. Total RNA for RT-qPCR was isolated using the hot phenol protocol [38] followed by purification with RNase-free DNase (Promega, USA), and finally subjected to first-strand cDNA synthesis using SuperScript III Reverse transcriptase (Invitrogen, USA). 1μl of the single-strand cDNA in 10-fold dilution was used as the template for real-time quantitative PCR (RT-qPCR) using SYBR-green (Toyobo, Japan) and ACT1 as the internal control. Every tested strain was independently cultured three times to gain three independent biologically replicated samples and each sample was assayed in triplicate in the qPCR assay.
AMN1Kl (XM_451967.1) was cloned from genomic DNA of K.lactis (NRRL Y-1140) using primers ATGTCTTGCGTCTCCAGTATTAG and TCAGGTCGCTTCGTTGAC, whilst AMN1Cg (XM_447123.1) was cloned from genomic DNA of C. glabrata (CBS 138) using primers ATGGTATTGCCTGATTCCAAC and TTATTCATTTTCTAATTGATTAAC. We used UAR3 pop-in/pop-out method to perform AMN1 ortholog replacement experiments. The AMN1 ORF in YL1C was firstly deleted using the URA3 marker, and then the URA3 marker was replaced with AMN1KL and AMN1CG separately.
Proteins were extracted using the protocol described by Kushnirov [39]. Briefly, 107 yeast cells were harvested by centrifugation and the pellet was washed with H2O before incubation in 200μl 0.2M NaOH for 5 minutes at room temperature. Cells were centrifuged again and the pellet was incubated in 50μl SDS-loading buffer (120mM Tris-Cl, 10% glycerol, 4% SDS, 8% β-mercaptoethanol, 0.005% bromophenol blue), and boiled for 5 minutes. Cells were centrifuged and 10μl supernatant was loaded and separated in SDS-PAGE. Western blotting was performed using the standard laboratory procedures [40].
Cell-cycle was synchronized by using of α–factor or nocodazole, as described previously [41]. Briefly, mating type of YL1C cells which were to be arrested with α–factor was switched to MATa using plasmid pTetra, and their BAR1 gene was deleted using natMX4 so that the strain could respond to a low density of α–factor. The logarithmic phase cells (OD600 = 0.5) were grown in YPD with added α–factor (50 ng/ml) for 1.5 hours. The G1-arrested cells were subsequently released after repeatedly washing with pre-warmed ddH2O at 30°C. The cells were then incubated in pre-warmed YPD medium with Pronase E (0.1 mg/ml). Similarly, the nocodazole-arrested assay was performed by adding 200 μl of nocodazole (1.5M) into 20 ml of logarithmic phase cells (OD600 = 0.5). The cells were grown for another 2.5 hours to be arrested in G2/M phase and then washed twice using pre-warmed YPD. The arrested cells were released in 40 ml fresh YPD medium. Finally, 2 ml samples of the cells treated either with α–factor or with nocodazole were harvested every 10 minutes over the next two hours for further analysis.
Yeast cell cultures in a volume of 50 ml were grown in YPD to an optical density of OD600 = 1.0. The cells were then harvested by centrifugation and the cell pellet was washed with ice-cold water. The cells were pelleted and re-suspended in 1000 μl of lysis buffer (50 mM HEPES-KOH at pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Na-deoxycholate, 1 tablet of the complete inhibitor cocktail supplied by Roche) and lysed with acid-washed glass beads for 15 minutes in a vortex on full output. Cell lysate was centrifuged (15 min at 13,000 rpm) at 4°C to remove cell debris. The gel filtration column (Superdex 200; Amersham Biosciences) was washed and equilibrated using cold PBS (4°C). Lysates were passed over the gel filtration column with a flow rate of 0.5 ml/min. Samples were collected every 0.5 ml per tube and analyzed by western blot.
Co-immuno-precipitation assay was performed for detecting physical interaction between Ace2 and Amn1. Cells carrying pGU-Myc-Ace2 were grown in SC-U medium overnight. Cell cultures were diluted to an OD600 of 0.2 using 20ml YPR and grown to stationary phase. Cells were then harvested by centrifugation. The cell pellets were resuspended in fresh 50ml YPR and continuously cultured until OD600 reached an approximate value of 0.8. Galactose was immediately added and MG132 (20 μM in final solution) was added two hours later into the cell culture. The cells were grown for another hour and then harvested by centrifugation. WCEs (whole cell extracts) were extracted using glass beads in IP lysis buffer (50 mM HEPES-KOH at pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 1 tablet of the complete inhibitor cocktail supplied by Roche) and quantified using NanoDrop. 5mg of WCEs was used as input and 50 mg (1000 μl) of WCEs were then added with magnetic dynabead protein G (Invitrogen, USA), which was pre-incubated with the monoclonal mouse anti-Myc (Transgene, China) antibody for 30 min at room temperature. The mixture was rotated for 1 hour at room temperature. The dynabeads conjugating protein complex was harvested by magnetic force and washed three times with IP lysis buffer. The beads were resuspended in SDS-loading buffer (120mM Tris-Cl, 10% glycerol, 4% SDS, 8% β-mercaptoethanol, 0.005% bromophenol blue) and boiled for 5 minutes. All samples were followed by western blotting, using anti-Myc antibody (Transgene, China) and anti-Flag (Sigma, USA). To detect the physical interactions of Cdc53/Amn1, Skp1/Amn1, Amn1/Tem1 and Cdc15/Tem1, yeast cells were grown in YPD until OD600 of 1.0. The following steps were performed as above.
in vitro ubiquitination assays were performed as previously described [42]. Briefly, YL1C cells expressing AMN1368D, AMN1368V or Δamn1 mutant cells were harvested when the OD600 reached 1.0. Cell pellets were resuspended in reaction buffer (50 mM Tris-HCl, pH 7.5, 5 mM MgCl2, 1 mM DTT, 2 mM ATP). Protein was extracted by the standard glass beads method and concentration of the extracts was measured and normalized using the Bradford protein assay. Equal amounts of extracts from YL1C cells expressing AMN1368D, AMN1368V or Damn1 mutant cells were added with 2 μg N-terminal histidine tagged human ubiquitin, and the cell extracts were incubated with Myc-Ace2 (purified from YL1C cells with AMN1 deleted) for 1 hour at 30°C. Myc-Ace2 were recovered using 20 μl of dynabeads protein G (Invitrogen, USA) that was pre-incubated with the monoclonal mouse anti-Myc (Transgene, China) antibody for 30 min at room temperature. After three washes with the lysis washing buffer (50 mM HEPES-KOH at pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100), ubiquitylated proteins were detected using antibodies against the His epitope.
Ste12 tagged with 6 tandem repeats of c-myc epitope was used in ChIP assays for both the haploid YL1C and the corresponding diploid cells. The ChIP assay was implemented according to the classic protocol [43]. In detail, cell cultures in a volume of 50 ml were grown in YPD to an optical density of OD600 = 1.0, and incubated with 1% FA for 15 minutes at room temperature to enable protein-DNA cross-linking. After addition of 125 mM glycine and incubation for a further 5 minutes at room temperature, the cells were harvested and washed three times with ice-cold 1× TBS at pH 7.5 (20 mM Tris-Cl at pH 7.5, 150 mM NaCl). The cells were pelleted and re-suspended in 1000 μl of lysis buffer (50 mM HEPES-KOH at pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Na-deoxycholate, 1 tablet of the complete inhibitor cocktail supplied by Roche) and lysed with acid-washed glass beads for 15 minutes in a vortex on full output. After removing the cell debris by centrifugation at 12000 rpm for 5 minutes at 4°C, the chromatin in the supernatant was sheared to a length of 200 bp to 500 bp using Covaris S220 (Duty Factor = 25%, Intensity Peak Incident Power = 400W, Cycles per Burst = 200, Processing Time = 20 minutes, Volume = 1ml in TC16 tubes).
Immuno-precipitation was performed with 2.5 mg (1000 μl) cell extract in 20 μl magnetic dynabeads protein G (Invitrogen, USA), which was pre-incubated with the monoclonal mouse anti-Myc (Sigma, 9E10, USA) antibody for 2 hours at room temperature. The precipitates were washed in order with lysis buffer, lysis buffer with 360 mM NaCl, washing buffer (10 mM Tris-Cl at pH 8.0, 250 mM LiCl, 0.5% NP-40, 0.5% Na-deoxycholate, 1 mM EDTA), and 1× TE at pH 7.5 using the magnetic device supplied by Toyobo (Japan). The precipitated DNA was eluted by heating in TES for 30 minutes at 65°C and de-cross linked by heating at 65°C overnight, and digesting with proteinase K (Merck, USA) for 1 hour at 37°C. DNA in the precipitate was purified using the PCR Purification Kit (QIAGEN, USA).
|
10.1371/journal.ppat.1004498 | A Gatekeeper Chaperone Complex Directs Translocator Secretion during Type Three Secretion | Many Gram-negative bacteria use Type Three Secretion Systems (T3SS) to deliver effector proteins into host cells. These protein delivery machines are composed of cytosolic components that recognize substrates and generate the force needed for translocation, the secretion conduit, formed by a needle complex and associated membrane spanning basal body, and translocators that form the pore in the target cell. A defined order of secretion in which needle component proteins are secreted first, followed by translocators, and finally effectors, is necessary for this system to be effective. While the secreted effectors vary significantly between organisms, the ∼20 individual protein components that form the T3SS are conserved in many pathogenic bacteria. One such conserved protein, referred to as either a plug or gatekeeper, is necessary to prevent unregulated effector release and to allow efficient translocator secretion. The mechanism by which translocator secretion is promoted while effector release is inhibited by gatekeepers is unknown. We present the structure of the Chlamydial gatekeeper, CopN, bound to a translocator-specific chaperone. The structure identifies a previously unknown interface between gatekeepers and translocator chaperones and reveals that in the gatekeeper-chaperone complex the canonical translocator-binding groove is free to bind translocators. Structure-based mutagenesis of the homologous complex in Shigella reveals that the gatekeeper-chaperone-translocator complex is essential for translocator secretion and for the ordered secretion of translocators prior to effectors.
| Type Three Secretion Systems (T3SS) are essential virulence factors found in many pathogenic Gram-negative bacteria. These machines aid infection by delivering bacterial proteins into host cells where these proteins modulate host processes and help establish a niche for the bacteria. Protein delivery occurs in a highly regulated manner in which proteins involved in early steps in infection, or necessary to build the secretion conduit, are typically secreted before other substrates, a phenomenon termed secretion hierarchy. This study presents the structure of a molecular complex that physically links one class of early substrates, components of the secretion pore termed translocators, to a gatekeeper protein, a protein that has been implicated in the secretion hierarchy. Disruption of this interaction in Shigella disrupts the secretion of translocators, while supporting increased secretion of effectors, resulting in phenotypes indistinguishable from a gatekeeper deletion, and leading to the conclusion that a gatekeeper-chaperone-translocator complex is a critical component of the T3SS.
| Type Three Secretion Systems (T3SS) are conserved bacterial protein delivery machines used by many pathogenic Gram-negative bacteria to deliver a diverse group of protein molecules, termed effectors, into cells [1]–[4]. The type three secretion (T3S) apparatus is a conserved molecular machine that forms a protein-conducting channel from the bacterial cytosol to the target cell cytosol. Major structural components of the T3SS include: a cytosolic-ring complex, which includes the ATPase that catalyzes protein unfolding and secretion; a basal-body, which forms a pore across both the inner and outer bacterial membranes; a needle complex, which extends from the basal-body to the host cell; and translocators, which form a pore in the target cell membrane, termed a translocon [5], [6].
The translocon, through which effectors enter the host cell, is an oligomeric pore formed by bacterial proteins termed translocators that are themselves secretion substrates of the T3SS [2]. Translocator secretion is regulated such that it occurs prior to effector secretion, ensuring that effector secretion occurs after a functional conduit from the bacterial cytoplasm to the target cell has been formed [7], [8]. Efficient secretion is dependent on the interaction of specialized chaperones with cytosolic T3SS components [9]. Molecular structures have revealed two architectures for T3S chaperones: a mixed α/β homo or heterodimer and an all α-helical tetratricopeptide repeat (TPR) chaperone [10]–[14]. These chaperones, termed class I (α/β) and class II (TPR), are specific for effectors and translocators, respectively [9]. Our understanding of how the T3SS switches from translocator to effector secretion is limited, but in multiple systems this process is known to involve a conserved gatekeeper protein [15]–[20]. Gatekeepers are encoded as one of two molecular architectures, either as two separate proteins (YopN-TyeA family), or as a gene fusion (MxiC family) [20]. The importance of this architectural distinction is unknown, and a protein fusion resulting from ribosomal frameshifting has been reported without an evident functional change [21].
In Chlamydia, the gatekeeper, CopN, is known to directly bind Scc3, a translocator-specific chaperone [22]–[26]. Translocator-specific chaperones (class II chaperones) bind directly to translocators, prevent their degradation, and maintain the translocators in a secretion competent state [9], [13]. Structures from homologous class II chaperone/translocator pairs have revealed the chaperone to be a TPR protein with a conserved binding groove that binds an amino terminal chaperone-binding motif on the translocator [12]–[14], [27]. In addition to binding translocators, Scc3 also binds CopN, although the nature of this interaction is unknown [22], [23], [25], [26]. It is not known if CopN and the Chlamydial translocators (CopB/B2) compete for the TPR binding groove on Scc3, or if different binding determinants are responsible for the Scc3-CopN interaction. More fundamentally, it is not known how gatekeepers promote translocator secretion.
Gatekeepers and translocator chaperones have been observed in immunopurified complexes from other systems, but only as components of large complexes that also include other components of the T3SS [15], [17], [28]. Because such complexes are not readily accessible to structural study, we have focused our structural studies on the gatekeeper-translocator chaperone complex from Chlamydia. We reasoned that the CopN-Scc3 complex is likely to be involved in the ordered secretion of translocators prior to effectors, a conserved phenomenon termed the translocator-effector secretion hierarchy.
The origin of the translocator-effector secretion hierarchy is not understood, but has been proposed to arise from differential affinities and competition for binding sites either between chaperones and their effector or translocator cargo or between chaperone-effector/translocator complexes and cytosolic components of the T3SS [10], [11], [29]–[32]. To assess the importance of gatekeeper-translocator chaperone interactions in diverse pathogens, and because adequate tools and reagents for functional analysis of CopN mutants are not available in Chlamydia, we have extended our structural analysis with functional studies of MxiC and IpgC, the gatekeeper and translocator-specific chaperone from Shigella.
We determined the crystal structure of the Scc3-CopNΔ84 complex and refined the structure to 2.2 Å (PDB ID 4NRH). The amino terminal 84 residues of CopN were not included in the construct used for crystallization because they are unstructured [25]. Data collection and structure refinement statistics are given in Table 1, and representative electron density is shown in Supporting Figure S1. Two nearly identical Scc3-CopNΔ84 complexes (RMS deviation 0.34 Å for all CopNΔ84 mainchain atoms and 0.49 Å for all Scc3 mainchain atoms) are present in the asymmetric unit. CopNΔ84 forms a long cylindrical structure composed of three helical domains (Figure 1). A search for structurally similar proteins using the DALI software [33], indicates structural homology to the globin fold, which, aside from the use described here, is used in bacteria both as a light harvesting complex and as a stress response sigma factor [34]–[36]. In gatekeeper proteins multiple domains are concatenated through elongated connecting helices, whereas globin domains typically oligomerize through lateral contacts. CopNΔ84 is structurally similar to other gatekeeper proteins, both MxiC from Shigella and the YopN-TyeA complex from Yersinia. The most substantial differences among family members relate to the position of the carboxy-terminal domain or subunit (Figure 1 and [37], [38]). In the Scc3-CopNΔ84 complex, this domain is translated ∼9.5 Å and rotated ∼50° relative to the YopN-TyeA complex (Figure 1). Similarly, Scc3 is structurally similar to other translocator chaperones. The striking result from the Scc3-CopNΔ84 is the unexpected assembly of the complex and the role of the Scc3 amino terminus in binding CopNΔ84.
T3SS chaperones bind the amino terminus of effectors and translocators, and class II chaperones (those specific for translocators) use a conserved peptide-in-groove binding mode, utilizing the TPR binding groove, in which translocators bind in the concave face of the chaperone [9]–[14], [27]. The structure reveals that Scc3 does not engage CopNΔ84 using this conserved binding groove. Instead, the amino terminus, referred to here as a gatekeeper-binding region (GBR), forms a relatively flat surface, adjacent to the convex side of the TPR and binds across the interdomain interface of the last two domains of CopNΔ84 (Figure 1). The interface formed by this interaction results from burial of 980 Å2 of surface area. The Scc3 side of the interface is formed exclusively by residues from the GBR, consistent with separate functions of translocator and gatekeeper binding for the TPR and GBR regions of Scc3. Despite minimal sequence conservation, other translocator chaperones also have an amino terminal extension (GBR) prior to the TPR domain (Figure 2). In the homologs from Shigella and Pseudomonas this region mediates homo-dimerization, although translocator binding is known to disrupt these dimers such that in translocator-chaperone complexes this region (GBR) is no longer involved in homodimerization [13], [29], [39]. In the homologs from Yersinia and Pseudomonas crystallization and structure determination required removal of the GBR [12], [14], [27].
Scc3 engages CopNΔ84 with residues from the GBR, which bind a contiguous surface on CopNΔ84. This surface is formed by two distant patches of sequence conservation, site 1 and site 2, and spans the second and third domains of CopNΔ84, requiring these two domains to be appropriately oriented (Figures 1, 2). In the YopN-TyeA family, the third domain is encoded as a separate protein, such that in these homologs one would predict the interdomain interface recognized by the Scc3 homolog to be a dimer.
Residues 16–23 of Scc3 interact principally with site 1 of CopNΔ84, whereas residues 24–43 form a much larger interface in which Scc3 projects a ring of hydrophobic sidechains toward CopNΔ84 to surround a highly conserved arginine (R365) on CopNΔ84 (Figures 2, Supporting Figure S2). This interaction includes three tyrosines, one of which, Y43 is oriented to allow a π-cation interaction. Peripheral to this ring of hydrophobic residues are a collection of inter-molecular salt bridges (Figure 2). The circumscribed arginine (R365) is conserved across diverse species, including species with two polypeptide gatekeepers and is among the most conserved surface exposed residues in this protein family (this residue is an arginine in homologs from Shigella, Vibrio, Pseudomonas, Bordetella, and Yersinia and glutamine in Salmonella) (Supporting Figure S2). Residues on the CopN side of this interface are better conserved than those on the GBR, despite the fact that they span two proteins in the Yersinia architecture and are on the same protein in the architecture presented here (Figure 2, Supporting Figure S2).
To assess the importance of the two binding regions, we disrupted each interaction by mutagenesis. We made an amino terminal 24 amino acid deletion to Scc3 (Scc3Δ24), which eliminates the GBR-site one interaction. We also mutated the central arginine and two adjacent residues (A362R, R365D, and G369R) in site 2 of CopNΔ84 (CopNΔ84-RDR). A362, R365, and G369 are buried by Scc3 and likely solvent exposed in unliganded CopNΔ84. We introduced charged residues at these sites with the expectation that the solvent exposed charges would not disturb CopN, but would disrupt the CopNΔ84-Scc3 complex. CopNΔ84-RDR and Scc3Δ24, are well-folded, as judged by circular dichroism spectra similar to CopNΔ84 and Scc3 (Supporting Figure S3, Methods S1). As judged by the inability of Scc3Δ24 to bind CopNΔ84 and the inability of CopNΔ84-RDR to bind Scc3 (Supporting Figure S4), both regions are important for Scc3-CopNΔ84 complex formation.
TPR family proteins are often unfolded when not bound to appropriate ligands and are considered to be somewhat flexible proteins [40], [41]. Scc3 has an appropriately organized but empty binding cleft when bound to CopN. Structural comparison with other class II chaperones, for which structures have been determined in complex with translocator-derived peptides, indicates that CopNΔ84 binding causes no significant reorganization of the translocator-binding site (Figure 3A). The Scc3-CopNΔ84 binding mode leaves the translocator-binding site on Scc3 unperturbed and available to bind translocators (Figure 3A). In support of this observation, the purified Scc3-CopNΔ84 complex is able to directly bind a translocator-derived peptide (presented as residues 158–177 from CopB fused to GST) and form a CopNΔ84-Scc3-CopB158-177 complex as judged by size exclusion chromatography (Figure 3B). Isothermal Titration Calorimetry (ITC) using a synthetic peptide (CopB residues 163–173) revealed Kd's of 79±16 µM for the Scc3-CopB peptide complex and 49±13 µM for the Scc3-CopNΔ84 peptide complex (Supporting Figure S5, Methods S1).
To determine the importance of the gatekeeper-chaperone interaction during T3S, we disrupted the homologous gatekeeper-translocator chaperone interface in Shigella. Shigella, unlike Chlamydia, are genetically tractable allowing disruption of the endogenous mxiC gene (the copN homolog) and rescue with a plasmid expressing mutant or wild-type MxiC. This is a well-established strategy that has been used to study other MxiC mutants [18], [28]. The CopN and Scc3 homologs in Shigella, MxiC and IpgC, form a complex that includes the T3SS ATPase [28]. The Scc3-CopNΔ84 structure was disrupted by mutation of A362R, R365D, and G369R on CopN (Supporting Figure S4), supporting the idea that the homologous mutations would disrupt IpgC-MxiC interface. We expressed the E331R/R334D/I338R MxiC mutant (MxiC-RDR) from a plasmid in a previously described mxiC null Shigella strain [18] and compared secretion profiles following Congo Red induction. MxiC-RDR is deficient in secretion of the translocators IpaB, IpaC, and IpaD, but efficiently secretes IpaA, an effector, and secretes elevated levels of the effectors OspC1-3 and IpgB (Figure 4A). IpaA is not secreted efficiently if wild type MxiC is present, but is secreted earlier and in greater quantities from ΔMxiC or MxiC-RDR strains (Figure 4B, compare IpaA secretion at 10 minutes and 60 minutes). The secretion profile of MxiC-RDR closely mimics that of the ΔMxiC strain (Figure 4A), highlighting the importance of gatekeeper-translocator chaperone complexes in translocator secretion. To further evaluate the importance of the conserved, central, arginine, we evaluated the secretion profile of a single point mutant, MxiC-R334D, revealing a phenotype similar to the triple mutant (Supporting Figure S6). Similar to CopN, MxiC is both the gatekeeper and a T3S substrate [18], [42], [43]. To verify that the mutations to MxiC did not prevent recognition and secretion of MxiC by the T3SS, we evaluated the secretion of MxiC-RDR, which is unaltered from wild-type MxiC (Figure 4A). To further confirm that the mutations didn't grossly alter the structure of MxiC, we compared circular dichroism specta of MxiCΔ74, MxiCΔ74-RDR, and MxiCΔ74-R334D, which indicated that all three proteins are similarly folded (Supporting Figure S3, Methods S1). In these constructs the first 74 residues of MxiC have been deleted to allow expression and purification from E. coli. MxiC-RDR is secreted in a similar manner to MxiC, indicating that the mutations do not disrupt its ability to interact with the T3SS, yet is unable to direct translocators for secretion, and is unable to prevent inappropriate secretion of the effectors, IpaA, OspC1-3, and IpgB.
In this study we have presented the structure of CopNΔ84 bound to Scc3. This structure provides the first description of a novel interaction for translocator-chaperones, which involves an amino-terminal extension, termed the GBR, binding across the YopN-TyeA-like domain interface in CopNΔ84. The translocator-chaperone-gatekeeper interaction involves a conserved arginine from the gatekeeper and the amino terminal GBR from the chaperone. Despite the lack of sequence conservation within the chaperone, the presence of a GBR in class II chaperones appears to be conserved (Figure 2). Scc3's GBR is ∼20 amino acids longer than others, implying that the extensive interactions made to the YopN-like domain are likely unique to the Scc3-CopNΔ84 complex. Although the unique α1 helix of Scc3 (amino acids 2–15) does not contact CopN, α1 does orient the α1–α2 loop, which contacts CopN. In organisms with shorter GBRs, other T3S apparatus components likely contribute to organizing this interface. This appears to be the case in Salmonella, where the gatekeeper interacts with the translocator-chaperone when it is bound to either translocator, but not to the empty chaperone [17]. In Shigella and Pseudomonas, the translocator-chaperones are known to homodimerize using the amino terminus, and translocator binding is necessary to disrupt this dimer [29], [39]. Although the monomeric chaperone is able to bind translocators, our data suggest a role for the chaperone amino terminus distinct from the role in homodimerization, namely a role in gatekeeper binding. We find that Scc3 is a monomer and the Scc3-CopNΔ84 complex is a heterodimer in solution (Supporting Figure S7). We suggest that disruption of chaperone homodimers, by translocator binding likely occurs prior to chaperone-gatekeeper complex formation, and that once the chaperone-translocator complex is formed, it is recruited to the T3S apparatus by the gatekeeper.
The binding mode observed in our structure leaves the canonical translocator-binding groove free and available to bind translocators. The Scc3-CopNΔ84 complex binds directly to CopB (Figure 3, Supporting Figure S5), a Chlamydial translocator, thus establishing a physical link between the gatekeeper and a translocator. By physically linking gatekeepers and translocators, the temporal order of secretion events in which translocators are secreted immediately after gatekeepers and prior to effectors is assured. The mechanism by which gatekeepers are recruited to the apparatus prior to effectors or translocators remains to be determined. The unusual CopN-Scc3 interface and relatively small buried-surface area (980 Å2) are consistent with the physiological role of this complex, in that the gatekeeper-translocator chaperone must assemble and disassemble each cycle of translocator secretion.
In Chlamydia the CopN-Scc3 interaction can be formed with purified proteins, suggesting that it is stronger than in homologous systems, likely because the GBR is longer and complex formation is not prevented by chaperone homodimerization. The precise advantage this affords Chlamydia is unclear, but T3S activation may represent an early “committed step” for Chlamydia infection. As obligate intracellular pathogens, the need for Chlamydia to enter their host subsequent to T3S activation seems absolute. Chlamydia are dependent on their hosts for ATP [44] and therefore a T3S event that doesn't result in entry is likely fatal for the bacterium. Consistent with this idea, Chlamydia also express a second translocator chaperone, Scc2, which is expressed during late stages of infection, after invasion, and does not bind CopN [22], [23].
Mutations shown to disrupt the CopN-Scc3 complex were evaluated in the MxiC-IpgC complex. The highly conserved central arginine and two additional charged residues, located at positions homologous to the sites mutated in CopN, were mutated (E331R/R334D/I338R) on a MxiC expressing plasmid and used to complement a mxiC deletion strain. This mutant MxiC (MxiC-RDR) was expressed and secreted normally, indicating that the mutations did not significantly disrupt MxiC. Strains harboring MxiC-RDR, however, mimicked deletion strains and were both deficient in translocator secretion and secreted elevated levels of effectors, resulting in significant effector secretion at early timepoints, prior to translocator secretion. These results support the conclusion that a key function of MxiC during secretion is to scaffold translocators and that the MxiC-IpgC complex is needed for this function. This arrangement, in which the gatekeeper directs translocators to the secretion apparatus, makes proper assembly of a gatekeeper-chaperone-translocator complex needed to promote translocator secretion. Gatekeepers are also needed to prevent effector secretion [15]–[20] indicating that the relevant “plug” that prevents premature effector secretion is likely the gatekeeper-chaperone-translocator complex. Consistent with this interpretation, disruption of IpaB (a Shigella translocator) or MxiC (the gatekeeper) both cause constitutive effector secretion [18]. In many systems, translocator chaperones have an additional role and act as transcription co-activators after translocator secretion thereby linking expression of some effector proteins to secretion of translocators [9].
Collectively, our results support a new mechanism for the translocator-effector hierarchy. We suggest that the translocators are recruited to the T3S pore as a molecular complex including the gatekeeper, translocator-chaperone, and translocator. There are likely multiple variations on this theme, one of which is evident in Pseudomonas aeruginosa, where translocators make multiple interactions with their chaperone, including regions outside the canonical chaperone binding region, at the extreme carboxy terminus [45]. The entire complex is needed both to promote translocator secretion and to prevent effector secretion. A triggering event from the tip of the T3S needle is known to induce gatekeeper release [18], which through the molecular complex described above is directly linked to translocator secretion. Secretion of the gatekeeper and translocator then allows effector secretion through a gatekeeper independent mechanism, similar to the efficient effector secretion seen in gatekeeper mutants.
CopNΔ84 and Scc3, from C. pneumoniae AR39, and MxiCΔ74 from S. flexneri were amplified and cloned into pET28 with an amino-terminal hexa-histadine tag. CopNΔ84 and MxiCΔ74 are amino terminal 84 and 74 amino acid deletions of CopN and MxiC, respectively. It was previously found that these were well-behaved variants [25], [37]. All CopN variants were made in pET28 using PCR based mutagenesis and verified by sequencing. CopN truncation mutants were designed from described limited proteolysis mass spectrometry analysis [25]. Proteins were expressed in BL21 (DE3) bacteria grown in Luria Broth at 37°C. Bacteria at an optical density (at 600 nm) of ∼0.6 were induced at 20°C with 0.1 mM isopropyl β-D-thiogalactopyranoside, and grown for ∼12 hours. Cultures were harvested by centrifugation and lysed with a French Press in phosphate buffered saline with ∼1 µg/mL chicken egg white lysozyme, ∼1 µg/mL bovine pancreatic deoxyribonuclease I, 10 µg/mL leupeptin, 1 mM PMSF, 0.7 µg/mL pepstatin. The lysate was clarified by centrifugation and proteins were purified by Co-NTA affinity using Talon resin. Eluted proteins were further purified by size exclusion chromatography, superdex 200, before being snap-frozen in liquid nitrogen and stored at −80°C until needed.
The Scc3 binding region of CopB, residues 158–178 based on sequence homology with IpaB from Shigella [13], was expressed as a GST fusion protein from PGEX-4T. Proteins were expressed and purified as for His tagged proteins with minor modifications; proteins were expressed from BL21 bacteria and purified with Glutathione Sepharose 4B (GE Healthcare).
Chaperone, and translocator peptide binding assays were performed by gel-filtration, using a 24 mL Superdex 200 10/300 GL (GE Healthcare), equilibrated in 10 mM Tris-HCl pH 7.5, 150 mM NaCl, run at 0.5 mL/min, and maintained at 4°C. Equivalent molar concentrations, determined from calculated extinction coefficients, of proteins were applied to the gel filtration column. Protein complexes were incubated for 15 minutes prior to analysis. Molecular weight determination was done under the same conditions using gel-filtration standards from BioRad.
Scc3-CopNΔ84 crystals were grown via vapor diffusion from a reservoir containing 0.2 M Na/K tartrate and 18–22% PEG 3350. Crystals were obtained from a 1∶1 mixture of reservoir and 15 mg/MmL Scc3-CopNΔ84. KAu(CN)2 derivatives were prepared by adding 100 mM KAu(CN)2 to the crystal drop to a final concentration of 2 mM KAu(CN)2. After two days of incubation derivative crystals were harvested. Native and derivative crystals were cryoprotected with 15% glycerol and flash cooled.
Diffraction data were collected from single crystals on stations D and F at LS-CAT beam line at the Advanced Photon Source. Data were indexed, integrated and scaled with HKL2000 [46]. Three gold atoms were located and refined using Phenix [47]. The initial figure of merit for these sites was 0.46, which improved to 0.70 following density modification. The model was traced with a combination of automated and manual building in Phenix and COOT [47], [48]. Multiple rounds of refinement were done using Phenix. Refinement included simulated annealing, coordinate, individual B-factor, and TLS refinement as implemented in Phenix [47]. Non-crystallographic symmetry constraints were included in all rounds of positional refinement. Data collection and refinement statistics are given in Table S1. Figures were prepared using Pymol, ClustalW, ESPript, the DALI server, the PISA server, and the Consurf server [33], [49]–[52].
Sequences used in the Consurf alignments of CopN homologs were chosen to represent the sequence diversity within genera shown in Figure 1 and included five Chlamydial sequences (NP_224529.1, NP_829326.1, YP_515466.1, 84785886, 332806765, YP_005809291.1), four Shigella sequences (YP_005712038, YP_313363.1, YP_001883209.1, YP_406185.1), three Salmonella sequences (1236849, 75349427, NP_461818.1), and three Bordetella sequences (NP_880900.1, WP_004568105.1, NP_884470.1). For the two-component gatekeepers, chimeric sequences were generated to agree with the spatial orientations in the YopN/TyeA structure (pdb accession code 1XL3). Yersinia YopN sequences used were NP_863522.1, NP_395173.1, and NP_052400.1. Because there is zero sequence diversity in TyeA from species evaluated here, we used YP_004210060.1 for all chimeras. An identical strategy was used for Pseudomonas and Vibrio. Pseudomonas sequences used were NP_250389.1, WP_003122865.1, and WP_010794024.1, WP_015648550.1, which were all matched with WP_009876220.1. Vibrio sequences used were NP_798046.1, YP_003285992.1, WP_005395115.1, WP_005377238.1, WP_005441804.1, WP_004745560.1, WP_005528936.1, which were all matched with WP_005395113.1.
Secretion assays were performed essentially as described [18], [53], with minor modifications. Shigella strain M90T was a gift from Marcia Goldberg. Shigella strain ΔmxiC as well as pUC19-mxiC have been previously described, and were gifts from Ariel Blocker [18]. The pmxiC-RDR and pmxiC-R334D were made by standard molecular biology methods and used to transform Shigella strain ΔmxiC. Strains were grown on tryptic soy broth (TSB) plates containing 100 µg/mL congo red, with appropriate antibiotics. Colonies were selected and grown overnight in liquid TSB broth at 37°C and harvested by centrifugation. Pellets were resuspended in 5 mLs of fresh liquid medium. A fraction, ∼1∶25 final dilution, of the resuspended cultures was added to 50 mL TSB cultures and grown to an optical density of 1.0 (600 nm). Cultures were harvested by centrifugation, washed with warm media, and resuspended to a final OD600 of 5.0 in PBS+100 mg/mL Congo Red at 37°C for 10 min and 30 min. Samples were analyzed by SDS-Page using both coomassie and silver staining, as well as western blotting. Western blotting was done with an α-MxiC antibody primary, which was a gift from Ariel Blocker, and goat anti-rabbit secondary antibody (LI-COR Biosciences). Blots were developed with an Odyssey fluorescent scanner. Protein bands were identified from Mass Spectrometry of trypsin-digested bands excised from coomassie-stained gels and was performed by the Vanderbilt University Proteomics Core.
|
10.1371/journal.pntd.0004467 | IL-33-Dependent Endothelial Activation Contributes to Apoptosis and Renal Injury in Orientia tsutsugamushi-Infected Mice | Endothelial cells (EC) are the main target for Orientia tsutsugamushi infection and EC dysfunction is a hallmark of severe scrub typhus in patients. However, the molecular basis of EC dysfunction and its impact on infection outcome are poorly understood. We found that C57BL/6 mice that received a lethal dose of O. tsutsugamushi Karp strain had a significant increase in the expression of IL-33 and its receptor ST2L in the kidneys and liver, but a rapid reduction of IL-33 in the lungs. We also found exacerbated EC stress and activation in the kidneys of infected mice, as evidenced by elevated angiopoietin (Ang) 2/Ang1 ratio, increased endothelin 1 (ET-1) and endothelial nitric oxide synthase (eNOS) expression. Such responses were significantly attenuated in the IL-33-/- mice. Importantly, IL-33-/- mice also had markedly attenuated disease due to reduced EC stress and cellular apoptosis. To confirm the biological role of IL-33, we challenged wild-type (WT) mice with a sub-lethal dose of O. tsutsugamushi and gave mice recombinant IL-33 (rIL-33) every 2 days for 10 days. Exogenous IL-33 significantly increased disease severity and lethality, which correlated with increased EC stress and activation, increased CXCL1 and CXCL2 chemokines, but decreased anti-apoptotic gene BCL-2 in the kidneys. To further examine the role of EC stress, we infected human umbilical vein endothelial cells (HUVEC) in vitro. We found an infection dose-dependent increase in the expression of IL-33, ST2L soluble ST2 (sST2), and the Ang2/Ang1 ratio at 24 and 48 hours post-infection. This study indicates a pathogenic role of alarmin IL-33 in a murine model of scrub typhus and highlights infection-triggered EC damage and IL-33-mediated pathological changes during the course of Orientia infection.
| Scrub typhus is a life-threatening disease, caused by infection with O. tsutsugamushi, a Gram-negative bacterium that preferentially infects and replicates in the endothelium. Every year, approximately one million people are infected globally, especially in the Asia-Pacific region. However, the molecular mechanism(s) of tissue pathogenesis and immune responses in scrub typhus remain poorly understood. IL-33 is a damage-associated molecular pattern factor, which can modulate host inflammatory responses in several infectious diseases. In this study, we compared the severity of disease between wild-type (WT) and IL-33-/- mice infected with O. tsutsugamushi and used exogenous IL-33 to further examine the function of IL-33 during the infection. Our studies in mouse models, as well as in vitro studies in human endothelial cells, have revealed a pathogenic role of IL-33 in promoting endothelial cell stress, cellular apoptosis, tissue damage, and host death. This study will help us understand the pathogenesis of severe scrub typhus.
| Orientia tsutsugamushi is an obligately intracellular bacterium and the etiological agent of scrub typhus with a geographical distribution that encompasses much of the Asia-Pacific region [1]. Scrub typhus is a neglected but important tropical disease, which puts one-third of the world’s population at risk. The disease is transmitted by the bite of an infected larval Leptotrombidium mite or chigger. After 7–14 days of incubation, patients exhibit signs of infection such as an inoculation site eschar followed by fever and rash accompanied by non-specific flu-like symptoms. Although the endothelial tropism of Orientia can lead to disseminated endothelial infection that affects all organs; macrophages, dendritic cells and cardiac myocytes are also the targets of infection [2, 3]. Primary characteristics of fatal scrub typhus pathology include diffuse interstitial pneumonia, hepatic lesions, glomerulonephritis, meningoencephalitis, and coagulation disorders [3–6]. Scrub typhus often presents as an acute febrile illness [1, 7]. Without appropriate treatment, scrub typhus can cause severe multi-organ failure with a relatively high mortality rate [8]. Several antibiotics (doxycycline, azithromycin, rifampicin, chloramphenicol, etc.) have been used to treat Orientia infection. Although these antibiotics are effective if given early [9–12], misdiagnosis, inappropriate antibiotic treatment, and antibiotic failures have occurred, emphasizing the need for a vaccine and alternative therapeutics [1]. Understanding the molecular mechanism of the infection will be beneficial for vaccine design and future therapeutic strategies.
Infection-induced renal dysfunction, as well as acute kidney injury, has often been described in moderate-to-severe scrub typhus [13–18]. The severity of disease is often correlated with the extent of renal dysfunction [13, 17, 19, 20]; however, the molecular mechanism that accounts for such renal dysfunction is poorly understood. The endothelium provides a crucial interface between tissues and circulating inflammatory cells. During tissue damage, endothelial cells (ECs) become activated, expressing adhesion molecules that alert circulating leukocytes to possible insults and further allow the leukocyte to transmigrate across the endothelial layer. In the case of scrub typhus, ECs are the primary target cells once the bacteria has disseminated [3, 21]. During infection the ECs become activated, attracting inflammatory cells, resulting in the observed pathology. Endothelial activation and dysregulation can lead to tissue damage and organ dysfunction. Understanding the molecules that are released during infection is crucial to understanding the role of ECs in the host response during scrub typhus.
Damage-associated molecular pattern molecules (DAMPs) are molecules that can initiate and perpetuate an immune response within the noninfectious and infectious inflammatory responses. Among them, IL-33, a member of interleukin-1 family, locates in the nucleus as a chromatin-associated nuclear factor. IL-33 can modulate inflammatory responses when released [22, 23]. In damaged tissues, necrotic cells can directly release endogenous IL-33, which can signal through its receptor IL-33R/ST2L on target cells [24, 25]. IL-33 has pro- or anti-inflammatory roles, depending on the disease models and tissues involved [26]. For example, recombinant IL-33 (rIL-33) treatment can exacerbate cisplatin-induced acute kidney injury by increasing CD4+ T cell infiltration, CXCL1 production, and acute tubular necrosis [27]. IL-33 also mediates inflammatory responses in human lung tissue cells involved in the chronic allergic inflammation of the asthmatic airway [28]. However, IL-33 can be hepatoprotective in viral infection and ischemia/reperfusion-induced acute liver injury [29, 30]. At present, the role of IL-33 in severe scrub typhus is unclear.
In this study, we found that mice infected with O. tsutsugamushi Karp strain had significantly increased expression levels of IL-33 and its receptor in the kidneys and liver, but not in the lungs. IL-33 deficiency resulted in decreased kidney cellular infiltration and apoptotic cells, as well as delayed bodyweight loss. Compared to WT mice, the endothelium stress and activation in the kidneys of IL-33-/- mice were significantly attenuated, as evidenced by increased angiopoietin (Ang) 1 and endothelial nitric oxide synthase (eNOS), and decreased endothelin-1 (ET-1). To further confirm the role of IL-33 in scrub typhus, we injected rIL-33 to sub-lethally infected mice, and observed the exacerbated illness and increased mortality. Moreover, rIL-33 treatment resulted in increased vascular dysregulation in the kidneys of infected mice. In vitro, Orientia infection significantly stimulated IL-33 and ST2 expression in human EC and increased EC activation. These data suggest that IL-33 plays a significant role in modulating host immune and endothelial responses during scrub typhus infection.
We have recently reported a strong type 1 immune response, but a repressed type 2 response, accompanied with severe tissue damage in multiple organs during lethal infection with O. tsutsugamushi Karp strain in B6 mice [31, 32]. Since Orientia lacks the classical ligands for TLR2/4 stimulation, we speculated that the host DAMP molecule, IL-33 plays a role in modulating inflammation responses in this infection. B6 mice were challenged with a lethal dose of Orientia and serially sampled on 2, 6 and 10 days post-infection (dpi). IL-33 expression began to rise at 6 dpi, and the increase was significant at 10 dpi in the kidneys (Fig 1A). Along with this increase, its receptor ST2L also significantly increased at 6 and 10 dpi in the kidneys (Fig 1A). As IL-33 increased as early as 2 dpi in the livers and remained at similar levels throughout the course of infection (Fig 1B), ST2L expression in the liver was also significantly increased at 6 and 10 dpi (Fig 1B). Unlike the kidneys and liver, the lungs had significantly decreased IL-33 expression at 6 and 10 dpi and no statistically significant changes in ST2L expression (Fig 1C). These findings suggest tissue-specific expression of IL-33/ST2L during the infection.
To assess the role of IL-33 in scrub typhus progression, we challenged IL-33-/- and WT mice with a lethal dose of Orientia and monitored them daily for disease manifestations. IL-33-/- and WT mice both lost body weight from 4 to 8 dpi; however, IL-33-/- mice had significantly less weight loss than did WT mice (Fig 2A) and were considerably more active throughout the course of infection as compared to WT mice. Nevertheless, both groups of mice were moribund at 9 dpi, with similar bacterial loads in the kidneys, livers and lungs (Fig 2B–2D). We also examined a panel of cytokines and chemokines at 9 dpi at the RNA and protein levels (Fig 3). IL-33 deficiency led to significantly higher gene expressions of IL-6, IL-10, and IFN-γ in the kidneys, but no major changes for the expression of Th2 cytokines (IL-4 and IL-13) or chemokines (CXCL9 and CXCL10) (Fig 3A). Similar trends by these cytokines were confirmed in protein levels from the kidneys (Fig 3B). The gene expression levels for IFN-γ, TNF-α, and IL-4 in lung and livers were comparable, while IL-13 expression was undetectable (S1 and S2 Figs).
IL-33 appears to play a role in pathogenesis of the kidneys [27]. To further gauge inflammatory responses and renal pathology in IL-33-/- mice, we examined and found fewer cellular infiltrates in the kidneys of IL-33-/- mice. Intertubular infiltration was evident, as well as cellular infiltrations in and around the glomeruli in WT mice, whereas IL-33-/- mice had very few instances of infiltration in the kidneys (Fig 4A). A considerable number of dense/fragmented nuclei resembling apoptotic cells were observed in endothelial locations in WT mice (Fig 4A, WT-box). To confirm that these fragmented nuclei were indeed apoptotic cells, TUNEL immunohistochemistry was performed allowing quantification and comparison of apoptotic cell numbers. The kidneys of WT animals had more intense positive staining compared to that in IL-33-/- mice (Fig 4B), especially in the endothelium (Fig 4C). While apoptotic ECs constituted approximately 50% of the total number of apoptotic cells in kidneys from both groups, the WT kidneys had 5-fold more apoptotic cells than did their IL-33-/- counterparts (Fig 4D). To verify the TUNEL findings, we compared the expression of the anti-apoptotic gene BCL-2 and found a significantly higher level of BCL-2 transcripts in IL-33-/- kidneys than seen in WT controls (Fig 4E). There were no major differences of pathology in lungs and livers between infected WT and IL-33-/- mice (S1 and S2 Figs).
Orientia infection can result in EC stress and activation in lung and liver tissue, as judged by the changes in the Ang2/Ang1 ratio [32]. In the kidneys, elevation in the Ang2/Ang1 ratios were evident as early as 2 dpi and peaked at 10 dpi (Fig 5A and 5B). When EC activation between WT and IL-33-/- mice in the kidneys was compared at 9 dpi (Fig 5C–5E), we found that IL-33-/- mice had significantly attenuated EC stress and activation compared to that in WT controls, as judged by higher levels of Ang1 (an EC-stabilizing factor) and eNOS (a synthase of EC-relaxing factor nitric oxide (NO) [33]), as well as a lower Ang2/Ang1 ratio and lower levels of ET-1 (an important factor in the development of vascular dysfunction by inhibiting eNOS and NO production [34]). The kidney eNOS/ET-1 ratio was significantly higher in the infected IL-33-/- mice (Fig 5F), implying that deficiency in DAMP molecule IL-33 alleviated the endothelial dysfunction in the kidneys of Orientia-infected mice. In the liver and lungs, the Ang2/Ang1 ratios were comparable in infected WT and IL-33-/- mice, indicating similar levels of EC activation in these organs (S1 and S2 Figs).
The above data indicated that the absence of IL-33 during Orientia infection resulted in an attenuated weight loss and cellular apoptosis in the kidneys during lethal challenge, but it was not sufficient to increase mouse survival. To validate the function of IL-33, we infected WT mice with a sub-lethal dose of Orientia and then i.p. delivered rIL-33 or PBS every other day for 10 days. As shown in Fig 6A, the rIL-33 group lost weight more rapidly starting on 8 dpi than did their PBS-injected counterparts. While the PBS-injected mice recovered part of their body weight after 10 dpi, rIL-33-injected mice exhibited severe signs of disease, with a 64.7% mortality rate (Fig 6B). This increased mortality in the IL-33-injected mice seemed not due to an increase in bacterial loads in the livers or kidneys (S3 Fig). To examine the underlying mechanisms, we examined endothelial markers in the kidneys. We found evidence for increased EC stress and endothelial dysfunction in the kidneys of Orientia-infected, IL-33-injected mice, including a significantly reduced Ang1 expression and a near 2-fold increase in Ang2/Ang1 ratio (Fig 7A), which was accompanied with a significantly reduced eNOS/ET-1 ratio (Fig 7B). Exogenous IL-33 also increased the liver inflammation and EC stress, as evidenced by increased liver Ang2/Ang1 ratios (S4 Fig). The down-regulated BCL-2 expression, plus increased CXCL1 expression, in the kidneys of rIL-33-treated mice (Fig 7C) suggested an increased cellular apoptosis and increased IL-33-mediated pro-inflammatory reaction, as previously reported [27].
To further examine the role of EC stress, we infected human umbilical vein endothelial cells (HUVEC) in vitro at 3 and 10 multiples of infection (MOI), respectively. At 24 hours post-infection (hpi), we found an infectious dose-dependent increase in the expression of IL-33, soluble ST2 (sST2), membrane-bound ST2L, and the Ang2/Ang1 ratio (Fig 8A). At 48 hpi, IL-33 levels were similar to those in controls, but elevation in sST2, ST2L, and Ang2/Ang1 ratio remained significant, especially for high-dose infection groups (Fig 8B). We also examined the secretion of IL-33 proteins in culture supernatants of infected HUVECs (MOI 3 and MOI 10) at 0, 3, 24 and 48 hpi by using an ELISA assay. Appreciable IL-33 was detected among high-dose infection groups, rather than in samples infected with 3 MOI and control samples (S5 Fig). Our in vitro data were consistent with those from mouse studies in vivo, implying an important role for IL-33/ST2-mediated responses during Orientia infection.
Invasion of O. tsutsugamushi can cause acute tubular necrosis, leading to renal failure in patients [35, 36]; however, the underlying mechanism is unclear. The DAMP molecule IL-33 is known to be a potent endothelial activator, promoting angiogenesis and vascular permeability [37], and can also selectively target the non-quiescent ECs, driving pro-inflammatory cell activation [38]. IL-33 also contributes to the pathogenesis of cisplatin-induced acute kidney injury [27]. However, it is unclear whether IL-33 modulates tissue injury and progression of scrub typhus. Here, we demonstrate that there was a significant increase in IL-33 and its ST2L receptor expression in the kidneys and liver during O. tsutsugamushi infection in mice and that IL-33 contributed to renal pathology and endothelial damage during experimental scrub typhus infection. The absence of IL-33 signaling during lethal infection attenuated cellular and tissue damage and delayed the onset of disease (i.e. weight loss), although such changes were not sufficient to rescue the mice from death. Conversely, addition of exogenous IL-33 during a sub-lethal infection exacerbated the activation of renal endothelium and lethality. Moreover, Orientia infection alone was capable of inducing gene expression of IL-33 and its receptors, as well as endothelial activation, in human endothelial cells. We have proposed a pathogenic role of IL-33 in endothelial dysregulation during the infection (Fig 9). This is the first study to address the role of IL-33 in a mouse model of scrub typhus.
IL-33 has crucial and diverse roles in infectious diseases, depending on the type of infectious agents, acute or chronic infection stages, tissues involved, and host immune microenvironments [39]. The protective roles of the IL-33/ST2 axis have been reported during chronic viral infection in the liver, via promoting CD8+ T-cell responses [40], repressing inflammatory cytokine TNF-α, inducing type 2 innate lymphoid cells (ILC2), and protecting the liver in acute adenovirus infection [30]. IL-33-induced ILC2 also promotes lung tissue homeostasis in influenza virus infection [41]. However, IL-33 also plays deleterious roles during Cryptococcus neoformans-induced lung mycosis and allergic inflammation in the lungs [42, 43]. Research concerning the role of IL-33 in kidney infection is relatively limited, and a few reports are focused on cisplatin or Candida albicans-induced renal injury [13, 44]. Our data demonstrate that infection with O. tsutsugamushi Karp strain can increase gene expression of IL-33 and ST2L in the kidneys and liver (Fig 1). The cellular sources of IL-33 in Orientia-infected tissues was not examined in this study, due to technical issues with respect to cell isolation in the ABSL3 facility; but the possible candidates may include ECs in the kidneys [22] and hepatocytes in the liver [45]. The marked reduction of IL-33 expression in the lung tissues at 6 and 10 dpi may not be surprising, given the massive cellular necrosis and tissue damage [32]. Yet, the tendency of reduced IL-33 expression at 2 dpi in the lung was interesting. Regardless of the underlying mechanisms, our data suggest tissue-specific roles of endogenous IL-33 and highlight its contributions in renal injury, cellular apoptosis, and endothelial activation in mouse model of severe scrub typhus.
To further investigate how IL-33 regulates immune responses and why IL-33-/- mice have attenuated weight loss and kidney injury (Figs 2–4), we examined a panel of immune cytokines in the kidney after infection. We have demonstrated previously that Orientia infection induces strong type 1, but impaired, type 2 immune responses, in several tissues [32]. In the present study, we found that IL-33 deficiency or exogenous rIL-33 did not drastically change type 1 or type 2 cytokine expression during Orientia infection, as reported in a C. albicans-induced renal injury model [44]. While IL-33 did not reprogram type 1 vs. type 2 responses during Orientia infection, the IL-33/ST2 signaling significantly amplified the magnitude of pro-inflammatory responses (Fig 3), cellular apoptosis, EC stress and activation, and host death (Figs 4–7). On one hand, we observed the upregulated CXCL1, but decreased anti-apoptotic gene BCL-2, in the kidneys by rIL-33 treatment. This finding is similar to that in the cisplatin-induced renal failure model [27], suggesting the unique mechanism that IL-33/CXCL1 axis may play a critical role in renal injury not only in the toxic reagent-induced model but also in infectious diseases. On the other hand, we found that IL-33-/- mice have a higher expression of anti-inflammatory cytokine IL-10 at both the gene and protein levels in Orientia infection. This increased IL-10 may play a role in renal protection in infected IL-33-/- mice [46]. Some acute inflammatory mediators (e.g. IL-6, IL-12 and IFN-γ) may contribute to bacterial control [47, 48].
Orientia infection in vitro can activate ECs, leading to cell apoptosis [49, 50]. ECs are known to be the source of nucleus IL-33 [22]; however, few studies have focused on the interaction and DAMP molecule expression in ECs infected with Orientia. We have provided evidence that Orientia infection in vitro increased IL-33 and ST2L expression in and the activation of human ECs by 24 h. Prolonged stimulation (48 h), however, did not alter the IL-33 gene expression levels, but dramatically increased both the soluble and membrane-bound ST2 forms of receptors (Fig 8); this may partially explain our difficulty in detecting IL-33 proteins in the culture supernatants. Since sST2 can bind IL-33 and block intracellular IL-33/ST2L signaling [51], this increased sST2 level may counterbalance the excessive IL-33 signal and keep the homeostasis. In addition to being the source of IL-33, ECs are the target of IL-33 [38]. It was previously shown that angiogenesis in ECs was induced by stimulating endothelial NO production via the ST2/TRAF6-Akt-eNOS signaling pathway [37]. It will be interesting to further examine the intracellular signaling events that regulate IL-33/ST2 expression during O. tsutsugamushi infection.
Based on our in vitro studies in human ECs and in vivo studies in WT and IL-33-/- mice, we have proposed a pathogenic role of IL-33 in endothelial dysregulation during the infection (Fig 8). High-dose O. tsutsugamushi infection in ECs and other cell types can trigger EC stress, dysfunction, and apoptosis. In the WT mice, the increased IL-33 and ST2 expression on EC and IL-33 production may further exacerbate EC stress and damage. These IL-33/ST2-mediated effects are diminished or markedly reduced in O. tsutsugamushi-infected IL-33-/- mice, leading to attenuated renal endothelium activation but higher levels of Ang1 in the kidneys. We have provided evidence that endogenous IL-33 promotes EC inflammation during Orientia infection, via multiple mechanisms, which includes reduced Ang1 and eNOS expression, but increased Ang2 and ET-1 expression in the infected kidneys, as in reports for other models [37, 52]. The interplays among Ang1, eNOS and ET-1 in Orientia-infected ECs warrant further investigation [34]. As expected, our results reveal that IL-33 regulates the balance of Ang1/Ang2 as well as that of eNOS/ET-1, modulating the EC inflammation and tissue dysregulation in the kidneys during severe scrub typhus. Our findings are important in the context of a recent report, showing that IL-33 concentrations in human serum strongly correlated with the severity of Hantaan infection, another endotheliotropic pathogen [53]. Therefore, while IL-33 plays a protective role in other models such as viral hepatitis, it has a pathogenic role in endotheliotropic diseases.
Overall, this study indicates a significant role of IL-33 alarmin in endothelial activation and renal damage, highlighting infection-triggered EC damage and IL-33-mediated pathological changes during the course of O. tsutsugamushi infection. This study provides a better understanding of the pathogenesis and a potential biomarker for monitoring disease progression of scrub typhus cases.
Female WT B6 mice were purchased from Jackson Laboratory. IL-33-/- mice on the B6 background were kindly provided by Dr. Rene de Waal Malefyt (Merck, Palo Alto, CA). Mice were maintained under specific pathogen-free conditions and used at 8- to 12 weeks of age following protocols approved by the Institutional Animal Care and Use Committee (protocol # 1302003) at the University of Texas Medical Branch (UTMB) in Galveston, TX. All mouse infection studies were performed in the ABSL3 facility in the Galveston National Laboratory located at UTMB; all tissue processing and analysis procedures were performed in the BSL2 or BSL3 facilities. All procedures were approved by the Institutional Biosafety Committee, in accordance with Guidelines for Biosafety in Microbiological and Biomedical Laboratories. UTMB operates to comply with the USDA Animal Welfare Act (Public Law 89–544), the Health Research Extension Act of 1985 (Public Law 99–158), the Public Health Service Policy on Humane Care and Use of Laboratory Animals, and the NAS Guide for the Care and Use of Laboratory Animals (ISBN-13). UTMB is a registered Research Facility under the Animal Welfare Act and has a current assurance on file with the Office of Laboratory Animal Welfare, in compliance with NIH Policy.
O. tsutsugamushi Karp strain was used herein, and all infection studies were performed with the same bacterial stock prepared from liver extracts pooled from several infected mice. Infectious organisms were then quantified via a focus forming assay as described previously [31, 32]. WT and IL-33-/- mice were inoculated intravenously (i.v.) with a lethal dose of O. tsutsugamushi (4.5 x 106 FFU in 200 μl). Control mice were similarly injected with PBS. At 9 dpi, serum and tissue samples were collected and inactivated for subsequent analyses. To study the effect of excess IL-33 mice received a mostly sub-lethal injection of 8.5 x 105 organisms. After infection mice were injected intraperitoneally with either PBS or 1 μg of rIL-33 at 2, 4, 6, 8, and 10 dpi, respectively. Animals were monitored for signs of disease progression daily until the end of the experiment (13 dpi).
HUVECs were cultured as described previously [54]. Briefly, HUVECs (Cell Application, San Diego, CA) were cultivated in Prigrow I medium supplemented with 10% (vol/vol) heat-inactivated FBS in 5% (vol/vol) CO2 at 37°C. All experiments were performed between passages 5 and 7, and cells were maintained in Prigrow I medium with 3% (vol/vol) FBS. When HUVECs were confluent, they were collected and seeded onto 24-well plates (Corning Inc., Corning, NY). Once all wells were confluent, the HUVEC monolayers were infected with either 3 MOI, 10 MOI, or media only. Total RNA was extracted from each plate at 3, 24, and 48 hours post-infection (hpi) by using an RNeasy mini kit (Qiagen, Valencia, CA) and digested with RNase-free DNase (Qiagen). Gene expression was determined as described below. Cell-free culture supernatants were collected and stored in -80°C until protein analysis.
IL-33 concentrations in supernatants of control and infected HUVECs were determined by using human IL-33 Quantikine ELISA kits (R&D Systems, Minneapolis, MN) following the manufacturer’s protocol. Briefly, 100 μl of supernatant was added to each well of the anti-hIL-33-coated, 96-well ELISA plate. The plate was analyzed using a Versamax Turntable Microplate Reader (Molecular Devices, Sunnyvale, CA) and Softmax Pro V.4.0. All procedures were performed in the BSL3 facility.
Mouse tissues were collected in an RNALater solution (Ambion, Austin, TX) at 4°C overnight to inactivate infectious bacteria and stored at -80°C for subsequent analyses. Total RNA was extracted from tissue by using an RNeasy mini kit (Qiagen, Valencia, CA) and digested with RNase-free DNase (Qiagen). cDNA was synthesized with the iScript cDNA synthesis kit (Bio-Rad Laboratories, Hercules, CA). The abundance of target genes was measured by qRT-PCR by using a Bio-Rad CFX96 real-time PCR apparatus, and a SYBR Green Master mix (Bio-Rad) was used for all PCR reactions. PCR reactions were started at 95°C for 3 min, followed by 39 cycles of 95°C for 10 sec, and 60°C for 10 sec, and ended with an elongation step at 72°C for 10 sec. Dissociation melting curves were obtained after each reaction to confirm the purity of PCR products. Relative abundance of mRNA expression was calculated by using the 2-ΔΔCT method. Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and β-actin were used as the housekeeping genes. Primer sequences are listed in S1 Table.
Bacterial loads were assessed by quantitative real-time PCR as described previously [31, 32]. DNA was extracted by using a DNeasy Kit (Qiagen, Gaithersburg, MD) from the tissue samples, and the bacterial load at each time point and for each organ sampled was determined by quantitative real-time PCR. The gene for a 47-kDa protein was amplified by using specific primers (OtsuF630 and OtsuR747 (IDT, Coralville, IA). PCR products were detected with a specific probe OtsuPr665 (Applied Biosystems, Foster City, CA). Bacterial loads were normalized to total nanogram (ng) of DNA per μL for the same sample, and data are expressed as the gene copy number of 47-kDa protein per picogram (pg) of DNA. 47-kDa gene copy number was determined by known concentrations of control plasmid containing single-copy inserts of the gene. The plasmid concentration was determined and serially diluted 10-fold for the standards.
All tissues were fixed in 10% neutral-buffered formalin and embedded in paraffin, and sections (5-μm thickness) were stained with hematoxylin and eosin. Apoptosis was detected by using a Millipore ApopTag Peroxidase In Situ Apoptosis Detection kit. Kidneys were assessed for positive staining; five, 40x images were taken on an Olympus BX53 microscope. Images were used so that multiple observers could assess the same fields for apoptosis. DAB-positive cells were counted as apoptotic and divided into endothelial cells, based on cellular and nuclear morphology, and other cells. Cells that were rounded or otherwise not recognizable as ECs were counted as other cells. The number of apoptotic cells was counted per 40x field-of-view. The observers’ counts were pooled and averaged and then numbers compared as total number of apoptotic cells per view and apoptotic ECs only.
Cytokine profiles in the tissues were characterized by using Procarta Plex Mouse Cytokine Panel (eBioscience, San Diego, CA). Briefly, kidney protein was extracted by using RIPA (Cell Signaling Technology, Danvers, MA) plus Protease Inhibitor Cocktails (Sigma, St. Louis, MO). The concentration of protein was determined by a Pierce BCA Protein Assay kit (Thermo Scientific, Waltham, MA). Colored magnetic beads coated with different antigens were mixed together with kidney protein samples, and then allowed to incubate for overnight at 2–8°C. After three wash cycles, detection antibody was added and allowed to incubate for 1 h at room RT, followed by incubation with Streptavidin-Phycoerythrin for 30 min at RT. After removal of excess conjugate, 150 μl of sheath fluid was added to each well. The beads were read on a Bio-Rad Bio-Plex 200 System. Raw data were measured as the relative fluorescence intensity and then converted to the concentration according to the standard curve.
Data were presented as mean ± standard errors of the mean (SEM). Differences between individual treatment and control groups were determined by using Student’s t test. One-way ANOVA was used for multiple group comparisons. Statistically significant values are referred to as *, p < 0.05; **, p < 0.01, ***, p < 0.001, ****, p < 0.0001; NS, no significance.
|
10.1371/journal.pgen.1002794 | Gene Expression Profiles in Parkinson Disease Prefrontal Cortex Implicate FOXO1 and Genes under Its Transcriptional Regulation | Parkinson disease (PD) is a complex neurodegenerative disorder with largely unknown genetic mechanisms. While the degeneration of dopaminergic neurons in PD mainly takes place in the substantia nigra pars compacta (SN) region, other brain areas, including the prefrontal cortex, develop Lewy bodies, the neuropathological hallmark of PD. We generated and analyzed expression data from the prefrontal cortex Brodmann Area 9 (BA9) of 27 PD and 26 control samples using the 44K One-Color Agilent 60-mer Whole Human Genome Microarray. All samples were male, without significant Alzheimer disease pathology and with extensive pathological annotation available. 507 of the 39,122 analyzed expression probes were different between PD and control samples at false discovery rate (FDR) of 5%. One of the genes with significantly increased expression in PD was the forkhead box O1 (FOXO1) transcription factor. Notably, genes carrying the FoxO1 binding site were significantly enriched in the FDR–significant group of genes (177 genes covered by 189 probes), suggesting a role for FoxO1 upstream of the observed expression changes. Single-nucleotide polymorphisms (SNPs) selected from a recent meta-analysis of PD genome-wide association studies (GWAS) were successfully genotyped in 50 out of the 53 microarray brains, allowing a targeted expression–SNP (eSNP) analysis for 52 SNPs associated with PD affection at genome-wide significance and the 189 probes from FoxO1 regulated genes. A significant association was observed between a SNP in the cyclin G associated kinase (GAK) gene and a probe in the spermine oxidase (SMOX) gene. Further examination of the FOXO1 region in a meta-analysis of six available GWAS showed two SNPs significantly associated with age at onset of PD. These results implicate FOXO1 as a PD–relevant gene and warrant further functional analyses of its transcriptional regulatory mechanisms.
| Parkinson disease (PD) is a neurodegenerative disease, which impairs the motor and cognitive abilities of affected individuals. Although the involvement of specific genes in the disease process has been recognized, the underlying genetic mechanisms are not yet understood. One common investigation approach for PD has been the comparison of gene expression levels in brain tissue from PD cases with those from neurologically healthy controls. We performed such an expression analysis in prefrontal cortex tissue from a set of 27 PD and 26 control samples. One of the 489 differentially expressed genes, forkhead box O1 (FOXO1), is involved in transcriptional regulation. Notably, the set of differentially expressed genes identified in our study was enriched for genes regulated by the FoxO1 protein. Analyses of DNA sequence variants known as single-nucleotide polymorphisms (SNPs) in the FOXO1 region, as well as of PD–relevant SNPs across the genome, suggest functional connections between this gene and 1) the age at onset in PD, and 2) the spermine oxidase (SMOX) gene. These findings implicate the involvement of FOXO1 in PD pathogenesis.
| Parkinson disease (PD, OMIM #168600) is a neurodegenerative disorder, which affects primarily motor function (difficulty in movement initiation, tremor, slowness of movement), and secondarily cognitive capabilities of affected individuals. The lifetime risk for the disease is 1.5%, with a median age at onset of 60 and 1.5 increased risk in men compared to women. While a minority of PD cases has been attributed to rare monogenic forms, most cases are likely to be attributed to both genetic and environmental influences [1]. PD has an established pathology, with depletion of up to 60% of dopaminergic neurons in the substantia nigra pars compacta (SN) brain region prior to the onset of motor symptoms, and with protein inclusion aggregates known as Lewy bodies. Nevertheless, the specific cellular mechanisms involved in the onset and propagation of PD are still largely undetermined [2].
A common strategy for studying neurodegenerative diseases has been the analysis of gene expression differences between diseased and neurologically healthy control brain samples using microarray technologies. Given its strong pathology in PD, the region of choice for assessing disease-specific expression changes has been SN. Whole SN tissue samples, as well as individually captured dopaminergic neurons from this brain region, have been used in prior microarray studies [3]. Nonetheless, the significant loss of dopaminergic neurons and the likely reactive responses present in surviving neurons at the time of patient death make the interpretation of expression data from SN challenging. The Sutherland et al. study [4] compared results from multiple SN PD microarrays and found low concordance among the implicated genes and pathways. Possible reasons for the inconsistent results might have been the small sample sizes used in individual experiments, the pronounced loss of pigmented SN neurons in PD cases, other types of cellular heterogeneity within and between disease and control specimens, and the large variability attributable to gender, age, RNA quality, post-mortem interval, and co-occurrence of other neurological disorders (e.g. Alzheimer disease pathology). Recently, the Zheng et al. study [3] used a gene-set enrichment meta-analysis approach to analyze expression data from a total of 17 studies (mostly SN, but also studies from other brain regions, as well as blood and human lymphoblastoid cells). They found 10 gene sets to be consistently associated with PD, including the gene set corresponding to 425 PGC-1α-responsive nuclear-encoded mitochondrial genes. Given this result and additional expression results from cellular disease models, the authors concluded that PGC-1α (PPARGC1A, peroxisome proliferator-activated receptor gamma, coactivator 1 alpha, Entrez ID = 10891) is implicated in PD and is a potential therapeutic target for the disease. Additionally, strategies for the integration of different types of data sources for the study of PD have emerged; a recent study by Edwards et al. (2011) combined expression and GWAS data from non-overlapping samples to detect biological pathways that might be relevant for PD [5].
In the current study, we sought to analyze expression differences between PD and neurologically healthy controls in a manner that would maximize our control of possible technical and design confounders, to the extent possible for a tissue homogenate microarray study. Using the One-Color Agilent 60-mer Whole Human Genome Microarray, we investigated expression differences in the prefrontal cortex Brodmann Area 9 (BA9) in the largest PD brain study to date (27 PD and 26 control samples, E-MTAB-812 ArrayExpress dataset). For the microarray experiment we used prefrontal cortex, a brain region which contains dopaminergic neuron projections, does not show the pronounced cell death observed in SN, while still being molecularly and pathologically affected by the disease [2], [6], [7]. The samples included in our study were highly homogenous: all were from males, with high pH values, and none showed significant Alzheimer disease pathology (e.g. the sample is that of pure Lewy body pathology for cases). To our knowledge, this sample is the most homogenous ever studied for PD (Table 1, Table S1). In addition to our microarray expression data, we had genotyping data available for 50 of the 53 samples, consisting of all 56 genome-wide significant SNPs derived from the US-PD GWAS consortium meta-analysis [8]. We combined the 52 genome-wide significant SNPs with minor allele frequencies greater than 0.1 with 189 microarray probes with false discovery rate (FDR) less than 0.05 and located in genes with common FoxO1 regulation in a targeted expression-SNP (eSNP) study. The performed analyses implicate the forkhead box O1 (FOXO1) gene as having an important regulatory role for PD. Furthermore, support for FOXO1 was found in its association to age at onset (AAO) in the US PD-GWAS consortium data [8].
Differential expression analysis for 27 PD and 26 control prefrontal cortex samples (see Materials and Methods, Microarray QC and differential expression analysis section) revealed 507 mRNA probes, within 489 expressed regions (known genes, as well as non-genic expressed genomic elements), with FDR-adjusted p<0.05. Among these differentially expressed probes, 50 had fold changes greater than 1.5. These 50 probes are displayed in Figure 1 and all the FDR significant probes are presented in Table S2. Since three of the available microarray samples had RIN values below 6, we performed a secondary differential expression analysis after removal of these samples [9]. The obtained FDR-significant results, consisting of 912 mRNA probes, are displayed in Table S5. The 36 probes that reached FDR-level of significance when the entire set of brains was used, but not after removal of low RIN samples are indicated in Table S2. Notably, the fold changes between PD and control samples were generally small, with few probes having fold changes larger than 2. This result differs from some of the previously published SN studies, where the contrasts between the two groups displayed large fold changes [10], [11], which may be attributable to artifacts introduced in the study of SN.
Functional analyses for the FDR-significant genes that were present in the Database for Annotation, Visualization and Integrated Discovery (DAVID) v6.7 and in Ingenuity's Pathway Analysis software (IPA) were performed. 418 of the 507 FDR-significant probes (82.4%) were mapped to 395 genes present in DAVID's database. None of the gene ontology (GO) terms present in DAVID reached Bonferroni or FDR-adjusted statistical significance at α = 0.05 for this set of genes. Nevertheless, brain-specific GO terms with nominal enrichment were observed, such as “neuron development”, “neuron projection development”, “gliogenesis”, and “neuron differentiation”. The DAVID analysis showed ten transcription factor-binding sites (TFBS) enriched at a Bonferroni level of significance less than 0.001 and with a fold enrichment of at least 1.2 for the mapped FDR-significant genes. The fold enrichment represents the ratio between the percentage of genes in the mapped gene list with a specific TFBS and the percentage of genes in the entire DAVID database with the specific TFBS. Among the ten enriched TFBS, the FoxO1 site (Bonferroni p = 1.3E-4, fold enrichment = 1.4) was the only one that corresponded to a gene that was also differentially expressed at FDR significance in the microarray (Table S2). 177 genes with the FoxO1 TFBS, including FOXO1 itself, were present among the 395 DAVID mapped genes (44.8%); these genes were covered by 189 FDR-significant probes. Enrichment for the FoxO1 TFBS among FDR significant genes continued to be observed when limiting analysis to only the genes studied in the microarray, and not all genes in the DAVID database (χ2 p<2.2E-16, odds ratio = 1.37). Notably, although PPARGC1A, gene implicated in the Zheng et al. study [3], was not among the differentially expressed genes in our microarray, this gene was determined to protect dopaminergic neurons when deacetylated by the Sirt1 (sirtuin 1) protein in the MPTP mouse model of PD [12]; SIRT1 expression was increased in PD samples compared to controls at an FDR-level of significance in our microarray.
The FOXO1 gene had two different, but strongly correlated (r = 0.75, p = 4.8E-11) probes in the microarray, both with FDR-significantly increased expression in the PD group. Among the FDR-significant probes corresponding to the genes with FoxO1 TFBS (and excluding the two FOXO1 probes), 78.07% were also associated with an increase in expression for the PD group compared with the control group. This percentage is significantly greater than that observed in the remaining 318 FDR significant probes, where only 66.35% were associated with increased expression in the PD group (χ2 = 7.81; p = 0.0052).
We used Ingenuity's IPA software to identify functional categories enriched for significantly associated genes and to build a functional network based on identified categories related to neurological diseases and processes. The network was constructed by starting with 412 unique genes with at least one FDR significant probe, and that were included in the IPA database. The genes present in two of the top nominally enriched functional categories, “Nervous System Development and Function” (individual functions annotation p-values between 2.56E-4 and 2.39E-2) and “Neurological Disease” (individual functions annotation p-values between 5.68E-4 and 2.39E-2), were merged to form a custom network. The FOXO1 gene was added to this network, and the largest connected component of the network was retained (Figure S1). Among the 31 genes included in this merger of neurologically relevant functional categories (without FOXO1), 24 genes with FoxO1 sites were present (77.4%), which is significantly more than the 44.8% observed in all FDR-significant genes.
To validate our microarray results, we used the QuantiGene Plex 2.0 gene expression assay (Affymetrix, Santa Clara, CA, see Materials and Methods, Microarray validation experiment section) for a subset of 8 PD and 9 control samples included in the microarray. We analyzed the expression of 10 genes that contained probes with FDR significance in the microarray. A microarray gene was validated if the fold change obtained for the analysis of the QuantiGene expression data was in the same direction as the fold change obtained for the analysis of the microarray data in the same subset of 17 samples. By this criterion, nine out of the ten considered genes, including FOXO1, were validated (Table S3). The difference in expression of FOXO1 was nominally significant for the validation study.
We compared our results with those obtained using the Affymetrix HG-U133A microarray data published by Zhang et al. [11], which was the only prior PD microarray performed in prefrontal cortex BA9 tissue. With 27 out of the 11,191 genes present on both microarray platforms showing consistent expression dysregulation, we could not detect a significant overlap between the top genes identified by the two BA9 studies (χ2 test p = 0.61, see Materials and Methods, Analysis of prior PD prefrontal cortex and substantia nigra microarray studies section). The top genes were defined as the set of genes with FDR adjusted p-values below 0.05 for our Agilent microarray (278 out of the 11,191 genes), and the set of genes with unadjusted p-values smaller than 0.05 for the Affymetrix microarray (1,012 out of the 11,191 genes). Despite the lack of significant overlap between the two studies, FOXO1 was one of the replicated genes, showing increased expression in PD samples in the Zhang et al. BA9 data (probe = 202723_s_at, p = 0.004, FC = 1.48). The FDR-significant genes from our microarray study with positive nominal signal in the Zhang et al. study are presented in Table S2.
To further investigate the observed enrichment for genes containing the FoxO1 TFBS among those identified as FDR significant microarray results, we performed a targeted eSNP analysis in the microarray brain samples. For the eSNP analysis, we evaluated the presence of potential regulatory effects of PD associated SNPs on differentially expressed probes from genes with FoxO1 TFBS (see Materials and Methods, SNP genotyping and eSNP analysis sections). We detected a single eSNP relationship (p = 8.1E-6) that met the Bonferroni corrected p-value threshold of 5.36E-5 for the used effective number of SNPs (Figure 2). This finding involved a PD GWAS genome-wide significant SNP, rs11731387, present intronically in the GAK gene (cyclin G associated kinase, Entrez ID = 2580) and a probe present in the 3′-UTR of the SMOX gene (spermine oxidase, Entrez ID = 54498). The rs11731387 minor allele was associated with higher risk for PD in the US PD-GWAS consortium meta-analysis (p = 8.81E-9, beta = 0.3018, [8]) and with decreased SMOX expression. Although stronger in the PD subgroup, the eSNP relationship was present in both PD (p = 7.47E-5, odds ratio = −0.727) and controls (p = 0.037, odds ratio = −0.494). SMOX probe expression was increased in the PD group. Interestingly, SMOX is a gene involved in the dopamine receptor signaling pathway, which is a process that has had evidence for involvement in PD [13], [14], [15].
Additionally, we tried to reduce the group of FDR significant, FoxO1 TFBS genes to a subset of genes which act as mediators for the relationship between FOXO1 expression and disease status (see Materials and Methods, Mediation analysis section). Twenty-nine genes, including SMOX, showed evidence that suggested they may act as mediators for the FOXO1 effect on PD (Table 2). Therefore, these genes may be most relevant to the FOXO1 pathway relating to PD, and could be important gene candidates for further analyses of FOXO1 involvement in the disease.
Finally, we analyzed genome-wide SNP data from the US PD-GWAS consortium meta-analysis [8] to further investigate the role of SNPs in the FOXO1 region for PD affection or age at onset (AAO). While no SNP in the FOXO1 region reached the required p-value for significance of 8.68E-5 in the affection meta-analysis (see Materials and Methods, PD affection and age at onset meta-analysis for the FOXO1 gene region section), two SNPs in the region reached this level of significance in the AAO meta-analysis (Table 3). AAO data were available for most PD brains in the microarray, so we investigated the FOXO1 probe expression – age at onset relationship (Table S1), while adjusting for age at death, post-mortem interval (PMI), and RNA integrity (RIN). FOXO1 expression was not significantly associated with age at onset (A_23_P151426: p = 0.46, beta = 0.008; A_24_P22079: p = 0.90, beta = −0.0007).
We performed a microarray study in prefrontal cortex Brodmann Area 9 (BA9), in a set of homogenous (male, non-significant Alzheimer disease pathology) and high quality (high pH, good RNA integrity) PD and control brain samples. To our knowledge, this is the largest and most uniform microarray brain study to date in PD, and we expect the expression data and available covariate information to represent an invaluable resource for the PD community (ArrayExpress E-MTAB-812 dataset). While the microarray was not performed in the most-involved brain region in PD, the substantia nigra pars compacta (SN), we propose that the use of prefrontal cortex tissue, or of other brain regions with neuropathological involvement of disease, but reduced neuronal cell death, has the potential to overcome limitations associated with the use of severely disease affected tissues; this is especially the case when whole tissue homogenate samples are considered. While the SN is almost completely depleted of dopaminergic neurons by the time of autopsy [16], prefrontal cortex tissue does not show such dramatic neuronal death. Nevertheless, prefrontal cortex is very frequently neuropathologically involved in PD [7] (74% of the cases investigated in the Beach et al. study showed Lewy bodies and associated fibers in this brain region), and shows biochemical alterations related to the disease process [2], [17]. Since the Lewy bodies and associated fibers appear later during the disease in the prefrontal cortex [16], the study of BA9 may reveal pathogenically relevant disease changes.
This study found evidence of a significant role for the transcription factor gene FOXO1 and genes under its transcriptional regulation: (1) FOXO1 expression was significantly increased in PD samples in our study, and (2) the top microarray results were enriched for genes containing the FoxO1 transcription factor-binding site. The increased FOXO1 expression in PD samples is consistent with a previous PD BA9 microarray study reported by Zhang et al. [11], the only prior PD expression study performed in this brain region. The majority of prior PD microarray studies performed in SN tissue also reported increased FOXO1 expression and enrichment of FoxO1 TFBS genes in their top results, with significant meta-analysis p-values for the two FOXO1 probes present on the Affymetrix HG-U133A chip ranging from 4.1E-3 to 3.2E-4 (Table S4).
To further explore the significance of FoxO1 TFBS enrichment, we performed a targeted eSNP study for FDR-significant microarray probes located in genes with FoxO1 regulation and SNPs known to be associated with PD affection at genome-wide level of significance [8]. This analysis revealed a significant relationship between the GAK SNP rs11731387 and probe A_23_P102731 in the SMOX gene. The rs11731387 minor allele was associated with both increased PD risk and decreased SMOX expression. Given the observed increase in SMOX expression in PD compared to control samples, we propose that elevated SMOX expression in the brain is required as a protective mechanism against the biochemical changes that lead to and are present in PD, and is not a direct cause of the disease. This would explain why a SNP that prevents sufficiently elevated SMOX expression levels could enhance sensitivity to PD. When both the GAK SNP and the expression of the A_23_P151426 FOXO1 probe were included as predictors for the expression of the SMOX probe, the magnitude of the effect for these two predictors changed only slightly from the initial results, both of them remaining significant. Given this evidence, we propose that FOXO1, GAK, and SMOX are involved in a common biological pathway, with FOXO1 and GAK independently influencing SMOX expression and, consequently, PD risk.
Multiple sources of evidence have recently implicated the cyclin G associated kinase (GAK) gene in PD [8], [18], [19], [20], [21], [22], [23], although it has been unclear how this gene influences the disease. The SMOX enzyme plays a role in polyamine catabolism, and it is known to be involved in the response to drugs, stressful stimuli, and apoptosis. High expression levels of SMOX have been found in the brain and the polyamine catabolism system has been implicated in psychiatric conditions [24]. Notably, the SMOX protein is a component of the dopamine receptor signaling pathway, where, together with the MAOA (monoamine oxidase A), MAOB (monoamine oxidase B), and IL4I1 (interleukin 4 induced 1) proteins, it makes up the MAO complex (Ingenuity Knowledge Base). Although inconsistently, MAOA and MAOB have been linked to PD in several genetic studies [15], [25], [26], [27], [28], and the dopamine receptor signaling pathway has been implicated in the disease [13], [14], [15].
Further, we performed a mediation analysis to determine if the expression of any of the FDR-significant FoxO1 TFBS genes shows evidence to potentially act as an intermediary step for the observed relationship between FOXO1 expression and PD case/control status. This analysis showed the expression values of 29 genes (Table 2) to act as mediating variables. Notably, SMOX was one of these genes, with the second highest percentage of total mediated effect. This set of 29 genes is of particular interest for future studies of FOXO1 involvement in PD.
Finally, we investigated the FOXO1 region for SNPs linked to PD affection or PD age at onset (AAO) using imputed data from the US PD-GWAS consortium [8]. While no SNP reached statistical significance for affection, two SNPs from the FOXO1 region, rs4509910 and rs9532809, were significantly associated with increased PD AAO. The association results for these two SNPs and six additional ones with p-values<1E-3 for the AAO analysis are displayed in Table 3. Using the SNPExpress database [29], we tried to determine whether or not there is evidence for any relationship between these SNPs and FOXO1 expression in brain tissue. Only three of the SNPs in Table 3 were present in the SNPExpress database; the two SNPs with significant AAO association were not among them. From the three SNPs present in the database, rs7987856 showed evidence for association with FOXO1 expression: its minor allele was associated with increased expression of the 3′-UTR exon of the FOXO1 NM_002015 transcript (beta = 98.91, p = 3.5E-3). This SNP is in high LD with the top AAO SNP, rs4509910, with an R2 of 0.76 in the release 22 of the HapMap CEU population, as determined in Haploview [30]. This result might be indicative of the existence of alternative FOXO1 transcripts in the brain, which could have an effect on the progression of PD.
The Forkhead box, subgroup O (FOXO) transcription factors have been implicated recently in studies of known PD genes and aspects of PD neurodegeneration. The FOXO3a protein was determined to control PINK1 (PTEN induced putative kinase 1, Entrez ID = 65018) transcription in mouse and human cells subjected to growth factor deprivation [31], and it was found to localize to Lewy bodies and Lewy neurites [32]. Additionally, the homologues of human FOXO1 have been recently involved in Drosophila melanogaster or Caenorhabditis elegans models of PD, with different post-translational modifications of the protein showing either protective or harmful effects. Drosophila PINK1 null mutants display mitochondrial dysfunction and dopaminergic neuron loss. Koh et al. [33] showed that Sir2 (the homologue or the human SIRT1/sirtuin 1 protein) and FOXO protect mitochondria and dopaminergic neurons downstream of PINK1. Sir2 and SIRT1 are deacetylases, which have FOXO and FoxO1, respectively, as one of their targets. This protective effect was observed to take place through overexpression of two FOXO target genes, SOD2 (superoxide dismutase 2, mitochondrial) and Thor. Kuwahara et al. recently [34] studied the role of Serine-129 phosporylation of α-synuclein in the transgenic C. elegans (Tg worm) model of synucleinopathy. The pan-neuronal overexpression of nonphosphorylatable (S129A) α-synuclein showed severe defects in the Tg worm. Gene expression profiling of S129A-Tg worms showed strong upregulation of Daf-16/FOXO pathway genes, which the authors proposed to act against the dysfunction caused by the S129A-α-synuclein. Two additional studies [35], [36] reported that phosporylation of FOXO at the same amino acid residue by the LRRK2 (leucine-rich repeat kinase 2) protein or the PRKG2 (protein kinase, cGMP-dependent type II) protein in Drosophila reduces dopaminergic neuron survival.
With our current study, we bring further evidence for the importance of the FOXO1 gene in PD. In addition to the differential expression of this gene in PD versus control BA9 and SN tissue, an increased number of genes under the transcriptional regulation of the FoxO1 transcription factor also have altered expression in BA9 tissue in our study. Finally, SNPs in the FOXO1 region are associated with the age at onset for PD. The results of our study warrant further investigation of the FOXO1 gene and of its protein product in the pathogenesis of PD, and we consider the exploration of the relationship between FOXO1, SMOX, and GAK in various PD models as a possible follow-up step. Although we presented multiple sources of evidence for the involvement of FOXO1 in PD, we cannot rule out that the change in FOXO1 expression may be a secondary effect seen mainly in prefrontal cortex and that this may not be primarily involved in the pathogenesis of PD.
While FOXO1 represents our main finding, additional genes with FDR-significant microarray probes and prior evidence for involvement in PD analyses are worth mentioning. A few of these genes are: HGF (hepatocyte growth factor, Entrez ID = 3082), which encodes a protein that promotes the survival and migration of immature neurons [37], [38], SLC41A1 (solute carrier family 41, member 1, Entrez ID = 254428), which was recently implicated in PD genome-wide association and genotyping studies [39], [40], EGFR (epidermal growth factor receptor, Entrez ID = 1956), a gene shown to play a crucial role in the dopamine-induced proliferation of adult neural precursor cells of subgranular, subventricular, and subependymal zones [41], [42], AQP4 (aquaporin 4, Entrez ID = 361), which encodes for the predominant aquaporin found in the brain, water channel involved in the pathophysiology of cerebral disorders [43], and NEDD4 (neural precursor cell expressed, developmentally down-regulated 4, Entrez ID = 4734), a gene that encodes for a ubiquitin ligase involved in the endosomal-lysosomal pathway and ubiquitinates alpha-synuclein [44]. These and other genes with prior evidence for involvement in PD-related processes are promising targets for further studies.
Finally, it is worthwhile to note the lack of overlap that we observed between our study and the BA9 study performed by Zhang et al. [11]. Some of the observed inconsistency may be due to significant differences between these two microarray analyses. For example, the different microarray platforms might assess different transcripts for the considered genes, gender and disease pathology might have a significant impact on the expression levels of a large number of genes, and the different available sets of covariates might affect the expression results (e.g. RIN is not available for the Zhang et al. data). Even with this apparent lack of overlap, we believe that transcriptome data are relevant and can help bring significant insights in the study of PD. A possible way to alleviate incoherent results could be the establishment of standard protocols for expression studies in brain samples, which is an important, yet overlooked objective. Nonetheless, those findings that do replicate, even with the existent microarray data (e.g. FOXO1), may be pointing to important disease-related pathways.
Brain tissue from the prefrontal cortex Brodmann Area 9 (BA9) was obtained from three different brain banks: the Harvard Brain Tissue Resource Center McLean Hospital, Belmont, Massachusetts, the Human Brain and Spinal Fluid Resource Center VA, West Los Angeles Healthcare Center, California, and the National Brain and Tissue Resource for Parkinson's Disease and Related Disorders at Banner Sun Health Research Institute, Sun City, Arizona [45]. Thirty-three Parkinson disease (PD) and 29 control samples were selected for the microarray study. The samples were selected based on the following criteria: (1) no significant Alzheimer disease pathology (specified by neuropathology reports), (2) tissue pH>6.25, (3) similar ages of death for PD cases and controls, and (4) male.
Total RNA for the 33 PD and 29 control samples was extracted with TRIzol (Invitrogen, Carlsbad, CA). RNA was purified using the RNeasy MinElute Cleanup columns (Qiagen Sciences Inc, Germantown, MD) and its quality was assessed with an Agilent Bioanalyzer Nano Chip 2100 (Agilent, Foster City, CA). 1.65 µg of each RNA sample were labeled and hybridized to the One-Color Agilent 60-mer Whole Human Genome Microarray (#G4112A) at the Agilent Microarray Facility of the Whitehead Institute for Biomedical Research (Cambridge, MA). The dye-normalized and post surrogate processed signal for the green channel, gProcessedSignal, obtained from Agilent's Feature Extraction Software was used for downstream analyses. The raw expression data for the 62 samples were evaluated for individual array quality (MA plots), array intensity distributions (box plots and density plots) and between-array differences (heat maps representing the distance between arrays) using the arrayQualityMetrics Bioconductor package. Nine outlier samples were detected based on the arrayQualityMetrics default criteria [46] and were dropped from further analyses. Table 1 describes the retained microarray samples. Post-mortem interval was the only significantly different covariate between the retained cases and controls (p = 0.02).
Microarray probes were removed if they had expression values outside the detectable spike-in range in more than 50% of the control arrays and more than 50% of the PD arrays, or if they had any of the Agilent flags IsWellAboveBG = 0, gIsSaturated = 1, gIsFeatPopnOL = 1, gIsFeatNonUnifOL = 1 in more than 75% of the arrays. The median expression value was used for replicated probes that passed the above filtering criteria. A total of 39,122 probes out of the total 45,015 probes present on the microarray chips were analyzed in the expression and eSNP studies. The expression data for the retained probes of the 53 arrays (E-MTAB-812 ArrayExpress dataset) were quantile normalized, and the obtained values were base 2 logarithm transformed. All the microarray processing analyses were performed in R (http://www.R-project.org), using the Agi4x44PreProcess and the limma Bioconductor packages.
The relationship of PD/control status to probe expression levels was determined using linear regression in R. The normalized and log 2 transformed mRNA levels were modeled as the dependent variable and the association of PD/control status was adjusted for RNA integrity (RIN), post-mortem interval (PMI) and age at death. The RIN and pH were the most highly correlated variables in our data (Spearman correlation coefficient = 0.403, p-value = 0.001), and we decided to include in the linear regression model only one of these two variables, to avoid the problem of over-adjustment. We chose the RIN variable, given its larger range of values compared with pH (Table 1). False discovery rate (FDR) adjustment was applied to the obtained p-values for the PD/control-probe expression relationship to account for multiple comparisons.
The Agilent identifiers of the FDR significant probes were uploaded and mapped to genes in the Database for Annotation, Visualization and Integrated Discovery (DAVID v6.7, http://david.abcc.ncifcrf.gov/, [47], [48]) for functional annotation. All available functional categories were considered, including Gene_Ontology, Pathways, and Protein_Interactions (contains the transcription factor binding site data from the UCSC database).
The genes corresponding to FDR significant microarray probes were analyzed through the use of Ingenuity Pathways Analysis (Ingenuity Systems, www.ingenuity.com). A data set containing FDR significant Agilent probe identifiers and corresponding fold changes was uploaded into the application. Each identifier was mapped to its corresponding gene in the Ingenuity Knowledge Base. These genes were overlaid onto a molecular network developed from information contained in the Ingenuity Knowledge Base. A network of genes with involvement in neurological diseases and processes was created (Figure S1).
The QuantiGene Plex 2.0 gene expression assay was used for the validation of the microarray (Affymetrix, Santa Clara, CA). The expression levels of 10 genes containing microarray probes with FDR-adjusted p-values smaller than 0.05 (Table S3) and of two control genes (TUBG1, tubulin, gamma 1; HPRT1, hypoxanthine phosphoribosyltransferase 1) were evaluated in a subset of 8 PD and 9 control samples from the Agilent microarray experiment (Table S1). The QuantiGene probes designed by Affymetrix targeted the exact transcripts as the ones targeted by the considered Agilent probes (as defined by transcripts present in the UCSC Genes, RefSeq Genes, and Ensembl Gene Predictions tracks from the UCSC genome browser). Gene expression measurements were performed in triplicates in lysed brain tissue, without prior RNA extraction (the RIN covariate was not available). To evaluate the gene expression differences between the PD and control samples, the following procedure was used: 1) for each gene expression measurement, the background value was extracted from the raw expression count; 2) given the 3 different background-extracted expression measurements for each gene, in each sample, average expression values were calculated; 3) the mean expression values for the ten genes were normalized by the geometric mean of the two control genes in each sample; 4) the base 2 logarithm of the obtained normalized values was calculated; 5) a linear model that included age and PMI was used to determine the difference in expression between the two groups.
The prefrontal cortex Brodmann Area 9 (BA9) microarray expression data published by Zhang et al. [11] were used as a replication study for our microarray results. The Affymetrix CEL files for 14 PD and 16 control samples (Affymetrix Human Genome U133A Array) and the corresponding annotation file were downloaded from ArrayExpress (http://www.ebi.ac.uk/arrayexpress/, E-GEOD-20168). The gcrma method was used to background correct, normalize, and summarize probes for the 30 brain samples. The obtained normalized and base 2 logarithm transformed expression values for the 22,283 available probes were modeled as the dependent variable and the association of PD/control status was adjusted for sex, age at death, PMI, and pH. One control sample lacked covariate information and was removed (GSM506036_1134_BA9_Cm.CEL). After adjustment for covariates, none of the analyzed probes reached FDR significance. To compare our Agilent microarray results with the Zhang et al. Affymetrix results, we annotated the probes for the two microarrays, and assigned the probe with the best p-value to each gene. For the Affymetrix data, 17,564 of the available probes could be assigned to 11,441 Entrez identifiers, while for the Agilent data, 29,927 probes could be assigned to 20,474 Entrez identifiers. There were 11,191 Entrez identifiers common for the two microarrays, and we used a χ2 test to evaluate if the overlap between the top genes observed in the two microarrays was larger than expected by chance. For the purpose of the χ2 test, we defined the top genes as the set of FDR significant genes (FDR adjusted p-value<0.05) for our Agilent microarray [278 genes with Entrez IDs and in common with the Affymetrix array], and the set of genes with unadjusted p-values smaller than 0.05 for the smaller Affymetrix microarray [1,012 genes with Entrez IDs and in common with the Agilent array]. The overlap between these two sets of genes consisted of 27 genes, which are highlighted in Table S2.
Additional Affymetrix PD expression studies performed in the substantia nigra (SN) brain region and present in the ArrayExpress or in the National Brain Databank (NBR, http://national_databank.mclean.harvard.edu/brainbank/Main) public repositories were analyzed similarly to the Zhang et al. data. Since the available covariates for each of the studies varied, we present the results obtained when 1) no covariate was added to the used linear model, and 2) the covariates age or sex (depending on availability) were included. Only the expression studies containing at least one of the two FOXO1 probes present in the BA9 Affymetrix study, 202723_s_at and 202724_s_at, were considered (Table S4). This includes PD studies performed on the following Affymetrix chips: HG-U133A, HG-U133_Plus_2, and HG-Focus (E-GEOD-8397, E-GEOD-20163, E-GEOD-20164, E-GEOD-20186, E-GEOD-20295, E-GEOD-7621, E-GEOD-20141, E-GEOD-20333, and the Simunovic et al. PD study [10] present in NBR). We meta-analyzed the results obtained using 1) no covariates and 2) the covariates age or sex (depending on availability) for the 3 FOXO1 probes included in all or part of the 9 studies using the weighted Z-score approach [49]. This method was chosen since it takes into account both the direction of association and the sample size of the individual studies. The Z-scores for the microarray probes of each expression study were obtained from p-values using the standard normal distribution. This conversion was performed in R by using the function qnorm(p-value/2) and changing the sign of the Z-statistic to match the direction of the estimate of association.
A mediation analysis was performed to assess whether or not the observed association between PD and FOXO1 expression acted through a pathway containing any of the FDR significant FoxO1 TFBS genes. Mediation is assessed by a multistep analysis [50], in which the total effect of FOXO1 is broken down into a direct effect and an indirect effect, acting through the intervening gene. The three analysis steps were: 1) in order to decompose the effects, a logistic regression was performed with PD as the dependent variable and the more strongly associated FOXO1 probe (A_23_P151426) as the predictor to establish the total effect (in the original microarray analysis, expression was used as the dependent variable); 2) a linear regression was performed to establish association between FOXO1 expression and the expression of each of the PD-associated FoxO1 TFBS genes; 3) a logistic regression was performed using each of the FoxO1 TFBS genes as a predictor of PD including FOXO1 expression in the model. All regressions were adjusted for age, PMI, and RIN. The direct effect is determined from the beta estimate of FOXO1 in step 3 of the analysis, while the indirect effect is the product of the beta estimates for the relation between FOXO1 and the FoxO1 TFBS gene and the relation between the TFBS gene and PD after standardization of the betas to account for combination of linear and logistic regressions [51]. Finally, the null hypothesis that the indirect effect equals zero is tested using a Z test [52]. The results are displayed in Table 2.
The 53 retained microarray samples were included among 5,849 PD cases and controls genotyped in the US PD-GWAS consortium meta-analysis replication sample [8]. The samples were genotyped using a custom Illumina genotyping array of 768 SNPs, and 56 SNPs provided genome-wide level of significance in the combined discovery and replication phases, and were considered for functional eSNP analyses of the microarray data. Three of the microarray brain samples failed to genotype at the accepted 98% success rate and were removed from the eSNP analysis.
We performed a targeted trans-effect eSNP analysis in the microarray brain samples for: 1) 52 of the 56 genome-wide significant SNPs from the US PD-GWAS consortium study with minor allele frequencies (MAF) of at least 0.1, and 2) a set of 189 microarray probes with FDR-adjusted p-values<0.05 and which mapped to genes with FoxO1 TFBS (Table S2). Many of the genome-wide significant SNPs in the PD associated regions were in strong to moderate linkage disequilibrium (LD); therefore, we used the program SimpleM [53] to determine the effective number of SNPs tested after accounting for LD in each of the different regions to be N = 34. A modified Bonferroni correction method [54] was used to calculate the required eSNP p-value for a 0.05 alpha level as 5.36E-5. Association between the SNPs and probe expression levels was evaluated in the 26 cases and 24 controls using a 2-degree of freedom (df) linear regression model implemented in Plink [55]. The 2-df model permits a simultaneous test of association between genotype and expression and between genotype and difference in association between cases and controls. This method has been used previously for eQTL studies involving mixed case and control samples and increases the power to detect effects of SNPs on expression levels that may be unique to disease [56]. In addition to including SNP, case status and the SNP x case status interaction term, the linear models were adjusted for RNA integrity number (RIN), post-mortem interval (PMI), and age at death. All SNPs were coded using a dominant model.
We considered the region on chromosome 13 covering the FOXO1 gene, as well as the areas up to 1 Mb away from the 3′ and 5′ ends of the gene (chr13: 39,027,801–41,138,734, hg18). In this region, there were 2,103 imputed SNPs present in the Pankratz et al. [8] meta-analysis of PD affection. Using the SimpleM program [53] and the imputed data for the NGRC GWAS, the largest study included in the meta-analysis, we determined the corresponding number of effective SNPs in the FOXO1 region to be N = 576. Since SimpleM uses only genotype data, each SNP was assigned the imputed genotype with the highest confidence for the purpose of this analysis. Given this number of effective SNPs, a p-value of 8.68E-5 was required for an alpha level of 0.05. In addition to the PD risk meta-analysis conducted by the US PD-GWAS consortium [8], a meta-analysis of age at onset of PD was conducted in 6 PD GWAS studies: the five studies present in the US PD-GWAS consortium (PROGENI/GenePD, NIA Phase I, NIA Phase II, HIHG, NGRC) and the LEAPS study [57]. Prior to meta-analysis, results were filtered for imputation efficiency and any study with a MACH-derived Rsq<0.30 did not contribute a result for that SNP to the meta-analysis. Meta-analysis was performed with METAL ([49], http://www.sph.umich.edu/csg/abecasis/Metal/) using an inverse-variance weighting scheme. This allowed an overall effect size to be estimated. Genomic control was employed so that results were down-weighted if the study's lambda exceeded 1.00.
|
10.1371/journal.ppat.1003046 | Intracellular Vesicle Acidification Promotes Maturation of Infectious Poliovirus Particles | The autophagic pathway acts as part of the immune response against a variety of pathogens. However, several pathogens subvert autophagic signaling to promote their own replication. In many cases it has been demonstrated that these pathogens inhibit or delay the degradative aspect of autophagy. Here, using poliovirus as a model virus, we report for the first time bona fide autophagic degradation occurring during infection with a virus whose replication is promoted by autophagy. We found that this degradation is not required to promote poliovirus replication. However, vesicular acidification, which in the case of autophagy precedes delivery of cargo to lysosomes, is required for normal levels of virus production. We show that blocking autophagosome formation inhibits viral RNA synthesis and subsequent steps in the virus cycle, while inhibiting vesicle acidification only inhibits the final maturation cleavage of virus particles. We suggest that particle assembly, genome encapsidation, and virion maturation may occur in a cellular compartment, and we propose the acidic mature autophagosome as a candidate vesicle. We discuss the implications of our findings in understanding the late stages of poliovirus replication, including the formation and maturation of virions and egress of infectious virus from cells.
| The autophagic degradation pathway is a well-known agent of innate immunity. Several pathogens, including poliovirus (PV), a model for several medically important RNA viruses, subvert this pathway for their own benefit. In doing so, pathogens often inhibit the degradative portion of the pathway, presumably to prevent their own destruction. We show here that, surprisingly, PV infection results in high levels of degradative autophagy. However, we find that autophagic degradation is dispensable for PV replication. Inhibiting the formation of autophagosomes inhibits virus RNA replication and subsequent steps in virus production. Inhibiting the acidification of vesicles, which in the case of autophagosomes precedes fusion with lysosomes and autophagic degradation, inhibits a much later step in virus production. Our data suggest an important role for an acidic compartment of the cell in the final maturation step, cleaving a capsid protein to generate infectious virus. Importantly, these data also call into question the long-standing hypothesis that all steps in the production of infectious poliovirus are cytosolic.
| The Picornaviridae, a family of non-enveloped viruses with a small positive strand RNA genome, includes numerous known and emerging pathogens of medical, veterinary and agricultural importance [1]. Poliovirus (PV) is the most extensively studied virus in this family in terms of our collective understanding of its molecular and cellular biology, biochemistry, structure, life cycle, and pathogenesis and is therefore an important model system. Infection with PV results in numerous changes to the host cell, and perhaps one of the most notable is the massive accumulation of cytosolic double-membraned vesicles [2], [3]. These vesicles are the hallmark of autophagy, a degradative pathway of homeostasis and stress response [4].
Autophagy begins with generation of novel membrane crescents which, as they expand and self-fuse, sequester cytoplasmic contents in double-membraned vesicles, referred to as autophagosomes [5]. As autophagy is induced, increasing amounts of the cellular autophagy protein LC3 become conjugated to the lipid phosphatidylethanolamine. This conjugation confers membrane association and is required for autophagosome formation and membrane expansion [6], [7]. Autophagosomes fuse with endosomes to form amphisomes. This fusion event provides amphisomes with vacuolar ATPases, and results in their acidification [8], [9]. Subsequent to acidification, amphisomes fuse with lysosomes to form single-membraned autolysosomes and cargo is degraded [9]–[11].
Poliovirus specifically induces autophagic signaling, and virus production correlates with the level of autophagic activity in cells. LC3 lipidation is evident as early as 3 hours post-infection (h.p.i.), and by 5 h.p.i. double-membraned vesicles are found throughout the cytoplasm [12], [13]. These vesicles are positive for both LC3 and the endosomal marker LAMP1, indicating that these vesicles have likely fused with late endosomes. PV-induced vesicles also stain with monodansylcadaverine (MDC), a lysosomotropic agent is concentrated in acidic compartments by an ion-trapping mechanism [12], [14]. MDC staining co-localizes with LC3 and LAMP1, indicating that the induced vesicles are likely to be acidic amphisomes. Components of the PV replication complex are located on these vesicles, leading to speculation that they may be the sites of genome replication [12], [15].
Although autophagic signaling is specifically induced by the action of PV proteins, the fate of the induced autophagosomes has not been investigated [12], [13], [15]. The regulation of autophagy can be broadly placed into two classes. One class controls the initiation of autophagosome formation by regulating mamalian target of rapamycin (mTOR) inhibition and subsequent LC3 lipid conjugation [16]. The second class controls the stepwise maturation of autophagosomes into degradative autolysosomes. Inhibitors of vesicle acidification, including the weak base ammonium chloride and the vacuolar ATPase inhibitor bafilomycin A1, have been shown to inhibit amphisome fusion with the lysosome and subsequent degradation of cargo [17]–[19].
The presence of late regulatory mechanisms mean that the formation of autophagosomes can occur without leading to active degradation. During infection with the picornavirus Coxsackievirus B3 (CVB3) double-membraned autophagosomes are observed, but autophagic degradation does not occur [20], [21]. Similarly, the bacteria Legionella pneumophila induces autophagosomes for use as replicative vesicles, but the bacterium secretes factors that delay maturation and fusion with lysosomes [22]–[24]. By inhibiting the degradative portion of the pathway, these pathogens are thought to maximize the benefits of autophagosome formation.
The production of the flavivirus Dengue Virus 2 also correlates with the level of autophagic activity in the cell. Unlike the previous examples, Dengue virus does not appear to replicate its RNA on or within autophagosomes [25]. A series of elegant experiments demonstrated that the virus benefits from the selective autophagic degradation of lipid droplets, known as lipophagy [26]. When lipophagy is inhibited, virus production is reduced. This effect is reversed when cells are supplemented with the products of lipophagy.
These data highlight the remarkable diversity in the ways that viruses subvert the autophagic pathway, and raise the possibility that autophagic degradation could itself promote virus production. Inhibitors of vesicle acidification, which would be expected to inhibit autophagic degradation, have been shown to inhibit infection with several viruses including Semliki Forest virus and human rhinovirus 2 [27], [28]. However, these effects are thought to be primarily associated with elevated pH of the endocytic entry vesicles and not related to autophagy. Previous studies have shown that PV entry, translation, and polyprotein processing are unaffected by these inhibitors [29]. These studies did not investigate overall infectious virus production.
Here we show that PV induces bona fide autophagic degradation, although the degradation is not required for normal virus production. We go on to show that formation of autophagosomes promotes viral RNA replication while acidification of cellular vesicles promotes a post-RNA replication step of infectious virus production. Specifically, we find that maturation of assembled particles into infectious virions is promoted by acidic compartments. We suggest that particles which assemble within, or those captured by, autophagosome-like vesicles are exposed to a low-pH environment, facilitating maturation of infectious virus.
Poliovirus Mahoney type 1 was isolated following transfection with an infectious cDNA [30] and propagated as previously described [31]. Poliovirus stocks were titered on H1-Hela cells. H1-Hela cells were maintained in MEM+10% calf serum (CS). 293T cells were maintained in DMEM+10% fetal bovine serum. For collection of intracellular virus cells were washed with PBS, then collected in 1 mL PBS+ 100 µg/mL MgCl2 and 100 µg/mL CaCl2. Cells were lysed by three cycles of freeze/thawing. Virus was added to monolayers of H1-Hela cells for a 30 minute absorption, after which cells were overlaid with 1% agar in MEM. Plaques were allowed to develop for 48 h, agar overlay was removed and cells stained with crystal violet.
MG132, ammonium chloride (NH4Cl), E64d, pepstatin A, 3-methyladenine (3-MA), and guanidine HCl were purchased from Sigma. Bafilomycin A1 and a polyclonal antibody against p62 were obtained from Santa Cruz Biotechnology. Polyclonal antibodies against GAPDH and LC3 were purchased from Cell Signaling and MBL respectively. LC3A and LC3B as well as irrelevant siRNA was purchased from Dharmacon and were transfected using Lipofectamine 2000 (Invitrogen) according to the manufacturers' instructions.
Total RNA was harvested from infected cells using Trizol (Invitrogen) according to the manufacturers instructions. Four micrograms of RNA were treated with DNA-free DNase treatment (Ambion) and split into two reactions. One half was subjected to reverse transcription using RevertAid First Strand cDNA Synthesis Kit (Fermentas) and oligo-dT primers. The second half of DNase treated RNA was subjected to mock reverse transcription in the absence of the enzyme (“No RT” control). cDNA was serially diluted (8-fold), including “No RT” controls, and measured, in triplicate for each dilution, by real time PCR using iCycler (Bio-Rad). Virus-specific primers were designed by PrimerQuestSM (Integrated DNA Technologies) and used to amplify cDNA. Primers were to the PV genome (TATGATGCATCTAGCCCTGCT and ACAGGTGGTGTGAGTGGT TTAGGT) and GAPDH (TGTGATGGGTGTGAACCACGAGAA and GAGCCCTTCCACAATGCCAAAGTT). The delta-Ct method was used to quantify relative abundance of viral cDNA. Ct values of “No RT” controls did not exceed background levels.
Cells were harvested with phosphate-buffered saline (PBS) containing 55 mM EDTA (pH 8.0) and collected by centrifugation at 1500 rpm for 5 min. Cell pellets were washed in PBS and pelleted by centrifugation at 7500 rpm. Cell pellets were resuspended in RSB-NP-40 (10 mM Tris-HCl [pH 7.5], 10 mM NaCl, 1.5 mM MgCl2, 1% NP-40) supplemented with Mini-complete EDTA-free protease inhibitor cocktail (Roche Applied Science, Indianapolis, IN), and incubated on ice for 15 min. Nuclei and insoluble debris were pelleted by centrifugation at 7500 rpm for 5 min. Cell extracts were then stored at −20°C or subjected immediately to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). Immediately following separation proteins were transferred to PVDF membranes (Thermo Scientific). Membranes were blocked for 1 h with 5% nonfat dry milk solution in Tris-buffered saline containing 1.0% Tween 20 (TBST). Blots were then incubated with the primary antibody, washed with TBST, and incubated for 1 hr with secondary antibody. Immunoreactive bands were visualized by enhanced chemiluminescence (HyBlot Cl, Deville Scientific).
Cells were infected at an MOI of 50 pfu/cell. Protein labeling was performed with 20 µCi of [35S] methionine per mL in methionine-free medium. The radiolabeled cell monolayers were collected in 1 mL PBS and centrifuged at 4500 rpm for 5 minutes. Cells were then resuspended in 30 µL RSB-NP-40 (10 mM Tris-HCl [pH 7.5], 10 mM NaCl, 1.5 mM MgCl2, 1% NP-40) supplemented with Mini-complete EDTA-free protease inhibitor cocktail (Roche Applied Science, Indianapolis, IN), and incubated on ice for 15 min. Nuclei and insoluble debris were pelleted by centrifugation at 7500 rpm for 5 min. Radiolabeled cell lysates were then stored at −20°C. Proteins were separated on a 13% polyacrylamide gel. Images obtained with the Typhoon FLA 9500 (GE Healthcare Life Sciences) and ImageQuant software (Amersham Biosciences).
After 3 h of incubation, the cells were washed twice with DME lacking methionine (GIBCO), 3.0 ml of DME lacking methionine, containing 100 µCi of [35S]methionine (ICN) per ml was added to each plate. Cells were incubated for either 2 or 3 hours at 37.0°C. Cells were washed and harvested by scraping into 1 ml of a solution containing 10 mM Tris (pH 7.4), 10 mM NaCl, 1.5 mM MgCl2, 1% Nonidet P-40, 1 uM phenylmethylsulfonyl fluoride (Sigma), and 40 U of placental RNasin (Promega) per ml. Nuclei were pelleted by centrifugation at 1,600×g for 10 min at 4°C, and 0.5 ml of the resulting cytoplasmic extract was loaded directly on 11-ml gradients containing 15 to 30% sucrose in 10 mM Tris (pH 7.4)-10 mM NaCl-1.5 mM MgCl2. The particles were sedimented through the sucrose gradients by centrifugation for 3 h at 27,500 rpm at 15°C, with a Beckman SW41 rotor. Viral proteins were separated on a 15% polyacrylamide gel. Images obtained with the Typhoon FLA 9500 (GE Healthcare Life Sciences) and ImageQuant software (Amersham Biosciences).
Poliovirus induces autophagic signaling during infection [12], [32]. Several other pathogens induce autophagic signaling during infection, but subsequently inhibit the degradation of autophagic cargo [21], [33]–[35]. To ask if poliovirus infection results in autophagic degradation, we monitored the steady-state level of a cellular protein, p62, over the course of infection. p62 is incorporated into autophagosomes though direct interaction with LC3 on the autophagosome membrane [36]. p62 is degraded following fusion of the autophagosome with the lysosome; therefore, a decrease in cellular levels of p62 reflects increased autophagic degradation [37]. We found that p62 levels begin to decrease by 3 h.p.i. The signal continues to decline over the course of infection, with over 90% of p62 signal gone by 6 h.p.i. (Figure 1A) Since poliovirus inhibits cellular translation, we considered that the loss of p62 signal could be the result of normal protein turnover. To test this, mock-infected cells were incubated with cycloheximide for 4 hours to inhibit cellular translation. We observed approximately 47% reduction in p62 levels, likely due to background autophagy, which is high in HeLa cells [38]. However, this background autophagy is insufficient to account for the p62 reduction observed during poliovirus infection. To further elucidate the mechanism of poliovirus-induced p62 degradation, infections were performed following inhibition of either the ubiquitin-proteasome or the lysosomal degradation pathway. Treatment with MG132, a specific inhibitor of the 26S proteasome, resulted in only a 13% rescue of p62 degradation following PV infection. Treatment of cells with bafilomycin A1, an inhibitor of vacuolar ATPases, inhibits fusion of amphisomes with lysosomes [18]. In the presence of bafilomycin A1, PV infection did not reduce p62 levels (Figure 1B).
To determine if p62 degradation is specific to autophagy, we transfected cells with siRNA specific to LC3A and LC3B, the predominant splice variants of LC3 used in cellular autophagy (Figure 2A). As a control, we transfected with scrambled, irrelevant siRNA. We achieved a 78% knockdown of LC3 protein, as estimated by Western blot in mock-infected cells. In PV-infected cells, the levels of membrane-bound LC3-II increase due to induction of autophagic signaling [32]. LC3 siRNA treatment reduced LC3-II to levels comparable to that of uninfected cells, likely due to a reduction in the overall amount of LC3 available for modification. This level of knockdown resulted in restoration of 55% of wild-type p62 during infection. To confirm these results, we used the autophagy inhibitor 3-methyladenine (3-MA), which reduces autophagy signaling by inhibiting type-III PI-3 kinase activity [39]. As seen in Figure 2B, treatment with 3-MA restored over 50% of p62 levels during infection. These data show that PV, which benefits from autophagosome formation, also allows active autophagic degradation during infection.
We next wanted to investigate if the virus benefits from active autophagic degradation. PV infections were performed at low (Figure 3A, C) or high (Figure 3B, D) multiplicity of infection (MOI) in the presence of lysosomal protease inhibitors. No effect on virus yield was observed when cells were pre-treated with leupeptin, a thiol protease inhibitor, and infected at an MOI of 0.1 pfu/cell [40]. To ensure that leupeptin was inhibiting autophagic degradation, lysates from parallel samples were probed for the autophagosome marker LC3. (Figure 3A). Lipidated LC3 (LC3-II) is incorporated into the autophagosome membrane and is specifically degraded by the autolysosome [41]–[43]. A decrease in background autophagic turnover results in increased LC3-II levels, so, as expected, cells treated with leupeptin showed an increase in LC3-II [38].
We wanted to test the effect of inhibiting autophagic degradation at a high MOI, so, as shown in Figure 3B, we repeated the leupeptin experiment at an MOI of 50 pfu/cell. For high MOI infections, it is possible to directly assay autophagic degradation by turnover of p62. Our Western blot demonstrates inhibition of PV-induced p62 degradation. To confirm that these results were not limited to leupeptin treatment, infections were performed in the presence of a combination of protease inhibitors E64d, a calpain and cathepsin B inhibitor, and pepstatin A, an inhibitor of aspartic proteases [40], [44]. As shown in Figure 3C, cells were infected at MOI of 0.1 pfu/cell and treated with both E64d and pepstatin A, and no difference in virus yield was observed. LC3-II levels significantly increased following treatment with the protease inhibitors, indicating that lysosomal degradation had effectively been restricted. Figure 3D demonstrates treatment with E64d and pepstatin A during high MOI infections, with no effect of viral yield. As in Figure 3B, a high MOI infection allows us to assay p62 levels. As expected, E64d/Pep.A abrogates PV-induced degradation of p62. These results show that inhibiting lysosomal degradation does not inhibit virus replication regardless of the nature of the inhibitors or the MOI of the infection.
The lumen of the amphisome reaches an acidic pH of approximately 5.7 [9]. To investigate whether this acidification is important for virus production, cells were treated with inhibitors of vesicle acidification during infection. Ammonium chloride (NH4Cl) is a weak base that is taken up by intracellular vesicles [45]. The acidic environment results in protonation of the base, which raises the lumenal pH as it diffuses out of the vesicle [46]. Cells were infected with PV, then treated with NH4Cl immediately after absorption of the virus. NH4Cl treatment reduced the production of infectious virus progeny by approximately one log at later time points. (Figure 4A) Since these are low MOI infections, we wanted to determine if the effect of NH4Cl treatment is due to a delay of the infectious cycle or an overall reduction in PV production. In Figure 4B, we carry the infection out to 16 hours and find that the reduction of PV production in the presence of NH4Cl is maintained throughout.
We next wanted to determine if the effect we were seeing was specific to low MOI infections. H1-HeLa cells were infected at an MOI of 50 pfu/cell and cell-associated virus was collected at 6 h.p.i. (Figure 4C) The results were strikingly similar to our low MOI infection experiments, with an approximately one-log reduction in viral titer, indicating that the effect is independent of MOI. We also wanted to investigate if the effects of acidification inhibitors would be diminished if they were added following the peaks of both viral translation and transcription. NH4Cl was added to cells at 3.5 h.p.i. to limit any possible effects on virus entry or the initiation of virus transcription or translation [47]. Cells were then lysed at 7 h.p.i., and plaque assays revealed an 8-fold reduction of infectious virus production. (Figure 4D) These data are similar to the results in Figure 4A, indicating the effect of NH4Cl on the virus life cycle primarily takes place later than 3.5 h.p.i.
To ensure that the effect observed was not specific to treatment with NH4Cl, cells were pre-treated with bafilomycin A1, a specific inhibitor of the vacuolar ATPase responsible for vesicle acidification [48]. Infection with poliovirus following bafilomycin A1 treatment resulted in a 6-fold reduction in infectious virus at 6 h.p.i. (Figure 4E) We carried this time course out to 12 h.p.i., beyond a single viral replication cycle, and found that the effect on viral titer is even more pronounced than in the 6 hour infection, although this may be due to compounding effects that result from the extended pre-treatment with bafilomycin A1 prior to infection. We conclude that inhibiting vesicle acidification reduces infectious poliovirus production significantly.
We wanted to understand the reason for the decrease in viral yield when vesicle acidification is inhibited. Poliovirus does not require a pH change in the entry vesicle to enter cells and release its genome, nor to translate and process the viral polyprotein [29]. However, we wanted to test this using our inhibitors and protocols to ensure that acidification of the entry vesicle was not affecting these early steps in the virus replication cycle. To do this, we used pulse labeling to detect translation of viral proteins. Because PV efficiently inhibits protein synthesis, 35S-Methionine labeling at discrete time points during infection should reveal that viral proteins make up the vast majority of protein production within a few hours of infection [49], [50]. This indicates that virus was able to enter cells, release its genome, and initiate translation. By comparing the pattern of proteins produced to untreated infected cells, we can identify changes in polyprotein processing as well. We observed no significant differences in the pattern of protein labeling from infections done in the absence or presence of NH4Cl. (Figure 5A) We conclude that virus entry, translation and polyprotein processing are not affected by NH4Cl treatment.
To ask if defects in viral RNA replication can explain the decrease in virus yield following NH4Cl treatment, qRT-PCR was performed using poliovirus specific primers to detect changes in viral genome replication. To ensure that our assay was able to detect changes in viral RNA levels, infections were also done in the presence of guanidine HCl, a specific inhibitor of poliovirus RNA replication [51]. Treatment with guanidine HCl resulted in significantly decreased viral RNA levels at 6 h.p.i. No change in viral RNA levels was observed at 6 h.p.i. when infections were carried out in the presence of NH4Cl. (Figure 5B) The reduced titer of intracellular virus from each replicate shows that NH4Cl reduces the yield of intracellular virus at 6 h.p.i., despite normal levels of viral genome replication.
Although tools do not exist to specifically inhibit acidification of the autophagic subset of vesicles, we can inhibit their formation using 3-MA. H1-HeLa cells have high levels of autophagy, and it is difficult to achieve significant reduction of autophagic levels. To examine cells in which autophagy can be efficiently inhibited, we turned to the 293T cell line, in which baseline autophagy levels are much lower [52]. (Figure S1A) We examined LC3 modification following PV infection and found that 3-MA was significantly inhibiting viral induction of the autophagic pathway in 293T cells but not in H1-HeLa cells. (Figure S1B) It is worth noting that although we could consistently detect slightly lower amounts of virus when H1-HeLa cells were infected in the presence of 3-MA, we were unable to identify changes in RNA levels or any other step in virus production. (Figure S1C–E) In 293T cells, 3-MA treatment led to a two-log reduction in viral genomic RNA levels at 6 hours post infection and a one-log reduction in virus titer. (Figure 5C) Our data indicate that inhibiting autophagosome formation and inhibiting vesicle acidification have different effects on the virus life cycle.
There are, essentially, three events which must occur to produce infectious virus from genomic RNA and capsid proteins. The first, assembly of pentamers from capsid protein, is thought to occur spontaneously [53]. The second step, association of the genomic RNA with pentamers, is thought to nucleate assembly of 150S provirions, named after their sedimentation coefficient on a sucrose gradient [54]. After virions assemble, the final step is a cleavage maturation of one of the capsid proteins, VP0, which is cleaved into VP4 and VP2 resulting in the infectious 150S virion. This is believed to be an autocatalytic reaction and there is no published evidence of its pH-dependence [55], [56]. The specific mechanism of virus particle maturation is unclear.
We wanted to determine which of these post-RNA replication steps is promoted by vesicle acidification. To analyze virions, we infected cells at an MOI of 50 pfu/cell and labeled viral proteins with 35S-Methionine beginning at 3 h.p.i. At 5 or 6 h.p.i., cells were collected and gently lysed, and lysates were separated on a continuous 15–30% sucrose gradient. Fractions were collected and analyzed for radioactive content. Figure 6A shows the 35S counts per minute (CPM) of fractions from cultures infected in the presence or absence of NH4Cl. In each case, we observe the two expected peaks, 150S and 75S. The 75S peak consists of empty capsids, while the 150S peak fractions contain encapsidated genomes, only some of which are mature and infectious. We found no significant change in the size or magnitude of the 150S peak, although there was a small but consistent shift in the NH4Cl-treated samples, with both peaks having slightly higher mobility in the gradient. At 5 h.p.i. we observe essentially no change in the peaks resulting from NH4Cl treatment. At 6 h.p.i., we observe an increase in the magnitude of the 75S peak, but little change in the 150S peak, which contains the infectious virions. We conclude that the loss of pfu resulting from NH4Cl treatment is likely not due to defects in particle assembly genome incorporation, as such defects would likely result in a reduction in the amount of 150S particles.
To confirm that NH4Cl was having the expected effect on viral titer, we measured pfu from the pooled 150S peak fractions. (Figure 6B) As in previous experiments, pfu was reduced by about one log in the presence of NH4Cl. Since the size of the peak was not altered by NH4Cl treatment, we hypothesized that the maturation of 150S particles is promoted by vesicle acidification. We pooled the three 150S and three 75S peak fractions and ran them on an SDS-PAGE. The 75S peak should contain exclusively VP0, VP1, and VP3, whereas in the 150S peak VP0 levels will be depleted and replaced with the cleavage products VP2 and VP4 as provirions mature to infectious virions. For reasons that are unclear, labeled VP4 is difficult to detect by SDS-PAGE [56]–[59]. In Figure 6C we identify three labeled bands in 75S fractions in both treated and untreated cells, which we have labeled as VP0, VP1, and VP3 based on classical relative molecular weight observations [55], [57]. In the 150S fraction, we identify the VP2 band at its expected migration between VP1 and VP3. In the presence of NH4Cl, the VP2 band is reduced and the VP0 band is more easily detected. We have quantitated these bands from four independent experiments and graphed the ratio of VP0 protein to the levels of VP3, which is not cleaved during maturation. We find that NH4Cl treatment significantly increases the amount of VP0 present, indicating that vesicle acidification promotes the maturation of encapsidated virus particles into mature, infectious virions.
To confirm that the results were not specific to NH4Cl treatment, we repeated these sucrose gradient purification experiments using bafilomycin A1 to inhibit vesicle acidification. We again see no change in the 150S peak containing encapsidated genomes, although the 75S peak appeared to increase in size. (Figure 7A) As with NH4Cl treatment, we found an increase in VP0 abundance when vesicle acidification is inhibited, indicating that VP0 cleavage is inhibited. (Figure 7B) These data are consistent with the data in Figure 6, and indicate that NH4Cl is having the expected effect on vesicle acidification. Taken together, our data show that autophagosome formation promotes viral RNA replication, while vesicle acidification promotes maturation of the assembled, encapsidated virus particles into infectious virions.
Many pathogens that induce autophagic signaling to promote their own replication also inhibit autophagic degradation, presumably to prevent their own destruction by the autophagy machinery. This idea is based on experiments using autophagy-subverting pathogens such as Coxsackievirus and Legionella pneumophila [20], [21], [60]. In contrast, we have shown that poliovirus, a pathogen whose replication is promoted by autophagic signaling, induces bona fide autophagic protein degradation. However, autophagic degradation does not itself promote poliovirus production. Instead, we find that acidification of cellular compartments is required for normal virus production. Specifically, we have shown that inhibitors of acidification inhibit maturation of virus particles into infectious virions. However, we do not know which cellular compartment is acidifying to promote PV maturation. Since cellular autophagosomes promote PV replication and become acidic prior to their fusion with the lysosome, autophagosomes are an attractive candidate for an acidic compartment to promote virion maturation [18].
During poliovirus infection, autophagic signaling can be identified by 3 h.p.i., and by 5 h.p.i. the cytoplasm is filled with double-membraned vesicles [15], [32], [47]. These autophagosome-like objects, abundant late in infection, appear to be full of viral particles, a phenomenon which has never been explained [2]. Virus-induced autophagosome-like vesicles are marked with proteins from the viral RNA replication complex, and poliovirus, like all positive-strand RNA viruses, replicates its RNA in association with cellular membranes [12], [15]. These data have lead to the hypothesis that autophagic vesicles are the sites of viral RNA replication. An alternative hypothesis for the site of RNA replication suggests that vesicles derived from the secretory pathway provide a substrate for replication complexes [61], [62]. For Coxsackievirus B3, the reorganization of secretory vesicles generates a novel membrane structure to act as a substrate for viral RNA replication complexes [63]. However, since CVB3 does not induce autophagic degradation, there may be important differences between the PV and CVB3 structures [20], [21].
Recently, EM tomography carried out on PV-infected cells fixed at different h.p.i. suggested that single-membraned vesicles, seen early in infected cells, may be connected to the double-membraned vesicles seen later in infected cells [64]. Research in the autophagy field suggests that vesicles of the secretory pathway are used in the generation of autophagosomes [65], [66]. Therefore, it is possible that the two hypotheses for the origin of viral replication vesicles are both correct. We propose such a “unified model”, in which single-membraned secretory-derived vesicles predominate at the peak of RNA replication (2.5 h.p.i.–3.5 h.p.i.) and act as the substrate for RNA replication complexes. (Figure 8) Then, using the cell's autophagic machinery, these vesicles morph into the double-membraned vesicles observed later during infection (4 h.p.i.–6 h.p.i.).
If this model is correct, then it is likely that during the change from single-membraned vesicles to double-membraned vesicles, a significant amount of newly synthesized viral RNA and viral proteins would be taken up into the lumen of the autophagosome-like structures. This would explain the observation of virus particles inside autophagosomes. If the virus only needs the cytosolic face of membranes as a substrate for RNA replication, then the effect of acidification inhibitors would not be significant. However, if characteristics unique to the environment within the vesicle, such as low pH, are important for virus production, then acidification inhibitors would be expected to reduce viral yield. We have provided the first experimental evidence to support this second scenario.
Our data demonstrate that inhibition of autophagosome formation by 3-MA affects poliovirus RNA replication. In the context of our model, there are two possible roles for 3-MA. (Figure 8) One, it could inhibit production of single-membraned secretory pathway-derived vesicles, which represent an early step in formation of the autophagosome. This makes sense in context of recent data showing single-membraned vesicles physically associating with double-membraned vesicles as infection progresses [64]. Alternatively, 3-MA could allow single-membraned vesicles to form, but inhibit the formation of double-membraned vesicles, which then act as primary sites of RNA replication. Although this is a formal possibility, we think it less likely in the context of the data indicating an important role for secretory-derived vesicles in viral RNA replication [61], [63]. In either case, 3-MA inhibits viral RNA replication and all downstream steps in virus production. NH4Cl, however, allows production of early autophagic vesicles, inhibiting only their acidification and maturation, which specifically reduces VP0 cleavage. Inhibiting acidification has no effect on viral genomic RNA replication. These data make it clear that vesicle acidification is important for maturation of newly-assembled poliovirus provirions into infectious particles.
There are several possibilities to explain why vesicle acidification promotes virus maturation. It is important to note that because we do not have the technical ability to inhibit acidification of specific compartments, it is possible that acidification of another cellular compartment, not amphisomes, is key to virus production. It is also possible that treating cells with bafilomycin A1 or NH4Cl alters organelle structure or membrane trafficking in a way that inhibits virus production. We think these explanations are unlikely due to the important role autophagosomes play in virus production and the observed presence of virus particles inside autophagosomes [2]. A second possibility is that autophagosomes act as “sponges” for ions, increasing the pH of the cytosol to promote virion maturation [67]. We believe this is unlikely because PV is an enteric virus, capable of remaining infectious in the low pH environment of the human gut, and should not be adversely affected by low pH.
Our favored hypothesis is that virions taken up inside acidic amphisomes have a greater likelihood of maturing into infectious virions. (Figure 8) Several pieces of data lead us to believe that the amphisome interior a likely site of virus maturation. First, autophagosome formation is required for normal viral RNA replication, indicating that newly synthesized viral genomes are likely to be associated with, or in close proximity to, autophagosomes. Second, recent published evidence showed that PV-induced double-membraned vesicles are derived from RNA replication-associated vesicles. Finally, PV-induced amphisomes, which stain with the viral RNA replication protein 3A, are clearly acidic due to their co-staining with the lysosomotropic agent monodansylcadaverine [12].
The majority of virus particles observed during infection are cytosolic, which appears to pose a problem. How can an important function of virus maturation occur in a compartment containing a small percentage of particles? First, the particle-to-pfu ratio of PV has been measured as anywhere from 30 to 1000 [68], [69]. This means that even at the lowest estimate, the vast majority of particles produced during an replication cycle are not infectious. We find that inhibiting acidification eliminates about 90% of pfu. (Figure 4) Therefore, it is not unreasonable to consider that the particles inside amphisomes may have a much better chance of maturing into infectious virus than those outside the amphisome. In our hands, 10% of pfu are resistant to treatment with acidification inhibitors, and we see evidence of residual VP0 cleavage in the presence of NH4Cl or bafilomycin A1. (Figures 6, 7) This could be for multiple reasons. First, our treatments do not completely block either autophagy or vesicle acidification. Second, there may certainly be residual cleavage of VP0 in the cytosol or at neutral pH. In any case, our data do not show that only particles inside acidic compartments will mature.
Our model is that poliovirus particle maturation is more efficient for particles within amphisomes. This study therefore has important implications for our understanding of the PV life cycle, because all steps in virus production have long been thought to occur in the cytoplasm. There are few reports in the literature showing that any stage in positive-strand virus production is not cytoplasmic. Picornavirus particles have been observed in autophagosomes, providing a precedent for our hypothesis [2], [70]. In addition, the alphavirus Brome Mosaic Virus (BMV) replicates its RNA within invaginations of the endoplasmic reticulum, and the passage of macromolecules through the neck of these compartments is believed to be tightly regulated by the virus [71]. However, it is clear from the data presented here that the compartmental requirement for PV is very different from BMV. Normal levels of poliovirus production require vesicle acidification for a post-RNA replication step or steps. (Figures 4–7)
Virus particles are not truly infectious until the mature particles exit the cell. Viral egress is a poorly understood process, often proposed to be the result of the cytopathic effect at the end of an infectious cycle [72]–[74] Previously it was found that the effect of modulating autophagy is greater on pre-lytic exit of infectious virus than on cell-associated virus [12], [75] Our data are consistent with the model that a higher percentage of packaged, infectious virus assembled and released as a result of increased autophagy.
Vesicles can be induced by the transfection of viral replication proteins into cells, but when those cells are superinfected with PV, the pre-formed vesicles are not used for viral RNA synthesis [76]. These data point to a coupling between RNA replication and vesicle generation and may indicate a viral strategy to ensure that at least some newly replicated RNA reaches the luminal side of the vesicles. If our model is correct and PV particles inside autophagosomes are more likely to be mature, infectious virus, then the viruses inside these double-membraned structures need a mechanism to exit the cell. It is also possible that the steps of virus assembly and encapsidation may occur in the cytosol, but maturation occurs inside acidic vesicles. However, we find this unlikely, as immature virions would face a topology problem, having to cross two membranes to enter the amphisome for efficient maturation cleavage. Therefore, we propose that it is most likely that virions are being assembled and encapsidated inside vesicles.
As a non-enveloped virus, poliovirus is not thought to be found in the wild surrounded by a lipid envelope. Therefore, if poliovirus genome encapsidation occurs in autophagosomes, then there must be some mechanism by which virus can exit from these double-membraned vesicles. We describe some possibilities in Figure 8. The simplest model is that the autophagosome fuses with the plasma membrane, releasing virus into the extracellular space. There are data for such a model in poliovirus, and more recent data has confirmed the existence of a minor secretory pathway linked to autophagy [77]–[80]. However, this fusion would likely have to occur after the double-membraned amphisome is converted to a single-membraned autolysosome, because fusion of a double-membraned vesicle with the plasma membrane would release a cytosol-filled single-membraned vesicle into the extracellular space. In the case of infection, this would release a vesicle presumably filled with multiple virions. Poliovirus is not found in extracellular vesicles, so any such structures would presumably be very short-lived and difficult to detect. Alternatively, there could be a mechanism by which virions re-enter the cytosol after the pH-dependent step. These cytoplasmic viruses would then be released by cell lysis.
Our data indicate that acidic amphisomes promote the late, post-RNA replication step of poliovirus particle maturation. The idea that the interior of these vesicles may be the site of virion assembly, genome packaging, maturation, and cell egress has the potential to alter the models in the field describing the latter part of picornavirus infection.
|
10.1371/journal.ppat.1002927 | Kaposi's Sarcoma Herpesvirus K15 Protein Contributes to Virus-Induced Angiogenesis by Recruiting PLCγ1 and Activating NFAT1-dependent RCAN1 Expression | Kaposi's Sarcoma (KS), caused by Kaposi's Sarcoma Herpesvirus (KSHV), is a highly vascularised angiogenic tumor of endothelial cells, characterized by latently KSHV-infected spindle cells and a pronounced inflammatory infiltrate. Several KSHV proteins, including LANA-1 (ORF73), vCyclin (ORF72), vGPCR (ORF74), vIL6 (ORF-K2), vCCL-1 (ORF-K6), vCCL-2 (ORF-K4) and K1 have been shown to exert effects that can lead to the proliferation and atypical differentiation of endothelial cells and/or the secretion of cytokines with angiogenic and inflammatory properties (VEGF, bFGF, IL6, IL8, GROα, and TNFβ). To investigate a role of the KSHV K15 protein in KSHV-mediated angiogenesis, we carried out a genome wide gene expression analysis on primary endothelial cells infected with KSHV wildtype (KSHVwt) and a KSHV K15 deletion mutant (KSHVΔK15). We found RCAN1/DSCR1 (Regulator of Calcineurin 1/Down Syndrome critical region 1), a cellular gene involved in angiogenesis, to be differentially expressed in KSHVwt- vs KSHVΔK15-infected cells. During physiological angiogenesis, expression of RCAN1 in endothelial cells is regulated by VEGF (vascular endothelial growth factor) through a pathway involving the activation of PLCγ1, Calcineurin and NFAT1. We found that K15 directly recruits PLCγ1, and thereby activates Calcineurin/NFAT1-dependent RCAN1 expression which results in the formation of angiogenic tubes. Primary endothelial cells infected with KSHVwt form angiogenic tubes upon activation of the lytic replication cycle. This effect is abrogated when K15 is deleted (KSHVΔK15) or silenced by an siRNA targeting the K15 expression. Our study establishes K15 as one of the KSHV proteins that contribute to KSHV-induced angiogenesis.
| Kaposi's Sarcoma Herpesvirus (KSHV) causes a multifocal angio-proliferative neoplasm, Kaposi's Sarcoma (KS), whose development involves angiogenic growth factors and cytokines. The K15 protein of KSHV upregulates the host factor RCAN1/DSCR1. RCAN1/DSCR1 has been implicated in angiogenesis but its role in KS has never been investigated. In this study we show that the increased expression of RCAN1/DSCR1 in KSHV-infected endothelial cells depends on K15 and that K15, by recruiting PLCγ1, activates PLCγ1, Calcineurin and NFAT1 to induce RCAN1/DSCR1 expression and capillary tube formation. Deleting the K15 gene from the viral genome, or silencing its expression with siRNA, reduces the ability of KSHV to induce angiogenesis in infected endothelial cells in tissue culture. These findings suggest that the K15 protein contributes to the angiogenic properties of this virus.
| Kaposi's Sarcoma Herpesvirus (KSHV) or Human Herpesvirus 8 (HHV8) is a gammaherpesvirus first identified in Kaposi's Sarcoma (KS) biopsies [1]. Apart from being the etiological agent of the classic, AIDS-associated, endemic (African) and iatrogenic forms of Kaposi's sarcoma, it is also associated with two lymphoproliferative disorders, primary effusion lymphoma (PEL) [2] and multicentric Castleman's disease (MCD) [3]. KS is an angio-proliferative disease and the histology of this tumor is characterized by KSHV-infected spindle-shaped activated endothelial cells, vascular spaces and infiltrating inflammatory cells, in particular, monocytes and eosinophils. Atypical endothelial cells and infiltrating inflammatory cells predominate in the early stage of KS (“patch, plaque”), whereas the endothelial spindle cells, the hallmark of KS lesions, become more numerous in the later nodular stage (reviewed in [4], [5]). Angiogenic and inflammatory cytokines are thought to play an important role in KS pathogenesis [6]. In cell culture, KSHV-infected primary endothelial cells adopt a spindle like morphology, reminiscent of KS spindle cells seen in tumor biopsies [7]–[12]. Both in vivo and in vitro, the vast majority of endothelial spindle cells are latently infected with KSHV, but a small proportion of endothelial cells shows signs of productive viral replication [8], [9]. The formation of spindle cells is mainly due to the latent viral protein vFLIP [13], [14]. In addition, KSHV-infected primary vascular endothelial cells show evidence of differentiation into lymphatic endothelial cells, whereas lymphatic endothelial cells change their transcriptional program to resemble vascular endothelial cells [15]–[17]. Experimentally, KSHV has also been shown to induce neo-angiogenesis by inducing the formation of vascular tubes when KSHV-infected primary endothelial cells are plated on matrigel [18]. This phenomenon appears to be more pronounced after activation of the productive replication cycle [18]–[20]. When expressed in isolation, several viral proteins have been shown to play a role in the induction of angiogenesis, including latency-associated nuclear antigen 1 (LANA; open reading frame (ORF) 73), [21], [22], vGPCR, the viral homologue of a G-protein coupled receptor (ORF74) [23]–[29], viral interleukin 6 (vIL6; ORF-K2) [29]–[31], and two viral chemokine homologues (vCCL-1; ORF-K6, and vCCL-2; ORF-K4) [32]. In addition, the viral K1 protein has also been implicated in KSHV-mediated angiogenesis [19], [20], [33]. Furthermore, KSHV infection leads to the production of angiogenic and inflammatory cytokines like VEGF, bFGF, IL6, IL8, GRO-α and TNF-β/LTAα [31], [34]–[37]. Vascular endothelial growth factor (VEGF) is an important angiogenic growth factor that is expressed in KS lesions and can induce endothelial cell growth and angiogenesis. The biological effects of VEGF are mediated by cell surface receptors, VEGFR-1 (FLT1), VEGFR-2 (KDR), and VEGFR-3 (FLT4) [38]. Endothelial cells express both VEGFR-1 and VEGFR-2, with VEGFR-2 being the principal receptor through which VEGF signals are transmitted [39]. Binding of VEGF to VEGFR-2 results in the activation of intracellular signaling pathways including MAPKs and PI3K [38], [40]. In addition, VEGFR-2 also activates the Phospholipase C γ1 (PLCγ1) pathway leading to an increase in intracellular calcium and activation of protein kinase C (PKC), thereby inducing the expression of NFAT-dependent genes like DSCR1 (Down syndrome critical region 1, also called RCAN1 (Regulator of Calcineurin 1)) and Cox-2 (Cyclooxygenase 2) that have been shown to be involved in angiogenesis [41]–[45]. The human RCAN1/DSCR1 gene comprises 7 exons and exons 1–4 can be alternatively spliced to yield four transcripts (RCAN1.1 through RCAN1.4) [42]. Knockdown of RCAN1 has been shown to inhibit VEGF-mediated migration of endothelial cells in vitro, angiogenesis in vivo and thereby tumor growth [43], [44], [46].
In our previous studies we investigated the functional role of the KSHV K15 protein, a non-structural viral membrane protein. The ORF K15 is located between the terminal repeat region and ORF 75 at the ‘right’ end of the long unique coding region of the viral genome. ORF K15 consists of eight alternatively spliced exons. The main K15 protein is predicted to feature 12 transmembrane segments and a C-terminal cytoplasmic domain, which contains several putative signaling motifs such as two SH2-binding sites (Y431ASIL and Y481EEVL), a proline-rich SH3-binding site (P387PLP) and a tumor necrosis factor receptor-associated factor (TRAF)-binding site (A473TQPTDD) [47]–[49]. We have shown before that the K15 protein interacts with cellular proteins like TRAFs and members of the Src family of protein tyrosine kinases via its C-terminal domain [50], [51], thereby activating the MAP kinases c-jun-N-terminal kinase (JNK) 1 and extracellular signal-regulated kinase (ERK2), as well as the NFκB pathway resulting in the activation of AP-1 and NFAT-dependent gene expression [50], [52], [53]. When expressed in epithelial cells, K15 induces the production of several cytokines and chemokines, as well as cellular genes known to be involved in angiogenesis and cell invasion (e.g. RCAN1/DSCR1, MMP1, MMP2 and IL8) [52].
Here we investigated whether RCAN1/DSCR1 is regulated by K15 in the context of virus-infected endothelial cells and if K15 plays a role in KSHV-induced angiogenesis. Our results show that expression of RCAN1/DSCR1 is upregulated in KSHVwt-infected endothelial cells but not in cells infected with a K15 deletion mutant of KSHV. We further found that K15 directly interacts with PLCγ1 to activate the PLCγ1-Calcineurin-NFAT pathway and induces RCAN1/DSCR1 expression and thereby tubular morphogenesis in KSHV-infected human umbilical vein endothelial cells (HUVECs). Deletion of K15 from the viral genome, or silencing its expression with siRNA, reduces the ability of KSHV to induce angiogenesis in cultured endothelial cells. Our study establishes K15 as one of the KSHV proteins that contribute to KSHV-induced angiogenesis.
HEK293 and HEK293T cells were cultured in Dulbecco's modified Eagle medium (DMEM) (Gibco) and Vero cells in Eagle minimal essential medium (MEM) (Biochrom), each supplemented with 10% FCS (Gibco), 50 IU/ml penicillin and 50 µg/ml streptomycin (Cytogen) in a 5% CO2 incubator. Human umbilical vein endothelial cells (HUVECs) were isolated from freshly obtained human umbilical cords by collagenase digestion of the interior of the umbilical vein as described previously [54] and were cultured in EGM2MV medium (Lonza) at 37°C in a 5% CO2 incubator. HEK293 and Vero clones, stably transfected (HEK293) or infected (Vero) with a bacterial artificial chromosome (BAC) carrying a wildtype or K15-deleted KSHV genome (see below), were cultured with additional 150 µg/ml hygromycin B (PAA). For infection with KSHV, HUVECs were cultured in medium without FCS. KSHV virus stocks, prepared as described below, were titrated on HEK293 cells. HUVECs were infected with a multiplicity of infection (m.o.i.) of 10 and were incubated on a shaker for 20 minutes followed by centrifugation for 30 minutes at 450×g. On the following day the medium was changed to medium containing FCS.
A K15 deletion mutant (KSHVΔK15) was constructed in the KSHV bacterial artificial chromosome 36 (KSHVBAC36) [55] using ET recombination essentially as described in [14] by replacing nucleotide 135338 to 136900 of the KSHV genome with an rpsL/neomycin cassette. Details of the construction of KSHVΔK15 are available on request. To verify the integrity of the KSHVΔK15 genome, the entire BAC was sequenced on a Roche/454 next generation sequencer as previously described [56]. The sequence of KSHVΔK15 has been deposited in GenBank with the accession number JX228174.
4 µg of BAC DNA were used to transfect HEK293 cells at 60% confluence in a 6 well plate. For transfection, Lipofectamine 2000 (Invitrogen) was used according to the manufacturer's instructions. Transfected cells were monitored daily for the percentage of green fluorescent protein (GFP)-positive cells and two days after transfection cells were split 1∶2. One day later, 150 µg/mL hygromycin B was added to the medium, and the medium was changed every 2–3 days. Single clones were picked and propagated further under selection with hygromycin B. The clones were named 293-KSHVwt (for the wild type) and 293-KSHVΔK15 (for the deletion mutant of K15). For the establishment of KSHVΔK15-harboring Vero cells, cells in 48 well plates were infected with virus produced from HEK293 stable cell lines and kept under hygromycin B selection after reaching a confluence of 70–90%. Propagation of Vero cells for the establishment of stable cell lines was carried out as described above for HEK293 cells.
To produce virus stocks for infection of HUVECs, Vero cells containing KSHVwt or KSHVΔK15 genome in a BAC vector, or a recombinant KSHV (rKSHV.219) [57], were plated at 30 to 40% confluence in T-175 flasks. The following day, the medium was replaced with induction medium containing 1.25 mM sodium butyrate (Na-Bu) (Sigma) and 10% SF9 cell supernatant containing a baculovirus (a kind gift from J. Vieira) coding for KSHV RTA [52]. The supernatant was harvested 48 to 72 hours later and centrifuged at 5,000 g for 10 minutes and cell debris was cleared by filtration using 0.45 µm filters. For heat-inactivated virus stocks, the supernatant was heated at 65°C in a thermomixer for 1 hour and then centrifuged at 14,000 rpm for 10 minutes at 4°C to remove precipitates. The cleared supernatant was collected in centrifuge bottles (230 ml/bottle) and centrifuged at 15,000 rpm for 5 hours using a Type 19 rotor in a Beckman ultracentrifuge. The supernatant was then discarded and the pellet was resuspended in 300 µl medium overnight at 4°C. The resuspended virus was kept at 4°C.
For detection and quantification of the KSHV titers, 2×104 HEK293 cells were plated per well of a 96-well plate and infected on the second day with serial dilutions of the centrifuged supernatants. GFP-positive cells were counted on day 2, and the infectious virus titer was calculated per ml.
The retroviral vector pSF91, a gift from Dr. C. Baum from the Department of Experimental Haematology, MHH, contains an internal ribosome entry site (IRES) upstream of the gene expressing enhanced GFP (eGFP). A K15 cDNA, codon-optimized for mammalian cells, was purchased from Gene Art Ltd, Regensburg (Germany) and cloned in the pSF91 vector using Not1 sites to generate pSF91-K15-IRES-GFP.
One day prior to transfection, 4.2×106 HEK293T cells were plated in a 10 cm dish. Transfection was carried out using a transfection medium containing HEPES (1 M) and chloroquine (25 µM). Subsequently, retroviral vector plasmids pSF91-IRES-GFP or pSF91-K15-IRES-GFP (5 µg) were co-transfected with packaging plasmids pM57DAW gag/pol (15 µg) and pRD114 vector (5 µg) expressing the envelope protein, using the calcium-phosphate transfection method. Twelve hours later the medium was replaced with fresh medium. Culture supernatant was harvested 36 and 48 hours post-transfection. Cell debris from the supernatant was cleared by filtration through 0.45 µm filters and then concentrated by ultracentrifugation at 10,000 rpm at 4°C for 16–18 hours using a SW 28 rotor in a Beckman ultracentrifuge. The pellet was resuspended in 200 µl of EGM2MV medium and aliquots were stored at −80°C. To infect HUVECs, 10 µl aliquots of pSF91-IRES-GFP and 50 µl aliquots of pSF91-K15-IRES-GFP virus stocks were inoculated on 1×105 HUVECs in 6 well plates in the presence of polybrene (5 µg/ml). The medium was changed after 4 hours and 2 days after infection, the percentage of infected GFP-expressing HUVECs was estimated under a fluorescence microscope.
Protein lysates from HUVECs were prepared in SDS sample buffer (62.5 mM Tris-HCl, pH 6.8, 2% (w/v) SDS, 10% glycerol, 50 mM DTT, 0.01% (w/v) bromophenol blue) containing β- mercaptoethanol for the detection of both PLCγ1 and phospho PLCγ1. For detection of all other proteins, lysates were prepared in RIPA100 buffer (20 mM Tris pH 7.5; 1 mM EDTA; 100 mM NaCl; 1% Triton-X100; 0.5% sodium deoxycholate, 0.1% SDS). For K15 proteins, lysates were not boiled prior to SDS-PAGE. Proteins were resolved by SDS-PAGE, and transferred onto nitrocellulose membranes (Amersham). Membranes were blocked with PBS-T 5% (w/v) milk. Proteins were detected using the following primary antibodies; a rabbit antibody to the cytoplasmic domain of K15 [48], [50], rabbit polyclonal to RCAN1/DSCR1 (D6694; Sigma), mouse monoclonal to NFAT1 (610702; BD Transductions Labs), rabbit polyclonal to Calcineurin pan A (#07-1491, Millipore), rabbit polyclonal to phospho-PLCγ1 (Tyr 783; #2821, Cell Signaling), rabbit polyclonal to total PLCγ1 (# 2822; Cell Signaling), rabbit polyclonal to VEGFR-1 (#2893, Cell Signaling), rabbit polyclonal to VEGFR-2 (#2479, Cell Signaling), rabbit polyclonal to VEGFR-3 (#2638, Cell Signaling) and mouse monoclonal to actin (Chemicon). Membranes were washed 3 times with PBS-T or TBS-T (for phospho-specific antibodies), and incubated with peroxidase-conjugated secondary antibodies. Proteins were detected using a standard enhanced chemiluminescence (ECL) detection kit (Thermo scientific).
Total RNA was extracted with the RNeasy Micro Kit (Qiagen) according to the manufacturer's recommendation and was subjected to microarray analysis using “Whole Human Genome Microarray” (G4112F, AMADID 014850, Agilent Technologies). This microarray contains 45015 oligonucleotide probes covering roughly 31000 human transcripts. Synthesis of cRNA was performed with the “Quick Amp Labeling kit, one color” (Agilent Technologies) according to the manufacturer's recommendation. cRNA fragmentation, hybridization and washing steps were performed exactly as recommended: “One-Color Microarray-Based Gene Expression Analysis V5.7” (see http://www.agilent.com for details) except that 2.5 µg of each labeled cRNA sample were used for hybridization. Slides were scanned on the Agilent Micro Array Scanner G2505 B at two different PMT settings (100% and 5%) to increase the dynamic range of the measurements (extended dynamic range mode). Data extraction was performed with the “Feature Extraction Software V9.5.3.1” by using the recommended default extraction protocol file: GE1-v5_95_Feb07.xml.
Processed intensity values of the green channel (“gProcessedSignal” or “gPS”) were normalized by global linear scaling: All gPS values of one sample were multiplied by an array-specific scaling factor. This scaling factor was calculated by dividing a “reference 75th Percentile value” (set as 1500 for the whole series) by the 75th Percentile value of the particular Microarray (“Array i” in the formula shown below). Accordingly, normalized gPS values for all samples (microarray data sets) were calculated by the following formula: normalized gPSArray i = gPSArray i×(1500/75th PercentileArray i). A lower intensity threshold was defined as 1% of the reference 75th Percentile value ( = 15). All of those normalized gPS values that fell below this intensity border, were substituted by the respective surrogate value of 15. Calculation of ratio values of relative gene expression and data filtering were performed using excel macros or R-Scripts. Filters were set to exclude all technically impaired spots and poorly annotated, poorly characterized, or non-coding transcripts. For K15-overexpression studies, ratio values were calculated from processed signal intensities of “K15 overexpressing/empty vector transduced” HUVEC samples, for experiments in the context of KSHV infection, ratio values were calculated from “KSHVwt-infected/KSHVΔK15-infected” HUVEC samples. Relative fold differences in mRNA expression were color-coded appropriately as indicated in the legend of figure 1.
The same total RNA samples as used in microarray experiments were reverse transcribed using 50 U of BioScript RNase H Low reverse transcriptase (BIO-27036, Bioline) in 20 µl reactions. The enzyme was finally inactivated for 10 minutes at 70°C. Aliquots of generated cDNA samples were used for real-time PCR with the ABI7500 FAST real-time PCR system (Applied Biosystems). Specific amplification was assured utilizing TaqMan probes and gene specific primers. Amplification was performed in 10 µl reactions with TaqMan Universal PCR Master Mix under recommended conditions (Applied Biosystems; #4364341). The following TaqMan gene expression assays (Applied Biosystems: #4331182) were used: Hs01120954_m1 (RCAN1); Hs00173626_m1 (VEGFA); Hs99999905_m1 (GAPDH). To detect K15 mRNA, primers 5′-CGGAAGAATCACGTGAAC-3′ (sense) and 5′-CGGTGTCTATACGGAAGG-3′ (antisense) and a dually labeled probe 5′-FAM-TCACCACAGCCAGACCAATCA-TAMRA-3′ were used. The average cycle-threshold value (Ct) for each individual amplification reaction was calculated from triplicate measurements by means of the instrument's software in “auto Ct” mode (7500 FAST System Software v.1.3.0).
For most siRNA transfection experiments, HUVECs were transduced with pSF91-IRES-GFP (control vector) or pSF91-K15-IRES-GFP. Thirty hours after transduction, cells were transfected with 100 pmol siRNA using the Neon transfection system (Invitrogen) according to the manufacturer's instructions. The following siRNAs (siGENOME SMARTpool) were ordered from Dharmacon: Control siRNA (Non-Targeting siRNA pool #1, D-001206-13-05), siRCAN1 (M-004268-02), siNFATc2 (M-003603-02), siPLCG1 (M-003559-01), siRELA/p65 (M-003533-02), siCalcineurin (PPP3CA) (M-008300-02), siFLT1 (M-003136-03), siFLT4 (M-003138-02) and siKDR (M-003148-01). To selectively target RCAN1.1, siRNA was designed to target the exon 1 of RCAN1 (GCUUCAUUGACUGCGAGAUU). The RCAN1.4 isoform was specifically targeted by siRNA against exon 4 of RCAN1 (AGUGAUAUCUUCAGCGAAAUU). siRNA against K15 protein was targeted to exon 8 of K15 (CAACCACCUUGGCAAUAAU) and was also purchased from Dharmacon.
HUVECs were transduced with pSF91-K15-IRES-GFP or pSF91-IRES-GFP retrovirus stocks. Thirty hours later cells were washed twice with PBS and then incubated in basal medium (EBM2+2% FBS) overnight to starve the cells. Growth factor-reduced Matrigel (BD Biosciences) was added to the wells of a prechilled 96-well plate (65 µl/well). The plate was placed in a 37°C incubator for 30 min and the matrigel was allowed to solidify. To determine whether HUVECs, untransduced or transduced with retroviral vectors, could form angiogenic tubes in the absence or presence of VEGF, the cells were resuspended either in basal medium (EBM2+2% FBS) or basal medium containing VEGF-A (50 ng/ml). 1.6×104 cells/ml (in 100 µl/well medium) were then plated on the solidified matrigel and incubated at 37°C for up to 6 hours. Images were taken with a Nikon T200 fluorescence microscope. The angiogenic index was calculated as the number of branch points in a visual field. Four different fields were counted for each treatment and the number of branch points was averaged. Error bars were calculated from means ± SD. For drug treatment experiments, cells were resuspended in medium containing 1 µM Cyclosporin A (Calcineurin inhibitor, Calbiochem) or 20 µM U73122 (PLCγ inhibitor, Calbiochem).
To investigate capillary tube formation in HUVECs harbouring the entire KSHV genome, HUVECs were infected with either a recombinant KSHV wildtype virus cloned in BAC (KSHVBACwt) or a K15 deletion mutant (KSHVBACΔK15) at an m.o.i. of 10 for 3 days to establish latency and 70–80% of GFP-expressing cells were obtained. In a similar way, HUVECs were infected with another recombinant KSHV virus (rKSHV.219) at an m.o.i of 10 for 72 hours and cells were then transfected with 100 pmol siRNA against K15 or control siRNA using the Neon transfection system. Twenty-four hours later the lytic cycle was activated with Na-Bu and a recombinant baculovirus expressing KSHV RTA (replication and transcription activator; see above) and 36 hours after induction of the lytic cycle, infected HUVECs were plated on matrigel and scored for capillary tube formation after 4–6 hours.
One day prior to transfection, 4.2×106 HEK293 cells were plated in a 10 cm dish. The cells were transiently transfected with retroviral vector plasmids pSF91-IRES-GFP, pSF91K15-IRES-GFP, pSF91K15/YF-IRES-GFP, pSF91K15/ΔSH3-IRES-GFP, pSF91K15/YF/ΔSH3-IRES-GFP (5 µg) using Fugene 6 transfection reagent (Promega) according to manufacturer's instructions. Forty-eight hours after transfection, cells were washed once in PBS and lysed in 500 µl of HEPES lysis buffer (20 mM HEPES, 150 mM NaCl, 0.5 mM EDTA, 1% Triton-X100, 1 mM DTT, 10% glycerol) containing protease and phosphatase inhibitors. 20 µl of the lysate was used as input control. Primary rabbit anti PLCγ1 (1∶50 dilution) was added to the rest of the lysate and incubated overnight with gentle shaking at 4°C. The next day the protein-antibody complexes were incubated with 20 µl of protein A Sepharose beads at 4°C with gentle shaking for 2–3 hours. Beads were washed three times with HEPES lysis buffer, and bound proteins were eluted and analyzed by Western blotting as described above, using a polyclonal antibody to K15. For co-immunoprecipitation in HEK293T cells, the expression vectors pFJ-K15P and pFJ-K15M [50], [52] were transfected with Fugene 6 transfection reagent and a polyclonal antibody to FLAG was used to detect the expression of K15.
For GST pulldown experiments, E. coli Rosetta cultures transformed with GST-K15 expression or GST plasmids [50], [51] were grown at 37°C in LB-medium plus ampicillin and chloramphenicol. Cultures were induced at an optical density (600 nm) of 0.4–0.6 with 1 mM isopropyl-β-D-thiogalactopyranoside (IPTG) and harvested by centrifugation 5 hours after induction. Cell pellets were resuspended in 500 µl PBS with protease inhibitors, sonicated for 1 min on ice, supplemented with 1% Triton-X100 and incubated for 1 hour at 4°C. After centrifugation the supernatant was incubated with 100 µl glutathione sepharose beads (Amersham Biosciences) overnight at 4°C. Beads were washed twice with TBS-T containing protease inhibitors and run on an SDS polyacrylamide gel stained with Coomassie blue to estimate equal amounts of GST fusion proteins. HEK293T cells were washed once with PBS and lysed with TBS-T containing protease inhibitors for 10 min on ice. 100 µl of cleared lysates were incubated with pre-calculated amounts of either beads coated with GST fusion proteins or GST alone overnight at 4°C. Beads were washed three times with TBS-T containing protease inhibitors and analysed by SDS-PAGE and Western blotting.
Statistical analysis was performed using GraphPad Prism software. For the comparison of more than two groups a Kruskal-Wallis test with Dunn's post-test was applied. In the corresponding graphs, statistical significance was indicated with asterisks: p-value less than 0.05 (*), 0.01 (**), or 0.001 (***), (ns: not significant). Error bars were calculated from means ±SD.
In previous studies we had shown that, in epithelial cells, transiently transfected K15 induces the expression of chemokines, cytokines, anti-apoptotic genes and genes involved in signaling, inflammation as well as angiogenesis [51], [52], [58]. We next wanted to investigate a possible role of K15 in inflammation and/or angiogenesis in KSHV-infected primary endothelial cells, the origin of Kaposi's sarcoma. To this end, we compared the cellular transcriptome of primary human umbilical vein endothelial cells (HUVECs) infected with a recombinant KSHVwt, derived from the BAC36 genome [55], [56] or a K15 deletion mutant of BAC36, KSHVΔK15. HUVECs isolated from 2 donors were infected with recombinant KSHVwt or KSHVΔK15 at an approximate m.o.i. of 10 titered on HEK293, thus achieving an infection rate of approx. 70% in HUVECs as judged by the percentage of GFP expressing cells. As a control, the same amount of virus was heat-inactivated and used to mock-infect HUVECs. In parallel, HUVECs were also transduced with a retroviral vector expressing a codon optimized K15 cDNA [51].
Data were filtered for mRNAs showing an at least two-fold induction by K15 overexpression relative to empty vector transduced cells in each of two independent experiments performed. 199 appropriately annotated transcripts fulfilled the applied filter criteria. The top-ranking 50 transcripts were presented as a heatmap (figure 1A, lanes 1–2). In lanes 3 and 4, the relative expression in KSHVwt versus KSHVΔK15 infected cells is depicted for the same set of genes. Only six of the genes induced by overexpressed K15 (lanes 1–2) i.e. pro-melanin concentrating enzyme (PMCH), H2.0 like homeo-box (HLX1), nuclear receptor family 4, subfamily A, group 1 (NR4A1), Regulator of Calcineurin 1/Down syndrome critical region 1 (RCAN1/DSCR1), v-myc myelocytomatosis viral related oncogene (MYCN) and ndrg family member 1 (NDRG1) were found to be induced upon KSHVwt infection and were differentially regulated between KSHVwt- and KSHVΔK15-infected HUVECs in both experiments when a ratio of 2.0 was used as a cut off (figure 1A, lanes 3–4; genes indicated by arrows). Applying the same selection criteria to all 199 K15-induced genes revealed a total of eight genes, impaired in induction when K15 is deleted from the viral genome. In addition to those mentioned above, these are EH-domain containing 3 (EHD3) and Rho GTPase activating protein 25 (ARHGAP25) (not shown).
RCAN1/DSCR1 has been shown to play a role in angiogenesis in tissue culture based models of vascular tube formation [41], [59] and in transgenic models [44], [46]. Located in the genomic region duplicated in patients with Down syndrome, it is thought to affect tumor angiogenesis. Expression of RCAN1/DSCR1 is induced by VEGF-A through a pathway involving PLCγ1, Calcineurin and NFAT1 and it acts as a feedback modulator on Calcineurin [42]. We had previously noted its increased expression after transient transfection of a K15 expression vector in epithelial cells [52], [58] and therefore proceeded to investigate its role in KSHV infection of endothelial cells.
We first confirmed the upregulation of RCAN1/DSCR1 by K15 in endothelial cells using quantitative PCR and the samples tested by gene expression microarray. To confirm the expression of K15 in cells infected with KSHVwt and KSHVΔK15, we performed quantitative RT-PCR on the same RNA and could detect the expression of K15 in KSHVwt-infected cells but not in cells infected with KSHVΔK15 (figure 1B). Expression of K15 in HUVECs from a retroviral vector resulted in a marked increase in the mRNA levels of RCAN1 (15 and 19 fold in two experiments; see figure 1C), while in KSHVwt-infected cells we observed an increase of 3.2 and 6.4 fold relative to cells exposed to heat inactivated virus (figure 1D). In contrast, cells infected with KSHVΔK15 showed no increase in both independent experiments (1.1 fold and 1.0 fold) (figure 1D). As RCAN1/DSCR1 is known to be upregulated by vascular endothelial growth factor (VEGF-A) we explored if K15 also induces VEGF gene expression. Only a minimal (2.8-fold) increase of VEGF-A mRNA from a low constitutive level was detectable in one of the two microarray experiments, while the second experiment did not show any significant change (not shown). These data were confirmed by quantitative PCR (figure 1C). Furthermore, no change of VEGF-A mRNA expression in response to KSHVwt-infection was seen in microarrays and confirmatory qPCR results (figure 1D). These observations could suggest that K15 might induce RCAN1/DSCR1 expression in a VEGF-independent manner. We therefore explored a role of K15 in RCAN1/DSCR1 expression and KSHV mediated angiogenesis. RCAN1/DSCR1 will be referred to as RCAN1 from this point onwards.
As a measure of its angiogenic properties, VEGF can induce the branching of endothelial cells and formation of capillary ‘tubes’ in primary endothelial cells plated on matrigel [41], [59]. This is illustrated in the two left panels of figure 2A and can be quantitated using an angiogenic index, based on the number of branch points counted in 4 different fields (see figure 2, central panel, bottom row and Materials & Methods). To investigate a possible role of K15 in capillary tube formation, HUVECs were transduced with the retroviral vector encoding K15 (pSF91K15) or the control vector (pSF91) and plated on matrigel 30 hours after transduction in the absence or presence of 50 ng/ml VEGF. As compared to untransduced cells and the cells transduced with the control vector, cells expressing K15 could form tubes even in the absence of exogenously supplied VEGF (figure 2A) and the angiogenic index showed a statistically significant increase for K15-transduced cells relative to the control vector-transduced cells in the absence of VEGF (figure 2B, 3B, 4C, 5A, 5C, 6C, 7C). Western blot analysis (figure 2C) showed that stimulation of HUVEC with VEGF-A increases the levels of both the RCAN1.1 and RCAN1.4 protein isoforms (lane 1 and 3 relative to lane 2 and 4). K15, on the other hand, could upregulate both isoforms of RCAN1 even in the absence of exogenously supplied VEGF (lane 6 relative to lanes 2,4). These experimental results are in accordance with a role for K15 in angiogenesis that involves increased expression of RCAN1 isoforms.
As other KSHV lytic proteins such as vGPCR and K1 have been reported to be involved in angiogenesis and to induce VEGF secretion [19], [20], [23], we next explored whether the induction of capillary tube formation by K15 could occur as a result of VEGF secretion. Fresh HUVECs were treated with supernatant from HUVECs transduced with the retroviral vector for K15 or the control vector and were then assessed for their ability to form capillary tubes on matrigel. While exogenously added VEGF induces capillary tube formation, conditioned supernatants from K15 or control vector, collected after 48 hours, did not (figure 2D). Taken together with the fact that expression of K15 in primary endothelial cells does not consistently induce expression of the VEGF-A gene (figure 1), this observation is compatible with the interpretation that K15 does not increase RCAN1 expression nor induces capillary tube formation by augmenting VEGF secretion. Additional evidence in support of this interpretation is provided below. This experiment does not exclude other paracrine effects mediated by K15, e.g. as a result of its induction of IL-8 secretion [52], [58].
To further investigate if the K15-induced capillary tube formation in the absence of VEGF requires RCAN1 expression, we targeted RCAN1 expression in both control vector and K15-transduced cells using a pool of RCAN1 siRNA that recognizes all the known isoforms of RCAN1. In HUVECs transduced with a GFP expressing retroviral vector pSF91, knockdown of RCAN1 significantly abrogates VEGF-induced tube formation in contrast to cells transfected with a scrambled siRNA (not shown) in line with earlier reports that RCAN1 is required for VEGF-induced capillary tube formation in HUVECs [41], [42]. Knockdown of RCAN1 also abrogated K15-induced capillary tube formation in the absence of VEGF (figure 3A and 3B). The Western blot in figure 3C confirms the knockdown of RCAN1. We next selectively abrogated the expression of RCAN1 isoforms, RCAN1.1 and RCAN1.4, with specific siRNA. As shown in figure 3A–C, expression of RCAN1.1 and RCAN 1.4 could be reduced selectively, and the absence of either isoform reduced the ability of K15 to induce angiogenic tube formation. Together, these experiments suggest that K15 plays an important role in the induction of capillary tube formation that might be dependent on both RCAN1 isoforms.
Previous reports have shown that VEGF-A regulates RCAN1.4 expression by activation of the classical Ca2+/Calcineurin pathway leading to activation of the transcription factor NFAT in endothelial cells [41]. Following VEGF-A stimulation, phosphorylation of the tyrosine 1175 site in VEGFR-2 allows the binding and subsequent phosphorylation of PLCγ1. PLCγ1 is then able to catalyze the hydrolysis of the membrane phospholipid phosphatidylinositol (4,5)-bisphosphate (PIP2) resulting in the generation of diacylglycerol (DAG) and inositol 1,4,5-trisphosphate (IP3). DAG is a physiological activator of PKC, whilst IP3 acts upon the endoplasmic reticulum to release calcium, thus inducing a rise in intracellular calcium (Ca2+). Increased Ca2+ levels activate the phosphatase Calcineurin, which dephosphorylates members of NFAT family. This dephosphorylation allows the translocation of NFAT into the nucleus, where it binds to NFAT consensus sequences present in the promoter region of several genes involved in angiogenesis and cellular migration, such as RCAN1 or COX-2, resulting in their increased expression. Since our findings so far suggest that K15 mimics the effect of VEGF in inducing the expression of RCAN1 and increasing capillary tube formation, and as we had observed earlier that K15 activates an NFAT-responsive promoter [52], we further investigated whether the Calcineurin-NFAT pathway was involved in the upregulation of RCAN1.
As expected from previous reports [42], [59], we could inhibit VEGF-induced capillary tube formation in HUVEC (figure 4A), NFAT1 dephosphorylation and the resulting increased RCAN1.1/1.4 expression (figure 4B) with chemical inhibitors of Calcineurin (cyclosporin A; CsA) and PLCγ (U73122). In addition, as shown in figure 4C, K15-induced capillary tube formation was significantly abrogated in the presence of CsA and U73122, suggesting that it follows the same pathway as VEGF. Western blot analysis (figure 4D) showed that K15 induced PLCγ1 phosphorylation and dephosphorylation of NFAT1. Treatment with CsA and U73122 reduced the levels of K15-induced RCAN1.1/1.4 expression, abolished dephosphorylation of NFAT1 and (in the case of U73122) reduced levels of PLCγ1 phosphorylated on Tyr 783.
To further investigate the involvement of Calcineurin, NFAT1 and PLCγ1 in K15-induced tube formation we targeted their expression using siRNA. As shown in figure 5A and 5B, suppressing protein levels of Calcineurin and NFAT1 significantly abrogated angiogenic tube formation induced by VEGF, as well as by K15 in the absence of VEGF. Western blots of the same samples (figure 5B) showed again that K15 expression led to an enhanced dephosphorylation of NFAT1 and that this effect was abolished by silencing the expression of Calcineurin by means of siRNA. Silencing of NFAT1 and Calcineurin also reduced K15-induced RCAN1.1 and RCAN1.4 levels (figure 5B).
Having observed that expression of K15 results in an increased phosphorylation of PLCγ1 and that a chemical inhibitor of PLCγ (U73122) reduced K15-induced capillary tube formation (figure 4C and 4D), we next examined if silencing of PLCγ1 by siRNA would also antagonize these effects of K15. As shown in figure 5C, knockdown of PLCγ1 decreased VEGF-induced capillary tube formation in the absence of K15 (pSF91-transduced cells in figure 5C), as well as K15-induced capillary tube formation (figure 5C) and RCAN1.1/1.4 expression (figure 5D) in the absence of VEGF.
We have shown earlier that K15, through its cytoplasmic C terminal domain activates the NFκB pathway [52], [58], [60] and that this involves the NFκB component RelA/p65 (unpublished). To investigate the involvement of the NFκB pathway in K15-induced capillary tube formation, we silenced RelA/p65 using siRNA. As shown in figure 5C and 5D, silencing of RelA/p65 does not antagonize RCAN1 expression and capillary tube formation induced by K15. Taken together, these results suggest that the activation of the PLCγ1-Calcineurin-NFAT pathway is required for the K15-induced tube formation and that this effect is independent of the NFκB pathway.
We next studied the kinetics of the activation of the PLCγ1 pathway by K15. We found that, in contrast to VEGF, which induces a transient increase in PLCγ1 phosphorylation that peaks at 5–10 minutes and disappears after approx 30 minutes (figure 6A), the K15-induced PLCγ1 phosphorylation lasted for up to 48 hours (figure 6B). Treatment with VEGF results in a rapid increase of RCAN1.1/1.4 protein levels, which however decrease after 30 minutes, despite the continuous presence of VEGF in this experiment. This suggests a transient VEGF-dependent signaling and rapid turnover of RCAN1 proteins (figure 6A). In contrast, in K15-transduced cells RCAN1.1/1.4 levels increased in parallel with K15 protein levels and remained high, along with levels of phosphorylated PLCγ1, for the duration of this experiment (48 hours) (figure 6B). These results indicate that VEGF induces a short pulse of PLCγ1 activation, which leads to a transient expression of RCAN1. K15, on the other hand, as seen in figure 6B, activates PLCγ1 and RCAN1 expression in a protracted constitutive manner.
We also explored if the effects of VEGF and K15 on capillary tube formation are additive. As shown in figure 6C, increasing concentrations of VEGF led to a gradual increase in angiogenic tube formation in vector-transduced, but not in K15-transduced cells. This observation suggests that, in K15 expressing endothelial cells, the PLCγ1-Calcineurin-NFAT pathway is maximally activated and cannot be further stimulated by exogenous VEGF.
As reported earlier, K15 is a trans-membrane protein containing eight predicted trans-membrane regions and a C- terminal cytoplasmic domain [47], [48], [50], [52], [58]. The cytoplasmic domain of the P-type of K15 used here contains two SH2-binding sites (Y431ASIL and Y481EEVL), a proline-rich SH3-binding site (P387PLP) and a TRAF-binding site (A473TQPTDD) [50], [52]. The Y481EEVL motif has been shown to activate various signaling pathways such as MAPK, JNK, NFκB [47], [50], [52]. To test which of these domains is involved in the activation of the PLCγ1 pathway, we used previously reported mutants of SH2 and SH3 binding motifs [51] in which the tyrosine at position 481 had been changed to phenylalanine (YEEVL to FEEVL), or all the prolines in the SH3 binding motif had been changed to alanine (PPLP to AALA). We also employed a double mutant in which both of these domains had been mutated [51] (figure 7A). We tested these mutants for their ability to induce capillary tube formation and to activate PLCγ1 phosphorylation. As shown in figure 7B, wildtype K15 and the ΔSH3 mutant could induce PLCγ1 phosphorylation and increase RCAN1.1/1.4 expression, whereas the tyrosine 481 (YF) mutant of K15 could not. As shown in figure 7C, mutants K15/YF and K15/YF/ΔSH3 could not induce capillary tube formation, while mutant K15ΔSH3 showed only a moderately (not significantly) lower angiogenic index than K15wt. This observation suggests that the tyrosine residue at position 481 in the YEEV motif is important for activation of the PLCγ1 pathway, induction of RCAN1 and capillary tube formation.
VEGF-induced activation of PLCγ1 in endothelial cells involves recruitment of PLCγ1 to the phosphorylated cytoplasmic tail (on tyrosine 1175) of VEGFR2 and subsequent phosphorylation of PLCγ1 on tyrosine 783 [39]. Since we had found that activation of PLCγ1 by K15 requires the K15 Y481 residue, which had been previously shown to be phosphorylated by Src family tyrosine kinases [47], [52], we next investigated if K15 can directly recruit PLCγ1 via its Y481EEV SH2 binding site. To test this hypothesis, we performed co-immunoprecipitation experiments in HUVECs transduced with K15 and control vector. Endogenous PLCγ1 was immunoprecipitated with an antibody to PLCγ1 and immunoprecipitates were probed on Western blots with an antibody for K15. We found that K15 can interact with PLCγ1 (figure 8A), and that the K15YF mutant and the K15ΔSH3 mutants showed a reduced interaction with PLCγ1. The K15 double mutant (K15YF/ΔSH3) showed no interaction. This suggests that K15 interacts with PLCγ1, and that the SH2 binding site (Y481EEV) as well as the SH3 binding site (P387LPP) contribute to the interaction. The interaction between the cytoplasmic domain of K15 and endogenous PLCγ1 was also confirmed in a GST pull-down assay using the cytoplasmic domain of K15 fused to GST (figure 8B). In this assay, the Y481F and ΔSH3 mutants also showed a decreased interaction while virtually no interaction could be seen in the case of the double mutant (YFΔSH3).
The K15 gene occurs in several allelic isoforms, thought to originate from recombination events with related rhadinoviruses [48], [49], [58]. In the experiments described so far, we had used the ‘predominant’ P allele of K15 (K15-P). To explore if other K15 isoforms also recruit PLCγ1, we tested the most divergent isoform, the M-type of K15 (K15-M). As shown in figure 8C, D, both K15-P and K15-M co-immunoprecipitate with PLCγ1 and interact with PLCγ1 in a GST pull-down assay. In addition, transfection of both K15-P and K15-M induces phosphorylation of PLCγ1 on tyrosine 783 (figure 8E). As K15-M is more divergent from K15-P than other isoforms [61], it is likely that all K15 isoforms are able to activate PLCγ1.
To exclude that low levels of VEGF might contribute to K15-dependent PLCγ1 activation by activating VEGF receptors that could, in turn, phosphorylate PLCγ1, we used siRNA to silence the expression of all three VEGF receptors. As shown in figure 9A, capillary tube formation in response to VEGF treatment is completely abrogated when the expression of KDR/VEGFR2 is silenced in HUVECs transduced with the parental retroviral vector, while no significant reduction is seen following silencing of the other two VEGF receptors, FLT1/VEGFR1 and FLT4/VEGFR3. This is in line with previous reports that VEGFR2 (KDR) is the prime receptor for VEGF-induced pathways in HUVECs [40]. In contrast, in K15-transduced cells, the K15-induced capillary tube formation in the absence of VEGF is not reduced when KDR/VEGFR2 is silenced suggesting that the K15 mediated activation of PLCγ1 is independent of VEGFR2. Likewise, neither FLT1/VEGFR1 nor FLT4/VEGFR3 appear to be involved in K15-induced capillary tube formation (figure 9A, B). This supports our previous result (figure 1C, 2D and 6C) that K15-induced tube formation is independent of VEGF.
To explore if Src family kinase members are involved in the K15-mediated activation of PLCγ1, we treated K15-transduced HUVECs with the Src family kinase inhibitors Su6656, PP1, PP2 and MNS. We could not observe a strong reduction of K15-induced PLCγ1 phosphorylation or angiogenic tube formation in these experiments (data not shown). The identity of the tyrosine kinase involved in the phosphorylation of K15 tyrosine 481 (required for PLCγ1 docking) or PLCγ1 tyrosine 783 therefore remains to be established.
Infection of primary endothelial cells with KSHV has previously been shown to induce the formation of capillary tubes when infected cells are plated on matrigel. Several other KSHV proteins including K1, vGPCR, and vIL6 have been reported to have angiogenic effects and/or induce the secretion of VEGF [19], [20], [25], [29]. As we had found that the increased expression of RCAN1, a mediator of VEGF-mediated angiogenesis [41], in KSHV-infected cells depends on the presence of K15 in the viral genome (figure 1), we next explored if K15 contributes to the angiogenic properties of KSHV in KSHV-infected cells. We infected HUVECs with a recombinant KSHV virus (rKSHV.219) for 3 days to obtain 70–80% GFP expressing cells before microporating these cells with siRNA against K15. Twenty-four hours later the lytic replication cycle was activated with Na-Bu and a recombinant baculovirus expressing KSHV RTA (regulator of transcriptional activator) and 36 hours after activation of the lytic cycle, infected HUVECs were plated on matrigel and scored for capillary tube formation after 6 hours. As shown in figure 10A, capillary tube formation is visible in cells transfected with control siRNA following the induction of the lytic cycle. In contrast, angiogenic tube formation is reduced upon silencing of K15, suggesting a role of K15 in virus-induced angiogenesis. Figure 10B shows that both the increase of the angiogenic index upon activation of the lytic cycle, as well as its decrease upon silencing of K15, are statistically significant. The top panel of figure 10C shows the effective silencing of K15 in this experiment and its increased expression following induction of the lytic cycle in cells treated with control siRNA. As reported previously [52], the upper band seen in the western blot is not affected by K15 siRNA, while the lower band represents the 43 kDa K15 protein seen in KSHV infected cells. Expression of the two isoforms of the KSHV structural K8.1 glycoprotein is used as a marker for lytic cycle induction (panels 2,3 of figure 10C). As shown in panel 4 of figure 10C, induction of the lytic replication cycle results in increased PLCγ1 phosphorylation, which, along with capillary tube formation, is strongly suppressed by silencing of K15. In addition, activation of the lytic replication cycle and increased PLCγ1 phosphorylation is accompanied by NFAT1 dephosphorylation and increased levels of RCAN1.1 and RCAN1.4, which are reversed by silencing of K15.
In an additional experiment, we compared the ability of BAC-derived KSHVwt and KSHVΔK15 to induce angiogenic tube formation. We infected HUVEC with cell-free KSHVwt and KSHVΔK15 to obtain approximately 60% of GFP expressing cells. As shown in figure 10 D, E, BAC-derived KSHV induced moderate angiogenic tube formation in latently infected compared to uninfected cells, and more extensive angiogenic tube formation following activation of the lytic cycle, whereas the angiogenic index was reduced in KSHVΔK15- relative to KSHVwt-infected lytically induced cells. Western blots (figure 10F) confirmed the absence of K15 protein in KSHVΔK15- compared to KSHVwt-infected cells (top panel) and comparable levels of lytic glycoprotein K8.1 expression in KSHVwt and KSHVΔK15-infected cells following lytic cycle activation. These results suggest that the increased formation of capillary tubes that occurs in KSHV infected endothelial cells depends to a significant extent on K15 and the K15-dependent activation of the PLCγ1-NFAT1-RCAN1 pathway.
Kaposi's Sarcoma Herpesvirus is the cause of Kaposi's Sarcoma (KS), a vascular tumor with pronounced inflammatory histological features that arises from infected endothelial cells. Early stages of KS (“patch, plaque”) are characterized by inflammatory infiltration and features of aberrant angiogenesis [4], [5], [62]. In advanced lesions (nodular KS), KSHV-infected endothelial spindle cells predominate and are thought to represent the neoplastic component of this tumor [4], [5], [62]. Unlike most cancers, KS represents, in most cases, an oligo- or polyclonal proliferation of infected endothelial cells [63], [64], suggesting that independent infection events, rather than the spread of a clonal tumor are responsible for the dissemination of KS to different organs and body sites [63], [64]. KS lesions may regress following the treatment of HIV with anti-retroviral therapy, or moderating iatrogenic immune suppression (in transplant recipients), indicating that this tumor requires the continued presence of KSHV and the expression of viral genes. Thus KS resembles other virus-driven proliferative diseases such as, in particular, EBV-associated post-transplant lymphoproliferative disease (PTLD) (reviewed in [4], [5], [62]).
KSHV induces a transcriptional reprogramming of vascular endothelial cells towards a pattern typical of lymphatic endothelial cells, as well as inducing cellular genes typical of vascular endothelial cells in infected lymphatic endothelial cells [15]–[17] In addition, KSHV causes endothelial to mesenchymal transition (EndoMT) in infected lymphatic endothelial cells, thereby accounting for the expression of mesenchymal markers frequently observed on KS spindle cells [65]. Both the latent viral protein vFLIP and the lytic protein vGPCR appear to contribute to this phenotype [65]. In cultured primary endothelial cells, KSHV can induce the formation of capillary junctions when infected cells are plated on matrigel [18]. It is likely that the KSHV-induced formation of capillary junctions may explain the aberrant angiogenesis seen in KS lesions, rather than the formation of spindle cells, which is mainly due to the latent vFLIP protein [13], [14].
In this study we wanted to define a possible role of the KSHV K15 protein in KSHV infected spindle cells. Starting from an earlier observation [52], that transfection of K15 into epithelial cells induced the expression of cellular genes that are known to be activated by VEGF, but did not appear to induce VEGF expression itself, we compared the effect of expressing K15 in primary endothelial cells on cellular gene expression with the pattern of cellular genes induced by infecting endothelial cells with KSHV, or a KSHV K15 deletion mutant. Only a small subset of cellular genes that can be induced by over-expressed K15 appear to be differentially regulated in KSHVwt- and KSHVΔK15-infected endothelial cells (figure 1). This likely reflects the fact that K15 is only weakly expressed in infected endothelial cells prior to the activation of the lytic cycle (figure 10F) and is reminiscent of our recent observation that only a small selection of cellular genes activated by over-expressed vFLIP are measurably increased in endothelial cells infected with KSHVwt compared to a deletion mutant of vFLIP, KSHVΔvFLIP [14].
Cellular genes that are differentially regulated between endothelial cells infected with KSHVwt and a KSHV deletion mutant are likely to be those whose expression is most strongly affected by the deleted viral gene in the context of the entire virus genome. In this study we found RCAN1/DSCR1, a regulator of the Calcineurin-NFAT pathway, to be differentially expressed in endothelial cells infected with KSHVwt and KSHVΔK15. The activation of RCAN1 by K15 in KSHVwt- but not in KSHVΔK15-infected endothelial cells, as well as in endothelial cells transduced with K15, is not accompanied by a consistently increased expression of VEGF family members (figure 1A–D).
Expression of RCAN1 is normally induced by VEGF (figure 2C) and is required for VEGF-induced angiogenic tube formation (figure 2A,B). Similarly, RCAN1 expression is required for K15-induced angiogenesis in the absence of exogenously added VEGF (figure 3B,C). Although the expression of RCAN1 is normally regulated by the binding of VEGF to its receptor, the subsequent activation (phosphorylation) of PLCγ1 and the ensuing influx of Ca2+, our observations suggest that the K15-induced expression of RCAN1 and angiogenic tube formation is not dependent on VEGF. The expression of VEGF-A is not consistently increased by overexpressed K15 whereas, the expression of RCAN1 is (figure 1C). Conditioned supernatants from K15 transduced cells do not induce the angiogenic tube formation caused by VEGF (figure 2D) and the silencing of any of the three VEGF receptors FLT1, KDR, FLT4 by siRNA does not reduce K15-induced angiogenic tube formation while VEGF-induced angiogenesis is inhibited by silencing KDR (figure 9A,B). We therefore conclude that K15-induced angiogenesis is independent of VEGF and VEGF receptors, in spite of K15-induced angiogenesis requiring several of the downstream components of the VEGF signaling pathway.
Like VEGF, expression of K15 induces the dephosphorylation of NFAT (figure 4) and this requires PLCγ1 and Calcineurin, since small molecule inhibition of PLCγ1 (U73122) and Calcineurin (cyclosporin A) (figure 4), as well as the siRNA-mediated silencing of PLCγ1, Calcineurin and NFAT1 (figure 5) inhibit K15-induced RCAN1 expression and K15-dependent angiogenesis. Both RCAN1 isoforms RCAN1.1 and RCAN1.4 appear to contribute to K15-induced angiogenic tube formation. These results suggest that K15 activates a PLCγ1-Calcineurin-NFAT1-dependent pathway, normally utilized by VEGF. In contrast to VEGF, which induces a short burst of PLCγ1 phosphorylation and a transient RCAN1 expression, K15 activates PLCγ1 phosphorylation and RCAN1 expression in a protracted manner (figure 6). Whether this is due to K15 not being subject to the same negative feedback regulation that operates in the case of VEGFR-dependent signaling, or to other mechanisms, remains to be established. Whatever the molecular basis of this observation, a constitutive activation of PLCγ1 by K15 could contribute to the sustained angiogenic phenotype seen in KS lesions. In K15-expressing cells, the PLCγ1-Calcineurin-NFAT pathway seems to be fully activated and cannot be further enhanced by exogenous VEGF, in keeping with the interpretation that K15 and VEGF use the same pathway.
In an attempt to understand how K15 would achieve the activation of this VEGF-dependent pathway without involving VEGF or VEGF receptors, we found that K15, a membrane protein, can directly interact with PLCγ1 (figure 8A,B). Both allelic isoforms of K15, K15-P and K15-M, are able to recruit and activate PLCγ1 (figure 8C–E), suggesting that KSHV isolates with divergent K15 variants are able to activate angiogenesis through the mechanism described in this report. An SH2 binding site in the cytoplasmic domain of K15, YEEVL, is required for K15-induced phosphorylation of PLCγ1 (figure 7B), and this site, as well as an SH3-binding site in the same cytoplasmic K15 region, contributes to the recruitment of PLCγ1 to K15 (figure 8A,B). It is therefore conceivable, but remains to be shown, that SH2 and SH3 domains, as they occur in the γ specific array of PLCγ1, may be involved in the recruitment of PLCγ1 to K15. Unlike the VEGF receptors, K15 is not known to have tyrosine kinase activity. Since silencing of VEGF receptors does not affect K15-induced angiogenesis and since several Src kinase inhibitors did not show a strong effect on K15-induced PLCγ1 phosphorylation and angiogenic tube formation (data not shown), the phosphorylation of PLCγ1 recruited to K15 most likely involves another, yet to be determined, tyrosine kinase.
The ability of K15 to induce the phosphorylation of PLCγ1 and to induce angiogenic tube formation is relevant in virus-infected cells, since silencing of K15 by siRNA in HUVEC infected with recombinant KSHV, KSHV.219, inhibited phosphorylation of PLCγ1, NFAT1 dephosphorylation, RCAN1.1/1.4 expression and angiogenic tube formation following activation of the lytic cycle (figure 10). Furthermore, KSHVΔK15-infected HUVECs showed a significantly reduced angiogenic tube formation compared to KSHVwt-infected cells (figure 10 D–F). We therefore conclude that, in addition to previously reported viral mediators of angiogenesis such as K1 and vGPCR, the membrane protein K15 contributes to KSHV-induced angiogenesis. We acknowledge the limitations of this tissue culture based angiogenesis assay, which is however, widely used to study the effects of VEGF on angiogenesis. Recruitment and activation of PLCγ1 by K15 may represent an upstream signaling event that not only feeds into IP3/Ca2+ influx-mediated activation of Calcineurin and subsequent NFAT activation (figure 11), but could also be responsible, via DAG (figure 11), for the previously reported activation of the MEK/ERK and JNK pathways by K15 [50], [52].
Although we can detect K15 expression at the mRNA (figure 1) and protein (figure 10) level in KSHV-infected endothelial cells prior to the induction of the lytic cycle, we only observed moderate, or no, angiogenic tube formation in HUVECs that were infected with, respectively, BAC36-derived KSHVwt (figure 10 D, E) or rKSHV.219 (figure 10 A, B) prior to the activation of the lytic cycle. In contrast, we observed significant angiogenic tube formation upon activation of the lytic cycle along with increased K15 expression (figure 10). We therefore assessed the impact of K15, using siRNA and a KSHV K15 deletion mutant, in infected cells following activation of the lytic cycle in order to obtain more robust measurements. Others [18], [66] have reported the formation of angiogenic tubes in latently infected primary or immortalized endothelial cells. We believe that these differences are due to slight variations in the experimental protocols used: in our experiments, HUVECs were infected with rKSHV.219, BAC36-derived KSHVwt or KSHVΔK15 for 3–5 days before measuring angiogenic tube formation, since, in our experience, a significant number of ‘lytic’ cells (as assessed by RFP expression in the case of the rKSHV.219 virus [57]) is seen during the first 1–2 days after infection. We note that in the report by Sharma-Walia et al. [66], angiogenic tube formation was assessed 24 hours after infection. We attribute the fact that we observed moderate angiogenic tube formation in HUVECs infected with BAC36-derived KSHVwt prior to the induction of the lytic cycle, but did not in cells infected with rKSHV.219, to differences in the virus preparations used: owing to the low yield of BAC36 producer cell lines, these virus preparations have to be concentrated (ultracentrifugation) to a higher degree and may have resulted in a higher proportion of lytically infected cells three days after infection.
This raises the question whether K15 exerts its impact on angiogenesis in KS lesions in latently infected cells, cells that show a restricted pattern of lytic gene expression, or cells undergoing the full productive replication cycle. The K15 locus in the KSHV genome appears to be occupied by ‘active’ chromatin characterized by specific histone acetylation and methylation patterns (H3K9/K14ac, H3K4me3) [67] and K15 was originally described as expressed during latency, but upregulated during lytic replication [48]. This suggests that ‘basal’ expression of K15 in infected cells that do not undergo full lytic replication may be possible and that it is this type of cells with a ‘restricted’ lytic program that contributes to KSHV-mediated angiogenesis. K15 may therefore exert its angiogenic effect in the early stages of KS (when aberrant angiogenesis rather than spindle cells dominate the histology and limited lytic gene expression may play a role) or in a subpopulation of cells showing a restricted pattern of lytic gene expression. Answering this question would require the development of new reagents, in particular of monoclonal antibodies capable of detecting K15 protein in paraffin-embedded histology sections with high sensitivity and specificity.
Our observation that K15 ‘usurps’ the PLCγ1-Calcineurin-NFAT pathway and renders K15-expressing cells refractory to additional stimulation by VEGF (figure 6C) does not argue against an effect of VEGF, induced by other viral proteins, on KSHV-infected or neighbouring cells that do not express K15. Our results therefore do not exclude other, paracrine effects mediated by KSHV and K15, e.g. via K15-induced IL8 secretion [52], [58].
In summary our findings suggest that recruitment and activation of PLCγ1 to the SH2 and SH3- binding sites of K15 may represent an important facet in KSHV-induced angiogenesis and a putative target for the development of inhibitors that do not interfere with the physiological activation of PLCγ1 and that would offer the opportunity to inhibit the angiogenic effects of this virus.
|
10.1371/journal.pbio.0050246 | Chromatin Structure Regulates Gene Conversion | Homology-directed repair is a powerful mechanism for maintaining and altering genomic structure. We asked how chromatin structure contributes to the use of homologous sequences as donors for repair using the chicken B cell line DT40 as a model. In DT40, immunoglobulin genes undergo regulated sequence diversification by gene conversion templated by pseudogene donors. We found that the immunoglobulin Vλ pseudogene array is characterized by histone modifications associated with active chromatin. We directly demonstrated the importance of chromatin structure for gene conversion, using a regulatable experimental system in which the heterochromatin protein HP1 (Drosophila melanogaster Su[var]205), expressed as a fusion to Escherichia coli lactose repressor, is tethered to polymerized lactose operators integrated within the pseudo-Vλ donor array. Tethered HP1 diminished histone acetylation within the pseudo-Vλ array, and altered the outcome of Vλ diversification, so that nontemplated mutations rather than templated mutations predominated. Thus, chromatin structure regulates homology-directed repair. These results suggest that histone modifications may contribute to maintaining genomic stability by preventing recombination between repetitive sequences.
| Homologous recombination promotes genetic exchange between regions containing identical or highly related sequences. This is useful in repairing damaged DNA, or in reassorting genes in meiosis, but uncontrolled homologous recombination can create genomic instability. Chromosomes are made up of a complex of DNA and protein, called chromatin. DNA within chromatin is packed tightly in order to fit the entire genome inside a cell; but chromatin structure may become relaxed to allow access to enzymes that regulate gene expression, transcribe genes into mesenger RNA, or carry out gene replication. We asked if chromatin packing regulates homologous recombination. To do this, we tethered a factor associated with compact chromatin, called HP1, adjacent to an immunoglobulin gene locus at which homologous recombination occurs constitutively, in order to produce a diverse repertoire of antibodies. We found that the compact, repressive chromatin structure produced by HP1 prevents homologous recombination. This finding suggests that regulated changes in chromatin structure may contribute to maintaining genomic stability by preventing recombination between repetitive sequences.
| Homologous recombination provides a pathway for restoring or altering DNA sequence and structure [1–7]. Homologous recombination can recreate the original DNA sequence at a DNA break, and predominates in S/G2 phases of cell cycle, when sister chromatids can serve as donors for faithful repair [8,9]. Homologous recombination can also have a mutagenic outcome by promoting recombination between nonallelic repeated sequences, leading to genomic instability, or by templating repair from a homolog rather than a sister chromatid, leading to loss of heterozygosity (LOH). In a living cell, multiple pathways compete to repair the same kinds of damage. For example, double-strand breaks (DSBs) can be repaired by nonhomologous end-joining, which may be accompanied by sequence loss or translocation [5,10,11]. Nicks can be efficiently repaired in situ, or by short- or long-patch repair pathways that use the complementary strand as a template [12,13].
Chromatin structure plays an important role in repair at the site of DNA damage. A critical signal for DSB repair is C-terminal phosphorylation of the variant histone H2AX by ATM and ATR, to generate γ-H2AX [14–16]. γ-H2AX is recruited to the break and extends over a large region surrounding the break site, creating a boundary of modified chromatin, and recruits the cohesin complex to sites of damage, to promote DSB repair using the sister chromatid as a template [17,18]. γ-H2AX alerts DNA damage checkpoints, and is recognized by histone acetyltransferases and chromatin remodeling complexes [19–22]. Changes in chromatin structure also facilitate synapsis of severed DNA ends for nonhomologous end-joining [23].
Homologous recombination involves two DNA molecules, the recipient, which is the site of the DNA lesion, and the donor. Two lines of evidence suggest that donor chromatin structure may contribute to homologous recombination. At the yeast mating type locus, changes in histone acetylation occur at the donor locus that are distinct from those at or near the DNA break [24]. In human cells, transcription of a donor promotes its use in gene conversion [25]. However, the role of donor chromatin structure in regulating recombination has not been directly tested in vertebrate cells.
Immunoglobulin (Ig) gene diversification in chicken B cells provides a powerful model for studying homologous recombination. Chickens have a limited number of functional heavy and light chain variable (V) regions, which undergo V(D)J recombination early in B cell development [26,27]. The rearranged V genes then undergo sequence diversification by gene conversion, using an array of homeologous upstream pseudo-V (ψV) regions as donors (Figure 1A). The ψV regions are nonfunctional, as they lack promoters and cannot be transcribed. The mechanism of Ig gene conversion is readily studied in the DT40 cell line, which derives from a bursal lymphoma and constitutively diversifies its Ig heavy and Ig light chain (Igλ) genes by gene conversion [28–30]. DT40 also supports very high levels of homologous gene targeting, thought to reflect elevated expression or activity of factors that promote recombinational repair [31–33].
Ig gene conversion in DT40 depends upon ubiquitous and conserved factors. The substrate for repair is a nick produced by successive action of three factors. The B-cell-specific enzyme activation-induced deaminase (AID) [34–37] deaminates cytosine to uracil in transcribed and targeted genes [38–41]; uracil DNA glycosylase excises the uracil produced by AID to generate an abasic site [42–46]; and the MRE11 abasic lyase, functioning within the MRE11/RAD50/NBS1 complex, nicks at the abasic site [47,48]. Strand transfer and new DNA synthesis are carried out by ubiquitous DNA repair factors including the RAD51 paralogs, BRCA2, FANCC, FANCD2, and polη [49–55]. Deficiencies in some of these factors, particularly the RAD51 paralogs, or targeted deletion of some or all of the ψV donors [56], impair gene conversion and can contribute to a shift in the processing of AID-initiated breaks so that templated repair is accompanied or even supplanted by nontemplated mutagenesis. ψV regions preferentially used as donors are in opposite orientation to the functional V region, suggesting that local chromosomal architecture may guide templated repair [57]. However, nothing is known about how epigenetic features of the donors affect recombination.
To understand how donor chromatin structure affects gene conversion in particular, and homologous recombination more generally, we characterized and experimentally manipulated chromatin structure at the Igλ locus in the chicken B cell line DT40. We found that the ψVλ donors contain acetylated histones, consistent with an open chromatin structure. To test whether this reflects requirements of gene conversion, we tethered HP1 (Drosophila melanogaster Su[var]205) to the ψVλ array in a DT40 derivative in which polymerized lactose operator (PolyLacO) has been inserted into that chromosomal region. HP1 is known to promote heterochromatic gene silencing [58–60]. Tethered HP1 caused a local transition of the donor sequences from an open to a nonpermissive state, and a switch from templated to nontemplated diversification, evident as point mutations. These observations demonstrate that permissive chromatin structure at the donor is a key regulator of gene conversion, and that nonpermissive chromatin structure can prevent homologous recombination and result in point mutagenesis. These results have implications for our understanding of homologous recombination and of the mechanisms that promote LOH, leading to tumorigenesis and nonallelic recombination between repeats. These results should also inform design of donor constructs for targeted gene therapy.
In DT40 B cells, the Vλ gene is rearranged and expressed at one Igλ allele, but it is unrearranged at the other allele. We characterized chromatin structure at the rearranged (VλR) and unrearranged (VλU) alleles and the ψVλ array by chromatin immunoprecipitation (ChIP). ChIP was carried out with antibodies specific for lysine acetylation at the N-termini of histones H3 and H4. Recovered DNA was amplified in duplex PCR reactions; recovery was normalized to an amplicon from the ovalbumin (Ova) gene, which is not expressed in B cells; and enrichment was normalized to a total DNA input control (see Materials and Methods for details). The distinct genomic structure of VλR and VλU permit them to be distinguished by PCR with specific primers. ChIP demonstrated considerable enrichment of acetylated histones H3 and H4 (AcH3 and AcH4) at the rearranged VλR gene. In a typical experiment, AcH3 was enriched more than 80-fold at VλR, and AcH4 more than 30-fold (Figure 1B). In contrast, at the VλU allele, the levels of AcH3 and AcH4 were much lower than at VλR (16-fold and 7-fold lower, respectively), and only a few fold enriched relative to input DNA.
Chromatin structure within the ψVλ array was assayed by amplification with primers that interrogated seven sites, including a region between ψVλ1 and the Vλ gene, ψVλ1, ψVλ5, ψVλ13, ψVλ18, ψVλ24, ψVλ25, and the upstream flanking region. (Because of a paucity of polymorphisms, the ψVλ arrays at the two Igλ alleles in DT40 cannot be readily distinguished by PCR.) Strikingly, we observed considerable enrichment of AcH3 and AcH4 throughout the ψVλ array (Figure 1B). Enrichment was not proportional to distance from the transcribed VλR gene, as sites distant from VλR did not consistently display lower levels of enrichment than proximal sites (Figure 1B). Thus, enrichment of acetylated histones within the ψVλ array does not simply represent a graded spreading of chromatin modification from the transcribed VλR gene to sites upstream. The nonuniform chromatin structure of the locus suggests the presence of cis-elements that regulate chromatin structure at the ψVλ array.
Local modification of chromatin structure can be achieved by tethering regulators to DNA binding sites as appropriate fusion proteins. This strategy has, for example, been used to show that the heterochromatin protein HP1, expressed as a fusion with Escherichia coli lactose repressor (LacI-HP1), promotes a closed chromatin structure and inactivation of reporter genes neighboring a LacO repeat in Drosophila [61,62], and to show that tethering of the vertebrate G9a histone methyltransferase to a GAL4 binding site within V(D)J minigene reporter impairs nonhomologous-mediated recombination of that construct [63]. Our laboratory has recently constructed a cell line, DT40 PolyLacO-λR, that is a DT40 derivative in which PolyLacO has been inserted by homologous gene targeting between ψVλ17 and ψVλ20, 17 kb upstream of the transcribed VλR (Figure 2A; M. Yabuki, E. C. Ordinario, W. J. Cummings, R. P. Larson, M. M. Fujii, et al., unpublished data). The PolyLacO insert is 3.8 kb in length and composed of approximately 100 copies of a 20-mer operator [64]. Using this cell line, it is possible to assay the effects of tethered regulatory factors on homologous recombination in a physiological process within an endogenous locus, avoiding the need for a transgene reporter. Control experiments have shown that the PolyLacO tag does not affect cell proliferation, cell cycle, or Ig gene diversification (M. Yabuki, E. C. Ordinario, W. J. Cummings, R. P. Larson, M. M. Fujii, et al., unpublished data).
In DT40 PolyLacO-λR GFP-LacI cells, which stably express enhanced green fluorescent protein (GFP) fused to LacI (GFP-LacI), the tagged λR allele can be directly imaged by fluorescence microscopy and appears as a distinct dot in each cell (Figure 2B, center). Tethering is reversible, as bright dots are not evident following overnight culture with 100 μM isopropyl-β-D-thiogalactoside (IPTG), which prevents LacI from binding to PolyLacO (Figure 2B, right).
To manipulate chromatin structure at the ψVλ array, we generated stable transfectants of DT40 PolyLacO-λR that express the D. melanogaster HP1 protein fused to LacI (LacI-HP1). HP1 is a nonhistone heterochromatin protein that functions in heterochromatic gene silencing, the spreading of heterochromatin, and histone deacetylation [58–60]. Tethered HP1 has been shown to promote a closed chromatin structure at adjacent genes [61,62,65–67]. Staining DT40 PolyLacO-λR LacI-HP1 transfectants with anti-LacI antibodies showed that LacI-HP1 colocalized with DAPI-dense regions corresponding to pericentric heterochromatin (Figure 3A), behaving as a functional marker of heterochromatin [68].
To ask if tethered LacI-HP1 altered chromatin structure, we assayed chromatin modifications at ψVλ17. This is the only site in the ψVλ array at which the rearranged and unrearranged alleles could be readily distinguished by use of specific PCR primers. Following ChIP, DNA was amplified with PCR primers specific for the targeted rearranged allele (ψVλ17R). Enrichment of ψVλ17R was compared to the nonexpressed Ova gene as an internal control, and normalized to the ψVλ17R:Ova enrichment ratio in total input DNA (see Material and Methods). AcH3 and AcH4 were enriched 2.2-fold and 5.9-fold, respectively, at ψVλ17R in DT40 PolyLacO-λR GFP-LacI controls (Figure 3B and 3C). These levels of enrichment are comparable to those documented in DT40 (Figure 1B). (Note that analysis of modification at ψVλ in the survey of the parental DT40 line necessarily included both alleles, which may underestimate activating modifications at the rearranged allele. In contrast, analysis of modifications at ψVλ17R interrogates only the active allele.) AcH3 and AcH4 were not enriched at ψVλ17R in DT40 PolyLacO-λR LacI-HP1 transfectants (0.6- and 1.0-fold, respectively; Figure 3B and 3C), consistent with HP1-mediated silencing. HP1 can effect silencing by recruitment of a histone methyltransferase that modifies lysine 9 of histone H3 [65–67], but may also promote silencing independently of this modification [61]. ChIP using antibodies against either di- and trimethylated H3 (lysine 9) did not reveal clear enrichment of the H3 lysine 9 methylation modification (data not shown). Dimethylation of lysine 4 of histone H3 (diMeK4[H3]) is associated with transcription and generally exhibits an overlapping distribution with acetylation [69,70]. Assays of diMeK4(H3) at ψVλ17R demonstrated that this modification was 18.9-fold enriched in DT40 PolyLacO-λR GFP-LacI cells, but at background levels in DT40 PolyLacO-λR LacI-HP1 cells (Figure 3B and 3C).
HP1 promotes maintenance and spreading of heterochromatin [65]. To verify that changes in chromatin structure promoted by tethered HP1 did not spread throughout the chromosome, we examined another site near the Igλ locus on Chromosome 15, the gene encoding the catalytic subunit of DNA polε. DNA polε is ubiquitously expressed and essential for chromosomal replication in eukaryotes [71], and it is encoded by a gene mapping approximately 2.1 Mb from Igλ. We found no difference in enrichment of AcH3 at the polε promoter region in the DT40 PolyLacO-λR LacI-HP1 transfectants relative to DT40 PolyLacO-λR GFP-LacI controls (polε/Ova enrichment 8.5-fold and 8.4-fold, respectively; Figure 3C). Similarly, there was no difference in AcH4 at the polε promoter in the DT40 PolyLacO-λR LacI-HP1 transfectants relative to DT40 PolyLacO-λR GFP-LacI controls (polε/Ova enrichment 1.9-fold and 1.7-fold, respectively; Figure 3C). Thus, tethering of LacI-HP1 at ψVλ caused local modifications in chromatin structure, diminishing the AcH3, AcH4, and diMeK4(H3) modifications characteristic of open chromatin at ψVλ17R, and causing chromatin to adopt a less permissive state.
We asked how tethered HP1 affected AcH3 and AcH4 levels at the expressed VλR by comparing these modifications in DT40 PolyLacO-λR LacI-HP1 cells and the DT40 PolyLacO-λR GFP-LacI control transfectants (Figure 4A). Tethered HP1 diminished AcH3 and AcH4 levels to approximately 40% and 20% of the control levels, respectively. To ask if this affected gene expression, we assayed both surface IgM (sIgM) expression and Vλ transcript levels. Staining cells with mouse anti-chicken IgM showed that sIgM expression was comparable in DT40 PolyLacO-λR GFP-LacI and DT40 PolyLacO-λR LacI-HP1 lines, cultured in either the presence or absence of IPTG (Figure 4B). Vλ transcript levels were assayed in RNA harvested from DT40 PolyLacO-λR GFP-LacI and DT40 PolyLacO-λR LacI-HP1 cells, and normalized to β-actin as a control (Figure 4C). No significant difference was observed between Vλ transcript levels in the two cell lines, demonstrating that transcription is not affected by tethering of HP1 within the ψVλ array. Thus tethered LacI-HP1 did not affect expression of the downstream Ig gene, although it did diminish AcH3 and AcH4 levels at VλR. The very high AcH3 and AcH4 levels characteristic of Vλ (Figures 1B and 4A) are therefore not essential to maintain high levels of gene expression.
To assess how extensive the chromatin effects of LacI-HP1 were, we examined AcH3 and AcH4 levels throughout the Igλ locus at the same amplicons examined in Figure 1, including one in the flank, six in the ψVλ array, and one at the expressed Vλ. Levels of modification were determined by comparing ψVλ17R:Ova ratios of immunoprecipitated and input conditions, as in Figure 3B. AcH3 modifications at the sites surveyed ranged from 24% to 63% of the levels at the same sites in the controls (Figure 5A, dark bars), and the average level of H3 acetylation across all of the sites was 38% of the DT40 PolyLacO-λR GFP-LacI control. Culture of DT40 PolyLacO-λR LacI-HP1 transfectants for 3 d with 250 μM IPTG increased acetylation of H3 at all eight sites surveyed (Figure 5A, compare dark and light bars). The effects of IPTG culture were somewhat variable, but at most sites IPTG culture restored levels of AcH3 to at least 45% of the level in the DT40 PolyLacO-λR GFP-LacI control cells, with an average of over 80%. Thus, the chromatin modifications at ψVλ17R in DT40 PolyLacO-λR LacI-HP1 cells resulted directly from tethered LacI-HP1, and were largely reversible.
H4 acetylation was surveyed at the same eight sites (Figure 5B, dark bars). AcH4 modifications were found to range from 18% to 42% of control levels, and the average level was 29% of that of the control cell line. Culture with IPTG for 3 d increased acetylation of H4 at all eight sites surveyed (Figure 5B, compare dark and light bars), restoring H4 acetylation to at least 57% of the level in the DT40 PolyLacO-λR GFP-LacI control cells, with an average of over 80%. Moreover, IPTG can at least partially reverse the effects of LacI-HP1.
These results show that the observed chromatin modifications in the ψVλ array are due to tethering of HP1. Moreover, the fact that these modifications are reversible shows that an active mechanism reverses histone modifications imposed by tethering chromatin modification factors at ψVλ.
The ability to manipulate chromatin structure at ψVλ by tethering LacI-HP1 (Figures 3–5) enabled us to directly ask whether and how chromatin structure influences Ig gene conversion. We used the sIgM loss variant assay to determine if tethered LacI-HP1 affected the clonal rate of sequence diversification of the rearranged VλR gene. This fluctuation assay measures the fraction of variant cells that no longer express structurally intact sIgM, and thus scores mutation events resulting from either gene conversion or point mutagenesis [47,50]. Independent clonal derivatives of DT40 PolyLacO-λR GFP-LacI and DT40 PolyLacO-λR LacI-HP1 were established by limiting dilution cloning of sIgM+ cells; the fraction of sIgM− cells in each population was determined by flow cytometry of cells cultured for 4 wk and then stained with anti-IgM antibody. The median sIgM loss rate was 0.5% for DT40 PolyLacO-λR GFP-LacI cells and 2.8% for DT40 PolyLacO-λR LacI-HP1 cells (Figure 6A). This corresponds to a 5.6-fold acceleration of clonal diversification rates in LacI-HP1 transfectants relative to GFP-LacI controls.
Ig gene diversification in chicken B cells occurs predominantly by gene conversion (templated mutation), but if gene conversion is impaired, for example by the absence of essential factors, repair can create a significant fraction of nontemplated mutations [50–55]. This is typically accompanied by an increase in the clonal diversification rate, because the ψVλ templates for gene conversion are about 80% identical to the rearranged gene, and a significant fraction of DNA lesions that are repaired by gene conversion do not undergo any alteration of sequence; in contrast, repair by a mutagenic polymerase is more likely to alter DNA sequence. To determine how tethering of HP1 accelerated diversification, we sorted single sIgM− cells from the DT40 PolyLacO-λR GFP-LacI and DT40 PolyLacO-λR LacI-HP1 transfectants, amplified expressed Vλ regions by single-cell PCR, and sequenced these regions. Sequence changes were categorized as templated if they were within a tract containing two or more base changes and the tract was an exact match to at least 9 bp of a donor ψVλ sequence, and as ambiguous if they consisted of only a single base change while matching at least 9 bp of a donor ψVλ sequence. Nontemplated events, consisting of point mutations, deletions, and insertions, were also scored. In the control DT40 PolyLacO-λR GFP-LacI transfectants, 55 templated events and two ambiguous events were documented among 71 mutations; thus, most events (77%) were templated, and a small fraction of events (20%) were point mutations (Figure 6B, left; Figure S1A). Strikingly, in DT40 PolyLacO-λR LacI-HP1 cells, point mutations predominated (58%), accompanied by deletions (8%) and insertions (14%), while only one clearly templated event and six ambiguous events were documented among 36 mutations (Figure 6B, right; Figure S1B). Thus, only 3% of mutations were clearly templated, and even including the ambiguous class of potentially templated mutations, templating could account for no more than 19% of mutation. Statistical comparisons showed that the difference between the fraction of clearly templated mutations in DT40 PolyLacO-λR GFP-LacI control cells and DT40 PolyLacO-λR LacI-HP1 transfectants (77% compared to 3%) was highly significant (p = 7.5 × 10−7, Fisher's exact test). The difference in the fraction of ambiguous, potentially templated mutations in the control cells (3%) and HP1 transfectants (17%) is also significant (p = 0.05, Fisher's exact test). This suggests that some mutations in this category may arise as a result of limitations on the length of a gene conversion tract imposed by nonpermissive donor chromatin. Thus, tethering of HP1 accelerated clonal rates of mutagenesis by impairing templated mutation.
Gene conversion at the chicken Ig loci uses an array of upstream ψV donors as templates for homology-directed repair of lesions targeted to the rearranged and transcribed V genes. We have shown that in chicken B cells carrying out active Ig gene conversion, chromatin within the donor ψVλ array is characterized by enrichment of AcH3 and AcH4, modifications that correlate with an open chromatin structure. We directly demonstrated the importance of permissive chromatin structure for Ig gene conversion by showing that tethering the heterochromatin protein HP1 to the ψVλ donor array caused local changes in chromatin structure, diminishing the AcH3, AcH4, and diMeK4(H3) modifications characteristic of open chromatin. Although these changes were not accompanied by the lysine 9 methylation (H3) modification characteristic of closed chromatin, they caused the region to adopt a state less permissive for gene conversion. Tethering of HP1 was accompanied by a dramatic shift in the Ig Vλ mutation spectrum, so that templated mutations were in the minority and point mutations predominated. Importantly, this effect on mutagenesis was correlated with a change in chromatin structure and not changes in expression of the locus. Thus, chromatin structure can dictate whether gene conversion occurs at an endogenously generated DNA lesion.
Gene conversion at Vλ results from priming of new DNA synthesis at the 3′ end of a break using a ψVλ region as template. Gene conversion requires synapsis between the donor and recipient DNA, as well as access to the donor by factors that carry out homology-directed repair. The elevated levels of H3 and H4 acetylation characteristic of the ψVλ array in wild-type DT40 are evidence of a relaxed chromatin structure, which would increase the accessibility of the ψVλ genes to trans-acting factors and also create a three-dimensional architecture that is favorable for sequence synapsis.
HP1 tethered within the ψVλ donor array impaired gene conversion at the rearranged VλR, without affecting Vλ gene expression. Chromatin changes caused by tethered HP1 may impair gene conversion by impeding access of repair factors and the invading strand to the donor template. Tethered HP1 may also contribute to larger chromosomal architecture that affects the mechanics of DNA repair pathways, such as looping necessary to juxtapose donor and recipient sequences. The point mutations that accumulated in LacI-HP1 transfectants are typical of thwarted recombinational repair, and are characteristic of cells lacking either trans-acting factors essential for recombination [49–55] or some or all of the ψV donor array [56]. HP1 regulates chromatin structure and heterochromatic gene silencing in two ways, by partnering with a histone methyltransferase [65] and by recruiting histone deacetylases [60]. Tethered HP1 caused modification changes characteristic of a nonpermissive chromatin structure within ψVλ.
Histone acetylation has been documented at actively transcribed mammalian Ig genes undergoing somatic hypermutation and class switch recombination, but whether hyperacetylation contributes to targeting of diversification has yet to be resolved [72–76]. A connection between histone acetylation and gene conversion was suggested by experiments showing that treatment of DT40 cells with the histone deacetylase inhibitor trichostatin A promotes genome-wide histone deacetylation accompanied by increased gene conversion at VλR [77]. However, the interpretation of those results is complicated by the fact that the effects of trichostatin A are genome-wide, and not specific. The DT40 PolyLacO-λR cell line permits local manipulation of chromatin structure, avoiding that complication. Moreover, we were able to demonstrate that the effects of tethering a LacI-HP1 fusion protein were largely reversed upon culture with IPTG, so an active mechanism must determine chromatin modification at ψVλ. For studies of homologous recombination, the DT40 PolyLacO-λR B cell line has the further advantage that Ig gene conversion is a physiological process within an endogenous locus, avoiding the need for a transgene reporter.
The importance of chromatin structure to the outcome of homologous recombination has implications for understanding the mechanisms that normally maintain genomic stability. There are vast numbers of repetitive elements distributed throughout the vertebrate genome, and recombination between these elements can lead to genomic instability [78]. In the human genome, there are approximately one million Alu elements, and recombination between Alu elements can cause duplications leading to tumorigenesis and genetic disease [79,80]. Histones carrying repressive modifications are enriched at repetitive elements [81]. These modifications undoubtedly maintain transcriptional repression; our results suggest they may also contribute to suppression of recombination.
LOH occurs as a result of unequal mitotic recombination between homologs at allelic sites. The mechanism of LOH is of particular interest, because it contributes to loss of tumor suppressor gene function, leading to tumorigenesis [82]. Recent experiments have demonstrated an age-dependent increase in LOH in Saccharomyces cerevisiae [83] and in reporter genes in Drosophila germ cells [84], and an increase in homologous recombination in mouse pancreatic cells [85]. Mechanisms proposed to explain age-associated LOH include elevated rates of DNA damage, changes in the cell cycle distribution, and inactivation of homology-independent repair pathways with aging. Our results suggest another possibility, that relaxation of chromatin structure may accompany aging and promote a genome-wide increase in homologous recombination in aging cells. This possibility is supported by recent analysis of Drosophila [86], as well as by recent evidence that the mutant form lamin A polypeptide (product of the human LMNA gene) responsible for Hutchinson-Gilford progeria syndrome leads to a genome-wide loss of H3 methylation [87].
The finding that chromatin structure regulates homologous recombination also has practical ramifications. Considerable current effort is directed toward developing strategies that harness a cell's capacity for homology-dependent repair to promote gene therapy, by providing an intact donor gene to replace a deficient target gene [88]. Our results suggest that permissive structure at the donor will be an important design parameter in developing donor genes for therapeutic applications.
ChIP was carried out as previously described [48,89]. For all experiments at least two chromatin preparations from at least two independent stably transfected lines were analyzed. Figures present one representative experiment in which results from analysis of four separate amplifications were used to calculate a standard deviation. Enrichment of the experimental amplicon was normalized to enrichment of an internal control amplicon from the Ova gene, amplified in the same tube by duplex PCR, and enrichment upon ChIP with specific antibodies was normalized to parallel experiments in which ChIP was carried out with total input DNA controls. Inclusion of the Ova internal control amplicon enabled us to normalize for immunoprecipitation efficiency, background carryover, and differences in gel loading. Enrichment equaled [(ψVλ/Ova)Ab]/[(ψVλ/Ova)Input]. As an additional control, the ratio of the experimental and control amplicons in the total input control was compared to a control ChIP with polyspecific IgG; in all cases, enrichment in input and IgG controls were essentially equal.
Antibodies used were as follows: anti-AcH3 (06–599), anti-AcH4 (06–866), and diMeK4(H3) (07–030) from Upstate ( http://www.upstate.com/). PCR primers for ChIP were as follows: VλR, 5′-GCCGTCACTGATTGCCGTTTTCTCCCCTC-3′ and 5′-CGAGACGAGGTCAGCGACTCACCTAGGAC-3′; region between ψVλ1 and Vλ, 5′-CTGTGGCCTGTCAGTGCTTA-3′ and 5′-GCAGGGAACCACAAGAACAT-3′; ψVλ1, 5′-GGGACTTGTGTCACCAGGAT-3′ and 5′-CGCAGTCACATGTGGAATATC-3′; ψVλ5, 5′-GAGCCCCATTTTCTCTCCTC-3′ and 5′-GAGATGTGCAGCAACAAGGA-3′; ψVλ13, 5′-CCCTCTCCCTATGCAGGTTC-3′ and 5′-CCCCTATCACCATACCAGGA-3′; ψVλ18, 5′-CCATTTTCTCCCCTCTCTCC-3′ and 5′-TCACCCTACAGCTTCAGTGC-3′; ψVλ24, 5′-CCATTTTCTCCCCTCTCTCC-3′ and 5′-CAGCCCATCACTCCCTCTTA-3′; ψVλ25, 5′-TCTGTTGGTTTCAGCACAGC-3′ and 5′-GCAGTTCTGTGGGATGAGGT-3′; ψVλ upstream flank, 5′-GGCTCCTGTAGCTGATCCTG-3′ and 5′-GTTCTTTGCTCTTCGGTTGC-3′; ψVλ17 at the PolyLacO-targeted allele, 5′-TAGATAGGGATAACAGGGTAATAGC-3′ and 5′-AGGGCTGTACCTCAGTTTCAC-3′; Ova, 5′-ATTGCGCATTGTTATCCACA-3′ and 5′-TAAGCCCTGCCAGTTCTCAT-3′; and polε, 5′-GGGCTGGCTCATCAACAT-3′ and 5′-CTGGGTGGCCACATAGAAGT-3′.
The LacI-HP1 expression plasmid was created by substituting LacI-HP1 from a construct provided by L. Wallrath (University of Iowa) for AID in pAIDPuro (from H. Arakawa, Munich, Germany), to position LacI-HP1 downstream of the chicken β-actin promoter. The GFP-LacI expression plasmid (p3′ss-EGFP-LacI) was provided by A. Belmont (University of Illinois). Cell culture and transfection were carried out as previously described [47]. DT40 PolyLacO-λR was generated by homologous gene targeting, using a construct carrying approximately 3.8 kb of PolyLacO flanked by arms designed for targeting the region between ψVλ17 and ψVλ20, 17 kb upstream of the transcribed VλR (M. Yabuki, E. C. Ordinario, W. J. Cummings, R. P. Larson, M. M. Fujii, et al., unpublished data). In brief, homologous integrants were identified by PCR, and the selectable marker deleted by Cre expression. The DT40 bursal lymphoma derives from B cells in which only one Igλ allele is rearranged, and in which the two parental chromosomes are distinguished by a polymorphism near ψVλ17. This enabled us to determine whether the rearranged or unrearranged allele had been targeted by PCR. Control experiments established that cell cycle distribution was comparable in DT40 PolyLacO-λR, DT40 PolyLacO-λR GFP-LacI, and DT40 PolyLacO-λR LacI-HP1 cells, and that culture of cells with up to 500 μM IPTG for 7 d did not affect proliferation rate or chromatin modifications at ψVλ17R in DT40 PolyLacO-λR GFP-LacI control cells. Oligonucleotides for Vλ sequence analysis have been described [47].
For fluorescence imaging, cells (2 × 105) were cytospun onto glass slides and fixed with 2% paraformaldehyde for 20 min, permeabilized with 0.1% NP-40 for 15 min, and stained as previously described [90]. Primary staining was with an anti-LacI monoclonal antibody (1:500 dilution; Upstate), and the secondary antibody was donkey anti-mouse IgG Alexa Fluor 594 (1:2,000; Molecular Probes, http://probes.invitrogen.com/). To visualize the nucleus, cells were stained with DAPI (Sigma-Aldrich, http://www.sigmaaldrich.com/). Fluorescent images were acquired using the DeltaVision microscopy system (Applied Precision, http://www.appliedprecision.com/) and processed with softWoRx software (Applied Precision).
RNA was harvested from cells using TRIzol Reagent (Invitrogen, http://www.invitrogen.com/), purified with a PreAnalytiX column (Qiagen, http://www1.qiagen.com/), and subject to one round of reverse transcription prior to PCR. Vλ transcripts were PCR-amplified following dilution of the template (1:1,300), and β-actin was PCR-amplified from an undiluted sample. The primers for amplification of Vλ were 5′-GTCAGCAAACCCAGGAGAAAC-3′ and 5′-AATCCACAGTCACTGGGCTG-3′. The primers for amplification of β-actin have been described [36].
The sIgM loss variant assay, which measures the accumulated sIgM loss variants resulting from frameshift or nonsense mutations in mutated V regions, was used to quantitate Ig V region diversification [47,50]. In brief, sIgM+ cells were isolated by flow cytometry followed by limiting dilution cloning, and expanded for 4 wk. To quantitate the fraction of sIgM− cells, approximately 1 × 106 cells were stained with anti-chicken IgM-RPE (SouthernBiotech, http://www.southernbiotech.com/) and analyzed on a FACScan with CellQuest software (BD Biosciences, http://www.bdbiosciences.com/).
Single-cell PCR and sequence analysis were performed as described [47]. In brief, sIgM− cells were sorted and aliquoted to single wells, Vλ regions were amplified and sequenced, and their sequences were compared to those of the ψVλ donors to determine if mutations were templated or nontemplated. The criterion for a templated mutation was that nine consecutive bases must be an exact match in donor and recipient. Sequences were derived from two independently transfected lines. Only unique sequences were included for classification of the mutations.
|
10.1371/journal.pmed.1002581 | Geographic and sociodemographic variation of cardiovascular disease risk in India: A cross-sectional study of 797,540 adults | Cardiovascular disease (CVD) is the leading cause of mortality in India. Yet, evidence on the CVD risk of India’s population is limited. To inform health system planning and effective targeting of interventions, this study aimed to determine how CVD risk—and the factors that determine risk—varies among states in India, by rural–urban location, and by individual-level sociodemographic characteristics.
We used 2 large household surveys carried out between 2012 and 2014, which included a sample of 797,540 adults aged 30 to 74 years across India. The main outcome variable was the predicted 10-year risk of a CVD event as calculated with the Framingham risk score. The Harvard–NHANES, Globorisk, and WHO–ISH scores were used in secondary analyses. CVD risk and the prevalence of CVD risk factors were examined by state, rural–urban residence, age, sex, household wealth, and education. Mean CVD risk varied from 13.2% (95% CI: 12.7%–13.6%) in Jharkhand to 19.5% (95% CI: 19.1%–19.9%) in Kerala. CVD risk tended to be highest in North, Northeast, and South India. District-level wealth quintile (based on median household wealth in a district) and urbanization were both positively associated with CVD risk. Similarly, household wealth quintile and living in an urban area were positively associated with CVD risk among both sexes, but the associations were stronger among women than men. Smoking was more prevalent in poorer household wealth quintiles and in rural areas, whereas body mass index, high blood glucose, and systolic blood pressure were positively associated with household wealth and urban location. Men had a substantially higher (age-standardized) smoking prevalence (26.2% [95% CI: 25.7%–26.7%] versus 1.8% [95% CI: 1.7%–1.9%]) and mean systolic blood pressure (126.9 mm Hg [95% CI: 126.7–127.1] versus 124.3 mm Hg [95% CI: 124.1–124.5]) than women. Important limitations of this analysis are the high proportion of missing values (27.1%) in the main outcome variable, assessment of diabetes through a 1-time capillary blood glucose measurement, and the inability to exclude participants with a current or previous CVD event.
This study identified substantial variation in CVD risk among states and sociodemographic groups in India—findings that can facilitate effective targeting of CVD programs to those most at risk and most in need. While the CVD risk scores used have not been validated in South Asian populations, the patterns of variation in CVD risk among the Indian population were similar across all 4 risk scoring systems.
| Cardiovascular disease (CVD) is thought to cause a large and increasing health and economic burden in India.
Understanding how CVD risk varies among India’s population groups could inform health system planning and the targeting of CVD programs to those most in need.
Yet, to date, there has not, to our knowledge, been a large-scale population-based study that examines how CVD risk varies among India’s states and sociodemographic groups.
This analysis pooled data from 797,540 participants aged 30 to 74 years across 2 large population-based household surveys, which jointly covered 27 of 29 states and 5 of 7 union territories in India.
The average 10-year risk of a fatal or nonfatal CVD event varied widely among states in India, ranging from 13.2% in Jharkhand to 19.5% in Kerala.
In addition, adults living in urban areas, as well as those with a higher household wealth or education, tended to have a greater CVD risk.
This study identified important variation in composite CVD risk (as well as in the prevalence of individual CVD risk factors) among India’s population, which can inform effective targeting of CVD programs to those most at risk and most in need.
While the absolute CVD risk levels predicted by any 1 of the 4 CVD risk calculators we used should be interpreted with caution (because of the absence of CVD risk equations that have been validated among South Asian populations), the relative variation of CVD risk among India’s population groups was similar across all 4 risk calculators.
| Cardiovascular disease (CVD) is the leading cause of mortality worldwide, including in low- and middle-income countries [1]. While the Global Burden of Disease project has recently highlighted the limited data availability for India [2], it nonetheless estimated that the country contributed almost one-fifth (18.6%) of the global CVD burden, as measured by disability-adjusted life years, in 2016 [3]. Although this proportion is only slightly above the share of the world’s population that lives in India (17.7% in 2015) [4], it is likely to increase in the future for 3 main reasons. First, India is expected to make the greatest contribution to global population growth of any country until at least 2050 [5]. Second, India’s population is aging and urbanizing: the share of people aged more than 60 years is estimated to double from 8.9% to 19.4% between 2015 and 2050 [5], and the percentage of Indians living in cities is projected to grow from 30.9% in 2010 to 50.3% in 2050 [6]. Third, the rise in living standards and socio-cultural transitions in India are likely to lead to more obesogenic lifestyles [7]. Evidence indicates that urban South Asians, especially those living in North America and Western Europe, have a higher prevalence of CVD and type 2 diabetes than local white populations [8–10]. While the reasons for this phenomenon are not clear (although some explanatory models have been proposed in the literature) [10,11], this susceptibility for CVD among South Asians living in urban, high-income settings suggests that increasing urbanization and the spread of obesogenic environments might raise the prevalence of CVD even more in India (and South Asia in general) than it has already in other world regions.
Given the detrimental effects of CVD on health outcomes [12], financial risk protection [13], and economic growth [14], the course of India’s CVD epidemic will directly impact several Sustainable Development Goals (SDGs). These include SDG 1 (“End poverty in all its forms everywhere”) and SDG 3 (“Ensure healthy lives and promote well-being for all at all ages”) as well as their corresponding targets SDG 3.4 (“By 2030, reduce by one-third premature mortality from NCDs [noncommunicable diseases]”) and SDG 3.8 on achieving universal health coverage. Considering the size and growth of India’s population [5], the development of its CVD epidemic over the next decade will also have a decisive impact on the world’s ability to achieve the SDGs [15].
Many studies have focused on providing the best possible prevalence estimates for CVD and its risk factors at the national level in India [16–19]. However, much less is known about the distribution of these risk factors within India—both geographically and by individuals’ sociodemographic characteristics. Given that India’s health system is largely decentralized to the state level [20], understanding the variation of CVD risk within India is highly relevant not only to identify target groups for CVD prevention, screening, and treatment programs but also for health system planning at the state and district level. Using data from a sample of 797,540 adults aged 30–74 years, this study therefore aimed to determine how CVD risk varies by geography and individual-level sociodemographic characteristics across India.
We pooled data from 2 large household surveys in India, the District Level Household Survey–4 (DLHS-4) and the second update of the Annual Health Survey (AHS), both of which were conducted between 2012 and 2014. These 2 surveys were combined because they (i) jointly covered most states (27 of 29) and union territories (5 of 7) of India (and no areas in India were covered by both surveys), (ii) were conducted simultaneously, (iii) are both representative at the district level, and (iv) used the same questionnaire and methodology to collect clinical, anthropometric, and biomarker (CAB) measurements. The states covered by each of the surveys are shown in Fig A in S1 Fig.
In both surveys, all non-pregnant household members aged 18 years and older were eligible for blood glucose, blood pressure (BP), height, and weight measurements. The analyses in this study were restricted to those aged 30 to 74 years because the CVD risk equations used in this study were developed among adults of this age range only [21–23]. Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters. Participants’ blood glucose was measured using a capillary blood sample (from a finger prick) taken using a handheld blood glucose meter (SD CodeFree), which multiplied capillary blood glucose readings by 1.11 to display their plasma equivalent [24]. Participants were instructed to fast for at least 8 hours before the time of the measurement. BP was measured twice, with each measurement 10 minutes apart, using an electronic upper arm BP monitor (Rossmax AW150).
All data collectors for the AHS and DLHS-4 were trained in the collection of sociodemographic as well as the CAB data. In the AHS and DLHS-4, training sessions were organized for 12–15 and 15–20 data collectors at a time, respectively. Trainings for anthropometric and biomarker measurements lasted for 7 days, with 4 days of training conducted in the classroom and 3 days in the field. The following mechanisms were put in place for both the AHS and DLHS-4 to ensure good data quality: (i) establishment of standard protocols for questionnaire administration, anthropometry, BP measurement, and blood glucose measurement, (ii) the field supervisor conducted a second CAB measurement on 10% of participants each day to identify poor-quality measurements, (iii) a medical consultant (who received additional training for the CAB component) visited 10% of all sampled households and conducted a second CAB assessment to identify poor-quality measurements, (iv) continuous data monitoring by the implementing organization, (v) immediate replacement of faulty equipment, and (vi) regular checks of the accuracy of digital BP monitors and the handheld blood glucometers. More details on the data collection procedures can be found in the CAB manuals of the AHS and DLHS-4 [25,26]. The documents can be obtained from the corresponding author.
The AHS and DLHS-4 jointly cover all 29 states of India apart from Jammu and Kashmir (where data were not collected due to violent conflicts) and Gujarat (where data were not available in the public domain). The datasets also include all union territories of India except Dadra and Nagar Haveli, and Lakshadweep. The 2 states and 2 union territories not included in this analysis accounted for 6% of India’s population at the time of the last census (2011) [27].
This analysis of an existing dataset in the public domain received a determination of “not human subjects research” by the institutional review board of the Harvard T.H. Chan School of Public Health on 23 November 2016 (protocol number: IRB16-1915). All participants provided written informed consent to participate in the AHS and DLHS-4.
Throughout this analysis, we used the predicted 10-year risk of a CVD event to summarize CVD risk as computed by risk calculators across different risk factors. However, we also “disaggregated” predicted CVD risk by examining the geographic and sociodemographic variation of each of the risk factors included in these risk calculators: (i) BMI, (ii) high blood glucose, (iii) systolic BP, and (iv) smoking. Results on diastolic BP are presented in supplementary files for completeness (Figs D and G in S1 Fig).
We primarily used continuous predicted 10-year CVD risk as an outcome. However, in secondary analyses, we dichotomized predicted 10-year risk of a CVD event into high and low risk whereby “high CVD risk” was defined as a 10-year CVD risk ≥ 30%. This threshold was chosen because it is the cutoff used in the World Health Organization’s NCD Global Action Plan targets to decide who is eligible for drug therapy and counseling [30]. We primarily used the Framingham risk score (the version not requiring total cholesterol measurements) to calculate CVD risk because it is the most widely used CVD risk scoring system internationally [21,31]. However, in secondary analyses, we also show results using CVD risk calculated with 3 other risk scores that do not require blood lipid measurements, namely Harvard–NHANES [23], Globorisk [32], and the risk score developed by WHO and the International Society for Hypertension (WHO–ISH) [33]. None of these risk scores have been validated among South Asian populations. Because data on participants’ medical history were unavailable, we did not exclude participants with a previous or current CVD.
All 4 risk scores used predict the risk of a fatal or nonfatal CVD event, but each score defines a CVD event differently (Table 1). The Framingham risk score uses the broadest [21], and Globorisk [32] and WHO–ISH [33] the narrowest, range of CVD events as outcome. The Globorisk project has calibrated its risk equation to 182 countries, including India, as described by Ueda et al. [32]. Similarly, WHO has calibrated its risk score to each WHO sub-region [33]. The Framingham and Harvard–NHANES risk scores were calibrated to India using the incidence rate (by 5-year age group) of peripheral artery disease (Framingham only), ischemic heart disease, and cerebrovascular disease in 2015 as estimated by the Global Burden of Disease project [12].
The 4 risk scores predict CVD risk by sex using the following inputs: age, BMI (except WHO–ISH), presence of diabetes (except the office-based version of the Globorisk score), current smoking, systolic BP, and treatment for hypertension (except Globorisk and WHO–ISH). Diabetes was defined as having a blood glucose ≥7.0 mmol/l if reporting to have fasted or ≥11.1 mmol/l if reporting not to have fasted, or reporting to be on regular treatment for diabetes. Because the survey only measured blood glucose to assess diabetes, which is insufficient for a clinical diagnosis of this condition, we refer to this outcome as “high blood glucose” for the remainder of the paper. For systolic BP, we used the average of the 2 systolic BP readings recorded.
Explanatory variables were household wealth quintile, education, and whether the household was located in a rural or urban area. We used a principal component analysis to create a household wealth index based on 5 key housing characteristics (water supply, type of toilet and whether it is shared, cooking fuel, housing material, and source of lighting) and household ownership of 12 assets (radio, TV, computer, phone, refrigerator, bicycle, scooter, car, washing machine, sewing machine, house, and land). The first component in the principal component analysis (using the methodology developed by Filmer and Pritchett [34,35]) was used to combine these variables into a single measure, separately for urban and rural areas. This index was then divided into quintiles (again, separately for rural and urban areas) based on the distribution in the national (aggregate) dataset.
CVD risk was computed for each study participant aged 30 to 74 years. Using sampling weights to account for the complex survey design, we then calculated the mean 10-year CVD risk at the national level, by state, and by individual-level sociodemographic characteristics. All mean risk values (and prevalence estimates) are unadjusted for individuals’ sociodemographic characteristics (other than age standardization where explicitly indicated). In addition, we used ordinary least squares regressions to regress the natural logarithm of the CVD risk score on sociodemographic characteristics and a fixed effect for district (i.e., a binary indicator for each district to adjust for unobserved differences between districts). The natural logarithm of CVD risk was used in all regression models to allow for a more intuitive interpretation of the regression coefficients as percentage changes in CVD risk. The regressions were run separately for males and females because each CVD risk score provides sex-specific risks. Two different regression models were fitted for each CVD risk score (except WHO–ISH because it only provides risk categories rather than a continuous risk variable [33]) and sex: (i) a model that included only 1 sociodemographic characteristic, age group, and a district-level fixed effect and (ii) a model that included all sociodemographic characteristics and a district-level fixed effect as explanatory variables. Standard errors were adjusted for clustering at the level of the PSU. The mean (for BMI and systolic BP) or the prevalence (for high blood glucose and smoking) of each CVD risk factor was plotted by state and sociodemographic characteristics to help explain observed patterns in the CVD risk scores.
This study did not have a prospective analysis plan. The analysis outlined above was conceived by the authors prior to embarking on data analysis. None of the analyses were unplanned with the exception that reviewer comments led us to add (i) additional maps to examine state-level variation (specifically, to stratify variation not only by sex but also by age group and rural–urban residence) and (ii) multi-level modeling to examine the association of CVD risk with district-level wealth and urbanization. Regarding the latter, the peer reviewer comments prompted us to further investigate area-level predictors of CVD risk because we identified wide geographic variation in CVD risk in our initial analysis. To do so, we computed a measure of district-level wealth by calculating (separately for rural and urban areas within districts because household wealth was also computed separately for rural and urban areas) the median of the continuous household wealth index in a district, and then categorizing the district-level median into quintiles (henceforth referred as “district wealth quintiles”). Another potential area-level predictor that we examined was the level of urbanization of a district assessed through the proportion of participants in a district who were residing in an urban area. These 2 area-level predictors were chosen because they could be calculated directly from the data. We, thus, did not have to rely on the accuracy of other data sources, and—unlike other indicators—these district-level indicators were automatically available for all districts in the sample for the time of the survey. The association of these 2 district-level predictors with CVD risk were studied using a multivariable linear regression model with the natural logarithm of 10-year CVD risk as the dependent variable, random intercepts by district, and individual-level sociodemographic characteristics (5-year age group, sex, educational attainment, and household wealth quintile) as independent variables.
We conducted a complete case analysis for all analyses presented in this paper. The Global Burden of Disease project’s 2013 population for India was used for age standardization [36]. This study is reported as per STROBE guidelines (S1 Checklist). Statistical analyses were run in R version 3.3.2 (2016) [37], and the WHO–ISH score was calculated using the whoishRisk package [38].
Sociodemographic information was available for a total of 1,094,754 adults aged 30–74 years, which included individuals who were not present at the time of the household visit (as sociodemographic information was collected for all household members from the household head). In total, 797,540 (72.9% [797,540/1,094,754]) survey participants who had all the values for the variables needed to calculate each CVD risk score (i.e., blood glucose, systolic BP, height and weight, age, sex, and smoking status) were included in the analysis. While mean BMI was similar between males and females (22.6 kg/m2 and 22.3 kg/m2, respectively), females were more likely to have BMI < 18.5 kg/m2 or BMI ≥ 25 kg/m2 than males (Table 2). In all, 10.0% (42,066/420,691) of females and 10.7% (40,444/376,849) of males had high blood glucose. Smoking prevalence and mean systolic BP were higher among men than women (27.1% [102,182/376,849] versus 2.6% [10,992/420,691] and 129.1 mm Hg versus 126.7 mm Hg, respectively). In all, 56.2% (236,555/420,691) of females and 34.0% (128,183/376,849) of males had not completed primary school, and approximately one-third of participants lived in urban areas. Table A in S1 Table shows that those who were excluded from the analysis (27.1% of participants) because they had a missing value for at least 1 of the variables needed to calculate predicted CVD risk had a similar prevalence of CVD risk factors as those who were included in the analysis.
Overall, the mean 10-year risk of a CVD event in the (not age-standardized) population aged 30–74 years was 12.7% (95% CI: 12.7%–12.8%) among females and 21.4% (95% CI: 21.3%–21.6%) among males (Table B in S1 Table). The (not age-standardized) prevalence of a high CVD risk (10-year risk ≥ 30%) in those aged 30 to 74 years was 14.6% (95% CI: 14.4%–14.8%) among females and 31.7% (95% CI: 31.4%–32.0%) among males. The Framingham risk score yielded similar risk estimates to Harvard–NHANES, but substantially higher estimates than Globorisk and WHO–ISH (Table C in S1 Table). As an alternative measure of need for treatment and counseling to reduce CVD risk, we show the (not age-standardized) proportion of participants who were current smokers, had a high blood glucose, had hypertension, or were overweight in Table D in S1 Table.
The age-standardized state-level mean 10-year CVD risk (across all age groups) varied from 10.2% (95% CI: 9.8%–10.7%) among females in Assam to 24.2% among males in Nagaland (95% CI: 23.5%–25.0%) and Himachal Pradesh (95% CI: 23.6%–24.9%) (Fig 1). Similarly, the age-standardized prevalence of a high CVD risk varied from 5.0% (95% CI: 4.5%–5.6%) among females in Assam to 30.4% (95% CI: 28.8%–32.0%) among males in Kerala (Fig B in S1 Fig). Among both males and females, CVD risk tended to be highest in South India (including Goa), the 3 most northern states in the dataset (Himachal Pradesh, Punjab, and Uttarakhand), the northeastern states (except Assam), and West Bengal (particularly among males). This pattern across states, as well as the wide degree of variation in CVD risk between states, largely remained when examining state-level prevalence within only certain age groups (Fig 2) and within rural and urban areas (Fig 3). While the absolute risk levels depended strongly on the choice of CVD risk calculator, the relative variation across states was similar regardless of the CVD risk score used (Fig C in S1 Fig).
Fig 4 shows differences between states in the age-standardized mean (for BMI and systolic BP) or prevalence (for high blood glucose and smoking) for each of the CVD risk factors that are included in the CVD risk score. Mean BMI was high in both northern (Haryana, Himachal Pradesh, Punjab, and Uttarakhand) and southern states (Andhra Pradesh, Goa, Karnataka, Kerala, Tamil Nadu), ranging from 22.8 kg/m2 among males in Uttarakhand to 25.1 kg/m2 among females in Punjab. High blood glucose prevalence, however, was relatively low in the northern states (ranging from 4.4% among males in Himachal Pradesh to 10.9% among females in Punjab). Mean systolic BP was highest in the northern states (ranging from 123.7 mm Hg among females in Haryana to 136.2 mm Hg among males in Punjab) as well as in Nagaland and Sikkim (130.7 mm Hg and 132.8 mm Hg among females and 133.6 mm Hg and 133.1 mm Hg among males, respectively). Smoking was most prevalent among males in the northeastern states of Arunachal Pradesh (46.4%), Manipur (60.3%), Meghalaya (59.7%), and Mizoram (71.7%) and the eastern state of West Bengal (49.5%). As with the CVD risk score, these patterns across states and the wide variation between states remained when examining the state-level distribution of these variables only within certain age groups and within rural and urban areas (Fig D in S1 Fig).
We found a positive association between the mean CVD risk in a district and the district’s wealth when plotting the district-level mean Framingham risk score against the district-level median (categorized into quintiles) of the continuous household wealth index (Fig 5). Similarly, mean CVD risk was positively associated with the proportion of the sample in a district that was living in an urban area (Fig 6).
Confirming the impression from the plotting of our data in Figs 5 and 6, our multivariable linear regressions revealed that district wealth quintile was positively associated with CVD risk in both rural and urban areas, with the association stronger in rural areas (Table 3). Specifically, among participants residing in rural areas, living in the wealthiest 20% of districts in India was associated with a relative increase in the 10-year CVD risk of 13.1% (95% CI: 10.7%–15.6%; p < 0.001) compared to the poorest 20% of districts. In urban areas, the corresponding increase was only 4.3% (95% CI: 1.5%–7.1%; p = 0.003). In addition, as shown in Table 4, living in an entirely urbanized district was associated with a relative increase in the 10-year CVD risk of 16.9% (95% CI: 12.7%–21.1%; p < 0.001) compared with living in an entirely rural district. The associations shown in Tables 3 and 4 were similar regardless of the CVD risk calculator used (Tables G–J in S1 Table).
Stratifying mean 10-year CVD risk by age group, sex, rural versus urban location, and household wealth quintile shows that (i) those living in urban areas generally had a higher CVD risk than those living in rural areas, (ii) irrespective of sex and location, mean CVD risk was higher in the wealthiest than in the poorest quintile in all age groups (except the youngest age group), and (iii) both the relative and absolute differences in mean CVD risk between wealth quintiles tended to be larger in rural than in urban areas (Fig 7). These patterns were generally similar when using Harvard–NHANES or Globorisk (WHO–ISH does not yield a continuous risk score) instead of the Framingham risk score (Fig E in S1 Fig), and when examining the prevalence of a high 10-year CVD risk (≥30%) as opposed to mean CVD risk (Fig F in S1 Fig).
Table 5 shows the regression coefficients (which can be interpreted as approximations of the percentage change in CVD risk) when regressing the natural logarithm of the Framingham risk score on individuals’ sociodemographic characteristics and a fixed effect for district. Household wealth quintile, education, and living in an urban area were positively associated with CVD risk among both sexes, but for all 3 variables the coefficients for males were substantially smaller than those for females. The association between education and CVD risk was weak once the regressions were adjusted for other sociodemographic characteristics. The regression results were similar when using Harvard–NHANES or Globorisk (WHO–ISH does not yield a continuous risk score) (Tables K and L in S1 Table).
Fig 8 shows that while mean BMI, high blood glucose, and mean systolic BP were all positively associated with household wealth and living in an urban area, the prevalence of high blood glucose and mean systolic BP were nonetheless high in middle and old age among the poorest wealth quintiles and in rural areas. Smoking, on the other hand, was more common in poorer quintiles, in rural areas, and among males.
Pooling and analyzing data on CVD risk for 797,540 adults across India (a country that accounts for more than one-sixth of the world’s population [4]), we identified important variation in risk among states (with CVD risk tending to be highest in the northern, northeastern, and southern states) and by individuals’ sociodemographic characteristics. In particular, we found that (i) CVD risk was higher in urban areas and among males, (ii) while mean BMI was substantially higher among wealthy than poor individuals, high blood glucose and high systolic BP were common among poor individuals in middle and old age, and (iii) smoking was most prevalent among men, in poorer wealth quintiles, and in rural areas. Thus, while a major investment in CVD and risk factor prevention, screening, and treatment is needed across India, this study provides important new insights on the distribution of CVD risk to effectively target health system resources for CVD management to those most at risk and most in need. Given that we found that district-level mean CVD risk was positively associated with district wealth and urbanization, such investments may be crucial to minimize further rises in CVD risk as socioeconomic development and urbanization in India progress over the coming decades.
Even though the Globorisk and WHO–ISH scores were developed specifically with the goal of providing CVD risk estimates in populations for which no validated CVD risk calculator exists [22,32,33], the absence of a CVD risk equation that has been validated in South Asian cohorts is a major limitation of this study. Nonetheless, CVD risk calculators are used routinely in clinical settings (where they are employed in conjunction with a clinical assessment) in India [40]. Although this does not necessarily justify their employment at the population level, there has been a recent move to applying these risk equations to entire populations. For instance, one of the WHO’s NCD Global Action Plan targets (that “at least 50% of eligible people receive drug therapy and counselling to prevent heart attacks and strokes” by 2020, for which the WHO defined eligibility as a 10-year CVD risk ≥ 30% [30]) is based on the concept of applying CVD risk equations to the population level. In addition, several recent studies have used CVD risk calculators for population-level assessments of CVD risk [22,32,41]. Nevertheless, we wish to emphasize here that the absolute risk predictions provided in this study should be interpreted with caution. Indeed, the lack of validation in South Asian populations may be one reason that our risk estimates varied widely across CVD calculators. Specifically, the Framingham and Harvard–NHANES risk scores yielded substantially higher estimates than Globorisk and WHO–ISH. This observed difference in estimates was expected to some degree given that Globorisk and WHO–ISH predict the risk of (fatal or nonfatal) myocardial infarction or stroke, whereas the Framingham and Harvard–NHANES risk scores include a broader set of outcomes (Table 1). Having acknowledged this limitation, we do believe that the CVD risk predictions are useful as a summary measure of CVD risk when assessing variation of risk among population groups. In this regard, it is important to highlight that the patterns of variation in CVD risk by state, rural versus urban residence, and individual-level sociodemographic characteristics were very similar across the 4 different risk calculators used in this study.
This study has several additional limitations. First, a relatively high percentage (27.1%) of participants had a missing value for at least 1 variable needed to calculate their CVD risk. While we show that participants excluded because of a missing value had similar summary statistics for CVD risk factors as those included in the analysis, there is nonetheless potential for selection bias. Second, a 1-time capillary blood glucose measurement is not recommended for the diagnosis of diabetes in clinical settings [42]. However, this screening method has been shown to have an acceptable sensitivity and specificity for defining diabetes in population-based research, and is hence the recommended method for monitoring diabetes prevalence in the WHO’s STEPwise Approach to Noncommunicable Disease Risk Factor Surveillance [43–45]. Nonetheless, to be clear about this limitation of our data, we refer to high blood glucose values (or being on treatment for diabetes) as “high blood glucose” in this paper rather than “diabetes.” Third, the questionnaire used in both the DLHS-4 and AHS was designed such that only those who answered in the affirmative to having had symptoms (of any type) lasting for more than 1 month during the last 1 year were asked whether they were “getting regular treatment” for the condition. Our data thus likely underestimate the number of participants who were on treatment for hypertension and diabetes. Fourth, we were unable to exclude participants with a current or previous CVD (e.g., a previous myocardial infarction) because data on participants’ medical history were not collected. Since those with a previous or current CVD tend to have a higher CVD risk than predicted with a CVD risk score, this limitation biases our CVD risk estimates for the population of India downwards. Lastly, the CVD risk scores used here do not take into account consumption of smokeless tobacco, which is common in India and may increase CVD risk [46,47].
In conclusion, this study identified important variation in CVD risk and risk factor prevalence among states and population groups in India—information that will be essential for effective targeting of resources and interventions for prevention, screening, and treatment to those most at risk and most in need. Such investments in targeted CVD care programs as well as relevant health policy measures are urgently needed—particularly in states with a high CVD risk—if India is to minimize CVD’s adverse consequences for health, well-being, financial risk protection, and economic growth. Given the size and projected growth of India’s population, the determination and effectiveness of the country’s measures to prevent and treat CVD over the coming years will have an important bearing on the achievement of the SDGs at the global level.
|
10.1371/journal.ppat.1001246 | Critical Role of IRF-5 in the Development of T helper 1 responses to Leishmania donovani infection | The transcription factor Interferon Regulatory Factor 5 (IRF-5) has been shown to be involved in the induction of proinflammatory cytokines in response to viral infections and TLR activation and to play an essential role in the innate inflammatory response. In this study, we used the experimental model of visceral leishmaniasis to investigate the role of IRF-5 in the generation of Th1 responses and in the formation of Th1-type liver granulomas in Leishmania donovani infected mice. We show that TLR7-mediated activation of IRF-5 is essential for the development of Th1 responses to L. donovani in the spleen during chronic infection. We also demonstrate that IRF-5 deficiency leads to the incapacity to control L. donovani infection in the liver and to the formation of smaller granulomas. Granulomas in Irf5-/- mice are characterized by an increased IL-4 and IL-10 response and concomitant low iNOS expression. Collectively, these results identify IRF-5 as a critical molecular switch for the development of Th1 immune responses following L. donovani infections and reveal an indirect role of IRF-5 in the regulation of iNOS expression.
| Leishmania donovani is a parasite that currently infects 12 million people around the world. In order to better understand why this parasite causes incurable disease we chose to investigate how the immune system sees L. donovani. The immune system sees infecting organisms by the recognition of molecules that are specifically expressed by pathogens. This is done by a family of receptors expressed by cells called Toll Like Receptors (TLRs). When TLRs recognize a pathogen it leads to a molecular chain reaction within the cell resulting in the release of cytokines. Interferon Regulatory Factors (IRFs) are a very important part of this signaling chain. The protein we have studied, IRF-5, has been identified as having a key role in inducing pro-inflammatory cytokines following the recognition of viruses by TLRs. However, whether it plays a role in the immune response to parasitic disease has not yet been examined. In this study we infected mice deficient of IRF-5 with L. donovani and demonstrate for the first time that IRF-5 is essential to develop a protective response against this parasite. These results are important as they help us to understand the molecular mechanisms required for an immune response to fight L. donovani.
| The protozoan parasite Leishmania donovani is the causative agent of visceral leishmaniasis (VL), a chronic life threatening disease if untreated. In the experimental model of VL, the two main target organs are the liver and the spleen [1]. While the spleen stays chronically infected, infection in the liver is self-resolving within 6-8 weeks due to the development of a Th1-dominated granulomatous response, which is characterized by high IFNγ production. This response is induced by IL-12 secreted by dendritic cells (DC) [2], [3], [4] and is crucial for parasite control and disease resolution in the liver, together with TNFα production and expression of inducible nitric oxide synthase (iNOS) by macrophages [1]. Studies using Myd88-/- mice have highlighted the importance of toll like receptors (TLRs) in the induction of IL-12 production by DC and the development of Th1 immune responses in Leishmania infection [5]. More recently, TLR9 has been shown to be required for IL-12 production by DC in a model of cutaneous leishmaniasis [6], [7] and also in Trypanosoma cruzi infected mice [8]. However, in contrast to T. cruzi infections, TLR9 deficiency in mice infected with L. major did not prevent the development of Th1 responses and only resulted in a transient disease exacerbation [6], [9]. As MyD88-/- mice are highly susceptible to Leishmania infection [5], this suggests that in addition to TLR9, other TLRs as well as IL-1 and IL-18 may also be involved in the generation of Th1 responses and in the induction of host protective immunity. Since Leishmania parasites reside in the phagolysosomes of the host cells, other endosomally localized TLRs, such as TLR 7 and 8 could be involved in the recognition of this pathogen [10], [11].
Interferon Regulatory Factor 5 (IRF-5) has been shown to be involved in the transcriptional activation of both Type I IFN genes and genes encoding key proinflammatory cytokines such as IL-12, TNFα and IL-6 [12], [13], [14], [15]. This transcription factor can be activated by TLR7 and TLR9 via the MyD88 signaling pathway and/or directly by viral infections and Type I interferon [16]. In vivo, IRF-5 has been shown to play a role in the innate antiviral immune response. Indeed, lack of IRF-5 expression in genetically modified Irf5-/- mice resulted in attenuation of Type I IFN, TNFα and IL-6 production in response to viral infection [13], [17], [18]. However, the antiviral effect of IRF-5 deficiency appeared to be cell type specific and mainly affected DCs and plasmacytoid DCs (pDCs), rather than macrophages [16], [17]. More recently, IRF-5 was also shown to cooperate with, among others, NOD2 and TBK1 in triggering expression of Type I interferon in response to Mycobacterium tuberculosis [19].
The aim of this study was to examine whether IRF-5 also plays a role in the regulation of the immune response to parasitic infections. Here we demonstrate that IRF-5 deficiency results in severe impairment in the development of Th1 immune responses following L. donovani infection. Moreover, Irf5-/- mice failed to develop typical Th1-type granulomas and to control infection in the liver, demonstrating a vital role for IRF-5 in the induction of the anti-parasitic response.
The transcription factor IRF-5 is an important downstream regulator of the TLR/MyD88 signaling pathway and is involved in the induction of several key proinflammatory cytokines [13], [16], [17]. As TLRs have been implicated in the recognition of Leishmania parasites [5], [6], [7], [20], [21], [22], [23], we first wanted to assess whether IRF-5 was at all involved in the generation of protective immunity against L. donovani. Hence, we infected wild type (WT) and Irf5-/- mice and monitored the course of infection at several time points. We observed similar parasite burdens for WT and Irf5-/- mice in the spleen (Fig. 1A). In contrast, the disease was exacerbated in the livers of Irf5-/- mice, with a parasite burden of approximately two times higher at d14 pi and almost 4-fold higher than in the WT controls by d28 pi (Fig. 1B). These data suggest that the requirements for IRF-5 in the immune response to L. donovani are strictly organ specific, at least until day 28p.i..
As IFNγ-producing CD4+ T helper cells are crucial for the control of L. donovani infection in the liver, we next investigated whether these responses were being generated in Irf5-/- mice. IFNγ–producing CD4+ T cells were already detected at d14 of infection in both the liver and spleen of WT mice, although still at low frequency. The frequency of IFNγ+ CD4+ T cells increased in the later stages of infection and by d28 pi, 39.1% of CD4+ T cells in the liver (Fig. 2A and B) and 7.6% in the spleen (Fig. 2C and D) of WT mice were producing IFNγ. In marked contrast, the generation of IFNγ-producing CD4+ T cells was significantly impaired in the livers of Irf5-/- mice (Fig. 2A), where only 13.2% of CD4+ T cells were IFNγ+ at d28 (a 60% reduction), and in the spleen (Fig. 2C), where only 2.8% of CD4+ T cells were found to be secreting IFNγ (a 60% reduction compared to WT controls). Interestingly, at day 14p.i., the frequency of IFNγ+ CD4+ T-cells in the liver and spleens of Irf5-/- mice was comparable, if not better than in WT mice (Fig. 2). This suggests that initially Th1 responses are being generated in the absence of IRF-5, however this transcription factor is essential for the maintenance and further expansion of Th1 responses.
In order to determine whether the defective generation of Th1 responses detected in Irf5-/- mice had an impact on the Th1-type granuloma formation normally observed in the livers of L. donovani infected WT mice [1], we proceeded to analyze H&E stained sections of livers from WT and Irf5-/- mice. Despite an approximately 4-fold higher liver parasite burden in Irf5-/- mice (Fig. 1B), the number of granulomas/inflammatory foci observed in Irf5-/- mice at d28 p.i. was reduced approximately by 77% compared with WT mice (406.7±15.7 granuloma/100 microscopic fields in WT mice vs. 92±38 granuloma/100 microscopic fields in Irf5-/- mice). A closer histological analysis also revealed that the granulomas in Irf5-/- mice were smaller in size and had a different cellular composition, characterized by a marked infiltration of polymorphonucleated cells that were identified as neutrophils (Fig. 3A). Neutrophils were also observed at higher frequencies in the spleens of infected Irf5-/- mice compared to WT controls (Fig. S1 A and B). Increased neutrophil infiltration correlated with higher mRNA levels for CXCL1, a chemokine receptor known for its neutrophil chemoactractant activity, in the liver (day 14 p.i. only) and spleens of Irf5-/- compared to WT mice (Fig. 3 B and C). Of additional interest is the observation that the mononuclear cell infiltrates in the livers of Irf5-/- mice were significantly reduced at day 28 p.i. compared to WT mice (data not shown and Fig. 2B). Moreover, in Irf5-/- mice we could not observe any splenomegaly, which is a typical symptom of VL in WT mice (Fig. 3D). These results suggest that IRF-5 is an essential factor in the maintenance of the inflammatory responses generated during L. donovani infection. IL-23 has recently been shown to be involved in immune cell homing to infected target cells [24], hence we also assessed whether this cytokine was expressed at different levels in the livers of Irf5-/- mice. To our surprise, IL-23 p19 was slightly upregulated in Irf5-/- mice at day 14 p.i., but was induced at similar levels at later stages of infection in both WT and Irf5-/- mice (Fig. 3E).
Since Irf5-/- granulomas were not the typical Th1-type granulomas which develop in infected WT mice and, unlike WT mice, Irf5-/- mice failed to eliminate the parasites, we next examined the cytokine environment in the livers of both groups of mice. IRF-5 deficiency resulted not only in a very weak IFNγ responses (Fig. 2B), but also in a significantly higher expression of IL-4 and IL-10 mRNA compared to WT mice (Fig. 3 F). Interestingly, the differences in cytokine mRNA levels were observed only at d28 pi but not during the first 2 weeks of infection. Surprisingly, however, we did not observe any significant differences in the relative levels of TNFα (Fig. 3E), IL-13, or IL-5 (data not shown) between the two groups. Moreover, Irf5-/- and WT mice showed comparable frequencies of IL-10- and IL-17-producing CD4+ T cells at d28 pi (data not shown). Taken together, these data suggest that unlike in WT mice, L. donovani infection in Irf5-/- mice induces a very small inflammatory infiltration in the liver and results in the generation of an IL-4-dominated response, resulting in a failure to control parasite growth.
It is well established that the production of IL-12 by DCs is crucial for the development of Th1 cells [25], although an IL-12-independent mechanism for the induction Th1 responses has also been described [26]. A study in Myd88-/- mice infected with L. major has highlighted the importance of TLRs, IL-1 and/or IL-18 in the induction of IL-12 and the generation of Th1 responses; MyD88 deficiency also resulted in complete abrogation of IFNγ production by CD4+ T cells and an inability to control infection [5]. More recently, TLR9 activation has been shown to play a crucial role in the induction of IL-12 secretion by DCs following T. cruzi and Leishmania infections [6], [7]. As the development of IFNγ-producing CD4+ T cell in L. donovani infected mice was shown to depend upon IL-12 production by conventional CD11chi DC (cDC), we examined whether IRF-5 deficient DC were able to produce IL-12 in vivo following L. donovani infection. In agreement with the literature [27], we detected a moderate increase in IL-12p35 mRNA levels in cDCs isolated from infected WT mice (Fig. 4). However, induction of IL-12p35 was not detected in IRF-5 deficient cDCs. In contrast, IL-12p40 mRNA expression was induced in cDCs of Irf5-/- at similar levels as in cDC of WT mice (data not shown). Since the p40 component of IL-12 is shared with IL-23, the upregulation of p40 may be caused by an increased expression of IL-23 (s. Fig. 3E).
Inducible nitric oxide synthase (iNOS) is a key enzyme involved in the production of nitric oxide (NO), which has direct microbial toxicity and is also involved in the regulation of cytokine gene expression and cytokine responsiveness. NO is typically produced by classically activated macrophages upon triggering of IFN and TLR pathways that enhance the expression of iNOS [28], [29], [30]. In the absence of a strong IFNγ response and the presence of an IL-4 dominated immune response in infected Irf5-/- mice, we were curious as to whether iNOS (NOS2) was still able to be induced at all in Irf5-/- mice. To our surprise, we found no difference in the induction of iNOS mRNA in the livers of Irf5-/- and WT mice (Fig. 5A). As the absence of IRF-5 appeared to be leading to a Th2-like state in the liver, we were also interested in determining whether markers for the alternative activation of macrophages were being induced in mice deficient in IRF-5. Using real time PCR we only found a slight increase in the induction of Arg1 (Fig. 5B) and Fizz1 (Fig. 5C) in the livers of infected Irf5-/- mice compared to WT mice.
Since iNOS mRNA levels were similar in Irf5-/- and WT mice, we next investigated whether this message was being translated into protein. iNOS protein was expressed in the livers of infected WT mice, mainly in granulomas (Fig. 5D). In contrast, it was not detected in infected Irf5-/- mice at any time point after infection (Fig. 5D).
To determine whether IRF-5 directly regulates transcription of iNOS, we used a luciferase reporter assay, where luciferase was under the control of the NOS2 promoter cotransfected with MyD88 and/or IRF-5 expression plasmids. As shown in Figure 5E and in agreement with the literature [28], [30], MyD88 expression induced a 7.86 fold increase in luciferase activity. In contrast, IRF-5 failed to induce luciferase activity, suggesting that overexpression of inactivated IRF-5 alone does not stimulate the transcriptional activity of the NOS2 promoter. Co- transfection with both MyD88 and IRF-5 expression plasmids resulted only in a slight increase in the transcriptional activity of the NOS2 promoter. Taken together, these results indicate that IRF-5 does not stimulate transcription of iNOS but may play a role to indirectly regulate the iNOS response to L. donovani at the posttranscriptional level.
We next infected WT and Irf5-/- bone marrow-derived macrophages in vitro with L. donovani and monitored the iNOS mRNA expression by real time PCR. In agreement with the results shown in Fig. 5E, IRF-5 deficiency did not affect the iNOS mRNA level (Fig. 5F) or the NO production (Fig. S2 A) following L. donovani infection, suggesting that IRF-5 does not directly induce transcription of iNOS. However, when we stimulated WT and Irf5-/- macrophages with LPS, we could see a marked decrease in both iNOS mRNA transcipts (Fig. 5F) and NO production (Fig. S2 A). This was possibly caused by a defective pro-inflammatory response in Irf5-/- macrophages (Fig. S2 B). This defect was not observed in Irf5-/- macrophages infected with L. donovani (Fig. S2 B), suggesting that the induction of pro-inflammatory cytokines in macrophages infected in vitro with L. donovani is IRF-5 independent. This data implies that the inflammatory signals downstream of TLR triggering largely contribute to the induction of iNOS. Thus it is possible that the defective pro-inflammatory response observed in L. donovani infected Irf5-/- mice (Fig. 2 and 3) is mainly responsible for the reduction in iNOS expression in these mice.
Since IRF-5 seems to be involved in the maintenance of the inflammatory response to L. donovani and in the development of protective Th1-responses, we were interested in which cells IRF-5 expression is essential for displaying and/or inducing anti-leishmanial effector function. Thus, we analyzed the cell-specific expression pattern of IRF-5 mRNA in the liver and the spleen at different time points during infection (Fig. 6). Interestingly, in the spleen IRF-5 mRNA was only upregulated in the CD5+ population, corresponding to T-cells. The increase in IRF-5 mRNA levels was only detected at d28 p.i. (Fig. 6A). Splenic B-cells, macrophages, dendritic cells, and neutrophils did not show increased expression of IRF-5 mRNA at any time point analyzed (data not shown). In the liver, only the CD5 positive fraction (T-cells) had upregulated IRF-5 mRNA at d14 p.i. (Fig. 6B). However by d28 p.i., only few T-cells were present in the liver (Fig. 2 and 3) and at this time point T-cells did not express IRF-5 mRNA at levels greater than in naïve mice.
We and others have previously shown that IRF-5 can be activated by TLR7 [16] and TLR9 [13] via the MyD88 signaling pathway. Indeed, it was recently shown that Leishmania infections in Tlr9-/- mice induced transiently deficient Th1 responses [7], [9], suggesting that the regulation of these responses in Leishmania infections might also be governed by other pathogen recognition pathways. Thus, we next investigated which TLR was involved in the activation of IRF-5 and consequently in the modulation of Th1 responses following L. donovani infection.
Hence we proceeded to assess whether TLR7 was involved in the recognition of L. donovani and in the generation of parasite-specific Th1 responses. To our surprise, L. donovani infection in Tlr7-/- mice resulted in a approximately 3-fold higher hepatic parasite burden at day 28pi compared to WT mice (Fig. 7A); as seen in Irf5-/- mice, the splenic parasite burden in Tlr7-/- mice was similar to the WT control group (data not shown). These results suggest that TLR7 plays a role in the recognition of L. donovani parasites and in the generation of protective immune responses against Leishmania.
We then determined the frequency of IFNγ+ CD4+ T-cells in infected Tlr7-/- mice and WT controls. At day 14 p.i the frequency of IFNγ producing CD4+ T-cells was similar in the liver of Tlr7-/- mice compared to the WT control group. However, at day 28, only 6.4% of CD4+ T-cells in the liver of Tlr7-/- mice was producing IFNγ compared to 14.5% in WT mice (Fig. 7B and S3). A similar reduction in IFNγ production was observed in the spleen (data not shown). Interestingly, the extent of the inflammatory cell infiltration in the livers of Tlr7-/- mice at day 28 p.i. was comparable to WT mice (data not shown). Taken together, these results suggest that the activation of TLR7 is crucial for the development of Th1 responses following L. donovani infection.
Since TLR7 not only signals through IRF5 but also through IRF7 [31] and IRF7 has recently been shown to regulate killing of intracellular Leishmania in marginal zone macrophages [32], we wanted to determine whether IRF7 contributed to the defect in Th1 responses we observed in Tlr7-/- mice. Hence, we infected Irf7-/- mice with L. donovani amastigotes and compared the development of Th1 responses in these mice to infected WT controls. The results revealed that in Irf7-/- mice, L. donovani induced IFNγ-producing CD4+ T-cells at frequencies comparable to infected WT mice (Fig. 7C and D).
Finally, we compared the cell-specific IRF-5 mRNA expression pattern in WT and Tlr7-/- mice at d14 and d28 p.i.. As previously shown (Fig. 6), only the CD5+ fraction (T-cells) had upregulated IRF-5 mRNA in the spleen of infected WT mice (Fig. 7E). The IRF-5 mRNA expression was not significantly higher in B-cells and non-B/non-T-cells from infected WT mice compared to the corresponding naïve WT cell populations (data not shown). Interestingly, we could not detect any upregulation of IRF-5 mRNA expression in splenic T-cells from infected Tlr7-/- mice at d28 p.i., suggesting that TLR7 is essential for the induction of IRF-5 in T-cells in the spleen at later stages of infection. In contrast, in the liver, IRF-5 mRNA expression was only upregulated in T-cells purified at day 14 p.i. from infected WT and Tlr7-/- mice (Fig. 7F). This suggests that increased IRF-5 expression in T-cells is independent of TLR7 during the early stages of infection in the liver. No IRF-5 mRNA could be detected at d28 p.i. in the liver of WT and Tlr7-/- mice (Fig. 7F).
Taken together, these data suggest that IRF-5 is a key molecular switch for the development of Th1 responses following L. donovani infection, and that TLR7-mediated IRF-5 activation plays a critical role during chronic infection.
In the present study we have demonstrated that the TLR7-mediated activation of IRF-5 is required for the development of host-protective Th1 responses to L. donovani at later stages of infection. Moreover, IRF-5 deficiency resulted in an IL-4 dominated response, reduced iNOS expression, and failure to control parasite growth in the liver. To our knowledge, the role of IRF-5 in modulating adaptive T-cell responses is a novel and previously undescribed finding.
One of the key points in the induction of a protective immunity to Leishmania parasites is the generation of IFNγ-producing CD4+ T-cells. Although IL-12 production by DC is crucial for the development of Th1 cells [25], the mechanism leading to the generation of these responses remains elusive. A study in Myd88-/- mice infected with Leishmania has highlighted the importance of TLRs, IL-1 and/or IL-18 in the induction of IL-12 and the generation of Th1 responses. Indeed, MyD88 deficiency resulted in complete abrogation of IFNγ production by CD4+ T-cells and an inability to control infection [5]. More recently, TLR9 activation was shown to play a crucial role in the induction of IL-12 secretion by DCs following T. cruzi and/or Leishmania infections [6], [7]. However, unlike T. cruzi infections, attenuation of Th1 responses to Leishmania infection in Tlr9-/- mice was only transient [7], [9]. These studies suggest that the regulation of Th1 responses in Leishmania infections is not exclusively mediated by TLR9 and might be governed by other pathogen recognition pathways as well. Our study pinpoints IRF-5, a transcription factor in the MyD88-mediated TLR pathways, as an essential factor in the development of adaptive Th1 responses following Leishmania infection. Interestingly, while Th1 responses in Irf5-/- mice were severely impaired 4 weeks after infection, expression of IRF-5 was not required during the first 2 weeks of infection, since the frequency of IFNγ-producing CD4+ T-cells in Irf5-/- mice was comparable to WT mice. This observation suggests that IRF-5 is not essential for the early induction of Th1 responses but is crucial for their further development and expansion. Although IRF-5 has been shown to have a major impact on the innate immune response to viral infections [17], [18], its role in shaping the development of adaptive Th1 responses has not been previously demonstrated.
Defective Th1 responses were also observed in Tlr7-/- mice infected with L. donovani. IRF-5 and IRF-7 are both known to be activated by the MyD88-dependant TLR7 signaling pathway [16], [31], however IRF-7 deficiency did not result in defective Th1 responses to L. donovani. TLR7 has been shown to recognize single-stranded RNA [33], [34] and to play an essential role in the immunity to ssRNA virus [24], [35]. Recognition of a DNA virus, the murine cytomegalovirus, by TLR7 has also been reported [36]. To date there is no indication of TLR7 activation by parasites.
The analysis of cell-specific IRF-5 mRNA expression pattern revealed that the only cells that upregulated IRF-5 during L. donovani infection at d14 and 28 p.i. were T-cells. This does not exclude, though, that other cell types such as DCs and macrophages may express IRF5 during earlier stages of infection. Nevertheless, IRF-5 expression in T-cells differed between the liver and the spleen: in the liver IRF5 expression is upregulated during the first 2 weeks of infection in a TLR7-independent manner; in contrast, in the spleen the upregulation occurs 4 weeks into the infection and appears to be mediated by TLR7. Since the defect in the Th1 responses seems to mainly affect the hepatic infection, at least until d28 p.i., and IRF-5 only appears to be upregulated in T-cells in the spleen during chronic infection, it is tempting to speculate that effector Th1 cells generated in the spleen are required for controlling parasite growth in the liver. Future experiments involving splenectomized mice should be able to prove this hypothesis.
It is possible that the upregulation of IRF-5 in T-cells may be required depending on the subsets and activation status of the cells. A role for IRF-5 in T-cells is yet unknown. Whether IRF-5 expression by T-cells is directly mediated by TLR 7/9 triggering or indirectly induced by Type I IFN, produced by APC following TLR7/9 signaling, is still an open question. Recent studies have highlighted a role for MyD88 in T-cells [37], [38], supporting the possibility of a direct TLR7/9 –mediated IRF-5 induction. Mice with MyD88 deficient T-cells only develop defective Th1 responses in a Toxoplasma gondii model [37], suggesting that MyD88 signaling in T-cells is essential for Th1 cell development. Human T-cells purified from HIV [39] and from Hepatitis C [40] patients were also shown to express TLR7 and/or TLR9. Why TLR7/IRF-5 activation is only required at later stages of infection and what role it plays in various T-cell subsets are two more questions that remain yet to be answered. Future investigations will address these questions using T-cell-specific IRF-5 and TLR7 knockout mice.
Interestingly, the severe defect in Th1 responses seen in Irf5-/- mice affected the hepatic but did not seem to have any effect on the splenic parasite burden, at least until day 28 p.i.. This is in agreement with a study by Engwerda and colleagues who have demonstrated that IL-12 neutralization did exacerbate infection in the spleen during the first 4 weeks of infection, even though IFNγ production, which is mainly derived from CD4+ T-cells, was severely reduced [2]. Thus, in contrast to the liver, Th1 responses are not critical for controlling parasite growth in the spleen during the first 28 days of infection. These observations imply that requirements for protection against L. donovani in the liver and in the spleen are very distinct, and underline again the organ-specific nature of the immune response during VL.
IRF-5 deficiency also resulted in a dramatic decrease in the extent of the inflammatory cell infiltration in the liver at day 28 p.i.. Moreover, L. donovani failed to induce splenomegaly in Irf5-/- mice, which is characteristic for L. donovani infection in WT mice. Since TLR7 deficiency did not significantly impair the recruitment of inflammatory cells to the liver of L. donovani infected mice, we can assume that there is some redundancy between different TLRs recognizing different PAMPs, but commonly utilizing IRF-5 in the induction of the inflammatory response during VL. Such overlapping effects between TLR7 and TLR9 have been observed during murine cytomegalovirus infection [36]. However, TLR7 and TLR9 were recently shown to have distinct effects in a murine model of Lupus [41] and also during experimental West Nile Encephalitis, where Tlr7-/- mice, but not Tlr9-/- mice, showed an impaired CD45+ leukocyte and macrophage infiltration at the site of infection [24]. Failure of these cells to migrate to infected target organs was caused by a significant reduction in IL-23 responses [24]. In L. donovani infected Irf5-/- mice, the level of IL-23 p19 were slightly higher compared to WT mice, suggesting that the lack of lymphocyte infiltration in the liver was not caused by the decreased induction of IL-23, but by some other yet unidentified pathways. Furthermore, the frequency of IL-17 producing cells, which are typically induced by IL-23 [42], was comparable in both groups of mice. This indicates that IRF-5 deficiency only affected the development of Th1 responses, but not the generation of Th17 cells in L. donovani infected mice.
NO is an important leishmanicidal effector molecule. It has direct microbial toxicity and it is also involved in the regulation of cytokine gene expression and cytokine responsiveness. NO is typically produced by classically activated macrophages upon triggering of interferon and TLR pathways that enhance expression of iNOS [28], [29], [30]. Expression of iNOS in mice infected with L. donovani appears to be tissue-specific: this enzyme is induced in the liver, however only limited iNOS expression is observed in the spleen [43]. Despite defective IFNγ responses in Irf5-/- mice the level of iNOS mRNA expression was comparable to the WT control group. Since iNOS expression can also be triggered by TLR pathways, high levels of iNOS mRNA in Irf5-/- mice may most likely be due to the high parasite burden in the liver at d28pi. Interestingly though, iNOS protein was not detected in the livers of Irf5-/- mice. A possible explanation for the lack of iNOS protein in Irf5-/- mice is that iNOS may be competing with arginase 1. This enzyme is commonly linked to alternative pathway of macrophage activation [44]. However, arginase 1 can also be induced in classically activated macrophages and can reduce NO production through competition with iNOS for their common substrate arginine [45]. Arginase 1 was also shown in vitro to suppress translation and enzymatic activity of iNOS without affecting mRNA levels [46], [47]. Nevertheless, arginase 1 mRNA was only slightly increased in Irf5-/- mice compared to WT mice, an increase that is probably not sufficient to inhibit translation of iNOS protein. Another possible explanation is that the severe defect in pro-inflammatory and the concomitant increase in Th2 responses observed in Irf5-/- mice may be responsible for the lack of amplification of the iNOS expression. Further investigations are needed to clarify the role of IRF-5 in the molecular mechanisms involved in the regulation of iNOS production.
In conclusion, we have indentified IRF-5 as a critical component for the development of Th1 responses to Leishmania infection. Furthermore, IRF-5 is an essential factor for the maintenance of the inflammatory response and plays a role in the indirect regulation of iNOS expression during L. donovani infection. Further experiments have yet to determine the molecular mechanism by which IRF-5 affects the development of host protective immunity to L. donovani.
Ethics statement: All experiments were approved by and conducted in accordance with guidelines of the Animal Care and Use Committee of the Johns Hopkins University School of Medicine.
C57BL/6J mice were obtained from The National Cancer Institute (Frederick, MD, USA), and B6.129S7-Rag1tm1Mom/J from The Jackson Laboratory. All mice were housed in the Johns Hopkins University animal facilities (Baltimore, MD) under specific pathogen-free conditions and used at 6–8 weeks of age. Irf5-/- mice were a generous gift from Dr. T. Mak (University of Toronto, Canada) and were backcrossed to C57BL/6J for at least 10 generations. The SNP analysis confirmed that they are 100% C57BL/6. We have previously reported age related splenomegaly in Irf5-/- mice that was associated with changes in splenic architecture [48]. The background of these mice was only 92% C57BL/6. In this study we have used 6–8 weeks old Irf5-/- mice that have normal spleen size and splenic architecture. Tlr7-/- mice were a generous gift from Dr S. Akira (Osaka University, Japan). Tlr7-/-were backcrossed to C57BL/6J for 9 generations. Irf7-/- mice were a kind gift from Dr. T. Taniguchi (University of Tokyo, Japan). Irf7-/- were backcrossed to C57BL/6J for 8 generations. Leishmania donovani (strain LV9) parasites were maintained by serial passage in B6.129S7-Rag1tm1Mom/J mice, and amastigotes were isolated from the spleens of infected animals. Mice were infected by injecting 2×107 amastigotes intravenously via the lateral tail vein. Hepatic and splenic parasite burdens were determined by examining methanol-fixed, Giemsa stained tissue impression smears. Data are presented as Leishman Donovan Units (LDU).
Mice were euthanized at indicated time points. Mononuclear cells were purified from the liver as previously described [49]. Hepatic mononuclear cells and splenocytes were restimulated for 2 h at 37°C in the presence of bone marrow derived dendritic cells (BMDC) previously pulsed with paraformaldehyde fixed amastigotes. Brefeldin A was then added for a further 4 h. Cells were then stained with biotinylated anti-CD3 followed by PerCP-conjugated strepatvidin, FITC-conjugated anti-CD4, and APC-conjugated anti-IFNγ (all BD Bioscience). Flow cytometric analysis was performed with a LSRII flow cytometer (Becton Dickinson). 350,000 cells per sample were acquired and analyzed with the FACSDiva software.
Spleens from infected and naive C57BL/6 mice were divided into 2 groups. Group A: B-cells were first enriched from splenic single cell suspension using anti-B220 beads (Miltenyi Biotec) following manufacturer instructions. B-cells were then sorted to >98% purity using FACSVantage (Becton Dickinson) based on their expression of CD19 and B220. The B220 negative fraction was then incubated with anti-CD5 beads for the enrichment of T-cells. T-cells were then sorted to >98% purity using FACSVantage based on their expression of CD4+, CD8+ and NK1.1-. Group B: splenocytes were first incubated with anti-CD11c beads and conventional splenic DCs were then sorted to >98% purity based on their expression of CD11c and MHCII. The CD11c negative fraction was then incubated with anti-CD11b beads. CD11b+ cells were sorted into different populations based on their expression of Gr1, MHCII and CD11b.
Livers from infected and naïve C57BL/6 mice and from infected Tlr7-/- mice were incubated with anti-CD5 beads in order to enrich T-cells. The purity of the T-cell preparation was >87%. Spleens from infected WT and Tlr7-/- mice were first incubated with anti-B220 beads; the B220 negative fraction was then incubated with anti-CD5 beads. >82% of the B220+ cells were B-cells; >94% of the CD5+ cells were T-cells.
Macrophages were differentiated from the bone marrow following red blood cell lysis. Cells were incubated with 30 ng/ml M-CSF (RnDSystems) for 5 days. Prior to use cells were counted and seeded at the 5×105 cells/well in a 24 well plate, rested in the absence of M-CSF for several hours and then treated with LPS (1 µg/ml) (Sigma-Aldrich) or L. donovani (MOI 10) for 1 h. Media was then changed to remove any extracellular parasites and cells incubated in fresh media for a further 18 h. At experimental end point supernatant was collected and cells washed in PBS and lysed for RNA extraction.
For the analysis of the IL-12p35/IL-12p40 mRNA induction, WT and Irf5-/- mice were infected with 2×107 L.donovani amastigotes. 5 h later, mice were euthanized and CD11c+ dendritic cells were isolated from splenic single cell suspensions using anti-CD11c beads (Miltenyi Biotec) following manufacturer instructions. The purity of the preparation was about 85%. RNA was extracted using Trizol (Invitrogen) as per manufacturer' instructions. For purified cell populations listed in the previous section RNA was extracted using the RNEasy Mini Kit (Qiagen). Reverse transcription was performed using the QuantiTect Reverse Transcription kit (Qiagen). SybrGreen was used to assay beta actin [17], IL-12p40 and IL-12p35. The following primers were used: IL-12p35 F- CCACCCTTGCCCTCCTAAAC and R-GGCAGCTCCCTCTTGTTGTG; IL-12p40 F- CTTGCAGATGAAGCCTTTGAAGA and R- GGAACGCACCTTTCTGGTTACA. For infected mice, livers and spleens were collected at indicated time points and whole tissue RNA extracted using Trizol, cDNA generated using the QuantiTect Reverse Transcription kit. Other than IL-12p35 and IL-12/23p40 all gene expression was analyzed using Taqman primers with a StepOnePlus cycler (Applied Biosystems).
Nitric oxide was quantitated using the Greiss assay (Promega) according to manufacturer's instructions.
For frozen sections, livers from infected mice and uninfected controls were embedded in OCT (TissueTek), snap frozen and 5 µm sections cut. Sections were stained for iNOS/NOS2 using an anti-iNOS polyclonal antibody generated in rabbit (Chemicon/Millipore) and detected using anti-rabbit IgG conjugated to HRP and DAB substrate (both Vector Laboratories). Paraffin-embedded sections were prepared from liver fixed in neutral buffered formalin, cut to 5 µm and stained by H&E. All immunohistochemistry was analyzed by light microscopy and photographs taken at the indicated magnifications.
HEK293 cells were seeded at 1×105 cells/well in 96 well plates. Cells were transfected with the NOS2-luc reporter construct (50 ng) (Addgene) and Renilla luciferase (10 ng) with either MyD88 (50 ng) and/or murine IRF-5 (100 ng) expression plasmids. 24 hours after transfection cells were lysed and assayed for luciferase activity using the Promega Dual-luciferase reporter assay. Luciferase activity was normalized against Renilla. Data shown is the mean± SEM for triplicate samples from 2 independent experiments.
Results were analyzed using an unpaired Student t-test. P<0.05 was considered significant. Real time PCR results were analyzed using an unpaired Student t-test or the Mann-Whitney test. For bone marrow derived macrophages the non parametric t-test with Welch's correction was used. P<0.05 was considered significant. Experiments were repeated at least twice.
|
10.1371/journal.pcbi.1002552 | Simple Epidemiological Dynamics Explain Phylogenetic Clustering of HIV from Patients with Recent Infection | Phylogenies of highly genetically variable viruses such as HIV-1 are potentially informative of epidemiological dynamics. Several studies have demonstrated the presence of clusters of highly related HIV-1 sequences, particularly among recently HIV-infected individuals, which have been used to argue for a high transmission rate during acute infection. Using a large set of HIV-1 subtype B pol sequences collected from men who have sex with men, we demonstrate that virus from recent infections tend to be phylogenetically clustered at a greater rate than virus from patients with chronic infection (‘excess clustering’) and also tend to cluster with other recent HIV infections rather than chronic, established infections (‘excess co-clustering’), consistent with previous reports. To determine the role that a higher infectivity during acute infection may play in excess clustering and co-clustering, we developed a simple model of HIV infection that incorporates an early period of intensified transmission, and explicitly considers the dynamics of phylogenetic clusters alongside the dynamics of acute and chronic infected cases. We explored the potential for clustering statistics to be used for inference of acute stage transmission rates and found that no single statistic explains very much variance in parameters controlling acute stage transmission rates. We demonstrate that high transmission rates during the acute stage is not the main cause of excess clustering of virus from patients with early/acute infection compared to chronic infection, which may simply reflect the shorter time since transmission in acute infection. Higher transmission during acute infection can result in excess co-clustering of sequences, while the extent of clustering observed is most sensitive to the fraction of infections sampled.
| Diversity of viral genetic sequences depends on epidemiological mechanisms and dynamics, however the exact mechanisms responsible for patterns observed in phylogenies of HIV remain poorly understood. We observe that virus taken from patients with early/acute HIV infection are more likely to be closely related. By developing a mathematical model of HIV transmission, we show how these and other patterns arise as a simple consequence of intensified transmission during the early/acute stage of HIV infection, however observing these patterns is highly dependent on sampling a significant fraction of prevalent infections.
| Phylogenetic clusters of closely related virus such as HIV arise from the epidemiological dynamics and transmission by infected hosts. If virus is phylogenetically clustered, it is an indication that the hosts are connected by a short chain of transmissions [1].
If super-infection is rare, and assuming an extreme bottleneck at the point of transmission, each lineage in a phylogenetic tree corresponds to a single infected individual with its own unique viral population [2], [3]. A transmission event between hosts causes an extreme bottleneck in the population of virus in the new hosts. For infections between MSM, it is estimated that infection is initiated by one or several virions [4], [5]. At the time of transmission, the quasispecies of virus within the transmitting host diverges and can thereby generate a new branch in the phylogeny of consensus viral isolates from infected individuals [6]. Transmissions in the recent past should be reflected by recently diverged lineages, and transmissions from long ago should reflect branches close to the root of a tree. [7]. Viruses such as HIV which have a high mutation rate relative to epidemiological spread can generate epidemics such that the correspondence between transmission and phylogenetic branching is most clear [2].
Given a phylogeny of virus reconstructed from samples, the phylogenetic clusters are a partition of the sample units into disjoint sets as a function of the tree topology. A cluster will consist of all taxa of the tree that are descended from a given lineage on the interior of the tree. There are many variations of this idea, and there is no general agreement about how to choose interior lineages for defining clusters. The most common algorithms require strong statistical support for a monophyletic clade among all taxa in a cluster [8]–[14]. These definitions may additionally require all taxa in a cluster to be connected by short branches with less than a threshold length [11], or similarly require that the genetic sequences corresponding to each taxon be separated by a genetic distance less than a given threshold [8], [14]. Definitions of clustering based on statistical support for monophyly are very difficult to operationalize in a mathematical model, and in particular, it is not clear how the statistical significance of internal nodes relates to population dynamics. Consequently, we have devised a conceptually similar definition of clusters that relies on the estimated time to most recent common ancestor (TMRCA) of a set of taxa [15]. A formal definition is provided below.
The sizes of the groupings that arise from a clustering algorithm have been interpreted as a reflection of the heterogeneity of epidemiological transmission. The distribution of cluster sizes of HIV is often skewed right, and depending on the definition of clustering used, can have a heavy tail [14], [15]. This is consistent with the prevailing view among modelers of sexually transmitted infections that there is a skewed and in some cases power-law distribution in the number of risky sexual contacts in the population, however it is not straightforward to make inferences about sexual network properties from cluster size distributions [16]. In the case of HIV, the distribution of branch lengths within clusters may also reflect the disproportionate impact of early and acute HIV infection on forward transmission, which is due to higher viral loads in the early stages of infection, higher transmissibility per act [17], and fluctuating risk behavior [18].
When the taxa of the phylogeny are labeled, such as with the demographic, behavioral or clinical attributes of the the individuals from whom the virus was sampled, one can further analyze statistical properties of clustered taxa. Similar taxa, such as those arising from acute infections, may cluster together (or co-cluster) at greater rates. Patterns of co-clustering might be informative about the fraction of transmissions that occur at different stages of infection or between different demographic categories. HIV phylogenies from men who have sex with men (MSM) have been widely observed [12], [13], [19] to have individuals with early/acute HIV infection that are much more likely to appear in a phylogenetic cluster. And moreover, if early-stage individuals are in a cluster, they are much more likely to be clustered with other early infections. Both Lewis et al. and Brenner et al. [8], [9] have hypothesized that co-clustering of early infection is caused by higher transmissibility per act during early infection. For example, in phylogenies with time-scaled branch lengths, if a large fraction of clusters have a maximum branch length of six months [8], [15], this suggests that at least that fraction of transmissions also occur within six months. In this article we demonstrate that the mechanisms that generate co-clustering of early infections are complex, and involve many attributes of the epidemic in addition to higher transmissibility per act [17]. To summarize, several features of the phylogenetic structure of HIV in MSM have been independently observed by several investigators:
Below, we illustrate these clustering patterns using 1235 HIV-1 subtype B pol sequences collected between 2004 and 2010 in Detroit, Michigan, USA.
These common clustering features motivate several questions. How informative are clustering patters about the underlying epidemic? In particular, how does higher transmissibility per act during early infection shape the phylogeny of virus ? To address these questions, we have developed a simple mathematical framework that demonstrates the connection between epidemiological dynamics and the expected patterns of clustering from a transmission tree and the corresponding phylogeny.
Our modeling work suggests that common features of HIV phylogenies are not coincidences, but universal features of certain viral phylogenies. We expect to see similar patterns for any disease such that the natural history features an early period of intensified transmission. High transmission rates during early infection may be a consequence of higher transmissibility per act due to high viral loads, but are also influenced by behavioral factors, such as fluctuating risk behavior [18], concurrency [20], and a lack of awareness of the infection. We do not explicitly model immunological or behavioral factors, but rather consider a compound parameter that describes the rate of transmission during the early/acute period. We find that while higher transmission rates increase the frequency of early/acute clustering, virus collected from early/acute patients clusters at a higher rate even when transmission rates are uniform over the infectious period.
This research was reviewed by the Institutional Review Board at the University of Michigan. Data used in this research was originally collected for HIV surveillance purposes. Data were anonymized by staff at the Michigan Department of Community Health before being provided to investigators. Because this research falls under the original mandate for HIV surveillance, it was not classified as human subjects research.
Our analysis consists of an empirical component which establishes clustering patterns for a geographically and temporally delineated set of HIV sequences, and an analytical component which establishes a possible mechanism that could generate the observed patterns.
We examined the phylogenetic relationships of 1235 HIV-1 subtype B partial-pol sequences originally collected for drug-resistance testing. All sequences were collected in the Detroit metropolitan statistical area between 2004 and 2010. Sequences were tested for quality and subtype using the LANL quality control tool [21]–[23], and aligned against a subtype-B reference (HXB2).Drug resistance sites [24] were treated as missing data.
A maximum clade credibility phylogeny was estimated with BEAST 1.6.2 [25]. The phylogeny was estimated using a relaxed molecular clock and and HKY85 model of nucleotide substitution with Gamma rate variation between sites (4 categories). The MCMC was run for 50 million iterations with sampling every iterations. The first million iterations were discarded. The effective sample size of all parameters exceeded 50.
The phylogeny was converted into a matrix of pairwise distances between taxa expressed in units of calendar time. The distance between a pair of taxa was the TMRCA estimated by BEAST. Taxa were then classified into clusters using hierarchical clustering algorithms. A pair of taxa were considered to be clustered if the estimated TMRCA did not exceed a given threshold, and a range of thresholds was examined, from 0.5% of the maximum distance to the distance corresponding to the point where 90% of taxa are clustered with at least one other taxon.
Co-clustering of early/acute infections was investigated using a clinical variable (CD4 count) and a measure of genetic diversity of the virus. Both CD4 and sequence diversity are imprecise indicators of stage of infection. Nevertheless, with a large population-based sample, even noisy indicators of stage of infection are useful for illustrating phylodynamic patterns.
In most cases, CD4 counts were assessed contemporaneously with samples collected for sequencing. The CD4 cell counts can be informative about disease progression and can be used as a noisy predictor of the unknown date of infection [26]. Individuals with very high cell counts are unlikely to represent late/chronic infections, and we hypothesize that virus from these patients will be more likely to be phylogenetically clustered. Clustering of patients with high CD4 was previously observed by Pao et al. [10]
Recent work [27] has also highlighted the potential for sequence diversity to be informative of the date of infection. The frequency of ambiguous sites (FAS) in consensus sequences provides an approximate measure of sequence diversity in the host. HIV infection is initiated by one or a few founder lineages [4], [5]; initially the diversity of the viral population within the host is low, but diversity increases steadily over the course of infection [28]. By convention, consensus sequences report ambiguous sites as those where the most frequent nucleotide is read with a frequency less than 80%. We hypothesize that having few ambiguous sites is an indicator of early/acute infection; sequences with fewer ambiguous sites will be more likely to be in a phylogenetic cluster and to be clustered with other sequences with few ambiguous sites.
A simple analysis was conducted to establish the existence of excess clustering and co-clustering in the Michigan sequences. This analysis is not designed to classify our sample into a early/acute component or to estimate the date of infection for each unit.
To illustrate excess clustering of early/acute infections, we calculated the mean CD4 cell count and FAS for each sample unit in a phylogenetic cluster. Because all clustering thresholds are arbitrary, we explored a large range of values, up to the point where 90% of the sample was clustered with at least one other unit. The standard error of the estimated mean was calculated assuming simple random sampling. For small threshold distances, very few taxa are clustered, and the standard error is large, but decreases monotonically as the threshold is increased and more taxa are clustered.
To illustrate excess co-clustering, we classified taxa into three categories of CD4: those with CD4 , representing AIDS cases; those with CD4 , and those with CD4 between 200 and 800. Taxa were also classified into quartiles by FAS. We then counted the number of pairwise clusterings of taxa within and between each category. These counts were arranged in a matrix. Large counts along the diagonal (within categories) represent co-clustering by stage of infection. To establish excess co-clustering, we compared the counts to the expectation if clusters were being formed at random, e.g. if two taxa were selected uniformly at random without replacement.We denote the symmetric matrix of co-clustering counts as , so that represents the number of times that a taxon in category is clustered with a taxon in category . The sum of counts in the 'th row of will be denoted . Following the methods described in [29], the expected value of under random pair formation isBelow, we illustrate the difference . We can also calculate the assortativity coefficient [29], , which describes the total amount of co-clustering in the matrix. To construct the co-clustering matrices, we selected the value of the distance threshold which maximized the assortativity coefficient.
Following the approach outlined in [6] and [30], we develop a deterministic coalescent model derived from a compartmental susceptible-infected-recovered (SIR) model. A system of several ordinary differential equations describe the dynamics of prevalence of early and late HIV infection. Individuals pass from a susceptible state, to an early/acute infection state, to a chronic infection state followed by removal (treatment or death). , and will denote the numbers susceptible, acute, and chronically infected respectively, and the population size will be denoted . For didactic purposes, we will suppose that treatment is completely effective at preventing forward transmissions. The HIV model is described by the following equations:(1)In these equations, and are respectively the frequency-dependent transmission rates for early and chronic infected individuals. The average duration of early and chronic infection are respectively and . Natural mortality occurs at the rate and immigration into the susceptible state occurs at the rate , which maintains a constant population size . is a term which modulates the way incidence of infection scales with prevalence. For the results presented below, we choose . This term corrects for observed patterns of decreasing incidence with prevalence; this can occur as a result of population heterogeneities (including sexual network structure) or as the result of decreasing risk behavior as knowledge of the epidemic spread. Many more relevant details could be included in a model of the HIV epidemic in MSM, however our purpose is to demonstrate how these simple dynamics lead to observed phylogenetic patterns.
In [6], a similar HIV model was presented along with a method to fit such models to a sequence of phylogenetic divergence times (the heights of nodes in a time-scaled phylogeny). Where possible, we will use the parameter estimates from [6]. The parameters are reported in table 1. Together, these parameters imply and that 41% of transmissions occur during the acute stage.
Corresponding to an epidemic model of the form 1, we can define a coalescent process [31], [32] that describes the properties of the transmission tree and by extension the phylogeny of virus. The taxa descended from a lineage at time in the past form a clade, which we will also call a cluster. The number of taxa in a randomly selected cluster will be a random variable. The cluster size distribution (CSD) is a function of a threshold TMRCA , and describes the probability of having a size cluster if a lineage (i.e. branch) at time is selected uniformly at random from the set of all lineages at and the size of the cluster descended from that branch is counted. A schematic of how clusters and the CSD are constructed given a tree and a threshold is shown in figure S5. In [6] we derived differential equations that describe the moments of the CSD.
Some of the properties of phylogenies that we seek to reproduce with the model developed below are:
Figure 1 shows a simple genealogy that could be generated by the HIV model. Four events can occur in this genealogy representing coalescence or the changing stage of a lineage. By quantifying the rate that these events occur using a coalescent model, we can calculate the clustering properties of these genealogies. These methods are described below and in detail in supporting Text S1.
The ancestor function is strictly decreasing in reverse time and converges to one (a single lineage) when the most recent common ancestor of the sample is reached. The initial value of the ancestor function (when the population is sampled) is equal to the sample size . For the purposes of modeling phylogenetic properties of HIV, we will be interested in phylogenies such that the taxa are labeled with the state of the sampled individual (e.g. the individual will have early or late infection corresponding to the states in equation 1). In this case, we will have two ancestor functions, since a lineage may correspond to an infected individual with either early or late infection.
The ancestor functions derived from equations 1, and which are derived in the Text S1 are as follows:(2)In these equations, is the number of lineages corresponding to early infections and is the number of lineages corresponding to late infections. These equations provide a deterministic approximation to the NLFT, which is . Each term in these equations accounts for loss or gain of lineages due to the concurrent processes of transmission (at rates and ) and transition between states (at rates ). This approximation becomes exact in the limit of large sample and population size. Note that since the model is continuous in both time and state variables, the ancestor functions are not integers in contrast to most coalescent frameworks based on discrete mathematics.
Real epidemics in a finite population will have transmission trees such that the number of lineages at any time is a random variable. The mean-field model presented in equation 1 can be viewed as a description of the dynamics of a stochastic system in the limit of large population size. In this case, we can adapt the coalescent to make approximate descriptions of the stochastic properties of the transmission tree in large populations. The ancestor functions will reflect the approximation of the actual (random) number of lineages. Previous work has demonstrated that deterministic descriptions can be excellent approximations for the number of lineages over time [6], [33]. In the following section, we compare our deterministic coalescent to stochastic simulations, confirming that it is a good approximation over a wide range of parameters.
Given a clustering threshold TMRCA , the random variable will be the number of stage- taxa descended from a given lineage that is extant at time in the past. As before, will be the number of type lineages at the time in the past. In our model, infected can be of two types (early/acute and chronic infected), so there are only two types: corresponds to earl/acute and corresponds to chronic. We will denote the set of all lineages of type at time in the past as . Then we define the and 'th moment of cluster sizes descended from a type lineage to be(3)
Many summary statistics that are potentially informative about transmission dynamics can be derived from these moments. The moments are difficult to interpret, so in practice we use them to calculate summary statistics such as variance and skew of the CSD. Below, we examine 30 summary statistics derived from the first three moments and multiple clustering thresholds.
For example, the variance of cluster sizes counting only type taxa descended from type lineages is(4)The total variance of cluster sizes counting only stage 1 taxa is found with the weighted average over lineage types:(5)A similar set of equations can be developed for the cluster sizes aggregated over taxon types, that is, for . Detailed derivations are provided in Text S1 for differential equations that describe these moments as function of the threshold .
Event-driven stochastic simulations were conducted to verify the suitability of the deterministic approximations for inference. Simulations implemented a variation on the Gillespie algorithm [34]. Populations consisted of agents, and were simulated for 15 or 30 years starting with one hundred initial infections. At the end of each simulation, a sample of either 20% or 100% of prevalent infections was taken and used to reconstruct a transmission tree. Five hundred simulations were conducted for each sample fraction and sample time. Corresponding to each simulation, 10 transmission trees were generated based on a random sampling of using distinct clustering thresholds. The CSDs were then estimated from each tree and the moments of these distributions were compared to the moment equations (3–5).
We have further conducted an investigation into the potential of various summary statistics of the viral phylogeny for inference of underlying epidemiological parameters. Of particular interest is the fraction of transmissions that occur during early HIV infection. As indicated above, it is possible that phylogenetic clustering of early infections reflects elevated transmission during early/acute HIV infection, which we will define as the infectious period from zero to six months. The following simulation experiment was carried out to identify informative statistics:
The coalescent tree was simulated such that the sample size matched that of the Detroit MSM phylogeny, and the heterochronous sampling of that phylogeny was reproduced in the coalescent tree. Furthermore, the number of early/acute versus chronic taxa sampled was determined using the BED test for recency of infection for each patient [36], and simulations were also made to match the numbers of early/acute and chronic taxa sampled. Virus from patients with early/acute infection accounted for 24% of the samples.
Summary statistics were centralized around the mean and rescaled by their standard deviation (). The dependent variable of interest is the fraction of transmissions attributable to the acute stage at the beginning of the epidemic, which may be defined(6)where is the expected number of transmissions generated during early/acute infection at the beginning of the epidemic, and is the expected number of transmissions over the entire infectious period. Pearson correlation coefficients were calculated for each statistic and . To give a better indication which statistics would be useful for estimating the ratio of acute to chronic transmission rates, we conducted a partial least-squares (PLS) regression [37], which has been used by other investigators when estimating parameters by approximate Bayesian computation (ABC) methods [38]. Prediction error was assessed with 10-fold cross validation. We controlled for the sample fraction by including the prevalence of infection at the time of the most recent sample as a covariate.
The mean CD4 cell count and FAS for clustered taxa is shown in figure 2. Consistent with our hypotheses, patients with higher CD4 count are more likely to yield phylogenetically clustered virus, and the mean CD4 count among clustered patients has an inverse relationship with the threshold TMRCA for clustering. Also consistent with our hypothesis, patients which yield virus with lower FAS (less diverse virus) are more likely to be phylogenetically clustered, and mean FAS has a positive relationship with the threshold TMRCA for clustering. Patients were strongly co-clustered within quantiles. Maximum assortativity values, which measures the similarity of co-clustered taxa were 13% for CD4 and 4.5% for FAS. The maximum assortativity also occurs at low threshold TMRCA for FAS and CD4 (1700 and 1467 days). Very little clustering is observed between the first and last quantiles.
In general, the deterministic model offers an excellent approximation to the stochastic system. All trajectories pass through or close to the median of simulation predictions. Figure 3 illustrates the prevalence of early/acute and chronic infections from a typical simulation of the HIV model and the corresponding deterministic approximations. This correspondence occurs despite large fluctuations in prevalence when the number of infections is small. In [6] it was shown that the correspondence between the stochastic and deterministic systems can be very good even if the epidemic is started from a single infection and the coalescent is fit to the resulting transmission tree.
In figure 3, late infections outnumber early infections by approximately 20 to 1. As a consequence, NLFT for late infections are more stable due to larger sample sizes, and the NLFT are more noisy for the sample of early infections. The prevalence of infection plateaus prior to the 15 year sample time, so there is not much difference in the phylogenetic features observed at 15 and 30 year sampling times.
Many summary statistics calculated from an HIV gene genealogy can be informative about the fraction of transmissions attributable to early/acute infection, (equation 6). Figure 4 shows the value of four statistics as is varied. The dependancy of these summary statistics on the sample fraction is also shown in figure S4. (upper left) is the Pearson correlation coefficient between the number of early/acute taxa and chronic taxa in a cluster and is most sensitive to . Also shown are the mean cluster size, the number of extant lineages at the threshold TMRCA, and the fraction of taxa in a phylogenetic cluster. As the fraction of transmissions from the early/acute stage is varied, transmission rates and are adjusted so that remains constant. The smallest value of shown in figure 4 corresponds to the point where , such that there is no excess transmission in the early/acute stage. The most recent sample is assumed to be at 35 years following the initial infection. Epidemic prevalence after 35 years is approximately constant. The threshold TMRCA was five years before the most recent sample. Sample size and distribution of samples over time was matched to the Detroit MSM phylogeny. Furthermore, the number of early/acute versus chronic taxa sampled was made to match the Detroit data by use of the BED test [36] for determining recency of infection.
The fraction of taxa which are phylogenetically clustered also varies with (figure 4, upper left). The fraction of early/acute taxa clustered is more sensitive to than the fraction from chronic taxa. Early/acute taxa are always clustered at a greater rate than chronic taxa, even when corresponding to the minimum value of . This is because virus from early/acute patients was recently transmitted, making it much more likely that the lineage will coalesce in the recent past regardless of the source of the infection.
Using the mathematical model, we explored many parameters including the threshold TMRCA for clustering, the sample fraction, and the time relative to the beginning of the epidemic at which sampling occurs. Figures S1, S2, S3 demonstrate that the deterministic model is capable of reproducing many phylogenetic signatures that have been associated with HIV epidemics in MSM. For example, figure S5 shows the fraction of the sample (both early and late infections) which remain unclustered with any other sample unit. When the threshold TMRCA is zero (corresponding to the far right of the time axis), the entire sample remains unclustered. As the threshold TMRCA increases (moving leftwards on the time axis), more sample units become clustered and the fraction of taxa remaining unclustered decreases.
The time of sampling makes little absolute difference to the qualitative nature of the tree statistics if sampling occurs after the peak epidemic prevalence (around 15 years). However the sample fraction (the fraction of prevalent infections sampled) has a large effect on all tree statistics. When the sample fraction is large, the fraction remaining unclustered drops much more precipitously than when it is small as the threshold TMRCA increases. This occurs because each transmission can cause a sample unit to become clustered; a large sample size implies that transmissions will have a greater probability of resulting in an observable coalescent event (e.g. it results in a larger ratio ).
Early infections become clustered at a much greater rate than late infections. This corresponds to the excess clustering of early/acute infections observed in many phylogenies. By virtue of being infected in the recent past, an acute infection inevitably has a very recent common ancestor with another infection who transmitted to that individual. Mathematically, this is reflected in transmission terms of the form which appear in the ancestor function for early, but not late infections.
When the sample fraction is non-negligible, the fraction of the sample in a cluster levels off for intermediate thresholds. Similar phenomena were noted by Lewis et al. [8] and Hughes et al. [14] who observed that the fraction of the sample in a cluster did not change substantially beyond a small threshold, though these studies probably had high sample fractions. The plateau is due to the bimodality of coalescence times induced by early infection dynamics. Many coalesce events occurs at thresholds close to the sampling time, which corresponds to lineages of early infection coalescing.A larger group of coalescence times occurs close to the beginning of the epidemic when the effective population size is small. We hypothesize that the amount of excess clustering of early infections can be informative for estimating the sample fraction when it is not known.
Figure S2 shows the Pearson correlation coefficient for the number of co-clustered early and chronic infections as a function of the clustering threshold (). Given that a sample unit is in a cluster, under certain circumstances, it is much more likely to be clustered with another unit of the same type. This is reflected by large negative correlation coefficients for the number of co-clustered early and late infections for small threshold TMRCA. But negative correlation between the number of early and late infections is only observed for small sample fractions and small threshold TMRCA. The region of negative correlation appears very briefly for a 100% sample fraction; the region is much longer for small samples. This implies that if a patient with early infection is clustered, it is much more likely to be clustered with another early infection than expected by chance alone.
The skewness of the CSD shows a similar trend (figure S3). The skewness is always positive (to the right) and rapidly decreases as the threshold TMRCA is increased reflecting greater probability mass in the tail of the distribution. Skew is greatest for small threshold TMRCA, when most clusters are of size 1. The distribution remains positively skewed, though it quickly levels off for intermediate threshold TMRCA. The mathematical model shows that all moments of the CSD are finite and diverge to infinity in the limit of large sample size and threshold TMRCA.
A practical consequence of having an intermediate to large sample fraction is that chains of acute-stage transmission will account for many of the clusters observed at low thresholds. If a taxon is clustered with an early infection, then it is more likely that the unit will be clustered with additional early infections since such cases are highly infectious and have likely transmitted in the recent past. This provides a justification for the theory expounded in Lewis et al. [8] that high clustering of cases with recent MRCA's indicates episodic transmission; chains of transmission by early infections are interrupted by occasional long intervals until a transmission by late stage infections.
Corroborating figure 4 which shows that many statistics are correlated with , the PLS regression did not single out any particular group of statistics as being informative of early/acute stage transmission rates. The first component distinguishes between statistics that describe co-clustering (correlation of the number of acute and chronic taxa in a cluster) and statistics that describe excess clustering (e.g. the fraction of early/acute taxa that are not clustered with any other taxa). Four principal components were required to explain 42% of the variance of the transmission fraction with additional components only explaining an additional 2%. All statistics were well represented in the model with four components.
We have used coalescent models to characterize the phylogenetic patterns of a virus which produces an early stage of intensified transmission followed by a long period of low infectiousness. These patterns have been observed in multiple phylogenies of HIV-1 from MSM and IDU, and our model suggests that these should be general features for epidemics which feature early and intense transmission. These patterns are not necessarily a consequence of complex sexual network structure [14]. Complex transmission dynamics driven by sexual networks are undoubtedly taking place, but detecting the phylogenetic signature of sexual network structure will require carefully-chosen summary statistics [15]. We have characterized phylogenies using the cluster size distribution (CSD) which is similar to commonly used clustering methods based on strong support for monophyly but is nevertheless tractable for mathematical modeling in a dynamical systems framework. Moments of the CSD reflect a wide range of tree topologies, such as the distribution of branch lengths and tree balance, and are potentially informative of a wide range population genetic processes. For example, a highly unbalanced tree would have produce very skewed CSD, and a very star-like tree would have a CSD that is insensitive to changes in the clustering threshold.
While there has been much discussion of how clustering of acute infections is caused by the intensity of transmission during the acute stage, the amount of excess clustering that will be observed is also very sensitive to the sample fraction. And even if transmission rates in the early/acute stage are equal to those in the late/chronic stage, we would still observe excess clustering of early/acute provided the sample fraction was large enough. This is a simple consequence of early/acute infections being connected by short branch lengths to the individual who transmitted infection. An advantage of the coalescent framework used in this investigation is that it is accurate even with large sample fractions [35].
Some of the statistics which are most informative of the underlying epidemiological processes are those based on co-clustering of labeled taxa, such as the correlation between the number of early and late infections in a cluster. Such statistics tend to be the most responsive to variation of the intensity of transmission during early infection, and are therefore good candidates for future estimation of the fraction of transmissions that occur during the first few months of infection with HIV. Knowing the frequency of early transmission is essential to prevention efforts, since these transmissions are the most difficult to prevent. Individuals with early and acute infection are usually not aware of the infection, and are therefore not susceptible to many interventions. Modeling to evaluate strategies such seek, test, and treat (STT) [39], [40] and pre-exposure prophylaxis(PrEP) [41] will require good estimates for the frequency of early-stage transmission in diverse populations, and phylogenetic data promise to refine these estimates.
Future work could focus on finding ways to use statistics derived from the CSD for estimation of epidemiological parameters within an approximate Bayesian framework [38], [42], [43]. Alternatively, advances [35] in coalescent theory may make it possible to calculate the likelihood of a gene genealogy conditional on a complex demographic history, such as those generated by the HIV model discussed here. Current techniques are limited in the amount of phylogenetic data that can be used for inference of demographic and epidemiological parameters. Estimation of the intensity of early stage transmission will likely require co-clustering statistics similar to the moments derived from the CSD. In cases where the simple compartmental models fail to reproduce phylogenetic patterns, a more complex transmission system model and its corresponding coalescent should be investigated which might involve sexual networks or geographical [44] and risk heterogeneity. We further conclude that care must be taken in using phylogenetic clusters for epidemiological inference. Mechanisms that generates clustering are often complex and counter-intuitive. We recommend that investigators shift from individual-based inference using small clusters to model-based inference using population-based surveys of sequence diversity.
|
10.1371/journal.pcbi.1002429 | Beyond the Binding Site: The Role of the β2 – β3 Loop and Extra-Domain Structures in PDZ Domains | A general paradigm to understand protein function is to look at properties of isolated well conserved domains, such as SH3 or PDZ domains. While common features of domain families are well understood, the role of subtle differences among members of these families is less clear. Here, molecular dynamics simulations indicate that the binding mechanism in PSD95-PDZ3 is critically regulated via interactions outside the canonical binding site, involving both the poorly conserved loop and an extra-domain helix. Using the CRIPT peptide as a prototypical ligand, our simulations suggest that a network of salt-bridges between the ligand and this loop is necessary for binding. These contacts interconvert between each other on a time scale of a few tens of nanoseconds, making them elusive to X-ray crystallography. The loop is stabilized by an extra-domain helix. The latter influences the global dynamics of the domain, considerably increasing binding affinity. We found that two key contacts between the helix and the domain, one involving the loop, provide an atomistic interpretation of the increased affinity. Our analysis indicates that both extra-domain segments and loosely conserved regions play critical roles in PDZ binding affinity and specificity.
| Protein interactions play crucial roles in all biological processes. A common way of studying them is to focus on sub-parts of proteins, called domains, that mediate specific types of interactions. For instance, it is known that most PDZ domains mediate protein interactions by binding to the C-terminus of other proteins. Humans have more than 200 slightly different copies of these domains. At the level of the binding site, PDZ domains look quite similar. This is in apparent contradiction with their heterogeneous binding specificity. Using detailed molecular dynamics simulations in conjunction with statistical analysis, we predict that contacts outside of the canonical binding site play important roles in regulating protein interactions. Some of these contacts influence the overall dynamics of PDZ domains, providing an explanation for their allosteric effect. These interactions involve regions of the PDZ domains that are much less conserved, suggesting that they can help in differentiating selectivity in this large domain family.
| PDZ domains are modular protein interaction domains specialized in binding short linear motifs at the C-terminus of their cognate protein partners [1], [2]. In human, they are found in hundreds of different proteins and are mostly involved in cell-cell adhesion and epithelial junctions [3]. PDZ domains are often classified on the basis of their preferred C-terminal ligand sequences. Early studies organized binding specificity in three canonical classes: class-I involving C-terminal motifs of the type [x–(s/t)–x–(v/i)cooh], class-II [cooh] and class-III [x–(d/e)–x–cooh], where is a hydrophobic residue and x any amino acid [2], [4]. This classification, though consistent with the highly conserved binding groove [2], does not explain the large selectivity observed both in naturally occurring C-terminal peptides and synthetic peptide library screening [5]–[8]. Systematic investigations of PDZ domain specificity revealed that more distal C-terminal peptide residues are involved in the binding process [7], [9], suggesting a role for the loop following the binding site [10]–[15]. For example, the solution structure of the second domain of the hPTP1E protein showed that this loop interacts with the sixth amino acid from the peptide C-terminus [10], while possible electrostatic contacts between the loop and peptide amino acids up to position eight were found in the Par3 PDZ3-VE-Cad domain [14], [15].
It was recently suggested that specificity beyond the canonical classes can be obtained by long-range interactions involving non-conserved structural motifs specific to the domain [16]. For instance, the extra-domain helical extension characterizing the third PDZ domain of PSD95 (also called DLG4 or SAP90) was shown to influence binding [17]. Although this helix is away from the binding groove, affinity is reduced by 21-fold upon truncation of this non-conserved structural motif. Titration calorimetry measurements indicated that the free-energy penalty is entropic in nature. It was proposed that enhanced side-chain flexibility upon helix truncation, which is subsequently quenched by peptide binding, might be the main reason for this effect. This exquisitely dynamical behavior, calling for a hidden dynamic allostery [17], [18], pinpointed the importance of conformational entropy upon binding mediated by structural elements not directly evident from structural inspection alone [19], [20].
Here, we investigate the set of interactions beyond the binding site influencing peptide binding in the PSD95-PDZ3:CRIPT complex. Molecular dynamics (MD) simulations indicate that residues upstream of the 4th C-terminal amino acid are crucial for binding. Specifically, lysines residues at position −4 and −7 in the CRIPT peptide are observed to dynamically interact with the loop. Shorter peptides spontaneously unbind from the domain, indicating that canonical interactions within the binding site are not sufficient for binding. Further simulations of the DLG1-PDZ2:E6 complex suggest a wide spread presence of such peptide-loop interactions in the PDZ family. Finally, we find that the extra-domain helix of PSD95-PDZ3 helps stabilizing the loop via ionic interactions. Our results provide direct evidence of the role played by peptide amino acids away from the C-terminus and the interplay with previously unrecognized PDZ structural motifs.
Seminal X-ray crystallography experiments on the third PDZ domain of PSD95 in complex with the CRIPT C-terminal peptide indicated that peptide binding is realized through the last four residues (peptide positions 0 to −3), while the rest of the peptide is mostly disordered [1] (the system was crystallized with a 9-mer peptide, see below). This observation suggested a minor role of residues upstream of the last four ones for binding. To test this hypothesis, four MD simulation runs were carried out using a 5-mer peptide from CRIPT (-KQTSV-COOH, CRIPT5), a natural class-I binder of PSD95-PDZ3 (see Methods) [1], [17]. Unexpectedly, all the four runs showed spontaneous unbinding within the first 110 ns (see blue and light-blue lines of Fig. 1 for two unbinding trajectories and Table S1 for specific unbinding times and simulation lengths). Weak affinity was a somewhat surprising result, suggesting that canonical class-I interactions alone are not sufficient for binding. Interestingly, one of the runs showed rebinding from a partially unbound state. This event was mediated by the interaction of on the peptide with on the loop following the binding site as shown in Fig. S1. The same peptide with a charged N-terminus (CRIPT5*), which can reinforce this type of electrostatic interactions, remained anchored to the binding site for the total simulation time [21]. However, the peptide canonical contacts were only partially formed (see Fig. S2).
These observations suggested that interactions beyond the canonical class-I motif are needed to achieve stable binding in native conditions (i.e. without an artificially charged N-terminal peptide), possibly with a major role of the loop. To elucidate this point, four simulations with a longer 9-mer CRIPT peptide (-TKNYKQTSV-COOH, CRIPT9) were performed for a total of roughly 700 ns. The peptide remained bound to the original X-ray configuration in all runs (see red curve in Fig. 1 for a typical RMSD time trace). Strikingly, the four extra amino acids strongly influenced binding. The two lysines at peptide positions −4 and −7 transiently formed specific salt-bridges with two negatively charged loop residues, and . These contacts are dynamic, interconverting between each other on the ns time scale. On the other hand, their cumulative contribution is large: the loop and the ligand are in contact via salt-bridges for 44% of the time. These results indicate an unexpected and biologically relevant role of this loop, going beyond class-I interactions.
Structural cluster analysis provides a quantitative classification of the non-canonical interactions (see Methods for details). In Fig. 2, structural ensembles characterizing the three most populated peptide-loop configurations are shown. We used a simplified code to classify the peptide-loop interactions. At the first, second and third position there is a “1” if interactions −7∶331, −7∶332 or −4∶331 are formed, respectively; “0” otherwise (these three contacts are the statistically more relevant ones). For example, “110” indicates that peptide is in contact with both and , as shown in Fig. 2e–f. The most observed configurations are “110”, “001” and “100”, having a relative population of 13%, 10% and 8%, respectively (see Fig. 2 for their structural characterization; the cumulative 44% is obtained by summing up the remaining peptide-loop interacting conformations).
This scenario is represented in Fig. 3 by the transition network of the different peptide-loop configurations (see Methods). Multiple pathways are present, where a quite unspecific network of conformational changes stabilizes peptide-loop interactions on a time scale which is faster than unbinding (for example, was measured for another member of the PDZ family [22]). Interestingly, the presence of peptide-loop interactions strongly influence the propensity to form canonical class-I contacts (see Fig. S2).
The dynamic nature of the interactions explains why peptide-loop contacts were difficult to detect by previous structural experimental investigations [1], [13]. For instance, both the original PDZ3 X-ray structure reported by McKinnon and collaborators [1] as well as further attempts by other groups (e.g. PDB-ID:1TP3) indicated that only a four residue C-terminal stretch (positions 0 to −3) is directly involved in binding. However, this observation is not supported by in vitro evolution and mutagenesis studies [7], [9], [15]. Along the same line, titration calorimetry experiments provided evidence for the role of peptide positions beyond −3 for both affinity and specificity [23], while water-mediated interactions were found when bound to the oncogenic E6 peptide [13]. Our observations reconcile these two views, providing a unifying picture for peptide binding to PSD95-PDZ3. While a dominant configuration characterizing the interactions between the peptide and the loop is absent, the cumulative effect of these interactions is necessary for binding. This effect is mostly dynamical, indicating that structure alone does not suffice to understand function in this case.
PSD95-PDZ3 is characterized by an extra-domain helix at the C-terminus [1], [17]. Structural analysis of our MD data showed that the helix directly interacts with the loop as well as with a region distant from the binding site, via two salt-bridges (red dashed lines in Fig. 4a). The first one involves at the end of the helix and a negatively charged amino acid on the loop, . The second ionic interaction is between helix and , which is located in a region of the domain without specific secondary structure. This region (blue in Fig. 4a), in turn, is in spatial contact with the carboxylate binding loop. No specific helix-peptide interactions were found, only unstable hydrophobic contacts.
Recent experiments indicated that the extra-domain helix strongly influences the dynamics of the domain [17]. Binding affinity to the 9-mer CRIPT peptide was shown to decrease by 21-folds upon helix truncation through a purely entropic effect. The truncated form of PDS95-PDZ3 is defined by residues 306–395, and referred to as throughout the text [17]. To provide atomistic insights into this mechanism, MD simulations of bound to CRIPT5 and CRIPT9 were performed (see Table. S1). The short 5-mer peptide unbound very quickly () from the domain in all the four simulation runs, while CRIPT9 remained in the binding site. As observed for the WT, binding is stabilized by a network of dynamic salt-bridges between the ligand and the loop (see Fig. S2).
Analysis of the backbone root-mean-square-fluctuations (RMSF) in the WT showed that the flexibility of the bound form is not affected by helix truncation (Fig. 4b). However, it affects the unliganded (apo) form, enhancing the overall domain backbone flexibility (Fig. 4c). The enhanced flexibility is mainly localized in three regions: the carboxylate binding loop (residues 318–323), the loop (residues 330–336) and residues 341–356. The latter corresponds to the region where the helix is forming the salt-bridges with . In our simulations for the WT, this interaction is present 49% and 41% of the time in the apo and bound forms, respectively. Given the spatial vicinity between this region (i.e., 341–356) and the carboxylate binding loop, we assume that the peaks relative to these two regions are coupled, arising from the missing interaction with the helix. Similarly, the enhanced flexibility of the loop is induced by the missing interaction with the extra-domain helix through the salt-bridge between and . This interaction is very stable in both the apo and peptide-bound states, being formed 83% and 82% of the time, respectively.
These observations have important consequences for the interpretation of the entropic penalty upon binding to . Given that the flexibility of the bound form is unaffected by helix truncation, while it is much larger in the apo form, peptide binding to requires the quenching of the three regions reported in Fig. 4c and described above. Hence, our results suggest that the quenching of both the carboxylate and loops is responsible for the entropic penalty. Nevertheless, we cannot fully exclude other effects like a contribution from side chain dynamics, since decoupling entropy into local terms is a controversial and unsolved problem [24], [25].
The important role of backbone dynamics is in contrast with recent NMR relaxation experiments which found a negligible contribution of the backbone compared to side chains flexibility [17]. We suggest that this apparent contrast can be solved by looking at the time scales of the fluctuations reported in Fig. 4c. RMSF differences peaks vanish when the time windows used for the calculations are similar to the ones relevant for NMR measurements, i.e. of the order of 10 ns or less (grey and black lines in Fig. 4c). Our data indicates that the relevant backbone fluctuations are on the 100 ns time scale. Such dynamics is, on the one hand too slow to be detected by NMR spin-relaxation techniques (i.e. ) [19], [25], [26] and, on the other hand, too fast to show up as a separate subpopulation in NMR relaxation-dispersion experiments (i.e. ).
Stabilization of the extra-domain helix is further mediated by a hydrophobic patch, formed by and on the PDZ domain, and and on the helix, as shown in Fig. 5a. Analysis of all human PDZ domains (see Methods) revealed that, while position 337 largely consists (i.e. 86%) of aliphatic or aromatic residues, position 328 is less conserved, with a large portion of aliphatic amino acids (see Fig. S3). Free-energy calculations between this helix and the PDZ domain performed with FoldX [27] (see Methods) predict that V328A and V328I mutants in the apo-form have a of 1.35 and −0.79 kcal/mol, respectively. Hence, mutation to ALA destabilizes the domain. MD simulations of both mutants are consistent with this scenario. Given the direct interaction between the extra-domain helix and the loop (Fig. 5b), it is found that bulkier aliphatics make this loop more rigid, avoiding the peptide induced quenching upon binding described in the previous section. Reversely, loop flexibility of the V328A mutant increases, approaching the one obtained in absence of the extra-domain helix (, blue line). These results suggest a correlation between bulkier aliphatics at position 328 and the presence of an extra-domain helix.
To further investigate this hypothesis, we used PSIPRED [28] to compute the helical propensity of C-terminal segments in all 258 human PDZ domains (see Methods). A larger helical propensity is found for domains with ILE, LEU or VAL at position 328, compared to the ones with ALA (see Fig. 5c). For instance, around 10 residues downstream of the C-terminus, an helical propensity twice as large is found (P-value of 0.02, see Methods). These results correlate very well with our previous findings, indicating that large aliphatic side chains at position 328 can serve as anchors for extra-domain segments, stabilizing the loop. Consequently, domains with an alanine at position 328 are less likely to have an extra-domain helix and we expect that in those cases the loop would be structured differently with respect to PSD95-PDZ3. This is in agreement, for example, with both PDZ1 and PDZ2 of PSD95. These domains are known to lack the C-terminal extra helix, possess an alanine at position 328 and have a different composition of the loop (see next section).
The PSD95-PDZ3 loop (together with V328) and the extra-domain alpha-helix are remarkably well conserved in orthologs up to fly (and even partially conserved in worm), as well as in human paralogs such as SAP97 (DLG1), PSD93 (DLG2) or SAP102 (DLG3), see Fig. S4. In particular, the three charged residues involved in peptide binding and helix contact are conserved in almost all cases, providing indirect evidence that the same loop-mediated protein/ligand recognition is taking place in distant organisms. This is not the case when looking at the entire PDZ family, where the loop is highly heterogeneous both in length and amino acid composition. For instance, the loop of the PSD95-PDZ2 is more rigid, making self-interactions with the main domain body in a region close to the hydrophobic patch mentioned earlier [29]. Despite these differences, there are studies suggesting a role of the loop in binding to PDZ2. Large chemical-shifts were measured in the loop region upon binding, substantially contributing to affinity [29]. Finally, X-ray crystallography of PDZ2 from the human paralog DLG1 in complex with the oncogenic E6 peptide pointed out to an asparagine on the loop () interacting with the ligand backbone at position (using our notation) [13].
To provide a dynamical picture of the process, we performed additional simulations of the DLG1-PDZ2:E6 complex (see Methods). Our calculations reiterate the importance of for binding to PDZ2. It is found that the E6 peptide is in contact with the loop through mainly three interactions, :, : and :, for a total of 69% of time. An example structure is shown in Fig. 6. These contacts interconvert on a ns time scale. Together with the results obtained for PDZ3, these observations suggest that the loop is actively involved in binding specificity: a property that would need to be consistently explored throughout the entire PDZ family.
In PDZ binding, the relatively limited information about peptide amino acids more distant from the C-terminus prevented a clear structural understanding of the effect and importance of these upstream side chains. Our work aims to fill this gap by providing calculations with both a canonical 5-mer CRIPT peptide as well as a longer 9-mer peptide in complex with PSD95-PDZ3. Three main results emerge from our work.
First, we observe in our simulations that peptide binding is mediated by ionic interactions with the loop following the binding site, referred to here as the loop. These contacts are found with the 9-mer peptide, while the shorter 5-mer unbinds spontaneously after a few tens of ns. Recent experimental results on several PDZ domains support our interpretation [23], [30]. Strong differences between short and long peptides were found for negatively charged loops (e.g. MAGI1-PDZ2) [30]. Peptide-loop contacts are dynamic, where multiple specific interactions interconvert on a fast time scale of tens of ns (i.e. much faster than unbinding [22], [31]). Such dynamic interactions are likely to characterize several other PDZ domains. Further calculations on another member of the PDZ family, the DLG1-PDZ2, which is characterized by a different loop, support our hypothesis. Moreover, unresolved side chains away from the C-terminus are often found in other PDZ-ligand X-ray structures (see examples in Table S2), indicating that these side chains can adopt multiple conformations. We note that the presence of positively charged residues downstream of the fourth C-terminal positions of PDZ peptide ligands is well attested by recent experimental specificity profiles [7]. These charged residues are not necessarily always at the same positions, even within ligands of the same domain [30]. This is likely so because the peptide is flexible at these positions (as shown in Fig. 2). Consistently, loops display a clear over-representation of negatively charged residues compared to other regions in PDZ domains: 11.6% of D/E in entire PDZ domains, 15.2% for D/E in loops (according to the Fisher's test the probability to have this difference by chance is as low as , see Methods). Many of these residues on the loop provide clusters of negatively charged side chains that are ideally suited to recruit ligands with positive charges at any position between −4 and −7.
Second, we propose a mechanistic explanation for the microscopic origin of the binding entropic penalty in absence of the extra-domain helix of PSD95-PDZ3. In the apo form, the helix plays a crucial role in stabilizing both the carboxylate binding loop and the loop. Hence, these two loops are more flexible in the helix truncated domain. In this case, the peptide quenches the two regions upon binding, resulting in the observed entropic penalty. This quench does not take place when the extra-domain helix is present. Our findings suggest that extra-domain regions might play a more important role than mere linkers between functional domains [16], reiterating that the reductionist approach that protein domains can be studied in isolation should be always validated. This is especially important because several segments adjacent to domains show little sequence specificity (and thus are often not included in domain definition), although they adopt well-defined secondary structures such as the in the third PDZ domain of PSD95.
Third, analysis of 258 human PDZ domains as well as MD simulations of single-mutants allowed for the identification of an amino acid at the beginning of the loop, VAL in PSD95, that correlates with the presence of the extra-domain helix in other PDZ domains. Prediction of helical propensities at positions following the C-terminus of the domain showed enhanced probability for those domains presenting bulkier aliphatic side chains other than alanine at that position. This analysis suggests that a binding mechanism, indirectly involving the extra domain helix as in PSD95-PDZ3, might be relevant for a significant portion of the PDZ domain family.
Molecular dynamics simulations were performed using the GROMACS implementation [32] of the CHARMM27 force field [33], [34] at constant temperature and pressure with reference values equal to 300 K and 1 atm, respectively. The use of hydrogen virtual sites and fixed covalent bonds allowed a 4 fs integration time-step [35]. All systems were solvated in a dodecahedric box with an average of roughly 5000 tip3p water molecules (see Table S1 for details of each simulation setup). In the case of PDZ3, the system was equilibrated from the deposited X-ray structures 1BE9 and 1BFE [1] for the bound and apo forms, respectively, using residues 306–402 for the WT and 306–395 for . The PDZ2 starting structure is 2I0L [13] (from DLG1/SAP97). Each molecular setup was sampled by four independent runs of approximately 200 ns each for a total of (Table S1). The first 50 ns of each trajectory were neglected in the analysis to reduce the bias from the starting configuration. Snapshots were saved every ps. The peptide N-terminus was neutralized in all cases, except CRIPT5*. The sequences of the 9-mer peptides are -TKNYKQTSV-COOH and -LQRRRETQV-COOH for PDZ3 and PDZ2, respectively. The first 5 peptide residues (i.e., positions from −4 to −8) as well as mutations at position 328 and the truncation of the extra domain helix were modeled using PyMol [36]. For each run, backbone RMSF values were calculated per residue as an average over the atoms C, and N. Final RMSF values were averaged over the four runs. Molecular trajectories were analyzed with the programs WORDOM [37], [38] and GROMACS [39]. Hydrogen bonds were determined based on cutoffs for the angle Acceptor - Donor - Hydrogen () and the distance Donor - Acceptor (3.6 Å). Ionic interactions are considered to occur when the two last carbons before the charged atoms are closer than 5 Å.
Each protein-ligand snapshot was labeled by a four-digits code. The first three digits describe the peptide-loop interactions, e.g. “110”. The last digit represents an id, encoding the peptide structural conformation (i.e., the internal degrees of freedom). The latter was obtained by running a leader-based cluster-analysis on the ligand backbone (atoms C, and N) with a 2 Å cutoff, using the program WORDOM [37], [38]. This digit distinguishes between different peptide conformations characterized by the same contacts with the loop. Each four-digit string represents a microstate of the protein-ligand complex. This decomposition is used to build a conformation-space-network [40]–[42], where each microstate is a node and a link between two nodes is placed if there is a direct transition between them during the MD simulation. Basins of attraction are defined using a gradient-cluster analysis [43], [44], where multiple microstates are lumped together if they interconvert rapidly. Each gradient-cluster represents a metastable configuration, which can contain heterogeneous peptide-loop contacts. Connectivity between these metastable configurations is represented as a coarse-grained network as shown in Fig. 3 (see also Fig. S2 in the Supp. Mat.). The gradient-cluster algorithm is freely available in the program PYNORAMIX (GPL license, available at the website raolab.com).
Predictions of free-energy differences upon mutations were done with FoldX using the BuildModel option after properly repairing the structures with the RepairPDB command [27]. The initial structure (PDB 1BFE) was first minimized with GROMACS in explicit water. This structure was originally crystallized with an ILE at position 328. We mutated it both to VAL (WT) and ALA to compute the free-energy differences.
The set of all human PDZ domains was retrieved from PFAM [45] and SMART [46] databases. A first multiple sequence alignment was generated with MUSCLE [47]. The alignment was manually curated, removing PDZ domains that could not be unambiguously aligned (most of them are unconventional PDZ domains). This resulted in a total number of 258 PDZ domains (see Table S3). The loop was mapped by homology starting form the structure of PSD95-PDZ3. Several PDZ domains are close paralogs, and this can result in strong biases when computing frequencies or correlation patterns. To account for this effect, we always grouped paralogs together (see Table S4). Groups of paralogs were defined using a cut-off of 50% on the sequence identity. The contribution of each member of a group was weighted by the inverse of the group size. For instance, to compute the amino acid frequency at a given position, residues from a group of 5 paralogs only contributed 1/5 each to the total frequencies. The helical propensity of C-terminal extensions of PDZ domains was computed with PSIPRED [28] for up to 20 residues downstream of the domains. If the protein C-terminus was reached before the 20 residues, a helix propensity of 0 was used. Here again, the contribution of paralogs was weighted to prevent purely phylogenetic correlations. P-values were computed by reshuffling the amino acid composition at position 328 in all PDZ domains of Table S3. The Fisher's test was used to compute the probability to have a given number of negative residues within all loop residues, knowing the total number of negative residues within the sequences of all PDZ domains [48].
|
10.1371/journal.ppat.1002203 | Complex Recombination Patterns Arising during Geminivirus Coinfections Preserve and Demarcate Biologically Important Intra-Genome Interaction Networks | Genetic recombination is an important process during the evolution of many virus species and occurs particularly frequently amongst begomoviruses in the single stranded DNA virus family, Geminiviridae. As in many other recombining viruses it is apparent that non-random recombination breakpoint distributions observable within begomovirus genomes sampled from nature are the product of variations both in basal recombination rates across genomes and in the over-all viability of different recombinant genomes. Whereas factors influencing basal recombination rates might include local degrees of sequence similarity between recombining genomes, nucleic acid secondary structures and genomic sensitivity to nuclease attack or breakage, the viability of recombinant genomes could be influenced by the degree to which their co-evolved protein-protein and protein-nucleotide and nucleotide-nucleotide interactions are disreputable by recombination. Here we investigate patterns of recombination that occur over 120 day long experimental infections of tomato plants with the begomoviruses Tomato yellow leaf curl virus and Tomato leaf curl Comoros virus. We show that patterns of sequence exchange between these viruses can be extraordinarily complex and present clear evidence that factors such as local degrees of sequence similarity but not genomic secondary structure strongly influence where recombination breakpoints occur. It is also apparent from our experiment that over-all patterns of recombination are strongly influenced by selection against individual recombinants displaying disrupted intra-genomic interactions such as those required for proper protein and nucleic acid folding. Crucially, we find that selection favoring the preservation of co-evolved longer-range protein-protein and protein DNA interactions is so strong that its imprint can even be used to identify the exact sequence tracts involved in these interactions.
| Genetic recombination between viruses is a form of parasexual reproduction during which two parental viruses each contribute genetic information to an offspring, or recombinant, virus. Unlike with sexual reproduction, however, recombination in viruses can even involve the transfer of sequences between the members of distantly related species. When parental genomes are very distantly related, it is anticipated that recombination between them runs the risk of producing defective offspring. The reason for this is that the interactions between different parts of genomes and the proteins they encode (such as between different viral proteins or between viral proteins and the virus genomic DNA or RNA) often depend on particular co-evolved binding sites that recognize one another. When in a recombinant genome the partners in a binding site pair are each inherited from different parents there is a possibility that they will not interact with one another properly. Here we examine recombinant genomes arising during experimental mixed infections of two distantly related viruses to detect evidence that intra-genome interaction networks are broadly preserved in these genomes. We show this preservation is so strict that patterns of recombination in these viruses can even be used to identify the interacting regions within their genomes.
| Although variations in the basal mechanistic predispositions of different regions of nucleic acid molecules to recombine is certainly a primary determinant of recombination patterns detectable within some viral genomes [1]–[5], it is becoming increasingly apparent that an important secondary determinant is natural selection [6]–[8]. Functional analyses of recombinant genes [9], [10] and genomes [11]–[13] has indicated that a large proportion (and possibly the vast majority) of recombination events between distantly related genomes probably yield defective progeny.
Analyses of bacterial [14]–[16] and viral recombination [6], [12], [13], [17] and protein engineering studies utilising DNA-shuffling methodologies [9], [10], [15], [18] have indicated that the probability of a given recombination event being deleterious depends on the modularity of the specific gene(s) or sub-gene module(s) that are transferred and tends to increase with decreasing parental sequence relatedness. This effect is caused, at least in part, by the tendency of recombination to disrupt the networks of genome encoded sequence specific interactions that underpin the biology of all organisms. Examples of such encoded interactions might include inter-amino-acid contacts that determine and maintain proper protein folding, Watson-Crick base pairing within functionally meaningful secondary structures that form in single stranded DNA (ssDNA) and RNA molecules, and DNA and amino acid sequence motifs that mediate protein-protein, protein-DNA and protein-RNA interactions. It might therefore be expected that the survival of recombinant genomes under natural conditions will be largely dependent on how severely recombination has impacted such interactions.
To a degree, accurate a priori inference of the approximate fitness costs of recombination events has already been achieved within the context of individual proteins for which inter-amino acid interaction networks can be inferred from atomic resolution 3D structural information [9], [19]. Such approximations have been successfully used to increase the efficiency of DNA shuffling based protein engineering approaches [20]–[22]. However, replicating these successes at a whole-organism scale to, for example, assess the risks of novel pathogens emerging via recombination between different viruses co-infecting a particular host species, would require high resolution information on multiple genome-wide interaction networks – information that is not available for even the most well studied virus species. Also, rather than trying to use such interaction networks to explain recombination patterns, it might actually be more productive to use recombination patterns to infer the interaction networks. Given that recombination patterns are probably strongly influenced by these networks, it is reasonable to suspect that recombination patterns might encode information on their architectures.
We investigated this possibility in experimentally constituted mixed begomovirus infections. Species in the genus Begomovirus (Family Geminiviridae) are ideal test subjects in this regard not only because the emergence of numerous begomoviral crop pathogens has been attributed to natural inter-species recombination [23]–[27], but also because begomovirus genomes contain various well characterised intra-protein [28], [29], inter-protein [30] and protein-DNA interactions [31]–[34]. Using a genome-wide association approach and a variety of permutation tests to analyse data from controlled evolution experiments we demonstrate how recombination patterns can be used to: (1) identify mechanistic causes of variations in basal recombination rates across begomovirus genomes, (2) identify host adaptive nucleotide polymorphisms, (3) identify the various ways in which recombination can disrupted intra-genome interactions, and (4) indirectly visualise long-range intra-genome interaction networks.
Agroinfectious clones of Tomato yellow leaf curl virus - Mild [Reunion:2002] (AJ865337, referred to here as TYX) and Tomato leaf curl Comoros virus – [Mayotte:Dembeni:2003] (AJ86539, referred to here as TOX) are described in Urbino et al. [35]. Relative to one-another the two viruses display nucleotide polymorphisms at 491 genome sites (i.e. at ∼17.8% of their total genomic nucleotides) and differ at an additional 13 sites displaying between 1 and 14 nucleotide insertions/deletions.
A total of 89 seven day old tomato seedlings (Farmer variety, Known-you Seed) were co-agroinoculated with both TYX and TOX and were checked fourteen days later for evidence of co-infection using specific PCR primers (TYX forward: CCCAATTTTCAAGGATATG; TYX reverse: GCGCTTCCAAATAAAATTGC; TOX forward: AGGCTTTCAGGGGTGCA; TOX reverse: GTCGTTTCAGCATCAAAGC). Out of 33 successfully co-infected plants, ten randomly selected plants were grown for a total of four months before leaf samples were collected and stored as previously described by Bos [36].
Total DNA was extracted from leaf samples using the DNeasy Plant miniprep kit (Qiagen) according to the manufacturer's instructions. Circular viral DNA molecules were amplified using the TempliPhi Kit (GE Healthcare) as described by Inoue-Nagata et al., [37]. Full genomes were cloned into either pGEM-3Zf or pGEM-7Zf (Promega, USA) vectors at their BamHI restriction sites. A total of 362 complete DNA-A-like components from the ten plants (Table S1) were cloned and sequenced using the Macrogen Inc. sequencing service (Korea).
Identification of potential recombination breakpoint regions and the origins of recombinant fragments was carried out for each of the cloned sequences using the computer program RDP3 ([38]; available from http://darwin.uvigo.es/rdp/rdp.html) which was set up to perform direct nucleotide-by-nucleotide comparisons of potential recombinants and a pair of known parental sequences (in this case the input TYX and TOX sequences). A breakpoint map was compiled denoting the positions of all clearly identifiable unique recombination breakpoints (breakpoints falling at identical sites were considered different only if they were observed in viruses cloned from different plants).
Breakpoint density plots were then constructed from this map as described in [39]. Briefly, 200, 100, 75 or 50-nt windows were moved one nucleotide at a time along the length of the map and at each window position counts of all clearly identifiable unique recombination breakpoints falling within the window were plotted at the central window position. Significant clustering of breakpoint positions within each window was tested using the recombination hot-spot permutation test described in Heath et al. [39].
To test whether local degrees of shared sequence similarity between TYX and TOX influenced where recombination occurred between their genomes, we sorted the observed recombination breakpoint positions based on the numbers of contiguous identical nucleotides shared between the two genomes at the sites where the breakpoints were detected. We used a simple permutation test to determine whether the observed frequency of recombination breakpoints at each of these site categories was significantly different from those expected at these sites assuming that recombination breakpoints were randomly distributed and uninfluenced by local degrees of sequence similarity (see Protocol S1 for details).
Given that the secondary structures of the TYX and TOX genomes are unknown, we inferred an ensemble of near minimum free energy (MFE) folds for each genome using the program UNAFold version 3.8 [40], with sequences treated as circular ssDNA, the annealing temperature set to 25°C and sodium and magnesium concentrations respectively set to 1M and 0M. UNAFold was additionally set-up to only allow Watson-Crick (GC, AU) and Wobble (GU) base pairings. From these we produced, for each parental genome, a unified fold by scoring each predicted base-pairing interaction as the proportion of times that interaction was found within the 22 and 19 near minimum free energy (MFE) predicted folds of the aligned TYX and TOX genomes respectively.
Previous investigations into the impact of ssDNA secondary structure on begomovirus recombination frequencies have indicated that recombination breakpoints tend to occur at genome sites where one parental genome has a stable secondary structure and the other does not [26]. We therefore tested for evidence of increased recombination breakpoint frequencies within four separate sets of sites: (1) those at which nucleotides were predicted to be involved in base-pairing within either the TYX or TOX genomes, (2) those at which nucleotides were predicted to be base paired in the TYX but not the TOX genome and (3) those at which nucleotides were predicted to be base paired in the TOX but not the TYX genome. In order to avoid the biasing influences of the presumed (and strongly predicted) stem-loop sequence at the virion strand origin of replication (a known major recombination hotspot) we excluded from these sets all nucleotides falling within the 41 100% identical nucleotides shared by TYX and TOX around this region.
Evidence of breakpoint clustering within the various structured nucleotide sets (i.e. nucleotides inferred to be paired in both/either/one-or-the-other parents) was tested using a slightly modified version of the permutation tests described in Simon-Loriere et al. [5] and Lefeuvre et al. [41] (see Protocol S1 for details). Briefly, for the four different sets of sites tested, this involved dividing sites within the alignment into those falling within the test-set and those falling outside the test set, and then testing for significantly increased recombination breakpoint frequencies within the test set.
The fundamental aim of our study was to determine the influence of encoded intra-genome interaction networks on inter-species recombination patterns arising during co-infections. To achieve this we sought to test for evidence of selection acting on recombinants in such a way as to maintain the stability of various hypothetical interaction networks including those occurring between: (1) amino acids within expressed proteins, (2) viral nucleotides within predicted genomic DNA secondary structures, and (3) viral genomic regions (e.g. long-range interactions occurring between either different viral proteins or between virus proteins and viral DNA). Specifically, we attempted to statistically test whether, and to what extent, maintenance of these various interactions influences the recombination patterns that arise during co-infections. An important assumption that we made throughout is that recombinant viruses sampled from these co-infections (or at least a subset of these viruses) have been acted upon by selection and are reasonably fit members of the populations from which they were sampled. However, considering both the high evolution rates of begomoviruses [42]–[44] and the probable quasi-species and/or genetic complementation dynamics at play during infections involving genetically diverse virus populations [45]–[47], we expected this assumption to be somewhat violated for many of the low frequency genetic variants that were discovered during extensive sampling of genomes within the various co-infections analysed. In order to both limit and quantify potential biases caused by the sampling of low fitness recombinant variants, we analysed on the one hand the entire set of sampled recombinants – called our FULL dataset – and on the other only those presumably viable and reasonably fit recombinants that were sampled multiple times from individual populations – called our FIT dataset. We must stress here that even when recombinants were sampled multiple times from infected plants the possibility remained that these might still have been defective and that their higher prevalence could have been the result of either random drift or complementation by fit genomes. However, rather than biasing analyses of the FIT dataset in favour of detecting selection against defective genomes, inadvertent inclusion of subtly defective recombinants in the FIT dataset would have degraded our ability to use this dataset to detect selection against defective recombinants.
Probably as a consequence of the fact that recombination breakpoints that occur within protein coding sequences can potentially disrupt protein folding, it has been noted that recombination breakpoints detectable in viruses sampled from nature tend to cluster either within non-coding regions or at the boundaries of genes [7], [41], [48], [49]. We firstly tested whether recombination breakpoints within the recombinants arising during co-infections displayed a tendency to fall outside coding regions or on the periphery of genes using the same tests described in Lefeuvre et al. [41] (implemented in RDP3). We then applied the SCHEMA based tests described in Lefeuvre et al. [8] (also implemented in RDP3; [9]) to determine whether recombination breakpoints falling within the genome regions encoding the structurally modelled portions of Rep (the 118 N-terminal amino acids; [28]) and CP (196 amino acids; [29]) were less disruptive of protein folding than was expected by chance. For full details of these tests see Protocol S1.
In the same way that recombination has the potential to disrupt protein folding, it could potentially disrupt ssDNA folding. We used two separate permutation tests of ssDNA folding disruption to determine whether there was any evidence of recombinant sequences displaying significantly lower degrees of estimated ssDNA folding perturbation than that observed in randomly generated recombinant sequences. The permuted datasets used in both tests were identical. For each of 50 unique recombinant sequences we produced 100 recombinants with randomly placed breakpoints containing (1) the same number of recombination breakpoints as the real recombinant, (2) the same number of TOX and TYX derived polymorphisms as the real recombinant, and (3) the same numbers of polymorphic sites between recombination breakpoints as the real recombinant. These simulated datasets were compared to the real one using (1) a test of global fold stability comparing MFE values and (2) a SCHEMA-like test very similar to that devised in Voigt et al. [9] for comparing conservation of nucleotide pairing in recombinant genomes (see Protocol S1 for details). In both ssDNA folding disruption tests we predicted for every recombinant (both real and simulated) an ensemble of ssDNA folds displaying free energies close to that of the MFE fold. As before the ssDNA/RNA folding program UNAFold [40] was used throughout with the same settings as those mentioned above.
We sought to determine whether there was a signal of recombination having preserved known long-range intra genome interactions such as those occurring between: (1) the DNA binding domain of Rep and iterated Rep binding sites within the C1 proximal side of the intergenic region (a DNA-protein interaction; [31], [33], [34]), (2) the V1 promoter and the transcription activator protein (TrAP; encoded by the C2 open reading frame (ORF); another DNA-protein interaction; [32]), (3) Rep and the replication enhancer protein (REn) (a protein-protein interaction; [30]) and (4) Rep and the coat protein [50].
However, rather than directly testing whether these known long-range begomovirus intra-genome interactions were preserved in the recombinants arising during our experiments, we considered the problem from an alternative perspective. We hypothesised that if selection within our experiment favoured the survival of recombinants in which co-evolved long-range intra-genome interactions remained intact, then interacting sites within the genome should preferentially be inherited in pairs from one parent or the other. We were specifically interested in whether blindly examining all possible pairwise associations between sites would identify these known interactions with a high degree of statistical confidence.
To achieve this we designed a simple Chi-square based permutation test capable of detecting potential associations between pairs of polymorphic nucleotide sites within TYX-TOX recombinants (see Figure S1). Briefly, this involved counting, for all possible pairs of polymorphic nucleotide sites (i, j; TYX and TOX differ at 491 sites meaning that “i” and “j” are integers between 1 and 491 and that there are (491×490)/2 = 120295 possible polymorphic nucleotide site pairs), the numbers of times the four different possible combinations of TYX and TOX derived sites (TYXi-TYXj, TYXi-TOXj, TOXi-TYXj and TOXi-TOXj respectively encoded as site-pair categories 1, 2, 3 and 4) were observed across all of the analyzed recombinants. For each of the 120295 site pairs these four numbers were used to calculate a 2x2 Chi-square value that was used as a primary indicator of possible associations between the site pairs – whereas pairs of sites yielding low Chi-square values are more likely to have been independently inherited, the inheritance of site-pairs yielding high Chi-square values is more likely to have been non-independent. Non-independence could be due to close physical linkage within the genome or, if it occurs between distantly separated sites, it might be due to the need to maintain co-evolved encoded interactions between the sites.
In each of 10 000 permuted datasets, the experimentally observed recombination events were reconstructed from the parental sequences so as to maintain the numbers of polymorphic nucleotides separating observed breakpoint positions (i.e. maintaining the degree of polymorphic nucleotide shuffling) observed in the real recombinants but with the actual genomic positions of recombination breakpoints randomly shuffled. Sets of 120295 2x2 association tables were then determined for each of these 10 000 datasets. Instances when a site pair in the real dataset had an associated Chi-square value higher than 95% of the corresponding values in the 10 000 permuted datasets were interpreted as being indicative of there being a potential interaction between the sites. These potentially interacting site-pairs were then shortlisted and subjected to second set of tests aimed at classifying the exact natures of their associations.
These secondary association tests were again permutation based rank tests but, rather than being based on the over-all patterns of values in the 2x2 tables (as was the case with the first Chi-square based test), they were based on the magnitudes of individual values within the 2x2 tables formulated for the real dataset relative to values in corresponding tables from the permutation datasets. For example, with the table representing the polymorphic sites 1 (site 22 in TYX) and 9 (site 52 in TYX) there are 6 instances where recombinants had both sites from TOX, 43 instances where recombinants had both from TYX, 1 where site 1 was from TOX and site 9 was from TYX, 0 instances where site 1 was from TYX and site 9 was from TOX. The Chi value associated with this table is 41.88 – a number which is larger than 7 793/10 000 of the Chi values determined for the same site pair in the permuted datasets (i.e. the sites do not display an association that cannot be accounted for by their close physical linkage). Out of 10 000 permuted datasets, 4 104 had more than 6 instances where both sites at positions 1 and 9 were from TOX and 1 589 had more than 43 instances where both sites were from TYX. Thus assuming (1) random recombination, and (2) no selection on either individual sites or pairs of sites, the p-value for the test of whether both sites are not preferentially derived from TOX = 0.4104 (4104/10 000) and the p-value for the test of whether both sites are not preferentially derived from TYX = 0.1589.
Here, however, we were specifically most interested in identifying lower than expected frequencies of (1) TOX derived polymorphisms at the first site and TYX derived polymorphisms at the second (i.e. site pair category 2), (2) TYX derived polymorphisms at the first site and TOX derived polymorphisms at the other (i.e. site-pair category 3), and (3) combinations of (1) and (2) (i.e. site pair categories 2+3).
Finally, we identified potentially interacting site pairs as those at which (1) the Chi-square based test indicated a potential interaction with an associated permutation p-value <0.05 and (2) where mixtures of TYX and TOX derived polymorphisms at either or both of the sites occurred less frequently in the real dataset than in >95% of the permuted ones.
Of the 362 sequences obtained from ten independent four month old TYX-TOX tomato co-infections, 106 (∼ 29%) displayed evidence of having undergone one or more recombination events. Among these 106 were 50 unique recombinant genotypes (Figure 1), 18 of which were sampled multiple times and were therefore probably reasonably fit. The 32 recombinant genotypes for which only single genomes were sampled, were potentially minor population members in their respective plants and were therefore of uncertain viability (if these genomes had arisen immediately prior to sampling they would have avoided the influences of natural selection).
Within the 50 unique recombinant genotypes we collectively identified 452 clearly identifiable unique recombination breakpoints implying a mean of 9.04 detectable recombination events per recombinant sequence. Although complex recombinant genomes with up to 18 recombination breakpoints have been observed previously in maize streak virus recombination experiments [51], many of the recombinants that we observed in our experiment are extraordinarily complex (for example see recombinants P2R4 and P9R5 in Figure 1), containing as many as 58 detectable recombination breakpoints.
Interestingly, across the ten co-infection plants studied, the genetic distances between the recombinant genomes and their parents were not randomly distributed (Chi-square test p-value = 4.66.10−16). Whereas in six plants, the recombinant sequences were most similar to the TYX parent, in three plants, they were most similar to the TOX parent (Table S1). Only one plant presented viruses with balanced similarity to both parents. These differences between plants possibly indicate an experimental bias where stochastic differences in the initial numbers of TYX and TOX genomes co-infecting the plants eventually yielded recombinants carrying mostly TYX or TOX derived polymorphisms. Whereas in previously described geminivirus recombination experiments [26], [51], [52] independent replicated experiments appeared to deterministically yield very similar recombinant genomes, the widely varying outcomes observed in our experiment most likely reflect the complexity of the fitness landscape encountered within tomato by TYX, TOX and their recombinant progeny. In further analyses, we decided to pool viruses from different plants to improve the statistical power of the tests we performed. It should be borne in mind, however, that in so doing we ignored the potentially confounding factors underlying the over-all differences between recombinant virus populations arising in different plants.
Within the recombinant genomes, of the 491 polymorphic nucleotide sites differentiating TYX and TOX, 121 (24.6%) almost exclusively situated within the intergenic region (IR) and the region of C1 encoding the Rep catalytic domain (Figure 2), were preferentially derived (p<0.001) from TYX. TOX derived nucleotides were significantly more common (p<0.001) at 192 sites (39.1%), mostly spread throughout the V1, C1, C2 and C3 ORFs (Figure 2). This indicates that both parental genomes possibly contained particular polymorphisms that performed better within the recombinants in tomato than those derived from the other parent.
The IR is the origin of virion and complementary strand replication in begomoviruses and is also involved in the regulation of gene expression. That this region is predominantly derived from TYX suggests that the TYX IR sequences are better adapted than their corresponding TOX analogues when it comes to productively interacting with the DNA synthesis and transcription machinery of tomato cells. It would be interesting to experimentally test this hypothesis to determine whether apparent adaptations within the TYX IR also provide a survival advantage either in other tomato genotypes or other begomovirus host species.
To visualize the overall distribution of the 452 unique recombination breakpoints detectable within the recombinant genomes that arose during the co-infections, we plotted them on a density map and tested this for evidence of recombination hot- and cold-spots. This analysis revealed that the distribution of breakpoints was non-random, with there being clear evidence of statistically significant recombination hot- and cold-spots (Figure 3 and Figure S2). Irrespective of the window size used in these permutation based recombination hot- and cold-spot tests, a single highly significant recombination hot-spot was apparent in the IR around the origin of virion strand replication. Depending on the window sizes used, either one or two highly significant breakpoint hotspots were also detectable within the C1 ORF on either side of the region overlapping the C4 ORF. Besides these recombination hot-spots we also detected cold-spots within the IR (upstream of the C1 start codon), at the overlapping part of the V1 and V2 ORFs and at the interface between the V1 and C3 ORFs (see Figure S2).
Consistent with these results, previous experimental studies and analyses of recombinant genomes sampled from nature have revealed the presence of recombination hot-spots within geminivirus IR and complementary sense gene sequences and cold-spots within virion sense gene sequences [53], [54]. It has been proposed that complementary sense gene transcription, which occurs in the opposite direction to rolling circle replication, may increase rates of replication complex displacement during replication of the complementary sense gene encoding genome regions [55]. If multiple similar but non-identical geminivirus genomes are present within the same nucleus then re-initiation of replication from partially replicated virion strands on a new template molecule (occurring via the recombination dependent pathway of geminivirus replication [53]) could result in an increased prevalence of detectable recombination events across the C-sense ORFs and the IR.
Surprisingly, the recombination cold-spots within the IR and at the V1/C3 interface are at sites that are clearly detectable as recombination hot-spots within studies of natural monopartite begomoviruses ([8]; compare the blue shaded regions of the breakpoint distribution plot with the pink bars above the plot in Figure 3). These differences may reflect the fact that our experiments involve only a single pair of parents within a single host whereas the distribution of breakpoints seen in field isolated viruses is attributable to hundreds of parent pairs recombining in hundreds of different host species. If selection influenced both the experimental and natural recombination patterns to a similar degree, the differences between the patterns would indicate that for our particular experimental host-virus combination, recombination breakpoints falling at the V1/C3 interface and at the IR cold-spot are either (1) unusually deleterious relative to analogous events between other begomovirus pairs in different host species or (2) simply occur at a mechanistically lower frequency than is generally the case with other virus-pair and host combinations.
It is obvious that the viruses in our experiment experienced selective processes that were probably substantially different to those of natural viruses which, besides infecting a variety of different host types, must generally also persistently maintain a high degree of transmission fitness. Nevertheless, the fact that the recombination patterns detected in our experiment largely match those observed for natural recombinants implies that, despite its simplicity, the experiment clearly recaptures many of the mechanistic and selective processes shaping patterns of recombination seen in natural viruses.
Before considering the influence of natural selection on the recombination patterns that we observed, we sought to identify mechanistic factors that might have contributed to variations in recombination frequencies observed across the TYX and TOX genomes. Two factors that have been frequently associated with variations in basal recombination rates across viral genomes are local degrees of sequence similarity [1], [2], [4], [56]–[58] and the presence of nucleic acid secondary structures [5], [26]. Whereas degrees of similarity shared by potentially recombining sequences will obviously impact the efficiency of homologous recombination [1], [2], [4], [56]–[58], secondary structures may stall replication and increase the probability of recombination breakpoints falling within structured genome regions [5], [26].
We therefore tested the influence of the length of identical nucleotide tracts between the parental viruses on the numbers of breakpoints falling within those tracts and detected four different tract-size classes: under conditions of completely random recombination, (1) zero length tract sizes (i.e. when breakpoints occur between two sites that are polymorphic) are statistically far less prone to recombination than can be accounted for by chance (p<0.00001, Figure 4); (2) intermediate tracts of perfect identity between 2 and 12 nucleotides long tended to be more recombinogenic than expected (there are seven p-values <0.05 for tract lengths in this size class when zero length tract sizes are considered and three when they are excluded); (3) large tracts of perfect identity greater than 16 nucleotides long tend to be less recombinogenic than expected (five p-values for tract lengths in this size class < 0.05 when zero length tract sizes are considered and six when they are excluded) and; (4) the 41 nucleotide tract of 100% TYX and TOX identity surrounding the stem-loop region encounters more recombination breakpoints than can be accounted for by chance (p = 0.0092 when zero length tracts are included in the analysis and p = 0. 0388 when they are excluded; Figure 4). The recombination hot-spot at the virion strand replication origin and the cold-spots at zero length tracts were entirely expected given that recombination in geminiviruses is believed to occur by a strongly homology dependent double stranded DNA break repair mechanism during which monomeric genomes are replicationally released from multimeric genome concatemers [55]. Although a similar tendency for decreasing recombination frequencies in fragments over a certain length having been reported in retroviruses (which recombine by a replicational copy-choice mechanism; [56]), we were unable to find any plausible explanation as to why recombination in begomoviruses should occur more frequently at 5–12 nucleotides long identical tracts than it is does at identical tracts >16 nucleotides long.
Mounting evidence indicates that recombination breakpoints observable in many single stranded viral genomes tend to co-localise with base-paired nucleotides within genomic secondary structures [2], [5], [26]. We therefore tested whether in our experiment recombination breakpoints fell more frequently than could be accounted for by chance at sites predicted to be base-paired within the thermodynamically most favourable secondary structures of the ssDNA TYX and TOX genomes. Applying the same test used in Simon-Loriere et al. [7] to detect a clear association between breakpoint sites and secondary structures within HIV genomes, we analysed breakpoint distributions in relation to the distributions of TYX and TOX genomic sites inferred using the program UNAFold to be involved in base-pairing within genomic secondary structures [40]. Regardless of the subsets of sites that we examined (see M&M and Table S2) we detected no clear tendency for recombination breakpoints to fall at sites associated with predicted ssDNA genomic secondary structures. Although we observed a marginally significant tendency for breakpoints to cluster at sites where TOX had a strongly predicted secondary structure but TYX did not (p = 0.0579; Table S2), it is important to stress that we performed three separate tests and that we would have expected an approximately one in six chance of finding at least one p-value this close to significance even in the absence of any real associations.
It is perhaps not surprising that we found no strong correlation between the positions of recombination breakpoints and secondary structures in that much (if not most) begomovirus recombination is probably a by-product of double stranded break repair pathways [55]. Clear associations seen between secondary structures and recombination breakpoints positions observed in some RNA viruses [2], [5] have been attributed to copy-choice recombination mechanisms where secondary structures cause either reverse transcriptase or RNA dependent polymerase complexes to dissociate from template molecules. If these stalled complexes reengage a template other than those on which they had originally been situated then the synthesised RNA or DNA strands will be recombinant. Although the possibility remains that in begomoviruses copy-choice type recombination occurs due to secondary structure induced stalling of DNA replication complexes, the breakpoint distributions attributable to this mechanism would be largely obscured by those attributable to double stranded break repair mechanisms.
Besides the potential influences of mechanistic factors such as local degrees of sequence similarity and nucleic acid secondary structure on the recombination breakpoint distributions observed in our experiment, we were also interested in determining whether there existed evidence of this distribution having been influenced by natural selection. As with all life on Earth co-evolved sequence specific DNA-DNA, protein-protein and protein-DNA interactions have a central role in the molecular biology of begomoviruses and it is reasonable to suspect that natural selection will strongly disfavour the survival of recombinant begomovirus genomes in which these interactions are disrupted. For example, while intra-protein amino acid interactions determine how proteins fold, inter-protein amino acid interactions determine how proteins oligomerise or form complexes. Importantly, recombination can potentially disrupt such co-evolved interactions by bringing together pairs of amino acids that do not interact properly during protein folding, oligomerisation and complex formation [9].
We therefore tested for evidence of observed recombination events that occurred during our experiments being generally less disruptive of co-evolved intra-genome interaction networks than could be accounted for by chance. Specifically the interactions that we considered included (1) amino acid interactions within the 118 N-terminal amino acids of Rep and within the 196 amino acids of CP, (2) nucleotide interactions within predicted genomic ssDNA structures and (3) known protein-protein and protein DNA interactions between different genome regions.
Previous analyses of begomoviruses [8], other ssDNA viruses [41] and HIV [5], [7], have indicated that, assuming random recombination, natural recombinants tend to display lower degrees of predicted recombination induced protein folding disruption than can be accounted for by chance. Whereas recombination breakpoints in these viruses display a marked tendency to fall outside of coding regions, when they do fall within genes they tend to occur either on the edges of genes [5], [41], or near domain boundaries where they have a minimal impact on protein folding [7], [8].
Using the breakpoint clustering tests of Lefeuvre et al. [41], we detected a clear tendency for recombination breakpoints to preferentially fall outside of coding regions (p = 0.0001 and p = 0.1174 for the 50 sequence FULL and the 18 sequence FIT datasets respectively). It was also clearly evident that recombination breakpoints that did fall within genes tended to fall within the 5′ 12.5% and 3′ 12.5% of nucleotides of the genes significantly more frequently than they did within the rest of the genes (p<0.0001 and p = 0.0061 for the FULL and FIT datasets respectively). Although differences in the p-values obtained with the FULL and FIT datasets most likely reflect the greater sample size in the FULL dataset, it is possible that they indicate that selection has not been entirely responsible for these observed breakpoint patterns. For example, the recombination breakpoint hotspot at the virion strand replication origin automatically predisposes the IR to have more recombination breakpoints than the remainder of the genome.
We therefore focused on the C1 and V1 genes for which corresponding atomic resolution 3D protein structure models are available to allow us to more directly test for evidence of selection acting against recombinants expressing misfolded proteins. Specifically, we tested whether recombinants observed in our experiment tended to express chimaeric Rep and CP molecules with less potential folding disruption than what would have been expected if recombination breakpoints within these coding regions occurred at random. Considering all 50 unique recombination patterns within the FULL dataset we detected no evidence of the recombinants expressing Rep and CP molecules with better preservation of intra-protein amino acid interactions than would have been expected if recombination breakpoints occurred at random (p = 0.13 and 0.89 for Rep and CP respectively). While this result implied that selection against protein folding disruption had not obviously influenced the over-all distribution of observed recombination breakpoints falling within the C1 and V1 ORFS, it was anticipated that low frequency, potentially transient/defective recombinant forms that we included amongst the 50 FULL dataset recombinants may have obscured this signal of natural selection.
When we repeated the analysis with only the FIT dataset containing 18 recombinants that were viable enough to have been sampled multiple times during the experiment we detected a clear signal indicating that the recombination breakpoint patterns within their C1 ORFs were less disruptive of protein folding than expected by chance (p = 0.016). Although we detected no such signal within the V1 ORF (p = 0.47), this negative result was expected in that (1) TYX and TOX express CPs that differ at only two potentially interacting amino acid sites (implying that even the most disruptive recombinant would only result in two potential interacting residue perturbations); and (2) among the FIT recombinants there were only 8 detected recombination breakpoints within V1.
It is nevertheless interesting that we detected significant avoidance of Rep folding disruption amongst what we assume to be the fittest subset of recombinants and no such signal amongst the total set of recombinants. This implies that the selective processes acting against recombinants expressing improperly folded chimaeric proteins that are detectable in global begomovirus populations, are clearly operational during the short-term evolution of recombinants within the hosts where they originate.
The stability and distribution of DNA and RNA secondary structures are important fitness determinants in many viral species [59]–[62], including geminiviruses [63], [64]. Since nucleotide interactions within evolutionarily conserved secondary structures within the TYX and TOX genomes are another potential subset of co-evolved intra-genome interactions that might be disrupted by recombination, we investigated whether the recombinant genomes that arose during our experiment displayed evidence of selection favouring the preservation of base-pairing interactions within genomic secondary structures.
We first investigated whether the over-all stability of secondary structures within the 50 unique recombinant genomes observed in our experiment were significantly different from what would have been expected given random recombination. Specifically we compared the predicted MFE of the real recombinants to sets of simulated recombinants displaying both similar spacing between recombination breakpoints and the same numbers of TYX and TOX derived nucleotides as the real recombinants. Although we found some variation in the estimated MFEs of recombinants and parental viruses (i.e. recombination was predicted to have had some influence on MFE estimates), there were no significant differences in MFE estimates between the simulated and real recombinants (for either the total 50 sequence FULL set or the 18 sequence FIT set). This indicated that the degree to which overall secondary structural stability (as measured by genomic MFEs) is conserved between parental and recombinant sequences is not an overwhelmingly significant factor determining the short-term survival of recombinants in the hosts where they originate.
This result did not imply, however, that there was no evidence of selection favouring the maintenance among recombinants of specific nucleotide interactions (such as those within the predicted hairpin structure at the origin of virion strand replication). To test for such evidence we adopted exactly the same approach used in the SCHEMA-based protein folding disruption tests but applied it to predicted nucleotide-nucleotide contact maps inferred from predicted MFE secondary structures instead of amino acid-amino acid contact maps inferred from protein 3D structure models. Also, unlike the protein folding disruption tests we inferred separate secondary structures both for the parental TYX and TOX genomes and for the real and simulated recombinants. This latter point meant that we were able to differentiate between two distinct types of ssDNA folding disruption: (1) disruptions where the recombinants have potentially aberrant nucleotide base pairings that are not found in either the TYX or TOX secondary structures and (2) disruptions where recombinants are missing base pairings that are found within the TYX and TOX secondary structures.
Both of these SCHEMA based ssDNA folding disruption tests indicated clear evidence of the recombinants in our experiment displaying lower degrees of ssDNA folding disruption than was expected by chance (p = 4.9×10−3 for the FULL dataset and p = 3.7×10−2 for the FIT dataset with test (1) and p = 2.4×10−5 for the FULL dataset and p = 7.8×10−5 for the FIT dataset with test (2). It is very noteworthy, however, that by far the strongest signal of secondary structure preservation came from the test looking at the preservation of parent-like base-pairing patterns. While it is not acceptable to equate these estimated p-values with the strength of the selective process that have probably generated these signals, the vastly different strengths of the signals in tests 1 and 2 nevertheless suggests that selection might not disfavour the survival of recombinants with novel non-parental base pairings nearly as strongly as it disfavours those with broken parental-base pairings.
The simple fact that we have detected evidence that observed recombinants tend to have lower degrees of ssDNA folding disruption that might be expected by chance also implies that in begomoviruses (but possibly also in other ssDNA viruses too) many of the genomic secondary structures that are readily predictable using currently available DNA folding methods are probably evolutionarily relevant. Besides the probable hairpin structure at the virion strand origin of replication of Tomato golden mosaic virus [63] and one other structure within the rep gene of Maize streak virus [64], there remains no direct experimental evidence for the existence of widespread and biologically relevant base-pairing within single stranded geminivirus genomes. Our results, however, strongly suggest that if such evidence is sought it will probably be found.
While our previous analyses attempted to determine the influences of short range interactions on recombination patterns (such as those between amino acids within folded proteins and nucleotides within DNA secondary structures), we realised that the patterns of recombination that emerged within our experiment were potentially also influenced by longer-range inter-protein or protein-DNA interactions. We therefore devised a permutation-based genetic association test to both determine whether there existed any evidence that these known long-range interactions were preferentially preserved within the observed recombinants, and indicate whether these recombinants could be used to reveal other currently unknown intra-genome interactions.
If we assume that intra-genome interactions have co-evolved over time and that, if the partners of a co-evolved interacting pair (such as two encoded amino acids interacting within a folded protein or an encoded protein motif and the DNA sequence it recognizes) are separated by a recombination event, we would expect the recombination event to incur a fitness cost if the transferred half of the interacting pair did not work as well with its new interacting partner as it did with its old co-evolved one. We would therefore expect that recombinants in which different halves of interacting pairs have been derived from different parental sequences, would be selected against and would occur at lower than expected frequencies in the population.
We therefore sought to identify site pairs at which nucleotides tended to be co-inherited from the same parent more frequently than could be accounted for by chance such that mixtures of TYX and TOX derived polymorphisms were significantly under-represented at these sites in our recombinant population. The sites where these “mixed” polymorphism pairs were under-represented were compiled on an association map (Figure 5). Collectively we detected 4759 site-pairs (3.9% of the total site pairs queried) at which mixed TYX and TOX polymorphism combinations were significantly under-represented (i.e. with an associated permutation p-value < 0.05 for both of the association tests that we performed).
Most of these site pairs included one polymorphic site within the IR and the other in either (1) the overlapping region of the C2/C3 ORFs, (2) the overlapping region of the C4 and C1 ORFs, (3) the 5′ region of C1 and (4) the V1 ORF (Figure 5A). It is important to note here that detecting 3.9% of site-pairs with fewer than expected mixed TYX-TOX polymorphisms is in fact within the range expected by chance alone given a p-value cutoff of 0.05. It should be remembered, however, that these site-pairs were selected based on two separate tests both with a 0.05 p-value cutoff and that the anticipated rate of falsely inferred interactions is therefore expected to be substantially lower than 5%. Since we do not know what the expected false positive rate is with using the consensus of these two tests we suggest when consulting Figure 5 that only those site pairs identified with a p-value cutoff of 0.005 for both tests (orange and yellow lines) should be individually interpreted as displaying strong evidence of interacting with one another.
If the under-representation of particular “mixed” polymorphism pairs indicates that the genome regions carrying these pairs are probably involved in co-evolved intra-genome interactions, then at least some of the regions known to interact within geminivirus genomes should be among these site-pairs. This is in fact exactly what we find. Namely, (1) the inferred IR/C2/C3 interactions inferred from our recombination analysis match the TrAP-V1 promoter interaction sites recently mapped by Lacatus & Sunter [32] (Figure 5B), (2) the inferred C2/C3- C4/C1 interactions match with previously mapped REn and Rep interaction sites ([30]; Figure 5C), (3) the inferred 5′C1 - IR interaction correspond precisely to the Rep-iteron-related domain-iteron interaction sites proposed by Argüello-Astorga & Ruiz-Medrano [31] (Figure 5D) and (4) inferred V1-C1 interactions correspond with the Rep-CP interaction sites mapped by Malik et al. [50] (Figure 5E).
This preferential co-inheritance from a single parent of genome site pairs that are known to interact with one another clearly supports our hypothesis that, within our experiment, natural selection has probably disfavoured the survival of recombinants in which long-range intra-genome interactions have been disrupted.
Perhaps more importantly, however, the recovery by our association tests of these known interactions provides compelling evidence that applying these tests to recombination experiments such as we have performed can potentially also uncover unknown intra-genome interactions such as those indicated in Figure 5F.
When considering evidence provided by our association test for such unknown interactions it is important to stress that, as with most genetic association tests, ours suffers from an inescapable degree of genetic hitch-hiking that is expected to result in polymorphic sites which are physically closely linked on the genome displaying the same or very similar association patterns with sites in the remainder of the genome. This is because genome fragments that are exchanged by recombination usually carry multiple polymorphisms. This has two important consequences for our test. The first is that whereas the test might be capable of detecting long-range interactions, it is expected to have far less power to detect interactions between sites that are situated close to one another on the genome. The second is that many apparent interactions revealed by the test could be indirect. For example, the test might indicate that a pair of sites, A and B are always co-inherited from the same parent and are therefore probably interacting, but what is actually happening is that sites A and B are both in independently interacting with a third site, C. Importantly, however, as the number of analysed recombination events increases and the degree of genetic hitch-hiking decreases, one would expect both increases in the over-all accuracy with which interactions can be mapped by this approach, and increases in the power with which shorter-range interactions can be detected: Perhaps even including short range biologically relevant intra-protein amino acid contacts or structurally important nucleotide interactions.
Our recombination experiments yielded an array of recombinant genomes that far exceeded in their sheer numbers and complexity those encountered in other previously described geminiviral evolution experiments. Although polymorphisms derived from one or the other of the parental genomes were clearly selectively favoured at ∼64% of polymorphic sites within the arising recombinants, no highly deterministic single “recombinant solution” emerged from our experiment. Whereas a single prevalent recombinant solution such as has been found in other geminivirus recombination experiments [26], [51], [52] would have indicated the existence of a fitness landscape dominated by a single sharp peak, the unparalleled diversity of recombinants observed in our experiment possibly reflects the relative “flatness” of the tomato fitness-landscape over which the TYX and TOX recombinants evolved during our experiment. Although we cannot know whether over the long-term upon this flattish landscape some of the recombinants we observed could have out-competed their parental genomes under natural conditions, it is apparent that TYX and TOX are possibly more “genetically compatible” when it comes to making recombinants than other geminivirus combinations that have been tested in the past [13], [26], [65].
We show here that within begomovirus co-infections, local degrees of sequence similarity, but not genomic secondary structure, strongly influences the genesis of recombinant sequences. Once recombinants are produced, however, we additionally show that by determining which of these survive, natural selection profoundly influences the over-all patterns of recombination that emerge. While selection apparently favours the survival of recombinants with particular combinations of host adaptive polymorphisms, we show that it also clearly favours the survival of recombinant genomes within which various categories of intra-genome interaction are preserved. Relative to simulated recombinants, those emerging during our experiment tended to display (1) better preservation of amino acid interactions within their folded proteins, (2) better preservation of nucleotide interactions within predicted genomic secondary structures, and (3) better preservation of known long-range intra-genome protein-protein and protein-DNA interactions. We in fact show that the imprint left by natural selection on the patterns of recombination that emerged in these experiments was so profound that the patterns could be used to trace key features of the interaction networks encoded within the two parental genomes.
Finally, there is no reason that such recombination based approaches to high throughput mapping of either host adaptive polymorphisms or genomic interactions could not be applied to any recombinogenic virus species (for example most double stranded DNA viruses, retroviruses or positive strand RNA viruses). In cases where viruses do not naturally recombine (such as with many negative strand RNA viruses), random recombinants could be synthesised en masse and used to co-infect either host cells in culture or whole organisms. Besides illuminating the internal workings of viral genomes, systematic screening of different virus and host combinations would reveal the genetically compatible virus pair and host species combinations that are likely to produce the emergent recombinants of the future.
|
10.1371/journal.pgen.1008164 | Genome-wide association study of multisite chronic pain in UK Biobank | Chronic pain is highly prevalent worldwide and represents a significant socioeconomic and public health burden. Several aspects of chronic pain, for example back pain and a severity-related phenotype ‘chronic pain grade’, have been shown previously to be complex heritable traits with a polygenic component. Additional pain-related phenotypes capturing aspects of an individual’s overall sensitivity to experiencing and reporting chronic pain have also been suggested as a focus for investigation. We made use of a measure of the number of sites of chronic pain in individuals within the UK general population. This measure, termed Multisite Chronic Pain (MCP), is a complex trait and its genetic architecture has not previously been investigated. To address this, we carried out a large-scale genome-wide association study (GWAS) of MCP in ~380,000 UK Biobank participants. Our findings were consistent with MCP having a significant polygenic component, with a Single Nucleotide Polymorphism (SNP) heritability of 10.2%. In total 76 independent lead SNPs at 39 risk loci were associated with MCP. Additional gene-level association analyses identified neurogenesis, synaptic plasticity, nervous system development, cell-cycle progression and apoptosis genes as enriched for genetic association with MCP. Genetic correlations were observed between MCP and a range of psychiatric, autoimmune and anthropometric traits, including major depressive disorder (MDD), asthma and Body Mass Index (BMI). Furthermore, in Mendelian randomisation (MR) analyses a causal effect of MCP on MDD was observed. Additionally, a polygenic risk score (PRS) for MCP was found to significantly predict chronic widespread pain (pain all over the body), indicating the existence of genetic variants contributing to both of these pain phenotypes. Overall, our findings support the proposition that chronic pain involves a strong nervous system component with implications for our understanding of the physiology of chronic pain. These discoveries may also inform the future development of novel treatment approaches.
| Chronic pain is common worldwide and imposes a significant burden from a public health and socioeconomic perspective. The reasons why some individuals develop chronic pain and others do not are not fully understood. In this study we searched for genetic variants associated with chronic pain in a large general-population cohort. We also assessed how this genetic variation was correlated with a range of other diseases and traits, such as depression and BMI, and we tested for causal relationships between depression and chronic pain. We found that chronic pain was associated with several genes involved in brain function and development and was correlated with mental health and autoimmune traits (including depression, PTSD and asthma). We also found evidence for causal relationships between chronic pain and major depressive disorder. This work provides new insights into the genetics and underlying biology of chronic pain and may help to inform new treatment strategies.
| Chronic pain, conventionally defined as pain lasting longer than 3 months, has high global prevalence (~30%; [1]), imposes a significant socioeconomic burden, and contributes to excess mortality [2,3]. It is often associated with both specific and non-specific medical conditions such as cancers, HIV/AIDS, fibromyalgia and musculoskeletal conditions [4–6], and can be classified according to different grading systems, such as the Von Korff chronic pain grade [7]. Several aspects of chronic pain, such as chronic pain grade and back pain, have been studied from the genetic point of view, and several have been shown to be complex traits with moderate heritability [3,8]. In part due to the heterogeneity of pain assessment and pain experience, there are very few large-scale genetic studies of chronic pain and no genome-wide significant genetic variants have yet been identified [9,10].
Chronic pain and chronic pain disorders are often comorbid with psychiatric and neurodevelopmental disorders, including Major Depressive Disorder (MDD) [11]. The immune and nervous systems play a central role in chronic pain development and maintenance [12,13]. Similarly, obesity and chronic pain are often comorbid, with extrinsic factors such as sleep disturbance also impacting on chronic pain [14,15]. Altered sleep quality and reduced circadian rhythmicity are also common in those with chronic pain [16]. Chronic pain is also a common component of many neurological diseases [17].
The relationship between injury and other peripheral insult, consequent acute pain and the subsequent development of chronic pain has not been fully explained. Not everyone who undergoes major surgery or is badly injured will develop chronic pain, for example [18], and the degree of joint damage in osteoarthritis is not related to chronic pain severity [19]. Conversely, Complex Regional Pain Syndrome (CRPS) can be incited by minor peripheral insult such as insertion of a needle (reviewed by Denk, McMahon and Tracey, 2014). Structural and functional changes in the brain and spinal cord are associated with the development and maintenance of chronic pain, and affective brain regions are involved in chronic pain perception (this is in contrast to acute pain and even to prolonged acute pain experience) [20–24]. It is also unlikely that there are legitimate cut-off points or thresholds for localised and widespread chronic pain, with pain instead existing on a “continuum of widespreadness” [25]. It may, therefore, be more valuable and powerful to examine measures of chronic pain as complex neuropathological traits in themselves, rather than just to study disorders and conditions with chronic pain as a main feature or pain experienced only in specific bodily locations. Our aim in this study was predicated on the idea that predisposing biological processes might influence how many sites are affected in individuals that experience any chronic pain, and we carried out a genome-wide association study of number of chronic pain sites to look for predisposing loci, assess the degree of genetic overlap with related traits and disorders and generate insights into the genetic architecture of chronic pain.
To identify genetic risk loci influencing Multisite Chronic Pain (MCP), we performed a GWAS with adjustment for age, sex and genotyping array using BOLT-LMM (see Methods). No evidence was found for inflation of the test statistics due to hidden population stratification (λGC = 1.26; after adjustment for sample size λGC1000 = 1.001). LD-score regression (LDSR) analysis was consistent with a polygenic contribution to MCP (LDSR intercept = 1.0249, SE 0.0274; Fig 1) [26] and yielded a Single Nucleotide Polymorphism (SNP) heritability estimate of 10.2%. BOLT-LMM gave a similar SNP heritability estimate (pseudo-h2 = 10.3%). In total, 1, 748 SNPs associated with MCP level at genome-wide significance (p < 5 x 10−8) were identified. Conditional analysis of the association signals at each locus revealed 76 independent genome-wide significant lead SNPs across 39 risk loci located on chromosomes 1–11, 13–18 and 20 (Table 1). Sensitivity analysis additionally adjusting for BMI did not significantly alter these association analysis results.
Genomic risk loci are as defined by FUMA. Genomic Locus = numeric label (1–39), rsID = SNP rsID label, chr = chromosome, pos = position in base-pairs, Nearest Gene = nearest mapped gene, A1 = effect allele, A2 = non-effect allele, MAF = minor allele frequency (MAF here refers to A1 frequency as all values are < 0.5 i.e. A1 is the minor allele as well as the effect allele), r2 = imputation r-squared value, beta = association beta value, se = standard error of beta, P = P value of association (GWAS P value).
Post-GWAS analyses including gene expression and gene-level association testing was carried out using FUMA. Gene-level association tests (MAGMA gene-based test) revealed 113 genes across 39 genomic risk loci significantly associated with MCP (S1–S3 Figs), including genes with roles in neuronal adhesion and guidance, regulation of neural development and neurotransmitter receptor function.
Analysis of Gene Ontology (GO) annotations revealed 3 significant categories (Table 2: Bonferroni-corrected p < 0.05). The significant categories were enriched for terms including neurogenesis and synaptic plasticity, DCC-mediated attractive signalling, neuron projection guidance and central nervous system neuron differentiation, amongst others. Genes of interest (n = 35) designated based on gene-level association tests and on annotation of genes at the identified genomic loci (see S1 Text) are listed in S2 Table. Analysis of tissue-level expression showed significant enrichment of brain-expressed genes, particularly in the cortex and cerebellum (Fig 2),
Genetic correlations between MCP and 22 traits were estimated via LD-score regression using ldsc [28]. The psychiatric phenotype most significantly genetically correlated with MCP was MDD (Table 3: rg = 0.53, pFDR = 1.69e-78) while the largest significant genetic correlation coefficient was for MCP and depressive symptoms (Table 3: rg = 0.59, pFDR = 6.19e-65). MCP was also positively genetically correlated with neuroticism (Table 3: rg = 0.40), anxiety (Table 3: rg = 0.49), schizophrenia (Table 3:rg = 0.10), cross-disorder psychiatric phenotype (Table 3: rg = 0.13) and PTSD (Table 3: rg = 0.41). Significant negative genetic correlations were observed between MCP and subjective well-being (Table 3: rg = -0.36), autism spectrum disorder (Table 3: ASD; rg = -0.10) and between MCP and anorexia nervosa (Table 3: AN; rg = -0.06). There was no significant genetic correlation between MCP and Bipolar disorder (Table 3: BD; PFDR > 0.05). In relation to the immune-related disorders, rheumatoid arthritis (Table 3: rg = 0.16) and asthma (Table 3: rg = 0.22) were significantly positively genetically correlated with MCP, as was primary biliary cholangitis (Table 3: rg = 0.10), while systemic lupus erythematosus (SLE), ulcerative colitis and Crohn disease were not (PFDR > 0.05). BMI was significantly genetically correlated with MCP (Table 3: rg = 0.31), while low relative amplitude, a circadian rhythmicity phenotype, exhibited a significant negative genetic correlation with MCP (Table 3: rg = -0.30). There was no correlation between Parkinson disease and MCP (PFDR > 0.05). Non-significant genetic correlation results are shown in S3 Table.
Mendelian Randomisation with Robust Adjusted Profile Score (MR-RAPS) analysis was performed to investigate causal relationships between MDD and MCP, first with MDD as the exposure and MCP as the outcome. QQ plots, leave-one out versus t-value plots (S4 Fig) and Anderson-Darling/ Shapiro-Wilk test p values indicated that models without dispersion were best-fitting (S4 Table rows 1–3, pAD > 0.05, pSW > 0.05). Effects of outliers (idiosyncratic pleiotropy) are not ameliorated in models with dispersion despite robust regression (S4D, S4E and S4F Fig right-hand panels). The model allowing the greatest amelioration of pleiotropy is one without over-dispersion and with a Tukey loss function (S4 Table: row 3, S4C Fig). This indicates idiosyncratic pleiotropy (pleiotropy in some but not all instruments), i.e. that a subset of instruments may affect MCP through pathways other than via MDD (the exposure). The causal effect of MDD on MCP is positive and significant at beta = 0.019 and p = 0.0006, but the diagnostic plots show a ‘swapping’ of sign for the causal estimate (S4 Fig), suggesting that there is not a truly significant causal effect of MDD on MCP.
MR-RAPS analyses were then carried out with MCP as the exposure and MDD as the outcome. Models with dispersion are a better fit than those without (S5A, S5B, S5C vs S5D, S5E and S5F Fig, S5 Table: rows 4–6, pAD > 0.05, pSW > 0.05, pτ << 0.05). This indicates that effectively all instruments are pleiotropic (affecting MDD through pathways other than via MCP). The causal effect of MCP on MDD is positive and significant at beta = 0.16 and p = 0.047.
Overall, this analysis suggests a causal effect of MCP on MDD.
Polygenic Risk Score (PRS) analyses were carried out to examine the relationship between MCP and chronic widespread pain in UK Biobank. Increasing MCP PRS value was significantly associated with having chronic pain all over the body (S6 Table: p = 1.45 x 10−109), with each per-standard-deviation increase in PRS associated with a 63% increase in the odds of having chronic widespread pain.
A secondary GWAS of chronic widespread pain (CWP) was carried out, the results from which were used in LD score regression analysis to determine the genetic correlation between CWP and MCP. This was found to be large (rg = 0.83) and significant (p = 2.45 x 10−54). A lookup analysis was also carried out using the CWP GWAS summary statistics, and >90% of SNPs showed consistent direction of effect between MCP and CWP (S7 Table). In addition, a paired t-test of MCP versus CWP effect values showed that they are not significantly different overall (t = -1.82, p = 0.07).
LocusZoom plots for independent, genome-wide significant loci, calculated according to the supplementary methods detailed in S1 Text, are shown in S6 Fig.
We identified 76 independent genome-wide significant SNPs associated with MCP across 39 loci. The genes of interest had diverse functions, but many were implicated in nervous-system development, neural connectivity and neurogenesis.
Potentially interesting genes included DCC (Deleted in Colorectal Cancer a.k.a. DCC netrin 1 receptor) which encodes DCC, the receptor for the guidance cue netrin 1, which is important for nervous-system development [29]. SDK1 (Sidekick Cell Adhesion molecule 1) is implicated in HIV-related nephropathy in humans [30] and synaptic connectivity in vertebrates [31], and ASTN2 (Astrotactin 2) is involved in glial-guided neuronal migration during development of cortical mammalian brain regions [32].
MAML3 (Mastermind-Like Transcriptional coactivator 3) is a key component of the Notch signalling pathway [33,34], which regulates development and maintenance of a range of cell and tissue types in metazoans. During neurogenesis in development the inhibition of Notch signalling by Numb promotes neural differentiation [35]. Numb is encoded by NUMB (Endocytic Adaptor Protein), which was also associated with MCP. In the adult brain Notch signalling has been implicated in CNS plasticity across the lifespan [35].
CTNNA2 (Catenin Alpha 2) encodes a protein involved in cell-cell adhesion [36], found to play a role in synapse morphogenesis and plasticity [37,38]. CEP120 (Centrosomal Protein 120) encodes Cep120, vital for Interkinetic Nuclear Migration (INM) in neural progenitor cells of the cortex [39]. KNDC1 (Kinase Non-Catalytic C-Lobe Domain Containing 1) encodes v-KIND in mice, linked to neural morphogenesis in the cortex [40], and KNDC1 in humans, linked to neuronal dendrite development and cell senescence [41]. SOX6 (SRY-Box 6) is part of the Sox gene family, first characterised in mouse and human testis-determining gene Sry [42] and encoding transcription factors involved in a range of developmental processes [43,44]. SOX6 may be involved in development of skeletal muscle [43], maintenance of brain neural stem cells [45] and cortical interneuron development [46], and variants in this gene have been associated with bone mineral density in both white and Chinese populations [47]. CA10 (Carbonic Anhydrase 10) is predominantly expressed in the CNS, encoding a protein involved in development and maintenance of synapses [48]. DYNC1I1 (Dynein Cytoplasmic 1 Intermediate Chain 1) encodes a subunit of cytoplasmic dynein, a motor protein which plays a role in cargo transport along microtubules, including in the function of neuronal cells [49]. UTRN (Utrophin) is a homologue of Duchenne Muscular Dystrophy gene (DMD), encoding utrophin protein which is localised to the neuromuscular junction (NMJ) [50]. Utrophin has also been implicated in neutrophil activation [51], dystrophin-associated-protein (DPC)-like complex formation in the brain [52], and is expressed during early foetal brain development in neurons and astrocytes [53].
FOXP2 encodes a member of the FOX family of transcription factors, which are thought to regulate expression of hundreds of genes in both adult and foetal tissue, including the brain [54]. These transcription factors may play an important role in brain development, neurogenesis, signal transmission and synaptic plasticity [55]. FOXP2 is essential for normal speech and language development [56]. GABRB2 encodes a GABA (gamma-aminobutyric acid) type A receptor beta subunit. These pentameric chloride channels mediate fast inhibitory synaptic transmission and are extremely important for network function in many brain regions, with the b2 subunit forming part of the most widely expressed receptor across the mammalian brain [57,58].
Another group of genes associated with MCP were linked to cell-cycle progression, DNA replication and apoptosis such as EXD3 (Exonuclease 3’-5’ Domain Containing 3), which encodes a protein involved in maintaining DNA fidelity during replication (‘proof-reading’) [59]. BBX (HMG-Box Containing protein 2) encodes an HMG (high mobility group) box-containing protein necessary for cell-cycle progression from G1 to S phase [60]. STAG1 (Cohesin Subunit SA-1) encodes a cohesin-complex component–cohesin ensures sister chromatids are organised together until prometaphase [61–63]. ANAPC4 (Anaphase Promoting Complex Subunit 4) encodes a protein making up the anaphase promoting complex (APC), an essential ubiquitin ligase for eukaryotic cell-cycle progression [64]. PRC1 (Protein Regulator of Cytokinesis 1) is involved in the regulation of cytokinesis [65], the final stage of the cell cycle. Y RNA (Small Non-Coding RNA, Ro-Associated Y3) encodes a small non-coding Y RNA. These RNAs have been implicated in a wide range of processes, including cell stress response, DNA replication initiation and RNA stability [66]. FAM120A (Oxidative Stress-Associated Src Activator) encodes an RNA-binding protein which regulated Src-kinase activity during oxidative stress-induced apoptosis [67]. The protein encoded by MON1B (MON1 Homolog B, Secretory Trafficking Associated) is necessary for clearance of cell ‘corpses’ following apoptosis, with defects associated with autoimmune pathology [68]. FAF1 (Fas Associated Factor 1) encodes a protein which binds the Fas antigen to initiate or facilitate apoptosis, amongst a wide range of other biological processes (including neuronal cell survival) [69].
Several MCP associated genes have been previously implicated in diseases such as Brugada Syndrome 9 and Spinal ataxia 19 & 22 (KCND3) [70–72], Systemic lupus erythematosus (SLE) (Y RNAs) [66], Joubert syndrome 31 and short-rib thoracic dysplasia 13 (CEP120) [73], Amyotrophic lateral sclerosis (ALS) (FAF1) [74], Urbach-Wiethe disease (ECM1) [75,76], mental retardation and other cohesinopathies such as Cornelia de Lange Syndrome (STAG1) [77,78], split hand/ split foot malformation (DYNC1I1) [79,80], and a wide range of cancers (PRC1) [81]. Other disorders found to involve MCP-related genes include schizophrenia (FOXP2 and GABRB2) [82–88], intellectual disability and epilepsy (GABRB2) [89], and neuroleptic-induced tardive dyskinesia (GABRB2) [90].
Several GWASs of chronic pain at specific body sites, of specific pain types such as neuropathic pain, and of diseases and disorders where chronic pain is a defining symptom, have been carried out previously (reviewed by [10], [91]). DCC and SOX5 (which jointly functions with SOX6 in chondrogenesis) have been associated with chronic back pain [92], GABRB3 (encoding one of three beta subunits of the GABA A receptor along with GABRB2) has been associated with migraine and fibromyalgia [10], and ASTN2 and SLC24A3 have been associated with migraine [10,93]
Overall, this indicated that MCP, a chronic pain phenotype, involves structural and functional changes to the brain, including impact upon neurogenesis and synaptic plasticity both during development and in adulthood. Also implicated was regulation of cell-cycle progression and apoptosis. This is also supported by GO categories DCC-mediated attractive signalling, neuron projection guidance and CNS neuron differentiation being significantly associated with MCP. There was also evidence of pleiotropy, with genes associated with a range of neurodegenerative, psychiatric, developmental and autoimmune disease traits, as well as being associated with MCP.
Chronic pain and chronic pain disorders are often comorbid with psychiatric and neurodevelopmental disorders [11]. This has been observed for Major Depressive Disorder (MDD) [8,94], post-traumatic stress-disorder (PTSD) [95–99], schizophrenia [100–102] and bipolar disorder (BD) [94,103]. There are also reported differences in the perception of pain and interoception (sensing and integration of bodily signals) for people with schizophrenia [104,105], anorexia nervosa (AN) [106–108] and autism spectrum disorders (ASD) [109,110], with some evidence of an increase in pain thresholds for AN and ASD.
There is significant cross-talk between the immune system and nervous system in nociception and sensitisation leading to chronic pain [12,13], and many autoimmune disorders cause or have been associated with chronic pain including neuroinflammation implicated in development of neuropathic pain [111].
Similarly, obesity and chronic pain are often comorbid, with extrinstic factors such as MDD and sleep disturbance also impacting on chronic pain [14,15]. Obesity and related chronic inflammation may affect chronic pain [112], and adipose tissue is metabolically active in ways that can affect pain perception and inflammation [113–115].
Sleep changes and loss of circadian rhythm is common in those with chronic pain [16], and myriad chronic diseases, including chronic pain, have shown diurnal patterns in symptom severity, intensity and mortality [116,117]. Chronic pain is also a common component of many neurological diseases, particularly Parkinson’s disease [17], and disorders such as Multiple Sclerosis and migraines are considered neurological in nature.
MCP showed moderate positive genetic correlation with a range of psychiatric disorders including MDD, SCZ, and PTSD, along with traits anxiety and neuroticism. The magnitude of genetic correlation between MCP and MDD was similar to that shown for von Korff chronic pain grade (a chronic pain phenotype) and MDD by McIntosh et al via a mixed-modelling approach (ρ = 0.53) [8]. This is in line with previous observations of association and indicates that shared genetic risk factors exist between MCP and a range of psychiatric disorders, most notably MDD, and that the genetic correlation between MCP and MDD matches with that between MDD and von Korff CPG, a validated chronic-pain questionnaire-derived phenotype [7].
Autoimmune disorders rheumatoid arthritis, asthma and primary biliary cholangitis showed positive genetic correlation with MCP. However, gastrointestinal autoimmune disorders UC, IBD and Crohn’s Disease did not. This suggests separate genetic variation and mechanisms underlying chronic pain associated with these autoimmune disorders compared to those outwith the digestive system. Pain related to inflammatory bowel diseases may represent something less ‘chronic’ and more ‘on-going acute’, as stricture, abscesses and partial or complete obstruction of the small bowel result in pain [118]. Structural and functional brain changes associated with the transition to chronic pain may also play a less central role in gastrointestinal autoimmune disorder-associated pain, due to potential for the enteric nervous system (ENS) to act independently from the CNS, and the role of the gut-brain axis (GBA) [119,120].
There was significant negative genetic correlation between low relative amplitude, a circadian rhythmicity phenotype indicating poor rhythmicity [121]. Opposing direction of effect of genetic variants on MCP versus low RA may mean that insomnia and other sleep difficulties (for which low RA represents a proxy phenotype) associated with MCP are due to environmental and lifestyle factors related to chronic pain, rather than shared genetic factors predisposing to increased risk for both traits. There was also significant negative genetic correlation between MCP and both AN and ASD, which may be linked to changes in interoception and atypical pain experience seen in individuals with these conditions [106–110], and may suggest a genetic basis for increased pain thresholds.
LDSR analyses gave a heritability estimate of 10.2% for MCP, lower than the pseudo-h2 estimate of 10.3% given by BOLT-LMM. this suggests SNP-heritability (h2) of MCP to be roughly-10%, slightly lower than an estimate of ‘any chronic pain’ of 16%, and markedly lower than a heritability estimate of 30% for ‘severe chronic pain’ derived from a pedigree-based analyses [3].
Mendelian randomisation analyses indicated a causal effect of MCP on MDD, with widespread pleiotropy and a less significant causal estimate value for MCP as the exposure–this suggests most instruments for MCP are pleiotropic, affecting MDD through pathways other than directly through MCP. In contrast, only a small subset of instruments for MDD as the exposure were found to be pleiotropic.
It has been argued that CWP and other clinical syndromes involving chronic pain all over the body represent the upper end of a spectrum of centralisation of pain, or the extreme of a chronic pain state [122]. It has also been argued that there are not “natural cut-off points” when it comes to chronic widespread pain versus localised chronic pain [25]. In support of this view, the MCP PRS was significantly associated with increased odds of having chronic pain all over the body/ CWP, suggesting that chronic widespread pain may in fact represent the upper end of a spectrum of ‘widespreadness’ of chronic pain, as previously suggested [25,122], and that there are likely to be genetic variants that predispose both to MCP and to CWP.
Multisite chronic pain (MCP), a chronic pain phenotype defined as the number of sites at which chronic pain is experienced, is a complex trait with moderate heritability. To date, this study represents the largest GWAS of any chronic pain phenotype and elucidates potential underlying mechanisms of chronic pain development. Substantial genetic correlations with a range of psychiatric, personality, autoimmune, anthropometric and circadian traits were identified.
The genes potentially associated with MCP implicated neurogenesis, neuronal development and neural connectivity, along with cell-cycle and apoptotic processes, and expression was primarily within brain tissues. This is in line with theories of functional and structural changes to the brain contributing to development of chronic pain [21,24,123–125], and may also explain the genetic correlations observed. A causal effect of MCP on MDD was identified.
Although the phenotype was based on self-report, this study was very large in size and so likely had sufficient power to detect genetic variation associated with MCP. Replication of SNP associations was not possible due to the nature of chronic pain phenotyping and available cohort sizes, but several genes significantly associated with MCP have been previously associated with chronic pain conditions including chronic back pain, migraine and fibromyalgia, and genetic risk for MCP was found to be significantly associated with chronic widespread pain.
We carried out a GWAS of Multisite Chronic Pain (MCP), a derived chronic pain phenotype, in 387,649 UK Biobank participants (Table 4). UK Biobank is a general-population cohort of roughly 0.5 million participants aged 40–79 recruited across the UK between 2006 and 2010. Details on phenotyping, follow-up and genotyping have been described in detail elsewhere [126].
During the baseline investigations, UK Biobank participants were asked via a touchscreen questionnaire about “pain types experienced in the last month” (field ID 6159), with possible answers: ‘None of the above’; ‘Prefer not to answer’; pain at seven different body sites (head, face, neck/shoulder, back, stomach/abdomen, hip, knee); or ‘all over the body’. The seven individual body-site pain options were not mutually exclusive and participants could choose as many as they felt appropriate. Where patients reported recent pain at one or more body sites, or all over the body, they were additionally asked (category ID 100048) whether this pain had lasted for 3 months or longer. Those who chose ‘all over the body’ could not also select from the seven individual body sites.
Multisite Chronic Pain (MCP) was defined as the sum of body sites at which chronic pain (at least 3 months duration) was recorded: 0 to 7 sites. Those who answered that they had chronic pain ‘all over the body’ were excluded from the GWAS as there is some evidence that this phenotype relating to widespread pain can be substantially different from more localised chronic pain [94] and should not, therefore, be considered a logical extension of the multisite scale. 10,000 randomly-selected individuals reporting no chronic pain were excluded from the GWAS to use as controls in subsequent polygenic risk score (PRS) analyses.
SNPs with an imputation quality score of less than 0.3, Minor Allele Frequency (MAF) < 0.01 and Hardy-Weinberg equilibrium (HWE) test p < 10−6 were removed from the analyses. Participants whose self-reported sex did not match their genetically-determined sex, those who had putative sex-chromosome aneuploidy, those considered outliers due to missing heterozygosity, those with more than 10% missing genetic data and those who were not of self-reported white British ancestry were excluded from analyses.
An autosomal GWAS was run using BOLT-LMM [127], with the outcome variable, MCP, modelled as a linear quantitative trait under an infinitesimal model, and the model adjusted for age, sex and chip (genotyping array). Related individuals are included and accounted for, as are any population stratification effects, via use of a genetic relatedness matrix as part of the BOLT-LMM analysis [127]. The SNP-level summary statistics from the GWAS output were analysed using FUMA [128], which implements a number of the functions from MAGMA (gene-based association testing, gene-set analyses) [129]. Tissue expression (GTEx) analysis [130] and Gene Ontology [27] and ANNOVAR [131] annotation analysis with default settings was used to characterise lead SNPs further. LocusZoom [132] was used to plot association results at higher resolution (N = 47) (S1 Text). Genomic risk loci were identified using the definition deployed by FUMA [128].
Genetic correlations between MCP and 22 complex traits selected on the basis of prior phenotypic association evidence were calculated using linkage disequilibrium score regression (LDSR) analyses [28], implemented either using the ‘ldsc’ package [28] and downloaded publicly-available summary statistics and summary statistics from in-house analyses or using LD Hub [133]. LD Hub datasets from the categories Psychiatric, Personality, Autoimmune and Neurological were selected and datasets with the attached warning note ‘Caution: using this data may yield less robust results due to minor departure from LD structure’ were excluded from the analyses. Where multiple GWAS datasets were available for the same trait, the one with the largest sample size and/or European ancestry was retained with priority given to European ancestry.
Mendelian randomisation analysis was carried out with MR-RAPS (MR-Robust Adjusted Profile Score; [134] using the R package ‘mr-raps’. This method is appropriate when doing MR analysis of phenotypes that are moderately genetically correlated and likely to share some pleiotropic risk loci. MDD was chosen for MR analysis as this disorder represents an important and common comorbidity with chronic pain [2,8,135]. Summary statistics from the most recent MDD GWAS meta-analysis [136], with UK Biobank and 23andMe results removed, were harmonised with MCP GWAS summary statistics following guidelines [137] as closely as possible with the available data. Bi-allelic SNPs shared between the two datasets were identified and harmonised (by ‘flipping’) with respect to the strand used to designate alleles. Reciprocal MR analysis was carried out using subsets of SNPs associated with each of the exposure traits (MCP and MDD) at p < 10−5. This threshold is an order of magnitude lower than suggested as part of the MR-RAPS method [134] and was chosen in order to attempt to account for ‘winner’s curse’, as independently selecting and then testing association for instruments in separate GWAS datasets was not possible in this study. The harmonisation process also involved ensuring that the effect allele was trait-increasing in the exposure trait, and that the effect allele matched between the exposure and the outcome. These selected subsets of variants were then LD-pruned at a threshold of r2 < 0.01 using command-line PLINK using ‘indep-pairwise’ with a 50-SNP window and sliding window of 5 SNPs [138]. This resulted in a set of 200 instruments for MCP as the exposure, and a set of 99 instruments for MDD as the exposure.
Those who reported chronic pain all over the body were excluded from the MCP GWAS analyses above. This is because chronic pain all over the body, taken as a proxy for chronic widespread pain (CWP), may be a different clinical syndrome from more localised chronic pain, and does not necessarily directly reflect chronic pain at 7 bodily sites. To investigate the relationship between CWP and MCP, a polygenic risk score (PRS) approach was taken.
A PRS was constructed for MCP in individuals who reported chronic pain all over the body (n = 6,815; these individuals had all been excluded from the MCP GWAS), and in controls (n = 10,000 individuals reporting no chronic pain at any site, also excluded from the MCP GWAS). The PRS was calculated using SNPs associated with MCP at p < 0.01, weighting by MCP GWAS effect size (GWAS β) for each SNP. A standardised PRS (based on Z-scores) was used in all analyses, constructed by dividing the calculated PRS by its standard deviation across all samples. The ability of the standardised PRS to predict chronic widespread pain status was investigated in logistic regression models adjusted for age, sex, genotyping array and the first 8 genetic principal components.
Individual-level data are available via application to UK Biobank. Multisite chronic pain GWAS summary statistics are available via contacting the authors and will be submitted to UK Biobank for publication at their website.
|
10.1371/journal.pcbi.1003366 | Designing Molecular Dynamics Simulations to Shift Populations of the Conformational States of Calmodulin | We elucidate the mechanisms that lead to population shifts in the conformational states of calcium-loaded calmodulin (Ca2+-CaM). We design extensive molecular dynamics simulations to classify the effects that are responsible for adopting occupied conformations available in the ensemble of NMR structures. Electrostatic interactions amongst the different regions of the protein and with its vicinal water are herein mediated by lowering the ionic strength or the pH. Amino acid E31, which is one of the few charged residues whose ionization state is highly sensitive to pH differences in the physiological range, proves to be distinctive in its control of population shifts. E31A mutation at low ionic strength results in a distinct change from an extended to a compact Ca2+-CaM conformation within tens of nanoseconds, that otherwise occur on the time scales of microseconds. The kinked linker found in this particular compact form is observed in many of the target-bound forms of Ca2+-CaM, increasing the binding affinity. This mutation is unique in controlling C-lobe dynamics by affecting the fluctuations between the EF-hand motif helices. We also monitor the effect of the ionic strength on the conformational multiplicity of Ca2+-CaM. By lowering the ionic strength, the tendency of nonspecific anions in water to accumulate near the protein surface increases, especially in the vicinity of the linker. The change in the distribution of ions in the vicinal layer of water allows N- and C- lobes to span a wide variety of relative orientations that are otherwise not observed at physiological ionic strength. E31 protonation restores the conformations associated with physiological environmental conditions even at low ionic strength.
| Calmodulin (CaM) is involved in calcium signaling pathways in eukaryotic cells as an intracellular Ca2+ receptor. Exploiting pH differences in the cell, CaM performs a variety of functions by conveniently adopting different conformational states. We aim to reveal pH and ionic strength (IS) dependent shifts in the populations of conformational substates by modulating electrostatic interactions amongst the different regions of the protein and with its vicinal water. For this purpose, we design extensive molecular dynamics simulations to classify the effects that are responsible for adopting different conformations exhibited in the ensemble of NMR structures reported. Lowering the IS or pH, CaM experiences higher inter-lobe orientational flexibility caused by extreme change in the non-specific ion distribution in the vicinal solvent. Amongst the titratable groups sensitive to pH variations, E31 is unique in that its protonation has the same effect on the vicinal layer as increasing the IS. Furthermore, E31A mutation causes a large, reversible conformational change compatible with NMR ensemble structures populating the linker-kinked conformations. The mutation in the N lobe, at a significant distance, both modulates the electrostatic interactions in the central linker and alters the EF-hand helix orientations in the C lobe.
| Protein behavior in solution may be manipulated and controlled through tailored structural perturbations [1] and rational control of the solution conditions [2] http://www.pnas.org/content/109/50/E3454.full.pdfhtml. In the living cell, proteins adapt to particular subcellular compartments which pose different environmental variables such as pH and ionic strength (IS), adapting their biophysical characteristics to tolerate pH fluctuations that are caused by cellular function [3]. Furthermore, proteins interact with many other biological macromolecules while they are transferred from one compartment to another, with subtle control over protonation and pK changes upon binding to other proteins and ligands [4], [5]. Interactions with the environment and other molecules are closely related to local anisotropy and dynamical heterogeneity of proteins [6]. The dynamics may be electrostatically guided, perhaps through long-range electrostatic interactions that select and bring interacting partners together, steering the protein to alternative conformations [7].
The main perturbation effect of the long-range electrostatic interactions is manifested on the acidic/basic groups in the protein which can be charged or neutral in relation to their conformation dependent pKa values [8]. Interacting with other molecules and changes in the environmental variables such as subcellular localization can induce shifts in ionization states of charged groups on a protein by proton uptake/release. Such changes facilitate the protein to span a large conformational space and enable it to participate in diverse interaction scenarios. Any in depth understanding developed through studying the conformational changes in proteins induced by shifts in the charge states of select amino acids would contribute to our knowledge base on diverse functionality observed in promiscuous proteins [9].
In this study, we focus on the conformation-related effects of introducing perturbations on charged group(s) of calmodulin (CaM). CaM is a notorious example among proteins having the ability to change conformation upon binding to diverse ligands [10], [11]. It was shown that negatively charged side chains in calcium loaded CaM (Ca2+-CaM) are attracted to positively charged residues in many of its targets [12]. Another study showed that Ca2+-CaM changes conformation when introduced to a solvent at low pH and low ionic strength [13]. Also, it was proposed that electrostatic interactions between acidic residues in CaM contribute to determining the most populated conformation under varying solution conditions [13].
Previously, we have studied the conformational changes in Ca2+-CaM [14], ferric binding protein [15] and a set of 25 proteins that display a variety of conformational motions upon ligand binding (e.g., shear, hinge, allosteric) [16] using the perturbation response scanning method. This coarse grained methodology is based on the assumption that the equilibrium fluctuations at a given local free energy minimum of the protein possess information on other viable conformations when an external force is applied [15]–[17]. Our study on CaM determined key residues that lead to the experimentally observed conformational changes upon application of force in specific directions [14]. Several different servers (H++ [18], propKa 2.0 [19], pKd [20] and PHEMTO [21], [22]) showed that the pKa of E31 value is upshifted; furthermore, the equivalent position in Calbindin was measured to have pKa of 6.5 [23]. In a follow-up study, we focused on residues with upshifted pKa values and we made a systematic study of the dynamics of Ca2+-CaM on time scales up to 200 ns for three separate initial configurations; extended form, compact form and extended structure with 10 protonated residues (9 acidic residues and a histidine) [24]. We found that Ca2+-CaM with 10 protonated residues undergoes a large conformational shift from the extended structure to a relatively compact form on the time scale of tens of nanoseconds. The latter was compatible with other structures reported in a nuclear magnetic resonance (NMR) ensemble of CaM [25].
Experimental work investigating dynamical behavior of Ca2+-CaM has shown that it occupies a number of hierarchical set of substates even in the crystal form [26]. Dynamical information obtained from fluorescence resonance energy transfer (FRET) experiments measuring the distance distributions between labeled sites illustrate that at least two conformations exist in solution under physiological conditions [27]. More recently, pseudo contact shifts and residual dipolar couplings of the C-terminal domain obtained using NMR [28] revealed neither the dumbbell shaped conformation observed in early crystal structures of the molecule [29], nor the compact conformation determined later on [30] exist in significant proportions in solution. Ca2+-CaM is identified as a protein which populates multiple conformations [28], [31]. A shift between the distribution of populations is induced by changing environmental conditions such as pH, Ca2+ concentrations and ionic strength [13], [27]. Each of these manipulated properties has effects on the charged groups of Ca2+-CaM.
The presence of multiple conformations is a physical property of Ca2+-CaM, and it is likely that the heterogeneity of structure is at least partially responsible for the ability of Ca2+-CaM to recognize diverse targets. Squier and coworkers have suggested that association of the C-terminal domain of CaM with a target may disrupt a structurally important hydrogen bond involving the central linker, facilitating formation of a compact binding conformation of Ca2+-CaM [32]. More recently, through rather benign mutations such as E47D, they have determined noninterfacial residues important for molecular recognition through indirect effects – an increase in fluctuations stabilizes the bound state [33]. It was further hypothesized that pH and ionic strength dependent shifts in the populations of conformational substates result from changes in electrostatic interactions in the central linker [13], [27]. For example, the shift in favor of the more compact conformation at reduced pH may result from the loss of electrostatic interactions that serve as spacers at neutral pH. This hypothesis is corroborated by inspection of the proximity of side chains of glutamic and aspartic acid residues surrounding the hinge region in the compact Ca2+-CaM crystal structure [30].
In this manuscript, we report extensive molecular dynamics (MD) simulations of fully solvated, extended and compact Ca2+-CaM under different perturbation scenarios, with focus on E31. We have previously shown that E31 is located in a unique position to manipulate the overall structure; it also has an upshifted pKa into the physiological range and there are several experiments implicating its involvement in signaling coordination between the two lobes (see [14] and references cited therein). Structural perturbations are introduced as either E31A mutation or its protonation. We also perturb environmental factors such as pH and IS. We analyze the structural dynamics through identifiers based on reduced degrees of freedom defined specifically for Ca2+-CaM. Key events leading to or preventing conformational change are discussed. We elaborate on the events occurring along the path sampled between different conformational states identified by MD simulations and we evaluate the effect of charge balance on the conformations. The molecular mechanisms that lead to the observed effects, their relationship to the experimental data, and the consequences of the observations that enhance our understanding of the dynamics and function of Ca2+- CaM are outlined.
CaM consists of 148 amino acids made up of the N-lobe (residues 1–68), the C-lobe (residues 92 to 148) and a linker which is helical in many, but not all, of the reported structures. Each lobe in CaM has two helix-loop-helix (EF-hand motif) calcium binding sites connected by unstructured sequences. Structured elements include helices I (residues 5–17), II (residues 30–39), III (residues 46–54), IV (residues 69–73), V (residues 83–91), VI (residues 101–110), VII (residues 119–129), and VIII (residues 137–144). Ca2+ coordinating residues in each of the four EF-hands are D20-D22-D24-E31 in loop I, D56-D58-N60-E67 in loop II, D93-D95-N97-E104 in loop III, and D129-D131-D133-E140 in loop IV. All MD simulations reported in this work include the four Ca2+ ions.
The existing X-ray structures of Ca2+-loaded, peptide free calmodulin (Ca2+-CaM) are either in an extended or a compact form. There are many examples for the extended form in the protein data bank (PDB) and we utilize that with PDB code 3CLN whereby the coordinates of the first four and the last residue are not reported [29]. The compact form is represented by the 1PRW coded structure [30], and has a bent linker as do many ligand bound Ca+2-CaM conformations present in the PDB. These particular structures have been determined at 2.2 and 1.7 Å resolution, respectively, and were both crystallized at low pH conditions in the range of 5–6, by growth in water-organic mixture compounds. We have previously reported the RMSD comparison for the overall structure as well as the N- and C-lobes of various x-ray structures, including 3CLN, 1PRW and five ligand bound forms [14].
An ensemble of Ca2+-CaM structures have also been reported (PDB code 2K0E) [25] by using experimental NMR order parameters (S2) together with interproton distances derived from nuclear Overhauser effects (NOEs) as restraints in MD simulations using RDC-refined solution structure of Ca2+-CaM. The IS of the experimental setup is 10 mM and the pH is 7 (the conditions in ref. [25] are the same as in ref. [34]; personal communication). The ensemble has 160 structures and reveals that Ca2+-CaM state samples multitude of conformations including, but not limited to, the compact and extended ones. In particular, unlike in the X-ray structures, there also exist compact conformers where the linker is not bent, as we pointed out in our previous study [24].
Throughout this work, an efficient approach to distinguish between the different conformations of CaM proves to be useful: We define two low resolution degrees of freedom projecting the 3N-dimensional conformational space into a visually tractable two-dimensional one. These are the linker end-to-end distance (l) and torsion angle (φ). The former is defined as the distance between the Cα atoms of the two outermost residues of the linker, residues 69 and 91. The latter is the torsion angle defined by four points: the center of mass of the N-lobe (residues 5 to 68), linker beginning and end points (Cα atoms of residues 69 and 91), the center of mass of the C-lobe (residues 92 to 147). These points are schematically shown in Figure 1.
Six sets of simulations were performed with various initial starting conditions. Simulations are summarized in Table 1; IS values reported correspond to the equilibrated box dimensions. For some systems, we have performed independent runs to check the reproducibility of the results. Each condition has at least 150 ns of total sampling time. We prolong the simulation in case there are substantial changes in the relative positioning of the two lobes and/or the length of the linker, measured by the region sampled on the (l, φ) plane described in the previous subsection.
The details of each simulation are as follows; the label for each type of simulation is indicated in parentheses and will be used throughout the text:
() Initial coordinates are taken from the extended 3CLN pdb coded structure and all residues are assigned their standard protonation states to study the conformational dynamics of extended form in solution. 45 Na+ and 30 Cl- ions are added to attain IS = 150 mM at the physiological range. There is a one MD run of 150 ns and an additional control run of 50 ns.
() Initial coordinates are taken from the compact 1PRW and all residues are assigned their physiological protonation states to study the dynamics of compact structure in solution. The system is neutralized by 15 Na+ ions. Due to the smaller box dimensions formed for this more compact structure, this protocol leads to IS = 161 mM at the physiological range. There are two runs of 200 ns each for this system.
() Starting from 3CLN and all residues having the same protonation states as in (), the system is neutralized by 15 Na+ ions. This leads to a low IS of 82 mM. There is a on MD run of 400 ns and an additional control run of 50 ns.
() Starting from 3CLN structure, only E31 is protonated. The system is neutralized by 14 Na+ ions leading to IS = 91 mM; there are two runs for this system, one of length 400 ns and the other of 150 ns.
() Starting from 3CLN structure, E31A mutation is made. The system is neutralized by 14 Na+ ions leading to IS = 94 mM; there are three runs of 400 ns each for this system.
() In all the previously listed simulations, residues are assigned charge states according to pH = 7.4 using pKa values calculated and listed in ref. [24]. These systems are assumed to be at physiological pH. Acidic residues 11, 31, 67, 84, 93, 104, 122, 133 and 140, as well as H109 consistently have pKas shifted from their standard values to ∼5.5. In the system, these are protonated to mimic the low pH conditions. The reader is referred to ref. [24] for details on the calculation of pKa values. There are two runs for this system, one of length 200 ns and the other of 100 ns. The system is neutralized by 5 Na+ ions leading to IS = 43 mM
In addition, control runs of 100 ns duration have been carried out on the extended, low IS proteins, singly or doubly protonating other residues with upshifted pKa values. These are labeled , , and are not separately listed in Table 1; they are neutralized by 14, 13 and 13 Na+ ions, respectively.
We use the NAMD package to model the dynamics of the protein-water systems [35]. The protein is soaked in a water box with at least 10 Å of water from all directions using VMD 1.8.7 program with solvate plug-in version 1.2 [36]. The CharmM22 force field parameters are used for the protein and water molecules [37]. Water molecules are described by the TIP3P model. Each system is neutralized by using VMD autoionize plug-in. Long-range electrostatic interactions are calculated by the particle mesh Ewald sum method, with a cutoff distance of 12 Å and a switching function of 10 Å [38]. RATTLE algorithm is utilized and a step size of 2 fs is used in the Verlet algorithm [39]. Temperature control is carried out by Langevin dynamics with a damping coefficient of 5/ps. Pressure control is attained by a Langevin piston. Volumetric fluctuations are preset to be isotropic. The system is run in the NPT ensemble at 1 atm and 310 K. Equilibration of the pressure is achieved within 2 ns. The equilibrated box dimensions of each system are listed in Table 1. The coordinate sets are saved at 2 ps intervals for further analysis.
We seek the conditions under which the conformations sampled by Ca+2-CaM may be manipulated under different perturbation scenarios. In what follows, we show that a starting conformation may be destabilized by changing the pH of the environment as well as reducing the IS. Moreover, large conformational changes may be delimited or induced by targeting perturbation-sensitive residues selected based-on a coarse-grained analysis of relatively short trajectories.
We display in figure 2, the locations of experimental structures on the reduced conformational space for 3CLN, 1PRW (the extended and compact crystal structures, respectively) and 2K0E (160 NMR solution structures). There are few conformational representatives in solution that are close to the X-ray structures. Also displayed on the right of the figure is the superposition of the linker conformations of the 2K0E ensemble. We find that for linker lengths less than 30 Å, the bending is always in the same direction, and occurs at residues 79–81, although there is no local salt bridge stabilizing it.
Overlaid on these points are the regions sampled by the and MD simulations clearly showing that at physiological pH and IS, the Ca2+-CaM conformations are confined near their initial states in all four runs at 150 mM IS. Thus, these structures are deemed stable around the initial conformations under physiological conditions on the time scales of 100 ns. This is an expected outcome, since conformational jumps between states were measured to occur on the 100 µs time scale from single molecule experiments at physiological pH and IS [27]. Note that the sampled regions overlap with only fourteen of the 160 NMR solution structures that are compatible with experimental measurements for the extended and only one for the compact conformer.
Starting from 3CLN which represents the extended Ca+2-CaM structure captured in most X-ray studies, we externally perturb the physiological conditions for which results were displayed in figure 2: (i) We lower the IS while keeping the pH at 7.4; and (ii) we lower the IS as well as reducing the pH to 5.0 [24]. These systems are labeled and , respectively (Table 1).
The regions sampled by are displayed in figure 3. This is a continuation of the MD simulations from our previous work [24], where the run has now been extended from 200 ns to 400 ns. The RMSD values of the subunits as well as the overall structure are shown in figure 3a. The initial conformer is not stable in solution as confirmed by the protein RMSD change. While the linker and the two lobes each display low intra-domain motions (less than 3±1 Å RMSD), their relative orientations change substantially (up to 13 Å in the overall RMSD). When projected on the reduced degrees of freedom (figure 3b), the trajectories clearly display the three separate sampled states: The two lobes initially point towards each other and within the first 25 ns, the N- and C-lobes complete a ca. 120° torsional motion reaching state II which is then sampled for 195 ns after which a new state is reached (III) by a further torsional motion of 100°. In the last 180 ns of the trajectory, state III is sampled. Snapshots exemplifying these three distinct states are shown in figure 3c. We note that a prompt move into region II also occurs in the supplementary 50 ns run.
All the structures that are sampled throughout the MD trajectory are compared with the experimental ones. The protein spends the first 220 ns in intermediate states with linker length (l>31 Å), and φ = [−80°, 100°]. During the MD simulation, the initial (X-ray derived) structure 3CLN is only transient and the conformations sampled in regions I and II do not overlap with any of those from the NMR ensemble. This is consistent with other NMR and single molecule experiments where very low occupancy is assigned to the fully extended structure [31]. After 220 ns, the system eventually relaxes into a region with l = [28], [34] Å, φ = [−210°, −130°] which overlaps with many of the 2K0E NMR ensemble members (figure 3b). These observations imply that there is an energy barrier between the regions with φ = 90° and φ = 150° so that the system must counter-rotate by a large torsional angle, instead of flipping the 60° directly.
In the runs, we find that by mimicking low pH, low IS environment, the sampled regions in figure 3b do not change (see figure S1), but the sampling is accelerated. We do not go into the details of these runs since we have already published a detailed account of the conformations sampled and key events leading to the conformational change [24]. However, it suffices to say that the same sequence of states I→II→III are followed in both runs. The shift from state I to II occurs at ca. 20 ns similar to the time scales observed in , but that from II to III occurs at ca. 70 ns.
For both the and the systems, the main intra-domain conformational change occurs as a reorientation of helix II in the N-lobe. For example, the nearly right angle between helices I-II, that was always maintained at nearly right angles (80±6°) at physiological IS in the and runs, is now reduced within the first 10–20 ns of the trajectory. It is maintained at a value of 56±7° in and 60±5° in throughout the window of observation. The major event that stabilizes the closed conformation is the formation of salt bridge(s) between the N-lobe and the linker in each case: E7-K77, E11-K77 or E54-K75 in (established in both runs prior to 50 ns) and E7-R74 or E11-K77 in the runs (forming permanently at ca. 40 ns in both samples).
Despite being a counterintuitive observation, it was shown as a general result that two negatively charged nano-sized spheres may be put into close contact by utilizing the competition of hydrophobic and Coulombic interactions, provided that the charges are placed discretely along the surface [40]. At physiological IS and pH, there exists a high energy barrier between the extended and compact structures, corroborated by the 100 µs time scale of jumps between them, measured by single molecule experiments [27]. Thus, a direct passage between the black and blue shaded regions in figure 2 is not observed within the time window of observations of the MD simulations. However, once achieved, the compact conformation is stable despite the net repulsions between the two lobes (net charge on the N- and the C-lobes are -8 and -6, respectively).
One may argue that the pH of the X-ray experiment (5.4) may have contributed to the stability of the 1PRW crystal structure, since the acidic residues 7, 11, 14, 120, 127 are found to be neutral at this pH [41]. The interface between the two lobes involves E7 and E11 on the N-lobe interacting with E127 on the C-lobe, as well as pairing between E14-E120. Thus, it may well be that the acidic contacts do not repel each other in the crystal due to the loss of the charges. In contrast, our MD simulations starting from the 1PRW structure assigns their usual charges to these residues to mimic the physiological pH conditions. Nevertheless, the interface accommodates the repulsions between the closely located negative charges by slightly expanding around the adjacent helices and rotations in the side chains (figure S2). Thus, the initial state is maintained during the 200 ns window of MD observations, regardless of the charge states of the interface residues.
With this robust accommodation of charges in mind, we seek the reasons behind the relaxation of the initial X-ray structure to new conformations when the environmental conditions are perturbed. To be noted is the conformational plasticity in the MD runs at low IS (figure 3b), and the similarity between a subset of the NMR ensemble structures and state III structures. The ionized states of the acidic residues make the electrostatic component dominant and strongly oppose direct inter-domain association on the time scale of the simulations [41]. This fact does not keep the system from sampling a plethora of conformations in the φ space. Thus, to understand how the interfered charge distribution in the environment affects the vicinal solvent layer around the protein, we study the distribution of the solvated non-specific ions around the protein in each case.
We display in figure 4a, the radial distribution function (RDF) of the ions in the solvent around the side chain heavy atoms of the protein at low and high IS. The first peak belongs solely to the contact of Na+ ions with O− atoms of the negatively charged residues. The second peak is due to the solvent mediated interactions. Interestingly, although there are plenty of positively charged residues on the surface of the protein, Cl− ions (which only exist in the physiological IS run) rarely interact with them. At low IS, the Na+ ions strongly interact with the negative charges on the protein, thus screening the extreme repulsions between the two lobes and allowing rotational motions around the linker. To achieve physiological IS, Cl- ions as well as additional Na+ ions are added to the system. In the presence of these additional mobile negative charges the Na+ ions mainly reside away from the protein surface and in bulk water where they may also dynamically interact with Cl− ions (we check that there is no permanent ion pairing occurring between Na+ and Cl− ions). More interestingly, lack of salt in the solvent environment reduces the time scale of conformational change by three orders of magnitude, from sub-milliseconds to sub-microseconds.
Decomposed into the different regions on the protein at low IS (Figure 4a inset), the most significant interaction of the Na+ ions is with the linker residues, followed by those of the C-lobe and even less so with the N-lobe. We have also monitored the trajectories to find that the cations are mobile and they do not have a preferred position near the linker. These ionic distributions are contrary to expectations from the net charges, with that of the linker being only -1, whereas those of the N- and the C-lobes are -8 and -6, respectively. Thus, the ionic interactions are geometry specific, and designate the smooth surface of the linker (relative to the two lobes) as a region that has a tendency of binding non-specific ions. We conclude that the conformational plasticity of the torsional motions observed in altered charge environments is due to the clustering of the cations around the linker which screens the strong repulsions between the two lobes.
We check the effect of the change in the number densities of the ions at different ionic strengths on the values of the RDF peaks. We confirm that the reduction of the peaks exceeds that expected by the 2.3 fold increase in the number densities of the ions in the system (e.g. the linker peak is reduced 3.3 fold.) In terms of the absolute values, the average number of Na+ ions within the first coordination shell of the acidic residues of the linker is 0.5 and 1.13 for the and systems, respectively.
While the removal of Ca2+ ion from EF-hand loop I readily induced compaction of CaM in a previous MD simulation [42], we are interested in revealing its role in CaM dynamics in fully loaded state. We have previously shown that E31located in this loop is unique in that its perturbation in a given direction reproduces the closed form structure with high overlap [14]. In fact, unlike its positional counterparts on the other EF-hand loops of CaM, the role of E31 is not as central in Ca+2 ion coordination as its involvement in signaling coordination between the two lobes. This statement is supported by a series of experimental E→K point mutation studies at the four equivalent EF-hand positions (31, 67, 104 and 140) [43], [44]. Two results are striking: E31K mutation (i) has wild type activation on four different enzymes while the others do not; (ii) does not lead to apparent binding affinity changes while the rest lead to the loss of Ca+2 binding at one site. It was also shown that proton flux is an important factor affecting conformational changes in CaM and its enzyme targets [45]. We have therefore protonated E31 while keeping the IS low in the set of MD simulations.
Since the topology of the residue is the same except for the reduced charge on the side chain, it still interacts with the Ca2+ in the EF hand I. We have monitored this motif throughout the trajectories and ensured there is no loosening in the motif. Strikingly, we find that the net effect of this single point protonation on the sampled conformations is similar to increasing the IS, keeping them near the initial extended structure (compare figures 2b and S3b), with an average l value of 33 Å and torsional angle range φ = [80°,130°]. Despite the protonation of a single point, the RDFs measured in these runs are also more similar to the high IS system (figure 4b), significantly reducing the density of Na+ ion clustering around the protein, mainly affecting the linker region (value reduced to 3.1 from 10.3; see Table 2).
E31 is able to significantly reduce the ion density around the protein and the linker at the same time (Table 2). For example, protonation of D122, the second residue with the most significantly upshifted pKa, in a control run also leads to similar values. E31/D122 double protonation further reduces the ion density around the whole protein and the linker; while the E31/H107 double protonation does not bring in this additional effect. However, protonation of 10 residues to mimic the pH 5 environment in is effective in further reducing the charges around the linker environment, while its effect on the overall protein is less apparent.
The most drastic change in the extended conformation occurs in the system. As we discuss in detail below, the E31A mutation opens a direct path between the extended conformation and compact structures with a bent linker, accessing conformations not sampled by any of the other systems.
runs are characterized by increased mobility of the N-lobe (4 Å RMSD) accompanied by an additional stability in the C-lobe (RMSD<2 Å) as well as the linker. The stability in the latter two regions, not directly perturbed by the E31A mutation, contrasts the simulations discussed in the previous subsection. We emphasize that the Ca+2 ion coordination is never lost in any part of these MD simulations which total 0.55 µs and 1.2 µs in and , respectively. By inspecting the MD trajectories, we find that the main direct difference between and runs is that while the calcium binding motif is not disrupted in the former, residue 31 can no longer participate in the motif in the latter due to its short side chain and hydrophobic character. Interestingly, the E31A mutation restores some of the depleted charge distribution around the acidic residues that occurred upon its protonation (Figure 4b and Table 2).
The reorientation that takes place in the N-lobe is quantified by an increase of the RMSD value from 2 Å to 4 Å (Figure 5a). The angle between helices III–IV displays a drastic change, with helix III tilting towards helix IV. This is followed by the formation of a salt bridge between residues E47 and R86 at 60 ns which may be traced in the sharp decrease in l from 34 Å to 27 Å (Figure 5b). After the salt bridge formation, at 80 ns, the linker is further bent from residue 81 and l drops to 25 Å bearing a compact conformation. Snapshots are taken before and after transition and shown in Figure 5c. The observed conformational change is reversible, and the extended structure is restored at ca. 160 ns. No significant pKa shift appears for charged residues in any part of the trajectory.
We find that the transition state is well defined, occurring through the same point in both forward and reverse steps. Time intervals of the transitions between the extended and compact (forward transition) and between compact and extended conformations (reverse transition) are examined in more detail in figure 6. The positions of structures near the transition state in 200 ps intervals are plotted on the (l, φ) plane. Note that the axes have been zoomed in.
The transition between the extended and compact states is also examined via the tool Geometrical Pathways [46], [47]. This tool utilizes geometric targeting (GT) method that has recently been introduced [46] as a rapid way to generate all-atom pathways from one protein structure to some known target structure. GT is based on the philosophy that essential features of protein conformational changes can be captured by solely considering geometric relationships between atoms. The protein is modeled as a geometric system, with constraints established to enforce various aspects of structure quality such as preserving covalent bond geometry, preventing overlap of atoms, avoiding forbidden Ramachandran regions for backbone dihedral angles, avoiding eclipsed side-chain torsional angles, and maintaining hydrogen bonds and hydrophobic contacts. We note GT cannot predict relative timing of events.
Using Geometrical Pathways in Biomolecules server [46], we have generated 10 random pathways between representative structures collected structures at 50 ns (extended) and at 80 ns (compact) of the run 1. The RMSD step size is 0.05 Å. The structures generated in the forward pathway by Geometrical Pathways are also plotted on figure 6 with the median of the pathway and the standard error bars along both axes. The random pathways produced via Geometrical Pathways overlaps with that visited by MD. They are widely distributed along the interdomain torsional angle dimension, but have narrow distribution in end-to-end-linker distance. GT generated pathways take energetics into account indirectly, through geometric factors only. Their overlap with the MD pathway corroborates that the conformational change may be achieved as a series of geometrically viable sequential steps, if the energy barrier between the two states allows them to take place.
In fact, a stabilized conformational change between states I and II is observed in only one of the three runs. However, several attempted jumps occur with a kinked linker conformation in all simulations (l<30 Å). Thus, the crucial step stabilizing the bent conformation is not the bending that is facilitated after the mutation, but the formation of the salt bridge between the N-lobe and the linker. Such attempts occur neither in nor in physiological IS runs.
The relative positioning of the entering and exiting helices of the EF-hand motifs have been used to characterize the diverse conformations utilized by CaM for target recognition [48]. We therefore focus on the dynamics of the angles between the helices in EF hand motifs to understand how this “rare” conformational change is facilitated by E31A mutation. On the N-lobe, helices I and II in EF hand I, and helices III and IV in EF hand II are both initially posed at about right angles to each other, having values between 80–90°. In all three trajectories, the increase in the RMSD of the N-lobe occurs simultaneously with the loss of this perpendicular arrangement between the former pair of helices. The change in the orientation of the N-lobe helices provides E47 to make the salt bridge with R86. Perhaps most intriguing is the diminished fluctuations between helices V and VII upon E31A mutation. The values of 38±11° in the rest of the simulations of Table 1 are suppressed to 30±5° for all three .
The model for inter-lobe communication of CaM is E31 mediated. E31A mutation induces conformational changes between the two EF-hand motifs in the N-lobe, which simultaneously stabilize helix reorientational fluctuations in the C-lobe. No apparent communication pathway between the two lobes is found, thus lending support to the ensemble view of allostery [49]. As a result, the N-lobe performs an intra-lobe conformational search to establish a salt bridge with the linker. The stabilized C-lobe helps maintain the initial contact. The solvent participates in the fluctuations that establish the background for the conformations sampled by the protein.
Multiple conformations and cooperative conformational changes are an essential part of many enzyme mechanisms [50]. We explore the role of electrostatics in altering the conformation distributions as well as the dynamics of Ca2+-CaM using extensive MD simulations under physiological and low IS/pH conditions, and by mutating/protonating single residues. While the protein is stable in the initial state at physiological IS, lowering the IS or pH leads to conformational switching to more compact structures on sub-100 ns time scales (Figures 3 and S1).
Although the net charges on the N- and C-lobes are significantly higher, at low IS cations approach the linker due to its relatively smooth geometry [51], screening the repulsions between the lobes. This leads to conformational plasticity, enabling large torsional motions around the linker, eventually causing a compact conformation within 200 ns, albeit with a stiff linker. Lowering the pH in addition to the IS, which in effect deletes 10 charges on the protein surface, contributes to the process further, letting the protein achieve similar conformations within 60 ns.
At high IS, the Cl− counterions are repelled by the significantly negatively charged protein (figure 4a). Na+ ions are then driven into the bulk solvent, since the counterion interactions are favored over those with acidic residues both energetically and entropically. Repulsion between the two lobes is more pronounced because of ion depletion in the intervening region, restricting the protein conformation near the initial structure (figure 2). It would be interesting to make a systematic study of the effect of ions on the dynamics of CaM, including different ion types and concentrations. However, our scope here is to merely demonstrate that CaM dynamics is sensitive to ionic strength and may be moderated between rigid and very flexible. Such an observation has implications on both tuning and interpreting experimental results and on conditions selected for computer simulations.
Perhaps more interestingly, neutralizing the single residue E31 at a key location by protonation has a similar effect on the ion distributions to increasing the IS (figure 4b), confining the sampled conformations near the initial structure (SI figure S3). Alanine mutation of the same residue results in an intermediate distribution of ions, leading to partial mobilization of the protein. This facilitates a bending in the linker near the extended conformation (figure 5). E31A mutation accentuates the existing allosteric interactions, by introducing a change originating on the N-lobe whose action is detected on the C-lobe via the rearrangements of the helices in the EF-hands. The coupling between the two lobes is detected simultaneously, and the stabilizing salt bridge between the N-lobe and the linker is established later on in the simulation, right before the transition state is reached (figures 5c and 6). Thus, no pathway of structural distortions between the allosteric sites is observed, lending support to the ensemble view of allostery [49].
Finally, we note that the sampled structures have representatives in the NMR ensemble of conformations [25]. 40% of the 160 conformers have been visited in the cumulative MD simulations. On the downside, these are predominantly the ones with the straight linker, while the compact conformers (those with l<25 Å) have been rarely observed. The structures are compatible with the FRET results both in terms of the interlobe distances and the fact that lowering the pH to 5.0 leads to a single stable state as opposed to the presence of at least two distinct forms at pH 7.4 [13]. Our observations are also compatible with the findings of Bertini and coworkers [31]. Therein, low occupancies are assigned to fully compact structure of Ca2+-CaM (1PRW). Extended conformers similar to 3CLN in general have low occupancies although some other extended conformers have occupancies as high as 35%. This supports our findings that in our cumulative MD trajectories, the initial structures whether starting from compact or extended crystalline structures of Ca2+-CaM relaxes to less extended or less compact forms.
Our current view suggests that at physiological IS and pH, there exists a barrier between the extended and compact forms, leading to 100 µs time scales for conformational jumps [27]. Barrier crossing is prevented by the repulsive electrostatic interactions between the two negatively charged lobes. As we show in this work, the crossing may be facilitated by tuning environmental conditions or by perturbing single residues located at key positions. The current study contributes to the knowledge-base in the direction of methods that determine how proteins adapt to changes in their environment or structure.
|
10.1371/journal.pcbi.1002899 | Significantly Improved HIV Inhibitor Efficacy Prediction Employing Proteochemometric Models Generated From Antivirogram Data | Infection with HIV cannot currently be cured; however it can be controlled by combination treatment with multiple anti-retroviral drugs. Given different viral genotypes for virtually each individual patient, the question now arises which drug combination to use to achieve effective treatment. With the availability of viral genotypic data and clinical phenotypic data, it has become possible to create computational models able to predict an optimal treatment regimen for an individual patient. Current models are based only on sequence data derived from viral genotyping; chemical similarity of drugs is not considered. To explore the added value of chemical similarity inclusion we applied proteochemometric models, combining chemical and protein target properties in a single bioactivity model. Our dataset was a large scale clinical database of genotypic and phenotypic information (in total ca. 300,000 drug-mutant bioactivity data points, 4 (NNRTI), 8 (NRTI) or 9 (PI) drugs, and 10,700 (NNRTI) 10,500 (NRTI) or 27,000 (PI) mutants). Our models achieved a prediction error below 0.5 Log Fold Change. Moreover, when directly compared with previously published sequence data, derived models PCM performed better in resistance classification and prediction of Log Fold Change (0.76 log units versus 0.91). Furthermore, we were able to successfully confirm both known and identify previously unpublished, resistance-conferring mutations of HIV Reverse Transcriptase (e.g. K102Y, T216M) and HIV Protease (e.g. Q18N, N88G) from our dataset. Finally, we applied our models prospectively to the public HIV resistance database from Stanford University obtaining a correct resistance prediction rate of 84% on the full set (compared to 80% in previous work on a high quality subset). We conclude that proteochemometric models are able to accurately predict the phenotypic resistance based on genotypic data even for novel mutants and mixtures. Furthermore, we add an applicability domain to the prediction, informing the user about the reliability of predictions.
| Infection with the human immunodeficiency virus (HIV) currently cannot be cured. It can however be contained through treatment with a combination of several anti-viral drugs. Yet, during treatment resistance can occur which leads to drugs becoming ineffective. Through a combination of drugs, this resistance can be deferred indefinitely. The optimal combination of drugs depends on the specific strain of HIV with which the patient is infected. Previously, methods have been developed that predict a personalized treatment regimen based on the genetic sequence (genotype) of the virus via the use of computer modeling, corner stone of the methods is drug affinity prediction. Here we have applied proteochemometric modeling which takes this genetic information into account, but also includes chemical description of the drugs that are now clinically available. We show that this combined technique performs better than models that only include genetic information. Our approach leads to personalized treatment predictions with a higher reliability compared to the current state of the art. In addition, we include a reliability measure which allows each prediction to be assessed for reliability. Finally we describe mutations of the HIV genome that were not previously described in literature and lead to resistance to treatment.
| The Human Immunodeficiency Virus (HIV) was discovered and isolated as the cause of ‘Acquired Immuno Deficiency Syndrome’ (AIDS) in 1983. [1], [2] Over the following three decades HIV has turned into a global epidemic, the number of people living with HIV in 2010 being estimated at 34 million according to the World Health Organization. [3] Furthermore the number of people newly infected was approximately 2.7 million and 1.8 million HIV related deaths were reported, [3] hence illustrating that HIV represents one of the major illnesses of mankind today.
Infection with HIV can be contained, however not cured, by Highly Active Anti-Retroviral Therapy (HAART), which relies on a combination of three or more inhibitors from different drug classes. [4], [5] Currently more than 20 HIV inhibiting dugs are approved, [6] with the largest classes of drugs being formed by Protease Inhibitors (PIs), Nucleoside/Nucleotide Reverse Transcriptase Inhibitors (NRTIs) and Non-Nucleoside Reverse Transcriptase Inhibitors (NNRTIs). However, while a large number of drugs is accessible to the physician (thus rendering HIV in some sense a disease that is currently ‘under control’ regarding the treatment options available), the question of which drugs to use for which patient is an exercise where more guidance would also in the current situation be of tremendous practical relevance.
The process of replication by HIV is extremely error prone and therefore mutations in the viral genome occur frequently. [7], [8] It is these mutations that can be the basis for HIV resistance against therapy, [6] even single point mutations can cause insensitivity of HIV to treatment with all members from an entire drug class (e.g. K101P in the case of NNRTIs). [6], [9] Occurrence of these resistance conferring mutations can be contained or minimized by the nature of HAART therapy due to the combination of multiple drugs classes. [5] However, the occurrence of high impact mutations can cause treatment failure in HAART for certain specific drug regimens. It is therefore crucial that the drug regimen is tailored to the specific viral genotype. [10], [11]
What is required for a tailored drug regimen is knowledge of the effect of individual mutations on the efficacy of different drugs. A rough distinction can be made between assay based methods and computational methods, with assay based methods being available since the year 1998. [12], [13], [14] Conversely, various computational methods have become available over the last decade. [15], [16], [17], [18], [19], [20] Personalized prediction has been shown to perform equal to standard of care in treatment naïve patients but significantly (P = 0.02) better in patients experiencing drug failure. [17] Furthermore, computational approaches have been shown to perform equal to phenotypic assays. [21] Several methods that have been published previously, both assay-based and computational approaches, will be outlined briefly in the following.
Phenotypic assays measure the replication of HIV in vitro subsequent to genotype determination. Three common different phenotypic assays include: Antivirogram (AVG) by Virco (1998), [12] an assay by Walter et al. by the Universities of Erlangen-Nürnberg and Leuven (1999), [14] and Phenosense by Monogram Biosciences (2000). [13] Diverse readouts are employed in these assays: spectrophotometrical determination of diphenyltetrazolium bromide reduction (AVG), luminescence produced by secreted alkaline phosphatase (Walter et al.), [14] and luminescence by luciferase produced in the cell upon completion of one round of virus replication (Phenosense). All readouts respond in a dose dependent manner. Antiretroviral drug susceptibility is expressed as the base 10 logarithm of a numerical fold change (Log FC). Log FC is determined by dividing the IC50 for inhibition of the mutated virus by the IC50 for inhibition of a determined wild type virus (wt). Hence, a Log FC value of 1 for a given drug – mutant pair means that the drug IC50 value for that particular mutant is 10 times that of the IC50 value for the same drug on wt. Likewise, a Log FC value of 3 for a given drug – mutant pair represents an IC50 value 1,000 times higher. The sequences that are defined to be wt are the HXB2 strain (Uniprot accession P04585) for AVG, [22], [23] or a recombinant pNL4-3 strain (Genbank entry M19921) for Walter et al. and PhenoSense. [24]
From the data generated by the phenotypic assays, computational models have been produced that predict a virtual phenotype from a given genotype. Based on the large amount of Log FC data generated by AVG, Virco introduced their first computational prediction tool, Virtual Phenotype in 2000 superseded by VircoTYPE HIV-1 in 2004. [25] VircoTYPE creates linear regression models based on the presence of mutations and pairs of mutations. Each mutation and mutation pair is given a weight factor in model training based on measured data (6,000 to 40,000 samples per drug). The sum of all weight factors for relevant mutations present in a mutant combined with the wild type weight factor then provides the predicted log FC. In a randomized clinical trial, VircoTYPE HIV-1 has been shown to perform slightly better than conventional phenotypic assays in decreasing HIV RNA concentration over a follow up period of 48 weeks (39% of the phenotypic assay group reached HIV RNA below 400 copies/ml compared to 51% of the VircoTYPE HIV-1 group). [21]
Next to VircoTYPE HIV-1, another implementation of a virtual phenotype has been developed at the Max Planck Institute, called Geno2Pheno. [20] This tool has been trained on smaller dataset compared to VircoTYPE. However, it has been retrospectively validated on the Stanford HIV Drug Resistance Database (Stanford Set) in 2009. [19] In this study Geno2Pheno outperformed state-of-the-art-expert based systems by finding 16.2–19.8% more successful regimens.
Nevertheless, what the computational methods described here have in common is that they are solely trained on the mutation patterns and the effect these patterns have on a single drug. [26], [27], [28] Therefore a separate model is created for every drug. Similarity between individual amino acids is not considered (how similar are two amino acids to each other and hence how big is the impact of a mutation). Furthermore, the chemical similarity between compounds is not considered in the models. Both types of similarity information have the potential to lead to better models and prompted us to apply ‘proteochemometric models’, described in the following, to improve upon the current situation.
Given that previous models did not take into account chemical information, the individual models mentioned above fail to acknowledge the chemical similarity between drugs that belong to a single class, thereby discarding very valuable information. This is the case because molecular similarity has been shown to have great predictive power when it comes to identifying which kind of related structures could also show activity against a given target. [29] Hence it is likely that also for established drugs, chemical similarity can improve models by explicitly taking the concept of drug – target interaction into account, which is then combined with mutational information of the drug target itself. This technique is called proteochemometric (PCM) modeling. The concept of PCM as we applied it has been summarized in Figure 1. This flow chart shows how we combine both mutant data and drug data and link it to a Log FC value. The authors have previously reviewed the technique and it has already been successfully applied to NNRTI inhibitors of HIV Reverse Transcriptase before. [30], [31], [32], [33]
Yet, the most important difference between this previous work and the current study is the scale of the mutant database used to train the models on. Previous work focused on a total of 4,792 data points, [30] 386 data points, [34] 654 data points, [31] 4,495 data points, [35] or 4,024 data points, [33] whereas here a total of 288,138 data points are used. Hence, we expect a more generally applicable model resulting from the current study. Furthermore, previous work included a larger number of compounds (451 compounds) on the chemical side, and their biological activity on a total of 14 mutants. Therefore, these models described a relatively large chemical space compared to the target space, while in the current work we have reversed this situation and the models now describe a relatively large target space compared (approximately 37,000 mutants) to the chemical space (21 drugs). In addition, what is lacking in previously published PCM approaches is the power to extrapolate, thereby able to also produce a reliable prediction for novel (unknown) mutants while including a reliability measure for these predictions. These are the points we are addressing in the current work.
In the current project it is our hypothesis that we can train a single PCM model for each of the following major HIV drug classes using the AVG data: Protease Inhibitors (PIs), Nucleoside/Nucleotide Reverse Transcriptase Inhibitors (NRTIs) and Non-Nucleoside Reverse Transcriptase Inhibitors (NNRTIs). As no PCM model has ever been trained on such a large dataset (the current dataset is 60 times larger than the largest published HIV PCM model) our hypothesis was on the one hand to arrive at better model performance, and on the other hand to unravel more reliable rules such as the influence of point mutations on compound activity. Scientifically interesting is also the reversal in the ratio between chemical space and target space in the model training set described above.
Given the wealth of training data present, the resulting bioactivity models can be used to predict the activity of clinical ARV drugs on mutants not present (untested) in the dataset (corresponding to a patient with a new, previously unseen genotype that needs to be treated in the clinic). For this purpose, an additional 7,798 data points have been used as a prospective validation set, in order to gauge predictive performance of the model in a real-world situation. These data points have been retrieved from the Stanford University database after model training and validation was completed.
We first validated by creating a learning curve for each drug class. Learning curves plot the quality of models that are created on an increasing fraction of the data. Concurrently these models are validated on the remainder (and hence decreasing) part of the dataset (supporting Figure S1). When given enough measures of model reliability in respect to the training set size, an estimate can be made of the optimal performance possible on said dataset.
We found that all models should reach a root mean square error (RMSE) <0.5 units Log FC (see Methods section and supporting Figure S1), which was subsequently confirmed in the external validation which was performed per drug rather than per drug class below (See supporting Table S1 for the used abbreviations for each drug).
The second step was the generation of models on 70% of the dataset as the learning curves showed this to be the optimal split size to get a reliable performance estimate for these models. While these 70% models give an estimate of the ability of the models to perform future predictions successfully, other additional forms of validation should also be included as we will show later on. [36] The RMSE for sequences that were present in the training set, however not in combination with the same drug, was 0.27 (PIs, Figure 2C), 0.31 (NRTIs, Figure 2B) and 0.45 (NNRTIs, Figure 2A), with an R02 0.89 (PIs, Figure 2C), 0.79 (NNRTIs, Figure 2A), and 0.75 (NRTIs, Figure 2B). Hence, we found that PCM was overall able to extrapolate the Log FC values for individual pairs of drug and mutant not encountered in the training set with a reliability that slightly better than the assay reliability of the current dataset (approximately 0.5 log units). Hence, PCM is on this dataset able to extrapolate to novel drug-mutant pairs when the drug and mutant in question are only present in the training set individually, and not in the combination, as shown in the test set (internal validation).
For sequences not present in the training set (representing predictions for previously unseen patients, or genotypes) the RMSE obtained by the model was 0.43 (PIs, Figure 2F), 0.49 (NNRTIs, Figure 2D) and 0.52 (NRTIs, Figure 2E) with an R02 of 0.74 (NNRTIs, Figure 2D), 0.71 (PIs, Figure 2F) and 0.33 (NRTIs, Figure 2E), respectively. Hence, PCM is on the current dataset also able to extrapolate the Log FC values for individual pairs of drug and mutant not encountered in the training set with reliability comparable to assay reliability when the mutant in question is not present in the training set (External validation, for validation plots per individual drug please see Figures S2, S3, S4).
We also pursued two alternative approaches to model the current dataset. These steps provide an estimate of the added value of PCM itself rather than size of the dataset. Our first validation was calculating the average of the Log FC values of mutants for which we had multiple drug Log FC measurements. Subsequently this average Log FC value was used as a predictor for the drugs that were left out. This method was called Log FC scaling and is similar to a benchmark we used in previous work. [33] By comparing the average Log FC value to the value measured on the sequence left out, the chemical descriptor component of PCM is removed. Moreover the main contributors to changes in Log FC in this method are those causing cross resistance as the effects on individual Log FC values are ignored. We aimed to leave out 30% when multiple measurements were possible, when only three or two measurements were available we left out one drug measurement. The results are included as Supporting table S2, where PCM on average has a 50% lower RMSE (on average 0.38 log units for PCM versus 0.56 log units for scaling).
Likewise we trained individual models per drug using only the sequence descriptors, hence this approach is conceptually identical to Virtual Phenotype or Geno2Pheno models (shown as ‘sequence only’ in supporting Table S2). The goal here was to ensure that including chemical (ligand) information indeed improves model performance. Indeed, we found that also here PCM outperforms sequence only models in all drug classes. In all cases the prediction error improves by approximately 11% (with a similar improvement of correlation coefficient). This improvement is significant for the NRTIs when performing a paired t-test (RMSE, P<0.01; R02, P<0.01) and PIs (RMSE, P<0.01; R02, P<0.05). The difference was not significant for the NNRTIs, while PCM did outperform sequence only models (RMSE, P = 0.33; R02, P = 0.14). We think this is mainly due to the large chemical diversity of the NNRTI drug class, which are similar in pharmacophoric properties but display a diverse collection of scaffolds. Since we use two dimensional chemical descriptors rather than three dimensional, PCM cannot reach the large performance difference shown for PI and NRTI. This is supported by the fact that the chemically most different NNRTI, ETR, is the only one that has a lower performance in PCM models (similarity on average 0.36, Supporting Table S7). Yet, the combination of the bioactivity space for individual NNRTIs is successful as NNRTIs are known to be sensitive to cross resistance, this is captured by PCM.
In order to investigate clinical relevance of our work, we next incorporated the actual clinical cut-off (CCO) values. These values describe the expected response of a patient to treatment with a certain drug based on the HIV genotype (the used clinical CCO values are given in supporting Table S10 and S11). When we apply the CCOs to our model predictions, our models achieve an overall correctly classified percentage (CCP) of 95% for the inhibition of mutant sequences present by a drug not present for that sequence in the dataset (Figure 2).
For the sequences not present in the training set, 91% was predicted correctly (supporting Table S3 and S4). More specifically per class, the PI scored the best (93% correct for internal validation and 90% correct for external validation), followed by NNRTIs (93% correct for internal validation and 91% correct for external validation), and lastly the NRTIs (80% correct for internal validation and 68% for external validation). However, it should be noted that for the NRTIs only a small number of sequences was available as validation, and all were not very resistant, possibly leading to a biased validation.
We can conclude that even prediction on sequences not present in the training set was possible, albeit slightly less than the internal validation (RMSE 0.34 log units when the sequence is known versus 0.48 when it is not). To further find the limitations of this extrapolation we employed leave-one-sequence-out (LOSO) validation.
LOSO validation is unique to proteochemometric approaches, since it enables the prediction of compound activities for entirely novel genotypes (or patients), hence estimating which treatment would be most likely to succeed in a given treatment situation. For computational reasons, our approach used a subset of approximately 1,000 mutants from the full set (4% (PR) and 9% (RT) of the total dataset, respectively). Each of these sequences was left out, and a model was trained on the remaining sequences; results are shown in Figure 3. Again, the PCM technique overall provides rather robust in modeling the current dataset. Best performance can be observed for the PI model (with an RMSE of 0.40 log units, R02 of 0.76 and CCP 90%), followed by the NNRTIs (RMSE of 0.67 log units, R02 of 0.53 and CCP 84%) and the NRTIs (RMSE of 0.45 log units, R02 of 0.50 and CCP 71%). The finding that PIs and NRTIs are easier to model than NNRTIs is in line with our finding above. What should be noted is that the NNRTI model tends to slightly underpredict the Log FC values that have been measured with a Log FC above 3.0. While those values are correctly predicted to be above 1.0 (which is an important prediction to have by itself in practice), the numerical correlation between predicted and experimental values leads to a slight, but consistent under prediction of activities in this value range.
Crucial for the application of computational models is an estimate in which cases the model can be trusted, and where it is likely to fail. In this spirit, the ‘Applicability Domain’ of computational models has become an important topic recently;[37] however, so far it was mainly applied to the chemical domain. This concept was extended in the current work, given the nature of PCM models, also to the protein target or biological domain where special considerations need to be taken into account. Since we are dealing with a large set of viral mutants we are unable to define a single similarity to a WT to get an idea of the applicability domain. Therefore, we chose to define the applicability domain based not only on the distance to the training set, but also on the density of neighbors in the training set (See Methods section for details). At a similarity threshold of 97% each sequence is hence assigned a density score between 0 and 1 (0 corresponding to no sequences with a similarity of at least 97%, and 1 corresponding to all sequences in the dataset having more than 97% similarity to the sequence under consideration).
Figure 3 visualizes the ‘Neighborhood Behavior’;[38] if the fraction of sequences having this similarity of 97% (X-axis) is larger (closer to 1), the maximal encountered prediction error (RMSE, y-axis) is lower (closer to 0 log units). This means that if the model can extrapolate from a larger number of sequences having a similarity of 97% or higher, the predictions become more reliable. Performance of a practically useful model would require the largest error to be below 1 log unit; hence, given this requirement, the density of sequences in the training set should be larger than 0.15 (for PIs and NRTIs) and larger than 0.25 (for NNRTIs), respectively. Due to this numerical quantification of the ‘Applicability Domain’ of the model, in biological space, we are now able to judge in which situations the model will be applicable (i.e. is likely to generate reliable results), and in which situations it is not which is of crucial importance in order to gain trust into computational models.
Further exploring the clinical relevance of this work, the CCO's were again applied to model predictions also in the case of the LOSO experiments. Overall the model reached a CCP of 81% of the individual mutant – drug pairs. Moreover, 12% of the total predictions were overpredicted, and only 7% underpredicted. Hence our models perform robust also on sequences that are entirely novel to the model (supporting Table S5). For the individual classes, the image is very similar to that in the external validation, the PIs perform the best (90% correct), followed by the NNRTIs (84% correct) and lastly the NRTIs (71% correct).
In the text above we have thoroughly validated our models and they have shown to be robust in modeling HIV resistance to PIs, NNRTIs and NRTIs. This was confirmed for known sequences in an unknown combination with a drug but also for unknown sequences in an unknown combination with a drug. Hence we conclude that our models describe the drug – target interaction space, therefore it is very interesting to investigate how our models actually derive these Log FC values from the contributions individual mutations make.
To compare the performance of our PCM models with state of the art models trained on sequence data only and to place the results of our work in perspective, we used a dataset previously published by Van der Borght et al. [39] We explicitly selected for each class the 150 sequences that were predicted most inaccurate, representing the most difficult sequences to predict (these were mutants that seem to exhibit a different resistance profile). Moreover, most of these sequences contained mixtures (several mutations present on a single position) that had been discarded from our PCM training set. The purpose of this validation was therefore twofold, to assess the performance of PCM when compared to sequence only models, and secondly to assess if the PCM models can deconvolute the effect of individual mutations to make accurate predictions for mixture sequences.
The results of this validation are shown in Figure 4, Table 1 and Supporting Table S2. Our PCM models clearly outperform sequence only models. For each class the PCM models predict the Log FC more accurately. This is indicated by the smaller RMSE (0.53 log units versus 0.68 log units for the NRTIs; 0.65 log units versus 0.75 log units for the PIs, and 0.85 log units versus 1.3 log units for the NNRTIs) and also by a higher CCP (68% versus 54% for the NRTIs, 78% versus 75% for the PIs, and 89% versus 78% for the NNRTIs). For several PIs, the sequence only models perform marginally better when measuring by the correlation coefficient; however as these values are systematically slightly overpredicted in the sequence only models, PCM still performs better.
Furthermore, when we limit ourselves to only predicting the Log FC for mutant mixtures, PCM still outperforms sequence only models (supporting Table S2, supporting Table S6 and supporting Figure S5). This is even the case while our PCM models were trained without mixture sequences in the training set whereas these were present in the training for the sequence only models. A large fraction of these mixtures sequences show a low value for the 97% similarity density, hence we would expect the models to perform suboptimal on these sequences. The applicability domain measure therefore also holds in this case. These results underline the added value of PCM models over sequence only models and hence we also wanted to interpret these models.
The aim of this feature importance investigation was to explain the average reduction in drug affinity that the presence of an individual mutation causes. Firstly, we investigated the effect of several known mutations from literature. To this end we compared the features selected as being significant by our model to the mutational overviews published by Johnson et.al. [6], [40]
Figure 5 shows the impact of selected mutations on NNRTI affinity. Overall, while there is a significant amount of cross-resistance, each of the NNRTIs still possesses its own distinct resistance profile, in agreement with the importance of personalized HIV treatment approaches. Furthermore, the impact of individual mutations varies per drug and is in line with literature data. [6], [18] Red indicates that the presence of this mutation leads to a higher Log FC on average, whereas green indicated that the presence of this mutations leads to a lower Log FC on average, and white indicates that this mutation has little effect on the Log FC (For an explanation of the abbreviations see supporting Table S1). For instance, mutation K103N has a rather specific pattern as it confers resistance to Nevirapine, Efavirenz, and Delavirdine but not to Etravirine. [6], [18] This pattern is reproduced by our model. Furthermore, V179F is known to lead to Etravirine resistance but to have less effect on Nevirapine, Efavirenz, and Delavirdine, [6], [18] a resistance profile that can also be reproduced based on our dataset. Some mutations are slightly underestimated, these include V90I and V106I. Another interesting observation is that mutations Y188C and G190A are predicted to render HIV more sensitive to Etravirine according to our model. This finding is in agreement with work by Vingerhoets et al. [41]
Related analyses for NRTI resistance and PI resistance have been included in the supporting information (supporting Figure S6 and supporting Figure S7). Specific NRTI mutations that were accurately reproduced include K65R, Q151M, and T215Y, while mutations M41L and M184V are slightly underestimated, compared to previous studies. [6] For the PIs mutations that are accurately reproduced include D30N, I50L, V82S, and I84, while the I64L and I93M mutations are assigned less importance than in previous work. [6]
Hence, the PCM models applied in this study are able to reproduce known resistance patterns as outlined above. This led us to the next step of the study, the identification of novel mutations (present in our dataset but not previously published) which are found to confer cross resistance to antiretroviral treatments. This work is similar to previous work by Van der Borght et al. [39] but here we focus on both cross resistance conferring mutations and drug specific mutations. Furthermore we apply the method to all three major classes of anti-HIV drugs rather than one and can do so directly from our models.
To identify cross-resistance as part of the current study, we were limiting ourselves to mutations that have a negative effect on the majority of drugs in a single class. However, in case of particular interest in the resistance profile of a particular drug this analysis can also be performed on the individual-drug level subsequently.
We selected mutants based on the following conditions: occurrence in the dataset more than once; average Log FC for all compounds above 0.4; standard deviation over this average below 0.4. Known mutations as published in literature were discarded. [6], [40], [42], [43] With these filters a number of novel resistance conferring mutations could successfully be identified which are listed in Table 2–4 (For an explanation of the abbreviations see supporting Table S1). Mutations identified have a high impact on drug affinity and which lend themselves to experimental validation, for instance in the case of NNRTI and NRTI resistance conferring mutation T216M. The full set of individual mutations (both known and novel) and their effect is included in the supporting Information as delimited text files (Dataset S1).
We furthermore analyzed not only mutations that cause cross-resistance, but also those with a particular effect on a specific drug treatment alone. The goal here was to identify mutations that lead to large resistance for one drug but are still sensitive for another drug from the same class. Hence this knowledge can be of high importance in a clinical setting. For the PIs the 30 most interesting mutations (defined as those mutations that have the most diverse effect on the different drugs), are shown in Figure 6 (while corresponding figures for the NNRTIs and NRTIs are included in supporting Figure S8 and supporting Figure S9). In those figures we can observe several mutations that lead to resistance for a single drug (Log FC on average >0.5) and at the same time lead to higher sensitivity for another drug (Log FC on average <0.0).
For instance, the G48W mutant is sensitive to Darunavir and Tipranavir, while showing some degree of resistance to all other PIs. Furthermore, R87K is resistant to Atazanavir, Darunavir, and Tipranavir, but sensitive to all other drugs in the dataset. This could indicate that at this point the mutant has over-adapted to the host environment, including the drug, hence rendering the mutant very sensitive to changes in this environment. Finally, N88G seems to only be sensitive to Darunavir, while conferring resistance to all other PIs in the dataset. Information of this type is of high relevance to prescribe the optimal drug for an individual patient, by being able to link the viral genotype to the clinical phenotype in a real-world situation. Applying these models in a real world situation on unseen clinical data is exactly what we implemented in the following paragraphs.
Given a sequence of PR and RT (and hence, a viral genotype of a patient to be treated), our models are able to predict which drugs will be least influenced by resistance, as measured via the lowest Log FC. To accurately estimate our model performance in prospective predictions, in the final step of this study we performed a validation on unseen data. Apart from only focusing on unseen data, in order to establish agreement of our modeling procedure with other approaches, we also employed data from an entirely different source – namely, for sequences obtained from the Stanford University HIV Drug Resistance Database (Stanford Set). [18], [44] The set we used has also been included as supporting information (Dataset S2).
Applied to the Stanford Set, the PCM models developed in the current work show an average RMSE of 0.52 log units, with the average R02 being 0.59. Compared to the models above, this is a slightly larger error compared to the validation on Virco data, which was below 0.50 log units. (It should be noted that this is very diverse data, including historical literature data of which we cannot estimate reliability.) The PI model again performs the best (with an RMSE of 0.43 log units and an R02 of 0.76), while the NNRTIs are predicted with the largest error (with an RMSE of 0.62 log units and an R02 of 0.66), which is the result of a number of outliers (see Figure 7 and explanation below). The NRTI model exhibits the lowest correlation coefficient (R02 0.39 and RMSE 0.61 log units), mostly due to the relatively small range of Log FCs present in the dataset. However, also in this case we observe a correlation between the density of sequences with a 97% similarity in the training set and modeling error, also allowing us to establish the Applicability Domain of the model throughout.
With the NRTIs and some NNRTIs there are outliers to the Applicability Domain we established, meaning that expected and observed errors exhibited some differences. (Note that this is usually the case, the Applicability Domain concept being a concept based on error distributions and likelihoods, not certainties, that a given error will be obtained in a given situation.) It was found that these outliers were obtained from only a small number of references (RefIDs) from the Stanford DB. References 369, 414, 649 all contained the M184V and T215Y mutations that are also known to differ between AVG and Phenosense. Furthermore there was a major discrepancy between the Log FC values reported for AZT on similar mutant which was >2000 (log value 3.3) in one reference, while being reported as low as 28 (log value 1.4) from another source. [45], [46], [47] Reference 789 contained a sequence carrying a deletion at position 69, which was not taken into account by our model. [48] Reference 947 linked to unpublished data and could therefore not be verified. Finally, reference 1261 is underpredicted for both the NRTI tested sequences and NNRTI tested sequences and we could not identify an apparent cause for this behaviour. [49] (More detailed results are listed in Table 5.) The table shows that performance per drug is very good with a low RMSE (an average RMSE of 0.54 log units; with two outliers, AZT and FTC, exhibiting an RMSE of >0.90 log units). Overall, when the results are grouped per literature reference number (which is included in the dataset) the average quality decreases and the standard deviation increases, indicating that differences between reported Log FC changes in literature exist and this could adversely affect model performance.
As each assay uses its own set of CCO values tuned for the respective assay we used values supplied by Virco and Rhee et al. for the Virco set and the Stanford set respectively. [44] Our model classifies the response correctly in 84% of the cases (Table 5). The average performance when grouped per individual drug class was very good (PI 85%, NNRTI 89% and NRTI 79%). Also noteworthy is that the model bias is towards over prediction rather than under prediction, something that is not always mentioned in literature but is especially relevant in a clinical setting.
Previous work on a high quality filtered subset of our Stanford DB set reached 80% correct predictions of phenotype from genotype on average (PI 78%, NNRTI 83% and NRTI 75%). [44] Other work indicates that an expert panel reaches up to 44% correct predictions. [50] The two outliers in the NRTI class are d4T and TDF, for which an apparent discrepancy between AVG data and Phenosense data has previously been described. [51]
In this work we report the construction of robust PCM models, based on 300,000 bioactivity data points measured against different HIV genotypes. In total, the model contained information on a total of 4 (NNRTI), 8 (NRTI) or 9 (PI) drugs combined with 10,700 (NNRTI) 10,500 (NRTI) or 27,000 (PI) mutants. Given the nature of the PCM modeling procedure employed in this work, we were able to combine all resistance profiles of the three above drug classes in three single models, hence focusing on very large target space (tens of thousands of different proteins) in this work. Both in internal validation and validation on unseen data our model showed performance comparable to assay reliability and better than sequence only models; moreover, model interpretation has been performed to identify novel resistance-conferring mutations that lead to resistance to all drugs in a class, such as T216M in the case of RT. In addition, we can use these models to find mutations that lead specific sensitivity (G48W in PR) or resistance (G68R in PR) to a single drug within a class.
Another application of our models is the support of personalized drug regimen predictions. We have shown that our models are able to predict clinical resistance with a high degree of reliability. This reliability is formed by a 95% CCP when predicting clinical response for Antivirogram data, which is the assay models were trained on, similar studies reached 80% CCP when predicting values for the assays they trained on. Furthermore, the CCP and is as high as 81% when predicting clinical response for unknown mutants. The novelty is formed by reliable predictions on unknown mutants and even unknown mixtures. Finally, the CCP is 84% when predicting clinical response for clinical isolates obtained from very diverse sources (including historical literature data and data from different assays), indicating that the model is robust and predictive.
We attribute the better performance of PCM to two reasons. Firstly our models are trained on a very large co-linked dataset. This large training set not only minimizes the influence and bias caused by single experimental error, it also allows the model to detect global patterns that are consistent over both genotype (sequence similarity) and chemo type (drug similarity). The second reason is related to the first, as the encoding of the full sequences using physicochemical properties rather than presence or absence of mutations allows for a better similarity measure between two sequences.
The main dataset was obtained from Virco (Beerse, Belgium) and consisted of mutants (both PR and RT sequences) and fold change (Log FC) in pIC50 (log units) data in the AVG assay collected by Virco up to January 2011. [12], [25], [28] Mixtures, consisting of multiple mutants that were identified in a single clinical isolate) were removed from the set and the total size of the dataset is listed in Table 6. The Log FC data was used as is, since it already consisted of log units difference to a single mutant defined as wild type. The wild type was defined as the HXB2 isolate (Uniprot accession P04585 and Genbank accession K03455), shown in Table 6 are the mean number of mutants present per sequence compared to HXB2. [22]
Furthermore, little duplicate sequences were actually in the dataset, specifically: NNRTI 1,501 duplicates (of 10,723 sequences), NRTI 1,411 duplicates (of 10,501 sequences), PI 9,803 duplicates (of 27,081 sequences).
Sequences were subsequently encoded using the first three Z-scales. [52], [53] The Z-scales are a previously published set of descriptors that characterize the physicochemical properties of the amino acids side chains. The resulting scales correlate to lipophilicity, size and polarity for each amino acid. For PR the full sequence was used and for RT only the first 400 amino acids were sequenced as the final 160 residues form an RnaseH domain and are not directly relevant in (N)NRTI resistance. These Z-scales were subsequently used to train models.
Structures of the drugs were normalized and ionized at pH 7.4, they were assigned 2D coordinates and subsequently converted to Scitegic circular fingerprints of type ECFP_8, ECFP_10 and ECFP_12 (depending on drug class, as described below). [54], [55] All this was done in Pipeline Pilot Student Edition version 6.1.5. [56] Circular fingerprints employ all possible substructures for a molecule up to a predefined maximal bond diameter. Each substructure is then encoded as a bit given the value ‘1’ when present and ‘0’ when absent. The reason for employing circular fingerprints is that they have previously been shown to give very high retrieval rates in comparative studies. [57] The NRTI dataset employed ECFP_10 fingerprints while the NNRTI dataset used ECFP_8 and the PI dataset ECFP_12 fingerprints.
In order to create a numeric descriptor for each drug, a similarity matrix was constructed using the fingerprints and based upon the Tversky Similarity coefficient. [58] Here fingerprints were converted to a fixed length array of counts with maximal length of 256 bits where the most descriptive bits were sorted to be at the beginning of the array. The value for α was 0.1 and the value for β was 0.9, putting more weight on the unique features of the target molecules compared to the reference molecule. For each drug in a class, the similarities to all other drugs from that class were then used as a descriptor (Tables S7, S8, S9).
Models were constructed in the academic version of Pipeline Pilot 6.1.5 using the R-statistics package. [59] Support vector machines (SVM) with a radial basis function kernel as coded in the e1071 package were used for model creation. [60] Parameters gamma and cost were tuned over an exponential range and epsilon was set at 0.25. It has been shown that setting epsilon to the approximate data error is the optimal value for training. [61] The optimal model was determined using 5 fold cross validation before proceeding to external validation of the model. The parameters used for validation were R02, R2, and RMSE. [36], [62]
As our models are trained on a database of different HIV mutants, applicability domains based on a single wild type sequence are expected to perform sub optimal. Rather we choose to determine the applicability domain based on the density of the nearest neighbors in the training set. This density was expressed as the fraction of the total number of sequences meeting a certain similarity criterion. Therefore this density score will be between 0 (0%, no sequences meeting the similarity criterion) and 1 (100%, all sequences meeting the similarity criterion). We calculated the density at a large number of similarity thresholds between 99% and 70%. Optimal performance was reached at 97%, similarity defined as 1 minus the euclidean distance. Furthermore, this similarity was based on full sequence similarity rather than binding site similarity.
Hence for each sequence, the total number of sequences being 97% similar or more can be between 0.0 (none) and 1.0 (all). We found that in practice the total fraction did not exceed 0.3 (30% of the sequences in the training set 97% similar or more).
The learning curves provide an estimate for the maximal performance that can be achieved on these datasets, simultaneously they represent external validation. The learning curves show that the models gradually improve when trained on a larger dataset. The results show that PCM is not only able to create models on this data, but also that these models are robust with good validation parameters. The PI model shows the best performance, RMSE = 0.42 log units when trained on 5% of the full set and <0.30 log units when trained on 70% of the dataset. The NNRTI model the worst performance, RMSE = 0.70 log units when trained on 5% of the full set and <0.50 log units when trained on 70% of the dataset (supporting Figure S1).
Subsequent to learning curve creation, y-scrambled models were created. Here the measured value (i.e. Log FC) was randomly permutated over the drug – mutant combination. The rationale being that no correlation should remain as the presence of a certain mutation will no longer be associated with a lower Log FC value but with mixed Log FC values. Supporting Table S2 display the lack of correlation between measured and scrambled values.
Models that were trained on this scrambled set and validated on 30% the data that was kept unscrambled produced very high RMSE values. These values were (in log units); 0.83 (PIs, versus 0.27 for predictive models), 1.10 (NRTIs, versus 0.31 for predictive models) and 1.11 (NNRTIs, versus 0.45 for predictive models). Furthermore, the values for the R02 were very low; −0.06 (PIs, versus 0.89 for predictive models), −0.20 (NRTIs, versus 0.75 for predictive models) and −0.21 (NNRTIs, versus 0.79 for predictive models) (Supporting Table S2). Finally the cross validation parameters for the models trained on these scrambled sets demonstrated a lack of correlation; RMSE in log units was highly similar to the external validation; 0.87 (PIs), 1.11 (NRTIs) and 1.12 (NNRTIs). The corresponding correlation coefficient was 0.00 for all three models.
To determine the effect of individual residues, for each sequence each residue was mutated back to wild type in silico by replacing the descriptors of the mutant amino acid with the descriptors of the corresponding wild type residue as was done previously. [33] Subsequently for all drugs the model prediction on the original mutant sequence was compared with the prediction of the model on the in silico changed mutant sequence. The difference was interpreted as the change in pIC50 induced by that particular residue, hence providing model interpretability. Changes that led to a 0 value shift in pIC50 were removed in the calculation of the average influence of mutations in a particular position, since in all cases this was caused by substitution of identical amino acids.
Known resistant mutations were retrieved from earlier publications by Johnson et al. and compared to our model interpretation. [6], [40] While these papers only mention high impact mutations and are gathered over the full population, they are a good frame of reference for our model interpretation. We used both the most recent publication and one from 2006 as Delavirdine (DLV) has been removed from these overviews due to the fact that it is only used rarely.
Mutations were filtered using the following parameters: have a negative effect on the majority of drugs in a single class; occurrence in the dataset more than once; average Log FC for all compounds >0.4; standard deviation over this average <0.4. This provided us with a number of mutations that lead to an increase in fold change on average, again using literature we discarded any previously known mutations and kept those mutations that were novel. [6], [18], [40]
For all interpretable mutations, the standard deviation was calculated over the average Log FC values per drug within a class. Subsequently all mutations were ranked and the top 30 were retained here. The goal here was to find mutations that have the most diverse effect over the different drugs within a class.
The dataset we used to compare the performance of PCM models with sequence only models was obtained from Van der Borght et al. [39] From the paper the 150 sequences with the largest prediction error were selected per drug class. For mixtures present in this set the average value of each z-scale for each of the present variants at a single position was used as descriptor. Mixtures with more than four possible variants at a single position were discarded leading to a total of 146 NNRTI sequences, 146 NRTI sequences, and 149 PI sequences.
Prediction of the Stanford University set is of is of particular interest since the correlation between Phenosense and AVG has previously been shown not to be very strong. [63], [64] Yet it should also be noted that the Phenosense assay is in fact more quantitative. AVG measures cell death which can be sensitive to slight differences in the state of the host cells used to grow virus. In particular for mutations M41L, M184V, and T215Y there are differences in Phenosense predictions compared with AVG. [65] While the correlation between Phenosense and VircoTYPE (trained on AVG) is slightly better, there are discrepancies. For instance the resistance profile of d4T and TDF, have been shown to have a Pearson's correlation coefficient <0.8 between the two assays. [51] The reference set was downloaded from the Stanford website (version 5.0, July 30, 2010), from this set the sequences by Virco were removed (as they are presumed to be in the training set, and this would artificially boost the results). The mixtures were removed and this provided us with the following numbers of sequence – compound pairs: 1,252 (NNRTI), 2,190, (NRTI), and 4,356 (PI).
After we predicted the Log FC values for individual drug – mutant pairs using our models, the validation parameters were calculated grouped by: Sequence ID (average and standard deviation), per Isolate (average and standard deviation), per Reference ID (average and standard deviation), per drug (average and standard deviation), per class (total), and per individual drug (total) (Table 5). The predictions per class are also included in Figure 7. Note that the raw data was used and no selection for high quality data was made, furthermore, the data was gathered at different labs, using different assays.
Resistance was also classified using clinical cut-offs (CCOs), here we used the values provided on the Stanford website and the values from AVG were obtained from Virco (supporting Table S10 and supporting Table S11). Subsequently CCP was calculated as a fraction of the total, in addition the fraction of overpredicted clinical response (resistance is predicted higher than measured experimentally) and underpredicted clinical response (resistance is predicted lower than measured experimentally) is included.
|
10.1371/journal.pgen.1003796 | Meta-Analysis of Genome-Wide Association Studies Identifies Six New Loci for Serum Calcium Concentrations | Calcium is vital to the normal functioning of multiple organ systems and its serum concentration is tightly regulated. Apart from CASR, the genes associated with serum calcium are largely unknown. We conducted a genome-wide association meta-analysis of 39,400 individuals from 17 population-based cohorts and investigated the 14 most strongly associated loci in ≤21,679 additional individuals. Seven loci (six new regions) in association with serum calcium were identified and replicated. Rs1570669 near CYP24A1 (P = 9.1E-12), rs10491003 upstream of GATA3 (P = 4.8E-09) and rs7481584 in CARS (P = 1.2E-10) implicate regions involved in Mendelian calcemic disorders: Rs1550532 in DGKD (P = 8.2E-11), also associated with bone density, and rs7336933 near DGKH/KIAA0564 (P = 9.1E-10) are near genes that encode distinct isoforms of diacylglycerol kinase. Rs780094 is in GCKR. We characterized the expression of these genes in gut, kidney, and bone, and demonstrate modulation of gene expression in bone in response to dietary calcium in mice. Our results shed new light on the genetics of calcium homeostasis.
| Calcium is vital to many biological processes and its serum concentration is tightly regulated. Family studies have shown that serum calcium is under strong genetic control. Apart from CASR, the genes associated with serum calcium are largely unknown. We conducted a genome-wide association meta-analysis of 39,400 individuals from 17 population-based cohorts and investigated the 14 most strongly associated loci in ≤21,679 additional individuals. We identified seven loci (six new regions) as being robustly associated with serum calcium. Three loci implicate regions involved in rare monogenic diseases including disturbances of serum calcium levels. Several of the newly identified loci harbor genes linked to the hormonal control of serum calcium. In mice experiments, we characterized the expression of these genes in gut, kidney, and bone, and explored the influence of dietary calcium intake on the expression of these genes in these organs. Our results shed new light on the genetics of calcium homeostasis and suggest a role for dietary calcium intake in bone-specific gene expression.
| Normal calcium homeostasis is regulated by three major hormones acting on their corresponding receptors in gut, kidney, and bone: parathyroid hormone (PTH) release governed by the calcium-sensing receptor (CASR), calcitonin, and the active metabolite of vitamin D, 1,25(OH)2-D. Despite heritability estimates of 33–78%, the genetic determinants of serum calcium are poorly understood [1], [2], [3]. We have previously reported a variant in CASR associated with calcium concentrations in European-ancestry individuals [4], [5]. To detect additional loci, we conducted a two-stage genome-wide association meta-analysis of serum calcium and studied expression of identified genes in key calcium homeostatic organs in the mouse under various calcium diets.
The discovery analysis consisted of 39,400 individuals from 17 population-based cohorts of European descent (Table 1 and Table S1). There was little evidence for population stratification at study level (median genomic inflation factor, λ = 1.006) or meta-analysis level (λ = 1.03), and we detected an excess of association signals beyond those expected by chance (Figure S1).
The CASR locus, previously identified in Europeans, was confirmed in our meta-analysis (P = 6.5E-59, Figure S2) [4], [5]. In addition, SNPs from five independent regions reached genome-wide significance (P<5E-08) in the overall discovery meta-analysis (Figure 1, Table 1, Table S2): rs1550532 (in DGKD, P = 4.60E-08), rs780094 (in GCKR; P = 3.69E-11), rs17711722 (near VKORC1L1, P = 2.78E-11), rs7481584 (in CARS, P = 9.21E-10) and rs1570669 (near CYP24A1; P = 3.98E-08).
Fourteen SNPs from Stage 1 were sent for Stage 2 validation in ≤21,679 additional Europeans: the twelve independent (≥1 Mb apart) SNPs with lowest P values (6.5E-59 to 8.1E-06) in Europeans and two additional genome-wide significant loci (rs9447004 and rs10491003) from a combined sample including 8318 Indian-Asians (Table 1). Of the fourteen SNPs, seven were considered successfully replicated (i.e. were in the same direction of effect as the discovery meta-analysis, had a one-side replication P<0.05 and were genome-wide significant (P<5E-8) in combined meta-analysis of discovery and replication sets). These were rs1801725 in CASR, rs1550532 in DGKD, rs780094 in GCKR, rs7336933 near KIAA0564 and DGKH, rs10491003 (closest gene GATA3), rs7481584 in CARS and rs1570669 near CYP24A1 (Table 1). Regional association plots are presented in Figure S3. Details on the seven SNPs that did not replicate are presented in Table S2. Association results for serum calcium in Caucasians for all SNPs with P value<5*E-5 are listed in Table S3. In a secondary analysis, all SNPs identified in the primary analysis showed consistent and significant association with serum calcium adjusted for serum albumin (Table S4, Figure S4), as well as an excess of association signals beyond those expected by chance (Figure S5); no additional locus was identified using albumin-corrected serum calcium (Table S5).
We found no significant association of the 7 replicated SNPs known to provide reliable tags for copy number variations (CNVs) in people of European-descent from the Hypergene dataset. For all the SNPs, the calculated correlation was below 0.002. We also explored a list of SNPs tagging CNVs from the GIANT consortium. Out the 7 SNPs tested, only the rs1570669 was in slight linkage disequilibrium (r2 = 0.54) with one SNP of the WTCCC2 list (rs927651). The corresponding SNP tags the CNVR7875.1 CNV located 455b from the SNP of interest.
For each of the 7 replicated SNPs, we identified all proxy SNPs with r2>0.8 in HapMap CEU (releases 21, 22, and HapMap 3 version 2) using the online SNAP database (http://www.broadinstitute.org/mpg/snap/). This led to the identification of 40 SNPs. We then queried each of these SNPs in the eQTL database of the University of Chicago (http://eqtl.uchicago.edu/cgi-bin/gbrowse/eqtl/). Three of the seven SNPs are in strong linkage disequilibrium with an eQTL, as illustrated in Table S6.
Proposed functions of the genes mapping into the associated intervals (±250 kb) are in Box 1 and in Table S7 for the gene-rich GCKR region. We report in Table S8 the mechanism and/or location of all available biological processes, cellular components and molecular functions related to the genes mapping into the associated intervals from the AmiGo 1.8 gene ontology database. We also queried the OMIM database for each genes located within ±250 kb of the replicated loci (Table S9)
In Indian-Asians, all 7 replicated SNPs had beta-coefficients that were direction-consistent with the primary analysis and 3 were statistically significant (P<0.05): rs1801725 (CASR, P = 1.4E-31), rs1550532 (DGKD, P = 0.002) and rs10491003 (GATA3, P = 0.009) (Table S10). In Japanese, 3 SNPs had betas that were direction-consistent with the primary analysis, but only rs1801725 (CASR) was associated with serum calcium (P = 0.001) (Table S10).
We conducted analyses of related bone mineral and endocrine phenotypic traits for the 7 replicated loci (Table 2). Several SNPs were associated (P<0.05) with bone mineral density (BMD) in the GEFOS consortium [6]: rs1801725 at CASR (P = 0.025; previously reported [4], [5]) and rs780094 (GCKR) at the lumbar spine (P = 0.006), rs1570669 at CYP24A1 at the femoral neck (P = 0.04), and rs1550532 at DGKD at both the lumbar spine (P = 0.003) and the femoral neck (P = 0.003). For endocrine phenotypes, rs1570669 at CYP24A1 was associated with higher PTH concentrations (P = 0.0005) and rs1801725 at CASR with higher serum PTH concentrations (P = 0.028) and lower serum phosphate concentrations, as previously reported [4], [5]. No SNP was associated significantly with circulating 25-OH vitamin D concentrations (all P>0.05) in the SUNLIGHT consortium [7].
We selected biologically plausible gene(s) at each locus for in vivo studies in a mouse model as described in Methods' section. We first analyzed gene expression in the three primary calcium-handling organs: duodenum, kidney and bone (tibia). CASR for the rs1801725 locus, DGKD for the rs1550532 locus, GATA3 for the rs10491003 locus, CARS, NAP1L4 and CDKN1C for the rs7481584 locus, DGKH and KIAA0564 for the rs7336933 locus, were expressed in all organs, whereas CYP24A1 (rs1570669 locus) was solely, and PHLDA2 (rs7481584 locus) mainly, expressed in the kidney (Figure 2). No significant expression of GCKR (rs780094 locus) was observed in any organ tested, which is of interest considering the strong attenuation of the association of rs780094 with serum calcium after adjustment for albumin (Table S4). In micro-dissection of nephron segments [8], [9], DGKD, DGKH, CARS, KIAA0564 and CYP24A1 were primarily transcribed in the proximal tubule, CASR in the thick ascending limb, and GATA3 predominantly in the distal nephron and collecting duct (Figure 3).
In order to determine regulation of gene expression by calcium intake, we measured gene expression levels in mice fed low and high calcium diets (0.17% vs. 1.69% calcium) for one week, with normal diet as control (0.82%) (Figure 4 and Table S11). In the kidney, both DGKD and DGKH were upregulated in response to low calcium diet (P≤0.05; Figure 4). In the tibia, CASR was markedly upregulated in response to low calcium diet (2.5-fold increased expression), as were GATA3, KIAA0564 and CARS (P≤0.05 for all; Figure 4), findings that suggest regulation by 1,25(OH)2-D. DGKD and DGKH were upregulated in the tibia in response to high and low calcium diet (P≤0.05 for all; Figure 4). The expression in duodenum of the majority of genes was not modified by dietary calcium, with the exception of NAP1L4 and CDKN1C.
We have identified and replicated one known and six new loci for serum calcium near genes linked to bone metabolism and endocrine control of calcium. Of these, 4 loci (DGKD, GCKR, CASR, and CYP24A1) were nominally associated with BMD in the general population. In supporting mouse studies, we demonstrate expression of several of these genes in tibia, and show regulation of gene expression in response to dietary calcium intake. We also demonstrate expression in nephron segments known to regulate calcium homeostasis. Taken together, these results shed new light on the genetics of calcium balance.
The vast majority of total body calcium is bound in the skeleton as hydroxyapatite and other calcium-phosphate complexes [10]. Apart from providing skeletal strength, bone serves as a calcium reservoir to maintain tightly controlled circulating concentrations vital to cellular signaling, muscle contraction and coagulation [10]. However, the genetic basis of the dynamic cross talk that occurs between these compartments is poorly understood. Our results advance our understanding in this area. Eight genes identified in the GWAS are constitutively expressed in bone and are regulated in response to dietary calcium, in particular low calcium diet, whereas no clear change was observed in kidney or duodenum. This bone reactivity in response to dietary calcium intake is consistent with what was recently reported for CASR [11]. Further, of the eight genes expressed in bone and regulated in response to dietary calcium, we show that rs1550532 (DGKD) and rs1801725 (CASR) are associated with BMD in humans, the primary determinant of fracture risk.
The A allele of rs1570669 (CYP24A1 locus) was associated with reduced BMD at the femoral neck although CYP24A1 was not found to be expressed in bone in mice experiment, which suggests an indirect role in bone mineralization. This may occur via its documented role in vitamin D metabolism, discussed below, and/or its association with higher PTH concentrations identified in the present analysis.
We observed specific expression patterns of several genes in the mouse nephron: DGKD, DGKH, CARS, KIAA0564 and CYP24A1 were primarily transcribed in the proximal tubule, CASR expression was mostly localized to the thick ascending limb, whereas GATA3 was predominantly found in the distal part of the nephron and the collecting duct. This pattern of expression in segments known to be involved in calcium reabsorption suggests a role in renal calcium handling and is consistent with previous exploratory transcriptome analyses in humans and mice [12], [13]. Both DGKD and DGKH were significantly upregulated in the kidney in response to low calcium diet, suggesting specific involvement of these genes in renal calcium handling.
Several of the newly identified loci harbor genes linked to the hormonal control of serum calcium. First, the association of CASR with PTH concentrations is consistent with its known role in PTH signaling. Second, several lines of evidence implicate rs1570669 (CYP24A1) in the vitamin D pathway: its association with serum calcium and PTH concentrations, its selective expression in the proximal tubule where 1,25(OH)2-D metabolism occurs, and that loss-of-function CYP24A1 mutations cause vitamin D-induced hypercalcemia in children (idiopathic infantile hypercalcemia). Third, we identified variants linked to 2 chromosomally distinct isoforms of diacylglycerol kinase, part of the phosphoinositol second messenger system, that may interact with each other at the protein level [14], [15].
Strengths of this study are the large sample size and consistent mouse studies to support the statistical associations and advance our knowledge of the biology at these loci. Human and mice largely share physiological processes linked to calcium metabolism, including tissue-specific gene expression. Limitations include the lack of a direct marker of bone remodeling and the potential for bias in gene selection for experimental follow-up. Mice may display subtle differences in the regulation of the genes tested compared to humans.
We have identified and replicated one known and six new loci for serum calcium near genes linked to bone metabolism and endocrine control of serum calcium. Supporting experimental mouse studies suggest a role for dietary calcium in bone-specific gene expression. Further work is needed to identify the causal variants and to understand how they influence calcium homeostasis.
In each human study, the local institutional review board approved the study and participants signed written informed consent, including for DNA analyses. The experimental protocol in mice was approved by the local veterinarian authorities and fulfilled Swiss federal regulations for experiences with animals.
Detailed information on the genotyping plateforms and data cleaning procedures for each discovery and replication cohort can be found in Table S13. De novo replication genotyping was perfomed in 4670 participants to the Bus Santé Study using KASPar v4.0 after whole genome amplification by primer extension pre-amplification (PEP) using thermostable DNA polymerases.
In each discovery study, genotyping was performed using a genome-wide chip and nearly 2.5 million SNPs were genotyped or imputed using the HapMap CEU panels release 22 or 21 as the reference. Each study applied quality control before imputation. Detailed imputation information is provided in Table S13. Each SNP was modeled using an additive genetic effect (allele dosage for imputed SNPs), including age and sex as covariates in the model as well as study-specific covariates if needed (e.g. principal components, study center). The primary dependent variable in each discovery study was untransformed and uncorrected serum calcium expressed in mg/dL. Beta regression coefficients and standard errors were used with at least 5 decimal places. For secondary analyses, albumin-corrected serum calcium was computed using the following formula: ([4-plasma albumin in g/dL]×0.8+serum calcium in mg/dL) and the same model as for the primary analyses was used. Each file of genome-wide summary statistics underwent extensive quality control prior to meta-analysis both for primary and secondary analyses, including (1) boxplots of all beta coefficients, as well as all standard errors multiplied by the square-root of the sample size, for each study separately; (2) the range of P values, MAF, imputation qualities, call rates and Hardy-Weinberg equilibrium P values and (3) QQ plots. In addition, we checked the direction and magnitude of effect at the previously reported rs1801725 CASR variant. Genome-wide meta-analyses were conducted in duplicate by two independent analysts. For each SNP, we used a fixed effect meta-analysis using inverse-variance weights as implemented in the meta-analysis utility Metal [16]. Results were confirmed by a z-score based meta-analysis. Data were available for 2,612,817 genotyped or imputed autosomal SNPs for the primary and secondary analyses. After the meta-analysis, genomic control correction was applied (λGC was 1.03 for both uncorrected and corrected serum calcium). Our pre-specified criterion to declare genome-wide significance was P value<5E-8 to account for 1 million independent tests according to the Bonferroni correction. We choose to move forward for replication all SNPs with discovery P value<1E-7 in the European sample or genome-wide significant SNP in the overall sample that included Indian Asians. To choose a single SNP per genome-wide associated region for replication, we merged all SNPs within 1 Mb region and selected the lowest P value for each region. Altogether, fourteen SNPs were moved forward for replication. Up to 17,205 participants contributed information to the replication analyses in silico and 4,670 participants provided data for de novo genotyping. We used fixed-effects inverse-variance weighted meta-analysis to combine discovery and replication meta-analysis results. Replication was considered as present whenever a combined P value<5E-8 together with an effect-concordant one-sided replication P value<0.05 were obtained.
We conducted look-ups for femoral and lumbar bone density in the GEnetic Factors of OSteoporosis (GEFOS) dataset [17]. Bone mineral density (BMD) is used in clinical practice for the diagnosis of osteoporosis and bone density at different skeletal sites is predictive of fracture risk. BMD was measured in all cohorts at the lumbar spine (either at L1–L4 or L2–L4) and femoral neck using dual-energy X-ray absorptiometry following standard manufacturer protocols [17]. Serum phosphorus was looked up from a previously published GWAS meta-analysis, including 16,264 participants of European ancestry [18]. Serum phosphorus concentrations were quantified using an automated platform in which inorganic phosphorus reacts with ammonium molybdate in an acidic solution to form a colored phosphomolybdate complex [18]. The 25-hydroxyvitamin D was looked-up in the SUNLIGHT consortium [7], which includes data from 33,996 individuals of European descent from 15 cohorts. 25-hydroxyvitamin D concentrations were measured by radioimmunoassay, chemiluminescent assay, ELISA, or mass spectrometry [7]. PTH was looked-up in the SHIP and SHIP-Trend studies. The serum parathyroid hormone concentration was measured on the IDS-iSYS Multi-Discipline Automated Analyser with the IDS-iSYS Intact PTH assay (Immunodiagnostic Systems Limited, Frankfurt am Main, Germany) according to the instructions for use. This chemiluminescence immunoassay detects the full-length parathyroid hormone (amino acids 1–84) and the large parathyroid hormone fragment (amino acids 7–84). The measurement range of the assay was 5–5000 pg/mL. The limits of blank, detection and quantitation were 1.3 pg/mL, 1.4 pg/mL, and 3.6 pg/mL, respectively. As recommended by the manufacturer, three levels of control material were measured in order to verify a decent working mode. During the course of the study, the coefficients of variation were 14.02% at low, 6.64% at medium, and 6.84% at high serum parathyroid hormone concentrations in the control material in SHIP and the corresponding percentages were 16.8% at low, 10.7% at medium, and 9.0% at high serum parathyroid hormone concentrations in the control material in SHIP-Trend.
The Hypergene dataset (a 4206 samples case-control study concerning hypertension genotyped using the Illumina 1M chip) has been used to call CNVs and to check their correlation with the SNPs of interest. The CNVs calls have been done using pennCNV software [19]. A SNP by sample matrix with the copy number status was created. Then the square correlation (Pearson correlation) between value of each SNP of interest and the SNPs copy number status in a +/−2 Mb region was calculated. The SNPs of interest for which no correspondence has been found in the Hypergene dataset have been replaced by the closest SNPs in high linkage disequilibrium (LD) and present in the Hypergene dataset. LD between the SNPs of interest and a list of SNPs tagging CNVs from the GIANT consortium has also been calculated. The SNPs from the GIANT list are in LD higher than 0.8 with their corresponding CNV.
We queried the AmiGo 1.8 gene ontology database for each gene located within ±250 kb of the seven replicated SNPs, including rs1801725 (CASR). (http://amigo.geneontology.org/cgi-bin/amigo/go.cgi, last accessed November 6, 2012). We used Homo sapiens as a filter for species.
For each of the 7 replicated SNPs, we identified all proxy SNPs with r2>0.8 in HapMap CEU (releases 21, 22, and HapMap 3 vers. 2) using the online SNAP database (http://www.broadinstitute.org/mpg/snap/). We then queried each of these 40 SNPs in the eQTL database of the University of Chicago (http://eqtl.uchicago.edu/cgi-bin/gbrowse/eqtl/).
The rs1801725 SNP encodes a missense variant in exon 7 of the CASR gene leading to an alanine to serine substitution (A986S). Given the key physiological role of CASR in calcium homeostasis (monogenic disorders of calcium balance), this gene was the logical candidate for analysis in mouse at this previously identified locus.
For the 6 newly identified loci, the precise rationale for gene selection varied from one locus to the other, but the main criteria was to focus on the most biologically relevant gene. Rs1550532 on chromosome 2 is an intronic SNP of DGKD, which was the most likely biological candidate for this locus and was therefore selected for analysis in mouse. None of the other genes located in this region (±250 Kb) has a known link with calcium homeostasis (Box 1) and rs1550532 is not in strong linkage disequilibrium with an eQTL (Table S6). We also took into account the fact that another member of the DGK family, namely DGKH was located near one of the other replicated loci, on chromosome 13.
Rs780094, on chromosome 2, is located in intro 16 of GCKR and is in strong linkage disequilibrium (r2 = 0.93) in Caucasians [20], with a common non-synonymous SNP (P446L, rs1260326) associated with glucokinase activity in vitro [20], [21]. This SNP has been associated with multiple other phenotypes in previous GWAS and it is in strong linkage disequilibrium with an eQTL (Table S6). Previous fine mapping analysis of this locus has attributed the signal from rs780094 to the functional rs1260326 variant [20]. The GCKR locus may indirectly influence calcium concentrations via its association with albumin levels [22]. In line with this, we observed an attenuation of the association of rs780094 with albumin-corrected serum calcium compared to the association with uncorrected serum calcium and we found GCKR not to be expressed in any of the key organs involved in calcium homeostasis that we tested in mice. We selected GCKR for analysis in mouse at this locus.
Rs10491003 on chromosome 10 is located within a long non-coding RNA. For this locus, we selected GATA3, the nearest and only gene located within this region, for analysis in mouse. GATA3 is implicated in monogenic disorders of calcium balance.
Rs7481584 is located within CARS (intronic SNP) in an imprinted region known to play a role in multiple cancers, which makes this locus a plausible candidate for malignancy-related hypercalcemia. Other plausible biological candidates in this locus are NAP1L4, PHLDA2 and CKDN1C (Box 1). Rs7481584 is in strong LD with 2 eQTLs, one associated with the expression of NAP1L4 (rs2583435) and the other one associated with the expressions of SLC22A18 and SLC22A18AS. We selected CARS, NAP1L4, PHLDA2 and CKDN1C for analyses in mouse.
For rs7336933, we selected the two only genes (DGKH and KIAA0564) located under this association peak on chromosome 13 for analyses in mouse.
Finally, rs1570669 is an intronic SNP of CYP24A1, a strong biological candidate implicated in monogenic disorders of calcium balance. The two other genes of this region (BCAS1 and PFDN4) have no known link with calcium homeostasis. Furthermore, rs1570669 and PFDN4 are separated by a recombination hot spot. We selected CYP24A1 for analysis in mouse.
As animal experiments started while the replication process was underway, we had also initially selected the following genes for analysis in mouse: RSG14 and SLC34A1 at locus rs4074995 (discovery P value = 2.4E-07), VKORC1L1 at locus rs17711722 (discovery P value = 2.8E-11), PYGB at locus rs2281558 (discovery P value = 6.4E-07), CD109 at locus rs9447004 (discovery P value = 8.1E-06). No gene was selected for the rs2885836 and rs11967485 and rs12150338 loci in the absence of obvious candidate. Results for these unreplicated loci can be found in Figures S6, S7 and S8. We present these results for quality control purposes: SLC34A1 (also known as NAPI-3 or NPT2), which encodes solute carrier family 34 (sodium phosphate), member 1, was expressed in the kidney, but neither in duodenum nor in bone, as expected based on current knowledge on this phosphate transporter. In the kidney SLC34A1 was mainly expressed proximally and SLC34A1 expression was upregulated under low calcium diet, which is in line with the known function of this gene.
Five C57bl/6 mice (Janvier) per group were fed, for one week, three different diets in which the percentage of calcium were 0.17% (low calcium diet), 0.82% (normal calcium diet) and 1.69% (high calcium diet) and had free access to water. 12∶12 hours light/dark alternance was imposed. At the end of the week of the specific diet, spot urine were collected and mice were anesthetized. Blood was collected by retro-orbital puncture. Organs were immediately harvested and snap frozen. RNA was extracted using Trizol (Invitrogen) and reversed transcribed with PrimeScriptTM RT reagent Kit (Takara Bio Inc). Calcium, sodium, phosphate and creatinine in plasma and urine were analyzed at the central lab of the Lausanne University hospital using a Cobas-Mira analyzer (Roche).
|
10.1371/journal.pntd.0004303 | Genetic Diversity and Population Structure of Leishmania infantum from Southeastern France: Evaluation Using Multi-Locus Microsatellite Typing | In the south of France, Leishmania infantum is responsible for numerous cases of canine leishmaniasis (CanL), sporadic cases of human visceral leishmaniasis (VL) and rare cases of cutaneous and muco-cutaneous leishmaniasis (CL and MCL, respectively). Several endemic areas have been clearly identified in the south of France including the Pyrénées-Orientales, Cévennes (CE), Provence (P), Alpes-Maritimes (AM) and Corsica (CO). Within these endemic areas, the two cities of Nice (AM) and Marseille (P), which are located 150 km apart, and their surroundings, concentrate the greatest number of French autochthonous leishmaniasis cases. In this study, 270 L. infantum isolates from an extended time period (1978–2011) from four endemic areas, AM, P, CE and CO, were assessed using Multi-Locus Microsatellite Typing (MLMT). MLMT revealed a total of 121 different genotypes with 91 unique genotypes and 30 repeated genotypes. Substantial genetic diversity was found with a strong genetic differentiation between the Leishmania populations from AM and P. However, exchanges were observed between these two endemic areas in which it seems that strains spread from AM to P. The genetic differentiations in these areas suggest strong epidemiological structuring. A model-based analysis using STRUCTURE revealed two main populations: population A (consisting of samples primarily from the P and AM endemic areas with MON-1 and non-MON-1 strains) and population B consisting of only MON-1 strains essentially from the AM endemic area. For four patients, we observed several isolates from different biological samples which provided insight into disease relapse and re-infection. These findings shed light on the transmission dynamics of parasites in humans. However, further data are required to confirm this hypothesis based on a limited sample set. This study represents the most extensive population analysis of L. infantum strains using MLMT conducted in France.
| In the south of France, the parasite Leishmania infantum is responsible for diseases that primarily affect dogs but can also impact humans. Several endemic areas have been clearly identified in the south of France including the Pyrénées-Orientales, Cévennes (CE), Provence (P), Alpes-Maritimes (AM) and Corsica (CO). In this study, 270 L. infantum isolates from four endemic areas, AM, P, CE and CO, were assessed using Multi-Locus Microsatellite Typing (MLMT), a tool applied for population genetic studies. MLMT revealed a strong genetic differentiation between the Leishmania populations from AM and P with exchanges observed between these two endemic areas. For four patients, the occurrence of disease relapses and re-infections was examined. These findings shed light on the transmission dynamics of parasites in humans. This study represents the most extensive population analysis of L. infantum isolates using MLMT conducted in France.
| Leishmaniases are a group of diseases caused by obligatory intracellular protozoan parasites of the genus Leishmania. Among the species of Leishmania, Leishmania infantum is mostly responsible for canine leishmaniasis (CanL), although it also causes sporadic cases of human visceral leishmaniasis (VL), and rare cases of cutaneous and muco-cutaneous leishmaniasis (CL and MCL) throughout the Mediterranean basin [1]. Transmission to humans is caused by the bite of infected phlebotomine sandflies, and dogs are considered to be the principal domestic reservoir. In France, the parasite is currently only endemic in the south of France, along the Mediterranean coast, where several foci have been clearly identified: Pyrénées-Orientales, Cévennes (CE), Provence (P), Alpes-Maritimes (AM) and Corsica (CO) [2]. In the Provence-Alpes-Côtes d’Azur (PACA) region, which comprises the AM and P endemic areas, transmission has been reported for 100 years [3,4]. The two cities of Nice and Marseille, which are located 150 km apart, and their surroundings concentrate the greatest number of French autochthonous leishmaniasis cases [2,5]. Although the same species of L. infantum (primarily zymodeme MON-1), the same predominant vector (Phlebotomus perniciosus) and the same and unique reservoir (dog) are found in both regions, the transmission environment of VL is heterogeneous in these two foci [5]. Disease transmission in Nice and the surrounding area is associated with scattered habitation and mixed forest in the foothills [5]. In contrast, around Marseille, VL transmission is associated with an urban environment [5]. Regarding the main vector; Phlebotomus pernicious; the population is quite homogeneous and belongs mainly to the same haplogroup (for 88% pern01) in Provence, France [6]. The isolates of L. infantum from AM and P endemic foci have been characterized using Multi-Locus Enzyme Electrophoresis (MLEE), which is the current reference method. However, MLEE based analyses are limited at the intrinsic level of polymorphisms. Thus, differentiating between isolates in PACA region is impossible using the MLEE method [7]. Epidemiological studies on L. infantum require the use of highly discriminative techniques that can differentiate between MON-1 strains. Multi-Locus Microsatellite Typing (MLMT) has been shown to be a powerful tool for population genetics and epidemiological studies of Leishmania spp. [8]. This tool has been already applied to genotype L. infantum isolates from healthy blood donors, sandflies, dogs and human patients in Southern France [9]. Genetic differentiations were evidenced between asymptomatic carrier strains and non-asymptomatic carrier strains and especially between asymptomatic carrier and HIV+ populations [9]. However, due to the weak sample size, these results must be confirmed on a larger sample set [9].
In the current study, microsatellite markers were used to analyze the genetic diversity of L. infantum parasites from Southeast France, with a focus on the PACA region. We assessed an extensive panel of isolates from an extended time period (1978–2011), from the two endemic regions of AM and P. The geographical and temporal distributions of genotypes were examined. The microsatellite profiles were used to assess relapse and re-infection among patients as well as the association between genotype and the various clinical forms of the disease.
The L. infantum isolates used in this study were obtained from the collection of the Centre National de Référence des Leishmania (Leishmania collection, BRC-Leish, Montpellier, France, BioBank N° BB-0033-00052). All human and animal samples had been isolated from patients and animals as part of routine diagnosis and treatment with no unnecessary invasive procedures. A total of 270 L. infantum isolates from the south of France were included in this study (Table 1). This panel included 247 human isolates from 239 patients (four patients harbored more than one isolate), 20 from CanL, one from feline leishmaniasis and two from sandflies. Among the 239 patients, there were 154 adult VL cases, 58 infant VL cases, 13 CL cases (five infants, seven adults and one unknown), three MCL adult cases, nine asymptomatic carriers (in adults) and two unknown cases (one adult and one unknown). Among the 247 human samples, 139 were isolated from immunocompetent patients, 95 from immunocompromised HIV+ patients, 11 from immunocompromised patients other than HIV+ (e.g., renal transplantation, lymphoma, auto immunity disease and cancer) and two from unknown cases. The location of isolates based on their position relative to the Vars River (S1 Table) was used for genetic differentiation analysis.
Concerning the geographical distribution (Fig 1), the samples were collected from the endemic areas of AM (n = 178), P (n = 75), CO (n = 9) and CE (n = 8). The samples were isolated at the University Hospitals of Nice, Marseille and Montpellier (France) between 1978 and 2011. MLEE typing and cryoconservation were performed at the Centre National de Référence des Leishmania (Leishmania collection, BRC-Leish, Montpellier, France, BioBank N° BB-0033-00052). Overall, 259 isolates were characterized as zymodeme MON-1, six were MON-24, one was MON-11, one was MON-80 and one was MON-108. The data were unavailable for two isolates. For this study, the primary cultures from the patient stored at -80°C were thawed and the cells were cultured for six days before harvesting for DNA extraction.
The microsatellite data indicated with an asterisk in Table 1 were obtained from a previous study [9].
DNA of the remaining isolates was extracted from promastigotes grown in Schneider’s insect medium (Sigma Aldrich, France) supplemented with serum calf, urine, penicillin, streptomycin and L-glutamine (Sigma Aldrich). Promastigotes were harvested on the sixth day of culture, and DNA was extracted from a pellet of 2X108 parasites using a QIAamp DNA mini kit (Qiagen, France) according to the manufacturer’s instructions. DNA extracted from the strain MHOM/FR/85/LEM716 was used as a control to determine the size of amplified microsatellite fragments, as this microsatellite data have been published [9].
Twelve microsatellite loci were amplified using the PCR conditions as previously described: LiBTG, LiBTA, LIST7021, LIST7025, LIST7026, LIST7031, LIST7033, Li22-25, Li45-24, TubCA, Li71-5/2 and Rossi2 [9–12]. The amplification products were analyzed using an automated capillary ABI Prism 3130XL Genetic Analyzer (Applied Biosystems, France). The data were stored and analyzed using GeneMapper analysis software (version 4.0, Applied Biosystems). PCR fragment sizes were determined using the internal size standard GeneScan 500 LIZ (Applied Biosystems). All 270 Leishmania isolates were genotyped at each of the 12 loci. With the PCR fragment size of the control strain MHOM/FR/85/LEM716, we were able to include microsatellite data from the Hide et al. study (data indicated with an asterisk in Table 1). Four isolates from this previous study (MCAN/FR/95/LPN122, MCAN/FR/95/LPN123, MCAN/FR/95/LPN124, and MCAN/FR/95/LPN124) were re-extracted from culture and re-genotyped blindly. The same microsatellite results described by Hide et al. were obtained [8].
Descriptive statistics for the observed genetic populations were calculated using Genetix version 4.05.2 (2004) and FSTAT Version 2.9.3.2 [13]. Using these programs, we calculated allelic diversity (number of allelic variants per maker), expected (He) and observed (Ho) heterozygosity, genetic diversity within subsamples Hs, inbreeding coefficient (FIS) the migration rates (gene flow) (Nm) [14]. The Fst value, which indicates the degree of genetic differentiation and gene flow among populations was also calculated. Fst values above 0.25 with significant p- values (<0.05) indicated strong genetic differentiation [15].
Phylogenetic analyses were performed based on the microsatellite profiles. A distance matrix was calculated using the Chord distance (Cavalli-Sforza and Edwards 1967) setting in the POPULATIONS 1.2.31 software with bootstrap values determined for 1,000 replicates (http://bioinformatics.org/~tryphon/populations/) [16]. The resulting distance matrix was processed using MEGA 4.0.2 to construct an unrooted Neighbor-Joining (NJ) tree [17].
The genetic characteristics of the Leishmania samples were also investigated using a model-based Bayesian clustering method implemented in STRUCTURE v 2.3.4 [18]. This algorithm simultaneously estimates the allele frequencies to assign individuals into genetically distinct populations (K) and each probability for the identification of the most likely number of populations. A series of ten independent runs was performed for each K value between one and ten. The following parameters were used: burn in period of 20,000 iterations, 200,000 Markov Chain Monte Carlo iterations and admixture model. The most probable number of clusters was identified via calculation of the Delta K (ΔK), which is based on the rate of change in the log probability of data between successive K values. The peak of the ΔK graph corresponds to the most probable number of populations in the data set [19].
A chi-square statistical test was performed to determine whether the observed data differed significantly from the expected ratios. The chi-square value was considered significant when p≤0.05. This test was used to compare the proportion of HIV+ patients in Populations A and B.
In total, 270 isolates were typed at 12 microsatellite markers with one or two alleles at each locus. All markers were polymorphic. The number of alleles per locus (Na) ranged from 3 to 13. Li22-35, LIST7021, LiBTG, LiBTA and LIST7026 were the most polymorphic markers with 13, 12, 10, 9 and 8 alleles, respectively. LIST7031, Li71-5/2 and LIST7025 were the least polymorphic markers with three different alleles (Table 2).
Genetic variability was analyzed among the 12 microsatellite loci (Table 2). The Ho was weak and ranged from 0.011 to 0.174 for Li71-5/2 and LiBTA, respectively, with an overall Ho at 0.044. The mean intra-population Hs was 0.546 (0.224–0.811) for the entire sample set and 0.531 (0.208–0.804) for the MON-1 population (Table 2). The FIS for the entire population was 0.920, thereby indicating a considerable degree of inbreeding. A separate analysis was performed to investigate the genetic polymorphisms among the four geographically determined populations (Table 3). Extensive inbreeding in the four populations was observed, with the highest inbreeding coefficient found in the populations of the CO and AM endemic areas. The genetic differentiation among the four endemic areas was tested using FSTAT version 2.9.3.2 (Table 4). The Fst values ranged from 0.067 to 0.321. All Fst values between the four endemic areas were significant. We obtained lower values for P versus CE and P versus CO and higher values for AM versus CE and CO versus CE (which may be due to the low number of isolates collected from CE and CO). When comparing AM and P samples from 1993 to 2009 (corresponding to the time period of the isolation of samples from the P endemic area used in this study), the Fst value obtained was similar to the Fst value corresponding to the complete period of sample collection (1978 to 2011).
Sub-populations were also defined according to their position in relation to Vars River (S1 Table). This river is located in the southeast of France and flows in the Alpes-Maritimes Department. The Fst values highlighted a gradient of differentiation from the East Vars to the West Vars and up to the P endemic area. A high and significant genetic differentiation (Fst = 0.308) was obtained when comparing isolates from east of the Vars River and the P endemic area. The comparison of sub-populations isolated from the P endemic area failed to show any genetic differentiation (Marseille versus Toulon or Marseille versus other cities in the P endemic area), with the limitation that few samples were collected from Toulon (n = 7).
A total of 121 different genotypes were identified from the 270 isolates corresponding to a genotype frequency of 0.037. The entire sample set comprised 91 unique genotypes (75%). Among the 30 repeated genotypes, seven were common to two endemic areas, one was common to three endemic areas and 22 pertained to the same focus (Fig 2). The genotypes 21 (primarily) and 46 were found in both endemic areas of AM and P (Fig 2). Within these endemic areas, Faucher et al. have described high- and low-risk sub-areas of VL [5]. In our study, the repeated genotypes were almost exclusively found in high-risk areas, with the only exception of genotype 21 which was found in high and low risk sub-areas (Fig 3A and 3B). Among the 30 repeated genotypes, 11 were found in four or more samples. These 11 genotypes represented 132 isolates.
The isolates were collected over the course of a 33-year period from 1978 to 2011 (Fig 4). The repeated genotypes, found in four or more samples, were isolated over a period of six (genotype 68) to 29 years (genotypes 57 and 55). All genotypes except 68, 36, 34 and 43 were isolated from both humans and dogs. Genotype 21 was found to be present in the AM and P endemic areas over the course of 28 years. This genotype was first isolated in 1981 in the AM endemic area and then in 1996, at which time it was first identified in both the AM and P endemic areas. After 1996, genotype 21 was only isolated in the P. Finally, in 2009, this genotype was found in both the AM and P endemic areas. Genotype 21 was isolated from a variety of patients including asymptomatic carriers (AC), adult VL, infant VL and HIV+ VL cases.
Four patients (patients 4, 64, 86 and 190) experienced more than one episode of VL (Table 1). Three of these patients were HIV+ (patients 4, 86 and 190). The isolates derived from patient 4 and 64 (HIV+ and HIV-, respectively), which were collected during a first VL episode (2003 for patient 4 and 2001 for patient 64) and second VL episode (one year later), presented identical MLMT profiles. In both cases, relapse was suspected. Patient 190 (HIV+) presented three episodes of VL, for which the second and third (2001 for both episodes) isolates differed from the first isolate (1996) at only one marker (LIST7021). Re-infection was suspected for this patient; however, further studies are warranted to confirm whether only one allelic change can lead to the suspicion of a re-infection or reflects the evolution of the isolate over time. Patient 86 (HIV+) presented five different episodes of VL due to active chronic VL [20]. Isolates from the second (2000), third (2001) and last episode (2003) presented with the same MLMT profile compared with the isolate from the initial infection (1997). However, the MLMT profile of the fourth episode (2002) isolate varied at 11 loci; only Rossi2 remained the same. Re-infection and relapse were suspected for this patient. Each isolate of these multiple episodes was characterized zymodeme MON-1.
Genetic differentiation among the various populations was tested using FSTAT Version 2.9.3.2.
In France, at the end of the 1990s and early 2000s, the repellent collar had been widely use to protect dogs from parasitic transmission [21,22]. To determine whether this had an impact on genetic differentiation, we compared the isolates before and after the introduction of repellent collar in the AM and P endemic areas. To minimize the temporal effect on genetic variability and determine whether the repellent collar led to a bottleneck effect, the samples collected between 1996 and 2004 were excluded (Table 5). No differentiation was found in both endemic areas, thereby suggesting no bottleneck effect due to repellent collar use.
Nine isolates from asymptomatic carriers were compared with 11 isolates from HIV patients collected between 1994 and 1998 in the same restricted endemic area. To avoid bias, we selected isolates collected two years before and after the date of the collection of the asymptomatic carrier isolates (1996). Unlike the findings reported by Hide et al., no genetic differentiation was observed between isolates from asymptomatic carriers and those derived from HIV+ patients (Table 5) [9]. However, additional isolates from asymptomatic carriers are required to strengthen these findings. A genetic differentiation was observed between isolates from HIV+ and VL adult patients in AM (Table 5).
Bayesian model-based analysis of the 270 isolates using STRUCTURE (with calculation of ΔK) indicated two distinct genetic populations (Fig 5A). Population A consisted of 148 samples: 73 from P, 61 from AM, seven from CO and seven from CE. This population consisted of zymodemes MON-1 and nine isolates with zymodemes other than MON-1. Population B consisted of 109 isolates: 104 from AM, two from P, two from CO and one from CE. Isolates from Population B were all characterized zymodeme MON-1. Among the Populations A and B, 13 isolates had mixed genotypes (11 isolates corresponding to seven genotypes and two isolates corresponding to two genotypes in Populations A and B, respectively). The isolates with mixed genotypes shared allele characteristics of each population. All isolates with mixed genotypes were collected in the AM endemic area and were characterized zymodeme MON-1. Thus, Population A is a mixed population with almost an equivalent number of isolates from AM and P corresponding to MON-1 and all non-MON-1 isolates. The estimated gene flow between isolates from AM and P within Population A (Nm value) was 8.38. Population A also displayed a marked proportion of isolates from HIV+ patients (44.7%) compared with Population B (21.6%) (p<0.05). The two populations A and B defined by STRUCTURE were significantly different as shown by the Fst value 0.503 and p-value equal to 0.05. The estimated gene flow between the populations (Nm value) was 0.25, thereby indicating very few exchanges between those two populations.
The STRUCTURE analysis of Population A separately from Population B, excluding admixed isolates (13 isolates), revealed a delta K graph with two peaks at K = 4 and K = 6 (Fig 5B). The main difference between K = 4 and K = 6 was that sub-population A1’ at K = 4 was split into three populations at K = 6: sub-population A2 (four isolates from sub-population A1’), A3 (16 isolates from sub-population A1’) and A4 (15 isolates from sub-population A1’). Whether at K = 4 or K = 6, the sub-population defined contained isolates from the AM and P endemic areas and from HIV+ patients. The non-MON-1 isolates were grouped into sub-population A1’ and A3’ with K = 4 and into sub-population A2, A4 and A6 with K = 6. No cluster based on clinical data, geographic area or zymodeme profile was found within the defined sub-populations. Some isolates displayed mixed genotypes within the A sub-populations. At K = 4 and K = 6, 26 and 44 isolates had mixed genotypes, respectively. At K = 6, isolates with mixed genotypes came from P (n = 33; 75%), AM (n = 8; 18.2%) and CE (n = 3; 6.8%).
When excluding mixed genotypes, the sub-populations A1, A2, A3, A4, A5 and A6 as well as A1’, A2’, A3’ and A4’ were significantly different, as shown by the significant Fst values ranging from 0.249 to 0.833 and from 0.322 to 0.813 for K = 4 and K = 6, respectively (Table 6). At K = 4, the highest Nm value was obtained between sub-population A1’ and sub-population A3’ (0.53), whereas at K = 6, the highest Nm value was 0.75 between sub-population A2 and sub-population A6, thereby indicating only limited genotype flow between these sub-populations (Table 6).
Two main sub-populations were defined using STRUCTURE for Population B: sub-population B1 and sub-population B2. No mixed genotypes were present. Sub-population B1 consisted of 48 isolates from AM, two from P and one from CO. Sub-population B2 consisted of 56 isolates from AM, one from CO and one from CE. These two sub-populations were genetically different as evidenced by the Fst value (0.537 and p-value = 0.05). Few exchanges occurred between these two sub-populations (Nm = 0.22 was obtained between sub-population B1 and sub-population B2).
Sub-populations A3’ (sub-population A K = 4) and A6 (sub-population A K = 6) displayed the highest number of alleles per population (Table 7). The expected heterozygosity (He), a measure of genetic diversity, was higher in the A sub-populations with non-MON-1 isolates compared with the A sub-populations with only MON-1 isolates (Table 7). All A sub-populations displayed a high inbreeding coefficient (FIS) (> 0.7), whereas the sub-population B2 displayed a low inbreeding coefficient (0.343) (Table 7).
The NJ tree presented in Fig 6 provides a graphic representation of the data. The bootstrap values based on the re-sampling of loci were low and therefore not included in the NJ tree (Fig 6). This was due to the presence of admixed genotypes and the high number of shared alleles even if the allelic frequencies are different between the populations and sub-populations. Two main clusters were found to correspond to the two populations obtained using STRUCTURE at K = 2. Population B formed a separate cluster from Population A. The two sub-clusters defined using STRUCTURE for Populations B1 and B2 are shown on the NJ tree. However, for Population A, the clusters defined by the NJ tree did not perfectly correlate with the A sub-populations defined using STRUCTURE for neither K = 4 nor K = 6. The sub-populations A1, A2 and A3 (A1: four genotypes from AM and seven genotypes from P; A2: one genotype from AM, one from CO and three from P; and A3: eight genotypes from AM, two from P and one from CO) are dispersed throughout the NJ tree. Within Population A, the isolates from AM and P are dispersed throughout the cluster with no correlation with endemic area, clinical form or host background. The mixed genotypes were present between the two main clusters of Populations A and B as well as within the cluster of A sub-populations. Some non-MON-1 isolates grouped together as a paraphyletic group in the sub-population A6, while the others were dispersed among the MON-1 isolates and are present at the end of the branches. As previously described, the MON-108 isolate is closely related to the MON-1 isolates [23–25]. Regarding the nine isolates from AC, two (genotype 21) were present in sub-population A5, whereas seven (genotype 43) belonged to sub-population B1 and grouped with isolates from infant VL and HIV + VL patients. None of the isolates from AC has a mixed genotype. Samples isolated from CanL grouped together with human isolates, and no correlation was found between host and MLMT profile.
Leishmaniasis due to L. infantum is endemic in Southern France. In this study, we used MLMT, a molecular tool useful for population genetic studies, to analyze an extensive set of isolates from four endemic areas in Southern France (AM, P, CE and CO). To the best of our knowledge, this study is the first to investigate a large number of L. infantum isolates from different endemic areas in Southern France. We also focused on the AM and P endemic areas over an extended period of time. A greater number of samples came from AM than P (AM = 178 versus P = 75) because AM is the most active foci in France with the greatest number of leishmaniasis cases per year [2]. The study period was also longer for the AM area than for the P endemic area (AM: 1978–2011; P: 1993–2009). This aspect may generate a sampling bias, although no significant genetic differentiation was found when comparing isolates from AM and P during the same time period. Although MON-1 is the most prevalent zymodeme, other zymodemes also circulate in the south of France [7]. Microsatellite characterization of L. infantum isolates revealed a total of 121 different genotypes. Overall, 91 unique genotypes and 30 repeated genotypes were found. A greater number of repeated genotypes were observed in AM compared with P for the same period, thereby suggesting variations in the transmission cycle between the two areas such as outbreak, vector diversity or density, or host density.
In the AM endemic area, the isolates belonged to two main populations as defined by STRUCTURE: Population A and Population B. Within Population A, gene flow occurred between the AM and P endemic areas. The spread of isolates seems to be from AM to P, as indicated by the results of genotype 21, which was found 15 years later in P.
Substantial genetic diversity was found to be comparable to other endemic areas, even within zymodeme MON-1, thereby confirming previous analyses assessed by other markers [26]. As genetic differentiation depends on the area, our findings suggest strong epidemiological structuring. This is in agreement with the known mechanism of Leishmania transmission and spread in micro-foci and the entomologic data that have demonstrated limited sandfly dispersion [27–31]. Indeed, considering the behavior of the phlebotomine sandfly, it seems less likely that the spread of isolates from AM to P is due to sandfly movement [32]. Unfortunately, due to the small sample size of the phlebotomine sandfly isolates, we cannot investigate the transmission between sandflies, humans and canine hosts in further detail. We suspect that people traveling with their infected dogs between the endemic areas plays a possible role in the etiology of these exchanges. This has been already described for the emergence of L. infantum in South America probably via Conquistadores infected dogs from Portugal [33]. More recently, an intercontinental transportation from France to French Guiana was also reported due to the probable importation of L. infantum from an infected dog [34]. Notably, some repeated genotypes in the AM endemic area are well settled and continue to spread through the area over time. Indeed, some repeated genotypes were detected during a limited period, ranging from 6 to 15 years, and are no longer detected (genotypes 36, 43, 46, 53, 68 and 77). However, other genotypes were still detected in 2011 (genotypes 34, 55 and 57) and one genotype spread to P endemic area (genotype 21). The mixed genotypes between Population A and Population B were isolated in the AM endemic area, whereas in the A sub-populations, 75% of the mixed genotypes came from the P endemic area. We also observed a predominance of isolates from HIV+ patients in Population A (44.7%) compared with Population B (21.6%), which may indicate a variation in virulence. Indeed, these isolates from HIV+ patients may produce leishmaniasis in immunocompromised patients, whereas affected immunocompetent patients may develop only an asymptomatic infection [35]. This hypothesis is in agreement with the distribution of isolates from AC in both Population A and Population B. These isolates belonged to the genotypes 21 (2 AC) and 43 (7 AC), with samples from HIV+ VL, IVL, VL and CanL cases and samples from infant VL and HIV+ VL patients, respectively. The comparison of isolates from AC and those from HIV+ patients isolated during the same period and within the same restricted endemic area revealed no genetic differentiation between these populations. This finding contrasts with previous data reported by Hide et al. [9] and as suspected, does not reflect a difference in virulence. However, further isolates from asymptomatic carriers must be assessed to confirm our hypotheses. This is critical in the endemic areas of Southern France (as well as all endemic foci of leishmaniasis), where the isolates responsible for leishmaniasis represent only the tip of the iceberg [36]. Indeed, depending on the test used to detect asymptomatic carriage, prevalence varies from 30% to 46.8% in the AM endemic area [36]. Although our study provides important insight into leishmaniasis epidemiology in AM and P, our panel represents only a small proportion of the L. infantum population circulating in Southern France as samples from asymptomatic carriers, dogs and sandflies are underrepresented.
Microsatellite analyses may be useful to estimate relapse and re-infection rates, which is important to evaluate anti-Leishmania drug efficacy and transmission dynamics, respectively. This aspect is particularly important for the follow-up of patient treatment. MLMT may also be a useful tool to differentiate between relapses from re-infection cases [11,25,26,35,37,38]. Moreover, Bourgeois et al. have described “active chronic visceral leishmaniasis” in patients with several episodes of VL [20]. In this particular form of the disease, identifying the MLMT profile of each isolate responsible of each episode of VL could be useful to monitor and optimize treatment regimes. In our study, we detected probable treatment failure in HIV+ and non-HIV patients, as the MLMT profiles were indistinguishable from one episode to another. Certain patients likely experienced re-infection, as isolates from two different episodes of leishmaniasis in same patient displayed different MLMT profiles. However, we cannot exclude the possibility of a mixed infection with differential strain isolation depending on the time of sampling. Due to the small number of patients with isolates from several biological samples at different times, the rate of relapse and re-infection needs to be confirmed on a larger sample set. Thus, the results on relapse and re-infection should be interpreted with caution. Further investigations are required to assess these hypotheses in further detail.
The high ratio of repeated genotypes in HIV patients (81.6%) compared with the remaining population (41%) may be due to an outbreak amongst this fragile human population. Outbreaks have already been reported among intravenous drug users, a population also frequently affected by HIV infection [39,40]. Nevertheless, we have no information concerning this aspect of the case population in our study.
Faucher et al. have highlighted the heterogeneity of environments associated with VL transmission in Southeastern France [5]. The authors showed two distinct foci strongly associated with specific environments. One focus, corresponding to the AM endemic area, was characterized by scattered habitation and mixed forest in the foothills. In contrast, the other focus in the P endemic area was centered in urban areas of Marseille. These environmental differences correlate with the strong genetic differentiation we found between the Leishmania populations from AM and P. Indeed, the ecosystem influences the transmission cycle and thus the population dynamics of parasites. Moreover, in the P endemic area, Toscana virus, which is responsible for summer meningitis, and L. infantum share the same vector, Phlebotomus perniciosus. A recent study has described dogs co-infected by these two organisms [41]. Although cases of co-infection in humans or vectors have not been reported, we suspect that is also possible. This phenomenon of co-infection may have an impact on Leishmania transmission and should be addressed in future studies to understand whether this may also influence parasite evolution.
Other wild reservoirs of L. infantum have been demonstrated in Europe such as fox, rats and hare [42]. These wild reservoirs are able to transmit L. infantum to sandflies. However, the isolates from wild reservoirs have indistinguishable genotypes from those derived from domestic dogs and humans [26,43]. No isolates from wild animals were included in our study. The only uncommon host included in our study was a cat, and the isolate from this animal shared the genotype 55 with isolates from dog, IVL, HIV+ VL and VL samples.
In some studies, the isolates with a zymodeme other than MON-1 grouped together either via neighbor joining tree or STRUCTURE analysis [25,26]. We did not find such correlations with our data which is probably due to the high number of strains with mixed genotypes. Indeed the nine non-MON-1 isolates did not group into a separate population but rather clustered into Population A with a majority of the MON-1 isolates. In the NJ tree, some non-MON-1 isolates appeared as a paraphyletic group, while others were either isolated or dispersed among other zymodemes [23–25]. The zymodeme MON-108 (genotype 92) isolate appeared very close to MON-1 isolates with the genotype 37 [23–25].
In our study, no correlation was found between MLMT profile, clinical expression of the disease, immune status and host. Finally, MLMT is more discriminant and thus more appropriate than MLEE to evaluate epidemiological changes among parasite population in Southern France. MLMT data provided a better understanding of gene flow between L. infantum populations within the Southeastern France endemic area.
|
10.1371/journal.pbio.3000106 | A census-based estimate of Earth's bacterial and archaeal diversity | The global diversity of Bacteria and Archaea, the most ancient and most widespread forms of life on Earth, is a subject of intense controversy. This controversy stems largely from the fact that existing estimates are entirely based on theoretical models or extrapolations from small and biased data sets. Here, in an attempt to census the bulk of Earth's bacterial and archaeal ("prokaryotic") clades and to estimate their overall global richness, we analyzed over 1.7 billion 16S ribosomal RNA amplicon sequences in the V4 hypervariable region obtained from 492 studies worldwide, covering a multitude of environments and using multiple alternative primers. From this data set, we recovered 739,880 prokaryotic operational taxonomic units (OTUs, 16S-V4 gene clusters at 97% similarity), a commonly used measure of microbial richness. Using several statistical approaches, we estimate that there exist globally about 0.8–1.6 million prokaryotic OTUs, of which we recovered somewhere between 47%–96%, representing >99.98% of prokaryotic cells. Consistent with this conclusion, our data set independently "recaptured" 91%–93% of 16S sequences from multiple previous global surveys, including PCR-independent metagenomic surveys. The distribution of relative OTU abundances is consistent with a log-normal model commonly observed in larger organisms; the total number of OTUs predicted by this model is also consistent with our global richness estimates. By combining our estimates with the ratio of full-length versus partial-length (V4) sequence diversity in the SILVA sequence database, we further estimate that there exist about 2.2–4.3 million full-length OTUs worldwide. When restricting our analysis to the Americas, while controlling for the number of studies, we obtain similar richness estimates as for the global data set, suggesting that most OTUs are globally distributed. Qualitatively similar results are also obtained for other 16S similarity thresholds (90%, 95%, and 99%). Our estimates constrain the extent of a poorly quantified rare microbial biosphere and refute recent predictions that there exist trillions of prokaryotic OTUs.
| The global diversity of Bacteria and Archaea ("prokaryotes"), the most ancient and most widespread forms of life on Earth, is subject to high uncertainty. Here, to estimate the global diversity of prokaryotes, we analyzed a large number of 16S ribosomal RNA gene sequences, found in all prokaryotes and commonly used to catalogue prokaryotic diversity. Sequences were obtained from a multitude of environments across thousands of geographic locations worldwide. From this data set, we recovered 739,880 prokaryotic operational taxonomic units (OTUs), i.e., 16S gene clusters sharing 97% similarity, roughly corresponding to prokaryotic species. Using several statistical approaches and through comparison with existing databases and previous independent surveys, we estimate that there exist globally between 0.8 and 1.6 million prokaryotic OTUs. When restricting our analysis to the Americas, while controlling for the number of studies, we obtain similar estimates as for the global data set, suggesting that most OTUs are not restricted to a single continent but are instead globally distributed. Our estimates constrain the extent of a commonly hypothesized but poorly quantified rare prokaryotic biosphere and refute recent predictions that there exists trillions of prokaryotic OTUs. Our findings also indicate that, contrary to common speculation, extinctions may strongly influence global prokaryotic diversity.
| Microorganisms are the most ancient and the most widespread form of life on Earth, inhabiting virtually every ecosystem and driving the bulk of global biogeochemical cycles. Culture-independent methods such as amplicon sequencing of 16S ribosomal RNA genes revealed the existence of a potentially vast undescribed microbial diversity, the full extent of which, however, remains highly controversial [1–9]. Determining the extent of this diversity remains an important but challenging task in our overall understanding of life, with major implications for ecological and evolutionary theory, environmental sciences and industry. Notably, a global census of microbial phylogenetic diversity, or at least knowledge of its full extent, is essential for reconstructing microbial evolution over geological time [10]. Estimates of global microbial diversity are also needed for scrutinizing proposed biodiversity scaling laws and macroecological theories [2,6,11]. Finally, undiscovered microorganisms may exhibit a large breadth of metabolic capabilities of particular interest to industry and medicine. An efficient exploration of this potential and realistic assessment of the feasibility of such an endeavor requires knowledge of the gaps in existing diversity databases [12–14].
The extent of global microbial diversity remains subject to intense controversy and widely diverging speculations [1–9]. This controversy stems largely from the fact that existing estimates are either based on extrapolations of empirical scaling laws [6], on theoretical biodiversity models [2], on data sets covering only a small fraction of global diversity [1,9], or on taxonomically biased databases, including mostly organisms that have been cultured or are of particular medical/industrial interest [3–5]. For example, Mora and colleagues [3] used the subset of currently named prokaryotic species to estimate that there exist approximately 10,000 bacterial species worldwide; this is clearly a strong underestimate, given that the SILVA sequence database [14] alone now contains hundreds of thousands of bacterial operational taxonomic units (OTUs), i.e., clusters of the 16S gene at 97% similarity—a traditional microbial species analog. Yarza and colleagues [4] and Schloss and colleagues [5] estimated that there exist a few million bacterial and archaeal ("prokaryotic") OTUs based on sequence discovery statistics in SILVA; however, environmental and taxonomic biases in SILVA [15] compromise the reliability of these estimates [16]. Larsen and colleagues [9] estimated that there exist billions of host-associated bacterial OTUs based on a heuristic and mathematically flawed extrapolation of bacterial OTU counts in typical insect species to all animal species (see the "Implications" section below for a detailed discussion). Locey and colleagues [6] even predicted that there exist trillions of microbial OTUs (at 97% similarity) based on an extrapolation of empirical scaling laws of local diversity in individual communities to global scales. Locey's estimate has fueled discussions about a potentially immense undiscovered microbial diversity and its uncertain ecological roles [16–20]. Locey's extrapolation of empirical scaling laws from local to global scales and across several orders of magnitude has been criticized and remains controversial [8,21].
Here, to address the above shortcomings, we attempted to explicitly census a large fraction of extant prokaryotic clades and used our census to estimate and chart total global prokaryotic OTU richness. For this census, we compiled massive publicly available raw Illumina 16S amplicon sequencing data from 34,368 samples across 492 studies, covering a wide range of environments from over 2,800 distinct geographical locations worldwide (S1 Fig). Environments covered include the surface and deep ocean, oxygen minimum zones, freshwater and hypersaline lakes, rivers, groundwater, marine surface and deep subsurface sediments, agricultural and forest soils, peats, permafrost, deserts, animal hosts and feces, plant leafs and rhizospheres, salt marshes, bioreactors, processed food, methane seeps, mine drainages, sewages, hydrothermal vents, and hot springs (overview in S1 Data). Particular effort was put into representing soils (14,242 samples across 100 studies), sediments (3,198 samples across 37 studies), and animal guts (8,646 samples across 52 studies), which likely harbor a large fraction of Earth's prokaryotic diversity [22]. Sequences in this composite data set cover at least 200 basepairs in the V4 hypervariable region of the 16S gene, a commonly targeted region in microbial ecology [22–24]. By clustering the pooled sequences at 97% similarity, a commonly used threshold in microbial ecology [2,5,6,9], we recovered hundreds of thousands of OTUs. Based on the recovered OTUs, henceforth referred to as Global Prokaryotic Census (GPC), and through comparisons to previous surveys and existing databases, we estimate global prokaryotic OTU richness and highlight major implications for microbial ecology and evolution. We emphasize that our main objective was to estimate global prokaryotic richness using as deep of a census and covering as many environments and geographic locations as possible; as a trade off, our data set does not offer the same level of experimental standardization across samples nor the amount of metadata included in projects such as the Earth Microbiome Project (EMP) [22].
Here, we focus on OTUs clustered using the conventional 97% similarity threshold so as to facilitate comparison with existing prokaryotic richness estimates [2,5,6,9]. Recent work, however, suggests that a greater similarity threshold (approximately 99%) is often required for distinguishing ecologically differentiated organisms [25–28]. We thus also repeated our analyses using a 99% clustering threshold, which yielded qualitatively comparable results. That said, we point out that clusters of the 16S gene—regardless of similarity threshold and even if completely free of sequencing errors—only provide an approximate "species" analog to sexually reproducing organisms. Indeed, even strains with identical 16S sequences may exhibit different genomic content and ecological strategies; hence, the 16S gene is not always sufficient for distinguishing ecologically differentiated organisms, even when considering exact sequence variants [29–30]. Whether and how prokaryotic "species" can—or even need to—ever be reasonably defined remains highly debated [30–33]. To date, the 16S gene remains an important and the most popular marker for cataloguing prokaryotic diversity and for describing evolutionary relationships in a well-defined and reproducible manner [4,27]. We stress that prokaryotic 16S diversity detected and estimated based on amplicon sequences, as in this and most previous studies, is limited to clades detectable by the PCR primers used. As discussed below, the GPC partly resolves the issue of limited primer scope by using multiple alternative primers; however, it is in principle still possible that some clades are completely missed.
To ensure maximal phylogenetic coverage, the raw sequencing data from each study was considered as input to our analyses. After stringent quality- and chimera-filtering, the data set comprised 1,734,042,763 high-quality reads, which were pooled and clustered into OTUs at 97% similarity. To avoid spurious (i.e., nonbiological) OTUs generated by sequencing errors or PCR chimeras, only OTUs found in at least two samples of the same study were kept. While this additional quality filter may also remove some biological OTUs, aggressive filtering is necessary for eliminating spurious OTUs, a common and serious problem in amplicon sequencing studies [34–37]. The resulting GPC comprises 739,880 prokaryotic OTUs (690,474 bacterial and 49,406 archaeal), accounting for 1,349,766,275 reads. Accumulation curves of bacterial and archaeal OTUs discovered by the GPC, as a function of studies included, clearly show a deceleration with increasing number of studies (Fig 1A and 1B) and provide an estimate of how many novel OTUs would be discovered in subsequent studies. Specifically, on average, about 93% of bacterial OTUs and 83% of archaeal OTUs found in any additional study are expected to be already included in the GPC. As we show below, this estimate is consistent with the fractions of other independent data sets and databases covered (rediscovered) by the GPC. Most OTUs were matched by at least three reads (88%) and most were found in at least three samples (81%, S2 Fig). Based on the fraction of reads matched to the rarest OTUs (i.e., with only two reads), we estimate that any new random 16S amplicon sequence (i.e., from a randomly chosen prokaryotic cell) would hit an OTU in the GPC at 97% similarity with a probability ≥99.98% (using the Good–Turing frequency formula [38]; see Methods for details). This probability is sometimes referred to as "Good's coverage" and corresponds to the proportion of living or recently deceased prokaryotic cells, detectable by current 16S amplicon sequencing techniques, which is represented by OTUs in the GPC. We emphasize that Good's coverage should not be interpreted as the fraction of global OTU richness represented by the GPC; indeed, estimation of the latter requires additional statistical reasoning, as presented below.
To estimate the total number of extant prokaryotic OTUs globally (discovered plus undiscovered), we used statistical approaches based on the number of OTUs that have been discovered in exactly one study (Q1), the number of OTUs discovered in exactly two studies (Q2), and so on. Indeed, the recommended (and only statistically admissible) way to estimate OTU richness is by modeling the incidence frequency counts Qi in order to predict the number of unobserved OTUs Q0 [21,39–41]. These methods date back to mathematical theorems for cryptographic analyses during World War II and have been used for microbial as well as macrobial richness estimates [40, 42–44]. Intuitively, widely distributed and abundant OTUs—which are almost certain to be detected—contain very little information about undetected OTUs, while rarely detected OTUs (e.g., detected only once or twice) carry the most information about undetected OTUs; hence, estimators typically rely on the low-frequency counts Q1, Q2, etc. [40]. To ensure the robustness of our estimations, we considered several alternative estimation methods, each of which is based on a different frequency model and relies on different assumptions: the improved-Chao2 ("iChao2") richness estimator [45], based on the frequency counts Q1–Q4; the incidence coverage-based estimator (ICE) [41], based on the frequency counts Q1–Q10; the CatchAll estimator [46], based on frequency counts Q1–Qτ, in which τ is chosen adaptively based on internal quality criteria; the transformed weighted linear regression model (tWLRM), which uses a linear regression model for the ratios of consecutive log-transformed frequency counts to predict Q0 [46,47]; and the breakaway estimator [48], based on a nonlinear regression model for the ratios of consecutive frequency counts. All of the above estimators have been designed to account for heterogeneities in detection frequencies among OTUs (i.e., the presence of rare and frequent OTUs), and breakaway is particularly optimized for efficiently dealing with high fractions of undiscovered diversity. We note that the majority of existing richness estimators, including the ones described above, are based on models in which individual sampling units are assumed to be equivalent (e.g., of the same "effort"); however, studies included in the GPC differ in terms of the environment sampled and the techniques used. To check whether our estimates are sensitive to this caveat, we also deployed an estimation approach whereby we randomly assigned studies to four complementary and equally sized groups (representing four statistically equivalent global "sampling units") and used the iChao2 estimator based on the number of OTUs found in exactly one, two, three of four sampling units ("iChao2split," illustration in Fig 1E). All of the above methods yielded comparable estimates for global prokaryotic OTU richness, with the lowest estimate obtained using tWLRM (901,902 OTUs), and the highest estimate obtained using breakaway (1,588,567 OTUs). The majority of prokaryotic OTUs are estimated to be bacterial, with bacterial richness (Fig 1C) being roughly 10 times greater than archaeal richness (Fig 1D). Importantly, all of the above estimates suggest that there only exist in the order of approximately 1–2 million prokaryotic OTUs, a substantial portion of which is represented by the GPC (47%–82%). We point out that even at a finer phylogenetic resolution (99% clustering similarity), we estimate that there exist only approximately 3–9 million prokaryotic clusters worldwide (S4 Fig and S1 Table), which is six orders of magnitude lower than estimated previously via extrapolation of empirical scaling laws [6].
To further scrutinize our estimates of global OTU richness and to verify whether a substantial fraction of that richness is indeed covered by the GPC, we determined the fraction of 16S sequences from previous global surveys or existing databases that was rediscovered ("recaptured") by the GPC. We found that at 97% similarity, the GPC recaptured 96% of prokaryotic sequences in the SILVA database (nonredundant set, release 132) [14], 89% of prokaryotic sequences in the Ribosomal Database Project (RDP release 11) [12], and 93% of prokaryotic sequences in the Genome Taxonomic Database (GTDB release 86.1) [49] (domain-specific coverages in S2 and S3 Tables). Using these coverages as a proxy for the fraction of global OTU richness covered by the GPC and combining this coverage fraction with the total number of OTUs in the GPC yields additional independent estimates of global prokaryotic OTU richness (771,234–832,420 OTUs, Fig 1F), roughly consistent with our previous estimates. We also found that at 97% similarity, the GPC recaptured 92% of unique noise-filtered ("deblurred") 16S amplicon sequences from another recent independent massive global survey, the EMP [22]. The high fraction of EMP sequences recaptured by the GPC further supports our conclusion that the GPC covers a substantial portion of extant prokaryotic OTUs.
While our statistical richness estimators (Fig 1C and 1D) were designed to account for variable detection probabilities among OTUs, the potential risk of neglecting a large number of extremely rare OTUs cannot be overemphasized. To further assess this risk, we also explicitly investigated the global distribution of relative OTU abundances. Specifically, for each OTU, we estimated its relative abundance in each sample (using the Good–Turing formula) [38] and then took the average across all samples to obtain its mean relative abundance (MRA). We then created a frequency histogram of MRAs by grouping OTUs into equally sized MRA intervals on a logarithmic axis. We note that this empirical histogram only includes OTUs discovered by the GPC and may thus be skewed toward more abundant OTUs. We therefore reconstructed the total number of extant OTUs in each MRA interval (blue continuous curve in Fig 2A) using a probabilistic model of OTU discovery. This model accounted for our quality filtering and finite sequencing depths and was calibrated by comparing OTU discovery rates in the GPC with those in a rarefied variant of the GPC (i.e., using only half of the original sequences). Following recommendations by Shoemaker and colleagues [11], we then fitted a log-normal model to the reconstructed distribution of MRAs of extant OTUs. We found that the latter was well described by the log-normal model (blue dashed curve in Fig 2A), resembling analogous observations commonly made for larger organisms. We point out that the log-normal model is largely phenomenological, although it is sometimes derived from certain stochastic population models [50]. Hence, we make no assertion as to which mechanisms could possibly lead to the observed log-normal–like distribution of MRAs and as to whether other (potentially yet to be discovered) models may be even more suitable. Based on the reconstructed distribution of MRAs, as well as based on the fitted log-normal model, we estimate that the majority of extant OTUs exhibit an MRA across samples between 5 ×10-10 and 5 ×10-8 (mode approximately 5 ×10-9). For lower MRAs, the number of OTUs declines rapidly toward zero. The rapid decline of the number of OTUs for lower MRAs suggests that the number of much more rare OTUs (specifically, with an MRA lower than the OTUs detected by the GPC) is relatively small and that the GPC did not miss vast numbers of extremely rare OTUs. This conclusion contrasts previous speculations that there exists a vast number of extremely rare and largely undetected OTUs, sometimes referred to as "rare microbial biosphere" [6,17,51]. According to the fitted log-normal model, there exist only approximately 886,291 prokaryotic OTUs across the entire range of MRAs, further supporting our other estimates.
Since OTUs are inevitably taxonomically identified through comparison with reference databases (here, SILVA was used to identify OTUs at the kingdom level), censuses such as the GPC may in principle miss clades lacking a close relative in the databases. To investigate this potential caveat, we calculated the phylogenetic distance of each OTU to its closest match in SILVA in terms of 16S sequence divergence and created a frequency histogram of these distances that shows the overall distribution of OTUs in comparison to SILVA (Fig 2B). We found that the vast majority of OTUs in the GPC has a distance to SILVA that is far below the threshold allowed for taxonomic identification (maximum 40%) and that the frequency of OTUs drops rapidly toward that threshold. This suggests that our taxonomic identification algorithm did not miss a substantial number of biological sequences at larger phylogenetic distances (omitted sequences at greater distances are likely spurious, see Methods for details).
Primer "blind spots," i.e., clades not captured by PCR primers, could in principle lead to an underestimation or a phylogenetically biased assessment of prokaryotic diversity by the GPC. For example, recent studies suggest that roughly 10% of prokaryotic 16S sequences may be missed by any given existing primer pair [52–54]. To investigate this caveat and to check whether a large fraction of diversity may have been missed by the GPC due to primer blind spots, we calculated the fraction of 16S sequences recovered from a multitude of environments using primer-independent (metagenomics-based) methods that were rediscovered by the GPC. We found that, at 97% similarity, the GPC recaptured 91% of 16S sequences in prokaryotic genomes previously assembled from metagenomes (Uncultivated Bacteria or Archaea [UBA]) [55] and 93% of bacterial 16S sequences extracted from thousands of public metagenomes [56]. These recapture fractions are comparable to the fraction recovered from the EMP, suggesting that the fraction of OTUs missed by the GPC due to primer blind spots is small. One reason may be that the GPC comprises sequences obtained using a multitude of alternative primers optimized for different clades, therefore partly alleviating the problem of primer nonuniversality. In particular, 16S sequences currently not detectable by any primers may only represent a minority of prokaryotic diversity, even if any given primer set has limited sensitivity scope. It is thus improbable that primer-independent methods will reveal a prokaryotic richness much (i.e., orders of magnitude) higher than composite multiprimer-based surveys such as the GPC.
When we repeated our analyses using only studies from the Americas or near American coasts (165 studies across 14 countries, see map in S1 Fig) instead of the full GPC, OTU discovery rates for any given number of studies remained almost unchanged (Fig 1A and 1B). Hence, for the same "sampling effort," the same OTU richness is recovered from the Americas as from the full GPC, and importantly, the restriction to the Americas does not cause a stronger deceleration of OTU discovery rates. This suggests that the majority of global prokaryotic OTUs could have been censused from a single hemisphere, if sufficient samples had been available. Consistent with this conclusion, when controlling for the number of studies included and using the same methods as above, we found that prokaryotic OTU richness estimated for the Americas was very similar to estimates based on an equal number of studies randomly chosen from across the world (0.7–1.3 million OTUs, S5A Fig). Similar results were also obtained at a higher 16S similarity threshold of 99% (S4A and S4B and S5B Figs). Our findings extend previous observations that for any given number of samples, similar prokaryotic OTU richness is recovered from soil in New York Central Park as from distinct soil samples worldwide [57]. Most prokaryotic OTUs thus appear to exhibit low geographic endemism and global dispersal ranges at geological time scales, i.e., at time scales needed for 16S to diverge by more than 1% [58,59]. A global distribution of prokaryotic OTUs has long been a central but controversial hypothesis [60,61]. Our finding provides strong support for this hypothesis and is also consistent with previous findings that most marine bacterial OTUs can be recovered from a single location in the ocean with sufficiently deep sequencing [62,63] and with findings that salt-marsh Nitrosomonadales OTUs are globally distributed [64]. That said, we point out that a global distribution of OTUs does not rule out geographic endemism at finer phylogenetic resolutions since younger clades, e.g., recently differentiated ecotypes with identical 16S, may not have had time to overcome dispersal barriers at global scales [65,66].
Our census allows an unprecedentedly precise assessment of the diversity covered by existing 16S databases such as SILVA [14] or the RDP [12]. Based on the fraction of GPC OTUs matched to entries in SILVA (release 132, nonredundant set) at 97% similarity, we estimate that SILVA represents about 29% of bacterial and 14% of archaeal OTUs globally. Similarly, we find that the RDP (release 11) represents about 42% of bacterial and 20% of archaeal OTUs. Our findings confirm recent estimates that SILVA covers about 30%–40% of global prokaryotic OTU richness [5,67]. We point out that Bacteria are currently overrepresented in SILVA and the RDP relative to Archaea. The uneven representation of various taxonomic groups is generally more pronounced at lower taxonomic levels, with some phyla being strongly overrepresented compared to others (Fig 3B). In addition, about 7% of prokaryotic OTUs in the GPC could not be reliably assigned to any phylum listed in SILVA. This indicates that some phyla are not represented in SILVA at all, consistent with conclusions from metagenomic studies [56,68].
Our estimates also highlight strong differences in the OTU richness specific to different phyla, with Proteobacteria (mostly Gammaproteobacteria and Deltaproteobacteria) clearly dominating global richness, followed by the Firmicutes (mostly Clostridia), Bacteroidetes (mostly Bacteroidia), Nanoarchaeota (mostly Woesearchaeia), Patescibacteria, and Planctomycetes (mostly Planctomycetacia) (Figs 3A and S11A). Hence, the large representation of Proteobacteria in reference databases and among cultured species [69] is not just the result of a biased discovery rate (e.g., due to ease of culturing) but also partly reflects their general ability to expand and persist in a multitude of ecological niches [70]. Similarly, the large richness of Firmicutes may be explained by their ability to colonize a wide range of animal hosts [56]. Interestingly, the Nanoarchaeota are known as a deeply branching and poorly characterized ancient clade [71], which has been suggested to comprise a largely underestimated diversity [72]. The few isolated Nanoarchaeota indicate that they share a common history of adaptation to ectosymbiosis [73], and this may have contributed to the difficulty of isolating representatives. In contrast, while the Actinobacteria phylum contains the second largest number of cultured strains [69,74], it only ranks eigth in terms of estimated total OTU richness (Fig 3A), suggesting a strong culturing bias for this phylum, consistent with previous findings [69]. We point out that extant prokaryotic diversity is the result of diversification and extinction processes operating over billions of years and throughout geological transitions [15]. It is thus possible that the relative richness of various taxa varied strongly over time.
Our work suggests that global prokaryotic OTU richness is about six orders of magnitude lower than previously predicted via extrapolation of diversity scaling laws and OTU abundance distributions fitted to individual microbial communities [6,8]. While we find support for a log-normal distribution of mean relative OTU abundances consistent with assumptions made by Locey and colleagues [6], at least two aspects differentiate our approach from Locey and colleagues. First, we fitted the log-normal model to a global data set comprising thousands of samples across hundreds of environments rather than to individual local communities, thus obtaining a description of relative abundances that is more suitable for global richness estimates. Second, we did not assume or extrapolate any phenomenological scaling relationships between different parameters of the model, thus relying on fewer questionable assumptions. The discrepancy between our estimates and those by Locey and colleagues [6] suggests that phenomenological scaling relationships of microbial diversity cannot be extrapolated to global scales when these relationships were fitted solely to individual communities. This conclusion also supports arguments by [21] that the extrapolations performed by Locey and colleagues [6] have no predictive power and are statistically unsound. Our estimates also contrast extrapolations by Larsen and colleagues [9], who argued that there exist billions of animal-associated bacterial OTUs based on the number of OTUs typically found in individual insect species and the estimated total number of animal species. One reason for this discrepancy may be that Larsen's extrapolation did not properly account for the overlap of microbiomes between animal taxa (detailed discussion in S1 Text). Our much lower bacterial richness estimates suggest that many symbiotic OTUs are found in multiple host species that may or may not be closely related, potentially due to host trait convergences, consistent with recent observations [75–77]. Since the microbiome of only a minuscule fraction of animal species has been examined so far, it is quite possible that many allegedly "host-specific" bacteria are shared by a broader spectrum of host species than currently known. This could explain why overall bacterial richness (at the OTU level) appears to have been largely unaffected by past mass animal extinctions, as recently suggested based on phylogenetic analyses [15].
Given the long evolutionary history and ubiquity of prokaryotes, a richness of only approximately 0.8–1.6 million OTUs may seem surprisingly low. To put this finding into perspective, we considered a steady state null model, in which global prokaryotic cell counts (N) are constant over time, in which cells are replaced randomly and regardless of phylogenetic relationships via births and deaths, and in which the 16S-V4 region evolves neutrally [59] at some constant drift rate (r, measured in mutations per site per generation) and independently at each site. Note that one important and potentially wrong assumption of this model is that cell turnover is statistically independent of phylogeny. A similar model was recently proposed by Straub and colleagues [78] as a null model for 16S phylogenies. Based on our model, we predict that there should exist about 2rN/0.03∼1022-1023 OTUs (assuming N = 1030 [79] and r=4×10-9-5×10-10 [80,81], details in S2 Text). This extreme discrepancy between the model and our global richness estimates persists regardless of the similarity threshold used (97% or 99%). The discrepancy also persists even if currently estimated 16S mutation rates (r) or global cell counts (N) were off by 10 orders of magnitude or even if global cell counts varied drastically (e.g., by 1–10 orders of magnitude) over recent time. One explanation for this discrepancy could be that the evolution of the 16S-V4 region along a lineage is subject to strong constraints that favor some mutations or sequence variants more than others, thus effectively reducing the "permissible" sequence space [82–84]. This would suggest that only about 10−14% of the theoretically possible 16S variants are actually biologically viable and attainable over the course of approximately 4 billion years [85]. Alternatively, some processes not captured by the model may eliminate all but just a small fraction of 16S sequence variants emerging over time. Phylogenetically correlated turnover, i.e., closely related organisms experiencing birth or death simultaneously more frequently than expected by chance (e.g., due to their greater ecological similarity), would lead to increased removal of sequence variants from the pool compared to the above null model and may also be an explanation for the relatively sparse filling of 16S sequence space found here. This would imply that extinction plays a central role in prokaryotic diversification, as recently suggested by [15] and contrasting common speculations that prokaryotic OTUs are unlikely to go extinct [1,86–88].
We emphasize that our results are specific to the similarity threshold used (97% similarity in 16S) and the gene region targeted (V4), although these choices are a popular combination in microbial ecology [22]. For example, at coarser phylogenetic resolutions (e.g., 95% and 90% similarity, roughly corresponding to genera and families [89,90]), we estimate that there exist substantially fewer 16S clusters and that the GPC covers a greater fraction of those clusters (50%–98% and 51%–99.5% of extant clusters, respectively, S1 Table). Consistent with these estimates, we found that at these coarser resolutions the GPC recaptured 95%–96% and 98%, respectively, of previous global 16S surveys (S3B, S3C, S8 and S9 Figs and S4 and S5 Tables). Reciprocally, when we analyzed a subset of our data (approximately 0.2 billion reads across 111 studies) at the finest possible phylogenetic resolution (100% identity) using the recent Divisive Amplicon Denoising Algorithm (DADA2) software [91], we obtained about 3.4 times as many exact amplicon sequence variants (ASVs) as 97%-OTUs and about 1.5 times as many ASVs as 99%-OTUs (S13 Fig). This suggests that the global richness of exact sequence variants is at most an order of magnitude larger than the number of OTUs. The sequence length considered may also affect global richness measures. For example, full-length 16S diversity (currently much harder to census) is expected to be greater than partial-length (V4) 16S diversity [4] because short gene regions may cluster as one OTU due to the stochasticity of mutations even if the full genes differ by more than 3%. For example, when restricted to the V4 region or when considering the full 16S gene, at 97% similarity, 16S sequences in SILVA cluster into 102,416 or 270,788 OTUs, respectively, suggesting that the number of extant full-length OTUs may exceed the number of V4 OTUs by a factor of approximately 2.7. When combined with our V4-based richness estimates, this suggests that there exist 2.2–4.3 million full-length OTUs worldwide. A similar ratio between full-length and partial-length clusters is also obtained at 99% similarity (S6 Table). Unfortunately, while full-length sequencing undoubtedly improves phylogenetic resolution, technical complications and a higher cost currently prevent the wide adoption of full-length 16S sequencing in microbial community surveys. Finally, we stress that 16S diversity only provides a coarse surrogate for prokaryotic genomic and phenotypic diversity [29,30], and it is probable that the global number of prokaryote ecotypes greatly exceeds the number of OTUs. Cataloguing the phenotypic and genomic diversity of prokaryotes will undoubtedly be an important but much more challenging future task.
In 2002, Curtis and colleagues [2] hypothesized that experimental approaches to directly enumerating extant prokaryotic diversity will remain fruitless due to logistical challenges. Almost two decades later, we demonstrated that publicly available sequencing data from 492 studies around the world are sufficient to recover a substantial fraction (47%–96%) of global prokaryotic diversity in the 16S-V4 region, the very extent of which has long been a topic of speculation [1,2,4–7]. Our composite data set, covering a multitude of environments worldwide, enabled us to strongly constrain global prokaryotic OTU richness. Indeed, our global richness estimates are similar across a multitude of statistical estimators (Fig 1C and 1D), all of which are based on different models of OTU detection probabilities and, in most cases, use a different set of OTU incidence frequency counts. The high fraction of 16S sequences from other amplicon- and metagenomic-sequencing surveys (e.g., the EMP [22] or UBA [55]) and large databases (e.g., SILVA [14] and RDP [12]), recaptured independently by the GPC (91%–93%), further supports our global prokaryotic richness estimates and our assessment that the GPC covers a substantial portion of that richness.
While no particular 16S similarity threshold provides an ideal species analog, OTUs provide an operational and clearly defined measure of richness that can be compared across studies, environments, and geological time [15]. For example, our work revealed that global prokaryotic OTU richness is orders of magnitude lower than often predicted [1,6,9], regardless of the considered similarity threshold (97% and 99%). Further, the fact that our global richness estimates are approximately 16–17 orders of magnitude lower than predicted by a null model for neutral OTU emergence, regardless of the similarity threshold used, suggests that extinction played a major role in prokaryotic evolution [15] and/or that the attainable 16S-V4 sequence space is extremely constrained. Our work also showed that at the phylogenetic resolutions considered here (≥1% divergence in 16S), most prokaryotic OTUs are globally distributed, yielding insight into the time scales involved in global-scale microbial dispersal.
We reiterate that the goal of the GPC was to enable a more robust estimate of total extant prokaryotic richness than previous studies. Indeed, our estimates are based on an unprecedentedly large and environmentally broad composite sequencing data set, assembled from hundreds of studies utilizing alternative primers and alternative sampling techniques, and using a wide array of alternative statistical estimation methods for increased robustness. The GPC can thus facilitate future efforts to catalogue and phenotypically describe Earth's extant prokaryotes. The GPC also opens up new avenues for reconstructing prokaryotic evolution over geological time using massive phylogenetic trees and for refining macroecological theories. While long considered an unseen majority [79], thanks to ongoing technological revolutions, prokaryotes could one day become one of the most exhaustively characterized and best understood forms of life.
Publicly available 16S rRNA amplicon sequences (V4 region) from various environmental and clinical studies were downloaded from the European Nucleotide Archive (https://www.ebi.ac.uk/ena). Only Illumina sequences were downloaded to ensure sequence qualities en par with current standards and because Illumina-based studies typically achieve much deeper sequencing than studies using previous-generation (e.g., 454) technology. We only considered sequences covering the V4 hypervariable region for three reasons. First, use of the same gene region in all samples is necessary for clustering sequences into nonredundant OTUs. Second, the V4 region is one of the most popular regions targeted in microbial surveys, including the EMP [22], making it easier to find publicly available data sets and allowing for comparison with the EMP. Third, the V4 region was shown to be the most suitable single hypervariable region for reconstructing bacterial phylogenetic relationships [24]. Studies were chosen to represent as wide of an environmental spectrum as possible. A total of 34,368 samples from 492 studies were downloaded (description and accession numbers in S1 Data). Geographical sample locations (where available) are shown in S1 Fig.
We mention that sequencing data from the EMP [22] were omitted from the GPC because this allowed us to use the EMP as an independent reference data set for assessing the fraction of OTUs rediscovered by the GPC and because the much shorter read lengths in the EMP (122 bp on average) compared to the GPC (246 bp on average) would reduce the available phylogenetic resolution [92–96]. Indeed, as we expected the EMP to be less phylogenetically biased than reference databases such as SILVA and RDP, the EMP provided a valuable means to further evaluate the overall coverage of extant prokaryotic diversity by the GPC (see main text and Methods below).
Paired-end reads with sufficient overlap were merged using flash v1.2.11 [97] (options—min-overlap = 10—max-mismatch-density 0.01—phred-offset 33—allow-outies). Of the nonsufficiently overlapping pairs, forward reads were kept and reverse reads discarded. Single-end reads, merged paired-end reads, and nonmerged forward reads were subsequently processed in the same way, as follows. Reads were trimmed and quality filtered using vsearch v2.6.2 [98], keeping only reads that were at least 200 bp long after trimming (options—fastq_ascii 33—fastq_minlen 200—fastq_qmin 0—fastq_maxee 0.5—fastq_truncee 0.5—fastq_maxee_rate 0.002—fastq_stripleft 7—fastq_trunclen_keep 250—fastq_qmax 64). Any samples with more than 106 quality-filtered reads were subsampled down to 106 randomly chosen reads to reduce computational requirements; samples with fewer quality-filtered reads were not subsampled. The 1,988,445,238 kept reads were then chimera-filtered de novo using vsearch (options—abskew 1.9 –mindiv 0.5 –minh 0.1) separately for each sample. About 10% of reads were identified as chimeric (on average, 8.6% of reads per sample), yielding in total 1,734,042,763 quality-filtered and chimera-filtered reads with a mean length of 246 bp. Reads from all samples were pooled and subsequently clustered de novo at 97% similarity using cd-hit-otu v0.0.1 [99]. We chose cd-hit-otu because—in contrast to most other OTU-clustering algorithms—it scales relatively well to massive data sets such as ours. For a comparison between cd-hit-otu and other clustering algorithms, we refer to [99–102]. For consistency with our own downstream error filters (removal of spurious OTUs), we set the minimum size for a cluster of duplicates in the cd-hit-otu algorithm to 2 (step clstr_sort_trim_rep) and the primary cluster size cutoff to 1 (disabling cd-hit-otu's noise removal algorithm). De novo clustering yielded 1,545,602 clusters. Because primers of the various studies included did not all cover exactly the same regions and due to the clustering algorithm implemented by cd-hit-otu, a small number of clusters was redundant, i.e., the representative sequences of some clusters were slightly shifted versions of others. To remove this redundancy, we further clustered representative sequences using vsearch (command—cluster_fast—usersort–id 1.0—iddef 2—strand plus), thereby obtaining 1,386,686 nonredundant OTUs. To further avoid spurious (i.e., nonbiological) OTUs, we only kept OTUs that were found in at least two samples of the same study (944,863 OTUs). While we cannot completely rule out the inclusion of some spurious OTUs in the GPC, we point out that a hypothetical removal of these OTUs would only further decrease our estimates of global prokaryotic OTU richness. Representative sequences for the final set of prokaryotic GPC OTUs (at 97% and 99% clustering threshold) and OTU tables are available online at www.loucalab.com/archive/GPC.
The taxonomic identity of each OTU was determined based on its similarity to entries in the SILVA database [14] and by using a consensus approach, as follows. Each OTU was mapped to SILVA's nonredundant (NR99) SSU sequences using vsearch [98], at a similarity threshold of 60% and keeping only the top 10 hits (options "—id 0.6—strand both—iddef 2—maxaccepts 20—maxhits 10"). If at least one hit had a similarity 100%, then all hits with similarity 100% were used to form a consensus taxonomy. Otherwise, if the best hit had a similarity s≥60%, then all hits with similarity ≥(s-5%) were used to form a consensus taxonomy. In either case, the consensus taxonomy of a set of hits was defined as the taxon at the lowest taxonomic possible level (e.g., domain, phylum, etc.) containing all of the hits. If an OTU did not have any hit in SILVA at or above a threshold of 60% similarity or did not have a consensus taxonomy even at the domain level, it was considered unidentified and was subsequently omitted (see justification in the next paragraph). The overwhelming majority (87%) of OTUs had at least one hit in SILVA at similarity ≥60%, and almost all of these OTUs (>99.9%) could be identified at some taxonomic level. Any OTUs identified as eukaryotes, chloroplasts, and mitochondria were omitted from subsequent analyses.
We note that our imposed similarity threshold of 60% to SILVA is much lower than the thresholds commonly suggested for delineating phyla (e.g., 75% similarity according to [4]), thus the bulk of biological (i.e., nonspurious) sequences is expected to pass this threshold. While the 75% similarity threshold by [4] referred to the full-length 16S gene, the same study also showed that partial gene regions (e.g., "R3" in that paper, roughly corresponding to V4) exhibit less richness than the full-length gene for any given clustering threshold. Hence, organisms that are >75% similar to a SILVA entry in the full 16S are even more likely to be >75% similar in the V4 region; consequently, a similarity threshold of 60% in V4 is probably more permissive than a similarity threshold of 75% in the full gene. OTUs with a similarity to SILVA below 60% (or equivalently, a distance above 40%) are likely largely spurious. To confirm this expectation and to further investigate the nature of these omitted OTUs, we calculated the distribution of distances of OTUs to SILVA as well as the fraction of OTUs that could be matched to SILVA, as the similarity threshold decreased below 60% all the way to zero (S14A and S14B Fig). We found that as one approaches the 60% similarity threshold, the fraction of OTUs matched to SILVA levels off; that is, very few OTUs lie in the 60%–65% range, while the majority of OTUs lies in the 80%–100% range (as discussed in the main article). Strikingly, for slightly lower similarity thresholds, there exists a sharp peak of OTUs within the 50%–60% similarity range and virtually no OTUs below that range. This agglomeration of a small fraction of OTUs in the 50%–60% range is likely mostly spurious, specifically consisting of bichimeras (the most common type of chimeras). Indeed, bichimeras inevitably include a biological segment that makes up at least 50% of their length, and that biological segment will likely match SILVA at considerable similarity. Thus, most bichimeras are expected to aggregate within the 50%–60% similarity interval, as observed in our case. When we repeated the above analysis for clusters at 99% identity, we observed that the peak within the 50%–60% similarity range decreased substantially (S14C and S14D Fig). This is consistent with the expectation that chimeric sequences clustered at 99% identity are easier to detect than when clustered at 97% identity, since the variance around representative sequences hinders a reliable identification of parent sequences by chimera detectors. In fact, when we considered exact ASVs generated and chimera-filtered with DADA2 [91] for a subset of our data (subset "AG," see below), the peak in the 50%–60% similarity interval disappeared nearly completely (S15C Fig). In other words, almost all ASVs had a similarity to SILVA above 60%. This is consistent with the expectation that chimeric ASVs are easier to detect than chimeric OTUs [91] and further supports our conclusion that most of the removed OTUs (all falling within the 50%–60% similarity interval) are likely bichimeras that have escaped our previous chimera filters. The alternative explanation that this agglomerate at 50%–60% similarity represents biological sequences is much less probable, since this would beg the question as to why these sequences aggregate within the similarity interval 50%–60% and why they disappear at higher clustering identities.
To calculate the fraction of prokaryotic 16S diversity recovered by the EMP [22] that was recaptured by the GPC, we proceeded as follows. We dowloaded the EMP's set of unique quality- and chimera-filtered 16S sequences (202,540 "deblurred" sequences, covering 150 bp of the V4 region) from the EMP FTP repository (ftp://ftp.microbio.me/emp/release1/otu_info/deblur/emp.150.min25.deblur.seq.fa). EMP sequences were taxonomically identified using the same methods as for the GPC, and any sequences identified as eukaryotes, chloroplasts, or mitochondria were omitted. EMP sequences were then mapped to GPC OTUs using vsearch at a similarity threshold of 97% whenever possible (options "—id 0.97—iddef 2—strand both"). For any given taxon, the fraction of recaptured EMP sequences was calculated as Nm/NEMP, in which NEMP is the number of EMP sequences identified to be within the focal taxon and Nm is the number of EMP sequences in the focal taxon matched to a GPC OTU. An overview of recapture fractions is provided in S2 Table.
To calculate the fraction of prokaryotic 16S diversity in the RDP (release 11) [12] that was recaptured by the GPC, we proceeded as follows. Nonaligned bacterial and archaeal 16S sequences were downloaded as fasta files from the RDP website (https://rdp.cme.msu.edu/misc/resources.jsp). The RDP's original taxonomic annotations were assumed for each RDP sequence. The fraction of RDP sequences recaptured by the GPC was calculated for various taxa, as described above for the EMP (overview in S2 Table).
To calculate the fraction of prokaryotic 16S diversity in the GTDB (release 86.1) [49] that was recaptured by the GPC, we proceeded as follows. Bacterial and archaeal 16S sequences, extracted from the GTDB genomes, were downloaded as fasta files from the GTDB website (http://gtdb.ecogenomic.org/downloads). Only sequences at least 1,000 bp long were kept. The fraction of GTDB sequences recaptured by the GPC was calculated for various taxa, as described above for the EMP (overview in S2 Table).
To calculate the fraction of 16S sequences from metagenome-assembled UBA genomes [55] that was recaptured by our GPC data set, we proceeded as follows. Fully or partly assembled 16S sequences for 2,853 metagenome-assembled genomes were downloaded from https://data.ace.uq.edu.au/public/misc_downloads/uba_genomes/ on October 25, 2017. Only UBA sequences longer than 1,000 bp were considered to increase the probability of adequate overlap with the V4 region, leaving us with 620 sequences. UBA sequences were taxonomically identified using the same methods as for the GPC, and any sequences identified as eukaryotes, chloroplasts, or mitochondria were omitted. The fraction of UBA sequences recaptured by the GPC was calculated for various taxa, as described above for the EMP (overview in S2 Table).
To calculate the fraction of bacterial 16S sequences previously extracted from metagenomes in the Integrated Microbial Genomes and Microbiomes (IMG/M) database [56] that was recaptured by the GPC, we proceeded as follows. Aligned SSU sequences (≥1,200 bp long) extracted from IMG/M were downloaded as a fasta file from https://bitbucket.org/berkeleylab/bacterialdiversity/downloads on February 13, 2018 (file IMGG_SSU1200.fasta). Only sequences obtained from metagenomes were kept (tag "MTGBAC," 63,367 sequences). Aligned sequences were dealigned (gaps removed); taxonomically identified, as described above for the GPC; and any sequences identified as eukaryotes, chloroplasts, or mitochondria were omitted. The fraction of IMG/M sequences recaptured by the GPC was calculated for various taxa, as described above for the EMP (overview in S2 Table).
Unless otherwise mentioned, sequences in SILVA classified as eukaryotes, mitochondria, or chloroplasts were omitted from all analyses. To calculate the fraction of 16S diversity in the SILVA database [14] that was covered ("recaptured") by the GPC (Figs 3C and S11C), we proceeded as follows. Nonredundant (NR99) SSU alignments in SILVA release 132 were downloaded from the SILVA website (https://www.arb-silva.de/fileadmin/silva_databases/release_132/Exports/SILVA_132_SSURef_Nr99_tax_silva_full_align_trunc.fasta.gz) and subsequently dealigned (gap characters removed). Dealigned SILVA NR99 sequences were then mapped to GPC OTUs via global alignment using vsearch, at a similarity threshold of 97% (options "—id 0.97—iddef 2—strand both"). For any given taxon (domain, phylum, or class), we calculated the coverage by the GPC (Figs 3C and S11C) as the ratio (ρ) of mapped SILVA sequences in that taxon divided by the total number of SILVA sequences in that taxon. The total number of extant OTUs within the taxon (Figs 3A and S11A) was estimated as NGPC/ρ, in which NGPC is the number of GPC OTUs in the taxon.
To estimate the coverage of various prokaryotic taxa (domains, phyla, or classes) by SILVA (Figs 3B and S11B), we proceeded as follows. For any given taxon, we mapped GPC OTUs within that taxon to the dealigned SILVA NR99 sequences via global alignment using vsearch at a similarity threshold of 97% (options "—id 0.97—iddef 2—strand both"). The fraction of OTU richness covered by SILVA was estimated as the ratio of mapped GPC OTUs within that taxon divided by the total number of GPC OTUs in that taxon.
To calculate the 16S diversity in SILVA, in terms of OTUs comparable to the GPC (clusters at 97% identity in the V4 region), we proceeded as follows. We downloaded the full set of SSU alignments from the SILVA website (https://www.arb-silva.de/fileadmin/silva_databases/release_132/Exports/SILVA_132_SSURef_tax_silva_full_align_trunc.fasta.gz). We then aligned GPC OTUs to SILVA using the QIIME script parallel_align_seqs_pynast.py [103] and using a random subset (1%) of the SILVA alignments as a template. We identified the first nucleotide position in the GPC alignments that had a gap fraction below 0.9 (Escherichia coli position 516) and extracted the part starting at that nucleotide position and extending 200 bp in the 5'→3' direction (excluding gaps) from the SILVA alignments. Extracted partial SILVA alignments were then dealigned (gaps removed) and clustered at 97% similarity using uclust v1.2.22 [104], yielding 102,416 prokaryotic OTUs ("SILVA V4-OTUs"). To calculate the 16S diversity in SILVA in terms of full-length OTUs, we also clustered the full-length dealigned SILVA sequences using uclust at 97% similarity, obtaining 270,788 prokaryotic OTUs.
To calculate the distances between GPC's OTUs and SILVA (Fig 2B), we proceeded as follows. OTUs were globally aligned against SILVA NR99 sequences using vsearch, keeping only the top hit (options "—id 0.6—iddef 2—strand both—maxaccepts 1000—maxhits 1—top_hits_only"). For any OTU, its distance to SILVA was defined as 100−I, in which I is the percentage identity to the top hit. The histogram in Fig 2B was obtained after binning distances into intervals of 2%.
In order to obtain a rough estimate of the global richness expected in terms of ASVs (which, in the absence of errors, are equivalent to sequence clusters at 100% similarity) when compared to OTU richness—the standard richness measure considered in previous studies—we investigated the density of exact ASVs in the GPC. ASVs were determined using DADA2 v1.10.0 [91], a tool that fits a stochastic error model to the available sequencing data in order to then distinguish between likely sequencing errors and true biological sequence variants. To limit computational requirements, we only considered a pseudo-randomly chosen subset of GPC studies (111 studies with paired-end reads and whose names started with the letter "A" through "G"), henceforth referred to as "AG" subset. (This subset was chosen for convenience of file handling, and an alphabetical choice of projects is practically random for our purposes.) Any samples with more than 106 raw reads were subsampled down to 106 randomly chosen reads to reduce computational requirements. Reads were quality-filtered and trimmed using the DADA2 function filterAndTrim, with options maxEE = 0.5, minLen = 160, truncQ = 0, trimLeft = 7, truncLen = 167 for forward reads and options maxEE = 1, minLen = 140, truncQ = 0, trimLeft = 7, truncLen = 147 for reverse reads. This yielded 357,738,981 quality-filtered nonmerged paired-end reads. Error rate models were fitted using the DADA2 function learnErrors, separately for each study and separately for forward and reverse reads. ASVs were then inferred for each sample using the DADA2 functions derepFastq and dada (with options pool = FALSE, selfConsist = FALSE), and paired-end denoised reads were subsequently merged using the DADA2 function mergePairs (with options minOverlap = 10, maxMismatch = 0). A preliminary ASV table was created using the DADA2 function makeSequenceTable, yielding an ASV table comprising 258,448,458 reads across 2,319,542 ASVs. Chimeric sequences (specifically, bichimeras) were subsequently removed using the DADA2 function removeBimeraDenovo (with options method = "concensus"), separately for each study. The resulting chimera-filtered ASV table comprised 206,982,673 reads across 725,682 ASVs. Only ASVs matched by at least two reads (across all samples) were kept for downstream analyses in order to eliminate spurious sequences. Because we were mainly interested to check if the number of detected ASVs would be substantially (i.e., orders of magnitude) higher than the number of detected OTUs and because the DADA2 pipeline includes an algorithm for removing sequencing errors, we did not filter out ASVs found only in a single sample so as not risk underestimating the number of exact sequence variants. ASVs were taxonomically identified using SILVA and a consensus approach, as described above for OTUs, resulting in 580,965 prokaryotic ASVs, accounting for 181,673,137 reads across 5,584 samples. (Note that some samples did not pass the various filtering/merging steps.) A summary of AG samples, including sequence accession numbers, is provided as S3 Data.
To compare the number of ASVs and OTUs detected, we also analyzed the same set of quality-filtered reads as used for the above DADA2 analysis using our OTU-clustering approach utilized for the full GPC. Specifically, quality-filtered nonmerged paired-end reads, produced by the first step in the DADA2 pipeline, were used as input to the GPC clustering pipeline described above. This yielded 390,893 prokaryotic sequence clusters at 99% similarity accounting for 190,247,727 reads or 173,166 prokaryotic sequence clusters at 97% similarity accounting for 192,718,873 reads. For a comparison of ASVs and sequence clusters obtained for various numbers of studies included, see S13 Fig.
Accumulation curves of OTUs discovered, as a function of studies included, were calculated as follows. For any given number of studies N, we randomly chose N studies in the GPC and counted the number of OTUs detected in at least one of the chosen studies. We repeated this step 100 independent times and averaged the number of OTUs counted each time. By performing this process for various N (from 1 to 492), we obtained the accumulation curves shown in Fig 1A and 1B.
To estimate the total number of OTUs globally using the statistical estimators described in the main text (iChao2, ICE, CatchAll, breakaway, tWLRM), we considered each study as an independent sampling unit and counted the number of OTUs found in exactly one sampling unit (Q1), in exactly two sampling units (Q2), and so on. Note that since our last quality filter, by which we only kept OTUs found in at least two samples of the same study, was applied separately for each study, every study can indeed be considered as an independent sampling unit. Estimates and standard errors were either calculated using the R package breakaway (breakaway and tWLRM [48]), the R package SpadeR (iChao2 and ICE [105]), or the CatchAll software (CatchAll with 3-mixed exponential model [46]).
The assumption of the above estimators that sampling units are equivalent (e.g., of similar effort) is potentially violated by the GPC, since each included study was performed in a different environment and by using different techniques. To check whether our estimates are affected by this caveat, we also used a variant of iChao2 ("iChao2split"), whereby we randomly assigned studies to four complementary and equally sized groups and considered each group as a single independent global sampling unit. Hence, iChao2split considered the number of OTUs found in only one study group (Q1), in exactly two study groups (Q2), in three study groups (Q3), and in all four study groups (Q4). The splitting was randomly repeated 100 times, and the obtained estimates were averaged (Fig 1E); the standard error was set to the standard deviation of estimates across repeated splittings.
We mention that analogous estimators exist (e.g., "iChao1") for estimating richness in a community based on the observed OTU abundances (such as sequencing read counts) in a single reference sample [41]. Such abundance-based estimators are not suited for our data set for two reasons: first, to obtain a single globally ranging reference sample, we would need to pool all GPC samples so as to obtain a measure of abundance for the various OTUs. However, read counts from separate amplicon-sequencing samples cannot be combined to obtain a measure of global OTU abundances since the total number of cells that was present in each sample is unknown and sequencing depths varied between samples. Second, typical abundance-based estimators such as iChao1 rely on knowing the number of singleton OTUs (i.e., comprising only one read); however, singleton OTUs have a high probability of being spurious and can thus not be reliably used to estimate OTU richness [37]. In fact, singleton OTUs, as well as OTUs found in at most one sample, were omitted from the GPC to minimize spurious OTUs. Note that this filter corresponds to increasing the OTU detection threshold in each study, just as sequencing depth affects detection thresholds. Since the incidence-based richness estimators used in this study all account for finite (a priori unknown and potentially variable) detection probabilities, their applicability is not expected to be substantially compromised by a systematic application of this filter. This is roughly analogous to performing a mark-recapture–based assessment of wildlife population size; a systematic decrease of capturing effort may increase the variance of the resulting estimate, but it will not affect the expected value of that estimate.
To estimate the fraction of prokaryotic cells currently detectable by 16S amplicon sequencing that is represented by GPC OTUs (i.e., at 97% similarity in 16S), we calculated the probability (P) that a single additional read would hit a GPC OTU, as follows. Based on the number of OTUs with exactly two reads (N2 = 87,940) as well as the total number of reads (N = 1,734,042,763) and using the Good–Turing frequency formula [38], we estimate the total probability of hitting an OTU with one read in the GPC to be P1=2N2/N=0.000101. (Note that OTUs with one hit were omitted from the final GPC.) Using the fact that the total estimated probability of hitting an OTU with zero reads in the GPC (P0) is not greater than P1 (it is more probable to rehit some OTU with one read than to hit some OTU with zero reads) and the fact that P≈1−(P0+P1), we obtain the lower bound P≥1-2P1=99.98%. Hence, the probability of a single additional amplicon sequence hitting an OTU with ≥2 reads in the GPC is estimated to be P≥99.98%. An overview of computed probabilities for various clustering thresholds is given in S7 Table.
To estimate the distribution of relative OTU abundances in the GPC, we proceeded as follows. First, for each OTU in the GPC, we estimated its relative abundance (α) in each sample based on the number of assigned reads and using the Good–Turing frequency estimator [38,106]:
α=(r+1)NNr+1Nr,
(1)
in which r is the number of reads assigned to the OTU, Nr is the number of OTUs in the sample with exactly r reads, and N is the total number of reads in the sample. We note that the Good–Turing frequency estimator is widely used in biological statistics and has been repeatedly shown to be more robust than simply using the fraction of assigned reads [106,107]. Next, we averaged the relative abundances of each OTU across all samples to obtain its MRA in the GPC. We emphasize that we calculated MRAs separately for each sample, even though MRAs from shallower sequenced samples may be less accurate. This approach was preferred over the alternative of simply calculating the fraction of reads assigned to an OTU when all samples are pooled because samples differ drastically in sequencing depth; thus, OTUs that happen to occur in deeply sequenced samples would appear to be more abundant than OTUs in shallowly sequenced samples. Similarly, pooling within studies was also avoided because sequencing depth varied widely even among samples of the same study, and samples were usually not technical replicates; hence, MRAs calculated for a given study (after pooling) would be biased toward organisms that happened to be present in deeply sequenced samples. By calculating MRAs separately for each sample prior to averaging, we avoid biases toward OTUs in more deeply sequenced samples.
Next, we grouped OTUs into small, equally sized MRA intervals (on a logarithmic scale) to calculate a frequency histogram of MRAs in the GPC. We note that the resulting frequency histogram should not be interpreted as a true OTU abundance distribution because it only includes OTUs discovered by the GPC and may thus be artificially positively skewed [108]. To estimate the probability that an extant OTU in an MRA interval was included in the GPC (P(α), in which α is the center of the MRA interval) and, from that, the total number of extant OTUs in each MRA interval, we proceeded as follows. We randomly removed half of the quality- and chimera-filtered reads and repeated the OTU clustering and analyses described above, thus obtaining a rarefied variant of the GPC (rGPC). A total of 514,432 high-fidelity prokaryotic OTUs were retrieved from the rGPC. We then calculated the frequency histogram of MRAs for the rGPC and compared it to the one obtained from the GPC to estimate P(α) for each MRA interval. Specifically, we assumed that the number of reads assigned to an OTU in any given MRA interval was Poisson-distributed and that the probability of being discovered was given by the probability of being matched by at least two reads, i.e.,
P(α)=1−e−λ(α)−λ(α)e−λ(α),
(2)
in which λ(α) is the unknown rate of the Poisson distribution for that MRA interval. Since the rGPC includes half the reads of the GPC, the probability of OTU discovery by the rGPC is Pr(α)=1−e−λr(α)−λr(α)e−λr(α), in which λr = λ/2. For each MRA interval, we estimated λ(α) by numerically solving the equation
f(α)fr(α)=1−e−λ(α)−λ(α)e−λ(α)1−e−12λ(α)−12λ(α)e−12λ(α),
(3)
in which f(α) and fr(α) is the number of OTUs in the focal MRA interval, discovered by the GPC and the rGPC, respectively. From the estimated λ(α), we thus obtained P(α) via Eq 2 and the total number of extant OTUs in the MRA interval as F(α)=f(α)/P(α).
Following suggestions by Shoemaker and colleagues [11], who concluded that microbial communities are often well described by log-normal species abundance distributions, a log-normal model was fitted to the reconstructed OTU MRA distribution F:
F(α)∼S2πσ2exp[−(log(α)−μ)22σ2],
(4)
in which μ, σ, and S are fitted parameters. Fitting was performed via least-squares. The fitted log-normal model was integrated over the entire real axis to obtain an estimate for the total number of extant prokaryotic OTUs.
|
10.1371/journal.pgen.1003102 | Selective Pressure Causes an RNA Virus to Trade Reproductive Fitness for Increased Structural and Thermal Stability of a Viral Enzyme | The modulation of fitness by single mutational substitutions during environmental change is the most fundamental consequence of natural selection. The antagonistic tradeoffs of pleiotropic mutations that can be selected under changing environments therefore lie at the foundation of evolutionary biology. However, the molecular basis of fitness tradeoffs is rarely determined in terms of how these pleiotropic mutations affect protein structure. Here we use an interdisciplinary approach to study how antagonistic pleiotropy and protein function dictate a fitness tradeoff. We challenged populations of an RNA virus, bacteriophage Φ6, to evolve in a novel temperature environment where heat shock imposed extreme virus mortality. A single amino acid substitution in the viral lysin protein P5 (V207F) favored improved stability, and hence survival of challenged viruses, despite a concomitant tradeoff that decreased viral reproduction. This mutation increased the thermostability of P5. Crystal structures of wild-type, mutant, and ligand-bound P5 reveal the molecular basis of this thermostabilization—the Phe207 side chain fills a hydrophobic cavity that is unoccupied in the wild-type—and identify P5 as a lytic transglycosylase. The mutation did not reduce the enzymatic activity of P5, suggesting that the reproduction tradeoff stems from other factors such as inefficient capsid assembly or disassembly. Our study demonstrates how combining experimental evolution, biochemistry, and structural biology can identify the mechanisms that drive the antagonistic pleiotropic phenotypes of an individual point mutation in the classic evolutionary tug-of-war between survival and reproduction.
| The most fundamental mechanism of natural selection in a changing environment is the modulation of fitness by mutations. It is the tradeoffs offered by these mutations that drive evolution. However, fitness tradeoffs are rarely understood at the molecular level, in terms of how the selected mutations affect protein structure and function. Here, we merge experimental evolution and structural biology to study the fundamental tradeoff between survival and reproduction. We challenged populations of an RNA virus to evolve in a novel temperature environment where heat shock imposed extreme virus mortality. A single mutation in a specific viral protein increased the stability, and hence survival of challenged viruses, despite a concomitant tradeoff that decreased viral reproduction. This mutation increased the thermal stability of the mutant protein. Atomic structures of the wild-type and mutant protein reveal the molecular basis of this stabilization. The mutation did not reduce the enzymatic activity of the protein, suggesting that the reproduction tradeoff stems from other factors, such as inefficient virus assembly or disassembly. Our study uncovers the mechanism that drives the antagonistic effects of an individual point mutation in the classic evolutionary tug-of-war between survival and reproduction.
| The ability of a single mutational substitution to modulate fitness across environments is the most important consequence of natural selection under environmental change. Understanding the antagonistic tradeoffs of pleiotropic mutations that promote survival in changing environments is therefore essential for a complete understanding of evolution. However, the molecular basis of fitness tradeoffs caused by pleiotropic mutations is rarely determined in terms of how the mutations affect protein structure. Perhaps the main reason for this intellectual gap is because the fields of structural biology and experimental evolution do not often intersect. Structural studies tend to focus on proximate explanations for protein function stemming directly from structural features, without determining the ultimate consequences of evolved protein changes for fitness across environments at the system level. In contrast, experimental evolution studies have identified that point mutations can be consequential for determining fitness tradeoffs in independently evolving populations facing the same environmental change [1], [2], without elucidating the structural details of how such trade-offs are mediated by functional changes at the protein level. It has been argued that interdisciplinary approaches are necessary for the ‘functional synthesis’ that will advance our understanding of evolutionary biology [3], [4], especially to reveal the mechanistic details of evolutionary novelty and adaptive constraint; however, the necessary mergers between disciplines remain rare [5], [6], [7].
Perhaps the most fundamentally important tradeoff in evolutionary biology is that between survival and reproduction, the cornerstones of evolution by natural selection [8]. It is often assumed that natural selection is driven by genetic changes that promote relative differences in offspring production, or reproduction in close relatives [9]. However, the need for organisms to survive in the face of depleted resources or environmental stressors can be of equal or greater importance for dictating relative differences in fitness. It is evident that the functional properties of proteins could bridge tradeoffs in survival versus reproduction, because the genetic changes underlying a protein may simultaneously affect its stability (survival) as well as operational (reproductive) properties across environments. Thus, adaptive evolution in a changing environment provides a key context for studying how protein changes might mediate the interplay of survival versus reproduction, and for determining which variants are favored to evolve under natural selection. ‘Life-history’ tradeoffs between survival and reproduction have been invoked in the adaptive evolution in a variety of organisms [10], but these examples often hinge on statistical correlations between traits, without attempting to identify the molecular basis of changes in protein function that cause such tradeoffs to arise.
Here we challenged populations of an RNA virus, bacteriophage Φ6 of the cystovirus genus [11], to evolve in a novel temperature environment where heat shock imposed extreme virus mortality. A single amino acid substitution in the viral lysin protein P5 favored improved stability (and hence, survival) of challenged viruses, despite a concomitant tradeoff that decreased viral reproduction. Lysins are lytic proteins that locally degrade the cell wall, either to provide access to the inner bacterial membrane during infection or to release virus progeny by cell lysis [12], [13]. An electron microscopy image reconstruction of the bacteriophage Φ12, a cystovirus related to Φ6, suggests that Φ12 P5 is part of the icosahedral nucleocapsid shell of Φ12 and that P5 may interact with the lipid membrane, which constitutes the outer virus layer [14]. It remains to be confirmed whether P5 has an analogous location in Φ6. Nevertheless, the selected mutation in the Φ6 P5 gene increased the thermostability of P5. Crystal structures of the wildtype and mutant P5 reveal the molecular basis of this thermostabilization and identify P5 as a lytic transglycosylase. We show that loss of P5 enzymatic activity is not the source of the viral reproduction tradeoff, which may instead result from inefficient capsid assembly or disassembly. Our study demonstrates how a combination of experimental evolution and biophysical approaches can be used to discover the mechanistic details that drive antagonistic pleiotropic effects of an individual point mutation in the classic evolutionary tug-of-war between survival and reproduction.
The P. phaseolicola host bacteria for phage Φ6 cannot grow and survive at temperatures greater than 30°C. In contrast, virions of wildtype phage Φ6 can withstand exposure to temperatures between 30°C and 40°C, but suffer ‘high mortality’ (subsequent inability to productively infect cells) when subjected to 5 min heat shock at temperatures ranging between 40°C and 50°C [15]. Three populations of wildtype phage Φ6 were evolved independently for 20 days (100 generations) under selection involving 5 min heat shock at 50°C every fifth generation; three control populations were evolved identically, but experienced periodic ‘mock’ heat shocks of 25°C (Figure S1). Subsequently, we conducted repeated (n = 3) survival assays at 42.5°C, 45°C, 47.5°C and 50°C for each of the endpoint treatment and control populations. Each treatment population improved in survival in the 50°C selective environment, relative to wildtype phage Φ6 (independent samples t-tests with 5 df, P<0.0001). Moreover, the treatment populations were significantly advantaged in survival at temperatures above 45°C relative to the controls, which did not differ in thermotolerance from the ancestor [15] (Figure 1A). These differing thermotolerance ‘reaction norms’ (Figure 1A) clearly demonstrated that evolutionary history affected the evolved ability for treatment versus control populations to withstand elevated temperatures.
To examine whether particular molecular substitutions were associated with the differing thermotolerance phenotypes of the evolved populations (Figure 1A), we obtained the consensus genome sequence for each evolved population. Each evolved lineage differed from the wildtype ancestor by 1 to 6 substitutions; overall we observed 18 substitutions at 9 sites across the 3 genomic segments (Table S1). Interestingly, all three treatment populations showed an identical non-synonymous mutation (G2238T) that was the only molecular change on the small RNA segment, and which was not present in the controls. These data strongly suggested that the mutation was somehow beneficial for adaptation to withstand the 50°C heat shock environment, because it arose spontaneously and fixed in all of the independently evolved treatment (but not control) populations. The G2238T mutation corresponds to a V207F amino-acid substitution in the gene for the lysin protein P5, which locally degrades the cell wall, either to provide access to the inner bacterial membrane during infection or to release virus progeny by cell lysis [12], [13]. Hereafter, we refer to the wildtype protein as P5wt and the mutant protein as P5V207F.
Laboratory culture of phage Φ6 on agar always occurs at 25°C, because this incubation temperature allows the P. phaseolicola host bacteria to produce a confluent lawn, which supports robust plaque formation of the virus. Intriguingly, we observed that the viruses from the treatment populations showed a novel plaquing phenotype, which is generally referred to as a ‘bull's-eye’ plaque morphology (Figure 1B). Whereas plaques formed by the wildtype virus were clear, the bull's-eye plaques appeared turbid due to residual bacterial growth within the plaque, indicating that mutant viruses were less efficient at killing bacteria at the ordinary growth temperature of 25°C. To confirm that the P5V207F mutation was antagonistically pleiotropic (i.e., caused improved extracellular survival at 50°C but reduced intracellular growth at 25°C), we isolated viruses containing only the P5V207F mutation (Figure 1C). To do so, we conducted a classic genetic cross [1], [2], where an evolved strain bearing only this mutation on the small segment was ‘back-crossed’ with its wildtype Φ6 ancestor, to obtain a hybrid reassortant with a mutated small segment and the ancestral medium and large segments; genotype of the hybrid was confirmed through sequencing. Plaque assays at 25°C showed that the hybrid produced the same bull's-eye plaque morphology that was characteristic of the treatment populations (Figure 1D). In addition, survival assays showed that the thermotolerance reaction norm for the hybrid was qualitatively similar to data observed in the evolved treatment populations (Figure 1A, 1C); at all elevated temperatures survival of the P5V207F mutant significantly exceeded that of the wildtype. Although percent survival of both strains was modest at the extreme 50°C temperature, survival of P5V207F (0.355±0.241 std. dev.) was still greater than the wildtype (0.013±0.007 std. dev.) (t-test with t = 3.42, df = 14, P = 0.004; Figure 1C). These results indicated that the P5V207F substitution caused both the unique bull's-eye plaque morphology when grown at 25°C, as well as the improved extracellular survival at elevated temperatures.
Because the bull's-eye plaque morphology of the V207F mutant suggested that this genotype less efficiently killed bacteria (relative to the wildtype) at 25°C, we hypothesized that the mutation was deleterious for growth at 25°C. To examine this idea, we conducted paired-growth assays at 25°C, which measured reproduction of the hybrid strain and of the wildtype ancestor under benign conditions, relative to a genetically-marked common competitor virus; (hybrid: n = 28 replicates, wildtype: n = 18 replicates). After adjusting for the cost of the genetic marker on the common competitor, results showed that the mean log reproductive fitness of the mutant relative to the wildtype was −0.341 (±0.324 std. dev.; Figure 1E). Reproductive fitness of the hybrid was significantly less than that of the wildtype based on a two-tailed t-test (with t = 2.878, df = 44, P = 0.006). Together, we observed that the P5V207F mutation caused a ∼1.5-fold decrease in reproduction at 25°C (Figure 1E), but a ∼27-fold increase in extracellular survival at 50°C (Figure 1C). Results of repeated (n = 5) fitness assays modified to match the treatment conditions (i.e., imposing 50°C heat shock prior to growth on agar at 25°C; Figure S1) also demonstrated the net positive effect of the V207F mutation in the selective environment: mean log fitness of the mutant relative to the wildtype was 3.920 (±0.280 std. err.), which significantly exceeded ancestral log fitness of zero (t-test with t = 13.98, df = 5, P = 0.0002). We conclude that the P5V207F mutation is an antagonistically pleiotropic allele that produces a survival/reproduction tradeoff across the two portions of the selective environment experienced by the treatment populations; although the mutation is beneficial for extracellular survival under the brief 50°C heat shock, it is deleterious for reproduction occurring at 25°C.
The evolved thermotolerance of the viruses carrying the P5V207F mutation could result from an increase in the inherent thermostability of P5, or an increase in the stability of a protein-protein interaction directly or indirectly involving P5. To determine how the V207F mutation in protein P5 induced thermotolerance, we first determined the effect of the mutation on the thermostability of P5. We purified recombinant P5wt and P5V207F from Escherichia coli. Thermal melting curves of the two proteins were measured by circular dichroism (CD) spectrometry and differential scanning calorimetry (DSC). P5wt and P5V207F both began to unfold cooperatively as the temperature reached 50°C and 55°C, respectively, as judged from the sharp loss of CD signal from α-helical secondary structure at 220 nm (Figure 2A). Melting temperatures for P5wt and P5V207F calculated from the CD melting curves were 55.3°C and 62.9°C, respectively (Figure 2A). Similarly, the melting temperatures of P5wt and P5V207F determined by DSC were 52.6°C and 58.3°C, respectively (Figure 2). These data indicate that the V207F mutation increases the melting temperature of P5 by between 5.7–7.6°C. Additionally, the heights and areas of the DSC curves were greater for the V207F mutants than for their wildtype counterparts. Together, these data indicate that the higher thermotolerance of the mutant virus is due to increased thermostability of P5V207F relative to P5wt. Since melting temperature and DSC peak area depend on the change in entropy and enthalpy, respectively, we conclude that the additional free energy of stabilization from the V207F mutation (ΔG) derives from both the entropic term (TΔS) and the enthalpy (ΔH) in the free energy equation (ΔG = ΔH−TΔS).
The P5 proteins eluted from a size-exclusion column as 40 kDa proteins despite a calculated molecular mass of 25 kDa. We concluded that P5 was either an elongated monomer or a compact monomer with disordered regions that increased the hydrodynamic radius of the protein. Φ6 P5 was previously reported to be a monomer in solution [16]. To determine whether P5 contained regions of disorder or internal flexibility, we subjected P5wt and P5V207F to limited proteolysis with various proteases. Treatment with V8 protease (Staphylococcus aureus endoproteinase Glu-C) produced three fragments, which were identified by mass spectroscopy as consisting of residues 1–39, 40–47 and 48–220, respectively (Figure 2B). We will refer to the largest fragment, residues 48–220, as P5ΔV8. Additional proteolytic products were observed for P5wt but not for P5V207F (Figure 2B), revealing a higher overall protease sensitivity of the wildtype protein. P5ΔV8 eluted from a size-exclusion column as would be expected for a globular 19 kDa protein. To investigate the state of residues 1–47 further we calculated the difference CD signal between P5wt and P5ΔV8wt (Figure S2). The minimum at 190 nm indicates that residues 1–47 have essentially no secondary structure. The melting temperatures of P5ΔV8wt and P5ΔV8V207F were similar to the full-length proteins as determined by CD (56.6°C and 61.0°C) and DSC (54.1°C and 59.4°C; Figure 2C, 2D). CD melting curves of the truncated proteins also showed greater stability at temperatures below 50°C than the full-length proteins, suggesting that residues 1–47 are responsible for the observed non-cooperative unfolding of the full-length proteins in this temperature range (Figure 2A, 2C). Together, these data indicate that the first 47 residues of P5 are either disordered in solution or fold separately from the rest of the protein. In the virion, however, residues 1–47 may adopt a stable conformation upon binding other viral components.
To understand the molecular basis of the thermostabilization of P5 by the V207F mutation we determined the crystal structures of P5ΔV8wt and P5ΔV8V207F at 1.4 Å resolution (we could not obtain crystals of P5wt or P5V207F). Crystallographic data collection and refinement statistics are provided in Table S2. P5ΔV8 adopts a lysozyme superfamily fold consisting of an N-terminal lobe (NTL) and a C-terminal lobe (CTL) connected by a central helix (Figure 3A). The two lobes create a substrate binding cleft containing the predicted catalytic residue Glu95 (Figure 3A) [17]. The central helix and the CTL of P5ΔV8 have similar conformations as various other lysozyme structures. However, the NTL in P5 differs from other lysozyme structures in its secondary structure content and relative orientation to the CTL (Figure S3). We note that P5ΔV8 forms an unusual crystal-packing interaction in which three N-terminal residues (residues 48–50) of one of the two subunits in the asymmetric unit insert into a shallow groove in a molecule in the adjacent asymmetric unit (Figure S3F). Given that residues 48–50 make specific crystal contacts it was surprising that residues 53–59 were disordered and only residues 48–52 and 60–220 could be modeled.
The structures of P5ΔV8wt and P5ΔV8V207F are almost identical, with a root mean square deviation (rmsd) of the Cα positions of 0.135 Å (Figure 3B, Figure S3G). Residue 207 is located on helix α8, facing the hydrophobic core of the protein. In P5ΔV8wt, V207 and adjacent residues 153, 176, 179–180, 194, 199, and 203–204 create a small (30.28 Å3) cavity lined with hydrophobic side chains (Figure 3B, Figure S4). The lack of electron density within the cavity suggests that it is unoccupied. The larger phenylalanine side chain in P5ΔV8V207F neatly fills the cavity, with no significant changes in the protein. The B-factors near residue 207 are similar in both structures (15–16 Å2). The only significant side chain shift observed in P5ΔV8V207F is an adjustment in the N179 side chain to accommodate the F207 side chain, resulting in a van der Waals interaction between these two side chains. By filling an unoccupied hydrophobic cavity, the V207F substitution increases the buried hydrophobic surface area in the protein. This provides a likely explanation for the increased thermostability of the mutant, because the hydrophobic surface area that is buried within a folded protein contributes directly to its free energy of stabilization [18], [19]. In support of this explanation, the hydrophobic cavity created by the replacement L99A in T4 lysozyme was large enough to bind benzene, and binding of benzene to the L99A mutant increased the melting temperature of T4 lysozyme by 6.0°C [20].
The structure of Φ6 P5 bears the closest resemblance to G-type lysozymes and lytic transglycosylases (LTs) such as gp144 of phage ΦKZ [13] and the catalytic domain of E. coli slt70 with rmsds of 1.2 Å and 1.6 Å, respectively (Figure S3A–S3C). Like G-type lysozymes and LTs [13], [21], [22], [23], Φ6 P5 has just one glutamate, Glu95, in its predicted active site. To gain additional insight into the enzymatic activity of P5, we determined the structure of P5ΔV8wt in complex with the substrate analog chitotetraose (NAG4), at 1.23 Å resolution. Chitotetraose is the most similar commercially available ligand to the natural substrates of LTs and ΦKZ gp144 binds chitotetraose [13]. The structure of ligand-bound P5ΔV8 is similar to that of the apo-enzyme with a few exceptions (Figure 4A, Figure S3G). The major difference is that residues 199–220 are missing from the ligand-bound structure. This region spans the last α-helix (α8) and includes the site of the selected V207F mutation. Helix α7 and the following linker are also in slightly different positions in the ligand-bound structure (Figure S3G). Additionally, the ligand displaces the side chain of Tyr196 out of the substrate-binding site, where the side chain is located in the apo-P5 structure (Figure 4). The ligand-bound P5 crystals belonged to a different space group with different crystal packing than the apo-P5 crystals. The substrate binding cleft is solvent-exposed in the ligand-bound crystals but mostly buried by non-crystallographic symmetry contacts in the apo-P5 crystals. Because the crystals of ligand-bound P5 took twice as long to grow as the apo-P5 crystals, we speculate that the lack of electron density for residues 199–220 in the ligand-bound structure is due to proteolytic cleavage by residual V8 protease or another contaminating protease in the P5 preparation. This cleavage may be favored by the increased solvent exposure of the C-terminal region in the ligand-bound crystals.
The substrate-binding sites of LTs and lysozymes have six sugar binding subsites (A–F) and substrate cleavage occurs between an N-acetyl muramic acid (NAM) bound in subsite D and a N-acetyl glucosamine (NAG) bound in subsite E [24], [25], [26]. In the ligand-bound P5 structure, the four NAG residues of the ligand bind to subsites A–D (Figure 4A). A water molecule is observed between Glu95 and the NAG in subsite D (Figure 4B), supporting a role of Glu95 as the acid/base in the catalytic mechanism. The ligand binds in the same manner as the identical chitotetraose ligand in the structure of the ΦKZ gp144 LT [13]. In both the P5 and gp144 structures, the mode of chitotetraose binding is consistent with the presence of the additional lactic acid and peptidyl moieties that are present in the NAM residues of natural substrates at subsites B and D. Moreover, a superposition of ligand-bound P5 onto the structure of E. coli slt70 LT bound to a glycan product in subsites E and F shows that the geometry and electrostatics of the P5 surface should allow binding of such a product in the same manner (Figure 4C, 4D). Together, the structural features of the P5 active site and glycan binding site and their similarity to the ΦKZ and slt70 LTs strongly support that P5 is a lytic transglycosylase, as previously predicted [17].
To confirm that Φ6 P5 could lyse bacterial cell walls, we examined the cell lysis activity of P5wt and P5V207F using a turbidity assay at the normal growth temperature (25°C). Both proteins lysed cells with the same efficiency (Figure 5A). The maximal rates of cell lysis of P5wt and P5V207F were similar and were directly proportional to the enzyme concentration (Figure 5B).
Merging experimental evolution and structural biology, we used RNA phage Φ6 as a model to demonstrate how a single antagonistically pleiotropic mutation caused a survival/reproduction tradeoff in evolving populations. The V207F mutation conferred better survival of viruses under 50°C heat-shock, despite reducing their reproduction at 25°C; these data demonstrated that thermotolerance was the most important fitness component for dictating the overall evolutionary success of the treatment populations.
We are aware of few other studies that have examined survival/reproduction tradeoffs in viruses. de Paepe and Taddei compared lytic phages of E. coli and suggested a survival/reproduction tradeoff mediated by capsid structure [27]. The proposed mechanism was that denser-packaging of viral DNA within capsids affords increased stability, but tends to slow the rate of phage genome replication. Moreover, in experimental evolution studies where E. coli bacteria and phage ΦX174 or the related phage ID11 were exposed to elevated temperatures, results showed that changes in viral capsid proteins were likely stabilizing [28], [29]. Our results differed from these previous studies because we found the tradeoff was governed by an enzyme, rather than changes to the phage capsid. Also, we observed strong convergence evidenced by a single mutation that fixed in the independently evolved treatment populations, whereas the experiment with phage ID11 showed multiple possible first-step substitutions [28], [29]. One possibility is that our heat-shock regime selected strongly for structural stability, a single fitness component, whereas the phage ID11 study required growth of phage and bacteria at high temperature, thus selecting on multiple fitness components allowing various beneficial mutations to fix. Apparently, our selective regime was so stringent that the convergent mutation fixed despite strong antagonism for growth at ordinary temperature, which constituted environmental conditions aside from the 5 min heat shock. Further research could explore how this tradeoff may be lessened (or even eliminated) via further molecular change(s) resulting from fixation of mutations that compensate for the growth deficit.
A previous report suggested that the Φ6 P5 lysin might have endopeptidase activity rather than the glycanhydrolase activity of lysozyme [16]. However, the cell lysis activity assay used in that study cannot distinguish between cell wall lysis due to endopeptidase activity from cell wall lysis due to glycosidase activity. Conversely, a bioinformatics study by Pei and Grishin classified Φ6 P5 as a distant relative of the lytic transglycosylase (LT) subfamily within the lysozyme superfamily [17]. LTs are enzymes that can degrade the peptidoglycan layer of the bacterial cell wall by cleaving a β(1,4)-glycosidic bond between NAM and NAG residues and forming a new glycosidic bond between the O6 and C1 atoms of the NAM residue [30], [31]. In general, lysozymes possess a catalytic dyad of glutamic and aspartic acid residues that catalyze the hydrolysis of the same substrate by using a water molecule from the solvent [32], [33], [34], [35]. However, in LTs, there is only one acidic residue, typically a glutamic acid, in the vicinity of the substrate cleavage site [36]. Φ6 P5 adopts a lysozyme superfamily fold and our structure of P5 bound to a tetrasaccharide substrate analog shows that the enzyme contains a single glutamate (Glu95) in its active site. Together, our structural data unambiguously identify Φ6 P5 as an LT.
The lysozyme superfamily has been extensively studied as a model system for protein folding and stability, as reviewed in ref. [37]. Several hydrophobic cavities have been identified in lysozymes and it has been proposed that mutations filling these cavities should stabilize the protein by increasing the hydrophobic surface area buried within the fold [38]. However, attempts to fill two different hydrophobic cavities in T4 lysozyme by site-directed mutagenesis resulted unexpectedly in a slight decrease in the melting point of the protein, because stabilization from the increased hydrophobic contacts was offset by strain within the mutant side chains [38]. In contrast, the naturally selected V207F mutation in Φ6 P5 achieves the stabilization expected from the addition of four carbon atoms with a surface area of 35–40 Å2, ∼1 kcal/mol (or a 4°C increase in melting temperature), without the introduction of strain within the protein. Consistent with the average increase in the melting temperature of P5V207F of 5.7°C reported here, binding of benzene to the cavity created by the L99A substitution in T4 lysozyme resulted in a 6.0°C increase in the melting temperature of T4 lysozyme [20]. We conclude that the enhanced thermal stability of P5V207F is responsible for the survival of the mutant phages after the heat-shock challenge.
The V207F mutation did not affect the structure of the P5 active site, allowing the mutant enzyme to fully maintain its functional role as a lysin, which is essential in the viral lifecycle [12], [13]. Thus, it is not clear why the mutation adversely affected viral reproduction. We speculate that the mutation may reduce the structural plasticity of P5, and may hence reduce the efficiency of the assembly or disassembly of the viral capsid. Indeed, the architecture of quasiequivalent icosahedral viral capsids, such as that of Φ6, necessarily depends on structural plasticity within the capsid proteins to form the contacts that hold the capsid together in multiple nonequivalent environments. In support of this hypothesis, an electron cryomicroscopy structure of phage Φ12 (a cystovirus closely related to Φ6) at 10 Å resolution suggests that P5 in an integral part of the viral capsid [14]. Moreover, the weak electron density for P5 in the Φ12 structure suggests that P5 has some flexibility relative to the rest of the capsid, and this flexibility of P5 has been proposed to allow other proteins to access the capsid during virus replication [14]. The selected V207F mutation may therefore impair viral replication, and hence reproduction, by reducing the flexibility of P5.
Although it is widely recognized that infectious viruses can differ markedly in terms of their stability in the face of environmental stress, the associated effects of individual viral proteins remain largely unexplored. Medically and agriculturally important viruses sometimes show an inherent tendency to survive for extended periods outside of their hosts, suggesting that survivability should factor heavily in the relative transmission success of virus genotypes. This is likely to be true for variants of viruses that are transmitted between hosts via inert objects such as transmission of Hepatitis C Viruses between injection-drug users that share needle syringes [39]; better studied examples include differing ability of Influenza A Virus genotypes to withstand exposure to cold water when undergoing fecal-oral transmission in avian hosts [40]. Fever is generally assumed to be a useful innate defense against infecting viruses because these pathogens can degrade when exposed to elevated temperature; although this assumption is critical to the current debate of whether fever-reducing drugs ultimately harm or benefit infected hosts, it is perhaps surprising that evolution of temperature tolerance in viruses is seldom studied [28], [29], [41]. Our study indicated that evolved thermotolerance is rapidly acquired in RNA viruses selected under temperatures much higher than those they normally encounter, strongly suggesting that simple solutions (i.e., point mutations) may govern this adaptation in other virus systems.
Cultures of Pseudomonas syringae pathovar phaseolicola (ATCC #21781) host bacteria were initiated by a single colony grown at 25°C in LC medium: Luria-Bertani broth at pH 7.5. Phage were grown by mixing ∼100 particles with 200 µl of overnight bacterial culture in 3 ml 0.7% LC top agar, overlaid on a 1.5% LC agar plate. After 24 h, phage lysates were prepared by harvesting viral plaques into LC broth, followed by centrifugation and filtration to remove bacteria. Viral stocks were stored at −20°C in 2∶3 glycerol/LC (v/v). Bacterial stocks were stored in 2∶3 glycerol/LC(v/v) at −80°C.
Clones of wildtype Φ6 (strain #PT522) were used to found three treatment and three control populations. Treatment populations were incubated in the absence of cells for 5 min at 50°C followed by 24 h of growth (5 virus generations) on a lawn of P. phaseolicola at 25°C. Viral progeny were harvested as described above. This process was repeated for a total of 20 passages (100 generations [42]) while monitoring the bottleneck population size of evolving lineages to ensure that they experienced equal generation numbers. Control populations were maintained identically but experienced periodic mock heat shocks at 25°C.
Relative reproduction of virus strains was estimated in paired-growth assays as described [43]. Reproduction was gauged relative to common competitor phage: wildtype phage Φ6 containing an engineered mutation (fragment of the Escherichia coli lacZ gene for beta-galactosidase) on segment L [15]. We mixed the test phage and marked competitor at a 1∶1 volumetric ratio, and then plated a dilution of this mixture containing ∼200 viruses onto a host lawn of bacteria. After 24 h incubation, the resulting plaques were harvested and filtered to obtain a cell free lysate. We tracked the ratio of test virus to marked competitor in the starting mixture (R0) and in the harvested lysate (R1) by plating on lawns of LM1034: P. phaseolicola containing the complementing fragment of the E. coli lacZ gene. LM1034 allows the marked competitor to produce blue plaques on agar containing X-gal (0.4% w/v), whereas unmarked phage produce colorless plaques. We defined reproductive fitness (W) as the relative change in ratios, W = R1/R0. After log-transforming fitness estimates, mean log fitness of the wildtype strain was calculated and this value was subtracted from all fitness estimates to adjust for cost of the genetic marker on the common competitor. Fitness assays conducted under the treatment conditions used in experimental evolution (Figure S1) incorporated 5 min heat shock at 50°C. Plating a sample of the starting mixture onto a bacterial lawn confirmed the 1∶1 initial ratio (R0). An additional sample of the mixture was subjected to 5 min heat shock at 50°C, followed by plating on a lawn for plaque growth at 25°C; resulting plaques were harvested and titered to estimate the final ratio (R1). These data were analyzed as above, to estimate fitness in the treatment environment.
Survival was assayed as described [15]; 120 µl of a virus lysate containing ∼108 particles was diluted onto a P. phaseolicola lawn to confirm the initial virus titer (Ni). The lysate was then heated for 5 min and the final titer (Nf) was measured. Percent survival equaled (Nf/Ni) * 100. Thus, survival under heat shock was gauged by tracking pfu viable for growth at 25°C.
Genomic RNA was extracted (QiaAMP viral RNA extraction kit; Qiagen) and converted to cDNA by RT-PCR with Superscript polymerase and random hexamer primers (Invitrogen). Standard PCR methods were used to amplify 93.2% of the genome excluding the single-stranded ends of each segment [2]. PCR products were purified for sequencing with ExoSAP-It (US Biological). Virus genomes were sequenced with double coverage of every nucleotide. Sequences were analyzed with CLC DNA Workbench 6 (www.clcbio.com).
Genes encoding P5wt and P5V207F were cloned into the pET-28 vector (Novagen) in frame with an N-terminal six-histidine tag followed by a tobacco etch virus (TEV) protease cleavage site. P5wt and P5V207F were expressed in E. coli Rosetta (DE3) cells and (Novagen) purified by nickel-affinity and size-exclusion chromatography. The histidine tag was removed with 1∶100 (w/w) TEV protease (12 h at 16°C). Uncleaved P5 and TEV protease were removed with nickel-agarose beads. The proteins were stored at −80°C in 10 mM Tris pH 8, 0.1 M NaCl. P5ΔV8wt and P5ΔV8V207F were prepared by treating purified P5wt and P5V207F with 200∶1 (w/w) S. aureus V8 protease (Worthington) for 3 h. The protease was inactivated with 3 mM PMSF and Complete protease inhibitors (Roche). The P5ΔV8 proteins were then purified by size-exclusion chromatography.
V8 (Glu-C) protease (Worthington) was added to 0.3 g/l of P5wt or P5V207F to a molar ratio of 200∶1 P5∶V8, incubated on ice for 0.5–18 h and heat inactivated (95°C, 5 min in SDS-PAGE loading buffer). Proteolytic products were purified by reverse phase chromatography with a C4 column (Vydac) in 0.05% trifluoroacetic acid using a 10–80% acetonitrile gradient. Peak elution fractions were analyzed by MALDI mass spectroscopy at the Yale Chemical Instrumentation Center.
Circular dichroism measurements were performed on an Aviv 202 spectrometer using 1mm path length cell. Protein samples were diluted to 0.3 g/l in 5 mM sodium phosphate pH 8 to give a reading of approximately −30 millidegrees at 220 nm. For melting curves the temperature was increased from 4°C to 95°C in 1° increments. Readings were taken every degree and were averaged over 3 s after 3 min of temperature equilibration.
Spectra were measured between 180 nm and 260 nm at scan rate of 1 nm/s. P5wt and P5ΔV8wt concentration was 5.5 µM and 4 µM, respectively. The raw data were corrected by subtracting the contribution of the buffer to CD signal. Data were smoothed and converted to molar ellipticity. The measurements were taken at a constant temperature of 16°C. The signal of residues 1–47 was calculated by subtracting the signal of P5ΔV8wt from that of P5wt after correcting for concentration and number of amino acid residues in terms of molar ellipticity.
For differential scanning calorimetry, P5 proteins were diluted to 20 µM in 10 mM Tris pH 8, 0.2 M NaCl and subjected to thermal scans from 10°C to 100°C at a rate of 60°C/h in a MicroCal VP calorimeter with a 15 min pre-equilibration time. Protein-free buffer was used as the reference. Data were collected in triplicate and analyzed with Origin 7 (OriginLab).
Crystals were grown by vapor diffusion at 16°C. P5ΔV8 at 7 g/l in 10 mM Tris pH 8, 0.1 M NaCl was mixed with a half-volume of reservoir solution (1.6 M sodium acetate, 0.1 M sodium citrate pH 6.5). Crystals were frozen in mother liquor. For the ligand-bound structure, P5ΔV8wt was co-crystallized with a 10-fold molar excess of chitotetraose (Sigma) added 3 h prior to mixing with a half-volume of reservoir solution (15% PEG 3350, 0.2 M KNO3 pH 6.9). Two rounds of streak seeding into pre-equilibrated drops of reservoir solution were required to obtain single ligand-bound crystals, which were frozen in reservoir solution plus 25% (v/v) glycerol. P5ΔV8wt crystals were derivatized by soaking in reservoir solution plus 0.2 M NaI for 45 s followed by freezing at 100 K. The structure was determined by single-wavelength anomalous diffraction with HKL2MAP [44]. The atomic model was built with ARP/wARP [45] and refined with Coot [46] and PHENIX [47]. The structure of ligand-bound P5ΔV8wt was determined by molecular replacement with PHENIX using P5ΔV8wt as the search model. Cavities were identified and analyzed with VOIDOO [48] using a 1.1 Å probe radius. See Table S2 for data collection and refinement statistics. Atomic coordinates and structure factors for P5ΔV8wt, P5ΔV8V207F and ligand-bound P5ΔV8wt were deposited in the Protein Data Bank (ID codes 4DQ5, 4DQ7 and 4DQJ).
To assay the cell lysis activities of P5wt and P5V207F, the decrease in turbidity of a chloroform-treated E. coli culture was measured as described [16] by tracking absorbance at 450 nm and 25°C for 20 min after addition of 10–50 ng of protein. For additional details, see Extended Materials and Methods (Text S1).
|
10.1371/journal.pcbi.1005683 | Towards a category theory approach to analogy: Analyzing re-representation and acquisition of numerical knowledge | Category Theory, a branch of mathematics, has shown promise as a modeling framework for higher-level cognition. We introduce an algebraic model for analogy that uses the language of category theory to explore analogy-related cognitive phenomena. To illustrate the potential of this approach, we use this model to explore three objects of study in cognitive literature. First, (a) we use commutative diagrams to analyze an effect of playing particular educational board games on the learning of numbers. Second, (b) we employ a notion called coequalizer as a formal model of re-representation that explains a property of computational models of analogy called “flexibility” whereby non-similar representational elements are considered matches and placed in structural correspondence. Finally, (c) we build a formal learning model which shows that re-representation, language processing and analogy making can explain the acquisition of knowledge of rational numbers. These objects of study provide a picture of acquisition of numerical knowledge that is compatible with empirical evidence and offers insights on possible connections between notions such as relational knowledge, analogy, learning, conceptual knowledge, re-representation and procedural knowledge. This suggests that the approach presented here facilitates mathematical modeling of cognition and provides novel ways to think about analogy-related cognitive phenomena.
| Analogy is claimed to be the core of human cognition due to its pervasive involvement in phenomena such as language, reasoning and learning. However, this phenomenon has mainly been studied in isolation through computational methods, which has made it difficult to appreciate its different roles in diverse cognitive phenomena. Recent studies have signaled that abstract concepts from category theory are able to describe constructions carried out by the human mind, thus presenting this formal theory as a general framework to study cognition. Our contribution here is to provide a model of analogy that bridges a gap between the formal notions of category theory and the psychological notions of cognition. We illustrate the usefulness of this approach by using our model to represent and analyze three different cognitive phenomena. Besides showing that some abstract mathematical concepts can describe concrete cognitive notions such as re-representation, learning, conceptual knowledge, and procedural knowledge, the process of mathematical modeling is presented here as an alternative for outlining relations between cognitive notions and for developing novel conceptualizations of the involved cognitive phenomena.
| More than five decades ago, a formal notion of “isomorphism” was used to define “representations” that sustained measurement theory and models of cognitive systems [1, 2]. Afterwards, it was proposed to describe how certain processes in a cognitive symbol system are able to reflect corresponding processes in an environment [3]. More recently, this formal notion was proposed to describe associations between “mental representations” and “environments” as a general framework to approach cognition [4, 5]. In regard to analogy making, this notion satisfied the structural consistency principle proposed by Dedre Gentner at the core of analogy [6, 7], but it was pointed out that the constraints imposed by this formalism would be too strong to be useful since isomorphisms identify entities that share the exact same internal structure. This notion was then weakened as a framework for modeling analogies because “the kinds of analogies of psychological interest, virtually never have the structure of a valid isomorphism” [5, p. 300].
But the proposal of similar but more flexible notions allowed novel conceptualizations of analogies and other cognitive phenomena. For example, the notion of structure preserving mappings between two domains—called morphisms—motivated proposals such as the quasi-homomorphism, namely a morphism endowed with “exceptions” where the map would not preserve the structure [4]. Additionally, the notion of a local homomorphism is a partial mapping that preserves the structure only where the map is defined i.e. a map that does not preserve the structure of the entire object [8]. These ideas have proven valuable as cognitive modeling tools, but they still have not taken full advantage of the richness provided by the notion of morphisms.
There are claims pointing to the advantages of studying cognitive phenomena through a branch of mathematics called Category Theory which is a theory of structure based on formal notions such as morphisms, limits, colimits, products, adjuntions and other concepts developed out of a need to formalize commonalities between various mathematical structures [9–11]. It has been pointed out that the human mind has the ability to carry out a large number of constructions that seem so universal that they must be somehow severely constrained. And that these constraints might well be adequately formalized by the notions proposed by category theory which describe the natural constraints of mathematical constructions [12]. Furthermore, category theory has been proposed as the foundation for a theory of cognitive developmental stages [3, 13]
Along this path, it is emphasized that the use of category theory could lead to developing new and fruitful analyses of classical cognitive notions [13]. For example, systematicity is the feature of human cognition whereby cognitive capacity comes in groups of related behaviours. The properties of this phenomenon cannot be explained neither by classical nor connectionist theories because these theories must make ad-hoc assumptions on their respective representations of knowledge i.e. grammars and neural networks [14]. However, when it is examined under a categorical approach, by using notions such as products and adjunctions, the systematicity properties turn out to be uniquely determined. This approach thus explains systematicity without making ad-hoc assumptions [15]. Further support for the potential of category theory as an analytic tool is provided by similar studies of cognitive phenomena [16–19].
On another aspect, category theory seems to be also useful as a language for formulating cognitive processes [20]. Formal notions such as limits and colimits have been used in models of reasoning about space and time [21, p. 25], in semantic models for neural networks [22], and in theories about brain’s spatial representation [23, 24]. Similarly, the notion of morphism has been used to describe structure preserving paths between artificial perceptions [25, 26], and to formalize the functional relation between man-machine interfaces and their machine-functionality (in approaches to user interface development) [27, 28].
There are various applications of category theory to research cognitive psychology [12, 15–18] but applications of category theory to research analogy have been almost non-existent (see however [19, 26]). This is odd because category theory has been used extensively in computer science for the analysis of computation [29–31], and computationalism has been the main tool to research analogy. Our goal here is to give a first step into developing a category theory-based approach for analyzing and modeling cognitive phenomena directly related to analogy.
In the present study, we introduce a simple mathematical model of analogy (MMA) and illustrate how its application allows using notions of category theory into studying cognition: (1) We use the model as a device to elaborate a theory able to describe the effects of playing board games on children’s learning of mathematics [32, 33]. (2) We use the formal notion of a coequalizer to explain a property of symbolic models of analogy called flexibility whereby non-similar representational elements are considered matches and placed in structural correspondence during the analogical comparison [34]. (3) We formalize a frequently used method for teaching fractions to children thus obtaining a formal learning model that suggests that the abilities of re-representation, language processing and analogy making can explain the acquisition of knowledge of rational numbers.
The analyses performed on the aforementioned three objects of study offer insight on the role of analogy in these cognitive phenomena while illustrating the potentialities of applying category theory into the study of cognition. In these analyses, we show how some abstract structures provided by category theory can help us to arrange cognitive notions into formal theories capable of explaining, analyzing and organizing cognitive material. Above all, these three objects of study illustrate how this approach can help in developing novel ways to think about cognitive phenomena. In this way, we argue here that the MMA (see Definition 0.3 below) provides a convenient bridge between the formal notions provided by category theory and the psychological notions necessary to develop theories of cognition.
Analogy enables humans to gain understanding of unknown structures (target domains) by using knowledge of previously known structures (source domains). Evidence supporting this point of view, together with the observation that analogy is pervasive in language and thought, suggests a key role for analogical processes at the core of human cognition [35–40]. This suggests that analogy plays a key role in diverse fields such as linguistics, psychology, cognitive science, education and artificial intelligence [5, 6, 41–45] among others.
The study of analogy has been mainly done through computational models and simulations of the phenomenon. Some of these models are called symbolic because they describe domains of knowledge as sets of formulas of symbolic languages [7, 46]. Other models are called connectionist because they represent semantic knowledge through neural networks and distributed representations [43, 47, 48]. If these two features are present in a computational model, it is called hybrid because it represents knowledge through the integration of syntax and semantics [42, 49]. A review of these families of models has been presented in [34] and hence we present here only a short review of analogy models that can be regarded as the predecessors of the formal model presented in the next section.
COPYCAT solves proportional analogies in a domain of strings of characters: Its authors would ask “suppose the letter-string aabc were changed to aabd; how would you change the letter-string ijkk in ‘the same way’? [42]. To analyze this kind of problems, copycat has perceptual agents named codelets which, by using a stochastic method called simulated annealing, combine the primitives stored in its slipnet (the letters a, b, c, d, …, z) using operators such as “succesor”, “predecessor”, “same” and others that permit it to construct an internal representation of the problem (see also TABLETOP in [49]). In our example, the codelets choose the answer ijll by generating representations such as the one described by Fig 1. Since copycat uses a non-deterministic algorithm to solve the combinatorial problem, other answers such as ijkl or hjkk are also given.
Around the same time, a formal treatise of analogy was proposed as the basis of a theory of cognition inspired in algebraic concepts [40]. To represent knowledge, the notion of conceptual network is defined as a finitely generated algebra i.e. a set of objects along with some operations defined on that set. An analogy from conceptual network A to conceptual network B is called a cognitive relation which is defined as a subset of the product A × B.
These ideas were then used to solve proportional analogies in string domains: Dastani and Indurkhya [8] developed a computational model using the algebra determined by the characters a, b, c, d, …, z and some operators (or gestalts) such as “iteration”, “succesor”, “symmetry” and “alternation”. Problems such as abc: abcd::zyx:? were solved by computing local homomorphisms i.e. partial maps preserving the algebraic structure only in the subset where the partial map is defined. For example, to solve the mentioned problem, the core algorithm of this model capitalizes on Indurkhya’s ideas by: (1) generating subalgebras A1 and A2 such that abc, abcd belong to A1 and zyx belongs to A2, (2) generating identical representations for abc and zyx (see Fig 2 to see what “identical” means) and (3) generating a local homomorphism h: A1 → A2 that satisfies h(abc) = zyx. This last step entails the solution h(abcd) = zyxw.
The reviewed models blend together representational processes and analogy-making processes by using an algebraic technique: fixed primitives are combined to generate representations of the relevant domains. Their algorithms are designed to build representations of source and target that share the same structure. This way, the analogy is constructed by forcible interaction of processes of analogy-making and representation. This short review sets an appropriate context to introduce our formal model of analogy because the model proposed in the next section can be thought as an abstraction of these computational models. The mathematical definitions introduced below will be exemplified through computing solutions for analogy problems similar to the ones depicted in the figures above.
Let us first introduce some mathematical tools to formalize concepts such as the source, the target and the analogical map. These key components of the model will be illustrated through formalizing the two following proportional analogy problems borrowed from [8]:
1. abba : abab : : pqrrqp : ?(pqrpqr) 2. abba : abbbbba : : pqrrpq : ?(pqrrrrrpq)
A signature is a set K of function symbols where a nonnegative integer n is assigned to each f ∈ K making f an n -ary function symbol. A K-algebra is a nonempty set C where it is defined a family of finitary operations indexed by K i.e. to each n-ary function symbol f ∈ K corresponds exactly one n-ary operation fC defined on C.
Example 0.1. Let C be the set of all non-null strings on the letters a, b, …., z. Let C be endowed with two operations: “append” which is associated to the binary function symbol λ, and “symmetry” whose associated unary function symbol is σ. The “append” operation inputs two strings s1 and s2 and outputs the single string s1 s2. The “symmetry” operation inverts the order of letters in its input string i.e. σ(l1 l2…ln−1 ln) = ln ln−1…l2 l1. This structured set C is a K-algebra when considering the signature K = {λ, σ} □.
Now, let us consider a countable set V = {x1, x2, …} of variables and let us generate the set of K-terms in the standard recursive way:
This recursive generation of terms has some well known consequences (see [50–52]): First, each assignment of variables α: V → C can be extended to every term t = f(t1, …tn) by the recursive definition α(t) = fC(α(t1), …., α(tn)). And second, if {x1, x2, …, xn} is the set of variables occurring in the term t, then an n-ary operation ft is determined on C by ft(a1, a2, …, an) = α(t) where the assignment α satisfies a1 = α(x1), …, an = α(xn). Notice that any nonempty set Π of K-terms determines a family of operations {ft}t ∈ Π in the K-algebra C. The closure of A0 ⊆ C with respect to {ft}t ∈ Π is the countable union ⋃ i = 0 ∞ A i = A 0 ∪ A 1 ∪ A 2 . . . where Ai = {ft(a1, a2, …, an) | t ∈ Π and a1, a2, …, an ∈ Ai−1}.
Definition 0.1. Let C be a K-algebra, let A ⊆ C, and let Π be a set of K-terms. The pair (A, Π) is a domain with context C if the set A is closed with respect to the family of operations {ft}t∈Π. Also, we say that A0 ⊆ A generates (A, Π) if A is the closure of A0 with respect to {ft}t∈Π.
Example 0.2. Let us consider the K-algebra from example 0.1 and consider two K-terms: t = λ(x1, x2) and s = λ(x1, σ(x2)). Notice that ft represents the append operation. In contrast, the operation fs takes two strings and appends to the first one, the result of applying symmetry to the second one. Given Π = {t, s}, let us consider the domain (A, Π) generated by the two-string set A0 = {ab, p}. For illustrative purposes, let us emphasize that A is an infinite set containing elements like ab, p, abp = ft(ab, p), pba = fs(p, ab), and abpabp = ft(abp, abp) = fs(abp, pba). □.
We say that a ∈ A is an element of (A, Π) and that ft is an operation of (A, Π). In Example 0.2 the elements are strings of letters and the operations are syntactic operators. To introduce our formalism, let us first consider a similar, widely known algebraic notion. Let A and B be two K-algebras. A map h: A → B is called a homomorphism from A to B if for each n-ary function symbol f ∈ K and every a1, …, an ∈ A,
h ( f A ( a 1 , . . . , a n ) ) = f B ( h ( a 1 ) , . . . , h ( a n ) ) (1)
This algebraic definition formalizes the preservation of the structure of the K-algebra A into the K-algebra B. As mentioned in the introduction, notions such as quasi-homomorphisms [4] and local-homomorphisms [8] have been proposed as models for analogy. A key problem with these approaches is that many mathematical constructions—such as products, quotients, limits, adjunctions and others whose construction depends on homomorphisms—lose their key properties when “exceptions” are allowed (the composition of two partial-functions or quasi-homomorphisms is not well defined in general, hence these notions may not satisfy the composition axiom required by category theory [9, pp. 7]).
To preserve the potential usefulness of these constructions in modeling of cognitive phenomena, we propose a model based on homomorphisms. But we need a strategy to achieve certain flexibility for formalizing at least some of the analogies with psychological interest. To this aim, we have included a set of parameters (the set Π of Definition 0.1) that gives us a margin of action in the modeling process. A first version of our model is presented now.
Definition 0.2. (RMA/Restricted Model for Analogies) Let (A, Π) and (B, Π) be two domains whose context is the K-algebra C. A map h: A → B is called a homomorphism from (A, Π) to (B, Π) and denoted by h: (A, Π) → (B, Π) if for each t(x1, x2, …, xn) ∈ Π and every a1, a2, …, an ∈ A,
h ( f t A ( a 1 , . . . , a n ) ) = f t B ( h ( a 1 ) , . . . , h ( a n ) ) (2)
The only difference between conditions (1) and (2) is that the last condition ensures the preservation of the structure determined by the family of operations {ft}t ∈ Π instead of the operations that are indexed by the signature K. This subtle characteristic together with the notion of “generating a domain” endow our model with certain flexibility when formalizing analogies as it is shown by the next examples.
Example 0.3. Let us model the analogy problem abba: abab::pqrrqp:? and propose a solution through the model.
Let us consider the set of terms Π = {t, s} from Example 0.2 where the term t = λ(x1, x2) gives the operation “append” and s = λ(x1, σ(x2)) gives an operation which appends a first string to the result of applying symmetry to a second string. Let (A, Π) be the domain generated by the singleton A0 = {ab} and let (B, Π) be the domain generated by the singleton B0 = {pqr}. Observe that abba, abab ∈ A and that pqrrqp ∈ B. We need to look for a homomorphism h: A → B that extends the partial map h′(ab) = pqr. There exists a (unique) homomorphism h from the domain (A, Π) to the domain (B, Π) that extends h′ (see S1 Note). Since h is a homomorphism of domains, it must satisfy that h(abba) = pqrrqp and h(abab) = pqrpqr which is the answer that this particular modeling of the problem provides. □
The example above formalizes a relation between the two key letter-strings in the source domain. This relation allows determining the missing item in the target domain via analogical transfer: abba and abab are built from the same element (ab) by applying the operations ft and fs imposed on the source domain (see S2 Note for more details about this relation and its transfer). Notice also that the definition of (A, Π) gives these two operations a key role, while downplaying the role of the original “symmetry” operation. These observations are consistent with empirical data suggesting that analogy is based on the mapping of relations [6, 53] and is able to emphasize certain features of domains while downplaying certain others [54].
Clearly, our definition of domains is reminiscent of similar proposals in the literature (see [5, 8, 40]) where a domain is a mathematical set together with a collection of operations defined on it. However, the syntactic parameters in the set Π, besides giving some flexibility to our model, gives us a way of “quantifying” the representational structure that a domain imposes over an analogy problem. To see this, observe that under the perspective of the K-algebra C (Example 0.1), the structure imposed on the string abba is “low” because this string can be built up in many ways from the alphabet and the primitives λ and σ. But under the perspective of (A, Π) in Example 0.3, the structure imposed on abba is “high” because there is exactly one way of building such a string: the only way in which the domain (A, Π) can “understand” the letter-string abba is fs(ab, ab). Next example displays another domain that “understands” the string abba in a unique way.
Example 0.4. Let us model the proportional analogy a b b a : a b b b b b a : : p q r r p q : ? (3) Let us consider the set Π = {t1, t2} = {λ(λ(x1, x2), λ(x2, x1)), λ(λ(λ(λ(λ(x1, x2), x2), x2), x2), λ(x2, x1))} and observe that f t 1 ( a , b ) = a b b a. At the same values, the operation determined by t2 returns the string abbbbba. Therefore, when considering (A, Π) as the domain generated by the set A0 = {a, b} we conclude that abba, abbbbba ∈ A. Also, if (B, Π) is the domain generated by the set B0 = {pq, r}, then pqrrpq ∈ B. We now look for a homomorphism h from (A, Π) to (B, Π) that extends the partial map given by h′(a) = pq and h′(b) = r. It can be shown (see S3 Note) that there exists a (unique) homomorphism h from the domain (A, Π) to the domain (B, Π) extending h′. This map must satisfy h(abba) = pqrrpq and h(abbbbba) = pqrrrrrpq. This last string is the answer given by this particular model of the problem. □
The last two examples show the parameters available in our model of analogy: The source and the target of the analogy are modeled through domains (A, Π) and (B, Π) respectively. A partial map h′—considered as the “analogical map”—characterizes the relation between the source and the target. The model takes this information and provides us with the “analogical transfer” given by h: (A, Π) → (B, Π).
We have been assuming that the source and the target of an analogy share the same context (the same K-algebra). And that the set Π determines the family of operations in source and target i.e. the operations in both domains share the same “syntactic structure”. These assumptions are systematically violated by most analogies because, in general, the source and the target of an analogy differ radically. To account for this observation, we need a way to relate terms of the two domains—because each one may have its own language.
Let K1, K2 be two signatures, let Π be a set of K1-terms and let Ψ be a set of K2-terms. A map F: Π → Ψ is called a term translation if it preserves the variables. More precisely, for each t ∈ Π, the set of variables occurring in t is the exact same set of variables occurring in F(t).
Definition 0.3 (MMA/Mathematical Model for Analogies) Let C1 be a K1-algebra and let C2 be a K2-algebra. Let (A, Π) be a domain with context C1, let (B, Ψ) be a domain with context C2 and let F: Π → Ψ be a term translation. A map h: A → B is called a F -homomorphism from (A, Π) to (B, Ψ) and denoted by h: (A, Π) → (B, Ψ) if for each term t(x1, x2, …, xn) ∈ Π and every a1, a2, …, an ∈ A,
h ( f t A ( a 1 , . . . , a n ) ) = f F ( t ) B ( h ( a 1 ) , . . . , h ( a n ) ) . (4)
This is our core definition, and we refer to it as the mathematical model for analogy (MMA). Notice that Definition 0.2 is just a particular case of the MMA with C1 = C2, Π = Ψ and F equal to the identity. This model has two main components, namely, the “structure-preserving” map h: A → B and the “symbol system” determined by the map F: Π → Ψ. These two components are in line with the explanation given by Dedre Gentner for the striking abilities of human cognition when she says “…analogical ability is the key factor in our prodigious capacity, and, … possession of a symbol system is crucial to the full expression of analogical ability” [55].
In this sense, the MMA abstracts a process whereby chunks of information (from context C1) are combined and coded in (A, Π) for representing the source domain, and concurrently, other chunks of information (from context C2) are combined and coded in (B, Ψ) for representing the target domain. These two processes are coordinated in a way that both domains end up sharing a common structure characterized by the map F: Π → Ψ so that the analogical transfer can be performed by the map h: A → B. Each domain has two dimensions. The symbolic dimension is conformed by sets of terms (Π and Ψ) that represent symbolic information and embedded grammars. And the semantic dimension is conformed by sets (A and B) that abstract possibly non-symbolic elements such as conceptual objects or spatial coordinates.
Some features of this definition can be illustrated through the example: “suppose the letter-string aabc were changed to aabd; how would you change the letter string ijkk in the same way?” [42]. The copycat model reports that the most preferred answer is the letter string ijll followed by ijkl and then by hjkk. The next example uses the MMA to provide us with insight about the algebraic nature of the solution hjkk.
Example 0.5. Let us model the proportional analogy a a b c : a a b d : : i j k k : ? (5) Let us endow the K-algebra C (Example 0.1) with the additional operation “successor” and its inverse “predecessor”, denoted by γ and γ−1 respectively. They are defined in a way that agrees with the alphabet ordering i.e. γ(a) = b, γ(b) = c, …, γ(y) = z, γ(z) = a, γ−1(z) = y, γ−1(y) = x, etc. Let F: Π → Ψ be the only term translation that can be defined between the following sets of terms.
Π = { λ ( x 1 , x 2 ) , γ ( x 3 ) } Ψ = { λ ( x 2 , x 1 ) , γ - 1 ( x 3 ) } Now, let (A, Π) and (B, Ψ) be the two domains generated by the singletons A0 = {a} and B0 = {k}, respectively. Observe that aabc, aabd ∈ A, and that ijkk ∈ B. We look for an F-homomorphism h: A → B satisfying h(a) = k. It exists (see S4 Note ) and can be defined recursively by h(γ(c)) = γ−1(h(c)) and h(λS(c1, c2)) = λS(h(c2), h(c1)) for every c, c1, c2 ∈ A. This model gives the intended answer of the analogy problem since h(aabc) = ijkk and h(aabd) = hjkk. □
The example above illustrates how the complete model is more flexible than the restricted model. However, the constraints imposed by the MMA on the modeling of analogies are still strong. The nature of these constraints is illustrated in the next example where the model cannot give a solution.
Example 0.6. Let us model the proportional analogy a b a b a : a b b a a : : c d c d g : ? (6) Let us set Π = {λ(λ(x1, x1), x2), λ(λ(x1, σ(x1)), x2)}. Let (A, Π) be the domain generated by the set A0 = {ab, a} and let (B, Π) be the domain generated by the set B0 = {cd, g}. Notice that ababa, abbaa ∈ A and cdcdg ∈ B. Now, let us notice that the MMA cannot provide an answer to this problem because an F-homomorphism h that extends h′ can not exist. Let us write Π = {t1, t2} and observe that f t 1 A ( a b a b a , a ) = f t 2 A ( a b a b a , a ) = a b a b a a b a b a a, and then such an h would satisfy h ( a b a b a a b a b a a ) = f t 1 B ( c d c d g , g ) = c d c d g c d c d g g and also h ( a b a b a a b a b a a ) = f t 2 B ( c d c d g , g ) = c d c d g g d c d c g. This would imply that cdcdgcdcdgg = cdcdggdcdcg which is a contradiction. Therefore, h can not exist. □
The lack of a solution is due to the fact that each binary operation ft along with each pair of elements x, y ∈ A determine one constraint, namely
h ( f t ( x , y ) ) = f t ( h ( x ) , h ( y ) )
All these constraints must be satisfied to ensure that (A, Π) and (B, Ψ) share the same structure. The last example illustrates a case where one of these constraints is violated, meaning that the source and target domains have different structures. This suggests that if one does not plan using category theory to analyze phenomena, one might want to consider instead only a small number of these constraints—as in the definition of local homomorphisms [8, Definition 10].
All the examples until now suggest that the partial map h′ somehow forces the unique way in which h must be defined. This does not mean that the MMA predicts only one “valid” answer for each analogy problem. We can use the same framework, even the same K-algebra, and come up with a variety of different domains that yield different answers for the same problem. We illustrate this situation in the next example motivated by results given by Copycat: kjh, kjj and lji are solutions to the problem abc: abd::kji:?
Example 0.7. Let us model the proportional analogy a b c : a b d : : k j i : ? (7) Let us consider the K-algebra C from the last example: the operation “successor” is denoted by γ and its inverse “predecessor” by γ−1. Now, let F1: Π1 → Ψ1 be the term translation between the sets of terms Π1 = {λ(x1, x2), γ(x3)} and Ψ1 = {λ(x1, x2), γ−1(x3)}. Also, let (A, Π1) and (B, Ψ1) be generated by the singletons A0 = {a} and B0 = {k}, respectively. Notice that abc, abd ∈ A and kji ∈ B. The F-homomorphism h: A → B satisfying h(a) = k exists. It can be defined recursively by h(γ(c)) = γ−1(h(c)) and h(λS(c1, c2)) = λS(h(c1), h(c2)) for every c, c1, c2 ∈ A. This first model entails h(abc) = kji and the answer h(abd) = kjh.
Now, let F2: Π2 → Ψ2 be the term translation between Π2 = {λ(x1, x2), γ(x3)} and Ψ2 = {λ(x2, x1), γ(x3)}. Also, let (A, Π2) and (B, Ψ2) be the two domains generated by the singletons A0 = {a} and B0 = {i}, respectively. Notice that abc, abd ∈ A, and that kji ∈ B. The technique used in Example 0.5 can be used to show that the F-homomorphism f: A → B satisfying f(a) = i exists. It can be defined recursively by f(γ(c)) = γ(f(c)) and f(λS(c1, c2)) = λS(f(c2), f(c1)) for every c, c1, c2 ∈ A. This second model entails f(abc) = kji and the answer f(abd) = lji. □
In the next section we present three objects of study that illustrate how this model enables the application of abstract structures from category theory for studying analogy-related cognitive phenomena. But first, we need to present an adaptation of this model to build a representation of a symbolic model of analogy which will be used in our analysis of re-representation and flexibility.
According to literature, symbolic models are characterized by the following features [34, 56, 57]:
We want to emulate these features in a mathematical model that captures the behavior of symbolic models along with their key symbolic mechanisms to represent domains of knowledge. To this aim, let X = {x1, x2, …} be any set (whose elements are thought here as variables), let K be a signature, t, t1, t2, …., tn be K-terms on X, and let us denote by t t 1 , . . . , t n x 1 , . . . , x n the “term” obtained by substitution i.e. simultaneously replacing each occurrence of xi in t by the term ti. We can now define the set of terms Π*(X) that is “freely” generated by X (on a set of terms Π) as the minimal set that satisfies two conditions:
Notice that each K-term t with variables x1, …, xn determines a n-ary, symbolic operation ft on this set Π*(X) as follows:
f t ( t 1 , . . . . , t n ) = t t 1 , . . . , t n x 1 , . . . , x n
Hence the pair (Π*(X), Π) is a domain, namely the free Π-domain on the set X. This allows us to define a symbolic version of a domain of knowledge. The following definition is a particular case of Definition 0.1.
Definition 0.4 (Symbolic Domain). Let K be a signature and let Π be a set of K-terms on the set of symbolic variables X. The pair (Π*(X), Π) is a symbolic domain with signature K. When the set X of symbolic variables is clear from the context we just call it a domain and denote it by (Π*, Π).
We can think of the set Π* as the language that the symbolic model uses to describe a domain of knowledge. We regard the terms in Π and the operations in {ft}t∈Π as the symbols and the rules (grammar) that determines such language. We now adapt the notion of a F-homomorphism with the aim to capture the behaviour of symbolic models of analogy and their aforementioned characteristics.
Definition 0.5 (SMA/Symbolic Model for Analogy). Let Π and Ψ be a set of K1-terms and K2-terms respectively. Let (Π*, Π) and (Ψ*, Ψ) be two symbolic domains and let F: Π → Ψ be a term translation. A map F*: Π* → Ψ* is called a term morphism when (1) it extends F and (2) it is a F-homomorphism from (Π*, Π) to (Ψ*, Ψ).
The features (a), (b) and (c) of symbolic models of analogy are satisfied by a term morphism: The sets Π and Ψ are symbolic descriptions of the source and target that are provided as inputs to the symbolic analogy model. The term translation F: Π → Ψ associates source’s symbols with target’s symbols—it is a description of the characteristic “analogy map” that any symbolic model computes heuristically. The domains (Π*, Π) and (Ψ*, Ψ) are larger descriptions of the source and target created by the symbolic model by recursively using syntactic rules on the descriptions Π and Ψ. Finally, since F* translates the “larger description” Π* to the “larger description” Ψ* in a way that agrees with F, the notion of a term morphism formalizes the “analogical transfer” of knowledge that is based on the preservation of the syntactic structure of the representations of source and target.
The two last definitions comprise a formal model of computational symbolic models of analogy that captures a key mechanism implicit in the nature of symbolic models, namely that the analogical matching is guided by the syntax of the representational elements associated to the source and target domains. Clearly, computational symbolic models are more complex than our formal model and thus we do not expect a full description of them. Particular features, such as architectures or specific algorithms, are lost. Furthermore, our mathematical description does not accurately reflect a key feature of symbolic models: they represent knowledge by using higher-order formal languages whereas our formal description is restricted to using terms of first order languages. Still, the analyses performed on it provide (we believe) some conceptually interesting insights about computational symbolic models of analogy.
In order to analyze the role of re-representation in these computational models, we will explore a relation that appears between a “symbolic model” represented by a (term) morphism F*: (Π*, Π) → (Ψ*, Ψ) (Definition 0.5) and its underlying “conceptual analogy” represented by an F-homomorphism h: (A, Π) → (B, Ψ) (Definition 0.3). The core of this analysis is carried through the formal notion of a coequalizer—a generalization of a quotient object—that is introduced below in the context of category theory.
In category theory, the importance of diagrams and diagram chasing method has been emphasized as a notational method and as a device to organize ideas because “their use provides extensive savings in space and in mental effort” [58]. Our analysis of the cognitive design of board games in our first object of study is based on the use of diagrams that give a precise formal interpretation to the analogy between the spatial domain embedded in these board games and the numerical domain that is the mental representation to be learned. To this aim, we will need the notions of object, morphism and commutative diagram that are exemplified below. Although these notions are not directly defined in category theory, the examples below illustrate how these notions are instantiated in different applications of the theory.
The distinction between the syntactic and semantic versions of the MMA was leveraged in the last section to provide an account of procedural and conceptual knowledge. This distinction is similar to the one proposed by Halford and Wilson who employed category theory to develop a theory of cognitive development [3]. They pointed out that representations in thought must be general so that they can be transferred to situations not previously experienced and argued that representations in the form of relational knowledge are necessary. They described how symbol systems and environmental elements must be set in structural correspondence by building a formal model: a cognitive system is defined as a symbol system (a n-ary operation f defined on a set of symbols S), an environment system (a n-ary operation g defined on a set of environmental elements E) together with a mapping a: S → E that makes the next diagram commute:
(20)
A cognitive system: the mapping a: S → E assigns symbolic representations to environment elements. The n-ary symbolic process f transforms a n-tuple of symbols into another symbol. The n-ary environmental process g transforms a n-tuple of semantic elements into another element. The commutativity of the diagram means that symbolic processes exactly reflect the corresponding processes in the environment system.
This definition allowed its authors to predict age-related differences in performance of certain cognitive tasks. This prediction was based on that the greater the number of symbols involved in a cognitive process (i.e. the arity associated to f or g), the greater the cognitive demands imposed on short term memory. These kind of thinking guided Halford, Wilson and Phillips to create an account of cognitive processing capacity in terms of the relational complexity of symbolic representations [66]. This is pointed out here because of the close relation between the MMA and a cognitive system: the mapping a in Diagram 20 is an F-homomorphism
a : ( S , { f ( x 1 , … , x n ) } ) → ( E , { g ( x 1 , … , x n ) } ) .
This may open a path to research how the relational complexity of representations influences analogical processing: the MMA extends the notion of a cognitive system to include the case of analogical processing where two symbol systems are interacting. In particular, Lemma 0.1 contributes by providing two “orthogonal” structure-preserving mappings: (1) between syntax and semantics within a domain of interest, and (2) between domains of interest (see Diagram 10).
It is important to point out here that we do not claim that symbolically structured cognition necessarily requires symbolic languages as part of their representational systems. This point is made by a model of analogy and schema induction (DORA) [48] which represents relations through four layers of units in a neural network. Such relations are composed by a relation symbol linked to roles which are bounded to fillers. Units representing a role oscillate in close temporal proximity with units representing the filler bound to that role, and out of phase with units representing other roles and fillers. The relational instance less_than(two, three) would be represented by units representing the first-object role of less_than oscillating in synchrony with units representing the number two, while units representing the second-object role of less_than oscillating in synchrony with units representing the number three. Units representing the first-object oscillate out of synchrony with units representing the second-object. Structure-consistent mapping occurs by concurrent activation of units in two analogues i.e. superimposing another relational instance (e.g. smaller_than) by having corresponding roles of all relational instances oscillate in synchrony.
In our perspective, the importance of these achievements lies partly in showing that the integration of semantic and syntactic information must not necessarily rely on the grammar of a symbolic language but rather it may be based on mechanisms similar to neural nets with synchronous oscillation. More importantly, DORA provides a precise account of how children transition from representing the world in terms of unstructured representations of objects to representing the world in a structured fashion where relational knowledge and structure-preserving mappings are explicit. The work presented here pushes a bit further these achievements by showing how similar premises can also provide a basis for the emergence of more complex structures such as those of procedural and conceptual knowledge in the learning of fractions. Although our models cannot give a precise account of the training process (in the way that DORA does), they at least suggest the core data structures and algorithms that may be used in a computational implementation aimed to simulate the human learning of fractions (similar to what is done in [94] where number meanings and other related concepts are learned from naturalistic data).
Category Theory can help us to understand how the core constructions of mathematics are systematically related to one another and how they arise from one another according to simple and general basic principles. By bridging a gap between the formal notions of category theory and the psychological notions of cognition, the MMA helped us to exploit these principles for thinking about the role of analogy in cognition: We used commutative diagrams to describe the learning by analogy that underlies the playing of board games. Also, we used a free domain and a coequalizer to explain the arising of flexibility in symbolic models of analogy. And we built formal models that suggest that the human abilities of re-representation, symbolic language processing and analogy making can explain the acquisition of knowledge of rational numbers. The coherence between the theoretical results and the empirical observations in literature supports that the approach presented here serves as a framework for modeling and analyzing cognitive phenomena related to analogy.
The approach seems to have at least two limitations. The first one is that the MMA imposes conditions that may be too strong for capturing analogical behaviour in certain contexts. But we have presented here three objects of study showing the existence of psychologically interesting phenomena that are well suited to be studied within this framework. The second limitation is that the MMA has less expressive power than computational frameworks of analogy that represent knowledge by using higher order languages. Although this comparison might be unfair because computational models pursue goals conceptually different from the ones pursued in this study, we acknowledge that the presented framework could be improved with the addition of higher order logics. In the meanwhile, this lack of expressiveness may well be compensated by the variety of formal notions (such as limits, colimits, adjuntions, functors and others) from category theory that can be used as construction blocks into the building of new cognitive models.
Even though the analyses presented in this work are based on simple models of complex phenomena, these models allowed us to articulate analogy-related cognitive theories and to exploit them into analyzing, explaining and organizing cognitive material. Hence, category theory not only helped us to re-conceptualize cognitive notions, but also to hypothesize on how these notions are connected to each other. This interplay between mathematics and cognitive theories yielded results that are mathematically interesting, conceptually revealing and potentially useful for the cases of re-representation and acquisition of numerical knowledge.
But, what other cognitive phenomena can be studied in this manner? We believe the approach presented here might help us to inquire into more fundamental aspects of analogy. For example, a problem of central concern to analogy researchers is to understand why a particular analogy is chosen over the possibly many other alternatives. It might be interesting to apply the MMA along with a theory of relational complexity and cognitive processing [66] in order to investigate whether the relational complexity of the analogical candidates can determine (or influence) the final selection. We suggest to move forward by using formal notions such as products, functors, limits and colimits into exploring more cognitive phenomena. Non-trivial results of category theory might become relevant in future research. For example, the free domain (Π*, Π) presented in this study is associated to a free functor that appears as the left adjoint to a forgetful functor defined on certain sub-categories of Dom. Hence, adjoint functor theorems might turn out to be suitable tools when studying the category of domains introduced here. We expect the full development of this framework will provide a large collection of mathematical tools to formulate theories that exploit the advantages of formal analysis in the study of the human cognitive architecture.
|
10.1371/journal.pgen.1001215 | RACK-1 Acts with Rac GTPase Signaling and UNC-115/abLIM in Caenorhabditis elegans Axon Pathfinding and Cell Migration | Migrating cells and growth cones extend lamellipodial and filopodial protrusions that are required for outgrowth and guidance. The mechanisms of cytoskeletal regulation that underlie cell and growth cone migration are of much interest to developmental biologists. Previous studies have shown that the Arp2/3 complex and UNC-115/abLIM act redundantly to mediate growth cone lamellipodia and filopodia formation and axon pathfinding. While much is known about the regulation of Arp2/3, less is known about regulators of UNC-115/abLIM. Here we show that the Caenorhabditis elegans counterpart of the Receptor for Activated C Kinase (RACK-1) interacts physically with the actin-binding protein UNC-115/abLIM and that RACK-1 is required for axon pathfinding. Genetic interactions indicate that RACK-1 acts cell-autonomously in the UNC-115/abLIM pathway in axon pathfinding and lamellipodia and filopodia formation, downstream of the CED-10/Rac GTPase and in parallel to MIG-2/RhoG. Furthermore, we show that RACK-1 is involved in migration of the gonadal distal tip cells and that the signaling pathways involved in this process might be distinct from those involved in axon pathfinding. In sum, these studies pinpoint RACK-1 as a component of a novel signaling pathway involving Rac GTPases and UNC-115/abLIM and suggest that RACK-1 might be involved in the regulation of the actin cytoskeleton and lamellipodia and filopodia formation in migrating cells and growth cones.
| In the developing nervous system, the growth cone guides axons of neurons to their correct targets in the organism. The growth cone is a highly dynamic specialization at the tip of the axon that senses cues and responds by crawling toward its target, leaving the axon behind. Key to growth cone motility are dynamic cellular protrusions called lamellipodia and filopodia. These protrusions are required for growth cone movement and steering. The genes that are involved in lamellipodia and filopodia formation in the growth cone are still being discovered, and studies to understand how these genes act together in cell signaling events that control growth cone movement are in their infancy. Here we report discovery of a new gene necessary for growth cone movement in Caenorhabditis elegans called rack-1. This gene is conserved in vertebrates and is involved in cellular signaling. We show that it interacts in a novel manner with other cell signaling genes (the Rac GTPase genes) and a gene involved in lamellipodia and filopodia formation, called unc-115/abLIM. We think that rack-1 is involved in a novel cellular signaling event involving Rac GTPases that regulates lamellipodia and filopodia protrusion in the growth cone during nervous system development.
| The actin cytoskeleton is necessary for the formation of cellular protrusions, lamellipodia and filopodia, that underlie morphogenetic events such as cell migration and axon pathfinding [1]–[4]. Unraveling the complex molecular events that regulate actin structure and dynamics in migrating cells and growth cones will be central to understanding the development of multicellular organisms and the nervous system in particular. Migrating cells and growth cones display dynamic lamellipodial and filopodial protrusions consisting of a meshwork of actin filaments and bundles of actin filaments, respectively [4]–[8]. Lamellipodia and filopodia serve to guide cells and growth cones and also provide in part the motile force necessary for cell migration and growth cone advance [9]. A complex interplay of filopodial and lamellipodial dynamics controlled by guidance receptors and their ligands is the basis for guidance outgrowth and migration.
In cultured cells, the actin-nucleating Arp2/3 complex controls the formation of lamellipodial networks [1], [10], [11], whereas the anti-capping protein Enabled controls filopodial formation [12], [13]. Enabled also affects axon pathfinding in Caenorhabditis elegans [14], [15]. In migrating growth cones in C. elegans, the Arp2/3 complex is required for both lamellipodial and filopodial formation [16], likely due to the contribution of Arp2/3-nucleated actin filaments to filopodial bundles [7]. The actin-binding protein UNC-115/abLIM [17] also controls lamellipodial and filopodial formation in C. elegans growth cones [16], and acts in parallel to the Arp2/3 complex in axon pathfinding [16]–[18], indicating that UNC-115/abLIM may be contributing to both lamellipodial and filopodial formation in growth cones. The signaling pathways that control Arp2/3 activation are well documented. The Arp2/3 activators WASP and WAVE act downstream of Cdc42 and Rac GTPases respectively to regulate Arp2/3 activity [11], [19]–[22]. In C. elegans axon pathfinding, WVE-1/WAVE acts downstream of CED-10/Rac and WSP-1/WASP acts downstream of the MIG-2/RhoG GTPase to regulate Arp2/3 [18].
While much is known about the Arp2/3 signaling pathway, less is known about the control of UNC-115/abLIM in lamellipodia and filopodia formation. The conserved UNC-115/abLIM proteins have multiple LIM domains at the N terminus and an actin-binding villin headpiece domain at the C terminus [17], [23], [24]. The central region of the molecule contains a short region of similarity shared with the dematin protein, which also contains a C terminal actin-binding villin headpiece domain. Previous studies in C. elegans showed that UNC-115/abLIM acts downstream of the CED-10/Rac GTPase in neuronal lamellipodia and filopodia formation [25]. The conserved seven-WD repeat molecule SWAN-1 physically interacts with the UNC-115 LIM domains and with Rac GTPases, and is normally required to attenuate Rac GTPase signaling [26], indicating that SWAN-1 might be a link between Rac signaling and UNC-115/abLIM.
A two-hybrid screen with the central region of UNC-115 identified the C. elegans Receptor for Activated C Kinase molecule (Rack1), called RACK-1 in C. elegans [27], [28]. Rack1 molecules are composed of seven WD repeats, which form a seven-bladed beta propeller structure that serves as a scaffold for protein-protein interactions [29]. Rack1 was first identified as a molecule that bound to activated protein kinase C and mediated its plasma membrane translocation [30], [31]. Further studies have shown that Rack1 acts with a very diverse set of signaling complexes and can mediate their sub cellular distributions and shuttling (reviewed in [32]). This diversity of interaction leads to a diversity of function for Rack1, including transcriptional and translational regulation, regulation of membrane trafficking, regulation of signal transduction, and cell adhesion [32]. Interestingly, Rack1 controls cell motility via its interaction with the Src tyrosine kinase [31], [33]. Rack1 is a substrate for Src tyrosine phosphorylation and acts as a repressor of Src in response to active PKC [31], [34]–[37]. Rack1 inhibits Src-induced cell motility in cultured 3T3 fibroblasts, and inhibits Src phosphorylation of p190RhoGAP [38], a modulator of Rho GTPase signaling and actin organization. Rack1 is also phosphorylated on tyrosine 52 by c-Abl, which is involved in Rack1 regulation of focal adhesion kinase and integrin function [39]. In C. elegans, RACK-1 has been shown to be involved in embryonic cytokinesis [27]. C. elegans RACK-1 regulates membrane trafficking and recycling endosome distribution via interaction with dynactin, and thus might regulate the microtubule motor dynein. As a consequence, rack-1 loss of function leads to defects in cytokinesis and chromosome separation in the early embryo.
Here we show that RACK-1 interacts with the actin-binding protein UNC-115/abLIM, and that RACK-1 is required for axon pathfinding. Genetic interactions indicate that RACK-1 acts in the UNC-115/abLIM pathway in axon pathfinding, downstream of the CED-10/Rac GTPase and in parallel to MIG-2/RhoG and the UNC-34/Enabled. Neuron-specific expression of RACK-1 is sufficient to rescue the axon pathfinding defects of rack-1 mutants, indicating that RACK-1 acts cell autonomously in axon pathfinding. Furthermore, we show that RACK-1 is involved in migration of the gonadal distal tip cells, and that the signaling pathways involved in this process might be distinct from those involved in axon pathfinding. In sum, these studies pinpoint RACK-1 as a component of a signaling pathway involving Rac GTPases and UNC-115/abLIM, and suggest that RACK-1 might be involved in the regulation of the actin cytoskeleton and lamellipodia and filopodia formation in migrating cells and growth cones.
The actin-binding protein UNC-115/abLIM has three LIM domains in the N-terminus, a villin headpiece domain (VHD) in the C-terminus, and a middle region with unknown function that contains a highly conserved region across species, the UAD domain (UNC-115, abLIM, dematin) (Figure 1A) [17]. The VHD physically interacts with F-actin [24], [25], while the LIM domains are thought to mediate protein-protein interactions. Previous studies showed that the seven WD-repeat protein SWAN-1, a negative regulator of UNC-115 activity, interacts with the LIM domains of UNC-115 [26].
In order to identify molecules that interact physically with the non-LIM-domain region UNC-115, the central region of UNC-115 (residues 243 to 553 of the F09B9.2b molecule as described on Wormbase) was used as bait in a yeast two-hybrid screen (Figure 1A). This two-hybrid screen was performed at the Molecular Interaction Facility at the University of Wisconsin-Madison. The screen involved activation of β-galactosidase activity and HIS5 expression in a liquid-based microtiter screening procedure (see Materials and Methods). From a total of 36 million C. elegans poly-A primed cDNAs screened, seven cDNAs that corresponded to the K04D7.1 gene (as annotated on Wormbase) were found. All seven cDNAs were found to activate when retested, and all seven cDNAs displayed bait-dependence and did not activate in the absence of the UNC-115 bait (data not shown).
The seven cDNAs represented five independent isolates (i.e. represented five different 5′ ends), with two of the isolates having two representatives each (Figure 1E). Three of the cDNAs contained the entire predicted K04D7.1 open reading frame, and two were missing some of the predicted 5′ open reading frame. All five cDNAs were in frame to the GAL4 activation domain in the pACT two-hybrid vector.
The K04D7.1 cDNAs were predicted to encode a molecule similar to vertebrate Receptor for Activated C Kinase (Rack1), called RACK-1 in C. elegans (Figure 1D and 1E) [27]. RACK-1 is predicted to contain seven WD repeats that form a seven-bladed beta-propeller similar to the beta subunit of G proteins [40]. Rack1 molecules define a conserved family of seven-WD repeat proteins, and are distinct from other families such as Gβ and AN11/SWAN-1 [26], [40]. Rack1 molecules are defined by two conserved regions that interact with protein kinase C, a conserved tyrosine residue that is phosphorylated by the Src tyrosine kinase, and a tyrosine residue at position 52 that is phosphorylated by c-Abl. The PKC interaction sites and the Src phosphorylated tyrosine are conserved in the C. elegans RACK-1 protein, but the c-Abl phosphorylated tyrosine at position 52 in human Rack1 is not present in C. elegans RACK-1 (Figure 1E). Two of the cDNAs isolated in the two-hybrid screen were missing coding region for the first predicted WD repeat and one of them was missing part of the second predicted WD repeat (Figure 1E).
To confirm that RACK-1 and UNC-115 interact in a complex, we determined if RACK-1 and UNC-115 co-immunoprecipitated from C. elegans extracts. We generated a transgene expressing MYC-tagged RACK-1 under its endogenous promoter and made animals transgenic for this construct. This transgene produced functional RACK-1::MYC, as it rescued the sterility, gonadal distal tip cell migration defects, and axon pathfinding defects caused by the rack-1(tm2262) deletion (see below). We immunoprecipitated MYC-tagged RACK-1 (RACK-1::MYC) using an anti-MYC antibody from animals harboring a rack-1::myc integrated gene (see Materials and Methods). Using anti-MYC western blots, we found that RACK-1::MYC (36 kD) was expressed in C. elegans extracts and that it was immunoprecipitated by this treatment (Figure 1B). Western blots using anti-UNC-115 antibody [26] showed the specific co-immunoprecipitation of UNC-115 (72 kD) with RACK-1::MYC (Figure 1B). In the absence of the anti-MYC antibody, RACK-1::MYC did not precipitate, and neither did UNC-115 (Figure 1B). Furthermore, we could detect no UNC-115 when extracts from C. elegans not expressing RACK-1::MYC were immunoprecipitated with the MYC antibody (data not shown). We repeated this co-immunoprecipitation two additional times, and the results of one representative experiment are shown in Figure 1.
C. elegans RACK-1 is a 325 amino-acid protein that has two regions similar to the PKC binding sites of vertebrate RACK and a conserved tyrosine that is phosphorylated by Src in vertebrate RACK. C. elegans PKC and Src isoforms are expressed in the nervous system, and both PKC and Src have been implicated in growth cone pathfinding and cell migration [41], [42]. Furthermore, we show above that RACK-1 interacts with UNC-115, a molecule that controls axon pathfinding in C. elegans [17]. Thus, we determined if RACK-1 was also involved in axon pathfinding in C. elegans.
The VD and DD motor neurons are GABAergic neurons that control the coordination and movement of the nematode [43], [44]. The VD and DD cell bodies reside on the right side of the ventral nerve cord. Axons extend anteriorly, branch, and extend dorsally to form axon commissures (Figure 2). Upon reaching the dorsal cord, the axons branch again and extend posteriorly and anteriorly. We used an unc-25::gfp transgene (juIs76) to image the VD/DD neurons and their axons [45]. unc-115(ky275) disrupts axon pathfinding in these neurons, yielding in an uncoordinated movement phenotype [17]. We perturbed rack-1 function using RNAi by injection (see Materials and Methods). In 22% of injected animals (n>100), rack-1(RNAi) disrupted the proper pathfinding of the VD and DD commissural axons (Figure 2A and 2B). The defects seen, such as axon misguidance, branching and premature termination, resembled the defects observed in unc-115(ky275) [17] and were never observed in wild-type animals. These results suggest that RACK-1 might be involved in axon pathfinding, similar to UNC-115.
A deletion of the rack-1 locus, called tm2262, was isolated and kindly provided by the National Bioresource Project for the Experimental Animal “Nematode C. elegans” (S. Mitani). The tm2262 deletion was an in-frame deletion that removed part of the first WD repeat, all of the second, and most of the third, including the predicted PKC interaction site in WD3 (Figure 1C–1E). Since tm2262 is an in-frame deletion, tm2262 animals might still produce truncated RACK-1 protein and rack-1(tm2262) might be a hypomorph. However, RNAi did not worsen the low brood size or axon defects of rack-1(tm2262) (see below; data not shown), indicating that it might be a strong loss of function allele.
Similar to RNAi of rack-1, the deletion allele rack-1(tm2262) caused pathfinding defects in the VD and DD motor neurons (Figure 2). Normally, all VD/DD commissures extend on the right side of the animal except DD1/VD2, which form a single commissure in the anterior (arrow in Figure 2C). rack-1(tm2262) displayed VD/DD commissures aberrantly extending up the left side of the animal (Figure 1D), and VD/DD axons that were misguided on their dorsal migrations (Figure 1D). In our hands, 27% of wild type animals harboring the unc-25:gfp transgene juIs76 had VD/DD commissures on the left side in addition to DD1/VD2. However, 60% of rack-1(tm2262); juIs76 showed VD/DD commissures on the left side (p<0.001) (Figure 3). In juIs76 animals, generally only one or two left-side VD/DD were observed, whereas multiple axons on the left side were often observed in rack-1(tm2262); juIs76 animals (Figure 2D). In addition, 42% of rack-1(tm2262); juIs76 animals displayed VD/DD axon guidance and outgrowth defects such as axonal wandering, branching or termination (Figure 3), whereas juIs76 alone showed no strong defects but did display some minor axon wandering.
To ensure that the axon guidance defects observed in rack-1(tm2262) were due to rack-1 perturbation and not a background mutation, we rescued the VD/DD axon defects with a rack-1::myc transgene. rack-1::myc rescued both left-right defects and commissural guidance defects (60% to 32% (p<0.001) and 42% to 10% (p<0.001)) (Figure 3). Together, these results indicate that RACK-1 is required for VD/DD axon pathfinding.
rack-1(tm2262) animals were slow growing and had very low brood size. In a progeny count, ten wild type and ten rack-1(tm2262) animals were individually plated and then transferred to a new plate every day until egg laying ceased. The number of viable adult progeny resulting from each animal were counted and averaged. The average progeny count for a wild-type N2 animal was of 278.4 (s.d. = 32.72), while for rack-1(tm2262) the count dropped to 23.3 (s.d. = 9.9) (p<0.0001). A transgene containing the rack-1 gene under its native promoter fused to the gfp coding region (rack-1::gfp) (see Materials and Methods) increased brood size in rack-1(tm2262) animals to 73.78 (s.d. = 17.18) (p<0.0001), suggesting that rack-1::gfp was functional and could rescue the brood size defect in rack-1 animals. rack-1::myc could also rescue the low brood size of rack(tm2262) (data not shown). Thus, the reduction in brood size was due to rack-1 and not due to genetic background in the tm2262 strain.
The reduced brood size of rack-1(tm2262) seems to be predominantly due to decreased production of fertilized embryos. rack-1 might affect oogenesis or spermatogenesis, but the nature of this sterility has not been explored. Previous studies indicate that rack-1 also affects embryogenesis by regulating membrane trafficking and recycling endosome distribution via interaction with dynactin to control cytokinesis and chromosome separation in the early embryo [27].
In order to determine where the rack-1 gene is expressed, we constructed a reporter transgene consisting of the promoter region of rack-1 fused to the gfp coding region (see Materials and Methods). rack-1 promoter::gfp was expressed in most if not all tissues. Due to mitotic loss of the transgene-bearing extrachromosomal array, we were able to analyze rack-1 promoter::gfp expression in mosaic animals in which we could discern specific cell types. rack-1 promoter::gfp was expressed in neurons as well as the distal tip cells of the gonad (Figure 4A and 4B).
In order to determine the subcellular localization of RACK-1 protein, we constructed a full-length rack-1::gfp fusion. This transgene is predicted to encode a full-length RACK-1 protein with GFP at the C-terminus (RACK-1::GFP). rack-1::gfp rescued the sterility and gonadal distal tip cell migration defects of rack-1(tm2262) mutants. RACK-1::GFP was present in the cytoplasm of cells and showed little if any nuclear accumulation (Figure 4C and 4D), although low levels of RACK-1::GFP in the nucleus cannot be excluded. RACK-1::GFP was present in the growth cones of extending VD commissural axons, but was present in the axons and cell bodies as well (Figure 5A and 5B).
rack-1 was expressed in most if not all tissues in the animal, including neurons. To determine if RACK-1 is required in the VD/DD neurons themselves for axon pathfinding, we drove expression of rack-1::gfp specifically in the VD/DD neurons using the GABAergic neuron-specific unc-25 promoter. The wild-type rack-1(+) coding region lacking the upstream promoter region was fused to gfp downstream of the unc-25 promoter. The Ex[unc-25 promoter::rack-1::gfp] transgene was expressed specifically in the GABAergic neurons including the VD/DD neurons and nowhere else (Figure 5A). This transgene did not rescue the fertility defects and DTC migration defects of rack-1(tm2262) as did the genomic rack-1(+) transgene (data not shown), indicating that expression was specific to the VD/DD neurons. Ex[unc-25 promoter::rack-1(+)] rescued the lateral asymmetry defects and axon wandering defects of rack-1(tm2262) animals (Figure 3) (60% to 33% for lateral asymmetry defects and 16% to 8% for axon wandering defects; p<0.001 in both cases). In this experiment, individual axons were scored, due to the mosaic nature of the Ex[unc-25 promoter::rack-1(+)] transgene. These data indicate that rack-1 acts cell autonomously in neurons in axon pathfinding.
The above results show that RACK-1 physically interacted with UNC-115/abLIM and that rack-1 loss of function caused axon pathfinding defects similar to unc-115. Previous studies showed that UNC-115/abLIM acts downstream of the Rac GTPase CED-10/Rac and in parallel to MIG-2/RhoG in axon pathfinding [25], [46]. We next set out to determine if RACK-1 interacts with UNC-115/abLIM and the Rac GTPases in axon pathfinding. To analyze genetic interactions between these molecules, we used the PDE neurons, which are located at the post-deiridic region of the animal. These neurons are a good model for axon pathfinding since the reporter construct osm-6::gfp is expressed only in the PDEs in the post-deirid [25], [47], allowing unambiguous identification and scoring of the simple PDE axon morphology. Furthermore, the defects in PDE axon pathfinding in single mutants were weak, allowing for discrimination of genetic interactions in double mutants.
In wild type, the PDE cell body extends an axonal projection toward the ventral nerve cord in a straight line, where the axon then branches and extends anteriorly and posteriorly (Figure 6A) [43]. Pathfinding defects were defined as axons that were prematurely terminated or that wandered at a greater than 45 degree angle relative to the normal PDE axon (for example, Figure 6B). As shown previously, mig-2(mu28), ced-10(n1993), and unc-115(ky275), alone had low-penetrance defects in PDE axon pathfinding on their own (3%–7%; Figure 6C). We found that rack-1(tm2262) also had very few defects in PDE axon pathfinding (1%; Figure 6C).
Previous results show that CED-10/Rac and MIG-2/RhoG act redundantly in PDE axon pathfinding, and UNC-115/abLIM works downstream of CED-10/Rac, in parallel to MIG-2/RhoG in PDE pathfinding [25], [48]. If RACK-1 works in the same pathway as UNC-115/abLIM, we expect that loss of function of both rack-1 and unc-115 would be no more severe than either mutant alone. Indeed, rack-1(tm2262M+); unc-115(ky275) double mutants (M+ denotes that the homozygous animal was derived from a balanced heterozygote and has wild-type maternal contribution) displayed levels of PDE axon pathfinding defects (6%; Figure 6C) that were not significantly different from unc-115(ky275) and rack-1(tm2262) alone, suggesting that UNC-115/abLIM and RACK-1 might act in the same pathway. In contrast, rack-1(tm2262M+); mig-2(mu28) double mutants showed significantly increased levels of defects compared to either single alone (28%; Figure 6C). This result demonstrates that rack-1(tm2262) can synergize with other mutants to cause axon defects, and that RACK-1 and MIG-2/RhoG might act in parallel pathways in axon pathfinding.
CED-10/Rac and UNC-115/abLIM have previously been shown to act in the same pathway in parallel to MIG-2/RhoG [25]. rack-1(tm2262M+) ced-10(n1993M+) double mutants displayed no significant increase in PDE defects compared to either single alone (6%; Figure 6C), consistent with the idea that RACK-1, CED-10/Rac, and UNC-115/abLIM act in a common pathway in parallel to MIG-2/RhoG in axon pathfinding. If this is the case, we would expect the rack-1(tm2262M+) ced-10(n1993M+); unc-115(ky275) triple mutant to be no more severe than any double mutant combination alone. As previously reported, ced-10(n1993); unc-115(ky275) double mutants were significantly more severe than either single alone (35%; Figure 6C) [25], [48]. This is likely due to the fact that CED-10/Rac also regulates the Arp2/3 complex in parallel to UNC-115/abLIM [16], [18]. rack-1(tm2262M+) ced-10(n1993M+) mutants did not show this interaction. Possibly, RACK-1 is not the only molecule regulating UNC-115 and has a weaker effect. In any case, the rack-1(tm2262M+) ced-10(n1993M+); unc-115(ky275) triple mutant was not significantly more severe than ced-10(n1993); unc-115(ky275) alone (34% compared to 35%; Figure 6C). Taken together, these results are consistent with the idea that RACK-1, CED-10/Rac, and UNC-115/abLIM act in a common pathway in axon pathfinding in parallel to MIG-2/RhoG.
UNC-34/Enabled has been shown to act in parallel to both CED-10/Rac and MIG-2/RhoG in axon pathfinding [15]. Indeed, rack-1(tm2262M+); unc-34(e951M+) double mutants displayed significantly increased pathfinding defects compared to unc-34(e951) alone (35% compared to 19%; Figure 6C). This result indicates that RACK-1 acts in parallel to UNC-34/Enabled and is consistent with RACK-1 acting with CED-10/Rac and UNC-115/abLIM in axon pathfinding.
Loss-of-function studies described above provide evidence that RACK-1 might act with CED-10/Rac and UNC-115/abLIM in axon pathfinding. In order to further test the relationship between RACK-1 and the Rac GTPases we next asked what effect rack-1(tm2262) loss of function might have on overactive Rac GTPases. Constitutively-activated Rac GTPases transgenes harbor a guanine-12-valine mutation in the GTPase binding pocket, which favors the active GTP-bound state of the GTPases. Previous studies showed that CED-10(G12V) and MIG-2(G16V) (the G12V equivalent) both caused axon pathfinding defects and drove the formation of ectopic neurites, lamellipodia, and filopodia when expressed in PDE neurons (Figure 7A), and that UNC-115/abLIM was required for ectopic lamellipodia and filopodia induced by CED-10(G12V) but not MIG-2(G16V) [25].
We determined if RACK-1 was required for the effects of CED-10(G12V) and MIG-2(G16V). CED-10(G12V) alone caused 66% of PDE neurons to have ectopic lamellipodia and filopodia in young adults (Figure 7B). rack-1(tm2262); ced-10(G12V) animals displayed 45% ectopic lamellipodia and filopodia, a significant reduction (p = 0.004) from CED-10(G12V) alone (Figure 7B). These data indicate that rack-1(tm2262) partially suppressed activated CED-10(G12V) and that functional RACK-1 might be required for the formation of ectopic lamellipodia and filopodia induced by activated CED-10. In contrast, rack-1(tm2262) did not suppress ectopic lamellipodia and filopodia associated with MIG-2(G16V) and in fact slightly enhanced these defects (Figure 7B; p = 0.03), indicating that this suppression is specific to CED-10(G12V). These effects are similar to those observed with the wve-1/WAVE mutant, which suppressed CED-10(G12V) and slightly enhanced MIG-2(G16V) [18]. These data are consistent with the idea that RACK-1 acts downstream of CED-10/Rac in parallel to MIG-2/RhoG in axon pathfinding, similar to UNC-115/abLIM.
Previous studies showed that UNC-115 tagged with an N-terminal myristylation sequence (MYR) caused activation of the molecule [49]. MYR::UNC-115 localized to the plasma membrane and other membranes as expected for a myristylated protein and induced the formation of ectopic lamellipodia, filopodia and neurites in C. elegans neurons and in cultured mammalian fibroblasts [49]. The formation of these ectopic lamellipodia and filopodia was dependent upon the actin-binding domain of UNC-115, suggesting that the molecule was constitutively active [49].
To further dissect the interaction of RACK-1 with UNC-115, we assayed the effects of MYR::UNC-115 in a rack-1(tm2262) loss of function background. MYR::UNC-115 was expressed from the unc-115 promoter (the lqIs62 transgene), which drives expression in most neurons including PDE and the VD/DDs [49]. The myr::unc-115 transgene scored in [49] was maintained as an extrachromosomal array. We integrated this transgene into the genome for these studies (lqIs62). We found that lqIs62[MYR::UNC-115] caused 8% ectopic lamellipodia and filopodia in PDE neurons (Figure 7B), similar to but weaker than the extrachromosomal array effects reported in [49]. The ectopic lamellipodia and filopodia induced by MYR::UNC-115 were not significantly altered by rack-1(tm2262) mutation (Figure 7B), indicating that RACK-1 is not required for lamellipodia and filopodia induced by MYR::UNC-115. This result suggests that RACK-1 might act upstream of UNC-115 or together with UNC-115, or that the MYR::UNC-115 molecule acts independently of RACK-1 activity.
To study the interactions of rack-1 with myr::unc-115 in more detail, we analyzed the VD/DD motor neurons as described above. myr::unc-115 expression caused left-right lateral asymmetry defects and commissural axon pathfinding defects as described for rack-1(tm2262) in Figure 2 and Figure 3 (Figure 8A). rack-1(tm2262); myr::unc-115 animals displayed lateral asymmetry defects similar to each alone (Figure 8A; 45%–55%, not significant). This is consistent with RACK-1 acting upstream of or together with UNC-115 in the same pathway. In VD/DD commissural axon pathfinding, rack-1(tm2262); myr::unc-115 displayed significantly increased defects compared to the additive effects of each alone (Figure 8A). Thus, rack-1(tm2262) might enhance myr::unc-115 in VD/DD pathfinding.
In summary, we detected no strong suppression of myr::unc-115 by rack-1(tm2262) in the PDE neurons and the VD/DD neurons. The results are consistent with RACK-1 acting upstream of or together with UNC-115 in axon pathfinding. Some context-specific interactions were observed, such as rack-1(tm2262) enhancing VD/DD commissural axon pathfinding, indicating that RACK-1 and UNC-115 might have distinct interactions in different contexts or developmental events.
The above results indicate that RACK-1 is not required for the effects of MYR::UNC-115, suggesting that RACK-1 might act upstream of UNC-115. To test this idea, we constructed a myristylated version of RACK-1, similar to MYR::UNC-115. We reasoned that constitutive membrane localization might activate RACK-1 as it does UNC-115. myr::rack-1::gfp was expressed in the PDE neurons by the osm-6 promoter. MYR::RACK-1::GFP displayed a membrane-associated distribution (arrowhead in Figure 9A), as did MYR::UNC-115 [49]. MYR::RACK-1 animals displayed ectopic lamellipodial protrusions along the cell body, dendrite, and axon, similar to MYR::UNC-115 and activated Rac GTPases (arrow in Figure 9A). The putative null unc-115 alleles ky275 and ky274 suppressed this effect (11% in myr::rack-1 reduced to 0% and 4% in unc-115(ky275); myr::rack-1 and unc-115(ky274); myr::rack-1, respectively) (Figure 9B). The hypomorphic unc-115(mn481) allele [17], which retains some UNC-115 activity, did not suppress myr::rack-1, indicating that possibly only a small amount of UNC-115 activity is required for MYR::RACK-1 to drive ectopic lamellipodia. These studies support the model that RACK-1 acts upstream of UNC-115 in lamellipodia formation.
Despite the strong genetic interactions of rack-1 and unc-115, we could detect no change in distribution of UNC-115::GFP in loss of function rack-1(tm2262) or in the activated myr::unc-115 transgenics (data not shown). In each case, UNC-115::GFP was present uniformly throughout the cytoplasm, similar to wild-type animals, and showed no membrane localization.
Activated myr::unc-115 caused lateral displacement of GABAergic motor neuron cell bodies such that they often were found outside of the ventral nerve cord (Figure 8B). The VD GABAergic neurons are descendants of the P cells. The P cells are born laterally and the P nuclei migrate ventrally to the ventral nerve cord, where the P cells divide to produce ventral hypodermal cells including the vulva and the ventral VD neurons [50]–[51]. Failure of the ventral migration of the P nuclei can result in laterally displaced VD neuron cell bodies. This phenotype is observed in mig-2; ced-10 double mutants, but not in unc-115 mutants [48]. Possibly, ectopic activity from MYR::UNC-115 impedes P nucleus migration. rack-1(tm2262) suppressed the displaced VD cell body defect of myr::unc-115 (Figure 8A): 30% of myr::unc-115 animals had misplaced VD cell bodies compared to 15% of rack-1(tm2262); myr::unc-115 (p = 0.011). This result suggests that RACK-1 might act downstream of or participate with MYR::UNC-115 in impeding P nucleus migration, again suggesting context-dependent interactions of UNC-115 and RACK-1.
The C. elegans gonad is derived from two somatic cells (Z1 and Z4) surrounding the two germ cells (Z2 and Z3) [52]. Z1 and Z4 divide to produce the somatic cells of the gonad. Before morphogenesis, the gonad is oval shaped and located ventrally in the middle of the animal. The distal tip cells (DTCs) at the anterior and posterior tips of the gonad begin migration, and as they migrate they lead the gonad behind them. The DTCs migrate anteriorly and posteriorly, turn dorsally and migrate to the dorsal region of the animal, and then migrate posteriorly and anteriorly back toward the middle of the animal. DTC migration results in the U-shaped bi-lobed gonad of C. elegans (Figure 10A). If DTC migration is perturbed, misrouted and misshapen gonads result. The gonads of 32% of rack-1(tm2262) animals were misrouted (Figure 10B and 10C). Misrouting defects included failure to turn dorsally as well as extra turns, such as turning back ventrally after the dorsal migration. We did not observe gonads that had extended past their normal stopping point near the middle of the animal, as has been observed in other mutations that affect DTC migration [53], [54]. A rack-1::gfp transgene rescued gonad misrouting defects in rack-1(tm2262) (32% to 5%; p<0.005) (Figure 10C), indicating that the gonad defects were due to rack-1 perturbation. It should be noted that rack-1(tm2262) homozygotes from a heterozygous mother (rack-1(tm2262M+)) had less severe DTC migration defects compared to the rack-1(tm2262) animals without maternal contribution (21% compared to 32%; p = 0.03). This indicates that DTC migration defects were partially rescued by wild-type maternal rack-1(+) activity. unc-115 mutants displayed no defects in DTC migration, and the gonad defects of rack-1 were not affected by unc-115 (data not shown). rack-1::myc also rescued the DTC migration defects of rack-1(tm2262) (data not shown). Thus, DTC migration is controlled by RACK-1 and is independent of UNC-115/abLIM.
CED-10/Rac and MIG-2/RhoG have previously been shown to control gonad distal tip cell migration [48], and we have shown here that RACK-1 also controls DTC migration. To determine if RACK-1 interacts with Rac signaling in DTC migration, we analyzed DTC migration in double mutants.
As reported previously, both ced-10(n1993) and mig-2(mu28) mutations caused defects in DTC migration (27% for mig-2(mu28), 12% for ced-10(n1993)) (Figure 10C) [48]. We found that double mutants of ced-10 and mig-2 with rack-1 showed no significant difference in defects compared to the stronger singles alone. rack-1(tm2262M+); mig-2(mu28) showed no significant difference (26%) compared to mig-2(mu28) (27%), and rack-1(tm2262M+) ced-10(n1993) showed no significant difference compared to rack-1(tm2262M+) (21% in each case).
CED-10/Rac and MIG-2/RhoG interact differently in the DTCs than they do in other tissues such as axons, in which they act in parallel redundant pathways. ced-10; mig-2 double mutants did not display enhanced DTC migration defects compared to either single alone [48], suggesting that they might act in the same pathway or in independent pathways that each control a distinct aspect of DTC migration. Our results suggest the same for RACK-1, that it might act in a pathway independent of CED-10 and MIG-2, or that CED-10, MIG-2 and RACK-1 might all act in the same pathway in DTC migration.
In summary, we have presented data indicating that RACK-1 is required cell autonomously for axon pathfinding, and that RACK-1 is required for migration of the distal tip cells of the gonad. We show that RACK-1 interacts physically with UNC-115/abLIM, and that RACK-1 and UNC-115/abLIM might act in the same pathway in axon pathfinding. Consistent with this idea, RACK-1 was required for ectopic lamellipodia and filopodia induced by the activated CED-10/Rac GTPase, similar to UNC-115/abLIM. RACK-1-like molecules have been implicated in a wide variety of cellular events and have been shown to interact with a large number of distinct protein complexes, consistent with the idea of RACK molecules as scaffolds and integrators [32]. RACK molecules have been shown to control cell adhesion and migration and interact with Src, c-Abl, Rho GTPase regulators, and integrins in these events [31], [33], [39]. Our studies here suggest that RACK-1 interacts with the actin-binding protein UNC-115/abLIM and Rac GTPases in the control of axon pathfinding and cell migration.
Previous studies showed that in C. elegans, RACK-1 depletion by RNAi resulted in embryos with cytokinesis defects, including shorter astral microtubules, defects in chromosome separation, and defects in membrane organization and recycling endosome distribution [27]. We found that the deletion rack-1(tm2262) was very sick and slow growing, and gave very few progeny. rack-1(tm2262) produced very few embryos, suggesting that the animals had defects in sperm and/or oocyte production. Indeed, rack-1 RNAi resulted in defects in germline membrane organization [27], consistent with the sterility that we observed in the rack-1(tm2262) mutant.
rack-1(tm2262) mutants displayed a variety of axon pathfinding defects, including left-right choice and guidance defects of the VD and DD commissural motor axons and guidance defects of the PDE axons. VD/DD axons are dorsally directed, and PDE axons are ventrally directed, indicating that rack-1 is not specific for any particular guidance direction. VD/DD axon pathfinding defects were rescued when rack-1(+) was expressed under a promoter that specifically drives expression in the GABAergic neurons including VD/DD and nowhere else, demonstrating that RACK-1 is required cell-autonomously for axon pathfinding. As RACK-1 is likely involved in many different developmental events, this result shows that the effects of RACK-1 on axon pathfinding are due to defects in the neuron itself and not a substrate or guidepost tissue such as the hypodermis or other neurons. Indeed, RACK-1 was expressed in neurons, and functional RACK-1::GFP fusion protein accumulated in the growth cones of neurons, consistent with a role of RACK-1 in growth cone cytoskeletal regulation. RACK-1::GFP also accumulated in the cell bodies and axons of neurons.
We have shown that rack-1(tm2262) mutants display defects in the structure of the gonad arms consistent with a defect in distal tip cell migration. rack-1 is expressed in the migrating distal tip cells. The Rac GTPases CED-10/Rac and MIG-2/RhoG also each affect distal tip cell migration, but do not show the phenotypic synergy in DTC migration as is observed in axon pathfinding [48]. Thus, CED-10/Rac and MIG-2/RhoG might act in independent pathways that control distinct aspects of DTC migration.
DTC migration defects in mig-2; rack-1 and ced-10 rack-1 double mutants were not significantly different than the stronger single mutants alone. This result suggests that RACK-1 might act in a pathway independent of MIG-2 and CED-10, or that all three act in a common pathway. This again points to context dependent differences in the function of RACK-1 and suggests that RACK-1 might interact with different effectors in different ways in different cells and cellular events. Indeed, the effect of RACK-1 on DTC migration is likely to be independent of UNC-115/abLIM, as unc-115 mutants have no effect on DTC migration alone or in any double mutant combination analyzed so far, including ced-10 and mig-2.
A model of RACK-1 interaction with CED-10/Rac and UNC-115/abLIM is shown in Figure 10A. Double mutant analysis showed that rack-1(tm2262) synergized with mig-2/RhoG and unc-34/Enabled in PDE axon pathfinding, similar to unc-115/abLIM and ced-10/Rac. rack-1(tm2262) did not synergize with unc-115/abLIM or ced-10/Rac, consistent with the idea that they act in the same pathway in parallel to mig-2/RhoG and unc-34/Enabled.
Activated CED-10(G12V) drives the formation of ectopic lamellipodia and filopodia in PDE neurons, and unc-115 loss of function suppresses this effect [25]. We show here that rack-1(tm2262) also partially suppressed ectopic lamellipodia and filopodia caused by CED-10(G12V), indicating that RACK-1 is required downstream of CED-10/Rac in lamellipodia and filopodia formation (Figure 10A). This result suggests that RACK-1 might normally be required for lamellipodia and filopodia formation. This is in contrast to the seven-WD repeat protein SWAN-1, which physically interacts with the UNC-115 LIM domains and with CED-10/Rac but which is normally required to inhibit CED-10/Rac signaling in lamellipodia and filopodia formation [26]. Thus, these two seven-WD repeat proteins SWAN-1 and RACK-1 might have opposite effect on CED-10/Rac signaling: SWAN-1 inhibits it, and RACK-1 is required downstream of it to form lamellipodia and filopodia. That RACK-1 is required for lamellipodia and filopodia formation downstream of CED-10/Rac suggests that RACK-1 might be acting directly in cytoskeletal regulation. It is also possible that RACK-1 exerts its effects downstream of Rac GTPases through transcriptional or translational control, but the fact that RACK-1 interacts physically with the actin-binding protein UNC-115/abLIM supports the idea that RACK-1 directly controls cytoskeletal signaling.
RACK-1 physically interacts with UNC-115/abLIM and genetically acts in the same pathway in axon pathfinding. UNC-115 can be activated constitutively by the addition of an N-terminal myristylation sequence [25], which mediates the covalent attachment of a fatty acid myristyl residue to the protein and drives localization to membranes, including the plasma membrane. MYR::UNC-115 also drives the formation of ectopic lamellipodia and filopodia, similar to but weaker than CED-10(G12V) [25]. No strong suppression or enhancement of axon pathfinding defects were observed in double mutants of rack-1(tm2262) and myr::unc-115. One interpretation of these data is that RACK-1 does not act downstream of UNC-115/abLIM and instead might act together with or upstream of UNC-115/abLIM. Indeed, unc-115 mutations suppressed the ectopic lamellipodia caused by MYR::RACK-1, indicating that UNC-115 acts downstream of RACK-1. These results are consistent with a model in which RACK-1 acts downstream of CED-10/Rac and upstream of UNC-115/abLIM in axon pathfinding (Figure 11).
RACK-1 and UNC-115 displayed context-dependent interactions in addition to those described in the PDE neurons above. First, rack-1 slightly but significantly increased VD/DD commissural pathfinding defects caused by myr::unc-115. We do not understand the nature of the VD/DD axon pathfinding defects caused by MYR::UNC-115, but it is possible that they are due to excessive lamellipodial and filopodial protrusion, possibly in the growth cone. If this is the case, these effects were enhanced by rack-1 loss of function, suggesting that RACK-1 might negatively regulate MYR::UNC-115 in this context, possibly by excluding MYR::UNC-115 from regions in which it induces lamellipodia and filopodia.
Second, rack-1(tm2262) suppressed the lateral displacement of VD cell bodies caused by myr::unc-115. Laterally misplaced VD cell bodies are indicative of a defect in the ventral migration of the nuclei of the P cells. UNC-115 is not normally involved in P nucleus migration, but ced-10/Rac and mig-2/RhoG act redundantly in the process [48]. Possibly, myr::unc-15 ectopically interferes with P nucleus migration, and RACK-1 is required for this effect. In this case RACK-1 might normally positively regulate MYR::UNC-115. In any case, these data indicate that RACK-1 and UNC-115 might have distinct interactions in different cellular contexts.
In summary, these studies suggest that RACK-1 acts in a common pathway with CED-10/Rac and UNC-115/abLIM in axon pathfinding (Figure 11). These studies implicate the Receptor of Activated C Kinase as a new Rac GTPase effector molecule, as RACK-1 acts downstream of CED-10 and upstream of UNC-115/abLIM in axon pathfinding. Future studies will be directed at understanding the roles of plasma membrane localization and phosphorylation in the regulation of this pathway.
C. elegans culture and techniques were performed using standard protocols [55]–[56]. All experiments were performed at 20°C. The rack-1(tm2262) allele was provided to us by the National Bioresource Project for the Experimental Animal “Nematode C. elegans” (S. Mitani), and was outcrossed to wild-type N2 animals three times before analysis. Polymerase chain reaction (PCR) was used to verify the homozygosity of rack-1(tm2262) in strains. The following mutations and genetic constructs were used: LGII: juIs76[unc-25::gfp]; LGIV: ced-10(n1993), rack-1(tm2262), nT1 IV:V, lqIs3[osm-6::gfp]; LGV: unc-34(e951); LGX: unc-115(ky275), mig-2(mu28), lqIs2[osm-6::gfp]; LG?: lqIs62[myr::unc-115(+)]. C. elegans transformation was performed by standard techniques using DNA microinjection into the syncytial germline of hermaphrodites [57]. Transgenes were integrated into the genome using trimethylpsoralen and standard techniques [58]–[59].
All micrographs were obtained on a Leica DMRE microscope with a Qimaging Rolera MGi EMCCD camera or a Qimaging Retiga CCD camera. Openlab and IPlab software were used to obtain images.
All coding regions amplified by PCR were sequenced to ensure the absence of mutations in the sequence. PCR, recombinant DNA and other molecular biology techniques were performed according to standard techniques [60]. Primer and plasmid sequences are available upon request.
Axon pathfinding defects were scored with fluorescence microscopy of hermaphrodite animals in the fourth larval stage (L4) or young adults expressing a green fluorescent protein transgene for specific cells. To visualize and score the axons of VDs and DDs, animals harboring an unc-25 promoter::gfp integrated transgene (juIs76 II) were used [45]. To visualize and score PDE axons, animals harboring an osm-6 promoter::gfp integrated transgene (lqIs2 X or lqIs3 IV) were used [25], [47].
ced-10(G12V) and mig-2(G16V) transgenes under the control of the osm-6 promoter were used as described previously [25]. A myr::unc-115 transgene under the control of the unc-115 promoter was used as described previously [49].
Gonadal distal tip cell migration defects were scored by Differential Interference Microscopy in young adult hermaphrodite animals. Any deviation from the normal U-shape of gonad arms was scored as defective, including failure to migrate fully, failure to make a dorsal turn, failure to make an anterior or posterior turn, or extra dorsal-ventral or anterior-posterior turns. Significances of differences (p values) were determined using Fisher Exact Analysis.
A full-length rack-1(+) transgene was generated by PCR from genomic DNA (based on the Wormbase gene model K07D7.1) and included the entire upstream rack-1 region (∼2.5 kb), the coding region, and the downstream region past the poly-A addition site (Figure 1E). rack-1::gfp and rack-1::myc fusion constructs were generated by amplifying the entire rack-1 upstream region and coding region lacking the stop codon fused in frame to gfp or myc. The unc-25 promoter::rack-1::gfp fusion protein was generated by amplifying the rack-1 coding region lacking the upstream region. This fragment was placed downstream of the unc-25 promoter and fused in frame to gfp at the 3′ end.
The two-hybrid screen was conducted at the Molecular Interaction Facility at the University of Wisconsin-Madison (thanks to E. Maher). In a liquid multi-well format, approximately 36 million C. elegans cDNA clones representing both oligo dT and random-primed libraries were screened via mating. UNC-115 was fused to the GAL4 DNA-binding domain in the pBUTE plasmid and the prey cDNAs were fused to the GAL4 activation domain in the pACT plasmid. In the yeast strain, the bacterial lacZ gene and the HIS5 gene were under the control of a GAL4-regulated promoter. The interaction screen consisted of assaying β-galactosidase (β-gal) activity (for lacZ) and growth on 25 mM 3-aminotriazole (3-AT) (for HIS5). This analysis identified 244 potential interacting cDNAs that had β-gal activity and grew on 25 mM 3-AT. From these 244, 142 isolates activated both lacZ and HIS5 similarly when re-tested. Of these, 124 were bait-specific and did not activate when the bait plasmid was removed. These cDNAs were sequenced, and seven of these were found to represent the K07D7.1 gene in Wormbase (rack-1) [27].
In order to obtain large amounts of C. elegans protein extract, animals carrying an integrated rack-1::myc transgene were raised at room temperature in a liquid culture containing 2.5 mg cholesterol, 0.05 mg/mL streptomycin, Escherichia coli strain HB101 and M9 buffer (up to 500 mL). After about a week, these animals were harvested and snap-frozen in liquid nitrogen. We then added lysis buffer (1× PBS, 10% glycerol, 0.1% NP40, 0.1% Tween) in a 1∶1 ratio, and then 1 mM of phenylmethanesulphonylfluoride. We lysed the animals with glass beads in a beater for two cycles of 1 minute each. The supernatant was then collected and stored at −80°C for further experiments. We based our immunoprecipitation assays in Clonetech Laboratories' protocol No. PT3407-1 (Clonetech). We performed the standard immunoprecipitation assays as described in [26] using protein G (Zymed) and anti-Myc monoclonal antibody (Clontech). The anti-UNC-115 antibody is described in [26].
|
10.1371/journal.pcbi.1002219 | Receptive Field Inference with Localized Priors | The linear receptive field describes a mapping from sensory stimuli to a one-dimensional variable governing a neuron's spike response. However, traditional receptive field estimators such as the spike-triggered average converge slowly and often require large amounts of data. Bayesian methods seek to overcome this problem by biasing estimates towards solutions that are more likely a priori, typically those with small, smooth, or sparse coefficients. Here we introduce a novel Bayesian receptive field estimator designed to incorporate locality, a powerful form of prior information about receptive field structure. The key to our approach is a hierarchical receptive field model that flexibly adapts to localized structure in both spacetime and spatiotemporal frequency, using an inference method known as empirical Bayes. We refer to our method as automatic locality determination (ALD), and show that it can accurately recover various types of smooth, sparse, and localized receptive fields. We apply ALD to neural data from retinal ganglion cells and V1 simple cells, and find it achieves error rates several times lower than standard estimators. Thus, estimates of comparable accuracy can be achieved with substantially less data. Finally, we introduce a computationally efficient Markov Chain Monte Carlo (MCMC) algorithm for fully Bayesian inference under the ALD prior, yielding accurate Bayesian confidence intervals for small or noisy datasets.
| A central problem in systems neuroscience is to understand how sensory neurons convert environmental stimuli into spike trains. The receptive field (RF) provides a simple model for the first stage in this encoding process: it is a linear filter that describes how the neuron integrates the stimulus over time and space. A neuron's RF can be estimated using responses to white noise or naturalistic stimuli, but traditional estimators such as the spike-triggered average tend to be noisy and require large amounts of data to converge. Here, we introduce a novel estimator that can accurately determine RFs with far less data. The key insight is that RFs tend to be localized in spacetime and spatiotemporal frequency. We introduce a family of prior distributions that flexibly incorporate these tendencies, using an approach known as empirical Bayes. These methods will allow experimentalists to characterize RFs more accurately and more rapidly, freeing more time for other experiments. We argue that locality, which is a structured form of sparsity, may play an important role in a wide variety of biological inference problems.
| A fundamental problem in systems neuroscience is to determine how sensory stimuli are functionally related to a neuron's response. A popular mathematical description of this encoding relationship is the “cascade” model, which consists of a linear filter followed by a noisy nonlinear spiking process. The linear stage in this model is commonly identified as the neuron's spatiotemporal receptive field, which we will refer to simply as the receptive field (RF) or “filter”. The RF describes how a neuron sums up its inputs across space and time. It can also be conceived as the spatiotemporal stimulus pattern that optimally drives the neuron to spike. A large body of literature in sensory neuroscience has addressed the problem of estimating a neuron's RF from its responses to a rapidly fluctuating stimulus, a problem known generally as “neural characterization” [1]–[17].
Here we focus on a highly simplified encoding model that describes neural responses in terms of a linear filter and additive Gaussian noise [5], [11], [18]. Although this model gives an imperfect description of real neural responses, the RF estimators that arise from it (such as the spike-triggered average) are consistent under a much larger class of models [7], [19], [20]. The maximum likelihood filter estimate under the linear-Gaussian model is the whitened spike-triggered average (STA), also known as linear regression, reverse correlation, or the first-order Weiner kernel [1]–[3]. The STA has an extensive history in neuroscience and has been used to characterize RFs in a wide variety of areas, including retina [4], [7], [13], [21], [22], lateral geniculate nucleus [23], [24], primary visual cortex [5], [25], and peripheral as well as central auditory brain areas [8], [9], [11], [26]–[28].
The STA is often high-dimensional (containing tens to hundreds of parameters) and generally requires large amounts of data to converge. With naturalistic stimuli, the whitened STA is often corrupted by high-frequency noise because natural scenes contain little power at high frequencies. A common solution is to regularize the filter estimate by penalizing unlikely parameter settings, generally by biasing parameters towards zero (also known as “shrinkage”). Statisticians have long known that biased estimators can achieve substantially lower error rates in high-dimensional inference problems [29], [30], and Bayesian methods formalize such biases in terms of a prior distribution over the parameter space. In neuroscience applications, priors for sparse (having many zeros) or smooth (having small pairwise differences) filter coefficients have been used to obtain substantially more accurate RF estimates [9], [11], [12], [15], [31].
However, neural receptive fields are more than simply sparse or smooth. They are localized in both spacetime and spatiotemporal frequency. This is a structured form of sparsity: RFs contain many zeros, but these zeros are not uniformly distributed across the filter. Rather, the zeros tend to occur outside some region of spacetime and, in the Fourier domain, outside some region of spatiotemporal frequency. Although this property of receptive fields is well-known [32], [33], it has not to our knowledge been previously exploited for receptive field inference. Here we introduce a family of priors that can flexibly encode locality. Our approach is to first estimate a localized prior from the data, and then find the maximum a posteriori (MAP) filter estimate under this prior. This general approach is known in statistics as parametric empirical Bayes [34], [35]. Our method is directly inspired by previous empirical Bayes estimators designed to incorporate sparsity [36] and smoothness [11]. We show that locality can be an even more powerful source of prior information about neural receptive fields, and introduce a method for simultaneously inferring locality in two different bases, yielding filter estimates that are both sparse (local in a spacetime basis) and smooth (local in a Fourier basis).
The results section is organized as follows. First, we will describe the linear-Gaussian encoding model and the empirical Bayes framework for receptive field estimation. Second, we will review several previous empirical Bayes RF estimators, to which we will compare our method. Third, we will derive three new receptive field estimators that we collectively refer to as automatic locality determination (ALD). We will apply ALD to simulated data and to neural data recorded in primate V1 and primate retina. Finally, we will describe an extension from empirical Bayes to “fully Bayesian” inference under the ALD prior.
A typical neural characterization experiment involves rapidly presenting stimuli from some statistical ensemble and recording the neuron's response in discrete time bins. Let denote the (vector) stimulus and the neuron's (scalar) spike response at time bin . Here, is a vector of spacetime stimulus intensities over some preceding time window that affects the spike response at time bin .
We will model the neuron's response as a linear function of the stimulus plus Gaussian noise:(1)where denotes the neuron's receptive field and is a sample of zero-mean, independent Gaussian noise with variance . This model is the simplest type of cascade encoding model (depicted in Fig. 1 A), and plays an important role in the theory of neural encoding and decoding [5], [11], [17], [28], [37], [38]. For a complete dataset with stimulus-response pairs, likelihood is given by(2)where is a column vector of neural responses and is the stimulus design matrix, with 'th row equal to . The maximum likelihood (ML) receptive field estimate is:(3)This estimate, also known as the whitened spike-triggered average, and is proportional to the ordinary spike-triggered average if the stimulus ensemble is uncorrelated, meaning .
A major drawback of the maximum likelihood estimator is that it typically requires large amounts of data to converge, especially when is high-dimensional. This problem is exacerbated for correlated or naturalistic stimulus ensembles, because the high-frequency components of are not well constrained by the data. In the Bayesian framework, regularization is formalized in terms of a prior distribution , which tells us that we should bias our estimate of toward regions of parameter space that are more probable a priori. The posterior distribution, which captures the combination of likelihood and prior information, is given by Bayes' rule:(4)The most probable filter given the data and prior is known as the maximum a posteriori (MAP) estimator:(5)The log prior behaves as a “penalty” on the solution to an ordinary least-squares problem, forcing a tradeoff between minimizing the sum of squared prediction errors and maximizing .
Biased estimators can achieve substantial improvements over the maximum likelihood, particularly for high-dimensional problems, without giving up desirable features such as consistency (i.e., converging to the correct value in the limit of infinite data). However, the important question arises: how should one select a prior distribution? (Choosing the wrong prior can certainly lead to a worse estimate!)
One common method is to set the prior (or “penalty”) by cross-validation. This involves dividing the data into a “training” and “test” set, and selecting the prior for which (estimated on the training set) achieves maximal performance on the test set. However, this approach is computationally expensive and may be intractable for a prior with multiple hyperparameters. Empirical Bayes is an alternative method for prior selection that does not require separate training and test data.
Empirical Bayes can be viewed as a maximum-likelihood procedure for estimating the prior distribution from data. It is also known in the literature as evidence optimization, Type II maximum likelihood, and maximum marginal likelihood [11], [34], [39]–[41]. The basic idea is that we can compute the probability of the data given a set of hyperparameters governing the prior by “integrating out” the model parameters. This probability is really just a likelihood function for the hyperparameters, so maximizing it results in a maximum-likelihood estimate for the hyperparameters. (Technically, this is parametric empirical Bayes, since we will assume a particular parametric form for the prior; see [34], [35], [42] for a more general discussion).
Let denote a set of hyperparameters controlling the prior distribution over , which we will henceforth denote . The posterior distribution over the RF (eq.4) can now be written:(6)The denominator in this expression is known as the evidence or marginal likelihood. (Note that we ignored this denominator when finding the MAP estimate (eq.5), since it does not involve ). The evidence is the probability of the responses given the stimuli and the hyperparameters , which we can compute by integrating the numerator (eq.6) with respect to :(7)where is the parameter space for . Maximizing the evidence for therefore amounts to a maximum likelihood estimate of the hyperparameters. The MAP estimate for under this prior is an empirical Bayes estimate, since the prior is learned “empirically” from the data.
Empirical Bayes can therefore be described as a two-stage procedure: (1) Maximize the evidence to obtain ; (2) Find the MAP estimate for under the prior . Fig. 1 shows a diagram for this hierarchical receptive field model the steps for empirical Bayesian inference.
Following earlier work [11], [36], [43], [44], we will take the prior distribution to be a Gaussian centered at zero:(8)where is a covariance matrix that depends on hyperparameters in some yet-to-be-specified manner. This Gaussian prior together with a Gaussian likelihood (eq.2) ensures the posterior is also Gaussian:(9)where and are the posterior mean and covariance. The MAP filter estimate is simply the posterior mean , since the mean and maximum of a Gaussian are the same. Moreover, the evidence (eq.7) can be computed in closed form, since it is the integral of a product of two Gaussians. This allows for rapid optimization of . We will in practice maximize the log-evidence, given by:(10)where is the number of samples (rows) in and . All that remains is to specify the prior covariance , which we will explore in detail below.
Before continuing, we wish to distinguish two distinct notions of “dimensionality” for a receptive field. First, dimensionality may refer to the number of parameters or coefficients in . We will refer to this as the parameter dimensionality of the filter, denoted . Second, dimensionality may refer to the dimensionality of the coordinate space in which the filter is defined. In this sense, a filter with elements arranged as a vector is 1-dimensional (e.g., a temporal filter), while a filter with the same number of elements arranged in a matrix is 2-dimensional (e.g., an image filter). We will refer to this as the coordinate dimensionality of the filter, denoted .
We will examine three empirical Bayes RF estimators from the literature: ridge regression [45], Automatic Relevance Determination (ARD) [36], [43], [44], and Automatic Smoothness Determination (ASD) [11]. Fig. 2 provides an illustrative comparison of these methods, using a simulated example consisting with a 100-element vector filter (), stimulated with correlated (“1/F”) Gaussian noise stimuli. The true filter was a difference of two Gaussians, and the maximum likelihood estimate (middle left) is badly corrupted by high frequency noise.
First, ridge regression assumes a prior with covariance matrix proportional to the identity matrix: . This treats the filter coefficients as drawn i.i.d. from a zero-mean Gaussian prior with precision (“inverse variance”) . Ridge regression is penalized least-squares estimate with a penalty (eq.5) on the squared norm of the filter, given by . This penalty shrinks the coefficients of towards zero. Larger yields smaller filter coefficients, and in the limit of infinite , the MAP estimate shrinks to all-zeros. Set correctly, the ridge prior can provide substantial improvement over maximum likelihood, especially when the stimulus autocovariance is ill-conditioned, as it is for naturalistic stimuli (see Fig. 2). Ridge regression is perhaps the most popular and well-known regularization method. Although it is not usually employed in an empirical Bayes framework, it is straightforward (and fast) to maximize the evidence for the ridge parameter using a fixed-point rule [36], [45]. (See Methods).
Second, Automatic Relevance Determination (ARD) [36] assumes a diagonal prior covariance matrix with a distinct hyperparameter for each element of the diagonal. This resembles the ridge prior covariance except that the prior variance of each filter coefficient is set independently. The prior covariance matrix can be written , where ranges over the number of elements in . It would be intractable to use cross-validation to estimate all the elements in (a 100-element vector in Fig. 2), so empirical Bayes plays a critical role for inference. In practice, evidence maximization drives many of the prior variances to zero, making the posterior a delta function at zero for those coefficients. The MAP estimate for these coefficients is therefore zero, making the ARD estimate sparse. The ARD estimate can be computed rapidly using fixed-point methods, expectation-maximization, or variational methods [43], [44], [46]–[49]. Fig. 2 (middle column) shows the ARD and the lasso estimate [50], the latter of which is the MAP estimate under an exponential (or ) prior. We set the lasso parameter here by cross-validation. Both estimates are sparse. The ARD estimate is actually sparser and less biased towards zero for large coefficients, but both fail to provide a close match to the smooth filter used in this example.
Third, Automatic Smoothness Determination (ASD) [11] assumes a non-diagonal prior covariance, given by a Gaussian kernel [51], which is parametrized so that the correlation between filter coefficients falls off as a function of their separation distance. The rationale here is that RFs are smooth in both space and time, so nearby coefficients should be highly correlated, while more distant ones should be more nearly independent. For a 1D filter, the ASD prior covariance takes the form of a “fuzzy ridge”, with Gaussian decay on either side of the diagonal. The 'th element is given by , where is the squared distance between the filter coefficients and in pixel space, and the hyperparameters control the scale (analogous to the ridge parameter) and smoothness (the width of the fuzzy ridge), respectively. For filters with higher coordinate dimension (e.g., a 2D spatial filter), the hyperparameters include additional hyperparameters to control smoothness in each direction. Optimization of can be achieved by gradient ascent of the log-evidence (see Methods). For our simulated example (Fig. 2, bottom middle), the ASD estimate is indeed smooth due to the correlations in the inferred prior.
Note that for smooth RFs, the ASD prior covariance matrix becomes ill-conditioned, as some of its eigenvalues are very close to zero. This implies that the ASD estimate is sparse, but (unlike ARD) it is not sparse in the pixel basis. Rather, the ASD estimate is sparse in a basis that depends on the hyperparameters (since the eigenvectors of the ASD prior covariance vary with the hyperparameters). The small-eigenvalue eigenvectors tend to have high-frequency oscillations, meaning that the ASD estimate is sparse in a Fourier-like basis, with the prior variance of high-frequency modes set near to zero. In our view, ASD is the current state-of-the-art method for linear filter estimation and indeed (as shown in Fig. 2) it performs far better than previous methods for realistic neural RFs.
The motivation for our approach is the observation that neural receptive fields tend to be localized in space, time, and spatiotemporal frequency (i.e., Fourier space). Neurons in the visual pathway, for example, tend to integrate light only within some restricted region of visual space and some finite window of time, and respond only to some finite range of spatiotemporal frequencies [25], [32], [52], [53]. This is tantamount to a structured form of sparsity: large groups of coefficients (e.g., those outside some spacetime region) that fall to zero in a dependent manner. Here we describe three prior distributions for exploiting this structure. We refer to these methods collectively as automatic locality determination (ALD).
To compare performance with previous receptive field estimators, we began with simulated data. We generated six different 2D spatial receptive fields with varying degrees of locality in space and frequency. Each filter consisted of a 2D array of pixels, making for a parameter space of dimensions. Noisy responses were simulated using 1600 samples of 1/F correlated Gaussian noise according to (eq.1). Results are shown in Fig. 4.
Each row of Fig. 4 shows one of the six filters, and the estimates provided by maximum likelihood (ML), ridge regression, ARD, ASD, and (highlighted in blue) ALDsf. The numbers in red below each estimate indicate the mean squared error between the true filter and the estimate. (We did not show ALDs or ALDf because ALDsf always performed best of the three new methods). The simulated examples included: (A) a large Gabor filter; (B) a small Gabor filter; (C) a retina-like center-surround RF; (D) a grid cell RF with several non-zero regions; (E) circularly windowed Gaussian white noise; and (F) a pure Gaussian white noise filter. The grid cell filter did not exhibit strong locality in space, while the windowed white noise did not exhibit locality in frequency, and the pure white noise filter did not exhibit locality in either space nor frequency. Nevertheless, the ALDsf estimate had the smallest error by a substantial margin for all examples except the white noise filter. For the white noise filter, the ridge prior (i.i.d. zero-mean Gaussian) was in fact the “correct” prior. For this example, the ASD and ALDsf estimates were not distinguishable from the ridge regression estimate, consistent with the expectation that both should default to the ridge prior when the evidence did not favor smoothness (ASD) nor locality (ALDsf).
We examined the convergence properties of the various estimators as a function of the amount of data collected. We simulated responses from the first filter in Fig. 4A according to (eq.1), using two kinds of stimuli: Gaussian white noise, and 1/F correlated Gaussian noise, which more closely resembles natural stimuli. The results (Fig. 5) show that the ALDsf estimate achieved the smallest error for both kinds of stimuli, regardless of the number of training samples. The upper plots in Fig. 5 show that for white noise stimuli, traditional estimators (ML and ridge regression) needed more than four times more data than ALDsf to achieve the same error rate. For naturalistic stimuli, traditional estimators needed twenty to thirty times more data. The bottom row of plots shows the ratio of the average mean-squared error (MSE) for each estimate to the average MSE for the ALDsf estimate, showing that the next best method (ASD) exhibits errors nearly 1.8 times larger than ALDsf.
Next, we compared the various estimators using neural data recorded from simple cells in primate V1 [53]. The stimuli consisted of 16 “flickering bars” aligned with each cell's preferred orientation. We took the receptive field to have a length of 16 time bins, resulting in a filter with two coordinate dimensions (spacetime), resulting in a -dimensional parameter space. Because the “true” filter was not known, we quantified performance using relative cross-validation error, defined as the prediction error on an 8-minute test set (See Methods). We varied the amount of data used for training, and performed 100 repetitions with randomly selected subsets of the full training data to obtain accurate estimates for each size training set.
Fig. 6 (left) shows ML, ridge regression and ALDsf estimates for an example cell with a 1, 2 or 4 minutes of training data. Numbers in red indicate the average cross-validation error of each estimate. Note that with only 1 minute of data, ALDsf performed nearly as well as ML and ridge regression with 4 minutes of data. The middle panel shows a summary of cross-validation error for each of the five empirical Bayes estimators discussed previously, as a function of the amount of training data. ALDsf once again achieved substantially lower error than other methods. The right panel shows how many times more data were required to achieve the same level of cross-validation error as ALDsf. On average, ALDsf required 1.7 times less data than the next best method (ASD) and five times less data than maximum likelihood.
Fig. 7 shows the ML and ALDsf estimates for all 16 V1 simple cells in the population obtained with 1 minute of training data, as well as the ML estimate obtained using all the data available for each cell (40 minutes of data, on average). Note that for ALDsf recovers the qualitative structure of these RFs even when the underlying RF structure is barely discernible in the 1-minute ML estimate. Also note that the population exhibits substantial variability in RF shape, with many neurons whose RFs would not be well described by a fixed parametric form such as a Gabor filter.
We examined a second dataset of retinal ganglion cells (RGCs) in primate retina, which stimulated with 2D spatiotemporal white noise (“binary flicker”) [54], [55]. The RFs considered had 3 coordinate dimensions (spacespacetime), and a 2500-dimensional parameter space ( pixels in space25 8.33 ms-bins in time). Fig. 8 shows the spatial (2D) and the temporal (1D) slices through the estimated 3D RFs (schematized at left). Even with only 1 minute of training data, the ALDsf estimate recovered the qualitative structure of the RF at all time points, including the filters' departure from spacetime separability (i.e., the center pixel has different timecourse than surround). By contrast, the ML estimate is indistinguishable from noise in many places, indicating that ALDsf can reveal qualitative structure that is not visible in the ML estimate. We examined 3 ON and 3 OFF RGCs, and found that error was 18 times higher in ML estimates and 6 times higher in ridge regression estimates than in ALDsf (where error was computed with respect to the ML estimate using a full 20 minutes of data).
How can we quantify uncertainty in a receptive field estimate? The error bars shown in Figs. 5 and 6 represent variability in across resampled or permuted datasets. However, we would like to be able to measure the uncertainty in a single estimate given a single set of training data. Given the hyperparameters , the model specifies a Gaussian posterior (eq.9) with mean and covariance . The diagonal of specifies the posterior variance for each element of , giving us 95% credible intervals (Bayesian confidence intervals) of the form(14)The interpretation of these credible intervals is that, given the data and , . More generally, for any unit vector , the credible interval of size () for the projection is , where is the inverse normal cumulative density function.
However, these credible intervals, and the associated Gaussian posterior for , are conditioned on maximum-evidence estimate of the hyper-parameters . These intervals fail to take into account uncertainty in , which may be substantial if the evidence is not tightly concentrated around its maximum. The true uncertainty in will therefore generally be greater than that captured by the posterior covariance .
To accurately quantify uncertainty, we may wish to perform fully Bayesian inference under the priors introduced above. Empirical Bayes (EB) inference can be interpreted as an approximate form of fully Bayesian (FB) inference in a hierarchical model [35], [45]. If we incorporate a prior over the hyperparameters at the top level of the graphical model shown in Fig. 1 B, also known as a hyperprior, we will have a complete hierarchical model of the neural response. The difference between EB and FB inference for comes down to the fact that the FB prior involves marginalizing over :(15)while the EB prior is just the conditional distribution . When are these priors equivalent or, more importantly, when do the EB and FB estimates agree?
The relationship between EB and FB inference can be understood by examining the posterior distribution over . The full posterior is(16)where is the posterior over given , and is proportional to the evidence (i.e. the exponential of (eq.10)) times the hyperprior:(17)where is a normalizing constant. Note that if the evidence is proportional to a delta function at its maximum, then the posterior over is itself a delta function, . The full posterior then reduces to(18)which is the EB posterior (i.e., the posterior over conditioned on ). Thus, EB and FB inference are identical when the evidence is proportional to a delta function, and the two methods will in general give similar results whenever the evidence is highly concentrated around its maximum [45]. In general, EB and FB estimates will always agree given enough data, since by central limit theorem, the evidence will concentrate around its maximum with variance that falls as . However, for finite datasets, the two may differ.
To examine the proximity of EB and FB estimates and credible intervals, we developed a sampling-based algorithm to perform FB inference under the ALD prior. The factorization shown in (eq.16) suggests an efficient method for sampling from via Markov Chain Monte Carlo (MCMC), using a Markov chain over the space of the hyperparameters whose stationary distribution is proportional to the evidence. The summary of the algorithm for sampling is as follows:(19)A nice feature of this approach is that the hyperparameters live in relatively low-dimension (e.g., for a 1D filter and for a 2D filter under ALDsf). The Markov Chain therefore only has to explore this low-dimensional space, instead of the high-dimensional space of , which contains tens to thousands of parameters in typical cases [57]. Samples are obtained by drawing from the Gaussian conditioned on each MCMC sample . These samples may be averaged to the posterior mean , also known as the Bayes Least-Squares estimate, and their quantiles provide credible intervals. (See Method).
Fig. 9 shows a comparison of EB and FB estimates and credible intervals for the 1D simulated example shown previously. The hyperprior was taken to be uniform over a large region (See Methods). For a small dataset, the FB credible intervals were noticeably larger than the EB credible intervals, as expected, owing to the effects of uncertainty in [35]. For larger datasets, this discrepancy was much smaller, and was smaller in general for ALDsf than ALDs or ALDf intervals. The EB and FB (Bayes least-squares) filter estimates, however, did not differ noticeably even for small amounts of data. Fig. 10 shows a comparison of EB and FB inference for the V1 neural data presented in Fig. 6. For small datasets, the FB credible intervals were larger than EB intervals, but cross-validation error did not differ noticeably across dataset sizes. This suggests that the higher computational cost of FB inference may not be justified unless one is interested in obtaining accurate quantification of uncertainty from a small or noisy dataset.
We have described a new family of priors for Bayesian receptive field estimation that seek to simultaneously exploit locality in spacetime and spatiotemporal frequency. We have shown that empirical Bayes estimates under a localized prior are more accurate than those obtained under alternative priors designed to incorporate sparsity and smoothness. Although the ALD prior does not explicitly impose sparseness or smoothness, the estimates obtained with realistic neural data were both sparse and smooth. Sparsity arises from the fact that pixels outside a central region fall to zero, while smoothness arises from the fact that Fourier coefficients outside some low-frequency region fall to zero. However, for a receptive field dominated by high frequency components, ALD should outperform ASD and other smoothed estimates (e.g., smooth RVM [47], fused lasso [58]), since it can also select regions centered on high frequencies.
We have also derived an algorithm for performing fully Bayesian inference under ALD, ASD, and ridge regression priors. The algorithm exploits the low-dimensionality of the hyperparameter space and the tractability of the evidence to perform MCMC sampling of the posterior over hyperparameters. The full prior takes the form of a Gaussian scale mixture [59], [60], a mixture of zero-mean Gaussians with covariances and mixing weights , resulting in a Gaussian posterior over given that is trivial to sample. MCMC sampling allows for the calculation of fully Bayesian credible intervals over RF coefficients, which we found to be systematically larger than empirical Bayesian intervals. Nevertheless, we found no differences in the quantitative performance of EB and FB receptive field estimates with either simulated or real neural data (Figs. 9 and 10). Of course, both intervals rely on the linear-Gaussian model of the neural response, which may be inaccurate in cases where the neural response noise is highly non-Gaussian (e.g., heavy-tailed).
More generally, this work highlights the advantages of locality as an additional source of prior information in biological inference problems. Shrinkage and sparsity have attracted considerable attention in statistics, and they have advantageous properties for a variety of high-dimensional inference problems [29], [50], [61], [62]. ALD exploits a stronger form of prior information, assuming that large groups of coefficients go to zero in a correlated manner. This may not hold for generic regression problems; for a sparse filter with randomly distributed non-zero coefficients, the ARD estimate substantially outperforms ALD (not shown), but such filters are unlikely to arise in neural systems.
Two general ideas that arise from ALD may be useful for thinking about statistical inference in other biological and non-biological systems. The first is the idea of exploiting an underlying coordinate system or topography. Whenever the regression coefficients can be arranged topographically (e.g., temporally, spatially, spectrally), it may be possible to design a prior that exploits dependencies within this topography using a small number of hyperparameters. This idea is central to ALD as well as to ASD, which uses the distances between RF pixels to set their prior correlation. But other coordinates and prior parameterizations are possible. For example, although ALD performs reasonably well for a simulated grid cell (Fig. 4 D), locality in space does not hold for grid cells, and a prior that exploits the “natural” parameters of grid cell responses (e.g., grid spacing, size, orientation, phase) might perform even better. Optimizing the hyperparameters governing such a prior is tractable with empirical Bayes. The second idea that arises from ALD is that of simultaneously constraining a set of regression coefficients in two (or more) different bases. The ALDsf method combines a local prior in a spacetime basis and a local prior in Fourier basis via a “sandwich matrix” (eq.13), which effectively applies prior constraints in series: first in spacetime and then in frequency. Another solution would be to combine the two priors symmetrically, e.g., using prior covariance . (This is the covariance that results from taking the product of the ALDs and ALDf Gaussian priors). We found this formulation to perform slightly worse on test data, but results were similar. Note that the sum of prior covariances would not achieve the desired goal of imposing the prior constraints simultaneously, since it would prune only those coefficients in the (effective) null space of both and . A large literature has examined regularization and feature selection in overcomplete dictionaries (e.g., “basis pursuit”) [62]–[65], but combining structured prior information defined in different bases poses an intriguing open problem.
One potential criticism of ALD is that the linear-Gaussian encoding model (eq.1) is overly simplistic. Despite its simplicity, this model has a long history in the neural characterization literature [5], [11], [18], and the estimators considered here are consistent (i.e., converge asymptotically) for responses generated by any linear-nonlinear response model, so long as the stimuli are elliptically symmetric and the expected STA is non-zero [20]. We addressed whether the linear-Gaussian modeling assumption undermines our results by re-analyzing the V1 simple cell data with maximally informative dimensions (MID) [66], an information-theoretic estimator that incorporates neural nonlinearities and Poisson spiking. The results (shown in Supporting Information (Text S1), Fig. S1), indicate that MID errors were large, comparable in size to those of the maximum likelihood (linear regression) estimate. Even when comparing to the MID filter computed from test data, ALDsf outperformed MID by a substantial margin. This shows that the limitations of the linear-Gaussian model do not substantially undermine its performance on simple cells. However, we have applied ALD only to neurons whose responses exhibit a quasi-linear relationship to the stimulus. ALD would indeed fail for a neuron with a symmetric nonlinearity (e.g., squaring) and cannot recover multiple filters (e.g., those driving a complex cell). A variety of techniques exist estimating multi-dimensional feature spaces (e.g., spike-triggered covariance (STC) [67]–[69], MID [20], [66], iSTAC [70], spike-triggered ICA [71]). However, the “kernel trick” [17], [41], which involves using linear methods on nonlinearly transformed stimuli, provides the simplest method for extending ALD to nonlinear response models. Many nonlinear transformations (e.g., transforming the stimulus to its Fourier power [72]) preserve the topography of the underlying stimulus, making this approach directly applicable to ALD.
One advantage of the linear-Gaussian model is its computational tractability. ALD is fast because the evidence can be calculated and optimized entirely from the sufficient statistics , , and (the raw stimulus covariance, the STA, and sum of squared responses, respectively). This means that the computational cost does not scale with the amount of data (unlike MID and maximum-likelihood point process methods). Evidence optimization is also much faster than cross-validation, particularly with the hyperparameters employed by ALDsf. The computational cost of ALD is still at least in the number of filter coefficients, since evidence evaluation requires left-division by matrices of size . However, the number of approximately zero coefficients often falls considerably during optimization, and eliminating these coefficients by thresholding small eigenvalues of can speed convergence considerably.
Given the hyperparameters, the log-posterior over is concave, with a single maximum that can be computed in closed form (eq.5). Although the log-evidence (eq.10) is not concave in the hyperparameters , there are far fewer hyperparameters than parameters, making ALD far easier than non-convex optimization in the full space of (e.g., as in MID). We can maximize the evidence more rapidly by using its first and second derivatives, which we can compute analytically (see Methods). We also exploit a heuristic strategy for initializing the ALDsf hyperparameters using the estimates from ridge regression (to identify the scale), ALDs (to identify a spatiotemporal region) and ALDf (to identify a Fourier region). Although it is substantially more computationally expensive, the fully Bayesian estimate based on MCMC avoids the issue of local maxima because it explores the entire evidence surface, not just its modes.
However, we do not ultimately view ALD and other model-based or information-based methods as in conflict. Rather, we regard ALD as providing a prior distribution over RFs that can be combined with any likelihood. Computing and optimizing the evidence under nonlinear models with non-Gaussian noise represents an important direction for future work. We suggest that locality is a general feature of neural information processing and anticipate that it will be useful for neural characterization in a wide variety of brain areas, including those where response properties are not yet well understood [73]. We expect hierarchical models and empirical and fully Bayesian inference methods to find application to a wide range of problems where structured prior information can be usefully defined.
For the simulated data shown in Fig. 5 , we used a 2-dimensional Gabor filter (shown in Fig. 4 A) and two types of stimuli: Gaussian white noise and “naturalistic spectrum” noise–Gaussian noise with a power spectrum. Simulations were carried out with various numbers of stimulus samples , noise variance , signal variance of 1, and a pixel filter (coordinate dimension , filter dimension ). To quantify performance, we defined the filter error as , where is the true filter and is an estimate. To obtain reliable estimates of mean error, we ran 100 simulations at each sample size. To calculate the relative error (Fig. 5 B and D), we computed the error for each method, and then computed the geometric mean of the error ratio across datasets.
For V1 data shown in Fig. 6 , the data and experimental methods are described in [53]. Briefly, cells were stimulated with 1D spatiotemporal binary white noise stimuli (“flickering bars”) aligned with each neuron's preferred orientation. Stimuli were presented at a frame rate of 100 Hz. The number of bars varied for different neurons, . The linear receptive field was assumed to extend over a time window of frames before a spike (a 160 ms time interval). The full dimensionality of the filter was thus , ranging from 192 to 384 parameters.
For retinal ganglion cell data shown in Fig. 8 , the data and experimental methods are described in [54], [55]. Briefly, cells were stimulated with the spatiotemporal binary white noise stimuli presented at a frame rate of 120 Hz, contained in 10×10 pixels in space. We assumed the size of the linear receptive field to be pixel 25 time bin, making for total coefficients in the RF.
We used cross-validation to quantify the performance of the various estimators (Fig. 6), and resampled the training data to examine performance as a function of training sample size. To quantify error reliably, we performed 100 repetitions for each sample size, drawing the training data randomly without replacement in blocks of size 2s, which helped to minimize the effects of non-stationarities in the data. To quantify cross-validation performance, we used relative cross-validation , defined as , where is the number of samples of test data, is a spike count in the 'th time bin in the test set, is the th row of the design matrix , is the RF estimate obtained by each method (from training data), and is the ML estimate obtained on the test data. Essentially, this is the ordinary test error minus the error of the ML estimator trained on test data (which provides an absolute lower bound on the performance of any linear model). We computed the relative cross-validation errors from five methods (ML, Ridge, ARD, ASD, and ALDsf) using 8 minutes of test data. In Fig. 6, we normalized the errors by dividing them by maximum average error across methods (the ML estimate using 30 seconds of data yielded the maximum cross-validation error). We computed the standard deviation of the normalized cross-validation error across 100 different training sets for each dataset size.
To perform fully Bayesian inference, we used Metropolis-Hastings (MH) sampling to sample from the distribution over hyperparameters given the data . We used an isotropic Gaussian proposal distribution with variance given by the largest eigenvalue of inverse Hessian of the log-evidence around . (More advanced proposal distributions and sampling methods are found in [76], [77], but this simple proposal sufficed for our purposes and mixed reasonably quickly). Thus, we first optimized the evidence to obtain the mode of , which is the mode of . We assumed a non-informative hyperprior , taken to be uniform over the range of values permitted during constrained optimization of the log-evidence (see above).
To carry out MH sampling, we sampled from the Gaussian proposal distribution centered on the current state of the Markov chain, , then computed , with the . We accepted the proposal randomly with probability , setting , and otherwise rejected it, setting . Given each sample , we drew a sample of the receptive field . These samples were averaged to compute the posterior mean (or Bayes Least Squares estimator). Their quantiles were used to compute credible intervals for each filter coefficient.
In Fig. 10 , we compared fully Bayesian (FB) and empirical Bayes (EB) filter estimates obtained from V1 simple cell data [53]. For each set of training data, we drew 5000 samples using MH to compute the posterior mean and credible intervals. The average acceptance rate of the MH sampler was 0.12. For Fig. 10 A, we computed the average of the EB and FB error from 100 repetitions with independently drawn sets of training data. We computed the average cross-validation error of both estimates of the example cell (in red). For Fig. 10 B, we computed the average posterior variance by averaging the posterior variances in the estimates from the 100 iterations in each cell, which we then averaged across all 16 cells. For Fig. 10 C, we computed the average cross-validation error by averaging the errors from the 100 iterations in each cell, and we averaged these across 16 cells. The same 8 minutes of held out test data was used for cross-validation, for all training iterations.
|
10.1371/journal.pcbi.0030052 | Positive and Negative Design in Stability and Thermal Adaptation of Natural Proteins | The aim of this work is to elucidate how physical principles of protein design are reflected in natural sequences that evolved in response to the thermal conditions of the environment. Using an exactly solvable lattice model, we design sequences with selected thermal properties. Compositional analysis of designed model sequences and natural proteomes reveals a specific trend in amino acid compositions in response to the requirement of stability at elevated environmental temperature: the increase of fractions of hydrophobic and charged amino acid residues at the expense of polar ones. We show that this “from both ends of the hydrophobicity scale” trend is due to positive (to stabilize the native state) and negative (to destabilize misfolded states) components of protein design. Negative design strengthens specific repulsive non-native interactions that appear in misfolded structures. A pressure to preserve specific repulsive interactions in non-native conformations may result in correlated mutations between amino acids that are far apart in the native state but may be in contact in misfolded conformations. Such correlated mutations are indeed found in TIM barrel and other proteins.
| What mechanisms does Nature use in her quest for thermophilic proteins? It is known that stability of a protein is mainly determined by the energy gap, or the difference in energy, between native state and a set of incorrectly folded (misfolded) conformations. Here we show that Nature makes thermophilic proteins by widening this gap from both ends. The energy of the native state of a protein is decreased by selecting strongly attractive amino acids at positions that are in contact in the native state (positive design). Simultaneously, energies of the misfolded conformations are increased by selection of strongly repulsive amino acids at positions that are distant in native structure; however, these amino acids will interact repulsively in the misfolded conformations (negative design). These fundamental principles of protein design are manifested in the “from both ends of the hydrophobicity scale” trend observed in thermophilic adaptation, whereby proteomes of thermophilic proteins are enriched in extreme amino acids—hydrophobic and charged—at the expense of polar ones. Hydrophobic amino acids contribute mostly to the positive design, while charged amino acids that repel each other in non-native conformations of proteins contribute to negative design. Our results provide guidance in rational design of proteins with selected thermal properties.
| Despite recent advances in computational protein design [1], there is no complete understanding of basic principles that govern design and selection of naturally occurring proteins [2]. In particular, the physical basis for the ability of proteins to achieve an adaptation to a wide variety of external conditions is still poorly understood. While several attempts to design proteins with a desired fold were successful [1,3], rational design of proteins with desired thermal properties is still an elusive goal. However, Nature apparently succeeds in doing so by “designing” proteins in hyperthermophiles that are stable and functional up to 110 °C. Thus, in the absence of the complete solution of the protein design problem, it is tempting to get clues from Nature as to how thermal properties of proteins can be modulated by proper sequence selection and which physical factors play a role in this process.
A clear manifestation of thermophilic adaptation can be found in a highly statistically significant variation of amino acid compositions of proteomes between meso- and thermophilic organisms [4–7]. Recently, we showed that the total concentration of seven amino acids, I, V, Y, W, R, E, and L, is highly correlated with optimal growth temperature (OGT) of an organism (R = 0.93) [8]. The total concentration of IVYWREL combination of amino acids serves as a predictor of OGT with mean accuracy of 8.9 °C [8]. In this work we seek a fundamental theoretical explanation as to why Nature requires an elevated concentration of both hydrophobic and charged amino acids to design hyperthermostable proteins.
Our first goal here is to develop a minimalistic physical model of protein design that could help us to rationalize comparative proteomic analysis of thermo- and mesophiles. A crucial question is how to incorporate the environmental temperature in the model of protein design. Two factors may play a role. The first effect is due to fundamental statistical mechanics of proteins that posit that stable and foldable proteins should have an “energy gap” [9–12]. Specifically, the stability of the native state of a protein is determined by the Boltzmann factor exp(−ΔE/kBT), where ΔE is the energy gap between the native state and lowest energy completely misfolded structures [11–15]. Therefore, to maintain their stability at elevated temperature, the thermophilic proteins should have a greater energy gap. In principle, the increase of the energy gap can be achieved by lowering the energy of the native state (positive design), raising the energy of misfolds (negative design), or both. Another factor that may affect protein thermostability is a possible dependence of fundamental interactions (e.g., hydrophobic forces) on temperature. However, the temperature dependence of different types of interactions may be very complex, and it remains a subject of controversy as to how and to what extent it influences the stability of proteins [16–19]. Our approach to this complex issue is simple: consider first how far one can go based on purely statistical–mechanical analysis of protein thermostability without resorting to explanations based on temperature dependence of various interactions. Specifically, here we use the 27-mer cubic lattice model of proteins [20,21]. The model features 20 types of amino acids that interact when they are nearest neighbors on the lattice; interaction energy depends on types of amino acids involved. The potential is derived from known protein structures and is temperature-independent [22]. For this lattice model all compact conformations can be enumerated [20] and, therefore, exact statistical–mechanical analysis is possible. Previously, protein thermodynamics [9,23], folding [24,25], and evolution [11,26,27] were extensively studied by using this model. We simulate the process of thermal adaptation by the design of 27-mer sequences with selected (at a given environmental temperature Tenv) thermal properties [14,15]. The algorithm of design (see Methods) carries out simultaneous unrestricted search in conformational, sequence, and amino acid composition spaces. In our analysis we will focus on the amino acid composition of designed sequences as a function of the environmental temperature and we will compare the model findings with amino acid trends in real proteomes. Our main result is that thermal adaptation utilizes both positive and negative design. We show that by increasing the content of amino acids from both extremes of the hydrophobicity scale, thermostable proteins achieve exactly that goal: hydrophobic residues help with positive design while elevated concentration of charged residues helps to achieve stronger negative design. Further, we find an interesting and potentially important aspect of negative design: similar to positive design that strengthens certain native interactions, negative design can make specific non-native interactions strongly repulsive. This, in turn, may lead to emergence of correlated mutations between amino acids that are not in contact in native structure.
We design lattice model proteins with selected thermostability as a first step toward modeling thermal adaptation of organisms. There is a direct connection between OGT (environmental temperature, or Tenv) of an organism and the melting temperature of its proteins [28,29]. We used the P–design procedure to create model 27-mer sequences that are stable at selected Tenv (see Methods). We designed sets of 5,000 model proteins for each Tenv in the range 0.3 < Tenv < 0.8 in Miyazawa–Jernigan dimensionless units. The average melting temperature <Tmelt> of lattice proteins is strongly correlated with Tenv (see Figure S1) suggesting that the P-design procedure does work. It provides model proteins with desired stability in response to the increase of environmental temperature. The dependence of <Tmelt> on Tenv is close to linear and qualitatively matches the empirical linear relationship, Tmelt = 24.4 + 0.93 Tenv, between the average living temperature of the organism and melting temperature of its proteins [28].
As expected, the amino acid composition of designed proteins does depend on Tenv for which they were designed. To quantify the differences between “low-temperature” and “high-temperature” amino acid compositions, we plotted temperature dependencies of the fractions of hydrophobic (LVWIFMPC), weak hydrophophobic and polar (AGNQSTHY), and charged (DEKR) amino acids for designed lattice proteins (Figure 1A) and natural (Figure 1B) proteomes. Figure 1A shows a significant increase in the amount of charged residues (red triangles) and a slight increase in hydrophobic amino acids (green squares) at the expense of polar ones (black dots). Remarkably, the results shown in Figure 1 suggest that increase of thermostability is accompanied by growth of amino acid content from both extremes of the hydrophobicity scale, adding both charged and hydrophobic residues. This observation is further highlighted in Figure 2, which shows—amino acid by amino acid—how compositions of model proteins with Tenv in designed model proteomes for all 20 amino acids are ranked by their hydrophobicity according to the Miyazawa–Jernigan set of interaction parameters (see Methods and Figure S2A and S2B for more detailed explanation). Figure 2 clearly shows that addition of amino acids to thermophilic model proteomes occurs from the extremes of the hydrophobicity scale while the middle is depressed. The content of charged (Asp, Glu, Lys, Arg; DEKR) and four of the hydrophobic (Ile, Leu, Phe, Cys; ILFC) residues is increased with temperature at the expense of other residues, mostly polar ones. This observation shows that combining amino acids with maximum variance in their hydrophobicity is crucial for creating hyperthermostable model proteins. We refer to this effect as the “from both ends of hydrophobicity scale” trend.
For comparison, we analyzed the variation of amino acid composition in fully sequenced bacterial proteomes (83 species in total, see complete list, Table S1) of psycho-, meso-, thermo-, and hyperthermophilic prokaryotes (habitat temperatures from −10 to +110 °C, see Table S1). Importantly, amino acid composition of 83 natural prokaryotic proteomes reveals similar trends, an increase of the contents of hydrophobic and charged residues, and a decrease of the content of polar ones (Figure 1B). For a more direct comparison of the predictions of our model with the properties of natural proteomes, in Figure 3 we plotted the temperature derivative of the fraction of each of the amino acids in designed lattice proteins against the corresponding temperature derivative calculated over the 83 natural proteomes. The observed positive significant correlation (R = 0.56, p = 0.01) suggests that generic physical factors captured by this simple statistical–mechanical model played a major role in shaping the amino acid composition patterns across a wide range of habitat temperatures.
We hypothesize that the generic character of the “from both ends” trend that is universally observed in the model and in natural proteins is related to the positive and negative elements of design. In this case, one (hydrophobic) end of the scale is responsible for positive design while another (hydrophilic) end provides negative design. To test this hypothesis, we first studied how the energy gap between the energy of the native state and that of misfolded conformations for the designed model proteins depends on Tenv (Figure 4). Positive design is the major contributor to the effect (the slope of the temperature-dependent energy decrease of the native state with growth of Tenv is −5.22; Figure 4, black line), while the increase of the average energy of decoys with Tenv (slope +1.64; Figure 4, orange line) is pronounced, but less significant. Nevertheless, the results presented in Figure 4 provide clear evidence that negative design works, along with positive design, in the selection of thermostable model proteins.
The findings shown in Figure 4 demonstrate that indeed both positive and negative design act in enhancing thermostability of model proteins. However, the question remains as to how positive and negative design is related to the “from both ends” trend in amino acid compositions, as shown in Figure 2. To address this question, we plot the number of contacts between amino acids whose content grows with Tenv, according to Figure 2. Figure 5 shows how the average number of contacts (per structure) within both groups of amino acids, FILC and DEKR, in native conformations and in misfolded decoys, depends on Tenv. Remarkably, we see that in decoy structures the growth of the number of contacts occurs only within the “charged” group DEKR, some of which—according to the Miyazawa–Jernigan potential—repel one another. On the other hand, the number of contacts in the hydrophobic group in decoys do not change despite an overall increase of concentration of these amino acids in sequences designed at higher Tenv. This result shows that while strongly mutually attractive hydrophobic groups provide lower energies of native states for hyperthermophilic model proteins, the growth in concentration of “charged” (DEKR) groups mainly contributes to the negative design factor by raising average energy of misfolded conformations. Remarkably, the average number of contacts between hydrophobic groups (FILC) in misfolded conformations remain roughly the same in meso- and hyperthermophilic model proteins despite significant growth in overall concentration of these groups in hyperthermophiles. Therefore, the data shown in Figure 5 indicate that the “from both ends” trend in amino acid composition is directly related to positive and negative design in stabilization of hyperthermophilic model proteins.
The data presented so far provide insight into averaged (over many model proteins) contributions to the energies of native conformation and decoys. However, a question arises whether negative design works by increasing “average” non-native interactions or by strengthening certain specific repulsive non-native interactions. Indeed, negative design may be based on introducing a few energetically disadvantageous non-native contacts that are persistent in many decoy structures, increasing their energy [2,30]. Therefore, non-native contacts responsible for negative design may well be specific for each sequence, making this effect more detectable if individual proteins are considered.
The exact nature of the lattice model makes a detailed residue-by-residue analysis of the action of both positive and negative design possible. To this end, it is instructive to identify interactions, native and non-native contacts, between residues that play especially important roles in stabilization of the native state and destabilization of decoys. The key idea here is that such important interactions should be conserved in all sequences that fold into a given structure. While identities of amino acids that form such a contact may vary from sequence to sequence, the strength (or energy) of key native or non-native contacts will be preserved: it will be either strongly repulsive or strongly attractive for all sequences that fold into a given structure [31]. Therefore, to identify such key contacts, distributions of energies of native and non-native contacts in multiple sequences that fold into the same native structure should be considered. Such analysis can reveal not only conserved strong native contacts but also possible conserved strong repulsive non-native contacts. To investigate such possibility, we designed 5,000 lattice proteins that all fold into the same (randomly chosen) native structure. To achieve that, we used the design algorithm similar to P-design (see Methods), but for a fixed native structure, and checked a posteriori that the target structure is indeed the native state for all 5,000 sequences. We designed a set of 5,000 mesophilic sequences at Tenv = 0.2 and 5,000 hyperthermophilic sequences that fold to the same structure but are much more stable (Tenv = 0.8).
The concept of native and non-native contacts for our lattice model is illustrated in Figure 6A. It is a cartoon with a zoom-in into the contact matrix of the lattice structure used in simulations. The contact matrix of any compact lattice conformation contains all native (green, total 28 in any structure) and all possible non-native (blue, total 128) contacts. All other contacts (red) are prohibited according to the properties of the cubic lattice. To identify important native and non-native contacts whose energies are conserved, we applied the following procedure. First, for each of the 5,000 sequences that fold into selected structure, we calculated energies of 28 native and 128 possible non-native contacts in this structure (using the identities of residues and Myazawa–Jernigan potentials that were employed to design sequences). Next, for each contact we calculated average energy and its standard deviation over all 5,000 designed sequences (see Figure 6B for illustration of this calculation). Contacts whose energy shows a very low standard deviation over all designed sequences are apparently the ones that are most important for stability. This procedure was carried out both for mesophilic sequences (Tenv = 0.2) and for thermophilic sequences (Tenv = 0.8). The results are shown in Figure 7, which presents standard deviation of interaction energies of each native and non-native contact over all 5,000 designed mesophilic sequences (A) and hyperthermophilic sequences (B), plotted against the average (over 5,000 designed sequences) energy of that contact. The plot consists of 28 + 128 = 156 points, covering all native and all possible non-native interactions. The native state clearly defines conserved low- and high-energy native contacts (shown in black) in most of the sequences, as the standard deviation is the lowest at the extreme values of the energy. Conserved attractive interactions are in the protein interior, corresponding to the lattice analog of the hydrophobic core; apparently, they emerge due to the action of positive design. The non-native contacts (red dots) follow a different pattern, with only a few conserved attractive interactions, suggesting the diversity of decoy structures. What is surprising to see, however, is that energies of certain most-repulsive (high-energy) non-native contacts show a very low standard deviation, indicating that such contacts may be as important for protein stability as conserved native ones. Comparison of meso- and hyperthermophilic sequences shows clearly that emergence of strong and conserved attractive and repulsive interactions in key native and non-native contacts is directly related to sequence design that generates stable sequences: design of hyperthermostable sequences (Figure 7B) results in stronger and more conserved (lower dispersion of energy) attractive and repulsive specific native and non-native interactions. The only reason that repulsive energies of non-native contacts are conserved is that such contacts persist in certain frequent decoy structures and contribute to the widening of the gap between the native state and decoys. Such repulsive contacts are indirectly (via the sequence) related to a particular native state and are not numerous. Their role may be completely obscured in a “high-throughput” analysis where sequences with different native states are considered together, as in Figure 4. Therefore, we conclude that negative design involves a very specific strategic placement of repulsive contacts in certain decoy structures.
The results and analysis presented in Figure 7 have very important implications for real proteins. The requirement to conserve the energy of key contacts in multiple sequences that fold into the same structure implies that amino acids forming such contacts can mutate in a correlated way, for example by swaps. The observation that mutations may occur often as swaps to preserve specific attractive native and specific repulsive non-native interactions leads to a prediction of a peculiar dependence between frequency of amino acid substitutions (as in, e.g., BLOSUM matrices [32]) and interaction energy between amino acids. Indeed, as illustrated in Figure 8, a correlated mutation in the form of a swap can manifest itself in sequence alignment as a substitution between amino acids that are making the swap. The implication is that frequent substitutions will be observed between amino acids that strongly attract each other (to preserve specific stabilizing native contacts). More interesting and perhaps more surprising, frequent substitutions are also predicted between amino acids that strongly repel each other (to preserve specific non-native repulsive contacts). In other words, we predict that the scatter plot between elements of amino acid interaction energy matrix and substitution matrix will be non-monotonic with maxima at both extremes.
We tested this prediction by plotting the dependence of elements of the substitution matrix BLOSUM62 [32] for 190 pairs of amino acids (synonymous substitutions are excluded) versus their interaction energy as approximated by the knowledge-based Miyazawa–Jernigan potential [22] (Figure 9). This analysis indeed reveals a non-monotonic shape: the parabolic fit in Figure 9A highlights the highly significant non-monotonic nature of the dependence. The striking feature of this dependence is that most frequent substitutions are observed not only between the most attractive amino acids but also between the most repulsive ones. One could argue, however, that the high frequency of substitutions between amino acids that repel each other may be a trivial consequence of conserved substitutions that preserve the charge (R to K and E to D). However, a detailed inspection of the upper right part of the plot in Figure 9A shows that this is not the case (Figure 9B). Indeed, frequent substitutions are observed between mutually repulsive amino acids with vastly different physical–chemical properties and encoded by very dissimilar codons, such as Serine to Asparagine, Glutamin to Arginine, etc. Several highly nonconservative substitutions show about “random”frequencies (element of BLOSUM matrix close to zero, e.g., for Asn to Lys), but this may be due to compensation of two opposite effects: suppression of highly nonconservative substitutions (e.g., that change charge) and facilitation of correlated substitutions such as the ones in the form of swaps as illustrated here.
Use of correlated mutations as predictors of spatial proximity of amino acids in the native structure has been proposed by many authors [33–36]. Indeed, statistical analysis shows that overall correlation between distance between amino acids and degree of correlation in multiple sequence alignments does exist [37]. However sometimes correlated mutations are observed between amino acids that are distant in native structure [33,38]. While sometimes such observations are discarded as false positives in the prediction algorithm [33], our analysis predicts that indeed residues that are distant in structure but may form important repulsive contacts in misfolded conformations may exhibit correlated mutations as illustrated in Figures 8 and 9.
As an illustration of the significance of correlated mutations between amino acids that are far apart in structure, we consider a TIM–barrel fold protein triosephosphate isomerase. Guided by the results of statistical analysis shown in Figure 9, we looked for pairs of residues with strong repulsion according to the Miyazawa–Jernigan potential, random or higher substitution rates between these residues according to the BLOSUM matrix, and highly correlated substitutions of these residues in two positions of the protein sequence in multiple sequence alignment for 7tim (see Methods). To distinguish the effect that we seek from functional conservation, these residues should not be in contact in the native state and should not be involved in the functional site or in the protein–protein interactions
We found correlated substitutions in the sequence of TIM–barrel fold (7tim, chain a), according to the physicochemical characteristic hydropathy [39], by using the CRASP program, which “estimates the contribution of the coordinated substitutions to invariance or variability of integral protein physicochemical characteristics” [40]. Four pairs of residues (Table 1) that have highly coordinated substitutions and repel each other according to the Miyazawa–Jernigan energy matrix were selected. None of those residues belong to the functional site of triosephosphate isomerase, and the protein itself is a single-domain protein not involved into protein–protein interactions [41]. Four pairs of polar and charged residues and one pair of charged residues were identified (see Figure 10A and Table 1). The shortest contact distance (Cα − Cα, 8.7 Å) is between charged Lys(84) and polar Gln(119), which excludes the stabilizing interaction between them in the native structure. We found that correlated mutations between some of these amino acids occur as swaps in the TIM–barrel fold, possibly accompanied by conservative mutation, e.g., surface Lys 120 and Gln 85 in 7tim swap to Gln 120 Lys 85 in Thermotoga maritima thermophilic ortholog of triosephosphate isomerase. Even more striking, the Gln85, Lys 213 pair in 7tim (distance in native structure 30 Å) is replaced by Lys 85, Asn 213 in 1b9b (Figure 10B). This pair of residues shows a highly correlated substitution pattern in TIM–barrel multiple sequence alignment despite the fact that these are very distinct amino acids.
Stabilization of thermophilic proteins is achieved by negative and positive design working together, i.e., the gap “opens” from both sides, decreasing energy of the native state and at the same time increasing the energy of misfolded conformations. This factor is responsible for the “from both ends of hydrophobicity scale” trend observed in model and real [8] thermophilic proteomes. In particular, our recent analysis of complete bacterial proteomes [8] revealed that proteomes of thermophilic bacteria are enriched in both hydrophobic residues (IVYLW) and charged ones (ER), while all polar residues are suppressed. Discrepancies between different hydrophobicity scales [42], the statistical nature of knowledge-based Miyazawa–Jernigan potential [22], and limitations of the lattice model make it impossible to quantitatively compare the content of individual amino acids in lattice and natural proteomes or exactly predict the amino acid composition of thermophilic proteomes with very high accuracy from lattice model calculations. Nevertheless our lattice calculations are in semiquantitative agreement with data on natural proteomes, (see Figures 1 and 3) and exhibit the same “from both ends of hydrophobicity scale” trend in amino acid composition adaptation in response to elevated habitat temperature.
The knowledge-based Miyazawa–Jernigan potentials, derived from native structures of proteins, are certainly a crude approximation to real protein energetics [43]. A question arises as to whether our observations are generic or are due to the specific potential used to design model proteins. A detailed comparison of several potentials—all atom and group-based derived by different methods—was carried out recently in our lab [44]. Remarkably, we found that despite differences in detail all these potentials reflect the same dominant contributions to protein stabilization. It appears that dominant contributions to energy gaps in proteins come principally from two types of interactions: hydrophobic interactions and electrostatics [44]. Further, it was found that knowledge-based potentials derived using structures of meso- and thermophilic proteins are virtually indistinguishable (KZ and ES, unpublished data). These observations suggest that the “from both ends of hydrophobicity scale” trend observed in model calculations and in real proteome is a robust phenomenon, reflecting basic physical principles of protein design, rather than a consequence of a specific potential set used in calculations.
While positive design [45] is universally used in experiments, the role and omnipresence of negative design are still under discussion [46]. The main challenge in the study of negative design stems from the difficulties in the modeling of relevant misfolded conformations and energetic effects of mutations that destabilize them [46]. It was shown that charged residues can be effectively used in negative design [30]. Another indirect evidence of the contribution of charged residues to negative design emerges from site-directed mutagenesis, where mutations of polar groups to charged ones on the surface of a protein lead to protein stabilization even in the absence of salt–bridge partners of the mutated group [47–49]. In a series of experiments [47,48,50], surface electrostatic interactions were shown to provide a marginal contribution to stability of the native structure, hence their possible importance for making unfavorable high-energy contact in decoys. An alternative view, proposed recently by Makhatadze et al., suggests that long-range electrostatic interactions may contribute to stability of the native state [51]. However, at normal physiological conditions the range of electrostatic interactions is limited due to Debye screening and hardly exceeds 8 Å. Our simulations and proteomic analysis point to a possible role of some surface charged residues as contributing to destabilization of misfolded structures through a negative design mechanism.
Positive and negative elements of design affect the evolution of protein sequences. The dependence of substitution rates in sequences of natural proteins (BLOSUM62 substitution matrix) on interaction energies according to knowledge-based Miyazawa–Jernigan potential has a peculiar nonmonotonic shape showing elevated substitution rates between residues that attract each other as well as between residues that repel each other. The physical reason for this phenomenon is the same as for the “both ends of hydrophobicity scale” trend: simultaneous action of positive and negative design. Upon substitutions, energy of attractive contacts in native states should be preserved as well as energies of specific repulsive contacts in misfolded conformations. Apparently both these factors act in concert to preserve the energy gap in proteins.
Our study deepens an understanding of correlated mutations in proteins. With regard to native contacts, the fact that amino acids making strongly attractive native interactions should exhibit correlated mutations had been realized long ago. Several authors proposed to use correlated mutations as a tool to determine possible native contacts from multiple sequence alignment [33–36]. However, this suggestion is complicated by the observation that correlated mutations are often found between residues that have no obvious functional role and are distant in structure [33,38,52,53]. Using the double mutant technique, Horovitz et al. [33] suggested a relation between correlated mutations and energetic connectivity (i.e., nonadditivity of stability effects in double mutation cycles) between corresponding amino acids. Green and Shortle [54] showed that amino acids that are distant in structure may indeed be “energetically coupled,” attributing this effect to influence of mutations on the unfolded state of proteins, consistent with our findings. Lockless and Ranganathan [38] suggested that a “pathway of energetic connectivity” exists between distant residues that exhibit correlated mutations. Fodor and Aldrich [37], however, examined several other proteins and argued against the “general principle of isolated pathways of evolutionarily conserved energetic connectivity in proteins.” Here we show that negative design that destabilizes misfolded conformations of proteins may be responsible for correlated mutations between residues that are far apart in native structures.
In this work, we developed a simple exact model of thermophilic adaptation and discovered fundamental statistical–mechanical rules that Nature uses in her quest to enhance protein stability. While many other factors, including dependence of hydrophobic and other interactions on temperature, certainly play a role in protein stabilization, the action of positive and negative design found and described here in a minimalistic model appears to be a basic universal principle determining evolution of sequences of thermostable proteins. A better understanding of fundamental principles of protein design and stability makes it possible to decipher peculiar signals that emerge in the analysis of meso- and thermophilic genomes and proteomes [8] and in many studies of correlated mutations in proteins [33,35,53].
We use the standard lattice model of proteins as compact 27-unit polymers on a 3 × 3 × 3 lattice [20]. The residues interact with each other via the Miyazawa–Jernigan pairwise contact potential [22]. It is possible to calculate the energy of a sequence in each of the 103,346 compact conformations allowed by the 3 × 3 × 3 lattice, and the Boltzmann probability of being in the lowest energy (native) conformation,
where E0 is the lowest energy among the 103,346 conformations, and Tenv is the environmental temperature. The melting temperature Tmelt is found numerically from the condition Pnat(Tmelt) = 0.5. Note that if the energy spectrum Ei is sparse enough at low energies, the value of Pnat is determined chiefly by the energy gap E1 – E0 between the native state and the closest decoy structure that has no structural relation to the native state.
To design lattice proteins, we use here a Monte-Carlo procedure (P-design, [14,15]) that maximizes the Boltzmann probability Pnat of the native state by introducing mutations in the amino acid sequence and accepting or rejecting them according to the Metropolis criterion. As this procedure takes the environmental temperature Tenv as an input physical parameter, and generates amino acid sequences designed to be stable at Tenv, it is an obvious choice for modeling the thermophilic adaptation.
Initially, the sequence is chosen at random; the frequencies of all amino acid residues in the initial sequences are equal to 5%. At each Monte-Carlo step, a random mutation of one amino acid in a sequence is attempted and Pnat of the mutated protein is determined. The native structure is determined at every step of the simulation; generally, the native state changes upon mutation of the sequence. If the value of Pnat increased, the mutation is always accepted; if Pnat decreased, the mutation is accepted with the probability exp[−(Pnat(old) − Pnat(new))/p], with p = 0.05 (a Metropolis-like criterion). We chose p = 0.05 so that the average melting temperature of designed proteins is higher than the environmental temperature (see Figure S2), in agreement with experimental observations [29,55]. The design procedure is stopped after 2,000 Monte-Carlo iterations. Such length of design runs is sufficient to overcome any possible effects of the initial composition of the sequences, so the amino acid composition of the designed sequences depends only on the environmental temperature Tenv.
To relate the trends in amino acid composition with the physical properties and interaction energies of individual amino acids, we use hydrophobicity as a generic parameter characterizing an amino acid [42]. To characterize the hydrophobicity of amino acids in the simulations, we make use of the fact that the Miyazawa–Jernigan interaction energy matrix is very well approximated by its spectral decomposition [43]. Interestingly, it is sufficient to use only one eigenvector q, corresponding to the largest eigenvalue, so the interaction (contact) energy Eij between amino acids of types i and j reads Eij ≈ E0 + λqiqj [43]. In this representation, hydrophobic residues have the largest values of q, while hydrophilic (charged) residues correspond to small q.
All sequences of TIM–barrel folds with length less than 300 amino acid residues were extracted according to the SCOP database description [56]. Identical sequences were excluded from further consideration. Remaining sequences (total 39) were aligned against the sequence of the triosephosphate isomerase (7tim.pdb, chain a) by using Kalign Web-server for multiple alignment of protein sequences (http://msa.cgb.ki.se/cgi-bin/msa.cgi, [57]).
Correlated substitutions in the multiple alignments were determined by using the CRASP program (http://wwwmgs.bionet.nsc.ru/mgs/programs/crasp, [40]). The CRASP program gives the correlation coefficient between the values of physicochemical parameters at a pair of positions of sequence alignment. We chose hydropathy [39] as a physicochemical characteristic appropriate for establishing correlated mutations of interest. Only significant correlations, with the correlation coefficient higher than the critical threshold (0.311), were considered.
The complete genomes were downloaded from the National Center for Biotechnology Information Genome database at http://www.ncbi.nih.gov/entrez/query.fcgi?db=Genome (see Table S1).
The accession numbers from the Protein Data Bank (http://www.rcsb.org/pdb/) used in this paper are: TIM–barrel fold protein triosephosphate isomerase (7tim); T. maritima thermophilic ortholog of triosephosphate isomerase (1b9b). |
10.1371/journal.pntd.0001943 | Calodium hepaticum: Household Clustering Transmission and the Finding of a Source of Human Spurious Infection in a Community of the Amazon Region | Background: Calodium hepaticum (syn. Capillaria hepatica) is a worldwide helminth parasite of which several aspects of transmission still remain unclear. In the Amazon region, the mechanism of transmission based on the ingestion of eggs present in the liver of wild mammals has been suggested as the cause of the spurious infections described. We performed an epidemiological investigation to determine the incidence, risk of spurious infection and the dynamics of transmission of C. hepaticum in a community of the Brazilian Amazon. Methodology/Principal Findings: Stool samples of 135 individuals, two dog feces and liver tissue from a peccary (captured and eaten by the residents) were analyzed by conventional microscopy. Dog feces were collected from the gardens of households presenting human cases of spurious C. hepaticum infections. Community practices and feeding habits related to the transmission of the parasite were investigated. The individual incidence of spurious infection was 6.7% (95% CI: 2.08–11.24). Cases of spurious infection were observed in 7.5% of the families and the household incidence was from 50% to 83.3%. The risk of spurious infection was 10-fold greater in persons consuming the liver of wild mammals (p = 0.02). The liver tissue of a peccary and one feces sample of a dog presented eggs of C. hepaticum. The consumption of the infected liver was the cause of the spurious infections reported in one household. Conclusions/Significance: This is the first identification of a source of spurious infection by C. hepaticum in humans and we describe a high rate of incidence in household clusters related to game liver alimentary habits. The finding of a dog feces contaminating peridomiciliary ground suggests the risk of new infections. We conclude that the mechanism of transmission based on the ingestion of liver is important for the dynamics of transmission of C. hepaticum in the studied area.
| The zoonotic parasite Calodium hepaticum is the causative agent of rarely reported liver disease (hepatic calodiasis) and spurious infections in humans. In spurious infections eggs of this parasite are excreted in the stools without causing disease. It has been suggested that the cause of this type of infection in Amazonian areas is the ingestion of liver of wild mammals infected with the eggs of the parasite. Nonetheless, studies are needed to confirm this mechanism of transmission and investigate its epidemiological importance. In the present study we report the high individual (6.7%) and household incidence (50%–83.3%) of spurious infection in a rural community of the Brazilian Amazon. We found a high risk of spurious infection among subjects who usually ate the liver of wild mammals and detected a source of spurious infection in humans (peccary liver) as well as, for the first time, ground contamination with infected dog feces in a household presenting human cases. We confirm the existence of this mechanism of transmission of C. hepaticum and suggest that it is important for transmission not only in this area but probably also in other areas of the Amazon with similar sociocultural characteristics.
| Calodium hepaticum (syn. Capillaria hepatica) is a zoonotic nematode of the Trichinellidae family found worldwide. This helminth infects the hepatic parenchyma of rodents (principle hosts) and various other mammals (e.g. carnivores, humans) of different families [1]. In humans infection may cause hepatic calodiasis (syn. hepatic capillariasis), a rare liver disease (72 cases reported around the world, 5 being found in Brazil) which may have a severe clinical course [1]–[5].
Infection by C. hepaticum occurs following the ingestion of embryonated eggs (true or hepatic infection) which pass through the intestinal tract. Larvae hatch at the level of the cecum, pass through the intestinal wall and reach the liver via the portal-hepatic system. The larvae mature in the hepatic parenchyma, transforming into adults 28 days after the infection. Females lay the eggs in the parenchyma and these develop only to the eight-cell stage. Eggs reach the environment through the decay of the host carcass or when a predator or cannibal ingests the host and releases the eggs through the stools. Over a 5–8 week period in optimal conditions of temperature, humidity and air exposure, the eggs embryonate in the ground and may infect a new host. Ingestion of non embryonated eggs leads to untrue (or spurious) infection in which the eggs pass through the intestinal tract and exit with the stools without causing liver disease [6]–[8].
The dynamics of the transmission of C. hepaticum and the risk factors associated with infection remain unclear [9], [10]. In urban areas transmission is related to the presence of small rodents (e.g. Rattus novergicus and Mus musculus) and poor hygienic and sanitary conditions [1], [6], [11]. In small rodents, characteristics such as the high prevalence of natural infection [7], [11], [12], the rapid populational turnover and the habit of cannibalism may explain the elevated transmission of the parasite among these rodents and their involvement in environmental contamination by eggs [13], [14]. The ingestion of eggs present in the ground or in contaminated foods has been accredited as the mode of transmission to humans in urban areas. It has been suggested that domestic animals (cats and dogs) may also contaminate the peridomiciliary ground with infected stools [1], [14] after eating small rodents, carcasses or infected liver of other mammals [15]. The participation of domestic animals in the domiciliary cycles has not, as yet, been well defined.
Spurious infection has predominantly been described in tribal or immigrant communities around the world [5]. Several authors have suggested that the cause of this infection in determined populations is the mechanism of transmission based on the ingestion of non embryonated eggs present in the liver of mammals [15]–[20]. Foster & Johnson related the occurrence of spurious infection in natives of Panama to the encounter of three new hosts (Tayassu pecari, Ateles geoffroyi and Cebus capucinus) commonly used by the natives as food [16]. In a rural community in the Brazilian Amazon a case of spurious infection was associated with the reported consumption of liver of tapir [18]. Recently, 41 cases of spurious infection and the true infection of a peccary (T. pecari) and a monkey (Ateles paniscus) were reported in an indigenous amazonian population from Brazil suggesting the potential of these animals as local reservoirs [21]. However, studies are needed to confirm the mechanisms of transmission of C. hepaticum to humans as well as provide evidence of the cycles potentiating this transmission.
More than half of the spurious infections by C. hepaticum reported worldwide in the last decade have been found in Brazil [5]. Ninety-eight percent (81/82) of these cases are from indigenous tribes or rural communities of the Amazon region (from the States of Mato Grosso and Rondônia) [17]–[19], [21]–[25]. Nonetheless, no case of disease has, to date, been described in this region. The probable explanation is diagnostic difficulties in the Amazon that may be attributed to factors such as scarce access to health care services, unawareness of health professionals of the existence of the pathogen and the co-existence of tropical diseases (such as malaria, viral hepatitis, arbovirosis, toxocariasis, among others) [19] which share the same clinical symptoms and signs (typical syndrome for C. hepaticum: persistent fever, hepatomegaly and leukocytosis with eosinophilia) [14] suggesting that hepatic calodiasis is probably neglected in this region [18], [19].
The aim of the present study was to determine the incidence and risk of spurious infection as well as the dynamics of transmission of C. hepaticum in a community in the Brazilian Amazon region.
This study was approved by the Ethics Committee in Investigation of the Oswaldo Cruz Foundation (Protocol 384/07 of 20/08/2007). Written Informed consent was obtained from all the study participants. According to the current regulations of the Brazilian legislation and of the Commission of Ethics in the Use of Animals (CEUA) of the Oswaldo Cruz Foundation, the study of dog feces samples collected from the gardens of households does not require ethical approval because the dogs were not handled or manipulated by the researchers. Dog owners provided prior permission for the collection of dog feces samples from their gardens.
This study was carried out in the agricultural community of Rio Pardo of the municipality of Presidente Figueiredo, located ∼160 Km to the north of the city of Manaus (∼1°48′S; 60°19′W), Amazonas State, Brazil (Figure 1). This community was officially created in 1996 by the National Institute of Colonization and Agricultural Reform (INCRA), in an area of tropical jungle. It is composed of 7 unpaved roads, known locally as “Ramal”, which includes households on both sides of these roads surrounded by tropical rain forest. The community also includes a riverine population living along the Rio Pardo stream known as “Igarapé”.
A population census (October–September of 2008) identified 701 inhabitants in the Rio Pardo community, with 360 (51.4%) living in the Ramal area and 341 (48.6%) in the Igarapé area. Most of the incomers are natives from the Amazon Region and make their livings from subsistence farming, plant harvesting (wood, chestnuts, medicinal herbs), hunting and fishing. Most of the households present precarious basic sewage systems. Health care services are sparsely available in the community.
A cross-sectional coproparasitologic study (Text S1) of 40 randomly selected households was performed in the community in August 2009. One stool sample was collected from each participant and evaluated 1–6 times by the Lutz [26] and/or Paratest (Diagnostek, São Paulo, Brazil) techniques. In addition, feces samples of dogs collected from the gardens of households presenting human cases of C. hepaticum and a liver tissue sample of a wild mammal (captured and eaten by the residents) were analyzed by the Lutz technique. The liver tissue was manually shredded in a NaCl solution at 0.85% prior to performing the diagnostic technique.
Identification of the eggs of C. hepaticum was based on morphologic and morphometric analysis of 20–50 eggs per sample. The morphologic analysis was based on aspects of the structure of the eggshells [27], [28]. Photomicrographs were made with a Leica microscope.
A questionnaire was applied to obtain socio-demographic and epidemiologic information, especially community activities (hunting), individual risk factors (habits of ingestion of game meat) and family practices related to the transmission of the parasite (the habit of sharing game meat with dogs).
The characteristics of the population and the eggs of C. hepaticum were described using tables of frequencies if the variables were qualitative and calculating means, standard deviations, maximum and minimum values if the values were quantitative. Comparisons of groups and the associations among variables were evaluated with Chi-square or Fisher exact tests. Estimations of incidence and relative risk (RR) were made with a 95% confidence interval (CI). The analyses were performed using the SPSS v.18 statistical package and the EPIDAT 3.1. A level of significance of 5% was set.
A total of 135 individuals residing in 40 households in the community participated in the study. The study population was characterized by a predominance of males (60.7%) and adults (65.2%).
Nine cases of spurious infection were identified, representing an incidence of 6.7% (95% CI: 2.08–11.24). The eggs presented morphologic and morphometric characteristics compatible with the species of C. hepaticum, being yellowish-brown in color, barrel-shaped, with shallow polar plugs and radial striations and measuring an average of 64.4 µm in length and 36.7 µm in width (Figure 2, Table 1). The cases were from households located in the area of Igarapé and in one of the Ramals of the community. Of the individuals infected, 55.5% were women and 55.5% children (<14 years of age). The rate of households with spurious infection was 7.5% (95% CI: 1.50–20.38).
Eight out of nine (88.9%) of the cases were found in two households of the Ramal. The rate of intradomiciliary spurious infection was 83.3% (5/6) in one household and 75% (3/4) in the other. All the cases were asymptomatic with the exception of two individuals in the same household who presented diarrhea and were both co-infected by Blastocystis hominis and Salmonella spp. The case of spurious infection from the area of Igarapé was an adult woman, the only participant that could not be found to do the questionnaire. In this latter case, the rate of intradomiciliary spurious infection was 50%.
The habit of game intake was reported by 94.8% (127/134) of the individuals. The animals most frequently consumed were paca (85%), peccary (57.5%), armadillo (42.5%), agouti (37.5%) and deer (37.5%). Game was eaten at least once a week by 25.6%, with the liver of game being eaten by 57.5%. The risk of spurious infection was 10-fold greater in those eating the liver of wild mammals [10% vs. 0% (p = 0.02)].
After undertaking the epidemiological investigation the complete history of the spurious infections in the Ramal was obtained. The residents reported that a few days prior to the coproparasitologic study a group of hunters captured several peccaries and shared the entrails and meat among the residents of the Ramal for food. The two families presenting cases of C. hepaticum reported having eaten the liver of the hunted peccaries.
In addition, one of the families reported that raw meat remaining from the peccary liver that had been consumed was still stored in the freezer of their home. This piece of liver tissue was provided and analyzed in the laboratory, being positive for the presence of typical eggs of C. hepaticum. On average the eggs measured 63.1 µm in length and 36.3 µm in width (Figure 2, Table 1). In this household the consumption of infected liver was the cause of the spurious infection reported in 83.3% (5/6) of the residents. The host was probably a Pecari tajacu or T. pecari since there are only two species of peccary in the study area.
Some families reported the habit of giving game meat (raw) with their dogs as food. We estimated that 7.5% (3/40) of the families surveyed did this. Two dog feces samples were collected from the gardens of the two Ramal households presenting human cases of C. hepaticum. One of the samples analyzed presented eggs with characteristics compatible with species C. hepaticum, measuring an average of 61.1 µm in length and 35.4 µm in width. (Figure 2, Table 1).
In the present study we describe a rate of spurious infection of 6.7% in a rural community of the Amazon, being, to our knowledge, one of the highest reported to date. This rate was similar to that estimated for indigenous people of the northwest of State of Mato Grosso (8.6%) [21] and of the Suruí etnia in Rondônia (5.2%), in the Brazilian Amazon [25], indicating that the Amazon region has the highest incidence of spurious infection worldwide. Other studies have reported lower rates ranging from 0.2% to 2.3% [17], [22]–[24]. It should be noted that the rate estimated here might have been lower than that presented if all the samples had been evaluated only once.
Three capillarid species of zoonotic importance are known: C. hepaticum, Eucoleus aerophilus (syn. Capillaria aerophila) and Paracapillaria (Crossicapillaria) philippinensis (syn. Capillaria philippinensis) [5]. E. aerophilus is widespread and parasitizes the trachea and mainly the bronchi of dogs, cats, wild carnivores and, occasionally, humans [29], [30]. P. philippinensis is a parasite of fish, endemic in Philippines and Thailand and is the etiologic agent of human intestinal capillariasis [31]. Only the species C. hepaticum has been reported in Brazil.
Eggs of C. hepaticum, E. aerophilus and P. philippinensis can be found in human feces and can be differentiated. In capillarids, different aspects of the eggshell structure can be used as a taxonomic clue [27], [28]. The combination of morphologic and morphometric analysis of the eggs allows the identification of species of capillarids at a light microscopy level [27], [32]. Morphologic characteristics of the bipolar plug (asymmetric in E. aerophilus, inconspicuous flattened in P. philippinensis), the shell (with a network of anastomosing ridges in E. aerophilus, striated in P. philippinensis) and the shape (peanut like in P. philippinensis) can be used for differentiation of the eggs [30], [33]. The morphology of the eggs found in this study (from dog feces, human stools and liver tissue) was compatible with the species C. hepaticum (presence of shallow polar plugs and radial striations) with dimensions according to those described by previous authors (40–75 µm in length×27–41.3 µm in width) [8], [14], [18], [21], [27], [34].
We report a frequent habit of wild mammal meat (94.8%) and liver (57.5%) intake similar to previous studies in Amazon populations and in indigenous tribes [18], [19], [25]. Recently, in a river-side population from the State of Rondônia (western Brazilian Amazon) with a high consumption (91.7%) of meat from wild mammals (paca, agouti or peccary), the serum prevalence of C. hepaticum was 34.1% at a dilution of 1∶150, suggesting frequent contact with eggs of C. hepaticum [19].
Mild diarrhea has been reported in spurious infection of C. hepaticum, although this type of infection appeared to be asymptomatic in most cases [35]. In this study most individuals were asymptomatic, but the occurrence of diarrhea in two subjects could not be attributed to spurious infection by C. hepaticum due to the concomitant presence of two potential agents of diarrhea (B. hominis and Salmonella spp.).
This is the first report of a causative source of spurious infection of humans by C. hepaticum, that of peccary liver. Peccaries of the species T. pecari and P. tajacu are natural reservoirs of C. hepaticum [16], [21], [36], are widely distributed in Brazil [37], and are one of the wild mammals most frequently used as food in Brazilian amazonian communities [19]. For these reasons we suggest that these animals can be an important source of spurious infection for humans in the Amazon region. In Brazil, liver infection by C. hepaticum has been described in domestic dogs and cats and other mammals of the subfamilies Murinae (R. novergicus, Rattus rattus and M. musculus), Sciurinae (Sciurus aestuans), Caninae (Lycalopex gymnocercus, Cerdocyon thous and Chrysocyon brachyurus), Tayassuinae (P. tajacu and T. pecari), Felinae (Puma concolor) and Atelinae (A. paniscus) [21], [34], [36], [38]–[40].
We estimated, for the first time, that individuals who usually eat the liver of wild mammals present a 10-fold higher risk of presenting spurious infection than those without this habit. As a consequence of this alimentary habit the spurious infection showed previously unreported high intradomiciliary rates (50% to 83.3%), characterized as infection by household clusters. The present results confirm the suspicion of several authors as to the existence of the mechanism of transmission by the ingestion of non embryonated eggs present in the liver of mammals and their involvement as a cause of spurious infection in humans. This thereby allows the conclusion that this is an important mechanism of transmission of eggs of C. hepaticum in this area and probably also in other areas of the Amazon with similar sociocultural characteristics.
Eggs characteristic of the species C. hepaticum were found in a sample of dog feces collected from the garden of one household presenting cases of spurious infection. It is known that domestic dogs are susceptible to infection by C. hepaticum [40], [41] and other capillarid species (E. aerophilus and Eucoleus boehmi) [32]. E. boehmi (syn. Capillaria boehmi) is a parasite of the nasal cavities and sinuses of wild canines (e.g. foxes and wolves) and domestic dogs, and its eggs can also be found in feces. Eggs present asymmetrical plugs, tiny pits on the surface of the wall and measure 50–60 µm×30–35 µm [32]. E. aerophilus have been described in dogs from Europe, North America and Australia and E. boehmi in dogs from Europe and North America [32]. Only the species C. hepaticum has been described in domestic dogs from Brazil.
The spurious infection by C. hepaticum of a pet within a setting presenting human spurious infections has not been previously described. This finding may be related to the report of the families about having given raw game meat to the dogs. The practice of feeding pets with raw meat and close living relationships between humans and pets have previously been suggested as having an important role in the transmission of zoonotic pathogens [42], [43].
This suggests that dogs may potentiate the emergence of a peridomestic cycle of C. hepaticum in this area. Since the dogs usually deposit their feces around the household, a new epizootic focus could be established very close to the family thereby increasing the risk of spurious and hepatic infections and even the development of cases of disease, especially among children. Children are more likely to be infected because of pica (especially geophagia) [5]. The deficient sanitary conditions in the community studied may be another important factor contributing to the risk of further infections. This last characteristic is common in rural communities which routinely hunt in the Amazon region [18], [19], [25], suggesting the risk of the emergence of cases in other populations.
We therefore recommend the implementation of an epidemiologic surveillance system for the diagnosis of spurious infection (with correct microscopic identification of the parasite) in areas in which the population has the habit of eating game meat. To prevent mix-ups, laboratory technicians could be trained to differentiate the eggs of Trichuris trichiura from those of capillarids [5], taking into account morphologic and morphometric characteristics. Since Trichuris spp. eggs have smooth walls they can be distinguished from the mainly ornamented eggs of the capillarids [44].
Moreover, in areas presenting spurious infections, we recommend the investigation of C. hepaticum in subjects with clinical suspicion of hepatic disease by serology and, if necessary, histopathological examination of liver biopsy samples [5]. To date, there are no molecular tools for the detection of C. hepaticum. As measures of prevention it should be recommended that families should cook the liver well prior to ingestion and should not feed dogs with raw entrails. Improvements in local sanitary conditions should also be implemented.
Investigation of the sources of infection in areas in which the presence of spurious infection has been confirmed is advisable, including the mammals most frequently consumed and small rodents. In the latter case, several studies have described the adaptation of some small rodents (Rhipidomys spp. and Mesomys spp.) to villages and households located in deforested areas of the Amazon invaded by man [45], [46]. Thus, their role in the dynamics of peridomiciliary transmission in rural Amazon areas should also be evaluated. In addition, the species M. musculus and R. rattus, which are widely distributed reservoirs of C. hepaticum in Brazil (that adopts the human household or its proximities as its habitat), have already been described in an area of the Amazon biome with recent human occupation [47]. Near the location of the present study, in an area of forest reserves (Minimum Critical Size of Ecosystems reserves), small rodents of some subfamilies, such as Sigmodontinae (e.g. Euryoryzomys macconnelli, Hylaeamys megacephalus and Rhipidomys nitela) and Eumysopinae (e.g. Proechimys cuvieri) have been found. Moreover, known C. hepaticum reservoirs, such as peccaries (P. tajacu and T. pecari), A. paniscus and P. concolor have been described in the area [48].
This is the first study to identify a source of spurious infection of C. hepaticum in humans (peccary liver) in a rural community of the Brazilian Amazon. A high rate of incidence in household clusters is described in relation to the habit of the ingestion of liver of wild mammals. The finding of contaminated peridomiciliary ground with an infected dog feces suggests greater risk of new infections without the participation of a wild agent. The dynamics of transmission found in the community studied led to the conclusion that the mechanism of transmission following the ingestion of liver of wild mammals is an important mechanism in this area.
|
10.1371/journal.pmed.1002418 | Comparison of two cash transfer strategies to prevent catastrophic costs for poor tuberculosis-affected households in low- and middle-income countries: An economic modelling study | Illness-related costs for patients with tuberculosis (TB) ≥20% of pre-illness annual household income predict adverse treatment outcomes and have been termed “catastrophic.” Social protection initiatives, including cash transfers, are endorsed to help prevent catastrophic costs. With this aim, cash transfers may either be provided to defray TB-related costs of households with a confirmed TB diagnosis (termed a “TB-specific” approach); or to increase income of households with high TB risk to strengthen their economic resilience (termed a “TB-sensitive” approach). The impact of cash transfers provided with each of these approaches might vary. We undertook an economic modelling study from the patient perspective to compare the potential of these 2 cash transfer approaches to prevent catastrophic costs.
Model inputs for 7 low- and middle-income countries (Brazil, Colombia, Ecuador, Ghana, Mexico, Tanzania, and Yemen) were retrieved by literature review and included countries' mean patient TB-related costs, mean household income, mean cash transfers, and estimated TB-specific and TB-sensitive target populations. Analyses were completed for drug-susceptible (DS) TB-related costs in all 7 out of 7 countries, and additionally for drug-resistant (DR) TB-related costs in 1 of the 7 countries with available data. All cost data were reported in 2013 international dollars ($). The target population for TB-specific cash transfers was poor households with a confirmed TB diagnosis, and for TB-sensitive cash transfers was poor households already targeted by countries’ established poverty-reduction cash transfer programme. Cash transfers offered in countries, unrelated to TB, ranged from $217 to $1,091/year/household. Before cash transfers, DS TB-related costs were catastrophic in 6 out of 7 countries. If cash transfers were provided with a TB-specific approach, alone they would be insufficient to prevent DS TB catastrophic costs in 4 out of 6 countries, and when increased enough to prevent DS TB catastrophic costs would require a budget between $3.8 million (95% CI: $3.8 million–$3.8 million) and $75 million (95% CI: $50 million–$100 million) per country. If instead cash transfers were provided with a TB-sensitive approach, alone they would be insufficient to prevent DS TB-related catastrophic costs in any of the 6 countries, and when increased enough to prevent DS TB catastrophic costs would require a budget between $298 million (95% CI: $219 million–$378 million) and $165,367 million (95% CI: $134,085 million–$196,425 million) per country. DR TB-related costs were catastrophic before and after TB-specific or TB-sensitive cash transfers in 1 out of 1 countries. Sensitivity analyses showed our findings to be robust to imputation of missing TB-related cost components, and use of 10% or 30% instead of 20% as the threshold for measuring catastrophic costs. Key limitations were using national average data and not considering other health and social benefits of cash transfers.
A TB-sensitive cash transfer approach to increase all poor households’ income may have broad benefits by reducing poverty, but is unlikely to be as effective or affordable for preventing TB catastrophic costs as a TB-specific cash transfer approach to defray TB-related costs only in poor households with a confirmed TB diagnosis. Preventing DR TB-related catastrophic costs will require considerable additional investment whether a TB-sensitive or a TB-specific cash transfer approach is used.
| Household costs related to active drug-susceptible (DS) or drug-resistant (DR) tuberculosis (TB) disease include costs for consultations, transport to and from clinics, increased food needs and lost income. If these costs are greater than or equal to one-fifth (20%) of the household’s annual income, then the patient is at risk of unsuccessful TB treatment and these high costs are termed “catastrophic costs.”
The World Health Organization’s End TB Strategy prioritises preventing TB-affected households from facing catastrophic costs and proposes cash transfers as one way to achieve this. However, there are at least 2 approaches by which cash transfers could be provided to TB-affected households. In the first, they are provided to defray/reimburse households’ TB-related costs (termed a “TB-specific” approach). In the second, they are provided to increase households’ pre-illness income to prevent poverty and strengthen their economic resilience (termed a “TB-sensitive” approach).
Lack of available individual-level data sources has meant that no studies have compared a TB-specific versus a TB-sensitive cash transfer approach. A literature review combined with a secondary data analysis was an effective way to bring together relevant data from several different sources and model the potential of cash transfers provided by these 2 approaches to prevent catastrophic costs.
We performed a rigorous review of public data sources available on the internet, extracting national average data published between 2005 and 2013 for the 7 low- and middle-income economy countries of Brazil, Colombia, Ecuador, Ghana, Mexico, Tanzania, and Yemen. The data values we extracted included the countries’ mean value of TB patient costs, mean household income, mean cash transfers, and the expected size of the population that would be targeted with either TB-specific or TB-sensitive cash transfers. In all 7 countries these analyses were completed for DS TB, and in 1 of the 7 countries we were also able to complete these analyses for DR TB.
Expressing TB patient costs as a percentage of household income, we found that average DS TB costs were catastrophic in 6 out of the 7 countries included in the study. In these 6 countries, TB-specific cash transfers prevented DS TB catastrophic costs in only 2, whilst TB-sensitive cash transfers did not prevent DS TB catastrophic costs in any of them. In the 1 country with available data, average DR TB costs were catastrophic, and neither TB-specific nor TB-sensitive cash transfers were sufficient to prevent this.
For both TB-specific and TB-sensitive approaches, we then estimated the total value that cash transfers would need to be increased to in order to prevent the countries’ average DS or DR TB costs from being catastrophic for DS or DR TB-affected household. Based on this, we also estimated the average budget that each country would need to prevent catastrophic costs for all DS or DR TB-affected households. We found that a TB-specific approach was much more affordable than a TB-sensitive approach.
The potential of cash transfers to prevent TB-related catastrophic costs is greater if they are provided to defray/reimburse poor households’ costs (TB-specific approach) rather than to increase the income and strengthen the economic resilience of poor households with high TB risk (TB-sensitive approach).
Where cash transfers are insufficient to prevent catastrophic costs, it will be cheaper to supplement their value to achieve this objective using a TB-specific approach rather than a TB-sensitive approach.
Important study limitations were that the study was at the country-level, so we might have underestimated the potential of cash transfers to prevent catastrophic costs. Also, we did not assess other health and social benefits of cash transfers, so the impact of TB-specific versus TB-sensitive cash transfers was only judged from the perspective of preventing catastrophic costs.
| Tuberculosis (TB) disproportionately affects poor households in low- and middle-income countries that are least able to afford the burden that TB-related costs represent relative to their income [1–6]. Even when diagnosis and treatment is available free of direct charges, TB-affected households are known to incur hidden “out of pocket” direct medical costs (e.g., for consultations) and direct nonmedical costs (e.g., for transport, additional food and symptomatic medicines), as well as indirect costs from lost income [7,8]. Combined, these costs can have severe consequences for affected households. They hinder patients’ access to care and increase their odds of adverse TB treatment outcomes, which are abandoning or failing treatment, dying during treatment, or having recurrent TB within 30 months of starting TB treatment [9–15]. They also force some households to engage in damaging financial coping strategies, which sometimes referred to collectively as dissaving, include taking out a loan, selling productive assets, reducing consumption expenditure to below basic needs, taking children out of education, and/or taking out large loans [16]. Two groups of households that are especially vulnerable to TB-related costs are those in the countries’ poorest population quintile and those affected by drug-resistant (DR) TB [7].
Addressing households’ TB-related costs is essential for ensuring that people with active TB disease are able to access TB diagnosis and treatment. Acknowledging this, the World Health Organization’s End TB Strategy includes a high-level financial risk protection milestone for 2020: “zero TB-affected households facing catastrophic costs due to TB” [17,18]. In this milestone, “catastrophic costs” refers to a combination of direct medical, direct nonmedical, and indirect costs excessive enough to increase a patient’s risk of adverse TB treatment outcome and/or force their household to engage in damaging financial coping strategies [19]. By encompassing all 3 cost components, the term “catastrophic costs” is distinct from the term “catastrophic health expenditure,” which only considers direct medical costs and is used to monitor progress towards financial risk protection as part of universal health coverage [19]. As part of the End TB Strategy, research has focussed on developing an empirical measure of catastrophic costs. Recently, total TB-related costs greater than or equal to 20% of TB-affected households’ pre-illness annual income have been found to significantly increase the likelihood of TB patients experiencing an adverse treatment outcome and their household engaging in damaging coping strategies [14,15]. As the only indicator established to be clinically and financially relevant for assessing a household’s ability to pay for TB care, this measurement of catastrophic costs has tentatively been included by the Global TB Programme in a pilot tool to monitor catastrophic costs of TB-affected households worldwide [19].
Preventing catastrophic costs for TB-affected households is a priority for facilitating individuals’ access to TB diagnosis and treatment, increasing their likelihood of treatment success and reducing onwards TB transmission [18]. With this objective, the Global TB Programme endorses social protection initiatives including cash transfers, food baskets, social insurance and labour market measures to complement universal health coverage initiatives like prepayment, resource pooling, and patient-friendly service delivery [19]. In the TB literature, evidence from a randomized trial in Peru shows that when provided as incentives to support TB treatment, cash transfers reduce poor TB-affected households’ likelihood of incurring catastrophic costs, as well as improve patients’ likelihood of TB treatment success, and uptake of preventative therapy amongst people they are in close contact with (e.g., family, friends, care giver) [15,20]. Outside of the TB literature, synthesised evidence from governmental poverty-reduction policies in several low- and middle-income countries provides evidence that cash transfers increase poor households’ income and consumption expenditure, help them cope with livelihood risks (e.g., illness and unemployment), and support family investments in the human capital of their children (e.g., sending them to school and taking them to regular health checks) [21–23].
Currently, there are at least 2 alternative approaches proposed in the TB literature for providing cash transfers to TB-affected households [24]. The first is termed a “TB-specific” approach, whereby cash transfers would be targeted to poor households with a confirmed TB diagnosis to incentivise and enable TB treatment by defraying their TB-related costs [24]. This approach is exemplified by the cash transfer component of the Community Randomized Evaluation of a Socioeconomic Intervention to Prevent TB (CRESIPT) trial in Peru [25,26]. The second is termed a “TB-sensitive” approach, whereby cash transfers would be targeted to poor households at high risk of developing active TB disease to increase their income, thereby protecting them from poverty-related risk factors for TB infection, progression, and adverse treatment outcomes (e.g., poor living conditions and undernutrition), as well as strengthen their economic resilience to TB-related costs [24]. This approach already exists in many low- and middle-income countries, and is exemplified by governmental poverty-reduction cash transfer programmes like Programa Bolsa Familia in Brazil [27,28].
Depending on whether cash transfers are provided with a TB-specific or a TB-sensitive approach, their impact might vary [24]. We aimed to investigate how this might relate to the potential of cash transfers to prevent catastrophic costs.
With no known data sources for investigating if the potential of TB-specific and TB-sensitive cash transfers to prevent catastrophic costs varies, we undertook an economic modelling study using published national average data gathered from a rigorous review of the literature. Our economic modelling study was aggregated at the country level. The setting was low- and middle-income countries where over 95% of TB cases live and where formal institutions to protect households from the social and economic impacts of illness are weakest [29]. The intervention being investigated was cash transfers paid to poor households, and the alternative approaches being compared were: (1) cash transfers provided to defray TB-related costs of households with a confirmed TB diagnosis (termed a “TB-specific” approach); versus (2) cash transfers provided to increase income of households with high TB risk and strengthen their economic resilience (termed a “TB-sensitive” approach). These approaches were compared because of current uncertainty about the potential of each approach to prevent catastrophic costs. Using only TB-related costs incurred by patients, study outcomes were assessed from the patient perspective.
Primary study outcomes were an indicator for catastrophic costs after TB-specific versus TB-sensitive cash transfers, and the countries’ country-level cash transfer budget needed to prevent catastrophic costs for each of these approaches. Catastrophic costs were estimated over a time horizon from the onset of TB symptoms to TB treatment completion. The countries’ country-level cash transfer budgets were estimated over a time horizon of 1 year. In the 1 country with available data, outcomes were investigated separately for drug-susceptible (DS) TB and DR TB. The reason for this is that treatment of DR TB versus DS TB is longer and more intensive and is therefore associated with much higher TB-related costs [7]. The study used cross-sectional data drawn from secondary sources. Data inputs were countries’ mean patient TB-related cost, mean pre-illness household income, mean poverty-reduction cash transfer, and TB-specific versus TB-sensitive target populations. Inputs were retrieved by reviewing TB-related cost and cash transfer literature and countries’ national statistics. Because there was insufficient data across low- and middle-income countries on programmes providing cash transfers with a TB-specific approach, this study compared cash transfers offered by existing governmental poverty-reduction programmes as if they were provided with a TB-specific versus a TB-sensitive approach.
For transparency, the study was reported according to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) checklist [30]. The completed checklist is provided in S1 CHEERS checklist. The study’s prospective analysis plan is provided in S1 Text. In the present analysis, we did not attempt to model the potential of TB-inclusive cash transfers to prevent catastrophic costs, and results from key informant interviews are reported elsewhere [31]. Extraction of cash transfer target population data, estimation of 95% confidence intervals (95% CIs), and our sensitivity analyses were added in the peer review process. Key study definitions are listed in Box 1.
In this study, the target population for cash transfers provided with a TB-specific approach was households in countries’ poorest population quintile with a confirmed TB diagnosis. Guidance is not currently available for which TB-affected households should be targeted with a TB-specific approach. We chose to focus on TB-affected households in countries’ poorest population quintile because they are typically at greater risk of incurring catastrophic costs [14]. Whilst it might have been preferable to focus on all TB-affected households that incur catastrophic costs, at the time of analysis no estimates of the size of this population were available in any countries included in this study. The target population for cash transfers provided with a TB-sensitive approach was households in poverty already targeted by countries’ established governmental poverty-reduction cash transfer programme.
To allow comparison of monetary data extracted in different currencies and measured in different years, all extracted monetary values were inflated and converted to 2013 international dollars using the purchasing power parity conversion factor that accounts for differences in the cost of living across countries [53,54].
In countries that had missing values for direct or indirect costs pre- or during-treatment, we estimated their value. To do this, we assumed that average TB-related costs followed a make-up of cost components equivalent to the one synthesised by Tanimura et al. in their systematic review of TB-related costs in low- and middle-income countries, which is that direct and indirect costs are equivalent to 40% and 60% of total costs respectively, and pre- and during-treatment costs are each equivalent to 50% of total costs respectively [7]. Because only 1 included cost survey reported the standard deviation of total costs [33], we also assumed that average TB-related costs had a standard deviation with the same ratio to total costs as the one estimated by Tanimura et al. for average total costs across all low- and middle-income countries, which was 1.1 [7]. We used the assumed standard deviation and the sample size of countries’ cost surveys to calculate 95% CIs for estimated TB-related costs.
All analyses used published mean national data. To account for uncertainty in the value of extracted TB-related costs, annual household income, and cash transfers, we conducted a multiway analysis that allowed all 3 of these inputs to vary simultaneously according to their sampling distributions. Sampling distributions were simulated from 10,000 computationally generated random samples and were all assumed to have normal distributions according to the central limit theorem. Random samples were generated for TB-related costs using a standard deviation with a ratio of 1.1 to mean estimates, which was the ratio estimated by Tanimura et al. for average total costs across all low- and middle-income countries, and a sample size equivalent to countries’ cost surveys [7]. For annual household income, we used a standard deviation with a ratio of 0.8 to mean estimates, which was the average observed across 2 studies investigating the household-level income effect of poverty-reduction cash transfer programmes in Brazil and Colombia [37,44] and a sample size equivalent to countries’ household income surveys [55–61]. For cash transfers, we used a standard deviation with a ratio to mean estimates equivalent to a quarter of maximum cash transfers minus minimum cash transfers, and a sample size equivalent to the one reported in studies from which we extracted mean cash transfers. In Ecuador and Ghana, we did not simulate sampling distributions for cash transfers because, respectively, all beneficiary households receive the same flat cash transfer, and the mean cash transfer we extracted was estimated from all beneficiary households. Throughout our analysis, 95% CIs were calculated for model estimates using the quantile method. All analyses were run in R version 3.3.0.
We tested the sensitivity of our results in Brazil, Colombia, Tanzania, and Mexico to imputation of missing DS TB-related cost components by repeating our analysis omitting rather than imputing the value of missing DS TB-related cost components [7]. We separately tested the sensitivity of our results across all countries included in the study to the use of 20% as the threshold for measuring countries’ TB-related cost burden as catastrophic. We did this by repeating our analyses instead using a 10% and 30% threshold.
Fig 1 is a flow chart of the review process for assessing the eligibility of countries for inclusion in this study. Argentina, Bangladesh, and South Africa had to be excluded after insufficient publically available background information was identified for eligible cash transfer programmes in these countries. Consequently, 7 countries were included in the data analysis.
Conducted in Brazil, Colombia, Ecuador, Ghana, Mexico, Tanzania, and Yemen between 2006 and 2012, survey sample sizes ranged from 94 to 320 patients with active DS TB disease (Table 1). Surveys collected data on DS TB-related costs incurred pre- and during-treatment, except in Brazil [33], Colombia [62], and Tanzania [63], where they only collected data during-treatment (Table 1). Surveys collected both direct and indirect cost data, except in Mexico [64] where no data was collected characterising indirect costs (Table 1). In countries where data was collected, methods for estimating indirect costs varied in 2 ways: 1) reported time lost travelling and waiting to receive TB care was multiplied by patients’ reported income [33,65]; or 2) reported time lost travelling and waiting to receive TB care was multiplied by an estimate of national average income (gross national income per capita or official wage rate) [33,62,63,66]. In Ecuador [66], data was collected on additional costs described in the publication as referring to “loans, paying for additional help and other impacts throughout the course of TB illness.” The ambiguity of this cost category meant that it could not be classified as either direct or indirect costs and was thus reported as its own subcategory. Reported mean DS TB-related total costs for the complete TB illness ranged from $387 to $2,382 (Table 1). After imputing missing TB-related cost components in Brazil, Colombia, Mexico and Tanzania, estimated mean DS TB-related total costs ranged from $774 (95% CI: $618–$930) to $5,954 (95% CI: $4,997–$6,911), Table 1.
Conducted in Ecuador in 2007, the survey sample size was 14 patients with active multidrug-resistant TB disease, Table 1. The survey reported mean DR TB-related costs incurred pre- and during-treatment (Table 1). The survey collected both direct and indirect cost data (Table 1). Cost data was also collected on additional costs. This category of costs was reported as its own subcategory. Indirect costs were estimated by multiplying reported time lost travelling and waiting to receive TB care by the estimated hourly wage in Ecuador. Mean DR TB-related total costs were $16,667 (95% CI: $7,063–$26,271), Table 1.
All extracted cash transfer data refer to programmes’ status in 2013. Mean cumulative annual cash transfers were greatest in Brazil, Colombia, Ecuador, Mexico, and Yemen varying between $823 (range: $239–$1,084) and $1,091 (range: $1,091–$1,091); and lowest in Ghana and Tanzania where they were $217 (range: $150–$299) and $451 (range: $349–$655), respectively (Table 2). Across countries, cash transfers ranged from 7.7% (95% CI: 7.6%–7.9%) to 43% (95% CI: 42%–44%) of annual household income. In Colombia, Ecuador, Ghana, Mexico, and Tanzania they varied between 13% (95% CI: 11%–17%) and 59% (95% CI: 50%–72%) of DS TB-related costs, and in Brazil and Yemen, respectively, they were 104% (95% CI: 93%–119%) and 106% (95% CI: 88%–133%) of DS-TB-related costs (Table 2). In Ecuador, cash transfers represented 7.3% (95% CI: 4.2%–15%) of DR TB-related costs (Table 2). A summary of cash transfer data sources and additional extracted data is provided in S1 Table.
Conducted between 2005 and 2011, survey sample sizes ranged from 8,687 to 55,970 households [55–61]. Surveys reported mean household income, except in Tanzania where mean household expenditure was reported. Estimated mean annual household income in countries’ poorest population quintiles was highest in Brazil, Ecuador, and Mexico varying between $4,755 and $8,692, and lowest in Colombia, Ghana, Tanzania, and Yemen varying between $1,617 and $2,812. A summary of annual household income data sources and extracted data is provided in S1 Table.
For DS TB, the size of countries’ estimated TB-specific target population, which was equivalent to 40% of countries’ TB burden, ranged from 3,520 to 67,600 households, and the size of countries’ estimated TB-sensitive target population, which was equivalent to the number of households in poverty already targeted by countries’ established poverty-reduction cash transfer programme, ranged from 70,000 to 26 million households (Table 2). For DR TB, the size of Ecuador’s estimated TB-specific target population was 300 households, and the size of its estimated TB-sensitive target population was 450,000 households (Table 2).
Before cash transfers, estimated DS TB-related cost burdens varied between 15% (95% CI: 12%–18%) and 125% (95% CI: 105%–145%) of annual household income, and were catastrophic in Colombia, Ecuador, Ghana, Mexico, Tanzania, and Yemen where they varied between 27% (95% CI: 21%–32%) and 125% (95% CI: 105%–145%) of annual household income (Fig 2).
If cash transfers were applied using a TB-specific approach to defray TB-related costs incurred by households with a confirmed DS TB diagnosis, then on average, they were sufficient to prevent catastrophic costs in Ecuador and Yemen, but insufficient to prevent them in either Colombia, Ghana, Mexico, or Tanzania (Fig 2). In Colombia, Ghana, Mexico, or Tanzania, the DS TB-related cost burden after TB-specific cash transfers varied between 26% (95% CI: 15%–38%) and 106% (95% CI: 86%–126%), and the estimated value of household-level additional TB-specific cash transfer needed to prevent DS TB catastrophic costs varied between $144 (95% CI: $0.0–$387) and $4,071 (95% CI: $3,122–$5,014), Table 3. In the 6 countries where TB-related costs were originally catastrophic, the estimated value of household-level total TB-specific cash transfer needed to prevent DS TB catastrophic costs varied between $850 (95% CI: $627–$1,079) and $5,011 (95% CI: $4,063–$5,952), Table 3. According to the size of countries’ TB-specific target populations, this value translated into a TB-specific country-level cash transfer budget needed to prevent DS TB catastrophic costs varying between $3.8 million (95% CI: $3.8 million–$3.8 million) and $75 million (95% CI: $50 million–$100 million), Fig 3.
If cash transfers were provided using a TB-sensitive approach to increase pre-illness income of poor households with high risk of developing active TB disease, then on average, for those that later develop active DS TB disease, this would not be sufficient to prevent them from incurring catastrophic costs in any of the 6 countries where DS TB-related costs were originally catastrophic (Fig 2). In these 6 countries, the DS TB-related cost burden after TB-sensitive cash transfers varied between 24% (95% CI: 19%–29%) and 105% (95% CI: 88%–121%), and the estimated value of household-level additional TB-sensitive cash transfer needed to prevent DS TB catastrophic costs varied between $1,360 (95% CI: $821–$1,897) and $24,115 (95% CI: $19,374–$28,817), Table 3. The estimated value of household-level total TB-sensitive cash transfer needed to prevent DS TB catastrophic costs varied between $2,282 (95% CI: $1,743–$2,819) and $25,055 (95% CI: $20,316–$29,761), Table 3. According to the size of countries’ TB-sensitive target populations, this value translated into a TB-sensitive country-level cash transfer budget needed to prevent DS TB catastrophic costs varying between $298 million (95% CI: $219 million–$378 million) and $165,367 million (95% CI: $134,085 million–$196,425 million), Fig 3.
In Ecuador, the DR TB-related cost burden before cash transfers was 192% (95% CI: 86%–299%), Fig 2. Here, cash transfers provided with either a TB-specific or a TB-sensitive approach were, on average, insufficient to prevent DR TB catastrophic costs (Fig 2). The estimated value of TB-specific versus TB-sensitive additional cash transfer needed to achieve this objective was $13,782 (95% CI: $4,274–$23,376) versus $73,275 (95% CI: $25,736–$121,246); and the estimated value of household-level total TB-specific versus TB-sensitive cash transfers needed was $14,877 (95% CI: $5,365–$24,467) versus $74,375 (95% CI: $26,827–$122,337), Table 3. According to the size of Ecuador’s DR TB-specific and DR TB-sensitive target population, this value translated into a country-level cash transfer budget needed to prevent DR TB catastrophic costs of $4.5 million (95% CI: $1.6 million–$7.3 million) with a TB-specific approach versus $33,469 million (95% CI: $12,072 million–$55,052 million) with a TB-sensitive approach (Fig 3).
Before cash transfers, the TB-related cost burden remained catastrophic in the same countries as when missing TB-related cost components were imputed, and the only difference after cash transfers was that TB-specific cash transfers prevented catastrophic costs in Colombia (S4 Table). Across countries, TB-specific cash transfers remained more affordable at preventing catastrophic costs compared to TB-sensitive cash transfers both at the household and country level (S5 Table).
Before cash transfers, in addition to Colombia, Ecuador, Ghana, Mexico, Tanzania, and Yemen, the DS TB-related cost burden was also catastrophic in Brazil. In Ecuador, the DR TB-related cost burden before cash transfers remained catastrophic. Across countries, TB-specific cash transfers remained more affordable than TB-sensitive cash transfers for preventing DS and DR TB catastrophic costs both at the household and country level (S6 Table).
Before cash transfers, the DS TB-related cost burden remained catastrophic in Colombia, Ghana, Mexico, Tanzania, and Yemen, but ceased to be catastrophic in Ecuador. In Ecuador, the DR TB-related cost burden before cash transfers remained catastrophic. Across countries, TB-specific cash transfers remained more affordable than TB-sensitive cash transfers for preventing DS and DR TB catastrophic costs both at the household and country level (S7 Table).
In the 7 countries that met our inclusion criteria, our analysis of national average data suggests that DS TB-related costs would be catastrophic for the average poor TB-affected household in most low- and middle-income countries. This is concerning and concordant with the limited evidence that is already available [7]. If cash transfers were provided with a TB-specific approach to defray TB-related costs of poor households with a confirmed DS TB diagnosis, then in some low- and middle-income countries, they would likely be sufficient to prevent the average household incurring DS TB catastrophic costs. Alternatively, if the same value of cash transfers were provided with a TB-sensitive approach to increase the income and strengthen the economic resilience of poor households at high risk of developing active TB disease, then across low- and middle-income countries, they would likely be insufficient to prevent the average household that later developed active DS TB disease incurring DS TB catastrophic costs. In countries where neither TB-specific nor TB-sensitive cash transfers would be sufficient to prevent DS TB catastrophic costs, the average value of household-level additional cash transfer needed to achieve this objective would be much lower using a TB-specific approach compared to a TB-sensitive approach. Further, by only targeting poor households with a confirmed TB diagnosis, a TB-specific approach would, on average, require a much smaller country-level budget than using a TB-sensitive approach to target much larger numbers of poor households at high risk of developing active TB disease.
Although DR TB is rare, it is associated with extreme TB-related costs [7]. Neither TB-specific nor TB-sensitive cash transfers would be sufficient to prevent DR TB catastrophic costs for the average poor DR TB-affected household. The value of household-level additional cash transfer needed to achieve this objective would be very high. Because few poor households are affected by DR TB, countries’ county-level cash transfer budget needed to prevent DR TB catastrophic costs would, on average, be much lower using a TB-specific approach compared to a TB-sensitive approach. Given that so few households are affected by DR TB, it may not be rational for TB-sensitive cash transfer programmes to aim to increase households’ annual income sufficiently to make all poor households resilient to the rare and extreme costs associated with DR TB.
To our knowledge, our study is the first to compare the potential of cash transfers provided with a TB-specific versus a TB-sensitive approach to prevent catastrophic costs. The contrasting effects of defraying TB-related costs using a TB-specific approach versus increasing households’ pre-illness income using a TB-sensitive approach has important and novel implications for protecting TB-affected households from catastrophic costs. We believe our study is also the first to compare the country-level cash transfer budget that would be needed to prevent catastrophic costs for poor TB-affected households using a TB-specific versus a TB-sensitive approach. We show that by being more effective and aiming to reach fewer households, a TB-specific approach would cost less than a TB-sensitive approach. It is important to emphasize that these findings are only valid when preventing catastrophic costs is the only outcome of interest. Cash transfers provided to poor households with a TB-sensitive approach might have far-reaching effects on wellbeing, health promotion, and disease prevention, and further evaluation is needed to study the costs versus benefits of each approach [24,67]. Nevertheless, the End TB Strategy prioritises ensuring that 0 TB-affected households experience catastrophic costs [17]. For achieving this specific milestone, the implications of our study are clear: cash transfers provided with a TB-specific approach are likely to achieve this goal more affordably than if they were provided with a TB-sensitive approach.
Our study adds to limited evidence informing the best targeting strategy for cash transfers aimed at enhancing TB care and prevention [24]. At the country-level, showing that in Latin America and Central Asia a TB-sensitive approach might reach between 12% and 35% of countries’ population, whereas in some parts of sub-Saharan Africa it might only reach between 1% and 2% of countries population, this study supports speculation that the potential of countries to provide cash transfers with a TB-sensitive approach might follow an inverse care law [68], whereby poorer countries with higher TB burdens have less well established poverty-reduction cash transfer programmes [24]. Showing also that approximately 40% of TB-affected households might be in the poorest population quintile, this study highlights the need to consider how cash transfers might be targeted to households that incur catastrophic costs but are outside of this population category [50,51]. With a TB-specific approach, it would be relatively easy to modify programmes’ target population to include more TB-affected households, whilst with a TB-sensitive approach, it might be harder to modify the target population of existing poverty-reduction programmes, which are usually well-established parts of national social protection systems [35].
This study has several limitations, and conclusions should be drawn cautiously. Insufficient data forced us to estimate the potential of TB-specific versus TB-sensitive cash transfers to prevent catastrophic costs using the value of cash transfers offered by existing governmental poverty-reduction cash transfer programmes. Whilst the only solution, it will have nonetheless under- or overestimated the potential of TB-specific cash transfers depending on how their actual value compares to governmental poverty-reduction cash transfers. Our inputs were all associated with some uncertainty, especially TB-related costs, which were mostly extracted from small subnational cost surveys [33,34,62–66]. We attempted to account for this using a multiway analysis that allowed inputs to vary by their simulated sampling distributions. Inconsistent reporting of standard deviations for mean TB-related costs, household income, and cash transfers forced us to make assumptions about the amount of variance around extracted values and to generalise these across countries. Whilst we ensured that our estimates of variance were as accurate as possible by drawing from relevant literature [7,37,44], this approach will have ignored any country-specific skewness or kurtosis in input parameters. Inconsistent reporting of TB-related costs disaggregated by income quintile meant that we had to assume that estimated TB-related costs were representative of those incurred by affected households in countries’ poorest population quintile. Because poorer households usually incur lower TB-related costs compared to less poor households, this is likely to have underestimated the potential of cash transfers to prevent catastrophic costs, and overestimated the country-level budget needed to achieve this [14]. Whilst our analysis should provide an accurate estimate of the effect of cash transfers on countries’ mean TB-related cost burden, because sample distributions of TB-related costs are known to be positively skewed [7], the aggregate-level nature of our study means that our results are unlikely to be representative of the majority of TB-affected households. Two sources of error in this study were imputation of missing TB-related cost components in Brazil [33], Colombia [62], Tanzania [63], and Mexico [64], and measuring catastrophic costs using a threshold TB-related cost burden that still hasn’t been assessed to determine its clinical or financial relevance for TB-affected households in any of the countries included in the study. Sensitivity analysis showed that the potential of cash transfers to prevent catastrophic costs was robust to these sources of error, but the precise estimates of countries’ household-level additional and total cash transfer and country-level cash transfer budget needed to prevent catastrophic costs were dependent on them. Therefore, whilst a TB-specific approach was consistently more effective and affordable for preventing catastrophic costs compared to a TB-sensitive approach in all our analyses, further research is needed to precisely estimate the cost of each of these approaches.
Our study is consistent with, and adds to, individual level evidence supporting the potential of TB-specific cash transfers to prevent catastrophic costs for poor TB-affected households in Peru [15]. Whilst our study questions the ability of TB-sensitive cash transfers to prevent households from engaging in damaging coping strategies, it supports individual evidence from Latin America for their ability to improve households’ coping capacity in response to severe livelihood risks [21]. By focussing on preventing catastrophic costs in low- and middle-income countries, it adds another dimension to the 2015 Cochrane review of the use of cash transfers in TB control, which did not find evidence on this outcome and mostly examined the use of cash transfers in the United States [69]. For future research, the validity of our results should be tested using individual-level primary data from future TB-related cost surveys, secondary data that includes information on households’ TB exposure, income and social protection status [70], and/or experimental data from interventions like the ongoing CRESIPT trial in Peru [25,26]. This work should also look to explore the effect of TB-specific and TB-sensitive cash transfers on other proxy measures of catastrophic costs like household dissaving (e.g., taking out a loan and/or selling household items) [15,16]. For a more complete understanding of the impact of providing cash transfers with a TB-specific versus a TB-sensitive approach, future research should also aim to incorporate this study’s data into an epidemiological model that accounts for their respective effects on TB-related catastrophic costs, and the additional effects of a TB-sensitive approach on individuals’ risk of TB infection and progression by addressing poverty-related risk factors (e.g., poor living conditions and undernutrition) [71,72].
In addition to studying the potential of cash transfers to prevent TB-related catastrophic costs, future research should also prioritise investigating the effect of other forms of social protection on this outcome. For example, in Mexico [64], food assistance might effectively defray households’ high direct nonmedical food costs [73], and in Ghana [65], facilitating patients’ access to sickness benefits and their prompt reintegration into the labour market might help to avoid high indirect costs [74]. Obviously, social protection should not be implemented in isolation of other healthcare initiatives to reduce costs. Further research should also aim to evaluate the complementary effects of social protection and efforts to reduce out of pocket direct medical costs. For example, combining social protection with further investments to maximize ambulatory community-based care might be especially effective for preventing catastrophic costs in Ecuador, where patients incur high direct medical costs for hospitalisation [66,75,76]. Multidisciplinary research platforms like the Health and Social Protection Action Research & Knowledge Sharing (SPARKS) Network will be key for facilitating this sort of research [77].
Our analysis compares cash transfers provided with a TB-specific versus TB-sensitive approach. Interestingly, it has been proposed that one efficient and cost-effective strategy might be to integrate both TB-specific and TB-sensitive approaches into a so-called “TB-inclusive” approach [24]. Results from our study demonstrating the greater potential of TB-specific cash transfers to prevent catastrophic costs, and the existing coverage of TB-sensitive cash transfer programmes, may support this integrated approach. For a TB-inclusive approach, existing poverty-reduction programmes could be adapted to include an additional variable TB-specific benefit, which beneficiary households would be eligible to receive upon receipt of a confirmed TB diagnosis. To finance such an approach, stakeholders from across TB control, development, and finance sectors could coordinate to determine how much each would be willing to contribute given their separate objectives [78,79]. From a TB prevention perspective, such an investment would be expected to reduce delays for TB diagnosis [10], reduce the risk of adverse treatment outcomes [14], and thus potentially contribute to reduced national TB incidence [80]. From the perspective of social development, reduced national TB incidence would be expected to enable previously vulnerable households to invest more in human capital, increase their labour productivity, and thus contribute to long-term sustainable economic growth [81]. Because households affected by human immunodeficiency virus (HIV), mental health issues, diabetes, and other noncommunicable diseases are also known to incur high direct and indirect costs [82,83], any efforts to prevent catastrophic costs for TB-affected households should aim to collaborate with these other areas of public health. Whichever approach for providing cash transfers to prevent TB-related catastrophic costs is chosen, it will be key to ensure that it is not implemented in isolation from universal health coverage initiatives including more decentralised and patient-friendly TB service delivery. Social protection initiatives and universal health coverage initiatives should be developed and implemented hand-in-hand [19].
Our finding that neither a TB-specific nor a TB-sensitive approach might be sufficient to prevent DR TB catastrophic costs highlights the urgent need for considerable investments in social protection and universal health coverage initiatives targeted to households affected by this disease. Globally, only 20% of people with DR TB are estimated to begin treatment, and only 52% of those that start treatment are estimated to successfully complete it [1]. Therefore, DR TB-affected households should constitute a “special case” for future investments to prevent catastrophic costs.
Reviewing and analysing the literature on TB-related costs and poverty-reduction cash transfer programmes in low- and middle-income countries, our study compares the potential of cash transfers provided with a TB-specific versus a TB-sensitive approach to prevent catastrophic costs for poor TB-affected households. Our findings suggest that providing cash transfers with a TB-specific approach to defray TB-related costs of households with a confirmed TB diagnosis will be more effective and affordable for achieving this objective compared to a TB-sensitive approach that increases the income and strengthens the economic resilience of households at high risk of developing active TB disease. Our findings also highlight an urgent need for investments to prevent catastrophic costs for households having to confront the severe medical, social, and economic challenges caused by DR TB.
|
10.1371/journal.pntd.0005671 | Minimal genetic change in Vibrio cholerae in Mozambique over time: Multilocus variable number tandem repeat analysis and whole genome sequencing | Although cholera is a major public health concern in Mozambique, its transmission patterns remain unknown. We surveyed the genetic relatedness of 75 Vibrio cholerae isolates from patients at Manhiça District Hospital between 2002–2012 and 3 isolates from river using multilocus variable-number tandem-repeat analysis (MLVA) and whole genome sequencing (WGS). MLVA revealed 22 genotypes in two clonal complexes and four unrelated genotypes. WGS revealed i) the presence of recombination, ii) 67 isolates descended monophyletically from a single source connected to Wave 3 of the Seventh Pandemic, and iii) four clinical isolates lacking the cholera toxin gene. This Wave 3 strain persisted for at least eight years in either an environmental reservoir or circulating within the human population. Our data raises important questions related to where these isolates persist and how identical isolates can be collected years apart despite our understanding of high change rate of MLVA loci and the V. cholerae molecular clock.
| Cholera is a deadly disease caused by the bacterium Vibrio cholerae. The ancestral home of cholera is around the Bay of Bengal, but recently cholera has moved to Africa. In Africa, cholera occurs in sporadic outbreaks. In order prevent cases of cholera, we want to understand the transmission of cholera in Africa, does it stay in one place or does it move around. To gain insight into these questions, we have examined the DNA of the bacteria. The DNA provides a identity for each isolate and we can infer how the isolates are related to each other based on the number and type of DNA changes. In our study, we examined the DNA of cholera isolates from southern Mozambique. We were surprised how similar all the Mozambique isolates were even though that were collected up to eight years apart. Based on previous work, we would have expected much more change in the DNA. Our data raises several important questions that relate to where these cholera isolates persist, possibly in local refuges, and how seemingly identical isolates can be collected years apart despite our understanding of the high rate of change of the molecular clock.
| Cholera remains a public health concern in developing countries with an estimated burden of 1.2–4.3 million cases and 28,000–142,000 deaths per year, worldwide [1]. South Asia and sub-Saharan Africa account with the majority of cases and deaths. Between 01 January and 03 June 2013, a total of 25,762 cholera cases and 490 deaths were reported from 18 African countries. Mozambique accounted for 7% (1,861/25,762) of cases and 4% (19/490) of deaths, being the third most affected country after the Democratic Republic of the Congo and Angola [2]. The most recent cholera outbreak in Mozambique started in December 2014 in Nampula [3], where there were 8,835 cases with case fatality rate of 0.7% (65 deaths) in 5 northern and central provinces in five months [4,5]. The most recent cholera outbreak in the south was reported in 2011 [6]. The peak of a cholera epidemic is often preceded by increasing prevalence of the pathogenic strains in the environment [7] where V. cholerae are harbored in aquatic reservoirs during extended periods between outbreaks [8].
Vibrio cholerae O1 is associated with most epidemic and pandemic outbreaks [9]. Whole genome sequence (WGS) analysis of V. cholerae isolates from around the world [9–12] demonstrated that the current (Seventh) Pandemic is monophyletic and originated from a single source with a clonal expansion of the lineage, with a strong temporal signature [13]. The seventh pandemic has been divided into 3 waves beginning in 1952, 1981 and 1988, respectively [13]. Each wave appears to be a selective sweep, as was the second wave of V. cholerae that swept across Haiti after the initial introduction [14]. In Africa, isolates from all three waves have been identified [13] with wave 3 isolates forming two distinct clades in Kenya between 2005 to 2010 [15].
The epidemiology and transmission patterns of outbreaks of V. cholerae have been explored using both multilocus variable number tandem repeat analysis (MLVA) and WGS [13,16]. In rural Bangladesh, MLVA revealed that multiple genetic lineages of V. cholerae occur naturally in the environment with geographic and seasonal genetic variation and identical genotypes can be found in the environment and humans [12]. In Kenya, MLVA demonstrated that several distinct genetic lineages emerged simultaneously during outbreaks in a single cholera season, linked to local environmental reservoirs [17]. WGS revealed all lineages were part of the Seventh Pandemic expansion [15]. In central and western Africa, MLVA revealed distinct clusters of isolates from different countries: Democratic Republic of the Congo, Zambia, Togo, and Guinea [18]. A second study in Guinea demonstrated spread and differentiation of V. cholerae during an outbreak [19].
In Mozambique, previous analysis of V. cholerae O1 showed these isolates had the typical traits of the El Tor biotype overall except that they carried a tandem array of classical CTX prophage [20] located on the small chromosome [21,22]. However, there are no data available on the genetic relatedness of V. cholerae circulating in Manhiça in southern Mozambique. Here, we characterized clinical and environmental V. cholerae isolates from Mozambique using MLVA and WGS to determine the genetic relatedness of strains isolated from patients with diarrhea in Manhiça District Hospital.
Manhiça District Hospital (MDH) is a 110 bed referral health facility for Manhiça District, a rural area of Maputo Province in southern Mozambique. The characteristics of the area have been described in detail elsewhere [23,24]. Briefly, the climate is subtropical with two distinct seasons: a warm, rainy season between November and April, and a cool and dry season during the rest of the year. Manhiça has 160,000 inhabitants, who are mostly subsistence farmers or workers in two large sugar- and fruit-processing factories. The Manhiça Health Research Centre (Centro de Investigação em Saúde da Manhiça[CISM]) is adjacent to the MDH and has been conducting continuous demographic surveillance for vital events and migrations since 1996 [23], currently covering 165,000 individuals.
The strain collection of V. cholerae described in the study was isolated from cholera surveillance and other studies conducted in the Manhiça community by CISM approved by the National BioEthic Committee (CNBS). Any and all patient data were anonymized/de-identified. The IRB at University of Maryland School of Medicine approved the use of anonymized strains.
A total of 75 V. cholerae isolates were collected from MDH, between 2002 and 2012. The strains were isolated from stool of patients admitted to the MDH with suspicion of cholera, presenting with watery diarrhea. In addition, three isolates collected from the Incomati River were included. V. cholerae isolates were identified by standard biochemical tests; and confirmed by API-20E biochemical test strips (bioMérieux SA, Marcy-l'Etoile, France). Serotypes were determined using commercially available poly- and mono-clonal slide agglutination antisera (Mast Group Ltd., Merseyside, UK) according to the manufacturer’s instructions. All the isolates were stored at -80°C in tryptone soya broth (TSB) with 15% glycerol, and retrieved at the time for molecular characterization.
A pure culture of V. cholerae was plated in Thiosulfate Citrate Bile Sucrose (TCBS) agar and incubated overnight at 37°C. DNA was extracted using the Qiagen QIAamp DNA Mini Kit (Hilden, Germany). The DNA template was sent to the University of Maryland Baltimore, Baltimore, Maryland, USA for molecular typing by MLVA and WGS.
DNA from each isolate was amplified by PCR using the conditions and primers previously described for 5 loci containing variable length tandem repeats [11]. The amplified products were separated and detected using a model 3730xl Automatic Sequencer (ABI) and their sizes were determined using internal lane standards (Liz600; ABI, Foster City, CA) with the Gene Mapper v4.0 program (ABI). Genotypes were determined according to the published formulas to calculate the number of repeats from the length of each allele and identify the alleles at the 5 loci [11]. The 5 loci, in order, are VC0147, VC0436–7 (intergenic), VC1650, VCA0171, and VCA0283; thus, the genotype 9,4,6,19,11 indicates that the isolate has alleles of 9, 4, 6, 19, and 11 repeats at the 5 loci, respectively. Relatedness of the strains was assessed by eBURSTv3 (http://eburst.mlst.net), in which genetically related genotypes were defined as those possessing at least 4 identical alleles of the 5 loci. An alternative analysis was performed using Network 2.x (http://www.fluxus-engineering.com/sharefaq.htm).
The DNA concentration was quantified by NanoDrop 2000 Spectrophotometer (Thermofisher Scientific, Waltham, MA, USA) and only specimens with sufficient concentration (n = 71) were submitted to WGS. DNA was prepared for Illumina sequencing using the KAPA High Throughput Library Preparation Kit (KapaBiosystems, Wilmington, MA). DNA was fragmented with the Covaris E210. Libraries were prepared using a modified version of manufacturer’s with-bead protocol (KapaBiosystems, Wilmington, MA). The libraries were enriched and barcoded by ten cycles of PCR amplification step with primers containing an index sequence seven nucleotides in length. The libraries were sequenced on a 100 bp paired-end run on an Illumina HiSeq2500 (Illumina, San Diego, CA).
The quality of the 101-base paired-end reads was confirmed using Fastqc (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Kmergenie (http://kmergenie.bx.psu.edu/) was used for choice of the best peak and the assembly was performed using the SPAdes software [25]. CSI Phylogeny 1.0a (http://cge.cbs.dtu.dk/services/CSIPhylogeny/) was used to generate a tree of genetic relatedness based on high quality nucleotide variants and then compared to V. cholerae O1 El Tor reference strain N16961 (NCBI accession numbers AE003852 and AE003853). Splitstree [26] was used to determine networks. Previous work in our lab demonstrated that this pipeline produces identical results to SMALT [27].
A high resolution SNP based phylogeny for the 67 7th pandemic strains was placed in context of a globally representative collection of 274 isolates (S1 Table) by mapping the reads to the V. cholerae 01 El Tor reference N16961 using SMALT (http://www.sanger.ac.uk/resources/software/smalt) as previously described [28]. Gubbins [29] was used to simultaneously remove regions of high SNP density and putative recombination sites in the alignment and infer the phylogenetic tree. The pre-seventh pandemic strain M66 (NCBI accession numbers CP001233 and CP001234) was used to root the phylogenetic tree.
To accurately place the four non-01 V. cholerae into a phylogenetic context we calculated a core genome alignment of 1093 genes (1,055,747 bp) using Roary [30] from a set of diverse V. cholerae genomes along with genomes of the closely related species Vibrio metoecus and Vibrio parilis. The resulting alignment was used to reconstruct the phylogenetic relationship using RAxML v. 7.8.6 [31] under the GTR model with 100 bootstrap replicates. The resulting phylogenetic trees were visualized using FigTree v.1.4.2 (http://tree.bio.ed.ac.uk/software/figtree/).
Among the 78 isolates (Table 1), 26 distinct genotypes were identified by MLVA using all five VNTR loci. The numbers of distinct alleles at loci VC0147, VC0437, VC1650, VCA0171, and VCA0283 were 10, 2, 4, 7 and 7, respectively.
Two clonal complexes of genetically related genotypes (each complex comprising genotypes that differed by an allelic change at a single locus) and four singleton genotypes unrelated to any other (differed by ≥2 loci) were identified (Fig 1). The first clonal complex (CC1) consisted of twenty of the twenty-six genotypes and comprised 91% (71/78) of the isolates (including all 3 isolates from river water). In CC1 the most common genotype, identified as the “founder” genotype (defined by eBURST as the genotype that differed from the largest number of other genotypes at a single locus), comprised 41% (29/71) of the isolates. In this complex, the founder genotype radiated into 9 other genotypes, and 5 of those differentiated further. Two of the three river water isolates shared a common genotype with the clinical isolates, one of which was of the founder genotype. Interestingly we found isolates with identical MLVA genotypes up to 8 years apart. In general, isolates from the same year were genetically related, all four genotypes found in 2002 were single locus variants, as were the two genotypes in 2008, two in 2009, and three genotypes from 2010. In contrast to similarity within a year, in months (May 2002, March 2003, April 2003, May 2003, June 2010) where multiple cases presented with V. cholerae, we found that isolates belonged to more than one genotype.
The other V. cholerae isolates consisted of a second clonal complex CC2 and four singletons (Fig 1). CC2 had two genotypes and comprised only 4% (3/78) of the analyzed isolates. The four singleton genotypes contained one isolate each corresponding to 5% (4/78) of the isolates. Three of them were isolated in 2008 and the fourth in 2012.
We successful sequenced the whole genomes of 71 V. cholerae O1 isolates (Table 1). The average number of high quality reads was 15,375,291 which upon assembly produced an average of 143 contigs (range: 99 to 443) or scaffolds with an average depth of 374. The assembled genome was 4.09 Mb (3.95 to 4.23 Mb) in length and had an average N50 of 232,090 bp (144,507 to 339,296 bp).
Of the 71 isolates, the genomes of sixty-seven differed by less than 100 SNVs. Consistent with this, the core genome of all the wave 3 isolates shared 2802 genes. To look for evidence of recombination in these sequences we examined 2543 single copy conserved genes. We removed all of the invariant nucleotides, any SNVs in locally collinear blocks smaller than 200 bp, and examined the remaining variant sites in Clustal W (Fig 2). These data showed evidence of homoplasy by visual examination: for example there were multiple examples of dinucleotide sequences at the same site in different isolate genomes being present in all four possible combinations (e.g. AA, AG, GA, GG). This was despite the high level of pairwise nucleotide conservation between isolate genomes, usually differing by only 18 to 62 nucleotides. Among the 148 SNVs, 360 pairs were determined to have both alleles in all possible combinations. The probability of this occurring by mutation alone is vanishingly small (the probability of the same mutation on two different genomes to the 360th power). The most parsimonious explanation for this observation is recombination within this population of V. cholerae O1s.
The genetic relatedness, including the recombinant loci, was estimated from WGS using a phylogenetic network [26]. As shown in Fig 3, while all of the genomes were distinct, there was some clustering by year of isolation, four of the five isolated in 2010 clustered together as did the four from 2008 and the two from 2009. The largest number of isolates in our collection were collected in 2003. It is evident from Fig 3 that they occupy positions throughout the network. This was also true for the three river isolates. In the network, every sequence had some nucleotide variants that distinguish it from every other sequence, thus in contrast to the MLVA network, there are no central genotypes that might be construed as the founder.
Fig 4 shows the phylogenetic tree with genomic sequences from a global collection of 274 previously published sequences (S1 Table) including the 67 genomic sequences generated in this study. These data show that the sequences of the Manhiça, Mozambique isolates cluster in a monophyletic group distinct by 39 nucleotides from the backbone of the third wave of the Seventh Pandemic phylogeny (S2 Table). All Seventh Pandemic isolates harbor virulence associated genes, such as the CTX prophage, the genomic islands VSP-I and VSP-II, the toxin-coregulated pilus, the toxic linked cryptic element, and the integrative conjugative element SXT harboring multidrug resistance genes. The CTX phage harbours the ctxBcla allele within an otherwise El Tor biotype CTX phage sequence and is typical of isolates referred to as “atypical” El Tor biotypes [32].
The four sequences that differed by >24,000 nucleotides among the 3,237,973 basepairs conserved in all our sequenced genomes did not cluster with sequences from the Seventh Pandemic (Fig 5) and are quite distantly related to each other, but all were more closely related to the other V. cholerae sequences compared to other species. These isolates were taken from patients admitted to hospital with suspected cholera indicating that these divergent lineages are capable of causing clinical symptoms, as has been reported previously [13]. Unlike the Seventh Pandemic strains, these four strains do not contain any of the aforementioned virulence associated genes, although they contain a few genes, such as the RTX toxin gene cluster and the hemolysin hlyA, from some of the virulence associated islands.
The network based on WGS included isolates of both CCs (CC1 and CC2), however while the isolates of CC1 was distributed through the network, the isolates of CC2 clustered together (Fig 3). In the same way, the four singletons isolates by MLVA did not cluster in the network and were quite distantly related by WGS analysis.
Cholera remains important public health problem in Mozambique. We characterized V. cholerae isolates using MLVA and WGS to determine the genetic relatedness and transmission dynamics of cholera outbreaks in the Manhiça District. Our analyses of WGS data revealed that 94% of the isolates were a monophyletic group in the third wave of the Seventh Pandemic. These isolates have formed a locally evolving population that has persisted for at least eight years, either in a local environmental reservoir or circulating within the human population, and sporadically caused recurrent disease in southern Mozambique. It is yet to be shown if these isolates are representative of those circulating outside of the study site across Mozambique.
Previous studies have reported the role of environmental factors, such as seasonal fluctuations, that influence the dynamics of V. cholerae in environmental reservoirs [33,34]. Of note, all of our isolates were collected during the first half of the year, January through July, with a peak in May. However, our study does not allow establishing a clear relationship between environmental strains and those causing cholera outbreak. The presence of four isolates that are not part of the Seventh Pandemic (differing by ~ 0.75% of the genome sequence) demonstrates that a diverse set of V. cholerae not linked to the Seventh Pandemic are causing a background, low level of sporadic disease in southern Mozambique, despite the absence of many of the major virulence determinants. This has been seen elsewhere including in the Gulf coast of the USA particularly in the 1980’s and 1990’s [32].
We detected the presence of recombination in our WGS data. Most analyses of WGS data detect recombination as a large number of SNVs in short sequence of DNA. Our method of detection, the four-gamete test applied to haploid genomes, removes the constraint of the SNVs occurring in a small region by simply looking at dinucleotide pairs located at any distance from each other. Since mutations occur at random, a dinucleotide sequence, say GG, can mutate at either of two positions for example: AG and GA. In order to further mutate to form AA, one of the positions must mutate a second time. However, in this instance it also possible to replace the original sequence in a single recombination event. We found 360 pairs of nucleotide positions across all our genomes at which all four combinations of alleles in the dinucleotides were found. The probability of this occurring by mutation is extremely small (the rate of mutation at the same nucleotide to the 360th power). Our sample is unusual because it was collected in small area and in a short time frame. The geographical and temporal proximity of isolates is a prerequisite for recombination. V. cholerae meets other necessary prerequisites like having an intact mechanism for uptake of DNA and integration into the chromosome [38]. The clearest examples of the effectiveness of recombination can be found in serotype switching, a process known to be accelerated by chitin [39]. The limited amount of variation in our sample is consistent with the recombinant events occurring within this population of V. cholerae O1s.
Most isolates (91%) were distributed among the 20 genotypes of CC1 and 41% had the founder genotype, supporting the hypothesis of a common ancestor which subsequently differentiates into additional genotypes. In a study of 187 isolates conducted in Haiti, only 9 MLVA genotypes clustered in a single clonal complex and 53% had the founder genotype [35].
We found more alleles (ten) at the first locus (VC0147) on large chromosome than at the two loci on small chromosome (VCA0171 and VCA0283) with seven alleles for each one. Our findings are in contrast to previous descriptions that show the three loci on the large chromosome varying at a slower rate than those on the small chromosome [11]. Extensive genetic variation on both chromosomes has been reported [36], but not more variation at the large chromosome loci [35–37].
WGS and MLVA patterns were performed to discriminate V. cholerae isolates from various geographic locations and distinct populations [17,40]. But, none of the 26 MLVA genotypes that we found in this study has been reported in previous studies from Haiti, Thailand, Bangladesh, India, Vietnam and Mozambique [10,11,35,36,40,41]. However, the three loci on large chromosome are likely to be considered the best for estimating across large distances [35]. Thus, when we considered the MLVA profiles in terms of the three loci (VC0147, VC0437, and VC1650) on large chromosome, the isolates with profiles 8,4,6, (50%, 39/78) and 9,4,6, (10%, 8/78) were related to the Haiti and Bangladesh strains, respectively [11,35]. Our WGS analysis demonstrates that these Mozambican isolates formed a distinct lineage within a clade that includes El Tor variants from Bangladesh, China, Haiti, Nepal, India, and Kenya that belong to the current global radiation of the Seventh Pandemic. The Manhiça, Mozambique isolates differ from the wave 3 backbone by 39 nucleotides. Furthermore, previous reports demonstrated Wave 2 strains in Mozambique during the same time period [21,42]. Taken together with our analysis, it is clear that there were multiple, independent V. cholerae lineages from Wave 2 and Wave 3 which were circulating within Mozambique during this period of time. In Manhiça, although the relationships between isolates differed in detail between MLVA and WGS, both analyses demonstrated the isolates were very closely genetically related, even if they are collected several years apart.
Our findings further define the molecular epidemiology of V. cholerae in Mozambique. Our study demonstrates that Wave 3 isolates of the Seventh Pandemic have become established in Manhiça, Mozambique and have persisted in this region over the time of this study alongside V. cholerae Wave 2 strains in other regions of the country. The subsequent radiation of genotypes has been enriched by the process of recombination as detected in the WGS data. Our data raises several important questions that relate to where these V. cholerae isolates persist and how seemingly identical isolates can be collected years apart despite our understanding of the high rate of change of the MLVA loci and the V. cholerae molecular clock. Although the environmental triggers for the emergence of cholera are unknown in Manhiça, it is important to be vigilant to prevent an emergence from becoming an outbreak.
|
10.1371/journal.ppat.1005090 | AAV-Delivered Antibody Mediates Significant Protective Effects against SIVmac239 Challenge in the Absence of Neutralizing Activity | Long-term delivery of potent broadly-neutralizing antibodies is a promising approach for the prevention of HIV-1 infection. We used AAV vector intramuscularly to deliver anti-SIV monoclonal antibodies (mAbs) in IgG1 form to rhesus monkeys. Persisting levels of delivered mAb as high as 270 μg/ml were achieved. However, host antibody responses to the delivered antibody were observed in 9 of the 12 monkeys and these appeared to limit the concentration of delivered antibody that could be achieved. This is reflected in the wide range of delivered mAb concentrations that were achieved: 1–270 μg/ml. Following repeated, marginal dose, intravenous challenge with the difficult-to-neutralize SIVmac239, the six monkeys in the AAV-5L7 IgG1 mAb group showed clear protective effects despite the absence of detectable neutralizing activity against the challenge virus. The protective effects included: lowering of viral load at peak height; lowering of viral load at set point; delay in the time to peak viral load from the time of the infectious virus exposure. All of these effects were statistically significant. In addition, the monkey with the highest level of delivered 5L7 mAb completely resisted six successive SIVmac239 i.v. challenges, including a final challenge with a dose of 10 i.v. infectious units. Our results demonstrate the continued promise of this approach for the prevention of HIV-1 infection in people. However, the problem of anti-antibody responses will need to be understood and overcome for the promise of this approach to be effectively realized.
| AAV-mediated antibody delivery represents a promising alternative to classical vaccine approaches for the prevention of HIV/AIDS in humans. We used recombinant adeno-associated viral vectors (rAAV) as gene carrier to deliver anti-SIV (simian immunodeficiency virus) antibodies to monkeys. This non-classical immunization approach is independent of the host immune system and the AAV-transferred gene into muscle cells produces antibodies with known specificities. Our goal was to determine what serum levels of delivered antibodies can be achieved by using two different vector strategies, whether the delivered antibodies provoke an immune response, and to what degree the delivered non-neutralizing antibodies have the ability to protect against SIV virus challenge. This approach allowed for sustained levels of delivered antibodies in serum and exhibited significant protective effects against highly pathogenic SIVmac239 challenge. However, immune responses limited the concentration of circulating antibodies that could be achieved. This problem will need to be understood and overcome for the promise of this approach to be effectively realized.
| There are good reasons for believing that development of an effective preventive vaccine against HIV-1 is going to be a very difficult task. HIV-1 has evolved a variety of immune evasion strategies that allow continuous virus replication in the face of apparently strong host immune responses, both cellular and humoral [1,2]. For a vaccine to be effective it will likely need to either completely block the initial infection or to provide immune responses that are more effective than those resulting from infection.
One promising, creative approach is to use adeno-associated virus (AAV) vector to deliver antibodies with potent and broadly neutralizing activity [3–7]. This method is an alternative to classical immunization and independent of the host immune system. Proteins delivered by AAV vector can persist at stable levels for years [8,9], and AAV vectors have proven to be safe and effective in human gene therapy trials against a number of diseases [10–13]. More than a dozen potent broadly neutralizing anti-HIV-1 monoclonal antibodies have already been characterized; they are human-derived, they can effectively neutralize more than 90% of circulating HIV-1 strains, and they represent a range of different specificities [14–21].
In a previous pioneering study performed in monkeys, anti-SIV immunoadhesins (antibody-like molecules) were delivered by AAV vector and a reasonable percentage of the monkeys exhibited apparent sterilizing immunity against intravenous (i.v.) challenge by SIVmac316 [4]. However, SIVmac316 is an easy-to-neutralize SIV strain, one third of the immunized monkeys developed antibody responses (anti-anti) to the delivered immunoadhesin, and monkeys with anti-anti responses were not protected against the challenge. It was not clear from that initial study to what extent anti-anti responses may have resulted from delivery of immunoadhesins, which do not represent an authentic IgG structure. We decided to build upon these studies by using AAV to deliver authentic IgG and to examine protection against the difficult-to-neutralize strain SIVmac239. At the time these studies were initiated, there were no monoclonal antibodies available that could neutralize SIVmac239. We thus converted the previously reported immunoadhesins 4L6 and 5L7 [4] into authentic IgG versions. These anti-SIV mAbs in IgG1 form contain only authentic rhesus IgG sequences. In this study we describe the AAV-mediated delivery of these full-length immunoglobulins to rhesus macaques and the protective effects of the non-neutralizing antibody 5L7 IgG1 against SIVmac239 challenge.
Two fundamentally different types of AAV systems are available for vector-mediated delivery: single-stranded (ssAAV) vs. self-complementary (scAAV) [22]. scAAV has been reported to achieve considerably higher levels of transgene expression [23,24] but suffers from the drawback of not being able to accommodate sufficient genetic information for the expression of both heavy (H) and light (L) chains needed for authentic IgG [25]. We thus developed two strategies for AAV-mediated synthesis of authentic IgG for comparison purposes: ssAAV for the synthesis of matched H and L chains from a single vector vs. scAAV for the synthesis of H and L chains from separate vectors, an approach that we will call the two vector approach. The logic behind the two vector approach is that such a high quantity of AAV vector particles is inoculated into a highly localized area of muscle that it seems likely that many, perhaps even a majority, of cells will take up multiple AAV particles. We sought to deliver authentic IgG versions of the previously described monoclonal Fabs 346-16h and 347-23h [26] in order to allow comparison to the results of Johnson et al. that utilized immunoadhesins with these binding specificities (called 4L6 and 5L7, respectively) [4]. Since SIVmac239 was planned for challenge, this also allowed investigation of protective effects in the absence of virus-neutralizing activity. We employed IgG1 sequences in order to maximize antibody-dependent cellular cytotoxicity (ADCC) activity and other effector functions. We used AAV serotype 1 for intramuscular expression. The constructs shown in (S1 Fig) were demonstrated to express authentic rhesus IgG with expected properties prior to the initiation of monkey experiments.
In our first monkey study, three AAV1-negative monkeys were inoculated intramuscularly with 1.6 x 1013 ssAAV vector particles expressing both heavy and light chains of the authentic rhesus monkey IgG1 monoclonal antibody 5L7 (ssAAV-5L7 IgG1). On the same date, three AAV1-negative monkeys were inoculated with 0.8 x 1013 scAAV vector particles expressing 5L7 IgG1 heavy chain and 0.8 x 1013 scAAV vector particles expressing 5L7 kappa light chain (scAAV-5L7 IgG1 H + scAAV-5L7 kappa L). The levels of delivered 5L7 mAb in plasma were measured by a gp140 ELISA and purified mAb as standard. Considerable monkey-to-monkey variation was observed in the levels of 5L7 IgG1 that were achieved (Fig 1A). However, the levels in each individual monkey remained stable over the course of months leading up to the time of challenge. There were no clear differences in the persisting levels of 5L7 IgG1 mAb associated with the one vector vs. two vector delivery systems. While monkey 153–10 in the one vector group achieved persisting levels of 5L7 IgG1 mAb of only 1–5 μg/ml, monkey 84–05, also in the one vector group, achieved remarkable levels of approximately 270 μg/ml that have persisted now for more than two years. Three of the six monkeys in this first study had clear antibody responses to the delivered 5L7 IgG1 mAb (Fig 1C and S2 Fig). Monkey 153–10 with the lowest level of persisting 5L7 mAb had the strongest anti-anti response, while monkey 84–05 with the highest level of persisting 5L7 mAb had no detectable anti-anti response. The other two monkeys without detectable anti-anti responses (111–09 and 154–10) were in the two vector scAAV group.
In our second monkey study, we used the one vector ssAAV approach to deliver the 4L6 mAb in IgG1 form to six AAV1-negative monkeys. In this experiment, the levels of delivered 4L6 mAb peaked 3–4 weeks after administration and then fell precipitously (Fig 1B). All six of the recombinant AAV-4L6 recipients had antibody responses to the delivered 4L6 mAb whose appearance coincided with the precipitous decline in 4L6 mAb levels (Fig 1D and S3 Fig). The anti-anti responses in the 5L7 IgG1 recipients (Fig 1C) were largely specific for 5L7 while the anti-anti responses in the 4L6 IgG1 recipients (Fig 1D) were largely specific for 4L6; since the 5L7 and 4L6 IgG1s have the exact same constant regions, these results indicate that the responses were mostly directed to the variable domains (Fig 1E and 1F). By week 14, the time of SIV challenge, the levels of 4L6 mAb in two of the recipients (279–10 and 283–10) had dropped to undetectable levels, while the levels in the other four monkeys had dropped to 1–38 μg/ml (Fig 1B).
Six appropriately-matched control rhesus monkeys were enrolled in the challenge phase (S1 Table). Repeated, marginal dose, intravenous challenges were simultaneously initiated to control monkeys, to the 5L7 mAb group 44 weeks after their administration of recombinant AAV-5L7 IgG1, and to the 4L6 mAb group 14 weeks after their administration of recombinant AAV-4L6 IgG1. Administrations were performed every three weeks such that administrations were stopped when a monkey became infected from the previous exposure. A PBMC-grown stock of cloned SIVmac239 was used, one that has been carefully titered previously by the i.v. route in monkeys and has been used extensively by numerous investigators for controlled dose challenges [27–35]. One i.v. animal infectious dose was used through the first four administrations, 2 i.v. infectious doses for the fifth administration, and 10 i.v. infectious doses for the sixth administration.
Only one monkey of the 18 resisted all six challenges: monkey 84–05 with the highest levels of delivered mAb (averaging approximately 270 μg/ml of the 5L7 mAb through all of the challenges) (Fig 2A). Dozens of control monkeys have been challenged with the 10x dose of this exact same stock in numerous previous studies and not a single monkey has been refractory to infection [27–35]. The absence of anti-gp41 antibody responses in 84–05 even following the 10x challenge is consistent with the apparent sterilizing immunity in this animal (S4 Fig). Otherwise, no significant difference in the acquisition of SIV infection was observed among the test vs. control monkeys (Fig 2A).
Given that levels of delivered 4L6 IgG1 mAb were driven to low or undetectable levels by the time of first challenge in most monkeys in this group, it is not surprising that plasma virion SIV RNA levels in the 4L6 mAb group as a function of weeks following the infectious exposure were not significantly different from control monkeys, neither at peak viremia nor during chronic phase infection (Fig 2B and 2C; Fig 2E). However, SIV plasma RNA loads in the 5L7 mAb group were significantly lower than those in the control group both at peak viremia (Δ of 0.67 logs) and during chronic phase infection (Δ of 2.06 logs) (Fig 2B and 2C; Fig 2F). The differences were significant whether or not the undetectable viral loads of monkey 84–05 were included in the analyses. The viral loads in the 5L7 mAb group at set point were also significantly lower than those in the 4L6 mAb group (Fig 2C). The duration in weeks from the time of the infectious exposure to peak plasma viral loads was also significantly different between the 5L7 mAb group and the control group (Fig 2D). While there is a suggestion that viral loads at peak height were inversely correlated with the levels of delivered 5L7 mAb, this correlation did not reach statistical significance (S5 Fig). There was no obvious association of measured 4L6 IgG1 levels in serum at the time of infectious exposure with the time to peak viremia (S6 Fig).
Both the 5L7 and the 4L6 mAb IgG1s lack detectable neutralizing activity against the challenge strain SIVmac239 (Fig 3A). In fact, even at 1 mg/ml of the 5L7 mAb IgG1, 50% neutralization of SIVmac239 was not achieved (Fig 3B). However, both the 5L7 and 4L6 mAbs had neutralizing activity against the neutralization-sensitive derivative of SIVmac239 called SIVmac316 (Fig 3C). The half-maximal inhibitory concentrations (IC50s) against SIVmac316 were 0.0015 μg/ml (4L6 IgG1) and 0.003 μg/ml (5L7 IgG1), which are consistent with previously described IC50s for Fab molecule and immunoadhesin versions of these mAbs [4,26]. The concentrations of mAb in sera shown in (Fig 1A) were measured biochemically with a gp140 ELISA using varying amounts of purified mAb produced by transfection of 293T cells in culture as standards. In order to validate the biochemically measured mAb concentrations and to confirm the expected neutralizing activity in serum against SIVmac316, serum samples were diluted to contain a starting concentration of 4 μg/ml 5L7 mAb based on the biochemically-measured concentration and neutralizing activity against SIVmac316 was measured (Fig 3D). The neutralizing activity in serum agreed spot-on with the biochemically-measured concentration of the 5L7 mAb for all six samples. As expected, sera from 84–05 lacked detectable neutralizing activity against SIVmac239 (Fig 3E). HEK 293T-expressed purified immunoadhesins and mAbs showed equivalent binding to SIVmac239 gp140 as determined by ELISA (Fig 3F).
With this information in hand, we proceeded to measure ADCC activity in pre-challenge serum samples relative to the biochemically measured 5L7 mAb concentrations in them. We used an assay that measures natural killer (NK) cell-mediated ADCC activity of mAb or serum independent of neutralization or complement activity [36]. The assay uses a NK cell line that expresses macaque CD16 as effector cell and a CD4+ T cell line that expresses luciferase upon SIV infection as target. 5L7 IgG1 purified from 293T transfection was diluted to match the starting concentration of 5L7 IgG1 in the serum sample to which it was being compared. Serum from 84–05 was unusual in that it had markedly higher ADCC activity than the equivalent concentration of purified 5L7 IgG1 (Fig 4A and 4B). Pre-AAV serum from 84–05 contained no measurable activity capable of enhancing the ADCC activity of purified 5L7 IgG1 (S7 Fig). Four of the five remaining animals in the group had ADCC activity in their serum samples comparable to that of purified 5L7 IgG1 at equivalent concentrations (Fig 4A and 4B; S8 Fig). Serum from monkey 111–09, the monkey with the lowest post-challenge viremia, also had somewhat higher ADCC activity compared to purified 5L7 IgG1 at equivalent concentrations (S8 Fig).
Given the current availability of a spectacular array of human mAbs with potent broadly-neutralizing activity against HIV-1, it has become possible to envision prevention scenarios in which cocktails of such antibodies could be used to provide a sterilizing barrier to infection against the vast majority of HIV-1 strains circulating in the population. However, it is not practically feasible to consider passive delivery of purified mAbs over a period of years for prevention purposes. AAV vectors are ideally suited for this purpose. The only protein expressed from AAV vectors is derived from the inserted transgene put into it. As long as the protein is viewed as self, protein expression can continue for prolonged periods. In fact, there are already examples where proteins delivered by AAV vector have persisted at stable levels for years [4,8,9,37], including animal 84–05 in our study. Since muscle cells exhibit little or no turnover, they are a preferred source for long-term delivery purposes. Little or no integration of AAV vector DNA into host genome sequences has been observed [38–40] and AAV vectors have proven safe when used in clinical settings to reverse hereditary absence of gene function [10–13].
The molecularly cloned SIVmac239 is a difficult-to-neutralize strain with many of the characteristics of a primary isolate. It has been extensively used as a challenge strain to gauge efficacy of vaccine approaches in monkeys, and it has generally proven quite difficult to achieve protection against [29,30,41–43]. Being a molecular clone, it avoids complications of interpretation that may arise when a heterogeneous mixture of sequences in an uncloned virus stock are used and when differences in neutralization sensitivity/resistance may exist within the mixture [44,45]. The presence of AAV-delivered 5L7 mAb in our experiments significantly delayed the time to peak viral loads and significantly lowered SIVmac239 viral loads both at peak viremia and at viral load set point. The lowering of viral load set points in the 5L7 mAb group was also significantly different from the 4L6 mAb group, lending additional support to the significance of the protective effects. It is also likely that the lack of SIVmac239 acquisition in monkey 84–05 is also a real effect of the high concentration of 5L7 mAb in this animal. The 10x (6th) challenge of this monkey utilized the same dose of the same stock that has been used previously in dozens of control monkeys without a failure of infection acquisition [27–35]. The protective effects in the 5L7 mAb group occurred in the absence of any detectable neutralizing activity of the 5L7 mAb against the SIVmac239 molecular clone even out to 1 mg/ml. To our knowledge, this is the first demonstration of significant protective effects of a mAb in the absence of an ability to neutralize the challenge virus [46,47].
Recent studies have emphasized the importance of Fc-mediated effector functions in contributing to the protection afforded by mAbs capable of neutralizing the SHIV used for challenge [48–52]. These studies modulated the Fc-mediated effector functions of a neutralizing mAb up or down in order to examine the effects on protective capacity. Some vaccine studies have also found associations of ADCC activity with protective efficacy [53–57]. Effector functions potentially contributing to the protective effects of the 5L7 IgG1 observed in our current study include but are not limited to ADCC activity and complement-mediated virolysis. Along these lines, it is fascinating that the 5L7 mAb circulating in monkey 84–05 was not only present at extraordinarily high levels, it had extraordinarily high ADCC activity on a per μg basis. Because of the nature of the ADCC assay that was used, it seems likely that the extraordinarily high ADCC activity in the serum of this one animal results from differences in post-translational modifications of the 5L7 mAb. Decreased core fucosylation or terminal sialylation have been shown to dramatically increase Fc-mediated ADCC activity [58–62].
One question addressed in our current studies was whether the anti-anti responses to the immunoadhesin constructs observed in Johnson et al. [4] could be avoided by delivery of authentic IgG. The answer is a clear no. Immune responses to the delivered IgG1 antibody described here were readily detected in 9 of the 12 test monkeys and these appeared to limit the levels of delivered antibody that could be achieved. This is especially true of the 4L6 group in which the levels of delivered 4L6 IgG1 mAb dropped precipitously coincident with the appearance of the anti-anti responses. In separate studies, we have observed strong antibody responses in rhesus monkeys to AAV-delivered, rhesusized anti-HIV-1 mAbs in 24 of 24 opportunities using a ssAAV1 vector design identical to the one used for the studies described here. This contrasts strikingly with studies in humans in which 13 sequential administrations of 1g of anti-HIV mAb failed to elicit anti-anti responses to the three different antibodies that were used in each of the 12 test subjects [63,64]. With our 5L7 and 4L6 experiments, it is possible that a lack of natural pairing of IgG heavy and light chains contributed to their immunogenicity. With our anti-HIV experiments referred to above, it is possible that the failure to rhesusize the variable regions contributed to their immunogenicity. Nonetheless, our results serve as a warning signal for the potential of immunogenicity problems when AAV is used to deliver anti-HIV mAbs in human trials.
Coding sequences (heavy chain, light chain or bicistronic) were designed in silico, codon-optimized and gene-synthesized (Genscript). 4L6 and 5L7 immunoadhesin sequences[4] served as a template and full-length antibodies were constructed by adding CH1 domain and CL domain of rhesus IgG to the already known immunoadhesin sequences. 4L6 and 5L7 sequences originate from recombinant anti-SIV Fab sequences derived from the bone marrow of SIV-infected rhesus monkeys[26]. Amino acid sequence alignments of 4L6 and 5L7 variable regions to the most closely corresponding rhesus genomic regions are depicted in (S9 Fig). Rhesus IgG1 sequence is based on accession no. AAF14058 and AAQ57555, and rhesus IgG2 sequence is based on AAF14060 and AAQ57567. Rhesus kappa light chain was designed using CL domain sequence from AAD02577. Synthesized fragments were then cloned into NotI site of scAAV or ssAAV vector plasmids[4]. Both 4L6 and 5L7 recognize gp120 and gp140 forms of the envelope glycoprotein.
HEK293T cells (ATCC) were transfected with rAAV vector plasmids and proteins were purified from cell culture supernatant using Protein A Plus (Pierce). Concentration of purified proteins was determined by NanoDrop (Thermo Scientific) A280 measurement and purity and integrity was verified by Coomassie staining (Life Technologies).
Immunoadhesins and mAbs were tested for their ability to bind SIVmac239 envelope glycoprotein by ELISA. Test plates were coated with recombinant gp140 of SIVmac239 (Immune Technology) for 1 h at 37°C. Plates were washed using PBS-Tween20 (Sigma-Aldrich) and subsequently blocked with 5% nonfat dry milk in PBS (Bio-Rad). Immunoadhesins, mAbs, positive and negative sera from macaques were serially diluted 1:3 in blocking buffer and added to the test plate. After 1 h of incubation at 37°C the plates were washed again and a HRP-conjugated goat anti-rhesus IgG H+L (SouthernBiotech) was then added for detection. The reaction was stopped after 1 h at 37°C and plates were washed 10 times. Subsequently, TMB substrate and stop solutions (SouthernBiotech) were added and Absorbance at 450 nm was measured in a microplate reader (PerkinElmer).
Neutralization activity of mAbs against tested SIV strains was measured by either a luciferase-based assay as described previously (TZMbl)[36] or an assay that is based on secreted alkaline phosphatase (SEAP)[36]. Virus and a serial dilution of Abs or serum were incubated for 1 h at 37°C before adding it to the TZMbl or SEAP cells (ATCC). Neutralization was assessed two days later by measuring luminescence in a 96-well plate reader (Perkin Elmer).
Measurement of antibody effector function was performed by NK cell activity towards virus-infected target cells expressing luciferase[36]. A CD4+, CCR5+ T-cell line was infected with SIV four days prior assay readout. These target cells were washed three times on the day of the assay and co-cultured with a NK cell line expressing rhesus CD16 and a serial dilution of mAb or heat-inactivated serum for 8 to 10 hours. Cells were lysed afterwards and ADCC was measured as the loss of luciferase activity in the cell culture supernatant.
Production of rAAV was conducted as described previously[65]. In short, HEK 293 cells were transfected with rAAV vector plasmid and two helper plasmids to allow generation of infectious AAV particles. After harvesting transfected cells and cell culture supernatant, rAAV was purified by three sequential CsCl centrifugation steps. Vector genome number was assessed by Real-Time PCR, and the purity of the preparation was verified by electron microscopy and silver-stained SDS-PAGE.
The animals in our study were rhesus macaques of Indian-origin (Macaca mulatta). We purchased the 18 monkeys and housed them at the New England Primate Research Center of Harvard Medical School in a biocontainment facility in accordance with standards of the Association for Assessment and Accreditation of Laboratory Animal Care and the Harvard Medical School Animal Care and Use Committee. The animal samples used here were collected under experimental protocols approved by the Harvard Medical Area Standing Committee on Animals, and conducted in accordance to the Guide for the Care and Use of Laboratory Animals. All macaques were tested negative for the presence of antibodies to SIV and AAV1 capsid prior AAV administration. The weights of the animals ranged from 2.5 kg to 17.2 kg at the time of immunization. The six 5L7 IgG1 animals were split in two groups and each animal received a total of 1.6 x 1013 AAV vector genomes. One group received 1.6 x 1013 particles of ssAAV expressing the heavy and light chains of 5L7 IgG1, the other group received 0.8 x 1013 particles of scAAV expressing the heavy chain of 5L7 IgG1 plus 0.8 x 1013 particles of the scAAV expressing the kappa light chain of 5L7 IgG1. AAV administration was conducted by four equal and deep intramuscular injections, where each animal received two separate 0.5 ml injections into both quadriceps. All 6 animals that received 4L6 IgG1 were given a total of 2.5 x 1013 AAV vector genomes per monkey.
The animal management program of Harvard Medical School is accredited by the American Association for the Accreditation of Laboratory Animal Care, and meets National Institutes of Health standards as set forth in the Guide for the Care and Use of Laboratory Animals (DHHS Publication No. (NIH)85-23 Revised 1985). The Institute also accepts as mandatory the PHS Policy on Humane Care and Use of Laboratory Animals by Awardee Institutions and NIH Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training. There is on file with the Office for Protection from Research Risks an approved Assurance of Compliance.
Animal facilities were administrated by the state and met Harvard standards for humane care and use of animals through a program of veterinary care, inspection and oversight. Animal care and welfare is the charge of the Committee on Animals, appointed by the Dean, consisting of 22 members comprising 4 veterinarians, 2 public representatives, and 16 doctoral level representatives of principal sites of animal use by Harvard Faculty. Additionally, local facilities are guided and monitored in daily activities by 8 departmental animal use committees. The procedure to avoid unnecessary discomfort, pain, or injury to animals are those prescribed in the aforementioned NIH “Guide” and additional detailed protocols for anesthesia, analgesia, tranquilization, euthanasia, or restraint have been developed and circulated by the Committee on Animals.
Rhesus macaques on study were housed in the biocontainment facility of the New England Primate Research Center of Harvard Medical School under approved protocol 04655 of the Institutional Animal Care and Use Committee (IACUC) of Harvard Medical School. The Harvard Office of Microbiological Safety oversaw the design and construction of this facility. It includes individual Hepa-filtered caging for monkeys and additional features such as a pass-through autoclave, footpedal-operated sink, and a special safety cabinet for surgical and necropsy procedures. The facility has restricted access. Individuals entering the room must suit-up with disposable cap, mask, coverall, gloves and shoe covers. Procedures used in the care and feeding of these animals are posted on the entry door. The animals were provided ad lib access to municipal source water, offered commercial monkey chow twice daily, and offered fresh produce a minimum of three times weekly. Light cycle was controlled at 12/12 hours daily. Containment facilities and procedures for macaques experimentally infected with RRV and SIV have been reviewed and approved by the Harvard Committee on Microbiological Safety.
Animals housed in the biocontainment facilities received a daily health check by both animal care technicians and veterinary professional staff. All animals received a complete physical examination on the average of once every four to six weeks. A comprehensive environmental enrichment and psychological well-being plan was in place for primates in the described studies and is available for inspection by the United States Animal and Plant Health Inspection Service (APHIS) and to officials of any pertinent organization. Euthanasia took place at defined experimental endpoints using protocols consistent with the American Veterinary Medical Association (AVMA) guidelines. Animals were first sedated with intramuscular ketamine hydrochloride at 20 mg/kg body followed by sodium pentobarbital (≥100 mg/kg) intravenously to achieve euthanasia.
To measure the concentration of 5L7 IgG1 and 4L6 IgG1 in vivo we performed a SIVmac239 gp140 (Immune Tech)/anti-rhesus IgG ELISA (Southern Biotech). Absorbance at 450 nm was compared to a serial dilution of purified mAb produced in HEK 293T cells, and the amount of antibody in serum was extrapolated based on the mAb standard curve. Reference protein was quantified by NanoDrop (Thermo Scientific) A280 measurement and purity was verified by Coomassie staining (Life Technologies). Levels of AAV-delivered antibody were measured up to the time of SIV challenge: 44 weeks (5L7 IgG1) and 14 weeks (4L6 IgG1) after AAV administration. The appearance of anti-env antibody responses following SIV infection obviated our ability to measure levels of delivered mAb at post-infection time points.
Humoral responses to the AAV-delivered mAb were measured by an antibody capture ELISA. Plates were coated with purified 5L7 IgG1 and 4L6 IgG1. After coating and blocking, we incubated the plates with antisera from the AAV-immunized monkeys. For detection, we probed with a HRP-conjugated anti-human lambda light chain antibody (Southern Biotech). This secondary antibody did not cross-react with the coated mAb on the plates since 5L7 IgG1 and 4L6 IgG1 harbor a kappa light chain. For detecting kappa anti-anti responses we produced recombinant versions of mAb that harbor a lambda light chain and detected kappa chain anti-mAb antibodies by probing with a HRP-conjugated anti-human kappa light chain antibody (Southern Biotech). To assess the target of the anti-anti response we tested the reactivity of antisera against AAV-delivered mAb and its immunoadhesin version as well as against antibody or immunoadhesin that was not given to one group of animals. For example, reactivity of antisera from 5L7 IgG1 recipient was tested against purified 5L7 IgG1, 5L7 immunoadhesin, 4L6 IgG1 and 4L6 immunoadhesin.
Infectious SIVmac239 was produced in PBMCs as described previously [27]. SIV challenges were conducted by repeated low dose intravenous injection with SIVmac239. The challenge dose was 1x AID50 for the first four injection and was elevated to 2x AID50 and 10x AID50 for the fifth and sixth challenge, respectively. The animals were exposed to SIV every three weeks or until positive viremia was measured. Challenge virus in plasma was detected by quantitative real-time RT-PCR using primers specific for SIV gag as described previously[66].
Presence or absence of anti-SIV envelope antibodies following SIV challenge was assessed by measuring the reactivity of sera (before and after challenge) against purified SIVmac251 truncated gp41 protein (ImmunoDX) by ELISA.
|
10.1371/journal.pntd.0004347 | Dynamics of Vector-Host Interactions in Avian Communities in Four Eastern Equine Encephalitis Virus Foci in the Northeastern U.S. | Eastern equine encephalitis (EEE) virus (Togaviridae, Alphavirus) is a highly pathogenic mosquito-borne zoonosis that is responsible for occasional outbreaks of severe disease in humans and equines, resulting in high mortality and neurological impairment in most survivors. In the past, human disease outbreaks in the northeastern U.S. have occurred intermittently with no apparent pattern; however, during the last decade we have witnessed recurring annual emergence where EEE virus activity had been historically rare, and expansion into northern New England where the virus had been previously unknown. In the northeastern U.S., EEE virus is maintained in an enzootic cycle involving the ornithophagic mosquito, Culiseta melanura, and wild passerine (perching) birds in freshwater hardwood swamps. However, the identity of key avian species that serve as principal virus reservoir and amplification hosts has not been established. The efficiency with which pathogen transmission occurs within an avian community is largely determined by the relative reservoir competence of each species and by ecological factors that influence contact rates between these avian hosts and mosquito vectors.
Contacts between vector mosquitoes and potential avian hosts may be directly quantified by analyzing the blood meal contents of field-collected specimens. We used PCR-based molecular methods and direct sequencing of the mitochondrial cytochrome b gene for profiling of blood meals in Cs. melanura, in an effort to quantify its feeding behavior on specific vertebrate hosts, and to infer epidemiologic implications in four historic EEE virus foci in the northeastern U.S. Avian point count surveys were conducted to determine spatiotemporal host community composition. Of 1,127 blood meals successfully identified to species level, >99% of blood meals were from 65 avian hosts in 27 families and 11 orders, and only seven were from mammalian hosts representing three species. We developed an empirically informed mathematical model for EEE virus transmission using Cs. melanura abundance and preferred and non-preferred avian hosts. To our knowledge this is the first mathematical model for EEE virus, a pathogen with many potential hosts, in the northeastern U.S. We measured strong feeding preferences for a number of avian species based on the proportion of mosquito blood meals identified from these bird species in relation to their observed frequencies. These included: American Robin, Tufted Titmouse, Common Grackle, Wood Thrush, Chipping Sparrow, Black-capped Chickadee, Northern Cardinal, and Warbling Vireo. We found that these bird species, most notably Wood Thrush, play a dominant role in supporting EEE virus amplification. It is also noteworthy that the competence of some of the aforementioned avian species for EEE virus has not been established. Our findings indicate that heterogeneity induced by mosquito host preference, is a key mediator of the epizootic transmission of vector-borne pathogens.
Detailed knowledge of the vector-host interactions of mosquito populations in nature is essential for evaluating their vectorial capacity and for assessing the role of individual vertebrates as reservoir hosts involved in the maintenance and amplification of zoonotic agents of human diseases. Our study clarifies the host associations of Cs. melanura in four EEE virus foci in the northeastern U.S., identifies vector host preferences as the most important transmission parameter, and quantifies the contribution of preference-induced contact heterogeneity to enzootic transmission. Our study identifies Wood Thrush, American Robin and a few avian species that may serve as superspreaders of EEE virus. Our study elucidates spatiotemporal host species utilization by Cs. melanura in relation to avian host community. This research provides a basis to better understand the involvement of Cs. melanura and avian hosts in the transmission and ecology of EEE virus and the risk of human infection in virus foci.
| Eastern equine encephalitis (EEE) is a highly pathogenic mosquito-borne virus responsible for outbreaks of severe disease in humans and equines, causing high mortality and neurological impairment in most survivors. In the past, human disease outbreaks in the northeastern U.S. occurred sporadically with no apparent pattern; however, during the last decade, this region has experienced changes in the frequency of EEE virus activity with expansion into new localities. We studied vector-host interaction using molecular methods in order to: 1) quantify vector host-feeding preferences of Culiseta melanura, 2) identify key bird species that serve as frequent hosts for Cs. melanura and reservoir hosts for the virus, and 3) determine the extent to which these preferences shape virus transmission dynamics in four historic foci of EEE virus activity in Connecticut. We also examined avian population density and dynamics to determine spatiotemporal host community composition experienced by Cs. melanura and EEE virus. We developed an empirically informed mathematical model to describe enzootic transmission of EEE virus in a community of multiple avian hosts and the primary mosquito vector. To our knowledge this is the first mathematical model for EEE virus, a pathogen with many potential hosts, in the northeastern U.S. Our study clarifies the host associations of Cs. melanura in four foci of virus activity in the northeastern U.S., identifies vector host preferences as the most important transmission parameter, and quantifies the contribution of preference-induced contact heterogeneity to enzootic transmission. Our study also elucidates spatiotemporal host species utilization by Cs. melanura in relation to avian host abundance. We identified eight bird species (Wood Thrush, American Robin, Tufted Titmouse, Common Grackle, Chipping Sparrow, Black-capped Chickadee, Northern Cardinal, and Warbling Vireo) with a larger feeding index than the remaining birds, indicating these species were fed upon by Cs. melanura more frequently, and therefore are important in maintenance of EEE virus at these four foci. This research provides a basis for better understanding the involvement of Cs. melanura in the transmission and ecology of EEE virus, and the risk of human infection in virus foci.
| Eastern equine encephalitis (EEE) virus (Togaviridae, Alphavirus) is responsible for outbreaks of severe disease in humans and equines, causing high mortality and neurological sequelae in most survivors [1,2]. EEE virus is maintained in an enzootic transmission cycle involving ornithophagic mosquitoes, specifically Culiseta melanura (Coquillett) (Diptera: Culicidae), and passerine birds in freshwater swamp foci [1–6]. In the past, disease outbreaks have occurred intermittently with no apparent pattern. Since 2003, however, the northeastern U.S. and southeastern Canada have experienced a resurgence of EEE virus activity with expansion into new regions [7]. These outbreaks occur when ecological factors and environmental conditions favor virus amplification followed by overflow into human and equine populations.
It is widely acknowledged that Cs. melanura feeds predominately on birds; however, the identity of key bird species that may serve as superspreaders of EEE virus has not been established in various virus foci [8–11]. Regional differences exist in the proportion of blood meals by Cs. melanura from various avian species that may be due to the availability and abundance of these birds among other ecological and physiological factors. Vector-host interaction studies conducted in EEE virus foci in the northeastern U.S. have identified >50 bird species as hosts for Cs. melanura, among which Wood Thrush and American Robin were most common [12–14]. Serological surveys also indicate that many of these bird species were frequently exposed to EEE virus [2,15]. The percentage of viral antibody was the highest for Wood Thrush, followed by American Robin, Ovenbird, and Swamp Sparrow in studies conducted in New Jersey and Massachusetts [2,16].
Successful transmission of EEE virus in avian host communities is governed by the abilities of host species to maintain, amplify, and transmit the virus to mosquito vectors (mainly Cs. melanura), and by ecological factors that influence contact rates between competent avian hosts and the mosquito vector. Earlier studies have identified a number of avian species as hosts for Cs. melanura. However, the potential for vector host preference to affect transmission of EEE virus in multiple foci has not been fully explored [12–14]. In this study, we investigated vector-host contact rates between Cs. melanura and avian hosts by identifying host species from blood meals, and the potential for heterogeneity in vector host preference to influence EEE virus transmission dynamics. The main objectives of this study were to 1) quantify vector host-feeding preferences in EEE virus foci, 2) identify key bird species that serve as frequent hosts for Cs. melanura, and as reservoir hosts for the virus, and 3) determine the extent to which these preferences shape the virus transmission dynamics. To achieve these objectives, we used PCR-based molecular methods and direct sequencing of the mitochondrial cytochrome b gene for profiling of blood meals in Cs. melanura to quantify its contact with vertebrate hosts, and to infer epidemiologic implications of its feeding behavior in four historic EEE virus foci in the northeastern U.S. We conducted avian point count surveys to determine spatiotemporal host community composition experienced by Cs. melanura and EEE virus. Finally, we developed a novel empirically informed mathematical model to describe enzootic transmission of EEE virus in a community of multiple avian hosts and the mosquito vector, Cs. melanura.
Field studies were conducted in four historic EEE virus foci, Chester, Killingworth, Madison, and North Stonington, CT (Fig 1). These four locations were considered to be virus foci because from 1996 to 2014, EEE virus was detected in 115 pools of mosquitoes, including 74 (63%) pools of Cs. melanura, the enzootic vector for EEE virus. North Stonington had the greatest number of positive pools (n = 47, 40.9%), followed by Chester (n = 42, 36.5%), Killingworth (n = 13, 11.3%), and Madison (n = 13, 11.3%). Majority of the positive pools were identified during 2009 (n = 56, 48.7%), which contributed to the rationales for initiating the present study during 2010–2011 (Table A in S1 Text). In 2010, four mosquito pools tested positive in mosquitoes from North Stonington but none from the other three sites. Interestingly, positive mosquito pools were identified in North Stonington in 8 years of the nearly two decades during which active mosquito surveillance has been conducted for EEE and other arboviruses in Connecticut.
A total of 6,234 female Cs. melanura were collected from the four EEE virus foci using 120 resting boxes (or 11”x11” stackable fiber nursery pots) placed on dry forested uplands within sight of red maple/Atlantic white cedar swamp habitats, and along the edges of these swamps in 8 sites, Chester 3, Killingworth 1, Madison 3, and North Stonington 1, during May through October, 2010–2011, and according to the established protocol [17] (Table 1). Greater numbers of Cs. melanura were collected during 2011 (total n = 4390; Chester 1936, Killingworth 777, Madison 1129, and North Stonington 548) than in 2010 (total n = 1844; Chester 548, Killingworth 545, Madison 577, and North Stonington 174). Multiple collection peaks were observed during the trapping season, which suggested 2–3 generations of Cs. melanura each year (Fig A in S1 Text). Resting boxes were examined daily, and a battery-powered handheld aspirator was used to collect engorged mosquitoes. Specimens were transported in coolers containing dry ice to the laboratory. Mosquitoes were then identified to species using a dissecting microscope and an identification key [18]. Specimens with visible blood meals were transferred to 1.5 mL microtubes, labeled with a unique number, and stored in an ultra-low temperature freezer.
Mosquito abdomens were removed with the aid of a dissecting microscope and disposable razor blades for blood meal analysis. DNA was extracted from the abdominal content of engorged mosquitoes individually by using DNAzol BD (Molecular Research Center, Cincinnati, OH, USA) according to the manufacturer’s recommendation with some modifications as described elsewhere [13,19]. Extracted DNA from the mosquito blood meals served as DNA templates in subsequent polymerase chain reaction assays with primers based on vertebrate mitochondrial cytochrome b sequences according to published protocols [13,19]. Sequencing of both DNA strands was carried out on 3730xL DNA Analyzers, using Big Dye chemistries (Applied Biosystems Inc., Grand Island, NY) at the Keck Sequencing Facility, Yale University, New Haven, CT. Sequences were analyzed and annotated using ChromasPro version 1.7.5 (Technelysium Pty Ltd., Tewantin, Australia), and identified by comparison to the GenBank DNA sequence database utilizing the BLAST search (BLASTN) of the National Center for Biotechnology Information [20].
Avian point count surveys were conducted in the study sites weekly from April through October, 2010–2011 in order to assess species composition and relative spatial and temporal abundance of bird species, according to the previously described protocols [21,22] (Table 1). A skilled observer knowledgeable of the identification and vocalizations (i.e. songs and call notes) of the local bird fauna conducted the point count surveys. Three 100 m-diameter point count circles, located 50–200 m apart, were established at each resting box site with the aid of a Garmin eTerx 10 GPS unit (Garmin International, Inc., Olathe, KS) and a rangefinder (Bushnell Outdoor Products, Overland Park, KS). To estimate bird distances within the point count circles, flagging tape was affixed to trees or shrubs at distances of 15, 30, and 50 m. Landmarks within the point count circles were also used to estimate distances. Counts began shortly after sunrise under favorable weather conditions, when bird activity and vocalization are highest. The observer approached the point count with as little disturbance as possible and began counting birds upon arrival at the site. With the aid of binoculars and a stopwatch, the observer recorded the distance and number of bird species seen and heard during a 15-minute period. Status and habitat codes, including observations of nestlings, fledglings, and juvenile birds, and individuals detected outside or flying over the point count circles, were recorded. Temperature and wind speed was measured with a hand-held anemometer (La Crosse Technology, La Crosse, WI) and recorded on the form along with the start and finish times (Fig B in S1 Text). Bird nomenclature followed the 7th edition of the “American Ornithologists’ Union” Checklist [23].
The modeling methods are briefly described here, and more detailed information is provided in the “Supporting Information”. The data for each of the four EEE virus foci was pooled to increase sample size and to illustrate general trends. Across the four locations, eight bird species were selected based on their high abundance in bird counts or high prevalence as the source of blood meals, including Wood Thrush (Hylocichla mustelina), American Robin (Turdus migratorius), Tufted Titmouse (Baeolophus bicolor), Common Grackle (Quiscalus quiscula), Chipping Sparrow (Spizella passerina), Black-capped Chickadee (Poecile atricapillus), Northern Cardinal (Cardinalis cardinalis), and Warbling Vireo (Vireo gilvus). The remaining bird species at each location were combined into a ninth category of other birds. Our model simulates the dynamics of EEE virus in each of these eight bird species, the remaining birds and Cs. melanura, over a season of 180 days. We chose this time interval to simulate a typical mosquito activity season.
The feeding index was used to model the preference of Cs. melanura for the different bird species [24,25]. The feeding index assesses the proportion of blood meals from a host species relative to the abundance of that species in the host community, or the relative likelihood of a blood meal on a given bird species per bird of that species. Thus a feeding index shows the relative preference for one bird species compared to the other species examined. As a result, we chose the ninth category, other birds, as the frame of reference for the other species, assigning its feeding index the value of 1. We calculated the feeding index for the eight bird species from the bird count and blood meal data collected in this study.
The force of infection is defined as the per-capita rate, or hazard, at which susceptible hosts become infected, with a separate force of infection for each bird species and one for the mosquitoes. The number of female mosquitoes, the feeding index and bird abundances determine the relative number of mosquito bites on each bird species. The proportion of infectious mosquitoes and the vector-to-host transmission rate provide the force of infection for the bird species, while the proportions of each infectious bird and the host-to-vector transmission rate give the mosquito force of infection. We assumed that host-to-vector and vector-to-host transmission rates were constant across all bird species, due to a lack of comprehensive information on species-specific transmission rates. Statistical sampling models were used to account for sampling error due to small numbers of bird counts and blood meals for some species. Utilizing Markov chain Monte Carlo methods, 1000 samples were selected for both the counts and blood meals. For each of these samples, the feeding index of the selected bird species were calculated, and the median and inner 95% quantiles were reported.
For each of the selected bird species, we used a standard SIR sub-model to simulate the EEE virus transmission dynamics, with the population of each species divided into susceptible, infectious, and recovered compartments. We assumed that vectors do not recover from infection, using a SI sub-model. Each sub-model includes births and deaths, with birth balancing deaths leaving population sizes constant. The number of susceptible hosts increases according to their birth rates, and decreases due to infection and death. For simplicity, we assumed that there was no death due to infection in birds and mosquitoes, so that death is only due to background mortality, and that the population size of each bird species and mosquitoes remained constant throughout the simulation period. Mosquitoes and each bird species transition from susceptible to infectious according to their forces of infection; birds transition from infectious to recovered at a recovery rate of 1 per day, i.e. 1 day mean duration of infectious viremia. The model then simulated for a period of 180 days, starting with all bird species completely susceptible and one tenth of one percent of mosquitoes infectious.
Studies suggest that vector-to-host transmission is guaranteed when an infectious vector feeds from a host [26], and thus the vector-to-host transmission rate was set to 1. Limited data exists regarding the value of the host-to-vector transmission rate among the various host species. In order to focus the analysis on the effect of variable biting rates on the model's output, the host-to-vector transmission rate is assumed to be constant across all of the host species. The all-species host-to-vector transmission rate was determined by fitting the model proportion of birds that became infected over the model period to the observed proportion of seropositives amongst the various bird species from a previous study [15]. For bird species that were not present in the previous study, the mean proportion of seropositives from that study was used.
A total of 1,798 Cs. melanura with visible blood meals were collected from the four virus foci, and blood meal sources were identified in 1,127 (62.7%) specimens by DNA sequencing. These included Chester 348 of 581 (59.9%), Killingworth 249 of 364 (68.4%), Madison 361 of 598 (60.4%), and North Stonington 169 of 255 (66.3%). The remaining blood-fed Cs. melanura either did not produce visible amplification products or the sequencing results were insufficiently conclusive to assign a host species. In addition to Cs. melanura, 372 engorged specimens of 12 species in the genera of Aedes, Anopheles, Coquillettidia, Culex, Culiseta, and Ochlerotatus were collected, and blood meal analyses conducted (Table B in S1 Text). However, because the focus of the present study was on Cs. melanura, the principal mosquito vector of EEE virus, results of these analyses are not presented here.
Of the 1,127 engorged Cs. melanura blood meals that were successfully linked to a vertebrate host at the species level, 99.4% were from avian hosts, comprising 65 species from 27 families and 11 orders. Passeriformes constituted the most numerous hosts representing 97.5% of avian blood meals. Comparatively few Cuculiformes (1.0%), Columbiformes (0.6%), Accipitriformes (0.3%), Strigiformes (0.2%), and six other avian orders were additionally identified. Among taxonomic families, Turdidae (thrushes) served as the most frequent hosts (33.4%), followed by Paridae (chickadees and titmice, 18.6%), Cardinalidae (cardinals and tanagers, 10.9%), Icteridae (blackbirds, 9.5%), Vireonidae (vireos, 8.0%), and 22 other avian families (Table 2). Four mammalian species belonging to the Cervidae (White-tailed Deer), Bovidae (Domestic Cow and Sheep) and Sciuridae (Eastern Gray Squirrel) were also identified in individual or mixed blood meals.
A total of 37 avian families in 14 orders were encountered at the study sites during the point count survey with differing frequencies calculated over 7 months, April to October. The most frequent avian order was Passeriformes (86.2% of all birds) with 20 families including Paridae (chickadees and titmice, 17.7%), Icteridae (blackbirds, 9.7%), Turdidae (thrushes, 8.8%), Parulidae (wood warblers, 8.1%), and Emberizidae (New World Sparrow, 6.7%). Other relatively frequent avian orders included Piciformes ([Picidae (woodpeckers)], 5.2%), Anseriformes ([Anatidae (ducks, geese, and swans)], 3.3%), and Columbiformes ([Columbidae (doves and pigeons)], 1.6%) (Table G in S1 Text).
We compared percentage of avian-derived blood meals for Cs. melanura with average avian frequencies in the four EEE virus foci.
The eight selected bird species (Wood Thrush, American Robin, Tufted Titmouse, Common Grackle, Chipping Sparrow, Black-capped Chickadee, Northern Cardinal, and Warbling Vireo) were found to have a larger feeding index (Table L in S1 Text) than the remaining birds, indicating these species were fed upon more frequently by Cs. melanura (Tables 3–6 and C-F in S1 Text). The feeding index was highest for Wood Thrush followed by Warbling Vireo. As a result, both Wood Thrush and Warbling Vireo see early peaks in infections (Fig 4). The infections in Wood Thrush in turn increase the prevalence of infection in Cs. melanura (Fig C in S1 Text). As a result, the infection rate for some of the remaining bird species, which are less preferred as blood meal hosts, can be seen to increase slightly, allowing EEE virus to persist through the entire season.
Our study examines vector–host interactions of Cs. melanura, the principal vector of EEE virus in the northeastern U.S., and demonstrates how a relatively limited number of avian hosts may regulate the dynamics of pathogen transmission in complex host communities. We found that Wood Thrush, in particular, may effectively function as a principal reservoir host that serves to amplify EEE virus during the early summer months, whereas less preferred avian hosts facilitate sustained transmission and persistence of virus throughout the remainder of the summer and early fall. These results have broader implications for understanding the perpetuation of vector-borne pathogens in species-rich host communities. The concept of dilution effect has recently been proposed which purports that increases in host diversity may lead to a reduction in disease risk due to the dilution of competent host species [27]. This premise has been applied to tick-borne Lyme disease but its application and relevance to other vector-borne pathogens has been questioned [28]. EEE virus occurs in freshwater swamps where host species diversity is relatively high: in our study sites, we encountered up to 99 avian species. Nevertheless, our model demonstrates that EEE virus may readily amplify in these ecological settings because the mosquito vector Cs. melanura preferentially feeds on a few virus-competent bird species.
Wood Thrush served as the most frequent host for Cs. melanura in Madison (28.5%) and Killingworth (26.9%), and was relatively frequent in North Stonington (11.8%) and Chester (4.6%). Similarly, in earlier studies, Wood Thrush was identified as the most frequent host for Cs. melanura (23.6%) in New York, as the 2nd most frequent (12.5%) in Connecticut, and as a relatively frequent host (5.1%) in Massachusetts [12–14]. Wood Thrush breed in deciduous and mixed forests in the eastern U.S. where there are large trees, moderate understory, shade, and abundant leaf litter for foraging. The breeding range for these birds extends from Manitoba, Ontario, and Nova Scotia in southern Canada to northern Florida, and from the Atlantic coast to the Missouri River and the eastern Great Plains [29]. Wood Thrush was an important source of blood meal later in the season (August-September) where 85.1% (n = 57) of all blood meals were from this bird species in Killingworth, 80.6% (n = 83) in Madison, 80.0% (n = 16) in North Stonington, and 62.6% (n = 10) in Chester (August-October). The intensity of mosquito feeding on this bird species late in the season overlaps with the molt period in adult Wood Thrush that extends from July to early October, a period during which they lose flight feathers, and some individuals drop several primaries over a few days [30]. This extensive molting impairs flight efficiency and makes them remarkably cautious and difficult to observe [30].
Frequent infection of Wood Thrush with EEE virus has been reported. Of the 42 isolations of EEE virus from more than 3,000 birds bled in southern Alabama, there were more from Wood Thrush than any other bird species [31]. In New Jersey, early-season virus isolation from Wood Thrush and a few other bird species has been reported as evidence of a cryptic EEE virus cycle [2]. In Massachusetts, Wood Thrush had the highest EEE virus antibody prevalence rate (26.7%) among 20 avian species examined [16]. In a study conducted in a prominent EEE virus focus surrounding Toad Harbor Swamp in upstate New York, Wood Thrush had antibody prevalence rates of 7.0%, 9.6%, and 50% during 1978, 1979, and 1980, respectively [32]. EEE virus was also isolated from a migrating Wood Thrush in the Mississippi Delta [33].
American Robin served as the second most frequent host for Cs. melanura in Killingworth (18.5%), Madison (14.1%), and Chester (10.9%), and the third most frequent host in North Stonington (11.8%). Similarly, American Robin served as the most frequent source of blood meal for Cs. melanura in neighboring Massachusetts (21.7%), and in Connecticut in an earlier study (22.9%), and as the second most frequent host in New York (9.1%) [12–14]. American Robin has also been reported as a frequent host for other mosquito species such as Culex spp. throughout the Northeast and other regions of the U.S., highlighting the role of this bird species in the amplification of another avian arbovirus: West Nile virus [10,19,34–36].
American Robin is a common bird species throughout most of North America with permanent and migratory populations. Populations of this avian species inhabit a wide variety of open and forested habitats in urban/suburban and rural settings, riparian forests, early successional forests, and closed canopy forests and woodlands [37–40]. American Robin can occur in large flocks that roost communally in woodland habitats during summer months after nesting ends. Within these roosts, it can be the most prominent tree-roosting bird [41]. The first brood of American Robin emerges in late April through June in southern regions of the Northeast, and in May through early July in northern areas (e.g., northern Maine) providing temporal overlaps with the first generation of Cs. melanura.
American Robin is a competent amplifying host for EEE virus, and the virus has been isolated from this bird in Massachusetts and New Jersey [2,16,42]. Serosurveys indicate American Robin is frequently exposed to EEE virus throughout the region [2,16, 43–45]. Identification of American Robin as a frequent host in this study, in conjunction with its abundance and other evidence, suggests that this bird species contributes to EEE virus amplification in the region.
Tufted Titmouse was identified as the most frequent host for Cs. melanura in Chester (22.4%) and a frequent host in Killingworth (15.7%), North Stonington (5.9%), and Madison (5.8%). Our earlier vector-host interaction study in Massachusetts also reported that Tufted Titmouse served as the second most frequently identified host (8.7%) for Cs. melanura [14]. Tufted Titmouse is common east of the Great Plains in the woodlands of the southeastern, eastern, and midwestern U.S., and in southern Ontario, Canada. Tufted Titmouse prefers deciduous woods or mixed evergreen-deciduous woods, especially moist woodlands found in swamps and river basins, and areas with a dense canopy and many tree species [46]. Tufted Titmouse does not migrate extensively, and remains in residence throughout the winter [47]. Tufted Titmouse was shown to have particularly high antibody prevalence (44.2%) for EEE virus, and specimens of this species were also captured with active viremia in Cape May, New Jersey [2].
The mathematical model we developed is, to our knowledge, the first for EEE virus, a pathogen with many potential hosts, in the northeastern U.S. The most notable result of the model is the dominant role played by Wood Thrush in amplifying EEE virus. Relative to the other bird species, Wood Thrush had a small observed population in the bird counts but a high number of identified blood meals from field-collected mosquitoes. As a result, Wood Thrush had the largest feeding index, and hence played the largest role in the spreading of EEE virus amongst the various bird populations in the model.
We chose to focus on eight bird species due to both their large populations in the bird counts and the large number of blood meals collected from these species. To maintain simplicity in the model, and due to small sample sizes for the remaining bird species, the remaining observed species were combined into a single class. Future work could attempt to identify groups of bird species that are necessary to accurately model EEE virus dynamics. Combining bird species into groups like migratory vs non-migratory could give a clearer understanding of the role of various bird species provided such grouping accurately models EEE virus dynamics. Modeling migratory birds might even allow the data between different sites to be linked, and could potentially be used to track movement from more southerly regions where the virus circulates all year long.
In order to understand the role that vector feeding preference plays in the amplification of EEE virus, and given the limitations of the collected data, several assumptions were made. Data to estimate the transmission and recovery rates for the host species is limited. As a result, we chose these rates to be the same across the host species. With the assumption that transmission rates do not vary by species, limited sensitivity analysis has shown that species that act as amplifying hosts are insensitive to changes in transmission and recovery rates. If data were available to estimate these parameters separately for each host, then a wide range of substantially different dynamic patterns of infection between host species would be possible.
Pooling the samples over time leads to a model that does not include changes in the mosquito or bird populations over the 180-day model period. If the observed seasonal changes in bird species abundance were incorporated into the model, we expect that overall infection rates amongst the bird species would remain about the same, but the time when infections peak in each species could shift depending on the abundance over time. Further investigations are underway to better understand the potential influence of these factors. We have also assumed that the feeding index values remain constant across the 180-day model period. Wood Thrush may be more susceptible to being bitten during their molting period from late July through August; incorporating this shift in biting preference over time into our model would lead to Wood Thrush becoming infected later in the season than in our current results [30].
Several bird species had comparatively few observed bird counts or identified blood meals. In particular, some bird species in some sample periods had 0 observations in the bird count but a positive number of blood meals, which for our simple estimate would give an infinite feeding index in these periods. As a result, these feeding index values are extremely sensitive to small perturbations in the data. This sensitivity can be seen in the confidence intervals for the feeding index for both Wood Thrush and Warbling Vireo, as both species had relatively small observed bird counts, which led to high coefficients of variation in our Poisson sampling model for bird counts. In order to further examine previously excluded host species, more sophisticated sampling models would be needed.
Despite these limitations, the model provides valuable insight into the role that vector feeding preferences play in the transmission of EEE virus. In particular we notice that observed bird species with relatively small abundances could drive a majority of the infections across all bird species, due to the amplification of the virus in those species early in the season.
In conclusion, we found that Cs. melanura was exposed to diverse avian communities but preferentially focused feeding on Wood Thrush. The model suggests that this species may play a vital role in supporting EEE virus amplification, subject to confirmation that Wood Thrush is a competent host. Culiseta melanura had fed frequently on several other bird species, including American Robin, Tufted Titmouse, Common Grackle, Chipping Sparrow, Black-capped Chickadee, Northern Cardinal, and Warbling Vireo, that were shown to play a less important role in maintaining EEE virus transmission later in the season.
|
10.1371/journal.pcbi.1006100 | The AmP project: Comparing species on the basis of dynamic energy budget parameters | We developed new methods for parameter estimation-in-context and, with the help of 125 authors, built the AmP (Add-my-Pet) database of Dynamic Energy Budget (DEB) models, parameters and referenced underlying data for animals, where each species constitutes one database entry. The combination of DEB parameters covers all aspects of energetics throughout the full organism’s life cycle, from the start of embryo development to death by aging. The species-specific parameter values capture biodiversity and can now, for the first time, be compared between animals species. An important insight brought by the AmP project is the classification of animal energetics according to a family of related DEB models that is structured on the basis of the mode of metabolic acceleration, which links up with the development of larval stages. We discuss the evolution of metabolism in this context, among animals in general, and ray-finned fish, mollusks and crustaceans in particular. New DEBtool code for estimating DEB parameters from data has been written. AmPtool code for analyzing patterns in parameter values has also been created. A new web-interface supports multiple ways to visualize data, parameters, and implied properties from the entire collection as well as on an entry by entry basis. The DEB models proved to fit data well, the median relative error is only 0.07, for the 1035 animal species at 2018/03/12, including some extinct ones, from all large phyla and all chordate orders, spanning a range of body masses of 16 orders of magnitude. This study is a first step to include evolutionary aspects into parameter estimation, allowing to infer properties of species for which very little is known.
| We discovered that parameters of Dynamic Energy Budget (DEB) models can be estimated from a set of simple data on animal life history aspects, growth and reproduction, if treated in combination. Apart from goodness-of-fit as an estimation criterion, relations with parameter values of other species are important, since DEB parameters have a clear physiological interpretation and a good fit for the wrong reasons is always a risk to consider. We developed and optimized methods for this type of parameter estimation-in-context and organized the results of over 1000 animal species in the open-access Add-my-Pet (AmP) database, to which 125 authors contributed so far. We also developed software package AmPtool to compare parameter values in the collection, that builds on DEBtool to assist applications of DEB theory. A family of related DEB models, structured with respect to the modes of metabolic acceleration, captures biodiversity, including various life stages. We discuss some features of the family structure of DEB models in an evolutionary context. The AmP collection has a great potential for research on the role of biodiversity in ecosystem structure and functioning, which will grow with the size of the database.
| The role of biodiversity in ecosystem structure and functioning is central for conservation and environmental quality management, as well as biospherics and earth system studies. Biodiversity is not only about the number of species present, but also the number and nature of the different characteristics and functions which make up a community or an ecosystem, often referred to as traits. Scientists and managers are turning towards such trait-based approaches to measure the health and vitality of ecosystems. In this context of apprehending biodiversity on the basis of diversity of characteristics and functionalities we have been developing the AmP (Add-my-Pet) project.
AmP is a database of referenced data on animal energetics, parameter values of models based on Dynamic Energy Budget (DEB) theory [1–4], and properties derived from these parameters. Some 125 authors contributed to the database at 2018/03/12. The AmP project aims: (i) to find the simplest organization principles for metabolism upon which all life is based and (ii) to understand taxon-specific patterns as variations on this common organization. The development of DEB theory started in 1979 and meanwhile over 700 papers have been published on DEB theory, see www.zotero.org/groups/500643/deb_library/.
Partly based on the fact that a large number of popular empirical models turned out to be special cases of DEB models [4], we claim that DEB theory is presently the best tested quantitative theory in biology. The comparison of species on the basis of parameter values is an important aspect of the AmP project. Species-comparisons based on measured quantities suffer from the problem that these quantities typically have contributions from many underlying interacting processes, and were not measured for all species of interest. Parameters of mechanistic models, however, have much simpler links with such processes, which makes it easier to find explanations for differences between species. Moreover, the complete parameter set is available for each entry, allowing to predict, e.g. respiration, without any measured data on respiration.
Comparison of species is, however, not the only important application of the AmP website. Prediction of effects of global change [5], understanding the geographic distribution of species [6–8], the effects of (toxic) chemical compounds [9–12], the optimization of bio-production (e.g. aquaculture and agriculture [13–15]), stock management, the best re-introduction of endangered species or the control of invading species [16], are just examples of applications where detailed knowledge of energetics of species in a DEB context is very useful. Like many ecologists, we see energetics as the key to understand the ecological behavior of species [17], and as the root of population [18] and ecosystem dynamics [19, 20], with consequences at the planetary level [21]. This is the context that motivated the development of DEB theory, of which the AmP project is an application. In view of the rapid build-up of ecological stress all over the world, we think that the field is in urgent need of an online database like AmP.
AmP started in 2009 as an educational initiative, to teach researchers how to estimate DEB parameters from their data and animals have the simplest metabolisms (if we compare with say plants, bacteria or microalgae). The database grew and in 2013, (at about 300 entries) we teamed up, formed the first AmP curator board, and together developed the code and web-platform underlying AmP. We wanted to get an overview on: (i) how well the standard DEB model worked for describing animal metabolism, (ii) standardize and improve the parameter estimation procedure and (iii) improve our capacity to judge the realism of parameter values.
We start with a brief introduction of DEB theory. All applications of models, including testing of the model against data, start with knowledge of parameter values. The parameters of a DEB model must be estimated from a collection of data sets on the various aspects of energy budgets and life history, using all this information in combination. This involves new features in the methodology of parameter estimation. The next section of this paper describes improvements we implemented, based on the co-variation method [22, 23], which was used in an early phase of the database project.
Moreover we needed more administrative rigor and improved methods for detecting patterns in parameter values. Development of new routines and re-organization of the previous estimation procedure allowed us to include these new extensions. So the following section describes the new web-interface and structure of the database. Finally we present and discuss the results obtained after implementing the new method and reaching 1035 species in the collection at 2018/03/12. Given that these entries employ together 270 different types of data, in 585 combinations, the estimation of 14 parameters of each species not only illustrates the scope of the data reduction, but also the step-up in comparison potential.
Dynamic Energy Budget (DEB) theory, aims to specify commonalities underlying metabolic organization for all life. It does this by delimiting a small set of assumptions from which mathematical formulae for metabolism are derived, covering the start of embryo development to death by aging through a range of life stages [24–27]. DEB models are meant to apply to all life on earth and allow species comparisons on the basis of (functions of) parameters of that model. The standard DEB model (’std’) is the simplest non-degenerated DEB model implied by the theory and it applies to heterotrophic animals (see Fig 1 for the summary). We refer the reader to [28] for an accessible summary of the principles of DEB theory and a description of the standard DEB model. A detailed derivation of the model on the basis of underlying assumptions is presented in [4], chap. 2. Each DEB parameter of the std-DEB model (see Table 1) has a clear link with one underlying physiological process (specified on the arrows of Fig 1). The combination of parameters covers all aspects of energetics throughout the full life cycle of organisms, include feeding, digestion, storage, maintenance, growth, development, reproduction, aging. Parameter values are individual-specific in the context of DEB theory, but the difference between individuals are typically small enough to average for a species in a meaningful way. Parameter values determine how state variables of an individual (reserve, structure, maturity and reproduction buffer) change in time through all life stages (embryo, juvenile, adult). Life stages have specific definitions: Embryos do not assimilate; Juveniles assimilate and allocate to maturity but not to reproduction; Adults feed, no longer allocate to maturity but store energy/ mass allocated to reproduction into a reproduction buffer which is converted to offspring. To complete the background information needed for this study, we must point to the co-variation rules [29]. If there was no selection pressure of parameter values then particular parameters would co-vary in a simple way with the dimensionless zoom factor z (Table 1), purely based on parameters being either intensive or extensive. The estimation of parameters for each species allows comparing parameters between species and seeing to what extent each parameter is in fact either dependent or independent of body size for different taxonomic groups, thus revealing important patterns in how environmental pressures exert selection pressure on parameter values.
Prior to this work, DEB parameters for the AmP entries were obtained with the covariation-method for parameter estimation [22, 23]. All the code for parameter estimation is developed in the DEBtool_M package, which is frequently updated and freely available at https://github.com/add-my-pet/DEBtool_M/. In this section we describe improvements to the method which are implemented in DEBtool. The improvements comprise:
The overview of the parameter estimation procedure is presented in Fig 2. In the next section we describe the important new components.
One way to judge how good the parameter estimates are is to compute a goodness of fit measure which assesses how close the model predictions are to all of the data. Goodness of fit is not enough, one also needs to check for biological realism. The previous system used goodness of fit measures defined in section 2 of [23]. The new system uses the Mean Relative Error (MRE) and the Symmetric Mean Squared Error (SMSE) to quantify the goodness of fit. MRE = 1 n ′ ∑ i = 1 n RE i, where RE i = ∑ j = 1 n i w i j w i | p i j - d i j | | d i | (w i = ∑ j = 1 n i w i j > 0), that simplifies to R E i = | p i 1 d i 1 - 1 | for zero-variate data. n′ is the number of data sets with wi greater than 0. SMSE = 1 n ′ ∑ i = 1 n SSE i where SSE i = ∑ j = 1 n i w i j w i ( p i j - d i j ) 2 d i 2 + p i 2, that simplifies to SSE i = ( p i 1 / d i 1 - 1 ) 2 1 + ( p i 1 / d i 1 ) 2 for zero-variate data. See section on loss functions for the definition of the other symbols. MRE can have values in the interval [0, ∞], while SMSE has values in the interval [0, 1]. In both cases, 0 means predictions match data exactly. MRE assesses the differences between data and predictions additively, judging equally an overestimation and underestimation of the same relative size (e.g, +20% or −20% have the same contribution), while SMSE assesses the difference multiplicatively, judging overestimation and underestimation by the same factor equally (e.g. × 2 or ×1/2 have the same contribution). Notice that the result of the minimization of loss functions does not, generally, correspond with the minimum of MRE or SMSE (unless the fit is perfect).
AmPtool is a software package that is designed to analyse patterns in (functions of) parameter values in selected entries. It is available via github.com/add-my-pet/AmPtool/, and changes frequently since the AmP collection is rapidly expanding. Meta-data, parameters and implied properties (biologically relevant quantities which are functions of parameters as well as food and temperature) are collected in a single Matlab structure (allStat.mat). Advanced plot-routines were developed to plot (functions of) parameters against other (functions of) parameters, for selected species. These selections make use of the taxonomic tree. The tree actually consists of lists-of-lists, based on the newest insights in taxonomy, as presented in the Catalog of Life, the Encyclopedia of Life and Wikipedia. Legends can be created where different taxomonic groups in the tree are attributed user-defined markers. This legend functionality allows selecting taxa at a large number of taxonomic levels. We mentioned in the previous section that goodness of fit is not enough to judge a set of DEB parameters and that one must also check for biological realism. AmPtool is one way to do this. The user can visualize how parameters of closely and less closely related taxonomic groups relate to each other and see if the new parameter estimates are extreme outliers or not. Careful examination of the coherence of the implied properties within an entry is the other way. AmPtool can also be used to find entries on the basis of data types that have been used, or print values of parameters and statistics of selected species.
The AmP web-interface was developed as a result of this work. The interface allows examining each entry as well as obtaining on overview on how entries relate to each other.
From 2013 till 2017 the number of AmP entries (Fig 3, left) was dramatically increased from ca 300 in 2013 to 1035 at 2018/03/12 to find out whether DEB models do apply to all animals and determine how problematic it is to have limited amounts of data for estimating DEB parameters. AmP receives entries which are submitted by the international scientific community and to date some 125 authors have contributed to the collection. Every author and their associated entry (with links to the entries) are listed on the AmP web-interface. Over the course of this study, we included a bit over 700 extra species to the collection since 2013 as well as converted the previous 300 entries to this new format. We were careful to select species such that the collection had a broad taxonomic scope, as well as included species of commercial relevance and species relevant for conservation and toxicity testing. We aimed at maintaining a good balance between the different taxonomic groups (Fig 3, right) We were on the look-out for exceptional species (in terms of size or aging) that might show that the DEB model was not applicable. We did not yet find such an animal species.
The increasing number of DEB applications on animal species motivated the continuing amelioration of the method, making it more robust, more efficient and easier to apply. As the collection of species grew, it became evident that the standard DEB model (Fig 1) required simple extensions for particular taxa, to accommodate larval life stages, fetal development, various forms of metabolic acceleration [36], substantial programmed shrinking (as sported by Elopocephalai), etc. DEB models are classified as s-models, a-models and h-models according to the mode of metabolic acceleration. Since we see the structured model collection that resulted from the AmP project as an important insight into animal metabolism, we come back to it in the discussion section. We here present the typified models. They must be understood as variations on the standard DEB model (std), see Methods section on DEB theory. The general idea is that the choice of typified model depends more on higher-level classifications than the species-level. Delayed stage transitions are also accounted for in the different model families. Most mammals delay the start of fetal development during gestation. Some bivalves delay the start of metabolic acceleration; this phenomenon can prove to be more common with the increase of available data. The three sets of models are detailed below.
The AmP database not only allows us to test model against data and evaluate implications, but also to identify evolutionary and ecological patterns in parameter values. The number of papers on patterns in parameter values is now increasing rapidly [10, 15, 30, 31, 33, 36–41]. Apart from being of direct scientific relevance, future improvements of DEB parameter estimation methods might exploit these patterns, since lack of data is the rule rather than the exception and having some testable prediction is better than no prediction at all.
Where, for instance, (measured) maximum body weight is treated as an independent variable in the eco-physiological literature, DEB theory sees this as a property resulting from underlying processes that are quantified via parameter values. So maximum body weight is a function of parameter values (and food availability). Respiration (e.g. the use of dioxygen) is another function of parameter values, which has contributions from various underlying processes, such as maintenance, development, growth and assimilation overheads, etc. While respiration has been measured for only a small part of the species in the AmP collection, it is available for all species as a prediction (i.e. a function of parameter values). This is just an illustration of the power of comparison on the basis of parameter values, rather than on the basis of measurements. The Patterns page on the AmP website illustrates some patterns in parameter values, such as (predicted) respiration as function of (predicted) maximum body weight, confirming Kleiber’s empirical law stating that (measured) respiration is about proportional to (measured) maximum weight to the power 3/4. Two other patterns illustrate the explanation provided by DEB theory: relative reserve capacity is increasing and specific somatic maintenance is decreasing with body size. Reserve does not require maintenance, so does not contribute to respiration. The increase of reserve capacity [ E m ] = { p ˙ A m } / v ˙ with maximum (structural) length follows from co-variation rules [29], where energy conductance v ˙ (which quantifies reserve mobilization) is independent of maximum length, and specific maximum assimilation rate { p ˙ A m } is proportional to it. The increase of specific somatic maintenance with decreasing maximum (structural) length is seen as an ecological adaptation to exploit short-lasting peaks in food abundance [36]. Another pattern shows that, contrary to popular believe, the maximum growth rate of dinosaurs is just in line with other taxa, given their body size. The full analysis of all patterns in parameter values is beyond the scope of this work, and is an ongoing activity [33, 41].
The classification of DEB models into s–, a– and h–models has more or less clear links with the evolution of metabolism. This will be discussed here using a feature that sets a– and h–models apart from s–models: metabolic type M acceleration [36], where surface area temporarily scales with volume, while before and after this period it scales with volume to the power 2/3. The amount of type M acceleration is quantified by the ratio of body length at the end and the start of the acceleration period. The effect of this change in a DEB context is that both the specific assimilation and the energy conductance increase with length, while outside the acceleration period they stay the same. The ratio of the two, the specific reserve capacity, is not affected by this type of acceleration. The h–models differ mainly from the a–models by the fact that acceleration extends into the adult stage.
Fig 7 presents evolutionary relationships among animal taxa with a color code to indicate the amount of acceleration in their species. Since the oldest animal group, the Radiata, and the oldest deuterostomes, the echinoderms, accelerate, it might well be that acceleration became suppressed in several other groups and this suppression evolved several times in evolution [36].
The Ecdysozoa (n = 99 at 2018/01/01) beautifully illustrate the link between model type and taxonomic relationship. Fig 7 shows that Chaetognatha, Tardigrada, Nematoda and Entognatha hardly accelerate metabolism, just a factor 2 or less. The basic insects, the ephemeropterans and odonata accelerate a bit more, while the crown groups, the holometabolic insects, accelerate very much. All of the h–models are found in insecta, but very interestingly, springtails Entognatha (n = 6), which are no longer classified as insects, follow the abj model. Most insects seem to skip the juvenile phase and allocate to reproduction as larvae, which classifies them as adult by definition in DEB terms, while the imago neither grows, nor eats (frequently). Holometabolic insects insert a pupal phase between the larval and imago phases that behaves like an embryo with a reproduction buffer, where most of the larval structure is first converted to reserve [42] and imago structure is build from reserve. Crustacea (n = 62) sport a mix of s– and a–models Branchiopoda (n = 18) are described by std. Copepoda (n = 6) are hardly resolved and require more research. Calanaus does not accelerate (sbp), the others do (abp). Copepods are special with respect to the other crustaceans in that the κ-rule no longer applies to the adult stage. The 1 species of ostracod is described by abj. Malacostraca are better represented in the collection (n = 36) and are described by abj. While abp now applies to most copepods, it may be that it also applies to ostracods, arachnids and scorpions. The future will teach us. The results further show that among the spiralia only some mollusk species accelerate metabolically, see Fig 8. Ray-finned fish (n = 206) sport a wide range of acceleration factors, see Fig 9, but extreme forms of acceleration are confined to the Otomorpha, the Paracanthomorphacea and the crown groups of the Percomorphaceaei only. The coupling between the amount of acceleration in mediterranean perches with the spawning season, as reported in [39], shows that, apart from evolutionary aspects, ecological ones are involved as well and the two aspects cannot be fully separated. These examples beautifully show that the occurrence of acceleration is far from random.
Energy conductance, one of the two parameters that are affected by metabolic acceleration, controls reserve mobilisation, so dominates the incubation (or gestation time), since eggs start their development as a lump of reserve. We have several cases with data for embryo development in combination with post natal development, which show that energy conductance remains constant before and after birth. We met one convincing case, the Asian freshwater leech, Barbronia weberi [43], where energy conductance makes a jump up at birth. There are quite a few cases, however, where incubation time is under-estimated. These cases do not have data on embryo development, so we cannot be sure if energy conductance also makes a jump here, or that the start embryo development is delayed. The latter might be due to a variety of reasons. Most mammalian embryos, for instance, have a period to prepare for growth during which the fetus does not increase in size. The onset of growth is typically quite clear, since consistent with DEB expectations, structural length starts to increase linearly [44, 45]. Data on embryo development is relatively scarce.
The colors in the taxa names of Figs 7, 8 and 9 reflect the range of values of metabolic acceleration that were found in the various taxa. Although over 1000 entries is very large for a database of this type, compared to the 10 million existing animal species, it is close to nothing. Thus we cannot assume that the species in the collection are fully representative for the species in nature. Moreover, the number of species in each taxon, both in nature and in the database, varies enormously; some have just a single species. It is already an accomplishment to indicate the ranges this way, and we will not be surprised if insights on which taxa accelerate change somewhat as the collection grows. Since the database is online and freely accessible, the reader can easily check the values of metabolic acceleration for each species in each taxonomic group using the free package AmPtool (we refer the reader to the online manual for how to do this).
This study represents a large scale application of a general theory for metabolic organization of living organisms: the Dynamic Energy Budget theory. Although DEB theory applies to all organisms, the AmP collection only deals with animals. The reason is that animals eat other organisms, which do not vary much in chemical composition. As a first approximation, their environment can be characterized by the variables food availability and temperature. This characterization is hard to make “complete” for other organisms, which hampers comparison. And comparison is the most useful asset of this collection.
We contend that animal species can be compared on the basis of DEB parameters and that this offers a tractable means to study animal biodiversity in an ecological and evolutionary context. Moreover, by being mechanistic (= based on first principles), DEB models interpret data, rather than just describe it. They can therfore reveal inconsistencies in data and predict un-measured properties of species as functions of parameters.
We present and discuss how DEB parameters can be extracted from eco-physiological data: the AmP approach. The two associated software packages, DEBtool and AmPtool, are freely available via GitHub and have online user manuals. We demonstrate the applicability of DEB theory, by showing that it is possible to extract DEB parameters for animal species even when there is little data. We evaluate goodness of fit with respect to data completeness per species and overall the models fit data well. We found that a family of related DEB models, which share the same 14 DEB parameters, are needed to capture the diversity of life-cycles in the animal kingdom. A main metabolic feature which distinguishes the life-cycles is that some groups have ‘metabolic acceleration’, which has links with larval stages. We present the latest evolutionary overview on which groups were found to have metabolic acceleration. Knowledge gaps are highlighted.
The AmP project was initiated at 2009/02/12 and meanwhile 125 authors contributed with entries. It has 1035 entries at 2018/03/12, including all larger phyla and all chordate orders, and is both the smallest as well as the largest database of this kind, since it is unique. We expect that it will remain unique for a long time to come, in view of the huge amount of effort to arrive at the state we are presently in and because we think that DEB models will not have alternatives with matching generality, simplicity and realism. We hope that this study motivates the scientific community to contribute and use the AmP collection.
|
10.1371/journal.pcbi.1000962 | Evolution of a Signaling Nexus Constrained by Protein Interfaces and Conformational States | Heterotrimeric G proteins act as the physical nexus between numerous receptors that respond to extracellular signals and proteins that drive the cytoplasmic response. The Gα subunit of the G protein, in particular, is highly constrained due to its many interactions with proteins that control or react to its conformational state. Various organisms contain differing sets of Gα-interacting proteins, clearly indicating that shifts in sequence and associated Gα functionality were acquired over time. These numerous interactions constrained much of Gα evolution; yet Gα has diversified, through poorly understood processes, into several functionally specialized classes, each with a unique set of interacting proteins. Applying a synthetic sequence-based approach to mammalian Gα subunits, we established a set of seventy-five evolutionarily important class-distinctive residues, sites where a single Gα class is differentiated from the three other classes. We tested the hypothesis that shifts at these sites are important for class-specific functionality. Importantly, we mapped known and well-studied class-specific functionalities from all four mammalian classes to sixteen of our class-distinctive sites, validating the hypothesis. Our results show how unique functionality can evolve through the recruitment of residues that were ancestrally functional. We also studied acquisition of functionalities by following these evolutionarily important sites in non-mammalian organisms. Our results suggest that many class-distinctive sites were established early on in eukaryotic diversification and were critical for the establishment of new Gα classes, whereas others arose in punctuated bursts throughout metazoan evolution. These Gα class-distinctive residues are rational targets for future structural and functional studies.
| Proteins evolve new protein-protein interactions through changes to their residues. Many residue changes are harmful because they disrupt important existing interactions and functions. The more interactions a protein participates in, the more difficult it is to make changes that are not harmful to the protein. And yet, proteins with many existing interactions are also likely to evolve new functions or new interactions. How does evolution occur in the context of a well-constrained protein with many interactions? We studied the heterotrimeric G protein subunit Gα, a multi-functional protein that acts at the nexus between receptors responding to extracellular signals and the cytoplasmic proteins driving the response within the cell. The Gα subunit participates in numerous interactions that have constrained much of Gα evolution; yet Gα has diversified into four functionally specialized classes. We developed an approach that identifies key residue changes important to the evolution of Gα functionality and class, and gained insight into the types of residue changes that occurred both early and late in the evolution of Gα function. By studying these critical residues in Gα we can de-couple the many functionalities of this signaling nexus.
| How is functional novelty generated when a protein is highly constrained by its many interactions with other proteins and by its critical role in the cell? In these proteins, new mutations are likely to have deleterious consequences by disrupting some important function within the cell due to the high probability that the mutation interferes with at least one of the many interactions. The Gα subunit of the heterotrimeric G protein complex is a classic example of a highly constrained family of proteins. Heterotrimeric guanine nucleotide binding proteins (G proteins) serve as physical couplers between cell surface 7 transmembrane (7TM) G-protein coupled receptors (GPCRs) and downstream targets known as effectors. As such, they are critical for signal transduction in eukaryotes and act as a nexus of extracellular signaling and intracellular changes. The Gα subunit, therefore, is ideal for understanding how functional novelty arises when a protein that is highly constrained evolves.
G proteins have three subunits – Gα, Gβ and Gγ. In humans, there are 21 Gα, 6 Gβ and 12 Gγ subunits, which can be combined into many possible heterotrimers [1]. The human G protein signaling pathway is diverse and complex with approximately 850 GPCRs and dozens of G protein effectors (Jones and Assmann, 2004). These complex interactions mean that changing any residue of a G protein may have profound pleiotropic effects (Figure 1). For example, a promiscuous Gα subunit may interact with dozens of receptors and effectors [2], thus any mutation resulting in novel receptor or effector interactions potentially impacts many signaling pathways and can disrupt other interactions such as heterotrimer formation. The Gα subunit also has endogenous enzymatic activity, GTP hydrolysis, which puts further mechanistic constraints on the protein structure, and also drives a functional role where Gα acts as a “timer” with the intrinsic and regulated hydrolysis activity controlling the length of time the signaling pathways are activated as well as the amplitude of the response. Gα subunits are further constrained because they must cycle through multiple conformational states; any alteration of these states can disrupt the function of the G-protein and its interactions. The Gα structural core contains nucleotide-binding domains and switches that establish the basal, active and transition-state conformations. All Gα subunits bind GDP and GTP within a nucleotide pocket comprised of structural elements called a P-loop and an NKxD motif (Figure 1, center). The basal state occurs when Gα is bound to GDP, driving switch conformations compatible with interactions to Gβγ subunits. Nucleotide exchange of GDP for GTP generates switch conformations that define the active state and form an interface for target downstream effectors while driving the dissociation of Gα from Gβγ. The transition state for nucleotide hydrolysis is a third conformational state which is only recognized by a subset of interactors involved in regulation of G protein signaling. The Gα amino terminus, which is defined by an extended helix that affects associations with the Gβγ dimer, is involved in delimiting the subunit to the membrane through covalent attachment to lipids, and associates with a GPCR. Certain residues at both the amino and carboxyl termini of Gα interact with the activated receptor and are involved in nucleotide exchange [3], [4].
Besides receptors and effectors, a great many other proteins interact with the Gα subunit to control the activation state (Figure 1). Gα subunits are regulated by molecules that control its activation by acting as guanine-nucleotide exchange factors (GEFs) and guanine-nucleotide dissociation inhibitors (GDIs), and its deactivation by acting as GTPase activating proteins (GAPs). Thus, the surface of Gα evolved multiple, specific protein-protein interaction interfaces – such as those for the Gβγ dimer, the receptor, and the cognate effectors or regulators – many of which were partially or completely overlapping. The complexity of the Gα surface means that pleiotropic effects would most likely accompany any single mutation.
Given these enormous constraints, how did Gα evolve from a single ancestral subunit to form the four main classes in humans (G(io), G(q), G(s), G(12)) with multiple subtypes (Figure 2), each with distinguishing sets of sub-functionalities? Gene duplication clearly provided the raw genetic material, but how these nascent duplicates acquired class-distinctive functionality is unclear [5]. A confounding factor is the sporadic emergence of interacting proteins throughout evolution (Figure 2). Three new developments enabled us to answer these questions. First, plants, in contrast to animals, have a greatly simplified G-protein signaling pathway [6] thus providing a working structure of an ancestral-like Gα subunit. Second, there is now a wealth of comparative genomic sequence data to track over evolutionary time how and when new functionality was added to the ancestral Gα subunit [7], [8]. Third, there are now atomic structures of Gα in three conformational states, in its heterotrimeric complex, as well as co-resolved with several different interacting proteins. These structures allow us to place the evolutionary changes that we observe into a spatial context. These spatial data then reveal which protein interfaces or conformational states provide the evolutionary pressure driving the emergence of class-distinctive amino acid changes.
To understand how the functional diversity of extant Gα subunits arose from a single ancestral core, we need to know how the intermolecular interactions of this signaling network constrained the evolution of structure to a set of core sub-functions associated with all the subunits, and how differentiated structural elements drove the emergence of unique sets of sub-functions within subgroups of Gα subunits. Typically, this type of analysis begins with a deduction of the ancestral structure along with ancestral core functionalities, followed by an analysis of retained modifications as subunits duplicated and diverged throughout class evolution [9]–[11]. This approach by itself is recalcitrant to dissecting structure-function relations in a signaling nexus like Gα because this large Gα family has members containing both partially overlapping and non-overlapping protein-protein interaction interfaces as well as multiple distinct conformations. For example, many interfaces are dependent on the nucleotide-bound state.
We developed a broadly-applicable, synthetic approach for identifying key functional sites in Gα using structural data from mammalian Gα subunits and sequence data from across the diversity of eukaryotes. We used mutual information theory [12], [13] to select functional sites and phylogenetic analyses to show when and how the ancestral Gα diverged. Mutual information theory is a statistically-robust method for identifying the subset of sites most critical to the preservation of the functional core of Gα and those evolved sites important to diversification among subclasses of Gα (class-distinctive sites). We used this strategy to select evolutionarily important sites, setting criteria to automatically select sites that are uniquely associated with the functional divergence of a single Gα class and therefore likely arose from modifications to parental functionality after gene duplications. We used the atomic structures of Gα complexes to place our class-distinctive sites in a three-dimensional context and to identify sources of constraints on certain class-distinctive sites. We traced changes in these class-distinctive sites over evolutionary time to show when and how each of these functional Gα classes emerged.
The initial impetus was to determine the structural requisites for Gα class-specific functionality to enable regulation of activation/deactivation, coupling, and specificities. However, given the broad and deep genomic resources available, the approach described here is applicable to any protein that is a member of a gene family that underwent divergence through multiple, closely-spaced gene duplications, such as phospholipase C proteins, kinases, GPCRs, etc. Our analysis yielded several surprising results regarding Gα. For example, class-distinctiveness within the functional core was conferred by relatively few sites per class. A closer look at these sites within a class revealed unique features, functions, and interfaces of that class. Class-distinctive sites were found to impact all Gα classes and functionalities in addition to protein-protein interactions, such as the nucleotide binding properties that control signaling pathway dynamics. We used these data to propose explanations for several intriguing questions about Gα functional divergence and to propose sites that are rational targets for generating class specific mutations.
We applied mutual information theory to 14 of the 17 vertebrate Gα subtypes listed in Table S1 to identify class-distinctive sites that contribute to functional differences among subgroups of Gα subunits. A robust multiple sequence alignment (MSA) (Text S2) was achieved by seeking consensus from sequence alignments generated by different MSA programs and also by structural comparisons of different Gα gene family members (see Materials and Methods).
We identified 106 invariant and 59 class-distinctive sites from an MSA of 58 mammalian Gα sequences encompassing all 4 major classes (Figure 3). Our criterion for labeling a site in the alignment as class-distinctive was that it had an invariant amino acid value in sequences of three of the Gα classes and a different amino acid value in sequences from the fourth Gα class (as defined using a reduced amino acid alphabet, see Materials and Methods). This criterion limited the analysis to those sites that reflect a modification within a single, given class of a parental sub-functionality after a gene duplication event. Our more restrictive criteria, as compared to the earlier sequence-based analyses on the Gα family [14], [15], allowed us to immediately hypothesize that each class-distinctive site contributed to a unique functionality of the given Gα class. A corollary is that any Gα class for which class-distinctive sites could not be identified would imply a Gα class that had conserved parental functionality without modification. This was not the case for mammalian Gα as we found class-distinctive sites for all 4 classes (14 G(io), 10 G(q), 16 G(s), and 19 G(12) sites; these sites are labeled with an ‘I’, ‘Q’, ‘S’, and ‘2’, respectively, in Figure 3). The distinct amino acid value (designated ∂ – distinct) was not required to be absolutely conserved within all sequences in the distinctive Gα class, thus allowing for sub-class variation at that site. In our initial analyses, we defined the conserved amino acid value (designated η – not distinct) to be invariant among all sequences in the remaining 3 classes—implying that these η amino acids were functionally constrained in the ancestor. Sites with different evolutionary histories are apparent in Figure 3. Some sites have a single ∂ amino acid value for all sequences within a class, or distinctive values in only a subclass or even in just a single sequence. At some sites, however, there is more than one ∂ amino acid value, implying subclass variation. There was no penalty placed on the occurrence of η amino acids within the distinctive class, allowing a site to be ancestral-like early in the evolution of a given Gα class but then later acquiring class-distinctness. Table S2 summarizes the class-distinctive sites, their η and ∂ residues, and their evolutionary histories. The class-distinctive sites are displayed on the Gαi1•Gβγ heterotrimeric complex structure in Figure S1 in both space-filling and cartoon rendering for relative positioning of the distinctive sites from different classes. (A comparison between the sites identified here and the evolutionarily important sites identified by a different method – Evolutionary Trace – is presented in Text S3).
Class distinctiveness at a few positions was not readily apparent because of the stringency of the η residue criterion. A case in point – the residue located at position Y261 in human Gαq, a residue flanked by two G(q)-class distinctive residues (“TYP” in Gαq in 4th row of alignment in Figure 3); this begs the question of why this position was not originally classified as G(q)-distinctive. This is because one subtype in the G(io) class, namely Gαt1, has an amino acid value of H instead of the η value of N, thus precluding designation as class-distinctive using the given stringent criteria. To optimize the utility of the class-distinctive sites, we looked for neighboring residues that would contribute specificity to the new sub-functionality gained with the class-distinctive site, but which may have also independently diverged in a second class and would not, therefore, have been identified in our first analysis. These sites are analogous to position Y261 in Gαq. We used a contact distance of less than 5 Å between the two residues in the active state crystal structure to define which sites were neighbor to a given class-distinctive site. We looked for variation in sequences in the distinct class to select neighboring sites that likely contributed to the same specificity associated with the class-distinctive site. We limited variation to within one additional class, otherwise it was impossible to confidently assign one residue as the η residue. In this second level of scrutiny, we identified 16 more class-distinctive sites (designated d in Figure 3). We summarize these d sites in Table S3 and indicate the neighboring ∂ class-distinctive sites that flagged the second round of analysis. The core residues we identified, including invariant (106), ∂ (59) and d (16) class-distinctive sites, encompass approximately half (46–52%) of the total number of residues in the Gα subunit.
Functional regions are enriched with both unchanging core residues and evolving class-distinctive sites (Figure 3); regions that are less critical for the known functionality of a particular subtype typically lack class-distinctive sites. For example, residues at both termini that comprise the GPCR coupling interface are enriched with class-distinctive sites (20 ∂ class-distinctive sites [plus 4 d sites] of 66 residues). Similarly, switches I, II, and III are also enriched for ∂ and d class-distinctive sites (11 [plus 2] of 44). Three of the most functionally important regions of a Gα subunit are the GPCR binding interface, the α5 helix with its β sheet enclosure, and the 3 switches. The GPCR, through its interactions with the Gα subunit, determines which extracellular signal is being received and which pathway will be stimulated by that signal. The α5 helix and surrounding residues are critical for receptor-mediated nucleotide exchange. The switches are critical for interactions with target effectors, GEFs, GAPs and GDIs that affect the response and state of the Gα. In most cases, changes at these sites are deleterious. The evolutionary shifts at these critical sites, however, suggest a fundamental alteration in the function for that Gα class.
After duplication, each Gα gene diverged by evolving class-distinctive sites in a subset of – but not all –functionally-important regions (Figure 3). Since a different subset of functional regions were modified within each gene subfamily, the set of regions selected for evolving class-distinct functionality become characteristic for that gene or gene subfamily. For example, G(12) is unique in that it has class-distinctive sites in switch I. In contrast, both G(s) and G(12) have distinctive sites in switch II. Switches I and II are involved in binding the Gβγ heterodimer [16], [17] and other proteins such as regulators of G protein signaling (RGS) that stimulate Gα GTPase activity [18], [19]. A change in a switch commonly adds or removes critical contacts between the Gα and its effectors or regulators, suggesting that the changes in switch I of G(12) altered the interactions between G(12) and some of its binding partners, potentially regulatory proteins (see below). Analogously, class-distinctive sites evolved in switch III and the upstream region in G(q) and G(s), with G(io) containing d sites that also have ∂ residues in G(q) or G(s) subunits. However, there are no G(12)-distinctive sites located in switch III. Switch II (helix 2) and the region upstream of switch III (helix 3 and loop) is an interface for cognate effectors [20], again suggesting altered interactions in these Gα classes, but with effectors this time, allowing for the partially overlapping interfaces of regulatory and effector proteins on the Gα subunit. The GPCR coupling region, distributed over both termini, contains class-distinctive residues from all 4 major classes. These differing patterns of class-distinctive sites between switches and the GPCR interaction region are consistent with the coupled receptor and effector class specificity noted by Lichtarge et al. [14], but also indicate that natural selection exploited Gα as an existing signaling nexus by independently modifying individual regions associated with sub-functionalities (such as switch I) so that new connections in the network between effectors, regulators and receptors formed. These changes to different subsets of functionally-important regions within the Gα subunit ultimately resulted in new proteins with altered function and in new class-specific signaling pathways.
Gα subunits are not mere scaffolds for protein-protein interactions; they also affect signaling dynamics. We propose that control of GPCR based signaling pathways occurs through sequence-based modifications to Gα that indirectly affect nucleotide binding by directly affecting interactions with regulators such as GPCRs, GDIs, GAPs and GEFs. The α5 helical region, discussed below, has been shown to be important for receptor mediated exchange [21]. The α5 helical region contains class-distinctive sites from all classes except for G(q). Other regions are also involved in nucleotide binding. G(q) subunits are unique because they contain a class-distinctive site in the P-loop associated with nucleotide binding and because G(q) subunits have an undetectable basal nucleotide exchange rate [22]. Directly modifying nucleotide exchange properties within the different classes throughout evolution enhances the functional role of Gα as a “timer” controlling the length of time of activation of the different signaling pathways.
We hypothesize that the evolutionary patterns associated with a small set of class-distinctive sites within these largely autonomous functional domains of Gα predict residues that are critical for the functional specificity of these domains within each Gα class. In the following analyses, we test this hypothesis and link these evolutionary changes with class specific functions showing how this analytical framework can explain several conundrums regarding the structure and function of specific Gα subunits. We will explain (1) how functional specificity evolved, (2) how Gα subunits evolved class-specific functionality in their active state without affecting their inactive state, (3) how different but structurally-related Gα subunits evolved opposing functional outcomes, and (4) how new functionality evolved by modifications to residues participating in intramolecular interactions – versus intermolecular interactions – thereby controlling activation of the Gα subunit, and (5) how Gα diversified throughout metazoan evolution within and between functional classes.
All three G(q) subtypes included in our study, Gαq, Gα11, Gα14, acquired two G(q)-distinctive sites that we propose are key to determining the specificity of the interaction with G protein-coupled receptor kinase 2 (GRK2). GRK2 inhibits GPCR signaling by phosphorylating activated GPCRs [23], and also by sequestering Gβγ and G(q) subunits through its pleckstrin homology (PH) [24] and RGS homology (RH) [25] domains, respectively. The atomic structure of an activated Gαi/q chimera and Gβγ in complex with GRK2 [26] revealed the structural elements by which G(q) subunits are sequestered. In this complex, the RH domain of GRK2 interacted with switch II and an adjacent helix in Gαi/q while the N-terminal helix of Gαi/q – the domain inherited from Gαi1 – was disordered (Figure 4A). G(q) family subunits have no distinctive residues in switch II to distinguish this family from members of the G(io) class. However, G(q) family subunits do have two G(q)-distinct residues in the helix bordering switch II that formed part of the interface. Gαq residue T260, labeled as the 9th G(q)-distinctive site in Figures 4A and 4D, formed a hydrogen bond with a GRK2 residue in the structure. P262, G(q) site 10, was found to pack into a hydrophobic pocket formed by GRK2 and Gαq residues. Tesmer et al. [26] reported that GRK2 binding to G(q) subunits was eliminated with a P262K mutation, which corresponded to a ∂ to η mutation at G(q)-distinctive site 10, and identified residues 261–263 as a specificity determinant region [26]. Residue Y261 was discussed earlier and is a d site with ∂ residues in both G(q) family members and in Gαt1 of the G(io) family. The role of G(q)-distinctive site 9 (T260) in contributing to specificity determination has not previously been recognized or verified experimentally.
We propose that G(q)-distinctive sites evolved to drive specificity of G(q) interactions to GRK2 but not p63RhoGEF, a G(q) specific effector that activates the small GTPase RhoA [27]–[29]. The atomic structure of p63RhoGEF complexed with activated Gαi/q [30] revealed this interface contains no direct interactions with G(q)-distinctive residues (Figure 4B). In addition, the modeled heterotrimeric G-protein complex containing Gαq (Figure 4C) revealed a parental interface on Gαq for the Gβγ heterodimer. At present, only the Gαq•GRK2 interface appears to constrain G(q)-distinctive sites 9 and 10.
The Gα12/13•p115RhoGEF interface is dense with G(12)-distinctive sites (Figures 5A, 5B). Class-distinctive sites, analogous to those in the interaction of Gαq with GRK2, contribute significantly to the specificity of interactions between the G(12) subunit family and p115RhoGEF. The G(12)-distinctive sites, however, lie in the switches, which are regions sensitive to the bound nucleotide. In contrast, the G(q)-distinctive sites driving the Gαq specificity lie in a helix neighboring switch II, a region not sensitive to the state of the bound nucleotide. We hypothesize that the G(12)-distinctive sites confer effector and regulator specificity in the active and transition states (Gα13 in Figure 5A and Gα12 in Figure 5B), yet do not disrupt interactions with Gβγ in the inactive state (Gα12 in Figure 5C) even though the sites are in switches I and II, regions important for binding both Gβγ and p115RhoGEF.
The G(12) story is complicated by significant differences in the functional outcomes that result when the two different vertebrate G(12) subunits interact with p115RhoGEF. Specifically, Gα13, but not Gα12, activates RhoA when in complex with p115RhoGEF. Several of the G(12)-distinctive sites in switch II, which form part of the interface, show subtype variation within the gene family. This subtype-specific variation at G(12)-distinctive sites in switch II may contribute to this G(12) subtype difference in effector functional outcome (below).
P115RhoGEF is a G(12) specific effector that binds members of the G(12) family in a nucleotide-dependent manner and acts as a GAP toward Gα12 and Gα13 [31], [32]. P115RhoGEF also stimulates GEF activity on Rho GTPase when bound to Gα13, activation exerted via its DH and PH domains [31], [33]. The structure of the N-terminal domains of p115RhoGEF bound to an activated Gα13/i1 chimera (Figure 5A) suggested the GAP activity was associated with an N-terminal βN-αN hairpin element that was conformationally distinct from canonical RGS domains, which had also been shown to possess GAP activity toward Gα proteins [20]. Mapping our class-distinctive sites onto the structure of the Gα13/i1•p115RhoGEF complex revealed an interface covering switches I and II of the Gα13 subunit, a region that possesses 7 G(12)-distinctive sites within these two switches. One (site 10, Figures 5A and 5D) of three G(12)-distinctive sites in switch I made a direct contact to the βN-αN structural element of p115RhoGEF. Mutating the ∂ amino acid value at site 10 (K204) diminished binding of p115RhoGEF to Gα13 [34], [35], verifying the importance of this G(12) site in the evolution of G(12) functional specificity. Chen et al. [20] also noted that R201, which is the ∂ amino acid in G(12)-distinctive site 9, acted as a tether between switch I and a Gα13-unique helical insert within the α-helical domain, suggesting that some distinctive sites may be important for switch conformation and intra-domain contacts rather than direct interactions at an interface.
The Gα13/i1•p115RhoGEF structure revealed that the RGS-like box of p115RhoGEF bound to the Gα effector interface (switch II) rather than the typical regulator interface of Gα13. Chen et al. [20] proposed that, based on the effector-like interactions between switch II and the RGS-like box, Gα13 may act indirectly on the DH and PH domains of p115RhoGEF through the RGS-like box to exert the GEF activity on RhoA. We show that two (sites 12 and 13) of three G(12)-distinctive sites in switch II made direct contacts to residues of the RGS-like box of p115RhoGEF. Distinctive sites 11 and 12 in switch II show subtype variation, with the η amino acids evident in Gα13 at these sites (Figure 5A) and the ∂ amino acids in a model of Gα12 [36] bound to p115RhoGEF (Figure 5B). P115RhoGEF also acted as a GAP toward Gα12 [32], but Gα12, unlike Gα13, did not mediate RhoA activation [31]. It is possible that the subclass sequence variation at these two sites account for this subtype specific loss of p115RhoGEF activity, but the story may be more complex (see Text S4 for additional discussion).
Although these G(12)-distinctive sites confer specificity to the interaction with p115RhoGEF when the Gα subunits are in the active/transition state (Gα13 in Figure 5A, Gα12 in Figure 5B), the G(12) subunits still bind Gβγ in the inactive state (Gα12 in Figure 5C). The switches in the Gα12•GDP conformation form a ledge with Gβγ binding to the side of the ledge shaped by conserved residues (Figure 5C, right view and inset). The G(12)-distinctive residues (sites 10–13) are on the opposite side of the ledge, positioned away from the Gβγ interface, and thus do not disrupt G(12) family members binding to the Gβγ heterodimer. Gβγ is not the only macromolecule which binds the inactive conformation. GoLoco motifs found in several proteins also bind the Gα•GDP conformation (see below and also Figure 1), but GoLoco motifs bind in the concavity formed by the G(12) sites 10–13 and the main Gα structure (Figure 5C, right view and inset). Several of the G(12) ∂ residues in the switches are positioned to discriminate among molecules that utilize this surface (data not shown), emphasizing the pleiotropic effects that arise whenever shifts are made in a molecule highly constrained by so many interactions.
Two Gα subunits interact with adenylyl cyclase (AC) with opposite functional outcomes. Gαs stimulates AC, whereas Gαi inhibits AC. Comparisons of the crystal structures of Gαi1•GTPγS [37], [38] with those of Gαs•GTPγS [39], and the Gαs•GTPγS•VC1•IIC2•forskolin [40] complex prompted Sunahara et al. [39] to suggest that the interface on Gαs for AC (Figure 6A), which is comprised of switch II (Figure 6B “sw II”) and its neighboring loop (Figure 6B “neigh”), was similar in sequence but dissimilar in shape to the same region on Gαi1 (Figure 6B; Gαs, gray cartoon; Gαi1, green cartoon). They concluded that disparately-shaped binding surfaces, not sequence differences, drove the distinct functional outcomes [39], [40]. With their model in mind, we noted three G(s)-distinctive sites (sites 7, 8, and 9 in Figures 6A and 6D) are in or near switch II. Site 9 is the only one of these sites that has a direct interaction with AC (Figure 6B), but a mutation of three residues at the interface that also included site 9 resulted in only a threefold reduction in AC activation [39], [41], consistent with the proposed role of conformational differences, not sequence differences, as the source of discrimination between Gαi1 and Gαs.
There are two other G(s)-distinctive sites near the interface: G(s) site 11 that lies in the neighboring loop that forms part of the interface, and site 13 that lies in a loop abutting the interface (Figure 6B “abut”). The η amino acid in Gαi1 at G(s) site 13 is a solvent-exposed lysine, whereas the ∂ amino acid in Gαs at the same site is a buried histidine. Adjustments to the backbone in the abutting loop allow for these different side chain orientations (Figure 6B) in the two Gα subunits. The abutting loop is different in sequence and length between G(s) and G(io) family members, which contributes to the conformational differences in this loop between the two families [39]. In contrast, the loop neighboring switch II containing G(s)-distinctive site 11 is similar in sequence and length between G(s) and G(io) family members [39], except for the single G(s) class-distinctive site, even though it adopts slightly different conformations in the two family members. Conformational differences in this neighboring loop may be driven both by sequence changes at site 11 – the bulky phenylalanine (η amino acid) in Gαi1 is shifted in position from the leucine (∂ amino acid) in Gαs [39] – and by the conformational changes in the abutting loop. Thus, conformational differences in these two loops leading to the opposite functional outcomes between Gαi1 and Gαs are potentially driven by the class-distinctive sites in G(s) subunits. Both G(s)-distinctive sites 11 and 13 were identified by Sunahara et al. [39] in a structural analysis as critical components driving structural differences, which is consistent with earlier mutational studies replacing entire loops in the two Gα families [41]. This analysis suggests that G(s)-distinctive sites could influence the conformational changes that affect the interactions of Gα subunits with AC.
Though there are two d sites in switch III of G(io) family members, there are no G(io)-distinctive sites in the switches of all three subtypes (Gαi, Gαo, Gαt) of the G(io) family members (Figure 7A, left view); all of the G(io)-distinctive sites lie on the opposite face of the molecule (Figure 7A, right view) or are buried. The lack of G(io)-distinctive sites on the switch side of the molecule implies that this family of Gα subunits has maintained the parental functionality in all switches and, therefore, continues to interact with the primordial set of effectors and regulators. While the interface remained ancestral, new effectors – such as GoLoco motifs [42] (Figure 7A, left view) or PDEγ [18] – that utilized surface areas of the parental structure emerged in metaozoans.
Most of the G(io)-distinctive sites on the opposite face of the subunit (Figure 7A, right view) tend to lie in regions associated with binding to the GPCR and with GPCR-driven GDP release (7 sites out of 14 total G(io)-distinctive sites) implying modifications to GPCR specificity and Gα nucleotide binding properties. An N-terminal peptide from Gαt1 of the G(io) class was reported to competitively inhibit Gαt1-rhodopsin interactions [43]. This N-terminal region contains class-distinctive sites from all four classes (Figure 3). A site-specific fluorescence labeling study reported the greatest receptor activation induced intensity changes and emission shifts – indications of a less aqueous accessible environment – at 3 residues within the Gαi1 N-terminal helix [44]. Two of these three residues are class-distinctive sites: G(q)-distinctive site 2 and G(12)-distinctive site 3 (Figure 3). Furthermore, another study identified G(s)-distinctive site 1 as being a key determinant of GPCR selectivity in G(q) family subunits [45]. In yet another study that further refined the GPCR contact surface on Gα subunits, we find the first three G(io)-distinctive residues (sites 1, 2 and 3) lie within the 10-amino acid region in the N-terminal helix identified by covalent cross-linking as a site of contact on Gαt1 by the GPCR rhodopsin [46] (Figure 3). We hypothesize these class-distinctive sites are key determinants in Gα-GPCR coupling, although subtle and cooperative interactions are also involved [47].
Similarly, previous studies found key sites of GPCR interaction on the C-terminal region of Gα subunits [48]–[52], a region with several class-distinctive sites. However, several of the sites important for GPCR specificity in the C-terminus rapidly evolved and are thus unique to each Gα subtype.
In contrast to Gα-GPCR coupling in which class specificity was conferred by changing intermolecular interfaces, we hypothesize Gαi evolved class specific functionality by changing an intramolecular interface. Three G(io)-distinctive sites within a helix reported to undergo conformational shifts during activation are likely responsible for mediating that conformational shift during the GPCR driven release of GDP in a class specific manner (Figure 7B). Oldham et al. [21] proposed, based on their measured changes in mobility for residues within helix 5, that helix 5 rotates and translates during GPCR induced activation and in conjunction with the release of GDP. All three of the G(io)-distinctive sites in helix 5 were mutated [21]; mutations at G(io)-distinctive sites 12, 13 and 14 decreased rate of receptor-catalyzed exchange, especially site 14, while the mutation at site 13 also affected the basal exchange rate. In a different study, Kapoor et al. [53] reported that mutation V332A in helix 5 of Gαi1 increased basal exchange rates. This residue corresponds to d site 15 that lies between G(io)-distinctive sites 12 and 13. An additional G(io) site, site 11, that lies in β-strand 6 (Figure 7B) shows subtype variation within the G(io) class, although no experimental evidence yet links modifications at this site to class-specific functionality. Intriguingly, G(io) sites 12 and 13 also show subtype variation within the G(io) class, in stark contrast to the conservation evident at these same sites in the other three classes. These results are consistent with our proposal that the G(io) class of Gα subunits evolved unique properties for this conformational shift.
We proposed earlier that G(q)-distinctive sites evolved to drive specificity of G(q) interactions to GRK2 (Figure 4A) but not to p63RhoGEF (Figure 4B). In contrast to the Gαi/q•GRK2 interface in which 80% of the Gα residues within 4 Å of GRK2 are either invariant or G(q)-distinctive sites, only 50% of the Gα residues at the Gαi/q•p63RhoGEF interface have core functionality, either through invariance or G(q)-distinctive sites. This suggests the Gαq•p63RhoGEF interaction arose through de novo evolution of neo-functionality with the acquired utilization of residues that were not previously functional, rather than primarily through modifications to parental functionality like the Gαq•GRK2 interaction. This hypothesis is further supported by noting that the Gαq interface with p63RhoGEF, containing 30 residues, is twice the size of the Gαq•GRK2 interface, which contains only 15 residues. Therefore, evolving the Gαq•p63RhoGEF interface required fixing an additional 15 residues beyond those used in parental functionality, enough residues to warrant identifying this as de novo evolution of an interaction interface.
Gα subunits of G proteins are essential for signal transduction in all eukaryotes. As eukaryotes diversified and became more functionally complex, so did Gα subunits. Extant Gα subunits arose through multiple rounds of duplication and divergence [54]. How these gene duplicates functionally diversified, however, is not well understood. Because Gα resides at the nexus of many signaling pathways and interacts with many effectors, any change can have profound negative pleiotropic effects. How then do highly constrained proteins like Gα evolve the functional complexity we see today? We hypothesize that a narrow subset of class-distinctive sites has the evolutionary potential to confer class-distinctive function with minimal evolutionary cost. From a structural point of view, these sites are those that can mutate and shift the class functionality with a minimal deleterious effect on other aspects of the signaling nexus. At first blush, this idea that highly constrained sites are the ones that confer class specificity is counter-intuitive. Part of the explanation is simple; we found class specificity in the core because we seeded this analysis with the conserved sites within and between classes and avoided highly-labile sites because lack of conservation provides little information about the molecular evolution. Another reason we focused on conserved residues is because changes in these residues have known functional consequences, thereby making any observed class-distinctive change in these sites likely critical for class specific function. We identified 59 of these sites spread across the 14 mammalian Gα. In several instances, these class-distinctive sites associate with known class-specific properties. We also identified many more uncharacterized sites that likely play a role in the sub-functionalization of mammalian Gα. While we have probably not identified all of the residues important for class-distinctive behavior, we identified an important subset of these residues. Mutations at our selected sites will likely disrupt a class-distinctive functionality, but are not likely to be sufficient to confer a full gain of functionality. Other residues, both neutral and restrictive [10], [11], most likely occurred but would not have been identified by our approach.
Our analyses suggested several interesting evolutionary patterns. We showed how two changes at G(q)-distinctive sites determine the specificity of the GRK2 interaction with Gαq and how changes at G(q)-distinctive sites are effector specific, driving specificity of the Gαq•GRK2 interaction but not the Gαq•p63RhoGEF interaction. We also highlighted the role of G(12)-distinctive sites in the specificity of the Gα13•p115RhoGEF interactions. Two of these three examples illustrates how functional diversity within and between classes was driven by changes to parental functionality at class-distinctive sites, while the third example, p63RhoGEF, showed emergence of new functionality by utilizing previously non-constrained residues. In this case, it is possible that the interface evolved in two stages – originally the extended PH domain of p63RhoGEF could have bound to the parental structure of the switch region and the DH domain could have evolved contacts over time to a non-parental surface area (Figure 4B). This process would be mechanistically similar to that speculated for the phosducin interaction with the Gβ subunit [7]. We also showed how evolution can overcome the complexity of G protein interactions by producing structurally-related Gα subunits with opposing functional outcomes. For instance, we showed that by evolving class-distinctive sites that induced conformational changes in Gα, Gα proteins shift from inhibiting AC to stimulating AC. All of these changes affected only the active/transition states of Gα while leaving the inactive state intact and able to interact with the heterodimeric Gβγ complex (Figures 4C, 5C, 6C and 7C). Although two distinctive residues, G(12)-D site 10 and G(s)-D site 8 lie at the interface with the Gβγ complex and have the potential to confer specificity to the interaction of Gα with Gβγ, we are not aware of any published data suggesting there is specificity in this interaction. Finally, we illustrated how novel functionality evolves by variations at sites involved in functionally important conformational changes related to Gα activation rather than through evolution of new interfaces.
All four Gα classes were formed early in metazoan evolution. From the number of distinctive sites established in the lowest metazoans and the correlations of these changes with class-specific function, our data suggest that the four major Gα classes were established by the split with sponges, in agreement with two earlier studies of Gα evolution [55], [56]. Gα evolution is characterized by bursts of duplication and diversification followed by long quiescent periods [55] and this is also true for class-distinctive sites (Figure 8). For example, our data suggest that the evolution of the class-distinctive sites critical for the GRK2 interaction with Gαq (Figure 8B, G(q) sites 9, 10) occurred around the time of emergence of the G(q) class. However, the GRK2 interaction was not the only sub-functionality driving the emergence of G(q) as several other class-distinctive sites not likely involved in the Gαq•GRK2 interaction also appeared at the time of emergence (Figure 8B, G(q) sites 4, 5, 7), and other sites clearly became class-distinctive at later times (Figure 8B, site 1). For G(s) subunits, the G(s) sites associated with AC functionality were also established at the time of emergence of the G(s) class (Figure 8C, G(s) sites 9, 11, 13). At the same time, we see ∂ amino acids became fixed at G(s) sites 1, 15, and 16, sites which are structurally adjacent (Figure 6C, right view). This set of sites is in the GPCR coupling region. We speculate that the evolution of new GPCR specificity was linked to AC activation, resulting in a new signaling network. G(s) site 14 in helix 5, the helix associated with GPCR induced activation, also became distinctive with emergence of the G(s) class, potentially imparting new exchange properties to this signaling pathway. Two additional G(s) sites that currently are not correlated with any known function – G(s) sites 3 and 4 – were also established early, whereas several G(s) sites became distinctive later in evolution. We see similar patterns within the G(12) class.
Interestingly, G(io) sites 12, 13, and 14 – the 3 G(io) sites in helix 5 discussed above – show variance in early metazoans in G(io) and G(12) subunits but not in G(q) or G(s) subunits (Figure 8). This implies an exploration of nucleotide-binding properties in early metazoan ancestors or lineage-specific modifications in G(io) and G(12), but not the other two classes of Gα subunits. Lastly, we see that two G(12)-distinctive sites in switch II (sites 11, 12) possessed ∂ amino acids in the single invertebrate G(12) family member throughout early metazoan evolution but reverted to η amino acid values in Gα13 after a gene duplication that occurred with the emergence of vertebrates (Figure 8D). This accepted change went from ∂ to η – an unusual direction – rather than the canonical direction of η to ∂.
Our data show that the four Gα classes acquired class-distinctive sites throughout metazoan evolution, usually along with the evolution of an expanded or novel class-specific function. One particularly significant period for distinctive site acquisition was before nematodes split from the mammalian lineage (Figure 8) a time when the olfactory system greatly evolved. Interestingly, the second most significant period occurred during the emergence of vertebrates, when all four Gα classes experienced gene duplications leading to an explosion of Gα subtypes, a period when endocrine system complexity dramatically increased. Both olfaction and hormone signaling rely on G protein coupled signaling. This observation may mean that Gα diversification played a critical role in the morphological and physiological evolution of the modern vertebrate.
We propose that specific sequence changes that occurred early in the acquisition of class-specific functionality arose from modifications to parental functionality. Most class-distinctive sites were in regions that were already constrained by functional demands – such as the switches – where modifications to surprisingly few residues could complement existing functionality while simultaneously contributing to divergence. As suggested by Conant and Wolfe [57], the “new” function may have been a secondary property that was always present in the ancestor, similar to the property recently revealed for steroid hormone receptors [11]. Our observation that many of the class-distinctive sites arose from highly conserved residues and essential structural components suggests that gene duplication was essential for the diversification of the Gα. Reduction of the functional constraint on a new paralog following duplication allowed that copy of Gα to convert its secondary property into its primary. This is not, however, the sole mechanism for evolutionary divergence. Both the lack of sequence conservation in the Gαq•p63RhoGEF interface and the presence of residues with different evolutionary rates in the α-helical domain [15] imply divergence by evolution of neo-functionality of a previously unspecialized but highly constrained domain following gene duplication.
We believe that our comprehensive view of Gα evolution shows us the amino acid changes that allowed G proteins – despite the constraints put upon them by their myriad of interactions – to become functionally diversified proteins. With this view, we explained several conundrums regarding the structure and function of specific Gα. We produced a partial list (Table S2 and Table S3) of the sites that are likely to contribute to class specific function, and are important for the role of Gα as a signaling nexus. Translating the patterns of evolution at Gα class-distinctive sites into predictions for future structural and functional studies is the next challenge. We will achieve this by uncovering additional details in metazoans of class divergence and the acquisition of neo-functionality in Gα and also by defining the characteristics of the primordial Gα through analysis of the pre-metazoan plant and fungal Gα.
G-protein sequences were collected from the UniProtKB/Swiss-Prot/TrEMBL Knowledgebase [58] available on the ExPASy Proteomics Server (www.expasy.org) [59] using BLAST [60] and filtered for redundancy using the Ensembl Genome Browser (www.ensembl.org) [61]. Sequences were aligned using ClustalX [62] and adjusted using a T-Coffee alignment program [63] and finally by eye using known atomic structure data as a guide. The final multiple sequence alignment (MSA) contained 347 sequences. Four Gα classes and 16 subclasses were tabulated (Table S1). A subset of sequences was selected for distinctive site determination based on the following criteria: 1) must be mammalian, 2) must be at least human and rodent sequences available for every subclass included in the analysis (Gαt3 had only a rat sequence at the time of the original analysis), and 3) subtypes must not be highly divergent (e.g. this excluded Gαz and Gα15). Ultimately, a total set of 58 mammalian sequences from all 4 major classes comprising 14 subclasses were culled from the full MSA and used for the final analysis.
Gα sequences from several lower metazoans were included in the analysis reported here. Geodia cydonium, a marine sponge, and Ephydatia fluviatilis, a freshwater sponge, belong to the phylum Porifera. G. cydonium has three Gα proteins: a G(io), a G(q) and a G(s), while E. fluviatilis has five proteins that are clear progenitors of mammalian proteins. E. fluviatilis has a single G(q), G(s) and G(12) family member, but two G(io)-like members (a G(i) and a G(o)). There are four additional Gα proteins specific to the Ephydatia lineage and these were not included in our analysis here. All E. fluviatilis sequences used in this study are fragments missing the first ∼50 residues of the amino terminus.
The Gα family Caenorhabditis elegans (nematode) expanded greatly, with 21 Gα proteins in total, but only four subunits – G(o), G(q), G(s) and G(12) – are clearly related to the progenitors of mammalian proteins and are thus included in this analysis. The fruit fly, D. melanogaster belonging to the phylum Arthropoda, also has G(i), G(o), G(q), G(s), and G(12) members with three additional Gα subunits that are specific to the insect lineage. Strongylocentrotus purpuratus (purple sea urchin) is an echinoderm and is the last invertebrate considered in this analysis. Four S. purpuratus Gα sequences were analyzed: a G(i), G(q), G(s) and G(12). A number of gene duplications occurred between invertebrates and vertebrates and, given the current available sequences in the databases, it appears that most vertebrates possess a full complement of the 16 mammalian Gα subunits with some having taxa-specific subunits. Xenopus laevis (frog) was the model vertebrate organism included in our analysis. Four X. laevis Gα sequences were analyzed, annotated as Gαi1, Gαo1, Gαq and Gαs. Current data suggest that the ancestral plant had a single Gα subunit while most extant fungi have two or three Gα proteins. An annotated version of the MSA containing the 58 mammalian sequences used for determining the class-distinctive residues highlighted in Figure 3 and the metazoan sequences included in the evolutionary analysis in Figure 8 is provided in FASTA format (Text S2).
Mutual information can be used to measure the correlation between amino acid value and protein family for a set of sequences subdivided into families with different functional specificity (Basharin, 1959). Positions in the alignment which exhibit conservation within each family and variation between families have high mutual information. Positions that exhibit conservation between families (such as invariant residues) or variation within families (such as non-conserved residues) have low mutual information. This method was used by [64] to detect putative specificity-determining residues for paralogous protein kinases. In their study, mutual information was defined asWhere i is the position in the alignment, x the amino acid value, and y the protein family number. The summations are over all families in the alignment (y) and amino acid values (x). Pi (x, y) is the probability of finding amino acid value x at position i and in family y; Pi (x) is the probability of finding amino acid value x at position i regardless of family; and P(y) is the fraction of proteins belonging to family y. In our mutual information calculation, we subdivided our sequences into four families: G(io), G(q), G(s), and G(12). We also treated amino acid residues with similar side chains as identical, resulting in an amino acid alphabet of 15 values (G, A, V, I = L, M, P, F = Y, W, S = T, N, Q, C, K = R, H, D = E). In addition, we normalized the mutual information scores to the range [0.0,1.0].
Sites of interest were characterized as either invariant or class-distinctive. Invariant sites contained the identical amino acid values for the 58 mammalian Gα subunits from all four of the major animal classes. These residues likely were constrained early in Gα evolution and formed part of the primordial Gα core. While invariant sites are important for our understanding of the structural/functional aspects of Gα subunits, they do not contribute to an understanding of the evolution of Gα classes. Each class-distinctive site is occupied by an invariant amino acid (designated η – not distinct) in all sequences except for those of a specific functional or distinctive class. Within the distinctive class, sequences contain a different amino acid value (designated ∂ – distinct). In 33 of our 59 sites, the distinctive amino acid ∂ is conserved across all subtypes within the class. Of our 59 sites, 21 show subtype variation within the class. In 5 of our 59 sites, the ∂ amino acid was not conserved for a single subtype and the variation potentially occurred in a non-human sequence.
A single mutual information calculation simultaneously using all four Gα classes cannot identify our selective distinctive sites. Therefore, we used a series of six pair-wise mutual information calculations covering all possible pairs of Gα classes [G(io) vs. G(q); G(io) vs. G(s); G(io) vs. G(12); G(q) vs. G(s); G(q) vs. G(12); G(s) vs. G(12)], then scanned for patterns in the scores to identify distinctive sites (see Table S4). Invariant sites corresponded to positions with the lowest Ii (0.0) for all 6 calculations. Distinctive sites corresponded to positions with the lowest Ii (0.0) for all Gα pairs not involving the distinctive class and higher Ii (>0.0) for all Gα pairs involving the distinctive class.
By accepting any nonzero Ii in the calculations involving determination of distinctive sites, residue positions with a wide range of properties were only tolerated in the distinctive class. The sites had scores that ran from low Ii (0<Ii≪1) for all Gα pairs involving the distinctive class where the position was almost invariant with many η and few ∂ values, to residues with high Ii (0<Ii≤1) for all Gα pairs involving the distinctive class and containing only ∂ amino acid values in the distinctive class.
The stringency of criteria for designation as class-distinctive is a function of the amino acids permitted to evolve at that site. Allowing unrestricted evolution, that is any site can evolve to any of the 20 amino acids, would yield only 30 class-distinctive sites instead of the 59 sites identified using a reduced, and more evolutionarily plausible set (see Table S2). Although some substitutions within our reduced amino acid set could result in unaccounted functional changes (e.g. incorporation of a tyrosine phosphorylation site), some sites with known class-distinctive functionality discussed would not have been identified with a 20 amino acid set. We also included class distinctive sites that were identified using an evolutionarily likely set of possible amino acids from our phylogenetic and structural analyses (discussed in the RESULTS section).
We used the Evolutionary Trace Server (ETS) at http://mordred.bioc.cam.ac.uk/~jiye/evoltrace/evoltrace.html to identify evolutionarily important sites for comparison to our class-distinctive sites [65]. We used the identical Gα MSA as utilized for identification of class-distinctive sites, along with chain A of PDB ID 1GP2 for the mapping [17]. Using 10 evenly spaced partitions of our phylogenetic tree, we computed the trace using the TraceSeq and TraceScript algorithms as implemented on the ETS. This revealed the functional patches on the surface of these highly related proteins that reside in similar regions of Gα, regardless of functional differences.
The homology models of Gαq•GDP, Gα12•GDP and Gαs•GDP each used two partial structures as templates. The first template was an active conformation structure for the given class with the switch regions removed. Structures used as templates – after the removal of the switch regions – were (PDB ID) 2BCJ (Gαq), 1AZS (Gαs), and 1ZCA (Gα12). The switch regions in the three inactive state homology models were built using the switch regions from the inactive Gαi1•GDP (PDB ID 1GP2) as the template. Models were generated using InsightII (www.accelrys.com). Side chain rotamer conformations were selected that minimized steric hindrance upon complex formation with the Gβγ subunits from 1GP2. The model of the Gα12/i1•p115RhoGEF complex was based on structures of Gα12/i1•GDP•Mg2+•AlF4− (PDB ID 1ZCA) and Gα13/i1•p115RhoGEF complex (PDB ID 1SHZ). Gα12 from 1ZCA was used directly for complex formation with p115RhoGEF except for the adjustment of one side chain conformation to reflect the conformation evident in the complex structure of Gα13/i1•p115RhoGEF.
|
10.1371/journal.pgen.1001042 | The Fitness Landscapes of cis-Acting Binding Sites in Different Promoter and Environmental Contexts | The biophysical nature of the interaction between a transcription factor and its target sequences in vitro is sufficiently well understood to allow for the effects of DNA sequence alterations on affinity to be predicted. But even in relatively simple in vivo systems, the complexities of promoter organization and activity have made it difficult to predict how altering specific interactions between a transcription factor and DNA will affect promoter output. To better understand this, we measured the relative fitness of nearly all Escherichia coli binding sites in different promoter and environmental contexts by competing four randomized promoter libraries controlling the expression of the tetracycline resistance gene (tet) against each other in increasing concentrations of drug. We sequenced populations after competition to determine the relative enrichment of each −35 sequence. We observed a consistent relationship between the frequency of recovery of each −35 binding site and its predicted affinity for that varied depending on the sequence context of the promoter and drug concentration. Overall the relative fitness of each promoter could be predicted by a simple thermodynamic model of transcriptional regulation, in which the rate of transcriptional initiation (and hence fitness) is dependent upon the overall stability of the initiation complex, which in turn is dependent upon the energetic contributions of all sites within the complex. As implied by this model, a decrease in the free energy of association at one site could be compensated for by an increase in the binding energy at another to produce a similar output. Furthermore, these data show that a large and continuous range of transcriptional outputs can be accessed by merely changing the , suggesting that evolved or engineered mutations at this site could allow for subtle and precise control over gene expression.
| A major challenge in molecular genetics has been to understand how cis-regulatory information is integrated to determine the amount of transcript generated. The difficulty has been that there are a large number of variables (known and unknown) that combine through an extensive array of possible mechanisms. Differences in the affinity of a binding site for its cognate binder within the initiation complex are known to account for significant differences in promoter output, but data for the activity of binding site variants in vivo has been limited. Here, we were able to map the fitness of nearly all E. coli binding sites in multiple promoter and environmental contexts using a novel method that utilizes the sequencing power of a next generation DNA sequencer. These data for the first time show the phenotypic range and continuity of a nearly complete set of possible binding targets in vivo, and they are useful in our ability to understand the mechanism, evolution, and designability of gene regulation.
| While we have a reasonable understanding of the biophysical forces that determine the affinity of a transcription factor to its target sequences [1]–[4], we still have a poor understanding of how the affinity of a factor for a site affects the output of the promoter in which it sits. The major challenge is that these relationships are highly context dependent. A high affinity site tightly bound in isolation will have no function in that it will not affect the rate of transcription of a gene, whereas a low affinity site weakly bound in the context of the initiation complex will. More subtly, a single base pair difference in the spacing between sites can affect the function of those sites [5], [6]. Here, we attempt to better understand how binding site affinity and context relate to promoter output by determining the relative fitness of binding sites within specific variations of an engineered promoter in the bacteria Escherichia coli.
The engineered promoter that we use contains three binding sites: one for the transcriptional activator MarA [6], and another for the and the that are recognized by [7]. In the simplest thermodynamic model of transcriptional regulation in prokaryotes, the rate of transcriptional output varies as a direct function of the stability of the initiation complex [8]–[11]. The stability of the initiation complex in turn is dependent upon the cooperative binding of multiple DNA-binding transcription factors, each of which recognizes a degenerate set of sequences with different affinities [4]. The binding strengths of these sites are distributed such that there is a single optimal site that is bound with the highest affinity (the consensus site) and an increasing number of sequences that are bound with lower affinities as the sequences deviate from the consensus [1]–[3]. At some point the deviation becomes so great, that the site is no longer specifically bound and all remaining sequences have the same non-specific binding energy. The general assumption has been that the greater the affinity that the factor has for a site, the greater the occupancy at that site and the greater the probability that it will affect transcription [10]. This has only recently been tested for large libraries of sequences, and indeed much of the variance in expression can be explained by differences in binding site affinity [12]. Given this relationship, the distribution of binding energies for a factor defines the range of regulatory phenotypes that can be selected [2], [13], the number of possible DNA sequences that can be used to generate that phenotype, and subsequently the likelihood of a sequence of that strength evolving.
How multiple binding sites combine to determine the stability of the initiation complex is poorly understood, mainly because there are a large number of proteins that can cooperate to regulate transcription through a variety of mechanisms [9], [14], including direct stabilization or destabilization of the initiation complex through protein-protein interactions or occlusion [15], [16] or by perturbations of DNA structure that affect promoter-DNA binding [17], [18]. MarA has been shown to modulate transcription through multiple mechanisms depending on its binding context [6]. Here we use MarA as a Class I activator that increases the rate of expression by stabilizing interactions with the carboxy-terminal domain of the alpha subunit (CTD) [6], [9], [19]. The ordering, spacing and orientation of binding sites can also mediate transcriptional regulation [11], [20]. Differences in the spacing between the and the [5], [21] and between MarA and the have been shown to affect the rate of transcription [6].
Here, we examine the effects of varying a binding site on promoter output by measuring the relative fitness of binding sites in different promoter and environmental contexts. To do this we placed the tetracycline resistance gene under control of the MarA-activated promoter on the plasmid pBR322. We generated four libraries that contained different strength and MarA binding sites, to yield four varied energetic contexts for selection. By increasing the tetracycline concentration, we can change the range of selected viable transcriptional outputs. We competed variants within a library in liquid culture for 24 hours, and sequenced the competed population with an Illumina Solexa sequencer. Using this approach, we were able to map the fitness of a large population of binding sites in multiple promoter and environmental contexts relatively easily.
We generated four plasmid libraries that contained the tetracycline resistance gene (tet) under the control of a MarA-activated promoter with a randomized binding site. Each library contained a different combination of and MarA binding sites (Figure 1). The was either the consensus (TATAAT) or the weaker variant (TTTAAT). The MarA binding site was either the one that regulates the mar operon [22], or the anti-consensus site, which is not expected to bind or be activated by MarA. We will refer to each library based on which MarA binding site (Mar or Anti), and which binding site (TAT or TTT) it contains. The four libraries therefore are named Mar:TAT, Anti:TAT, Mar:TTT and Anti:TTT.
To test the dependency of cell growth in tetracycline on the sequence at the , we created promoters that contained either the consensus TTGACA or the anti-consensus GCCGGC in the Mar:TTT context. The anti-consensus site did not allow growth at as low as 5 g/ml of tetracycline, where the consensus allowed for growth in tetracycline concentrations at least as high as 100 g/ml suggesting that cell survival is dependent upon the binding site (data not shown).
Promoter competitions were performed as described in Materials and Methods. Briefly we transformed each library into E. coli cells and grew the cells overnight. The following morning, fresh LB cultures containing increasing concentrations of tetracycline were inoculated with the overnight cultures. Cells were competed for 24 hours and the competed populations were sequenced on a Solexa sequencer to determine the relative frequency of each hexamer. We sequenced 24 competed populations that covered 20 distinct selection conditions. Each competed population is named based on the competed library and on the concentration of tetracycline used in the competition. We carried out two independent competitions with the Mar:TAT and Mar:TTT libraries. The first was performed over the range of 5 to 30 g/ml tetracycline. We expanded the range to 50 g/ml tetracycline for all other experiments. To distinguish between different competitions with the same library, each culture that came from the same starter is given a common number (1 or 2). For example, Mar:TAT Tet-5 (1) and Mar:TAT Tet-10 (1) came from the same Mar:TAT overnight culture, but Mar:TAT Tet-50 (2) came from a different one.
The number of sequencing reads are given in Table S1. Differences in read numbers are most likely a result of sample loss in the Solexa prep and to the lower cell density in higher tetracycline concentrations, especially with libraries containing the TTT . All but four of the sequenced competed populations had at least 25,000 reads. As expected, Mar:TAT Tet-5 (1) was the most variable, and appeared to show only a slight preference for the sequence at the binding site. We observed 3918 of the 4096 possible hexamers in this population, suggesting that the coverage of all sequences in our library is essentially complete.
We sequenced Anti:TAT Tet-5 (1) and Mar:TTT Tet-5 (2) on two independent sequencing runs to determine if the number of sequenced DNA molecules gave an accurate and reproducible representation of the competed promoter populations. These runs generated 29,803 and 93,863 reads for the Anti:TAT Tet-5 (1) library and 33,229 and 11,263 reads for the Mar:TTT Tet-5 (2) library. We compared the relative frequency of each as determined from sequencing run 1 against run 2 and observed an for both samples (data not shown). This suggested that for the more degenerate TAT libraries, as few reads sufficiently covers the distribution of binding sites. As few as reads are sufficient for the TTT libraries.
Sequence logos are shown for the population of binding sites from each promoter context at 5, 10, 20 and 50 g/ml tetracycline (Figure 2). Logos generated from the Mar:TAT (1) and Mar:TTT (1) competitions over the smaller range of 5 to 30 g/ml were similar (data not shown). We observed a decrease in the variability for each library as the amount of tetracycline used for selection was increased, with the population converging towards the consensus binding site TTGACA, suggesting that only stronger sites (those closer to the consensus) are viable under more stringent selection conditions. We observed a similar decrease in variability as we decreased the energetic contribution of the other components in the promoter, strongly suggesting that a decrease in the affinity of the or MarA binding sites can be compensated by an increase in the strength of the . The single base-pair mutation in the had a major effect on the population variability. Whereas completely destroying the MarA binding site by replacing it with the anti-consensus affected the population variability considerably less.
For most populations, the first position of the hexamer is the least variable, and the site increases in variability towards the end. The first three positions are much more conserved than the last three, and position 6 appears to be relatively non-specific for most populations. This is consistent with the logo made from naturally occurring sites [11]. Only at the most stringent selective condition (Anti:TTT Tet-50) does the consensus sequence dominate.
We compared the information content () [23] for each competed population as a function of tetracycline concentration for the Mar:TAT and Mar:TTT libraries (Figure 3). This figure includes data for both competition series with these libraries. Both libraries show a linear increase in information content from 5 to 30 g/ml, with a leveling at 50 g/ml. As apparent from the sequence logos in Figure 2, the information content of the Mar:TTT library is much greater than that of the Mar:TAT library at all concentrations of tetracycline, suggesting that a weaker needs to be compensated for by a stronger for the promoter to be viable. Duplicate selections at 5 and 10 g/ml showed similar information contents for both libraries.
We predicted the relative affinity () of to each using the information theory based approach described in [2], [4] and the model presented in [11] (see Materials and Methods). The sites ranged in strength from to bits of information. Conventionally, sites with more than 0 bits are thought to be specifically bound [24]. 418 of the 4096 binding sites were bits. The relative fitness of each in the population was calculated by dividing the number of occurrences of that by the number of occurrences of the most frequently observed . We ranked all binding sites according to their , and compared the relative frequency for each in each experiment in Figure 4, and only those sites with an bits in Figure S1.
The majority of hexamers were present in all libraries that contained the sequence TATAAT. As seen in Figure 2, there is a decrease in the variability of observed binding sites as we increased the concentration of tetracycline used in selection and as the strengths of the and MarA sites are decreased in the promoter. We also observed a convergence of the viable sites towards those with higher information (sites closer to the consensus sequence).
Several competitions contained scattered low affinity sites with significantly higher fitness than the sites around them. We ordered all hexamers alphabetically (AAAAAA, AAAAAC, AAAAAG … TTTTTT) to see if there were sets of binding sites close in sequence space that had a high relative fitness, but not a high predicted affinity (Figure S2). We identified clusters of hexamers that contained a strong shifted one base to right (orange boxes in Figure S2 and Figure 5). That is, the second base of the randomized hexamer was the first base of the binding site. Differences in spacing between the and have been shown to affect the rate of initiation [5]. We tried to limit the number of binding sites with sub-optimal spacings from our libraries by placing bases disfavored by the model at the positions flanking the randomized hexamer [11] (see Materials and Methods). Since the last two bases of the hexamer are fairly non-specific, it is difficult to exclude viable s with shorter spacings.
The fitnesses of the binding sites were reduced at shorter spacings compared to the larger optimal spacing, and only the strongest sites were viable and only under the mildest selection conditions (Figure 5). To quantify this, we calculated the average relative fitness of four sets of hexamers that had shifted binding sites (Table 1). These sets of binding sites contained the 16 sites that had the consensus ‘TTG’ at the first three positions (positions 2–4 of the randomized hexamer) and a ‘G’ at the sixth position (TTGNNG). This ‘G’ is the base immediately of the randomized region, and is therefore fixed. The four sets only varied in which base was of the , and should be the highest affinity sites at this spacing according to the binding site model [11]. The average relative fitness was calculated across all experiments for these sequences (Table 1). The four sets had a similar average fitness to each other and a significantly higher fitness relative to 100,000 randomly chosen 16 hexamers (p), but on average were half as fit as the same set of sites at the optimal spacing (TTGNNG) and one third as fit as the 16 binding sites closest to the consensus (TTGANN) (Table 1).
To directly compare sequence activity to and relative fitness, we measured the transcriptional output of 8 binding sites in the Mar:TTT promoter context and 7 in the Mar:TAT context by quantitative PCR (Figure 6). The sequences of these sites, their predicted affinities and their transcriptional activities are reported in Table 2. For both libraries, output generally increased with . The data was best fit by a single exponential curve, but weakly; and for Mar:TTT and Mar:TAT respectively (these values were only calculated for sites with an bits) (Figure 6A). Sites similar in sequence produced almost equivalent outputs. In the Mar:TTT context, TTGCGT, TTGCAG and TTGCTT vary only at their last two bases, and have similar activities (Table 2). In the Mar:TAT context, TGGAGC and TGGCTA vary at the last three bases and have the same output, and TTGCTC, TTGATG and TTGCTT have similar outputs. We suspect the model is slightly overestimating the contributions of the last 3 bases of the hexamer, and this can account for inconsistencies between our predicted affinity and transcriptional output.
Expression from the Mar:TAT context was much greater than from the Mar:TTT context. The weak TAGACG in conjunction with the consensus TATAAT produced an output greater than the strongest that we assayed in the Mar:TTT context, TTGACT. Additionally, the activity of the same sequence (TTGCTT) in both contexts was 2.8 fold greater with the stronger . As seen in Figure 2 and Figure 3, these results indicate that differences in the have a significant effect on transcriptional activity.
Two of the binding sites in the Mar:TTT context had an bits, and both produced the same weak expression level (Table 2). We expect all non-specifically bound s to have this same output. One of these sites (CTTGAC) contained a strong that was shifted one base closer to the , but showed no activity (blue triangle in Figure 6). Additionally, we characterized two hexamers in the Mar:TAT context with an bits. One of these sequences (CCGTTC) showed a significantly reduced output relative to all other Mar:TAT sequences, but a high output relative to the Mar:TTT sequences. We expect this to be the transcriptional output for all non-specific s in this context. The other sequence (CTTGCC) contained a strong that was shifted one base to the right (orange triangle in Figure 6), but unlike the shifted site in the Mar:TTT context displayed high activity. This suggests that s with shorter spacings are only functional with the stronger , as seen in Figure 5.
There was a strong correspondence between transcriptional output and relative fitness for the 8 characterized s in the Mar:TTT context (Figure 6B). At 5 g/ml of tetracycline, fitness increased as a function of output for the 5 lowest expressing s and then slightly decreased for the 3 highest expressing. At 10 g/ml, the increase in fitness extended to all but the strongest , and at 20 and 50 g/ml, fitness increased with output for all sequences. The cellular advantage for producing more of the tetracycline resistance protein may be outweighed by the cellular cost in low concentrations of drug [25]. This may explain this decrease in the overall fitness at greater outputs. The relationship between output and fitness for the Mar:TAT characterized s was less striking (Figure 6C). At 5 and 10 g/ml of tetracycline, we observed an initial increase in fitness from the lowest to the second lowest expressing , and then no consistent trend. It is important to note that the differences in fitness between variants in this context are relatively small, especially compared to the Mar:TTT examples, and there could possibly be no effect on fitness at these high expression levels in these low concentrations of drug. More data points are needed to determine this. At 20 and 50 g/ml of tetracycline, we observed a general increase in fitness with output. Unlike in the Mar:TTT context, there was a gradual increase in fitness across these sites.
Fitness landscapes for individual hexamers across 16 different conditions are shown in Figure 7. We chose a series of five hexamers that decrease in predicted binding affinity from the consensus TTGACA, and differ from their neighboring sequence by a single nucleotide mutation. We also show a fitness landscape for the anti-consensus binding site GCCGGC. As expected the anti-consensus is not viable under any condition. There is an interesting contrast in the fitness landscape of the consensus sequence (TTGACA) to the weaker site TTGTTG. The consensus sequence shows a general increase in fitness to more stringent selective conditions, with a relatively low fitness in weak selective conditions. Conversely, TTGTTG is most fit in the weakest conditions and not viable at stringent conditions. TTGACG like TTGACA shows low fitness in the TATAAT libraries, but has a greater fitness for most of the selections with the weaker TTTAAT binding site, except for the most stringent. The fitness profile for TTGATG is weaker than expected for a site of that strength suggesting that its actual affinity may be lower than predicted. Regardless of our prediction of site strength, the difference between the TTGACG and TTGATG landscapes is large, illustrating how a single nucleotide mutation can radically change the fitness landscape of a binding site.
To better understand how binding site strength correlates with relative fitness in different promoter and environmental contexts, we calculated the average relative fitness for all sites within 1 bit bins (Figure 8). For the Mar:TAT library (Figure 8A), we observed that the range that has the greatest average fitness is not the highest one. We did observe an increase in the strength of the optimal fitness range as we increased the selection concentration of tetracycline, but for all tetracycline concentrations we saw a decrease in fitness at the highest range of binding sites. For the Mar:TTT library, we observed a general increase in relative fitness as a function of binding site strength for all tetracycline concentrations. Interestingly we did not observe the decrease here as we observed in Figure 6B. We did observe a similar decrease in fitness at higher information sites for the Anti:TAT library at 5 g/ml tetracycline, but not at higher concentrations. The Anti:TTT library only showed an increase in fitness at higher binding site strengths (data not shown).
To decipher cis-regulatory information and subsequently understand how it evolves, we need to be able to experimentally associate expression phenotype to genotype for large libraries of sequences. While there has been some success in doing this [12], these datasets are still extremely challenging to generate because it is difficult to maintain genotypic information in bulk reactions, requiring a large number of independent assays. Here we were able to overcome this problem by measuring the abundance of a genotype in a competed population of promoters, where cellular fitness is a function of its transcriptional phenotype (production of the tet gene). Given a mapping of phenotype to genotype for large libraries of sequences, it is still difficult to parse out the effects of single nucleotide differences on transcription since the rate of initiation is dependent upon many variables. Here we reduced this problem by generating libraries of promoters that only differ by the sequence of a single binding site (the ). The method worked well. For the first time, we were able to generate experimentally determined fitness landscapes for a large set of sequences in multiple promoter and environmental contexts. These data give insight into both the mechanism and evolution of transcriptional regulation at the level of an individual binding site.
The fitness of the transcriptional output of a binding site is a complex function of the cellular gain and cost associated with the production of expressed gene [25]. The cellular gain in our synthetic system is the increased ability to export tetracycline from the cell. The cellular cost is the toxic effect of over-expressing the tetracycline efflux pump [26], [27]. While we do not fully understand the absolute relationship between binding site strength, transcriptional output and the fitness of that output, clearly these things are related (Figure 8, Figure 6) and highly context dependent (Figure 7).
The relative frequency of recovery of a binding site in a competed population is dependent upon two variables, (Minimum Viable Stability) and (Optimal Stability). is the minimum stability of the initiation complex needed to produce enough of the tet gene to survive. is the stability of the initiation complex that produces the maximally fit output given a concentration of tetracycline. For a to be viable in our selection, it must have an affinity that in combination with the other binding sites produces an initiation complex stability that is stronger than . As the strength of the other sites or the output requirement changes, so does the boundary of the minimum viable binding site strength. This is indeed what we observe in Figure 4 and Figure S1. As we increased the concentration of tetracycline (decrease ) or as we decreased the strength of the or MarA binding sites, only stronger s remained in the selected population. This is also illustrated in Figure 2 and Figure 3 as a decrease in the variability of the population and a convergence on the consensus sequence at more stringent (energetically demanding) selection conditions. Compensation in binding energies between sites to produce similar stabilities has been previously predicted computationally for binding sites [11] and is shown clearly here. Interestingly, the information content of the competed populations increases linearly as a function of tetracycline concentration over the range of 5 to 30 g/ml and levels off at 50 g/ml for both the Mar:TTT and Mar:TAT libraries (Figure 3). We are not sure why the information content levels off. One possibility is that we are approaching the maximum stability where the transcriptional initiation rate is limited by the stability of the closed complex.
The most fit in a given context should have an affinity, that in combination with the other binding sites, equals . We expect that fitness will increase with the overall stability of the initiation complex from to . We observe this qualitatively for libraries containing the weaker TTTAAT binding site or libraries selected at high concentration of tetracycline. Here, sites generally increase in fitness as a function of binding site strength (Figure 4, Figure S1). Some sequences show an unexpected high or low fitness compared to their neighboring sequences with similar predicted affinities. These could be partially explained by insufficient sequencing depth, but we expect to a small degree since technical replicates suggest that for most conditions our depth gives an accurate representation of the population. Another possibility could be that some promoters may be under or over-represented in the initial library. We expect that to some extent these discrepancies are due to inaccuracies in the binding model that we used. A comparison between and transcriptional output suggests that the model may be slightly overestimating the energetic contributions of the last three bases of the hexamer to binding site strength (Figure 6). A large number of sequence anomalies can also be attributed to binding sites with shifted spacings relative to the (Figure 5).
When the average fitness is calculated for binding sites with similar affinities (reducing the effects of anomalous s), we see a smooth relationship between fitness and binding site strength (Figure 8). In strong selection conditions (high tetracycline concentration, weak ), exceeds the maximum stability that can be accessed by only varying the binding site, so here an increase in binding affinity always increases fitness (Figure 6B and 6C, Figure 8B). In weak selection conditions (low tetracycline, strong ), the optimal binding site does not appear to be the strongest (Figure 6B, Figure 8A). That is, is within the range of affinities that can be accessed by changing the . The additional energy from the presumably shifts the distribution of outputs for the binding sites into a range where there is no longer an increased advantage or even a disadvantage for transcribing that much tet.
Overall, we observed a large and continuous range of fitnesses suggesting a similar scope of potential outputs can be evolved or engineered by solely mutating the . Fitness landscapes of individual sequences illustrate the large effect on fitness by even a single mutation (Figure 7). It is not clear what the maximum stability of the initiation complex is where increases in stability will no longer increase output (closed-complex stability is not limiting). It has been shown for some promoters that too strong of an interaction can actually decrease transcriptional output, presumably because it is difficult for the polymerase to dissociate from the DNA [28]. A decrease in fitness from the highest affinity consensus binding site compared to a single base pair mutation of the consensus in the Anti:TTT context (Figure 7), suggests that the range of affinities of binding sites alone does not exceed that maximum. There may have been selection on to keep the range of affinities below this maximum, to maximize its output range.
The relative contributions of the and MarA binding sites do not appear to be equivalent. A single mutation in the second position of the consensus greatly reduces the variability of the binding site populations. Whereas completely removing the MarA binding site has a significantly reduced effect. This suggests that binding at the contributes more to the stability of the initiation complex than does binding by MarA. The decrease in effect from the MarA site could be related to the energetics in the contact with the CTD which we do not understand [11], or MarA expression could be low resulting in a low occupancy of the site.
The significant effect of mutating the on transcript production is clearly shown in Figure 6A. The expression levels of all s in the Mar:TAT context, except for the non-specifically bound one, are greater than the expression from the most active in the Mar:TTT context that we characterized. This suggests that differences in the may contribute more than differences in the to the overall output. Open complex formation occurs through melting at the [11], [29], [30]. A mutation in the sequence could have a greater effect on the rate of initiation because it could lead to both a change in promoter stability and the rate of open complex formation. We expect that regardless of whether differences in the affect the stability of the closed complex or open complex formation, selection on the will be on its binding site strength. The larger range of outputs in the Mar:TAT context compared to the Mar:TTT context suggests some cooperativity between sites (Figure 6A). We do not have enough data to determine to what extent.
As previously mentioned, the spacing between the and can affect the rate of initiation [5]. While we tried to minimize the number of binding sites with alternative spacings from our library, this proved difficult because the last two positions of the hexamer are fairly non-specific. We observed that binding sites were viable with a 1 bp shorter spacing relative to the , but only in weak selective conditions (low tetracycline, strong and MarA binding sites) and only the strongest sites (Figure 5). This was confirmed by quantitative PCR, where we observed that only in the Mar:TAT context, could shifted sites produce an output above that of a non-specifically bound (Figure 6). The additional energy of the may be able to compensate for the energetic cost of binding the with a sub-optimal spacing [11]. We observed a similar average fitness for related sets of binding sites with a shifted (Table 1), suggesting that differences in the position of the do not affect transcriptional initiation. These sets of binding sites were on average about half as fit as the same set of sites with the larger optimal spacing, suggesting that differences in spacing significantly decrease transcriptional activity.
We placed the tetracycline resistance gene (tet) under control of a MarA-activated promoter on the E. coli plasmid pBR322. pBR322 has several advantages: (1) It confers resistance to both ampicillin and tetracycline, allowing for maintenance of the plasmid to be either independent of or dependent on the promoter of tet. (2) It is a relatively low copy plasmid (15–20 copies per cell) [31]. This eliminates the high expression of tet associated with large copy numbers. We generated four promoter libraries where the was randomized and contained either one of two MarA and binding sites (Figure 1).
Variability in the relative spacing between binding sites can affect the rate of transcription [5], [6]. We designed the promoter insert to strongly favor a single spacing between the and the to avoid having to consider spacing effects on the fitness of the promoter in the analyses. We used the optimal spacing between the and [11], where deviations from this spacing would result in a decrease in binding affinity. Additionally, the two bases immediately (‘CA’) and the two bases immediately (‘GC’) of the hexamer are disfavored at the first and last two positions of the respectively [11], further reducing the possibility of strong binding sites with different relative spacers. The sequence between the and MarA binding site is a slight variant of the sequence found between the MarA site and the in the mar promoter [22]. We shortened the spacer by one base at the end to have the disfavored ‘CA’ immediately adjacent to the . Martin et al. showed that this shortened spacing has a minimal effect on the degree of MarA activation [6]. We also changed three bases in the spacer to create a BstBI site (TTCATT is now TTCGAA).
The weaker (TTTAAT) in the promoter of the tet gene was mutated to the consensus (TATAAT) by QuickChange according to Zheng et al. [32]. These two pBR322 variants, pBR322 and pBR322, were used for subsequent library construction. The of the tet gene on pBR322 is flanked by two unique restriction sites, EcoRI and ClaI. These sites were used to clone in MarA binding site and variants as described below.
The randomized library inserts were created by DNA synthesis (Integrated DNA Technologies). Variation of the binding site was done by mixing equal quantities of each base at those positions. Two library inserts were synthesized that contained either the stronger mar MarA binding site [22], or the non-specific anti-consensus MarA binding site. The latter has the least frequently observed base at each position based on the MarA binding model (model not published but generated from sequences in [6]) and should not be bound. These inserts will be referred to as Ins and Ins. The DNA was made double stranded by second strand synthesis with Klenow (NEB), and the fragments were purified with a QIAquick PCR purification kit (Qiagen).
pBR322, pBR322, Ins, and Ins were cut with EcoRI and ClaI (New England Biolabs) for two hours at C and gel purified using a QIAquick gel extraction kit (Qiagen). All four combinations of plasmids and inserts were mixed and ligated overnight at C with T4 DNA ligase (NEB) generating 4 libraries (Mar:TAT, Anti:TAT, Mar:TTT and Anti:TTT). The ligated libraries were transformed by electroporation into DH10B cells (Gibco BRL), and plated on 100 ml LB+30 g/ml ampicillin plates. The number of transformants for each library was ca. . The colonies were suspended from the plate in 10 ml LB, and mini-prepped using a QIAquick miniprep kit (Qiagen).
Libraries were transformed by electroporation into the E. coli strain DH10B (Gibco BRL). The number of transformants was ca. as determined by plating. After transformation, cells were recovered in 500 l LB for 1 hour, and grown further in 5 ml LB+30 g/ml of ampicillin overnight at C, with shaking at 225 RPM. Fresh 5 ml LB cultures containing from 5 to 50 g/ml of tetracycline were inoculated with 100 l of the promoter libraries grown overnight. Promoter libraries were competed against each other for 24 hours at C, with shaking at 225 RPM. Plasmids were purified from the competed libraries using a QIAquick miniprep kit.
The Mar:TTT and Mar:TAT libraries were plated on LB agar plates containing 0 to 100 g/ml of tetracycline. Individual colonies were sequenced from these plates, and 8 variants in the Mar:TTT context and 7 variants in the Mar:TAT context were chosen that covered a large range of predicted binding strengths for further analysis. 5 ml LB cultures containing 30 g/ml of ampicillin were inoculated with E. coli containing a single binding site variant and grown overnight. A fresh 5 ml LB+30 g/ml ampicillin culture was started at and grown to an . cells were added to RNAprotect Bacteria reagent (Qiagen), and RNA was purified using the RNeasy Mini kit with on-column DNase digestion (Qiagen). cDNA was made from 2 g of RNA using the Superscript III RT kit (Invitrogen). QPCR was performed with the SYBR green mix from NEB. QPCR primers specific to the tet and gyrA gene were both used. The relative expression of the tet gene was determined by the ratio of tet transcript abundance over gyrA transcript abundance for each sample. A serial dilution of the Mar:TTT, TTGACT sample was used as a standard for both primer sets. The expression of the tet gene for all variants was calculated relative to this. All sequences used, their predicted affinity () and the expression values are reported in Table S2.
Conversion of pBR322 to pBR322 destroyed a HindIII site that overlapped the first two bases of the hexamer. Libraries that contained the wild type (TTTAAT) were digested with HindIII and PvuI (NEB) for 2 hours at C. pBR322 libraries were digested with ClaI and PvuI (NEB) for 2 hours at C. base pair fragments were gel purified for all four libraries using the QIAquick gel extraction kit. Excised fragments from all four promoter libraries, selected at a single tetracycline concentration, were mixed at equal concentration. Solexa libraries were then generated from this mixed population.
The Illumina genomic library protocol was slightly modified (Illumina, Inc.). We used a 1∶10 dilution of the Solexa genomic adapter, and ran the PCR for 16 rounds. We gel purified the final product after the PCR step instead of before as suggested. This allowed the removal of potential adapter contaminants. Sample purity and concentration were measured using a Bioanalyzer (Agilent Technologies). A 45 bp single-end run was performed on a GAII machine according to the Illumina protocol.
For each tetracycline concentration, the reads were identified as originating from one of the four promoter types. We used only those sequences that had an exact match to 14 or 21 specific bases that flanked the −35 region for the TAT and TTT libraries respectively. We did this to ensure that this sequence was not mutated, the spacing between the −10 and −35 was not changed, and to increase our confidence in the accuracy of the −35 sequence. We used 7 additional bases for the TTT libraries because those libraries were cut 7 bases further from the −35 than the TAT libraries. These additional bases were used to determine which variant was present for that sequence. Additionally, we required an additional 10 bases before and overlapping the MarA binding site to exactly match to confidently distinguish between the Mar and Anti libraries. The number of reads for each competition that pass these criteria are reported in Table S1.
Each was counted for each competed library at a tetracycline concentration. To determine the relative fitness of a in a competed population, the number of reads containing that was divided by the number of reads of the most frequently observed . For two of the competitions, Anti:TAT Tet-5 and Anti:TAT Tet-10, three hexamers (TGCCCA, TCCATT and CTGGAT) were disproportionally high relative to the others. Interestingly, if two of these hexamers are put in the context of the promoter sequence, CA-TCCATT-G is only one base different from the reverse complement of CA-CTGGAT-G (C-ATCCAG-TG). The hexamer sequence is separated from surrounding sequence by ‘-’. These sequences may encode for the binding site of some unknown factor which may explain their increased fitness. At greater tetracycline concentrations though, these were observed much less frequently. For these competitions, the fitness of the hexamers were calculated relative to the fourth most frequently observed hexamer.
Sequence logos were generated from the alignment of all reads for a single library at a single tetracycline concentration using the delila software [33].
We used the program scan to predict the relative affinity () of to each hexamer. Briefly, scan compares an individual sequence to an information theory based weight matrix and sums the information contribution of each base across all positions in a site [2]. The weight matrix that we used is the one generated from 401 experimentally verified promoters in E. coli presented in [11] and is given in the supplemental materials of this paper (Table S3).
There are several advantages to this approach. First, the weight matrix is generated from a large number of experimentally verified promoters, and should not be skewed by binding site selection biases [34]. Second, has been shown experimentally to be directly proportional to and more specifically [4]. Third, the information theory approach predicts a clear demarcation between specifically and non-specifically bound sites at 0 bits [24].
|
10.1371/journal.pgen.1003579 | Evidence for Two Different Regulatory Mechanisms Linking Replication and Segregation of Vibrio cholerae Chromosome II | Understanding the mechanisms that coordinate replication initiation with subsequent segregation of chromosomes is an important biological problem. Here we report two replication-control mechanisms mediated by a chromosome segregation protein, ParB2, encoded by chromosome II of the model multichromosome bacterium, Vibrio cholerae. We find by the ChIP-chip assay that ParB2, a centromere binding protein, spreads beyond the centromere and covers a replication inhibitory site (a 39-mer). Unexpectedly, without nucleation at the centromere, ParB2 could also bind directly to a related 39-mer. The 39-mers are the strongest inhibitors of chromosome II replication and they mediate inhibition by binding the replication initiator protein. ParB2 thus appears to promote replication by out-competing initiator binding to the 39-mers using two mechanisms: spreading into one and direct binding to the other. We suggest that both these are novel mechanisms to coordinate replication initiation with segregation of chromosomes.
| Replication and segregation are the two main processes that maintain chromosomes in growing cells. In eukaryotes, the two processes are restricted to distinct phases of the cell cycle. In bacteria, segregation follows replication initiation with a modest lag. Influences of one process on the other have been postulated. The act of replication has been suggested to provide a motive force in chromosome segregation. Moreover, segregation proteins (ParA) have been found to interact with and control the replication initiator, DnaA. Here we show that in V. cholerae chromosome II, which is believed to have originated from a plasmid, a centromere binding protein (ParB) could control replication by two distinct mechanisms: spreading from a centromeric site into the replication-control region, and direct binding to the primary replication-control site, which has limited homology to the centromeric site. These studies establish that Par proteins can influence replication by at least three mechanisms. Homologous Par proteins participate in plasmid segregation but they are not known to influence plasmid replication. The expanded role of Par proteins appears likely to have been warranted to coordinate chromosomal replication and segregation with the cell cycle, which appears less of an issue in plasmid maintenance.
| Studies in bacteria as well as in eukaryotes have shown that processes that maintain chromosomes, such as replication, recombination and repair, although able to occur independently of each other, often influence each other. Chromosome segregation is a major maintenance process but our knowledge of it in bacteria is relatively recent. This is because in well-studied bacteria such as Escherichia coli, genes dedicated to the segregation process have not been evident. Such genes were discovered in bacterial plasmids, called parA and parB. Subsequently, their homologs were found in a majority of sequenced bacterial chromosomes [1], [2]. Wherever tested, the chromosomal parAB genes were capable of conferring segregational stability on unstable plasmids bearing parS (“centromere” analogous) sites [3], [4], [5], [6], and made at least some contribution to chromosome segregation [7], [8]. In spite of limited study, it is becoming clear that chromosomal segregation systems can influence and be influenced by other chromosome maintenance processes.
In bacteria, replication and transcription have been proposed to provide motive force in chromosome segregation [9], [10], [11]. Coupled transcription-translation of membrane proteins is also thought to play an important role in chromosome segregation [12], [13]. One of the segregation proteins, ParB, can also spread and silence transcription of genes in its path [14], [15], [16]. The influence of segregation proteins in replication was suggested when ParB was found to load a condensin protein in the vicinity of the replication origin in Bacillus subtilis and in Streptococcus pneumoniae [17], [18], [19]. A more direct role was evident when ParA was found to influence the activity of the initiator DnaA in B. subtilis chromosome replication [20], [21] and in replication of Vibrio cholerae chromosome I (chrI) [22]. Recently, ParB encoded by V. cholerae chromosome II (chrII) was also found to influence chrII replication [23]. Here we report two distinct mechanisms for this ParB-mediated effect.
V. cholerae chrII replication is primarily controlled by its specific initiator protein, RctB [24], [25]. RctB binds to two kinds of site in the replication origin of chrII. One kind, the 11- or 12-mers, plays both essential and regulatory roles [26]. The other kind, two 39-mers and a 29-mer (a truncated 39-mer), plays only an inhibitory role in replication [26], [27]. One of the 39-mers is situated at a locus called rctA, at one end of the origin, and the other is more centrally located in the origin (Figure 1, top). The 29-mer is located in front of the rctB gene and is involved primarily in autorepression of the gene [27]. The rctA locus contains, in addition to a 39-mer, one of the ParB2 binding sites, parS2-B [28]. It has been reported recently that parS2-B alleviates some of the replication inhibitory activity of rctA in a ParB2-dependent fashion, but the mechanism is unknown [23]. Here we show that ParB2 spreads from parS2-B into the rctA 39-mer, and suggest that the spreading likely interferes with RctB binding to the 39-mer and thereby restrains the inhibitory activity of rctA. Unexpectedly, we also found ParB2 promotes replication by directly binding to the central 39-mer, without requiring spreading from parS2-B. We provide evidence that ParB2 competes with RctB for binding to the central 39-mer specifically and could thereby restrain its activity. In addition to revealing new ways by which a Par protein might influence replication, our results are significant in demonstrating that a segregation protein can bind specifically outside of centromeric sites.
The origin region of chrII comprises three functional units: A region required for controlling initiation (incII), a region minimally required for initiation (oriII) and a gene required for synthesizing the initiator protein (RctB) (Figure 1, top). The region covering the first two units, which consists mostly of sites for initiator binding, will be referred to as the origin. The rctA locus of incII exerts a strong inhibitory effect on chrII replication because of the 39-mer it contains [26]. The locus also has a site, parS2-B, for binding to the segregation protein ParB2, and the inhibitory effect of rctA is reduced in the presence of ParB2 [23]. Knowing that ParB proteins can spread out from their binding sites to neighboring sequences [14], we asked whether ParB2 spreading from parS2-B over the 39-mer might be a mechanism to control its inhibitory function. The extent of ParB2 spreading was tested by ChIP-chip analysis using antibody against ParB2 (Figure 1, bottom). Spreading was evident on either side of parS2-B (grey profile). In contrast, when the antibody was against RctB, the immunoprecipitated DNA in the origin region was restricted to where RctB has specific binding sites (black profile). These results suggest that ParB2 has the potential to modulate replication initiation activity by interfering with RctB binding.
Spreading of proteins across genes can silence them [14]. For example, spreading into plasmid replication genes suitably close to a parS site can be lethal when selective pressure for plasmid retention is applied. By such an experimental test, we found that ParB2 can spread from parS2-B (Figure S1A). Spreading is also suggested by the formation of ParB2-GFP fluorescent foci in parS2-B carrying plasmids (Figure S1B). ParB2 could also silence two promoters, PrctA and PrctB, in the origin of chrII (Figure 1, top). PrctA is proximal to parS2-B whereas PrctB is located about 1 kb away at the other end of the origin. The activity of the promoters was assayed by fusing them to a promoter-less lacZ gene present in a multicopy plasmid in E. coli (Figure 2). The promoter fragments fused to lacZ carried either the entire origin including parS2-B (1A and 1B), or the origin lacking parS2-B (2A and 2B) or no additional DNA (3A and 3B). ParB2 was supplied constitutively in trans at about an order of magnitude higher than the physiological level (monitored by Western blotting; Figure S2), using Ptrc promoter without an intact lac repressor binding-site (lacO1), which makes the promoter unresponsive to IPTG. The presence of ParB2 reduced the activities of PrctA and PrctB significantly, only when the parS2-B site was present (Figure 2, 1A and 1B). These results suggest that ParB2 can spread over the entire origin in the presence of parS2-B, and does not have a significant effect on either promoter in the absence of parS2-B.
Since RctB has numerous binding sites in the origin, it appeared possible that RctB binding to them could counteract ParB2 spreading and reduce silencing of the promoters. This possibility was addressed by supplying RctB from an arabinose-inducible promoter, PBAD [24], [26], [27], [29]. The induction of RctB alone (at about two-fold the physiological level) repressed PrctA marginally and PrctB about two fold (Figure S3, lanes 1 vs. 5). Silencing by ParB2 exceeded 90% for both the promoters (Figure S3, lanes 1 vs. 4). When RctB and ParB2 were supplied together, the repression of both the promoters was reduced about two fold compared to the level achieved with ParB2 alone (Figure S3, lanes 4 vs. 6; also inset). These results indicate that RctB can counteract the ParB2-mediated silencing.
In the results presented above (Figure 2 and Figure S3), the level of ParB2 was about 14-fold the level normally present in V. cholerae (Figure S2). When the concentration was reduced to about 10-fold, ParB2 could silence only the parS2-B proximal PrctA, but not the distal PrctB promoter (Figure S3, lanes 1 vs. 3). This reduced level of ParB2 was used in all subsequent experiments.
In order to determine how far ParB2 can spread beyond PrctA, progressively increasing lengths of incII were fused to a foreign reporter promoter, PrepA, itself fused to lacZ [30] (Figure 3). In these experiments, in addition to a plasmid supplying ParB2, another plasmid was used to supply RctB. The two proteins were expressed from inducible promoters, Plac and PBAD, respectively (Figure 3, cartoon at the top right corner).
As expected, neither ParB2 nor RctB influenced the activity of PrepA itself (Figure 3, top panel). In contrast, when rctA was present, ParB2 reduced the activity of PrepA by two-fold (second panel). We believe this is due to ParB2 spreading from parS2-B into PrepA, whose −35 box was only 167 bp away. RctB alone was ineffective, most likely because it does not spread, and its specific binding site, the rctA 39-mer, is well separated (by 85 bp) from the −35 box of PrepA. Supplying RctB was marginally effective in relieving the silencing by ParB2. The next extension of the incII fragment included the 3x11-mers (third panel). Neither ParB2 nor RctB could silence the reporter promoter in this case. This result suggests that the spreading may not extend too far beyond PrctA. A further extension of the incII fragment by only 74 bp that included the central 39-mer, restored ParB2-mediated repression of the reporter promoter (fourth panel). This result was surprising since parS2 sites were not found within incII [28]. The effect of ParB2 was significantly reduced when RctB was supplied, which is to be expected since RctB binds strongly to the central 39-mer [26]. This result suggests that ParB2 and RctB can compete for binding to the central 39-mer. The largest fragment (bottom panel) did not show a significant ParB2 effect on the reporter promoter, suggesting that ParB2 may not spread significantly from the central 39-mer. Together, the results suggest that under the conditions tested, ParB2 affects the origin primarily through interactions near rctA and the central 39-mer. The interaction near the central 39-mer suggests that ParB2 might bind there directly.
The possibility of site-specific binding of ParB2 within the origin but outside of parS2-B was tested by EMSA. Several fragments covering the origin were used. Fragments 1 and 2, carrying the parS2-B site (positive controls), showed maximal ParB2 binding (Figure 4). The next significant binding was with the fragment containing the central 39-mer (fragment 5). This fragment contained natural flanking sequences of only 3 bp and 32 bp beyond the central 39-mer. The sequences (3+39+32) are exactly those that were added to the incII fragment of the third panel to generate the silencing-proficient fragment of the fourth panel (Figure 3). We found that the flanking sequences do not contribute to the central 39-mer binding (Figure S4, fragment #1). This result supports the inference from in vivo studies that ParB2 can directly bind to the central 39-mer without requiring parS2-B. Binding to the rctA 39-mer (fragment 3) was considerably weaker, possibly because the two 39-mers have several mismatches between them (Figure 1 of [26]; discussion related to Figure 5 below). The level of binding seen with the rctA 39-mer was comparable to the levels seen with fragments 4, 6, 8 and 9, and the level was marginally above that of the negative control that lacks any chrII sequences, suggesting that ParB2 has significant non-specific DNA binding activity.
The sequence requirement for specific binding of ParB2 to the central 39-mer was tested by variously mutating the sequence. The 39-mer has two conserved 9 bp direct repeats (called A and B boxes) flanking a 19 bp AT-rich spacer (Figure S4). The presence of both of the repeats and their proper phasing are important for RctB binding [26]. The AT richness of the spacer is also important but not the sequence per se. The parS2 sites are AT-rich inverted repeats, only 15 bp long. Notably, the 39-mer spacer also contains an inverted repeat, which has some similarity to the consensus parS2 site. However, the 39-mer spacer by itself was not sufficient for ParB2 binding; the presence of one of the direct repeats was necessary (Figure S4, fragments 4–6). Either of the direct repeats alone was also not sufficient (fragments 2–3). The inverted repeat feature could also be destroyed without compromising the binding efficiency (fragment 12). When the same fragments were tested for RctB binding, only the ones with the intact 39-mer and 10 bp deletion or addition (fragments 9–11) showed significant binding, as was also found earlier (data not shown; [27]). It appears that while both ParB2 and RctB bind to the 39-mer, the presence of one of the direct repeats is not obligatory for ParB2 binding.
Specific binding of ParB2 to the 39-mer was also verified by DNase I footprinting (Figure S5). Protection by ParB2 was conspicuous at the junction of the first direct repeat (A-box) and the AT-rich spacer. At this junction an intact parS-2B half site, 5′-TGTAAA, is present. This sequence is fully conserved in all 10 parS2 sites that were competent in ParB2 binding [28]. In Figure S4, this half-site sequence was intact in all the binding positive fragments and mutant in all the fragments that failed to show specific binding. The half-site is also mutated to 5′-TTAAAC in the ParB2 binding-negative 39-mer in rctA (Figure 4, fragment #3). The half site thus appears to be necessary for 39-mer binding of ParB2. In further support of this inference, when we restored the original bases to some of the binding-negative 39-mer mutants to regenerate the half-site, binding proficiency was regained (Figure 5, fragments #3–6). Although necessary, the half site was not sufficient for binding ParB2 (fragment #7). We conclude that extension of the half site either to the left or right is necessary. This is not surprising since the affinity drops by orders of magnitude when one half of a dyad symmetric site is mutated [31], [32]. The minimal size of the extensions needed to regain binding activity of ParB2 remains to be determined.
ParB2 and RctB can bind to rctA simultaneously [23] (Figure 6, top panel). This is not surprising since there are 34 bp of spacer sequence between the binding sites of the two proteins, and that ParB2 does not spread in vitro. The sites also remain functional when isolated from each other [33]. On the other hand, at the central 39-mer, the binding sites for the two proteins appeared to be largely overlapping, suggesting that they could not bind simultaneously. This was indeed the result (Figure 6, bottom panel). Even at the higher protein concentrations (++), no new discrete species representative of dual binding was detected. The results indicate that ParB2 and RctB compete for binding to 39-mer, unlike the simultaneous binding that can occur on rctA.
We previously showed that the central 39-mer is the most potent replication inhibitory site in incII and it functions through RctB binding [26]. If ParB2 competes with RctB for binding to the central 39-mer, this competition appeared likely to influence oriII activity without requiring the parS2-B site. This prediction was tested by determining the copy number of oriII-driven plasmids (Figure 7). The copy number of oriII plasmids depends on the extent of the incII sequences present [26]. Although the 39-mers are always inhibitory to replication, the 11- and 12-mers can either promote or inhibit replication depending upon whether the 39-mers are present or not. In the present experiments also, the oriII plasmid copy number first decreased and then increased with increasing deletion of incII (Figure 7, − ParB2 column). When ParB2 was additionally present, the copy number increased significantly in the first two 39-mer-carrying plasmids, the increase being maximal for the plasmid with the lowest copy number (pTVC25). In this plasmid, we suggest that the 39-mer was unencumbered by the 3x11-mers, and was maximally available for binding to ParB2. Together, these results indicate that ParB2 has the potential to facilitate chrII replication by restraining the inhibitory activity of the incII sequences, and can do so whether parS2-B is present or not.
If ParB2 spreading is one of the mechanisms by which the protein stimulates chrII replication, it might be possible to restrain this activity by placing a roadblock in the path of spreading. To this end, we inserted an array of five P1 RepA binding sites (iterons) between parS2-B and the 39-mer in rctA (Figure 8). The effectiveness of the roadblock in preventing the spreading of P1 ParB protein was demonstrated earlier [34]. Comparison of the top two rows of the Table in Figure 8 shows that in the absence of RepA (that is in the absence of a roadblock), ParB2 was equally efficient in promoting cell growth that depended on the functioning of oriII plasmids. In other words, the P1 iterons in pBJH218 did not compromise ParB2 spreading in the absence of the roadblock. The same two plasmid-carrying cells behaved differently in the presence of RepA (the last two rows). Upon induction of ParB2 production by IPTG, cell growth improved more in the case of pTVC20 than in the case of pBJH218. In other words, ParB2 effect was compromised under the condition the roadblock was expected to be effective. These results are consistent with ParB2 spreading as a mechanism for stimulating chrII replication initiation. Note that some increase of growth rate was seen even when ParB spreading was inhibited by a roadblock (generation time decreased 7% for cells in row #4). This result is not surprising because ParB2 can bind to the central 39-mer without requiring spreading from parS2-B. Overall, the ParB2 effects were modest, which is to be expected because of the existence of multiple controls on chrII replication.
In chromosome and plasmid segregation, ParB proteins serve to couple centromeres to ParA proteins (NTPases) and modulate the NTPase activity that is believed to provide the movement required for segregation [8]. The binding ParB2 to a 39-mer raises the possibility of an inherent centromeric function of the site. This was tested by cloning the central 39-mer into a miniF plasmid, which is unstable due to deletion of its own segregation genes [35]. The stability of the miniF plasmid improved with the inclusion of the parS2-B site but not with the 39-mer, when ParA and ParB proteins were supplied in trans (Figure S6). This result suggests that ParB2 binding to parS2-B and the central 39-mer is different in an important respect.
Our knowledge of interactions between the processes of replication and segregation of chromosomes in bacteria is rather recent and limited. An interaction between the universal bacterial replication initiator, DnaA, and a segregation protein, ParA, was recently discovered in B. subtilis [20], [21], [22], [36] and later in V. cholerae where it was specific for chromosome I [22]. In both cases, replication initiation was modulated by ParA. A partner segregation protein, ParB, was found to affect replication of chromosome II (chrII) in V. cholerae. This parB, called ParB2, somehow promoted replication by binding to one of its cognate centromeric sites (parS2-B) [23]. Here we show that ParB2 spreads out of this centromeric site into the replication origin of chrII, and suggest that this spreading is a mechanism by which ParB2 promotes replication. We also report an additional mechanism by which ParB2 can promote chrII replication: direct binding to a replication inhibitory site in the origin. To our knowledge, the present results provide the first examples of a replication activation mechanism that is mediated by spreading of the activator from a distant site, and by the specific binding of a segregation protein outside of the centromere. These studies have revealed at least three ways by which segregation proteins can influence replication (Figure 9).
The spreading of ParB2 from a centromeric site into the origin of chrII was evident from in vivo cross-linking experiments (Figures 1,S7), from silencing of promoters within the origin (Figure 2) and from the reduction of reporter promoter activity when natural initiator (RctB) binding sites were present between the centromeric site and the promoter (Figure 3, third panel; Figure S3, insets). This latter result indicates that RctB binding could create a natural roadblock to ParB2 spreading. The fact that the span of silencing lengthened with increased ParB2 concentration (Figure S3) also supports the idea that the underlying mechanism involves spreading along the DNA. Finally, the results of placing an artificial roadblock were also consistent with the spreading mechanism (Figure 8). When a powerful replication inhibitory site (the rctA 39-mer) was present within the span of spreading, growth of cells dependent upon the functioning of the chrII origin improved. It was also reported earlier that ParB2 could increase replication of chrII origin carrying plasmids when they included the adjacent rctA region [23]. This increase was shown to be dependent on the presence of parS2-B. Together with the finding that ParB2 does not directly bind to the rctA 39-mer (Figure 4), and cannot spread from its binding site in the central 39-mer as discussed below (Figures 3 (last panel), S1A, S1B), the simplest explanation of these results is that by spreading from parS2-B, ParB2 compromises the inhibitory activity of the rctA 39-mer by interfering with RctB binding.
ParB2 was also found to reduce the activity of another potent replication inhibitor (the central 39-mer) without requiring the centromeric site and spreading (Figure 3, S1). The latter effect appears to be due to direct binding of ParB2 to the central 39-mer. This mode of ParB2 interaction with the 39-mer most likely also causes interference with RctB binding to this site (Figure 6). Since the 39-mers are the two sites most inhibitory to chrII replication and their activities are mediated through RctB binding, the reduction in binding suffices to explain how ParB2 could promote replication of chrII (Figures 7,8). Interference with binding of regulatory proteins to DNA by the spreading of a competing protein along DNA has also been invoked to explain transcriptional silencing, inhibition of DNA methylation and of DNA gyrase binding, and resistance to DNase I cleavage [37], [38], [39], [40].
Although the only model we have entertained so far to explain the ParB2 effect is interference with specific binding of RctB, we have also tested whether ParB2 and RctB could interact directly. This possibility was suggested by the finding that ParA influences replication by protein-protein interaction rather than DNA-protein interaction [20], [21], [22], [36]. However, ParB2 did not show any detectable interaction with RctB, as was also reported earlier (Figure S8) [23].
ParA participates in a number of processes involving ParB [8]. Here, we asked whether binding of ParB2 to the central 39-mer is also influenced by ParA2. The binding was assayed indirectly by fusing a foreign promoter close enough to the 39-mer that ParB2 binding to the site could interfere with the promoter activity. The promoter activity did not change significantly upon supply of ParA2 (Figure S9; data with PrepA). This suggests ParB2 binding to the 39-mer is not influenced by ParA2. We did find a minor influence of ParA2 on PrctA silencing by ParB2 spreading, the basis of which was not explored. In the case of P1 plasmid and B. subtilis chromosome, no ParA effect on ParB spreading was evident [14], [41].
Whereas deletion of parS2-B in V. cholerae was easily tolerated (Figure S7; [23]), deletion of the parB2 gene was essentially lethal [42]. In the absence of ParB2, chrII loss is evident at every cell division that causes a severe growth defect. We were therefore unable to test conveniently the role of ParB2 on replication of chrII in the native host. On the other hand, there was no obvious effect of ParB2 on the growth of E. coli, in which we did most of our experiments. The validity of extrapolating the observations in E. coli to the native host appears warranted by the observation that ParB2 effects were seen in the context of the entire origin (Figure 8; [23]), and by the finding that some of the inferences from the E. coli results were valid when tested in vitro (Figure 6). In the past, wherever chrII replication control was studied in both E. coli and V. cholerae, the results agreed [25], [26], [28], [29]. Nonetheless, the ParB2 concentration required for spreading to proceed just over rctA was an order of magnitude higher than that is normally present in the native host. The reason for this discrepancy is not understood but a possibility is that ParB2 when supplied from a trans source is much less effective. A discrepancy in the amount of protein required from a cis vs. trans source has been noted in the case V. cholerae ParA1 [22] and ParA of Pseudomonas aeruginosa [43]. The production of one of the Par proteins without its partner could also have altered the protein activity and stability. The importance of maintaining the stoichiometry of Par proteins has also been indicated in studies of B. subtilis [20], [44]. Another possibility is that higher protein concentration may be required to bind to a single parS site, as we have used here, than when there exists neighboring sites, as in the native host, that might allow cooperative binding.
We show that in wild type V. cholerae cells ParB2 can bind and spread over the entire origin (Figure 1). We detected considerable ParB2 spreading, even with the deletion of the origin proximal centromeric site (parS2-B) (Figure S7). Most likely, this spreading originates from the neighboring parS2-A site 5.7 kb away (Figure 1). The spreading could add an additional layer of control over PrctA by silencing the promoter, which is independently repressed by RctB (Figure S3) [24], [45]. The PrctA activity in turn controls RctB binding to the rctA 39-mer [29]. The multiple feedback loops that operate to control the initiation of replication from the origin of chrII appear securely interlocked with the specific segregation system of this chromosome. The presence of multiple layers of control could compensate for a deficiency in any one of the regulators, and help in homeostasis of origin copy number.
The finding that ParB2 could spread over the entire origin might suggest that it could be a mechanism to promote chromosome segregation. It might increase the effective size of the kinetochore, which might facilitate its interaction with ParA, the essential partner of ParB in chromosome segregation. However, this role has yet to be established [16], [34], [46].
ParB proteins of plasmids are known to be plasmid-specific and to bind to their cognate sites [47]. This helps to avoid segregation-mediated incompatibility if different plasmids happen to be present in the same host. By the same token, in multichromosome bacteria, the segregation systems should be chromosome-specific. Such is clearly the case in Burkholderia cenocepacia [5] and in V. cholerae [4], [48]. The same ParB protein has been found to bind to variant parS sites but the sites are believed to be descendants of a common ancestor [49]. In this context, it is noteworthy that although the central 39-mer is largely non-homologous to parS2-B, the region of the 39-mer crucial for ParB2 binding shares six bp of perfect identity with parS2-B (Figures 5,S4), suggesting the possibility of an evolutionary link between the sites here also.
Chromosome segregation begins soon after replication initiation, thereby compressing the total time for the completion of these two processes. Their close coordination also allows segregation to proceed in a more orderly fashion than if the substrate for segregation were a pair of completed and entangled sister chromosomes. Here, we have described interactions that might assist in coordinating replication initiation and segregation. In V. cholerae, following replication initiation, the majority of the RctB binding sites (11- and 12-mers) stay hemi-methylated and are unable to bind the initiator [50]. This stage of the cell cycle should favor spread of ParB2 into the origin (Figure S7), which is likely to favor origin segregation and at the same time discourage premature reinitiation. Spreading of ParB2 towards the origin is apparently prevented later in the cell cycle when the origin is remethylated, allowing RctB binding to 11- and 12-mers that eventually leads to initiation (Figure 3, 3rd panel). At these latter stages, when spreading is blocked, direct binding of ParB2 to the central 39-mer should favor initiation. Thus, depending upon the stage of the cell cycle, ParB2 appears to play opposite roles in controlling chrII replication, but in such a way as to promote the orderly sequence of chromosome replication followed by segregation. In the case of plasmids, which can complete replication in a tiny fraction of the cell division cycle, such coordination is neither necessary nor evident. We suggest that the acquisition of interactions such as we describe are a feature of the putative adaptation of an acquired plasmid to permanent residency as a second chromosome.
V. cholerae and E. coli strains, and plasmids used in this study are listed in Table S1. ChrII fragments were amplified from N16961 (CVC209) DNA by PCR using Phusion High-Fidelity polymerase (NEB, Beverly, MA). The sequences of primers used for PCR are shown in Table S2. For cloning sequences up to 100 bp, complementary oligonucleotides (IDT, Skokie, IL) were used after annealing the two [29]. The exact chrII coordinates of each cloned fragment are given in Table S1.
This was done in L broth cultures of E. coli strain BR8706 at OD600 between 0.4–0.5, as described [24]. To account for any effect that ParB2 might have on the replication of the lacZ-reporter plasmid, β-galactosidase activities were normalized for the plasmid copy number in all cases. The copy number variation was small; one standard deviation was within 20% of the mean. The copy numbers were measured (see below) from aliquots of the same cultures that were used for β-galactosidase measurements. Some of the cultures were simultaneously monitored for ParB2 amounts by Western blotting (Figure S2). Note that +/− ParB2 refer to cells carrying pTVC501 (that carries parB2 under IPTG control) with and without IPTG induction, respectively. In Figure 2 and Figure S3 (lanes 4,6), + refers to cells carrying pTVC236, which supplies ParB2 from a constitutive promoter and − refers to cells carrying the empty vector, pACYC184.
The copy number of lacZ-carrying plasmids (Figures 2,3, S3, S9) were measured exactly as described [32]. Briefly, different experimental cultures were grown to log phase and mixed with separately grown cells carrying pNEB193 before plasmid isolation. The latter plasmid helped to account for plasmid loss, if any, during plasmid isolation steps. The copy number of oriII plasmids (Figure 7) was determined similarly except that cells instead of growing in liquid cultures were obtained by washing out colonies from transformation plates directly, to avoid mutant accumulation, as described [26]. The origin fragments were first cloned in a plasmid vector driven by the γ-origin of plasmid R6K, and the clones were maintained in cells that supplied the cognate initiator (π) protein. The clones were electroporated into E. coli (BR8706) carrying pTVC499 that supplied RctB (but no π protein) and pTVC501 that supplied parB2.
The DNA probes were made from plasmids by PCR using oligonucleotides TVC286 (5′-TCCGATTACGGCACCAAATCGA-3′) and TVC287 (5′- AACGTGGATAAACTTCCTGTAAT-3′), which allowed amplification of extra 100 bp of vector sequences from each flank of the region of interest. The PCR products were labeled using 30 units T4 Polynucleotide Kinase (NEB) and 50 µCi of [γ32-ATP] (Perkin-Elmer) and purified by passing through G-50 columns (Roche diagnostics). Binding was done in the presence of 300 ng poly dI-dC. Other details are as described [25], except that the binding reactions were run in 0.5×TBE, which improved ParB2 binding. In Figure 5, the probe was non-radioactive and was visualized with SYBR Gold nucleic acid gel stain (Molecular Probes) at 0.5 mg/ml for 30 min. As non-specific competitor, supercoiled pUC19 DNA was used instead of poly dI-dC, as the former stayed at the top of gels and did not interfere with visualization of probe bands. The images were recorded using Fuji LAS-3000 imaging system.
ChIP assay was performed as described [50]. Briefly, cells of V. cholerae CVC209 were cultivated in L broth at 37°C to exponential phase and cross-linked with 1% formaldehyde. After cell lysis and sonication, RctB-DNA or ParB2-DNA complexes were immunoprecipitated using RctB or ParB2 antibody, respectively. The precipitated DNA was amplified, labeled and hybridized to a custom Agilent 8 X 60K V. cholerae oligonucleotide microarray representing the whole genome according to the manufacturer's protocol and as described [51]. The custom tiling array contained 60 bp probes specific for both the Crick and Watson strands. The consecutive probes were separated by 140 bp in each strand and by 10 bp between the Watson and Crick strands. Data was extracted using an Agilent scanner and Agilent Feature Extraction program. Individual ChIP (Cy5) and input (Cy3) signals were first normalized with respect to total Cy5 and Cy3 signals, respectively. Fold change was calculated by dividing the normalized Cy5 signals with normalized Cy3 signals. The values are mean from three independent experiments.
The effect of roadblock to ParB2 spreading was tested by cloning an array of five consensus P1 plasmid iterons in front of chrII origin of pTVC20, resulting in pBJH218 (Figure 8). Consensus iterons were used to avoid the PrepA promoter present within the array of natural iterons of P1ori. To clone the iterons, pTVC20 was modified by creating a NdeI restriction enzyme site between parS2-B and the 39-mer within rctA, using QuikChange II XL site-directed mutagenesis kit (Agilent Technologies) and oligonucleotides BJH472 and BJH473. To the resulting plasmid (pBJH217), the iteron array, amplified from pALA753 [52] using oligonucleotides BJH475 and BJH476, was cloned at the NdeI site, resulting in pBJH218. The plasmid pair, pTVC20 and pBJH218, was used in the same genetic background as used in Figure 7, and additionally, in an otherwise isogenic host that supplied P1RepA protein from a constitutive promoter (bla-p2 of pBR322). The P1repA gene and the adjoining bla-p2 promoter was cloned in a λD69 vector and the resulting phage (λDKC311 [53]) was used to lysogenize BR8706.
The copy numbers of pTVC20 and pBJH218 were nearly identical, about four-fold lower than pTVC22 used in Figure 7, as was found earlier for pTVC20 [29]. The copy number was estimated to be about one per cell when there should be four oriC. The low copy number of pTVC20 and pBJH218, and the lack of an active partitioning system, made the plasmids unstable (130/130 cells lost the plasmids after seven generations of growth without selection). The effect of ParB2 was therefore checked only under selection. Young colonies (≤1 mm) from selection plates were used to inoculate LB medium with appropriate drugs (ampicillin at 50 µg/ml, chloramphenicol at 25 µg/ml and spectinomycin at 40 µg/ml) and inducers (arabinose at 0.02% and, when desired, IPTG at 100 µM), and the cultures were grown to early log phase (OD600∼0.1) and stored in 0.01 M MgSO4 in ice. For calculating generation times, the cultures were diluted to OD600 = 0.002 and grown to early log phase. Generation times were calculated from OD600 values in the range 0.02–0.2. Saturation of growth was avoided to prevent accumulation of faster growing revertants. Relative colony sizes were also determined from the MgSO4 suspensions on plates with and without IPTG but otherwise identical in volume and contents.
|
10.1371/journal.pcbi.1006867 | Increasing consensus of context-specific metabolic models by integrating data-inferred cell functions | Genome-scale metabolic models provide a valuable context for analyzing data from diverse high-throughput experimental techniques. Models can quantify the activities of diverse pathways and cellular functions. Since some metabolic reactions are only catalyzed in specific environments, several algorithms exist that build context-specific models. However, these methods make differing assumptions that influence the content and associated predictive capacity of resulting models, such that model content varies more due to methods used than cell types. Here we overcome this problem with a novel framework for inferring the metabolic functions of a cell before model construction. For this, we curated a list of metabolic tasks and developed a framework to infer the activity of these functionalities from transcriptomic data. We protected the data-inferred tasks during the implementation of diverse context-specific model extraction algorithms for 44 cancer cell lines. We show that the protection of data-inferred metabolic tasks decreases the variability of models across extraction methods. Furthermore, resulting models better capture the actual biological variability across cell lines. This study highlights the potential of using biological knowledge, inferred from omics data, to obtain a better consensus between existing extraction algorithms. It further provides guidelines for the development of the next-generation of data contextualization methods.
| Genome-scale models of human metabolism have facilitated numerous exciting discoveries regarding human physiology and therapeutics. The accuracy of results from such studies requires that models capture the tissue or cell-type specific metabolism. In hopes to obtain accurate models, several algorithms have been developed to extract cell- or tissue-specific metabolic models. Each algorithm has provided useful insights into the metabolism of specific cell and tissue types. However, since each of these methods use different assumptions to guide reaction inclusion and removal, they result in considerable differences in size, functionality, accuracy, and ultimate biological interpretation, even when using the same data set. To overcome this, the enclosed research proposes an approach to infer the functionalities of a cell or tissue from omics data, and then protect these functions to guide the construction of a context-specific model. Through this study, we highlight the value of using experimental data to help infer the set of metabolic functions that should be included in a model, in an effort to obtain greater consensus across existing extraction algorithms. This study further provides guidelines for the development of the next-generation of data contextualization methods.
| Genome-scale metabolic models (GeMs) have been widely used for model-guided analysis of large omics datasets, since they provide cellular context to these data by establishing a mechanistic link from genotype to phenotype. GeMs include all reactions in an organism. Since not all enzymes are active in each cell type or culture condition, algorithms have been developed to build context-specific models using omics data to recapitulate the metabolism of specific cell types under specific conditions [1,2]. These algorithms have provided useful insights in the metabolism of specific cell and tissue types [1,3–10]. However, since each method uses different assumptions to guide reaction inclusion and removal, they result in considerable differences in size, functionality, accuracy, and ultimate biological interpretation, even when using the same data set [1,2,11].
The poor consensus in generated models requires increased caution in the interpretation of model-derived hypotheses of how metabolism is used under specific environments. Indeed, most generated models, upon construction, will be missing known metabolic functions and this varies considerably for models built using different approaches [1]. To gain confidence in model predictions and reconcile the differences across approaches, users can enforce the inclusion of known metabolic capabilities in the model. In this regard, the tINIT extraction algorithm introduced the possibility to enforce the capacity of context-specific models to represent some cellular functionalities by using a list of metabolic tasks known to occur in all cell types [12]. However, this protectionist approach requires one to know and predefine the functionalities of a specific cell line, tissue, or context.
To overcome this, we propose an approach to infer the functionalities of a cell or tissue from omics data, and then protect these functions to guide the construction of a context-specific model. To this end, we curated and standardized published lists of metabolic tasks [13,14], resulting in a collection of 210 tasks covering 7 major metabolic activities of a cell (energy generation, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism). We also developed a framework to directly predict the activity of these functionalities from transcriptomic data and subsequently use these for a protectionist approach to several existing extraction algorithms. Models resulting from this approach should more comprehensively capture the unique metabolic functions of a given cell type. We evaluated the validity and variation across models built with this approach, coupled to existing context-specific extraction methods. Specifically, we constructed hundreds of models for 44 cancer cell lines in which we built the models using standard approaches or protected a list of metabolic functions that have been inferred from the original transcriptomic data of each cell line. We also varied the reference human reconstruction and algorithms employed for the generation of cell line specific models, using two different reference models (iHsa [13] and Recon 2.2 [15]) and 6 different algorithms (mCADRE [16], fastCORE [5], GIMME [6], INIT [7], iMAT [4], and MBA [8]). We compared the sets of extracted models at the level of reaction content, metabolic functions, and capacity to predict essential genes identified in CRISPR-Cas9 loss-of-function screens. Through this study, we highlight the value of using experimental data to help infer the set of metabolic tasks that should be included in a model, in an effort to obtain greater consensus across existing extraction algorithms.
We built models from Recon 2.2 [17] and iHsa [13] using six model extraction methods (MEMs: mCADRE, fastCORE, GIMME, INIT, iMAT, MBA) for 44 different cell lines from the NCI-60 panel (S1 Table; 15 cell lines were not used due to the absence RNA-Seq data in [18] for these cell lines). Uptake and secretion rates of the input GeMs were quantitatively constrained using a list of experimentally measured metabolites (S2 Table)[19,20]. Furthermore, a biomass function, consisting of 56 metabolites required for growth, was added and constrained to the experimentally measured growth rate of the cell lines (S3 Table). The biomass function and constraints from exometabolomic data introduced in the GeMs were implemented as described in [1]. The extraction process of cell line specific models was done based on RNA-Seq data [18] to specify active genes in each cell line. Details on the implementation of MEMs tested and the preprocessing of gene expression data for the definition of gene activity are provided in the Methods section.
To assess the relative impact of algorithm and data source on model content, we conducted a principal component analysis (PCA) of the reactions in all models for each reference GeM. As observed previously [1,11], the decisions regarding algorithm choice significantly impact the content of our cell line-specific models. The first principal (PC1) component explains 38% of the overall variance in model reaction content, with >60% of the variation in PC1 explained by the choice of model extraction method (Fig 1A and 1B). Indeed, the different algorithms yielded cell line-specific models that varied considerably in size, with few reactions common to all models extracted from either Recon 2.2 or iHsa (Fig 1C, Fig A in S1 Text). Even among models extracted using the same algorithm, there is non-negligible variability in model reaction content (Fig 1D). This leads to the generation of models that are substantially different with respect to the cell line considered, while the transcriptomic data used to tailor the GeMs shows high consistency across most cell lines (Fig 1E).
Model reaction content is often evaluated to compare context-specific algorithms. Recently, approaches to benchmark models with their functionalities have been proposed [1,12,21]. Current approaches use repositories of known cellular tasks to assess the capacity of models to achieve specific modeling goals or to enable the representation of specific metabolic functions. This idea of assessing the quality of a metabolic network reconstruction using biological knowledge was introduced in Recon 1 through the characterization of the “human metabolic knowledge landscape” [22]. However, the concept of “metabolic tasks” (Fig 2A) was clearly defined in 2013 by Thiele and coworkers [14] to benchmark the improvements of Recon 2 compared to Recon 1, wherein they stated that “a metabolic task is defined as a nonzero flux through a reaction or through a pathway leading to the production of a metabolite B from a metabolite A”. Since then, additional lists of tasks have been published. To standardize these and develop a framework for their easy use with GeMs, we curated the existing lists of metabolic tasks (13,14) and obtained a collection of 210 tasks covering 7 major metabolic activities of a cell (energy generation, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism) (Fig 2B and 2C, S4 Table). We evaluated the task collection using genome-scale metabolic models for human [13,14,17,22,23], CHO cells [9], rat [13] and mouse [24] (Fig 2D, S5 Table). Details on our proposed formalism of the metabolic tasks and the associated computational framework for their use are presented in the Methods.
Metabolic tasks can be used to compare the performance of models extracted from different reference GeMs. As observed at the level of the reaction content, the extraction method strongly influences the model functions (explaining >50% of the overall variance in the first PC; Fig 3A). However, the reference model is the most prominent factor in the second PC underlying a non-negligible influence of this variable in the extraction process. This is mainly due to differences in gene, protein, reaction association (GPR) annotations and reaction content between Recon 2.2 and iHsa. Interestingly, Recon 2.2 captures more metabolic functions with fewer reactions (Fig 3B). However, the number of successful tasks increases proportionally with the number of reactions in a model. Furthermore, as the extraction method used influences the number of reactions removed, distinct patterns are seen from the ratio of the number of metabolic tasks to the number of reactions introduced by the different algorithms (Fig 3C). As for the reaction content, the number of tasks retained in each model varies substantially, depending on the cell lines considered. Surprisingly, only 8% of the tasks are present in all models (Fig 3D, Fig B in S1 Text), thus highlighting the large variation in metabolic functions a model will have, depending on algorithm choice.
We inferred active metabolic tasks directly from transcriptomic data using the whole genome-scale model. To this end, we computed the list of reactions associated with each task and used the GPR rules to determine the gene expression levels associated with each of these reactions. A metabolic score is attributed to each task by using the mean activity level of each reaction (Fig 4A; See Methods). We found that more than the half of the tasks should be conserved across all cell lines (Fig 4B), which is far more than those active using the algorithms in their standard format (i.e., without protecting tasks). Therefore, we generated a new set of models, wherein we also enforced the inclusion of reactions associated with tasks inferred for each of the 44 different cell lines (S7 Table). We focused on MBA-like algorithms (i.e., MBA, fastCORE and mCADRE), since they are directly amenable to use the protectionist approach with minor modifications to the algorithms (Fig 4C). Indeed, other algorithms do not ensure the inclusion of a reaction even if it is enforced. For example, iMAT relies on the definition of a core set of high-confidence reactions, but core reactions can be removed if it depends on many non-expressed non-core reactions (Fig 4C). See Methods for a detailed description of the implementation of the protectionist approach for each algorithm.
For equivalent extraction setups (i.e., same reference model, extraction method, and cell line), the number of reactions included in the extracted model was not considerably influenced by the protection of the metabolic task, while the number of active tasks clearly increases (Fig C in S1 Text). We performed PCA of the reaction content and the metabolic functions of the models with protected tasks. We observed that the protection of metabolic tasks inferred from data significantly decreased the influence of the extraction method on the final model content. However, the use of this approach remains sensitive to the choice of the reference model (Fig 5A and 5B). The reduction of the variability of model with respect to the extracted methods used can be explained by the increased number of shared tasks across cell lines, all supported by the transcriptomic data. This is seen in particular for tasks involved in amino acid and lipid metabolism (Fig D in S1 Text). Furthermore, we observed that the variation in model content was better explained by the cell lines (Fig E in S1 Text). Actually, the task protection increases the similarities between context-specific models with respect to the cell line (Fig F in S1 Text) but also with respect to the transcriptomic data (Fig G in S1 Text). Finally, all the models now share more than 64% of the metabolic tasks (Fig 5C).
Beyond model content, we evaluated how the protectionist approach influenced model predictions. Thus, we analyzed the influence of protecting inferred tasks on gene-essentiality predictions (i.e. prediction of the genes whose knockdown leads to a growth impairment). We systematically deleted each gene in all generated models, and then used flux balance analysis to test models for normal or impaired growth. Gene deletions associated with impaired growth are considered as essential. We observe that task protection reduces the number of genes predicted to be essential for all thresholds considered (i.e., percentage of the maximum wild type growth rate) for the various extraction methods used on both reference models (Fig 5D; Fig H in S1 Text).
We further evaluated the accuracy of essentiality predictions by comparing these to CRISPR-Cas9 loss-of-function screens for 20 cell lines [25–27]. In these screens, essential genes are identified based on gene scores attributed using single guide RNA (sgRNA) abundance for each knockout before and after growth selection. Gene scores that are more negative have a higher probability of being essential. Therefore, the agreement between model predictions and the CRISPR screen data can be quantified as the percentage of predicted essential genes that have a negative gene score [28]. Furthermore, the significance of the improvement gained from protecting data-inferred metabolic tasks can be computed using a 1-tailed Wilcoxon test. Consistent with previous reports [1,29], we found that the models, without protecting metabolic tasks, correctly predicted many essential genes. However, overall, the protectionist approach provided a small but significant improvement to gene-essentiality predictions (up to 5% improvement; Fig 5E; Fig I in S1 Text). However, the task protection reduced the number of predicted essential genes, which increased the proportion of true positives and reduced the number of false positives.
To further assess the identity of these true positives provided by task protection, we compared our model simulations to a collection of known anti-cancer drug targets (S11 Table) [30]. In this analysis, we found the protectionist approach better captured the gene essentiality related to known drug targets (Fig J in S1 Text). We further tested if the models contain cancer hallmark genes [31] and that models built using the protectionist approach increased the proportion of genes in the models that are associated with cancer hallmarks (Fig K in S1 Text; S12 Table).
Here we generated hundreds of models for 44 cell lines from the NCI-60 panel using multiple MEMs and two reference GeMs (Recon 2.2 and iHsa) using standard approaches or by protecting metabolic tasks that have been directly inferred from transcriptomic data. We presented a comparative analysis of these two sets of models. As previously observed, the analysis of the first set of extracted models (i.e., models generated without protecting metabolic functions) indicated that the choice of model extraction algorithm significantly influenced the model content at the reaction level [1,2,11]. This leads to considerable variability in context-specific model content, which dwarfed the biological variability across cell lines, otherwise seen in their transcriptomes.
We provided here a curated list of 210 tasks that were used to compare the functionalities of the extracted models. The evaluation of metabolic tasks has emerged as a valuable practice in metabolic modeling studies [12–14,22,32–35]. Such an approach allows one to evaluate the capacity of models to achieve specific modeling goals by capturing known metabolic features. Here we also demonstrated that the approach allows one to objectively compare models that may not share the same structure, such as different reference network reconstructions or models that have been extracted using different methods or parameters. We demonstrated that the selection of a reference model can significantly impact the resulting metabolic functions captured by extracted models, thus possibly impacting the results and interpretations from modeling studies. Indeed, the comparison of the functions of models extracted from both Recon 2.2 and iHsa demonstrated the non-negligible influence of these reference models. We found this is principally due to differences in the GPR annotations in both GeMs. However, these differences in GPR annotations do not considerably influence the inference of metabolic tasks from transcriptomic data. The functional similarity across cell lines captured using data-inferred metabolic tasks is highly consistent between both reference models (Fig L in S1 Text). While community initiatives to standardize the formal representation of GeMs will facilitate cross-comparison between diverse existing GeMs [36], these results highlight the potential of using the inference of functionalities directly from the transcriptome as a way to increase the consensus between extraction methods and reference models.
One challenge in the evaluation of metabolic models is the difficulty of comprehensively defining metabolic functions from a manual search of the literature. Thus, another strength of our approach is that it decreases the need for a priori knowledge or assumptions of the metabolic functions that should be included when building a cell or tissue specific model. Therefore, this list of metabolic tasks provides a framework for modelers to develop more physiologically accurate models by inferring the activity of metabolic tasks directly from omics data. Thus, key reactions that need to be included in a model can be protected, without requiring one to know what the cell does. However, the resulting models should still be curated to evaluate expected functionalities, such as for example auxotrophies.
Our protectionist approach can be implemented with diverse model extraction algorithms since it only requires the algorithms to prevent the removal of active metabolic tasks during the extraction process. However, some algorithms will require modifications to ensure the protection of all reactions related to a task. Current implementations of the GIMME-like and iMAT-like families do not favor this type of protection. By minimizing flux through reactions associated with low gene expression, GIMME-like extraction methods may remove low expression reactions one would want to retain for a validated metabolic task if there are high expression reactions that allow for growth. The iMAT-like methods are similar as they rely on finding an optimal trade-off between removing reactions associated with low gene expression, and keeping reactions whose genes/enzymes are highly expressed. Thus, modified implementations of these algorithms will be needed to allow the protection of reactions based on experimental observations. Finally, this approach can also be extended to any type of network complexity reduction that have been developed in the metabolic modeling field, such as the MILP-based approaches developed to tailor models based on exometabolomic data [37,38].
In our work, we also demonstrated that the models built with the protectionist approach are able to better capture cell-type specific metabolism and accurately predict many essential metabolic genes. Thus, these models may be invaluable for drug development strategies. The emergence of experimental techniques to assess the genetic vulnerabilities of a cell (e.g., CRISPR-Cas9, RNAi) allows researchers to identify sets of genes that should be essential for growth maintenance. These essential genes can further be used to evaluate the capacity of models to represent the interdependence between down-regulation of a gene and the concomitant impairment of growth. Thus, models can be used for interpreting the mechanisms underlying metabolic vulnerabilities that may be invaluable for new drug discoveries. Furthermore, many of the metabolic changes occurring in certain diseases, such as cancer, can be captured by the current list of tasks. Since many of the metabolic tasks are shorter paths, cases where the metabolic flux is redirected due to disease-related metabolic perturbation might be captured by the specific collection of tasks computed using our method. Finally, for the rare cases where a mutation to a specific enzyme leads to a change in the metabolic reaction catalysed by the enzyme change, such as mutations in isocitrate dehydrogenase leading to the production of oncometabolites [39–41]. With such knowledge, researchers are able to define such changes as new metabolic tasks associated with the mutations and incorporate them into their models.
Finally, the list of tasks presented in this study was constructed based on existing repositories. However, a community effort could be undertaken to extend the scope and the definition of these metabolic functions, including the development of tasks seen in plants and microbes and tasks associated with secondary metabolism and microbial gene clusters [42,43]. Furthermore enzymatic mutations leading to new metabolic functions [39] can be systematically defined and added, as currently efforts in the constraint-based modeling community do so on a model by model basis [41]. As these tasks are connected to their associated gene products, this repository of curated tasks would facilitate the description of genome-scale metabolic reconstructions as more than a network of reactions but rather as interconnected maps of cellular functions for diverse organisms. This would be invaluable for the development of algorithms using more relevant biological information and facilitate more comprehensive and accurate descriptions of metabolic adaptations that occur in cells facing a change of context.
Conclusively, context-specific extraction methods are powerful approaches that provide insights in the metabolic state of a cell in specific environments. However, the underlying assumptions used to tailor the GeM based on omics data vary across algorithms, with the consequence that drastically different models can be obtained based on the same data. The poor consensus in generated models may limit the use of context-specific methods for data-driven hypotheses. The definition of metabolic tasks can help with these concerns. Our curated list of tasks and computational framework will allow users to infer metabolic functions directly from transcriptomic data using the whole genome-scale model, and drive the development of improved context specific models. Such models will pave the way toward a better consensus between existing context-specific extraction algorithms, and facilitate the application of models for novel biomedical and engineering applications.
RNA-Seq data for the 44 cell lines from the NCI-60 panel were downloaded from [18]. We processed the gene expression data to attribute a gene activity score for each gene and define which genes are active in each cell line. A gene is defined as active in a sample if its expression value is above a threshold defined for this gene within the dataset considered. The threshold of a gene is defined by the mean value of its expression over all the samples coming from the same dataset with exceptions that the threshold needs to be higher or equal the 25th percentile of the overall gene expression value distribution and lower or equal to the 75th percentile. The gene score is computed as follows:
GeneScore=5∙log(1+ExpressionlevelThreshold)
These gene scores are mapped to the models by parsing the GPR rules associated with each reaction. The gene score for each reaction is selected by taking the minimum expression value amongst all the genes associated to an enzyme complex (AND rule) and the maximum expression value amongst all the genes associated to an isozyme (OR rule) [44]. Note that we have recently benchmarked the influence of preprocessing methods on the definition of the set of active genes and observed that this parameter combination presented the best performance [45].
Model extraction methods (MEMs) employ diverse algorithms to extract cell line- or tissue-specific models from a GeM. The MEMs we have considered can be categorized into three families: “GIMME-like” (i.e., GIMME), “iMAT-like” (i.e., iMAT and INIT) and “MBA-like” (i.e., MBA, FASTCORE, and mCADRE), as proposed previously [2]. The GIMME-like family minimizes flux through reactions associated with low gene expression. The iMAT-like family finds an optimal trade-off between removing reactions associated with low gene expression, and keeping reactions whose genes/enzymes are highly expressed. In the MBA-like family, the algorithms use sets of core reactions that should be retained and active, while removing other reactions if possible. All the algorithms used in this study have been implemented using the function createTissueSpecificModel available in the COBRA Toolbox 3.0 [46]. We describe below the list of required parameters needed to run the different methods, all optional parameters have been kept to their default setting.
FASTCORE [5]—The core reactions set (options.core) is determined by all the reactions associated to a gene score superior to 5log(2). Note that the biomass reaction was added to the core reactions sets.
GIMME [6]—The implementation of GIMME requires two parameters: the gene scores (options.expressionRxns) and a threshold value, the reactions associated with a gene score value below this threshold will be minimized (options.threshold = 5log(2)). Note that we manually attributed a gene score of 10log(2) to the biomass reaction to ensure its inclusion.
iMAT [4,47]—Three parameters need to be provided to run iMAT: the gene scores (options.expressionRxns), a lower threshold value (reactions with gene score below this value are considered as “non-expressed”) and a upper threshold value (reactions with gene score above this value are considered as “expressed”). To simplify the comparison across algorithms, we set both thresholds to the same value: options.threshold_lb = options.threshold_ub = 5log(2), as done in a previous benchmarking study (1). Note that we manually attributed a gene score of 10log(2) to the biomass reaction to ensure its inclusion.
INIT [7]—The implementation of INIT requires attributing positive weights (options.weights) to each reaction with high expression and negative weights for the ones with low expression. All the reactions associated with a gene score below 5log(2) have been assigned a weight of -8 while the weights of remaining reactions were defined as the ratio between the gene score for each reaction and 5log(2). The weight associated with the biomass reaction was put to the maximum of obtained reaction weights.
MBA [8]—The implementation of MBA requires the definition of two set of reactions: high confidence (options.high_set) due to their expression and others with medium confidence (options.medium_set). The set of reactions with high confidence is defined as reactions with a gene score above the 75th percentile of the distribution of all gene scores and the medium confidence set by all the reactions presenting score above 5log(2) and below the 75th percentile of the distribution of all gene scores. Note that the biomass reaction has been manually added to the high confidence set of reactions.
mCADRE [16]—The implementation of mCADRE requires a score quantifying how often a gene is expressed across samples (options.ubiquityScore) and a literature-based evidence score (options.confidenceScores). Since the confidence score identification used in the original paper is difficult to transpose in this study, we did not define the confidence score as preformed in the tutorial presenting the implementation of mCADRE in COBRA Toolbox 3.0 (46). Furthermore, as the gene scores are computed based on the knowledge of the gene expression of a gene across all samples, we used the gene scores as ubiquity scores.
The curation has been done by first taking the union of previously published lists of metabolic tasks [13,14]. We removed duplicated tasks and lumped tasks that rely on the description of similar metabolic functions. Each remaining task without strong biological evidence was removed. We also created 9 new tasks that were essential for the acquisition of already described metabolic functions (i.e., intermediate biosynthetic steps for the acquisition of other tasks). Doing so, we obtained a collection of 210 tasks associated with 7 systems (energy, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism). For each task, we provided its original source (Recon and/or iHsa) and comments on the biological evidence of this metabolic function (S4 Table).
In its original version, Thiele and coworkers (2013) [14] define a
“metabolic task as a nonzero flux through a reaction or through a pathway leading to the production of a metabolite B from a metabolite A. The metabolic capacity of the network was demonstrated by testing nonzero flux values for these metabolic tasks. For each of the simulations, a steady-state flux distribution was calculated. Each metabolic task was optimized individually by choosing the corresponding reaction in the model, if present, as objective function and maximized the flux through the reaction”.
In parallel, Agren and coworkers presented an alternative framework to compute the metabolic tasks present in a model within their RAVEN toolbox [48]. They defined a metabolic task through a list of inputs and outputs for which the pseudo-stationary assumption will be relaxed following a magnitude imposed by the user and assumed that a task successfully passes if the variation imposed to the inputs leads to the imposed variation of the outputs.
We also propose to define a metabolic task as the capacity of producing a defined list of output products when only a defined list of input substrates is available. However, we modified the way to implement it from the RAVEN toolbox. Instead of relying on the relaxation of the steady-state assumption, we take an approach more similar to that proposed by [14] by imposing constraints only at the flux level. Therefore, a model successfully passes a task if the associated LP problem is still solvable when the sole exchange reactions allowed carrying flux in the model are temporary sink reactions associated with each of the inputs and outputs listed in the task. This framework allows the use of known stoichiometry to fix the ratio between the fluxes of the sink reactions associated with each input and output of the task. We implemented the code to compute the tasks in Matlab, and the code, checkMetabolicTasks, has been contributed to the COBRA Toolbox3.0 [46].
We tested the list of tasks using published genome-scale models of human [13,14,17,22,23], Chinese hamster [9], rat [13] and mouse [24] cells (Fig 2D, S5 Table). All models successfully pass more than 90% of the tasks. For each failed task, we provided a reason of the failure (i.e. definition of the missing reaction to successfully pass the task) (S5 Table). As the definition of the metabolic tasks depends on the provision of the exact name of the metabolites in each model, we also provide a table of nomenclature compatibility between the different genome-scale models tested (S6 Table).
We developed a computational framework for attributing a score to each metabolic task in order to extend the application of the concept beyond the model benchmarking scope. If a task successfully passes in a model, one can compute the list of reactions associated with this task and, in doing so, access the list of genes that may contribute to the acquisition of this metabolic function based on the GPR rules. To this end, we used the parsimonious Flux Balance Analysis (pFBA) algorithm to define the set of reactions and associated genes required to pass a task within a specified model [49]. Thanks to the availability of this information, metabolic functions can now be directly assessed from transcriptomic data. The proposed computation of a metabolic score relies first on the preprocessing of the available transcriptomic data and the attribution of a gene activity score for each gene (see associated Methods section). We further used the GPR rules associated with each reaction required for a task to decide which gene will be the main determinant of the enzyme abundance associated with this reaction and attribute the corresponding gene activity level (i.e., selection of the minimum expression value among all the genes associated to an enzyme complex (AND rule) and the maximum expression value among all genes associated with an isoenzyme (OR rule)). Therefore, each reaction involved in a task is associated with a reaction activity level (RAL) that corresponds to the preprocessed gene expression value of the gene selected as the main determinant for this reaction. Finally, the metabolic score can be computed as the mean of the activity level of each reaction:
MTscore=sum(RAL)/numberofreactionsinvolvedinthetask
Doing so, a metabolic task will be considered as active if its MT score has a value greater than 5log(2). The list of active metabolic tasks for each of the 44 cell lines from the NCI-60 panel is available in S7 Table.
We used the list of active metabolic tasks (S7 Table) to determine the set of reactions that should be protected during the extraction process for each of the 44 cell lines. The protectionist approach has been implemented for each extraction method by using the same set of parameters as previously described with the following modification:
FASTCORE—The set of reactions associated with the metabolic tasks defined as active based on the transcriptomic data has been manually added to the core reactions set (options.core).
GIMME & iMAT—A gene score of 10log(2) (options.expressionRxns) has been attributed to all the reactions associated to the metabolic tasks defined as active based on the transcriptomic data.
INIT—The weights (options.weights) for all reactions associated with the metabolic tasks defined as active based on the transcriptomic data were put to the maximum of obtained reaction weights.
MBA—The reactions associated with the metabolic tasks defined as active based on the transcriptomic data have been manually added to the high confidence set of reactions.
mCADRE—A ubiquity score (options.ubiquityScore) of 1 has been attributed to all the reactions associated to the metabolic tasks defined as active based on the transcriptomic data.
For the reaction PCAs, a binary matrix is constructed in which each row represents an extracted model and each column represents a reaction, with each element representing the presence (1) or absence (0) of a reaction in a model. Reactions in all or no models were removed from the matrix. Similarly for the metabolic function PCA, the matrix had each row as an extracted model and each column as a metabolic task, with each element in the matrix representing if the task is present (1) or absent (0) in a model. For the PCAs, the matrix was centered to have zero mean within each row. PCA was done on this matrix. The variance explained by the different factors (MEM, cancer type and cell line) within each of the principal components is calculated as follows. Within one factor, the maximum Pearson correlation coefficient (R) of the component scores and categories is calculated across all possible orderings of the categories. Reported is the R2 scaled to percentages. The same procedure was used to perform the PCA on the model functionalities except that the binary matrix of reactions was replaced by the binary matrix representing the list of metabolic tasks that are successfully passed in each extracted model. The attributes of all extracted models (number of reactions and metabolites, number of successfully passed tasks and predicted growth rate) are available in S8 Table and the results of the extracted model benchmarking using the list of metabolic tasks is available in S9 Table.
To predict gene-essentiality, FBA was used to optimize biomass production following the removal of each reaction in the cell line-specific models that would be affected by gene removal based on the GPRs. The function used to perform this deletion analysis is available in COBRA Toolbox 3.0, singleGeneDeletion.m [46]. To test these essentiality predictions of the models against experimental data, we downloaded CRISPR-Cas9 loss-of-function screens data for 20 NCI-60 cell lines from depmap.org [25–27]. In these screens, essential genes are identified based on genes scores attributed using single guide RNA (sgRNA) abundance for each knockout before and after growth selection. A more negative gene score suggests a higher probability that the gene is essential. Therefore, the agreement between prediction and data can be analyzed by using the percentage of predicted essential genes that have a negative gene score [28]. A 1-tailed Wilcoxon rank sum test was used to test whether the percentage of predicted essential genes of the model extracted using the protectionist approach were significantly higher than the ones without protection. The results of the gene deletion study and prediction against CRISPR-Cas9 loss-of-function screens are available in S10 Table.
|
10.1371/journal.ppat.1002462 | CNS Recruitment of CD8+ T Lymphocytes Specific for a Peripheral Virus Infection Triggers Neuropathogenesis during Polymicrobial Challenge | Although viruses have been implicated in central nervous system (CNS) diseases of unknown etiology, including multiple sclerosis and amyotrophic lateral sclerosis, the reproducible identification of viral triggers in such diseases has been largely unsuccessful. Here, we explore the hypothesis that viruses need not replicate in the tissue in which they cause disease; specifically, that a peripheral infection might trigger CNS pathology. To test this idea, we utilized a transgenic mouse model in which we found that immune cells responding to a peripheral infection are recruited to the CNS, where they trigger neurological damage. In this model, mice are infected with both CNS-restricted measles virus (MV) and peripherally restricted lymphocytic choriomeningitis virus (LCMV). While infection with either virus alone resulted in no illness, infection with both viruses caused disease in all mice, with ∼50% dying following seizures. Co-infection resulted in a 12-fold increase in the number of CD8+ T cells in the brain as compared to MV infection alone. Tetramer analysis revealed that a substantial proportion (>35%) of these infiltrating CD8+ lymphocytes were LCMV-specific, despite no detectable LCMV in CNS tissues. Mechanistically, CNS disease was due to edema, induced in a CD8-dependent but perforin-independent manner, and brain herniation, similar to that observed in mice challenged intracerebrally with LCMV. These results indicate that T cell trafficking can be influenced by other ongoing immune challenges, and that CD8+ T cell recruitment to the brain can trigger CNS disease in the apparent absence of cognate antigen. By extrapolation, human CNS diseases of unknown etiology need not be associated with infection with any particular agent; rather, a condition that compromises and activates the blood-brain barrier and adjacent brain parenchyma can render the CNS susceptible to pathogen-independent immune attack.
| There are many CNS diseases, including multiple sclerosis and amyotrophic lateral sclerosis, which have an inflammatory component, though no direct link has been established between incidence and a CNS-resident infectious agent. We reasoned that peripheral immunogens could play a role in CNS disease by inducing an immune response that is “mis-targeted” to the brain. This hypothesis was based on the immunological principle that, while education and activation of naïve cells is an antigen-driven process, recruitment is primarily antigen-independent. We developed a viral co-infection model using measles virus (MV) as a CNS activator and recruiting signal and lymphocytic choriomeningitis (LCMV) as a peripheral immune response initiator. Co-infection with both viruses resulted in significant morbidity and mortality, coincident with LCMV-specific CD8+ T cell trafficking to the brain. Death occurred due to edema, despite an apparent absence of LCMV antigens within the brain, and pathogenesis was CD8+ T cell-dependent, but perforin-independent. Thus, recruitment of peripherally activated CD8+ T cells to the CNS can potentiate neuroinflammation. This work raises the possibility that concomitant immune challenges may be an important cause of the neuroinflammation of some human CNS diseases, perhaps accounting for the inability to identify a discrete pathogenic trigger within affected brain tissues.
| Despite the exquisitely specific activation of the adaptive immune response following antigenic encounter, recruitment of immune cells to the affected site is governed by relatively nonspecific factors, including chemokine gradients and adhesion molecule induction on barrier endothelia [1]–[3]. Indeed, some studies have shown that activated immune cells can be recruited to a tissue where no cognate antigen exists. For example, using a mouse model of influenza infection, it was shown that primed transgenic CD4+ T cells that were specific for ovalbumin (with no cross-reactivity to flu) migrated efficiently to the infected lung. Despite such recruitment, these cells did not proliferate [4], showing that T cell recruitment and proliferation can be uncoupled.
The complexity of concurrent immune challenges that humans are likely to encounter is staggering, including myriad combinations of pathogens, allergens, and vaccines. In fact, many human and animal diseases are caused by polymicrobial exposures, including human pneumonia, otitis media, peritonitis and periodontitis. Other diseases, such as hepatitis and Lyme's disease, though caused by a single pathogen, can have exacerbated symptoms when combined with a second pathogen [5], [6]. In light of the observed antigen-independent recruitment of activated immune cells, an understanding of the trafficking and function of immune cells beyond the traditional “single pathogen challenge” approach that most viral pathogenesis studies employ is paramount. Specifically in this report, we asked whether recruitment of activated immune cells to virus-negative tissues occurs in individuals who are challenged simultaneously with multiple pathogens/antigens of differing tropism, and if so, whether this affects the pathogenic outcome.
The studies reported here focus on the consequences of immune cell recruitment into the CNS, as the unique environment of the brain (e.g., restricted opportunity for inflammation, nonrenewable cell populations) may make this organ system particularly vulnerable. Moreover, given the number of CNS diseases of unknown etiology that have an inflammatory component, this work may be relevant to future translational efforts to deduce the origins of such conditions.
As a model, we infected permissive mice with neuron-restricted measles virus (MV) and peripherally-restricted lymphocytic choriomeningitis virus (LCMV). Measles is a member of the family Paramyxoviridae; natural infection is restricted to humans and primates. The identification of CD46 and SLAM as human MV receptors allowed for the establishment of transgenic mice that express these receptors in specific tissues. In NSE-CD46+ mice, CD46 receptor transcription is restricted to CNS neurons by control of the neuron-specific enolase (NSE) promoter [7]. Neuronal infection occurs in all NSE-CD46+ mice [8], though CNS disease occurs only in neonates and immunocompromised adults. Adult immunocompetent mice clear the infection without pathogenic consequences, chiefly via interferon gamma (IFN-γ) [9].
Lymphocytic choriomeningitis virus is a member of the family Arenaviridae, a natural pathogen of mice, and a historically important tool in uncovering basic aspects of viral immunity [10], [11]. One of the notable attributes of LCMV is the diversity of consequences that can be achieved in infected mice by varying viral parameters (including the route of inoculation, dose, and viral strain), as well as host features (including mouse strain, age, and immunocompetence). Two of these outcomes are relevant here. Following intraperitoneal (IP) inoculation of adult, immunocompetent mice, LCMV establishes infection of peripheral organs including the spleen, liver, and kidney. By 4-6 days post-infection (dpi), the host mounts a robust T cell response (principally CD8+ T cells, with 80–95% of the splenocytes being specific for LCMV [12]) that resolves the infection within 7–10 days in the absence of overt disease. In contrast, delivery of as few as 1 plaque forming unit (PFU) of LCMV into immunocompetent adults by an intracerebral (IC) route results in lethal choriomeningitis within 6–7 dpi [13]. By this route, LCMV rapidly establishes infection of the membranes of the brain, including the meninges, leptomeninges, and ependyma, as well as the cerebro-spinal fluid (CSF)-producing choroid plexus cells within the ventricles. As with the peripheral infection, virus-specific CD8+ T cells rapidly expand and migrate to the infected tissues. In the case of an LCMV IC infection, these cells migrate to the CNS, where they then cause damage to the ventricular membranes, resulting in edema and precipitous uncal herniation [14].
In this study, we have taken advantage of the unique tissue tropism of transgene-restricted neuronal infection with MV and route of inoculation-restricted peripheral infection with LCMV to evaluate the pathogenic consequences following the recruitment of activated, peripheral LCMV-specific lymphocytes to the CNS, despite a lack of LCMV infection in this tissue. We show that upon concomitant infection with these two viruses, activated immune cells specific for LCMV are recruited to a “primed” CNS where they trigger neurological disease in the absence of LCMV infection.
This study was carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was reviewed and approved by the Fox Chase Cancer Center Institutional Animal Care and Use Committee (Office of Laboratory Animal Welfare assurance number: A3285-01).
Inbred NSE-CD46+, NSE-CD46+ mice on a recombinase-activating gene-2 deficient background (RAG-2 KO) [15], and NSE-CD46+ mice on a perforin deficient background (PFN KO)[16], were maintained in the closed breeding colony at Fox Chase. All animals were on the H-2b background. LCMV Armstrong (LCMV-Arm; ATCC) was passaged in BHK-21 fibroblasts and plaque purified, and titers were determined on Vero fibroblasts. MV-Edmonston (MV-Ed; ATCC) was also passaged and titered on Vero cells.
Mice were infected IC along the midline with the indicated PFU of LCMV-Arm, MV-Ed, or phosphate buffered saline (PBS) as a control, in a total volume of 15–20 µl using a sterile 27-gauge needle. LCMV was delivered by an IP route in a volume of 200 µl using a sterile 25-gauge needle. All mice were anesthetized with metofane prior to inoculation and were monitored daily for signs of illness. For some experiments, weight change was determined by establishing the pre-infection baseline and subsequently calculating weight gain or loss. Moribund mice that lost >20% of their original body weight were euthanized. For some experiments, virus was inactivated by exposure to UV for 15 m; inactivation was verified by subsequent plaque assay. For CD8 depletion experiments, 150 µg of CD8 depleting antibody (purified at the FCCC hybridoma facility from hybridoma clone 2.43 [ATCC #TIB 210]) was injected IP 1d prior to infection and every 7d thereafter throughout the course of study.
To isolate RNA from infected tissues, brains were snap-frozen in liquid nitrogen and homogenized in TriReagent (Sigma). Thereafter, RNA was purified and quality-tested by gel electrophoresis. Contaminating DNA was removed using TURBO DNA-free (Ambion, Austin, TX). RNA was quantified using the Agilent 2100 BioAnalyzer in combination with a RNA 6000 Nano LabChip. RNA was reverse-transcribed using M-MLV reverse transcriptase (Ambion) and a mixture of anchored oligo-dT and random decamers. For each sample, 2 RT reactions were performed with inputs of 100 and 20 ng. An aliquot of the cDNA was used for 5′-nuclease assays using Taqman chemistry. LCMV specific primers, in combination with Universal Master mix, were run on a 7900 HT sequence detection system (Applied Biosystems, Foster City, CA). Cycling conditions were 95°C, 15m followed by 40 (2-step) cycles (95°C, 15s; 60°C, 60s). Relative quantification to the control was done using the comparative Ct method. The values plotted are the average of at least 2 PCR reactions, and were normalized to actin.
Measles nucleoprotein: fwd -- CGCAGGACAGTCGAAGGTC,
rev -- TTCCGAGATTCCTGCCATG,
probe -- 6fam-TGACGCCCTGCTTAGGCTGCAA-bhq1.
LCMV nucleoprotein: fwd -- CTAACTATGGCTTGTATGGCCAAA,
rev -- TAAAGCAAGCCAAGGTCTGTGA,
probe -- 6fam-CACAGACTCCGCTCAATGACGTTGTACA-bhq1.
Actb: fwd -- CCAGCAGATGTGGATCAGCA,
rev -- CTTGCGGTGCACGATGG,
probe -- 6fam-CAGGAGTACGATGAGTCCGGCCCC-bhq1.
For each timepoint, brains from 4–5 mice were removed and either immersed in 10% formalin for paraffin embedding and subsequent sectioning and staining with hemotoxylin/eosin, or immersed in tissue embedding compound, snap-frozen in a dry ice/isopentane bath and stored at −80°C for immunohistochemistry. Horizontal cryosections (10 µm) were air dried and stored at −80°C. On the day of staining, sections were fixed in ice-cold 95% ethanol, rehydrated in PBS, and blocked for 20 m with 0.1% BSA/PBS, followed by an avidin and biotin block (Vector Laboratories, Burlingame, CA). Rat anti-mouse CD4 (clone RM4-5; 1∶100; Pharmingen, San Diego, CA), or rat anti-mouse CD8a/CD8b.2 antibodies (clones 53-6.7 and 53-5.8, respectively; 1∶100 each; Pharmingen) were used to identify CD4+ and CD8+ T lymphocytes. Sections were then incubated for 1 h at room temperature with a biotinylated anti-rat IgG secondary antibody at a 1∶200 dilution (Vector). Sections labeled with biotinylated antibodies were treated for 30 m with a streptavidin-peroxidase conjugate (ABC Elite; Vector), followed by visualization with diaminobenzidine (DAB; 0.7 mg/ml in 60 mM Tris) and urea H2O2 (0.2 mg/ml), purchased as pre-weighed tablets (Sigma). All cells were counterstained with hematoxylin and preserved with an aqueous mounting medium. Uninfected tissues, use of a haplotype matched primary antibody, or omission of the primary antibody served as negative controls. For all histological analyses, at least 3 sections per brain were examined from 3 different horizontal levels, and at least 4 mice per experimental group were assessed.
On the indicated dpi, mice were deeply anesthetized with 400 µl 3.8% chloral hydrate in PBS, delivered IP. Once animals were confirmed to be nonresponsive, the mice were perfused with 30 ml PBS. Following perfusion, each brain and spleen was removed and pressed through a nylon mesh cell strainer in PBS. Dissociated tissue was run over a 30/70% discontinuous Percoll gradient for 20 m at 4°C. Mononuclear cells (MNCs) were recovered from the interface, washed with PBS, treated with 0.84% ammonium chloride to remove contaminating red blood cells (RBCs) and washed again. Collected MNCs were counted using a standard hemocytometer and plated into a v-bottom 96-well plate for subsequent antibody staining for multicolor flow cytometry. The following antibodies (eBioscience) were used: PE-Cy5-CD8a, APC-Ax750-CD4, PE-CD3e, FITC-CD11c, APC-CD161c (NK1.1), PB-CD19, PE-Cy5.5 Gr-1 (Ly-6G), PE-Cy7-CD49b (DX5), Ax700-CD11b, APC-NP396 and APC-GP33 tetramers (a generous gift from John Wherry, Univ. Pennsylvania). Cells were allowed to incubate with antibody for 1 h at 4°C and then washed following the incubation period. Pelleted, stained cells were resuspended and read in a BD LSR II system. Percentages obtained from flow cytometry were combined with hemacytometer counts in order to calculate total cell numbers.
CD46-MC57 or MC57 cells were infected, respectively, with either MV or LCMV at a multiplicity of infection (MOI) of 1 for 1 h, or left uninfected, and further cultured for 2d. 3×106 of these target cells were then labeled with 35 µl 51Cr for 1 h at 37°C. Single cell suspensions of splenocytes were isolated from mice infected with LCMV IP 6d previously and incubated with 51Cr labeled target cells for 6 h at 37°C. Effector and targets cell concentrations were calculated and cells were added at ratios of 100∶1, 50∶1, and 10∶1 in a total reaction volume of 200 µl, with each condition in triplicate. DMEM and 1% IGEPAL (Sigma) were used in place of splenocytes for calculating spontaneous release and maximum release, respectively. Following effector/target cell co-culture, 40 µl of collected supernatant for each condition was diluted in 60 µl MicroScint scintillation fluid (Perkin Elmer). Released 51Cr was measured in a gamma counter with the percentage of specific 51Cr release being calculated by the following formula: [(experimental release – spontaneous release) / (maximum release – spontaneous release)] x 100.
At 7d post-inoculation with LCMV-Arm IP, MV-Ed IC or PBS, mice were deeply anesthetized with 400 µl 3.8% chloral hydrate in PBS, delivered IP. Once animals were confirmed to be nonresponsive, 200 µl of a 2% Evan's Blue solution was then administered transcardially; 2 m post-injection, the mice were perfused with 30 ml PBS. For Evan's blue quantification, brains were dissected, placed in test tubes, snap-frozen in liquid nitrogen and stored at −80°C. Brain tissues were slowly thawed at RT, homogenized in 1 ml N,N-dimethyl formamide and incubated at 50°C for 2d. Following centrifugation at 2700 g for 10 m, the absorbance of the supernatant containing extracted Evan's blue was measured in a spectrophotometer at 620 nm. Each absorbance reading was normalized to wet brain weight. Evan's Blue uptake into the tissues of each infected animal was divided by the average uptake detected in similar tissues from PBS control mice and the results expressed as fold induction. Statistical significance was determined using the Wilcoxon signed-rank test.
To determine brain water weight, mice were euthanized with isofluorane at 7d post-inoculation with LCMV-Arm IP, MV-Ed IC or PBS. For CD8+ T cell depletion experiments, 150 µg of CD8 depleting antibody (purified at the FCCC hybridoma facility from hybridoma clone 2.43 [ATCC #TIB 210]) was injected IP 1d prior to infection. Brains were removed, placed in test tubes and then desiccated in a vacuum oven for 24 h at ∼80°C at 15 mm Hg. Brain weights were determined both prior to and immediately following desiccation, with relative percent water weight determined by the following formula: [(wet weight - dry weight)/wet weight] x 100. Water weight of the brain of each infected animal was divided by the average water weight detected in similar tissues from PBS control mice, and the results expressed as percentage change. Statistical significance was determined using the Wilcoxon signed-rank test.
Infection of mice with LCMV by an IP route generates a robust CD8+ T cell response that clears the virus within 7–10 dpi in the absence of overt disease. Similarly, MV infection of transgenic mice expressing a human measles receptor targeted to CNS neurons (NSE-CD46+ mice) results in activation of a protective adaptive response that resolves the infection in a similar timeframe, and without concomitant illness. Because these infections are tissue-restricted (LCMV to peripheral tissues and MV to the CNS), we explored the parameters of viral clearance, immune cell trafficking, and pathogenesis in mice infected simultaneously with both pathogens (Figure 1A).
NSE-CD46+ mice were challenged concurrently with MV by an IC route (1×104 PFU) and LCMV by an IP route (2×105 PFU) and monitored daily for weight loss and overt signs of illness. As shown in Figure 1B and Table 1, no weight loss was observed in any of the singly infected mice or uninfected controls over a 34d observation period and all mice appeared healthy throughout this timecourse. In contrast, approximately half of the mice challenged with both LCMV and MV showed dramatic weight loss (Figure 1B, white squares; Table 1, group 3), leading to death by 8–10 dpi. Within this group, all animals showed overt signs of CNS disease, including ataxia, tremors, and a hunched (kyphotic) and ruffled appearance. In the terminal stages of disease, mice developed severe seizures lasting >10s per episode; death was always coincident with these seizures. The co-infected mice that did not develop severe disease showed little change in weight (Figure 1B, white diamonds), but most did show moderate illness, including slight ataxia. Of note, most of the severely sick mice began to show signs of weight loss and illness within 3–4 dpi, whereas the mice that would go on to have mild disease and survive appeared asymptomatic at this timepoint.
To address whether the pathogenesis resulting from co-infection was attributable to increased viral load, we reduced the dosages of both MV and LCMV by 10-fold; however, identical results were obtained (Table 1, group 4). Moreover, when UV-inactivated MV was administered instead of replication-competent virus, no disease or weight loss was observed (Table 1, group 5). Interestingly, however, challenge of mice with UV-inactivated LCMV did result in morbidity in approximately 25% of infected mice, indicating that replicating LCMV was not required to elicit the pathogenic phenotype (Table 1, group 6). Finally, we found that the timing of the infections was critical: if the infections were separated temporally (7d and 30d tested, in which MV was given first, followed by LCMV), no pathogenesis was observed (Table 1, groups 7 and 8). Collectively, these data indicate that concurrent challenge of mice with two viruses triggers a novel pathogenic outcome that is not observed with either virus alone.
To determine the basis of fatal neurological disease in some co-infected mice and to define the factors that govern the differential pathogenic response, we first considered possible differences in the infected mice themselves. Consequently, we evaluated whether sex or age distinguished mice that developed severe CNS disease from those that survived the milder disease. Mice of both sexes between 3–6 months of age were co-infected, and neither age nor sex correlated with pathogenic outcome (data not shown).
We next asked if co-infection altered MV or LCMV biology with respect to titer, tropism, or rates of viral clearance. As shown in Figure 2A, which depicts quantitative RT-PCR results from co-infected mice at 4 dpi, MV RNA remained restricted to the CNS and LCMV RNA to peripheral tissues, including the spleen and liver. Likewise, no evidence of LCMV immunopositive cells was found by immunohistochemistry in brains of co-infected mice (data not shown). When mice were evaluated at 2, 7, 12 and 30 dpi, though viral loads differed as these viruses were being cleared, the same tissue restriction was observed (data not shown). Moreover, the viral load of MV in the CNS was not appreciably different between singly and co-infected mice (Figure 2B). RNase protection assays (TNFα, TNFβ, CCL2, CCL3, CCL4, CCL5, CXCL2, CXCL10, CXCR3, IFNβ, IFNγ, IFNγR, VCAM1, IL-4, IL-12, IL-15, IL-18, iNOS) were performed and no significant change, in either levels or profile, was observed between MV infection alone and co-infection or between symptomatic and asymptomatic co-infected mice (data not shown). These data indicate that key aspects of the viral life cycle for both MV and LCMV, including tissue tropism and viral load, were not changed upon co-infection.
We next evaluated the host immune response within brains of infected mice, first using standard immunohistochemistry for the presence of both CD4+ and CD8+ T cells in horizontal brain sections, followed by ex vivo isolation of brain-infiltrating immune cells combined with flow cytometry and hemocytometer counts. Representative horizontal plane images are shown in Figure 3A. As published previously [9], [17], NSE-CD46+ transgenic mice infected with MV IC show substantial, and approximately equivalent, infiltration of both CD4+ and CD8+ T cells at 7 dpi (Figure 3A, center panels). As expected, T cells typically accumulate in regions where MV antigen is prevalent (data not shown). Brains from LCMV-infected mice had no evidence of T cell infiltration into the CNS (Figure 3A, left panels), because the CNS is not a target organ for infection by LCMV via the IP route (Figure 2A). When mice were challenged with both LCMV and MV and examined histologically at 7 dpi (Figure 3A, right panels), the number and proportion of CD8+ T cells in the CNS was steeply elevated compared to singly infected mice (Figure 3A, 3B), increasing from approximately 14,000 cells in MV-infected brains to 170,000 cells in brains of mice challenged with both viruses. This represents a 12-fold increase over the single MV infection and a >250-fold increase compared to uninfected mice. As shown in the bottom right panel of Figure 3A, CD8+ T cells were found distributed throughout the brain, though somewhat concentrated in patches that correlated with sites of MV replication (data not shown), as well as surrounding vascular structures, ventricles and meninges. While there was a range in the number of lymphocytes isolated from individual brains, the increase in CD8+ T cells was observed in all co-infected mice that were tested, regardless of illness (Figure 3B). In contrast, the number of CD4+ T cells was only slightly increased in co-challenged mice, compared to levels seen in MV-infected mice alone (avg. 13,500 vs. 22,000; Figure 3A, 3B). Similarly, flow cytometry revealed no significant increases in other immune cell populations (B cells, NK cells, NKT cells, macrophages, neutrophils, dendritic cells; data not shown). Thus, the lethal CNS disease that occurs in approximately half of the co-infected mice is not simply attributable to the abundance of infiltrating cells within the brain parenchyma, and appears to involve CD8+ T cells.
LCMV potently activates CD8+ T cells; thus, we used LCMV-specific tetramers to determine the proportion of CNS-infiltrating CD8+ T cells of LCMV-specificity. Tetramers specific for the two major LCMV epitopes on the H-2b background (GP33 and NP396) were employed (provided by Dr. John Wherry). As shown in the representative data in Figure 4A, right panels, 8% (mean: 6.1 ± 2.0%) and 17% (mean: 15.2 ± 3.8%) of the lymphocytes extracted from co-infected brains were CD8+ T cells specific for LCMV GP33 and NP396, respectively. Based on these two immunodominant peptides, >35% (37.6 ± 5.8%) of the total infiltrating CD8+ T cells (>21% of the total brain-infiltrating immune cell population) were specific for LCMV within the brains of co-infected mice. This is likely a substantial underestimate, as this does not account for T cells specific for minor LCMV epitopes. LCMV tetramers did not label any lymphocytes isolated from brains of mice infected with MV alone, indicating that there is little to no cross-reactivity between LCMV and MV, as measured by this assay. This was further supported in standard CTL assays, in which splenocytes isolated from mice infected with LCMV IP did not recognize MV-infected target cells (Figure 4B).
Interestingly, the posture of mice that die following co-infection is identical to that seen in mice challenged with LCMV by an IC route (Figure 5A). IC LCMV infection of immunocompetent mice with as few as 1 PFU results in overt signs of illness by ∼6 dpi, including ruffled fur, ataxia and tremors, which invariably lead to seizures and death by 6.5–7 dpi. The decerebrate posture following LCMV death is unique to LCMV, as mice that succumb to other CNS infections (such as RAG-2 KO mice following MV infection; Figure 5A, second panel) or mice infected with poliovirus (data not shown), do not develop seizures or a similar post-mortem posture.
Recently, we showed that death following LCMV IC challenge coincided with apparent compression of the brain against the skull [14]. Moreover, all mice dying of LCMV by the IC route had unilateral pupillary dilation (mydriasis), which is likely attributable to uncal herniation of brain tissue through the foramen magnum. Similar observations were found in all moribund mice that were infected with both LCMV IP and MV IC (Figure 5B, right panels), implying a common mechanism of disease.
It is well established that mortality following IC challenge with LCMV is mediated by CD8+ T cells [18], [19]. Thus, given the massive infiltration of CD8+ T cells into the brains of co-infected animals, combined with the observed similarities in gross pathology between the co-infected mice that die and those challenged with LCMV IC, we next elucidated the role that CD8+ T cells play in the neuropathogenesis observed in the co-infected animals.
CD8-depleted mice were challenged with both MV IC and LCMV IP (Figure 6A). As expected, CD8+ T cell depletion delayed the onset of pathogenesis by approximately 6d, indicating that CD8+ T cells are important for the mechanism of pathogenesis. While some co-infected CD8-depleted mice did succumb at later timepoints, they displayed symptoms similar to MV IC infection of NSE-CD46+/RAG-2 KO mice (hunched/kyphotic posture and ruffled appearance) and not those observed with symptomatic co-infected immunocompetent mice (seizures leading to death and the characteristic decerebrate posture). Furthermore, a proportion of CD8-depleted animals challenged with MV IC alone also succumbed with MV-like symptoms, confirming that, in the absence of CD8+ T cells, mice are susceptible to an intracerebral MV infection.
To elucidate the mechanism of pathogenesis mediated by CD8+ T cells in the co-infection model, we infected mice which lack the pore-forming protein perforin, which is required by cytotoxic lymphocytes to mediate granzyme-driven lysis of target cells. Co-infected NSE-CD46+/PFN KO mice succumbed in a similar fashion to NSE-CD46+ immunocompetent mice, both temporally and symptomatically (Figure 6B). This indicates that perforin does not a play a central role in CD8+ T cell-mediated pathogenesis in co-infected mice. At later time points (>15 dpi), NSE-CD46+/PFN KO animals succumbed somewhat more rapidly than NSE-CD46+ immunocompetent mice; however, they did so with symptoms identical to those seen during MV infection of immunocompromised NSE-CD46+/RAG-2 KO mice [9]. This decline may be attributable to the observation that PFN KO animals succumb to a LCMV IP alone challenge at later timepoints, whereas immunocompetent mice survive such a challenge.
In LCMV IC-challenged mice, brain herniation results from an increase in intracranial pressure, which causes the brain tissue to shift from an area of higher pressure to one of lower pressure. To determine whether the co-infected mice were dying from brain herniation (as suggested by their postmortem posture in a CD8-dependent fashion), we examined various mechanisms that could result in increased intracranial pressure.
Edema increases intracranial pressure, which can then lead to herniation. To directly ascertain if moribund mice had evidence of cerebral edema (elevated brain water weight), brains were removed at 7 dpi from: i) control PBS IC mice, ii) mice that were challenged with LCMV by an IP route (not expected to lead to CNS water volume changes), iii) MV IC (which results in immune cell infiltration, but no CNS disease), or iv) mice infected with both viruses. For the co-infected group, mice were further separated into those with severe (fatal) illness and those with the milder, nonfatal disease. Mice challenged singly with either LCMV IP or MV IC had water volume values that were not appreciably changed from uninfected or PBS-challenged mice (Figure 7A). (Water weight raw values ranged from 72–81%). Similarly, co-infected mice with mild disease showed no substantial edema (average of 77%, and a range of 73–79%). In sharp contrast, severely sick, co-infected mice showed a statistically significant increase in water volume (>7% over baseline), with an average of 83%, ranging from 80–87%, strongly suggesting that edema is mechanistically linked to mortality in this model system. As further evidence of a role for CD8+ T cells in this process, CD8-depleted mice did not show a significant increase in edema.
We also evaluated the integrity of the blood-brain barrier (BBB), as determined by extravasation of albumin-bound Evan's Blue into the brains of infected mice, as a possible cause of increased brain edema and resulting herniation. As shown, all singly infected mice and co-infected mice with mild symptoms showed negligible increases in Evan's Blue levels, indicative of an intact barrier, at 7 dpi (Figure 7B). Severely sick, co-infected mice showed only modest BBB permeability. Though there was variability in the values within the severely sick group, with some mice exhibiting high levels of Evan's Blue, there was no statistical significance in the overall increase in permeability. This is consistent with the LCMV IC model: while the blood-brain barrier is permeabilized in many of these mice, the changes do not correlate with the onset of disease.
Finally, we assessed the histopathology in brains of moribund mice to determine potential causes for the observed cerebral edema (Figure 7C). In all animals evaluated, inflammation within the choroid plexus (i) and meninges (ii) was found, along with evidence of parenchymal damage, as determined by the extensive deposition of hemosiderin within the parenchyma (iii, arrows), indicative of capillary bleeding. Interestingly, the moribund co-infected animals had greater levels of meningitis than the animals which survived challenge with both viruses (data not shown).
Collectively, these data suggest that the processes that mediate pathogenesis in co-infected NSE-CD46+ mice are similar to those observed in mice challenged with LCMV by the lethal IC route. However, unlike the LCMV IC infection in which all mice succumb, co-infection results in mortality in only approximately 50% of infected individuals. Specifically, disease is CD8-dependent, yet perforin-independent, and correlates with increased cerebral edema and evidence of both meningitis and encephalitis.
This paper describes the development of a mouse model of tissue-restricted, simultaneous viral challenges: MV is restricted to CNS neurons due to transgenic specification of the receptor, and LCMV is limited to peripheral tissues due to the route of inoculation. While single infection with either virus alone results in immune-mediated protection in the absence of disease, co-infection results in a high degree of morbidity and mortality, apparently triggered by migration of LCMV-specific CD8+ T lymphocytes into the MV-infected CNS.
The pathogenesis observed in moribund co-infected mice is indistinguishable from that seen in mice challenged with LCMV by an IC route. We previously showed that disease in LCMV IC-infected mice is attributable to increased brain water weight (edema) and reactive periventricular disease, culminating in physical manifestations consistent with uncal herniation (seizures, decerebrate posture, mydriasis) [14]. In the present study, similar increases in edema were observed in moribund, co-infected mice. While changes in blood-brain barrier integrity were also noted in these mice (as also noted in LCMV IC infected mice), increased permeability was neither significant nor consistently correlated with the timing of disease. Therefore, though blood-brain barrier damage may contribute to neuropathogenesis, it does not appear to be the precipitating cause of mortality in our mouse model. Importantly, active infection of the CNS was a prerequisite for disease, as instillation of the CNS with UV-inactivated MV did not result in either elevated T cell recruitment or neuropathogenesis. Thus, the acute CNS infection (which is pathogenically inert when delivered alone) likely induces recruitment signals, including chemokines [20], and adhesion molecules on barrier endothelium, that then aid lymphocyte recruitment. Of relevance, the receptor for CXCL10, CXCR3, is expressed by a majority of activated, LCMV-specific CD8+ T cells, which would presumably allow them to traffic toward the MV-initiated recruitment signal [21], [22]. These chemoattractant signals would be expected to recruit all activated immune cells, regardless of specificity. While other models have shown that lymphocytes can be recruited to tissues that do not express cognate antigen [4], [23], [24], to our knowledge, the MV/LCMV co-infection model is the first to show that this phenomenon can have pathogenic consequences.
While the parallels between mortality following co-infection and the classic LCMV IC infection allowed us to define neuroanatomical events that were consistent with mortality, the precise mechanism by which LCMV-specific T cells trigger damage within the CNS remains unknown. It does not seem likely that the massive infiltration of T cells into the brain can, alone, account for disease, as there was no direct correlation between total lymphocyte counts in the CNS and disease severity. Nevertheless, the possibility that migration of T cells to a particular site within the brain may contribute to enhanced vulnerability cannot be excluded. Using immunodepleted and immune knockout mice, it is clear that the neuropathogenesis following co-infection in this model is dependent on CD8+ T cells, but independent of their effector function involving perforin.
A key question that is the basis of our ongoing work is to define what LCMV-specific T cells “see” to trigger disease, since we detected no cross-reactivity between MV and LCMV, and by multiple methods, no LCMV RNA or proteins were detected in the CNS at any timepoint tested, from 2–30 dpi. Moreover, co-infection with UV-LCMV IP together with MV IC triggered neuropathogenesis, indicating that replicating LCMV need not be present in the CNS. Classically, the expression of effector functions by CD8+ T cells is reliant on the recognition of antigen. However, together with IFN-γ, the innate cytokines, IL-12, IL-18, IL-23 and IL-27, have all been shown to be potent activators of T cells that act in an antigen-independent manner, as well as independent of TCR ligation and signaling [25], [26]. Binding of these interleukins to their respective receptors on T cells alone causes these CD8+ T cells to become activated and release IFN-γ and TNF-α. These early functions of CD8+ T cells are important components in the host response to such pathogens as influenza A virus, Lysteria monocytogenes, Cryptosporidium parvum, and tumors [26]–[30]. Together, these data suggest that CD8+ T cells may indeed play a role in immunopathogenesis if they have been independently activated and then ‘misrecruited’ to another, independent site of infection.
Neither changes in tropism nor changes in the extent of infection of either virus were seen to change following co-infection. While our PCR-based and IHC-based studies found no evidence of LCMV infection within the brain, work by others suggests that brain endothelial cells and epiplexus cells can present antigen in the absence of direct infection, similar to cross-presentation by antigen presenting cells [31]–[33]. Thus, it is possible that in these co-infected animals, brain endothelial cells or epiplexus cells present LCMV epitopes picked up from the CSF, in the absence of direct infection. Evaluation of this mechanism as a possible cause of neuropathogenesis is an ongoing effort, and adoptive transfer studies using activated LCMV-specific CD8+ T cells into a MV-infected recipient are underway. However, we are cautious about a result of a lack of neuropathogenesis in these experiments, as a “primed” microenvironment within the periphery may be an important component in activation of such cells.
What general implications can be inferred from these results? Mouse models have been essential for the study of viral pathogenesis, both in aiding our understanding of the viral life cycle in vivo and in revealing critical aspects of immune response induction, recruitment, and function. Nevertheless, the "single-challenge" approach that most viral pathogenesis studies employ--in which immunologically naïve adult animals are inoculated with only one pathogen and evaluated thereafter--does not approximate the complexity of immune history or host responses to concurrent immune challenges that occur in humans.
A prominent example of coinfection with enhanced pathology is human immunodeficiency virus (HIV) and hepatitis C virus. For pathogen combinations such as this, the mechanism of increased pathology is well understood: Pathogen A causes immunosuppression of the infected host, thereby setting up a permissive environment for Pathogen B to establish an unrestricted infection [34]. The issue of immunosuppression is one that we have not explored, even though it is known that MV reduces responses to a co-infecting agent (i.e. HIV) [35]. At this point, we do not believe that this is a major confounder for our experiments since all arms of the immune response examined in our model are either increased or remained at baseline, when compared to either single infection alone. For other diseases, such as Burkitt's lymphoma, polymicrobial infections (Epstein-Barr virus (EBV) and Plasmodium falciparum malaria) are clearly associated with the disease outcome, but the causative mechanism is still elusive [36]. Thus, polymicrobial infections – especially in developing regions of the world – are common, though how these pathogens and their respective immune responses interact to result in pathogenic outcomes remains unknown.
While making direct connections between this model system and human diseases would be premature, a similar phenomenon could account for both the reported increase in inflammatory cells found in the brain of patients presenting with MS and ALS [37]–[40], as well as the lack of consistent and convincing presence of any specific pathogen in the brains of affected individuals [41]–[44]. A large number of CNS diseases of unknown etiology have been proposed to have viral triggers [45]–[48], though no single pathogenic agent has been linked with any chronic CNS disease, perhaps because it is generally assumed that the viral trigger must be co-localized with the observed damage. Our co-infection model of immune cell recruitment suggests that this may not need to be the case – the virus infection and damage sites may be spatially distant, provided that a recruiting signal (e.g., a chemokine gradient) is present within the CNS.
While our studies focused on viral pathogens, the results can be considered more broadly in terms of any agent or situation that can trigger an immune response. For example, concurrent immune responses to other antigens (e.g., tumors, allergens, autoantigens) might also contribute to pathology resulting from infection-irrelevant immune effector trafficking. In summary, while single-pathogen mouse models will continue to be powerful tools to explore the induction and function of immune cells and their mediators, more complex animal models, which consider both pathogenic and non-pathogenic triggers of host immunity, will be required to fully reveal the diversity of ways by which immune-mediated neurological diseases occur.
|
10.1371/journal.pntd.0001698 | Footprint of Positive Selection in Treponema pallidum subsp. pallidum Genome Sequences Suggests Adaptive Microevolution of the Syphilis Pathogen | In the rabbit model of syphilis, infection phenotypes associated with the Nichols and Chicago strains of Treponema pallidum (T. pallidum), though similar, are not identical. Between these strains, significant differences are found in expression of, and antibody responses to some candidate virulence factors, suggesting the existence of functional genetic differences between isolates. The Chicago strain genome was therefore sequenced and compared to the Nichols genome, available since 1998. Initial comparative analysis suggested the presence of 44 single nucleotide polymorphisms (SNPs), 103 small (≤3 nucleotides) indels, and 1 large (1204 bp) insertion in the Chicago genome with respect to the Nichols genome. To confirm the above findings, Sanger sequencing was performed on most loci carrying differences using DNA from Chicago and the Nichols strain used in the original T. pallidum genome project. A majority of the previously identified differences were found to be due to errors in the published Nichols genome, while the accuracy of the Chicago genome was confirmed. However, 20 SNPs were confirmed between the two genomes, and 16 (80.0%) were found in coding regions, with all being of non-synonymous nature, strongly indicating action of positive selection. Sequencing of 16 genomic loci harboring SNPs in 12 additional T. pallidum strains, (SS14, Bal 3, Bal 7, Bal 9, Sea 81-3, Sea 81-8, Sea 86-1, Sea 87-1, Mexico A, UW231B, UW236B, and UW249C), was used to identify “Chicago-“ or “Nichols -specific” differences. All but one of the 16 SNPs were “Nichols-specific”, with Chicago having identical sequences at these positions to almost all of the additional strains examined. These mutations could reflect differential adaptation of the Nichols strain to the rabbit host or pathoadaptive mutations acquired during human infection. Our findings indicate that SNPs among T. pallidum strains emerge under positive selection and, therefore, are likely to be functional in nature.
| During infection, the agent of syphilis, Treponema pallidum subsp. pallidum (T. pallidum), successfully evades the host immune defenses and establishes a persistent infection that can cause blindness, paralysis, or even death in some individuals that progress to the tertiary stage of the disease. The study of the Nichols strain of T. pallidum, isolated over a century ago and continually propagated in rabbits, has been paramount to deepen our knowledge on the biology of the agent of syphilis and the pathogenesis of this complex disease. Nonetheless, when the more recent Chicago isolate of T. pallidum is compared to the Nichols strain, significant differences in gene expression, gene conversion rates, and antibody responses against virulence factor candidates are detected during experimental infection. To investigate whether differences at the genomic level between Nichols and Chicago might explain such phenotypic differences, we sequenced the Chicago strain genome and compared it to the previously sequenced T. pallidum Nichols strain. Our findings indicate that the genomic differences between these T. pallidum strains emerge under positive selection, and are likely to be functional in nature, thereby being involved in shaping the phenotypic diversity between the Chicago and Nichols strains.
| Syphilis continues to be a common and serious disease, affecting at least 25 million persons worldwide [1]. It is a recognized cofactor in the transmission and acquisition of HIV [2], [3], and is a major cause of stillbirth and perinatal morbidity particularly in the developing world [4], [5]. The peculiar biology of the causative agent of syphilis, Treponema pallidum subspecies pallidum (T. pallidum), along with the inability to grow this pathogen continually in vitro, has hindered progress in understanding the pathogenesis of this disease. Syphilis research however, greatly benefited from the elucidation of the T. pallidum Nichols strain genome sequence [6]. This 1.138 Mb genome is among the smallest characterized in prokaryotes. The lack of genes encoding for several metabolic pathways (i.e. Krebs' cycle, glyoxylate shunt, amino acid and fatty acid synthesis, etc.), restriction-modification enzymes, transposons, or prophages [6], strongly suggests that T. pallidum's evolution as a human pathogen exploited progressive genome reduction and loss of those functions now provided by the host.
Since its isolation in 1912 from the cerebrospinal fluid (CSF) of a patient with secondary syphilis [7], the Nichols strain of T. pallidum has been continually propagated in rabbits, and has become the reference strain in experimental syphilis. Thus, it was the obvious choice for the original T. pallidum genome project. The Chicago strain of T. pallidum, isolated in 1951 by Turner and Rodriguez from a primary chancre [8] and far less extensively propagated in rabbits, has become increasingly important in the study of the pathogenesis of syphilis. Despite the fact that Nichols and Chicago belong to the same T. pallidum molecular strain type (14a/a [9]), suggesting an elevated degree of genetic similarity, several phenotypic and genotypic differences have been highlighted between these strains during experimental infection. Important differences between the two strains were described regarding gene expression of candidate virulence factors [10], [11], as well as antibody and cellular responses against Nichols and Chicago antigens during experimental infection [10], [11]. An example of the above differences involves the 12-membered tpr (T. pallidum repeat) gene family [12]. The tpr genes and the antigens they encode have been the focus of intense research by our group, leading to the characterization of the immune response against these antigens during experimental syphilis [10], [11], and their potential as protective antigens [12]–[16]. The study of transcriptional patterns of these genes and the mechanisms that control expression of several tpr genes resulted in the identification of phase variation as a mechanism for controlling expression of at least five tpr genes [11], [17]. Another member of the tpr gene family, tprK, undergoes extensive sequence variation mediated by gene conversion during infection, resulting in changes in seven discrete variable (V) regions in the tprK ORF [18]. Chicago has been shown to diversify the sequence of tprK at a significantly higher baseline rate than Nichols, before onset of detectable specific immunity, during intratesticular (IT) passages and intradermal (ID) infections. The tprK gene in the Nichols strain remains virtually clonal, varying its sequence only after onset of an adaptive immune response against the initial TprK antigen [19], while variants arise throughout infection with the Chicago strain. In the presence of an adaptive immune response against the TprK antigen, the difference in accumulation of variants in Chicago is even more striking.
To investigate whether genomic differences could explain the biological differences between Nichols and Chicago, the genome of the Chicago strain was elucidated using next-generation Illumina sequencing, annotated, and compared to the published Nichols genome. Genomic differences were confirmed by dideoxy-terminator (DT) sequencing using template DNA from Chicago as well as the Nichols strain (Houston) used for the original T. pallidum genome project. All coding sequences carrying SNPs, as well as approximately one third of the loci carrying small indels, were amplified and sequenced, revealing a strikingly high frequency of sequencing errors in the available Nichols genome. Nonetheless, comparison of 16 Nichols and Chicago polymorphic loci with the corresponding genomic regions of 12 more recently isolated T. pallidum strains (SS14, Bal 3, Bal 7, Bal 9, Sea 81-3, Sea 81-8, Sea 86-1, Sea 87-1, Mexico A, UW231B, UW236B, and UW249C) suggested that the genetic differences between Chicago and Nichols were acquired under the action of positive selection, and allowed us to speculate on the pathoadaptive nature of these changes in these T. pallidum strains.
No investigations were undertaken using humans/human samples in this study. New Zealand white rabbits were used for T. pallidum propagation. Animal care was provided in accordance with the procedures outlined in the Guide for the Care and Use of Laboratory Animals under protocols approved by the University of Washington Institutional Animal Care and Use Committee (IACUC).
The Chicago strain of T. pallidum subsp. pallidum, initially supplied by Dr. Paul Hardy and Ellen Nell (Johns Hopkins University, Baltimore, MD), was propagated intratesticularly in seronegative New Zealand white rabbits as previously reported [20]. Briefly, three rabbits were injected with 5×107 T. pallidum cells per testis and checked daily for disease progression. Animals were euthanized approximately 10 days after infection, at peak orchitis, to recover the highest number of organisms before the onset of immune clearance.
Testes were minced in 20 ml of PBS for approximately 10 min and suspensions were centrifuged twice for 10 minutes at 1,000× G to remove large host cellular debris. The supernate was then centrifuged at 18,000× G for 15 minutes to pellet treponemes. Treponemes were resuspended in 1 ml of PBS and stored on ice as the gradients were prepared. Discontinuous sodium and meglumine diatrizoate (Renografin-60, Bracco Diagnostics, Princeton, NJ) gradients were prepared at room temperature by first diluting Renografin-60 stock solution (60%) to the desired concentrations with PBS [6]. To obtain the discontinuous gradient, a first layer of 60% Renografin-60 (1.5 ml total) was deposited in the bottom of a 10 ml Ultra-Clear Thinwall centrifuge tube (Beckman-Coulter, Fullerton, CA), followed by 1 ml each of 37.5%, 25%, and 19% Renografin-60 dilutions, respectively. Approximately 0.2 ml of ice cold treponemal suspension was carefully layered onto each gradient and tubes were centrifuged at 20°C for 45 min at 100,000× G in an Optima XL-100K ultracentrifuge (Beckman-Coulter) equipped with a SW-41 Ti rotor. Fractions of approximately 0.2 ml were recovered by drop from the bottom of the tube. Fractions containing high numbers of treponemes (identified by dark-field microscopy) were pooled together and treated with a total of 5 units of RQ1 DNaseI (Promega, Madison, WI) to reduce the contamination by rabbit DNA. After treatment and heat inactivation of the enzyme (10 min at 65°C), an appropriate volume of 50× lysis buffer for DNA purification (final concentration: 10 mM Tris, pH 8.0; 0.1 M EDTA; 0.5% w/v sodium dodecyl-sulfate) was added to the treponemal suspension. DNA extraction was performed using the QIAGEN Genomic-tip 100/G kit (Qiagen Inc., Chatsworth, CA), according to the manufacturer's instruction and the sample was stored at −20°C until use.
The list of the additional strains used here can be found in Table 1. Strain propagation, harvest and DNA isolation for amplification and DT-sequencing protocols were performed as previously described [11], [17]. Although we cannot formally evaluate treponemal growth rates for the strains used in this study, differences were found regarding the time between strain passage into the rabbit hosts, and the yield of treponemes at the time of harvest. Some strains were transferred every 10–12 days (Nichols, Chicago, Sea 81-8, SS14, and UW249C), others every 15–30 days (Bal 7, Sea 87-1, Mexico A, Bal 9, Sea 86-1, Sea 81-3, and Bal 3), while some required ≥30 days (Sea 87-1, UW231B, and UW236B). Treponemal yields varied from approximately ≥108 treponemal cells/ml of testicular extract (Nichols, Chicago, SS14), ∼107 cells/ml (Bal 7, Bal 9, Sea 86-1, Sea 81-3, Bal 3, UW231B, and UW249C), and ∼106 cells/ml (Sea 87-1, Mexico A, Sea 81-8, and UW236B).
The percentage of rabbit genomic DNA in the Chicago sample was determined by quantitating the copy number of the rabbit (Oryctolagus cuniculus) cystic fibrosis conductance transmembrane regulator (RCFTR) gene and the T. pallidum TP0574 gene (which encodes for the 47 kDa antigen) by quantitative real-time PCR (qRT-PCR). Primer sequence, amplification protocol and standard curve preparation for the TP0574 gene were previously reported in detail [11]. RCFTR-S (5′-gcgatctgtgagtcgagtctt-3′) and RCFTR-As (5′-cctctggccaggacttattg-3′) primers (Oligos Etc. Inc., Wilsonville, OR) were used to determine the rabbit CFTR gene copy number. Amplification was carried on for 45 cycles in a Roche LightCycler 2.1 instrument (Roche, Basel, Switzerland) using the Masterplus SYBR green kit (Roche) according to the manufacturer's instruction. The reaction conditions for these amplifications included a 10 sec denaturation step at 95°C, an 8 second annealing step at 60°C, and an extension step for 10 sec at 72°C. Acquisition temperature was set at 83°C upon amplicon melting curve analysis. The standard curve for the rabbit CFTR gene was prepared as for the TP0574 gene [11]. The sizes of the rabbit and T. pallidum genomes were taken into account to determine the percentage of rabbit DNA in the sample.
Genomic DNA isolated from the T. pallidum Chicago strain was further processed for Illumina-based sequence analysis using the Paired End DNA Sample Prep Kit (Illumina Inc., San Diego, CA) following the provided protocol. Genome sequencing was performed at the Center for Genome Research and Biocomputing (CGRB) at Oregon State University (Corvallis, OR) using a Genome Analyzer IIx System (Illumina Inc.). A first draft of the Chicago strain genome was assembled using the reference-guided assembly program Maq [21] with the T. pallidum Nichols strain genome [6] (GenBank accession number for Nichols is NC_000919) as reference. Regions in the reference-guided assembled genome where Maq could not resolve sequence were then compared to contiguous sequences assembled through the use of the de novo assembly software VCAKE [22], and a single contiguous draft sequence was then produced. Nucleotide differences between matched pairs were identified using the Diffseq program from the Emboss software suite. The locations and effects of individual differences were first determined using an in-house SNP parsing program (not currently online but available upon contacting the authors) and then re-evaluated after the annotation of the Chicago strain was completed.
Regions containing nucleotide differences between the Chicago and Nichols (Houston) strain were targeted by PCR amplification and conventional DT-sequencing to confirm the high-throughput sequencing data. DNA from both Chicago and the Nichols-Houston strain sequenced in the original T. pallidum genome project were used as template. A subset of these regions were selected randomly, and others were selected to confirm differences in genes possibly implicated in generation of diversity in the tprK gene (TPChic0897) or transcriptional control (such as TPChic0924, encoding the toxin expression gene, also known as tex). Overall, 41 loci (39.8% of the total originally reported [23]) carrying small indels were sequenced in both strains. Twenty six additional regions carrying small indels were amplified using DNA from the Chicago strain and sequenced to further confirm the reliability of the high-throughput sequencing approach.
Primers (designed using the Primer 3 software, http://frodo.wi.mit.edu/primer3/) are in File S1. All PCR amplifications were performed in 100 µl reactions containing 200 µM each dNTP, 20 mM Tris-HCl (pH 8.4), 1.5 mM MgCl2, 50 mM KCl, 400 nM of each primer, and 1.0 U of Taq DNA Polymerase (Promega, Madison, WI) with approximately 100 ng of DNA template in each reaction. Cycling conditions were denaturation for 5 min at 95°C, followed by 1 min at 95°C, annealing for 1 min at 60°C and extension for 1 min at 72°C for a total of 45 cycles. A final extension of 10 min at 72°C was included. Amplicons were purified using the QIAgen PCR purification Kit (Qiagen Inc.) according to the provided protocol, and the concentration of each sample was determined using a ND-1000 instrument (NanoDrop Technologies, Wilmington, DE). Sequencing was performed at the Department of Biochemistry DNA Sequencing Facility of the University of Washington, Seattle, WA. Electropherograms were analyzed using the BioEdit software (http://www.mbio.ncsu.edu/BioEdit/bioedit.html).
Amplification and sequencing of 16 ORF fragments found to carry authentic SNPs between the Chicago and Nichols strains were also performed on 12 additional T. pallidum strains (SS14, Bal 3, Bal 7, Bal 9, Sea 81-3, Sea 81-8, Sea 86-1, Sea 87-1, Mexico A, UW231B, UW236B, and UW249C). Sequencing of the TP0924 (tex) gene region (containing the C→A transversion that truncates the putative Tex protein in Chicago) was also performed using DNA template from various Nichols isolates maintained in different laboratories over the last two decades (Seattle, Farmington, Dallas, and UCLA), as described above.
The Chicago strain genome sequence was submitted to the J. Craig Venter Institute (JCVI) Annotation Service (http://www.jcvi.org/cgi-bin/annotation/service/submit/annengine.cgi), where it was processed through JCVI's prokaryotic annotation pipeline. Included in the pipeline are 1) a gene-finding function with Glimmer, HMM, and TMHMM (Hidden Markov Models and Trans Membrane Hidden Markov Models, respectively) searches; 2) frame shift mutation identification through Blast-Extend-Repraze (BER) searches; 3) SignalP predictions for identification of signal peptides; and 4) automatic annotations from AutoAnnotate. The manual annotation tool Artemis (www.sanger.ac.uk/Software/Artemis/v11/) was used to manually review the output from the JCVI Annotation Service and compare it with the Nichols strain genome annotation (Nichols GenBank accession number is NC_000919).
To assess the genome-wide nucleotide diversity of protein-coding genes in Chicago and Nichols genomes, each gene was subject to a modified version of ZPS [24] to perform in batch mode ClustalW-based sequence alignment [25], followed by calculation of the rates of nonsynonymous (dN) and synonymous (dS) mutations using the mutation-fraction method of Nei and Gojobori [26].
Paired-end sequencing yielded a single circular contig devoid of sequence gaps. The Chicago genome [23] was found to be 1,139,281 bp long, in contrast to 1,138,011 bp in the published Nichols genome, suggesting that genomic differences might contribute to explain the differences in infection phenotypes associated with the Nichols and Chicago strains. Next-generation Illumina sequencing was not adversely affected by residual rabbit DNA, corresponding to ∼18% of the total DNA content of the sample, and the coverage of the Chicago genome ranged from ∼50× to ∼100× (average depth coverage was 64×).
Based on the annotation service provided by the JCVI, there were some ORF assignment discrepancies between the Nichols and Chicago genomes that were due to differences in the annotation algorithm rather than any sequence differences. The Chicago genome annotation identified 96 putative ORFs not previously identified in Nichols (File S2). The size of these ORFs was relatively small, ranging from 111 to 399 bp (average length = 180 bp). To facilitate direct ORF comparisons between T. pallidum strains, we named the new ORFs based on their proximity to a coding sequence shared by both strains. (For example, according to our nomenclature, TPChic0005a is an ORF annotated only in Chicago and located immediately downstream, either on the plus or minus strand, of TPChic0005 that is homologous to Nichols TP0005. If multiple new ORFs follow a shared annotation, their order is reflected by the alphabetical letter following the ORF. TPChic1025a and TPChic1025b, for instance, follow TPChic1025 and precede TPChic1026. On the other hand, 21 published Nichols ORFs (File S3) were not identified by the JCVI annotation software in the Chicago genome sequence despite nucleotide sequence conservation between the two strains.
Also, the annotation service provided by the JCVI permitted the re-analysis of the possible functions of some T. pallidum ORFs shared by two genomes. Among a total of 842 genes with the same annotation, a total of 158 ORFs (File S4) previously listed as hypothetical or conserved hypothetical proteins in the Nichols annotation were now assigned putative identities. Newly annotated possible functions include tyrosine kinases (TPChic0024, TPChic0139), efflux pumps (TPChic0901, TPChic0965, TPChic0988), and permeases (TPChic0301, TPChic0302). New putative lipoproteins (TPChic0069, TPChic0087, TPChic0149, TPChic0625 TPChic0646, TPChic0693), outer membrane lipoprotein carriers and permeases (TPChic0333, TPChic0580, TPChic0582), and metal transporters with outer membrane subunits (TPChic0034, TPChic0035, TPChic0036) were also identified.
Over 99% of all predicted protein-coding genes shared between Chicago and Nichols strains were syntenic (having same relative position in both genomes), thereby arguing against any major role of gene shuffling in shaping the genotypic/phenotypic differences between these two strains. No gene inversions were identified.
Because the Chicago strain tprK is hypervariable with respect to Nichols, a consensus sequence for the seven variable (V) regions of this gene could not be obtained and, thus, are not accounted for in the nucleotide-based comparative analysis. In the complete Chicago genome sequence found in GenBank, the tprK V1–V7 region sequences are replaced by N's.
For the Chicago genome, comparison of Illumina sequencing data with traditional DT-sequencing of genomic regions carrying SNPs showed perfect agreement between the two sequencing methods. Although we previously reported [23] that preliminary comparison with the published Nichols genome [6] identified the presence of 44 SNPs between Chicago and Nichols, recent DT-sequencing of the regions carrying these SNPs in the Nichols (Houston) strain, revealed a high frequency of sequencing errors in the published Nichols genome sequence [6]. Overall, only 20 authentic SNPs are found between Chicago and the Nichols genome: four are located within intergenic regions and 16, all non-synonymous, within ORFs coding for putative proteins (Table 2). The SNPs were evenly split between C/T and A/G transitions and were not clustered, but distributed more or less evenly along the genome (File S5).
To further explore whether these genomic differences between Nichols and Chicago genomes could have been promoted by the extensive propagation of the Nichols strain in the rabbit host, we analyzed the identity of each ORF-associated mutation in 12 other T. pallidum strains (Table 3) which, like Chicago, were propagated in rabbits far less extensively than Nichols. As a result of these 14 genome cross-examinations of 16 SNP regions, we identified only one SNP accumulated in Chicago (in TPChic0746, Table 3). Because the other 12 T. pallidum strains were identical to Nichols for this nucleotide position, we define such a change as “Chicago-specific”. Interestingly, the remaining 15 SNPs were determined to be “Nichols-specific”, in that Chicago and the other 12 genomes had identical nucleotides in these polymorphic positions, with the exception of the tprJ gene (TpChic0621) where one of the 12 other strains (Bal 7, Table 3) showed a sequence identical to Nichols. These findings clearly demonstrate that 12 other strains analyzed here are significantly more similar to Chicago at the DNA level. Overall, these data strongly suggest that “Nichols-specific” SNPs were acquired through mutation, and not recombination; furthermore, because all the “Nichols-specific” SNPs predict amino acid changes in their respective putative proteins, such a significant predominance of “Nichols-specific” changes suggests functional adaptation of Nichols in the rabbit host. Of the 16 polymorphic genes targeted in our analysis, 12 (75%) genes were annotated with defined functions, equivalent to 729 (74%) total genes with defined functions in the annotated Chicago genome. Although this small set of genes did not permit us to statistically evaluate over-representation of functional categories, at least four of these polymorphic genes are known to encode putative virulence factors, possibly contributing to the phenotypic differences seen during infection between Nichols and Chicago. These genes are TPChic0488 (Methyl-Accepting Chemotaxis protein), TPChic0621 (TprJ protein), TpChic0922 (Tex protein, discussed later in more detail), and TpChic0978 (LspA Signal Peptidase II).
The most direct way to detect any action of positive selection in protein-coding genes is to evaluate whether the rate of amino acid replacement (dN, nonsynonymous mutation per non-synonymous nucleotide site) is significantly higher than the rate of silent, synonymous mutations (dS, synonymous mutation per synonymous nucleotide site), assuming silent mutations to be, in general, of a neutral nature. Due to the small number of SNPs and because all changes were non-synonymous, the dN/dS rate could not be evaluated directly either for individual genes or for all the polymorphic genes concatenated. If, however, for the sake of analysis we incorporate a synonymous change in the concatenated genes with SNPs, the resulting dN/dS value of 4.5 (0.00081/0.00018) shows that dN was significantly higher (P = 0.03) than dS. Therefore, the absence of any synonymous SNP in the observed dataset strongly indicates that the genetic changes are positively selected and, likely, of an adaptive nature.
Apart from a single large event involving a 1204 bp insertion in an intergenic region (position 148519–149723 in the Chicago genome), indel analysis at the time the Chicago genome was released on GenBank [23] identified 103 small (≤3 nt) insertions/deletions between the two genomes (involving a total of 109 nt, due to the presence of 4 di-nucleotide indels, and 1 tri-nucleotide indel). DT-sequencing of 41 loci carrying such indels in both Chicago and Nichols (Houston) revealed however that, with the exception of two loci (TPChic0667, and the IGR 3′ of TPChic0222), the above result was due to sequencing errors in the 1998 Nichols genome. DT-sequencing of “indel-carrying” loci using template DNA from the Chicago strain never showed discrepancies with the Illumina-based sequencing results. A list of erroneous differences (both SNPs and indels) between the Nichols and Chicago strains is reported in File S6. Although only 39.8% of the originally reported differences due to indels were re-analyzed using DT-sequencing in both strains, it is striking that only 2 indels out of 41 (4.8%) were confirmed as real. This indicates that the total extent of differences due to true indels between the Nichols and Chicago strains is likely to be significantly more limited than originally reported.
A single C nucleotide insertion within the TPChic0667 ORF (coordinates: 730194–731009) caused a frame shift and an early termination of the ORF with respect to Nichols' paralogous gene. As a result, when Nichols and Chicago annotations are compared, Nichols' TP0667 ORF (555 codons) encompasses both Chicago's TpChic0667 ORF (275 codons) and TPChic0667a (271 codons). This indel was found to be “Chicago-specific”, based upon analysis by DT-sequencing of the same locus in 12 more T. pallidum strains (data not shown). A single C insertion (position: 228663) was also confirmed in the intergenic region 3′ of TpChic0222 (Table 2). Indels that were identified by comparative genomic analysis but are not currently confirmed by DT-sequencing using the Nichols (Houston) strain are reported in Table 4. Indels falling within homopolymeric nucleotide sequences were found in three Chicago ORFs (TPChic0127, TPChic0479, and TPChic0618), and within 3 intergenic regions (3′ of TPChic0026, TPChic0121, TPChic0621).
Among the “Nichols-specific” indels, the only mutation targeting intergenic regions appeared to be the 1204 bp deletion corresponding to the region downstream of TPChic0126 and upstream of TPChic0127 (spanning the location of TPChic0126a/b/c regions in the reverse strand and TPChic0126d in the plus strand). Šmajs et al. [27] previously reported that a subpopulation of the Nichols (Houston) strain used in the original T. pallidum genome project does not carry such deletion, suggesting that this genomic region might not be stable within a single treponemal isolate. The 1204 bp insertion lies between two direct repeats of 24 bp (aatgtatttcagggtgtctttctc), suggesting a loop-out mechanism for this deletion.
Chicago and Nichols differ in their origins of isolation (primary chancre vs CSF), durations of propagation in the rabbit host, gene expression levels, induction of antibody and cellular immune responses to some antigens, and rates of TprK variation, the latter being higher in Chicago than in the Seattle Nichols [19]. With respect to the published Nichols genome sequence, a 1204 bp insertion was found in the intergenic region downstream of TPChic0126. This large insertion contains 19 putative donor sequences used by T. pallidum to generate variability within all of the seven tprK V regions, especially V3 and V6 [19]. Although this insertion might be speculated to be a reason for Chicago's higher tprK variability, this 1204 bp fragment is also present in the Nichols strain currently propagated in our laboratory [18], which is slow to develop tprK variants. Therefore, the number of donor sites alone cannot explain the relative hypervariability of Chicago tprK. The Nichols strain has been extensively propagated in rabbits and this might have selected for a tprK sequence that is optimal for survival and rapid growth in rabbit tissues. Frequent passage of the Nichols strain (every 9–12 days) for routine propagation, virtually in the absence of an adaptive immune response, might have permitted the reduction in Nichols' propensity to vary tprK. Comparative analysis between the two strains did not show differences in the genes coding for the recombination machinery typically involved in gene conversion (i.e. ruv and rec genes, genes encoding site-specific recombinases or hypermutation homologues; data not shown). Structural predictions of the TPChic0899 ORF obtained using the Bio Info Bank Metaserver (http://meta.bioinfo.pl) however, found the encoded protein to be similar to an AddB-like deoxyribonuclease, a component of the counterpart of the E. coli RecBCD enzyme in Gram positive bacteria. TPChic0899 spans Nichols' TP0899 and TP0900 (originally annotated as separate hypothetical proteins) [6]. The presence of two ORFs in Nichols is due to a single G deletion that puts in frame the TGA triplet introducing a premature stop codon. Because of the possible involvement of this enzyme in homologous recombination, we further explored this difference between Chicago and Nichols. DT-sequencing of the region containing the G insertion was performed in a total of 16 T. pallidum isolates, including Nichols strains obtained from several laboratories and the SS14 strain (also reported carrying the deletion; GenBank accession number CP000805.1) [28]. Our sequencing data revealed that the G nucleotide is actually present in all isolates (Figure 1) confirming that the annotation of two separate ORFs, TP0899 and TP0900, in Nichols [6] and SS14 [28] is indeed erroneous. Because this gene appears to be functional in all T. pallidum strains, it is, therefore, likely not associated with the increased rates of tprK variation that Chicago exhibits with respect to Nichols. Nonetheless, this example underscores the likelihood, when comparative genome-wide studies among T. pallidum strains are pursued, of encountering inaccuracies in available sequences.
TPChic0924, which encodes the Tex transcriptional regulator, could potentially explain reported differences in transcription of some tpr genes in Chicago vs. Nichols [11]. The Chicago Tex protein is predicted to be 250 aa shorter than in Nichols. Tex was first isolated and characterized in Bordetella pertussis by virtue of its negative effect on the transcription and expression of toxin genes ptx and cyaA [29]. Tex paralogs were then identified in a wide variety of bacterial species [30], [31] and were shown to contain domains involved in nucleic acid binding [31]. Interestingly, studies conducted on the Pseudomonas aeruginosa Tex protein showed that presence of the carboxyl-terminal domain (present in Nichols but not in Chicago) permits Tex to bind nucleic acids [31] and thus inhibit transcription. The presence or absence of a complete Tex protein in T. pallidum could affect a strain's ability to express virulence factors. To further support the “Nichols-specific” nature of this change, it is found that all examined non-Nichols T. pallidum isolates carry the same A/C transversion (Figure 2) that would truncate the Tex protein in Chicago, in sharp contrast with the five Nichols isolates (Seattle, Houston, Dallas, Farmington, and UCLA), where the ORF encoding the Tex protein would not be truncated.
When the Chicago genome was first released [23], we reported that 44 coding sequences, annotated as independent ORFs in Nichols, are fused in Chicago leading to 21 considerably longer genes. TPChic0006, for instance, was predicted to be 417 aa long, and to span Nichols' TP0006-0008 (51, 216, and 89 aa, respectively). It is however evident now that these initial observations were a result of sequencing errors in the original Nichols genome, and not the result, as initially postulated, of gene inactivation of original longer sequences by frame shift or nonsense mutations. Recently, Šmajs and collaborators [32] suggested that genomic decay might have played a central role in T. paraluiscuniculi's adaptation to the rabbit host and loss of infectivity to humans [33], and the hypothesis that gene inactivation in the Nichols strain could reflect its adaptation to rapid passage in rabbits for nearly a century, also appeared plausible. Resequencing of the Nichols (Houston) genomic regions containing mutations hypothetically responsible for inactivation of these genes, however, clearly revealed that these annotation differences are also due to sequencing errors in the Nichols genome. It is therefore very likely that reannotation of the resequenced Nichols genome will be significantly more similar to that currently reported for Chicago. Similar findings were described by Cejková et al. [34]. A complete list of predicted gene fusions is reported in File S6.
Indels falling within homopolymeric nucleotide sequences were found in three Chicago ORFs (TPChic0127, TPChic0479, and TPChic0618), and within 3 intergenic regions (3′ of TPChic0026, TPChic0121, TPChic0621). Growing evidence suggests that changes in the length of these homopolymeric repeats, likely induced by slipped-strand mispairing during DNA replication, might be involved in transcriptional or translational control of T. pallidum genes. For example, the poly-G repeat upstream of TPChic0621 (TprJ) was shown to control transcription of this gene through a phase variation mechanism that allows transcription only when the poly-G tract is eight (or fewer) nucleotide-long [17]. The poly-G repeat upstream of TPChic0026 (encoding the fliG1 gene) could have a similar role, although evidence of intra-strain variability of this homopolymeric tract is currently not available. Furthermore, recent evidence suggests that changes in the poly-G repeat within TpChic0127 could either cause a frameshift that prematurely truncates the putative TP0127 protein, or change its reading frame, resulting in a novel protein of approximately equal length but with a different amino acid sequence (unpublished data). Variation in the homopolymeric tracts associated with TPChic0479, and TPChic0618 can also influence the annotation of these ORFs.
Analysis of SNPs in protein-coding genes showed only nonsynonymous mutations, suggesting the presence of recent diversification favoring structural changes in T. pallidum genomes. Overall, significantly higher rates of nonsynonymous changes in the Nichols genome indicate positive selection pressures in 16 protein-coding genes throughout the genome. Limited frequency of polymorphic genes did not permit us to determine whether these genes with recent structural changes could be grouped into specific functional categories of proteins. However, we found a strong clustering of polymorphic genes into two functional groups – membrane proteins and DNA-binding proteins. Within the set of genes with defined functions, the single “Chicago-specific” SNP accumulated in an ATP-binding protein-coding gene, while most of “Nichols-specific” SNPs were found to be in membrane protein-coding genes mostly related to transport and proteolysis (Table 3).
Our study suggests that genetic variability likely influences the phenotypic differences seen between the Nichols and Chicago strains of T. pallidum [10], [11], [35], even though definitive evidence for the correlation between specific genomic change(s) and phenotypic differences will require further investigation. This study also raises an important concern regarding the selection process that led to these mutations, believed to result from the adaptation of the Nichols strain to the rabbit host. Our comparative analysis incorporating 12 more T. pallidum strains for the regions carrying SNP changes in Nichols and Chicago, indeed initially suggested that this might be the case, and that the SNPs identified in Chicago and Nichols might reflect pathoadaptive changes the Nichols strain acquired following years of growth in the laboratory animal where it has been propagated so far. Interestingly however, in the DAL-1 genome (GenBank accession number NC_016844) [34], a T. pallidum strain recently isolated from the amniotic fluid of a pregnant woman [36], most of the Chicago/Nichols polymorphic loci were identical to Nichols sequences. Based on this evidence, we cannot exclude that Nichols and DAL-1 represent a separate naturally-occurring clonal lineage within T. pallidum. The significant predominance of non-synonymous polymorphisms between Chicago and Nichols strains strongly suggests the likelihood of a role of positive selection in microevolution of T. pallidum strains, whether due to differential adaptation during rabbit passage or pathoadaptation of individual strains in the human host.
Support for the mutational evolution of Nichols from an ancestral T. pallidum lineage also comes from the published genome of T. paraluiscuniculi (Cuniculi A strain, GenBank accession number NC_015715.1), closely related to T. pallidum [37]. In the Cuniculi A strain, nine of the Chicago/Nichols polymorphic loci (TP0051, TP0265, TP0430, TP0443, TP0488, TP0584, TP0748, TP0790, and TP0978) are identical to non-Nichols strains that were analyzed here, confirming the “Nichols-specific” nature of the mutations. Ongoing research in our laboratories using comparative genomics on a population-wide scale will provide an insight into phylogenetic relationships of T. pallidum clonal populations and likely will help explain the role of such sequence changes during syphilis infection.
|
10.1371/journal.pcbi.1004313 | Detailed Per-residue Energetic Analysis Explains the Driving Force for Microtubule Disassembly | Microtubules are long filamentous hollow cylinders whose surfaces form lattice structures of αβ-tubulin heterodimers. They perform multiple physiological roles in eukaryotic cells and are targets for therapeutic interventions. In our study, we carried out all-atom molecular dynamics simulations for arbitrarily long microtubules that have either GDP or GTP molecules in the E-site of β-tubulin. A detailed energy balance of the MM/GBSA inter-dimer interaction energy per residue contributing to the overall lateral and longitudinal structural stability was performed. The obtained results identified the key residues and tubulin domains according to their energetic contributions. They also identified the molecular forces that drive microtubule disassembly. At the tip of the plus end of the microtubule, the uneven distribution of longitudinal interaction energies within a protofilament generates a torque that bends tubulin outwardly with respect to the cylinder's axis causing disassembly. In the presence of GTP, this torque is opposed by lateral interactions that prevent outward curling, thus stabilizing the whole microtubule. Once GTP hydrolysis reaches the tip of the microtubule (lateral cap), lateral interactions become much weaker, allowing tubulin dimers to bend outwards, causing disassembly. The role of magnesium in the process of outward curling has also been demonstrated. This study also showed that the microtubule seam is the most energetically labile inter-dimer interface and could serve as a trigger point for disassembly. Based on a detailed balance of the energetic contributions per amino acid residue in the microtubule, numerous other analyses could be performed to give additional insights into the properties of microtubule dynamic instability.
| The molecular machinery of chromosome segregation during cell division is one of the most sophisticated molecular biology mechanisms employing the interplay of different proteins and forces. The long filamentous tube-shaped microtubule structure is a central player in chromosome segregation and cell division, making it an important physiological and therapeutic target. However, the driving force for microtubule disassembly and dynamic instability, and hence force generation, is still not fully understood. In our all-atom molecular dynamics simulations we calculated the energy of interactions, within a microtubule cylinder, that is responsible for microtubule stability. We broke this energy down to individual contributions of every residue and domain. Different energy profiles enabled us to unravel the driving force behind microtubule disassembly and force generation, a longstanding unanswered biological question. We also elucidated the mechanism of disassembly and explained the effects of different factors on disassembly rates. Our list of energetic contribution of single amino acid residues could also serve in tailor-designing engineered microtubules that could be used for therapeutic and diagnostic purposes.
| Microtubules (MTs) are cellular organelles that participate in major cellular processes such as mitosis, cell shape maintenance, cell motility and motor protein transport and constitute a major target for a wide range of drugs, most notably anti-mitotic chemotherapy agents such as paclitaxel. Due to their importance in cell biology, MTs have been the topic of active research into their structure and function for several decades [1]. The pivotal role of MTs in cell division, by forming the mitotic spindle that segregates chromosomes, makes them an important target for antimitotic cancer chemotherapy drugs [2, 3].
The peanut-shaped αβ-tubulin heterodimer is the building block of MTs [4]. Tubulin heterodimers associate longitudinally to form protofilaments, which in turn associate laterally to form a left-handed three-start helix with a seam, that results in the most common microtubule structure, the so-called B lattice [5]. Since tubulin dimers polymerize end to end, MTs become polarized, meaning that one end has α-subunits exposed (minus end) while the other end where faster growth usually occurs has β-subunits exposed (plus end) (Fig 1A) [6]. Within a tubulin heterodimer, GTP binds at the α-tubulin N-site which occurs at the intra-dimer interface. This GTP molecule does not undergo hydrolysis. Another GTP molecule attaches at the β-tubulin E-site and undergoes hydrolysis to GDP and phosphate shortly after assembly [7], in a process which drives the stochastic switching between growth and shrinkage in MTs. This unique property of microtubules is commonly referred to as dynamic instability [8]. Mitchison and Kirschner proposed the so-called GTP-cap model, which states that as long as the plus end of an MT is capped with GTP, it continues to grow. However, if GTP hydrolysis is sufficiently fast to catch up to the growing tip of the MT, rapid shrinkage, called a catastrophe, results [9]. Upon binding to an MT, some pharmacological agents such as taxol or epothilone stabilize the system and inhibit shrinkage [10]. Several studies have been conducted to determine which specific structural transitions that accompany GTP hydrolysis or taxol binding are responsible for their effect on MT stability, especially the transition of the tubulin dimer between its straight and curved states [11–15]. In the most recent of these studies, Alushin et al. found that GTP hydrolysis leads to a compaction around the E-site nucleotide which is reversed upon taxol binding [15]. This compaction was proposed to generate a strain which is powered by the energy of GTP hydrolysis and is believed to be released only through outward curving of protofilaments, initiating disassembly [16]. A missing component in these studies, however, is the quantification of the free energy changes that accompany these structural transitions. Due to the difficulties related to its experimental measurements, many simulations have been conducted to study detailed MT energetics [17–22]. In a recent study we have analyzed the strength of hydrogen bonds that bring and hold tubulin subunits together within different lattice configurations [23]. However, in all of these simulations, several factors were still missing. Most importantly, the full energetics of a complete MT model, which is essential to understanding the thermodynamics of tubulin assembly, has not been estimated yet due to the high computational price associated with such analyses. A detailed energy balance involving contributions due to each residue, domain or subunit, to the best of our knowledge, was never considered.
As a result of recent advances in computational technology, GPU-based computations can now be implemented to perform very demanding calculations in a reasonable amount of time. With this technology readily available, we simulated two complete all-atom MT models and studied in detail their energetics. The models studied are: (a) an MT with GDP in the E-site (GDP-Model) and (b) an MT with GTP in the E-site (GTP-Model). We did not need to look for a non-hydrolyzable analogue of GTP as hydrolysis is not a problem in molecular dynamics simulations, in contrast to experimental procedures [16]. The MT model that we used was initially built by Wells and Aksimentiev [24] utilizing sophisticated theoretical techniques to combine experimental structural information from a cryo-electron microscopy map of MT at 8 Å resolution [25] and electron crystallography structure of tubulin at 3.5 Å resolution [26]. We combined this model with the recently published crystal structures [15] in order to generate an atomistic representation involving an infinite number of infinitely long MTs. This is possible due to the use of periodic boundary conditions. (see S1 Movie).
A 50-ns MD trajectory was analyzed for several equilibration aspects, the first of which is the root-mean-square deviation (RMSD) of the backbone atoms relative to the starting structure. In addition, two nearly perpendicular MT cylinder diameters, namely Dx and Dy, were also calculated along the trajectory. Referring to the tubulin dimer numbering in Fig 1B, the diameter Dx was defined as the distance between the center of mass of dimer 4 and the center of mass of dimer 10 and 11, while Dy was defined as the distance between the center of mass of dimer 1 and the center of mass of dimer 7 and 8. In both diameters, only the distance projection on the x-y plane was considered as this is what gives the cylinder diameter. Plots showing the change in RMSD of the backbone atoms, Dx and Dy over simulation time for the GDP- and GTP-Models are shown in Fig 2A and 2B. The two diagrams indicate a strong correlation between fluctuations in RMSD and in diameters which indicates that most of RMSD fluctuations are due to changes in the circular shape of MT cylinders rather than the rearrangement of domains. The two diagrams also show the flexibility of MT cylinders as they deform spontaneously from a circular to an oval shape and vice versa. Movies showing the change of the two diameters over simulation time can be found in Supporting Information (see S2 and S3 Movies).
Since our particular interest is in the MT energetics, we used the overall MT energy across lateral and longitudinal inter-dimer interfaces as an indication of whether the system is equilibrated or not. Hence, we calculated these energies using MM/GBSA and the formula in Eqs 1 and 2 and plotted the total energy per MT ring versus simulation time (Fig 2C and 2D). Both plots indicate that the overall lateral and longitudinal energies in both the GDP- and GTP-Models have already equilibrated at least before the last 20 ns of the MD simulation time. The plots also show that the large fluctuations in RMSD or Dx and Dy hardly affect the MT energetics at either of the two interfaces, which is a good indication of the energetic stability of our models.
Total breakdown of the predicted energy contributions enabled us to perform the analysis for different residues, domains, subunits, and dimers across both lateral and longitudinal inter-dimer interfaces. Before listing the results, it should be noted that energies calculated via the MM/GBSA method do not necessarily reflect absolute energy values. Rather, they are used for relative comparison within the same model [27]. It should also be noted that all energies listed here are calculated per MT ring, unless otherwise specified.
As Table 1 summarizes, the overall energy of interaction across the 13 lateral tubulin interfaces (see Fig 1B), E t o t l a t, was found to be −411±29 kcal/mol, nearly 60% of which is due to α-α interactions and the rest is due to β-β interactions. On the other hand, the contribution of the dimer acting as a receptor (see the explanation of the ligand/receptor convention in the methods section and in Fig 3A and 3B), E R l a t, was about 54% of the overall energy while the rest was attributed to the ligand, E L l a t, with the difference entirely attributed to solvation effects rather than direct interactions. It should be noted, however, that the α subunit of the ligand (Lα) and the β subunit of the receptor (Rβ) together contribute −312±29 kcal/mol which is nearly 75% of E t o t l a t, with the Lα contribution slightly larger than that due to Rβ. The contribution of Lβ and Rα was found to be much smaller, only 25% of E t o t l a t. Upon structural inspection, this 50% difference, being almost entirely due to electrostatic interactions, was attributed to diagonal interactions between subunits; although the interface between Lα and Rβ is dominated by oppositely-charged residues and thus stabilizing the interaction, the opposite is true at the destabilizing interface between Rα and Lβ which has, for example, residues Rα/Glu220 and Lβ/Asp130 destabilizing the lateral interface by 12±1 and 10±2 kcal/mol, respectively.
As to the energetic breakdown according to interaction types, the contribution of the van der Waals and non-polar solvation energy, E(vdW+SA), to the overall energy is largely stabilizing with an average value of −1476 kcal/mol, 85% of which is due to the vdW interactions. This stabilization is opposed by destabilization due to electrostatic interactions; the average sum of electrostatic and the polar solvation energy, E(ele+GB), is 1065 kcal/mol. This is expected since tubulin dimers are highly negatively charged and tend to repel each other.
Regarding the detailed energy contributions per individual residues, the most important residue across the lateral interface was found to be Rβ/Tyr283 followed by Rα/His283 and Lα/His88, with overall stabilization energies of −90±5, −47±5 and −42±3 kcal/mol per MT ring, respectively. Rβ/Tyr283 alone supplies more than 20% of lateral stability most of which is due to the vdW interactions. In fact, most of the stabilizing residues on top of our list were neutral ones with a strong stabilizing vdW component. On the other hand, almost all of the destabilizing residues were charged ones with a strong electrostatic component, most destabilizing of which is Lβ/Lys124 with an energy of 22±7 kcal/mol. A complete list of the different energetic contributions of each residue in the ligand and receptor per MT ring is provided in the Supporting Information.
Domain contributions to the overall energy per MT ring were also calculated and Fig 4A and 4B show the most relevant of them. The contribution of the M-loop in both α and β subunits is by far the largest, with values of −112±10 and −159±10 kcal/mol, respectively, making up about two thirds of the energy of the overall lateral interactions. This agrees well with previous predictions, although precise values of their energetic contributions were never calculated [25, 28, 29]. Other less important domains are the Lα/N-terminal loop, Lα/H2-S3 loop, Lα/H3 helix and Lα/H9 helix at the α interface with a stabilization of −72±6, −62±6, −57±10 and −16±7 kcal/mol, respectively [25, 28]. Lβ/H3 helix at the β-β interface, however, has a strongly destabilizing effect of 37±8 kcal/mol. This supports previous predictions based on structural analysis by Li et al. and Nogales et al; however, these authors did not specify if these interactions are stabilizing or not [25, 28]. Additionally, Lβ/H2” helix and Lβ/H1’-S2 loop also have relatively strong stabilizing contributions of −53±7 and −43±5 kcal/mol, respectively.
As Table 1 summarizes, the overall interaction energy across the lateral interface in the GTP-Model, E t o t l a t, was found to be −482±29 kcal/mol, nearly 60% of which is due to the α-α interactions. This average overall energy is 71 kcal/mol (nearly 20%) more stable than the overall energy of the GDP-Model which explains the role of GTP in stabilizing MTs as will be shown later. Nearly 90% of this difference in stability is solely attributed to enhancement of the contribution of the ligand, both α- and β-subunits, rather than the receptor. As was noticed in the GDP-Model, Lα and Rβ are also responsible for most of the lateral stabilization in the GTP-Model, −338±22 kcal/mol (70% of E t o t l a t).
Upon breakdown of the interaction energy to its individual components, we find that in the GTP-Model, the E(vdW+SA) contribution becomes −1432 kcal/mol while E(ele+GB) becomes 950 kcal/mol. Comparing this to the GDP-Model, it turns out that GTP destabilizes the vdW and non-polar solvation interactions by 44 kcal/mol and stabilizes electrostatic and polar solvation interactions by 115 kcal/mol, which results in the net stabilization of 71 kcal/mol as mentioned earlier. This difference becomes clear by analyzing Fig 4A and 4B for domain contributions and Fig 5 for residual contributions. It is apparent from Fig 4A that GTP strengthens the contributions of the Lα/H3 helix and Rα/H9 helix by 23±10 and 20±16 kcal/mol, respectively. Most of this helix stabilization can be attributed to interactions involving Rα/Glu290 (residue number in Fig 5, i, is 290), residue Lα/Asp127 (i = 998), and residue Lα/Arg123 (i = 994). These three residues stabilize the GTP-Model over the GDP-Model by energy values of 31, 20 and 19 kcal/mol, respectively, mostly due to electrostatic interactions. Upon structural analysis it is apparent that GTP slightly rotates the dimer acting as a ligand toward the one acting as a receptor, thus allowing stronger interactions between H3 and H9 helices with oppositely-charged residues. GTP also enhances the stability imparted by the Lα/H2-S3 loop and the Rα/H10-S9 loop, although it moderately decreases the role of the Lα/N-terminal loop as well as the Rα/M-loop in the overall MT stability.
Similar conclusions are reached in regard to the β-subunit and the effect of the Lβ/H2” helix through residue Lβ/Asp90 (i = 1401) and the Rβ/M-loop through residue Rβ/Arg284 (i = 724). Both domains are stabilized in the GTP-Model by extra 18±10 and 10±15 kcal/mol compared to the GDP-Model, respectively. The charged nature of all these residues explains why most of GTP stabilization is manifested in E(ele+GB) not E(vdW+SA). Fig 4B also shows that GTP reduces the destabilization caused by the Lβ/H3 helix and the Lβ/H3-S4 loop. On the other hand, GTP reduces stability imparted by the Lβ/H1’-S2 loop and the Rβ/H9 helix. Details of the contribution of each residue in the GTP-Model can be found in the Supporting Information.
Analysis of the strength of interactions across the longitudinal inter-dimer interface in the GDP-Model yielded, as summarized in Table 2, an overall energy of −1240±32 kcal/mol per MT ring, which is nearly three times the lateral interaction energy. This is in agreement with structural observations [28]. Due to the orientation of tubulin dimers at the longitudinal inter-dimer interface, the contributions of Lα and Rβ are essentially zero and will always be neglected here. On the other hand, the contribution of Lβ is 54% of the total value, and the remainder is contributed by Rα. The breakdown of this energy yields an average E(vdW+SA) of −2668 kcal/mol which is almost twice as large as the value across the lateral interface. This is obviously due to the tighter packing of the residues here as opposed to looser packing at the lateral interface. The average E(ele+GB) across the longitudinal inter-dimer interface is 1428 kcal/mol and it is 34% larger than its value at the lateral interface.
Per-residue energy analysis reveals the most important residues to longitudinal stability, the first of which is Lβ/Arg401 from the H11-H11’ loop which alone supplies −101±7 kcal/mol (nearly 10%) [23]. After that come residues Lβ/Phe404 and Lβ/Trp407 from the H11’ helix both of which support longitudinal stability by contributing −91±3 and −78±3 kcal/mol, respectively. This makes the two former domains, which constitute part of the tubulin C-terminal domain, the most critical for longitudinal stability in the β-subunit (Fig 4D). The figure also shows that the following domains in the Lβ subunit: the T5 loop, T3 loop, and T2 loop are also very important for longitudinal stability. The role of the GDP cofactor appears quite influential at the longitudinal inter-dimer interface, in contrast to the lateral one. It is primarily destabilizing with a large contribution of 79±11 kcal/mol due mainly to a strong electrostatic repulsion with the highly negative environment, despite its strong salt bridge with Rα/Lys352. Residual analysis of the Rα subunit also shows some relatively less important residues; Rα/Trp346, Rα/Tyr262 and Rα/Lys352 with energy contributions of nearly −60 kcal/mol for each of them. These and other residues are responsible for the following domains in the Rα subunit: the H10-S9 loop, H8-S7 loop, and the S9 strand being the top stabilizers in Fig 4C. The Rα/H8 and Rα/H10 helices are also relatively important for longitudinal stability. Both the Rα/C- and Rα/N-terminal domains are important as well, with the Rα/N-terminal loop being a destabilizer, in contrast to its role at the lateral interface.
As summarized in Table 2, the overall interaction energy across the longitudinal inter-dimer interface in the GTP-Model was found to be −1098±30 kcal/mol per MT ring, which is 141 kcal/mol (10%) less stable than the GDP-Model system. This difference is attributed to a 7% decrease in the Rα and 3% decrease in the Lβ interactions. Upon energetic breakdown we see that GTP destabilizes the vdW and non-polar solvation energy by nearly 250 kcal/mol, while stabilizing electrostatic and polar solvation energy by nearly 110 kcal/mol. This could be due to the longstanding observation that GTP leads to an expansion in the E-site and lengthening of the tubulin dimers. That is, axial dimer repeat changes from 81.20 Å in GDP-tubulin to 83.38 Å in GTP-tubulin [12, 15]. This reduces the packing of atoms at the interface and hence lowers both the vdW attraction and electrostatic repulsion, the former being affected most due to its stronger dependence on distance.
Looking into domain contributions in Fig 4C and 4D we see how GTP destabilization of longitudinal interactions can be subdivided. The most pronounced difference between the GDP- and the GTP-Model appears in regard to the cofactors at the E-site. Although GDP was largely destabilizing in the GDP-Model, GTP becomes relatively largely stabilizing, with an energy change from the GDP-Model of nearly −125±14 kcal/mol. However, this change should not be considered without taking into account the effect of the Mg2+ ion that accompanies GTP. This magnesium ion introduces an instability of 95±4 kcal/mol to the GTP-Model. Hence, the overall effect of replacing GDP by GTP and a magnesium ion is a stabilization of 30 kcal/mol on average. Other causes of the lack of stability in the GTP-Model Lβ include the decrease in the contribution of the H11’ helix because GTP offsets interactions by Lβ/His406 (i = 1709) by as much as 25 kcal/mol. This is because this histidine is protonated in the GTP-Model and neutral in the GDP-Model and therefore behaves differently in both cases. Being charged in the GTP-Model, it is distracted from the strong attractive vdW interactions it makes with the Rα/H8-S7 loop by electrostatic and hydrogen bonds with other residues within the Lβ subunit. GTP also causes longitudinal stabilization due to the domains: the H2 helix and the T2 loop to decline while causing stabilization due to the H11-H11’ loop and the T5 loop to rise. As to the Rα-subunit (Fig 4C), stabilization due to several domains declines in the GTP-Model. These domains include the T7 loop, the S9 strand, the C-terminal loop, the H10 helix, the H12 helix, the H8-S7 loop, and the H10-S9 loop. In short, the GTP-Model is longitudinally less stable than the GDP-Model in most of the domains occurring at the longitudinal inter-dimer interface. An exception to this rule is the increased stabilization due to the C-terminal tail, the N-terminus and the H8 helix, Fig 4C shows the extent of stabilization or destabilization imparted by GTP on each domain. We should also mention that the strong attraction of the Rα/T7 loop emerging after GTP hydrolysis (Fig 4C) could explain the proposed compaction of the E-site after GTP hydrolysis [15]. In fact, the overall increase in longitudinal dimer-dimer attraction after GTP hydrolysis, as shown by the different values of E t o t l o n g in both models, explains the driving force for this E-site compaction.
Among other important residues, Rα/Lys352 (i = 352) of the domain S9 strand has a largely reduced contribution in the GTP-Model, as shown in Fig 5, which is 37 kcal/mol less stabilizing than in the GDP-Model. While having comparable vdW contributions in the two models, this residue suffers strong repulsion probably due to the nearby Mg2+ ion in the GTP-Model. Another important residue is Rα/Val440, located in the C-terminus of the α-subunit in our model. GTP enhances the stabilization caused by this residue by nearly 33 kcal/mol over the GDP-Model. Additional important residues and their contributions are shown in the Supporting Information.
Depolymerizing MTs display protofilaments that peel into “ram’s horns” formations under high magnesium buffer conditions. The ends of MTs become frayed, however, under physiological concentrations of magnesium [11]. The energy profile throughout the longitudinal inter-dimer interface provides a clear explanation for the disassembly mechanism, its driving force, and its relation to Mg2+ concentration. We characterized each residue in the longitudinal subsystems by its radial distance from the MT lumen in Å, which was plotted on the x-axis. The interaction energies of residues, per MT ring, over half-closed intervals of [x, x+3) were summed up and plotted on the y-axis to produce the radial energy profiles in Fig 6A, 6B and 6C.
The diagram in Fig 6A leads to a striking observation that the energy distribution throughout the longitudinal inter-dimer interface is not even, with the outward portion (x > 30 Å) largely outweighing the inward portion (x < 30 Å), with the center of mass of tubulin being at x ≈ 30 Å. To mention specific values, in the GTP-Model, the outward portion provides nearly −956 kcal/mol while in the GDP-Model it provides −982 kcal/mol, both values being larger than 80% of the overall longitudinal interaction energy. This uneven distribution of energy, or forces of attraction, is proposed to yield a strong torque that tends to curl MT protofilaments outwardly, breaking lateral bonds and promoting disassembly as illustrated in Fig 7A.
Radial energy profiles of different components of the interaction energy are also shown in Fig 6B and 6C, where electrostatic interactions cause very strong repulsion through the inward portion and attraction only at the periphery where the H11-H11’ loop and particularly residue Lβ/Arg401 are located. We propose a pivotal role for this residue, and for the entire C-terminal domain, in regulating dynamic instability. Electrostatic repulsion by the inner domains and attraction by the outer C-terminal domain is the recipe for outward curling and disassembly in MTs. The vdW distribution will also work, as shown in Fig 6B for the GDP-Model and Fig 6C for the GTP-Model, to curl protofilaments outward until the vdW contacts, and other components, are balanced out.
The largely destabilizing Mg2+ ion (see Fig 4D) also plays an important role. Even though GDP at the E-site has low affinity for Mg2+ [30], it may still attract Mg2+ if it is present in high concentrations or Mg2+ may stay in the E-site after GTP hydrolysis. This largely destabilizes the inner portion of the protofilament (blue dashed arrow in Fig 6A), allowing outward forces to pull tubulin out with even less resistance from the other side, thus promoting outward curling and MT disassembly. This explains why large Mg2+ concentrations promote ram’s horns formations [31] and increase the rate of disassembly [32, 33], while its low concentrations produce frayed ends and lower rates of disassembly [11].
To explain MT disassembly from a free energy perspective, Fig 7A shows an illustration of the analyzed situation. As already established, uneven distribution of attractive interactions along the longitudinal inter-dimer interface favors outward curling. In the GTP-Model, outward curling is favored by −956 kcal/mol of interaction energy outwardly with respect to the center of mass of tubulin, as compared to −982 kcal/mol in the GDP-Model. These curl-favoring energies/forces are opposed by the lateral interaction energies which tend to pull protofilaments back from both sides, i.e. double the effect. The magnitude of this effect is 2 × E t o t l a t, giving −964 kcal/mol in the GTP-Model which is much larger than −822 kcal/mol in the GDP-Model, all energies given per MT ring. We propose that this lateral inward pull balances out the longitudinal outward push in case of the GTP-Model. That being said, the presence of a GTP cap at the tip of the MT would prevent outward curling and thus provide stability for the entire MT structure. After GTP hydrolysis reaches the cap, however, lateral bonds become weaker and longitudinal outward push manages to break the lateral contacts, causing outward curling and MT disassembly. High concentrations of Mg2+ may also increase outward curling and the disassembly rate, as explained earlier.
Similar observation could be made about the tangential energy profiles at the longitudinal inter-dimer interface. Fig 6D, 6E and 6F show the tangential energy profiles with the x-axis showing the distance from the laterally adjacent protofilament. On the x-axis, x < 30 is the tubulin intermediate domain while x > 30 is the nucleotide binding domain with x ≈ 30 being at the center of mass (see Fig 7B). Fig 6D shows that in The GTP-Model, the distribution is also uneven with right-side portion being −1023 kcal/mol (nearly 93% of the total) as compared to −887 kcal/mol (71% of the total) in the GDP-Model. This means that in the GTP-Model, there is a strong force tilting it sideways. However, after GTP hydrolysis and rearrangement of domains at the longitudinal inter-dimer interface, that force largely decreases and the uneven distribution starts to balance out, as shown in Fig 6D, decreasing the strain on lattice integrity. This is in perfect agreement with the recent findings of Alushin et al. [15] They observed that GTP hydrolysis and the release of an inorganic phosphate group leaves a hole within the longitudinal inter-dimer interface between tubulin dimers producing a strain which results in sideway tilting in the same direction [15, 16]. In the present work we show that this tilting is also driven by the uneven energy distribution along the same direction as in the work of Alushin et al. [15] (see Fig 7B). However, this sideway tilting should not be considered as the the driving force for disassembly since it is orthogonal to the outward curling. Combining the two effects together, we conclude that uneven distribution at the longitudinal inter-dimer interface generally leads to a large outward and slight sideway tilting of protofilaments, the former of which is responsible for disassembly of GDP-bound MTs.
As mentioned in the Methods section, the MT ring was divided into 13 subsystems of laterally adjacent tubulin dimers and another 13 subsystems of longitudinally adjacent tubulin dimers (see Fig 3). All of the energies presented earlier were expressed per MT ring, meaning that they were summed over the 13 subsystems. In this section, however, we focus on the interaction energy in each subsystem. Fig 8A and 8B show energy diagrams for lateral and longitudinal interactions superposed over the MT ring. We first note that the shape of the lateral interactions (Fig 8A) in the GDP-Model is very distorted with several “kinks” of very low energy. When compared to the GTP-Model, its shape is much less distorted. This could come as a straightforward consequence of the fact that GTP-Model is laterally more stable than the GDP-Model and hence suffers less “deformations”.
It is worth mentioning that the deepest of the kinks in the GDP-Model energy diagram, i.e. the interface with the weakest binding energy, is the one occurring at the seam (between dimer 13 and dimer 1), in contrast to its strength in the GTP-Model. It has a binding energy of −9±7 kcal/mol which is very low compared to the one at the interface between dimer 12 and 13, for example, which has an energy of interaction equal to −57±9 kcal/mol. We predict that protofilaments number 1 and 13 having very strong longitudinal contacts antagonized by very weak lateral contacts at the seam, will be the first to dissociate laterally and curve outwards. This should open the MT cylinder which should then trigger disassembly. Therefore, MT energetics suggest that the seam is the most labile inter-dimer interface in the MT structure and could act as a trigger point for disassembly. This is precisely what was reported recently [34].
The energy diagrams at the longitudinal inter-dimer interfaces (Fig 8B) appear to be more even than at the lateral interfaces. However, we see no major difference in the pattern between the GTP-Model and the GDP-Model except that longitudinal interactions in the GDP-Model are stronger, which was established earlier.
We used sophisticated all-atom molecular dynamics simulations to produce accurate MT models, combined with high resolution cryo-electron microscopy maps, to generate an infinite number of infinitely long MT representations. The MM/GBSA energy analysis that followed the simulations enabled an estimate of the contributions of individual residues, domains, subunits and dimers toward the lateral and longitudinal stability of a complete MT ring. We found that longitudinal interactions are about two to three times stronger than lateral interactions explaining the greater stability of the MT structure along its axis than radially. This finding agrees with previous structural observations [28] and computational estimations [18, 22]. We also found that interactions are not evenly distributed radially along the longitudinal inter-dimer interface. That is, attractive interactions are largely concentrated away from the MT lumen, producing a force that curls protofilaments outward and eventually causing MT disassembly. The GTP-Model was laterally more stable than the GDP-Model and the opposite was true for the longitudinal inter-dimer interface. Since lateral forces oppose outward curling while longitudinal forces support it, we expect the GTP-Model to be less prone to disassembly than the GDP-Model. With its lateral forces being strong enough to prevent outward curling caused by longitudinal forces, the GTP-cap at the plus end can stabilize an entire MT cylinder. After GTP hydrolysis reaches the cap, lateral forces are too weak to prevent outward curling, especially at the seam which has the weakest lateral contacts. This results in outward curling and microtubule disassembly.
We also confirmed that the MT seam is most likely to act as a trigger point for MT disassembly by being the most labile interface in the MT cylinder [34]. Magnesium ion was demonstrated to be an influential factor in MT stability. Being present at the inner portion of the longitudinal inter-dimer interface, the largely destabilizing Mg2+ ion repels the inward portion and enhances outward curling, the formation of ram’s horns structures and rapid disassembly, which is consistent with key experimental findings [11]. This action of Mg2+ at the E-site of tubulin is suppressed by GTP in GTP-capped MTs. As we showed earlier, the ensemble of Mg2+ and GTP at the E-site is collectively stabilizing. However, hydrolysis of GTP and release of inorganic phosphate create a gap at the longitudinal inter-dimer interface and leave the largely destabilizing ensemble of GDP and Mg2+ which rapidly promotes outward curling to fill this gap. This happens only at large Mg2+ concentrations since GDP at the E-site has low affinity for Mg2+ [30]. At low Mg2+ concentrations, disassembly becomes slower and outward curling becomes less pronounced [11].
Tangential energy profiles at the longitudinal inter-dimer interface were also shown to be uneven and confirmed the hypothesis that GTP hydrolysis produces a strain which promotes sideway titling [15, 16]. However, much of this strain could be tolerated within the lattice constraints and its orthogonality to the direction of outward curling rules out its role in disassembly.
We also identified the most important residues and domains with respect to MT stability at both interfaces and their energetic contributions. At the lateral interface, the α/M-loop, β/M-loop, α/H3 helix, α/N-terminal loop and the α/H2-S3 loop were shown to be most stabilizing while the β/H3 helix was actually destabilizing. This supports predictions based on structural studies [25, 28]. Residue α/Tyr283 was shown to form a very strong network of vdW interactions with neighboring residues and to provide the largest amount of stability at the lateral interface. At the longitudinal inter-dimer interface, the β/C-terminal domain was found to be of paramount importance not only to stability but also to the mechanism of MT disassembly. In particular, residues β/Arg401, β/Phe404, and β/Trp407 of the C-terminal H11 helix and the H11-H11’ loop were shown to provide more than 20% of longitudinal stability in both the GTP- and GDP-Models. The complete breakdown of MT energetics per every single residue was further analyzed in order to provide crucial insights into many aspects of MT dynamic instability. Of highest importance is the calculation of the amount of force generated through outward curling due to uneven longitudinal interactions. This could help unravel many aspect of the molecular machinery of cell division, in particular the force generation requirement for chromosome segregation.
As a future prospect, simulation of a free protofilament is necessary in order to find out about the effect of uneven longitudinal energy distribution on the extent of outward curling. By comparing the energy of a free protofilament to the energy of a protofilament constrained within our MT model, we can predict the amount of free energy released by outward curling and additional light could be shed on the mechanism and driving forces in MT disassembly. Also, simulating a GDP-Taxol case is necessary to understand the molecular mechanisms by which taxol bound to an MT prevents outward curling and MT disassembly.
The recent structures for GMPCPP and GDP bound MTs at resolutions of 4.7 and 4.9 Å, respectively [15], represented an excellent starting point for building the models presented here. The 3×3 lattice PDB structures of 3J6E (with GMPCPP) and 3J6F (with GDP) were processed using MOE software [35] by the addition of hydrogens and prediction of ionization states. The central tubulin dimer of the 3×3 lattice in each case was separated and was repaired by the addition of missing residues (Residue 1 in β-tubulin and residues 1,39 to 48, 440 in α-tubulin) from the PDB structure 1TUB [36], using MOE. We modified GMPCPP into GTP since in our simulations there is no need to use the nonhydrolyzable GTP analogue as hydrolysis is not expected in MD simulations. Next, for both GTP and GDP systems, the repaired tubulin was superimposed over the 13 tubulin dimers in the complete MT model built by Wells and Aksimentiev [24], producing a hybrid complete MT model for both systems. Thus, we produced two models, the GTP-Model and the GDP-Model, by combining the helical structural configuration developed by Wells and Aksimentiev with the lattice tubulin coordinates obtained from Alushin’s model. Several clashes existed at lateral interfaces between tubulin dimers and were resolved through a short minimization using the Generalized Born (GB) continuum model in Amber [37].
Each model, as shown in Fig 1B, has 13 tubulin dimers in an MT orientation. For the GDP-Model, each tubulin has GTP, Mg2+ and four coordinating water molecules at the α-tubulin N-site, and GDP at the β-tubulin E-site. For the GTP model, there was GTP, Mg2+ and four coordinating water molecules at both the N-site and the E-site. Solvation was carried out using box of dimensions 293.85 × 293.85 × 83.38 (or 81.20) Å3 for the GTP- and GDP-Models, respectively. The z-component was obtained from Alushin’s lattice structure [15] and ensures perfect longitudinal alignment of tubulin dimers in both systems (see Fig 1B). Both x and y components were obtained from Wells’ structure [24]. A total of 181,000 TIP3P water molecules were added in the solvation box. This number was obtained based on several optimization trials which guaranteed consistency in box dimensions and density throughout the simulations. A total number of 442 Na+ ions was needed for neutralizing the GTP-Model, versus 455 for the GDP-Model. An extra 327 Na+ and Cl− ions were added to bring the salt concentration to 0.1 M.
During the addition of water and ions, we made sure that no atoms were placed in positions which will be occupied by the periodic images of our system in both the positive and negative z direction (see the gaps in the water box of Fig 1B). Thus, exploiting the periodic boundary conditions, the mirroring of our nearly 720,000-atom system in all directions should effectively result in an infinite number of infinitely long MTs, (see S1 Movie). The AMBER Molecular Dynamics package was used for solvation, ionization, and dynamics [37].
The all-atom forcefield AMBERff12SB was used to parameterize the protein [38, 39]. Cofactors were parameterized utilizing the parameter set developed by Meagher et al. [40]. Each of the two systems was then minimized through nearly 1000 steps of the steepest descent algorithm followed by about 6000 steps of the conjugate gradient algorithm. Then, the systems were heated, with restraints of 10 kcal mol−1 Å−2 on the protein, to a temperature of 310 K using the Langevin thermostat over 20 ps under constant volume. This was followed by 200 ps of density equilibration under constant temperature and pressure, in which the restraints were eliminated gradually, followed by a production phase of 50 ns for each system. Simulations were performed using NVIDIA Tesla K20X GPU cards on the PharmaMatrix Cluster (University of Alberta) through AMBER GPU-accelerated code [41–43]. All simulations were performed using periodic boundary conditions employing the particle-mesh Ewald method for treating long-range electrostatics and a non-bonded cut off of 10.0 Å under constant pressure with anisotropic pressure scaling.
The 50-ns trajectory of each system was analyzed for several structural and conformational aspects. Most of the analysis was done utilizing the CPPTRAJ module in AMBER [44], MM/GBSA implementation in AMBER [45] plus several scripts that we designed to facilitate data analysis. The software VMD 1.9.1 was also used for viewing trajectories and image rendering [46].
Data analysis included calculating the total as well as the per-residue MM/GBSA binding energies [47] between pairs of tubulin dimers in lateral and longitudinal orientations. These calculations involved all the 13 heterodimers included in the simulations and would always give the energy per MT ring (Fig 1B). Hence, energetic contributions were assessed via the equation:
E x = ϵ x ( R 13 L 1 ′ ) + ∑ k = 1 12 ϵ x ( R k L k + 1 ) (1)
for lateral systems, and the equation:
E x = ∑ k = 1 13 ϵ x ( R k L k ′ ) (2)
for longitudinal systems. In both equations, Ex represents an energetic contribution of a given residue, domain or subunit x per MT ring of 13 tubulin dimers shown in Fig 1B. In Eq 1, ϵx(Rk Lk+1) is the energetic contribution of the same entity x in a subsystem composed of only tubulin k, treated as a “receptor”, and tubulin k+1, treated as a “ligand”. ϵ x ( R 13 L 1 ′ ) does the same but at the lateral seam, taking into account the flip between α- and β-subunits. In Eq 2, ϵ x ( R k L k ′ ) carries the same concept except that the ligand in a longitudinal subsystem is simply the periodic image of the receptor, hence the prime. Therefore, we ended up investigating 12 lateral subsystems plus 1 lateral subsystem at the seam and 13 longitudinal subsystems, for each model. An illustration of each subsystem is shown in Fig 3. Hence, our convention in this work is that the dimer whose M-loop is involved in lateral interactions is always termed “receptor” in lateral subsystems, and the dimer whose α-tubulin is involved in longitudinal interactions is always termed “receptor” in longitudinal subsystems. This distinction was necessary since we noticed that energetic contributions can vary between tubulin dimers acting as receptors and those acting as ligands.
All the energy calculations were performed on 200 evenly-spaced snapshots from the last 10 ns of the molecular dynamics trajectory where equilibration was confirmed. A solvent and solute dielectric constant of 80 and 1, respectively, were used for electrostatics in the Amber MM/GBSA implementation.
|
10.1371/journal.ppat.1004835 | Conserved Streptococcus pneumoniae Spirosomes Suggest a Single Type of Transformation Pilus in Competence | The success of S. pneumoniae as a major human pathogen is largely due to its remarkable genomic plasticity, allowing efficient escape from antimicrobials action and host immune response. Natural transformation, or the active uptake and chromosomal integration of exogenous DNA during the transitory differentiated state competence, is the main mechanism for horizontal gene transfer and genomic makeover in pneumococci. Although transforming DNA has been proposed to be captured by Type 4 pili (T4P) in Gram-negative bacteria, and a competence-inducible comG operon encoding proteins homologous to T4P-biogenesis components is present in transformable Gram-positive bacteria, a prevailing hypothesis has been that S. pneumoniae assembles only short pseudopili to destabilize the cell wall for DNA entry. We recently identified a micrometer-sized T4P-like pilus on competent pneumococci, which likely serves as initial DNA receptor. A subsequent study, however, visualized a different structure - short, ‘plaited’ polymers - released in the medium of competent S. pneumoniae. Biochemical observation of concurrent pilin secretion led the authors to propose that the ‘plaited’ structures correspond to transformation pili acting as peptidoglycan drills that leave DNA entry pores upon secretion. Here we show that the ‘plaited’ filaments are not related to natural transformation as they are released by non-competent pneumococci, as well as by cells with disrupted pilus biogenesis components. Combining electron microscopy visualization with structural, biochemical and proteomic analyses, we further identify the ‘plaited’ polymers as spirosomes: macromolecular assemblies of the fermentative acetaldehyde-alcohol dehydrogenase enzyme AdhE that is well conserved in a broad range of Gram-positive and Gram-negative bacteria.
| Streptococcus pneumoniae often escapes prevention and treatment through rapid horizontal gene transfer via natural transformation. Uptake of exogenous DNA requires expression of a transformation pilus but two markedly different models for pilus assembly and function have been proposed. We previously reported a long, Type 4 pilus-like appendage on the surface of competent pneumococci that binds extracellular DNA as initial receptor, while a separate study proposed that secreted short, ‘plaited’ transformation pili act simply as peptidoglycan drills to open DNA gateways. Here we show that the ‘plaited’ structures are not competence-specific or related to transformation. We further demonstrate that these are macromolecular assemblies of the metabolic enzyme acetaldehyde-alcohol dehydrogenase—or spirosomes—broadly conserved across the bacterial kingdom.
| Despite medical advances and vaccination campaigns, respiratory tract invasion by Streptococcus pneumoniae remains a leading mortality cause worldwide [1–3]. A particular challenge in the prevention and treatment of pneumococcal infections lies in the bacterium’s striking genomic plasticity, as it allows for efficient antibiotic resistance development, capsular serotype switching and vaccine escape [4]. Horizontal gene transfer and chromosomal rearrangements typically result from the avid uptake and recombination of exogenous DNA known as natural transformation. A strictly regulated event, it occurs during a transitory state of the bacterium’s life cycle—competence—and requires the timed expression of a dedicated set of genes [5]. Among these are the genes of the comG operon, which are conserved among naturally competent Gram-positive bacteria and are homologous to the ones encoding Type 4 pili (T4P) and Type 2 secretion system (T2SS) pseudo-pili components in Gram-negative bacteria [6,7]. Although mechanistic studies of structural determinants for DNA uptake—such as putative transformation-specific cellular appendages—hold promise for the development of novel antiinfectives and helper compounds, there have been only limited and contradictory reports on the initial steps of this important biological process [8–10].
As until recently no pilus-like structure had been observed in any transformable Gram-positive bacterium, it had been postulated that the pneumococcal comG operon encodes a short T2SS-like pseudo-pilus that serves to destabilize the cell wall peptidoglycan for DNA entry [6,9,11]. The main experimental evidence for this model comes from a different transformable organism, Bacillus subtilis, where pilus length was indirectly deduced from biochemical data [9].
Our team identified a long, micrometer-sized, T4P-like pilus protruding on the surface of competent cells from different pneumococcal strains with wild-type genotype (Fig 1A) [10]. Among these are two highly transformable laboratory strains of different genetic background—R6 and TCP1251—as well as a capsulated clinical isolate—the G54 strain [10]. We showed that major constituent of the transformation pilus is the ComGC pilin and that the pilus is sensitive to mechanical stress, which can lead to its detachment from the cell (shearing) [10,12]. Finally, we showed that this transformation pilus binds extracellular DNA and proposed that it acts as the initial DNA receptor on the surface of competent pneumococci [10].
A subsequent study visualized completely different structures—short, ‘plaited’ polymers—in the medium of competence-induced S. pneumoniae [8]. Biochemical observation of significant ComGC release in the medium during competence convinced the authors that the ‘plaited’ structures corresponded to secreted transformation pili. After failing to immunolabel these structures, they expressed heterologously the whole comG operon in Escherichia coli and visualized the release of similar polymers [8]. Finally, they proposed a model, which is consistent with the classical but speculative model of transformation pseudo-pili: rather than acting as a DNA receptor, the pneumococcal transformation pilus acts as a peptidoglycan-drilling device whose release leaves a gateway for transforming DNA to find the uptake machinery [8,10].
Here we show definitive experimental evidence that the short ‘plaited’ filaments are not transformation pili or other structural determinants of natural transformation. We further identify the structures as fermentative spirosomes, or macromolecular complexes of the acetaldehyde-alcohol dehydrogenase enzyme AdhE, which is widely conserved across the bacterial kingdom. Being aware of the limited view and resolution that observation by electron microscopy provides, we underscore the need for thorough validation by orthogonal approaches. Finally, we briefly synthesize the present-day published collective knowledge by proposing an updated model of pneumococcal transformation.
Perhaps the most intriguing aspect of the Balaban et al. study is the distinctive morphology of the reported ‘plaited’ filaments themselves [8]. As the authors point out, the genetic makeup of the comG operon resembles significantly that of operons encoding T4P or T2SS components in Gram-negative bacteria [5,7,8,13]. This includes from sequence homology of the individual genes through their intraoperon organization to the putative bioassembly platform and post-translational modifications of the encoded components. The structure of both T4 pili and T2SS pseudo-pili has been extensively studied [14,15]. Generally T4 pilins pack tightly into thin but extremely strong and several micrometers long surface-attached helical filaments [14]. Typical dimensions vary from 5–6 nm width for the T4aP of many bacteria (Pseudomonas aeurginosa, Neisseria gonorrhoeae and others) to the thicker, about 8 nm wide T4bP of enteropathogens such as Vibrio cholerae and Salmonella enterica serovar Typhi [14]. Conversely, structural models of T2SS pseudopili, which normally act as short, protein ejecting pistons in the periplasm, present an architecture that is very similar to that of gonococcal T4aP (Fig 1B) [15]. Both T4P and T2SS pseudo-pili feature a grooved surface with relatively small protuberations characteristic of the pilin helical packing [14,15].
The characteristic structural features of T4P were conserved in the transformation pili that we visualized previously on the surface of competence-induced pneumococci from several different wild-type strains [10]. In contrast, the ‘plaited’ structures visualized subsequently represent short, 40–200 nm long structures that are significantly thicker (~ 10 nm) and present large protuberant domains on their helical surface (Fig 1C) [8]. The authors proposed that the filaments are ‘plaited’, i.e. composed of two interlaced transformation pili [8]. Given the tight pilin packing in known T4P structures, however, we found it quite striking that homologous pili could sustain such a significant deformation to form an interlaced dimer. Although molecular dynamics simulations revealed that T2SS pseudo-pili can adopt a wide range of helical twist angles, they were not shown to have a propensity for short-range bending [16]. Moreover, while many T4P can form bundles, those can be hundreds of nanometers thick, contain a large number of pili and present much more limited distortion at the level of individual filaments [14].
Intrigued by the striking new features of the ‘plaited’ filaments we wanted to see whether the transformation pili we previously reported could form similar structures. We were able to visualize the ‘plaited’ polymers along with the T4P-like long transformation pili in wild-type pneumococci (Fig 1D). However, in a strain with an additional FLAG-tagged ectopic copy of comGC gene for the major pilin [10] (S1 Table), we were able to immunolabel only the long, surface-attached pili using an anti-FLAG antibody (Fig 1E). Similar results were reported by Balaban and colleagues who failed to immunolabel the ‘plaited’ structures with a different antibody raised against ComGC [8].
As negative immunolabeling results are difficult to interpret and can be due to a variety of technical and structural factors, we proceeded to investigate the role of the plaited filaments in competence. We tested three negative control strains carrying either single-gene deletions for essential pilus biogenesis components (S1 Table)—the assembly platform protein ComGB [17] or the associated powering ATP-ase ComGA—or expressing a preprocessing incompetent ComGC variant (ComGCE20V, or ComGCE5V in mature pilin residue numbering). As shown previously, although these mutants can express monomeric pilins, they cannot assemble surface exposed pili and are transformation-incompetent [8,10,17]. In all three ΔcomGA, ΔcomGB and comGCE20V strains we still detected release of ‘plaited’ filaments while expression of the long, T4P-like transformation pilus was abolished (Fig 1F and Fig 1G and S1). Finally, we observed significant release of these structures even in the absence of competence induction (Fig 2A), further confirming that they are not related to pilus biogenesis during natural transformation.
To identify the building subunits of the ‘plaited’ filaments, we developed an enrichment and purification protocol based on differential ultracentrifugation, microfiltration and size exclusion chromatography. Electron microscopy imaging of the purified polymers showed a homogeneous sample composed primarily of the characteristic coiled polymers with an average length of 100–300 nm (Fig 2A). SDS-PAGE analysis of the corresponding fraction showed the presence of a predominant protein species with a molecular weight of ~100 kDa. LC-MS/MS proteomic analyses on the excised, trypsin-digested gel band, as well as on the total purified filaments fraction (S2 Table), unambiguously identified the predominant protein as bifunctional acetaldehyde-alcohol dehydrogenase AdhE, and the result was validated biochemically by Western blot detection using an anti-AdhE antibody (Fig 2B). Heterologous expression of the S. pneumoniae adhE gene in Escherichia coli and affinity purification of the expressed protein showed spontaneous coiled filament formation in the eluted fraction (Fig 2C). Finally, the protein composition of the ‘plaited’ polymers was validated by affinity pull-down with anti-FLAG antibody-conjugated resin on samples from a S. pneumoniae strain carrying an additional ectopic adhE gene copy for competence-inducible expression of a C-terminally FLAG-tagged protein (Fig 2D).
AdhE is a 98 kDa protein with an N-terminal acetylating aldehyde dehydrogenase domain (AldDH) and a C-terminal Fe-dependent alcohol dehydrogenase domain (Fe-ADH) (Fig 2E) [18]. Homologous dual domain proteins are common among fermentative bacteria and are reported to catalyze the NADH-dependent conversion of acetyl-CoA to ethanol via an aldehyde intermediate. Most importantly, in many species AdhE has been shown to polymerize into fine helical filaments called spirosomes [19–26] that are morphologically consistent with the ‘plaited’ filaments discussed here and reported as self-secreting, ‘plaited’ transformation pili by Balaban and colleagues [8]. A high resolution structural model of spirosome assembly by the closely homologous AdhE of Geobacillus thermoglucosidasius shows multimeric arrangement of the individual subunits into a right-handed spiral filament with six protomers per helical turn and overall pitch and width parameters consistent with negatively stained class averages of its pneumococcal counterpart (Fig 2E and 2F) [19]. It is also important to note that the proposed spirosome structure—which is based on crystallographic and in-solution biophysical data, homology modeling and in silico macromolecular docking—corresponds to a single-start helix rather than a ‘plaited’ polymer of two or more interlaced filaments [19].
To examine a putative role of AdhE in natural transformation, we first followed the protein’s expression over the course of competence induction that was verified by the detection of a competence-inducible FLAG-tagged ectopic copy of ComGC. While we have shown competence-specific ComGC expression in wild-type genetic background previously [10], AdhE protein levels remained stable over the course of the experiment (Fig 3A). We next constructed an adhE-null Streptococcus pneumoniae R6 mutant (ΔadhE) and examined its transformation efficiency for uptake of resistance-encoding DNA cassette under challenge with the corresponding antibiotic. While the ΔadhE mutant shows slightly decreased transformation efficiency (~ 2-fold), this change is negligible compared to typical results under comGC disruption (~ 10 000-fold) and can be due to reduced metabolic fitness under the microaerobic conditions of the experiment (Fig 3B). Our data are consistent with a previous genome-wide study aiming to identify genes essential for natural transformation in Streptococcus pneumoniae, which have failed to identify AdhE as a requirement for DNA uptake [27]. Finally, while no spirosome release was detected for the competence-induced ΔadhE pneumococci, typical T4P-like transformation pili were observed (Fig 3C and 3D).
Formation of spirosomes has been reported in a variety of Gram-positive and Gram-negative bacteria, with the first studies dating back several decades and refering to the building protein, AdhE, as spirosin [19–26]. AdhE conservation across representative species with confirmed spirosome formation shows significant sequence homology even among relatively distant taxa (Table 1). Nevertheless, sequence similarity mapping along the AdhEG.thermoglucosidasius structural model reveals that highly conserved residues cluster in only few surface-exposed patches [19,28]. These correspond to the deep active site clefts of the two dehydrogenase domains, as well as sites at or near the interdomain linker. The latter would likely remain buried in the context of mature spirosomes, as they stabilize embrace-like interactions between AdhE monomers in the high-resolution structural model of G. thermoglucosidasius spirosomes [19] (S2 Fig). Thus the exposed spirosome surface would retain significant variability, which in turn could translate into differences in spirosome morphology and stability across species. Moreover, earlier reports have demonstrated that spirosome helix parameters can vary significantly depending on the presence and type of small molecule and metal ion cofactors [20,21].
In addition to S. pneumoniae, we observed spirosome release in cultures of Clostridium difficile, Streptococcus sanguinis, and E. coli (Fig 4A–4C and Table 1) [29,30]. An adhE null strain of S. sanguinis [30] showed no release of morphologically consistent filaments, serving as an additional control for correct target identification. For the two Gram-positive species, C. difficile and S. sanguinis, spirosome morphology was practically indistinguishable from that of S. pneumoniae. The helical filaments we observed in E. coli cultures, on the other hand, were visibly more tightly coiled (Fig 4D) To verify that those correspond to AdhE macromolecular complexes we purified an enriched spirosome fraction and validated its major component as the bifunctional dehydrogenase using proteomic and biochemical methods (Fig 4D and S3 Table). Thus, although we expect morphological variations to be commonplace across species and even sample handling protocols, we are confident that spirosome assembly and extracellular release can be detected in many more environmental, clinically isolated, or genetically engineered bacterial strains.
The major horizontal gene transfer mechanism in S. pneumoniae—natural transformation—requires regulated expression of the comG pilus biogenesis operon, homologous to operons encoding T4P and T2SS pseudopili in Gram-negative bacteria [5,7,8,13]. Recent studies have reported conflicting results regarding the morphology and function of pneumococcal transformation pili. One proposed mechanism is that S. pneumoniae expresses a long, DNA-binding, T4P-like cell surface appendage to ‘fish’ extracellular transforming DNA [10], while an alternative hypothesis argues that competent pneumococci express short, self-secreting T2SS plaited pili that perforate the cell wall peptidoglycan to allow for DNA entry [8].
Here we show that the ‘plaited’ filamentous polymers are not related to natural transformation or pilus biogenesis but are instead widely conserved and well documented macromolecular complexes of the fermentative enzyme bifunctional acetaldehyde-alcohol dehydrogenase AdhE. Its tandem domain architecture secures the two-step NADH-dependent reduction of acetyl-CoA to ethanol via an aldehyde intermediate [19–25,31]. Although the biological significance of AdhE polymerization in such massive structures remains enigmatic, it is plausible that spirosome assembly delivers spatially localized metabolic flux to limit diffusion of the highly reactive aldehyde species and secure optimized conversion kinetics [32]. In addition, the high resolution structural model of G. thermoglucosidasius spirosomes shows that polymer assembly buries ~ 6 500 Å2 of surface area per monomer, which could have dramatic effects on protein stability and function [19,31].
Extracellular spirosome release by cultured bacteria appears to be the result of random cell lysis as no biological function or secretion mechanism can be assigned to the phenomenon. As expected for a fermentative enzyme and consistent with reports in the literature, AdhE expression and spirosome assembly is expected to increase under microaerobic and anaerobic conditions as opposed to aerobic cultures [23,33]. Anaerobic growth and increased cell lysis are both common in cultures of competence-induced pneumococci, where the signaling process of fratricide kills non-competent cells to release extracellular DNA available for uptake [17,34]. This can explain why the spirosomes were initially associated to natural transformation in S. pneumoniae and reported to represent ‘plaited’ transformation pili [8]. Conversely, in their study Balaban and colleagues attempted to reconstitute expression of pneumococcal transformation pili in E. coli by heterologous expression of the entire comG operon [8]. As a result, extracellular release of spirosomes was readily detected and the AdhE polymers were again labeled erroneously. It remains unclear why the authors failed to detect spirosomes in their negative control culture. One possibility is that the structures were omitted in the limited observation field that high-magnification electron microscopy experiments provide. Conversely, it is possible that overexpression of several non-native proteins—among which a macromolecular complex targeted to the inner membrane—could have had destabilizing effects on the expression strain, leading to increased cell lysis and spirosome release [35,36].
In our quest to identify the structures observed by Balaban and colleagues [8], we initially hypothesized that they were randomly released RecA nucleofilaments due to their striking similarity to polymerized RecA homologs from other bacterial and eukaryotic species. Such structures have been extensively documented in the literature: forming characteristic helical coils, RecA filaments can be more or less extended depending on the presence and type of DNA and small-molecule ligands [37–39]. Essential but not exclusive to natural transformation, cytosolic RecA is massively expressed during competence and its polymerization on the incoming single-stranded DNA is key to DNA protection and subsequent integration in the genome (Fig 5) [40]. Nevertheless, while ΔrecA cells display normal DNA uptake during competence, RecA-based recombination plays pivotal role in DNA repair throughout the bacterial life cycle [40,41]. This indicates that pilus-dependent DNA uptake and RecA polymerization are intrinsically uncoupled and could have explained the continuous release of ‘plaited’ filaments in the pilus-defficient strains. We were indeed able to detect RecA release in the medium of competence-induced cultures of both wild-type and ΔcomGB cells (S3 Fig panel A). Also, although with slightly different parameters in terms of helical width and pitch, the ‘plaited’ filaments were structurally similar to both the coiled structure of a eukaryotic RecA homolog (S3 Fig panel B) [37], as well as to in vitro reconstituted RecAS.pneumoniae nucleofilaments (S3 Fig panel C). Nevertheless, release of the characteristic polymers persisted in a ΔrecA strain (S3 Fig panel D) and we were unable to detect the protein in the filament-enriched fractions following purification (S1 Table).
It is therefore important to note that macromolecular organization in helical filaments is not uncommon among proteins from both the bacterial and other taxa. These include but are not limited to nucleic acid-binding proteins, cytoskeletal elements, building blocks of cell surface appendages and phage capsid subunits [39,42–44]. This, together with the markedly different helical parameters of the highly conserved E. coli spirosomes shown here underscores the fact that limited-view, low-resolution morphology imaging and bulk biochemical experiments alone are often insufficient to deduce the nature of macromolecular assemblies. Rather, a combination of orthogonal approaches that spans the different resolution levels and integrates genetic, biochemical and structural data in a meaningful way is generally warranted to avoid false-positive or otherwise erroneous results.
Taken together, our data rule out the existence of short ‘plaited’ transformation pili in competent pneumococci and reassert the expression of a long, 5–6 nanometer wide appendage, structurally and compositionally similar to T4P in Gram-negative bacteria [10]. This finding bridges Gram-negative and Gram-positives DNA uptake systems and provides a comprehensive picture of this major lateral gene transfer event. Indeed, a recent study showed the existence of a T4-pilus on competent V. cholerae, which shares many features with the pneumococcal transformation pilus: competence-induced expression, prerequisite for DNA uptake, and roughly a single copy per cell [45].
Apart from morphology alone, however, it is interesting to discuss the probable mechanism through which the transformation pili secure DNA entry into competent pneumococci. Although expression of any type of pilus would require overcoming the physical barriers of cell-wall peptidoglycan and overlaying capsule—and thus possibly facilitate DNA entry—the similarities among transformation pili of Gram-positive and Gram-negative bacteria suggest that naturally transformable species might have evolved a conserved and more sophisticated mechanism of pilus function than simple cell-wall destabilization.
In agreement with a long-standing ‘pseudo-pilus’ hypothesis, Balaban and colleagues proposed a model in which the transformation pili self-secrete in the medium of competent S. pneumoniae, thus opening gateways in the cell wall peptidoglycan for passive exogenous DNA entry [8,9]. Their hypothesis was supported by the observation that ComGC found in the supernatant of different S. pneumoniae strains after centrifugation correlates with the peak of transformation efficiency [8]. Since we previously showed that transformation pilus expression is absolutely required for DNA uptake, it is not surprising to observe correlation between extracellular ComGC and transformation efficiency [10]. However, ComGC release in culture supernatants can be a result from both cell lysis and/or compromised pilus integrity. As we showed previously, pneumococcal transformation pili are fairly sensitive to mechanical stress and short vortexing and centrifugation are routinely used for their shearing and isolation [10,12]. Such mechanical forces, however, are unlikely to be exerted in nature, where competent pneumococci are typically cushioned in protective biofilm matrix [46].
As we have conducted only single time point visualization experiments, it is theoretically possible that the expresses transformation pili eventually detach from the cell to open entry pores for transforming DNA (Fig 5). However, the sheer size and ATP-dependent assembly of the transformation pilus makes such self-secretion hypothesis unlikely: the observed long native pili would be energetically taxing on the cells if their sole function were to be ejected prior to DNA uptake. Finally, it has been previously reported that native transformation pili bind and co-purify with DNA already present in the cell culture and that DNA binding at the surface of competent pneumococci is abolished in a pilus-deficient strain [10,47]. DNA-binding is also conserved in the homologous T4P of Gram-negative bacteria [14,48,49]. In such a DNA-binding context, pilus release would actually inhibit transformation by titrating out DNA available for uptake (Fig 5). This once again argues against a self-secreting mechanism of function and reinforces a cell-surface attached role for the pilus in transformation.
While no mechanistic or quantitative data on DNA binding by the pilus are available, electron microscopy showed extensive contact interfaces between the long transformation pili and DNA chains [10]. It is therefore plausible that multiple weak interactions along the helical pilus lattice stabilize this interaction and allow its reversal upon DNA uptake. Such a scenario would also explain why no DNA binding to a non-polymerizing ComGC truncation has ever been detected [8,50]. Even more interesting, however, is the question of how pilus-bound DNA gets brought to the DNA uptake machinery in the cell membrane. In Gram-negative bacteria, T4P-bound DNA is proposed to be actively hauled to the cell by rapid bottom-up pilus depolymerization powered by a dedicated retraction ATPase [14]. Although a similar mechanism has been proposed for S. pneumoniae and other transformable Gram-positive bacteria, pneumococci lack homologous retraction ATPase and are likely to utilize a distinct mechanism for DNA entry. In addition, transforming DNA uptake occurs at much lower speeds in Gram-positive bacteria than Gram-negative T4P retraction [51,52].
Many sequence-specific DNA binding proteins can scan DNA for their target sites at speeds several orders of magnitude higher than the upper limit for a three-dimensional diffusion-controlled process [53]. This can generally be achieved by at least two passive mechanisms, which involve sequence non-specific DNA binding and subsequent translocation of the protein along the DNA: 1) charge-based protein sliding, where the protein engages in a one-dimensional random walk along the DNA in search of its target, and 2) direct intersegment transfer, where the protein can bind and hop between two remote regions on the DNA without losing the non-specifically bound state [53]. Although we can not exclude the involvement of an unidentified retraction ATPase or additional receptor proteins in exogenous DNA uptake, we favor a model where the pneumococcal transformation pilus provides a similar facilitated diffusion framework (Fig 5). By preserving multiple dynamic non-specific interactions with the pilus, transforming DNA would overcome the thermodynamic limitations of a three-dimensional diffusion process until it passively finds the membrane associated uptake machinery and becomes actively pumped in the cell (Fig 5).
Streptococcus pneumoniae spirosomes were observed in both competent and non-competent cells. For competence induction cells were grown in microaerobic conditions, without agitation, at 37°C in Casamino Acid-Tryptone (CAT) medium supplemented with 0.2% glucose, 15mM dipotassium phosphate, 3mM sodium hydroxide and 1mM calcium chloride and adjusted to pH 7.8. Competence was triggered by the addition of Competence Stimulating Peptide (CSP) at OD600 = 0.15 for 10–30 min. Non-competent pneumococci and Escherichia coli cells were grown similarly in LB to OD600 = 0.3 and OD600 = 0.6, respectively. Clostridium difficile cells were grown at 37°C under strict anaerobic conditions on Tryptone-Yeast extract-Glucose (TYG) plates supplemented with 0.1% thioglycolate. Streptococcus sanguinis cells were grown anaerobically, without agitation, at 37°C in CAT medium to OD600 = 0.3. For spirosome visualization cells were scraped off the plates or pelleted by centrifugation and resuspended in TBS (50 mM Tris-HCl pH 7.6, 150 mM NaCl) at ~ 5 μl TBS per milliliter of culture at OD600 = 0.3. 5 μl drops of each suspension were then placed directly on glow discharged carbon coated grids (EMS, USA) for 1 minute. The grids were then blot-dried on filter paper, washed on a drop of ultrapure water, and negatively stained with 2% uranyl acetate in water. Specimens were examined on an FEI Tecnai T12 BioTWIN LaB6 electron microscope operating at 120 kV at nominal magnifications of 18500–68000 and 1–3 μm defocus. Images were recorded on a Gatan Ultrascan 4000 CCD camera.
An adhE deletion (strain AD001) was introduced in the R1501 genetic background by transformation with a DNA cassette carrying a kanamycin resistance gene inserted between two ~1000 base pair fragments corresponding to the S. pneumoniae genomic regions flanking adhE. Briefly, the genomic region upstream from the AdhE open reading frame was amplified using forward and reverse primers 5’-ACA TGG CAA TCC GAT TCA TAA GGG G-3’ and 5’-GCC ATC TAT GTG TCG GAA CGA TAT CCT TTG TTA ATT TTT TCA CAA GTT TAT TAT AAC G-3’, respectively, while the genomic region downstream of the adhE gene was amplified the following primer pair 5’-AAA ATG TGT TTT TCT TTG TTT TGT TTA TCA GTC TAG AAG CAA GAC AAA AAC TCA A-3’ and 5’-TTG CTA TTT ATG CAT GCA GAA GAC CAA ATG-3’. A third pcr reaction was used to amplify a kanamycin-resistance gene using the pR411 plasmid as template DNA [54] and forward and reverse primers 5’-AGG ATA TCG TTC CGA CAC ATA GAT GGC GTC GCT AGT-3’ and 5’-GCT TCT AGA CTG ATA AAC AAA ACA AAG AAA AAC ACA TTT TTT TGT CAA AAT TCG TTT-3’, carrying complementarity to the 3’-end of the adhE-upstream and 5’-end of the adhE-downstream fragments, respectively. The three pcr products were then assembled using overlap extension PCR and the purified DNA cassette was used for transformation of competence-induced S. pneumoniae R1501 cells. adhE-null mutants (strain AD001) were positively selected by growth in the presence of kanamycin (60 μg/ml) and adhE deletion was confirmed independently by DNA sequencing and western blot detection using an anti-AdhE antibody. For all transformation experiments, competence was triggered as above at OD600 = 0.15 for 10 minutes, followed by DNA addition and 20 minute incubation at 30°C. Transformants were selected on Columbia Agar supplemented with 5% horse blood and appropriate antibiotics. For the transformation efficiency assays, cells were transformed with 100 ng of a DNA cassette, amplified from S. pneumoniae R304 genomic DNA and containing the streptomycin resistance gene str41. Bacteria were plated in the presence and absence of streptomycin (100 μg/ml) and incubated at 37°C overnight before colony counting.
All steps of the purification protocol were performed at 4°C. 8L of S. pneumoniae culture grown in LB to OD600 = 0.3 or 4L of E. coli culture grown to OD600 = 0.6 were pelleted by centrifugation (20 min at 5000 g) and resuspended in 6 ml of cold TBS, vortexed briefly and centrifuged to remove the bulk of intact cells and debris (10 min at 12000 g followed by 15 min at 50000 g). Triton x-100 was added to the supernatant at final concentration of 0.25%. Following 30 minute agitation for solubilization of remaining membrane fragments, the samples were filtered through a 0.45 μm cellulose acetate filter (Corning) and centrifuged for 1h at 125000 g for spirosome pelleting. After careful removal of the supernatant, the pellet was resuspended in 50 μl TBS, re-filtered and loaded on a Superose 6 3.2/300 size exclusion column (GE Healthcare). Spirosome enriched fractions were found to elute with the void volume. Sample preparation for electron microscopy was performed as above.
Trypsin digestion was performed as described previously [55] and the digests were analyzed under standard conditions on an LTQ-Orbitrap Velos (Thermo Fisher, Bremen) equipped with Ultimate 3000 nano-HPLC (Dionex). Briefly, tryptic peptides were desalted and separated on a C-18 nano-HPLC column under a gradient of acetonitrile. Minimum signal threshold for triggering an MS/MS event was set to 5000 counts. After a survey scan, the 10 most intense precursor ions were selected for CID fragmentation (top10). Raw files were processed with MaxQuant 1.4.1.2 [56]. Protein identification was done using Andromeda against a Streptococcus pneumoniae (strain ATCC BAA-255 / R6) (Taxonomy 171101–1947 proteins) or Escherichia coli (strain K12 / MC4100 / BW2952) (Taxonomy 595496–4043 proteins) database. A false-discovery rate of 1% was used for both peptide and protein identification. Reverse and contaminant proteins were excluded and only proteins identified with a minimum of 2 peptides were considered.
Spirosomes and transformation pili were visualized by EM as above at nominal magnifications of 49000 and 68000, respectively. The contrast transfer function parameters were assessed using CTFFIND3 [57], and the phase flipping was done in SPIDER [58]. Linear filament segments were boxed with e2helixboxer and single particle stacks were generated for each dataset (EMAN2 [59]): 4309 particles for the transformation pilus (134x134 pixel box, 1.5 Å/pixel), 728 particles for the S. pneumoniae spirosomes (180x180 pixel box, 2.2 Å/pixel) and 555 particles for the E. coli spirosomes (180x180 pixel box, 2.2 Å/pixel). Normalization, centering, multi-reference alignment, multi-statistical analysis, and classification (15–40 particles per class) were done in IMAGIC-4D (Image Science Software, GmbH). IMAGIC-4D was also used for generation of two-dimensional reprojections of the T2SS pilus and the G. thermoglucosidasius spirosome.
The coding sequence for AdhES.pneumoniae was cloned in-frame in a pET21a vector for heterologous expression of C-terminally hexahistidine-tagged protein in E. coli. Briefly, the AdhES.pneumoniae open reading frame was PCR-amplified from S. pneumoniae R1501 genomic DNA using forward and reverse primers 5’-CAT ATG AAA GCT ATG GAG GAA AAT ATG GCT G-3’ and 5’-GCG GCC GCT TTA CGG CGT CCT GGT CTT TCT TTG-3’, respectively. The pET21a vector was pcr-amplified using forward and reverse primers 5’-GGA CGC CGT AAA GCG GCC GCA CTC GAG CAC CAC CAC-3’ and 5’-CAT ATT TTC CTC CAT AGC TTT CAT ATG TAT ATC TCC TTC TTA-3’, respectively. 90 ng of linearized vector DNA were incubated in 1:1 molar ratio with the AdhE pcr product in an In-Fusion cloning reaction (Clontech) and transformed into Top10 cells (Invitrogen). Plasmid DNA was purified from individual clones, verified for AdhE coding region insertion and used for transformation in BL21 (DE3) cells (Invitrogen). For expression, transformed BL21 (DE3) cells were grown under agitation at 37°C in LB to OD600 = 0.6 and expression was induced with 0.7 mM IPTG for 2h. Cells were pelleted by centrifugation (4500 g for 20 min), resuspended in TBS supplemented with cOmplete Protease Inhibitor Cocktail (Roche) and disrupted by sonication. Cellular debris were removed by centrifugation (20 000 g for 30 min) and the lysates were incubated with batch Talon metal affinity resin (Clontech) for 30 min at room temperature. The resin was extensively washed (TBS, 15 mM imidazole) and bound protein was eluted with TBS supplemented with 140 mM Imidazole. For EM observation eluted protein was immediately applied on glow-discharged continuous carbon grids, stained, and imaged as above.
For AdhE spirosome pull-down, a strain carrying an additional competence-inducible FLAG-tagged ectopic copy of the adhE gene was constructed. Briefly, the adhE gene was PCR-amplified using S. pneumoniae R1501 genomic DNA as template and primer pair 5’-GAG GAA GAA ACC ATG TTG AAA GCT ATG GAG GAA AAT ATG GCT GAT AAA AAA AC-3’ and 5’-AAA ATC AAA CGG ATC TTA CTT GTC ATC GTC ATC CTT GTA ATC TTT ACG GCG TCC TGG TCT TTC TTT G-3’, the latter designed to add a C-terminal FLAG tag to the encoded protein. In parallel, pCEPx vector DNA [60] was digested with NcoI and BamHI restriction enzymes. 90ng of the linearized vector were incubated in 1:1 molar ratio with the adhE-FLAG pcr product in an In-Fusion cloning reaction (Clontech) and transformed into Top10 E. coli cells (Invitrogen). Plasmid DNA was purified from individual clones, verified for AdhE-FLAG coding region insertion, and transformed into competence-induced R1501 cells, followed by selection with kanamycin (Kan). The resulting SO007 strain was sequence-verified for the pCEPx-derived adhE-FLAG–KanR cassette recombination [60] and cultured in CAT medium to OD600 = 0.15. CSP-induced and non-induced cells were pelleted by centrifugation and resuspended in TBS by brief vortexing in the presence of millimeter-sized glass beads for increased cell lysis. Cells were pelleted again, and the supernatants were incubated with anti-FLAG M2 affinity resin (Sigma-Aldrich A2220) for 1h at RT and under agitation. After washing with TBS, the resin-bound fraction was eluted by the addition of 100 μg/mL 3X FLAG peptide (Sigma Aldrich F4799) and mixing for additional 15 min at 4°C. The samples were then subjected to SDS-PAGE and EM analyses.
2.5 μM purified RecAS. pneumoniae was incubated with 50 mM KCl, 0.5 mM DTT, 10 mM Tris-HCl pH 7.5, 2 mM magnesium acetate, and 2 mM ATP-γ-S for 15 minutes at 37°C. RecA nucleofilament formation was induced by the addition of a 54-nucleotide-long single-stranded DNA fragment or single-stranded M13mp18 DNA (New England Biolabs) at nucleotide:RecA monomer ratio of 3:1. After additional 15 minutes at 37°C, the samples were prepared for EM observation as above.
|
10.1371/journal.pgen.1003437 | Extensive Natural Epigenetic Variation at a De Novo Originated Gene | Epigenetic variation, such as heritable changes of DNA methylation, can affect gene expression and thus phenotypes, but examples of natural epimutations are few and little is known about their stability and frequency in nature. Here, we report that the gene Qua-Quine Starch (QQS) of Arabidopsis thaliana, which is involved in starch metabolism and that originated de novo recently, is subject to frequent epigenetic variation in nature. Specifically, we show that expression of this gene varies considerably among natural accessions as well as within populations directly sampled from the wild, and we demonstrate that this variation correlates negatively with the DNA methylation level of repeated sequences located within the 5′end of the gene. Furthermore, we provide extensive evidence that DNA methylation and expression variants can be inherited for several generations and are not linked to DNA sequence changes. Taken together, these observations provide a first indication that de novo originated genes might be particularly prone to epigenetic variation in their initial stages of formation.
| Epigenetics is defined as the study of heritable changes in gene expression that are not linked to changes in the DNA sequence. In plants, these heritable variations are often associated with differences in DNA methylation. So far, very little is known about the extent and stability of epigenetic variation in nature. In this study, we report a case of extensive epigenetic variation in natural populations of the flowering plant Arabidopsis thaliana, which concerns a gene involved in starch metabolism, named Qua-Quine Starch (QQS). We show that in the wild QQS expression varies extensively and concomitantly with DNA methylation of the gene promoter. We also demonstrate that these variations are independent of DNA sequence changes and are stably inherited for several generations. In view of the recent evolutionary origin of QQS, we speculate that genes that emerge from scratch could be particularly prone to epigenetic variation. This would in turn endow epigenetic variation with a unique adaptive role in enabling de novo originated genes to adjust their expression pattern.
| DNA mutations are the main known source of heritable phenotypic variation, but epimutations, such as heritable changes of gene expression associated with gain or loss of DNA methylation, are also a source of phenotypic variability. Indeed, several stable DNA methylation variants affecting a wide range of characters have been described, mainly in plants [1]–[3]. In most instances, epimutations are linked to the presence of structural features near or within genes, such as direct [4]–[6] or inverted repeats [7], [8] or transposable element (TE) insertions [9], which act as units of DNA methylation through the production of small interfering RNAs (siRNAs) [3], [10]. Examples of epimutable loci in Arabidopsis thaliana (A. thaliana) include the PAI [7] and ATFOLT1 genes [8], which have suffered siRNA-producing duplication events in some accessions and also the well characterized FWA locus, which contains a set of SINE-derived siRNA-producing tandem repeats at its 5′end [4], [5]. Repeat-associated epimutable loci are almost invariably found in the methylated form [5]–[9] in nature, which reflects, at least in part, that DNA methylation is particularly well-maintained over repeats [11], [12]. Indeed, epigenetic variation at PAI, ATFOLT1 and FWA has only been observed in experimental settings. Similarly, sporadic gain or loss of DNA methylation associated with changes in gene expression has only been documented in A. thaliana mutation accumulation lines [13], [14] and examples of natural epigenetic variation in other plant species are few [15]–[17].
Here we report a case of prevalent natural epigenetic variation in A. thaliana, which concerns a de novo originated gene [18]. We show that expression of this gene, named Qua-Quine Starch (QQS), is inversely correlated with the DNA methylation level of its promoter and that these variations are stably inherited for several generations, independently of DNA sequence changes. Based on these findings, we speculate that epigenetic variation could be particularly beneficial for newly formed genes, as it would enable them to explore more effectively the expression landscape than through rare DNA sequence changes.
The A. thaliana Qua-Quine Starch (QQS, At3g30720) was first described as a gene involved in starch metabolism in leaves [19], [20]. Despite being functional and presumably already under purifying selection (dN/dS = 0.5868; p-value<0.045), QQS is likely a recent gene that emerged de novo. Indeed, QQS has no significant similarity to any other sequence present in GenBank [18], [19], suggesting that it originated from scratch since A.thaliana diverged from its closest sequenced relative A. lyrata around 5–10 million years ago. Furthermore, QQS encodes a short protein (59 amino acids) and it is differentially expressed under various abiotic stresses [18], which are also hallmarks of de novo originated genes [21]–[23].
As shown in Figure 1, QQS is surrounded by multiple transposable element sequences (Figure 1A) and contains several tandem repeats in its promoter region and 5′UTR (Figure 1B). In the Columbia (Col-0) accession, these tandem repeats are densely methylated and produce predominantly 24 nt-long siRNAs (Figure 1B, Figure S1A and S1B). Publically available transcriptome data [24], [25] and results of RT-qPCR analyses (Figure S1C) show that steady state levels of QQS mRNAs are higher in several mutants affected in the DNA methylation of repeat sequences, including met1 (DNA METHYLTRANSFERASE 1), ddc (DOMAINS REARRANGED METHYLTRANSFERASE 1 and 2 and CHROMOMETHYLASE 3), ddm1 (DECREASE IN DNA METHYLATION 1) and rdr2 (RNA-DEPENDENT RNA POLYMERASE 2), which abolishes the production of 24 nt-long siRNAs as well as most CHH methylation. These findings indicate that QQS expression is negatively controlled by DNA methylation and point to the siRNA-producing tandem repeats as being potentially involved in this repression.
We first observed epiallelic variation at QQS unexpectedly, in a Col-0 laboratory stock (hereafter referred to as Col-0*) with increased expression of the gene and decreased DNA methylation of its promoter and 5′UTR repeat elements (Figure 2A). No sequence change could be detected in the Col-0* stock within a 1.2 kb region covering the QQS gene (Figure 1B), which excluded local cis-regulatory DNA mutations at the QQS locus as being responsible for DNA methylation loss in Col-0*. Additionally, comparative genomic hybridization analysis as well as genome-wide DNA methylation profiling using methylated DNA imunoprecipitation assays revealed no major differences between Col-0 and Col-0* (Figure S2).
We next investigated the QQS epigenetic status in pooled seedlings (S1) derived from the selfing of 12 individual Col-0* plants (Figure S3). Results revealed a range of QQS epialleles and a strong negative correlation between DNA methylation and expression of the gene (Figure 2B and 2C). To explore further this variation, a single S1 individual was then selfed for each of the 12 lines and seedlings (S2) were analyzed in pool for each line, as above (Figure S3). Remarkably, the differences in QQS expression and DNA methylation observed at the S1 generation were also observed at the S2 generation (Figure 2B and 2C). Taken together, these results suggest therefore the existence of a range of epiallelic variants at QQS, which are stably inherited for at least one generation.
The inheritance of QQS hypomethylated epialleles was also examined in a random sample of 19 ddm1-derived epigenetic Recombinant Inbred Lines (epiRILs) obtained by crossing a Col-0 wild-type (wt) line with an hypomethylated Col-0 ddm1 line [26]. High DNA methylation/low expression and low DNA methylation/high expression of QQS were observed in 14 and 5 epiRILs, respectively (Figure 2D). This is consistent with Mendelian segregation of the highly methylated/lowly expressed Col-0 wt and lowly methylated/highly expressed Col-0 ddm1 parental QQS epialleles (75%/25% expected because of backcrossing rather than selfing of the F1; Chi2 = 0,017, p-value>0.05). Indeed, examination of the epi-haplotype obtained for 17 of these epiRILs [27] confirmed the wt or ddm-origin of the QQS locus in each case (data not shown). These results demonstrate therefore that, like many other ddm1-induced epialleles [28], [29], QQS hypomethylated epialleles can be stably inherited for at least eight generations and are not targets of paramutation.
We next investigated the degree to which DNA methylation of QQS and of flanking TEs are independent from each other. To this end, we first analyzed DNA methylation patterns of TE sequences flanking QQS in a series of epiRIL with contrasted QQS epialleles. Unlike for ddm1-derived QQS, hypomethylation was not inherited for the three TEs located immediately upstream of the gene, as they did systematically regain wt DNA methylation levels (Figure 3A and 3B), presumably because of their efficient targeting by RNA-directed DNA methylation (RdDM) [28]. In addition, although the TE just downstream of QQS was always hypomethylated when inherited from ddm1, hypomethylation was also observed in one epiRIL that inherited the QQS region from the wt parent. Thus, there is no strict correlation between DNA methylation at QQS and this downstream TE. We next examined the effect of several T-DNA and transposon insertions located ∼3.1 kb or 153 bp upstream of the transcription start site (TSS), 653 bp downstream of the 3′UTR and within the second coding exon of QQS. Whereas three of these insertions had no effect on DNA methylation and expression levels of QQS, the T-DNA insertion located closest to the TSS was associated with a drastic reduction of DNA methylation of both the promoter and 5′UTR of the gene, as well as with an increase in QQS expression (Figure 3A and 3C). However, this insertion had no impact on DNA methylation of upstream and downstream TEs (Figure 3A and 3D). Taken together, these results suggest that epigenetic variation at QQS is most likely determined by sequences within the promoter and 5′UTR of the gene, not by the TEs that are located immediately upstream or downstream.
We next investigated the possibility that QQS is subject to epigenetic variation in natural populations. To this end, we first analyzed QQS expression and DNA methylation in 36 accessions representing the worldwide diversity [30]. QQS was methylated and lowly expressed in 29 accessions, but unmethylated and highly expressed in seven accessions distributed over the entire geographic range (Figure 4A). This indicates that epigenetic variation at QQS is widespread in nature. In contrast, upstream and downstream TEs were consistently methylated in all accessions (Figure S4A and S4B), thus confirming that the epigenetic state at QQS is not determined by that of flanking TEs. We then sequenced a 2.8 kb interval encompassing the QQS gene and its flanking regions from the seven accessions carrying the hypomethylated/highly expressed epiallele as well as from three accessions carrying a methylated/lowly expressed epiallele. Although several SNPs and indels were identified (Figure S4C), no correlation between any specific sequence alterations and QQS DNA methylation or expression states could be established (Figure 4A). In addition, while Kondara and Shahdara have identical QQS sequences, they have contrasted DNA methylation/expression patterns at the locus (Figure 4A and Figure S4C), which provides further evidence that natural epiallelic variation at QQS is independent of local cis-DNA sequence polymorphisms and is thus most likely truly epigenetic. Analysis of a Cvi-0 vs. Col-0 Recombinant Inbred Line (RIL) population revealed in addition that QQS expression is controlled by a large-effect local-expression quantitative trait locus (local-eQTL; http://qtlstore.versailles.inra.fr/) [31]. This suggests that like the Col-0 wt and Col-0 ddm1 QQS epialleles (Figure 2D), the Cvi-0 hypomethylated QQS epiallele is stably inherited across multiple generations. This further demonstrates that epigenetic variation at QQS is not appreciably affected by sequence or DNA methylation polymorphisms located elsewhere in the genome and indicates also that QQS is not subjected to paramutation [29].
To validate experimentally the causal relationship between DNA methylation and repression at QQS, seedlings of several accessions were grown in the presence of the DNA methylation inhibitor 5-aza-2′-deoxycytidine (5-aza-dC). In the two accessions Col-0 and Shahdara, which harbor distinct methylated and lowly expressed QQS alleles, treatment resulted in reduced DNA methylation and increased expression of QQS (Figure S4D). In contrast, seedlings of Jea, Kondara and Cvi-0 accessions, all of which harbor a demethylated/highly expressed QQS allele, did not show further reduction of DNA methylation or increased expression when grown in the presence of the demethylating agent (Figure S4D). Moreover, whereas expression of QQS in F1 hybrids derived from crosses between Col-0 (methylated QQS) and Kondara (hypomethylated QQS), was always higher for the epiallele inherited from the hypomethylated parent, further confirming that QQS is not subjected to paramutation [29], treatment with 5-aza-dC reduced dramatically this expression imbalance, most likely as a consequence of demethylation of the Col-0-derived QQS allele (Figure S4E). Taken together, these results clearly demonstrate that DNA methylation at QQS is causal in repressing expression of the gene.
Finally, we asked whether epigenetic variation at QQS could be observed in natural settings or if such variation only emerged in the laboratory, where accessions are grown under controlled growth conditions. To this end, we analyzed QQS expression and DNA methylation in plants grown from seeds directly collected from wild populations in Tajikistan, Kyrgyzstan and Iran (NeoShahdara, Zalisky and Anzali populations, respectively). Widespread QQS epiallelic variation was observed, both between and within these diverse wild populations (Figure 4B). In addition, QQS epigenetic variation was examined in the offspring (after two single seed descent generations) of 25 NeoShahdara individuals. These individuals were randomly sampled among a single patch of several thousands of plants that presumably represent the direct descendants of the Shahdara accession. Based on 10 microsatellite markers and one InDel marker, two genetically distinct subpopulations could be identified. While QQS was highly methylated/lowly expressed in all 16 individuals of subpopulation #1, clear differences in DNA methylation and expression were detected among the 9 individuals of subpopulation #2 (Figure 4C). Whether epiallelic variation at QQS in subpopulation #2 reflects inherent fluctuations or an intermediary stage in the route to fixation of one of the two epiallelic forms remains to be determined.
QQS is a protein-coding gene that likely originated de novo in A. thaliana within a TE-rich region (Figure 1A). We have shown that this gene, which contains short repeat elements matching siRNAs (Figure 1B, Figure S1A and S1B), varies considerably in its DNA methylation and expression in the wild (Figure 4). We also show that these variations are heritable and independent of the DNA methylation status of neighboring TEs or of DNA sequence variation, either in cis or trans (Figure 2 and Figure 3, Figures S2 and S4). Thus, we can conclude that QQS is a hotspot of epigenetic variation in nature. Consistent with this, QQS is among the few genes for which spontaneous DNA methylation variation was observed in Col-0 mutation accumulation lines [13].
Cytosine methylation at QQS concerns CG, CHG and CHH sites, which is the pattern expected for sequences with matching siRNAs (Figure 1B, Figure S1B). All three types of methylation sites likely contribute to silencing of QQS, as judged by the reactivation of QQS in the met1, ddm1, ddc and rdr2 mutant backgrounds (Figure S1C; [24], [25]). Yet, among the different DNA methyltransferases targeting DNA methylation at QQS, MET1 may play a more prominent role, given that DNA methylation at this locus is only fully erased in met1 mutant plants [25]. QQS demethylated epiallelic variants may thus preferentially arise through spontaneous [13] or stress-induced [10] defects in DNA methylation maintenance and be stably inherited for multiple generations as a result of the concomitant loss of matching siRNAs, which would prevent efficient remethylation and silencing of the gene [28], [29]. Indeed, although we could not detect QQS siRNAs by Northern blot analysis, presumably because of their low abundance, deep sequencing data indicate that they accumulate less in met1 mutant plants than in wild type Col-0 [25].
Few genes have been shown so far to be subject to heritable epigenetic variation in A. thaliana [5]–[8], [13], [14], [32] and QQS is unique among these in exhibiting this type of variation in nature (Figure 4). This therefore raises the question as to what distinguishes QQS from other genes, such as FWA, for which epigenetic variation can be readily induced in the laboratory in advanced generations of ddm1 and met1 mutant plants [5], [33], but for which this type of variation is not observed among accessions [11], [34]. One possibility is that unlike QQS epivariants, fwa-hypomethylated epialleles are strongly counter-selected because of their potentially maladapted phenotype, namely late flowering [5]. Consistent with this explanation, epiallelic variation with no phenotypic consequences has been documented at FWA in other Arabidopsis species. In these cases, however, inheritance across multiple generations has not been rigorously tested [35]. Another possibility is that de novo originated genes, such as QQS, are particularly prone to heritable epigenetic variation. This is a reasonable assumption considering that these genes tend to lack proper regulatory sequences initially, unlike new gene duplicates, which by definition come fully equipped [21]. In turn, given that epigenetic variation enables genes to adjust their expression in a heritable manner much more rapidly than through mutation while preserving the possibility for rapid reversion, it could prove particularly beneficial in the case of genes that are created from scratch. Once the most adaptive expression state is reached, it could then become irreversibly stabilized (i.e. genetically assimilated) through DNA sequence changes. Although speculative, this proposed scenario could be highly significant given the recent discovery that de novo gene birth may be more prevalent than gene duplication [23].
A. thaliana accessions were obtained from the INRA Versailles collection (dbsgap.versailles.inra.fr/vnat/, www.inra.fr/vast/collections.htm) [30], [36], [37]. Insertion lines were obtained from the GABI-Kat at University of Bielefeld, Germany (GABI-Kat 755C03 and 522C07) [38], the ABRC at Ohio State University (SALK 003195C) and University of Wisconsin, Madison, US (WiscDsLoxHs077_09) [39]. Seeds of ddm1-2 [40], rdr2-1 [41] and ddm1-derived epiRIL lines [26] were provided by V.Colot. NeoShahdara individuals were genotyped with 10 microsatellite markers (NGA8, MSAT2.26, MSAT2.4, NGA172, MSAT3.19, ICE3, MSAT3.1, MSAT3.21, MSAT4.18, ICE5; http://www.inra.fr/vast/msat.php) and one InDel marker in MUM2 gene (MUM2_Del-LP TGGTCGTTATTGGGTCTCGT, MUM2 Del-RP TTAAGAACGCCCGAGGAATA). For expression and DNA methylation assays, seedlings were grown in vitro (MS/2 media supplemented with 0,7% sucrose) for eight days in a culture room (22°C, 16 hours light/8 hours dark cycle, 150 µmol s−1 m−2). Treatment with 5-aza-2′-deoxycytidine was performed as described in [8].
Total RNA was isolated as described in [42] and cDNA was synthetized using oligo(dT) primers and IMPROM II reverse transcriptase (Promega). Real time PCR reactions were run on an Applied Biosystems 7500 Real-Time PCR System using Platinum SYBR green (Invitrogen). QQS expression levels relative to Actin2/PP2A or PP2A/GAPDH internal references were calculated using the formula (2- (Ct QQS – mean Ct internal references))*100. Primers are listed in Table S1.
Total DNA was isolated using Qiagen Plant DNeasy kit following the manufacturer's recommendations. Digestion was carried out overnight at 37°C with 200 ng of genomic DNA and 2 to 8 units of McrBC enzyme (New England Biolabs). Quantitative PCR was performed as described above on equal amounts (2 ng) of digested and undigested DNA samples using the primers described in Table S1. Results were expressed as percentage of molecules lost through McrBC digestion (1-(2-(Ct digested sample - Ct undigested sample)))*100. As a control, the percentage of DNA methylation for At5g13440, which is unmethylated in wt, was estimated in all analyses.
To assess the relative contribution of each allele to the population of mRNA in F1 individuals from reciprocal crosses between Col and Kondara, a single pyrosequencing reaction using the primers QQS_pyro_F1 (PCR) - TCAAAATGAGGGTCATATC ATGG, QQS_pyro_R1-biotin (PCR) - ATTGGATACAATGGCCCTATAACT and QQS_pyro_S1 (Pyrosequencing) - GATATTGGGCCTTATCAC was set up on a SNP polymorphic between the QQS parental coding sequences (Figure S4C; position +285). Pyrosequencing was performed on F1 cDNA, as well as on 1/1 pools of parents cDNA to establish the allelic contribution to QQS expression. F1 genomic DNA is used as pyrosequencing control to normalize against possible pyrosequencing biases. Anything significantly driving allele-specific expression in hybrids is by definition acting in cis, since F1 nuclei contain a mix of all trans-acting factors [43], [44].
CGH experiments were performed for Col-0* vs. Col-0 using Arabidopsis whole-genome NimbleGen tiling arrays [45]. The normalmixEM function of the mixtools package on R was used to found the normal distribution for the distribution of the Col-0*/Col-0 ratio with an expected number of gaussians of two. A Hidden Markov model [46] was used to find regions with copy number variation.
DNA was extracted using DNeasy Qiagen kit and MeDIP-chip was performed on 1.8 µg of DNA as previously described in [47]. The methylated tiles were identified using the ChIPmix method [48]. Probes methylated in one line only (Col-0 or Col-0*) were used to create domains. Domains contain at least three consecutive or nearly consecutive (400 nt min, with one gap of 200 nt max) tiles with identical methylation patterns.
Available QQS coding-sequences (464 different accessions) were downloaded from the “Salk Arabidopsis 1001 Genomes” database (http://signal.salk.edu/atg1001/index.php). A. suecica QQS sequence (coming from the A. thaliana genome of this allotetraploid [49]) was also included in the analysis. The aligned sequences were used to calculate the probability of rejecting the null hypothesis (H0) of strict-neutrality (dN = dS; where dN = number of nonsynonymous and dS = number of synonymous substitutions per site) in favor of the alternative hypothesis of purifying selection (HA; dS>dN). The analysis was done using the MEGA5 software under the Nei-Gojobori method [50] with the variance of the difference calculated by the bootstrap method with 100 replicates. Our overall analysis of 465 sequences rejected H0 in favor of HA (dN/dS = 0.5868; p-value<0.045).
|
10.1371/journal.ppat.1000066 | The Uptake of Apoptotic Cells Drives Coxiella burnetii Replication and Macrophage Polarization: A Model for Q Fever Endocarditis | Patients with valvulopathy have the highest risk to develop infective endocarditis (IE), although the relationship between valvulopathy and IE is not clearly understood. Q fever endocarditis, an IE due to Coxiella burnetii, is accompanied by immune impairment. Patients with valvulopathy exhibited increased levels of circulating apoptotic leukocytes, as determined by the measurement of active caspases and nucleosome determination. The binding of apoptotic cells to monocytes and macrophages, the hosts of C. burnetii, may be responsible for the immune impairment observed in Q fever endocarditis. Apoptotic lymphocytes (AL) increased C. burnetii replication in monocytes and monocyte-derived macrophages in a cell-contact dependent manner, as determined by quantitative PCR and immunofluorescence. AL binding induced a M2 program in monocytes and macrophages stimulated with C. burnetii as determined by a cDNA chip containing 440 arrayed sequences and functional tests, but this program was in part different in monocytes and macrophages. While monocytes that had bound AL released high levels of IL-10 and IL-6, low levels of TNF and increased CD14 expression, macrophages that had bound AL released high levels of TGF-β1 and expressed mannose receptor. The neutralization of IL-10 and TGF-β1 prevented the replication of C. burnetii due to the binding of AL, suggesting that they were critically involved in bacterial replication. In contrast, the binding of necrotic cells to monocytes and macrophages led to C. burnetii killing and typical M1 polarization. Finally, interferon-γ corrected the immune deactivation induced by apoptotic cells: it prevented the replication of C. burnetii and re-directed monocytes and macrophages toward a M1 program, which was deleterious for C. burnetii. We suggest that leukocyte apoptosis associated with valvulopathy may be critical for the pathogenesis of Q fever endocarditis by deactivating immune cells and creating a favorable environment for bacterial persistence.
| Infective endocarditis (IE) is a problem of public health that still causes high mortality despite antibiotic treatments. Most of the patients who develop an IE have pre-existing cardiac lesions, although the relationship between IE and valvulopathy is not clearly understood. We showed here that patients with valvulopathy exhibited increased levels of circulating apoptotic leukocytes. As the binding of apoptotic cells to monocytes and macrophages is known to inhibit their inflammatory activity, we hypothesized that the high levels of circulating apoptotic leukocytes may be responsible for the immune impairment observed in Q fever endocarditis, an IE due to Coxiella burnetii, a bacterium that survives in monocytes and macrophages. The binding of apoptotic lymphocytes to monocytes and macrophages increased the replication of C. burnetii by stimulating their anti-inflammatory response. In contrast, the binding of necrotic lymphocytes to monocytes and macrophages induced C. burnetii killing and stimulated an inflammatory response. Interferon-γ, which is associated with the control of C. burnetii infection, prevented the replication of C. burnetii in monocytes and macrophages that have bound apoptotic lymphocytes by stimulating their inflammatory response. In conclusion, we suggest that leukocyte apoptosis associated with valvulopathy may be critical for the pathogenesis of Q fever endocarditis by deactivating immune cells and creating a favorable environment for pathogen persistence.
| Infective endocarditis (IE) has long been recognized as a fatal disease. Despite the availability of antimicrobial agents and cardiac surgery, IE still causes high morbidity and mortality. About 75% of patients with IE have pre-existing cardiac diseases [1], including congenital cardiac malformations, acquired valvular dysfunction and prosthetic cardiac valves [2]. Normal endocardium is resistant to colonization by bacteria [3] unless it exhibits pre-existing lesions. Lesions expose underlying extracellular matrix proteins and enable deposition of fibrin-platelet clots [4], bacterial adhesion [5] and recruitment of monocytes, which produce tissue factor and inflammatory cytokines [6]: this usually leads to the growth of vegetation. Cardiac valve lesions are associated with pathological fluid shear stress [7]. In vitro, fluid shear stress modifies the structure and the function of the endothelium [8] and increases apoptosis of neutrophils [9], platelets [10] and monocytes (Mo) [11], suggesting that leukocyte apoptosis may be related to cardiac valvulopathy.
IE associated with negative blood culture constitutes 5% of all IE cases. It is often caused by obligate intracellular pathogens, such as Coxiella burnetii [12]. This bacterium that resides in Mo and macrophages [13] is the agent of the so-called Q fever. In patients with acute Q fever and valve disease, chronic endocarditis will develop in 30% to 50% of cases [14]. Q fever endocarditis is characterized by the lack of vegetations [15] and granuloma formation and impaired systemic cell-mediated immune response [13], whereas acute Q fever is usually controlled by the cell-mediated immune system [13], including the interferon (IFN)-γ pathway [16],[17]. This suggests that mechanisms distinct from endothelial lesions are involved in the immune impairment associated with the development of Q fever endocarditis. It is tempting to speculate that leukocyte apoptosis induced by cardiac valvulopathy may impair the immune response to C. burnetii through the modulation of macrophage polarization induced by the binding of apoptotic cells. Indeed, the phagocytosis of apoptotic cells by phagocytes and neighboring cells results in a powerful anti-inflammatory and immunosuppressive response [18] via the secretion of anti-inflammatory molecules, such as interleukin (IL)-10 and transforming growth factor (TGF)-β1 [19]. Different activation states of macrophages induced by microbial products, cytokines, glucocorticoids or immune complexes have been described [20]. By referring to the Th1/Th2 nomenclature, many now refer to M1 and M2 macrophages. M1 macrophages, stimulated by lipopolysaccharide (LPS) and/or IFN-γ, have a high capacity for antigen presentation, express CCR7, exhibit high levels of inducible nitric oxide synthase (iNOS) and secrete inflammatory cytokines, such as tumor necrosis factor (TNF), IL-1, IL-6 and IL-12, and chemokines, such as CXCL8, CCL2 and CCL5. M2 macrophages, induced by IL-4, IL-13 or IL-10, express the Fcγ-receptor type 2 (Fcγ-R2, CD23), the mannose receptor (MR) and CD14 and secrete anti-inflammatory cytokines, such as IL-10, TGF-β1 and IL-1 receptor antagonist (IL-1ra), and specific chemokines, such as CCL16, CCL18 and CCL24 [21]. The expression of arginase-1 by M2 macrophages shifts L-arginine metabolism toward the production of ornithine and polyamines by arginase-1, which, in turn, contributes to blocking the iNOS pathway [22].
The aim of this study was to determine the mechanisms by which valvulopathy creates favorable conditions for the establishment of Q fever endocarditis. We showed that apoptosis of circulating leukocytes was increased in patients with valvulopathy, independently of Q fever. We also showed that the binding of apoptotic lymphocytes (AL) by Mo and Mo-derived macrophages (MDM) increased C. burnetii replication and polarized Mo and MDM toward a M2 profile. Interestingly, IFN-γ prevented C. burnetii replication and re-directed Mo and MDM toward a M1 program. Leukocyte apoptosis associated with valvulopathy may be critical in the pathogenesis of IE by creating a favorable environment for pathogen persistence through the deactivation of immune competent cells.
Apoptosis was investigated by measuring circulating nucleosomes and caspase activity in leukocytes from control subjects, patients with valvulopathy and patients with Q fever (Figure 1). In patients with valvulopathy, acute Q fever and valvulopathy or Q fever endocarditis, circulating nucleosomes (Figure 1A) and the percentage of active caspases in leukocytes (Figure 1B–F) significantly (p<0.001) increased as compared to control subjects. Interestingly, circulating nucleosomes (p<0.001) and the percentage of active caspases in leukocytes (p<0.01) was significantly higher in patients with acute Q fever and valvulopathy than in patients with acute Q fever without valvulopathy. These results showed that patients with valvulopathy exhibited increased levels of apoptotic leukocytes.
To determine which type of leukocytes was apoptotic, cells with active caspases were quantified in CD3- and CD14-gated populations (Figure 1G). In patients with valvulopathy, with or without acute Q fever, the percentages of CD3+ and CD14+ cells that expressed active caspases were significantly higher (p<0.01) as compared to control subjects and patients with acute Q fever without valvulopathy. They were similar to those found in patients with Q fever endocarditis (Figure 1G). Taken together, these results show that increased leukocyte apoptosis was associated with valvulopathy.
Impaired immune responses associated with Q fever endocarditis may result from the binding of apoptotic cells to professional phagocytes. Consequently, we created an experimental model to test in vitro this hypothesis. Apoptosis of lymphocytes was induced by 10−6 M dexamethasone, which stimulated apoptosis via a caspase-dependent pathway (Figure S1). As a control, necrosis of lymphocytes was induced by a 95°C shock (Figure S1).
As apoptotic cells stimulates cytoskeleton reorganization in phagocytic cells [23], we analyzed the morphological changes induced by AL and necrotic lymphocytes (NL) in Mo (Figure 2A) and MDM (Figure 2B). Resting Mo and MDM were rounded and did not present ruffles or filopodia. The F-actin distribution was roughly homogeneous in quiescent Mo and MDM. AL induced intense morphological changes consisting of membrane ruffles in Mo and MDM. In contrast to AL, NL induced the formation of numerous filopodial extensions in Mo and MDM. F-actin was redistributed around membrane ruffles and filopodial extensions induced by AL and NL, respectively. The percentage of Mo and MDM with membrane ruffles or filopodia was quantified: AL induced membrane ruffles in about 70% of Mo (Figure 2C) and MDM (Figure 2D) whereas NL induced filopodia in about 75% of Mo (Figure 2C) and MDM (Figure 2D). These results showed that AL and NL induced distinct cytoskeleton reorganization in Mo and MDM.
Finally, we determined the time course of AL and NL binding. The binding of AL and NL to Mo and MDM was maximal after 2 h. Indeed, about 80% of Mo (Figure 2E) and MDM (Figure 2F) had bound at least one AL or NL after 2 h. Thus, a 2-h incubation time of AL and NL with Mo and MDM was used in further experiments.
Mo and MDM were incubated with AL or NL for 2 h and infected with C. burnetii. The binding of AL and NL to Mo (Figure 3A, inset) and MDM (Figure 3B, inset) had no effect on C. burnetii phagocytosis. C. burnetii survived without replication in control Mo (Figure 3A) and moderately replicated in control MDM (Figure 3B), as determined by real-time quantitative PCR (qPCR). Mo that had bound AL allowed C. burnetii replication after 9 days of culture (29,000±3,000 vs. 12,600±2,000 bacterial DNA copies in control Mo, p<0.001; Figure 3A). In MDM that had bound AL, C. burnetii replication was detectable after 3 days and was significantly (p<0.05) higher at day 9 (45,000±2,000 vs. 24,280±2,200 bacterial DNA copies in control MDM, Figure 3B). The effect of AL was independent of the apoptosis-inducer agent since the replication of C. burnetii was increased in Mo and MDM that have bound apoptotic cells induced by either corticoids or staurosporine (data not shown). In contrast to AL binding, the binding of NL to Mo (Figure 3A) and MDM (Figure 3B) induced C. burnetii killing. After 9 days of culture, the bacterial load was significantly (p<0.05) decreased by 66% in Mo (Figure 3A) and by 50% in MDM (Figure 3B), as compared with control cells. Bacterial infection was also determined by immunofluorescence. The percentage of Mo (Figure 3C) and MDM (Figure 3D) that contained more than 5 bacteria increased after the binding of AL. It reached about 25% after 9 days of culture, while this percentage did not exceed 15% in control Mo and MDM (p<0.001). In contrast, the number of Mo and MDM containing more than 5 bacteria was significantly (p<0.001) lower after NL binding (Figure 3C and D). Taken together, these results showed that AL binding stimulated C. burnetii replication in Mo and MDM, while the NL binding led to C. burnetii killing. Finally, the effect of AL and NL on C. burnetii replication was cell-contact dependent since the culture of Mo (Figure 3E) and MDM (Figure 3F) with AL and NL in separate chambers had no effect on C. burnetii replication.
In macrophages, C. burnetii survives in a late phagosome that fails to fuse with lysosomes [17]. The effect of AL binding on the intracellular traffic of C. burnetii in MDM was studied by determining the colocalization of C. burnetii with Lamp-1 (Lysosomal associated membrane protein-1), a marker of the late endosomes-early lysosomes, and the lysosomal protease cathepsin D. In control MDM, C. burnetii resided in a late phagosome unable to fuse with lysosomes (Figure 4A). Indeed, 78±6% of C. burnetii phagosomes were Lamp-1+ and cathepsin D− while only 22±8% of phagosomes expressed both Lamp-1 and cathepsin D (Figure 4D). The binding of AL (Figure 4B) and NL (Figure 4C) to MDM increased the maturation of the C. burnetii phagosome toward a mature phagolysosome: about 60% of C. burnetii phagosomes were Lamp-1+ and cathepsin D+ after binding of AL or NL (Figure 4D). Thus, AL and NL binding to MDM stimulated the maturation of C. burnetii phagosomes in the early phase of infection.
We hypothesized that AL binding may orient Mo and MDM toward a M2 program that favors C. burnetii replication. Transcriptional patterns of AL-Mo and AL-MDM were compared by clustering algorithm analysis (Figure 5A). Mo and MDM stimulated with C. burnetii exhibited distinct transcriptional programs that combine M1 and M2 features. The expression of CCL18, CCL24 and Fcε-R2, associated with the M2 profile, and CD1B, CD1C and CD1D were up-regulated in both cell types. In Mo, C. burnetii up-regulated the expression of CCL20, which is associated with M2 polarization; CCR7 and TNF, which are associated with M1 polarization; and CD1A and Fcα-R. It only down-regulated the expression of IL-20. In C. burnetii-infected MDM, the expression of CCL16 and IL-1ra, which are associated with M2 polarization, and that of CXCL8, CXCL11 and IL-6, which are associated with M1 polarization, were up-regulated. The expression of the two M1 cytokines IFN-γ and TNF were down-regulated. AL binding directed C. burnetii-infected Mo and MDM toward a M2 program. In Mo, AL binding up-regulated the expression of CCL16, CD14, IL-10, IL-1ra, which are associated with M2 polarization, IL-6, IL-13 and the three members of the IL-10 family, IL-19, IL-20 and IL-24. AL binding also down-regulated numerous genes involved in M1 polarization, such as CCR7, IL-12p40, IL-1β, TNF and iNOS. In MDM, AL binding up-regulated the expression of IL-10, IL-1ra and TGF-β1, which are associated with M2 polarization; and CXCL10 and IL-6, which are associated with M1 polarization. The expression of the M1 markers CXCL8, CCR7, IL-12p40, IL-1β and TNF were down-regulated. In contrast to AL binding, NL binding stimulated a typical M1 profile in C. burnetii-infected Mo and MDM. Indeed, NL binding up-regulated the expression of the M1 molecules CCL5, CXCL8, CXCL11, CCR7, CD80, IFN-γ, TNF, IL-12p40, IL-1β and iNOS in both Mo and MDM (Figure 5A).
Functional tests confirmed the transcriptional studies. After AL binding, Mo exhibited a M2 program characterized by high levels of IL-10 and IL-6, decreased TNF release (Figure 5B) and increased expression of CD14 (p<0.05, Figure 5D). As found for transcriptional responses, the response of MDM was characterized by M2 features that were partly different from those of Mo. Indeed, AL-MDM released high levels of TGF-β1 (Figure 5C) and the percentage of MDM that expressed MR was significantly (p<0.001) increased (Figure 5E). The effect of AL binding was specific since NL binding induced a M1 program in both Mo and MDM stimulated with C. burnetii. Indeed, the release of TNF by Mo and MDM was increased whereas those of IL-6, TGF-β1 and IL-10 were inhibited (Figure 5B and C). In addition, the expression of CD14 in Mo (p<0.001, Figure 5D) and that of MR in MDM (p<0.01, Figure 5E) were significantly decreased. Finally, we showed that IL-10 was critically involved in the replication of C. burnetii since adding blocking anti-IL-10 antibodies (Abs) to AL-Mo (Figure 5F) and AL-MDM (Figure 5G) prevented bacterial replication. Anti-TGF-β1 blocking Abs prevented the replication of C. burnetii only in AL-MDM and had no effect in AL-Mo (Figure 5F and G). Taken together, these results showed that AL binding stimulated a clear-cut M2 profile in Mo and MDM, in which IL-10 appears critical in both Mo and MDM and TGF-β1 in MDM.
As IFN-γ stimulates C. burnetii killing in Mo and MDM [24], we wondered if it may correct the permissive effect of AL binding on C. burnetii replication. IFN-γ inhibited C. burnetii replication in AL-Mo (Figure 6A) and AL-MDM (Figure 6B). After 9 days of culture, the bacterial DNA copy number decreased from 31,100±1,330 in AL-Mo to 8,300±450 in IFN-γ-treated AL-Mo (p<0.001). In AL-MDM, the bacterial DNA copy number decreased from 47,340±2000 to 18,400±3,800 in IFN-γ-treated cells (p<0.05). The number of bacterial DNA copies was similar to that found in Mo and MDM treated with IFN-γ. The percentage of cells that contained more than 5 bacteria was also significantly (p<0.05) lower in IFN-γ-treated AL-Mo and AL-MDM (Figure S2), demonstrating that IFN-γ completely reverted the effect of AL binding on C. burnetii replication. In addition, IFN-γ increased the maturation of C. burnetii phagosomes toward phagolysosomes independently of AL binding: more than 90% of C. burnetii phagosomes were Lamp-1 and cathepsin D positive after IFN-γ treatment (Figure 6C–E). Finally, IFN-γ reverted the response of AL-Mo and AL-MDM toward a M1 profile. First, it caused up-regulation of CXCL8, TNF and iNOS genes and down-regulation of IL-6 and IL-10 genes in AL-Mo. It also down-regulated the expression of TGF-β1 in AL-MDM (Figure 6F). Second, the IFN-γ treatment of AL-Mo and AL-MDM induced the release of high levels of TNF and inhibited the release of IL-6, IL-10 and TGF-β1 (Figure 6G). These results showed that IFN-γ prevented the effect of AL binding on C. burnetii replication and polarized AL-Mo and AL-MDM toward a M1 profile.
Patients with cardiac valve lesions have the highest risk to develop IE [2], although the role of valvulopathy in the establishment of IE is not clearly elucidated. We showed here, for the first time, that patients with valvulopathy exhibited increased levels of circulating apoptotic leukocytes, suggesting a link between valvulopathy and leukocyte apoptosis. We have previously demonstrated that Q fever endocarditis is associated with immune impairment [25],[26]. The immune impairment may be due to lymphopenia induced by lymphocyte cell death but lymphopenia was not reported in Q fever endocarditis [14]. We hypothesized that the binding of apoptotic leukocytes to Mo and macrophages may be responsible for the immune impairment observed in Q fever endocarditis. To test this hypothesis, we have established an in vitro model to study the effect of apoptotic cells on C. burnetii replication.
Binding of AL to Mo and macrophages induced the formation of large membrane ruffles with local F-actin redistribution. This result is in accordance with previous reports that showed that adhesion of apoptotic cells to specific membrane receptors leads to the formation of large membrane ruffles and cytoskeleton reorganization [27],[28]. We showed for the first time that NL binding induced cytoskeletal reorganization characterized by the formation of filopodia with F-actin concentration at their basis. The differences in cytoskeletal changes induced by AL and NL in Mo and MDM were not due to differences in binding levels: about 80% of Mo and MDM ingested AL and NL after 2 hours. They may result from the engagement by AL and NL of specific receptors that further modulate cytoskeleton reorganization in different ways [18],[29].
The binding of AL increased C. burnetii replication in Mo and MDM as do IL-10, the only cytokine able to stimulate C. burnetii replication [30]. Our results emphasized recent studies on the impact of the binding of apoptotic cells to phagocytes on survival of intracellular pathogens. The binding of apoptotic cells to murine macrophages increases the replication of the avirulent form of C. burnetii [31] and the growth of Trypanosoma cruzi [32] and Leishmania major [33]. In addition, the replication of C. burnetii induced by AL is dependent on AL contact with Mo and MDM. It has been previously demonstrated that the contact between monocytes and apoptotic cells is required for inducing the immunosuppressive response of monocytes [34]. Finally, the receptors engaged by AL binding are likely critical for C. burnetii replication. Indeed, when AL were opsonized (Protocol S1) with specific Abs to engage the Fc-receptor (FcR) pathway, the replication of C. burnetii was prevented (Figure S3). These results are supported by the fact that apoptotic cells opsonized with antibodies, particularly IgG, are recognized by macrophage FcR and stimulate a pro-inflammatory response [35]. This also suggests that the entry pathway orients the intracellular fate of C. burnetii.
Interestingly, the binding of AL or NL to macrophages increased the maturation of C. burnetii phagosomes. It has been recently shown that the uptake of apoptotic cells by macrophages or fibroblasts results in a rapid maturation of phagosomes by stimulating Rho GTPases [36]. It is tempting to speculate that the maturation of C. burnetii phagosomes observed after NL binding results also in the activation of Rho GTPases in Mo and MDM. However, the increased maturation of C. burnetii phagosomes, which occurs early during the C. burnetii infection, did not seem to govern the intracellular fate of bacteria in the later stages of infection.
AL binding by C. burnetii-infected Mo and MDM reprogramed them toward a M2 profile, but the properties of the M2 programs were different in Mo and MDM. In AL-Mo, the expression of genes encoding four members of the IL-10 family, namely IL-10, IL-19, IL-20 and IL-24, were up-regulated. The four genes are expressed within a highly conserved cluster [37]. IL-10 is highly secreted by macrophages following ingestion of apoptotic cells [38]. IL-10 is also critically implicated in the persistence of microorganisms and the chronic evolution of Q fever [39]. IL-19 and IL-20 down-regulate IFN-γ expression and up-regulate that of IL-4 and IL-13 in T cells, supporting a Th2 polarization [40]. In Mo, AL binding up-regulated also the expression of IL-1ra, which is increased in patients with Q fever [41]. In MDM, AL binding up-regulated the expression of TGF-β1, which is known to interfere with the activities of IFN-γ, iNOS and superoxide anion [42]–[44], alter the expression of co-stimulatory molecules [45], inhibit Th1/Th2 differentiation [46] and favor the expansion of regulatory T cells [47]. Interestingly, patients with acute Q fever and valvulopathy or Q fever endocarditis exhibit higher numbers of regulatory T cells as compared to patients with acute Q fever and healthy persons (our unpublished data). We also found that IL-6 was up-regulated in both AL-Mo and AL-MDM. IL-6 is largely considered as an enhancer of the inflammatory response. However, IL-6 can also act as a modulator of inflammatory responses since it shifts the T cell response toward a Th2 response by inducing B cell maturation [48]. Recently, it has been reported that IL-6 and TGF-β1 act together in inducing IL-10 production in T cells [49]. We can suppose that, in Q fever, IL-6 may contribute to the defective control of C. burnetii infection by macrophages. Finally, the replication of C. burnetii stimulated by AL binding was strongly associated with the production of M2 cytokines since IL-10 and TGF-β1 neutralization abolished bacterial replication. These results suggest a direct role of IL-10 and TGF-β1 in the signaling pathway leading to C. burnetii replication.
In contrast to AL binding, NL binding induced C. burnetii killing and a M1 transcriptional program in Mo and MDM. These results confirmed previous studies on the bactericidal response of Mo and MDM against C. burnetii induced by inflammatory cytokines, such as IFN-γ [24]. In addition, the expression of genes associated with a M2 program was down-regulated: peculiarly, the gene encoding IL-10 was inhibited in NL-Mo while the gene encoding TGF-β1 was inhibited in NL-MDM. Our results are consistent with other reports. The binding of necrotic cells induces the release of inflammatory cytokines by MDM [50]. It also induces the expression of major histocompatibility complex class II molecules by antigen-presenting cells and increases their ability to activate T cells [51].
Finally, IFN-γ inhibited the effect of AL binding on the C. burnetii replication in Mo and MDM. It induced C. burnetii killing and complete maturation of C. burnetii phagosomes. In contrast to LPS that is unable to inhibit the anti-inflammatory response of peripheral blood mononuclear cells (PBMCs) and Mo after binding of apoptotic cells [19],[34], IFN-γ reverted the M2 program induced by AL. It is likely that C. burnetii infection amplifies the inflammatory signal induced by IFN-γ to engage Mo and MDM toward a M1 profile. We can also suppose that the engagement of a broad number of receptors by C. burnetii, such as TLR4 [52], TLR2 [53] and the αvβ3 integrin [54], in the presence of IFN-γ counter-balances the anti-inflammatory signals delivered by AL to phagocytic cells.
In conclusion, we showed that valvulopathy increased the rate of circulating apoptotic leukocytes and we provided a model in which apoptotic cells play a key role in the establishment of Q fever endocarditis (Figure 7). The binding of apoptotic cells increased C. burnetii replication by subverting phagocyte responses; Mo and MDM, polarized toward M2 profiles, stimulate type 2 responses that are non-protective against most pathogens. If the systemic apoptosis of leukocytes occurs in an inflammatory context, such as that found in the presence of IFN-γ, the effect of AL binding is inhibited; Mo and MDM, polarized toward a M1 program, are able to kill C. burnetii as do patients with acute Q fever without valvulopathy. Our results give new comprehensive insights into the pathological processes resulting from valvulopathy that are associated with the high risk to develop rare and atypical IE.
Informed and written consent was obtained from each subject and the study was approved by the Ethics Committee of the Université de la Méditerranée. Nine patients with valvulopathy (5 men and 4 women, mean age of 69.5 years), 10 with acute Q fever (3 men and 7 women, mean age of 46.3 years), 11 with acute Q fever and valvulopathy (6 men and 5 women, mean age of 51.3 years) and 11 with Q fever endocarditis (7 men and 4 women, mean age of 47.6) were included. Ten healthy persons (6 men and 4 women, mean age of 35.0 years) were included as controls. The diagnosis of acute and chronic Q fever was based on epidemiological and clinical features, as previously described [55].
Plasma levels of nucleosomes were measured using the ELISA cell death detection plus kit from Roche Diagnostics. This assay is based on a quantitative sandwich enzyme immunoassay that recognizes DNA and histones [56]. The specific enrichment factor in nucleosomes, expressed in arbitrary units, was calculated as previously described [57].
Caspase activity of leukocytes was measured with the Apoptosis Detection Polycaspase Assay Kit (Immunochemistry Technologies), as previously described [57]. Briefly, 100 µl of EDTA-collected blood were incubated with 5 µl of 30× FLICA solution for 1 h and then with 20 µl of CD3-PE or CD14-APC for 15 min. After washing, leukocytes were fixed and analyzed by flow cytometry (EPICS XL, Coulter Beckman). The percentage of CD3+ or CD14+ leukocytes with active caspases was determined using the Expo32 ADC and the WinMDI 2.8 software.
C. burnetii organisms (Nine Mile strain) were cultured as previously described [54]. Dilacerated spleens of BALB/c mice infected with 108 C. burnetii organisms for 7 days were added to L929 cells. Infected cells were sonicated and centrifuged at 300×g for 10 min. Supernatants were collected and centrifuged at 10,000×g for 10 min. Bacteria were then washed and stored at −80°C. The concentration of organisms was determined by Gimenez staining and the bacterial viability was assessed using the LIVE/DEAD BacLight bacterial viability kit (Molecular Probes).
PBMCs were isolated from leukopacks (Etablissement Français du Sang, Marseille, France) by Ficoll gradient (MSL, Eurobio) and suspended in RPMI 1640 containing 20 mM HEPES, 10% fetal calf serum (FCS), 2 mM L-glutamine, 100 U/ml penicillin and 100 µg/ml streptomycin (Invitrogen). PBMCs were incubated in flat-bottom 24-well plates (Nunc) for 60 min at 37°C. After washing, adherent cells were designated as Mo (90% of cells expressed CD14) and MDM were obtained by a 7-day culture, as recently described [58]. Non-adherent cells were designated as lymphocytes (90% of them expressed CD3). Lymphocyte apoptosis, induced by incubation with dexamethasone (Merck) for 24 h [59], and necrosis, induced by heat shock, were determined by flow cytometry using the annexin V/Propidium Iodide (PI) kit (Roche) (Figure S1). Mo and MDM were incubated with AL or NL (ratio of 1∶5) for different periods, washed to remove unbound AL or NL, and infected with C. burnetii. In some experiments, Mo and MDM were separated from AL or NL using culture inserts (0.4-µm pore size; Transwell, Costar) in 24-well plates before C. burnetii infection.
Mo and MDM were incubated with C. burnetii (bacterium-to-cell ratio of 200∶1) for 4 h. After washing to remove free bacteria (time designated as day 0), infected cells were cultured for 12 days. In some experiments, recombinant human IFN-γ (1000 U/ml, Peprotech Inc.), monoclonal anti-IL-10 (10 µg/ml, R&D Systems) or anti-TGF-β1 (10 µg/ml, R&D Systems) Abs were added every 3 days. Infection was quantified by immunofluorescence [52] and qPCR. DNA was extracted with the QIAamp Blood Mini Kit (Qiagen) and stored in a volume of 100 µl at −20°C. qPCR was performed using 5 µl of DNA extract and the LightCycler FastStart DNA SYBR green system (Roche), as previously described [60]. Primers (Table S1) amplified a 75-bp fragment of the C. burnetii com1 gene (GenBank accession no. AF318146).
MDM were infected with C. burnetii for 4 h (bacterium-to-cell ratio of 200∶1), washed and cultured for 24 h. Cells were fixed in 3% paraformaldehyde, permeabilized with 0.1% Triton X-100 and immunofluorescence labeling was performed according to standard procedures [61]. Briefly, MDM were incubated with human anti-C. burnetii (1∶4,000 dilution), mouse anti-Lamp-1 (1∶1,000 dilution, DHSB, Developmental Studies Hybridoma Bank) and rabbit anti-cathepsin D (1∶1,000 dilution, a gift from S. Kornfeld, Washington University School of Medicine, St. Louis, Missouri) Abs for 30 min. Bacteria were revealed by Alexa 546-conjugated F(ab')2 anti-human IgG Abs, Lamp-1 by Alexa 488-conjugated anti-mouse IgG Abs and cathepsin D by Alexa 647-conjugated anti-rabbit IgG Abs. All secondary Abs were used at a 1∶500 dilution. The colocalization of bacteria and intracellular markers was examined by laser scanning microscopy using a confocal microscope (Leica TCS SP2) with a 63×/1.32-0.6 oil CS lens and an electronic Zoom 3×. Optical sections of fluorescent images were collected at 0.25-µm intervals using the Leica Confocal software and processed using the Adobe Photoshop V5.5 software. At least 65 MDM were examined for each experimental condition and the results are expressed as the percentage of bacteria that colocalized with fluorescent markers.
Mo and MDM were stimulated with C. burnetii for 4 h. Total RNA was purified using the RNeasy Mini Kit (Qiagen), according to the manufacturer's protocol. DNase treatment was performed with the RNase-free DNase set (Qiagen). The transcriptional pattern of cells was studied using a cDNA chip containing 440 arrayed sequences (Oligo GEArray Human Hematology/Immunology, SuperArray). Ten µg of RNA were transcribed into biotin-labeled cDNA by the MMLV reverse transcriptase (RT). Membranes were then hybridized with biotin-labeled cDNA and incubated with streptavidin-conjugated alkaline phosphatase. Chemiluminescence was visualised by autoradiography. Datasets were analyzed with the GEArray Expression Analysis Suite software (SuperArray) and the TIGR's Multiexperiment Viewer. Data were submitted to the ArrayExpress database (MIAME Accession number E-MEXP-1289). For qRT-PCR studies, reverse transcription of RNA was performed with the MMLV-RT kit (Invitrogen), according to the manufacturer's protocol. The primers (Supplemental data, Table I) were designed using the primer3 tool (http://frodo.wi.mit.edu/cgi-bin/primer3/primer3_www.cgi). RT was omitted in negative controls. The fold change in target gene cDNA relative to the β-actin endogenous control was determined as follows: fold change = 2−ΔΔCt, where ΔΔCt = (CtTarget−CtActin)test condition−(CtTarget−CtActin)reference condition. Ct values were defined as the number of cycles for which the fluorescence signals were detected [62].
Mo and MDM were incubated with heat-killed (100°C for 30 min) C. burnetii organisms (bacterium-to-cell ratio of 10∶1) for 24 h. Supernatants were stored at -80°C before immunoassays. IL-10, TGF-β1 and TNF assays were purchased from R&D Systems. IL-6 assay was purchased from Beckman Coulter. The intra- and interspecific coefficients of variation ranged from 5% to 10%.
Mo and MDM (5×105 cells per well) were incubated with AL and NL for 2 h, washed and then stimulated with C. burnetii (bacterium to cell ratio of 10∶1) for 24 h. After washing, Mo and MDM were scrapped and washed once with ice-cold PBS. Mo and MDM were then incubated with 10 µl of CD14-PE (Beckman Coulter) and MR-FITC (BioLegend, San Diego, California, USA) Abs for 30 min at 4°C. Cells were washed three times in ice-cold PBS and resuspended in PBS containing 10% FCS and 1% sodium azide (Sigma-Aldrich). Cells were then stored at 4°C in the dark and analyzed by flow cytometry (EPICS XL, Beckman Coulter). Ten thousand events were acquired for each sample. The percentage of positive cells was determined using the Expo32 ADC and the WinMDI 2.8 software.
Results are expressed as medians or means±SEM and compared with the non-parametric Mann-Whitney U test. Differences were considered significant when p<0.05. |
10.1371/journal.pgen.1008139 | Functional robustness of adult spermatogonial stem cells after induction of hyperactive Hras | Accumulating evidence indicates that paternal age correlates with disease risk in children. De novo gain-of-function mutations in the FGF-RAS-MAPK signaling pathway are known to cause a subset of genetic diseases associated with advanced paternal age, such as Apert syndrome, achondroplasia, Noonan syndrome, and Costello syndrome. It has been hypothesized that adult spermatogonial stem cells with pathogenic mutations are clonally expanded over time and propagate the mutations to offspring. However, no model system exists to interrogate mammalian germline stem cell competition in vivo. In this study, we created a lineage tracing system, which enabled undifferentiated spermatogonia with endogenous expression of HrasG12V, a known pathogenic gain-of-function mutation in RAS-MAPK signaling, to compete with their wild-type counterparts in the mouse testis. Over a year of fate analysis, neither HrasG12V-positive germ cells nor sperm exhibited a significant expansion compared to wild-type neighbors. Short-term stem cell capacity as measured by transplantation analysis was also comparable between wild-type and mutant groups. Furthermore, although constitutively active HRAS was detectable in the mutant cell lines, they did not exhibit a proliferative advantage or an enhanced response to agonist-evoked pERK signaling. These in vivo and in vitro results suggest that mouse spermatogonial stem cells are functionally resistant to a heterozygous HrasG12V mutation in the endogenous locus and that mechanisms could exist to prevent such harmful mutations from being expanded and transmitted to the next generation.
| Recent research has found that advanced paternal age is associated with increased risk in children to develop a subset of congenital anomalies, such as Apert syndrome, achondroplasia, Noonan syndrome, and Costello syndrome. The causative genetic errors (mutations) in these disorders have been identified to originate from the fathers’ testicles and their numbers increase with fathers’ age. It has been hypothesized that the germline stem cells that continuously self-renew and differentiate to supply sperm (referred as spermatogonial stem cells [SSCs]) carry these mutations and have the ability to expand preferentially as compared to normal SSCs with advancing age of the father, thereby increasing the likelihood of transmission of mutant sperm to the next generation. To test this hypothesis, we created a mouse model, in which a mutation known to enhance cell proliferation is induced in a subset of SSCs, and these cells compete with the neighboring normal (i.e., wild-type) stem cells. However, surprisingly, the germline cell population carrying the mutation in the testis was stable over a year of observation, suggesting that mechanisms could exist to prevent such harmful mutations from being expanded and transmitted to the next generation.
| In order to propagate genetic information to the next generation with high fidelity, germline cells must maintain a low mutation rate. Nevertheless, maternal germline cells (human oocytes) are well known to transmit abnormal chromosomes to offspring, especially in advanced maternal age (reviewed in [1]). Surprisingly, recent high-throughput genome analyses have revealed that men contribute a much higher number of mutations, specifically de novo single nucleotide mutations, to their children than do women [2–4]. Most strikingly, the risk of certain genetic disorders increases with advancing age of the father at the time conception of the child, referred to as the paternal age effect (PAE). This phenomenon could be explained by the unique biology of paternal germline stem cells. The latter are termed spermatogonial stem cells (SSCs), and, once established in the post-natal period, continue to self-renew and differentiate to supply sperm in mammals throughout adult life. This continuous self-renewal and long-term survival of SSCs may underlie the increase in mutation burden with paternal age, due to a cumulative increase in copy errors or other DNA lesions, despite the fact that the baseline germline mutation rate is thought to be lower than that of somatic cells [5]. Although the natural history of mutations in the aging testis is poorly understood, pathogenic variants are occasionally transmitted to offspring, resulting in a wide range of disorders. Among these, de novo gain-of-function mutations in the growth factor receptor-RAS signaling pathway are classically known to cause so-called PAE disorders, such as Apert syndrome, achondroplasia, Noonan syndrome, and Costello syndrome (reviewed in [6]).
Direct quantification of such mutations in the sperm and testes of healthy men of different ages has revealed an age-dependent increase in the mutation burden, in a manner that exceeds what would be expected from cumulative copy errors [7–9]. Moreover, in human testes, Ras pathway-associated mutations have been reported to occur in a clustered manner, suggesting that SSCs with PAE mutations are positively selected and clonally expand in normal, otherwise healthy testes over time [10–12]. We previously showed that a gain-of-function mutation in FGFR2 that causes Apert syndrome is sufficient to confer a selective advantage to murine SSCs in vitro [13]. However, no model system has been developed to interrogate mammalian SSC competition in vivo. Furthermore, no cell biological or molecular mechanisms have been described to explain this phenomenon. Although clonal expansion of stem cells with oncogenic mutations has been observed in the mouse intestinal crypt model [14, 15], it is not clear whether the same holds true for SSCs in the adult mouse testis. To test this long-standing hypothesis for SSC competition, we sought to establish an inducible mosaic model in which a hyperactive form of Hras could be induced within the endogenous locus in a subset of SSCs so that their long-term fate could be followed.
The undifferentiated spermatogonia (Aundiff) represent a population of cells in the mammalian testes that is defined by morphology and function. Along with somewhat more committed cells, the Aundiff pool contains long-term self-renewing SSCs. Morphologically, the Aundiff in rodents comprises As (single), Apr (pair), and Aal (aligned) cells, which are remarkably interconvertible, with significant migratory capacity and cell fate plasticity when subject to stress [16, 17]. Those cells reside along the basement membrane in the seminiferous tubules and are heterogeneous with respect to expression of genetic markers. Hara et al. (2014) first employed a cre driver controlled by the endogenous promoter of Gfra1, one of the robustly expressed markers for Aundiff and demonstrated that the labeled population marked by Gfra1-creERT2 comprised the long-term stem cell fraction [16]. Therefore, in our current study, we chose the same Gfra1 cre driver to create a novel germline mosaic model.
HRAS, a member of the RAS oncogene superfamily, is a monomeric GTPase and relays signals from receptor tyrosine kinases to the cell interior. It serves as a molecular switch for a MAP kinase signaling module in which HRAS is “on” when GTP is bound and “off” when GDP is bound (as reviewed in [18]). The HrasG12V mutation encodes a hyperactive form of HRAS protein that is locked in a GTP-bound state and cannot hydrolyze its bound GTP to GDP [19, 20]. HrasG12V is a rare mutation found in patients with Costello syndrome, whereas the HrasG12S mutation comprises the majority of probands [21], and it has been demonstrated that the Hras mutation burden in sperm from healthy donors increases according to donors’ age [9]. In mice, heterozygous HrasG12V mice phenocopy human Costello syndrome [22]. It was observed that overexpression of HrasG12V using transgenes in cultured mouse SSCs caused tumor development [23]. However, the effect of one copy of hyperactive Hras expression on paternal SSCs in vivo, simulating the putative earliest events in the human gain-of-function mutation disorder, has never been addressed.
In order to understand how a hyperactive HRAS affects long-term paternal stem cell fate, we induced HrasG12V at the endogenous locus in Aundiff in a mosaic manner. We found that Gfra1-creERT2 successfully drives mosaic HrasG12V activation in the adult germline in vivo. This model system allowed us to track mutated cell fate for prolonged chase periods in a quantitative manner. Surprisingly, the mutated SSC fraction persisted stably without significant expansion, suggesting that robust mechanisms exist to protect the SSC pool from harmful expansion of mutated cells and prevent transmission of deleterious alleles.
Whereas cultured SSCs are reported to express Hras [23], the abundance of Hras transcripts in the Aundiff spermatogonia in vivo has not been reported. Therefore, we evaluated Hras expression in vivo in Aundiff, as a surrogate for SSCs. To isolate the Aundiff population by flow sorting, we stained dissociated testicular cells with an anti-MCAM antibody (Fig 1A and 1B). Although MCAM-based sorting has been previously shown to enrich the Aundiff population, MCAM is also expressed in somatic cells [24, 25]. To avoid somatic cell contamination in FACS experiments, we first obtained tdTomato-labeled germ line cells using Gfra1-creERT2; tdTomfl/- mice, in which tamoxifen induces tdTomato efficiently in the Aundiff and their progeny but not in testicular somatic cells (Fig 1G, iWT). Three months after tamoxifen induction, tdTomato+ cells that had high, medium, low, or absent (negative) expression of MCAM were isolated by FACS and analyzed for Hras by RT-PCR (Fig 1A and 1B). Hras transcripts were found throughout the different germ cell populations tested (Fig 1C).
To induce the HrasG12V mutation exclusively in Aundiff, we employed the Gfra1-creERT2 mouse line. A “flox-and-replace” (FR) model generated by Chen et al. (2006) was utilized to produce monoallelic HrasG12V (Fig 1D) [22]. In this mouse line, the endogenous Hras locus consists of 2 tandemly-arrayed Hras genes, the upstream of which is a WT allele flanked by loxP sites, such that the downstream HrasG12V allele is silent until removal of the upstream 2.5 kb WT gene by recombination. By breeding, we generated Gfra1-creERT2; tdTomfl/-; FR-HrasG12Vfl/- mice. Tamoxifen-inducible Cre-mediated recombination of the Hras locus resulted in conversion from wild-type Hras to HrasG12V and activation of tdTomato expression at the Rosa26 locus. (Fig 1D & 1G). To verify the presence of the mutation in the testes, cDNA was obtained from FACS-sorted tdTomato-positive testicular cells, which consisted of pure germline cells (since the Gfra1 promoter is not active in testicular somatic cells) and was amplified for sequencing. Sanger sequencing confirmed that the expected nucleotide substitution (C>A) was present (Fig 1E). From the chromatogram, not all the tdTomato-positive cells exhibited HrasG12V three months after tamoxifen induction, indicating that wild-type and HrasG12V-positive germline cells coexisted in a mosaic manner in the labeled germline cell population. Importantly, Hras mRNA levels of Aundiff in vivo (i.e., tdTomato+/MCAM-high population in Fig 1A) measured by qPCR were similar between induced WT and FR mice, validating that a physiological level of Hras expression (WT + G12V) was achieved in the induced FR (iFR) model by utilizing the endogenous locus (Fig 1F).
To measure the fractional abundance of HrasG12V-positive cells in germline cells, two independent methods were developed (Fig 2A–2F). First, Sanger sequencing-derived plots of amplified cDNA from RNA were employed to quantify the mutated nucleotide (cytosine→adenine) from sequencing traces and create a model equation (S1 Fig). The validity of the model was confirmed by amplifying cDNA samples of RNA of known standard concentrations from WT (HrasWT/WT or HrasFR/WT) and Costello syndrome (HrasG12V/WT) mice at varying ratios (0–100% Costello mRNA) (Fig 2A–2C). The relationship between the HrasG12V-positive cell fraction (x%) and the ratio of adenine/cytosine (A/C) peak heights (y) fitted the model curve (y = 1.371*x / [200-x]). In the second approach, sperm genomic DNA (gDNA) analysis was used to quantify the HrasG12V-positive cell fraction. GDNA qPCR primers were designed to detect an SV40 poly-A (PA) region exclusive to the FR-HrasG12Vfl locus but which is lost after recombination (Fig 2D–2F). By performing qPCR with mixed gDNA comprising heterozygotic FR-HrasG12V (i.e., without recombination) and wild-type sperm at different known ratios, we confirmed a linear relationship between the FR-HrasG12Vfl recombined fraction (x%) and the SV40 PA allelic fraction obtained by qPCR (y%; Fig 2E). Since the loss of SV40 PA or FR means a gain of HrasG12V, y = 100-z was obtained, where z is the HrasG12V-positive cell fraction (Fig 2F). Using these two methods, HrasG12V-positive cell fractions of tamoxifen-induced animals were measured. The calculated fractions from sperm gDNA qPCR correlated well with that from the Sanger sequencing (Fig 2G). Furthermore, we observed a tamoxifen dose-dependent increase in the HrasG12V-positive fraction in sperm gDNA (Fig 2H). These results indicate not only that HrasG12V was successfully induced in Aundiff by tamoxifen but also that HrasG12V Aundiff can undergo differentiation and produce sperm.
Previously, Chen et al. (2009) used their murine FR-HrasG12V allele to model human Costello syndrome offspring [22] (Fig 3A left). In that study, FR-HrasG12V mice were crossed with Caggs-Cre (CC) mice to obtain HrasG12V heterozygotes, which phenotypically recapitulated Costello syndrome (driven by HrasG12S). In the prior model, HrasG12V induction takes place after fertilization, rather than in the parental germ cells. On the other hand, our inducible system enabled us to test whether mosaic HrasG12V-positive Aundiff (in a context of neighboring wild-type germ cells and somatic cells) can differentiate into normal sperm and give rise to F1 offspring with Costello syndrome (Fig 3A right). Tamoxifen-induced Gfra1-creERT2; tdTomfl/-; FRfl/- males (n = 3) were crossed with C57Bl/6J females and their pups were sacrificed at post-natal day 0 and genotyped using RNA. Genotyping revealed that Costello offspring were indeed born (left two pups in Fig 3B). This result demonstrates that HrasG12V-positive Aundiff are able to self-renew, undergo differentiation, and give rise to offspring.
HrasG12V has been detected in human sperm of healthy donors and is allelic to HrasG12S, the most common mutation in Costello syndrome, which increases with sperm donor age [9]. Given that gain-of-function mutations in the RAS pathway mediate positive selection of human SSCs [26, 27] and that overexpression of an HrasG12V cDNA is tumorigenic [23], we hypothesized that an increasing burden of HrasG12V would be detectable over time following induction in the adult testis. To test this hypothesis, our inducible mutation model enabled us to follow the fate of HrasG12V-positive Aundiff over time. The proportion of HrasG12V-positive cells in the labeled germline population was assessed at different time points after tamoxifen administration (Fig 4A). Sanger sequencing showed that the HrasG12V allele fraction did not increase from 3 to 14 months (Fig 4B, left). Notably, despite the long chase period, no germ cell tumors were detected in these animals, and the testes were normal at the gross histological level (see Fig 1G).
Sperm mutation analysis serves as a more physiologically relevant readout, since human studies on mutation burden utilized sperm samples from different age individuals [7, 9]. Thus, to mirror the human studies, we collected sperm from different time points following tamoxifen induction. Using qPCR, we quantified the HrasG12V-positive cell fractions in gDNA of sperm obtained from the cauda epididymis at these time points. Similar to the HrasG12V-positive cell frequency among tdTomato-labeled germ cells, the HrasG12V-positive sperm proportion also did not show an increase over the 11-month chase period (Fig 4B, right).
Sperm studies have demonstrated large inter-individual variance in the mutation burden in the same age cohort [7, 9]. Therefore, such static observations require a large sample size, which is not always readily achieved. Furthermore, the time at which mutations first appear in the human testis is unknown, nor is the mutational load at relevant loci at the time of early adulthood. An ideal study would examine sperm serially from same individuals, in order to capture a temporal change in the proportion of the variant cell population using a relatively small cohort. If selection for the HrasG12V genotype were to occur, one would predict an increase in the same individual over time. Thus, we established a method for in vivo serial sperm sampling. We performed microsurgical epididymal sperm aspiration [28], in which sperm was withdrawn twice from a live single male from the ipsilateral testis (Fig 4C & S2 Fig). For this study, we employed the mTmG reporter mouse line, in which membrane tdTomato expression converts to membrane GFP (mGFP) upon cre-induced recombination [29]. We generated Gfra1-creERT2; mTmGfl/-; FR-HrasG12Vfl/- males and induced HrasG12V and mGFP in Aundiff (Fig 4C). Sperm from wild-type siblings (i.e., Gfra1-creERT2; mTmGfl/-) served as controls. In the control model, the recombined mTmG fraction obtained by measuring the loss of the tdTomato allele (as a reference) by qPCR should not change over time unless there is competition between tdTomato+ and GFP+ Aundiff. Both the HrasG12V and reference allelic fractions were measured by qPCR as above (see Fig 2D). From 4 months to 12 months, a minimal increase (~9% on average) in the HrasG12V-positive cell fraction was observed (Fig 4E). Of note, the reference alleles in the wild type did not show any increase over the same time period, suggesting that tdTomato+ and GFP+ germ cells are mutually neutral with respect to fitness (Fig 4F).
To confirm whether the HrasG12V-positive cell fraction increases, we used a different method to assess the ratios of sperm bearing mGFP vs. tdTomato, resulting from Gfra1-expressing progenitors, in which the reporter transgene either remained intact or, alternatively, underwent recombination. The mGFP+ and tdTomato+ sperm were counted at two time points, at 4 months and 12 months, and the ratio of mGFP-positive to total sperm (mGFP-positive plus tdTomato-positive) was calculated. There was a strong positive correlation between values obtained by sperm counting vs. qPCR with the r2 = 0.90 (p<0.0001), validating that the sperm counting method itself was as accurate as the qPCR (Fig 4D). Thus, if the HrasG12V-positive sperm increase over time, we should observe an increase in the GFP+ sperm population. However, no change in the fraction of mGFP+ sperm in the mutant was detected (Fig 4G). Finally, we analyzed successive litters from tamoxifen-induced Gfra1-creERT2; tdTomfl/-; FR-HrasG12Vfl/- sires; no increase was observed in the proportion of HrasG12V/WT offspring obtained at the early phase of the study interval compared to the late phase (S1 Table). Taken together, these data suggest that the population structure is relatively stable and do not support a substantial change in the ratios of controls and mutant cells within the mosaic pool.
To test stem cell capacity of Aundiff containing the HrasG12V mutation, competitive SSC transplant assays developed by Kanatsu-Shinohara et al. (2010) were carried out [30]. In general, transplantation assays serve to measure cell autonomous capacity for self-renewal, survival, and differentiation. By mixing neutral, unlabeled competitor cells with either labeled experimental or labeled control donor cells, respectively, this method is designed to force the donor cells in question to compete against unlabeled competitors and allow direct functional comparisons between the two different donor types. Following two courses of tamoxifen treatment in vivo, the FR locus recombined at high efficiency, concurrent with the Rosa26-LoxStopLox tdTomato locus, indicating that most of the Aundiff were positive for HrasG12V and tdTomato (see next paragraph below). From the previous result above (see Fig 4B), this recombination ratio did not change over an extended period, suggesting that the number of initially induced HrasG12V-positive Aundiff remained stable by balancing self-renewal and differentiation for the rest of the time course. Therefore, the timing of testicular cell collection after tamoxifen (i.e., length of chase) should not affect the size of Aundiff population. Based on these findings, we chose to perform the transplantation assay 2 weeks after the start of tamoxifen administration. Dissociated testicular cells from these tamoxifen-administered WT (iWT) and FR (iFR) were mixed with non-labeled wild-type competitor testicular cells from Gfra1-creERT2 mice at a ratio of 1:1 and microinjected into busulfan-treated recipient testes (Fig 5A). After 10 weeks, tdTomato-positive colonies were counted (Fig 5B). Colony numbers were comparable between iWT and iFR, suggesting that the short-term cell-intrinsic capacity of stem cells of HrasG12V-positive Aundiff is not greater than that of wild-type cells (Fig 5C).
Activating mutations in oncogenes such as RAS family members are thought to produce relatively robust changes in cell phenotype and gene expression, particularly in tumor cells [31, 32]. To understand the molecular features of HrasG12V+ Aundiff in vivo, we isolated tdTomato+ and MCAM bright Aundiff (Fig 1A) from WT and FR mice treated with high-dose tamoxifen (i.e., two courses) for RNA-Seq three months after induction. In order to confirm that the HrasG12V-positive Aundiff fraction was sufficiently high in the mutant (MUT) group, a variant calling analysis was performed at position chr7:141192906, where the HrasG12V (c.35G>T) mutation substitutes an adenine for a cytosine. This revealed that the mutation fraction was 46–49% in the MUT group, indicating that >92% of the isolated Aundiff had undergone recombination and were positive for HrasG12V (Fig 6A). Differential gene expression analysis detected only minimal differences between the two groups, suggesting that Aundiff from WT and MUT are highly similar at the transcriptional level (S2 Table). Among the few differentially expressed genes, Pax7 was found to be upregulated in MUT. This result was confirmed by RT-qPCR (Fig 6C). However, the remaining seven differentially expressed genes were either functionally elusive or apparently irrelevant in Aundiff. To further capture any subtle differences in phenotype between WT and HrasG12V+ Aundiff, we manually curated a gene list comprising 28 stem cell marker genes and 25 differentiation marker genes based on recent studies that employed single cell RNA sequencing of mouse Aundiff [33, 34] and created a heat map for the differential expression (log2 CPM) for each sample (Fig 6B). Apart from Pax7, these 52 genes, were not differentially expressed; yet, the stem cell marker genes trended toward slight down-regulation and the differentiation markers exhibited slight up-regulation in the MUT group (Fig 6B). Despite these trends, we concluded that HrasG12V does not confer large transcription alterations in mutants.
Cultured SSCs serve not only as a complementary model for a complex in vivo system but also enable facile manipulation of extrinsic stimuli to enhance cell phenotypes. Thus, to elaborate on the functional effects of HrasG12V in vitro, we derived SSC cell lines from different adult mouse lines (Fig 7A). Upon 4-OHT addition in vitro, WT cells activated tdTomato (iWT1) or mGFP (iWT2) and FR cells initiated expression of HrasG12V, in addition to tdTomato (iFR1) or mGFP (iFR2) (Fig 7A). First, the presence of Hras transcripts in WT, iWT, FR and iFR lines was measured by designing RT-qPCR primers that recognized a common sequence between the wild-type and HrasG12V alleles. Levels of Hras transcripts were similar in these cell lines, validating that the SV40 polyadenylation signal in FR does not affect the Hras transcript (Fig 7B). Since HRASG12V protein is locked in a constitutively active conformation, we examined the amount of the functionally active form of HRAS in iFR SSC lines using a pull-down assay, in which only the active form of HRAS is recognized by the RAS-binding domain of RAF1. Active HRAS was pulled down only in the iFR but not in the iWT, demonstrating the presence of functional HRASG12V protein in this model (Fig 7C).
We next studied changes in the growth trajectory of HrasG12V SSCs. In normal culture conditions, iFR did not grow faster than FR (Fig 7D). Several lines of evidence show that a fitness advantage may be more pronounced upon environmental challenge, such as growth factor deprivation [13, 35]. We therefore examined whether HrasG12V SSCs outcompete wild-type SSCs when cultured at an FGF2 concentration 100-fold lower than is used in normal SSC growth media. For the following experiments, FR2 and iFR2 cell lines were employed (Fig 7A & 7E). We mixed the non-induced parental wild-type line FR2 (mtdTomato+) with iFR (mGFP+) at a 1:1 ratio and cultured them on feeder cells in FGF2 (0.01 or 10 ng/ml) for up to five passages. iFR did not replace its wild-type parental line over time (Fig 7E). This result suggests that a low FGF2 environment does not confer a competitive advantage to HrasG12V SSCs.
Hras is an important mediator molecule of FGF-RAS-MAPK signaling [23], and FGF2 has been shown to control stem cell self-renewal and possibly differentiation in SSCs [36, 37]. In various RASopathy models, RAS-MAPK signaling pathways are highly dysregulated [22, 38, 39]. However, it is unclear whether this holds true for SSCs harboring endogenous gain-of-function mutations seen in RASopathies. To understand how MAPK signaling is altered in HrasG12V SSCs in response to FGF2, iWT and iFR SSC clones were stimulated with FGF2 and probed for pERK and total ERK. pERK was similarly phosphorylated in iWT and iFR in response to FGF2 (Fig 7F). However, basal level ERK phosphorylation prior to stimulation appeared to be slightly higher in iFR as compared to iWT. Lack of enhanced ERK activity in iFR following stimulation was unexpected, considering the presence of active HRAS protein in iFR cell lines. This result suggests that active HRAS at a physiological gene dosage does not augment the signal response to FGF2 in SSCs.
The competitive interactions of wild-type and neighboring mutant germline stem cells in the adult human testis are thought to produce uneven transmission of pathogenic alleles to children, but a dearth of model systems has hindered progress toward understanding the details of selection. In this study, we created a stem cell competition system in the mouse germline, in which undifferentiated spermatogonia with endogenous expression of an oncogenic mutation, HrasG12V, compete with their wild-type counterparts in the testis. We found that the HrasG12V mutation was transmitted from the paternal germline stem cells to offspring, whereas the mutated germ cells did not exhibit a significant expansion over a year of fate analysis.
For in vivo genetic manipulation of the Aundiff, which contain the stem cell pool, only a handful cre driver lines have been described, including Vasa-cre and Stra8-cre. These cre alleles are not only non-specific with respect to germ cell subpopulations but also are active during early gonadal development, making it difficult to study adult mouse SSCs. For these reasons, cultured SSCs have been employed for loss- or gain-of-function studies using knock-down/out or ectopic expression of genes [13, 23]. Hara et al. (2014) first leveraged Gfra1-creERT2 for lineage tracing of adult Aundiff; by measuring long-term self-renewal of the population, they demonstrated that genetically marked Aundiff are functional stem cells that persist at homeostatic levels [16]. In our mouse model, therefore, we adapted Gfra1-creERT2 to induce both a mutation and a label in the Aundiff and achieved robust recombination at the target loci in Aundiff. Conveniently, the size of the recombined fraction was adjustable from ~8 to almost 100% by dose of tamoxifen. This tunability makes Gfra1-creERT2 an appropriate cre-driver to study germline mosaicism, in which only a small subset of stem cells is genetically manipulated or marked and their fate is traceable thereafter. On the other hand, with a high dose of tamoxifen, Gfra1-creERT2 is suitable to obtain a highly enriched population of genetically manipulated germline cells in vivo. Indeed, by combining Gfra1-creERT2 with a reporter and anti-MCAM staining, we were able to enrich an Aundiff sample with 96% HrasG12V positivity, as measured at the RNA level. Although Gfra1-creERT2 seems to be best available inducible driver to genetically manipulate adult Aundiff, it is notable that Gfra1 is broadly expressed in the Aundiff and that the Gfra1-expressing population comprises a heterogeneous subset of Aundiff. It remains controversial whether Gfra1-negative stem cells exist in the Aundiff.
Complex engineered alleles are required to model precisely the genetics of human germline disorders. In the FR-HrasG12Vfl/- system, we sought to induce HrasG12V without affecting total gene dosage throughout the recombination process; accordingly, similar levels of Hras transcript were observed before and after cre induction. Thus, such an approach most closely recapitulates the occurrence of human de novo mutations in male germline stem cells and the resultant germline mosaicism. Also, since this genetic model entails two tandem Hras genes, both comprising endogenous exons and introns, the final transcription product is also an endogenously spliced variant, in case any alternatively spliced SSC-specific variant exists. In our model, we found that recombination at the Rosa26 reporter locus happened more efficiently than that of the FR-HrasG12V locus. This could be explained by accessibility of the locus to cre recombinase; the Rosa26 locus is known for its constitutively active promoter region [40], whereas the activity of the Hras locus may be tightly regulated by transcription factors and/or chromatin remodeling.
Classical PAE disorders, such as Noonan syndrome, Costello syndrome, and other RASopathies, are caused by de novo mutations that are found only in affected children but not in the somatic DNA of either parent. Although most such mutations are derived from the paternal germline (as reviewed in [6]), the process of transmission from paternal germline stem cells to the offspring has not been widely explored. A major strength of our strategy is that it allows one to interrogate the transmission of the mutated gene induced in male germline stem cells to subsequent generations. Here, we observed that the fate of HrasG12V stem cells resulted in HrasG12V offspring with neonatal mortality, consistent with Costello syndrome. This indicates that Aundiff carrying HrasG12V can differentiate into functional sperm, and the resultant Costello embryos are capable of developing to the neonatal stage. Similarly, our germline mosaic model could be applied to a variety of other de novo disorders with PAEs. Owing to technological advancements in genome sequencing, many de novo mutations that originate in the paternal germline have been implicated as drivers of neurodevelopmental disorders (e.g., autism). Yet, adult germline mosaicism as a major source of human disease is still controversial, and its natural course has not been addressed experimentally. By inducing specific autism-driving mutations in Aundiff and tracking their fate, our mosaic model could be useful going forward to uncover the causal relationship between paternal age and neurodevelopmental disorders.
A human sperm analysis revealed that healthy males carried different Hras mutations at codon 12 [9]. Although G12S was the most prevalent mutation, G12V was also substantially elevated in the sperm and the level was weakly correlated with donor age. Thus, we sought to investigate how G12V-laden germline cells expand in our experimental model. To perform stringent lineage tracing of HrasG12V-positive cells, we employed multiple independent quantification methods and their outcomes were well correlated: cDNA sequencing from labeled germ cells, sperm gDNA analysis, and sperm counting. Over a year of lineage tracing, we did not observe a significant increase in the HrasG12V cell and sperm populations. In the serial sperm analysis, a minimal increase (~9%) was observed over the 8–9 month period, yet we did not observe an increase in the proportion of HrasG12V offspring from mutant sires in successive litters. Despite the fact that the HrasG12V is the most potent gain-of-function mutation among various HRAS proteins at codon 12, these findings indicate that the heterozygous HrasG12V alone is not sufficient to drive SSC competition in the mouse testis.
There are several possible explanations for the absence of competition in our model. First, there could be fail-safe mechanisms to suppress aberrant cell signaling driven by hyperactive RAS, providing a robust system that protects stem cells from harmful consequences. Second, HRAS may not have a major role in FGF-RAS-MAPK signaling in the Aundiff, which would be unexpected given previously published data [23]. Other members of the RAS superfamily may be critical for relaying FGF signals in the SSCs. Third, the HrasG12V mutation may be detrimental, such that that cell turnover could be faster than usual. Fourth, induction experiments performed at extremely low starting (i.e., baseline) mutation levels (e.g., <1% recombination) might reveal competitive interactions that could have been obscured in our experiments, due to as yet uncharacterized paracrine effects. Fifth, key differences could exist in the microenvironment or cellular population structure between mice and humans. Sixth, the observation period in this study (~one year) could be too short to capture long-term cell competition because of the much shorter lifespan of mice than of humans, which entails decades of continuous stem cell survival and self-renewal in the testis. Regarding these last two points, age-related changes in the stem cell niche could be necessary for HrasG12V-mediated stem cell competition to become apparent. Finally, additional, yet unidentified genetic lesions could be required to confer enhanced competitiveness.
An absence of cell competition was also observed in vitro. In cultured SSCs with HrasG12V, although a functionally active HRAS was detected, ERK activation following FGF2 stimulation was not enhanced. Chen et al. (2009) made a similar observation, using mouse embryonic fibroblasts (MEFs) with heterozygous HrasG12V [22]. Only in later passaged HrasG12V MEFs was more pERK detected than in controls. This suggests that active HRAS from one copy of HrasG12V is not adequate to activate ERK in either cultured SSCs or MEFs, and more time may be required to gain additional gene mutational hits. Furthermore, a reduced FGF2 environment did not favor selection of the HrasG12V cell population, indicating that HRAS may not be a major GTPase mediating FGF2 signaling in SSCs. On the other hand, when HrasG12V is overexpressed in cultured SSCs, increased proliferation and oncogenic transformation were observed [23], again suggesting that the HrasG12V effect is dependent on gene dosage.
The transcriptional profile between wild-type and HrasG12V Aundiff did not reveal a significant difference in gene expression. Interestingly, among a few differentially expressed genes, Pax7 was up-regulated in iFR cells. Pax7 is a transcription factor identified as a conserved marker for a particularly rare subset of Aundiff in mammalian testes [33, 41]. Although its function in self-renewal of Aundiff is not well characterized, Pax7 may be one of the downstream genes upregulated by MAPK signaling via HRAS in a subset of the Aundiff population. Overall, the absence of large transcriptional changes in HrasG12V Aundiff could account for the fact that there was no obvious competitiveness or higher stem cell capacity in HrasG12V-positive Aundiff in the transplant assays.
In conclusion, these results revealed a stable and tolerant system to prevent normal germline stem cells from being replaced by mutated cells. This unanticipated resistance to hyper-active Hras suggests inherent mechanisms within germline stem cells to suppress harmful mutations that would otherwise be propagated to offspring. In contrast, we anticipate that future studies will likely uncover factors that overcome such protective mechanisms, leading to aberrant clonal expansion. Given the increasing number of disorders (e.g., autism) linked to germline mosaicism and PAEs, the inducible adult mosaic model will be invaluable to understand the earliest origins of such pathogenic gene variants.
This study was approved by the Weill Cornell Medical College IACUC (#2010–0028). Either isoflurane or Ketamine/Xylazine was used for anesthesia in combination with buprenorphine and meloxicam for analgesia. For euthanasia, mice were exposed to CO2 followed by cervical dislocation.
Gfra1-creERT2 mice were a gift from Dr. Sanjay Jain [16, 42]. Reporter mice, tdTomato (#007914) [43] and mT/mG (#007676) [29], were obtained from the Jackson Laboratory. The FR-HrasG12Vfl/fl mice were previously generated by Dr. James Fagin [22]. The controls were wild-type littermates that do not have the FR-HrasG12Vfl allele but contain Gfra1-creERT2 to induce tdTomato expression. These mice were maintained on a mixed genetic background of C57BL/6J (>50%), 129/Sv, and Swiss Black mice. All the experimental protocols were approved by the Weill Cornell Medicine Institutional Animal Care and Use Committee.
At 6–8 weeks of age, 100 mg/kg of tamoxifen (Sigma) dissolved in corn oil (Sigma) was administered intra-peritoneally for 4 days (one standard course), unless otherwise specified. At each time point, animals were euthanized and the testes and caudal epididymal sperm were harvested for downstream experiments.
Detunicated testes were fixed with 4% paraformaldehyde in phosphate buffered saline overnight at 4°C, immersed in 30% sucrose, and embedded in OCT compound. After cryosectioning the samples at 10 μm, DAPI was applied. For whole-mount staining, after overnight fixation, the seminiferous tubules were untangled, washed in PBS, and blocked with 3% BSA/PBS with 0.1% Tween for an hour. After incubation with anti-Gfra1 antibody (1:200, BD) overnight, an anti-rabbit biotinylated secondary antibody, followed by Alexa647-conjugated streptavidin was used for detection. DAPI was used for nuclear staining. Images were captured with a Zeiss LSM 800 confocal microscope.
Microsurgical epididymal sperm aspiration was performed as previously described [28]. Under deep anesthesia, through a small skin incision on the scrotum, the cauda epididymis was punctured by a syringe, and its contents were aspirated (S2 Fig). The procedure was performed on ipsilateral testis at 4 months and again at 12 months after tamoxifen administration. gDNA extraction was performed using AllPrep DNA/RNA Mini Kit (Qiagen) with a modified protocol [44].
A whole testis dissociate was prepared using a two-step enzymatic digestion [45]. For MCAM staining, testes from Gfra1-creERT2; tdTomato mice (n = 3) were dissociated >3 months after tamoxifen administration. The single-cell suspensions from two testes were incubated with Alexa Fluor 647 anti-MCAM antibody (ME-9F1, BioLegend) at a concentration of 6 g/ml for 45 minutes at 4°C. After exclusion of doublets and DAPI-positive cells, Alexa Fluor 647 and tdTomato double-positive cells were gated and collected using a BD Aria flow cytometer.
Total RNA was extracted from sorted testicular cells or feeder-free cultured SSCs using Arcturus PicoPure Kit (Applied Biosystems) or RNeasy Plus micro kit (Qiagen), respectively, with an on-column DNA digestion protocol (Qiagen). Reverse transcription was performed using qScript (Quanta Biosciences) followed by a real-time PCR using Sybr Select Master Mix (Applied Biosystems) with a LightCycler 480II (Roche). Each technical triplicate was normalized to Actb and relative expression levels to control conditions were calculated using 2-ΔΔCt method.
FACS-collected tdTomato+ cells were lysed in RLTplus buffer (Qiagen RNA mini kit) with 2% β-mercaptoethanol, and RNA was purified according to the manufacturer’s instructions. The RNA was reverse-transcribed to cDNA using qScript cDNA SuperMix (Quanta Biosciences). A targeted region flanking HrasG12V was amplified by PCR and Sanger-sequenced. To quantify the mutated nucleotide (cytosine→adenine) from Sanger sequencing traces, a model curve was proposed and confirmed by amplifying cDNA samples of known concentrations from WT (HrasWT/WT or FR/WT) and Costello offspring (HrasG12V/WT) RNA (x-axis: 0, 25, 50, 75, and 100% of Costello RNA) (Fig 2A). By using 4 pairs of different offspring samples, it was confirmed that the relationship between the HrasG12V-positive cell fraction (x%) and ratio of adenine/cytosine (A/C) peak heights (y) fitted the model curve y = 1.371*x / (200-x). All the calculations for HrasG12V-positive cell fraction were done using this equation. ImageJ was used to measure A/C ratio via peak height. All the primers used in the study are listed in S3 Table.
To obtain a fractional abundance of the HrasG12V allele, qPCR primers were designed, detecting the SV40 PA region that is exclusively present in the FR-HrasG12Vfl locus and lost after recombination (Fig 2B). By running qPCR with mixed gDNA from a FR-HrasG12V heterozygote (without recombination) and wild-type sperm collected from cauda epididymides at different known ratios (x: 0, 25, 50, 75, 100% of FRfl/- allele), we confirmed a linear relationship between the FR-HrasG12Vfl recombined fraction (x%) and the SV40 PA recombined fraction obtained by qPCR (y%) (Fig 2B left). Since the loss of SV40 PA or FR means a gain of HrasG12V, y = -z+100 was obtained, where z is the HrasG12V+ cell fraction (%) (Fig 2B, right). Per reaction, 60ng of sperm gDNA in 4 ul was used. For normalization of the quantity of gDNA, Ngn3 gene primer sets were used. Light Cycle480 Software was used for analysis. To quantify a recombined fraction of a reference gene, mTmG, we employed the same strategy as the quantification for HrasG12V+ cell fraction. Primer sets were designed to detect a region of the mTomato locus that is deleted after recombination (Fig 2B).
Sperm obtained by microsurgical epididymal sperm aspiration was mounted on slides and imaged using a BX50 fluorescence microscope (Olympus) with a Spot Pursuit CCD camera (Diagnostic Instruments Inc). Total of 100 to 300 sperm (> 3 fields) were counted.
To obtain the highest labeling after the maximum dose of tamoxifen, testis cell suspensions from Gfra1-creERT2; tdTom mice and Gfra1-creERT2; tdTom; FR-HrasG12V mice (littermate of the former) were mixed with that of non-labeled wild-type (Gfra1-creERT2) at a 1:1 ratio, respectively, and transplanted into adult busulfan-conditioned C57Bl6 recipient testes (n = 10 or 11 mice per experiment). A total of 1.2 x 106 (first two experiments) or 0.6 x 106 (the 3rd experiment) mixed cells were injected per testis. Site of injection (left vs. right side) was alternated per genotype. Three independent transplantations were performed (3 donor mice per group). After 10 weeks, colony numbers were quantified using stereomicroscopy.
WT and FR mice (n = 3/ group) were treated with tamoxifen for four days for twice over two weeks. The testes were dissociated (see Fig 1A) and RNA from tdTomato+ MCAM-bright Aundiff was isolated via Arcturus PicoPure kit with DNase I treatment. Mean RNA Integrity Number (RIN) was 9.47 (SD 0.31). Libraries were constructed using TruSeq Stranded mRNA. Sequencing was performed on an Illumina HiSeq 2500 (v4 chemistry) with a 50 bp paired-end protocol. Variant calling and filtering were carried out using Bam-Readcount and SAMtools mpileup at chr7:141192906, where samples with heterozygous HrasG12V (c.35G>T) mutation should have evidence of a C/A genotype while wild-type samples have only a C genotype. Differential expression was assessed using DESeq2 with a false discovery rate (FDR) of 0.1.
SSC lines were derived from two pairs of littermate adult mice (see Fig 7A for specific cell lines) and maintained on mitotically-inactivated JK1 feeders [46]. SSC growth media was StemPro-34 with additional supplements as described previously [45]. Treatment with 6 μM of 4-OHT for 4 days was repeated for a total of 3–4 times over 3–4 weeks until the HrasG12V population was confirmed to be more than 70–80% by the sequencing method (see Fig 2A–2C). Between tamoxifen treatments, tdTomato-positive cells were enriched once by FACS to enhance the recombination efficiency for HrasG12V. All the experiments were performed using cells with passage number 9 to 22. The stem cell activity was confirmed by transplantation assays.
An HRAS activation assay was performed using a GST-fusion protein of the RAS-binding domain (RBD) of RAF1, as instructed by the manufacturer (Pierce Biotechnology). Briefly, SSCs maintained in growth media were washed with cold TBS and lysed. Active RAS was pulled down with GST-RAF1-RBD along with glutathione agarose resin, followed by Western blot detection with an anti-Hras antibody (sc-520, Santa Cruz Biotechnology).
After feeder subtraction, SSCs were washed once with ice-cold TBS and lysed in lysis buffer containing PMSF (Sigma), protease inhibitor (Sigma), and phosphatase inhibitor cocktails (Sigma). Protein concentration was quantified by the BCA method. Immunoblotting was performed according to standard procedures. Denatured samples were subjected to 12% SDS/PAGE gel and transferred to PVDF membrane. The following antibodies were used: pERK (9101, Cell signaling), ERK (9102, Cell Signaling), CyclophilinB (Invitrogen), and Hras (sc-520). To obtain mouse HRAS protein control for western blotting, a fragment (606 bp) of Hras (NM_008284.2) was synthesized by Integrated DNA Technologies, inserted into BamHI/EcoRI sites of pCIG [47], and overexpressed in HEK293T cells.
Gfra1-creERT2; mTmGfl/-; FRfl/- (FR2: tdTomato+ wild-type cells) and its induced derivative iFR2 (GFP+ HrasG12V) cells were mixed at 1:1 ratio, cultured in low (0.1 ng/ml) and high (10 ng/ml) FGF2 media, and passaged 5 times. At each passage, GFP+ cell ratios were measured by FACS (BD Accuri).
Feeder-free SSCs were starved for 18 hours and then either left unstimulated (0’) or stimulated with FGF2 (10 ng/ml), and harvested 5, 15, and 30 minutes after the stimulation. Cell lysates were analyzed by immunoblotting.
Results are presented as mean ± SD. At least three biological replicates and three technical replicates were performed for each experiment unless otherwise indicated in the text. GraphPad Prism was used for statistical analyses and generating graphs.
|
10.1371/journal.ppat.1003055 | The Calmodulin-like Calcium Binding Protein EhCaBP3 of Entamoeba histolytica Regulates Phagocytosis and Is Involved in Actin Dynamics | Phagocytosis is required for proliferation and pathogenesis of Entamoeba histolytica and erythrophagocytosis is considered to be a marker of invasive amoebiasis. Ca2+ has been found to play a central role in the process of phagocytosis. However, the molecular mechanisms and the signalling mediated by Ca2+ still remain largely unknown. Here we show that Calmodulin-like calcium binding protein EhCaBP3 of E. histolytica is directly involved in disease pathomechanism by its capacity to participate in cytoskeleton dynamics and scission machinery during erythrophagocytosis. Using imaging techniques EhCaBP3 was found in phagocytic cups and newly formed phagosomes along with actin and myosin IB. In vitro studies confirmed that EhCaBP3 directly binds actin, and affected both its polymerization and bundling activity. Moreover, it also binds myosin 1B in the presence of Ca2+. In cells where EhCaBP3 expression was down regulated by antisense RNA, the level of RBC uptake was reduced, myosin IB was found to be absent at the site of pseudopod cup closure and the time taken for phagocytosis increased, suggesting that EhCaBP3 along with myosin 1B mediate the closure of phagocytic cups. Experiments with EhCaBP3 mutant defective in Ca2+ -binding showed that Ca2+ binding is required for phagosome formation. Liposome binding assay revealed that EhCaBP3 recruitment and enrichment to membrane is independent of any cellular protein as it binds directly to phosphatidylserine. Taken together, our results suggest a novel pathway mediating phagocytosis in E. histolytica, and an unusual mechanism of modulation of cytoskeleton dynamics by two calcium binding proteins, EhCaBP1 and EhCaBP3 with mostly non-overlapping functions.
| Entamoeba histolytica is one of the major causes of morbidity and mortality in developing countries. Phagocytosis plays an important role in both survival and virulence and has been used as a virulence marker. Inhibition of phagocytosis leads to a defect in cellular proliferation. Therefore, the molecules that participate in phagocytosis are good targets for developing new drugs. However, the molecular mechanism of the process is still largely unknown. Here, we demonstrate that Calmodulin-like calcium binding protein EhCaBP3 is involved in erythrophagocytosis. We show this by a number of different approaches including immunostaining of actin, myosin1B, EhCaBP1 and EhCaBP3 during uptake of RBC; over expression and down regulation of EhCaBP3, and over expression of calcium defective mutant of EhCaBP3. Our analysis suggests that EhCaBP3 can regulate actin dynamics. Along with actin and myosin 1B it can participate in both initiation and formation of phagosomes. The Ca2+-bound form of this protein is required only for progression from cups into early phagosomes but not for initiation. Our results demonstrate the complex role of Ca2+ binding proteins, EhCaBP1 and EhCaBP3 in regulation of phagocytosis in the protist parasite E. histolytica and the novel mechanisms of manipulating actin dynamics at multiple levels.
| A variety of cell types, such as macrophages and neutrophils and many unicellular eukaryotes have the ability to engulf particles of size greater than 0.5 µm through a process called phagocytosis. In the former this process has evolved as one of the critical elements of host defence, while in the latter it serves as a mode of nutrition. Entamoeba histolytica, a parasite that colonizes the human gut and causes dysentery, is endemic in many developing countries and causes a high level of morbidity and mortality [1], [2]. Phagocytosis is considered to be important in E. histolytica pathogenesis, as a phagocytosis-deficient mutant showed reduced virulence [3]. In another study, the virulence potential of E. histolytica isolates could be directly correlated with their ability to phagocytose red blood cells (RBCs) [4].
Phagocytosis is initiated when a particle binds to a cell surface receptor, leading to local reorganization of actin cytoskeleton and providing the necessary force needed for the formation of phagocytic cups and phagosomes [5]–[7]. The rim of filamentous (F) actin (periphagosomal F-actin), surrounds early phagosomes and then progressively depolymerizes as the phagosome matures [5], [8], [9]. It is believed that this disassembly of the F-actin rim is necessary for phagosome maturation, as it may act as a barrier for phagosome-vesicle fusion [8]–[11]. Therefore, spatial and temporal regulation of actin dynamics is the key to controlling phagocytosis. This is achieved through a number of actin binding proteins (ABPs) [12]. ABPs are involved in regulating actin cytoskeleton dynamics at multiple levels; for example, promotion of nucleation and polymerization of F-actin by Arp2/3 complex and profilin [13], [14] and depolymerization of F-actin by ADF/cofilin and gelsolin [15]. Ca2+ is a prominent regulator that can exert multiple effects on structure and dynamics of actin cytoskeleton. Ca2+ transients during phagocytosis initiate these processes in many systems [16]–[18] including E. histolytica [19]. Cytoskeletal remodelling by Ca2+ may occur through Ca2+ binding proteins (CaBPs) that can sense alteration in Ca2+ concentration and undergo conformational change [20]–[22]. In Dictyostelium discoideum, a 34 kDa protein is involved in actin bundling in a calcium-regulated manner [23] and a 40 KDa protein restricts the length of actin filaments in the presence of Ca2+ [24], [25]. Ca2+ is also involved in other processes related to cytoskeleton remodeling, for example Ca2+-Calmodulin regulates actin polymerization via Fesselin [26] and a low molecular weight protein CBP1 in D. discoideum has been shown to regulate the reorganization of actin cytoskeleton during cell aggregation [27].
The role of actin in endocytic/phagocytic processes has been studied in different systems and cell types using a number of different inhibitors or pharmacological compounds [28]. Some of the results of these studies suggest that clathrin-coated vesicle formation may not require actin dynamics [29]. However, its role in post vesicle processing cannot be ruled out. In a different approach, over expression of Y282F/Y298F-FcgR, a signaling- dead mutant receptor in COS-7 cells is unable to signal to the actin cytoskeleton, but specifically binds IgG ligand, had no effect on phagocytosis [29]–[31]. In some of these cases it is thought that phagocytosis takes place via passive zipper mechanism in which ligand-receptor binding remains specific and strong but reversible due to the absence of actin polymerization. Passive engulfment is generally slower and produces much more variable phagocytic cups [32].
The genome analysis of E. histolytica has revealed 27 CaBPs with multiple EF-hand calcium binding domains [33]. Of these, EhCaBP1 has been studied in much more detail and it is now clear that EhCaBP1 is a central molecule involved in initiation of erythrophagocytosis along with EhC2PK, a C2 domain containing protein kinase [34]. EhC2PK accumulates at the site of RBC attachment in a Ca2+- dependent step and recruits EhCaBP1, which in turn brings actin filaments resulting in initiation of phagocytosis [34]. Ca2+ has been shown to participate in the initiation process at two levels. Firstly, it is necessary for membrane localization of EhC2PK and secondly, Ca2+-EhCaBP1 is required for phagocytic cups to progress towards phagosomes [34]. Therefore, Ca2+ has an important role in regulating erythrophagocytosis in E. histolytica.
A calmodulin-like calcium binding protein EhCaBP3 has been identified and partially characterized in E. histolytica [35]. Three dimensional structure, using nuclear magnetic resonance (NMR) spectroscopy, suggests that EhCaBP3 has a well folded N-terminal domain and an unstructured C-terminal counterpart, somewhat similar to calmodulin and EhCaBP1. Interestingly, EhCaBP3 was found in all three major cellular compartments; nucleus, cytoplasm and membrane [35]. In this report we show that EhCaBP3 is involved in the process of phagocytosis at both initiation and phagosome formation stages. In vitro experiments suggest that EhCaBP3 binds actin, and affects its polymerization and bundling. Therefore it is likely that EhCaBP3 regulates phagocytosis by participating in actin dynamics. Our studies also show that EhCaBP3 and EhCaBP1 have different roles though both are recruited early during phagocytosis. We conclude that E. histolytica displays unique mechanism of regulating phagocytosis using a number of novel calcium binding proteins not observed in any other system.
Ca2+ is required for phagocytosis in E. histolytica as chelation of cytoplasmic Ca2+ blocks phagocytosis [19]. Therefore, it is expected that CaBPs may be participating in phagocytosis as Ca2+ sensors. We have earlier shown the involvement of one of the calcium sensing CaBPs of E. histolytica, EhCaBP1 in erythrophagocytosis [19], [21]. EhCaBP3 was identified as a calmodulin-like calcium binding protein of E. histolytica as its structure showed similarity with calmodulin [35]. Since multiple CaBPs are likely to be involved in different steps of phagocytosis, the subcellular localization of EhCaBP3 was checked during RBC uptake by immunostaining with specific anti-EhCaBP3 antibody. The results are shown in Figure 1. Fluorescence signals clearly showed that EhCaBP3 was present in phagocytic cups, as has been shown for EhCaBP1 [19]. Actin was also observed to line the cups and the complete superimposition of both EhCaBP3 and actin suggested that both proteins are colocalized at the phagocytic cups (Figure 1A). EhCaBP3 was also found on early phagosomes along with actin. Superimposition of both molecules suggested that both EhCaBP3 and actin are also co-localized at the newly formed phagosomes.
Our earlier studies had shown that EhCaBP1 was found only at cups and not on phagosomes. Therefore relative localization of EhCaBP1 and EhCaBP3 were studied in actively phagocytosing cells in order to see functional differences between the two CaBPs, using antibodies against EhCaBP1 (red) and EhCaBP3 (green). Since we wanted to see both phagocytic cups and phagosomes, amoebic cells were incubated with RBCs for different times. As expected EhCaBP1 was observed only in the phagocytic cups whereas EhCaBP3 was found in both phagocytic cups as well as in early phagosomes (Figure 1B). The results suggest that EhCaBP3 is likely to be involved in erythrophagocytosis and it may be functionally different from EhCaBP1.
In order to check if EhCaBP3 may also participate in phagocytosis of other particles, EhCaBP3 was immunostained during phagocytosis of CHO cells and the results are shown in Figure S1. Fluorescent signals were found in the cups that are in the process of phagocytosing CHO cells. However, it was not clear whether any significant signal was present around the phagosomes, as observed with RBCs (compare Figure S1 and Figure 1). Phagosome with low intensity staining could be discerned in some cases and these are marked with asterisk. Many CHO cells formed tunnel like structure during phagocytosis and EhCaBP3 was localized at the tip (marked by an arrow). These tunnel-like structures have also been observed before [36]. The results suggest that EhCaBP3 may also be involved in phagocytosis of CHO cells. However, the extent of participation and the exact roles may be different from that of RBCs. We have further characterized the role of EhCaBP3 in phagocytosis using RBC uptake as our model.
Dynamics of EhCaBP3 recruitment and release during erythrophagocytosis was studied by expressing EhCaBP3 in E. histolytica cells as a GFP fusion protein on a plasmid vector maintained in the presence of G418 (Figure 1C). While there was no change in the expression of endogenous EhCaBP3 (17 kDa), the expression of GFP-EhCaBP3 (43 kDa) increased with increasing concentration of G418 as seen by western blotting using anti-GFP antibodies which do not stain endogenous EhCaBP3 (Figure 1D). There was no change in the levels of endogenous EhCaBP3 visualized by anti-EhCaBP3 antibody under the same conditions. Since it is likely that GFP tagged proteins may not behave like native proteins we checked the localization of GFP-EhCaBP3 during erythrophagocytosis using anti-GFP antibodies. Confocal microscopy revealed that GFP-tagged EhCaBP3 (but not GFP alone) enriched at phagocytic cups and early phagosomes along with actin (Figure 1E, 1F, 1G), suggesting that GFP-EhCaBP3 behaves in a similar way as endogenous EhCaBP3.
The results reported so far show that EhCaBP3 is required both at the initiation and end stages of phagocytosis. It appears to redistribute during the whole process. In order to observe this dynamic behaviour of EhCaBP3, time-lapse fluorescence microscopy was used with cells expressing GFP-tagged molecules in the presence of RBCs. The results clearly showed that EhCaBP3 first accumulated rapidly at the site of RBC attachment before moving towards the tip of the cups (Figure 2). EhCaBP3 was present at the time of scission and remained even after complete phagosomes were formed and detached from the membrane. The whole process took about 3 min after addition of RBCs (supplementary movie S1).
The results shown earlier clearly indicate colocalization of EhCaBP3 with F-actin in the context of phagocytosis. This may be brought about by binding of EhCaBP3 to F-actin directly or indirectly through a third molecule. In order to check these possibilities a direct binding assay of EhCaBP3 to F-actin was carried out by co-sedimentation. Polymerized actin was incubated with recombinant purified EhCaBP3 or other indicated proteins, and the complex was centrifuged and analysed by SDS-PAGE. Actin alone was found in the pellet fraction suggesting that the preparation contained mainly polymerized or F-actin. In the absence of actin, the pellet did not contain EhCaBP3 (Figure 3A, lane10). However, when actin was present EhCaBP3 was found in the pellet fraction (Figure 3A, lane 8). EhCaBP1 was also present along with polymerized actin in the pellet, as expected, being an actin-binding protein (Figure 3A, lane 4). In contrast, EhCaBP2, a close homolog of EhCaBP1 with a different function did not co-sediment with F-actin [21] (Figure 3A, lane 6). When actin was incubated with both EhCaBP1 and EhCaBP3, interestingly a complex containing both CaBPs and actin was detected in the pellet (Figure 3A, lane12). This could be due to a ternary complex (Actin, EhCaBP1 and EhCaBP3) or two separate binary complexes (Actin and EhCaBP3; Actin and EhCaBP1), which cannot be distinguished at present. Our results suggest that EhCaBP3 can bind F-actin directly.
To test the binding of EhCaBP3 to G-actin, a solid-phase assay was performed in the presence or the absence of Ca2+. It was observed that EhCaBP3 bound G-actin in the presence of Ca2+. However, the binding was inhibited by 75% when EGTA was added (Figure 3B). This data suggests that EhCaBP3-G-actin interaction requires Ca2+.
We then checked if binding of EhCaBP3 affects properties of actin, as EhCaBP1 was shown to alter the bundling of actin but not its polymerization [21]. EhCaBP1 was used as a negative control as it does not have any effect on actin polymerization [21]. First we tested if EhCaBP3 has an effect on actin polymerization by using pyrene-labelled G-actin. The rate of actin polymerization increased on adding increasing amount of EhCaBP3 reaching a saturation at about 10 µM. At this concentration both the rate as well as the value at saturation was higher by 50% compared to the control. No change in the rate of polymerization was observed in the presence of EhCaBP1 as expected (Figure 3C). To test whether EhCaBP3 influences bundling property of actin, the assay was performed in the presence and the absence of Ca2+. Majority (91%) of the actin was found in the supernatant fraction when actin alone, or in the presence of BSA were incubated without EhCaBP3, suggesting that there was no significant amount of actin in the form of bundles (Figure 3D). However, incubation of actin with EhCaBP3 led to bundling of actin as the majority of actin was in the pellet fraction. The result with EhCaBP3 was similar to that with known actin bundling agents, such as EhCaBP1 and alpha actinin [21]. In both cases actin was recovered from the pellet fraction after incubation. Our data also shows that actin bundling property of EhCaBP3 is independent of Ca2+ as actin was seen in the pellet in the presence and the absence of Ca2+. Our results suggest that EhCaBP3 is an actin remodelling protein and that EhCaBP1 and EhCaBP3 have different functional effects on actin.
Myosin IB is thought to be one of the proteins that interact with actin and is involved during some of the cellular processes in E. histolytica, such as phagocytosis [37]. The relationship between EhCaBP3 and myosin IB was investigated in the context of erythrophagocyosis using GFP-EhCaBP3 expressing cells and anti-myosin IB antibodies. E. histolytica cells were incubated with RBCs for different time points so as to capture different stages of phagocytosis. In all stages, that is, from cups to newly formed phagosomes, GFP-EhCaBP3 and myosin IB were found to co-localize (Figure 4). The presence of both myosin IB and EhCaBP3 at the tip just before phagosome closure (denoted by star) suggests that EhCaBP3 along with myosin IB may be involved in the process of phagosome closure. Localization of EhCaBP3 with myosin 1B suggests that these proteins might interact with one another. To confirm this, co-immunoprecipitation was carried out using immobilized anti-EhCaBP3 antibody and total cell lysate of E. histolytica trophozoites. The result is shown in Figure 5. While anti-EhCaBP3 antibody precipitated myosin 1B along with EhCaBP3 in the presence of Ca2+ (Figure 5), no myosin 1B was observed when EGTA was added, suggesting that Ca2+ is essential for their interaction.
We then investigated the importance of Ca2+ binding in the functioning of EhCaBP3. It was achieved by generating a mutant of EhCaBP3 which could not bind Ca2+ (EhCaBP3mEF). This was done by D→A and E→A mutagenesis respectively of the first D residue of all EF-hand motifs and last E residue of EF-I and EF-III (Figure 6A). The recombinant mutant protein did not bind Ca2+ as shown by ruthenium red staining (Figure S2). EhCaBP3mEF was checked for its ability to bind both F and G actin (Figure 6B and 6C). Polymerized actin co-sedimentation assay revealed that both wild type and mutant EhCaBP3 bound F-actin (Figure 6B). Binding to G-actin was carried out using a plate binding assay. EhCaBP3mEF did not bind G-actin unlike the wild type protein (Figure 6C) suggesting that binding of EhCaBP3 to G-actin requires involvement of Ca2+ whereas binding to F-actin does not.
EhCaBP3mEF was also checked for its ability to get recruited in phagocytic cups and phagosomes. This was done by expressing a GFP-tagged mutant protein in E. histolytica cells (Figure 6D) and monitoring GFP as described in “materials and methods”. In order to mark the phagosomes properly, a plasma membrane marker (EhTMKB1–9) was used [34]. Immunofluorescence images revealed that while the mutant protein was observed in the cups (Figure 6E; upper panel), none of the phagosomes contained GFP-EhCaBP3mEF (Figures 6E and 6F; lower panel) unlike wild type protein (Figure 6F; upper panel). Further, actin was present in both cups and phagosomes in cells expressing the mutant protein (Figure 6E). However, myosin 1B enrichment and recruitment to those phagocytic cups was hampered where EhCaBP3mEF was present (Figure 6G), suggesting that Ca2+ is essential for recruitment of myosin 1B to phagocytic cups via EhCaBP3. This is supported by co-immunoprecipitation result as binding of EhCaBP3 to myosin 1B was inhibited in the presence of EGTA. Interestingly cells over expressing the mutant protein displayed only 20% reduction in phagocytic cups, while the reduction in phagosomes was 65% compared with cells over expressing the wild type EhCaBP3 (Figure 6H), suggesting a dominant negative effect of expression of the mutant protein. Since wild type EhCaBP3 continues to be expressed from the endogenous gene it is likely that these molecules help continuation of phagocytosis at a slower rate, even in the presence of EhCaBP3mEF.
The results presented so far suggest that EhCaBP3 is associated with phagocytic machinery. In order to show whether it was also required for phagocytosis to occur, the level of EhCaBP3 was reduced by expressing specific antisense RNA. We have been able to down regulate expression of a number of genes using tetracycline-induced whole gene antisense RNA and this system was also employed to study the role of EhCaBP3 in phagocytosis [38]. The vector used and details of different constructs is shown in Figure 7A. On tetracycline addition the level of EhCaBP3 was significantly (55%) reduced in cells carrying antisense construct (EhCaBP3AS) as compared to the cells carrying only the vector (Figure 7B). This effect was specific as the amount of EhCaBP1 did not change. When EhCaBP3 gene was over expressed using the cloned gene in the sense orientation (EhCaBP3S) the amount of EhCaBP3 increased by 30% in the presence of 10 µg/ml of tetracycline (Figure 7C). E. histolytica cells carrying the sense and antisense constructs were then checked for erythrophagocytosis using a spectrophotometric assay. There was a 70% reduction in cells expressing EhCaBP3 antisense RNA (that is, in the presence of tetracycline) as compared with cells carrying only the vector in the presence of tetracycline, and cells carrying EhCaBP3 antisense construct in the absence of tetracycline. Over expression of EhCaBP3, that is addition of tetracycline to cells carrying a sense construct displayed an increase (40%) in erythrophagocytosis as compared to cells without tetracycline or vector containing cells in the presence of tetracycline (Figure 7D). The results of immunostaining of these cells are shown in Figure S3. The data showed that in cells expressing anti-sense RNA the cup formation was greatly reduced in the presence of tetracycline, while cup formation took place normally in cells expressing EhCaBP3 in the sense orientation with or without tetracycline.
Reduction in phagocytosis on down regulation of EhCaBP3 expression may be due to either a reduction in initiation, progression or scission of phagosome formation. It is also possible that all steps may be affected. In order to identify the site(s) affected, cells expressing EhCaBP3 antisense RNA were incubated with RBCs for indicated time and analysed by immunostaining. The results are shown in Figure 7E. In cells expressing EhCaBP3 in sense orientation many phagocytic cups were observed at about 3 min of incubation with RBC. However, the process of cup formation was delayed in antisense expressing cells. A few cups were visible only at about 8 min of incubation (Figure 7E). We also noticed that there was a defect in the closure of the cups to form phagosomes when EhCaBP3AS cells were incubated with RBC for 20 min (data not shown).The statistical analysis of the above data showed that cups appear in EhCaBP3AS cells at about 7 min after addition of RBC and there was a 58% reduction in the number of cups formed (Figure 7F). Interestingly cells over expressing EhCaBP3 consistently showed increased number of cups. It is also clear from Figure 7E that the amount of phalloidin staining in the cups is substantially less in EhCaBP3AS cells as compared to control cells suggesting that F-actin recruitment may also be affected. Quantitation of phalloidin staining in the cups showed 41% reduction in the intensity of F-actin in the phagocytic cups as compared to cells carrying only vector (Figure S4). This suggests that EhCaBP3 participates both in the initiation as well as closing stages during phagosome formation and that actin dynamics plays a critical role in EhCaBP3 function.
We have observed colocalization of myosin IB with EhCaBP3 in phagocytic cups and phagosomes (Figure 4). Therefore the distribution of myosin IB in EhCaBP3AS was studied in order to further validate interaction of these two proteins during phagocytosis. The results showed the absence of myosin IB at the phagocytic cups even after 20 min of incubation with RBCs in cells expressing anti-sense RNA of EhCaBP3, suggesting that EhCaBP3 is required for recruitment of myosin IB (Figure 8A). We have also visualized distribution of EhCaBP3, myosin IB and actin in over expressing EhCaBP3S cells. EhCaBP3 and myosin IB were found to accumulate at the site of cup closure whereas actin was mainly present just at the neck (Figures 8B1 and B2; lower panel). There was no colocalization of EhCaBP3 and actin at the tip (Figure 8B1; lower panel), unlike myosin IB and EhCaBP3. Overall this data suggests that EhCaBP3 and myosin 1B are involved in phagocytosis and both these proteins may be needed for scission of vesicles.
To check whether EhCaBP3 level has any effect on recruitment of EhC2PK and EhCaBP1 in phagocytic cups; these proteins were immunostained in EhCaBP3 anti-sense cells (Figure 9). Reduced levels of EhCaBP1 and EhC2PK were observed in the phagocytic cups suggesting that EhCaBP3 may be involved in creating a macromolecular complex along with actin, EhC2PK and EhCaBP1.
We have shown earlier that EhC2PK binds liposomes in the presence of Ca2+ and recruits EhCaBP1 in a calcium dependent manner [34]. To test whether EhC2PK also recruits EhCaBP3 at the plasma membrane, we have used liposome sedimentation assay as described before [34]. The results are shown in Figure S5. The presence of a specific immunostained band in the pellet (which contains liposomes) is an indicator of interaction. Unlike EhCaBP1, EhCaBP3 bound liposomes directly without the involvement of EhC2PK in the presence of Ca2+. Actin was found in the pellet only when EhCaBP3 was present (Figure S5A). The interaction required the presence of Ca+2 as EGTA reduced the intensity of bands in western immunostaining. EhCaBP1 alone was not able to bind liposomes and EhC2PK-bound liposomes, as expected (Figure S5C, B). This suggests that EhCaBP3 alone can bind lipids, and consequently membranes, unlike EhCaBP1. These results are consistent with our previous finding that EhCaBP3 is also localized at the membrane in E. histolytica [35].
Regulated actin dynamics is required at different stages of phagocytosis and is achieved through participation of a number of molecules, many of which are actin binding proteins [12]. In mammalian system Arp2/3 complex, aphiphysin 2, coronin, cofilin, WASP and Scar (also called WAVE) are some of the molecules known to participate in regulating actin dynamics by manipulating different steps, such as nucleation, polymerization, bundling and depolymerization, including fragmentation of filaments [39]–[44]. Many processes involving actin dynamics, such as cell polarity, psuedopod formation and endocytosis in higher organisms have been studied in detail and the molecular mechanisms mediating different steps of actin dynamics have been worked out [45], [46]. However, the mechanism of initiation of phagocytosis is understood only in a few systems, of which the best studied, is opsonisation involving Fc receptors [47]. Our laboratory has shown that a C2 domain-containing protein kinase EhC2PK along with a calcium binding protein EhCaBP1 is involved in initiating a signal transduction pathway that eventually results in phagocytosis of RBCs in the protist parasite E. histolytica [34]. EhCaBP1 helps in recruiting actin at the site of phagocytosis by bridging with EhC2PK, a Ca2+-dependent membrane binding protein. This is one of the first examples of direct involvement of a calcium binding protein in actin dynamics and initiation of an endocytic process. In this report we show that E. histolytica erythrophagocytosis requires participation of yet another calcium binding protein EhCaBP3. Our results suggest that unlike EhCaBP1 which acts only at the initiation stage of phagocytosis, EhCaBP3 is likely to participate in both initiation and phagosome closure stages. It also appears from our results that the proposed mechanism may not be applicable for RBC phagocytosis alone, but also applicable in phagocytosis of CHO cells, though the detailed mechanisms may be somewhat different.
A number of our observations support the conclusion that EhCaBP3 is involved in phagocytosis. Firstly, EhCaBP3 was observed in phagocytic cups and phagosomes by fluorescence imaging of both fixed and live cells. Secondly, the rate and extent of phagocytosis was greatly reduced in cells where EhCaBP3 expression was down regulated by antisense RNA, and finally, over expression of a Ca2+ binding- defective mutant of EhCaBP3 reduced the rate of phagosome formation showing a dominant negative phenotype. Though the involvement of EhCaBP3 in the initiation of phagocytosis appears to be similar to that of EhCaBP1, there are important differences in their chemical and biological properties. For example, Ca2+ binding affinities of the two molecules are different. The overall Ca2+ binding affinity of EhCaBP1 was more than 700 fold that of EhCaBP3, and their dissociation constants (Kd) were 1.3 nM and 1.85 µM respectively [19], [35]. This would indicate that the two proteins function optimally at different Ca2+ concentrations. However, we are not in a position to correlate actual local transient Ca2+ concentrations during phagocytosis with the function and Ca2+ binding properties of the two proteins due to lack of data about Ca2+ concentrations in E. histolytica. However, we can speculate that attachment of RBC to the surface of E. histolytica generates a local Ca2+ spike. EhCaBP3 and EhCaBP1 are likely to get activated at different stages of a spike when Ca2+ concentration can vary about 2 orders of magnitude resulting in sequential activation of the two EhCaBPs. However, we do not have at present any evidence in support of this. Further the two proteins are functionally different since EhCaBP1 is absent in newly formed phagosomes, while EhCaBP3 is present. This is an indication that EhCaBP3 may be participating in the process of phagosome closure.
The presence of EhCaBP3 along with myosin IB at the tip of membranes before closure to form phagosomes strongly suggests that EhCaBP3 along with myosin IB may be involved in psuedopod extension, phagosome closure and finally release of the vesicle into the cytoplasm. Association of myosin 1B and EhCaBP3 has also been validated by a pull down assay. This appears to be similar to a mammalian long tail class 1 myosin that also localizes to phagosomes at late stages and participates in phagosome closure [48]. Further, transient localization of class 1 myosin to phagocytic cups has also been observed in Acanthamoeba, and in yeast myosin 1 facilitates different events of endocytosis, such as membrane fusion and vesicle scission [49], [50]. Myosins are also known to manipulate dynamics of actin filaments [51]. However, the interplay between myosins and actin in filament dynamics in relation to phagocytosis, psuedopod formation and motility in E. histolytica is not yet understood. Both EhCaBP1 and EhCaBP3 can bind G- and F-actin directly. However, the effect of this binding translates into different biochemical changes. EhCaBP1 alters bundling properties of actin filaments without changing polymerization [21]. On the other hand, as shown here, EhCaBP3 enhances polymerization in addition to enhancing bundling formation. Together these two calcium binding proteins modulate dynamic properties of actin cytoskeleton, a unique feature not seen in any other system. Though the effects of EhCaBP1 and EhCaBP3 on actin polymerization and bundling were studied in vitro, we believe that the same properties are likely to be seen in vivo. Our assumptions are based on colocalization of actin filaments with these two proteins during phagocytosis and the observation that there is a reduction in phagocytosis when either the expression is reduced or mutant proteins are present. The reduction in phagocytosis is likely to be due to a defect in actin filament formations and this has been seen in reduced amount of F-actin in the phagocytic cups of EhCaBP3AS cells. It is also likely that the process of initiation is achieved through multiple steps and that EhCaBP3 interacts with as yet unknown molecules (other than actin) that participate in these steps. The molecular details of sequential changes in the state of actin, and the possible recruitment of other proteins by the CaBPs need to be worked out. It is not clear how EhCaBP1 moves out of the phagocytic cups before phagosome closure while EhCaBP3 does not. We suspect that other proteins, such as myosin IB may be involved as both myosin IB and EhCaBP3 were seen at the tips before closure.
Since EhCaBP3 is a small Ca2+ binding protein and only contains Ca2+- binding EF-hand motifs, it is expected that its function must be executed through binding of Ca2+. However, we observed that EhCaBP3mEF, a mutant of EhCaBP3 that could not bind Ca2+, is present in the phagocytic cups and over expression of this mutant protein led to only a small reduction in phagocytic cup formation. This is not surprising as the mutant protein is capable of binding F-actin and can cause bundling of actin similar to the wild type protein. The reasons for reduction in cup formation, though small, as compared to phagosome formation, on overexpression of the mutant protein in spite of being recruited in the cups, may be due to involvement of other proteins. We need to characterize the initiation complex and identify all the players before we can answer this question. However, it is clear that Ca2+ binding of EhCaBP3 is necessary for phagosome formation as only Ca2+-bound form of EhCaBP3 interacts with myosin 1B, and the latter's recruitment in phagocytic cups requires the wild type protein. Therefore, it appears that Ca2+ has multiple facilitators in the form of different CaBPs, and a large number of different species (Ca2+ bound and free forms) participate at different steps in the process of phagocytosis. We are beginning to understand some of the steps as outlined here.
The mechanism of recruitment of EhCaBP3 during the process of initiation of phagocytosis is not clear. Since it does not bind EhC2PK it may require participation of yet other unknown molecule(s). Alternately, molecules that are present in the membrane or may be recruited to the membrane due to changes in local Ca2+ concentration could form initiation complexes along with EhC2PK, EhCaBP1 and actin, along with other participants. Support for this comes from our observations that EhCaBP3 can bind liposomes in the presence of Ca2+ and can also form a complex with liposomes and actin. Interestingly the requirement of a complex formation involving EhCaBP1, EhC2PK and EhCaBP3 for initiation of phagocytosis is evident from the reduced recruitment of both EhCaBP1 and EhC2PK in EhCaBP3 down regulated cells. However, it is not clear if EhCaBP3 present in phagocytic cups migrates from other parts of the membrane or from a pool of membrane-bound EhCaBP3, or from the cytoplasmic pool. Further studies are needed to work out the detailed mechanisms including the pathway involved in formation of a multimeric complex of these proteins.
EhCaBP3 is likely to participate in multiple processes other than phagocytosis and actin mobilization. It is also present in the nucleus and the function of nuclear EhCaBP3 is not clear. Our studies show that the process of RBC phagocytosis in the human parasite E. histolytica follows a unique mechanism involving a number of molecules that have been identified only in this organism. Deciphering this pathway will be highly useful in understanding evolution of phagocytic mechanisms in eukaryotic cells, as E. histolytica is an early branching eukaryote. Moreover, phagocytosis is essential for the growth and survival of this parasite and blocking this process leads to inhibition of cellular proliferation. Therefore, unique molecules involved in the pathway could be potential targets for developing newer drugs.
E. histolytica stain HM1: IMSS and the transformants were maintained and grown in TYI-S-33 medium as described before [52]. Neomycin or Hygromycin (Sigma) were added at 10 µg ml−1 for maintaining transgenic cell lines as indicated.
Transfection was performed by electroporation. Mid-log phase cells were harvested and washed first by PBS and then cytomix buffer (10 mM K2HPO4/KH2PO4 (pH 7.6), 120 mM KCl, 0.15 mM CaCl2, 25 mM HEPES (pH 7.4), 2 mM EGTA, 5 mM MgCl2). The washed cells were then re-suspended in 0.8 ml of cytomix buffer containing 4 mM adenosine triphosphate, 10 mM glutathione and 200 µg of plasmid DNA. The suspension was then subjected to two consecutive pulses of 3,000 V cm−1(1.2 kV) at 25 µF (Bio-Rad, electroporator). The transfectants were initially allowed to grow without any selection for 48 h. Selection was carried out by adding G418 or hygromycin B (10 µg ml−1) depending on the plasmid used.
EhCaBP3 gene was cloned in the BamH1 site of pEh-Neo-GFP vector. The vector has been previously constructed (Gullien, unpublished) by cloning the GFP mut3 allele of GFP [53] in the unique BamH1 site of the pExEhNeo plasmid [54]. Calcium binding defective mutant was also cloned in the pEh-Neo-GFP vector at the C-terminus of GFP. The CAT gene of the shuttle vector pEhHYG-tetR-O-CAT [55] was excised using KpnI and BamHI and EhCaBP3 gene was inserted in its place in either the sense or the antisense orientation. The sequences of oligonucleotides used for making the above stated constructs are described in the Supplementary Table S1. Standard molecular techniques were used for making all these constructs.
Co-sedimentation assay was carried out following published conditions [21]. Briefly, 5 µM of rabbit muscle G-actin (Sigma) was polymerized in a polymerization buffer containing 100 mM KCl and 2 mM MgCl2 at room temperature for 1 h. After polymerization, actin was mixed with 1 mM ATP and appropriate target protein (5 µ M) in a total volume of 150 µl of G-buffer (10 mM Tris-Cl, pH 7.5, 2 mM CaCl2, 2.5 mM β-Mercaptoethanol, 0.5 M KCl, 10 mM MgCl2) and incubated for 2 h at room temperature. The samples were centrifuged at 100,000 g for 45 min at 4°C. The supernatant (one-fourth of the total) and pellet fractions (total) were analysed by 14% SDS-PAGE followed by Coomassie blue staining. In addition to WT (EhCaBP3), mutant (EhCaBP3mEF), EhCaBP1 and EhCaBP2 were also used as positive and negative controls respectively.
Solid phase G-actin binding assay was carried out as described before [21]. Briefly, different wells of a 96-well plate were coated with 5 µM G-actin in PBS overnight at 4°C and were blocked with 3% BSA in PBS for an additional 24 h. After washing with PBS-T (PBS containing (0.05% Tween-20), EhCaBP3 and CaBP3mEF were added to the wells in duplicates at concentrations ranging from 0.5 µM to 10 µM. Bound protein was detected with anti-EhCaBP3 antibody followed by HRPO-linked anti-rabbit IgG using the colorimetric substrate TMB (Sigma). The absorbance was monitored at 405 nm with a microplate reader (Bio-Rad, USA) after stopping the reaction with 2 N H2SO4. The reaction was carried out in the presence of 5 mM CaCl2 or 2 mM EGTA as indicated.
Polymerization assay was done as per the protocol supplied by the manufacturer (www.cytoskeleton.com). Polymerization of actin was monitored by an increase in fluorescence of pyrene-labeled actin (cytoskeleton, USA) with excitation at 366 nm and emission at 407 nm. The assays were carried out at 20°C in a Safas Flx spectrofluorimeter. A 100 µl sample containing 3 µM pyrene-labelled G-actin, was saturated with increasing concentration of EhCaBP3 (3 µM, 5 µM, 10 µM and 15 µM). EhCaBP1 (5 µM) was used as a control and the reactions were carried out in polymerization buffer (5 mM Tris-HCl, pH 7.5, 1 mM dithiothreitol (DTT), 0.2 mM ATP, 0.1 mM CaCl2, 0.01% NaN3,0.1 M KCl and 1 mM MgCl2).
The assays were carried as described before [21] and details are given in Text S1.
CNBr-activated Sepharose-4B (1 g, Pharmacia) was conjugated with anti-EhCaBP3 antibody following a protocol supplied by the manufacturer. Briefly, crude immunoglobulins were collected from the immunized serum using 40% ammonium sulphate and subsequently dialysed in coupling buffer (bicarbonate buffer). Usually, 10 mg protein was added per gram of resin. The resin was mixed gently for 18 h at 4°C. After coupling the coupled resin was processed as per the manual provided by the manufacturer. The conjugated Sepharose beads were incubated with E. histolytica lysate for 6 h at 4°C. The beads were then washed with wash buffer (10 mM Tris-Cl (pH 7.5), 150 mM NaCl, 1 mM imidazole, 1 mM magnesium acetate, 2 mM β-ME and protease inhibitor cocktail) three times. Ca2+ and EGTA were maintained throughout the process as required. After incubation the beads were washed sequentially with 60 mM Tris-Cl (pH 6.8), 100 mM NaCl and with 60 mM Tris-Cl (pH 6.8). The pellet was suspended in 2× SDS polyacrylamide gel electrophoresis (PAGE) buffer and boiled for 5 min followed by centrifugation for 5 min. The proteins were then analyzed by western blotting.
E. histolytica cells were labelled for immunofluorescent imaging following methods described before [21]. Cells were first washed with PBS and incomplete TYI-S-33 medium, and then resuspended in the same medium before transferring onto acetone-cleaned coverslips placed in a Petri dish. The cells were allowed to adhere for 10 min at 37°C and then were fixed with 3.7% paraformaldehyde (PFA) for 30 min at 37°C after removing the culture medium. The fixed cells were then permeabilized with 0.1%Triton X-100/PBS for 5 min. Additional treatment using chilled methanol (−20°C) for 3 min was needed for staining myosin IB. Fixed cells were then washed with PBS and quenched with 50 mM NH4Cl for 30 min at 37°C, followed by blocking with 1%BSA-PBS for 1 h. The cells were then incubated with primary antibody for 1 h at 37°C, followed by washing with PBS and 1%BSA-PBS before incubation with secondary antibody for 45 min at 37°C. When F-actin was labelled with phalloidin, the methanol step was omitted. Antibody dilutions used were: EhCaBP3 at 1∶50 (purified antibody), EhCaBP1 at 1∶200, phalloidin (Sigma; 1 mg/ml) at 1∶250, GFP monoclonal (Molecular Probes, Cat no. A11120) at 1∶250, myosin IB at 1∶30 [56], anti-rabbit or mice Alexa 488 (Molecular Probes, Catalogue No. A-11008 or A-11001) at 1∶200, anti-rabbit or mice Alexa 555 (Molecular Probes, Cat. No. A-21428 or A-21422) at 1∶300. The preparations were further washed with PBS and mounted on a glass slide using DABCO [1, 4-diazbicyclo (2, 2, 2) octane (Sigma) 10 mg/ml in 80% glycerol]. The edges of the coverslips were sealed with nail-paint to avoid drying. Confocal images were visualized by using an Olympus Fluoview FV1000 laser scanning microscope.
CHO cells were stained for 30 min with 20 mM Cell tracker orange dye (Molecular probes, Eugene, OR) in F12 medium containing 10% FCS. After staining, CHO cells were washed three times with fresh BI-S-33 medium, and approximately 4×105 CHO cells were incubated with 2×105 cells of amoeba expressing GFP-CaBP3 for indicated time points at 37°C in 500 µl of TYI-33 medium.
The cells expressing GFP-EhCaBP3 were plated onto a 35 mm Mat Tek glass bottom culture dish (MatTek Corporation) at 37°C. The medium was then removed after the cells got settled at the bottom and the glass chamber was filled with pre-warmed PBS. The dish was kept on a platform with a temperature controller to maintain temperature at 37°C. High-resolution fluorescent time-lapse imaging (Nikon A1R) of a moving and phagocytosing amoeba was performed. The images were captured at 8 s interval and 60× objective was used. The raw images were processed using NIS element 3.20 or Image J software available freely on the web (http://rsb.info.nih.gov/ij/).
Samples were separated on a 14% SDS–PAGE and the gel was then transferred on to a polyvinylidine fluoride (PVDF) membrane by semi-dry method and processed using standard protocols. The antigens were detected with polyclonal anti-GFP (1∶5000, Molecular probes; Cat. No. A6455) or anti-EhCaBP1 or EhCaBP3 raised in mice and rabbits (1∶5000, raised in our laboratory), followed by secondary anti-rabbit and anti-mice immunoglobulins conjugated to HRPO at 1∶10,000 dilution (Sigma, Cat No's A6667 or A2554). ECL reagents were used for visualization (Millipore). The concentration of proteins in a sample was estimated by bicinchoninic acid (BCA) assay using BSA as a standard.
The assay was carried as described before [57] and details are given in Text S1.
The liposomes were prepared as described by Avanti Polar Lipid, Inc. http://avantilipids.com. The proteins were incubated with liposomes in binding buffer (Tris-Cl (pH 7.5) 10 mM, β-ME 0.25 mM, NaCl 50 mM). CaCl2 and EGTA were used at 2 mM and 5 mM respectively at 37°C for 2 h with intermittent tapping. The liposomes were centrifuged at 18,000 g for 30 min, followed by washing with binding buffer to remove the nonspecific-binding proteins. Liposomes were than dissolved in SDS buffer and separated on SDS–PAGE. Specific proteins were detected by western blotting. For actin-binding assay the liposomes were incubated in polymerization buffer (Tris-Cl (pH 7.5) 10 mM, MgCl2 2 mM, KCl 50 mM, ATP 2.5 mM, β-ME 2.5 mM) with EhCaBP3, and actin.
Both mice and rabbits used for generation of antibodies were approved by the Institutional Animal Ethics Committee (IAEC), Jawaharlal Nehru University (IAEC Code No.: 18/2010).
All animal experimentations were performed according to the National Regulatory Guidelines issued by CPSEA (Committee for the Purpose of Supervision of Experiments on Animals), Ministry of Environment and Forest, Govt. of India.
|
10.1371/journal.pbio.1000615 | A Quantitative Test of Hamilton's Rule for the Evolution of Altruism | The evolution of altruism is a fundamental and enduring puzzle in biology. In a seminal paper Hamilton showed that altruism can be selected for when rb − c>0, where c is the fitness cost to the altruist, b is the fitness benefit to the beneficiary, and r is their genetic relatedness. While many studies have provided qualitative support for Hamilton's rule, quantitative tests have not yet been possible due to the difficulty of quantifying the costs and benefits of helping acts. Here we use a simulated system of foraging robots to experimentally manipulate the costs and benefits of helping and determine the conditions under which altruism evolves. By conducting experimental evolution over hundreds of generations of selection in populations with different c/b ratios, we show that Hamilton's rule always accurately predicts the minimum relatedness necessary for altruism to evolve. This high accuracy is remarkable given the presence of pleiotropic and epistatic effects as well as mutations with strong effects on behavior and fitness (effects not directly taken into account in Hamilton's original 1964 rule). In addition to providing the first quantitative test of Hamilton's rule in a system with a complex mapping between genotype and phenotype, these experiments demonstrate the wide applicability of kin selection theory.
| One of the enduring puzzles in biology and the social sciences is the origin and persistence of altruism, whereby a behavior benefiting another individual incurs a direct cost for the individual performing the altruistic action. This apparent paradox was resolved by Hamilton's theory, known as kin selection, which states that individuals can transmit copies of their own genes not only directly through their own reproduction but also indirectly by favoring the reproduction of kin, such as siblings or cousins. While many studies have provided qualitative support for kin selection theory, quantitative tests have not yet been possible due to the difficulty of quantifying the costs and benefits of helping acts. In this study, we conduct simulations with the help of a simulated system of foraging robots to manipulate the costs and benefits of altruism and determine the conditions under which altruism evolves. By conducting experimental evolution over hundreds of generations of selection in populations with different costs and benefits of altruistic behavior, we show that kin selection theory always accurately predicts the minimum relatedness necessary for altruism to evolve. This high accuracy is remarkable given the presence of pleiotropic and epistatic effects, as well as mutations with strong effects on behavior and fitness. In addition to providing a quantitative test of kin selection theory in a system with a complex mapping between genotype and phenotype, this study reveals that a fundamental principle of natural selection also applies to synthetic organisms when these have heritable properties.
| One of the enduring puzzles in biology and the social sciences is the origin and persistence of altruism, whereby a behavior benefiting another individual incurs a direct cost for the individual performing the altruistic action. A solution to this apparent paradox was first provided by Hamilton [1], who showed that a behavior increases in frequency when rb − c>0, where c is the fitness cost to the altruist, b is the fitness benefit to the beneficiary, and r is their genetic relatedness. While this rule has provided an important framework in which to conceptualize social evolution [2]–[12], it is based on several assumptions, including weak selection, additivity of costs and benefits of fitness components, and a special definition of relatedness that uses statistical correlations among individuals rather than genealogy to describe similarity. Several studies investigated how violations to these assumptions may lead to failures of Hamilton's original 1964 rule [13]–[21], but it is yet unclear how the combined effects of these factors may affect the evolution of altruism in organisms with a complex mapping between genotype and phenotype. It also remains to be investigated to what extent Hamilton's original 1964 rule is influenced by factors such as drift and interactions between loci within genomes [22],[23].
To investigate how a complex mapping between genotype and phenotype can affect the course of social evolution, we conducted artificial evolution with groups of robots in simulations by modifying a system recently developed to investigate the evolution of cooperative transport [24]. Eight small (2×2×4 cm) Alice robots [25] and eight food items were placed in a foraging arena with one white wall and three black walls. The performance of robots was proportional to the number of food items successfully transported to the white wall and the robots were given the option to allocate the fitness rewards of successfully transported items to themselves (selfish behavior) or share them with other group members (altruistic behavior—in this case the fitness reward of the food item was shared equally between the seven other robots in the group). By choosing appropriate fitness values for shared and non-shared food items (see Materials and Methods), it was possible to precisely manipulate the benefits and cost of helping behavior (i.e., the c and b values of Hamilton's rule, see Materials and Methods).
The robots were equipped with two motorized wheels and three infrared distance sensors that could detect food items up to 3 cm away, a fourth infrared distance sensor with 6 cm range allowing to distinguish food items from robots, and two vision sensors mounted on top of the robot to perceive the color of the arena walls (Figure 1A). These six sensors were connected to a neural network comprising six input neurons, three hidden neurons, and three output neurons (Figure 1B). Two output neurons determined the speeds of the wheels, while the third neuron determined whether the food items successfully collected were shared or not. The genome of the robots (33 genes) encoded the 33 connection weights of the neural network (see Materials and Methods) and thus determined how sensory information was processed and how robots behaved. Our analyses reveal that this system resulted in both pleotropic and epistastatic effects as well as a high proportion of mutations having strong effects on behavioral traits (i.e., leading to deviations from the assumption of weak selection).
We conducted 500 generations of selection in a population consisting of 200 groups. The probability of robots to transmit their genomes from one generation to the next was proportional to their individual fitness (see Materials and Methods). The selected genomes were randomly assorted and subjected to crossovers and mutations to create the 1,600 new genomes (200 groups of 8 robots) forming the next generation [24].
This experimental setup allowed us to independently manipulate the relatedness between robots within a group and the cost-to-benefit ratios of helping. To quantitatively test Hamilton's rule for the evolution of altruism, we investigated how the level of altruism (defined as the proportion of food items shared with other group members) changed over generations in populations with five different c/b ratios and five relatedness values (see Materials and Methods). For each of these 25 treatments, the selection experiments were conducted in 20 independently evolving populations. Because of the impossibility to conduct hundreds of generations of selection with real robots, we used physics-based simulations that precisely model the dynamical and physical properties of the robots. We previously showed that evolved genomes can be successfully implemented in real robots [26] that display similar behavior to that observed in the simulations.
Because the 33 genes were initially set to random values, the robots' behaviors were completely arbitrary in the first generation. However, the robots' performance rapidly increased over the 500 generations of selection (Figure 2). The level of altruism also rapidly changed over generations with the final stable level of altruism varying greatly depending on the within-group relatedness and c/b ratio (Figure 3). When the c/b value was very small (0.01), the level of altruism was very high in the populations where within-group relatedness was positive (i.e., 0.25, 0.5, 0.75, and 1.00) and close to zero when robots were unrelated (Figure 4). In the treatments with other c/b values, the level of altruism was also very low when the relatedness was close to 0 and the level of altruism also rapidly increased when the relatedness became higher than a given value. In all cases, the transition occurred when r became greater than c/b, as predicted by Hamilton's rule.
When the relatedness was equal to c/b, there was an intermediate level of altruism with the frequency of altruistic acts not differing significantly from the initial value, which was 0.5 (four one-sample Wilcoxon tests, df = 19, all p>0.368). This is the expected pattern because the inclusive fitness of robots, comprising both their own fitness points and those gained from altruists, is independent of whether or not they behave altruistically when r = c/b. Under such conditions, the level of altruism should vary only as a result of drift over generations, thus leading to important between-population variation in the level of altruism. Consistent with this prediction, the standardized variance (F = Var(p)/pq) in altruism when r was equal to c/b (F = 0.204) was significantly higher than when r was greater than c/b (F = 0.018; Mann-Whitney, df = 13, p = 0.002) and when r was smaller than c/b (F = 0.015; Mann-Whitney, df = 13, p<0.003).
The fact that the level of altruism remained slightly greater than 0 when r was smaller than c/b and slightly lower than 1 when r was greater than c/b can be explained by mutations maintaining some behavioral variability in the population. In line with this view of the level of altruism being at mutation-selection equilibrium, the level of altruism became significantly closer to zero (Pearson's r = 0.643; Mann-Whitney, df = 13, p<0.001) as the strength of selection increased (i.e., when the value r − c/b became more negative, only negative values of r − c/b considered for the correlation). Similarly, the level of altruism became significantly closer to 1 (Pearson's r = 0.805; Mann-Whitney, df = 13, p<0.004) as the strength of selection for higher levels of altruism increased (i.e., when the value r − c/b increased, only positive values of r − c/b considered in the correlation).
To determine whether mutations in our neural network had pleiotropic and epistatic effects and whether there were departures from weak mutations effects, we conducted additional experiments at the last generation in two treatments with intermediate r and c/b values (treatment 1: r = 0.25, c/b = 0.75; treatment 2: r = 0.75, c/b = 0.25). First, for each treatment, we subjected 4,000 individuals (one in each group) to a single mutation of moderate effect (see Materials and Methods). In the first experiment, performance was significantly affected by a much higher proportion of the mutations than the level of altruism (Table 1). Importantly, 1.36% of the mutations affecting the level of altruism also translated into a significant change in performance, indicating widespread pleiotropic effects. Similar results were obtained in the second experiment with 4.91% of the mutations affecting the level of altruism also significantly affecting performance. Second, we tested for epistatic effects by comparing the effect of a single mutation in 4,000 individuals with two allelic variants at another locus (see Materials and Methods). The genetic background significantly influenced the effect of the mutation in 2,371 (59.3%) of the cases in the first treatment and 2,336 (58.4%) of the cases in the second treatment. These results demonstrate that epistatic interactions are also widespread. Finally, our experiments showed frequent departures from weak effects on behavior and fitness. Performance changed by more than 25% for 1,616 (40.4%) of the mutations in the first treatment and 1,776 (44.4%) of the mutations in the second treatment, and the level of altruism changed by more than 25% for 552 (13.8%) and 1,808 (45.2%) of the mutations in the first and second treatment, respectively.
Although Hamilton's original 1964 rule provides a general framework of how natural selection works [17],[27], its theoretical and empirical applications usually involve the limiting assumptions of weak selection and additivity of costs and benefits of fitness components as well as the absence of pleiotropic and epistatic gene interactions [15],[16],[28] (but see [13] for relaxations of some of these assumptions in concrete applications), leading to the conclusion that the rb − c>0 rule should be used with caution when there are pleiotropic, epistatic, and non-additive effects [29],[30]. Interestingly, the genetic architecture of the robots in our system also led to departure from all these assumptions with the exception of non-additivity of costs and benefits of fitness components. However, the occurrence of non-additive (epistatic) effects of mutations at several loci in the genome leads to a situation that is conceptually similar to non-additivity of costs and benefits of fitness components [22]. In both cases, the fitness depends non-additively on gene action, with the interaction involving alleles at two loci on the same genome in the case of non-additive (epistatic) gene effects, and alleles at two homologous loci on two different genomes in the case of non-additivity of costs and benefits of fitness components.
Despite the fact that the assumptions mentioned above were not fulfilled, Hamilton's original 1964 rule always accurately predicted the conditions under which altruism evolved in our system. Whatever the c/b value used, altruism always evolved in populations where r was greater than c/b. This finding is important given that the assumption of weak selection, additivity of costs and benefits of fitness components and absence of pleiotropic and epistatic gene interactions are also likely to be violated in real organisms that also have a complex mapping between genomes and phenotypes.
Another important issue relates to the measure of relatedness. There has been considerable confusion in the literature since relatedness coefficients actually measure more than pedigree coefficients and because different derivations of Hamilton's rule take as their focal trait a variety of different quantities [16],[17],[30]. In the original derivation of Hamilton's rule [1] and many that followed (e.g., [12],[31]), the trait of interest was the genetic value at a single gene position and the regression coefficient of relatedness corresponded to an identity in state relative to the population average [31]. The interest in social evolution where social partners tend to be genealogical kin [1] has led to the use of Wright's F statistics as a measure of relatedness (e.g. [12],[22],[32]). Alternatively, Hamilton's rule has been derived to express the change in the social behavior phenotype (e.g., [16],[22],[33],[34]), often considered as a quantitative trait with many underlying gene positions contributing. In this case the coefficient of relatedness represents a regression of some measure of the individual's genetic value for that trait such as a breeding value [17], p score [16], gene frequency [1],[12], or partner phenotype on its own phenotype value [34].
Interestingly, the simple genetic structure of our groups leads to all these measures of relatedness being identical. In all our experiments groups were started by individuals randomly chosen from the previous generations. The relatedness between these founding individuals is therefore zero as they are not more genetically or phenotypically similar within groups than between groups. Positive within-group relatedness was created by cloning the founding individuals. Thus, positive relatedness was only due to one-generation coancestry and the probability that benefits of altruism being provided to a clone compared to an unrelated individual. Such a breeding system is conceptually very similar to that Hamilton had in mind when trying to explain the evolution of reproductive altruism in social insects where the sterile (altruistic) workers are the offspring of their mother queen (the individual benefitting from the altruistic worker behavior). The relatedness in such a system can also be described in terms of identity by descent [35], which provides an approximation of identity in state for rare genetic variants (see [31] for a recent review). Of interest would be to test in future studies how the evolution of altruism is influenced by more complex population structures where the effect of strong selection may lead to variation in within-genome differences in the covariance between genes in different individuals.
Because the rewards provided by the food items were either assigned to the focal individual who successfully transported it (selfish behavior) or shared equally between all the other group members (altruistic behavior), the fitness effects were additive and there were no synergetic effects. Thus, the cost incurred by an individual sharing altruistically a food item and the benefits to the other group members was not dependent on the recipients' genotypes and the proportion of them being altruistic. The lack of such synergetic effects results in the costs and benefits associated with an altruistic act being independent of the genotypic composition of the groups and the overall level of altruism in the population (i.e., there are no frequency-dependent effects). In natural systems there are frequently synergetic effects and this is one of the main reasons why it is not possible to reliably quantify the cost and benefits associated with altruistic actions (e.g., [15],[16],[36],[37]).
From an empirical perspective, our study is therefore valuable because there have been many tests of Hamilton's rule, but these studies are usually not quantitative due to the impossibility of assessing the costs and benefits of altruistic acts, even in the most simple social systems such as those documented in some bacteria [10],[38], social amoebae [39], or even synthetic microbial systems [36]. Our study also demonstrates that contrary to some misunderstandings [3], kin selection does not require specific genes devoted to encode altruism or sophisticated cognitive abilities, as the neuronal network of our robots comprised only 33 neurons. More generally, this study reveals that a fundamental principle of natural selection also applies to synthetic organisms when these have heritable properties [40].
Groups of eight Alice micro-robots and eight food items were placed into a 50×50 cm foraging arena. We chose a collective foraging task to investigate the evolution of altruism, because foraging efficiency is a key factor for many biological social groups such as ant or bee colonies [41]. Foraging required robots to locate a food item, to position themselves in front of the item, and to push it into a 4-cm-wide target zone along the white wall of the arena (the three other walls were black).
Robots were controlled by a feed-forward neural network consisting of six sensory input neurons, one bias input neuron, and six neurons with sigmoid activation. The robots had four infrared distance sensors, three of them sensing objects within a 3 cm range and the fourth, which was placed higher, having a 6 cm range. These sensors allowed robots to locate the food items and distinguish them from robots. Robots were also equipped with two vision sensors to see the white wall [24].
These six sensory inputs were scaled to a range of [−1; 1]. In addition to the sensory inputs the neural network also comprised a bias input set to a constant value of −1, which was used to encode the neuron firing threshold. These seven inputs were connected to three neurons in a hidden layer, which in turn connected to three output neurons. The strength of these 33 connections was determined by 33 genes, whose values ranged from 0 to 255 (i.e., 8 bit resolution per gene). The activation of each of the six hidden and three output neurons was calculated by multiplying each of its input values by its associated connection weight, summing over all inputs, and passing the sum through the continuous tanh(x) function to obtain the neuron's activation value in the range of [−1; 1]. The activation value of the first output neuron controlled the left motor speed, the second the right motor speed, and the third whether or not the successfully pushed food items were shared with other group members.
We used five different levels of relatedness in the experiments. To create groups of unrelated individuals (r = 0), we randomly distributed the 1,600 individuals in the 200 groups. To obtain groups with a relatedness of r = 1, we cloned one individual 7 times and formed groups with 8 genetically identical individuals. To create groups with a relatedness of approximately r = 0.75, we used two individuals (A and B) and cloned one seven times (clone proportion A:B = 1:7). The resulting relatedness in these groups was thus r,0.7492. To create groups with a relatedness close to r = 0.5, we similarly composed each group of three types of clones but in proportions 6:1:1, which led to r,0.5357. To create groups with a relatedness close to r = 0.25, we again composed each group of three types of clones, but this time using proportions 3:3:2, which resulted in a relatedness of r,0.2468. The genetic composition of groups thus differed from that of most animal groups in that some individuals were clones (r = 1) rather than belonging to kin classes such as full siblings (r = 0.5) or cousins (r = 0.125). However, in the absence of preferential interactions between kin, social evolution should be influenced by the average group relatedness. This is because genetic relatedness depends on interaction probabilities of genes [4], which in our model is equivalent to interaction probabilities between clonal individuals. Our experimental setup prevented preferential interactions between individuals by randomizing starting positions, having all robots being identical, and using a neural network that did not allow individuals to memorize past interactions.
To manipulate the c and b value of Hamilton's rule we modified the fitness values for shared and non-shared food items that were successfully transported. When non-shared, a food item provided a reward c to the selfish individual. When shared, the food item provided no direct benefit to the focal individual but a benefit b equally shared by the seven other robots in the group. The c/b ratios used were calculated using Queller's approach [42]. We used a value of 0.01 for the smallest c/b ratio because with a value of c/b = 0, there is no selection for foraging efficiency when r = 0, hence resulting in many populations going extinct (because no items were successfully foraged).
The foraging efficiency of each group was evaluated 10 times for 60 seconds and the inclusive fitness of each individual was estimated according to the number of food items collected and not shared + the number of food items that other group members collected and shared (these values being multiplied by c and b/7, respectively). The probability of the genome of a given robot to contribute to the next generation was directly proportional to the robot's inclusive fitness (roulette wheel selection with replacement [43]). Selected genomes were paired to conduct a crossing over with a probability of 0.005. The resulting genomes were subjected to mutation (probability of 0.005 per bit; i.e., 0.04 per gene). This process of selection, recombination, and mutation was repeated until there were enough genomes for the 1,600 individuals (200 groups) of the next generation. The level of altruism was calculated for each group as the proportion of collected food items that was shared within a group: A = n(a)/(n(a) + n(s)), where n(a) was the number of collected food items individuals shared and n(s) the number of items individuals did not share.
All 25 selection experiments were repeated 20 times (20 independent replicates). Evolution lasted for 500 generations for each experimental condition. For statistic analyses, the fitness and the level of altruism of all 200 groups in each of the 20 replicates were averaged over the last 10 generations. Means were compared with Mann-Whitney tests as Shapiro-Wilk tests showed that in many treatments the data did not follow a normal distribution.
In each of the individually evolving populations, altruistic interactions always occurred within groups, while the reproductive competition occurred at the level of the population. To manipulate the relatedness, we cloned genomes for each group and formed groups of different proportions of clones. Each group was composed of k different types of clones with respective frequencies xi, i = 1 … k, .
The genetic relatedness r quantifies the greater (or smaller) genetic similarity between individuals compared to the population average. Using the regression definition of relatedness [16],[42],where j indexes the individuals in the population and l indexes the social partner of j.
In our system corresponds to the average probability of a focal individual being genetically identical to another member of the population and to the average probability of a focal individual being a genetically identical clone of another member of its group. Assuming that populations contain m groups with n individuals each,
In all experiments the independently evolving populations consisted of m = 200 groups, each composed of n = 8 individuals.
Given that the evolution of social behavior is influenced by the relative rather than the absolute values of costs and benefits, we arbitrarily set and calculated the costs c and benefits b for the expected transition from selfish to altruistic behavior as
To test for pleiotropic effects, we studied the outcome of a single mutation on two behavioral measures, performance and altruism. Performance was determined as the number of food items collected by an individual, and the level of altruism as the percentage of these food items shared with other group members. One mutation was performed on one individual in each of the 200 groups for each of the 20 replicates at the last generation for each of two treatments with intermediate values of relatedness and c/b ratio (treatment 1∶ r = 0.25, c/b = 0.75; treatment 2∶ r = 0.75, c/b = 0.25). All 8,000 individuals were subjected to a mutation of medium effect. This was achieved by flipping, for each individual, the third of the eight bits of a randomly chosen gene, hence always resulting in a mutation size ±32. We chose this value because it was the median value of the mutations (range ±128) the robots were subjected to in the 500 generations of selection. The performance and level of altruism of each mutated individual was then evaluated in 100 independent trials in its group and compared to its performance and level of altruism before the mutation (Wilcoxon rank sum tests using a 5% significance level).
For the first treatment (r = 0.25, c/b = 0.75), rank sum tests could be conducted for 3,961 out of the 4,000 individuals as 39 individuals did not collect any food item either before or after the mutation, hence preventing determination of the level of altruism. For the second treatment (r = 0.75, c/b = 0.25), rank sum tests could be conducted for 3,848 out of the 4,000 individuals, as 152 individuals did not collect any food item either before or after the mutation.
To test for epistatic effects, we used the same individuals as used in the experiment on pleiotropic effects and assessed the performances of individuals without a mutation F(0) and with a mutation F(A). We then subjected each of these 16,000 individuals (8,000 without and 8,000 with a mutation) to a new mutation B (also of median effect) and assessed their fitnesses F(B) and F(AB). We then compared whether this new mutation had a similar effect on the fitness of individuals with and without the first mutation by evaluating each of the resulting 32,000 individuals in 100 independent trials and calculating z scores based on the standard deviation (SD) and mean fitness
Z scores could be calculated for 3.998 and 3,978 out of the 4,000 individuals for the first and second treatment, respectively (2 and 22 individuals, respectively, did not collect any food items). Statistics used a 5% (z = 2) significance level.
Models of social evolution, as most models in evolutionary biology, usually resort to weak selection, where different individuals have very similar fitness. To test whether the mutations frequently had large effects (i.e., whether there was departure from weak selection), we determined how frequently a mutation of median effect resulted in a greater than 25% change in performance and the level of altruism (4,000 individuals per treatment). Note that the value of 25% was arbitrarily chosen as there is no convention of what change in fitness can be assumed to be a departure of weak selection. Again Wilcoxon rank sum tests were performed on the 100 trials per individual with a 5% significance level.
|
10.1371/journal.ppat.1004659 | Viral and Cellular Proteins Containing FGDF Motifs Bind G3BP to Block Stress Granule Formation | The Ras-GAP SH3 domain–binding proteins (G3BP) are essential regulators of the formation of stress granules (SG), cytosolic aggregates of proteins and RNA that are induced upon cellular stress, such as virus infection. Many viruses, including Semliki Forest virus (SFV), block SG induction by targeting G3BP. In this work, we demonstrate that the G3BP-binding motif of SFV nsP3 consists of two FGDF motifs, in which both phenylalanine and the glycine residue are essential for binding. In addition, we show that binding of the cellular G3BP-binding partner USP10 is also mediated by an FGDF motif. Overexpression of wt USP10, but not a mutant lacking the FGDF-motif, blocks SG assembly. Further, we identified FGDF-mediated G3BP binding site in herpes simplex virus (HSV) protein ICP8, and show that ICP8 binding to G3BP also inhibits SG formation, which is a novel function of HSV ICP8. We present a model of the three-dimensional structure of G3BP bound to an FGDF-containing peptide, likely representing a binding mode shared by many proteins to target G3BP.
| Stress granules (SGs) are dynamic aggregates of proteins and translationally silenced mRNA that are formed in cells upon various stress conditions, such as virus infection. SGs are thought to be antiviral, and many viruses have hence evolved countermeasures to prevent their formation, often targeting the essential SG protein G3BP. Here, we show that several otherwise unrelated viral and cellular proteins all bind G3BP with the sequence motif FGDF, and thereby repress SG formation: the non-structural protein 3 (nsP3) of the Old World alphavirus Semliki Forest virus (a close relative of the emerging, highly pathogenic Chikungunya virus); the protein ICP8 of herpes simplex virus; and in addition, the cellular protein USP10 (an SG component and protein deubiquitinase that stabilises e.g. the tumor suppressor p53). In this work, we also present and validate a model of the three-dimensional structure of G3BP bound to an FGDF-containing peptide. The FGDF-mediated G3BP binding represents an attractive target for therapeutic interventions against a range of diverse viral infections, and may also regulate the p53-stabilising function of USP10 in cancers.
| The Ras-GAP SH3 domain–binding proteins (G3BP) are multifunctional RNA-binding proteins, present in two forms, G3BP-1 and G3BP-2 (here collectively referred to as G3BP). They have a well-described importance in mediating the formation of RNA stress granules (SG), both in cells exposed to environmental stress and viral infections [1,2]. SGs are formed when translation initiation is compromised after phosphorylation of eukaryotic initiation factor eIF2α [3] or inhibition of eIF4A [4]. The assembly of SGs allows for rapid redirection of translation to stress response mRNAs or, in the case of viral infection, for inhibition of viral gene expression. The G3BP proteins possess RNA recognition motifs (RRM), which, together with protein/protein interaction domains, are required for SG induction [2]. The N-terminus of G3BP comprises a nuclear transport factor 2 (NTF2)-like domain [5], which is likely involved in dimerization [5,6], but little is known about the functional consequences of such dimerization. The G3BP NTF2-like domain forms complexes with a number of cellular proteins such as ubiquitin-specific protease 10 (USP10), caprin-1 and OGFOD-1 [7–9]. G3BP-binding regulates the activity of USP10, a predominantly cytoplasmic deubiquitinating enzyme (DUB) [8] which stabilizes several important proteins including the cystic fibrosis transmembrane conductance regulator (CFTR) [10], the tumor suppressor p53 [11], the autophagy regulator Beclin-1 [12], the sirtuin family histone deacetylase SIRT6 [13], the NF‐kB essential modulator (NEMO/IKKγ) [14] and the transporter associated with antigen processing (TAP1) [15]. The G3BP binding region of USP10 is found within its N-terminal 76 residues [16], and this interaction inhibits the DUB activity [8,17].
SGs are induced by many virus infections and in turn, viruses have evolved many countermeasures, often targeting G3BP [18]. SG assembly in poliovirus infection is inhibited by cleavage of G3BP between residues Q325 and G326 by the viral 3C protease [19] separating the NTF2-like and RRM domains and leading to the formation of compositionally distinct SGs, lacking G3BP [20]. For some viruses, G3BP is recruited to foci of viral protein accumulation and may be important for efficient completion of the viral life cycle. In vaccinia virus (VV)-infected cells, G3BP is recruited to the cytoplasmic viral factories [21]. However, it has also been reported to have an antiviral role in VV infection [22]. Likewise, G3BP has been implicated as a potential component of the hepatitis C virus (HCV) replication complex [23] and may play an important role in virus assembly [24]. We and others have shown that the G3BP NTF2-like domain is directly bound by L/ITFGDFD repeat motifs in the C-termini of non-structural protein (nsP)3 of the Old World alphaviruses, including Semliki Forest virus (SFV) and chikungunya virus (CHIKV) [25–28]. Subsequent sequestration of G3BP to foci of viral protein accumulation renders the infected cells unable to assemble SGs, despite sustained high levels of eIF2α phosphorylation [28,29].
In this work, we set out to precisely define the characteristics of the G3BP-binding motif in the Old World alphavirus nsP3. We found that the core binding motif consists of two phenylalanine residues separated by a glycine and an aspartate residue (FGDF). Mutation of either of these residues prevented binding to either G3BP-1 or G3BP-2. We also find FGDF motifs at the N-terminus of USP10 and at the C-terminus of herpes simplex virus (HSV)-1 protein ICP8, and confirmed that these motifs similarly bind G3BP and block SG induction. Finally, we generated a molecular model of the G3BP/FGDF peptide interaction and validated the model by site-directed mutagenesis. These results reveal the mechanism by which at least two pathogenic viruses target G3BP. Moreover, they also provide a molecular basis for the regulation of USP10 activity by assembly of the inhibitory USP10/G3BP complex.
The C-terminal L/ITFGDFD repeat motifs, which constitute the G3BP-binding site of nsP3, are well conserved in the Old World alphaviruses [30] with particularly strong conservation of the phenylalanine residues at positions 3 and 6 of the seven residue motif and of the glycine residue at position 4. In order to identify essential residues in the SFV sequence for G3BP binding, we constructed mutants in which each residue from T2 to D7 in both repeats was exchanged for alanine (Fig. 1A). These were constructed in the context of the pEGFP-nsP3-31 construct, containing amino acids 447–477 of SFV nsP3 fused to the C-terminus of EGFP. The empty pEGFP-C1 vector, encoding EGFP with a C-terminal extension of 21 residues was used as a control. We previously demonstrated that EGFP-nsP3-31 and an analogous EGFP fusion protein containing amino acids 475–523 of the CHIKV nsP3 efficiently bind G3BP in transient transfection experiments [27,28]. BHK cells were transfected with vectors encoding EGFP alone, EGFP-nsP3-31 or each of the alanine mutants. Lysates were immunoprecipitated with G3BP-1 antibody and immunoblotted with antibodies against G3BP-1, GFP or actin (Fig. 1B, upper panels). As expected, EGFP-nsP3-31, containing the wild type (wt) nsP3 sequence, efficiently bound G3BP-1, but EGFP alone did not. The construct carrying a mutation of the threonine residue T2 to an alanine exhibited weak but readily detectable binding to G3BP. However, mutation of either of the residues F3, G4 or F6 completely disrupted binding. Finally, while mutation of the aspartate residues at D5 to alanine substantially reduced binding, mutation of D7 had no effect. Lysates were also immunoprecipitated with anti-GFP and immunoblotted for G3BP-1, GFP or actin (Fig. 1B, middle panels), with similar results. Furthermore, GFP-immunoprecipitates were immunoblotted for G2BP-2 (S1 Fig.). These results show that both alternatively-spliced isoforms of G3BP-2 (a and b) interact with EGFP-31 constructs in a similar manner to G3BP-1. We conclude from these results that residues F3, G4 and F6 are essential for G3BP-binding and that T2 and D5 contribute significantly.
Previously we have shown that in Old World alphavirus-infected cells, the C-terminal repeat sequences of nsP3 bind and sequester both G3BP-1 and G3BP-2 into foci of viral protein accumulation including cytopathic vacuoles (CPV) [27,28,31]. In order to test the effects of the non-G3BP-binding mutations on the life cycle of SFV, we mutated both nsP3 residues F451 and F468 to alanines in the infectious clone of SFV (corresponding to the mutation F3A in the two (L/I)TFGDFD motifs, hereafter referred to as SFV-F3A). Previously, we showed that SFV-Δ789, lacking nsP3 residues 449–472, including the G3BP-binding sites displayed slower processing of the P34 precursor [28]. To determine if SFV-F3A also displays slow processing phenotype, we infected BHK cells with wt SFV, SFV-F3A or SFV-Δ789 and compared levels of P34. Lysates from WT SFV or SFV-F3A-infected cells did not contain detectable levels of P34 (S2A Fig.), and we concluded that the F3A mutations do not result in defective P34 processing. To demonstrate that nsP3-F3A does not bind G3BP-1 in the context of a viral infection, we infected BHK cells with wt SFV or SFV-F3A at MOI 10. Lysates were immunoprecipitated with G3BP-1 antibody and analyzed by immunoblotting for nsP3, G3BP-1 and actin. As expected, nsP3 from wt SFV-infected cells coprecipitated with G3BP-1, but nsP3 F3A did not (Fig. 2A). To determine whether SFV-F3A nsP3 colocalized with G3BP, cells were infected at low MOI with wt SFV or SFV-F3A. Cells were fixed and stained for SFV nsP3, G3BP-1 and TIA-1, another SG marker. As expected, wt SFV nsP3 colocalized well with G3BP-1, with very little colocalization of G3BP-1 with TIA-1 (Fig. 2B). However, nsP3 in SFV-F3A infected cells did not colocalize with G3BP-1, confirming immunoprecipitation data indicating that nsP3 F3A does not interact with G3BP-1. Instead, SFV-F3A infected cells often contained foci of G3BP-1 and TIA-1 co-staining, suggesting longer persistence of SGs in cells infected with the SFV-F3A mutant.
Previously we demonstrated that at late times in SFV infection, when most of the infection-induced SG have been disassembled, SGs cannot be re-induced with phospho-eIF2α- or eIF4A-dependent stress inducers (sodium arsenite or pateamine A (Pat A), respectively) because G3BP is sequestered by nsP3 [28,29]. We next tested whether cells infected with wt SFV or SFV-F3A could mount an SG response to an exogenous stress inducer. Cells were infected with wt SFV or SFV-F3A and stressed with Pat A at 7 hpi for 1 h before fixation and staining for TIA-1 to detect SGs and nsP3 to detect SFV infection (Fig. 2C). Pat A was used since sodium arsenite has little effect in SFV-infected cells as these already display sustained eIF2α phosphorylation [28]. Approximately 100% of mock-infected cells responded to Pat A by forming SGs. After infection with wt SFV, 5.2% of cells had SGs, and this number was not significantly altered by Pat A treatment, confirming that wt SFV–infected cells cannot mount a stress response to secondary stress signals. SFV-F3A–infected cells had a significantly higher number of SG-positive cells than wt SFV–infected cultures (18.4%), and this proportion was increased to 49.2% by the Pat A treatment, indicating that in the absence of nsP3/G3BP interaction, infection-induced SGs persist longer and additional SGs can be stress-induced in SFV-F3A-infected cells. To determine whether the specific ablation of G3BP recruitment affects viral replication, we performed single-step and multi-step growth curves comparing wt SFV and SFV-F3A in MEFs. These data revealed that SFV-F3A propagated to titers between 1.5 and 2 orders of magnitude lower that wt SFV after both low and high MOI infection (Fig. 2D) indicating that blocking the nsP3/G3BP interaction by point mutation of the N-terminal phenylalanine residue in both FGDF motifs attenuates viral infection. Single and multi-step growth curve experiments were also performed in BHK cells with similar results (S2B Fig.). Finally, to address whether the F3A mutations in SFV-nsP3 affect other functions than SG inhibition by G3BP sequestration, we performed viral growth curves in eIF2α-AA MEF cells [32]. These cells contain a mutation at the phosphorylation site (S51A) in eIF2α gene, such that eIF2α cannot be phosphorylated and SGs are not induced. SFV-F3A replication was very similar to WT SFV in those cells (Fig. 2E). This result strongly suggests that the attenuation of SFV-F3A in WT cells is due to the action of a G3BP-dependent process downstream of eIF2α phosphorylation, most likely SG formation. Hence, inhibition of SG formation (in response to infection-induced eIF2α phosphorylation) is probably the primary function of the FGDF motifs in Old World alphavirus nsP3.
We have previously demonstrated that at late times in infection with SFV, a large proportion of cellular G3BP is recruited to sites of viral protein accumulation [28]. We wondered whether this interaction displaces USP10, a cellular binding partner of G3BP that also binds to the NTF2-like domain [8,25]. When we determined the localization of USP10 in SFV-infected MEFs, we did not detect any significant overlap with G3BP (S3 Fig.). This suggests that USP10 is excluded from the complex with G3BP by the nsP3 interaction, possibly by competition for the same site on G3BP. Recently, the G3BP-binding domain of USP10 was shown to be located within the N-terminal 76 amino acid residues [16]. Having shown that the G3BP-binding motif of nsP3 consists of two FGDF motifs, we asked whether a similar motif exists in the G3BP-binding region of USP10. When we aligned the sequences of the N-terminal 80 residues of USP10 from several species (Fig. 3A), we identified several well conserved phenylalanine residues (F10, F13, F18, F21, F22 and F59 in the human sequence), as well as a conserved FG(D/E)F motif at residues 10–13. We observed particularly high species conservation in the N-terminal 40 residues. To determine whether this USP10 region binds G3BP, we fused the N-terminal 40 residues of human USP10 to EGFP (EGFP-USP101–40). BHK cells were transfected with vectors encoding EGFP alone or EGFP-USP101–40. Lysates were immunoprecipitated with anti-GFP and immunoblotted with sera to G3BP-1, G3BP-2, GFP and actin (Fig. 3B, left panels). Indeed, both G3BP proteins efficiently coprecipitated with EGFP-USP101–40 but not with EGFP alone. Lysates were also immunoprecipitated with G3BP-1 antibody to verify reciprocal coimmunoprecipitation of EGFP-USP101–40 (Fig. 3B, middle panels). To determine whether this binding is dependent on the FGDF motif at residues 10–13, a panel of mutated EGFP-USP101–40 variants was created in which each residue from F10 to F13 was substituted to alanine. Cells were transfected with vectors encoding wt EGFP, EGFP-USP101–40 or each variant. Lysates were immunoprecipitated with anti-GFP and immunoblotted with sera to G3BP-1, G3BP-2, GFP and actin (Fig. 3C). Alanine substitution of F10, G11 or F13 completely disrupted binding, while the D12A mutation allowed weak binding to both G3BP-1 and G3BP-2. Similar results were obtained after G3BP-1 immunoprecipitation (S4 Fig.). These results are remarkably similar to those obtained following mutagenesis of the G3BP-binding domain of SFV nsP3 (Fig. 1), suggesting that both proteins indeed bind G3BP via their FGDF-motifs.
To test if the overexpression and binding of full-length USP10 to G3BP affects the SG response, BHK cells were transfected with EGFP alone, EGFP-USP10 wt or EGFP-USP10-F10A. Transfected cells were mock stressed or stressed with sodium arsenite for 1 h before fixation and staining for G3BP-1 and TIA-1 to detect SGs. Approximately 97% of the mock transfected cells and 71% of the EGFP transfected cells had SGs upon sodium arsenite treatment (Fig. 3D). Overexpression of wt EGFP-USP10 efficiently blocked the ability of the cells to form SGs (just 1% of cells showed SGs), whereas 59% cells transfected with the G3BP-nonbinding mutant EGFP-USP10-F10A responded to sodium arsenite stress with SGs. Representative images are presented in S5 Fig.. Taken together, the results in Fig. 3 show that USP10 acts as a negative regulator of G3BP-dependent SG formation and that this regulation is dependent on G3BP binding by the FGDF motif at residues 10–13.
In contrast to USP10, nsP3 contains two FGDF motifs (residues 451–454 and 468–471). We therefore hypothesized that a minimal G3BP-binding motif could consist of only one FGDF motif flanked by surrounding residues, which would provide enough space between the two motifs to allow nsP3 to bind two G3BP molecules. Based on our finding that both G3BP-1 and G3BP-2 bind FGDF motifs, we further hypothesized that if EGFP-nsP3-31 binds two molecules of G3BP, it would bind any combination of G3BP-1 and -2, with frequencies depending on the relative abundance of the two proteins and potential differences in affinity. Therefore, G3BP-2 should be detectable in nsP3 complexes immunoprecipitated with anti-G3BP-1 and vice versa, but USP10 should form a complex with only one molecule of either G3BP-1 or -2 (Fig. 4A). To test this, BHK cells were transfected with vectors encoding EGFP, EGFP-USP101–40 or EGFP-nsP3-31. When lysates were immunoprecipitated with anti-GFP, both G3BP-1 and G3BP-2 were detected in complex with EGFP-nsP3-31 and EGFP-USP101–40 (Fig. 4B, left panels). However, when G3BP-1 was immunoprecipitated, G3BP-2 could be detected in EGFP-nsP3-31, but not in EGFP-USP101–40 bound complexes (Fig. 4B, middle panels). This suggests that the nsP3 sequence, containing two FGDF motifs, can form a ternary complex with both G3BP-1 and G3BP-2 simultaneously, while the USP10 sequence, containing only one FGDF motif, can bind either G3BP-1 or G3BP-2, but not both.
Next, we endeavored to confirm the 2:1 stoichiometry of nsP3:G3BP using biophysical in vitro measurements. We employed isothermal titration calorimetry (ITC) to determine the affinity, stoichiometry and thermodynamic signature of the interaction between the purified NTF2-like domain of human G3BP-1, expressed in E. coli (G3BP-NTF2), and a peptide spanning residues 449–473 of SFV nsP3 (nsP3-25), containing two FGDF motifs. Injection of nsP3-25 into a G3BP-NTF2 protein solution resulted in concentration-dependent exothermic heat changes (Fig. 4C, upper panel). The binding isotherm of the integrated heat changes was fitted using a simple one-to-one binding model, yielding an affinity value of 7 μM as well as 2.4 G3BP binding sites per nsP3-25 peptide molecule (Fig. 4C, lower panel). This indicates that one peptide efficiently binds two G3BP molecules. The interaction affinity was 16-fold stronger than the previously determined affinity value of 115 μM for the interaction between a similar G3BP-NTF2 fragment and a DSGFSFGSK peptide [33]. Injection of a control peptide in which both FGDF motifs were exchanged for AGDA (nsP3-25-mut) only produced injection-related heat fluctuations at baseline level (Fig. 4D).
Engagement of two G3BP molecules with each nsP3-25 peptide was further supported by analytical size exclusion chromatography (SEC) analysis. It is known from crystal structures and SEC analysis that the NTF2-like domains of rat and human G3BP form homo-dimers [33]. G3BP-NTF2 alone was eluted with an apparent molecular weight of 30 kDa, slightly lower than the calculated molecular weight of the dimer (36 kDa), while pre-incubation of G3BP with a ten-fold molar excess of nsP3-25 peptide shifted the elution peak, corresponding to the formation of a 60 kDa complex (Fig. 4E). Taking into account the 2:1 ratio for the G3BP-NTF2:nsP3-25 interaction as derived from ITC (Fig. 4C), we suggest that the formed 60 kDa complex comprised four G3BP-NTF2 molecules (two dimers) and two nsP3-25 peptides. There was no evidence for higher-order oligomers.
The crystal structures of the NTF2-like domain of G3BP-1 in its apo form and complexed with the nucleoporin (Nup)-derived peptide DSGFSFGSK have recently been determined [33], revealing that the ligand is bound in an extended conformation within a long and deep groove on the surface of G3BP (S6A Fig.). The peptide-binding site is amphipathic; while the base of the groove is highly hydrophobic, both walls lining the cleft of the protein and the positively charged N-terminus are polar with several basic residues (R32, K5, K123). Both phenylalanine residues of the peptide protrude towards well-defined pockets within the hydrophobic groove. While the side chain of residue F4 of the Nup-derived peptide is buried in a deep pocket formed by the G3BP-1 residues F15, F33 and F124, the side chain of residue F6 is localized in a shallower pocket formed by the G3BP-1 residues F124, V11 and L10 [33].
Our binding and mutagenesis studies reveal that G3BP-1 binds approximately 16 times stronger to FGDF-containing peptides compared to the previously described interaction with the DSGFSFGSK peptide. Both these sequences contain two crucial phenylalanines, which we hypothesized to be bound similarly in both peptides. We assessed this by creating a molecular model in which the octapeptide LTFGDFDE was manually docked into the peptide-binding groove of G3BP-1, using the crystal structure of the G3BP-NTF2 complexed with the DSGFSFGSK peptide as a template (Figs. 5 and S6B). The conformational flexibility of the additional glycine residue within the FGDF motif would allow the phenylalanine side-chains to take similar orientations as in the SGFSF-peptide. Furthermore, the molecular model indicates that the negatively charged residues D5, D7 and E8 in LTFGDFDE could interact with the positively charged side chains of residues K123 and K5 as well as with the positively charged N-terminal region of G3BP (Fig. 5B). It should be noted that these two lysine residues are conserved in human, mouse and Xenopus G3BP-1 and are also present in human G3BP-2 (S7 Fig.). The binding mode suggested by the molecular model of the LTFGDFDE/G3BP-1 complex is in agreement with our biochemical analyses, which showed that mutation of the conserved phenylalanine residues at positions 451, 454, 468 and 471 of nsP3 and 10 and 13 of USP10 eliminated G3BP-binding (Figs. 1 and 3). The model might also explain the elimination of binding of both nsP3 and USP10 to G3BP upon mutation of the glycine residue to alanine, which would restrict flexibility and thus hinder adequate positioning of the phenylalanine side-chains in the FGDF motifs.
In order to validate the proposed binding model, residue F33, localized within the hydrophobic pocket of G3BP and predicted to be proximal to residue F3 of the bound LTFGDFDE peptide, was mutated to tryptophan (G3BP-F33W). Residue F33 is buried at the bottom of the peptide-binding cleft, and we hypothesized that its substitution to tryptophan would reduce the size of the pocket, thus hindering adequate positioning of the benzene ring of peptide residue F3 into the cleft (Fig. 6A, upper panel). As a control, residue F124 was also mutated to tryptophan, since this phenylalanine residue is solvent-accessible and not localized within the binding groove (Fig. 6A, lower panel). Notably, F33 is conserved in human, mouse, Xenopus and Aedes mosquito G3BP-1, while F124 conserved in human, mouse and Xenopus, but a tyrosine residue in the Aedes sequence (S7A Fig.). Both are conserved in human G3BP-2 (S7B Fig.). To determine whether mutation of these residues affects binding of G3BP with FGDF motifs, nsP3 (tagged with a biotin acceptor peptide, BAP) was co-expressed with either pEGFP-C1, pEGFP-G3BP-wt, pEGFP-G3BP-F33W or pEGFP-G3BP-F124W and binding was analyzed by immunoprecipitation with anti-GFP (Fig. 6B). Consistent with our structural model, EGFP-G3BP-wt and the F124W variant efficiently bound to nsP3-BAP, while EGFP-G3BP-F33W did not detectably interact. Similarly to results with cotransfected nsP3, endogenous USP10 was also coimmunoprecipitated with EGFP-G3BP-wt and EGFP-G3BP-F124W but not with EGFP-G3BP-F33W (Fig. 6C).
The extreme N-terminus of G3BP (residues 1–11) forms a large part of the hydrophobic pocket for positioning of the peptide residue F6 and also contains a lysine residue (K5), possibly interacting with the negative charges of acidic residues downstream of the FGDF motif. The model predicts that a truncated version of G3BP (Δ1–11) would not be capable of binding the peptide (Fig. 6D). To evaluate this, 293T cells were transfected with pEGFP-C1, pEGFP-G3BP-wt or pEGFP-G3BP-∆1–11 and analyzed by immunoprecipitation with anti-GFP and immunoblotting for USP10, GFP or actin (Fig. 6E). Endogenous USP10 coprecipitated with EGFP-G3BP-wt but not with EGFP-G3BP-∆1–11. Taken together, the results in Fig. 6 strongly support our structural model of the G3BP/FGDF complex.
Next, we asked whether the FGDF motif, mediating G3BP binding, is present in proteins other than SFV nsP3 and USP10. To identify such proteins, we searched the UniProtKB human and virus databases for all proteins containing FGxF motifs, where x = D, E or S, and also containing at least two acidic residues within the downstream 5 residues (as in both SFV nsP3 and USP10). Glutamic acid was permitted in the third position due to its chemical similarity to aspartic acid, while serine was permitted since Sindbis virus nsP3, shown by several researchers to bind G3BP-1 [25,26], contains a serine at that position. We identified 34 human (S1 Table) and 32 viral (S2 Table) sequences that meet these criteria. In one of those proteins, the infected cell protein (ICP)8 of herpes simplex virus (HSV)-1, the FGDF motif is located between residues 1144 and 1147 in the C-terminal region of the 1196 aa-long protein that is not predicted to adopt any specific secondary structure [34] and contains 3 acidic residues within the downstream 5 residues. This is reminiscent of the context of the FGDF motifs in Old World alphavirus nsP3, an otherwise quite different protein. ICP8 is a single stranded DNA binding protein that is expressed in the lytic cycle of herpes simplex viral replication. It is one of seven viral proteins that are necessary for viral DNA replication [35]. The FGDF motif in ICP8 is conserved in several HSV-1 strains and also in HSV-2 (S2 Table). To determine whether ICP8 binds G3BP-1, 293T cells were cotransfected with expression plasmids encoding HSV-1 ICP8 together with plasmids encoding EGFP-C1, EGFP-G3BP-wt, -F33W or -F124W. Lysates were immunoprecipitated with anti-GFP and immunoblotted for ICP8, GFP or actin (Fig. 7A). The results indicate that ICP8 indeed forms a complex with EGFP-G3BP-wt and EGFP-G3BP-F124W, but not with EGFP-C1 or EGFP-G3BP-F33W. These results are strikingly similar to the binding profiles of SFV nsP3 and USP10 (Fig. 6B and C), and strongly suggest that HSV-1 ICP8 forms a complex with G3BP-1 via binding of the FGDF motif at residues 1144–1147.
Since the binding of nsP3 and USP10 to G3BP inhibits SGs, we hypothesized that ICP8 binding might do the same. To investigate this, BHK cells were transfected with an ICP8 expression plasmid, stressed with sodium arsenite and stained for ICP8 as well as for G3BP-1 and TIA-1 to detect SGs. We observed that in 59% of the transfected cells, ICP8 accumulated mainly in the cytoplasm, while in the remaining 41%, ICP8 was localized in the nucleus. In neither case was there a detectable change in the localization of either G3BP-1 or TIA-1 (S8 Fig.), with these proteins exhibiting diffuse cytoplasmic or nuclear staining, respectively. After sodium arsenite stress however, only 10% of cells with predominantly cytoplasmic ICP8 contained TIA-1 and G3BP-1-positive SGs, while 96% of the cells with nuclear ICP8 and a similar proportion of mock-transfected cells had SGs (Fig. 7B). Representative images are provided in S8 Fig., showing that, when localized in the cytoplasm, ICP8 blocks the induction of SGs after sodium arsenite treatment. Due to the diffuse staining pattern of cytoplasmic ICP8, it is not possible to discern the proportion that is interacting with G3BP, but the inhibition of SG assembly was profound. Taken together with the results in Fig. 7A, this suggests that ICP8, like the cellular protein USP10 and SFV nsP3, interacts with G3BP via its FGDF-motif in a manner which blocks the formation of SGs
This work describes the FGDF motif, shared by proteins from at least two evolutionarily distant viruses and one cellular protein, that mediates strong binding to the multifunctional G3BP proteins. Among its many roles, G3BP is a critical determinant of SG assembly and is one of few proteins whose overexpression induces SGs [2,36]. We have shown that the binding of SFV nsP3, HSV-1 ICP8 and USP10 to G3BP via their FGDF motifs blocks SG formation (Figs. 2, 3D and 7B). The SG-nucleating function of G3BP requires both the RRM domain and the NTF2-like domain [2]. During stress, the RRM domain likely binds to translationally silent mRNAs or stalled 40S ribosomal subunits, and together with other RNA-binding proteins, targets them for SG inclusion. The SGs are then assembled via protein/protein interactions between SG-critical molecules. The NTF2-like domain, which binds the FGDF motifs, is involved in interactions with another critical regulator of SGs, caprin-1 [7]. It is therefore likely that the FGDF-mediated interactions inhibit or alter the ability of G3BP to form complexes with other proteins during the early stages of SG nucleation, resulting in an inhibition of SG formation. This work therefore reveals a mechanism by which USP10 acts as a negative regulator of SG formation. Furthermore, since G3BP appears to be targeted by many different viruses, its neutralization seems to be important for viral pathogenesis. Our work presents a common mechanism whereby Old World alphaviruses and at least two members of the alphaherpesviruses disrupt SG assembly.
Biophysical analyses demonstrated that one 25-mer peptide containing the two FGDF-motifs of SFV nsP3 binds two molecules of G3BP in vitro, suggesting that nsP3 of SFV and the other Old World alphaviruses binds and recruits two molecules of G3BP per nsP3 molecule (Fig. 4). We have previously shown that early in SFV infection, SGs are quickly disassembled in the vicinity of the newly established viral replication complexes [29]. The stoichiometry of the interaction may explain the rapidity of SG disassembly by nsP3, despite its relatively low expression compared to G3BP-targeting proteins of other lytic RNA viruses, such as the picornaviruses. It should be noted that a mutated SFV containing only one FGDF motif (SFV-Δ78) exhibited lower levels of G3BP binding and slower replication kinetics compared to wt SFV, but higher levels of G3BP binding and faster replication kinetics compared to SFV-Δ789, lacking both FGDF motifs [28]. Unlike other proteins with FGDF motifs such as USP10 and ICP8, the Old World alphavirus nsP3 has evolved to comprise two consecutive FGDF motifs to ensure rapid and efficient SG disassembly. This appears to be the main function of G3BP sequestration by nsP3’s FGDF motifs and represents an efficient evasion strategy, beneficial for virus replication.
Herpes simplex viruses block the induction of SGs via multiple mechanisms, highlighting the potent anti-viral effect of SGs. Although early events in viral infection activate PKR, the viral ICP34.5 protein promotes the protein phosphatase 1 (PP1)-mediated dephosphorylation of eIF2α and reactivation of translation [37]. HSV-1 mutants lacking the virion host shutoff (vhs) protein, an endoribonuclease that degrades cellular and viral mRNA, induce SGs late in infection [38,39]. This suggests that vhs may itself have a role in the inhibition of SGs or may alter expression of other SG-modulating viral gene products. HSV-2 also blocks the formation of SGs induced by sodium arsenite but not Pat A [40]. Here, we identify an FGDF motif at the C-termini of the HSV-1 and HSV-2 ICP8 proteins, and demonstrate that HSV-1 ICP8 binds G3BP and blocks SG formation (Fig. 7). Although ICP8 is a predominantly nuclear protein during HSV-1 infection, a sizeable fraction of the protein remains in the cytoplasm [41]. Functions of that fraction are not well described. Our results suggest that the cytoplasmic fraction of ICP8 inhibits SG assembly or other functions of G3BP. It remains to be determined if ICP8 contributes to the inhibition of SGs during HSV infection.
USP10 is a nucleocytoplasmic deubiquitinating enzyme (DUB), originally shown to be regulated by its binding partner G3BP [8]. USP10 deubiquitinates many different proteins of importance in several human diseases and is also a resident SG protein [16,22]. Here, we have shown that overexpression of EGFP-USP10, but not a mutant lacking an intact FGDF-motif, efficiently blocks the formation of SGs (Fig. 3). Interestingly, recent work has shown that the compound resveratrol can inhibit the interaction of G3BP with USP10 by binding to the NTF2-like domain of G3BP, thereby stimulating USP10 DUB activity and stabilization of p53 [17]. It appears therefore that the FGDF-mediated G3BP/USP10 complex is mutually inhibitory, with G3BP inhibiting the DUB activity of USP10 and USP10 inhibiting the SG nucleating function of G3BP. Elucidation of these inhibitory mechanism(s) will require further studies.
A molecular model reveals that the FGDF peptide binds tightly into a hydrophobic groove on the surface of the G3BP NTF2-like domain (Fig. 5). Mutagenesis analyses demonstrated that both phenylalanine residues and the glycine residue are required for binding, with a strong preference for aspartate in the third position.
Both phenylalanine side chains fit snugly within the binding site, the glycine is required for flexibility and the aspartate binds to the G3BP residue K123. Both phenylalanine and the glycine residues are fully conserved in most Old World alphavirus nsP3 sequences and in the USP10 proteins of all higher eukaryotes. The aspartate residue in the third position of the motif is conserved in many of the Old World alphavirus nsP3 sequences except that of Sindbis virus, in which it is a serine. A serine is also found in this position in the Arabidopsis thaliana USP10 gene, while this residue is a glutamate in avian and fish USP10 genes. We note the congruence of our biochemical, phylogenetic and structural analyses and propose that the core binding motif consists of FGxF, in which x can be aspartate, glutamate or serine. While mutation of the aspartate (D7) immediately downstream of the FGDF motif in SFV nsP3 had little effect on binding (Fig. 1B), it is notable that the FGDF motifs in SFV nsP3, USP10 and HSV-1 ICP8 are all followed by at least two acidic residues within the downstream four residues. In our molecular model, we observe that these residues are likely involved in interactions with the basic residues localized the N-terminus of G3BP, which is required for FGDF binding (Fig. 6E) and we propose that they constitute an important part of the motif that further stabilizes the complex. Using these criteria, we present a list of human and viral proteins that contain this motif, and therefore are candidate G3BP-binding proteins (S1 and S2 Tables).
In conclusion, our work describes a motif shared by three otherwise very different proteins, and potentially others, that mediates binding to G3BP and thereby inhibits SG formation. Our three-dimensional model provides a structural understanding of the G3BP/FGDF interaction and will form the basis for the design of pharmaceuticals to target this interaction with a therapeutic potential for a range of viral infections as well as cancers.
Expression vectors for EGFP-nsP3-31 [28], nsP3-BAP [27,42], G3BP-NTF2 [27], and HSV-1 ICP8 [43] were described previously. pEGFP-USP101–40 wt sequences and corresponding alanine mutants were obtained from GeneArt, and ligated between the BglII and EcoRI sites of pEGFP-C1. Construction of the infectious clone pCMV-SFV-F3A: The PCR product derived from primers 1 and 2 (S3 Table) and the PCR product derived from primers 3 and 4 were fused by a one-step PCR in a molar ratio of 1:1. The DNA was denatured and annealed at 46˚C for 2 min. These partially double-stranded molecules were made fully double stranded by extension at 72˚C for 3 min. The fusion DNA was then amplified by using primer 1 and 4 for 25 cycles of PCR consisting of treatment at 95˚C for 30 s, 69˚C for 30 s, and 72˚C for 2 min, followed by a final extension at 72˚C for 5 min. The derived PCR product was purified and subcloned into pTZ57R/T plasmid (Thermo Scientific). The resulting pTZ57R/T-F3A plasmid was digested with XhoI and BglII and religated to the similar digested pCMV-SFV4 vector [44]. The presence of mutations was confirmed by sequencing. Construction of pEGFP-G3BP-F33W and -F124W: The PCR product derived from primers 5 and 6 (F33W) or primers 5 and 7 (F124W) and the product derived from primers 8 and 10 (F33W) or primers 9 and 10 (F124W) were fused by a one-step PCR in a molar ratio of 1:1. The DNA was denatured and annealed at 33˚C for 2 min. These partially double-stranded molecules were made fully double stranded by extension at 72˚C for 3 min. The fusion DNAs containing the F33W mutation or F124W mutation were then amplified by using primers 5 and 10 for 25 cycles of PCR consisting of treatment at 95˚C for 30 s, 56˚C for 30 s, and 72˚C for 2 min, followed by a final extension at 72˚C for 5 min. The derived PCR product was purified and subcloned into pTZ57R/T plasmid (Thermo Scientific). The resulting pTZ57R/T-G3BP1-F33W, -F124W plasmid was digested with BglII and EcoRI and religated to the similar digested pEGFP-C1-G3BP1 vector. The presence of mutations was confirmed by sequencing.
All cell lines were maintained as previously described [28,45,46]. Where indicated, cells were stressed by addition of sodium arsenite (0.5 mM) or pateamine A (100 nM) in complete medium for 60 min. Cells were transfected with Lipofectamine 2000 (Invitrogen) reagent according to the manufacturer’s instructions. Virus titration was performed by plaque assay, as previously described [28]
Immunofluorescence, immunoprecipitations and immunoblotting were performed as described previously (Panas et al., 2012). For details of all antibodies used, see S4 Table.
His-tagged G3BP-NTF2 was expressed in E. coli BL21 T7 Express cells and purified using HisTrap columns (GE Healthcare). Before ITC, G3BP-NTF2 was eluted from a Superdex 75 HiLoad 16/60 (GE Healthcare) SEC column equilibrated in ITC-buffer (25 mM HEPES, 150 mM NaCl, 10 mM MgCl2, 10% glycerol, pH 7.5) at a retention volume of 60 mL. The peptides nsP3-25 and nsP3-25-mut were dissolved in ITC buffer at a concentration of 500 μM and dialyzed extensively against the same buffer. ITC measurements were performed using an ITC200 titration calorimeter (GE Healthcare). The cell temperature was set to 37°C, the reference power to 7 μCal/sec and the syringe stirring speed to 1000 rpm. G3BP-NTF2 was loaded into the cell. The peptides were titrated in 48 injections, each injection with a volume of 750 nL, a duration time of 1.5 sec and a waiting time between the injections of 150 sec. The first injection was performed using a volume of 300 nL, a duration time of 0.6 sec and a spacing time of 120 sec. Background measurements were performed with buffer injected into the protein solution, and peptide into the buffer solution. Data were analyzed using the Origin software as included in the instrument package.
Before analysis, 40 μM G3BP-NTF2 and 400 μM nsP3-25 were mixed and incubated for 1h in SEC-buffer (25 mM HEPES, 300 mM NaCl, 5 mM MgCl2, 10% glycerol, pH 7). A sample volume of 100 μL was injected onto the Superdex 75 10/300 GL column equilibrated in SEC-buffer. G3BP-NTF2 alone was used a control. The apparent molecular weights of the eluted proteins were calculated from their retention volumes as described in the gel filtration LMW/HMW calibration kits assuming similar globular shapes for the analyzed proteins and calibration standards (GE Healthcare).
.
|
10.1371/journal.pcbi.1001061 | PhylOTU: A High-Throughput Procedure Quantifies Microbial Community Diversity and Resolves Novel Taxa from Metagenomic Data | Microbial diversity is typically characterized by clustering ribosomal RNA (SSU-rRNA) sequences into operational taxonomic units (OTUs). Targeted sequencing of environmental SSU-rRNA markers via PCR may fail to detect OTUs due to biases in priming and amplification. Analysis of shotgun sequenced environmental DNA, known as metagenomics, avoids amplification bias but generates fragmentary, non-overlapping sequence reads that cannot be clustered by existing OTU-finding methods. To circumvent these limitations, we developed PhylOTU, a computational workflow that identifies OTUs from metagenomic SSU-rRNA sequence data through the use of phylogenetic principles and probabilistic sequence profiles. Using simulated metagenomic data, we quantified the accuracy with which PhylOTU clusters reads into OTUs. Comparisons of PCR and shotgun sequenced SSU-rRNA markers derived from the global open ocean revealed that while PCR libraries identify more OTUs per sequenced residue, metagenomic libraries recover a greater taxonomic diversity of OTUs. In addition, we discover novel species, genera and families in the metagenomic libraries, including OTUs from phyla missed by analysis of PCR sequences. Taken together, these results suggest that PhylOTU enables characterization of part of the biosphere currently hidden from PCR-based surveys of diversity?
| Microorganisms comprise the majority of the biodiversity on the planet. Because the overwhelming majority of microbes are not readily cultured in the laboratory, researchers often rely on PCR-based investigations of genomic sequence to characterize microbial diversity. These analyses have dramatically expanded our understanding of biodiversity, but due to methodological biases PCR-based approaches may only reveal part of the microbial biosphere. Shotgun sequencing of environmental DNA, known as metagenomics, avoids the biases associated with targeted amplification of genomic sequence and can provide insight into the diversity hidden from traditional investigations. However, the fragmentary, non-overlapping nature of shotgun sequence data makes it intractable to analyze with existing tools. Here, we present PhylOTU, a novel computational method that enables accurate characterization of microbial diversity from metagenomic data. We process over 10 million metagenomic sequences obtained from the global open ocean to identify novel Bacterial taxa and reveal the presence of microorganisms overlooked by investigation of PCR-based sequences from the same samples. These results suggest that to fully characterize microbial biodiversity requires a novel bioinformatics toolbox for analysis of shotgun metagenomic data.
| A central goal of ecology and evolution is to understand the forces that shape biodiversity - the variety of life on Earth. It is becoming increasingly clear that global biodiversity is mostly microbial. It is estimated that there are millions of microbial species on the planet, relatively few of which have been isolated in culture [1]–[2]. Despite the recognized importance of microorganisms, we still know little about the magnitude and variability of microbial biodiversity in natural environments relative to what is known about plants and animals. This is a major knowledge gap, given that microbes are critical components of our planet, responsible for key ecosystems services including the production of agriculturally critical small molecules, the degradation of environmental contaminants, and the regulation of human host phenotypes.
Biodiversity science has traditionally focused on comparing species richness across space, time and environments. Out of necessity, microbial diversity studies usually examine the richness (i.e. number) of operational taxonomic units (OTUs), where OTUs are sequence similarity based surrogates for microbial taxa, which can be difficult to define. In addition to richness, OTUs have been used to characterize the abundance, range, and distribution of microbes, thereby improving our understanding of both natural ecosystems and human health [3]–[6]. OTUs are commonly identified by aligning sequences of the small subunit of ribosomal RNA (SSU-rRNA) from one or more samples and identifying groups of related sequences using a hierarchical clustering algorithm. This clustering is based upon a measure of distance between all pairs of sequences, which is typically defined using some variant of the percent sequence identify (PID) (e.g. [3], [7]–[8]). For example, researchers traditionally cluster sequences that are no more than 3% diverged into the same OTU. This designation has been proposed as being roughly equivalent to a species-level classification [9], though evidence suggests that it may result in an underestimate of the true number of species [10].
The SSU-rRNA sequences for OTU identification are traditionally amplified from a sample via polymerase chain reaction (PCR) using universal primers. Each PCR product is then individually sequenced. One of the biggest drawbacks of this targeted sequencing approach is that it leverages PCR, which has been shown to exhibit sequence-based biases at the level of priming and extension [11]–[13]. In addition, the so-called ‘universal’ PCR primers used in such assays will fail to amplify sequences sufficiently diverged from those used to design the primers. The result is that some taxa may be disproportionately amplified or even missed [14]. Metagenomic approaches eliminate this bias by sequencing randomly-sheared fragments (i.e., shotgun sequencing) of environmental DNA, and, despite having their own sources of bias [15], may therefore provide a potentially more accurate characterization of microbial diversity. For example, the analysis of metagenomic data from a relatively simple microbial community revealed the presence of low-abundance acidophilic Archaea overlooked by PCR-based surveys of diversity [16].
Because of the fragmentary nature of shotgun sequencing, metagenomic reads frequently exhibit minimal, if any, sequence overlap. PID-based evaluations using metagenomic data are thus restricted to the subset of reads that mutually overlap and can therefore be aligned to one another (e.g., [17] and [18]). Alternative approaches have been adopted to describe microbial diversity from non-overlapping metagenomic sequences, including the binning of reads into a reference taxon by comparing each read against reference sequence databases (e.g., [17], [19] and [20]) and using de novo sequence assemblers to build SSU-rRNA contigs (e.g., [21]). While these approaches have substantially advanced the field of microbial biodiversity, they exhibit significant limitations. The former is limited by the diversity encoded in sequence databases, most of which was obtained via targeted sequencing studies. The latter is restricted to the subset of high-confidence assemblies, which can be difficult to produce in many environments given that contig assembly may result in chimeric SSU-rRNA sequences from complex communities [22]). Despite the rapidly growing metagenomic data in microbial ecology and human microbiome studies, no method currently provides a means of characterizing microbial diversity directly from non-overlapping metagenomic data. There is a great need for new approaches that identify OTUs using metagenomic data.
We present PhylOTU, the first method that enables automated identification of microbial OTUs directly from non-overlapping metagenomic sequence reads. PhylOTU leverages a phylogenetic tree of metagenomic SSU-rRNA reads, constructed using probabilistic sequence profiles built from full-length SSU-rRNA sequences from completed genomes, to identify and characterize phylogenetic distances between SSU-rRNA reads in metagenomic data sets. This phylogenetic distance (PD), rather than PID, is then used to cluster reads into OTUs in a fashion similar to that utilized for targeted sequencing data. Because the enormous volume of sequence in most metagenomic libraries presents substantial challenges in the form of sequence-alignment quality and the rate of computational through-put, we developed and implemented within PhylOTU a series of data quality control filters and efficient data structures. We also developed an error rate metric for the analysis of clustered data and used simulated sequences to quantify the accuracy of PhylOTU. These investigations enabled us to derive corrections for biases in phylogenetic methods, producing a tool with similar accuracy to existing PID-based methods. We used PhylOTU to describe microbial diversity in the global open ocean by processing the 10,133,846 shotgun reads in the Global Ocean Survey sequence library [21]. In addition, we compared the OTUs identified by PCR-generated sequences to those identified by shotgun sequences from the same samples. We find that analysis of shotgun sequences reveals a novel part of the biosphere missed by analysis of PCR-generated sequences. PhylOTU is freely available for download at github (https://github.com/sharpton/PhylOTU) and BioTorrents (http://www.biotorrents.net) [23].
Traditionally, OTUs are identified from a PCR-generated targeted sequence library by aligning all pairs of sequences, calculating each pair's PID-based distance, and using this distance to group sequences using agglomerative hierarchical clustering. Due to the fragmentary nature of shotgun metagenomic reads, this traditional approach is limited to the subset of overlapping sequences; non-overlapping reads cannot be directly aligned to one another. Even when reads can be aligned (e.g., to full-length reference sequences), one still cannot calculate PID for sequences that do not overlap. To overcome these limitations, we designed PhylOTU, which uses a probabilistic sequence profile to align reads and a phylogenetic tree to infer their similarity.
The general strategy PhylOTU employs is to leverage full-length reference sequences to construct a probabilistic sequence profile of SSU-rRNA. The profile is used to align metagenomic reads and reference sequences, and this alignment is in turn used to compute the phylogenetic distance between every pair of reads for input into the clustering algorithm. A general workflow schematic of our method is illustrated in Figure 1.
First, probabilistic profiles that encode the evolutionary diversity and secondary structure of the SSU-rRNA sequence from Bacteria and Archaea [24] are constructed via high-quality reference alignments of full-length SSU-rRNA sequence [25]. These profiles are pre-computed for use in different metagenomic analyses. For a given metagenomic data set, SSU-rRNA homologous reads are identified from the shotgun sequencing data via a BLAST search of every metagenomic read against the small but phylogenetically diverse SSU-rRNA STAP databases [26]. This relatively fast search allows one to accurately differentiate SSU-rRNA homologs of Archaea from those of Bacteria, which in turn accelerates and improves downstream alignment and phylogenetic analysis. Multiple sequence alignments of metagenomic reads are created by aligning each SSU-rRNA read to the appropriate Bacterial or Archaeal SSU-rRNA profile, using profile alignment methods [24]. This read alignment is then mapped onto the reference alignment used to build the profile, resulting in a multiple sequence alignment of full-length reference sequences and metagenomic reads. The final step of the alignment process is a quality control filter that 1) ensures that only homologous SSU-rRNA sequences from the appropriate phylogenetic domain are included in the final alignment, and 2) masks highly gapped alignment columns (see Text S1).
We use this high quality alignment of metagenomic reads and references sequences to construct a fully-resolved, phylogenetic tree and hence determine the evolutionary relationships between the reads. Reference sequences are included in this stage of the analysis to guide the phylogenetic assignment of the relatively short metagenomic reads. While the software can be easily extended to incorporate a number of different phylogenetic tools capable of analyzing metagenomic data (e.g., RAxML [27], pplacer [28], etc.), PhylOTU currently employs FastTree as a default method due to its relatively high speed-to-performance ratio and its ability to construct accurate trees in the presence of highly-gapped data [29]. After construction of the phylogeny, lineages representing reference sequences are pruned from the tree. The resulting phylogeny of metagenomic reads is then used to compute a PD distance matrix in which the distance between a pair of reads is defined as the total tree path distance (i.e., branch length) separating the two reads [30]. This tree-based distance matrix is subsequently used to hierarchically cluster metagenomic reads via MOTHUR into OTUs in a fashion similar to traditional PID-based analysis [31]. As with PID clustering, the hierarchical algorithm can be tuned to produce finer or courser clusters, corresponding to different taxonomic levels, by adjusting the clustering threshold and linkage method.
To evaluate the performance of PhylOTU, we employed statistical comparisons of distance matrices and clustering results for a variety of data sets. These investigations aimed 1) to compare PD versus PID clustering, 2) to explore overlap between PhylOTU clusters and recognized taxonomic designations, and 3) to quantify the accuracy of PhylOTU clusters from shotgun reads relative to those obtained from full-length sequences.
We sought to identify how PD-based clustering compares to commonly employed PID-based clustering methods by applying the two methods to the same set of sequences. Both PID-based clustering and PhylOTU may be used to identify OTUs from overlapping sequences. Therefore we applied both methods to a dataset of 508 full-length bacterial SSU-rRNA sequences (reference sequences; see above) obtained from the Ribosomal Database Project (RDP) [25]. Recent work has demonstrated that PID is more accurately calculated from pairwise alignments than multiple sequence alignments [32]–[33], so we used ESPRIT, which implements pairwise alignments, to obtain a PID distance matrix for the reference sequences [32]. We used PhylOTU to compute a PD distance matrix for the same data. Then, we used MOTHUR to hierarchically cluster sequences into OTUs based on both PID and PD. For each of the two distance matrices, we employed a range of clustering thresholds and three different definitions of linkage in the hierarchical clustering algorithm: nearest-neighbor, average, and furthest-neighbor.
To statistically evaluate the similarity of cluster composition between of each pair of clustering results, we used two summary statistics that together capture the frequency with which sequences are co-clustered in both analyses: true conjunction rate (i.e., the proportion of pairs of sequences derived from the same cluster in the first analysis that also are clustered together in the second analysis) and true disjunction rate (i.e., the proportion of pairs of sequences derived from different clusters in the first analysis that also are not clustered together in the second analysis) (see Methods and Figure S1). PhylOTU exhibits high true conjunction and true disjunction rates at commonly employed PID thresholds (e.g., 0.03, 0.06), demonstrating that PD-based clustering accurately recapitulated PID-based clustering at the same threshold (Figure S2).
On the other hand, when applying the same clustering threshold to both distance matrices, PID-based clustering produces a higher richness estimate (i.e., total number of OTUs) than PD-based clustering (Table S1). Comparing the pairwise distance distributions obtained from the PID- and PD- based approaches finds that at relatively short distances (e.g., 0–0.03), PD-based pairwise distances are shorter than the corresponding PID-based distances, while at relatively long distances (e.g., greater than 0.1), PD-based pairwise distances are longer than the corresponding PID-based distances (Figure S3). These findings suggest that differences in richness estimates result from the fact that PD-based clustering tends to merge some clusters that are found to be distinct, but closely related, by PID-based clustering. However, the overall composition of the clusters is very similar: merging of closely related clusters results in a significant reduction in estimated richness, but can produce a relatively small number of conjunction and disjunction errors.
We subsequently investigated whether we could both maintain accuracy of PD-based clustering, while at the same time obtaining richness estimates more similar to PID-based results, which are thought to approximately correspond to the number of distinct microbial taxa in an environmental sample. First, we considered changing the hierarchical clustering algorithm. It has been shown that the choice of nearest-neighbor, average, or furthest-neighbor linkage in hierarchical clustering algorithms results in substantially different estimates of taxonomic richness, with average-linkage clustering performing the best for PID-based approaches [33]. In agreement with these earlier studies, we observed different OTU richness estimates when these three different linkage methods were employed in PhylOTU, with furthest-neighbor clustering producing richness estimates most similar to PID-based clustering for a given threshold (Table S1). But there is a trade-off: employing a different clustering algorithm generally reduces the accuracy with which PhylOTU clusters recapitulate PID-based OTUs, implying that while our estimate for richness may be improved by varying the clustering algorithm, we might be finding the right number of ‘wrong’ OTUs. We reach a similar conclusion if we lower the PD-clustering threshold. We naturally find a greater number of OTUs with a lower threshold, so a threshold that produces a PID-like OTU richness estimate can be identified. However, the accuracy of PD clustering relative to PID clustering becomes systematically lower as the PD threshold deviates from the PID threshold. Given these results, PhylOTU implements average-linkage and a threshold of 0.03 as default settings when clustering full-length SSU-rRNA sequences into OTUs (Table 1).
Overall, our results imply that PhylOTU finds OTUs very similar to PID-based methods in terms of cluster composition, but that recapitulating PID-based clusters with high accuracy will generally result in a lower richness estimate. We consider the accurate clustering of sequences to be more critical than matching OTU richness, given that an equal number of clusters may be optimized between two methods while the accuracy of cluster member composition is simultaneously low. Therefore, we recommend using the default PhylOTU settings, which optimize similarity to PID-based clusters, with the caveat that lower OTU richness estimates may be produced.
Next, we looked at how well PhylOTU clusters full-length sequences relative to taxonomy-guided clusters. We obtained the GenBank taxonomy information for each of the 508 full-length reference sequences and clustered them into taxonomic groups at the species level. We find that PhylOTU clusters sequences into their proper taxonomic group with high true conjunction (96.5%) and true disjunction (99.4%) rates at a clustering threshold of 0.03 (Table S2). However, similar to the results observed in the comparison with PID-based OTUs, PhylOTU tends to underestimate richness relative to GenBank taxonomy. To provide a reference for understanding these results, we conducted a similar comparison of PID-based OTUs and taxonomic groups. PID and PD clustering recapitulate taxonomic groups with similar accuracy at a clustering threshold of 0.03. But, PID clustering produces a slightly closer approximation of richness relative to the taxonomy clusters, consistent with our direct comparison between PhylOTU and PID-based OTUs (Table S2). The similarity between taxonomy and PID-based OTUs is not surprising given the fact many bacterial taxa were defined via PID-based clustering of SSU-rRNA sequences (see Discussion).
To investigate the performance of PhylOTU on metagenomic reads versus full-length sequences, we generated 25 distinct simulation data sets using metaPASSAGE (Riesenfeld et al., unpublished communication), a recently developed, highly parameterized simulation pipeline which expands the function of the MetaSim program [34]. For each simulation, 50 of the 508 reference SSU-rRNA sequences were drawn at random to represent taxa detectable in the sample. These 50 sequences are termed “source sequences” because they are used to generate the simulated metagenomic data. Since most taxa in nature do not have full-length SSU-rRNA sequences in current databases, we used only the remaining 458 non-sampled sequences as the reference sequences for each simulation. We designated the 50 source sequences as full-length PCR products to simulate a targeted sequencing study for each simulated sample. To simulate metagenomic sequencing of the same sample, we generated in silico shotgun reads from the 50 source sequences with a read length distribution chosen to be similar to a 454-sequence library (see Methods). We simulated exactly one read per source and did not simulate sampling or PCR bias to enable direct comparison of full-length and shotgun PhylOTU results. For each in silico sample, we separately applied PhylOTU to the 50 metagenomic reads and the 50 full-length sequences. We used two metrics to quantify the performance of PhylOTU on metagenomic reads: 1) similarity between the read and full-length sequence distance matrices, and 2) accuracy at which the algorithm clusters reads into OTUs relative to clusters built from full-length sequences.
Comparing the PD matrices from metagenomic and full-length data sets, we observe a strong correlation between the pairwise distances computed on reads and full-length sequences. For each of the 25 simulated samples, the read and corresponding full-length-sequence distance matrices show a positive and significant correlation (Mantel test, p<0.05; Figure S4). Having established that pairwise PD measurements are on average similar between metagenomic reads and full-length sequences, we next investigated whether specific properties of individual metagenomic reads systematically generate errors in metagenomic PD estimates compared to full-length PD measurements. We hypothesized that PD error might be higher in shorter reads, which contribute less phylogenetic information than longer sequences, and in reads from hyper-variable regions in the SSU-rRNA locus, which will have higher than expected substitution rates. To explore these hypotheses, we calculated, for each read, a measure of the relative contribution by that read to the total PD error (see Methods). This measure is designed to detect whether certain reads are placed on particularly poorly estimated parts of the phylogeny. We compared this relative error to read length, location within the SSU-rRNA locus (mapped through a read's midpoint position in the multiple sequence alignment), and the amount of alignment overlap the read shares with other reads. We detected no significant correlation between relative PD error and rate variation or alignment depth. We did find a slightly negative, but significant, correlation between relative PD error and read length, suggesting that short reads may contribute more error than long reads (Spearman's rho = −0.088, p = 0.0028). This signal disappeared when reads less than 100 base pairs (bp) were removed from the analysis. As a result, we incorporate a 100 bp read length cutoff in our method. Further analyses are required to comprehensively study the effects of read length and other attributes on PD estimates.
Next, we compared the OTUs produced from metagenomic and full-length sequences, using PhylOTU with identical clustering settings. As illustrated in Figure 2, this analysis reveals that even at low false conjunction rates (meaning that few reads whose corresponding full-length sequences are in separate OTUs are clustered together), PhylOTU tends to correctly put reads from the same OTU in the full-length analysis into the same cluster. This indicates that PhylOTU accurately discriminates between sequence-pair conjunctions: false conjunctions do not need to be tolerated at a high rate to identify true conjunctions. Additionally, PhylOTU clusters reads substantially better than randomly permuting reads into OTU clusters.
We then determined whether the performance of PhylOTU on metagenomic data could be improved by tuning the parameters of the clustering algorithm. Taking the OTUs from full-length sequences at a given clustering threshold as a gold standard, we explored how the true conjunction rate and true disjunction rate vary as functions of the threshold used to cluster the reads. There exists a tradeoff between the true conjunction and true disjunction rates as the threshold changes: at small threshold values, PhylOTU accurately separates reads into distinct OTUs, while at high threshold values, the algorithm accurately clusters sequences into the same OTU (see Figure S5). Maximizing the true disjunction rate subject to a minimum true conjunction rate of 80%, we observe that increasing the read threshold relative to the full-length sequence threshold greatly improves the agreement between the two sets of OTUs. Interestingly, we find a nearly linear relationship between the most accurate read clustering threshold and the full-length sequence threshold (Figure S6). This relationship and the accuracy of PhylOTU remains consistent up to relatively large full-length sequence clustering thresholds (e.g., 0.29, Figure S7). The linear relationship between read and full-length sequence thresholds enabled us to identify adjusted thresholds for metagenomic reads that accurately recapitulate OTUs from full-length sequences (Table S3). PhylOTU obtains 80% accuracy (true conjunction rate = 80%, true disjunction rate = 99.58%) at a read threshold of 0.09, and 90% accuracy (true conjunction rate = 90%, true disjunction rate = 98.73%) at a threshold of 0.18. Thus, simulations enabled us to select tuning parameters of the hierarchical clustering algorithm in PhylOTU so that the OTUs generated from shotgun read data closely resemble those that would be identified if full-length PCR products were available for each SSU-rRNA sequence in the read library.
Given this insight into the accuracy with which PhylOTU clusters metagenomic reads under relatively simple simulation parameters, we evaluated how PhylOTU performs using more rigorous parameters that are reflective of situations encountered during real studies. First, in some environmental samples, the average read may be quite diverged from its closest reference sequence. Second, in many studies the number of reads will be greater than the number of reference sequences. To investigate these two issues, we first used our simulated sequences to evaluate the relationship between the mean phylogenetic distance from each read to its nearest reference sequence (e.g., read-to-reference distance) and the true conjunction rate. We found no significant correlation (Spearman's test). Next, we conducted additional simulations based on sampling reads from full-length Bacterial SSU-rRNA sequences in the SILVA database [35]. This investigation allowed us to generate data sets with more reads than reference sequences and where read-to-reference distances exceeded those in our primary simulations. The latter property is important because of known phylogenetic sampling biases, especially for sequenced genomes [36]. For each of 15 independent simulations, we randomly sampled 1,000 SSU-rRNA sequences from the SILVA database, reflecting the approximate number of SSU-rRNA reads expected when performing one run of next-generation sequencing on a shotgun library. These 1,000 source sequences were then used to simulate metagenomic reads as described above. Reference sequences were pruned from both the source and simulation phylogenies and full-length source sequences and simulated reads were then clustered into OTUs. In these simulations, the average distance between each read and its nearest source is an order of magnitude greater than that observed in our previous simulation analysis (0.182 versus 0.010 mean read-to-reference distance), which is expected given that the SILVA database is highly populated and comprised of phylogenetically diverse sequence data. Evaluating the accuracy of PhylOTU under these conditions reveals high true disjunction rates, similar to those observed in the RDP reference library based simulations. True conjunction rates are somewhat lower, but still meet our accuracy standards. For example, at a read threshold of 0.15, PhylOTU clusters metagenomic reads with an 80% true conjunction rate and a 98.8% true disjunction rate (Figure S8, Table S4), when compared to full-length sequences clustered at a threshold of 0.03 (corresponding to an 86.8% true conjunction rate and a 98.8% true disjunction rate under RDP reference library based simulation parameters). This suggests that read library size and phylogenetic novelty do have a small impact on the accuracy of PhylOTU, but that they can generally be compensated for by appropriately tuning the clustering cutoff.
To demonstrate the utility of PID-based clustering of metagenomic data, we analyzed the pooled Global Ocean Survey (GOS) metagenomic read library [21] with PhylOTU. This data set represents the most extensive publicly available metagenomic sequence library generated to date, with the exception of the Illumina library generated by Qin et. al, which contains reads that are too short to process via PhylOTU [37]. Additionally, many of the GOS sampling sites were also explored with deep, targeted sequencing of the SSU-rRNA locus enabling comparisons of shotgun and PCR libraries. Despite the use of Sanger sequencing, the mean SSU-rRNA metagenomic read length is roughly similar to that used in our simulation analysis (518 bp). Thus, the GOS read library represents the best opportunity to explore PhylOTU's ability to discover novel taxa from metagenomic data. Of the 10,133,846 Sanger sequenced reads in the library, PhylOTU identifies 14,320 Bacterial SSU-rRNA homologs, of which 12,020 passed the method's filters and could be used for OTU discovery. Previous work using the same library was constrained to analysis of 4,125 high-confidence SSU-rRNA assemblies [21], the difference resulting from the fact that many of the SSU-rRNA reads identified by PhylOTU were either assembled in this prior analysis or excluded from this early work given assembly constraints. PhylOTU clusters the 12,020 SSU-rRNA reads into 833 OTUs at a PD threshold of 0.15, which, according to our SILVA-based simulation analysis, corresponds to a full-length threshold of 0.03. Applying a cutoff of 0.09, which was identified as the appropriate corresponding cutoff from the RDP reference library based simulations, identifies 1,078 OTUs. We also identify 192 Archaeal SSU-rRNA sequences, 79 of which pass the quality control filters and cluster into 7 OTUs when using the 0.15 threshold and 10 OTUs when using a threshold of 0.09. This compares to the 811 total OTUs identified by Rusch et. al. via analysis of assembled SSU-rRNA reads at the 97% identity level. We have made our designation of OTUs derived from GOS metagenomic reads and PCR sequences available at BioTorrents [23]. This comparison reveals the ability of assembly-free methods such as PhylOTU to identify novel taxa missed by approaches that rely upon assembled contigs.
The GOS project also generated 6,413 full-length SSU-rRNA sequences via targeted sequencing of PCR products from six of the 73 geographical sites surveyed [38]. We evaluated the ability of PhylOTU to discover novel taxa in shotgun data by comparing the OTUs identified from metagenomic reads to those identified from full-length PCR data from these six sites. We applied PhylOTU to both data sets and corrected for the difference in sequence types by adjusting the read threshold relative to the full-length sequence threshold according to our simulation analysis. Specifically, we used a read threshold of 0.15 and a full-length sequence threshold of 0.03 to evaluate diversity at approximately the species level. We compared the number of OTUs identified per sequence across methods by conducting a rarefaction analysis (Figure 3) [39]. For each method and for subsets of the full data set from one to the observed number of sequences, we drew 100 random subsets of sequences from each data set and calculated the average number of OTUs identified by each method for that number of sequences. This allowed a comparison of the effect of read threshold and sequencing method on the total number of OTUs and rate of OTU accumulation. While there are more PCR SSU-rRNA sequences (N = 6,413) and OTUs (N = 1,563) than metagenomic SSU-rRNA reads (N = 1,233) or OTUs (N = 242), when normalized for the number of sequences in each library, the number of OTUs identified per sequence are similar for the two libraries (0.24 for PCR sequences, 0.20 for shotgun sequences). After normalizing by the average sequence length for each library, however, the shotgun sequence data generates three times as many OTUs per sequenced SSU-rRNA base relative to PCR-generated sequences (4.63×10−4 and 1.66×10−4 OTUs per sequenced base, respectively).
Evaluating the intersection of OTUs identified by the two libraries when they were pooled together and processed by PhylOTU reveals a shared set of OTUs as well as unique OTUs missed by each method (Figure 4). Because this pooled data set contains both full-length sequences and shotgun reads, we evaluated the distribution of sequences across OTUs for a range of thresholds (Figure S9) and made comparisons between OTUs obtained at thresholds appropriate for full-length sequence (0.03) and shotgun reads (0.15). Specifically, at the 0.15 threshold, the metagenomic library contains 80 OTUs that are not revealed through analysis of the PCR library, while the PCR library contains 1,254 unique OTUs at the 0.03 threshold. Normalizing the number of unique OTUs by the number of sequences per library finds that the PCR-based sequences encode more unique OTUs per sequence (0.19) than shotgun sequences (0.06). However, comparing the change in the number of OTUs uniquely identified by shotgun sequence data to the change in the number of OTUs uniquely identified by PCR sequence data across thresholds suggests that shotgun sequences reveal unique OTUs that are highly diverged from those identified using PCR-based sequences (Figure S9). Despite the amount of sequencing conducted, the steep slopes of the rarefaction curves indicate that sampling has not been saturated at these geographical sites. Thus, deeper sequencing through either method is warranted and may either increase or reduce the number of unique OTUs.
We compared the sequences from the novel OTUs identified from metagenomic reads to the Greengenes SSU-rRNA sequence database to determine if any other PCR-based study revealed the existence of these taxa [40]. Using traditional percent identity cutoffs and the Greengenes database as a reference of nearest neighbor percent identity (e.g., DNAML distance), we find that many of the metagenomic read OTUs represent novel species, genera and families. We further characterized the taxonomic distribution of these novel OTUs via taxonomic classification through comparison of the sequences to the RDP database. OTUs unique to the metagenomic reads are predominantly members of the Alpha- (19%) and Gamma-proteobacteria (11%), Actinobacteria (15%), and Bacteroidetes (12%). We also find that the Bacteroidetes, Verrucomicrobia, Firmicutes, and Delta-proteobacteria are enriched in the OTUs unique to shotgun sequences relative to OTUs unique to PCR data or shared between metagenomic and PCR data (Table S5). In addition, several clades, including TM7, Planctomycetes, OD1, and WS3 were only identified via analysis of metagenomic sequence. Reasoning that the universal PCR primers traditionally employed in most targeted sequencing studies (i.e., 8F, 27F, 1525R, 1429R [41]–[43]), may inefficiently amplify or fail to amplify the SSU-rRNA sequences uniquely identified via shotgun sequences, we searched SSU-rRNA reads that overlap the universal priming sites for the presence of sequence complementary to universal SSU-rRNA primers. Of the shotgun reads that overlap a universal priming site (N = 6), we find two that share a unique point mutation relative to the remaining overlapping reads and the 8F and 27F primer sequences (Figure S10). Prior work demonstrated that differences between the primer and template sequences can result in PCR amplification bias [43]. Our findings support the use of universal-primer-sequence variants that include degenerate positions, such as those described in [43], to improve the resolution of lineages harboring this variant through PCR-based investigations. For the remaining reads that do contain a universal priming site, we do not know if the sequence they were generated from contains the anti-sense priming site because these reads do not span the length of the SSU-rRNA locus. Alternatively, these reads may have been obtained from discontinuous rRNA, such as the rRNA sequence found in the mitochondria of Chlamydomonas [44]. Should the priming sites be located in relatively disparate parts of the genome, discontinuous rRNA may fail to amplify even if the universal primer sites are highly conserved.
We have developed a novel method that enables comparison of non-overlapping metagenomic SSU-rRNA reads and their assignment into OTUs. This is the first automated procedure that identifies OTUs directly from non-overlapping metagenomic reads, which facilitates the identification of taxa potentially overlooked by targeted sequencing studies and leverages the vast quantities of shotgun sequencing data currently being produced by environmental and microbiome studies. The key innovation allowing us to compare non-overlapping reads is our use of phylogenetic distance (PD) to cluster reads into OTUs in place of PID. Building a phylogenetic tree requires that at least some of the sequences within the input alignment overlap. Thus, we incorporate high-quality, full-length reference sequences into the SSU-rRNA sequence alignment to guide the phylogenetic placement of metagenomic reads. The accuracy of this approach is constrained, at least in part, by the phylogenetic diversity of the reference sequences and the means by which the phylogenetic algorithm processes missing data. For example, it is challenging to assess distances between non-overlapping shotgun reads derived from a similar place in the phylogeny, even via comparison to full-length reference sequences. We determined the robustness of our method by evaluating the OTU assignment accuracy of simulated metagenomic reads relative to their full-length sources, finding that the relative PD between a pair of reads is on average highly consistent with the relative PD between full-length sources. This result indicates that metagenomic reads can be assigned to OTUs with high accuracy by simply scaling the clustering threshold.
We also tested whether clustering based on PD could accurately recapitulate clustering based on PID for full-length reads where both methods may be applied. Processing 508 full-length reference sequences via both algorithms reveals that PD accurately assigns sequences into OTUs when compared to the PID OTUs. However, this analysis also reveals that PD results in lower richness estimates relative to PID. This phenomenon appears to be due to a difference in the relative distances between sequences. Specifically, the phylogenetic approach appears to shorten the estimated distance between closely related sequences, relative to the PID approach. This is likely due to the fact that the PD approach employs a weighted substitution model when calculating distances, while the PID approach treats all substitutions with equal weight. Thus, while the hierarchical structure of the clusters is generally consistent between the two methods, as revealed by the cluster composition accuracy analysis, sister OTUs in the PID analysis tend to be merged together via the PD approach. For this reason, it may be necessary to take into account this systematic difference in order to compare the diversity results from a PD-based study with a PID-based study.
A similar pattern is observed when the PD-based and PID-based OTUs are compared to OTUs constructed from GenBank taxonomy terms. Specifically, both methods accurately cluster the 508 full-length reference sequences at the species and genus level. Both methods also tend to underestimate the richness, though PID produces an estimate more in line with the taxonomy-guided clusters. Though this analysis serves as a useful benchmark, a more thorough investigation of richness estimation may be warranted in future work for several reasons. First, GenBank taxonomy terms do not necessarily recapitulate the true taxonomic signal or correspond to monophyletic clades. Second, there are known errors in taxonomic assignment and annotation of GenBank sequences [45]–[46]. In addition, many of the taxonomy terms found in GenBank were identified by using the PID approach to classify sequence data. As a result, the reference used in this comparison is necessarily biased towards the PID approach. Regardless, this analysis exemplifies the fidelity with which PhylOTU clusters sequences relative to a commonly adopted interpretation of taxonomy.
Having demonstrated the accuracy with which sequences, both full-length and shotgun, are clustered into OTUs using PD, we applied PhylOTU to the Global Ocean Survey (GOS) metagenomic library. Previous characterizations of SSU-rRNA diversity found in the GOS library were limited to full-length sequences amplified via PCR and full-length contigs produced from high-confidence read assemblies [21]. To demonstrate the ability to discover novel taxa directly from metagenomic data, we compared the PD-based OTUs from full-length PCR sequence to those identified from metagenomic reads. Several conclusions can be drawn. First, targeted sequencing produces more SSU-rRNA sequence per sequenced base (since much of the metagenomic library targets other genes), but fewer OTUs per sequenced SSU-rRNA base compared to metagenomic sequencing. Second, metagenomic sequences analyzed via PD reveal taxa missed by the targeted sequencing study. In particular, PhylOTU clusters metagenomic reads into OTUs belonging to several Bacterial Phyla overlooked by the PCR-generated sequences. We were not able to detect the presence of completely conserved universal PCR priming sites for some of these sequences, which supports the theory that some faction of the microbial biosphere may be hidden from the view of PCR-based investigation. Deeper sequencing of either library could erode the signal of library-specific OTUs. Nonetheless, the distinct taxonomic composition of the metagenomic-only OTUs compared to the shared and PCR-only OTUs (Figure 4, Figure S9, and Table S5) supports the hypothesis that the shotgun libraries would continue to contain unique diversity even after deeper sequencing of both libraries. Thus, we conclude that there are real differences in the identified diversity and composition of these communities depending on the sequencing method employed.
Metagenomic sequencing is an increasingly common means of investigating microbial communities. We expect methods, such as PhylOTU, which enable analysis of unassembled, non-overlapping reads to play an important role in the progress of this field. Future developments will include robust characterization of sources of phylogenetic error to improve methodological accuracy, optimization of PD-based richness estimations in conjunction with optimized cluster composition, and the inclusion of more sophisticated phylogenetic algorithms. Additionally, because the output of PhylOTU includes estimates of abundances for the resulting OTUs, future developments will explore the possibility of using PhylOTU to conduct weighted analyses of community structure by incorporating these abundance estimates. We also anticipate that our phylogenetically-based framework can be expanded beyond its current application to improve OTU identification in several ways, including the incorporation of phylogenetic structure and the utilization of multiple loci when designating of OTUs. When coupled with PCR-based sequencing investigations, this type of bioinformatic analysis of metagenomic data should result in a more comprehensive view of microbial biodiversity.
|
10.1371/journal.pbio.0060255 | Diverse RNA-Binding Proteins Interact with Functionally Related Sets of RNAs, Suggesting an Extensive Regulatory System | RNA-binding proteins (RBPs) have roles in the regulation of many post-transcriptional steps in gene expression, but relatively few RBPs have been systematically studied. We searched for the RNA targets of 40 proteins in the yeast Saccharomyces cerevisiae: a selective sample of the approximately 600 annotated and predicted RBPs, as well as several proteins not annotated as RBPs. At least 33 of these 40 proteins, including three of the four proteins that were not previously known or predicted to be RBPs, were reproducibly associated with specific sets of a few to several hundred RNAs. Remarkably, many of the RBPs we studied bound mRNAs whose protein products share identifiable functional or cytotopic features. We identified specific sequences or predicted structures significantly enriched in target mRNAs of 16 RBPs. These potential RNA-recognition elements were diverse in sequence, structure, and location: some were found predominantly in 3′-untranslated regions, others in 5′-untranslated regions, some in coding sequences, and many in two or more of these features. Although this study only examined a small fraction of the universe of yeast RBPs, 70% of the mRNA transcriptome had significant associations with at least one of these RBPs, and on average, each distinct yeast mRNA interacted with three of the RBPs, suggesting the potential for a rich, multidimensional network of regulation. These results strongly suggest that combinatorial binding of RBPs to specific recognition elements in mRNAs is a pervasive mechanism for multi-dimensional regulation of their post-transcriptional fate.
| Regulation of gene transcription has been extensively studied, but much less is known about how the fates of the resulting mRNA transcripts are regulated. We were intrigued by the fact that while most eukaryotic genomes encode hundreds of RNA-binding proteins (RBPs), the targets and regulatory roles of only a small fraction of these proteins have been characterized. In this study, we systematically identified the RNAs associated with a select sample of 40 of the approximately 600 predicted RBPs in the budding yeast, Saccharomyces cerevisiae. We found that most of these RBPs bound specific sets of mRNAs whose protein products share physiological themes or similar locations within the cell. For 16 of the 40 RBPs, we identified sequence motifs significantly enriched in their RNA targets that presumably mediate recognition of the target by the RBP. The intricate, overlapping patterns of mRNAs associated with RBPs suggest an extensive combinatorial system for post-transcriptional regulation, involving dozens or even hundreds of RBPs. The organization and molecular mechanisms involved in this regulatory system, including how RBP–mRNA interactions are integrated with signal transduction systems and how they affect the fates of their RNA targets, provide abundant opportunities for investigation and discovery.
| Much of the regulation of eukaryotic gene expression programs is still unaccounted for. Although these programs are subject to regulation at many steps, most investigation has focused on regulation of transcription. There are clues, however, that a significant portion of undiscovered regulation might be post-transcriptional, acting to regulate mRNA processing, localization, translation, and decay [1–5]. For example, systematic phylogenetic comparison among yeast and mammalian genomes sequences have revealed that untranslated regions of many mRNAs are under purifying selection, and thus presumably carrying information important for fitness [6–8].
Biological regulation can be achieved by controlling any of a large number of steps in the lives of RNA molecules. Alternative splicing of transcripts can enable a single gene to encode numerous protein products, greatly expanding its molecular complexity [9]. Even in organisms with few introns, such as Saccharomyces cerevisiae, splicing is subject to regulation [10,11]. Notable examples of regulated RNA localization include mRNA export from the nucleus to the cytoplasm, partitioning of mRNAs to the rough endoplasmic reticulum (ER) membrane for cotranslational export, and the precise subcellular localization of thousands of specific mRNAs [12]. In a recent survey of mRNA localization in developing Drosophila embryos, more than 70% of the roughly 3,000 mRNAs examined showed distinct patterns of subcellular localization [13]. Widespread regulation of translation rates is evident in several observations. In yeast, despite extensive regulation of transcription and mRNA decay, only about 70% of the observed variance in protein abundance is accounted for by variation in mRNA abundance [14,15]. When cells are moved from rich media to minimal media, the abundance of hundreds of proteins change, but mRNA abundance changes parallel changes in the abundance for only about half of the cognate proteins [16,17]. The abundance of each RNA is determined jointly by regulated transcription and regulated degradation. Widespread, transcript-specific regulation of mRNA decay is evident from the closely matched decay rates of mRNAs encoding functionally related proteins [18–21], particularly evident in S. cerevisiae in sets of proteins that form stoichiometric complexes [19].
Increasing evidence points to extensive involvement of specific RNA-binding proteins (RBPs) in regulation of these post-transcriptional events [1–5]. Pioneering studies focusing on tens of predominantly nuclear mRNA RBPs (so-called heterogeneous ribonucleoprotein [hnRNP] proteins), revealed that these proteins recognize specific features in mRNAs, bind at overlapping, but distinct, times during RNA processing, and differentially associate with subsets of nascent transcripts [22]. Steps in RNA processing in the nucleus are functionally and physically coupled, providing an opportunity for coordinated control [23].
Investigations of regulation acting on RNA have usually focused on a few model RNAs, leaving unanswered the extent to which mRNAs are coordinated and differentially regulated, and this regulatory landscape is still largely unexplored. Recent studies have systematically identified the suite of mRNAs associated with some individual RBPs. Several RBPs implicated in RNA processing and nuclear export in S. cerevisiae were found to associate with distinct sets of hundreds of functionally related mRNAs [24,25]. Five members of the Puf family of RBPs in S. cerevisiae were each found to associate with distinct, overlapping sets of 40–250 mRNAs [26]. The specific sets of mRNAs associated with each Puf protein were significantly enriched for mRNAs encoding functionally and cytotopically related proteins. For instance, most of the approximately 220 mRNAs associated with Puf3 are transcribed from nuclear genes and encode proteins localized to the mitochondrion (p < 10−100). Puf3, Puf4, and Puf5 each recognize specific sequences in the 3′-untranslated regions (UTRs) of their targets. These results and others, from studies of a few selected RBPs, may be just a glimpse of a much larger and richer post-transcriptional regulatory network, involving dozens to hundreds of RBPs and a cognate suite of recognition elements in their RNA targets (e.g., [22,24–40]).
But does such a multidimensional post-transcriptional regulatory network exist? To test this hypothesis and to extend and deepen our understanding of RBP–RNA interactions, we systematically searched for the RNA targets of a select sample of 40 out of the more than 500 known and predicted RBPs in S. cerevisiae.
We first developed a list of candidate RBPs based on annotations in the Saccharomyces Genome Database (SGD) (http://www.yeastgenome.org), the Yeast Protein Database [41], and the Munich Information Center for Protein Sequences database [42] and on literature searches. From the assembled list of 561 genes (Table S1), we chose a set of 36 with diverse RNA-binding domains and diverse functional annotations (Table S2 and Text S1). Because many known RBPs lack recognizable RNA-binding domains, we also included two metabolic enzymes whose homologs in other species are known to associate with RNA, and two proteins that were not, a priori, expected to bind RNA, but which we suspected might have post-transcriptional regulatory functions (Table S2).
To identify RNAs associated with each putative RBP, C-terminal tandem affinity purification (TAP)-tagged proteins, expressed under control of their native promoters, were affinity purified from whole-cell extracts of cultures grown to mid-log phase in rich medium [14,26,43]. Extracts were incubated with immunoglobulin G (IgG) agarose beads, washed, and ribonuclear protein complexes were eluted by tobacco etch virus (TEV) protease treatment (Text S2). We performed two to four independent isolations with each tagged strain. As controls, we performed 13 immunoaffinity purifications (IPs) of untagged strains to identify and exclude potential false-positive RNA targets.
We purified total RNA from the whole-cell extracts and TEV-purified fractions, reverse transcribed with an amino-allyl-dUTP/dNTP mix, coupled the purified cDNA to Cy3 and Cy5 dyes, respectively, mixed the two differentially labeled cDNA pools, and then hybridized them to DNA microarrays (Dataset S1).
We identified RNAs specifically associated with each protein using the significance analysis of microarrays (SAM) algorithm [44]. Although it is not possible to perfectly distinguish targets from nontargets, and the best criterion for distinguishing targets from nontargets is unlikely to be the same for all proteins, for most proteins, we chose a 1% false discovery rate (FDR) as a criterion for identifying targets (Datasets S2 and S3). For many RBPs, the number of RNAs called significantly enriched has an inflection point near 1% FDR, suggesting that this threshold is a good balance between sensitivity and specificity, but undoubtedly our identification of specific RBP targets is not comprehensive. For two proteins in the survey (Ssd1 and Khd1), we used a more stringent 1% local FDR criterion [45] (details in Materials and Methods; Datasets S2 and S3). We also included mRNAs specifically associated with Puf1–5 from our previous work [26], (defined using a 1% local FDR), and previously identified She2 targets [32].
The 40 proteins in the survey (and also Puf1–5 and She2 from our previous work [26,32]) displayed diverse patterns of specificity with regard to the numbers and types of RNA targets and their enrichment profiles (Figures 1 and S1, and Text S3). The number of confidently identified RNA targets varied widely among the proteins surveyed, ranging from fewer than ten (Nce102, Nrp1, Idh1, Rib2, Nop13, Bud27, Rna15, Pbp2, Dhh1, Upf1, and Mex67) to more than a thousand (Pab1, Pub1, Scp160, Npl3, Nrd1, and Bfr1) (Figure 1A). The two “negative controls,” Nce102 and Bud27, were each associated with specific RNAs. Nce102 was associated with eight distinct RNAs, whereas Bud27 was associated with two putative mRNA targets; interestingly, one of these putative targets (RPA190) was reproducibly enriched more than 300-fold, and both targets were lost when immunopurifications were performed in the absence of Mg2+ (unpublished data). Because neither Nce102 nor Bud27 was known or expected to associate with RNA, the RNAs identified as their targets may be spurious, but we cannot exclude the possibility that the RNA interactions we found for these two proteins are real and significant. Regardless, they provide a benchmark estimate of the number of RNA targets falsely identified for other RBPs. Aconitase (Aco1) and glyceraldehyde-3 phosphate dehydrogenase (Tdh3), two metabolic enzymes whose human orthologs also function as RBPs [46,47], but which were not previously known to be RBPs in yeast, associated with 38 and 155 RNAs, respectively, at 1% FDR, indicating that these enzymes are also RBPs in yeast.
Fourteen of the proteins we surveyed specifically associated with RNAs other than mature mRNAs encoded by nuclear genes (Figure S2). Their specific targets included intron-containing transcripts (Cbc2, Msl5, Npl3, Hrb1, Pab1, and Pub1), H/ACA box small nucleolar RNAs (snoRNAs) (Cbf5, Nrd1, and Pub1), C/D box snoRNAs (Nop56, Sof1, Nab3, Nrd1, Pub1, and Pab1), and mitochondrial mRNAs (Aco1, Tdh3, and Nab2). Several of these proteins have previously been shown to be associated with specific classes of RNA (Cbc2, Msl5, Npl3, Cbf5, Nrd1, Nop56, Sof1, and Nab3), and therefore provide de facto positive controls (Table S2 and Text S4). Aco1, a TCA cycle enzyme [48], which has recently been implicated in maintaining mitochondrial genome integrity [49], selectively binds transcripts encoded by the mitochondrial genome (p < 10−38). Our results also suggest unexpected associations for several noncoding-RNA–binding proteins and suggest possible regulatory links between mRNA and noncoding RNA (ncRNA) processing (Text S4). However, the remainder of this report will focus mostly on mRNA targets.
To explore the interrelationships among RBPs and their RNA targets, we organized RNAs (Figure 1B, columns) and RBPs (Figure 1B, rows), respectively, by hierarchical clustering based on their patterns of mutual interactions, and visualized the results as a heat map representing the confidence of an RNA–RBP interaction with a black (>10% FDR) to yellow (0% FDR) scale. For the most part, each RBP had a unique profile of enrichment, with a few notable exceptions, including Scp160/Bfr1 and Nrd1/Nab3, which are pairs of proteins that act together in stable stoichiometric complexes [50,51] and were correspondingly associated with similar sets of mRNAs.
Altogether, we identified more than 12,000 mRNA–RBP interactions (at a 1% FDR), an average of at least 2.8 RBPs interacting with each of 4,300 distinct mRNAs; 31 proteins (including Puf1–5 and She2) reproducibly bound at least ten mRNAs (at a 1% FDR). Most mRNAs were bound by multiple RBPs (Figure 1C, black bars); 628 mRNAs were bound by five or more of this set of 31 RBPs; intriguingly, a disproportionate fraction of the mRNAs with the greatest number of identified interactions with this set of RBPs encode proteins localized to the cell wall (31, p < 10−4).
About 75% (∼9,000) of the mRNA–RBP interactions identified in this survey were accounted for by the nine proteins that targeted more than 500 mRNAs each, (Figure 1C, grey bars). Our conservative approach to target identification, emphasizing specificity over sensitivity, probably underestimates the number of targets of these broad-specificity RBPs; some of these proteins, such as Scp160 and Pab1, probably bind most or all mRNAs (Figure S1 and Text S3). The specificity and regulatory contributions of these “general” RBPs are still poorly understood.
Regulatory proteins, including both transcription factors and RBPs, typically regulate sets of targets that share identifiable functional relationships (e.g., [26–29,32,35,52–60]). As a first step toward identifying relationships among RNAs bound by specific RBPs, we searched for gene ontology (GO) terms [61] that were significantly enriched among the targets of each RBP. Twenty-five of the RBPs in this survey were consistently associated with at least ten mRNAs; 13 of these sets of RNA targets specific to an RBP were significantly enriched for at least one “cellular component” GO term (Figure 2A and Table S3), representing a shared subcellular localization or in some instances a protein complex, and 13 of these RBP-specific target sets were significantly enriched for at least one “biological process” GO term (Figure 2B and Table S3).
Diverse subcellular loci and biological processes were represented among the annotations enriched in the sets of RNA targets of these 15 RBPs (as well as the five Puf proteins and She2), including nearly all major subcellular compartments. Some subcellular sites and biological processes were found as shared attributes of the RNA targets associated with an unexpectedly large fraction of the RBPs in this study, perhaps highlighting processes or systems in which post-transcriptional regulation plays an especially important role. For instance, six RBPs (Pub1, Khd1, Nab6, Ssd1, Ypl184c, and Scp160) were specifically associated with mRNAs encoding cell wall proteins; six (Pub1, Puf1, Puf2, Khd1, Ypl184c, and Scp160) were specifically associated with mRNAs encoding plasma membrane proteins; five (Puf3, Nsr1, Pab1, Npl3, and Nrd1) were significantly associated with mRNAs encoding subunits of mitochondrial ribosome; and four (Scp160, Bfr1, Puf4, and Gbp2) were specifically associated with mRNAs encoding proteins localized to the nucleolus and involved in RNA processing and ribosome biogenesis.
For many RBPs, several distinct subcellular components or biological processes were overrepresented in the functional annotations of the associated transcripts; these subcellular loci or processes were often functionally linked. For example, RNAs associated with Ssd1 were enriched for transcripts encoding cell wall and bud proteins, whereas Gbp2-associated RNAs were enriched for transcripts encoding nuclear proteins with roles in ribosome biogenesis or chromatin remodeling. In many instances, the functional themes significantly overrepresented among the RNA targets of an RBP are congruent with previously published work on that RBP, such as phenotypes associated with mutation of altered expression (Table S2). A few examples are described in subsequent sections.
Although some appear to bind to most or all mRNAs (Figure S2 and Text S3), the nine RBPs that bind large (>500) sets of mRNAs display several distinct enrichment profiles (Figure 1B), with correspondingly different GO annotations overrepresented among the most highly enriched mRNAs (Figure 2). In addition, for each of these nine RBPs, immunoaffinity enrichment of mRNAs with the RBP was significantly correlated with either ribosome occupancy [62], abundance [19], half-life [19], 3′-UTR length [63], 5′-UTR length [63], mRNA length [63], coding sequence length, or in some cases, with more than one of these features (Figure S3). Quantitative differences in the enrichment of mRNAs in association with a given RBP could result from the number or affinity of the RBP molecules bound or differences in the fraction of its lifespan that an individual mRNA spends at the specific stage during which a particular RBP plays a role (Text S5).
Pab1 provides a simple and useful example of the possible functional significance of the differential enrichment; immunoaffinity enrichment of mRNAs associated with Pab1 was correlated with ribosome occupancy (Pearson correlation = 0.35). Pab1 is the major poly(A) binding protein in both the nucleus and cytoplasm [64]. In the cytoplasm, Pab1 binds to the poly(A) tails of mRNAs and interacts with eIF4-G to promote translation initiation [65]. Because longer poly(A) tails have been reported to increase translation efficiency [66], a possible interpretation of these results is that the observed enrichment could reflect the number of Pab1 proteins bound per mRNA and thus the length of the poly(A) tail [39].
In contrast, immunoaffinity enrichment with Khd1 was negatively correlated with ribosome occupancy (r = −0.26). Khd1 is implicated in repressing translation of ASH1 mRNA during the transport of the mRNA to the bud tip [67]. The negative correlation with global ribosome occupancy and the large number of mRNAs associated with Khd1 suggest that Khd1 may similarly repress translation initiation of hundreds to thousands of mRNAs, perhaps during their transport to specific cellular loci.
Many RBPs associate with mRNAs at a particular stage in their lives [2]. For the approximately 270 intron-containing genes, the relative enrichment of introns (i.e., unspliced pre-mRNAs and possibly uncleaved excised introns) versus exons (i.e., mature mRNAs and pre-mRNAs) should reveal whether the RBP is bound specifically to intron-containing transcripts, mature mRNAs, or both, and thus indicate when and where the RBP associates with its target RNAs. Linking these data to functional information on the RBP could then provide insights into timing and duration of specific stages in the lives of mRNAs.
To test this idea, we compared the enrichment of intron and exon sequences in association with RBPs. For the approximately 120 intron/exon probe pairs for which our data were most consistently reliable, the relative enrichment profiles vary greatly among RBPs (Figure 3 and Text S6). For example, Cbc2 (a component of the heterodimeric nuclear cap-binding protein) and Pab1 were preferentially associated with both intron-containing transcripts and mature mRNAs derived from intron-containing transcripts (Figure 3). Cbc2 was strongly associated with intron-containing transcripts (mean enrichment of intronic sequences = 6.8), and also, but to a considerably lesser extent, with exon sequences from intron-derived transcripts (mean enrichment of exonic sequences = 1.5). These results are consistent with Cbc2 binding during transcription, prior to splicing, and being displaced shortly after the mature mRNA reaches the cytoplasm [68,69]. The enrichment of intron-related transcripts and the paucity of significantly enriched mature mRNAs suggest that most mRNAs spend only a very small fraction of their lives in the nucleus. That Pab1, the major poly(A) binding protein, associated with intron-containing transcripts (mean enrichment of intronic sequences = 1.5), as well as sequences from exons (mean enrichment of exonic sequences = 3.9), is consistent with most splicing occurring after poly(A) tail addition [70].
The RBPs we analyzed bound overlapping sets of mRNAs, and many individual mRNAs were bound by more than one RBP (Figure 1B and 1C). This network of interactions could support a robust and multidimensional regulatory program.
To explore the relationships among the groups of RNAs bound by different RBPs, we determined the extent to which the overlaps between targets for each RBP pair differed from what would be expected by chance. The significance values from this analysis were used as a metric of similarity for hierarchical clustering to identify pairs and sets of RBPs with similar patterns of shared targets. The results are presented in Figure 4A as a heat map, in which the similarity between the target sets of each pair of RBPs is shown on a blue (significantly fewer shared targets than expected, p = 10−25) to white (p > 0.001) to red (significantly more shared targets than expected, p = 10−25) scale. At a p-value threshold of 0.001, 69 of 465 RBP pairs shared significantly more mRNA targets than expected by chance, whereas 11 RBP pairs shared significantly fewer mRNA targets than expected by chance. Several of the most significantly overlapping target sets belong to sets of RBPs that are known to physically interact, such as Scp160 and Bfr1 [50], Nrd1 and Nab3 [51], Nrd1/Nab3 and Npl3 [71], and Nrd1/Nab3 and Pab1 [72].
To further explore the interrelationships among RBPs and their mRNA targets, we used a supervised method to identify smaller subsets of mRNAs that shared interactions with several RBPs. We did this by selecting mRNAs bound by a common set of RBPs whose targets, in turn, were enriched for common GO terms (Figure 2).
The group of mRNAs, defined by interactions with at least four of a set of six RBPs (Pub1, Khd1, Nab6, Ssd1, Ypl184c, and Scp160), includes a significant excess of mRNAs encoding proteins localized to the cell wall (Figure 4B); indeed, 23 of the 78 mRNAs in this cluster encode cell-wall proteins (p < 10−19). This group also contains mRNAs that encode proteins that are secreted (5), localized to sites of polarized growth (4), or localized to the ER (14). It is important to recognize that the unifying theme in this group is not narrowly restricted to simple functions in cell-wall metabolism—many mRNAs in this group encode proteins with diverse roles in regulation of cell-wall metabolism. Fifteen mRNAs encode proteins involved in post-transcriptional regulation, including SSD1, DHH1, and PUF5, which are genetically implicated in cell-wall biogenesis and maintenance [73,74], and NGR1 and WHI3, which are involved in control of cell growth [75–77]. Fourteen of these mRNAs encode proteins involved in transcriptional control, including SFL1, which is implicated in cell-wall assembly [78], and NDD1, YOX1, and NRM1, which are involved in cell-cycle control [79–81]. Seven mRNAs encode signal transduction proteins, including MFA2, CLN2, GIC2, WSC2, and MSB2, which are implicated in cell-wall growth or cell-cycle regulation [82–88].
We identified candidates for the sequence elements that mediate regulatory interactions with specific RBPs using two related computational methods: “finding informative regulatory elements” (FIRE), which searches for motifs with informative patterns of enrichment [89], and a newly developed method, “relative filtering by nucleotide enrichment” (REFINE). In brief, REFINE identifies all hexamers that are significantly enriched in putative 5′- and 3′-UTR regions of targets over nontargets, filters out regions of target sequences that are relatively devoid of such hexamers, and then applies the “multiple expectation maximization for motif elicitation” (MEME) motif-finding algorithm [90]. A full description of the REFINE methodology and more detailed analyses of predicted motif sequences will be published separately (D. P. Riordan, D. Herschlag, and P. O. Brown, unpublished data). Herein, we combined the results from these two approaches.
Using stringent statistical criteria based on randomized simulations (details in Materials and Methods), we identified a total of 60 candidate RNA regulatory motifs significantly associated with 21 different RBPs; 35 motifs (for 21 RBPs) were predicted by REFINE, and 25 motifs (for 13 RBPs) were predicted by FIRE (Table S4). Since the same motifs were often predicted by both programs for the same RBP or for different RBPs with significantly overlapping target sets, we manually grouped motifs with similar consensus sequences and origins into classes (Table S4). We then included only the most significant motif from each class and for each RBP, resulting in a set of 14 nonredundant RNA motifs predicted with high confidence (Figure 5). We also evaluated the predicted RNA motifs by testing whether motif sites occurring in targets were more likely to be conserved than sites in nontargets, and whether they exhibited a forward strand bias by testing for significant enrichment of the reverse complementary motif in RBP targets (Table S4).
The motifs we identified for Puf3, Puf4, Puf5, Pub1, Nab2, Nrd1, and Nab3 match previously described binding sites for the corresponding RBPs, validating our approach and suggesting that many of the RBP–RNA interactions we measured are likely to be directly mediated by these elements (Text S7). Interestingly, the inferred recognition element for Nrd1, Nrd1–1 (UUCUUGUW), contains both an exact match to the reported Nrd1 binding site consensus “UCUU” and a partial match to the reported Nab3 recognition site consensus “GUAR” [91,92]. As Nrd1 and Nab3 are known to act as a complex to control transcriptional termination of nonpolyadenylated RNAs [93], and a nearly identical motif was identified in Nab3 targets (Table S4), it is possible that these motifs represent a favored orientation of adjacent Nrd1 and Nab3 RNA elements that facilitates specific binding of the Nrd1–Nab3 complex.
The most significant novel motif we identified, Puf2–1 (UAAUAAUUW), is enriched in the 3′-UTRs and coding sequences of Puf2 targets and demonstrates significant conservation and a forward strand bias (Figure 5). This motif is similar to a motif identified for the paralogous RBP Puf1, which associates with a subset of the Puf2 target mRNAs (Table S4). The next most significant novel motif, Ssd1–1 (AKUCAUUCCUU), is highly enriched in the 5′-UTRs of Ssd1 targets (Figure 5). Although its presence upstream of the coding sequences of Ssd1 target genes would also be consistent with a role as a transcription factor binding site, its tendency to occur within the annotated 5′-UTRs of targets (63% targets versus 19% nontargets, p < 10−6) [94], its dramatic enrichment in targets, and its forward strand bias suggest that this RNA motif is recognized by Ssd1.
A selective sample of 11 mRNAs provides an unfinished, but revealing, picture of the organization of the information that specifies interactions with, and perhaps regulation by, specific RBPs examined in this study (Figure 6). For each mRNA, the location of high-confidence RNA recognition elements for RBPs that interact with the mRNA are indicated, while RBPs that interact with the mRNA, but whose binding site is uncertain, are shown to the right of the mRNA. The relative lengths of the 5′-UTR, coding sequence, and 3′-UTR are drawn to scale, and the translation start and stop codons are depicted with the corresponding “traffic signal.” Each of these mRNAs has specific interactions with overlapping, but distinct, subsets of RBPs in the study. The putative binding patterns of specific RBPs, with respect to the number and locations of sites, vary considerably among the mRNAs, which may have important functional consequences. The first five mRNAs (SUN4, DSE2, CTS1, SCW4, and EGT2) encode cell-wall enzymes (Figure 6A–6E). Each of these mRNAs associated with five to nine RBPs in this study, including all five with Pub1, Khd1, and Ypl184c, four with Ssd1 (SUN4, DSE2, CTS1, and SCW4), three with Scp160 (CTS1, SCW4, and EGT2), and two with Nab6 (CTS1 and SCW4) and Nrd1 (DSE2 and EGT2). In addition to these overlapping interactions, most of these mRNAs associated with a unique set of additional RBPs; for instance, SUN4 contains two Puf5-binding sites in its 3′-UTR and EGT2 contains eight She2-binding sites in its coding sequence. CLN2 encodes a G1 cyclin and associated with many of the same RBPs as SUN4, DSE2, CTS1, SCW4, and EGT2 (Figure 6F). PUF2 associated with several RBPs, including its cognate protein, which is common among RBPs in this study (Text S8); there are 12 Puf2-binding sites in its coding sequence (Figure 6G). PMA1 associated with a similar set of RBPs as PUF2, including Pub1 and Puf2, but the locations and numbers of binding sites for these RBPs are very different in the two mRNAs (Figure 6H). The putative binding sites for Puf4 and Puf5 in the 3′-UTR of HHT1 partially overlap, suggesting these RBPs may compete for binding to this mRNA (Figure 6J). These diagrams represent only a partial picture of the RBP interactions with these mRNAs; the mRNA targets have only been defined for a small fraction of all yeast RBPs, and the sequence elements that specify many of the interactions we have identified are not yet known.
For many RBPs, our computational method did not identify any sequence motifs with statistically significant enrichment, the motifs identified significantly overlapped those associated with other RBP target sets, or the motif did not match previously reported binding preferences (Table S4 and Text S7). The large degree of motif coenrichment observed in our analysis is consistent with combinatorial regulation by a highly interconnected regulatory network and represents an important limitation of computational regulatory element identification. It is likely that some of the RBPs for which we failed to predict sequence motifs recognize RNA structural elements or features primarily present in coding sequences, which are difficult to detect with current methods for RNA motif prediction, because they are not suited to modeling structural features or handling the significant confounding sequence biases in coding sequences.
Vts1 illustrates some of the limitations of current RNA motif prediction methods. Vts1 is known to bind to a structural RNA motif called the Smaug recognition element (SRE), which consists of a short hairpin with the loop consensus sequence CNGGN(0–1) [95]. SRE sites are indeed significantly enriched in the coding sequences of Vts1 targets (65% targets versus 36% nontargets, p < 10−7) in agreement with previous results [96], suggesting that SRE elements are directly responsible for these interactions in vivo. However, neither REFINE nor FIRE succeeded in identifying the SRE. Instead, both programs identified a motif, Vts1–1 (UKWCGRGGN), which is indeed enriched in the 3′-UTRs of Vts1 targets but is unrelated to the SRE (Table S4). We suspect that the Vts1–1 motif may represent a binding site for an unknown factor that regulates a set of mRNAs that overlaps extensively with the targets of Vts1.
It is likely that direct high-resolution mapping of in vivo RBP binding sites and systematic in vitro characterization of binding preferences of RBPs will overcome some of the limitations in current methods for RNA motif identification [97,98].
The functional and cytotopic themes represented among the specific targets of each RBP have obvious implications for their possible regulatory roles, which can be integrated with previously reported information to derive further insights, and generate new hypotheses, as illustrated here for Ssd1 and Ypl184c (see Text S9 for descriptions of Khd1 and Gbp2).
Ssd1 is a large (140 kDa), ribonuclease-II domain–containing, predominantly cytoplasmic protein [99], genetically implicated in cell-wall biogenesis and function: mutant phenotypes include increased sensitivity to osmotic stress and caffeine, altered composition and structure of the cell wall, defects in germination and sporulation, premature aging, and pathogenicity [73,74,100–103]. Ssd1 physically and genetically interacts with numerous signaling proteins, many of which are genetically implicated in cell-wall function [71,102,104,105]. Ssd1 binds to the C-terminal domain of RNA polymerase II in vitro [106].
Of the 52 annotated mRNAs associated with Ssd1, 16 encode proteins localized to the cell wall (p < 10−15), and 11 encode proteins localized to the bud (p < 10−5). The proteins encoded by the Ssd1-associated transcripts have diverse functional and structural roles related to cell-wall biosynthesis, or remodeling and its regulation, cell-cycle progression, and protein trafficking. Ssd1 also appears to bind its own transcript (Text S8).
For both of the Ssd1 mRNA targets encoded by intron-containing genes (PUF5 and ECM33), the intron-containing primary transcripts are also enriched by Ssd1 IP, suggesting that Ssd1 binds its RNA targets in the nucleus, perhaps while they are being transcribed. A putative RNA-recognition motif is significantly enriched in the 5′-UTRs of Ssd1 targets (Figure 5). The numbers and positions of this motif in Ssd1-bound RNAs vary widely among its targets (Figure 6A–6D and 6F). These data lead us to speculate that Ssd1 binds its targets cotranscriptionally by recognizing a specific RNA motif and prevents their translation initiation until these mRNAs reach specific locations in the cell, such as the ER membrane, bud, or sites of cell-wall biosynthesis. The multiple phosphorylation sites on Ssd1 could regulate the localization, binding, and release of its RNA targets. Although Ssd1 is a ribonuclease-II domain–containing protein, it has no discernable nuclease activity [99]. Given that Ssd1 does not contain any other known RNA-binding domains, we suggest that the ribonuclease-II domain may have evolved into a sequence-specific RNA-binding domain in this protein family.
Ypl184c is a largely uncharacterized, predominantly cytoplasmic protein that contains three RNA recognition motifs (RRMs). Of the three proteins that have been found to physically interact with Ypl184c, two are among the other RBPs included in this survey: Pab1 and Nab6 [71].
A disproportionate fraction of the 321 annotated mRNAs we found to associate with Ypl184c encode proteins localized to the cell wall (38, p < 10−23), ER (50, p < 10−5), plasma membrane (32, p < 10−3), or extracellular milieu (8, p < 10−3). Transcripts encoding components of several protein complexes were associated with Ypl184c, including three of five components of the Cdc28 complex (CLB2, CLN3, and CLN2) for which we obtained high-quality measurements, three of three components of the plasma membrane H+ ATPase (PMP1, PMP2, and PMA1) for which we obtained high-quality measurements, and four of nine components of the oligosaccharyltransferase complex (OST4, SWP1, OST3, and OST5) [107]. Components of these complexes that were not defined as targets of Ypl184c (at a stringent 1% FDR) were nevertheless more likely to be overrepresented in Ypl184c IPs than expected by chance, suggesting that Ypl184c may actually associate with the mRNAs encoding most or all members of these complexes.
Ypl184c associated with many mRNAs that exhibit unusual modes of translation regulation. Ypl184c bound all five of the mRNAs that have experimentally confirmed short upstream open reading frames (uORFs) (GCN4, CPA1, LEU4, SCH9, and SCO1) [108–115] in their 5′-UTRs and for which we obtained high-quality measurements; uORFs have been shown to regulate the translation of the downstream coding sequence and the stability of the mRNA [116]. Ypl184c associated with all five of the S. cerevisiae mRNAs that have been shown to have internal ribosome entry sites (IRES) (HAP4, YMR181C, GPR1, NCE102, and GIC1) in their 5′-UTRs [117,118] for which we obtained high-quality measurements; these IRESs enable cap-independent translation, often in response to environmental stresses [119]. Ypl184c also bound the unspliced HAC1 transcript, which associates with the cytosolic side of the ER membrane and is not efficiently translated until it is spliced by IRE1 as part of the unfolded protein response pathway [120,121].
Given Ypl184c's association with Pab1 and its striking association with sets of mRNAs that are known to be subject to extensive translational regulation, we speculate that Ypl184c regulates translation. The sequence motifs that we found to be significantly enriched in the mRNA targets of Ypl184c closely match the ones we found for Pub1 (Table S4). Indeed, the RNA target sets of these two proteins overlap significantly (Figures 1B and 4A). Given the absence of evidence for direct interactions between Ypl184c and Pub1, perhaps they compete for binding to overlapping groups of mRNAs. We have named YPL184C, post-transcriptional regulator of 69 kDa (PTR69).
A large body of work has given us a general picture of the relationship between the several hundred transcription factors and thousands of genes in yeast (e.g., [26–29,32,35,52–60]). Among the key features of transcriptional regulation are that: (1) individual transcription factors characteristically regulate sets of genes with related biological roles, (2) transcription factors are recruited to the specific genes they regulate by binding to specific sequences in the vicinity of those genes, and (3) combinatorial regulation of individual genes by two or more distinct transcription factors provides multidimensional control and precision to their regulation. Our systematic identification of RNAs associated with each of 46 proteins in yeast suggests that a system that shares these three key features, likely involving dozens to hundreds of RBPs, may regulate the post-transcriptional fate of most or all RNAs in the yeast cell.
This glimpse into the landscape of RNA–protein interactions has provided tantalizing clues to its organization and role. The mRNA targets of most of the RBPs in the survey encoded sets of proteins that were significantly associated with one or several related subcellular sites or biological processes (Figure 2 and Table S3). Although the regulatory roles and molecular mechanisms of most of these interactions remain to be elucidated, it seems unlikely that they have a purely decorative function. The selective binding of RBPs to sets of mRNAs that encode functionally and cytotopically related proteins provides strong evidence for widespread regulation at the post-transcriptional level. The functional relevance of these interactions is further supported by their relationships to phenotypes associated with mutation or altered expression of the RBP (Table S2). Many RBPs, including those examined in our survey, have mutant phenotypes only in specific physiological and developmental programs, and they have diverse gene expression patterns (http://www.yeastgenome.org). Thus, the regulatory program mediated by RBPs may be reorganized in response to specific physiological and developmental cues.
The striking tendency of individual RBPs to bind to sets of mRNAs whose protein products are similarly localized in the cell hints at an important role for RBPs in establishing and maintaining spatial organization in the cell, perhaps through facilitating localized protein production and mRNA decay [13,32,122–131]. The cellular structures that were most often overrepresented among the mRNA targets of many RBPs were the cell wall, plasma membrane, and ER. Thus, in addition to the familiar role of the peptide signal sequence in mediating ER-localized translation [12], RBPs may have important roles in RNA partitioning between the cytoplasm and ER, and perhaps in localization to specific sites in the periphery of the cell, such as sites of cell-wall biogenesis, bud development, and endocytosis [32,132–135]. Two of the RBPs whose targets disproportionably encode proteins localized to the cell periphery, She2 and Khd1, have been shown to be involved in trafficking some of their mRNA targets to the bud tip during the G2/M phase of the cell cycle [32,67,136]. The particularly strong overrepresentation of RBPs that associate with mRNAs encoding cell-wall components may reflect the need for extensive multilayered regulation of the location and timing of assembly and remodeling of this dynamic subcellular structure.
Identification of the information that specifies mRNA–RBP interactions is still in its earliest stages. The sequence motifs overrepresented in RBP targets, identified with the recently developed FIRE and novel REFINE methodologies, are diverse in design and location (Figures 5 and 6). Many of these RBPs recognized short linear sequences in the 3′-UTRs, 5′-UTRs, or coding sequences, or two or more of these regions. For about half of the RBPs, however, we were unable to find a sequence motif enriched among its RNA targets. Some of these RBPs may recognize structural elements. In support of this idea, we found the SRE hairpin loop, previously recognized as important for specific recognition of RNA by Vts1 [95], significantly enriched in coding sequences of Vts1 targets. Another protein in this survey, She2, is believed to recognize a three-dimensional structure in its targets [137,138]. We found promoter elements that likely specify transcription factor interactions enriched in the upstream regions of several RBP target sets, e.g., Gbp2 (Table S4). It is possible these promoter elements play an indirect role in specifying RBP interactions, perhaps by cotranscriptional recruitment of an RBP to mRNA targets via interactions with specific transcription-associated factors [22,23,139]. Identification of the large amount of still-undiscovered RNA regulatory information is an essential step in uncovering the specific regulatory program of each gene.
We identified over 12,000 mRNA–RBP interactions with high confidence. Most mRNAs in the yeast transcriptome associated with at least one of the RBPs in our survey and many associated with multiple RBPs. Some of the RBPs in the survey appear to interact with most or all mRNAs at some point in their lifecycle (Figure S1 and Text S3). Naively extrapolating from our results to the estimated 600 RBPs in Saccharomyces suggests that each mRNA might interact with a dozen or more different RBPs, on average, during its lifetime. This extrapolation is highly speculative; the sample of RBPs that we investigated is biased towards RBPs that we suspected might have a regulatory function; we do not have a good estimate of the number of regulatory RBPs that bind discrete sets of mRNAs in the manner analogous to specific transcription factors; given that three of the four proteins in this survey that were not annotated as RBPs nevertheless gave reproducible interactions with specific sets of mRNAs (Bud27, Aco1, and Tdh3), the number of potential noncanonical, unannotated RBPs with regulatory roles may be large, perhaps even in the hundreds [140–144].
There is no reason to believe the system we have described is peculiar to yeast. Extensive post-transcriptional regulation by combinatorial binding of a large and diverse set of specific RBPs is likely to be a general feature of regulation in eukaryotes. Indeed, several lines of evidence suggest an even greater genomic investment in post-transcriptional regulation in humans (and other metazoans); the number and diversity of RBPs encoded by the human genome seems to far exceed that of yeast [145], untranslated regions of mRNAs are much longer in humans (∼1,300 bases on average) than in yeast (∼300 bases on average) and appear to contain much more regulatory information [6,146,147], and the architecture of animal cells is far more diverse and complex than that of the yeast cell, with a correspondingly greater potential role for specific RNA localization [13,130,148–151].
This work has provided a glimpse of a network of RBP–mRNA interactions that is likely to play an important, but still largely undiscovered, role in biological regulation. The genes and cis-regulatory elements implicated in this process represent a substantial fraction of the genome's investment in regulation, yet the specific details and molecular mechanisms of this network of RBP–mRNA interactions are still largely terra incognita—and fertile ground for further exploration and discovery.
We carried out immunopurifications of specific proteins, together with the associated RNAs, using specific strains expressing a TAP-tagged derivative of each selected protein (Open Biosystems Cat# YSC1177-OB), essentially as described in Gerber et al. [26]. After growing 1L cultures to an optical density at 600 nm (OD600) of 0.6–0.9 in YPAD, we harvested cells by centrifugation, chilled the cell pellets on ice, washed them twice with 25 ml of ice cold buffer A (20 mM Tris–HCl [pH 8.0], 140 mM KCl, 1.8 mM MgCl2, 0.1% Nonidet P-40, 0.02 mg/ml heparin), then froze them in LN2 and stored them at −80 °C. In a few instances, we proceeded to lyse the pelleted cells immediately without freezing. To lyse the cells, we first thawed the cell suspension at 4 °C, added 5 ml of buffer B (buffer A plus 0.5 mM DTT, 1 mM PMSF, 1 μg/ml leupeptin, 1 μg/ml pepstatin, 20 U/ml DNase I [Stratagene Cat# 600032], 50 U/ml Superasin [Ambion Cat# AM2696], and 0.2 mg/ml heparin), and then mechanically lysed the cells by vortexing in the presence of glass beads. We removed the beads by centrifugation at 1,000g for 5 min, then clarified the extracts by centrifuging them twice at 7,000g for 5 min each. We adjusted the volume of the extract to 5 ml with buffer B, removed a 100-μl aliquot for reference RNA isolation, and then incubated the remaining 4.9 ml with 400 μl of 50% (v/v) suspension of IgG-agarose beads (Sigma Cat# A2909) in Buffer A with gentle rotation for 2 h. We washed the beads once with 5 ml of buffer B for 15 min, and three times with 12 ml of buffer C (20 mM Tris-HCl [pH 8.0], 140 mM KCl, 1.8 mM MgCl2, 0.5 mM DTT, 0.01% NP-40, 15 U/ml Superasin, 1 μg/ml pepstatin, 1 μg/ml leupeptin, 1 mM PMSF) for 15 min with gentle rotation. We pelleted the beads by centrifugation for 5 min at 60g in a table-top centrifuge. We then transferred the beads to 1.2-ml micro-spin columns (BioRad Cat# 732-6204), centrifuged them briefly to pellet the beads, removed buffer C, and then added 1 volume of buffer C. We cleaved TAP-tagged proteins by incubation with 80 U acTEV protease (Invitrogen Cat# 12575023) or an equivalent amount of purified TEV [152] for 2 h at 15 °C. We collected the eluent by centrifugation into 2-ml tubes. We isolated reference RNA using RNeasy Mini Kit (Qiagen Cat# 74106), while we isolated RNA from the eluate by extraction with Phenol/Chloroform/Isoamyl Alcohol, 25:24:1 (Invitrogen Cat# 15593031) twice, and chloroform once, followed by ethanol precipitation with 15 μg of Glycoblue (Ambion Cat# AM9515) as carrier.
Starting with the Operon AROS 1.1 oligo set, which contains long oligonucleotides for almost all annotated S. cerevisiae nuclear and mitochondrial coding sequences, we added 3,072 additional probes designed to detect annotated noncoding RNAs, ribosomal RNA precursors, introns, exon-intron and exon-exon junctions, other sequences predicted to be expressed, additional probes for genes with high cross-hybridization potential, and hundreds of controls for array quality measurements and normalization. Details of oligonucleotide selection and probe sequences are available from the Operon Web site (https://www.operon.com/; S. cerevisiae YBOX V1.0).
Detailed methods for microarray experiments are available at the Brown lab Web site (http://rd.plos.org/pbio.0060255).
For oligonucleotide microarrays, we resuspended oligonucleotides in 3× SSC (1× SSC = 150 mM NaCl, 15 mM sodium citrate [pH 7.0]) at a final concentration of 25 μM and printed oligonucleotides on poly-lysine glass (Erie Scientific Cat# C41–5870-M20) (http://rd.plos.org/pbio.0060255a). We printed each oligonucleotide twice per array. For most arrays, the second print was in reverse orientation to the first print, such that oligonucleotide pairs were printed with different pins and thus located in different sectors of the array.
Prior to hybridization, the oligonucleotides were crosslinked to the poly-lysine–coated surface with 65 mJ of UV irradiation. Slides were then incubated in a 500-ml solution containing 3× SSX and 0.2% SDS for 5 min at 50 °C. Slides were washed for 2 min in a glass chamber containing 400 ml of water, dunked in a glass chamber containing 400 ml of 95% ethanol for 15 s, and then dried by centrifugation. Free poly-lysine groups were then succinylated by incubation with 5.5 g of succinic anhydride that was dissolved in 350 ml of anhydrous 1-methyl,2-pyrolidoinone (Sigma Cat# 328634) and 15 ml of 1 M sodium borate (pH 8.0) for 20 min [53]. Slides were washed for 2 min in a glass chamber containing 400 ml of room temperature water, dunked in a glass chamber containing 400 ml of 95% ethanol for 15 s, and then dried by centrifugation.
cDNA microarrays containing long double-stranded DNA (dsDNA) from PCR reactions were prepared as previously described [53].
A total of 3 μg of reference RNA from extract and up to 3 μg (or 50%) of affinity-purified RNA were reverse transcribed with Superscript II (Invitrogen Cat# 18064–014) in the presence of 5-(3-aminoallyl)-dUTP (Ambion Cat# AM8439) and natural dNTPs (GE Healthcare Life Sciences Cat# US77212) with a 1:1 mixture of N9 and dT20V primers (Invitrogen). Subsequently, amino-allyl–containing cDNAs were covalently linked to Cy3 and Cy5 NHS-monoesters (GE Healthcare Life Sciences Cat# RPN5661). Dye-labeled DNA was diluted in a 20–40-μl solution containing 3× SSC, 25 mM Hepes-NaOH (pH 7.0), 20 μg of poly(A) RNA (Sigma cat # P4303), and 0.3% SDS. The sample was incubated at 95 °C for 2 min, spun at 14,000 rpm for 10 min in a microcentrifuge, and then hybridized at 65 °C for 12–16 h. For most oligonucleotide microarray experiments, we hybridized microarrays inside sealed chambers in a water bath using the M-series lifterslip to contain the probe on the microarray (Erie Scientific Cat # 22x60I-M-5522). For some oligonucleotide microarray experiments, we hybridized microarrays using the MAUI hybridization system (BioMicro), which promotes active mixing during hybridization. We hybridized cDNA microarrays inside sealed chambers in a water bath using a coverslip to contain the probe on the microarray.
Following hybridization, microarrays were washed in a series of four solutions containing 400 ml of 2× SSC with 0.05% SDS, 2× SSC, 1× SSC, and 0.2× SSC, respectively. The first wash was performed for 5 min at 65 °C. The subsequent washes were performed at room temperature for 2 min each. Following the last wash, the microarrays were dried by centrifugation in a low-ozone environment (<5 ppb) to prevent destruction of Cy dyes [153,154]. Once dry, the microarrays were kept in a low-ozone environment during storage and scanning (see http://rd.plos.org/pbio.0060255).
Microarrays were scanned using either AxonScanner 4200, 4000B, or 4000A (Molecular Devices). PMT levels were adjusted to achieve 0.1%–0.5% pixel saturation. Each element was located and analyzed using GenePix Pro 5.0 (Molecular Devices). These data were submitted to the Stanford Microarray Database [155] for further analysis. Data were filtered, as described in Text S10, to remove low-confidence measurements. Oligonucleotide pairs that both passed filtering criteria were averaged, and the data were globally normalized per array such that the mean log2 (Cy5/Cy3 fluorescence) ratio was zero after normalization. We analyzed a total of 123 IPs by microarray hybridization (Dataset S1). During the course of this work, we continued to improve and optimize our protocols. These changes and the manufacturing differences in reagents (especially in the beads used in the IPs) led to systematic differences in the background distribution of RNAs between corresponding experiments. We minimized systematic differences among sets of experiments by deriving estimates of the background separately for each set of experiments. Each group was normalized by subtracting the median log2 ratio for each molecular features across the experiments in a group from the log2 ratio of the molecular feature in each experiment. The details of the group normalization are described in Text S10, and the groups are labeled in Table S5.
Hierarchical clustering was performed with Cluster 3.0 [156], and the results were visualized as heat maps with Java TreeView 1.0.12 [157]. Clustering of FDR values (Figures 1B and 4B) was performed using the centered Pearson correlation as a similarity metric. FDR values that were greater than or equal to 10 and missing values were set to 10 prior to clustering. Clustering of the significance values measuring the degree of overlap between RBP target sets (Figure 4A) was performed using the uncentered Pearson correlation as a similarity metric.
For SAM, unpaired two-class t-tests were performed with default settings. FDRs were generated from up to 1,000 permutations of group normalized data. Details of SAM analysis are described in Text S11.
The p-values of enrichment of specific classes of RNAs and GO terms in target sets were determined using the hypergeometric density distribution function and corrected for multiple hypothesis testing using the Bonferroni method. Enrichment of GO terms was performed with GO::TermFinder [158]. For noncoding RNAs, all RNAs for which we obtained reliable measurements on the microarray were used as background. For GO analysis, only probes that are meant to capture mature mRNAs were included in analyses. For oligonucleotide microarray experiments, this corresponds to probes that match the following regular expression: Y[A-P][RL][0–9]{3}[WC][-ABC]*_ORF (Datasets S1–S3). For cDNA microarray experiments, this corresponds to probes that match the following regular expression: Y[A-P][RL][0–9]{3}[WC][-ABC]* (Datasets S1–S3). mRNAs for which we obtained high-quality measurements were used as background.
Yeast sequence files orf_genomic_1000.fasta and orf_coding.fasta were downloaded from SGD (ftp://ftp.yeastgenome.org). The 200 nucleotides upstream and downstream of coding sequences containing proper start and stop codons were extracted to create 5′-UTR and 3′-UTR databases, and the coding sequences were used for the coding sequence database. All-by-all WU-BLAST [159] (http://blast.wustl.edu/) comparisons were performed for each database against itself to identify highly similar sequences (using options -e 1e-10 -b 5000 -S 1 -F F). WU-BLAST output files were parsed to identify alignments of greater than or equal to 80% identity extending over half the length of the query sequence, and all such sequence pairs were grouped into redundant classes. One sequence from each redundant class was retained to create nonredundant databases for each region.
The REFINE procedure was run using hexamers with significant (p < 10−3) enrichment in RBP targets, as measured by the hypergeometric distribution (using options –ss –f 3 –g 6 –ct 3 –max 15 –dust). MEME analysis (version 3.5.1) was performed on the REFINE output sequences with options –dna –minw 6 –maxw 15 –text –maxsize 200000 –evt 10 –nmotifs 3. Motif site sequences were extracted from MEME output and used to generate position-specific log-odds scoring matrices based on the observed frequencies and 0.25 pseudocounts per base, and null frequencies based on mononucleotide composition of all sequences in the corresponding (5′-UTR or 3′- UTR) nonredundant database. Cutoff scores for motif classification were chosen to maximize the significance of association of motif sites with RBP target membership as measured by hypergeometric p-values for enrichment. All subsequences with scores above the cutoff threshold were classified as motif sites, and the final significance was measured as the negative log of the p-value of motif enrichment in RBP targets. FIRE analysis was run on the nonredundant 5′- and 3′-UTR databases using binary data indicating RBP target membership with options –exptype=discrete –seqlen_rna=200 –nodups=1 –dodna=0.
For both REFINE and FIRE, statistical significance of the predicted motifs was assessed by randomly generating target sets of similar size and repeating each procedure 100 times on the simulated target data. We defined a test statistic as the negative log of the p-value for motif enrichment for REFINE; the reported motif z-score was used for FIRE motifs, and we compared the observed values of these test statistics to the distributions generated by the random simulations (Table S4). Motifs were declared as significant if the observed test statistic was greater than three standard deviations above the mean, or if there was significant enrichment (p < 10−4) of the motif in targets occurring in regions from which that motif was not derived.
Our microarray experiment data are publicly available from the Stanford Microarray Database and Gene Expression Omnibus.
|
10.1371/journal.pbio.2006062 | Repeated translocation of a gene cassette drives sex-chromosome turnover in strawberries | Turnovers of sex-determining systems represent important diversifying forces across eukaryotes. Shifts in sex chromosomes—but conservation of the master sex-determining genes—characterize distantly related animal lineages. Yet in plants, in which separate sexes have evolved repeatedly and sex chromosomes are typically homomorphic, we do not know whether such translocations drive sex-chromosome turnovers within closely related taxonomic groups. This phenomenon can only be demonstrated by identifying sex-associated nucleotide sequences, still largely unknown in plants. The wild North American octoploid strawberries (Fragaria) exhibit separate sexes (dioecy) with homomorphic, female heterogametic (ZW) inheritance, yet sex maps to three different chromosomes in different taxa. To characterize these turnovers, we identified sequences unique to females and assembled their reads into contigs. For most octoploid Fragaria taxa, a short (13 kb) sequence was observed in all females and never in males, implicating it as the sex-determining region (SDR). This female-specific “SDR cassette” contains both a gene with a known role in fruit and pollen production and a novel retrogene absent on Z and autosomal chromosomes. Phylogenetic comparison of SDR cassettes revealed three clades and a history of repeated translocation. Remarkably, the translocations can be ordered temporally due to the capture of adjacent sequence with each successive move. The accumulation of the “souvenir” sequence—and the resultant expansion of the hemizygous SDR over time—could have been adaptive by locking genes into linkage with sex. Terminal inverted repeats at the insertion borders suggest a means of movement. To our knowledge, this is the first plant SDR shown to be translocated, and it suggests a new mechanism (“move-lock-grow”) for expansion and diversification of incipient sex chromosomes.
| Sex chromosomes frequently restructure themselves during organismal evolution, often becoming highly differentiated. This dynamic process is poorly understood for most taxa, especially during the early stages typical of many dioecious flowering plants. We show that in wild strawberries, a female-specific region of DNA is associated with sex and has repeatedly changed its genomic location, each time increasing the size of the hemizygous female-specific sequence on the W sex chromosome. This observation shows, for the first time to our knowledge, that plant sex regions can “jump” and suggests that this phenomenon may be adaptive by gathering and locking new genes into linkage with sex. This conserved and presumed causal sex-determining sequence, which varies in both genomic location and degree of differentiation, will facilitate future studies to understand how sex chromosomes first begin to differentiate.
| Sex chromosomes can be a strikingly diverse and evolutionarily labile component of eukaryotic genomes [1]. The defining feature of a sex chromosome, the sex-determining region (SDR), has experienced similar restructuring in multiple independent instances of autosomes evolving into heteromorphic sex chromosomes [2]. Specifically, recombination is suppressed, and an increasingly greater proportion of the chromosome becomes hemizygous, which is thought to involve existing and/or newly acquired linkage to loci under sexually antagonistic selection [3]. The mechanisms of this chromosome restructuring may involve modifying crossover sites and/or successive inversions of the SDR or translocations of large or small sequences on and off the sex chromosome [3,4]. Turnovers that change the genomic location of the SDR have been revealed in the evolution of animal sex-determining systems [1,5–8], where they may be important drivers of sexual dimorphism and speciation [9,10]. While theory on the processes driving these transitions is growing [11–14], few systems exist in which the mechanisms of turnovers can be empirically inferred [15–17].
Fundamental questions about SDR turnovers therefore remain unanswered. Do turnovers typically involve mutations in new loci that take control of an existing sex-determining mechanism [18,19], functionally independent mutations [20], or translocations of the existing sex-determining gene(s) to new chromosomes [21–24]? Similarly, do turnovers typically restart the process of SDR divergence, maintaining “ever-young” sex chromosomes [25], or do they contribute to increasing chromosome heteromorphy via loss or gain of sequence [11,14,26]? And ultimately, is there an adaptive basis for these turnovers? Although master sex-determining genes like SRY and DMRT1 are highly conserved in some animal systems, the causal SDR loci or gene cassettes remain unknown for most dioecious eukaryotes [27]. Even less is known about the temporal order of turnovers in any taxon and thus directional trends in sex-chromosomal rearrangement [2].
Turnovers of SDRs are likely to be quite common in plants, in which genetic control of sex appears to be poorly conserved [28,29]. Flowering plant SDRs may be diverse because dioecy (separate males and females) has evolved repeatedly from hermaphroditism (combined male and female function) and many sex chromosomes are relatively young and homomorphic [28,29,30]. Additionally, approximately one-third of flowering plant species are estimated to have a recent polyploid ancestry [31]. These whole-genome duplications provide a larger substrate for potential sex-determining genes or rearrangements [32]. Yet despite the potential of dioecious plants for yielding evolutionary insights, there are few systems with mapped SDRs [28,29] or known causal genes [33,34], although long-standing theory predicts that two linked genes, one controlling male function and one controlling female function, are involved [1,35]. Moreover, even when observed, the pattern and mechanism of turnovers remain entirely unexplored.
The octoploid (8x) strawberries (Fragaria) stand out as model system for studying plant sex chromosomes [36–40] and polyploidy [36,41] in an evolutionary context because they show recently evolved dioecy from within a group of closely related, predominantly hermaphroditic diploid (2x) taxa. The octoploid taxa all possess homomorphic, female heterogametic (ZW) sex chromosomes with a single SDR explaining the majority of variation in male and female function, though the degree of sexual dimorphism varies across taxa [36,39,42–45]. Male function (sterile versus fertile), in particular, is a binary trait showing simple Mendelian inheritance (1:1). Male sterility (“female”) is dominant to male fertility (“male”), determined entirely by the SDR, and here we use it to define sex phenotype. All octoploid species share a recent polyploid origin involving four diploid ancestors (now coexisting as “subgenomes” [Av, Bi, B1, and B2] within the octoploid genome, Fig 1A) [41,46]. The homologous chromosomes from each subgenome (homoeologs) are genetically distinct and are inherited disomically. Nevertheless, homoeologs show high synteny with each other and with the Fragaria reference genome (“Fvb”) derived from the hermaphroditic diploid F. vesca with seven haploid chromosomes (named Fvb1 through Fvb7, Fig 1A) [41]. Therefore, the octoploids have seven homoeologous groups, each with eight chromosomes (2N = 8x = 56). The approximately 700 megabase (Mb) octoploid genome is slightly smaller, however, than four times the approximately 200 Mb diploid Fvb reference genome, likely due to numerous small deletions [41,47]. The diploid and octoploid genomes are largely collinear [41], and we refer to all genome positions by their location along Fvb chromosomes in Mb.
The SDR of Fragaria octoploids has been mapped in three geographically distinct octoploid taxa (here in order from eastern to western North America): F. virginiana ssp. virginiana [42], F. virginiana ssp. platypetala [40], and F. chiloensis [39,43] (Tables 1 and 2A). Each SDR occurs at a unique section of a chromosome from the same homoeologous group, i.e., the group that corresponds to Fvb6 in the diploid reference, but each from a different subgenome (Fig 1B) [40]. Specifically, the mapped SDR locations match Fvb6 position 1 Mb on subgenome B2 in a cross of two F. virginiana ssp. virginiana parents ([42], results herein), 13 Mb on subgenome B1 in a cross of two F. virginiana ssp. platypetala parents [40], and 37 Mb on subgenome Av in three crosses involving pairs of F. chiloensis parents [39,43] (Table 2A). Moreover, genetic maps in the natural hybrid (F. × ananassa ssp. cuneifolia) of two of these taxa corroborate these map locations [48] (Table 2A). Though the chromosomes harboring the various SDRs are all homoeologous, they are distinct: Fragaria subgenomes show little evidence of recombination with each other [41], and the positions of the various SDR locations are too far apart (several Mb) for normal recombination (Fig 1B). All SDRs occur far from centromeres in gene-dense regions, and although early stages of recombination suppression may be evolving, pseudoautosomal recombination still occurs between the Z and W along most of their lengths, allowing for fine-scale mapping [39]. The recent evolutionary origin of dioecy and the extensive recombination still occurring on the sex chromosomes suggest that there is very little sex-specific sequence other than the causal gene(s). However, despite extensive previous work mapping the chromosomal locations of Fragaria SDRs [39,40,42,43,48–50] as well as conjecture that autosome Fvb6 may possess sexually antagonistic genes that predispose it to become a sex chromosome [51]—as seen in other systems [52–54]—no candidate causal genes have been identified, and nothing is known of the molecular mechanism beyond very broad inferences (e.g., that control is nuclear rather than cytoplasmic). Therefore, identifying sex-determining gene(s) and inferring whether they are shared across the octoploid Fragaria will provide a unique opportunity for testing whether sex chromosome turnovers represent translocations of the same SDR.
Here, we use whole-genome sequencing and molecular evolutionary analysis of multiple octoploid Fragaria taxa to characterize and compare SDRs that are found on different chromosomes (Fig 1B and Table 2B). Our goal is to determine whether a single W-specific sequence has translocated among genomic locations. We find an “SDR cassette” shared by females across taxa and never detected in male plants. The SDR cassette contains two putatively functional sex-determining genes and has moved at least twice, together with flanking sequences that reveal the order of the translocation events. Because the moved regions are hemizygous, each translocation has created a wider hemizygous region than formerly existed. Therefore, we report the first case, to our knowledge, of a repeatedly translocating SDR in plants and propose a new hypothesis for sex-chromosome differentiation.
To identify sequence unique to the W chromosome(s), we sequenced the complete genomes of 31 female and 29 male plants in five octoploid taxa (Tables 1 and S1 and S1 Fig; range of coverage relative to the haploid reference genome = 16–57×; median = 33×). These represent the North American range of the octoploid Fragaria and include the parents of the crosses used to map sex determination (Table 2A and S2 Fig) [36,39,40,43,48]. From these reads, we then identified sequence (“31-mers”: 31 bp motifs, the longest computationally feasible size under our particular pipeline, S1 Fig) seen in females but never in males (S2 Table). Fewer than 5% of these female-specific 31-mers aligned to more than one location in the F. vesca reference genome (Fvb), suggesting that the female-specific sequence is not highly repetitive. In 29 out of 31 females, we observe similar female-specific sequence. The exceptions here (2 out of 31 females) are both F. virginiana ssp. glauca plants, which originated from a distinct geographic region from all other samples (i.e., the Rocky Mountains, S1 Table and S2 Fig) and could carry distinct versions of this sequence or possibly possess nonhomologous SDR(s). We did not observe all shared female-specific sequence in the remaining 29 females, as expected owing to missing data due to our low sequencing coverage (2–7× per octoploid chromosome). Still, these 29 females all possess female-specific 31-mers aligning to the same 2 kb window on Fvb7 position 18 Mb (S3 Fig). They also possess sequence overlapping a single site homologous to Fvb6 position 1 Mb, where these octoploid females possess a 23 bp “diagnostic deletion” not seen in the diploid hermaphrodite F. vesca (Fvb) or any of the 29 male plants (S3 Fig). In contrast to the female-specific sequence found, male-specific 31-mers were rarely seen (S2 Table), as expected because Z chromosomes are present in both males and females, suggesting that our method yields a very small number of false-positive 31-mers. Moreover, while female-specific sequence is shared across the octoploid taxa, the SDRs of these plants maps to three different genomic locations. This suggests that translocations are likely involved, which demands further characterization of female-specific sequence for confirmation (see below).
To assess and annotate the SDR, we assembled the shared female-specific sequence, generating three contigs totaling 13 kb in length. These contigs were ordered and oriented into a unified W-specific haplotype, the SDR cassette (Fig 2), by using highly similar autosomal and Z chromosome sequences as scaffolds. Specifically, most (10.4 kb) of the SDR cassette could be aligned (98% similarity) to Z chromosome bacterial artificial chromosomes (BACs) obtained from F. virginiana ssp. virginiana, originating from the maternal linkage cross parent at the fine-mapped SDR location from that cross (S4 and S5 Figs). Most of this sequence (8.6 kb) could also be aligned (93% similarity) to the diploid (F. vesca) reference genome at the fine-mapped location of Fvb6 position 1 Mb. A 1.2 kb segment of the SDR cassette was not homologous to Fvb6 but instead showed 99% similarity to Fvb7 position 18 Mb. Therefore, the W-specific SDR cassette is relatively short and shows homology to multiple sections of the genome.
Only two coding genes—annotated as GDP-mannose 3,5-epimerase 2 (here GMEW) and 60S acidic ribosomal protein P0 (here RPP0W)—were identified in the SDR cassette (Fig 2). GMEW homologs occur on the Z chromosome BACs (99% similarity) and at Fvb6 position 1 Mb (98% similarity). GDP-mannose 3,5-epimerase converts GDP-mannose to GDP-L-galactose in vitamin C and cell wall biosynthesis [55,56], affecting fruit development in Fragaria [57,58] and pollen production in other plants [56]. In some females, GMEW has a premature stop codon shortening the coding sequence from 376 to 222 residues. Whereas GMEW is a plausible sex-determining candidate, the stop codon polymorphism may suggest a variable role among females. In contrast, the second gene, RPP0W, falls within a 1.2 kb W-specific insertion that shows 99% similarity to a gene at Fvb7 position 18 Mb and is thus responsible for the female-specific 31-mers homologous to that location (S3 Fig). However, it lacks that gene’s four introns, suggesting that it is a cDNA resulting from retrotransposition. RPP0W sequences across the Fragaria taxa studied here form a monophyletic group with respect to this autosomal paralog and other autosomal paralogs (S6 Fig and S1 Data), a finding that is consistent with a single SDR origin. Ribosomal proteins are essential for polypeptide synthesis and are often retrotransposed [59]. In plants they can affect processes from development to stress response [60], with mutations sometimes acting dominantly [61], as expected for the first mutation in a female heterogamic (ZW) system [1]. In rice, the overaccumulation of ubiquitin fusion ribosomal protein L40 results in defective pollen and male sterility [62]. In diploid hermaphroditic F. vesca, both RPP0W and GMEW homologs show decreasing expression during anther development and even lower expression within pollen [63], but expression profiles in octoploids remain to be characterized. Neither gene family (of GMEW nor RPP0W) has been directly implicated in sex determination, but many pathways could potentially affect plant sex functions [64,65].
In classic two-gene SDR models, one gene affects male function and another female function [35]. Previous quantitative trait locus (QTL) mapping has shown that the Fragaria SDR affects both male and female function [36,39,40] and shows differential recombination rates in ZZ versus ZW individuals [39]. However, we cannot yet conclude that there are two functional, non-recombining genes because a single master regulator could also perform both roles [33] and additional modifiers of female function could have evolved. Moreover, in addition to the two genes, there is the diagnostic deletion and two repetitive unassembled gaps within the SDR (Fig 2), which, though apparently noncoding, could also be functional motifs. Regardless, what is striking here is that an SDR cassette (W-specific) is shared across females from different taxa and populations where it occurs at multiple genomic locations (Fig 1B).
To infer the evolutionary history of the shared SDR cassette, we reconstructed the phylogeny of a 2.7 kb portion overlapping RPP0W and the diagnostic deletion (Fig 2) in the 29 females with the SDR cassette (Fig 3 and S2 Data). These W-specific sequences resolved into three distinct and well-supported (≥75% Shimodaira-Hasegawa–like support) clades: α, β, and γ (Fig 3 and Table 2). Notably, each of the three SDR map locations (Fig 1B) is associated with a single clade (Fig 3). The SDRs from the F. virginiana ssp. virginiana female for which male sterility was fine-mapped to Fvb6 position 1 Mb (S4 Fig and Table 2B)—and those of most other F. virginiana ssp. virginiana females—were in the α clade. SDRs of the β clade include two from females for which male sterility has been previously mapped to Fvb6 position 13 Mb (Table 2B) [40,48]. The SDRs that form the γ clade included those from all three F. chiloensis females for which male sterility maps to Fvb6 position 37 Mb (Table 2B) [39,43] and the remaining six F. chiloensis females as well as a few from females of F. virginiana ssp. virginiana and F. virginiana ssp. platypetala. The overall topology, with F. chiloensis nested within F. virginiana, reflects the inferred evolutionary history of these taxa, which are not reciprocally monophyletic [46]. The β and γ clades are sister to each other with strong support (93% Shimodaira-Hasegawa–like support; 92% bootstrap support; Fig 3), suggesting that these SDRs (Fvb6 positions 13 Mb and 37 Mb) may be more closely related, whereas those in the α clade (Fvb6 position 1 Mb) are more distantly related and represent the source of the homologous sequence shared across all clades (S3 Fig).
Because F. chiloensis had the largest amount of female-specific sequence (S3 Fig) and because all SDRs from F. chiloensis females formed a monophyletic group (Fig 3), we constructed an extended SDR haplotype from female-specific sequence identified in this species. This assembly then served as a reference sequence of the full W-specific haplotype for further analyses involving the other taxa. We inferred that all female-specific sequence must be very tightly linked because—barring lethal genotype combinations, which would skew the sex ratio in ways that we do not observe [36,39]—there is no known mechanism by which multiple unlinked regions of the nuclear genome could all be female specific. This inference was validated by the constructed haplotype. Specifically, we assembled a 28 kb haplotype containing 89% of the female-specific 31-mers for this species and seven coding genes (Fig 4 and S3 Data and S3 Table). Within the full W-specific haplotype, the SDR cassette was nested within an additional 10 kb of “flanking” sequence on either side (Fig 4, middle) that included 5 kb homologous to Fvb6 position 13 Mb (split nearly evenly between left and right flanks), consistent with the SDR map location on subgenome B1 of homologous group 6 (Fig 1B and Table 2A) [40], as well as 2 kb homologous to Fvb4 position 21 Mb (right flank), accounting for the female-specific 31-mers homologous to that location (S3 Fig). These sections were nested within an additional 5 kb of “outer” sequence (Fig 4, middle) primarily showing homology to Fvb6 position 37 Mb, consistent with SDR map location on subgenome Av observed in F. chiloensis (Table 2A) [39]. An additional 7% of the F. chiloensis female-specific 31-mers do not align to this haplotype but are probably closely adjacent, as they also align to Fvb6 position 37 Mb. The outer section contained 31-mers that were female specific in our sample but probably not hemizygous. That is, orthologous Z pseudo-autosomal [2] sequence presumably exists with which it may potentially recombine, although ZW recombination rates are low near the F. chiloensis SDR [39]. In summary, the SDR at Fvb6 position 37 Mb encloses nested “souvenir” sequence matching the other known SDR locations (1 Mb and 13 Mb) in other taxa (Fig 1B), and this explains the greater proportion of female-specific 31-mers in F. chiloensis compared with the other taxa studied (S3 Fig). The female specificity of this SDR sequence, despite showing homology to disparate portions of the diploid reference genome, is consistent with movements having occurred from those locations to a new location carrying a female-determining factor.
Using the full W-specific haplotype in F. chiloensis as the reference, we characterized in detail the sequence neighboring the SDR cassette in each of the three phylogenetic clades (Fig 4). We did not assemble complete haplotypes for each clade independent of the F. chiloensis W haplotype assembly because the α clade had few female-specific 31-mers and the β clade had only three females and therefore we lacked the power to eliminate false-positive female-specific 31-mers. Instead, we identified portions of the assembled haplotype within clades that we could infer to be female specific using the following two parallel methods: alignment of female-specific 31-mers to the haplotype, and sites on the haplotype at which paired reads aligned on either side in females only (S7 Fig). These analyses revealed that distinct portions of the W haplotype were female specific in each clade (S7 Fig and S4 Table). In particular, in the α clade, only the SDR cassette is female specific. In contrast, the β clade shows female-specific sequence in both the SDR cassette and flanking sections, and the γ clade shows female-specific sequence in all three sections. The two females lacking the diagnostic deletion also did not possess any female-specific read pairs, further confirming that the SDR cassette is absent in these individuals and suggesting other mechanism(s) of male sterility [51].
The presence of sequence homologous to Fvb6 position 1 Mb within the SDR cassette in both β and γ clades (Fig 4) suggests that the SDR of the α clade and its location on Fvb6 position 1 Mb is ancestral (Fig 1B), with a translocation from Fvb6 position 1 Mb to position 13 Mb in the ancestor of the β and γ clades (Fig 3). A second translocation to Fvb6 position 37 Mb, specific to the γ clade (Fig 3), explains the SDR cassette and flanking sections retained in γ from its previous locations and also the outer sections unique to γ with homology to Fvb6 position 37 Mb (Fig 4), as well as the map location at Fvb6 position 37 Mb in three previously studied females in the γ clade (Fig 1B) [39, 43]. Therefore, based on the “souvenir” sequence that suggests that two translocations each carried adjacent sequence from their previous locations, we can propose a temporal order of SDR movements (Fig 3, black arrows).
A 2 kb portion of the downstream flanking section shows homology to Fvb4 position 21 Mb, which could be a souvenir from another prior SDR location or an independent translocation of sequence into the SDR in the β and γ ancestor; such events are commonly seen in sex chromosomes [2]. The proposed translocations must have occurred rapidly because octoploid Fragaria originated only approximately 1 million years ago (Mya) [46,66] and the aligned 2.7 kb portions of the SDR cassettes (Fig 2 and S2 Data) show >99% sequence similarity. This conjecture is also supported by incomplete lineage sorting of the SDR in F. virginiana (Fig 3), resulting in SDR polymorphism among females of this species. In contrast, F. chiloensis, which is monophyletic and is derived from F. virginiana ssp. platypetala [46], is apparently fixed for the derived SDR γ clade. All three SDR clades are found within F. virginiana ssp. platypetala (Fig 3), whose phylogenetic position [46,67] and geographic range (S2 Fig) lie between F. virginiana ssp. virginiana and F. chiloensis.
Although the mechanism of translocation of sex-determining sequence remains unknown, a striking sequence pattern suggests transposon-mediated movement. Specifically, we observe a 25 bp sequence that is inverted and repeated at the very distal ends of the flanking sections, where sequence homologous to Fvb6 13 Mb meets sequence homologous to Fvb6 37 Mb (Fig 4). On the distal end of each segment, we observe the dinucleotide motif TA. Pairs of terminal inverted repeats of 10 bp or more in length, adjacent to short duplications, are hallmarks of Class 2 transposable elements [68,69]. Therefore, this sequence signature is consistent with the hypothesis that a mobile element transported the 23 kb of SDR cassette and flanking sequence from the β clade location at Fvb6 13 Mb to the γ clade location at Fvb6 37 Mb. Terminal inverted repeats also occur in foldback elements, which can cause chromosomal rearrangement via ectopic recombination [70], and this mechanism could also facilitate movement of the SDR among homoeologs of Fvb6. We do not see terminal inverted repeats at the border between the SDR cassette and the flanking sequence, but this may have been lost, perhaps explaining why adjacent sequence was then also moved during the second translocation. Most transposable elements are under 23 kb in size, and we see no evidence of either an intact transposase, a Helitron transposon, or any known plant repetitive sequences other than stretches of dinucleotide repeats under 50 bp. Therefore, although the full W-specific haplotype remains incompletely assembled and could harbor a transposase (Fig 4), we hypothesize that the SDR movements do not involve a classic, active transposon but rather are relatively rare events that leverage active transposases that may be encoded elsewhere, as with miniature inverted-repeat transposable elements [68,69].
Consistent with the scenario of relatively few SDR movements, no female appears to have more than a single SDR cassette. Although we cannot assemble paralogous autosomal sequence due to high similarity among subgenomes, we can identify autosomal read pairs that align to the W haplotype but are spaced too far apart (>1 kb) to have originated in the SDR. The nonadjacent sections of the W haplotype where these paired reads align must therefore be contiguous in autosomes as they are in Fvb, though not in the SDR (S7 Fig and S4 Table). Coverage depth for these reads does not differ between males and females (Student t test, p > 0.1), and in females, coverage is 8-fold higher than for W-specific read pairs, suggesting that these reads originate from autosomal or pseudoautosomal regions on all four subgenomes. Therefore, there is no evidence that any autosomal homoeolog possesses an insertion representing a degraded or partial SDR. After an SDR translocation event, there would have been little or no co-occurrence of two SDR cassettes in the same female because the SDRs would occur on distinct subgenomes that segregate separately. Once separated, two SDR cassettes can never rejoin the same genome because two female plants cannot mate. Therefore, it appears that the former sex chromosomes, which have reverted to autosomes due to SDR turnover events, are descended from Z chromosomes and not W chromosomes (Fig 5).
Repeated translocation of the SDR is the only explanation consistent with all observations. Shared sequence across disparate SDRs could be explained if the shared sequences were repetitive motifs common throughout the genome, but this is not the case. Indeed, such motifs would be present in multiple copies in all individuals and thus would not be female specific. If female-specific 31-mers were false positives due to chance co-occurrence of some sequences in our female samples, we would expect to see a similar quantity of male-specific false positives, which we do not (S2 Table). Similarly, if control of sex were polygenic, then several distinct sequences could all show a correlation with sex without being physically adjacent, but this explanation can also be ruled out. Not only does sex map to a single genomic location in each of several linkage crosses [39,40,42,43,48], but under polygenic architecture, no one sequence would show a perfect correlation with sex. Furthermore, we observe sequencing reads spanning the junctions between distinct sections of the female-specific haplotype (S7 Fig), confirming that these sequences occur side by side. Therefore, the distinct sections of the SDR are adjacent only in females and must have been brought together by translocation.
Chromosomal rearrangements may be especially common in polyploids [71], and because these could disrupt and/or create linkage between genes essential to sex function, they may underlie the widespread association between dioecy and polyploidy [32]. We cannot infer whether the SDR translocations have occurred at an unusually high rate relative to selectively neutral sequence. However, because rapid turnovers of SDR locations are common in evolution across many taxa [4–6,10,11], the rearrangements we observe have been plausibly favored by selection. If so, the continued coexistence of multiple SDR locations suggests that adaptive replacement may be ongoing but incomplete across these geographically widespread populations. High-density linkage maps of these octoploids [39,41] indicate conserved synteny across the homoeologs of Fvb6, with no rearrangements of Mb-sized regions, suggesting that these chromosomes have only experienced relatively small translocations. SDR translocations have been suggested to be favored by selection because this allows escape from genetic load if deleterious mutations linked to the SDR in its original location cannot be effectively purged due to lack of recombination or selective forces maintaining sex-determining alleles [13]. Alternatively, it could be advantageous for the SDR to move in order to become linked with loci under either sexually antagonistic selection [14] or other types of balancing selection without a direct connection to sex [72]. Either of these could apply in Fragaria. However, a third adaptive explanation is suggested by the observation that each jump of the SDR increased the size of the hemizygous female-specific haplotype by moving adjacent sequences (Figs 4 and 5 and Table 2B). We present this explanation as a conceptual hypothesis, but further work is required to confirm it in Fragaria and test it in other taxa with sex-chromosome variations.
Although we have not assembled Z chromosome sequences other than the α clade BACs, we can infer hemizygosity by assuming that the Z chromosomes have the same composition as the reference genome (Fig 4, bottom). Therefore, the SDR in the α clade is hemizygous (and shows female specificity) only for the 1.2 kb insertion containing RPP0W. The SDR in the β clade is hemizygous for the 13 kb SDR cassette and its two genes (GMEW and RPP0W) that show complete linkage disequilibrium with the sex-determining factor because they have no Z orthologs with which to recombine. The SDR in the γ clade is hemizygous for the 23 kb of SDR cassette and flanking sections containing five genes (GMEW, RPP0W, and three additional genes, Fig 4) in complete linkage disequilibrium with sex for the same reason. The outer sections of the γ SDR in F. chiloensis with homology to Fvb6 position 37 Mb are presumed to not be hemizygous but contain female-specific 31-mers representing variants that are in linkage disequilibrium with the hemizygous insertion. If SDR includes separate genes that are under sexually antagonistic selection—as seen for some F. virginiana traits [44]—and if two such genes are maintained polymorphic, recombination will generate maladaptive combinations [35]; a hemizygous translocated copy could thus maintain the adaptive combinations (e.g., a female-determining sequence and a female-beneficial allele of a polymorphic sexually antagonistic gene) in complete linkage disequilibrium. Therefore, translocation could represent a means of recombination suppression during sex-chromosome evolution, perhaps explaining some genomic rearrangements involving incipient SDRs and the process—“move-lock-grow”—by which the difference between the two sex chromosomes increases. F. chiloensis shows greater phenotypic differences between the sexes than other Fragaria species [43], as well as sex differences in recombination rates [39], and is fixed for the γ clade SDR in our samples (Fig 3 and Table 2B), which represents the largest hemizygous region. In contrast, F. virginiana shows less pronounced and more variable sex phenotype differentiation [44,45] and harbors SDRs from all three clades, with the α SDR being most common (Fig 3), which has the smallest hemizygous region. This is consistent with a correlation between SDR size/content and sexual dimorphism. A similar growth mechanism may underlie other hemizygous supergenes [73]. Whereas the “move-lock-grow” hypothesis is suggested by our data, future studies should test whether SDR translocations tend to increase the size of the hemizygous segment in other taxa and also whether there is an adaptive benefit to locking souvenir sequence into linkage with sex.
A hemizygous SDR cassette, which contains both a gene with a known role in fruit and pollen production and a novel retrogene absent on Z and autosomal chromosomes, is conserved and has repeatedly changed genomic location across octoploid Fragaria, supporting a translocation model of sex-chromosome turnover (Fig 5). To our knowledge, this is the first unambiguous evidence of SDR translocation in flowering plants because it is rarely possible to distinguish translocations from de novo innovations unless putative causal sequences have been identified in more than one taxon [29,74]. In Salicaceae, SDRs occur on different chromosomes with no evidence of large-scale rearrangements, but data thus far are consistent with either master/slave regulatory dynamics [18] or SDR jumps [75,76]. Turnovers involving reversal of heterogamety, as seen in Silene [77], are more likely to be fusions of sex chromosomes to autosomes rather than translocations of SDR sequence to new chromosomes. Our discovery of a conserved yet mobile W-specific SDR helps to unify extensive and disparate research on the genetic basis of dioecy in Fragaria and across flowering plants [41,74]. It suggests that independent mechanisms of dioecy within closely related taxa may be rarer than they appear. Instead, SDR translocation can maintain the same genetic basis for sex while adjusting genomic location and accumulating sequence that may contain sexually antagonistic alleles as well as increasing recombination suppression within the growing hemizygous SDR. The “move-lock-grow” phenomenon may allow for rapid and extensive change in sex chromosomes, perhaps influencing sexual dimorphism, hybrid compatibility, recombination rates, or other traits of evolutionary or ecological importance.
We determined sex using our established method [51]. In brief, we grew plants with 513 mg granular Nutricote 13:13:13 N:P:K fertilizer (Chisso-Asahi Fertilizer) under 15:20°C night:day temperatures and 10 to 12-hour days and then exposed them to 8:12°C night:day temperatures with an 8-hour low-light day to initiate flowering. Fertilizer and pest control measures were applied as needed. Male function was scored as a binary trait: plants with large, bright-yellow anthers that visibly released pollen were “male (male-fertile),” and plants with vestigial white or small, pale-yellow anthers that neither dehisced nor showed mature pollen were “female (male-sterile).” Because of the tight correlation between male function and female function, male sterility serves as a good phenotypic marker of the SDR [39].
Genomic DNA was extracted from silica dried leaf tissues using Norgen Biotek Plant/Fungi DNA Isolation 96-Well Kit (Ontario, Canada) and by the service provider Ag-Biotech (Monterey, CA). An additional 100 μl 10% SDS and 10 μl β-mercaptoethanol were added to the lysis buffer to improve DNA yield. DNA was further purified with sodium acetate and ethanol precipitation. DNA concentration was quantified by Quant-iT PicoGreen (Invitrogen, Carlsbad, CA) assays at University of Pittsburgh Genomics and Proteomics Core Laboratories (GPCL).
For the whole-genome analysis, we examined 60 outbred, unrelated plants distributed across the geographic ranges of the octoploid Fragaria species (S1 Table). These samples were collected from the wild as clones or obtained from the USDA National Clonal Germplasm Repository. Genomic DNA extraction and library preparation were performed by the Oregon State University Center for Genome Research and Biocomputing (CGRB) and at University of Pittsburgh. We sheared DNA to 300 bp using a Bioruptor Pico (Diagenode, Denville, NJ) and used the NEBNext Ultra DNA Library Prep Kit for Illumina (New England BioLabs, Ipswich, MA) with individually indexed dual barcodes. We sequenced whole genomes of 60 Fragaria samples using four lanes of paired-end 150 bp on an Illumina HiSeq 3000, with 13 to 20 samples per lane (Table 1). Although reads were not aligned to the diploid F. vesca reference genome Fvb, we report coverage relative to this reference as the sum of lengths of all sequenced reads divided by the size of Fvb (e.g., 8× coverage relative to Fvb should mean approximately 1× coverage per chromosome in an octoploid).
We converted FASTQ files to FASTA and used Jellyfish 1.0.2 [78] to count 31-mers in each sample, the largest k-mer size allowed by Jellyfish (S1 Fig). We used the Linux “sort” and “join” functions to combine lists of 31-mers and generate lists of 31-mers shared by sets of females (defined taxonomically or as α, β+γ, or γ clade) and absent in all male plants. As a control to ensure this method was not yielding false positives or repetitive sequence (e.g., from heterochromatin), we also searched for male-specific 31-mers, which are not expected to exist because the Z chromosome is present in both sexes. As with the females, we searched for male-specific 31-mers within clades (for males, defined as the clade of the closest-related female plant, as determined by chloroplast phylogeny [46]). To aid the assembly of the full W-specific haplotype in F. chiloensis, we also generated lists of 31-mers that were female specific in that species (ignoring males of other species), as well as 31-mers shared in all but one female, assuming that a W-specific 31-mer could be absent due to insufficient coverage or a rare sequence variant. The assembly was feasible because these nearly female-specific 31-mers were densely spaced across the SDR haplotype, such that median span between nonadjacent 31-mers was 100 bp, and 90% of them were separated by less than 500 bp (excluding the two unassembled gaps), typically within the range spanned by paired-end reads. We aligned these 31-mers to Fvb using BLAT version 32x1 [79] and retained hits with at least 29 bp matching and gaps no larger than 30 bp. We extracted reads containing female-specific 31-mers and their mate pairs from the original FASTQ files. We assembled these manually in BioEdit version 7.2.5 [80], beginning at the diagnostic deletion and moving outward in both directions, when possible guiding the assembly with alignment to homologous Fvb or BAC sequences. Gaps between contigs containing female-specific 31-mers were manually joined with additional reads as possible. We assembled the central 2.7 kb of the SDR cassette, including the diagnostic deletion and RPP0W, for all females possessing it. We assembled a pseudo-outgroup sequence based on homologous portions of Fvb and BAC6 and used RAxML [81] with -m GTRCAT to generate a phylogeny of the W sequence. Major clades (α, β, and γ) were assigned visually. We used a consensus sequence of RPP0W (966 bp) from each of the three clades to generate a phylogeny with the four RPP0W paralogs in Fvb, again using RAxML [81] with -m GTRCAT.
We assigned portions of the W haplotype to Fvb regions using BLAST at GDR [82] (S3 Table). We identified genes using GENSCAN [83] and annotated them with BLAST to the NCBI database and to Fvb—which is annotated—using GDR [82]. Adjacent genes (S5 Table) were identified from the Fvb annotation. Gene expression data in F. vesca [63] were extracted from http://mb3.towson.edu/efp/cgi-bin/efpWeb.cgi. We looked for significant (E-value < 0.05) hits to repetitive sequence by BLASTing to the TIGR Plant Repeat Databases [84] with GDR [82]. We search for Helitron transposons using HelitronScanner [85].
A BAC library was prepared by Chris Saski, Clemson University Genomics Institute (CUGI) from 90 g leaf tissue collected at the University of Pittsburgh from Y33b2, the female parent of the F. virginiana ssp. virginiana linkage-mapping cross [42,51]. BAC construction methods followed Luo and Wing [86], with minor modifications. We designed overgo probes from the mapped male sterility region between Fvb6 positions 1.626 Mb and 1.794 Mb (S6 Table). We labeled probes individually with 32_P following the CUGI protocol (http://www.genome.clemson.edu/resources/protocols) and hybridized them to the BAC filters at 60°C overnight. This yielded 69 positive clones (S6 Table).
Genomic libraries from these 69 BACs were individually prepared and barcode indexed with the Illumina TruSeq DNA HT kit and sequenced with 150 bp paired-end reads on a single lane of Illumina MiSeq at Oregon State University CGRB. Reads were quality trimmed for both Q > 20 and Q > 30 with Trimmomatic [87] and merged, when possible, with the program FLASH [88]. We filtered merged reads and unmerged pairs by digital normalization at coverage of 100 using khmer [89]. For each library, both quality trimming sets were de novo assembled with Velvet [90] using a range of kmers from 31 to 91 bp. We selected the assembly with the longest contig for downstream analyses of each BAC (S6 Table).
We masked vectors with bedtools [91] and used BLAT to identify identical overlap of >1 kb among BACs. Groups of BACs representing putative homoeologs were imported into Geneious R7 [92] and further scaffolded manually. The resulting 11 assemblies were assigned to homoeologs (S4 Fig) by the presence of linkage-mapped SNPs observed in the target capture and microfluidic markers. We MAFFT [93]-aligned eight scaffolds (excluding non-overlapping scaffolds 6, 9, and 10) with a mean length of 47.993 kb. We removed all gap positions, resulting in a 19.819 kb alignment. We estimated a maximum likelihood tree with PhyML [94], confirming the identification of four pairs of homologous chromosomes (S5 Fig and S4 Data).
F1 offspring from the previously described F. virginiana ssp. virginiana cross “Y33b2×O477” [42,51] were sexed (N = 1,878) as described above and genotyped (N = 184) at sex-linked microsatellite markers [42] to identify possible recombinants, which were sequenced with targeted capture (N = 67) as previously described [41]. We designed Fluidigm microfluidic markers for fine-mapping Y33b2×O477 following our previous methodology [39]. We designed primer pairs for 48 amplicons with mean expected size of 385 bp—12 and 16 on the two BAC contigs corresponding to the Z homoeolog (S4 Fig) and 20 between Fvb6 positions 0.716 Mb to 17.605 Mb (S7 Table). We used the Fluidigm 48.48 Access Array Integrated Fluidic Circuits (IFCs) at the University of Idaho IBEST for amplicon library preparation following standard simplex reaction protocol. We pooled the amplicons of 190 F1 offspring and the two parents for paired-end 300 bp sequencing on a one-quarter lane of Illumina MiSeq. We trimmed reads as above, aligned them to Fvb and the BAC sequences using BWA version 0.7.12 [95], and called genotypes with POLiMAPS [41]. We identified recombinants and used these to define the narrowest possible window overlapping male function.
|
10.1371/journal.pgen.1002982 | Proteome-Wide Analysis of Disease-Associated SNPs That Show Allele-Specific Transcription Factor Binding | A causative role for single nucleotide polymorphisms (SNPs) in many genetic disorders has become evident through numerous genome-wide association studies. However, identification of these common causal variants and the molecular mechanisms underlying these associations remains a major challenge. Differential transcription factor binding at a SNP resulting in altered gene expression is one possible mechanism. Here we apply PWAS (“proteome-wide analysis of SNPs”), a methodology based on quantitative mass spectrometry that enables rapid screening of SNPs for differential transcription factor binding, to 12 SNPs that are highly associated with type 1 diabetes at the IL2RA locus, encoding the interleukin-2 receptor CD25. We report differential, allele-specific binding of the transcription factors RUNX1, LEF1, CREB, and TFAP4 to IL2RA SNPs rs12722508*A, rs12722522*C, rs41295061*A, and rs2104286*A and demonstrate the functional influence of RUNX1 at rs12722508 by reporter gene assay. Thus, PWAS may be able to contribute to our understanding of the molecular consequences of human genetic variability underpinning susceptibility to multi-factorial disease.
| Genome-wide association studies (GWAS) are a powerful approach to identifying genes contributing to risk of disease. However, individual mapped single nucleotide polymorphisms (SNPs) may not map close to a gene, and it can be difficult to distinguish marker SNPs from causal SNPs. Furthermore, the molecular mechanism responsible for disease association is usually not clear. Here we develop a method termed “proteome-wide analysis of SNPs” (PWAS) that identifies differentially binding transcription factors (TFs) and thereby helps to unravel the molecular mechanisms by which the SNPs may exert their effect on gene regulation. We use quantitative interaction proteomics to identify proteins with allele-specific binding. Applied to fine-mapped SNPs conferring risk in type 1 diabetes, PWAS revealed preferential binding of common transcription factors to certain disease-associated SNPs, suggesting they could be causal. In general, a proportion of causal SNPs are likely to function by mimicking binding motifs for transcription factors, increasing their occupancy and modulating gene expression. In addition, PWAS is streamlined and can be used as an informative follow-up approach to GWAS results.
| Genome-wide association studies (GWAS) of common diseases typically result in the identification of genomic susceptibility loci, in which several single nucleotide polymorphisms (SNPs) showing strong inter-marker linkage disequilibrium (LD) are equally associated with disease predisposition. Further fine-mapping and re-sequencing studies can then uncover additional SNPs, ideally including those that are causal in the disease etiopathogenesis [1], [2]. However, the SNPs that are most associated with the disease are commonly located in non-coding regions where they have no obvious function. Such SNPs presumably alter expression of a nearby gene via differential transcription factor (TF) binding or by influencing gene splicing. To date, there are few published examples in which a GWAS-identified SNP(s) is correlated with TF binding. We therefore set out to develop an unbiased, sensitive and streamlined method for detection of SNP sequences that differentially bind protein in an allele-specific manner. Affinity purification combined with mass spectrometry (AP-MS) can be a powerful tool to study protein interactions particularly when using a quantitative filter to distinguish specific interactors from the vast majority of background binders by their isotope ratio in the mass spectrometer [3], [4]. The binding of transcription factors to DNA is predominantly mediated by interactions with the phosphate backbone of the DNA. Analysis of differential interactions due to single nucleotide changes is challenging because SNP-related differences in binding affinity are typically low. As a consequence, binding differences are small, even for sequences mutated at multiple positions [5]–[8]. Here we describe PWAS, a technique to study differential transcription factor binding to nucleotide sequences in a streamlined manner. To demonstrate PWAS in a disease relevant context, we also report differential transcription factor binding to type 1 diabetes- (T1D-) associated SNPs at the IL2RA or CD25 locus.
We improved a recently described technology for DNA affinity capture by quantitative mass spectrometry [5] and developed a pipeline for routine screening of SNPs. To establish an automated protocol for SNP screening, we used a single low stringency buffer for immobilization of the oligonucleotides, incubation with the extracts, and washes. To counterbalance this increased complexity in the lysates, we increased the density of binding sites by concatenation of chemical synthesized DNA oligonucleotides resulting in a greater enrichment of transcription factors. A high enrichment is desirable in mass spectrometric experiments with data-dependent acquisition in order to ensure identification of the desired binding proteins among the majority of peptides originating from background proteins. The chosen TT/AA-overhang further allows incorporation of modified nucleotides by Klenow polymerase. Previously we had used a biotinylated oligonucleotide, which we removed by restriction enzyme cleavage [5], but this introduced a large amount of exogenous protein into the analyzed sample. Here, we performed strand-specific labeling with a desthiobiotin-analog that can be removed conveniently by competition with biotin (Figure 1). The desthiobiotinylated oligonucleotides of the two alleles were then incubated with either light or heavy nuclear extract in parallel. After mild washing, both bead fractions were combined prior to release of the desthiobiotin-labeled oligonucleotide by biotin. We found that PWAS detected differential binding to SNP alleles with great sensitivity. It employs approximately 40 bp of synthetic DNA containing either variant of the SNP, relatively small amounts of nuclear extract (200 µg) that are labeled by SILAC (stable isotope labeling by amino acid in cell culture) [9], [10], single, high resolution mass spectrometric runs, and proved sufficiently simple and robust to be automatable in a robotic format (Figure 1).
We benchmarked our system with a SELEX-derived TFAP2 binding site mutated at a single nucleotide position and a SNP (rs509813 C/G) in the promoter region of the muscarinic acetylcholine receptor M1 (CHRM1). This locus is associated with functional differences in gene expression and differential binding of an unidentified transcription factor [11]. While we were only able to visualize TFAP2 binding by immunostaining and not in a Coomassie stained gel (Figure 2A), by mass spectrometry we measured robust and reproducible differential binding of TFAP2 to its SELEX derived binding site with a SILAC ratio of 10 (Figure 2B). For rs509813, we found SP1 as well as SP3 as differential interactors. Both are predicted to bind to the sequence containing the C-allele SNP, but not SP2, which has a slightly different binding motif [12]. Furthermore, we detected binding of the transcription factor ZNF148 (also known as ZBP89) to this region containing rs509813. ZNF148 is a zinc finger protein which has not been predicted to interact with this site, but which has been reported to bind to SP1 binding sites in a mutually exclusive fashion [13] (Figure 2C).
Next, we applied the PWAS methodology to the complex, multi-SNP T1D susceptibility association of the IL2RA gene, encoding CD25 in the 10p14 region [1], [2] (Table 1). There are three SNPs in this region which together can be used to tag four common disease-associated haplotypes [1], representing a total of 12 SNPs. The tagging SNPs are rs12722495(A/G), rs11594656(A/T) and rs2104286(A/G). The haplotype (A,A,T) was associated with increased susceptibility to disease, whereas the three haplotypes (G,G,T); (A,A,A) and (A,G,T) were all associated with lower risk of type 1 diabetes. Importantly, these four common haplotypes have also been associated with differences in surface expression of CD25 in T cells, implying that IL2RA is a causal gene for T1D in this region [1]. Specifically, individuals with one or two T1D-protective rs12722495 alleles show 27% higher mean CD25 levels on their CD4+ memory T cells compared to fully susceptible individuals or donors with protective rs11594656 or rs2104286 alleles [1]. This is thought to be related to haplotype-dependent transcriptional differences altering CD25 expression, in turn leading to modulation of autoreactivity against pancreatic beta cells. However, it is not known which of the SNPs in the 10p14 region have a direct functional effect, and the identity of the specific transcription factor(s) responsible for this differential binding is equally unclear. Therefore, the precise causal variant(s) in this region has not been determined.
As preferential binding of transcription factors can occur on either allele, we performed two separate DNA pull-down experiments for each SNP. In the ‘forward’ experiment, the heavy SILAC labeled nuclear extract was incubated with one SNP allele and the light SILAC labeled extract with the other allele. In the ‘reverse’ experiment the SILAC label was switched. As visualized schematically in Figure 3, this strategy allows us to create a two dimensional interaction plot for each pull-down in which interaction partners are grouped into two of four quadrants. Contaminants such as keratins are unlabeled and are sorted into the lower left quadrant because they have a low SILAC ratio in both pull-downs.
We performed 48 SNP pull-downs with the 12 type 1 diabetes SNPs using extracts from the Jurkat T lymphocyte cell line, selected because type 1 diabetes is a T-cell mediated disease (Table S2). A very small number of proteins were clearly separated from the bulk of proteins that bound non-specifically to DNA or to the beads (Figure S1). These outliers were statistically significant in both pull-downs, with a combined forward and reverse pull-down p-value less than 10−7. For three SNPs we did not detect significant differential protein binding, suggesting that they may be non-causative and instead represent markers for the causal variant(s).
In the eight SNPs of group 1, we found RUNX1 (also known as CBFA) five-fold enriched at allele rs12722508*A (Figure 4B). RUNX1 is a transcriptional regulator likely to be involved in hematopoesis [14], which has already independently been linked to risk of autoimmune disease [15]. Notably, one of the two other significant binders to this SNP is CBFB, which is known to form a heterodimer with RUNX1, underscoring the specificity of our screen [16]. The third differential interactor is SAFB1, a less characterized transcription factor reported to be important in transcriptional regulation of HSP27 and ERα [17]. Additionally, LEF1, a key transcription factor in the Wnt signaling pathway [18] involved in regulating T cell specific genes [19], interacted with rs41295061*A (Figure 4C) in our SNP pull-down experiments. The transcriptional regulators CREB and TFAP4 differentially interacted with rs12722522*C (Figure 4A). The SNP variant rs11597367*G of the three SNP containing group 2 bound ZNF148 and CGGBP1 specifically. We hypothesize that these genotype-dependent interactions are part of the molecular mechanism responsible for the association between these SNPs and expression of CD25 in naïve T cells and stimulated monocytes [1].
Three of the identified transcription factors have known DNA consensus motifs and we therefore investigated whether these motifs were present in the DNA fragment to which they bound. For CREB1 this was indeed the case: the sequence around rs12722522*C (CGTCA) when reverse complemented reconstituted the binding motif TGACG [20], whereas the other allele was TGACA (Figure 4A). Interestingly, for the other two cases, the region around the SNP did not reconstitute the deposited consensus sequence completely, but generated an additional mismatch (Figure 4B, 4C). However, we note that for RUNX1, the consensus sequence (TGTGGBH) for its murine homologue obtained in a recently published ChIP-seq experiment [21] matches the sequence around rs12722508*A (ACCCACA) when reverse complemented (TGTGGGT).
RUNX1 has previously been implicated in other autoimmune diseases [15], which prompted us to further validate this transcription factor - SNP association. We reproduced the allele-specific binding of RUNX1 detected in our mass spectrometric assay by immunostaining (Figure 5A) and investigated the effects on transcription in a transactivation assay. To test whether changes in RUNX1 level would act differently on rs12722508, we reduced the level of this transcription factor by RNA interference (Figure 5B). Upon knock-down of RUNX1, we observed an allele-specific activation of our reporter construct which was not observed when expression levels of control transcription factors were reduced (Figure 5C). Since RUNX1 binds to both alleles, it upregulated both SNP variants; however, consistent with allele-specific differential binding, the upregulation was different between the alleles: rs12722508*A by 16±7 percent and rs12722508*T by 34±11 percent compared to mock (P = 0.016). These results link the allele-specific binding of RUNX1 detected in our mass spectrometry screen to functional differences in transcriptional activity.
Our results indicate that differential transcription factor binding to candidate causal SNPs can indicate which SNPs and which TFs might be involved in the causal mechanism(s) from gene-to-protein expression. In the case of SNP group 1 of the IL2RA type 1 diabetes locus [1], three of the SNPs differentially bind common transcription factors (rs12722508*A, rs12722522*C and rs41295061*A) implying that there may be more than one SNP within SNP group 1 affecting IL2RA transcription. Extrapolating from the SILAC ratios, occupancy of the three SNPs rs12722508*A, rs12722522*C and rs41295061*A was altered between four and eight-fold for these transcription factors. Using reporter assays, we have shown that RUNX1 can mediate allele-specific expression via rs12722508. The relatively small difference measured in our reporter gene assay, and the fact that several SNPs in the haplotype also show differential binding to common TFs, offer a plausible explanation for the observed expression difference of 30% in cell surface CD25 expression [1], [2]. Our results indicate that several sites in a haplotype may contribute to differential transcription factor binding in a cumulative manner, as suggested here for three of eight SNPs of group 1. Supporting such a scenario, genome-wide chromatin immunoprecipitation studies have shown that common transcription factors typically occupy thousands of sites in the genome [22]. We propose that multiple SNPs in a haplotype cooperatively modulate expression levels of nearby genes, contributing to individual traits, including risk of common diseases.
Large scale differential transcription factor binding at SNPs has previously been reported for NFkB and PolII using allele-specific ChIP [23]. Our unbiased, SNP-centered, PWAS approach is orthogonal and complementary to the protein-centered ChIP-seq method and links the detected SNPs from genomics studies directly to the protein without a priori knowledge. In conclusion, PWAS is specific, reproducible and generic and only requires synthesis of 40 mers of DNA and batch labeling of cells without the need to obtain cell lines with matching haplotypes. The throughput is currently up to five SNP pairs per day and per mass spectrometer. PWAS may contribute evidence that a given variant is causal. It can thus help to select a subset of polymorphisms from a much larger candidate set that cannot be distinguished by genetic association mapping. Positive results from PWAS directly suggest further gene-phenotype associations that can be investigated to extend the molecular chain of events at least one step further than the GWAS-mapped SNPs themselves.
Hela S3 and Jurkat cells were SILAC-labeled in RPMI 1640 (-Arg, -Lys) medium containing 10% dialyzed fetal bovine serum (Gibco) supplemented with 84 µg/ml 13C615N4 L-arginine and 40 µg/ml 13C615N2 L-lysine (Sigma Isotec or Cambridge Isotope Labs) or the corresponding non-labeled amino acids, respectively. Nuclear extracts were prepared essentially as described [24].
25 µg of corresponding pairs of oligonucleotides (Table S1) were annealed and phosphorylated for 2 h at 37°C in the presence of polynucleotide kinase (Fermentas) in 1× T4 ligase buffer. 20 Units T4 ligase (Fermentas) was added to the reaction and incubated at RT overnight. Polymerisation was monitored by agarose gel electrophoresis of a small aliquot of the reaction mixture. After subsequent chloroform-phenol-extraction, the concatemerized oligonucleotides were desthiobiotinylated with d-desthiobiotin-(N6-(6-Amino)hexyl)-dATP (custom synthesis, Jena Biosciences) using 30 units Klenow fragment (Fermentas) overnight at 37°C. Unreacted desthiobiotin nucleotides were removed by size exclusion using a G50 column (GE Healthcare) according to the manufacturer's protocol. The baits were stored at −20°C.
DNA oligonucleotides were immobilized on 50 µl Dynabeads MyOne Streptavidin C1 (Invitrogen) and subsequently incubated with 200 µg of SILAC-labeled nuclear extract (light and heavy separately) in PBB buffer (150 mM NaCl, 50 mM Tris/HCl pH 8.0, 10 mM MgCl2, 0.5 percent NP-40, Complete Protease Inhibitor-EDTA [Roche]) for 2 hours at 4°C in a rotation wheel. After three times washing with PBB, bead fractions were pooled and bound DNA-protein complexes were eluted at RT with 200 µl PBB containing 16 mM biotin. The identical steps were automated on a TECAN EVO workstation equipped with a 4 channel LiHAN, robotic arm, cooled buffer storage reservoirs, a temperature controlled shaker and a magnetic separator. Incubations were performed in a 96 well plate (Nunc). Instead of incubation on a rotation wheel, the automated TECAN shaker was operated at 8°C and 700 rpm. Proteins in the elution fraction were precipitated with ethanol and resolubilized in 20 µl 8 M urea for MS analysis.
Samples were reduced in 0.5 mM DTT (Sigma) for 30 min, alkylated with 3 mM iodoacetamide (Sigma) for another 30 min and subsequently digested with trypsin (Promega) overnight at room temperature. The digested protein mixture was diluted in 50 mM ammonium bicarbonate buffer pH 8/0.5% TFA and loaded onto a self-made stage tip. MS analysis was performed essentially as previously described [25]. In short, peptides were eluted from stage tips and analyzed by nanoflow liquid chromatography on an EASY-nLC system from Proxeon Biosystems into a LTQ-Orbitrap XL (Thermo Fisher Scientific). Peptides were separated on a C18-reversed phase column packed with Reprosil (Dr. Maisch) directly mounted on the electrospray ion source on an LTQ-Orbitrap XL. We used a 140 min gradient from 2% to 60% acetonitrile in 0.5% acetic acid at a flow of 200 nl/min. The LTQ-Orbitrap XL was operated with a Top5 MS/MS spectra acquisition method in the linear ion trap per MS full scan in the orbitrap.
The raw files were processed with MaxQuant [26] (version 1.0.12.27) and searched with the Mascot search engine (version 2.2.4.1, Matrix Science) against a IPI human v3.37 protein database concatenated with a decoy of reversed sequences. Carbamidomethylation was set as fixed modification while methionine oxidation and protein N-acetylation were included as variable modifications. The search was performed with an initial mass tolerance of 7 ppm for the precursor ion and 0.5 Da for the MS/MS spectra. Search results were processed with MaxQuant filtered with a false discovery rate of 0.01. Prior to statistical analysis, known contaminants and reverse hits were removed. Proteins identified with at least 1 unique peptide and minimum 2 quantitation events in a single pull-down were considered for analysis and plotted in R (prerelease version 2.8.0).
Transcription factors identified for differential SNP binding were checked for deposited recognition sequences in the JASPAR database (www.jaspar.com).
For western blotting, proteins were solubilized in LDS sample buffer (Invitrogen), boiled for 5 min at 85°C and fractionated on a 4–12 percent NOVEX gradient gel using MOPS buffer (Invitrogen). Proteins were transferred to a Protran 85 membrane (Whatman) in a blotting chamber (Biorad) at 300 mA for 1 h. The membrane was blocked with PBST (0.1%)+4 percent low fat milk (Roth) for 15 min prior to incubation with the primary antibody for 1 h at room temperature. The following antibodies were diluted in PBST (0.1%) with 4 percent low fat milk: SP1 [1∶1000], TFAP2 [1∶2000], RUNX1 [1∶1000] and LEF1 [1∶1000] (all Abcam). Membrane was washed with PBST (0.1%) three times prior to incubation with either HRP-anti-mouse or HRP-anti-rabbit antibody (both Amersham) for 1 h at room temperature. For detection ECL Western blotting detection reagent (Amersham) was used according to manufacturer's instruction. Chemiluminescence screens (GE Healthcare) were used to visualize the band patterns.
Endoribonuclease-prepared short interfering RNAs (esiRNAs) were produced as previously described [27] with the following primers: Rluc (for: GGATAACTGGTCCGCAGTGGT, rev: CCCATTCATCCCATGATTCAA); TFAP4 (for: GTGCCCTCTTTGCAACATTT, rev: TTCTCGTCCTCCCAGATGTC); RUNX1 (for: GGCTGGCAATGATGAAAACT, rev: GATGGTTGGATCTGCCTTGT); CREB1 (for: GGAGTGCCAAGGATTGAAGA, rev: CCTCTCTCTTTCGTGCTGCT); CBFB (for: TTTGAAGGCTCCCATGATTC, rev: CCATGGCAGTTTGTGATGTC).
The heat shock promoter was amplified (Hsp68_for: GAGAAAGCTTCAGGAACATCCAAACTGAGCA, Hsp68_rev: GAGAAAGCTTCGCTTGTCTCTGGATGGAAC) from genomic DNA prepared from C2C12 cells. The promoter was cloned into pGL3-basic (Promega) in front of the firefly luciferase gene at the HindIII site. A gateway cassette was inserted 5′ of the promoter to facilitate insertion of different DNA sequences, generating the pGL3/GW/mHsp68prom vector. A triple repeat of the sequence surrounding rs12722508 (AACCCACCCAC[A/T]GAAACTATCAGAG) was cloned into pCR8 and transferred into pGL3/GW/mHsp69prom using LR recombination.
100,000 cells HeLa Kyoto were reverse transfected with 150 ng esiRNA and 1 µl oligofectamine (Invitrogen) in complete medium in a 24 well plate. After 16 hours, 4 ng Rhenilla luciferase vector phRL-TK (Promega) were co-transfected with 200 ng firefly reporter (pGL3/GW3/mHsp69prom) containing either SNP variant with Optimem medium (Gibco). Medium was exchanged against complete medium 4 hours past transfection. Cells were harvest the next day using passive lysis buffer (Promega). Independent biological triplicates were measured on the GloMax luminometer (Promega) using the Dual Luciferase Assay kit (Promega) according to the manufacturer's protocol. For representation, the mean and the standard deviation were calculated. Transfection efficiency was monitored by parallel transfection of RLuc esiRNA targeting rhenilla luciferase.
Cells form one well (24 well plate) transfected with esiRNA and reporter assay constructs were trypsinized, washed in PBS and RNA extracted using a spin filter kit (UBS). Around 1 µg of pure RNA were reverse-transcribed with polyT primers using the Fast cDNA kit (Fermentas). The following primers were used for the amplification: TFAP4 (for: GCAGACAGCCGAGTACATCTT, rev: CCTATGCCTTCGTCCTTGTCC), RUNX1 (for: GATGGCACTCTGGTCACTGTGA, rev: CTTCATGGCTGCGGTAGCAT), CBFB (for: AGAAGCAAGTTCGAGAACGAG, rev: GAAGCCCGTGTACTTAATCTCAC), CREB (for: CACCTGCCATCACCACTGTAA, rev: GCTGCATTGGTCATGGTTAATGT), GAPDH (for: TGCACCACCAACTGCTTAGC, rev: GGCATGGACTGTGGTCATGAG). Three independent replicates using the IQ SybrGreen supermix (Biorad) were measured on a CFX96 qRT-PCR machine (Biorad). For assessment of knock-down the ΔΔCt method was used.
|
10.1371/journal.pgen.1006469 | Whole-Organism Developmental Expression Profiling Identifies RAB-28 as a Novel Ciliary GTPase Associated with the BBSome and Intraflagellar Transport | Primary cilia are specialised sensory and developmental signalling devices extending from the surface of most eukaryotic cells. Defects in these organelles cause inherited human disorders (ciliopathies) such as retinitis pigmentosa and Bardet-Biedl syndrome (BBS), frequently affecting many physiological and developmental processes across multiple organs. Cilium formation, maintenance and function depend on intracellular transport systems such as intraflagellar transport (IFT), which is driven by kinesin-2 and IFT-dynein motors and regulated by the Bardet-Biedl syndrome (BBS) cargo-adaptor protein complex, or BBSome. To identify new cilium-associated genes, we employed the nematode C. elegans, where ciliogenesis occurs within a short timespan during late embryogenesis when most sensory neurons differentiate. Using whole-organism RNA-Seq libraries, we discovered a signature expression profile highly enriched for transcripts of known ciliary proteins, including FAM-161 (FAM161A orthologue), CCDC-104 (CCDC104), and RPI-1 (RP1/RP1L1), which we confirm are cilium-localised in worms. From a list of 185 candidate ciliary genes, we uncover orthologues of human MAP9, YAP, CCDC149, and RAB28 as conserved cilium-associated components. Further analyses of C. elegans RAB-28, recently associated with autosomal-recessive cone-rod dystrophy, reveal that this small GTPase is exclusively expressed in ciliated neurons where it dynamically associates with IFT trains. Whereas inactive GDP-bound RAB-28 displays no IFT movement and diffuse localisation, GTP-bound (activated) RAB-28 concentrates at the periciliary membrane in a BBSome-dependent manner and undergoes bidirectional IFT. Functional analyses reveal that whilst cilium structure, sensory function and IFT are seemingly normal in a rab-28 null allele, overexpression of predicted GDP or GTP locked variants of RAB-28 perturbs cilium and sensory pore morphogenesis and function. Collectively, our findings present a new approach for identifying ciliary proteins, and unveil RAB28, a GTPase most closely related to the BBS protein RABL4/IFT27, as an IFT-associated cargo with BBSome-dependent cell autonomous and non-autonomous functions at the ciliary base.
| Ciliopathies are genetic disorders that arise from loss or mutation of genes that encode proteins which play roles in the biology of cilia, organelles found on most of the cells in the human body. Ciliopathy-associated ailments include–but are not limited to–kidney dysfunction, blindness, skeletal abnormalities, as well as brain disorders. Although a great number of cilium-targeted proteins are known, it is thought that a large proportion remain unidentified. Here, we use a developmental gene expression series to discover novel cilia genes in the nematode Caenorhabditis elegans. We present several cilium-localised proteins resulting from our analysis, including RAB-28, a GTPase previously implicated in the degenerative eye disease known as cone-rod dystrophy. Through live videomicroscopy, we show that RAB-28 undergoes bidirectional transport within the cilium. A RAB-28 inactivating mutation results in loss of transport, while an activating mutation results in stronger localisation at the ciliary base and robust transport, although overexpression results in a variety of cilia-related defects. Both the wild type and activating mutant proteins require the Bardet-Biedl Syndrome-related complex of proteins for their transport, linking RAB-28 to an established ciliary transport machinery.
| The cilium is a conserved organelle, inferred to have existed in the last eukaryotic common ancestor (LECA) and now present in most extant protists, as well as all multicellular animals. Motile cilia generate cell movement or fluid flow, whereas non-motile (primary) cilia have evolved as specialised ‘antennae’ that capture extracellular sensory cues and orchestrate extrinsic signal transduction pathways linked to development (e.g., Sonic hedgehog) [1,2]. Cilium dysfunction in humans is associated with a growing number of so-called ciliopathies that affect virtually all physiological and developmental functions [3]. For example, Bardet-Biedl syndrome (BBS) includes retinal degeneration, cystic kidneys, obesity and skeletal anomalies (polydactyly) as primary ailments [4].
Cilia are subdivided into distinct subcompartments, each with unique structural and functional features, as well as molecular compositions [5]. The canonical cilium of 9 doublet microtubules (MTs) extends from a mother centriole-derived basal body, which connects via distal appendages (transition fibers) to the plasma membrane. The proximal-most 0.2–1.0 μm of the axoneme, called the transition zone, functions in early ciliogenesis, and together with basal body structures provides a permeability barrier that separates the ciliary cytosol and membrane from the cell body [6–8]. Additional subregions include the inversin and distal tip compartments, as well as the ciliary pocket, which is a depression of the periciliary membrane where the basal body is rooted [5]. Many ciliopathy proteins and associated complexes localise to particular ciliary subcompartments, where they conduct subdomain-specific functions [5,9].
Cilia rely on various intracellular transport systems to sort and deliver the protein cargo required for cilium formation, maintenance and function [10]. The best understood is intraflagellar transport (IFT), which consists of large macro-molecular assemblies that move bidirectionally between the ciliary base and tip, driven by kinesin-2 anterograde (base to tip) and IFT-dynein retrograde (tip to base) motors [11–13]. Associated with the motors—and essential for IFT—are the IFT-A and IFT-B complexes, which likely serve as cargo adaptors [11]. The IFT-associated BBS complex (BBSome) also tethers ciliary cargo and regulates the coupling of IFT-A and IFT-B complexes [14]. Also important are membrane trafficking pathways that regulate vesicle formation and transport between post-Golgi sorting stations and the periciliary membrane, as well as endocytic retrieval and recycling events at the periciliary membrane and within the ciliary pocket [15–19]. Various IFT and ciliary membrane trafficking regulators have been identified, including small GTPases of the RAB, ARF and ARL families, that function during early cilium formation as well as transport events post-ciliogenesis [10,20,21].
Given the multifaceted roles of cilia, together with its prevalent disease association, there have been major efforts to identify the ‘ciliome’, or complete molecular parts list of cilia, using a wide range of cell types and organisms [22–24]. Approaches have included comparative genomics of ciliated versus non-ciliated species [25], identification of binding sites for the ciliogenic transcription factors DAF-19/RFX or FOXJ1 [26–31], expression analyses involving microarray, serial analysis of gene expression (SAGE) and RNA-Seq [25,28,32–37], as well as proteomics [38–41]. Data from such studies are compiled in the online ciliary database, Cildb [23,24]. Whilst the studies have contributed immensely to understanding cilia biology, each approach has limitations and additional ciliary components almost certainly remain unidentified.
C. elegans represents a powerful genetic model for investigating cilium formation and function [22]. Hermaphrodite worms possess 60 ciliated cells (of 960 total), all of which are sensory neurons. The non-motile sensory cilia extend from the dendritic tips and many are contained within bilateral chemo- and thermo-sensory cuticular organs, supported by glial cell (sheath and socket) processes that establish environmentally exposed channels [42–44]. C. elegans cilium morphologies range from the canonical rod-like to forked, multi-branched and membrane-expanded structures [22]. Worm cilia also possess ultrastructural features conserved in vertebrate/mammalian cilia; for example, amphid (head) and phasmid (tail) channel cilia possess long A-tubule extensions that establish a proximal axonemal region or ‘middle segment’ of 9 outer doublet MTs and a ‘distal segment’ of 9 outer singlet MTs [22]. Because many ciliary genes and pathways are conserved in nematodes, and complete loss of cilia is non-lethal [26], C. elegans has been a leading metazoan model for discovering new ciliary genes and uncovering new insight into ciliary transport, function and disease mechanisms.
In this study, we identified a unique expression profile for ciliary genes using a series of RNA-Seq libraries generated specifically to improve annotation of the transcriptome [45–47]. We confirmed that our clustering analysis identifies known ciliary proteins, including several not previously studied in C. elegans, and uncovers novel conserved ciliary proteins. One of these proteins is RAB-28, which is expressed exclusively in ciliated cells, where it associates with the periciliary membrane and behaves as an IFT cargo via BBSome- and nucleotide binding-dependent mechanisms. Overexpression of predicted active or inactive forms of RAB-28 leads to variant-specific ciliary and cell non-autonomous sensory pore morphogenesis defects. Together, our work provides a novel approach to finding new ciliary proteins, and uncovers a functional association between the BBSome, IFT and the orthologue of the cone-rod dystrophy protein, RAB28.
To provide a complementary approach to ciliary gene discovery, we took advantage of the temporally-invariant birth of all C. elegans cells and tissues, including ciliated neurons, during development [48] (Fig 1A). Nearly all of the 60 ciliated neuronal cell types in C. elegans hermaphrodites are born within a discrete embryonic time period 300–450 minutes post-fertilisation, with cilium formation occurring very shortly thereafter (Fig 1A). We hypothesised that ciliogenesis genes are highly expressed during this time period and therefore distinguishable from genes required for general neuronal formation and development, which are expressed during a broader time window (Fig 1A). Using an available whole-organism developmental series of RNA-seq libraries from C. elegans [45–47], we confirmed this hypothesis: many well-characterised cilia genes are highly expressed in the early embryo, display peak expression in the late embryo and first larval stage, and show greatly reduced expression during subsequent developmental stages (Fig 1B).
Next we sought to identify novel ciliary/ciliogenic genes displaying a similar cilia-related expression pattern. Using a set of 41 well-characterised ciliary component genes as “baits”, representing the cilia-related gene expression profile during development (sheet 1 in S1 Table), we queried our ciliary transcriptome to identify other gene “preys” with similar expression profiles across the RNA-Seq libraries (Fig 1B and sheet 2 in S1 Table). Hierarchical clustering of genes based on temporal expression reveals a tight cluster of 34 of the bait genes with 151 prey genes (cluster 1), many of which are uncharacterised (S1 Fig and sheet 3 in S1 Table). We also filtered the gene list to only include those with human orthologues [49]. To validate our dataset, we determined if filtered cluster 1 is enriched for genes with >12 hits in the ciliary database Cildb [23,24] (sheets 3 and 4 in S1 Table). We found that conserved Cildb-represented genes are significantly enriched (84.5 fold) in filtered cluster 1 compared to the entire genome (p<0.0001, chi-squared test; sheet 4 in S1 Table). We further compared the cluster 1 gene set to two previous expression studies [25,50] and see significant overlap between lists (p<0.0001; sheet 4 in S1 Table).
Because our genes were identified by shared temporal gene expression, we reasoned that they may share common promoter regulatory elements beyond the previously identified X-box motif [27,28] and so we performed a motif discovery analysis on the conserved genes in cluster 1 (sheet 3 in S1 Table). Not surprisingly, we find several motifs that are significantly enriched, including matches to the GAGA factor binding site (S2 Fig). As might be expected, the 14-bp X-box was not identified by the motif elucidation program despite its enrichment in our dataset (sheet 4 in S1 Table), likely due to the number of degenerate bases in its consensus sequence and because the program identifies motifs only up to 8-bp long [26–28,51]. None of the elements we identified have been previously associated with cilia gene expression.
We observed that chemoreceptors (G protein-coupled receptors or GPCRs) [52], many of which are known to be cilium-localised, are largely absent from our ciliary dataset (sheet 3 in S1 Table). We hypothesised that such genes might be transcribed post-ciliogenesis. Using a statistical predictive modelling strategy similar to that used above, we identified 80 genes sharing the expression profile of srg-36, a cilium-localised dauer pheromone receptor required for entry into the dauer larval diapause life stage (Fig 1C and sheet 5 in S1 Table) [53]. Of the 80 genes, 27 are serpentine transmembrane receptors, representing a significant 7-fold enrichment over all such genes in the genome (p<0.0001, chi-squared test). Analysis of the upstream regulatory regions revealed that 25 of the 28 chemoreceptor genes (including srg-36) possess an E-box regulatory element (S3 Fig) [54]. This suggests that their similar temporal expression profiles stem from regulation by the same, or similar, transcription factors such as the E-box-binding Basic Helix-Loop-Helix bHLH transcription factor [54]. Interestingly, 4 of the predicted serpentine receptors, including srg-36, are known to be expressed in ciliated sensory neurons [53,55,56]. We note that the rise and fall of chemosensory/GPCR gene expression trails that of the cilia-related genes (Fig 1C), indicating that these two expression profiles of related developmental processes are temporally discrete.
To confirm that our predictive expression profiling model identifies novel ciliary proteins in C. elegans, we used GFP reporters to investigate the cell and tissue expression patterns, and subcellular protein localisations, of seven previously uncharacterised C. elegans proteins from filtered cluster 1, namely FAM-161 (Y38H6C.14), CCDC-104 (Y108G3Al.3), RPI-1 (W07G1.5), RAB-28 (Y11D7A.4), CCDC-149 (F29G6.2), MAPH-9 (C34D4.1) and YAP-1 (F13E6.4). These were chosen because at the onset of this study there was little or no published evidence of ciliary associations for most of these proteins in any system. Specifically, we made transcriptional ‘promoter fusion’ reporters (endogenous gene promoter fused to GFP) for CCDC-104, CCDC-149 and RAB-28, and translational ‘protein fusion’ reporters (endogenous gene promoter + genomic exon/intron or cDNA sequence fused to GFP) for all 7 candidates. The only exception was the YAP-1 translational reporter, where the bbs-8 gene promoter active only in ciliated cells [57] was used because of the widespread expression of YAP-1 in multiple tissues and cell types [58]. We also made a second RAB-28 translational reporter driven by the bbs-8 gene promoter.
All reporters employing the endogenous gene promoter show enriched or almost exclusive expression within ciliated cells (S4A Fig), thus validating the predictions from our coexpression profiling. Specifically, maph-9, rpi-1, ccdc-104, rab-28, and fam-161 are exclusively expressed in ciliated neurons, with ccdc-149 expressed in most ciliated neurons as well as pharyngeal neurons, touch receptor neurons, and motor neurons as previously reported (Figs 2 and S4A) [59]. Furthermore, the translational reporters reveal that all 7 proteins localise at the base of, or within, cilia (Fig 2 and Fig 3A). FAM-161::GFP signals are found at the ciliary base and within the proximal part of the ciliary axoneme, including the transition zone (TZ) (Fig 2). This localisation is similar to that of the mammalian protein implicated in retinitis pigmentosa, and is consistent with the suggested role of FAM161b in cargo delivery to photoreceptor outer segments (cilia) [60,61]. Both GFP reporters for RAB-28 (driven by bbs-8 gene or endogenous promoter), the orthologue of the small ciliary GTPase RAB28 linked to human autosomal-recessive cone-rod dystrophy [62–64], localise along the entire cilium. (Fig 2 and Fig 3A). Ciliary axoneme localisations are also observed for the GFP-tagged RPI-1, MAPH-9, YAP-1 and CCDC-104 translational reporters; in contrast, CCDC-149::GFP is absent from the ciliary axoneme, although a pool of signal is evident at the ciliary base (Fig 2). The ciliary localisation of CCDC-104::GFP is consistent with the mammalian ciliary base and axonemal localisations reported for the ARL3 interacting protein, CCDC104/BARTL1 [65]. We also found that RPI-1::GFP appears to localise to ciliary and cytoplasmic microtubules, similar to its mammalian orthologue (RP1) implicated in retinitis pigmentosa (Fig 2) [66]. Our finding of ciliary localisations for YAP-1, MAPH-9 and CCDC-149 are the first in any cellular system or organism, and is consistent with a known ciliogenesis role for mammalian YAP, and cilia-related phenotypes in MAP9 (MAPH9 orthologue)-disrupted zebrafish and dachshunds [67–70]. YAP is regulated by the ciliary protein NPHP4, and MAP9 is phosphorylated by the cilia-disassembly/centrosomal kinase PLK1 [71,72].
We also examined genes excluded from filtered cluster 1 due to a reported lack of predicted human orthologues (according to the Ortholist database; [49]). Highly ranked in unfiltered cluster 1 is an uncharacterised gene with an X-box promoter element [26], tza-3, which we demonstrate is expressed exclusively in ciliated neurons, and encodes a ciliary TZ protein (S4B Fig). Whilst TZA-3 is not conserved, the TZ localisation of this predicted transmembrane protein depends on the highly conserved core TZ scaffolding protein MKS-5 and “MKS module” components MKSR-1 and MKSR-2, but not the “NPHP module” protein NPHP-4 (S4B Fig). These findings are consistent with TZA-3 associating with the MKS module, which plays a role in ciliary gating [6–8]. Thus, our expression profiling approach identifies non-conserved ciliary proteins in addition to conserved ciliary proteins.
Of the novel C. elegans ciliary proteins we identified, we sought to characterise RAB-28 further for the following reasons: (1) human RAB28 is associated with a possible ciliopathy (cone-rod dystrophy [62,63]), (2) many small GTPases play essential roles in ciliogenesis and ciliary membrane trafficking pathways [10], and (3) aside from reports of RAB28 localising at the basal body [62] or the cilium [64], there is a complete absence of molecular studies on this evolutionarily conserved protein that support a functional and mechanistic link to cilia.
We further investigated the ciliary localisation of GFP::RAB-28 using time-lapse imaging and found that this GTPase undergoes continuous, bidirectional, IFT-like movement along ciliary axonemes (S1 Movie). In contrast, the additional non-ciliary GFP::RAB-28 signals present throughout the neurons are diffuse, with no processive movement detected. In C. elegans amphid and phasmid cilia, anterograde IFT is driven by two kinesin-2 motors (kinesin-II and OSM-3) to yield distinct anterograde rates along the middle and distal segments [14,73]. Kymograph analysis of phasmid (tail) cilia confirm that ciliary GFP::RAB-28 moves at IFT-associated velocities, displaying characteristic average anterograde rates of ~0.7 μm/sec (middle segment; proximal part of axoneme) and ~1.2 μm/sec (distal segment; distal part of axoneme), and retrograde rates of ~1.4 μm/sec (along entire cilium length) (Fig 3A and 3B and S1 Movie) [73]. To examine if this ciliary trafficking is truly associated with IFT, we examined GFP::RAB-28 in worms with disrupted CHE-11 (IFT140), which is a component of IFT-A essential for IFT [28,74,75]. Although GFP::RAB-28 is observed within the truncated cilia of che-11 mutants, we could not detect processive movement of the GFP signals, indicating that the trafficking behaviour of RAB-28 within cilia is bona fide IFT (Fig 3A and S1 Movie). In contrast to observations for other IFT-associated proteins in IFT deficient-worms [28,74,75], GFP::RAB-28 does not abnormally accumulate at the ciliary base or tip, or along the axoneme of these animals (Fig 3A). This lack of accumulation, despite the active transport defect, suggests that our GFP::RAB-28 reporter also freely diffuses within (and between) the ciliary and dendritic compartments, in addition to undergoing IFT. We confirmed this hypothesis using a fluorescence recovery after photobleaching assay in wild type and che-11 mutant worms, which show that GFP::RAB-28 is highly mobile, displaying very rapid exchange kinetics between ciliary and dendritic pools (Fig 3C). Thus, in addition to active transport via IFT, GFP::RAB-28 undergoes IFT-independent free diffusion, which explains why this reporter does not abnormally accumulate within che-11 mutant cilia.
To determine if RAB-28 is required for IFT, we examined ciliary structure, function and protein transport in worms containing a deletion (gk1040) in rab-28. The gk1040 mutation removes the GTP-binding switch II domain and the farnesylated C-terminal CAAX motif (S5A Fig), both of which are critical for RAB protein function [76]; thus, gk1040 is likely a severe loss-of-function or null allele. We outcrossed the gk1040 background at least 3 times with wild type worms to remove background mutations unlinked to gk1040. We find that ciliated amphid and phasmid sensory neurons in rab-28 mutant worms display a normal dye filling response, which suggests that ciliary structures are intact (short cilia usually abrogate dye uptake due to lack of environmental exposure) [77] (S5B Fig). Localisation of an IFT protein, OSM-6 (IFT52 orthologue), throughout normal-length cilia confirms this finding (S5C Fig). In addition, transmission electron microscopy (TEM) analyses reveal that ciliary ultrastructures appear normal in rab-28 mutants (S5D Fig). Furthermore, these worms are normal for cilia-related sensory behaviours (S5E Fig), as well as cilium-dependent carbon dioxide avoidance and development (body size) (S5F Fig) [78–80]. Finally, the localisation (and movement) of BBSome (BBS-5), IFT-A/B and various ciliary membrane proteins is also grossly normal in rab-28 mutants (S5G Fig). Thus, unlike the disruption of IFT proteins, loss of RAB-28 does not affect ciliary structure, transport, or function, at least for those proteins and cilia that were analysed. We conclude, therefore, that RAB-28 behaves more like an IFT cargo, or peripherally-associated component of IFT complexes, rather than a central component of the machinery required for bidirectional movement (IFT).
Next, we assessed the requirements of GDP and GTP nucleotide binding for RAB-28 localisation and transport. We made GFP-tagged constructs containing T49N or Q95L mutations in RAB-28, which are predicted to trap the GTPase in GDP (inactive) or GTP (active)-locked states, respectively [81]. These constructs were injected into wild type worms at the same concentration (5 ng/μl) as the GFP::RAB-28(WT) construct described above (Fig 3), and all appear to be expressed at similar levels. Like the RAB-28(WT) reporter, RAB-28(T49N) (hereafter termed RAB-28(GDP)) displays diffuse localisation throughout the entire neuron, including the cilium and its associated periciliary membrane compartment (PCMC) (Fig 4A). However, unlike wild type RAB-28, RAB-28(GDP) fails to undergo detectable IFT, indicating a highly reduced ability to associate with IFT trains (Fig 4A and S1 Movie). In contrast, the localisation of the RAB-28(Q95L) (hereafter termed RAB-28(GTP)) is highly enriched at the PCMC compared with the rest of the cell, forming a ‘tunnel-like’ localisation indicative of association with the periciliary membrane (Fig 4A). RAB-28(GTP) also undergoes clearly detectable IFT, albeit at an apparently reduced frequency compared to the wild type reporter (Fig 4A and S1 Movie). Notably, the localisation and IFT behaviour of all three GFP::RAB-28 reporters (WT, GDP-locked, GTP-locked) is not altered when expressed in the rab-28(gk1040) null background (S6 Fig), indicating that endogenous RAB-28 does not affect the distribution and transport behaviour of the RAB-28 markers. Furthermore, transgenic worms expressing GFP::RAB-28(GTP) at a very low level (injected at 0.5 ng/μl; oqEx304) is exclusively localised at the periciliary membrane and undergoes IFT (S7 Fig and S2 Movie), thus confirming our observations with the higher expressing marker, and indicating that the periciliary membrane is the primary site of activated RAB-28 function in nematode sensory neurons (Fig 4A). Taken together, these data show that GTP binding targets RAB-28 to the periciliary membrane and facilitates its association with IFT trains.
These observations for C. elegans RAB-28 are reminiscent of those for mammalian RAB8, whose ciliary targeting also depends on GDP-GTP exchange, via a mechanism involving RABIN8 (RAB8 guanine nucleotide exchange factor) and BBSome functions at the base of cilia [82–84]. We therefore examined the localisation and transport behaviour of all three GFP-tagged RAB-28 reporters (WT, GDP-locked and GTP-locked) in a likely null allele (nx77) of a C. elegans BBSome gene orthologue, bbs-8. The bbs-8(nx77) mutation does not alter the ciliary targeting and distributions of RAB-28(WT) and RAB-28(GDP) compared to a wild type background (Fig 4B). In contrast, BBS-8 loss has a striking effect on both of our RAB-28(GTP) reporters, preventing their accumulation at periciliary membranes (Fig 4B and S7 Fig). Furthermore, the IFT movement of RAB-28(WT) and RAB-28(GTP) is abolished in bbs-8 mutant cilia (Fig 4B and S1 Movie). From these data, we conclude that the IFT-associated BBSome is required for targeting activated RAB-28 (GTP-locked) to periciliary membranes and loading onto IFT trains.
The differences in localisation and IFT behaviour between the constitutively inactive (GDP-locked) and active (GTP-locked) forms of RAB-28 prompted us to investigate cilium structure and function in worms overexpressing these small GTPase variants. Assessment of dye-filling reveals reduced levels of dye uptake in the amphid neurons of wild type worms expressing RAB-28(GTP), compared to normal dye-filling for wild type worms expressing RAB-28(WT) or RAB-28(GDP) (Fig 5A). Similar dye-filling results were obtained when these variants were expressed in rab-28(gk1040) worms, although worms expressing RAB-28(GDP) may display a very weakly penetrant phenotypic defect (S8A Fig). Despite the distinct dye-filling phenotypes, worms expressing RAB-28(GTP) or RAB-28(GDP) both show a reduced roaming behaviour (Fig 5B), suggestive of cilia-related sensory abnormalities [78].
To investigate the structural basis of these defects, we examined amphid pore ultrastructure in wild type (N2) worms expressing GFP-tagged RAB-28(GTP) or RAB-28(GDP). Somewhat surprisingly, despite the dye uptake phenotype, RAB-28(GTP)-expressing amphid channel cilia are of normal length and morphology, with intact middle and distal segments, as well as TZ compartments (Fig 5C). However, an increased number of ciliary axonemes possess B-tubule seam breaks (‘unzipped’ microtubules), indicative of incomplete protofilament formation (S8B Fig). A more striking phenotype is that the proximal region of the sheath cell-defined amphid channel is greatly enlarged, and the ciliary axonemes are less tightly bundled (Fig 5C). Also, the channels of RAB-28(GTP) expressing worms display reduced electron densities, potentially indicating abnormally low concentrations of matrix material, which is thought to be secreted by the sheath cell (Fig 5C) [43]. In contrast, although the amphid channel in RAB-28(GDP)-expressing worms is not expanded, these worms exhibit a number of abnormally large dense matrix-filled vesicles (MFVs) in the sheath, in that portion of the cell surrounding the channel (Fig 5C). Such electron dense vesicles are not found in the channel region of the amphid sheath cell of wild type (N2), rab-8(gk1040) or RAB-28(GTP) overexpressing worms (Fig 5C), indicating this phenotype is specific to RAB-28(GDP) overexpression. A second phenotype unique to RAB-28(GDP)-expressing worms is that one channel axoneme is missing from the amphid pore, and instead lies adjacent to the channel, embedded in the sheath (Fig 5C).
Together, these findings indicate that overexpression of constitutively active or inactive RAB-28 cause variant-specific defects in the amphid sheath (enlarged channel; dense MFV accumulation) and ciliated cells (dye-filling defect; misplaced amphid channel axoneme), although shared phenotypes are also observed (roaming defect). The sheath cell phenotypes likely represent a cell non-autonomous function because our transcriptional reporter for RAB-28 (GFP under the control of the rab-28 promoter) is not expressed in the sheath cell (S9 Fig). Thus, we conclude from these overexpression data that RAB-28 serves ciliated cell autonomous and sheath cell non-autonomous roles in sensory pore formation and function.
In this study, we identified a specific C. elegans ciliary gene expression profile using a set of RNA-Seq libraries, originally produced to annotate the transcriptome, that together provide a temporal landscape of whole-organism gene expression [45–47]. A major advantage of C. elegans as a model for ciliary biology is that ciliogenesis occurs during a temporally discrete window of embryonic development (Fig 1) [48], making it well suited for this type of temporal gene expression analysis. The expression profile for known cilia genes peaks at the onset of ciliogenesis and decreases post-ciliogenesis, likely to levels adequate to maintain cilium structure. Using this pattern, we uncover new ciliary associations for several conserved, but poorly characterised proteins, including the orthologues of YAP, MAP9 and CCDC149. The ciliary localisation of MAPH-9 (microtubule-associated protein 9) is consistent with a known functional association between MAP9 and the centrosomal regulatory kinase PLK1 [72], the centrosomal localisation and suggested ciliary functions for MAP9 in zebrafish [69], and a role for MAP9 as a modifier of retinal degeneration in Dachshunds [70]. YAP also has previous links to cilia, including a role in ciliogenesis [67,68] as well as being regulated by the ciliary gate protein, NPHP4 [71]. Although we do not understand the exact ciliary functions of MAP9 and YAP, their suggested roles in other contexts include regulation of ciliary microtubules (MAP9) and cilium-based signalling (YAP) [67–72]. Since 64 of the 185 genes in our predicted list of ciliary genes (cluster 1) possess known ciliary associations, we are confident that additional cilia-related components can be uncovered from this dataset.
One of the highest ranked hits in our candidate list of ciliary genes is RAB-28, a small GTPase whose human counterpart is linked to autosomal-recessive cone-rod dystrophy and vision impairment [62,63]. Although recent studies show that mammalian RAB28 is localised to the mammalian photoreceptor basal body and cilium [62,64], its role in photoreceptors and in other tissues and organisms remains unknown. Here, we show that C. elegans RAB-28 is specifically expressed in ciliated sensory neurons, and undergoes IFT. RAB-28 IFT association depends on its GTPase activity; whereas the GDP-locked inactive form localises diffusely with little or no observable IFT, the GTP-locked active form concentrates almost exclusively at the periciliary membrane and undergoes IFT. Moreover, the periciliary membrane association and IFT behaviour of active RAB-28 depends on the BBSome, which functions as an IFT cargo-adaptor [11,14,85]. We conclude that C. elegans BBSome complexes recruit activated RAB-28 to the periciliary membrane, and regulate docking to IFT trains (Fig 6).
Our findings also demonstrate that RAB-28 is probably not a ‘core’ component of the BBSome or IFT particles, since ciliary structures, as well as BBSome and IFT-subcomplex A/B protein localisations, are unaffected in rab-28 mutant worms. Instead, RAB-28 may associate with these complexes as part of a ‘peripheral’ submodule, performing auxiliary functions unrelated to ciliogenesis, such as those described for metazoan IFT27 (Rab-like 4; RABL4), ARL6 (BBS3) and RAB8, whose ciliary targeting or removal also depend on GTP binding and the BBSome (ARL6, RAB8) [86–89]. Consistent with a non-essential global requirement for ciliogenesis, IFT and BBSome integrity, patients carrying predicted strong loss-of-function or possibly null alleles in RAB28 do not present with wider ciliopathy symptoms beyond cone-rod dystrophy [62,63].
Another revealing observation from our work is that overexpression of constitutively active (GTP-locked) or inactive (GDP-locked) RAB-28 cause variant-specific defects in the non-ciliated amphid sheath cell that supports the ciliated endings of amphid neurons. Specifically, RAB-28(GTP) overexpression causes an enlargement of the sheath cell-defined portion of the amphid sensory pore, whereas RAB-28(GDP) overexpression causes the accumulation of abnormally dense matrix-filled vesicles (MFVs) in the sheath cell. The RAB-28 variant constructs, as well as the RAB-28(WT) construct, are expressed at similar levels, using the same promoter sequence; thus, the observed variant-specific phenotypes are not likely due to differences in expression level, but instead are linked to the nucleotide bound state of RAB-28. Most interestingly, the abnormal sheath cell phenotypes appear to reflect a cell-non autonomous effect because RAB-28 is expressed in ciliated sensory neurons, and not in the amphid sheath cell (S9 Fig).
In C. elegans, the amphid channel is fashioned from glial cell (sheath and socket) processes which extend to the nose tip in close proximity to the dendritic processes of the ciliated sensory neurons. The proximal part of the channel forms from a hole in the sheath process which the channel cilia fully penetrate, whereas the distal part of the channel derives from a doughnut-shaped ending of the socket process that fuses with the cuticle and the underlying sheath (Fig 5C). The formation and function of neurons and glia in the amphid pore appear tightly linked, with channel morphogenesis thought to depend on signals from both cell types [43,44,90]. It is also suggested that amphid channel size is regulated via a balance of ‘exocytic’ membrane delivery via MFVs (increases channel size) and ‘endocytic’ membrane retrieval (reduces channel size) at the sheath cell plasma membrane [91]. By integrating this model with our findings, we hypothesise that constitutively active RAB-28 may enhance MFV delivery to the sheath cell membrane, leading to an enlarged channel, whereas constitutively inactive RAB-28 blocks MFV delivery and results in abnormally dense MFVs in the sheath. One prediction of this model is that expression of constitutively inactive RAB-28 would reduce amphid channel size, and consistent with this idea, one channel axoneme fails to enter the amphid pore of RAB-28(GDP) expressing worms.
How RAB-28 expressed in ciliated amphid neurons might control morphogenesis events in the supporting non-ciliated sheath cell is unclear. One scenario is that RAB-28 regulates extracellular release of neuronal factors purported to signal to the sheath cell during channel morphogenesis in the embryo [42–44] (Fig 6). In support of this notion, activated RAB-28 is almost exclusively concentrated at the periciliary membrane, which is the site of ectosome release in certain nematode sensory neurons; also, RAB-28 expression is elevated in nematode ectosome-releasing cells [92,93]. Future work will be required to investigate possible roles for RAB-28 in cilia-related ectosome release, and to identify sheath cell morphogenic factors released by the ciliated neurons.
A surprising observation is that sensory pore structure and function appears grossly normal in the rab-28 deletion mutant, which is likely a null allele. This indicates that other genes or pathways can compensate for rab-28 loss-of-function, but that the phenotypic effects of the overexpressed dominant active and inactive RAB-28 variants are more severe and cannot be as easily compensated. One possibility is that the cilium structure, and cilium-dependent behavioural phenotypes in worms overexpressing RAB-28(GTP) and/or RAB-28(GDP) derive from excessive activation of downstream effectors or possibly aberrant sequestration of GTPase activating proteins (GAPs) or GDP-GTP exchange proteins (GEFs) common to RAB-28 and other ciliary GTPases. To fully decipher the mechanistic basis of the RAB-28 gain vs loss of function phenotypes, and the physiological relevance of our findings, future efforts should include an examination of worms expressing RAB-28 variants expressed at endogenous or near-endogenous levels and the identification of currently unknown regulators of RAB-28 (GEF and GAP proteins).
Given the close sequence similarity between RAB28 and IFT27, and some of their common functional properties discussed below, we suggest that these two RAB GTPases may have similar or partially overlapping ciliary functions. RAB28 is conserved across metazoans and many ciliated protists (S10 Fig), and inferred to be one of 23 founding RAB family members present in the last eukaryotic common ancestor (LECA) [94,95]. Phylogenetically, RAB28 is most closely related to IFT27 (RABL4), which likely also existed in LECA and is dispersed throughout the eukaryotic lineage [94,95] (S10 Fig). Like the majority of ciliary and IFT components, both RAB28 and IFT27 are restricted to organisms that have cilia (S10 Fig). Interestingly, although RAB28 and IFT27 are present in most metazoans and many protists (e.g., Chlamydomonas and Trypanosoma have both), neither protein appears to be present in Drosophila and IFT27 is missing from C. elegans (S10 Fig).
RAB28 and IFT27 appear to share overlapping functional properties. As we show for nematode RAB-28, Trypanosome and mammalian IFT27 associate with IFT in a GTP-dependent manner [87,88,96]. Also, cilium formation is unaffected in mammalian cells with null mutations in IFT27 (similar to C. elegans rab-28 mutants), although disruption of IFT27 in protists results in short flagella [74,96]. In addition, mutations in IFT27 cause BBS, displaying a broad spectrum of ciliopathy phenotypes, including the cardinal retinal degeneration phenotype observed in RAB28 patients [97,98]. Furthermore, both small GTPases are functionally associated with the BBSome, albeit in different ways. The localisation and IFT motility of nematode RAB-28 is BBSome-dependent, but not vice versa (i.e., BBS-5 localisation is normal in rab-28 mutants), whereas mammalian IFT27 regulates BBSome assembly and removal from the cilium. When taken together, the above observations could suggest that in vertebrates RAB28 and IFT27 function at distinct steps of a common pathway, potentially ‘upstream’ (IFT27) and ‘downstream’ (RAB-28) of the BBSome [87,88]. Since IFT27 and the BBSome remove proteins from cilia [85,87,88,99], RAB28 may perform a similar function by facilitating the BBSome-mediated removal of a more-restricted set of cargo.
In further support of a common, yet tissue-specific, transport role for related ciliary GTPases such as IFT27 and RAB28, it is noteworthy that disruption of the short ARL6/BBS3 isoform in mice and zebrafish results in broad spectrum BBS-related phenotypes, whereas disruption of the long isoform causes a retina-specific phenotype [100]. Transition zone-localised RPGRIP1 and RPGRIP1L represent additional examples of very closely related ciliopathy proteins linked to a broad tissue pathogenesis (RPGRIP1L), versus a retinal-specific disease (RPGRIP1) [101–104]. Finally, whereas IFT27 is known to regulate the ciliary localisation of sonic hedgehog proteins [87,88], the identity of RAB28-associated cargoes is currently unknown. Interestingly, RAB28 itself may be a cargo of the prenyl-binding protein PDE6D (cGMP phosphodiesterase delta subunit) which immunoprecipitates RAB28 [105]; in agreement with this possibility, loss of PDE6D in the mouse phenocopies the photoreceptor degeneration of RAB28 patients [106]. In light of our worm data, it is possible that PDE6D delta functions with the BBSome, either together or at distinct parts of a transport pathway, to target RAB28 to ciliary membranes.
Our study describes a novel method for identifying genes with ciliary functions that complements other studies (bioinformatics, genomics, proteomics) aimed at deciphering a complete ‘ciliome’. From our list of candidates, we uncover several novel, evolutionarily-conserved cilium-associated proteins. Our in-depth analysis of RAB-28 in C. elegans reveals it to be a novel IFT- and BBSome-associated protein with cell autonomous and non-autonomous roles in maintaining the structural and functional integrity of cilia, and of the sensory neuron-supporting cells. Further analyses of mammalian RAB28 are required to reveal its functional similarities and differences with IFT27, and possible functional associations with PDE6D and ARL6, to help explain its seemingly specific function in ciliary photoreceptors, and thus specific association with cone-rod dystrophy.
We curated a list of 41 C. elegans genes with established roles in ciliary function and used these to search for genes with correlated expression. These “bait” genes included: R148.1, F38G1.1, Y75B8A.12, F20D12.3, Y41G9A.1, R01H10.6, C54G7.4, C27A7.4, R31.3, C48B6.8, F32A6.2, C27H5.7, T28F3.6, Y37E3.5, F59C6.7, F19H8.3, ZC84.2, F35D2.4, H01G02.2, B0240.3, ZK520.3, C02H7.1, ZK418.3, K08D12.2, F23B2.4, Y105E8A.5, F46F6.4, K07G5.3, F33H1.1, F02D8.3, K03E6.4, C38D4.8, T26C12.4, T27B1.1, Y110A7A.20, F53A9.4, M04C9.5, C30B5.9, C09G5.8, R13H4.1 and F54C1.5. We used expression data derived from RNA-seq for seven developmental stages (EE = early embryo, LE = late embryo, L1 = mid-L1 larvae, L2 = mid-L2 larvae, L3 = mid-L3 larvae, L4 = mid-L4 larvae and YA = young adult) obtained from [45–47]. As the expression values were relative to individual isoforms, we first computed an average RPKM (reads per kilobase per million) value for each gene across all isoforms. Using these 41 genes, we next computed the pairwise Pearson correlation of each gene against all 20,363 C. elegans genes in the data set across all libraries and estimated a P value for each test. As we were interested in capturing genes with expression correlates to any of these known ciliary genes, we retained the smallest P value from these 41 tests for subsequent calculations. To avoid circularity, we retained the second-smallest P value for correlation tests involving one of the bait genes. In addition to direct correlation, we sought to identify genes exhibiting a developmental expression pattern similar to known ciliary genes. We observed that all ciliary genes exhibit at least a 10-fold reduction in expression level (relative to their peak expression level) in the latter three developmental stages. A total of 5960 genes showed this expression trend.
Following the identification of those genes exhibiting strongly correlated expression to any of the known ciliary genes, this set was further clustered to search for sub-groupings of genes with strongly coordinated expression. The distance between individual genes was calculated using the ‘dist’ function in R and hierarchical clustering was performed using single, average and complete linkage. S1 Fig shows a heat map representation of the expression for each of the candidate ciliary genes with the dendrogram derived from clustering using complete linkage and cluster 1 at the top.
Using the method described above, we used srg-36 as a bait to identify co-expressed genes. We found 80 genes (p < 0.001), including 27 other predicted chemoreceptors [52]. The 1000 bp upstream from the start codon of all 28 chemoreceptors (including the bait srg-36) were entered into the program MEME [107,108], which found a consensus that matches the published E-box sequence [54]. Separately, the 1000 bp upstream of all the conserved cluster 1 were entered into the MEME suite program DREME [107,108], which found several potential transcription factor binding sites. We used the MEME suite program TomTom to compare identified sites to known sites [108]. Identified promoter elements were compared to consensus splice and trans-splice sites and any that matched were subsequently removed.
All nematode strains were maintained and cultured at 20°C on nematode growth medium (NGM) plates seeded with OP50 Escherichia coli using standard techniques. Standard genetic crossing techniques were used to introduce transgenes into genetic backgrounds. The rab-28(gk1040) mksr-1(ok2092), mksr-2(tm2452), mks-5(tm3100), mks-3(tm2547), mks-6(gk674), and nphp-4(tm925) mutations were followed using genotyping PCR (primer sequences available upon request).
N2 (Bristol)
rab-28(gk1040)
che-11(e1810)
bbs-8(nx77)
N2;oqEx300[Prab-28::gfp + Punc-122::gfp]
N2;oqEx301[Prab-28::gfp::rab-28 + Punc-122::gfp]
N2;oqEx302[Prab-28::gfp::rab-28(T49N) + Punc-122::gfp]
N2;oqEx303[Prab-28::gfp::rab-28(Q95L) + Punc-122::gfp]
N2;oqEx304[Prab-28::gfp::rab-28(Q95L) + Punc-122::gfp]
bbs-8(nx77);oqEx301[Prab-28::gfp::rab-28 + Punc-122::gfp]
bbs-8(nx77);oqEx302[Prab-28::gfp::rab-28(T49N) + Punc-122::gfp]
bbs-8(nx77);oqEx303[Prab-28::gfp::rab-28(Q95L) + Punc-122::gfp]
bbs-8(nx77);oqEx304[Prab-28::gfp::rab-28(Q95L) + Punc-122::gfp]
che-11(e1810);oqEx301[Prab-28::gfp::rab-28 + Punc-122::gfp]
rab-28(gk1040);oqEx58[arl-13::gfp + rol-6(su1006)]
rab-28(gk1040);Is[osm-6::gfp]
rab-28(gk1040);myIs[pkd-2::gfp]
rab-28(gk1040);nxEx289[rpi-2::gfp + xbx-1::tdTomato + rol-6(su1006)]
rab-28(gk1040);nxEx475[bbs-5::gfp + pCeh361]
N2;nxEx869[fam-161::gfp + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx1157[Pbbs-8::gfp::rab-28 + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx250[rpi-1::gfp + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx1223[maph-9::gfp + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx1230[Pbbs-8::gfp::yap-1 + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx2619[Pccdc-149::gfp::ccdc-149 + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx2623[Pccdc-104::gfp::ccdc-104cDNA + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx2626[Pccdc-149::gfp + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx2628[Pccdc-104::gfp + xbx-1::tdTomato + rol-6(su1006)]
N2;nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
mks-6(gk674);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
nphp-4(tm925);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
mks-3(tm2547);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
mks-5(tm3100);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
mksr-1(ok2092);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
mksr-2(tm2452);nxEx654[tza-3::gfp + xbx-1::tdTomato + rol-6(su1006)]
For the dye-filling assay [109], worms were incubated in 200 μl of DiI solution (Invitrogen; diluted 1:200 with M9 buffer) for 30 min. After incubation, worms were recovered on seeded nematode growth medium plates for a further 30 min and then mounted on slides. Epifluorescence wide-field imaging under the red filter was used to image DiI uptake into the ciliated amphid and phasmid cells. For the roaming (foraging) assay, single worms were placed for 18 hours onto seeded NGM plates and track coverage assessed using a grid reference [109]. For the osmotic avoidance assay, 5–6 worms were placed within a ring of 8M glycerol (Sigma) supplemented with Bromophenol Blue (Alfa Aesar) on unseeded NGM plates and observed for 10 min. Worms that crossed the barrier were removed from the assay. Chemotaxis attraction assays (towards isoamyl alcohol) were performed as previously reported [109], with a chemotaxis index calculated at 30 min and 60 min.
All constructs were generated by PCR fusion [110]. For the transcriptional (promoter) Prab-28::gfp construct, gfp amplified from pPD95.67 was fused to an 432-bp fragment of the 5’UTR of rab-28 that included the first 14 bp of exon 1 (start codon thymine mutated to guanine). For the transcriptional reporters for Pccdc-104 and Pccdc-149, gfp amplified from pPD95.77 was fused to 1997 and 395 bp, respectively, of the 5’UTR for each gene. For the translational (promoter + protein) fam-161, rpi-1, maph-9, ccdc-149 and tza-3 reporters, the entire genomic sequence (exons and introns) were fused to gfp including promoter sequences of 1373 bp, 2317 bp, 973 bp, 395 bp and 457 bp, respectively. For the translational reporter for rab-28 and yap-1 in Fig 2, 341 bp of the bbs-8 promoter followed by gfp were fused to the genomic sequence including 3’ UTR for each gene. For the ccdc-104 translational reporter, 1997 bp of the bbs-8 promoter followed by gfp were fused to the ccdc-104 cDNA and 3’ UTR. For the translational Prab-28::gfp::rab-28 reporter (Figs 3 and 4), the entire genomic sequence and 970-bp of the 3’UTR of rab-28 was first fused 5’ to a gfp fragment, amplified from pPD95.77. The resulting gfp::rab-28 amplicon was subsequently fused to a 422-bp fragment consisting of the 5’UTR (promoter) sequence of rab-28. To make the Prab-28::gfp::rab-28(GDP) and Prab-28::gfp::rab-28(GTP) constructs, primers incorporating the corresponding T49N and Q95L mutations were first used to amplify 5’ and 3’ fragments of the rab-28 genomic sequence, and these were fused by PCR to establish the rab-28(T49N) and rab-28(Q95L) amplicons. These amplicons were subsequently fused to a Prab-28::gfp fragment, generated by fusing the rab-28 promoter sequence (see above) to gfp amplified from pPD95.77. Transgenic worms expressing the above constructs were generated by gonadal transformation of N2 hermaphrodites via microinjection and subsequent screening for transgenic progeny. rab-28 constructs were injected at a concentration of either 5 ng/μl (all translational constructs except oqEx304 which was injected at 0.5 ng/μl) or 50 ng/μl (transcriptional constructs), together with a coelomocyte (Punc-122::gfp) or pRF4 (rol-6(su1006)) co-injection marker at 50–100 ng/μl.
Strains in Fig 2 and S4B Fig were anaesthetised with 10 mM levamisole in M9 buffer, mounted on slides with 8% agarose pads, and observed by epifluorescence or spinning-disc (WaveFX; Quorum Technologies) confocal microscopy performed on an Zeiss Axio Observer Z1 with a Hamamatsu 9100 EMCCD camera. Image capture and visualisation were performed on Volocity (PerkinElmer). All other worms were immobilised with 40mM tetramisole (Sigma no. L9756) or microbeads (Polysciences no. 00876–15) and mounted on 4% or 10% agarose pads. Epifluorescence images were taken on an upright Leica DM5000B and confocal images on an inverted Nikon Eclipse Ti microscope with a Yokogawa spinning-disc unit (Andor Revolution). Images were acquired using a charge-coupled device camera (iXon+EM-CCD, Andor Technology) and analysed using Image J software. For IFT assays, time-lapse (multi tiff) movies of IFT along phasmid cilia were taken at 200 ms exposure and 4 fps. Separated anterograde and retrograde kymographs were generated from multi tiff files using Icy image analysis software (http://icy.bioimageanalysis.org/) and rates determined using ImageJ [111]. Fluorescence recovery after photobleaching (FRAP) assays were performed using the above confocal microscope with an attached FRAPPA unit (Andor Technology). Samples were bleached using a single pulse of the 488nm laser at 100% with a dwell time of 100 μs. Images were recorded immediately post-bleach and continuously thereafter. Fluorescence intensities measured with image J software following a recently described protocol [109]. Values were normalised to pre-bleach values and corrected for signal intensity loss during image acquisition.
Young adult worms were fixed, sectioned and imaged as previously reported [109].
|
10.1371/journal.pgen.1008093 | Destabilization of chromosome structure by histone H3 lysine 27 methylation | Chromosome and genome stability are important for normal cell function as instability often correlates with disease and dysfunction of DNA repair mechanisms. Many organisms maintain supernumerary or accessory chromosomes that deviate from standard chromosomes. The pathogenic fungus Zymoseptoria tritici has as many as eight accessory chromosomes, which are highly unstable during meiosis and mitosis, transcriptionally repressed, show enrichment of repetitive elements, and enrichment with heterochromatic histone methylation marks, e.g., trimethylation of H3 lysine 9 or lysine 27 (H3K9me3, H3K27me3). To elucidate the role of heterochromatin on genome stability in Z. tritici, we deleted the genes encoding the methyltransferases responsible for H3K9me3 and H3K27me3, kmt1 and kmt6, respectively, and generated a double mutant. We combined experimental evolution and genomic analyses to determine the impact of these deletions on chromosome and genome stability, both in vitro and in planta. We used whole genome sequencing, ChIP-seq, and RNA-seq to compare changes in genome and chromatin structure, and differences in gene expression between mutant and wildtype strains. Analyses of genome and ChIP-seq data in H3K9me3-deficient strains revealed dramatic chromatin reorganization, where H3K27me3 is mostly relocalized into regions that are enriched with H3K9me3 in wild type. Many genome rearrangements and formation of new chromosomes were found in the absence of H3K9me3, accompanied by activation of transposable elements. In stark contrast, loss of H3K27me3 actually increased the stability of accessory chromosomes under normal growth conditions in vitro, even without large scale changes in gene activity. We conclude that H3K9me3 is important for the maintenance of genome stability because it disallows H3K27me3 in regions considered constitutive heterochromatin. In this system, H3K27me3 reduces the overall stability of accessory chromosomes, generating a “metastable” state for these quasi-essential regions of the genome.
| Genome and chromosome stability are essential to maintain normal cell function and viability. However, differences in genome and chromosome structure are frequently found in organisms that undergo rapid adaptation to changing environmental conditions, and in humans are often found in cancer cells. We study genome instability in a fungal pathogen that exhibits a high degree of genetic diversity. Regions that show extraordinary diversity in this pathogen are the transposon-rich accessory chromosomes, which contain few genes that are of unknown benefit to the organism but maintained in the population and thus considered “quasi-essential”. Accessory chromosomes in all fungi studied so far are enriched with markers for heterochromatin, namely trimethylation of H3 lysine 9 and 27 (H3K9me3, H3K27me3). We show that loss of these heterochromatin marks has strong but opposing effects on genome stability. While loss of the transposon-associated mark H3K9me3 destabilizes the entire genome, presence of H3K27me3 favors instability of accessory chromosomes. Our study provides insight into the relationship between chromatin and genome stability and why some regions are more susceptible to genetic diversity than others.
| Chromatin structure plays an important role in genome organization and gene expression [1–3]. A well-studied hallmark of epigenetic regulation is the reversible modification of histone tails, which can alter chromatin structure [4]. Chromatin structure determines accessibility of the underlying DNA to regulatory elements, whereby tightly packed DNA, known as heterochromatin, is less accessible for DNA binding proteins and usually shows little transcriptional activity [5]. Heterochromatic regions often cluster together and are spatially separated from more transcriptionally active and accessible euchromatic regions [6]. Specific histone modifications are associated with either heterochromatic or euchromatic regions. Some of the most studied histone modifications are histone H3 lysine 9 di- or trimethylation (H3K9me2/3) and H3K27me2/3 as markers for heterochromatin and H3K4me2/3 as markers for euchromatin [7].
H3K9me2/3 is catalyzed by the histone methyltransferase KMT1 (Su[var]3–9) [8,9], in fungi also called Clr4 [10] or DIM-5 [11]. Previous studies demonstrated enrichment of this constitutive heterochromatin mark in repeat-rich regions and a clear link with the control of transposable elements (TE) and genome stability [12–14]. For example, H3K9 methylation has been shown to be involved in suppression of meiotic recombination in Arabidopsis thaliana [15] and the control of DNA methylation in Neurospora crassa [11].
H3K27me2/3, usually associated with “facultative heterochromatin”, is catalyzed by KMT6 (E[Z]) as part of the PRC2 complex [16]. In plants, fungi, and animals, this histone mark is used to generate “transcriptional memory” and is easily reversible when environmental or endogenous stimuli require organismal responses. In many organisms, H3K27 methylation is required for development and cell differentiation [17–23], and aberrant H3K7me3 distribution is prevalent in cancer cells [24–26]. In fungi, H3K27me3 correlates with subtelomeric gene silencing [22,23,27], and has been shown to play a role in development, pathogenicity, and transcriptional regulation of secondary metabolite gene clusters [21,28,29].
H3K27me3 is also a hallmark of accessory chromosomes, which are found in several fungal plant pathogens [28,30,31]. Accessory chromosomes are not essential for survival under all environmental conditions, and thus encode “quasi-essential” genes [32] that can confer selective advantages under some conditions e.g. in a specific host species, resulting in presence or absence of these chromosomes among specific individuals of a given species. They are also characterized by extensive structural rearrangements and length variation [33,34]. In some species (Fusarium oxysporum, Nectria haematococca, Alternaria alternata), accessory chromosomes increase virulence [35–38]. However, in the wheat pathogen Zymoseptoria tritici, some accessory chromosomes have been demonstrated to confer reduced fitness and virulence in planta [39], suggesting that there are other stages in the life cycle when they become important. Accessory chromosomes of fungi differ structurally from core chromosomes by higher repeat and lower gene density compared to core chromosomes and show little transcriptional activity [35,40–43]. Transcriptional silencing can be explained by their predominantly heterochromatic structure, with H3K27me3 enrichment on almost the entire chromosome and H3K9me3 covering repetitive sequences [28,30]. Centromeres and telomeres are important structural components of chromosomes. In plants, centromeres of B chromosomes, equivalents to fungal accessory chromosomes, differ from those of A chromosomes [44], but in Z. tritici centromeres, telomere repeats, and subtelomeric regions are so far by all measures near identical on core and accessory chromosomes [31]. Though accessory chromosomes are a frequent phenomenon in fungi, little is known about their origin and maintenance. Studies on chromosome stability revealed that accessory chromosomes are highly unstable, both during mitosis [36,45,46] and meiosis [47].
Here, we investigated to what extent the particular histone methylation pattern on accessory chromosomes contributes to the structural differences, transcriptional repression and instability. We shed light on the roles of H3K9me3 and H3K27me3 on genome stability in the hemi-biotrophic wheat pathogen Z. tritici that reproduces both asexually and sexually. By combining experimental evolution with genome, transcriptome and ChIP sequencing, we show that both heterochromatin-associated histone methylation marks contribute significantly, but in distinct ways, to chromosome stability and integrity. While the presence of H3K27me3 enhances chromosome loss and instability, loss of H3K9me3 promotes chromosome breakage, segmental duplications as well as the formation of new chromosomes − possibly resembling the emergence of accessory chromosomes. Taken together, our findings demonstrate the importance of constitutive heterochromatin for maintaining genome stability and gene silencing as well as an unexpected destabilizing influence of facultative heterochromatin on mitotic accessory chromosome transmission. The presence of eight accessory chromosomes in the reference isolate IPO323 makes Z. tritici an excellent model to study accessory chromosome characteristics and dynamics, which relates to general interest in chromosome maintenance in cancer or other aneuploid cell types.
To investigate the impact of heterochromatin on fitness, transcription and genome stability in Z. tritici, we generated mutants of two histone methyltransferases Kmt1 (S. pombe Clr4; N. crassa DIM-5, Fusarium KMT1, H. sapiens SUV39H1) and Kmt6 (N. crassa SET-7; Fusarium KMT6; H. sapiens EZH2). We identified the Z. tritici genes by BLAST searches with the N. crassa and F. graminearum protein coding sequences as baits. Kmt1 is encoded by kmt1 (Zt_chr_1_01919), and Kmt6 is encoded by kmt6 (Zt_chr_4_00551) [48]. We used Agrobacterium tumefaciens-mediated transformation [49] to delete both genes in a derivate of the Z. tritici reference isolate IPO323 that lost chromosome 18 during in vitro growth, here called Zt09 [31,40,41]. Correct integration of the hph gene, which confers hygromycin resistance [49], and kmt1 or kmt6 deletion were verified by PCR and Southern analyses (S1 and S2 Figs). We generated a double deletion mutant by deleting the kmt1 gene in a kmt6 deletion mutant background by using resistance to nourseothricin conferred by the nat gene [50] as an additional selection marker. We isolated several independent transformants, including eight Δkmt1, six Δkmt6 and ten Δkmt1 Δkmt6 double mutants (from here on abbreviated Δk1/k6). For further studies we selected two or three mutants of each type (S1 Table). Δkmt1 and Δkmt6 single mutants were complemented by re-integrating the previously deleted gene and a neo+ resistance marker that can confer G418 resistance at the native gene loci (S2 Fig).
We performed ChIP-seq on Zt09, Δkmt1 (Zt125-#68, -#80), Δkmt6 (Zt110-#283, -#285, -#365) and the double deletion mutant Δk1/k6 (Zt219-#23, -#116), which verified the absence of H3K9me3 in Δkmt1 and Δk1/k6, and the absence of H3K27me3 in Δkmt6 and Δk1/k6 mutants (S3 Fig), confirming that Kmt1 and Kmt6 are the only histone methyltransferases in Z. tritici responsible for H3K9 and H3K27 trimethylation, respectively. An overview of the subsequent experiments and key results are summarized in S1 Fig.
To assess if deletion of kmt1 and kmt6 has an impact on in vitro growth or pathogenicity on wheat, we performed comparative growth and virulence assays comparing the mutants to the wild type Zt09. To compare growth rates, the reference strain Zt09, deletion and complemented strains were grown in liquid YMS cultures and the OD600 was measured until cells reached stationary phase. Overall, the Δkmt1 strains and Δk1/k6 double deletion mutants showed significantly reduced growth in vitro (S4 Fig). The Δkmt6 mutants and both kmt1+ and kmt6+ complementation strains showed no significant differences in growth compared to Zt09 (Wilcoxon rank-sum test, p-values: Δkmt1 0.025; Δkmt6 0.42; Δk1/k6 0.005; kmt1+ 0.28; kmt6+ 0.63).
We furthermore assessed the tolerance of the Δkmt1, Δkmt6 and Δk1/k6 mutants to abiotic stress in vitro by testing temperature, cell wall, oxidative, and genotoxic stressors. As observed in the growth assays, the Δkmt1 and Δk1/k6 double deletion mutants showed overall reduced growth under all tested conditions (S5 Fig), especially under osmotic stress induced by high sorbitol concentrations. The Δkmt6 mutants showed little differences compared to Zt09; however, elevated temperatures often, but not always, led to increased melanization in the Δkmt6 mutants suggesting involvement of H3K27me3 in the response to temperature stress. This phenotype was reversed in the complemented kmt6+ strain (S6 Fig).
To study the effect of the histone methyltransferase deletions on the ability to infect wheat (Triticum aestivum), we inoculated leaves of the susceptible cultivar Obelisk with single cell cultures of Δkmt1, Δkmt6, the Δk1/k6 double deletion mutant and Zt09. The infection assays demonstrated significant impact of both H3K27me3 and H3K9me3 on virulence. While the number of pycnidia and necrotic leaf areas only decreased in the Δkmt6 mutants, wheat infection by Δkmt1 and Δk1/k6 mutants resulted in almost no symptoms (S7 Fig). If any symptoms developed, these appeared considerably later than symptoms caused by the reference Zt09 and the Δkmt6 mutants (S7 Fig).
We next addressed how the deletion of kmt1 and kmt6 impacts the distribution of three histone modifications (H3K4me2, H3K9me3, H3K27me3) by ChIP-seq (S2 Table). We previously found that H3K4me2 is associated with gene-rich, transcriptionally active regions on core chromosomes, that constitutive heterochromatin, enriched with H3K9me3, forms almost exclusively on repetitive elements, and that facultative heterochromatin, enriched with H3K27me3, forms nearly on the entire length of all accessory chromosomes and the subtelomeric regions of core chromosomes [31] (Fig 1).
We computed the sequence coverage of each histone modification per chromosome to estimate the global effects on chromatin structure. The absence of one histone methylation mark had differential effects on the distribution of the other two methylation marks on core and accessory chromosomes (Fig 1, Table 1). In the Δkmt1 mutants, the amount of sequences enriched with H3K27me3 decreases on the accessory chromosomes when compared to Zt09, representing the opposite trend to the observations made on the core chromosomes, where we observed an increased amount of sequences enriched in H3K27me3 (Fig 1, Table 1). However, this effect varies on different accessory chromosomes (Table 1, Fig 2). The difference in H3K27me3 distribution can be explained by relocation of H3K27me3 to former H3K9me3-associated sequences in the Δkmt1 mutant (Fig 1). While fewer genes are associated with H3K27me3 (Fig 2), more TEs show H3K27me3 enrichment in the Δkmt1 mutant (Fig 2) compared to Zt09. These observations reveal that loss of H3K9me3 promotes H3K27me3 relocation to TEs and confers simultaneous loss of H3K27me3 at positions with this histone mark in the reference strain. The subtelomeric H3K27me3 enrichment, however, is not affected by this relocation, which explains why we observe opposite effects on core and accessory chromosomes, as core chromosomes predominantly show H3K27me3 enrichment in subtelomeric regions while accessory chromosomes show overall enrichment with H3K27me3. H3K4me2 increases on both core and accessory chromosomes, with accessory chromosomes showing a considerably higher relative increase compared to H3K4me2 in Zt09 (Table 1).
Conversely, H3K9me3 is not affected by loss of H3K27me3 in the Δkmt6 mutants, and we did not detect relocation of H3K9me3 as well as only minor differences in coverage. H3K4me2 enrichment does increase on accessory chromosomes, but not to the same extent as observed in the Δkmt1 mutants and it slightly decreases on core chromosomes (Table 1), suggesting minor effects of Δkmt6 on transcriptional activation. In the Δk1/k6 double deletion mutants, where both H3K9me3 and H3K27me3 are not present, we detected an increase in H3K4me2, similar to the Δkmt1 single mutants on core chromosomes and slightly higher on the accessory chromosomes.
In summary, loss of H3K9me3 has a great impact on H3K27me3 distribution, while loss of H3K27me3 has little influence on H3K9me3. Deletion of kmt1 promotes large scale relocalization of other histone modifications, suggesting more dramatic effects on genome organization and transcriptional activation than deletion of kmt6.
In other species, H3K27me3 plays a crucial role in gene regulation, while H3K9me3 is involved in silencing of TEs [13,21,28]. Based on our observations from ChIP-seq data, we hypothesized that the two histone methylation marks have similar effects in Z. tritici. To test this hypothesis directly, we sequenced transcriptomes of two biological replicates of Zt09 and two independent transformants of the Δkmt1, Δkmt6, and Δk1/k6 deletion mutants after in vitro growth for 2 days representing exponential growth (S2 Table).
First, we compared the total number of expressed genes. In total, 11,839 genes are annotated in the reference isolate [48]. Out of these, 8,906 are expressed (RPKM >2) in Zt09 during in vitro growth. The number of expressed genes is higher in both the Δkmt1 (9,259) and the Δk1/k6 (9,459) mutants, but to our surprise, lower in the Δkmt6 (8,717) mutants (Fig 3A, S3 Table). This is in contrast to previous studies, where deletion of kmt6 resulted in activation of otherwise silenced gene clusters and overall transcriptional activation [21,23,27,28]. We focused on differential gene expression between core and accessory chromosomes because genes on accessory chromosomes are silent under most conditions that have been tested. While 80% of genes on core chromosomes are expressed in Zt09, only ~25% of genes located on accessory chromosomes display transcriptional activity. Transcription of genes on accessory chromosomes is higher in all mutant strains, ~40–50% (Fig 3A, S3 Table), revealing gene activation on accessory chromosomes specifically upon removal of H3K27me3 or H3K9me3.
We further explored patterns of differential gene expression. Genome wide, 1,365 predicted genes were associated with H3K27me3 and 258 genes with H3K9me3 in Zt09 and the vast majority of these genes shows little transcriptional activity. Interestingly, only a small fraction of genes associated with these histone marks were activated or differentially expressed in the mutants (S4 Table). This indicates that loss of any of these methylation marks is not sufficient for transcriptional activation suggesting additional mechanisms involved in the transcriptional regulation of these genes.
In other fungi, removal of H3K9me3 and especially H3K27me3 was linked to the activation of certain gene classes, in particular secondary metabolite gene clusters [21,28,29]. To assess if genes with a specific function are enriched amongst the activated genes, we performed Gene Ontology (GO) enrichment analysis (topGO, Fisher’s exact test, p-value < 0.01). Consistent with the higher total number of expressed genes, we found the majority of differentially expressed (DE) genes (DESeq2, Padj < 0.001, │log2 fold-change│ > 2) to be significantly upregulated in the Δkmt1 mutant (365 of 477) and in the Δk1/k6 mutant (368 of 477), whereas a majority of DE genes was downregulated in the Δkmt6 mutant (188 of 310) (S5 Table).
We found two GO categories enriched amongst upregulated genes in Δkmt1 and Δk1/k6 mutants: DNA integration (GO:0015074) and RNA-dependent DNA replication (GO:0006278). Predicted functions assessed by BLAST analyses of the proteins encoded by the upregulated genes in these categories include reverse transcriptases, integrases, recombinases and genes containing transposon- or virus-related domains (S6 Table). Consistent with these findings, we detected an increased number of transcripts originating from annotated TEs in the Δkmt1 and Δk1/k6 mutants, but not in Δkmt6 mutants (Fig 3B). This is in agreement with the strong association of TEs with H3K9me3 [31]. Transposons in subtelomeric regions and on accessory chromosomes show additional H3K27me3 enrichment. Removal of H3K9me3, but not of H3K27me3, appears to be responsible for transposon activation but transcription is further enhanced when both, H3K27me3 and H3K9me3 are removed in the Δk1/k6 mutant (Fig 3B; S7 Table).
Amongst the genes upregulated in the Δkmt6 mutant no GO categories were enriched but based on the previous finding of secondary metabolite activation, we further investigated possible roles of H3K9me3 and H3K27me3 in secondary metabolite gene regulation. Therefore, we identified putative secondary metabolite clusters in the Z. tritici reference genome using antiSMASH (antibiotics & Secondary Metabolite Analysis SHell) [51]. We found a total of 27 secondary metabolite clusters, all located on core chromosomes, and merged the identified genes with the existing gene annotation (S8 Table). Except for the activation of one putative cluster on chromosome 7 in the Δk1/k6 mutant, we did not identify any differential expression of genes in secondary metabolite clusters. Based on these findings, we conclude that, unlike in other fungi [21,29], H3K9me3 and H3K27me3 are not involved in transcriptional regulation of secondary metabolites in Z. tritici under the tested conditions.
Taken together, removal of these histone modifications has little consequences for the expression of the vast majority of associated genes. As expected from its localization, loss of H3K9me3 increases expression of TEs while absence of H3K27me3 by itself has very little impact on transcriptional activation, thus suggesting that in this organism H3K27me3 does not delineate stereotypical “facultative heterochromatin” as removal of H3K27me3 does not activate gene expression.
Chromosome landmarks, namely centromeric and pericentric regions, telomere repeats and subtelomeric regions are similar on core and accessory chromosomes in Z. tritici [31]. Accessory chromosomes are enriched with TEs but share the same TE families as core chromosomes [48]. Nevertheless, accessory chromosomes of Z. tritici are highly unstable, both during meiosis and vegetative growth in vitro and in planta [46,47]. The most striking feature that sets these chromosomes apart is almost chromosome-wide enrichment with H3K27me3 and, as a consequence of the higher TE content, increased enrichment with H3K9me3 [31]. To test whether loss of these modifications affects genome and chromosome stability in Z. tritici, we conducted two different long-term growth or “lab evolution” experiments to study genome stability and to detect dynamics of accessory chromosome losses in strains deficient for two important chromatin marks (S8 Fig).
To assess whether the specific histone methylation pattern on accessory chromosomes contributes to instability of accessory chromosomes, we performed a short-term in vitro growth experiment over four weeks, representing ~80 asexual generations. Zt09, Δkmt6, Δkmt1 and a Δk1/k6 double deletion mutant, as well as the complemented strains kmt1+ and kmt6+ were used as progenitors in the experiment. The presence of all accessory chromosomes in the progenitor strains was verified by PCR at the beginning of the experiment. Each strain was grown in three replicate cultures and ~4% of the cell population was transferred to fresh medium every three to four days. After four weeks of growth, we plated dilutions of each culture to obtain single colonies that were subsequently screened by a PCR assay for the presence of accessory chromosomes (Table 2).
Previously, we showed that accessory chromosomes are lost at a rate of ~7% in Zt09 and we documented that accessory chromosomes 14, 15 and 16 are more frequently lost than others [46]. Here we demonstrate that, in comparison to Zt09, the Δkmt1 mutant showed a significantly increased chromosome loss rate (one sided Fisher’s exact test for count data, p-value = 2.7 x 10−6). Interestingly, this was not due to an overall increase of accessory chromosome loss, but rather by the dramatically increased (Fisher’s exact test, p-value = 3.7 x 10−9) frequency of loss for chromosome 20 (Table 2). The chromosome loss rate of the other accessory chromosomes was either comparable to Zt09 (Chr. 14, 17, 19, 21) or even significantly lower (Chr. 15 and 16, Fisher’s exact test, p-values = 1.2 x 10−3 and 2.5 x 10−5). This suggests a special role of H3K9me3 for the maintenance of chromosome 20.
In contrast to the Δkmt1 mutants, we detected significantly fewer chromosome losses (Fisher’s exact test, p-value = 1.2 x 10−4) in the Δkmt6 mutants. Out of 576 tested colonies, only ten had lost an accessory chromosome. This represents a four times lower chromosome loss rate compared to wild type. Therefore, absence of H3K27me3 appears to promote stability of accessory chromosomes. Interestingly, chromosome 17 was lost with the highest frequency in this mutant (5/10) but was not lost in any of the other mutant strains or in the wild type.
The double deletion mutant displayed a similar chromosome loss rate as wild type but showed a chromosome loss distribution comparable to the Δkmt1 deletion strain with chromosome 20 being lost significantly more often (Fisher’s exact test, p-value = 1.23 10−4), and chromosomes 15 and 16 lost less frequently (Fisher’s exact test, p-values = 1.95 x 10−3 and 9 x 10−3, respectively). This suggests that the increase in chromosome stability in Δk1/k6 compared to Δkmt1 is due to the removal of the destabilizing H3K27me3.
We considered possible reasons for the high rates of loss of chromosome 20 in Δkmt1 and Δk1/k6 mutant strains by a detailed analysis of TE content, genes and histone methylation redistribution on this chromosome. The proportion of TEs on chromosome 20 is ~23% and thereby considerably lower than the average across the eight accessory chromosomes (33.6% TE content). Consistent with patterns found on the other accessory chromosomes we find no enrichment of specific TE families [48]. Out of the 91 genes on chromosome 20, none has known or even predicted functions; it is therefore difficult to correlate potential gene functions of any of these genes to the enhanced loss rate. Chromosome 20 is one of the chromosomes that exhibits the highest extent of H3K27me3 redistribution (Fig 2) and loss of H3K27me3 in the Δk1/k6 double mutant decreases the high loss rate. However, chromosome 20 is still lost at a higher rate in Δk1/k6 compared to Zt09, indicating that other, so far unknown, factors are involved in this instability. Moreover, not all accessory chromosomes that exhibit H3K27me3 redistribution (19 and 21) are lost at higher rates, and not all accessory chromosomes are lost at the same rate in the Zt09 wild type, despite being enriched with H3K27me3, further suggesting that additional mechanisms are involved in the instability of chromosome 20.
Both complemented strains, kmt1+ and kmt6+, showed chromosome loss rates similar to the wild type strain Zt09 and the highest loss rates for the largest accessory chromosomes (Table 2). This strongly suggests that indeed the absence of the two histone methyltransferases and the respective histone marks influence accessory chromosome dynamics.
In summary, we found that loss of H3K27me3 increases accessory chromosome stability, suggesting a mechanistic explanation for how the widespread H3K27me3 enrichment on accessory chromosomes in normal cells contributes to the previously observed extraordinary chromosome instability.
In a second evolution experiment, we addressed overall genome stability over a longer period of mitotic growth. The single mutants (Δkmt1 and Δkmt6) and Zt09 were grown in triplicate cultures for ~6 months, representing ~500 asexual generations. We sequenced full genomes of progenitors and the evolved populations after 50 transfers to identify structural variations that arose during the experiment. All strains were sequenced to ~100x coverage by Illumina sequencing, and paired-end reads were mapped to the reference genome of IPO323 and normalized to 1x coverage for visualization [40].
We focused our analysis on large scale chromosomal rearrangements such as duplications, deletions, and translocations. Structural variation was detected computationally from sequence alignments, validated experimentally by PFGE and Southern blotting, and additional rearrangements were identified by manual screening of mapped reads. Analysis of progenitor genomes revealed, except for the already known absence of chromosome 18 [41] and the previously described variations (point mutations and short indels) in Zt09 compared to the IPO323 reference genome [46], lower sequence coverage (~0.6x) on chromosome 17 in the Δkmt6 progenitor strain (Fig 4A). This difference can only be explained by a lower copy number in the sequenced pool of cells, suggesting loss of chromosome 17 in ~40% of the sequenced Δkmt6 cells, a chromosome loss that likely occurred at the very beginning of the experiment.
Unexpectedly, the Δkmt1 progenitor displayed a long high-coverage (~1.6x) region on chromosome 1 (Fig 4), suggesting that the region had been duplicated in ~60% of the sequenced Δkmt1 cells. Furthermore, this genome has a shorter chromosome 6 and does not contain chromosome 20 (Fig 4A, S9 Table). The presence of this kind of structural variation in the progenitor strain is indicative for a high degree of genome instability in absence of Kmt1. Analysis of discordant reads mapped to both ends of the ~1 Mb high-coverage region on chromosome 1 revealed telomeric repeats (TTAGGGn), suggesting the formation of de novo telomeres. Pulsed-field gel electrophoresis (PFGE) and Southern analyses confirmed the formation of two new independent chromosomes both containing the high-coverage region and either the right or left arm of chromosome 1 (Fig 4C). The breakpoint on the left side coincides with a large TE-rich region that is associated with H3K9me3 in the wild type. Both breakpoints coincide with or are in close proximity to regions that show enrichment of relocated H3K27me3 in the Δkmt1 mutant (Fig 4B), suggesting a possible link between relocated H3K27me3 and genome instability.
After six months of vegetative growth, we sequenced the pooled genomes of all nine ‘evolved’ populations. We found no evidence for large-scale genomic rearrangements in any of the evolved Zt09 or Δkmt6 populations (Fig 5A). Apart from seven small deletions or duplications (S9 Table), the largest structural variation found in one of the evolved Δkmt6 populations (Δkmt6 50–2), was a partial loss (~18 kb) at the right end of chromosome 15. However, we found variation in the read coverage of accessory chromosomes in all sequenced genomes indicating whole chromosome losses in individual cells of the sequenced population. The distinct dynamics of individual accessory chromosome losses were described in the previous section as part of the short-term growth results (Table 2).
In contrast to the few variations detected in the Zt09 and Δkmt6 populations, we found numerous large-scale high-coverage regions on different core chromosomes, chromosome breakages followed by de novo telomere formation, chromosomal fusions, as well as several smaller deletions and duplications in the evolved Δkmt1 populations (Fig 5A, S9 Table). All three evolved Δkmt1 populations have large duplicated regions on chromosome 1 (Fig 6), but their locations as well as the resulting structural variations differ from the one identified in the progenitor strain (Fig 6A). This can be explained by independent events, as not all Δkmt1 progenitor cells underwent the rearrangement of chromosome 1 (Fig 4C), or by continuous structural rearrangement events as a consequence of the presence of large duplicated regions in the genome. Analyses of the affected regions and breakpoints indicate a connection between the structural variations of progenitor (compared to the reference) and evolved strains. In all evolved Δkmt1 populations, duplicated regions fully or partially overlap with the high-coverage region of the progenitor strain (Fig 6A–6D).
Since populations reflect a mixture of distinct genotypes, we also sequenced three single Δkmt1 clones originating from the populations from transfer 50 to characterize the structural variation in more detail (Fig 5B). The single clones were selected based on different PFGE karyotypes (S9 Fig) and originated from population Δkmt1-50-1 (Δkmt1-50-1-1) and Δkmt1-50-2 (Δkmt1-50-2-1 and Δkmt1-50-2-2). As two of these single clones (Δkmt1-50-1-1 and Δkmt1-50-2-2) largely resemble the genotypes found in their respective populations, we conclude the presence of a predominant genotype in each evolved replicate population. However, Δkmt1-50-2-1 clearly differs from this genotype and therefore reveals the existence of additional, rarer genotypes in the evolved populations. Relatively small deletions and duplications (up to 30 kb) as well as chromosome breakage followed by de novo telomere formation were found on almost all chromosomes. These occurred mainly linked to annotated TEs (S10 Table) whereby loss of H3K9me3 likely promoted instability. However, major rearrangements, including chromosomal fusions, were always linked to large segmental duplications (S10 Fig). In two strains we detected higher coverage of entire core chromosomes indicating core chromosome duplications (Fig 5B). Results from read coverage (S10 Table) and PCR analyses indicate that Δkmt1-50-2-2, as well as the majority of the Δkmt1-50-2 population, may have undergone a whole genome duplication.
To investigate whether the underlying sequence is involved in the formation of large-scale rearrangements, we analyzed the breakpoints of each duplicated region. The location of breakpoints does not show a clear TE-associated pattern as observed for the smaller deletions or chromosome breakages. Out of 28 analyzed breakpoints, only seven are directly located within annotated TEs, while thirteen fall into genes, seven are intergenic and one is located in the centromere (S11 Table). Considering all structural rearrangements in the three sequenced single clones, we found that out of 62 events, 34 were associated (direct overlap or <5 kb distance) to regions that show enrichment for H3K27me3 (S12 Table). Based on these observations, we hypothesize that two non-exclusive pathways, namely TE-associated instability caused by loss of H3K9me3 or invasion of H3K27me3, may serve as initial events, which are followed by continuous rearrangements possibly caused by increased mitotic recombination activity and deficiency in DNA repair resulting in a spectrum of structural variation (S11 Fig).
We investigated the effects of loss of two important heterochromatin-associated histone modifications, H3K9me3 and H3K27me3, on chromatin organization, transcription and genome stability and characterized phenotypes of the deletion mutants. Loss of H3K9me3 allows relocalization of H3K27me3 in kmt1 deletion mutants, which has great impact on genome and chromosome stability, resulting in numerous large-scale rearrangements. In contrast, the genomes of evolved Δkmt6 and Zt09 strains revealed only few and relatively minor changes. Unexpectedly, the presence of H3K27me3 impacts chromosome stability by either destabilizing whole chromosomes in normal cells, supported by the high loss-rate in the reference strain compared to the Δkmt6 mutants, or by mislocalization as shown by the increased sequence instability in the Δkmt1 mutants. Taken together, enrichment with H3K27me3 in wild type cells is a main driver of mitotic chromosome instability.
We propose different scenarios for how chromosomes may get lost during mitosis and how H3K27me3 may be linked to these processes. For example, accessory chromosomes may not be accurately replicated whereby only one sister chromatid is transmitted. Alternatively, non-disjunction of sister chromatids during mitosis produces one cell with two copies and one cell lacking the respective chromosome. Previous cytology on Z. tritici strains expressing GFP-tagged CENPA/CenH3 protein suggested that core and accessory chromosomes may be physically separated in the nucleus [31]. Previous studies showed that H3K27me3-enriched chromatin localizes near the nuclear periphery, and loss of H3K27me3 enables movement of this chromatin to the nucleus core in mammals and fungi [52,53]. Proximity to the nuclear membrane and heterochromatic structure can furthermore result in differential, and often late, replication timing [54,55]. Loss of H3K27me3 and the correlated movement to the inner nuclear matrix may alter replication dynamics of accessory chromosomes resulting in higher rates of faithfully replicated chromosomes and lower rates of mitotic loss (Fig 7).
Heterochromatic regions, especially associated with H3K27me3, tend to cluster together and form distinct foci in the nucleus of Drosophila melanogaster visualized by cytology [56,57], and loss of H3K27me3 reduces interaction between these regions [58]. We hypothesize that enrichment of H3K27me3 on the entire accessory chromosomes maintains physical interactions that persist throughout mitosis. This may decrease the efficiency of separation of sister chromatids resulting in loss of the chromosome in one cell and a duplication in the other cell. So far, we have focused our screening on chromosome losses but determining the exact rates of accessory chromosome duplications is necessary to test this hypothesis. Genome sequencing of Z. tritici chromosome loss strains revealed that duplications of accessory chromosomes can occur [46]. Similarly, B chromosomes in rye are preferentially inherited during meiosis by non-disjunction of sister chromatids during the first pollen mitosis [59], indicating that deviation from normal chromosome segregation occurs. Accessory chromosomes are commonly found in natural isolates of Z. tritici, despite the high loss rates we demonstrated during mitotic growth [46]. This observation implies the presence of other mechanisms that counteract the frequent losses of accessory chromosomes. Recent analyses of meiotic transmission showed that unpaired accessory chromosomes are transmitted at higher rates in a uniparental way [60,61]. We propose that H3K27me3 is involved in accessory chromosome instability and transmission both during mitosis and meiosis by influencing nuclear localization of chromosomes and thereby altering replication or transmission (Fig 7). Future analyses with fluorescently tagged core and accessory chromosomes and by chromosome conformation capture (Hi-C) will shed light on nuclear interactions and chromosome transmission dynamics. As not all accessory chromosomes, despite being enriched with H3K27me3, are lost at the same rate, we note that additional mechanisms likely contribute to accessory chromosome dynamics.
While loss of H3K27me3 resulted in only minor differences to wild type growth and, unexpectedly, rather promoted than decreased genome stability, we detected a high number of smaller (up to 30 kb) deletions and duplications, chromosome breakages and several gross chromosomal rearrangements linked to large duplications in the Δkmt1 mutants. Absence of H3K9me2/3 has been associated with chromosome and genome instability in other organisms [13,14,62,63]. Smaller deletions, duplications and chromosome breakages resulting in shortened chromosomes due to loss of chromosome ends that we identified in the Δkmt1 mutants, correlate with TEs, enriched with H3K9me3 in wild type. Replication of heterochromatin-associated DNA is challenging for the cell as repetitive sequences may form secondary structures that can stall the replication machinery [64]. Consequently, instability of repeated sequences has been linked to errors during DNA replication [65–67]. Furthermore, the structural variation that arises depends on the mode of DNA repair following the DNA damage [68,69]. The structural rearrangements detected in the Δkmt1 mutants indicate that repair of double-strand breaks involves both non-homologous end joining and de novo telomere formation. We propose that the main factor for genome instability is replication-associated instability of repeated sequences subsequently promoting the formation of large-scale rearrangements (S11 Fig).
Not all breakpoints of rearrangements, especially of the large duplicated sequences, were associated with TEs, however. We found that duplicated sequences in the experimentally evolved Δkmt1 mutants fully or partially overlap with the duplicated regions of the Δkmt1 progenitor strain. This suggests that structural variations are subject to continuous rearrangements, resulting in rearrangements no longer directly linked to the initial event. We note that the rearrangements and genotypes we detected are the result of selection during our long-term growth experiments and thus do not necessarily reflect the full spectrum of rearrangements occurring in Δkmt1 mutants; many additional structural variants may have disappeared quickly from the population or included lethal events.
Concomitant with loss of H3K9me3 in the Δkmt1 strains, we found relocalization of H3K27me3 to former H3K9me3 regions. A similar redistribution of H3K27me3 in absence of heterochromatin factors has been reported in plants and animals [70–72] and other fungi [22,27,73]. In N. crassa, redistribution of H3K27me3 in a Δkmt1 (dim-5) mutant background results in severe growth defects and increased sensitivity to genotoxic stress that can be rescued by elimination of H3K27me3, indicating that aberrant H3K27me3 distribution severely impacts cell viability [27]. Although we did not see rescue of phenotypic defects observed in planta or in in vitro stress assays in the Δk1/k6 double mutants, the chromosome-loss rate was reduced compared to Δkmt1 mutants, suggesting a stabilizing effect when H3K27me3 is absent. We found that some breakpoints of the rearrangements in the Δkmt1 mutants without H3K9me3 also show enrichment with the invading H3K27me3. This finding also suggests that sequences associated with H3K27me3 are more susceptible to genome instability. Regions enriched with H3K27me3 have been shown to exhibit a high degree of genetic variability in form of mutations, increased recombination, or structural variation compared to the rest of the genome [21,23,30,31,74–76]. Experimental evolution in Fusarium fujikuroi showed that increased H3K27me3 levels in subtelomeric regions coincided with increased instability [77] and we previously detected a highly increased rate of chromosomal breakage under stress conditions in subtelomeric H3K27me3 regions in Z. tritici [46]. These observations together with our findings strongly indicate that H3K27me3 plays a pivotal role in decreasing genome stability.
In summary, the presence of Kmt1 and H3K9me3 respectively, is essential to maintain genome integrity in this fungus. TE-mediated rearrangements may be involved in the genetic variability detected in Z. tritici isolates [78–80] and have been suggested as drivers of genome evolution in various species [81–83]. Our findings concerning the role of H3K9me3 for genome stability provide a basis for future studies focusing on the influence of heterochromatin on structural genome rearrangements using Z. tritici as a model organism. We found that, unlike for H3K9me3, presence and not absence of H3K27me3 is linked to genome instability. Surprisingly, loss of H3K27me3 does not result in dramatic changes of overt phenotypes and is also not clearly linked to transcriptional activation in Z. tritici. This allowed us to uncouple the transcriptional and regulatory effects of H3K27me3 from the influence on chromatin stability and will in the future result in further mechanistic insights on the influence of histone modifications on chromosome stability.
Zymoseptoria tritici strains were cultivated on solid (2% [w/v] Bacto agar) or in liquid YMS medium (0.4% [w/v] yeast extract, 0.4% [w/v] malt extract, 0.4% [w/v] sucrose). Liquid cultures were inoculated from plate or directly from glycerol stocks and grown for 3–4 days at 18°C in a shaking incubator at 200 rpm. Plates were inoculated from glycerol stocks and grown for 5–6 days at 18°C. Escherichia coli TOP10 cells were grown overnight in dYT (1.6% [w/v] tryptone, 1% [w/v] yeast extract, 0.5% [w/v] NaCl and 2% Bacto agar for solid medium) supplemented with antibiotics for plasmid selection (40 μg/mL kanamycin) at 37°C and at 200 rpm for liquid cultures. Agrobacterium tumefaciens strain AGL1 was grown in dYT containing rifampicin (50 μg/mL) and carbenicillin (100 μg/mL) supplemented with antibiotics for plasmid selection (40 μg/mL kanamycin) at 28°C at 200 rpm in liquid culture for 18 h and on plate at 28°C for two days.
Z. tritici deletion and complementation strains were engineered using A. tumefaciens-mediated transformations as described before [49,84]. Flanking regions of the respective genes were used to facilitate homologous recombination for integration at the correct genomic location. The plasmid pES61 (a derivate of the binary vector pNOV-ABCD [49]) was used for targeted gene deletion and complementation. Plasmids were assembled using a restriction enzyme-based approach or Gibson assembly [85]. Plasmids were amplified in E. coli TOP10 cells and transformed in the A. tumefaciens strain AGL1 as described previously [86]. Gene deletions of kmt1 (Zt09_chr_1_01919) and kmt6 (Zt09_chr_4_00551) were facilitated by replacement of the respective ORF with a hygromycin resistance cassette (hph). The kmt1/kmt6 double deletion mutant was constructed by integrating a nourseothricin resistance cassette (nat) replacing kmt1 in a kmt6 deletion mutant background. Complementation constructs containing the respective gene and a G418 resistance cassette (neo) were integrated at the native loci in the deletion strains. All plasmids and strains constructed in this study are listed in S1 Table. Transformed strains were screened by PCR for correct integrations of the construct followed by Southern blot [87] with probes generated by DIG labeling (Roche, Mannheim, Germany) following manufacturer’s instructions.
For rapid PCR screening (candidates for transformation and chromosome loss), a single Z. tritici colony was resuspended in 50 μL of 25 mM NaOH, incubated at 98°C for 10 min and afterwards 50 μL of 40 mM Tris-HCl pH 5.5 were added. Four μl of the mix was used as template for PCRs. For DNA extraction for Southern blotting, we used a standard phenol-chloroform extraction protocol [88] for DNA isolation.
For the in vitro growth assays, liquid YMS cultures were inoculated with 100 cells/μL (OD600 = 0.01); cells were grown in 25 mL YMS at 18°C and 200 rpm. For each mutant and complementation strain, two transformants (biological replicates), and three replicate cultures per transformant (technical replicates) were used. For the reference strain Zt09, two separate pre-cultures were grown as biological replicates and each pre-culture was used to inoculate three replicate cultures. OD600 was measured at different time points throughout the experiment until the stationary phase was reached. The R package growthcurver [89] was used to fit the growth curve data enabling to compare in vitro growth of the different strains.
To test the tolerance of mutant and reference strains towards different stressors, we performed an in vitro stress assay on YMS plates. Each plate contained additives constituting different stress conditions. Cell suspensions containing 107/108 cells/mL and a tenfold dilution series down to 100 cells/mL were prepared; 3 μL of each dilution were pipetted on solid YMS containing the following additives: 0.5 M NaCl, 1 M NaCl, 1 M sorbitol, 1.5 M sorbitol, 1.5 mM H2O2, 2 mM H2O2, 300 μg/mL Congo red, 0.01% MMS (methyl methane sulfonate), 0.025% MMS, 1 μg/mL actinomycin D and 1.5 μg/mL actinomycin D. Furthermore, we included a H2O-agar (2% bacto agar) plate. All plates were incubated at 18°C for six days, except for one YMS plate that was incubated at 28°C to test for thermal stress responses.
Seedlings of the susceptible wheat cultivar Obelisk (Wiersum Plantbreeding BV, Winschoten, The Netherlands) were potted (three plants per pot) after four days of pre-germination and grown for seven more days. Single cell suspensions of mutant and reference strain were prepared (108 cells / mL in H2O with 0.1% Tween 20) and brush inoculated on a marked area of the second leaf. Following inoculation, the plants were incubated in sealed plastic bags containing ~ 1 L of H2O for 48 h providing high humidity to promote infections. Growth conditions for the plants throughout the complete growth phase and infection were 16 h light (200 μmol/m-2s-1) and 8 h dark at 20°C and 90% humidity. First appearances of symptoms, necrosis or pycnidia, were assessed by manual inspection of every treated leaf. 21 or 28 days post infection, inoculated leaves were finally screened for infection symptoms. Visual inspection of each leaf was performed to evaluate the percentage of leaf area covered by necrosis and pycnidia. Six different categories were differentiated based on the observed coverage (0: 0%, 1: 1–20%, 2: 21–40%, 3: 41–60%, 4: 61–80%, 5: 81–100%). Furthermore, automated symptom evaluation was performed by analysis of scanned images of infected leaf areas as described previously [90].
For the long-term evolution experiment (~6 months), cells were inoculated directly from the glycerol stocks into 20 mL liquid YMS cultures. We used Zt09, Δkmt6 (#285) and Δkmt1 (#68), each strain grown in triplicates. Every three to four days, cells were transferred to new YMS medium. Cells were grown at 18°C and 200 rpm. For every transfer, cell density of the cultures was measured by OD600 and the new cultures were inoculated with a cell density of ~ 100 cells / μL (correlating to a transfer of 0.1% of the population). After 50 transfers, the genomes of the evolved populations and each progenitor strain were sequenced. Additionally, three genomes of single clones derived from the Δkmt1 populations after 50 transfers were sequenced to characterize genome rearrangements in more detail.
For the short-term evolution experiment over a time period of four weeks, cultures were inoculated from single colonies grown on solid YMS. Zt09, Δkmt6 (#285), Δkmt1 (#80), Δk1/Δk6 (#23) double mutant, kmt1+ (#42) and kmt6+ (#11) were grown in triplicate YMS cultures. For this experiment we used a different independent Δkmt1 mutant clone (#80), as we discovered that the strain used in the previous long-term evolution experiment (#68) was missing chromosome 20. Every three to four days, 900 μL culture were transferred to 25 mL fresh YMS (correlating to a transfer of ~ 4% of the population). After four weeks of growth (including eight transfers to new medium) at 18°C and 200 rpm, cultures were diluted and plated on YMS agar to obtain single colonies. These single colonies were PCR screened for presence of accessory chromosomes as described in [46].
Cells were grown in YMS medium for five days and harvested by centrifugation for 10 min at 3,500 rpm. We used 5 x 108 cells for plug preparation that were washed twice with Tris-HCl, pH 7.5, resuspended in 1 mL TE buffer (pH 8) and mixed with 1 mL of 2.2% low range ultra agarose (Bio-Rad, Munich, Germany). The mixture was pipetted into plug casting molds and cooled for 1 h at 4°C. Plugs were placed to 50 mL screw cap Falcon tubes containing 5 mL of lysis buffer (1% SDS; 0.45 M EDTA; 1.5 mg/mL proteinase K [Roth, Karlsruhe, Germany]) and incubated for 48 h at 55°C while the buffer was replaced once after 24 h. Chromosomal plugs were washed three times for 20 min with 1 X TE buffer before storage in 0.5 M EDTA at 4°C. PFGE was performed with a CHEF-DR III pulsed-field electrophoresis system (BioRad, Munich, Germany). Separation of mid-size chromosomes was conducted with the settings: switching time 250 s– 1000 s, 3 V/cm, 106° angle, 1% pulsed-field agarose in 0.5 X TBE for 72 h. Large chromosomes were separated with the following settings: switching time 1000 s– 2000 s, 2 V/cm, 106° angle, 0.8% pulsed-field agarose in 1 X TAE for 96 h. Saccharomyces cerevisiae chromosomal DNA (BioRad, Munich, Germany) was used as size marker for the for mid-size chromosomes, Schizosaccharomyces pombe chromosomal DNA (BioRad, Munich, Germany) for the large chromosomes. Gels were stained in ethidium bromide staining solution (1 μg/mL ethidium bromide in H2O) for 30 min. Detection of chromosomal bands was performed with the GelDocTM XR+ system (Bio-Rad, Munich, Germany). Southern blotting was performed as described previously [87] but using DIG-labeled probes generated with the PCR DIG labeling Mix (Roche, Mannheim, Germany) following the manufacturer’s instructions.
Cells were grown in liquid YMS medium at 18°C for 2 days until an OD600 of ~ 1 was reached. Chromatin immunoprecipitation was performed as previously described [91] with minor modifications. We used antibodies against H3K4me2 (#07–030, Merck Millipore), H3K9me3 (#39161, Active Motif) and H3K27me3 (#39155, Active Motif). ChIP DNA was purified using SureBeads Protein G Magnetic Beads (Bio-Rad, Munich, Germany) and, replacing phenol/chloroform extractions, we used the ChIP DNA Clean & Concentrator Kit (Zymo Research, Freiburg, Germany). We sequenced two biological and one additional technical replicate for Zt09, Δkmt1, Δkmt6, and the Δk1/k6 strains. Sequencing was performed at the OSU Center for Genome Research and Biocomputing on an Illumina HiSeq2000 or HiSeq3000 to obtain 50-nt reads and at the Max Planck Genome Center, Cologne, Germany (https://mpgc.mpipz.mpg.de/home/) on an Illumina Hiseq3000 platform obtaining 150-nt reads (S2 Table).
For RNA extraction, cells were grown in liquid YMS at 18°C and 200 rpm for two days until an OD600 of ~ 1 was reached. Cells were harvested by centrifugation and ground in liquid nitrogen. Total RNA was extracted using TRIzol (Invitrogen, Karlsruhe, Germany) according to manufacturer’s instructions. The extracted RNA was further DNAse-treated and cleaned up using the RNA Clean & Concentrator-25 Kit (Zymo Research, Freiburg, Germany). RNA samples of two biological replicates of Zt09, Δkmt1, Δkmt6, and the Δk1/k6 double mutant were sequenced. Poly(A)-captured, stranded library preparation and sequencing were performed by the Max Planck-Genome-centre Cologne, Germany (https://mpgc.mpipz.mpg.de/home/) on an Illumina Hiseq3000 platform obtaining ~ 20 million 150-nt reads per sample (S2 Table).
Genomic DNA for sequencing was prepared as described previously [92]. Library preparation and genome sequencing of the progenitor strains used for the evolution experiments were performed at Aros, Skejby, Denmark using an Illumina HiSeq2500 platform obtaining 100-nt paired-end reads. Library preparation (PCR-free) and sequencing of the evolved populations and the three evolved single Δkmt1 mutants were performed by the Max Planck Genome Center, Cologne, Germany (https://mpgc.mpipz.mpg.de/home/) on an Illumina HiSeq3000 platform resulting in 150-nt paired-end reads (S2 Table).
A detailed list of all programs and commands used for mapping and sequencing data analyses can be found in the supplementary text S1. All sequencing data was quality filtered using the FastX toolkit ((http://hannonlab.cshl.edu/fastx_toolkit/) and Trimmomatic [93]. RNA-seq reads were mapped using hisat2 [94], mapping of ChIP and genome data was performed with Bowtie2 [95]. Conversion of sam to bam format, sorting and indexing of read alignments was done with samtools [96].
To detect enriched regions in the ChIP mappings, we used HOMER [97]. Peaks were called individually for replicates and merged with bedtools [98]. Only enriched regions found in all replicates were considered for further analyses. Genome coverage of enriched regions and overlap to genes and TEs was calculated using bedtools [98].
We used cuffdiff [99] to calculate RPKM values and to estimate expression in the different strains. Raw reads mapping on genes and TEs were counted by HTSeq [100], differential expression analysis was performed in R [101] with DESeq2 [102]. Cutoff for significantly differentially expressed genes was padj < 0.001 and |log2 fold-change| > 2. The R package topGO [103] was used to perform gene ontology enrichment analyses. Fisher’s exact test (p-value < 0.01) was applied to detect significantly enriched terms in the category ‘biological process’.
To detect structural variation in the sequenced genomes, we used SpeedSeq [104] and LUMPY [105]. All detected variation was further verified by manual visual inspection. Visualization was performed with the integrative genome browser (IGV) [106].
|
10.1371/journal.pntd.0005057 | Osteopontin Is Upregulated in Human and Murine Acute Schistosomiasis Mansoni | Symptomatic acute schistosomiasis mansoni is a systemic hypersensitivity reaction against the migrating schistosomula and mature eggs after a primary infection. The mechanisms involved in the pathogenesis of acute schistosomiasis are not fully elucidated. Osteopontin has been implicated in granulomatous reactions and in acute hepatic injury. Our aims were to evaluate if osteopontin plays a role in acute Schistosoma mansoni infection in both human and experimentally infected mice and if circulating OPN levels could be a novel biomarker of this infection.
Serum/plasma osteopontin levels were measured by ELISA in patients with acute (n = 28), hepatointestinal (n = 26), hepatosplenic (n = 39) schistosomiasis and in uninfected controls (n = 21). Liver osteopontin was assessed by immunohistochemistry in needle biopsies of 5 patients. Sera and hepatic osteopontin were quantified in the murine model of schistosomiasis mansoni during acute (7 and 8 weeks post infection, n = 10) and chronic (30 weeks post infection, n = 8) phase. Circulating osteopontin levels are increased in patients with acute schistosomiasis (p = 0.0001). The highest levels of OPN were observed during the peak of clinical symptoms (7–11 weeks post infection), returning to baseline level once the granulomas were modulated (>12 weeks post infection). The plasma levels in acute schistosomiasis were even higher than in hepatosplenic patients. The murine model mirrored the human disease. Macrophages were the major source of OPN in human and murine acute schistosomiasis, while the ductular reaction maintains OPN production in hepatosplenic disease. Soluble egg antigens from S. mansoni induced OPN expression in primary human kupffer cells.
S. mansoni egg antigens induce the production of OPN by macrophages in the necrotic-exudative granulomas characteristic of acute schistosomiasis mansoni. Circulating OPN levels are upregulated in human and murine acute schistosomiasis and could be a non-invasive biomarker of this form of disease.
| Schistosomiasis is a major health problem that affects over 200 million people. Symptomatic acute schistosomiasis is a systemic reaction to the worms and eggs in individuals from non-endemic areas after a primary infection. Tourists, military personnel and people who practice water sports are at risk. Although most cases resolve 90 days post infection, severe cases with massive distribution of eggs can be fatal. It is frequently misdiagnosed, under diagnosed or has delayed diagnosis because the signs and symptoms are nonspecific and eggs are usually present in stool only 6 weeks post-infection. The mechanisms underlying the pathogenesis of acute schistosomiasis are not fully elucidated and currently there is a lack of noninvasive biomarkers to diagnose this form of disease. We report that serum osteopontin levels are increased in patients with acute schistosomiasis and parallel the clinical symptoms, returning to baseline level once the granulomas were modulated and the symptoms resolve. Soluble egg antigens provoke macrophages to produce osteopontin, recruiting more macrophages to the site of injury and inducing the granulomatous reaction. This observation suggests that osteopontin plays an important role in acute schistosomiasis mansoni and could be a novel non-invasive biomarker for this form of the disease.
| Schistosomiasis is a severe tropical disease caused by Schistosoma spp. flatworms that affects over 200 million of people from 76 countries and territories [1]. S. mansoni is the only species in the Americas and Brazil holds the majority of infected individuals with 25 million living in endemic areas and 4–6 million infected [2].
Infected individuals have various clinical manifestations that generally cluster into three distinct forms of the disease: acute, hepatointestinal and hepatosplenic schistosomiasis [2–5]. In patients from endemic areas, the acute phase of schistosomiasis is rarely symptomatic (0.3%) due to infection early in life (3–4 years-old) and exposure to schistosoma antigens/antibodies against antigens in-utero and/or in breast milk [3]. The majority of chronically infected patients from endemic areas (90–96%) develop the hepatointestinal form of the disease, which is asymtomatic or oligosymptomatic in most cases and characterized by granulomatous inflammation in the liver and intestines, little or no hepatosplenomegaly, and minimal liver fibrosis without any sign of portal hypertension [2, 4–7]. A small proportion (4–10%) of infected individuals from endemic areas develops the hepatosplenic form of disease characterized by hepatosplenomegaly, severe liver fibrosis and portal hypertension [2, 4–7].
Among individuals from non-endemic areas, the acute form of schistosomiasis mansoni is a systemic hypersensitivity reaction against the migrating schistosomula (pre-postural phase of infection) and mature eggs (post-postural phase of infection). This typically develops within 16–90 days after a primary infection [2]. The burden of infection (and probably host genetic background) dictates the severity of the clinical manifestations: more worm couples produce more eggs and consequently, trigger an exacerbated host immune response [2, 8].
The pre-postural phase occurs during the initial 35 days after infection and is caused by immune modifications induced by the schistosomules, immature and adult worms before laying eggs [2]. Cercarial dermatitis may occur soon after infection, but symptoms are more evident when schistosomules/immature worms arrive/grow/mature in the hepatoportal veins (peak 15–21 days post infection) [2]. High fever (38–39°C), cough, abdominal pain, discrete hepatosplenomegaly and nonspecific symptoms such as muscular pain, arthralgia and headache, are observed [2, 9]. Blood eosinophilia (10–75% of eosinophils) is frequent [2, 9]. Liver biopsy reveals discrete inflammatory infiltrate consisting of lymphocytes, eosinophils, neutrophils and macrophages surrounding schistosomules/immature worms, non-specific portal hepatitis, and sparse focal intralobular necrosis [2]. During this phase a Th1 response is predominant and an increase in pro-inflammatory cytokines such as IL-2, gamma Interferon and TNF alpha is frequently observed [2, 3]. Other less frequent clinical manifestations may be present: transverse myelitis or pseudotumoral lesions in encephalon (neural schistosomiasis) [2, 9].
The post-postural phase is initiated by egg laying (approximately 35 days post-infection) and egg maturation (which begins about 6 days later) [2]. Symptoms are aggravated, episodes of diarrhea increase, and the patient experience severe weight loss [2, 4, 8]. Clinical symptoms can continue until 90 days after infection [2, 4, 8]. Severe, toxemic forms of acute disease in which there is massive dissemination of eggs throughout the intestines and lungs may be fatal [2]. Moderate to mild disease spontaneously resolves two to three months after infection [2].
During acute schistosomiasis intense miliary distribution of eggs occurs in the liver, colons, small intestines, visceral peritoneum, abdominal lymph nodes, pancreas and lungs [2]. Periovular granulomas localize on the serosal surface of affected organs and macroscopically appear as translucent granule or nodules [2]. Microscopically, the granulomas are large (over 100 times the size of the egg), necrotic-exudative, and enriched with eosinophils [2], due to the naïve hosts’ hyperergic reaction to novel parasite antigens. In the liver, granulomas are irregularly distributed through the parenchyma and portal tracts and non-specific inflammatory cells frequently surround portal tracts. Because hepatocellular lesions are relatively mild (loss of basophilia, hydropic degeneration and rare focal necrosis), the serum aminotransferases are usually normal or slightly elevated [2, 9]. An important feature of acute schistosomiasisis is that all the granulomas are uniformly in the same necrotic-exudative phase of formation, with prominent central necrosis [2]. This finding in liver biopsies is pathognomonic of acute infection.
With egg-laying the Th2 immune response starts to suppress the initial Th1 response and IL4, IL5, IL10 and IL13 are the most predominant cytokines [2, 3]. The hyperergic, massive granulomas are modulated as the infection evolves to the chronic phase. By around 90 days post-infection, liver granulomas are smaller [2, 3, 10–12] and progressively heal by fibrosis [2, 7, 10, 12]. The symptoms usually disappear due to the modulation of the immune response to the eggs [2]. Because the signs and symptoms of acute schistosomiasis are nonspecific and diagnosis is established by presence of eggs in stools that occurs only six weeks after infection, acute schistosomiasis mansoni is frequently misdiagnosed, under diagnosed or has delayed diagnosis [2, 9]. Efforts to develop tests for earlier diagnosis of the disease have been challenging. Unfortunately, lesions similar to those observed in pre-postural phase of human acute schistosomiasis are not observed in mouse models of schistosomiasis mansoni [11, 13, 14], likely because the granulomas that form in mice are generally less necrotic than those that occur in acutely infected humans [14].
Osteopontin (OPN), a pro-inflammatory cytokine and pro-fibrogenic molecule [15–17], was recently associated with hepatosplenic schistosomiasis mansoni [18]. Soluble egg antigens (SEA) directly induce liver cells to produce OPN. Moreover, serum and hepatic osteopontin levels correlate with the degree of liver fibrosis and the level of portal hypertension, suggesting that this molecule could be a novel biomarker for hepatosplenic schistosomiais mansoni [18]. The authors observed that macrophages, stellate cells and bile ductular cells in/around the granulomatous reaction are the major sources of OPN in schistosomiasis [18]. Osteopontin was also demonstrated to play a role in recruitment and activation of macrophages/Kupffer cells, neutrophils and lymphocytes [15–17, 19]. OPN-/- mice injected with S. mansoni eggs develop abnormal granuloma formation in the lung due to reduced macrophage accumulation [20]. Since in acute schistosomiasis the liver is enriched with necrotic-exudative granulomas and there is an exacerbated immune response, our aims were to evaluate if OPN increases in acute Schistosoma mansoni infection of both humans and mice, and to determine if circulating OPN levels might be a novel biomarker of this infection.
This was a comparative cross-sectional study. A total of 28 patients with acute schistosomiasis mansoni diagnosed at Tropical Diseases Outpatient Clinic of the University Hospital of Universidade Federal de Minas Gerais (Belo Horizonte, Brazil) from January 2014 to December 2015 were included in the study. Serum samples from acute patients (n = 28; age 19.8±11.8 years; 21 males/7 females) and uninfected controls (n = 21; age 27.86±9.45 years; 14 males/7 females) were collected for analysis. Formalin-fixed, paraffin-embedded liver needle biopsies were available in a subgroup of patients (n = 5). Plasma samples from uninfected controls (n = 21) and from patients with Hepatointestinal (n = 27; age 35.66±12.09 years; 16 males/7 females), Hepatosplenic (n = 39; age 38.25±9.4 years; 30 males/9 females) and Acute (n = 3; age 39±25.23; 3 males) schistosomiasis were also included in the analysis.
Diagnosis of acute schistosomiasis was based on epidemiological data (recent contact with stream water in an endemic area), clinical data (cercarial dermatitis, acute enterocolitis, fever, cough, malaise, paraplegia, pulmonary involvement, hepatomegaly and or splenomegaly), laboratory assays (eosinophilia, IgG antibodies against SWAP, S. mansoni eggs in stools or rectal biopsy fragments), and imaging techniques (Ultrasound to observe liver, spleen and intra abdominal lymph node enlargement; MRI to demonstrate spinal cord injury). To be considered as having acute schistosomiasis in the present study the participants had to have more than 1 or more symptoms/signs described above, evidence of infection (parasitologic or serologic) and reported contact with contaminated waters. All patients included in the study were residents of the metropolitan region of Belo Horizonte (capital of Minas Gerais state), a non-endemic area for schistosomiasis mansoni. No previous history of contact with S. mansoni was reported by the patients or parents/guardians.
The present study was conducted in accordance with the Declaration of Helsinki (2013) of the World Medical Association and was approved by the Ethics Committee of Universidade Federal de Minas Gerais, Belo Horizonte, Minas Gerais, Brazil (UFMG) (Protocol ETIC 204/06). Written informed consent was obtained from all participating subjects or their parents/guardians (on behalf of child participant). All data regarding human participants was anonymized.
Female Swiss Webster outbred mice were infected with 50 cercariae of S. mansoni (Feira de Santana strain, CPqGM/FIOCRUZ) for 6, 7, 8 weeks (acute phase, n = 15) and 30 weeks (chronic phase, severe fibrosis, n = 8). Uninfected, age- and strain-matched animals were used as controls (n = 8). Liver tissue and serum were collected for analysis. The present study protocol meet the regulation and guidelines of Brazil’s National Animal Experimentation Control Board (CONCEA) and was approved (Protocol 003/2010) by the Ethical Committee for Animal Research of Centro de Pesquisas Gonçalo Moniz, Oswaldo Cruz Foundation, Salvador, Bahia, Brazil (CPqGM/FIOCRUZ).
OPN was quantified in the serum (humans and mice) or plasma (humans) using OPN Quantikine ELISA kit (R&D Systems) according to the manufacturer’s protocol.
Liver sections were stained with H&E (haematoxylin and eosin) for general histology. Immunohistochemistry (IHC) analysis was performed to evaluate the expression of osteopontin (R&D Systems; Antigen retrieval: 3% pepsin digestion for 10 min at 37°C; 5ug/mL of primary antibody, incubation overnight at 4°C). To confirm that Macrophages produce osteopontin, double IHC was performed using the chromagen DAB (3,3_-diaminobenzidine) for OPN and the chromagen Vina Green for CD68 (a macrophage marker).
OPN staining was quantified in 15 x200 fields/sample by computer-assisted morphometry using MetaMorph (Universal Imaging Corp.). OPN (+) bile ducts were counted in 15 x200 fields/sample by three independent observers.
The SEA was prepared at Centers for Disease Control and Prevention (CDC) as previously described [21]. The amount of Gram-negative bacterial endotoxin present in the SEA preparation was quantified using the end-point chromogenic limulus amebocyte lysate assay (Lonza). To investigate if macrophages produce osteopontin, primary human Kupffer cells (from Thermo Fisher Scientific) were incubated with 10 μg/ml SEA or 0.0001 μg/ml LPS (lipopolysaccharide; control, same amount of endotoxin present in the SEA preparation) for 3,6,12 and 24 hours. RNA was collected for analysis.
RNA was extracted using RNeasy mini kit (Qiagen) according to the manufacture’s protocol. Reverse transcription was performed using the First Strand Superscript III kit (Life Technologies) using the random hexamers protocol. Osteopontin mRNA expression was evaluated by real-time PCR (Taqman, Thermo Fisher Scientific). Each sample was analysed in duplicate and target gene levels in treated cells are shown as a ratio to levels detected in corresponding control samples, according to the ΔCT method, relative to the housekeeping gene (18s). The probes were designed by Thermo Fisher Scientific.
Results are expressed as means ± S.E.M. (Standard Error of the Mean; for normal distribution variables) or as medians (for non-normal distribution variables). Comparisons between groups were performed using the oneway ANOVA and Student’s t test (parametric) or Kruskal–Wallis one-way ANOVA and Mann–Whitney U test (non-parametric). Significance was accepted at the 0.05 level; Bonferroni correction was applied when comparing more than two groups. Receiver operating characteristics (ROC) curve analysis was used to investigate if sera OPN levels could be a good biomarker for symptomatic acute schistosomiasis. All statistical analyses were performed using SPSS Statistics 22 (IBM) and Prism 6 (GraphPad).
Our cohort of patients consisted of classic cases of symptomatic acute schistosomiasis mansoni, fulfilling the criteria for case definition of the acute form of the disease. The most frequent symptoms in acute cases are depicted in Table 1.
Patients with acute schistosomiasis have increased circulating levels of osteopontin in the plasma (p = 0.0005 vs non-infected; p = 0.0005 vs HI and p = 0.0012 vs HS) (Fig 1A) and serum (p = 0.0001) (Fig 1B). The plasma OPN levels in acute schistosomiasis are even higher than in patients with hepatosplenic form of the disease (p = 0.0012) (Fig 1A). We observe that OPN starts to increase in the beginning of the post-postural phase (5–6 weeks post-infection, p = 0.0005 vs non-infected) and OPN levels peaked 7–11 weeks post-infection (p = 0.0001 vs uninfected; p = 0.04 vc 5–6 weeks; p = 0.001 vs 12 weeks and p = 0.0001 vs 24 weeks), when the livers are enriched with necrotic-exudative granulomas (Fig 1C). Twelve weeks after infection the symptoms start to disappear, the granulomas reach a modulated state and circulating OPN levels start to fall, reaching levels comparable to uninfected individuals 24 weeks post-infection (Fig 1C). Receiver operating characteristics (ROC) curve analysis demonstrated that serum OPN measurement could be a good biomarker to identify patients with symptomatic acute schistosomiasis mansoni (Area under the curve = 0.9959; p<0.0001; 95% confidence interval 0.9848–1.007; S1 Fig). In our study population OPN serum test >23.34 can detect a symptomatic acute patient with 95.65% sensitivity and 95.24% specificity (Likelihood ratio = 20.09).
Immunohistochemistry demonstrated that the inflammatory cells in the necrotic-exudative liver granulomas express OPN, especially in the macrophage (epithelioid cells) enriched area around the egg and central necrosis (Fig 1D and S2 Fig).
Similar to humans, mice in the acute phase of infection also have more circulating and hepatic OPN levels than mice in the chronic phase of infection where there is severe fibrosis (p = 0.001 vs Non-infected; p = 0.0124 vs chronic phase) (Fig 2A, 2C and 2D). OPN levels in mice also peaked in the liver (p = 0.0001 vs non-infected; p = 0.0245 vs 6 weeks and p = 0.0104 vs 30 weeks) and serum (p = 0.0286 vs non-infected; p = 0.0286 vs 6 weeks; p = 0.0286 vs 8 weeks and p = 0.004 vs 30 weeks) 7 weeks post-infection, at a time when the livers were enriched with necrotic-exudative granulomas and inflammatory cells (Fig 2B, 2C and 2D). During the acute phase of infection in both mice and humans, the majority of liver OPN producing cells are inflammatory cells (Figs 1D and 2C; S2 Fig), while the ductular reaction is the most important source of OPN in chronic schistosomiasis (Fig 2C and 2E).
OPN expression in both human and murine acute schistosomiasis is enriched in the macrophage area of the necrotic-exudative granulomas. Double immunohistochemistry for OPN and CD68 (a macrophage marker) confirmed that the macrophages in acute schistosomiasis express this pro-inflammatory cytokine (Fig 3A). Since the macrophages are in contact with egg antigens, we investigated if soluble egg antigens could stimulate OPN production in vitro. Primary human Kupffer cells incubated with SEA for 3 hours upregulated OPN mRNA (p = 0.0082) (Fig 3B), indicating that infection per se can directly increase macrophage expression of this proinflammatory and profibrogenic molecule.
We demonstrated for the first time that circulating osteopontin levels are increased in human acute schistosomiasis mansoni. Our results also suggest that serum OPN measurement could be a good biomarker to diagnose symptomatic acute schistosomiasis. The highest levels of OPN were observed in patients during the peak of clinical symptoms (7–11 weeks post infection). Once the granulomas were modulated (>12 weeks post infection) the OPN levels decrease significantly.
Circulating and hepatic OPN levels were also elevated in the acute phase of experimental murine schistosomiasis mansoni. Chen et al. (2011) demonstrated that liver OPN levels peaked at the acute phase of S. japonicum infection. As previously mentioned, the murine model has some limitations in regard to acute schistosomiasis [14]. However the model may be helpful to identify the factors related to the onset of the generalized reactive changes during the early course of a primary schistosomal infection [14]. Importantly, our new data in humans demonstrate that the mouse model mirrored the human disease with regards to the pattern of OPN expression, reinforcing that this model could be useful to understand the mechanisms related to the acute phase of schistosomiasis in humans.
Macrophages are the major OPN producing cell in acute schistosomiasis and SEA induces OPN expression in primary human Kupffer cells. Pereira et al. (2015) also observed that macrophages are one of the major sources of OPN in the early phases of infection in mice and in patients with hepatointestinal schistosomiasis, while bile ducts are the main producers of OPN in patients with hepatosplenic disease. We confirm that osteopontin is mostly expressed by the ductular reaction in mice in the late chronic phase of infection. Pereira et al. (2015) also observed that SEA stimulates primary mouse Kupffer cells, stellate cells and cholangiocytes to produce OPN, demonstrating that egg antigens directly induce the expression of this pro-inflammatory and pro-fibrogenic molecule by multiple types of cells that localize in schistosoma-infected livers.
Osteopontin has been previously associated with acute hepatic injury [16, 17, 22, 23]. Patients with acute liver failure of different etiologies such as acetominophen toxicity, ischemia (shock), idiosyncratic drug-induced liver injury, autoimmune hepatitis and viral hepatitis A and B, have increased OPN plasma levels [22, 23]. Recent findings indicate that OPN plays a central role in liver diseases associated with necrosis [16, 17, 23]. Liver injury triggers OPN production in Kupffer cells and NKT cells that attract neutrophils, lymphocytes and macrophages to affected areas [16, 17, 19, 24]. The recruited cells become activated and produce OPN and Th1 cytokines, exacerbating liver necrosis [16, 17, 19, 24]. In acute liver failure patients, OPN was particularly associated with hyperactute injury [23].
The role of OPN has been described in granulomatous reactions, especially Th1-mediated, [16, 19]. OPN is essential for Th1 polarization [25] and OPN from dendritic cells mediates granuloma formation against bacterial antigens [26]. OPN expression in sarcoidosis, tuberculosis and other Th1-mediated granulomas is more associated with macrophages than extracellular matrix [27]. Using the B-glucan model, Morimoto et al. (2004) demonstrated that OPN-/- mice have a reduction in granuloma size and number and a 2-fold decrease in macrophage accumulation [28]. Overexpression of OPN increased granuloma formation and delayed its resolution, promoting an exacerbated fibrotic response [28]. Similar findings were observed by O’Regan and coworkers (2008) in S. mansoni egg-induced lung granulomas, a typical Th2-mediated granuloma [20]. Our results confirm the pivotal role of OPN in the Th1 and Th2 mediated granulomas and demonstrate that pathogen antigens directly induce OPN production by macrophages.
Acute schistosomiasis is a systemic hypersensitivity reaction against S. mansoni and it is characterized by miliary distribution of hyperergic necrotic-exudative granulomas [2]. The live miracidia inside the egg secrete a series of antigens and lytic substances that can trigger OPN production, recruiting inflammatory cells and inducing the granulomatous reaction to prevent further liver damage (Th1 over Th2 response) [2, 3, 10, 18]. As disease progress (Th2 over Th1 response), the granulomas are modulated (decrease in IFN-gamma and increase in IL10), the antigens and lytic substances are sequestered, necrosis is no longer observed and OPN is down regulated [2, 3, 10, 12]. Patients that will develop hepatosplenic schistosomiasis continue to produce OPN, especially by the ductular reaction, promoting fibrosis and portal hypertension [18].
The plasma levels in acute schistosomiasis were even higher than observed in hepatosplenic patients. Although OPN was demonstrated to be stable in both serum and plasma, OPN levels in the serum are 3.8–4.8 times lower than in plasma [29]. The authors speculate that this phenomenon may reflect OPN sequestration by the clot or its cleavage by thrombin, leading to loss of immunoreactivity [29]. In our cohort of acute patients only a small number of individuals had both plasma and serum samples collected and we also observed a 4–4.5 times reduction of OPN levels in serum compared to plasma (S1 Table). Ideally, future studies should use plasma samples in order to measure the total amount of circulating osteopontin.
In conclusion, S. mansoni egg antigens induce the production of OPN by macrophages in the necrotic-exudative granulomas characteristic of acute schistosomiasis mansoni. Circulating OPN levels are upregulated in human and murine acute schistosomiasis and could be a non-invasive biomarker of this form of disease.
|
10.1371/journal.pcbi.1000931 | Role of Hsp70 ATPase Domain Intrinsic Dynamics and Sequence Evolution in Enabling its Functional Interactions with NEFs | Catalysis of ADP-ATP exchange by nucleotide exchange factors (NEFs) is central to the activity of Hsp70 molecular chaperones. Yet, the mechanism of interaction of this family of chaperones with NEFs is not well understood in the context of the sequence evolution and structural dynamics of Hsp70 ATPase domains. We studied the interactions of Hsp70 ATPase domains with four different NEFs on the basis of the evolutionary trace and co-evolution of the ATPase domain sequence, combined with elastic network modeling of the collective dynamics of the complexes. Our study reveals a subtle balance between the intrinsic (to the ATPase domain) and specific (to interactions with NEFs) mechanisms shared by the four complexes. Two classes of key residues are distinguished in the Hsp70 ATPase domain: (i) highly conserved residues, involved in nucleotide binding, which mediate, via a global hinge-bending, the ATPase domain opening irrespective of NEF binding, and (ii) not-conserved but co-evolved and highly mobile residues, engaged in specific interactions with NEFs (e.g., N57, R258, R262, E283, D285). The observed interplay between these respective intrinsic (pre-existing, structure-encoded) and specific (co-evolved, sequence-dependent) interactions provides us with insights into the allosteric dynamics and functional evolution of the modular Hsp70 ATPase domain.
| The heat shock protein 70 (Hsp70) serves as a housekeeper in the cell, assisting in the correct folding, trafficking, and degradation of many proteins. The ATPase domain is the control unit of this molecular machine and its efficient functioning requires interactions with co-chaperones, including, in particular, the nucleotide exchange factors (NEFs). We examined the molecular motions of the ATPase domain in both NEF-bound and -unbound forms. We found that the NEF-binding surface enjoys large global movements prior to NEF binding, which presumably facilitates NEF recognition and binding. NEF binding stabilizes the ATPase domain in an open form and thereby facilitates the nucleotide exchange step of the chaperone cycle. A series of highly correlated amino acids were distinguished at the NEF-binding sites of the Hsp70 ATPase domain, which highlights the adaptability of the ATPase domain, both structurally and sequentially, to recognize NEFs. In contrast, the nucleotide-binding residues are tightly held near a global hinge center and are highly conserved. The contrasting properties of these two groups of residues point to an evolutionarily optimized balance between conserved/constrained and co-evolved/mobile amino acids, which enables the functional interactions of the modular Hps70 ATPase domains with NEFs.
| Many proteins are molecular machines. They function because their three-dimensional structure allows them to undergo cooperative changes in conformation that maintain the native fold while enabling their biological functions. The changes have been pointed out to be structure-encoded, intrinsically accessible to proteins, as can be deduced from simple physics-based approaches [1]. Yet, amino acid specificity is another important property that selectively mediates the interactions with specific partners and ligands [2]. Overall, a subtle balance exists between structure-encoded mechanical properties and sequence-encoded specific properties, and this balance must be evolutionarily optimized to achieve precise functioning.
The interplay between these two effects becomes particularly important in the case of a number of proteins or domains that play a modular role in a variety of biomolecular interactions. The ATPase domain (also called nucleotide-binding domain) of the Hsp70 family of proteins is a typical example. This domain plays a critical role in regulating the activities of these molecular chaperones, which, in turn, promote accurate folding, and prevent unwanted aggregation by either unfolding and refolding misfolded proteins or regulating their intracellular trafficking to the protein degradation machinery [3]–[5].
Chaperones of the Hsp70 family contain two domains: the N-terminal ATPase domain and the C-terminal substrate-binding domain (SBD), which regulate each other's activity via allosteric effects. ATP hydrolysis at the ATPase domain increases the substrate-binding affinity of the SBD, thus lowering the substrate exchange rate; on the other hand, the dissociation of the ADP produced upon ATP hydrolysis and its replacement by a new ATP trigger the release of substrate by the SBD, and therefore enhance the substrate exchange rate [3]. Regulation of substrate-binding affinity by the ATPase domain forms the basis of the chaperone activity of Hsp70s [6], [7].
The precise functioning of the Hsp70 ATPase domain involves an interaction with two families of co-factors, also called co-chaperones: the J-domain proteins that catalyze ATP hydrolysis [8], and the nucleotide exchange factors (NEFs) that assist in the replacement of ADP with ATP, by significantly increasing the ADP dissociation rate [9]. A molecular understanding of Hsp70 function requires a systemic analysis of the structural basis and mechanism of interaction with these co-chaperones. Here we focus on the interaction of their ATPase domain with NEFs.
The Hsp70 ATPase domain is composed of four subdomains: IA and IB in lobe I, and, IIA and IIB in lobe II (Figure 1a). ATP binds the central cleft between the two lobes at the interface between subdomains IIA and IIB such that the geometric and energetic effects of its binding and hydrolysis are efficiently transmitted throughout the ATPase domain.
To date, four classes of NEFs have been identified: GrpE in prokaryotes [10], and BAG-1 [11], HspBP1 [12] and Hsp110 [13] in eukaryotes. Their diverse three-dimensional structures (Figure 1b–e) exhibit a variety of binding geometries and interfacial interactions with the Hsp70 ATPase domain. In the present study, we examine these interactions, using sequence-, structure- and dynamics-based computations and identify their shared features. Our analysis provides insights into the generic and specific aspects of ATPase domain-NEF interactions, as well as the molecular machinery and sequence design principles of this highly versatile module, the Hsp70 ATPase domain, thus reconciling robust structure-encoded cooperative dynamics properties and highly correlated amino acid changes that enable specific recognition.
We began with 4839 sequences retrieved from the Pfam database 22.0 [14] for the Hsp70 family of molecular chaperones (Pfam id: PF00012). We refined the generated multiple sequence alignment (MSA) by using the consensus sequence of the ATPase domain (380 residues) in the bovine cytosolic homolog of Hsp70 [15]. The refinement consists of three steps: (i) iterative implementation of Smith-Waterman algorithm (SW) for pairwise alignment [16] using our consensus sequence, and elimination of those sequences below a threshold SW score (or less than 40% sequence identity; see details in Supplementary Material (SM) Figure S1 and Text S1) to retrieve the closest orthologs to human (Hsc70) and bacterial (DnaK) chaperones; (ii) deletion of MSA columns that correspond to insertions with respect to the consensus sequence, and (iii) removal of the sequences containing more than 10 gaps. These three steps resulted in a MSA of 1627 sequences with N = 380 columns (corresponding to residues 6 to 385 in Hsc70 ATPase domain), which has been subjected to evolutionary trace (ET) and mutual information (MI) analyses for detecting residue conservation and co-evolution patterns, respectively.
We retrieved from the Protein Data Bank (PDB) [17] structural data for HSP70 ATPase domains complexed with GrpE (PDB id: 1DKG [10]), BAG-1 (PDB id: 1HX1 [18]), HspBP1 (PDB id: 1XQS [19]), and Sse1 (Hsp110, PDB id: 3D2E [20]), shown in Figure 1b–e. Additionally, the structure of the above mentioned bovine Hsc70 ATPase domain resolved at 1.7 Å resolution (PDB id: 1HPM [15]) was used for the unbound form, and the PDB structure 1S3X [21] of the human Hsp70 served as a template to reconstruct the lobe I missing in the complex with HspBP1 using the method described in the SM Figure S2 and Text S2.
We performed Gaussian Network Model (GNM) [22], [23] and anisotropic network model (ANM) [24], [25] analyses for elucidating the equilibrium dynamics of Hsp70 ATPase domain both in the unbound form and in the complexes with different NEFs, including the reconstructed complex with HspBP1. Details on the methods can be found in our previous work [22]–[26]. Mainly, knowledge of the distribution of inter-residue contacts in the native structure permits us to construct the Kirchhoff (GNM) and Hessian (ANM) matrices, which, upon eigenvalue decomposition, yield information on the collective modes spectra.
We focused on the low frequency modes, also called global modes, as the major determinant of functional movements. In the GNM, each mode k is represented by an N-dimensional eigenvector, u(k), and eigenvalue λk, describing the mode shape and frequency (squared), respectively. The ith element, [u(k)]i, of u(k) describes the displacement of residue i along the kth mode axis; the plot of [u(k)]i2 as a function of residue index i defines the mobility profile Mi(k) in mode k. See for example, the mobility profile Mi(1) for the first (lowest frequency) mode accessible to Hsp70 ATPase domain in Figure 2a. By definition, eigenvectors are normalized, i.e., the mobility profile also represents the normalized distribution of square displacements in mode k. The reciprocal λk−1 serves as the weight of mode k, such that the slower modes, also called softer modes, make larger contributions to observed dynamics. The mobility of residue i driven by a subset of m soft modes is found from the weighted average (Figure 2b)(1)
The modes predicted by the ANM for the ATPase domain both in NEF-bound and –free forms were compared to the experimentally measured changes in structure (designated by a 3N-dimensional deformation vector d) using two metrics: (i) the correlation cosine (|v(k)·d|/|d|) between the kth ANM eigenvector v(k) (k = 1,…, 3N-6) and d, and (ii) the cumulative overlap achieved by the m softest modes [26],(2)The deformation d, is obtained by superposing the known NEF-bound and -free structures of ATPase domain and evaluating the differences in the Cα-coordinates. Kabsch's algorithm [27] is used for optimal superposition that eliminates rigid-body translations and rotations.
The ET method [28] identifies conserved residues using the MSA-derived phylogenetic tree for a given family. The application of the procedure to the Hsp70 family of chaperones is outlined in Figure 3, and details can be found in previous work [28], [29]. In summary, the method consists of three steps: (i) the phylogenetic tree is partitioned into multiple levels [30] as indicated by the vertical bars in Figure 3a; (ii) at each level, sequences are grouped into classes, each being characterized by a “class consensus sequence”. The consensus sequences are cross-examined to identify fully conserved (across classes) and class-specific or trace residues (conserved within classes but not across classes). The ET sequence for the particular level lists the fully conserved residues by their single-letter code, the trace residues by the symbol ‘X’, and the remaining residues as blank; and (iii) The ET sequences generated at each level are organized in rows (Figure 3b). An ET rank (leftmost column) is assigned to each residue. A fully conserved residue is assigned the highest rank (rank of 1). In the present case, Gly201 is the only residue with ET rank 1, i.e., it is fully conserved among the set of 1627 sequences (see SM Figure S3 for a larger version of this panel).
The conservation of a given residue in all subfamilies is a very strict condition when large sets of aligned sequences are considered. This limitation restricted the previous applications of the ET method to MSA of 100 and 200 sequences [31]. To adapt the ET method and its variations [32]–[34] to our dataset of >1,600 sequences, we relaxed the condition for defining an ET residue from conservation across “all” members in a given level to “90%”of members, and we allowed for gaps [34].
The ET method identifies conserved residues, but does not provide information on co-evolutionary relations between residues. Co-evolving residues are usually indicative of structural or functional constraints [35]–[38]. We adopted the MI content as a measure of the degree of intra-molecular co-evolution between residues in the Hsp70 ATPase domain [37], [39]–[41]. In this method, each of the N columns of the MSA is considered as a discrete random variable that takes on one of the 20 amino-acid types, or an insertion (gap, as the 21st type), with some probability. The MI associated with the ith and jth sequence positions is defined as an N×N matrix (for a MSA of N columns) of the form(3)where P(xi, yj) is the joint probability of observing amino acid types x and y at the respective sequence positions i, and j; P(xi) is the marginal/singlet probability of amino acid of type x at the ith position. I(i, j) varies in the range [0, Imax], where the lower and upper limits correspond to fully uncorrelated and most correlated pairs of residues.
Here is a brief summary of the approach and rationale. First, we examine the structural properties of known Hsp70 ATPase domain-NEF complexes from different organisms to identify the interfacial residues. Second, we analyze the intrinsic (structure-encoded) dynamics of the ATPase domain using the GNM, with an eye on the dynamic characteristics of the NEF-binding residues, on the one hand, and ATP/ADP-binding residues, on the other. A clear difference emerges between these two groups of functional residues: the former is distinguished by enhanced mobility in the softest modes while the latter is severely restricted. Third, calculations repeated with NEF-bound ATPase domains reveal how the open form of the ATPase domain is stabilized in order to facilitate ADP release, which is enabled by the intrinsic mobility of the NEF-binding regions. Nucleotide-binding sites, on the other hand, are shown to maintain their generic structure and dynamics irrespective of NEF binding, pointing to the robustness of the ATP-regulation by the ATPase domain. Fourth, detailed sequence analysis of Hsp70 family members reveals the distinctive sequence properties of the two regions: NEF-binding sites exhibit highly correlated mutations, consistent with the recognition of specific NEFs. Nucleotide-binding sites on the other hand, are almost fully conserved. In a sense, sequence variability is accompanied by conformation variability and vice versa.
Overall, Hsp70 ATPase domains appear to have been evolutionarily optimized to acquire a dual character: functional variability accompanied by structural variability at the co-chaperone binding sites and conservation/robustness both in terms of sequence and structural dynamics at the nucleotide-binding sites. This dual character is proposed to be essential for adapting to interactions with different co-factors while maintaining ATPase activity.
Organisms comply with the evolutionary pressure to maintain their phenotype by genotypic variations that are compensated or correlated as needed, conserving certain sequence fragments vital to preserving their functions [57]. Understanding the co-evolving and conserved sequence patterns in modular domains is an interesting problem in its own right [58], [59]. Understanding these patterns in the light of structural data, if available, provides us with further insights into shared mechanisms of interactions that form the molecular basis of the biological function of such modular domains. The Hsp70 ATPase domain is such a modular protein common to functionally diverse actin, hexokinase, and Hsp70 protein families [60]. The present combined analysis of structure-encoded dynamics and sequence evolution for Hsp70 ATPase domain discloses a subtle interplay between conserved interactions and those involving co-evolved residues. Conserved interactions define generic properties of the Hsp70 ATPase domain: these include the concerted dynamics of its four subdomains, which allow for sampling functional conformations (e.g., that stabilized upon NEF binding, allowing for ADP release; shown in Figure 3), and the physicochemical events (ATP hydrolysis) at the nucleotide-binding site. Those residues involved in NEF recognition, on the other hand, show low-to-moderate conservation, but exhibit a remarkably high tendency to co-evolve, or undergo correlated mutations, again to achieve specific NEF-dependent recognition and binding activities. Interestingly, NEF residues that interact with the Hsp70 ATPase domain appear to be rather conserved (Figure S11) to maintain this specificity.
An observation of interest is the similarity between the interactions of the Hsp70 ATPase domain with different NEFs, in terms of structural dynamics. While Hsp70 ATPase domains are highly conserved both sequentially and structurally, the four NEFs examined have distinct structures and consequently different dynamics. The key point is that their binding to the ATPase domain involves in all cases the subdomain IIB of the ATPase domain, although not in exactly the same arrangement. Their binding to a common interfacial region on the ATPase domain point to a shared mechanism of interaction: The ATPase subdomain IIB is originally distinguished by its high mobility in the slowest mode, especially at the β-sheet E and the exposed loop connecting the two strands of this sheet; and after NEF binding, there is a significant suppression in its mobility. The conserved dynamics of the complexes suggests a role of subdomain IIB as an “adjustable handle”, which regulates the Hsp70 chaperone machine, to facilitate other proteins making use of its SBD.
Many applications using the ANM have shown that the substrate recognition involves a region distinguished by its enhanced mobility in the most cooperative (or softest) modes, which enables the molecule to optimize its interactions with the substrate. Here we can see that the C-terminal part of helix 8 and the loop of β-hairpin E enjoy this type of high mobility/adaptability. On the other hand, substrate ‘binding’ may also involve more constrained residues in the close neighborhood, which may play a role in transmitting allosteric effects. In the opposite case of a binding site composed exclusively of floppy residues, the structural changes induced upon substrate binding could dissipate locally and not efficiently transmitted. In this respect, we propose that the involvement of residues such as Arg258, Arg261 and Arg262 in subdomain IIB, or N57, A60 and M61 in subdomains IB is critically important in establishing the communication between subdomains and transmitting allosteric signals between NEF-binding and nucleotide binding sites.
A putative communication pathway that couples distant residues in different subdomains of the Hsp70 ATPase domain is suggested here by the structural mapping of correlated and conserved residues, which needs to be further established. Figure 6a displays those residues identified to be co-evolving. Notably, we observe several pairs making interdomain contacts, in addition to spatially distant residue pairs (e.g. H23 in subdomain IA and N57, A60 and M61 in subdomain IB correlated with R258, R261, E283 and D292 in subdomain IIB). In a recent study, R272, R261, Y15 and Y41 have been identified to play a central role in establishing the allosteric communication in the unbound Hsp70 ATPase domain, along with highly conserved residues K71, R72, E175 and H227 [45]. It remains to be seen if these central residues play a key role in mediating between these co-evolving, spatially distant residues. We also note that Smock et al. recently identified a sparse but structurally contiguous group of co-evolving residues at the interface between the ATPase domain and the SBD in Hsp70/110 protein family, which has been proposed to underlie the inter-domain allosteric coupling [61], in support of the role of co-evolved residues in mediating allosteric signaling.
Many recent studies have pointed out the validity of “pre-existing equilibrium” concept where a substrate or ligand simply selects from amongst an ensemble of conformations already accessible to the protein prior to binding [49], [62]–[67]. The present results, and recent applications of ENMs, suggest that more important than the pre-existence of these ‘states’, is the existence of energetically accessible ‘paths’ that provide access to those states, or the intrinsic tendency of the native structure to reconfigure towards such functional states. In terms of energy landscape description, what is needed is not the existence of multiple minima, the depths of which change upon ligand or substrate binding, but the existence of one or more directions of reconfigurations, or paths along the energy landscape, that are easily accessible to the protein and lead to the targeted (functional) conformer. The softest modes provide such paths. They define directions of motion in the space of collective coordinates, which incur a minimal energy ascent as the molecule moves away from its original energy minimum. They also present the best mechanisms of dissipating energy, if the system is perturbed. These are the modes that are being exploited when proteins bind ligands or substrates. Notably these functional conformations accessible near the native state can be observed by NMR residual dipolar coupling, as shown for Hsp70 ATPase domain by Zuiderweg and coworkers [52]. Figure 4 clearly shows that movements along a handful of modes satisfactorily ensures the passage to the alternative (functional) open form, and that the open form itself has a strong tendency to restore its conformation back to the closed form, in the absence of NEF.
Protein-ligand binding interfaces and protein-protein contact interfaces are characterized by different sequence variation patterns. The protein-protein contact interfaces usually expose larger contact areas [68] and exhibit high mutation rates. Moreover, if the contact interface is a common recognition site for multiple targets (possibly in different organisms), co-evolution is likely to occur among the binding residues to preserve specific interactions and conformations at the sequence motif. On the other hand, the protein-ligand interface is usually buried in the folded core of the protein; in contrast to protein-protein interaction, the protein-ligand interaction is usually characterized by higher specificity, requiring sequence conservation [28], [29].
The Hsp70 ATPase domain exhibits patterns in close agreement with these general features: Its ligand (nucleotide) binding site essentially consists of highly conserved residues, which not only precisely coordinate the ligand, but also take part in a global hinge-bending region so that they are both chemically and mechanically required to be highly conserved. NEF recognition sites, on the other hand, exhibit much lower conservation properties; and in addition to their sequence variability, the subdomain IIB, which is observed to be most often involved in NEF binding, enjoys enhanced mobility. Briefly, global dynamics requirements entail residue conservation, and specific recognition entails sequence variation along with enhanced mobility. However, neither the sequence variability, nor the conformational mobility at NEF recognition sites, is random. The sequence variability takes place under unique restrictions, compensating mutations, as unraveled by the MI map. Conformational variability, on the other hand, is uniquely defined by the ATPase architecture, and precisely adept to accommodate the passage to the functional open state that is stabilized upon NEF binding. The ATPase domain uniquely juxtaposes such structure-encoded dynamics and sequence-specific interactions, which underlie its ubiquitous activities.
In general, subdomains IA and IIA are more conserved and more rigid than subdomains IB and IIB [69], as also indicated by the ET in Figure 3b; notably, they also serve as binding site to a number of proteins. For example, subdomain IA accounts for the binding of J-domain proteins [70]; subdomain IIA is reported to contain a putative binding site near its interface with subdomain IA (V189-V195) to the chaperonin-containing TCP-1 [71], and it is connected to the SBD by an inter-domain linker, which is considered important for the allosteric interactions between the two domains [72], [73]. It remains to be seen if the correlated sites on Hsp70 ATPase domain emerging from the MI analysis play a role in the functional communication with other co-chaperones or the SBD. Extensive experimental studies have been performed to date with the E. coli Hsp70, DnaK, to understand the molecular mechanism of activity of the molecular chaperones in the Hsp70 family. The analysis in the present paper will guide our interpretation of the NMR, FRET, and EPR data on different states accessible to DnaK. Each of these methods gives us a different window into the ensemble of conformational states populated in response to ATP, ADP and NEFs. Excitingly, a detailed chemical shift analysis of six different ligand bound states for the nucleotide-binding domain of DnaK, with and without the linker that connects it to the substrate-binding domain (i.e., 12 NMR samples compared pairwise and as a group) has pointed to the same subdomain interface rearrangements indicated in the present study (Zhuravleva & Gierasch, in preparation). Moreover, the NMR results point to the fundamental feature that subdomain IIB can undergo a hinge-like movement to enable nucleotide entry and release. It is this fundamental movement, intrinsic to Hsp70 ATPase domains, that different NEFs have exploited. They bind in different, sequence-specific ways, but modulate the same fundamental movement. Further detailed analysis of the ensemble distributions and rates of interconversion between states can be achieved using a synergistic battery of computational and experimental tools.
|
10.1371/journal.pcbi.1000190 | Organization of Excitable Dynamics in Hierarchical Biological Networks | This study investigates the contributions of network topology features to the dynamic behavior of hierarchically organized excitable networks. Representatives of different types of hierarchical networks as well as two biological neural networks are explored with a three-state model of node activation for systematically varying levels of random background network stimulation. The results demonstrate that two principal topological aspects of hierarchical networks, node centrality and network modularity, correlate with the network activity patterns at different levels of spontaneous network activation. The approach also shows that the dynamic behavior of the cerebral cortical systems network in the cat is dominated by the network's modular organization, while the activation behavior of the cellular neuronal network of Caenorhabditis elegans is strongly influenced by hub nodes. These findings indicate the interaction of multiple topological features and dynamic states in the function of complex biological networks.
| Many complex biological networks are characterized by the coexistence of topological features such as modules and central hub nodes. What are the relative contributions of these structural features to the networks' dynamic behavior? We used a computational model to simulate the general activation and inactivation behavior of excitable nodes in neural networks and studied the spread of activity in hierarchically organized networks as well as specific biological neural networks. We then evaluated the impact of modules and hub nodes on the network dynamics by correlating the patterns of node activity with the network architecture at difference levels of spontaneous network activation. Two dynamic regimes were observed: waves propagating from central nodes and module-based synchronization. Remarkably, the dynamic behavior of hierarchical modular networks switched between these modes as the level of spontaneous activation changed. We also found that the two dynamic regimes have different significance in the neuronal network of C. elegans, where activity is mainly organized by hub nodes, and the systems network of the cat cerebral cortex, which is dominated by the network's modular organization. Our approach can be used to dynamically explore the organization of complex neural networks, beyond the structural characterizations that were available previously.
| The analysis of biological networks presents an intriguing challenge, due to the complex, non-random organization of these systems and the diverse dynamic behaviors that they express. The topology of several biological networks has been shown to be based on a scale-free degree distribution, which implies the existence of highly connected network hubs [1],[2]. Biological systems were also found to be organized in network modules [3],[4], or to contain characteristic circuits (motifs) that do not occur as frequently in other types of networks [5]. Hub nodes, which have been identified in several biological networks, such as protein-protein interaction networks or metabolic networks, may serve as central distributing elements or linkage point for many regions of a network [2],[6],[7]. Such hubs might also be present in neural systems networks [8]. A hub, for our purposes, can either be a node with a high degree or with a high centrality (i.e. with many shortest paths between nodes passing through). For our purposes, the latter definition is dynamically more relevant. Modules or network clusters, which are characterized by a higher frequency or density of connections within than between node clusters [9] have been identified in biological metabolic networks [10],[11], as well as neural networks at the cellular level [12] or the systems level [13]. These modules often represent a specific function, e.g. a specific synthesis pathway in a metabolic reaction network [14], and may shape the functional interactions within the networks at different scales [15]–[17]. It has also been argued that motifs may represent specific functional circuits [18]–[20].
In addition to the mentioned features, the organization of biological systems is often described as hierarchical. However, no formal definition of hierarchical topology appears to exist. Typical descriptions of hierarchical organization use a modules-within-modules view [10],[21], others focus on the coexistence of modules and central (hub) nodes [11],[22] or relate the concept of hierarchy to fractality [23]. The distinction between hubs which organize modules around them and hubs which connect modules on a higher topological level has been productive for understanding the functional roles of these hub categories in various empirical networks [11],[22],[24]. Note: (1) In [10] the algorithm for generating modules within modules, leading to a hierarchical network, also produces a hierarchy of hubs in the network; (2) it is not immediately clear, whether the fractal graphs discussed in [23] are also “fractal” from the perspective of the box-counting formalism developed in [25],[26]. Particularly the latter concept of fractality has interesting implications for the organization of dynamic processes on the graph [27].
In the present paper, we attempt to summarize current topological concepts, condense the spectrum of different network arrangements into a few salient topological features and, using a simple three-state model of excitable dynamics on graphs, study how these topological features organize dynamic behavior. While this approach and our findings are valid for a wide range of networks, we investigate the question and the implications of our findings particularly in the context of neural networks, which most clearly express diverse patterns of excitable dynamics.
From a combination of modular and hub features, various types of network topologies can arise. Classical Erdös-Rényi (ER) random graphs do not contain hubs or modules and may thus serve as a general null model. Scale-free Barabási-Albert (BA) graphs, on the other hand, contain only hubs and no modules. Within such graphs, projections from the hubs can reach many network regions, and the hub nodes thus have a more privileged role than nodes with fewer connections and a more restricted reach. On the other hand, networks that do not contain hubs, but are modular, may arise from linking many distributed, dense clusters with a small number of inter-cluster connections. Such clusters could exist at different levels (representing clusters of sub-clusters of sub-sub-clusters [21]), resulting in a hierarchical network organization, which has recently been termed “fractal” [23]. Finally, networks may be modular and also contain hubs, which are either contained within the modules serving as local hubs, or may form global hubs that integrated network modules at different scales of organization [10],[14],[24],[28]. The two latter networks combine features of scale-free and modular networks. Figure 1 summarizes the topology of modular and hub features and their combination in complex networks. While all feature combinations provide networks of complex organization, we are particular interested in the hierarchical networks shown in the last row of Figure 1, which form modular arrangements, with or without hubs, at different network scales.
For discussing the link between network topology and dynamics we use a simple three-state model of an excitable medium. The model consists of three discrete states for each node (susceptible S, excited E, refractory R), which are updated synchronously in discrete time steps according to the following rules: (1) A susceptible node becomes an excited node, if there is at least one excitation in its direct neighborhood. If not, spontaneous firing occurs with the probability f, which is the rate of spontaneous excitation; (2) an excited node enters the refractory state; (3) a node regenerates (R→S) with the recovery probability p (the inverse of which is the average refractory time of a node). This minimal model of an excitable system has a rich history in biological modeling. It has been first introduced in a simpler variant under the name “forest fire model” [29] and subsequently expanded by Drossel and Schwabl [30] who also introduced the rate of spontaneous excitations (the “lightning probability” in their terminology). In this form it was originally applied on regular architectures in studies of self-organized criticality. Other variants of three-state excitable dynamics have been used to describe epidemic spreading [31]–[34]. As discussed previously [35],[36], this general model can readily be implemented on arbitrary network architectures. It has been shown that short-cuts inserted into a regular (e.g., ring-like) architecture can mimic the dynamic effect of spontaneous excitations [35]. Using a similar model setup we have recently shown [36] that the distribution pattern of excitations is regulated by the connectivity as well as by the rate of spontaneous excitations. An increase in each of these two quantities leads to a sudden increase in the excitation density accompanied by a drastic change in the distribution pattern from a collective, synchronous firing of a large number of nodes in the graph (spikes) to more local, long-lasting and propagating excitation patterns (bursts). Further studies on the activity of integrate-and-fire neurons in the classical small-world model from [37] also revealed a distinct dependency of the dynamic behavior on the connectivity of the system [38].
Here, we take this investigation one step further by analyzing which topological properties determine the distribution patterns of excitations. In order to study these patterns, we consider the individual time series of all nodes and for each pair of nodes (s,t) compute the number C = Cst of simultaneous firing events. When applied to the whole network the resulting matrix C essentially represents the distribution pattern of excitations which we now can compare with a corresponding distribution pattern of some topological property.
Examining hub and modular aspects of topology separately we first investigate which of them explains best the observed pattern of simultaneous firing events. In particular, we show that in different parameter regimes (characterized by the rate of spontaneous excitations) different topological properties determine the observed synchronization patterns. Moreover, we show that small systematic changes in the graph architecture, designed to enhance or decrease the selected topological property, are reflected in the dynamics. In a second step, we extend our study to hierarchically structured artificial graphs and then to biological networks, in order to demonstrate that the distribution patterns of excitations change dramatically when both properties are represented to different degrees in the respective graphs. Finally, we summarize our results and discuss limitations of the present approach, and extend our observations to describe general principles of pattern formation on graphs.
In this study, we focus on two structural properties of networks and use them in terms of topological references. These properties are modularity and node centrality and they are represented by the topological modularity (TM) reference and the central-node based (CN) reference, respectively. To highlight the individual impact of each topological property on dynamic pattern formation we first probe different types of artificial networks dynamically and compare the results with the respective topological reference. We then validate our results with modified versions of these networks (see Figure S1 and Figure S2 in Text S1: Analysis of randomized network topologies) and with different types of hierarchically structured graphs, which represent the two topological properties to different extents. We finally transfer our analysis to more densely connected networks and to different hierarchically structured real-world topologies (see Methods for details on the construction of the respective references, the dynamic models and the different types of graph architectures and graph randomization processes). Figures 2 and 3 summarize our strategy of comparing the pattern of simultaneous excitations (correlation matrix C) with the corresponding topological feature, namely the topological modules (TM, Figure 2) and the central-node based reference (CN, Figure 3). Both, the graph and the simulated “space-time” pattern are converted into matrices giving the pairwise distances and the number of simultaneous excitations, respectively. The two matrices are processed further to yield the respective clustering trees, which then are sorted, color-coded and systematically compared (see Methods for a detailed description of this procedure.)
We start our analysis with the modular scale-free network in order to test the explanatory power of the TM reference. As a first step we visualize for a single value of f how well the dynamically detected clusters follow the topological modules. We can map the clustering tree obtained from the correlation matrix onto the graph by thresholding it to yield the same number of modules μ as detected topologically and assign colors as labels to the modules. Figure 4 displays the corresponding graph with the modules colored exclusively on the basis of the dynamically detected clusters (DDCs), resulting from a simulation with f = 0.01. In this case, the dynamic clusters have a large overlap with the modules found topologically.
As a next step, we analyze the whole range of the parameter f. This is summarized in Figure 5. The color bar on the left-hand side represents the color-coded TM reference. The sequence of color bars from left to right are the color-coded DDC vectors for increasing values of f. There are three distinct ranges in f characterized by different patterns of the DDC vectors. Above a value of f = 0.1 any regularity is replaced by a random distribution of colors. Here, the random excitation events dominate the dynamics, thus leading to uncorrelated excitations and to a formation of unsystematic dynamically detected clusters. For lower values of f two different forms of node integration into dynamic clusters can be discriminated. Up to a value of f = 10−3 the DDC vectors are a mixture of homogeneous regions (representing well detected topological modules) on the one hand (in the bottom part of each DDC vector in this f range) and regions with smaller scale homogeneities on the other (top part of the DDC vectors). In this range the topological modules coincide partly with the dynamic clusters, but the dynamic integration fails to comply with the topological hierarchy of the modules. The middle range in f = 0.01 (10−3<f<0.1) is characterized by a very high order of the DDC vectors and an almost perfect agreement with the TM reference. Besides this systematic dynamic retrieval of the topological modules the DDC vectors in this f -range are also characterized by a strong consistency with the hierarchy of the modules on the level of the whole graph. The separation of the DDC vectors into two regimes with respect to f (omitting here the noise-driven high f-regime) is basically driven by the three-state model's behavior under spontaneous excitations. As pointed out in our previous work [36], the model displays a transition in the distribution patterns of excitations from a global (spike) to a more local (burst) regime with an increasing rate of spontaneous excitations f. While a spike (low-f regime) is able to reach most of the system (depending on the excess of nodes in the excitable state S), the burst (higher-f regime) is characterized by one or more excitation spots which propagate through the system on a localized level due to a more balanced distribution of the states S and R (Video S1 in Text S1 illustrates the propagation of excitations on a modular graph architecture during the burst regime). Consequently, the DDC vectors separate rather precisely at the position where the burst dynamics outbalances the spike dynamics. In this sense the burst dynamics provides a suitable tool for the dynamic retrieval of topological modules.
The results for f<10−3 suggest that another form of dynamic integration of nodes takes place beyond the module level. Groups of nodes which belong to different topological modules (see e.g. the blue and red labels in Figure 5) are placed in close dynamic proximity (that is, they are integrated into the same dynamic cluster). For testing this new principle of dynamic integration we repeat this simulation with a non-modular scale-free BA graph (see Methods) and the CN reference discussed in Figure 3. In Figure 6 the BA graph representation has been color-coded according to the dynamically detected clusters (with a preset value of 7 clusters, which determines the threshold applied to the corresponding clustering tree) at f = 10−5. One observes a rather clear ring-like arrangement of colors around a central node which is one of the hubs in the graph. This distribution of the dynamic clusters around a central node h (displayed in black) confirms our hypothesis that another topological feature is shaping the distribution of excitations in this low-f regime.
Studying the agreement between the CN reference and the DDC vectors for the BA graph over a whole range in f leads to the result shown in Figure 7. The CN reference (left-hand side) clusters all nodes t according to their distances d to the central node h with d = Lht. Up to a value of f = 10−3 all equidistant nodes assemble more or less in the same dynamic cluster and even the distance order is maintained (except for d = 1 and d = 2). Above f = 10−3 the homogeneity of the DDC vector drops rapidly finally reaching a random composition. Again this decrease of dynamic order is accompanied by a decrease of the spike regimes in the overall dynamics. The recurrent simultaneous excitations which lead to the observed pattern are caused by global properties of the graph's topology. We assume that such networks are able to channel the excitations produced by random events into their centers, which are composed of one or a few nodes displaying the highest betweenness centrality (as given by the number of shortest paths leading to the node; see Methods). From there, the excitation waves pass through the rest of the system reaching all equidistant nodes (seen from the center) at about the same time and thus integrate them dynamically. The dynamics in Video S2 in Text S1 contains several spike events which demonstrate the typical propagation of excitations in a BA graph. In addition Figure S5 illustrates the consistency between the sequential arrangement of ring-shaped modules (as seen from the central node) and the chronology of excitations showing the fraction of simultaneous excitations within each of these modules at the same time.
An integration of both topological properties (modularity and hub dominance) into one system has been accomplished via the introduction of the hierarchical scale-free graph [10],[28]. We expect from the previous discussion that both levels of dynamic organization are present in such a network. As other network designs exhibit hierarchical properties as well, we contrasted different types of hierarchical graphs, also considering densely connected graph structures which, for instance, characterize many neuronal systems. To allow for the analysis of highly connected networks we extended our dynamic model with the additional node degree-dependent parameter ω (which regulates the excitability of a node, i.e. the number of excitations needed in order to trigger a firing event; see Methods for the exact definition of ω).
All hierarchical networks introduced here share a hierarchical fashion of linking the modules, but some of them lack the hubs and the scale-free degree distribution. One would expect that such graphs are not able to produce consistent ring-like excitation patterns as observed in the BA graph. In the following we will investigate how these topological properties determine the distribution pattern of excitations. We checked, however, that this general phenomenon does not depend on the exact method of generating a particular topological property.
We tested four different hierarchical networks, i) the hierarchical scale-free graph [10],[28], ii) a variant of the hierarchical scale-free model (which permits the construction of densely connected graphs), iii) the fractal modular network [23], and iv) the hierarchical cluster network [21]. We generated 10 networks of each graph type, simulated the dynamics, and computed Qdyn from the resulting dynamic clustering trees, as before. Densely connected networks were simulated with a threshold of κ = 0.1, as described in Methods.
In the following the results are limited to the hierarchical scale-free graph [10] and the mapped fractal graph [23] as both other results agree well with their respective counterparts. Figure 8 displays averaged over all networks as a function of f for the TM reference (blue ▵) and the CN reference (red ○). In the hierarchical graph (Figure 8A) the dynamic detection of the topological modules based on the TM reference works very well for high values of f. Increase and decline of depend on the transition from spike dynamics to burst dynamics and on the increasing noise intensity f, respectively (Figure S3 displays the corresponding time series of the excitation density ρF for three different values of f). This increase is accompanied by decreasing values of for the CN-dependent results which display their maximum in the low f-regime. Here, the high values of indicate a strong dominance of the hubs and their importance for the formation of the excitation waves. Indeed, this graph structure facilitates the emergence of both forms of dynamic organization. This observation, that certain types of hierarchical graphs can host both dynamic patterns with the rate of spontaneous excitations inducing a switch from one to the other, will be discussed in detail elsewhere.
In the mapped fractal graph (Figure 8B the absence of hubs prevents the generation of ring-like excitation patterns (as seen in the low values of ) with the effect that the range of dynamically detected topological clusters () enlarges towards low values of f.
By an adjustment of the dynamic model the consistency to the more sparsely connected networks demonstrates that (i.e. by rescaling the excitability; see Methods) it is still possible to retrieve both dynamic regimes even in densely connected graph architectures, similarly to the more sparsely connected networks. Rescaling the excitability (by requiring more than one excitation in the neighborhood for exciting a node) thus provides a consistent extension of our original dynamics to higher connectivities.
Compared to metabolic reaction networks or protein-protein interaction networks, the architecture of many neuronal systems is characterized by a high density of connections [13],[39],[40]. We studied neuronal networks of two organisms at two fundamentally different levels of organization, namely the cortical systems network of the cat and the cellular neuronal network of the nematode C. elegans.
First, we analyzed the cortical network of the cat, which has a well-characterized topology [8],[13] and has been the basis of previous dynamical simulations [16],[41],[42]. We focused at connectivity at the systems level, which is more reliably established than cellular cortical connectivity. At the systems level, all the neurons of a cortical area are integrated into a single node. This coarse-graining approach scales the cortical network representation down to n = 55 nodes and 238 directed edges and 327 undirected edges which originate from 892 cortico-cortical connections.
Second, we considered the cellular neuronal connectivity of the nematode C. elegans, which has also been studied extensively. Due to the fixed number of nodes, the neuronal network of C. elegans serves as an excellent neuronal model system [43]. This version of the cellular neuronal network of C. elegans contained n = 277 nodes and 1731 directed edges and 187 undirected edges.
The connection density of the cat cortex representation is comparatively high (z = 0.3), while the connection density of the neuronal network of C. elegans is about tenfold smaller (z = 0.028). Therefore, we decided to use the modified DE model for the cat cortex with κ = 0.15 and p = 0.1 and the original DE model for C. elegans with p = 0.01. We analyzed both networks in the range of 10−6<f<1. The TM references consist of 4 modules (cat) and 8 modules (C. elegans), respectively. The four modules in the cat systems network correspond to those previously identified by other clustering approaches [13], and represent sets of visual, auditory, sensory-motor and fronto-limbic cortical areas. The diagrams (Figure 9 top) display the analysis of the dynamic modularity for both topological references. The diagrams on the bottom show corresponding curves with highlighted markers on the top. They display the TM-dependent DDC vectors for the Cat (Figure 9A bottom) and the CN-dependent DDC vectors for C. elegans (Figure 9B bottom).
Examining the relation between topology and dynamic properties independently of the organism, both networks show certain characteristics of a hierarchical scale-free network [10],[28], that is, the typical differences in the dynamic dominance of modular and hub features for different levels of spontaneous activation (as indicated in the f-dependent course of Qdyn in Figure 9 top), which implicate the existence of a complex hierarchical structure. However, both organisms also exhibit great differences in their dynamic regimes.
For low levels of spontaneous excitation in the cat cerebral network (Figure 9A top), the CN and TM references are equally well related to the network's dynamic behavior. The strong correlation between dynamics and the modular topology is reflected in a high consistency between the TM reference and DDC vectors in the high f-regime (Figure 9A bottom) also indicated in Figure 9A top in , while there seems to be only a marginal influence of hubs. If we exchange the TM reference by the modules previously identified for the cat cortical network [13], the general features of Qdyn(f) remain intact (in particular the clear peak in f; see Figure S4).
On the other hand, the dynamic behavior of the cellular network of C. elegans is for all but the highest levels of activation dominated by the distance to a central node (Figure 9B). Betweenness analysis revealed two nodes in direct neighborhood, which display the highest node degrees of the neuronal network, and which may serve as an initial point of circular excitation waves. Nodes 52 (AVAL) and 53 (AVAR) display the highest node betweenness (and the highest node degrees). The distance between both nodes is 1, as they are mirror-symmetric versions of the same neuron, AVA, on the L and R sides of the nematode's body.
The current paper presents some aspects of a pattern-based computational approach for linking network topology and dynamics. This approach proved useful in probing the functional organization of complex biological networks. The comparison of topological features and simulated network dynamics demonstrated that features such as central hub nodes and network modularity can strongly and systematically shape a network's dynamic behavior. Moreover, in hierarchical modular networks, where multiple of these features were present, the network dynamics exhibit a functional switch for different levels of spontaneous network activation between the dynamic organization through a central node or through modular features.
The method also reveals the dynamic impact of different topological characteristics in biological neural networks. In particular, the dynamics in the cellular neuronal network of C. elegans appears organized by the topological distance to a central hub node, whereas the dynamic behavior of the cat cerebral cortical network appears more strongly influenced by network modularity. Both topological features, however, contribute to the organization of the networks synchronization dynamics. Given the restricted size of the biological networks, the functional implications of the features would have been difficult to derive from a conventional analysis of the networks' degree distributions. These findings have implications for understanding the relationship of network topology and dynamics in complex neural networks, as detailed in the following sections.
The presented approach draws on a simple dynamic model for describing excitable elements. This model only represents node activation, inactivation, as well as a refractory period, with discrete time steps. Given the complex dynamic behavior of neurons and neuronal systems, the model may appear overly simplistic. However, we believe that the model captures essential features of excitable elements, such as the principal activation cycle of neurons. Moreover, at the moment it is far from clear how much detail is required to realistically describe the interaction of excitable elements in networks. A good starting point for analyzing such pattern-formation aspects also in more sophisticated models could be built upon the parallel to a recent simulation study of the cat cortical network, which uses a more sophisticated population oscillator model to describe the activity of individual cells within the cortical areas [16]. This study led to a similar finding of a modular dynamic organization that strongly followed the modular topological organization. There are also precedents for the successful application of highly simplified models of cortical networks. For example [41] used a simple spreading model to infer basic properties of the relationship between node lesions and network activity in the thalamo-cortical network of the cat. Similarly, [42] replicated epileptiform steady-state activation patterns in the cat cortical network with the help of a simple thresholded spreading model. In addition, in the present work the model parameters were varied over a wide range; however, the different simulations resulted in similar principal behavior.
When applied to biological neural networks, our approach revealed that the dynamic behavior of neural networks may be coordinated via different topological features. While activity in the neuronal network of C. elegans is shaped by excitation spreading from central hub nodes, the dynamic behavior of the cat cortical network is largely dominated by the network's modular organization. Moreover, the cortical network may switch from modular to hub dominance for low levels of spontaneous activation.
The current analysis applies to network dynamics with spontaneous node activations, as observed in tonic neural activity, but without explicit external (sensory) input. This description corresponds to the experimental case of so-called resting state connectivity, a type of functional connectivity that persists in the absence of specific external stimulation. Resting state networks have been studied intensively over the last years and have been considered as default frameworks of neural dynamics [44]. Resting state connectivity can be derived experimentally from time-series correlations between large-scale brain regions, such as cortical areas. The regions' activity is estimated from different functional imaging techniques (e.g., EEG, fMRI); and typically, the coupling occurs at very low frequencies, around or below 0.1 Hz [45]. The slow-frequency coupling may be a reflection of faster electrophysiological coupling among distributed neuronal populations [17]. Experimental resting state data are currently available for cortical networks in humans and non-human primates, but not for the cat cortical network studied here. However, the present theoretical findings largely agree with what is known from the available experimental data. For instance, resting state data for human and primate cortical networks at the systems level show a strongly modular organization [46],[47]. Earlier experimental findings, based on activity spreading after local cortical disinhibition, also suggest that primate cortical areas co-activate, in groups that closely match the known topological clusters [15]. In addition, previous theoretical studies also support the conclusion that the dynamic organization of large-scale cortical networks in the absence of external stimuli is strongly shaped by the networks' modular structural connectivity [16].
However, it was also suggested that hub-like areas exist in cortical networks which possess a relatively large number of connections and which can be identified implicitly from the networks' behavior after simulated node lesions [8],[24],[48]. The leading central nodes identified here for the cat cortical network by node betweenness, multimodal areas 35 and AES, are also among those suggested previously by degree and lesion impact [8],[24]. For low rates of spontaneous activation, the cortical dynamics became somewhat more strongly correlated to hub distance than network modules. This dynamic switch characterizes the cortical connectivity as a complex hierarchical network and indicates the possibility that particular cat cortical areas act as hub-like nodes for the organization of low-noise dynamic regimes. This point still needs to be investigated in more detail. Importantly, only coarse large-scale activations can be resolved with the current neuroimaging techniques. Nonetheless, it is clear that cortical networks have a multi-level modular organization (forming clusters of sub-clusters of excitable nodes [21], with modules spanning from cellular cortical circuits and columns to clusters of strongly interlinked areas). Therefore, it can be speculated that, once data for additional scales of cortical networks are available, switches of the dynamic behavior between different topological features become more clearly apparent.
In contrast to the cortical network the dynamic behavior of the C. elegans network was dominated by central node distance for all levels of spontaneous activation. Experimental findings also indicate that neuronal dynamics in C. elegans are coordinated by central pattern organizers [49],[50] rather than through network modules. Indeed, the pair of AVA neurons, which have the highest degree and highest node betweenness in the C. elegans network, and which therefore may be considered as network hubs, have been implicated as a component in a central pattern generator responsible for locomotion control [49]. Specifically, AVA is thought to be responsible for backward movements. The present results suggest that this node may also have a more general function in coordinating dynamic activity in the nematode nervous system.
The finding of dynamic organization through network modules in large-scale cortical networks, versus organization through few central nodes in cellular neuronal networks, makes intuitive sense. Given the small size of its nervous system, the functional specialization in C. elegans occurs at the level of individual cells, which exert their roles globally across the network. On the other hand, specialization in the mammalian cortex arises for whole brain regions (e.g., visual cortex, sensory-motor cortex) comprising several cortical areas which are closely cooperating within modules to perform the various aspects of their functional subdivision.
When studying dynamics on networks, the synchronization behavior of each single node is a suitable indicator to estimate the dynamic scope provided by a graph's topology. Different forms of synchronization require different structural properties. By the application of a simple excitable medium (the DE model) we were able to generate two distinct forms of synchronization via the regulation of a single dynamic parameter, the amount of spontaneous excitations f. This noise level f also defines the (length) scale on which a specific dynamic process will predominantly be situated. Consequently the (larger-scale) wave-like propagation (consistency with CN reference) is dominant at lower levels of f, while the local module-based synchronization (consistency with the TM reference) is situated preferentially at higher f.
Via comparison to two different topological references representing the elementary graph properties modularity and hub dominance the dynamic results were attributed to the respective synchronization behavior. In the burst range of f, networks exclusively featuring modular properties with decentralized hubs display synchronization behavior predominantly within their communities as indicated by the consistency to a module-based topological reference. If a graph is dominated by one or a few hubs in its center (a feature of the BA graph) a global (ring-like) synchronization phenomenon is visible due to the formation of excitation waves which reach the whole system from the graph's center. In contrast to our modularity definition it is more difficult to decide whether a node is the center of a graph or not. Here, we used the betweenness centrality (B) definition, but the results indicate that B does not alone account for the unifying topological quantity for different networks. The analysis of different hub categories [10],[11],[12] and their involvement in organizing the dynamics [24] is an important next step of the study described here. We did not do this so far, because it would require simulating substantially larger networks to obtain reliable results. We would also like to point out that the prototypes of pattern formation we identify, might serve as minimal models of the brain activity regimes reported by Izhikevich and Edelman in their model of mammalian thalamocortical systems, which emerge spontaneously as a result of interactions between architectural features and the dynamics [51]. An important challenge for the future will be to activate modeled neural networks more selectively with patterns representing functional inputs, and to observe the interactions of stimulus-related activity with default activity.
In summary, by using a simple dynamic model we could determine a “network equivalent” of pattern formation, where patterns are represented by correlations between topology and dynamics. Specific topological features give rise to and regulate quantitatively certain elementary forms of patterns. We believe that this correspondence is not restricted to the specific dynamics considered here. The recent findings on synchronization of phase oscillators [52],[53] show similar matches between topology and dynamics as the results reported for an excitable system. In this light a comparison of these systems in detail (our discrete excitable three-state model and the continuous phase oscillator model) would be very interesting and could point towards common links between topology and dynamics far beyond individual dynamical systems. It is particularly interesting that the authors employ phase oscillators and their synchronization properties also to determine functional groups in the neural system of C. elegans [54].
We applied the analysis approach to two sets of neural network data at different scales of organization. The first data set describes systems level connections between different areas of the cat cerebral cortex, and is based on a global collation of cat cortical connectivity (892 interconnections of 55 areas). This collation was developed from the data set described in Scannell et al. (1995) [57] and forms part of a larger database of thalamo-cortical connectivity of the cat [39]. The database was created by the interpretation of a large number of reports of tract-tracing experiments from the anatomical literature.
The second data set represents cellular neuronal connectivity of the nematode C. elegans (277 neurons and 2,105 synaptic connections). This data set was adapted from Achacoso and Yamamoto (1992) [43]. That compilation is largely based on the dataset of White et al. [58] in which connections were identified by electron microscope reconstructions. The previously presented connectivity data [43] was modified in the following way. Neurons of the pharyngeal ring, for which there was no internal connection information, were removed from the network, leaving 280 neurons. In addition, three neurons (AIBL, AIYL, and SMDVL) were removed, because of lacking spatial information. Eventually 277 neurons were included in the analysis. The size of the global and local C. elegans datasets analyzed here was comparable to that used in previous studies. For example, studies of the small-world properties [37] or characteristic motifs [12] of C. elegans considered 282 and 187 neurons, respectively. Both chemical and electric synapses (gap junctions) were included as connections in the analysis.
In order to understand how topological properties and dynamic observations are related, we will address our quantification schemes for topology and dynamics separately at first.
We determine two topological references which are both based on the pairwise distances of all nodes within a network. Let the distance Lst be the shortest path connecting node s with node t The first reference is based on the topological modules (topological module reference, TM, see Figure 2 top). It is computed from the distance matrix L = Lst which is then analyzed with a standard hierarchical clustering method. We tested single-linkage, complete-linkage and average-linkage approaches and found basically no differences between these methods for the task at hand. In the following, we used UPGMA (Unweighted Pair Group Method with Arithmetic mean) clustering, that is, the pair-wise combination of nodes or groups of nodes with minimal distance which is determined by the arithmetic means of the respective groups. The relative positions of the nodes which are the leaves of the topological reference tree obtained in this fashion are a condensed representation of all distance relations within the network. A similar way of analyzing the module structure uses the topological overlap [14]. The modules predicted with this method can be recovered from the topological reference tree by horizontally cutting the tree at a certain hight. The tree fragments resulting from this thresholding procedure serve as module predictions. In principle one has to analyze the dependence of the module predictions on threshold variation or conversely one can determine the threshold by prescribing the number of modules μ. Assigning a label (e.g. a color) to each node within a particular module leads to the final result, the TM reference, for which agreement with the distribution patterns of excitations can be checked.
The second topological reference is based on the central node h of the network (central node reference, CN, see Figure 3 top). Although many properties can in principle contribute to the centrality of a node, we will here select node h to be the one displaying the highest node betweenness B [59]–[61]. The distances between h and all other nodes form a distance vector. All nodes with the same entry in the distance vector (e.g. equidistant nodes from h) are taken to form a cluster, representing this topological reference (CN clusters). Resorting the distance vector accordingly yields the color-coded CN reference. Here, the number of clusters μ is given by the maximal distance from node h.
Dynamics were simulated on the graph architectures using the discrete excitable (DE) model described in the introduction. We used 35000 update steps (first 10000 updates were discarded) with the following parameter constellation: the rate of spontaneous excitations f was varied in the range of 10−6<f<1 to systematically study the impact of noise on the formation of the excitation patterns; recovery probability p was set to a constant value of p = 0.1; the initial condition was a random equipartition of the states E and T. This parameter constellation will be used in all of the studies presented here.
In the basic DE model highly connected networks are in principle characterized by burst dynamics. Indeed, spikes emerge at very low values of f even here, but with a sufficiently high simulation time they are outbalanced by burst dynamics. We solved this problem by introducing parameter ω in our excitable model system. This threshold depends on the degree ks of a node S and determines the number of excitations necessary to turn a susceptible node into the excited state. In this variant all incoming excitations are stored in node S until ωs = ks·κ (with a minimum value of ω = 1) is reached.
In order to allow for a direct comparison with topology, we base our analysis of the dynamics on pairwise node comparisons: for each pair of nodes we count the number of simultaneous excitations σst in the given time interval. Properly normalizing these quantities to arrange between 0 and 1 (σ ̃st) and converting the corresponding matrix into a distance matrix C = Cst = 1−σ ̃st leads to the correlation matrix C which represents the distribution patterns of excitations for a given graph and a given parameter constellation of the DE model. We aimed at understanding to what extent a selected topological reference is capable of explaining the patterns in the correlation matrix. To this end, the matrix can now be converted into a clustering tree (again by using UPGMA: see Topological references). The idea is now to rearrange the branches in the tree to best fit a given reference vector. The corresponding sequence of nodes constitutes the final result for the dynamics, namely the vector of dynamically detected clusters (DDC vector). The reference of the sorting vector can be any of the two topological references discussed above. Figures 2 and 3 summarize our analysis strategy. For the sorting we use an alignment algorithm which switches two neighboring branches at any position in the tree (obtained from the excitation patterns) as long as the similarity to the topological reference is increased. The decisive factor concerning the comparison of a pair of branches is the individual module composition of the respective leaves indicated by the mixture of (color) labels. A similar technique for the comparison of clustering trees has been introduced in [62].
For computation of our new quantity assessing the match between topology and dynamics, the dynamic modularity Qdyn, we compare two clustering trees, one coming from topology (with the clusters in the tree matching the modules in the graph), the other coming from the dynamics (more specifically: the matrix of simultaneous excitations). Cutting the first tree at a certain height (given by the module number, which is a parameter in our analysis) yields a set of modules, which we label by colors. Copying these node labels in the topological tree to the dynamic tree, and sorting for as many matching colors as the tree structure allows, permits us to quantify the color matches and mismatches between the two trees. Our null model is randomly distributing color labels on the graph (i.e. a sorting task of the dynamics tree to a random topological reference). As all these quantities depend strongly on the numbers of nodes in each module (or reference class), we normalize them to these sizes. In practice, this normalization is only important when we have very different sizes of modules in a graph. In this way we can assess whether the matching between a topological feature (here: the modules) and the dynamics (represented by the matrix of simultaneous excitations) is higher (or, in principle, even lower) than expected at random.
The same holds for the other topological reference, the CN reference, where the labels are provided not by a clustering tree, but by the distance from the central node. The possible values of for a topological reference R lie between zero and unity with indicating the strongest agreement to the topological reference. Values below unity hint at a deviating distribution of nodes in the dynamic cluster tree.For both the topological reference and each DDC vector the distribution values θ are determined via comparison of the scattering of nodes π belonging to the same topological module i (as indicated by the color) with a null-hypothesis of this color distribution which is the average standard deviation (in l = 1,000 realizations) of the same amount of nodes randomly scattered over the whole network size n. The resulting quotient is normalized to the size of each module nmod.
|
10.1371/journal.pgen.1000199 | Deletion of the Pluripotency-Associated Tex19.1 Gene Causes Activation of Endogenous Retroviruses and Defective Spermatogenesis in Mice | As genetic information is transmitted through successive generations, it passes between pluripotent cells in the early embryo and germ cells in the developing foetus and adult animal. Tex19.1 encodes a protein of unknown function, whose expression is restricted to germ cells and pluripotent cells. During male spermatogenesis, Tex19.1 expression is highest in mitotic spermatogonia and diminishes as these cells differentiate and progress through meiosis. In pluripotent stem cells, Tex19.1 expression is also downregulated upon differentiation. However, it is not clear whether Tex19.1 has an essential function in germ cells or pluripotent stem cells, or what that function might be. To analyse the potential role of Tex19.1 in pluripotency or germ cell function we have generated Tex19.1−/− knockout mice and analysed the Tex19.1−/− mutant phenotype. Adult Tex19.1−/− knockout males exhibit impaired spermatogenesis. Immunostaining and histological analysis revealed defects in meiotic chromosome synapsis, the persistence of DNA double-strand breaks during meiosis, and a loss of post-meiotic germ cells in the testis. Furthermore, expression of a class of endogenous retroviruses is upregulated during meiosis in the Tex19.1−/− testes. Increased transposition of endogenous retroviruses in the germline of Tex19.1−/− mutant mice, and the concomitant increase in DNA damage, may be sufficient to disrupt the normal processes of recombination and chromosome synapsis during meiosis and cause defects in spermatogenesis. Our results suggest that Tex19.1 is part of a specialised mechanism that operates in the germline to repress transposable genetic elements and maintain genomic stability through successive generations.
| The germ cells—eggs in females and sperm in males—are responsible for passing genetic information from one generation to the next. As any genetic changes that arise in the germ cells can be transmitted to the next generation, germ cells are a prime target for the activity of mobile genetic elements. Mobile genetic elements make up around 40% of a mammalian genome, and many of these elements are derived from retroviruses that have infected germ cells, or early embryonic precursors to germ cells, and have integrated into the genome. Here, we characterise the function of Tex19.1, a gene whose expression is restricted to germ cells and the pluripotent cells that are early embryonic precursors to germ cells. We show that when Tex19.1 is deleted from mice, germ cells have problems progressing through meiosis, and sperm production is impaired. Furthermore, we show that, in the absence of Tex19.1, endogenous retroviruses are activated in male germ cells attempting to go through meiosis. Our results suggest that Tex19.1 is part of a specialised mechanism that guards against mutagenic endogenous retrovirus activity in germ cells and pluripotent cells and thus helps to maintain the integrity and stability of the genome through successive generations.
| The germ cells of sexually reproducing organisms have a unique role in generating genetic diversity and transmitting genetic information from one generation to the next. Establishment of the germline in mammals involves the induction of germ cells from pluripotent epiblast cells through the action of extra-embryonic ectoderm-derived bone morphogenetic proteins and occurs comparatively late in development, commencing around day 6.25 days post coitum (dpc) in mice [1]–[4]. At around 12.5 dpc -13.5 dpc the sexually dimorphic germ cells become committed to develop along either a male or a female pathway and start to initiate sex-specific differentiation [5]. Although there are numerous differences in the differentiation of the germline and in the timing and regulation of meiosis between the sexes, the fundamental events of meiosis that increase genetic diversity and reduce ploidy of the gametes are common to both.
The main group of genes that have been shown to be required for mammalian meiosis are those involved in the recombination and synapsis of homologous chromosomes. Mice carrying loss-of-function mutations in these genes, such as Atm, Dmc1, γH2AX, Mlh1, Msh5, Rec8, Rad51, Smc1β, Spo11, Sycp1, Sycp2, Sycp3, Syce2 and Tex14, typically exhibit defects in chromosome synapsis in both sexes, although male and female germ cells can exhibit different responses to these defects [6]–[8].
A second group of genes that are required for progression through meiosis are those involved in repression of transposable genetic elements. Retrotransposons, for example long interspersed repeats (LINEs), short interspersed repeats (SINEs), and endogenous retroviruses such as intracisternal A-particles (IAPs) are the major class of transposable genetic elements in mammals and comprise around 37.5% of the mouse genome [9]. To allow new transposition events to propagate through subsequent generations, retrotransposons have evolved to be active in the germline. Accordingly, germ cells appear to have evolved mechanisms to reduce the mutational load of retrotransposon activity. Mutations in genes involved in mediating DNA methylation-dependent transcriptional repression of retrotransposons cause increased expression of retrotransposons and defects in chromosome synapsis during meiosis in either male or female germ cells [10],[11]. For example Dnmt3L is a catalytically inactive member of the DNA methyltransferase family that is expressed in foetal germ cells but is absent by 6 days post partum (dpp) [11]. Male mice null for this gene do not methylate dispersed repeat DNA during foetal germ cell development, and express LINEs and a class of endogenous retroviruses known as intracisternal A-particles (IAPs) in the germline [11]. Dnmt3L mutant male mice also exhibit meiotic abnormalities that result in a loss of post-meiotic germ cells in the testis [11]. The Dnmt3L mutant phenotype suggests that epigenetic changes that occur in foetal germ cells can cause meiotic defects later in germ cell development. A second example is provided by the murine piwi-related genes, which encode germline-specific proteins that are associated with a class of small germline-specific piwi-interacting RNAs (piRNAs) and are also required to repress retrotransposons during spermatogenesis [12]–[14]. Mili and Miwi2 both appear to be involved in de novo methylation of LINE and IAP elements during germ cell development in male embryos, and both mutants exhibit reduced DNA methylation and increased expression of LINE and IAP element in the testis, and defects in chromosome synapsis during meiosis in male germ cells [12]–[14]. The mechanism by which increased expression of retrotransposons results in the defects in chromosome synapsis seen in these mutant mice is unknown, but the phenotype of these mutant mice suggests that repression of transposable genetic elements is required to allow germ cells to progress through meiosis.
A set of testis expressed (Tex) genes has been identified in a subtractive hybridisation screen for genes expressed in spermatogonia but not somatic tissue [15]. One of these genes, Tex19.1 (AAH53492.1), was found in a screen for potential RNA-targets of the germline-specific RNA binding protein Dazl by immunoprecipitation and microarray analysis [16]. Further unpublished work in this laboratory and work recently published by Kuntz et al. [17] confirmed an earlier report that Tex19.1 is a “pluripotent cell expressed gene” [18]. Humans and primates possess a single Tex19 gene in their genome, but in rodents a recent duplication has produced a two-gene family arranged as divergently transcribed genes separated by 29kb of DNA [17]. While expression of murine Tex19.1 is restricted to pluripotent stem cells and developing germ cells, Tex19.2 is expressed in the testis somatic tissues and does not appear to be restricted to germ cells or pluripotent stem cells in mice [17].
The expression pattern of Tex19.1 suggests that the protein could have an important role in pluripotency or germ cell function. Since the sequence of Tex19.1 gives no clue to the biochemical function of this protein we decided to take a genetic approach to determining the function of this gene in the germline. In this paper we report that targeted deletion of Tex19.1 in mice results in upregulation of endogenous retrovirus expression in testicular germ cells, perturbed chromosome synapsis during meiosis, and impaired spermatogenesis.
Tex19.1 knockout mice were generated by replacing the Tex19.1 open reading frame with a neomycin selection cassette by homologous recombination in E14 embryonic stem cells [19]. Homologous regions were cloned by PCR from E14 embryonic stem cell genomic DNA using primers listed in Supplementary Table S1. The Tex19.1 targeting vector was linearised, electroporated into E14 embryonic stem cells, and neomycin-resistant clones screened for the desired integration event by PCR and Southern blot. Tex19.1+/− ES cells were used to generate Tex19.1−/− knockout mice by blastocyst injection and breeding as described [19]. Mice were genotyped by multiplex PCR using primers listed in Supplementary Table S1. Phenotypic analysis was performed on mice with a 129/Ola x CD1 mixed genetic background. Generation and analysis of Tex19.1−/− knockout mice was performed under a UK Home Office project licence with approval from an institutional ethics committee.
Anti-Tex19.1 antibodies were raised in rabbits using the synthetic peptide 78ESEQEPGPEQDAWRG92 (Eurogentec). This peptide was designed to be specific to Tex19.1 and is not present in the Tex19.2 protein sequence. Antibodies were affinity purified from sera using the immunising peptide immobilised on a Sulfolink column (Pierce) according to manufacturer's instructions.
Testes were recovered from mice, fixed at 4°C overnight in 4% paraformaldehyde in phosphate buffered saline (PBS) and embedded in paraffin wax. 6 µm-thick sections were dewaxed in xylene, rehydrated, and antigen retrieval performed by boiling slides for 15 minutes in 0.01 M sodium citrate, pH 6.0. Sections were blocked, incubated with rabbit anti-Tex19.1 primary antibody at 1∶50, and bound antibody detected using the DAKOvision ABC diaminobenzidine (DAB) kit as described by the manufacturer (DakoCytomation). For peptide competition, anti-Tex19.1 antibodies were pre-incubated with 5 nM immunising peptide. For immunostaining of cultured cells, cells were fixed for 30 minutes at room temperature with 3.7% formaldehyde in PBS, then blocked with PBS containing 5% serum and 0.01% Tween-20. Cells were incubated with rabbit anti-Tex19.1 primary antibody at 1∶100, then fluorescently labelled secondary antibodies at 1 µg/mL (Invitrogen). DNA was counterstained with 2 µg/mL DAPI.
E14 embryonic stem cells or 13 dpp postnatal testes were lysed in cytoplasmic lysis buffer (10 mM Hepes pH 7.6, 3 mM MgCl2, 40 mM KCl, 50 mM β-glycerophosphate, 5% glycerol, 0.5% Igepal CA-630, 2 mM NaF, 1 mM Na3VO4, 2 mM DTT and protease inhibitors) for 5 minutes on ice and the whole cell lysate centrifuged for 5 minutes at 1000g at 4°C. The nuclear pellet was resuspended in Laemmli buffer, boiled for 5 minutes and sonicated to disrupt genomic DNA. The cytosolic supernatant was mixed with Laemmli buffer and boiled for 5 minutes. Equivalent proportions of each fraction were separated by SDS-PAGE then Western blotted.
Testis was homogenised in Laemmli buffer, boiled for 5 min and sonicated to disrupt genomic DNA. Western blotting was performed using standard procedures [20]. Tex19.1 was detected with rabbit anti-Tex19.1 polyclonal antibodies used at a 1∶200 dilution, mouse anti-Gapdh antibodies (Abcam) were used at 1∶1000, mouse anti-HP1α antibodies (Chemicon) at 1∶2500, and rabbit anti-histone H3 antibodies (Abcam) at 1∶20000. Peroxidase-conjugated secondary antibodies and enhanced chemiluminescence were used to detect primary antibodies.
Non-radioactive Southern blots were performed using a digoxigenin-labeled DNA probe generated using primers listed in Supplementary Table S1, and alkaline phosphatase-conjugated anti-digoxigenin antibodies, essentially as described by the manufacturer (Roche).
Testis RNA was isolated with Trizol (Invitrogen) according to the manufacturer's protocol and reverse transcription performed with Superscript III (Invitrogen) on 1 µg RNA per reaction using oligo dT primer. Primers for RT-PCR are listed in Supplementary Table S1.
For quantitative PCR (qPCR), random-primed cDNA was generated from total RNA using Superscript III (Invitrogen). qPCR was performed using SYBR Green PCR System (Applied Biosystems) and a PTC-200 thermal cycler equipped with a Chromo4 continuous fluorescence detector and Opticon Monitor software (MJ Research). Primers for qPCR are listed in Supplementary Table S1. Five technical replicates were performed for each biological sample, and the relative changes in gene expression determined using the ΔΔ−2Ct method as described [21]. As Tex19.1 is expressed in the germ cells in the testis the Sertoli cell marker Sdmg1 [22] was used to normalise cDNAs prepared from different animals to reduce the probability that the cDNAs were being normalised to a transcript whose level could be influenced by loss of Tex19.1.
Both testes from each adult animal (6–36 weeks old) were weighed, and the mean testis weight was used for statistical comparison. For sperm count one epididymis from each animal was homogenised in 1 mL 1% sodium citrate and incubated for 5 minutes at room temperature to allow the debris to settle. Sperm in the supernatant was then counted with a hemocytometer.
Testes were fixed for 4–6 hours in Bouin's solution (Sigma-Aldrich) at room temperature, then embedded in wax. For histological analysis 6 µm sections were dewaxed with xylene, rehydrated, then stained with hematoxylin and eosin.
Immunostaining of chromosome spreads from meiotic spermatocytes was performed essentially as described [23]. Briefly, testes were homogenized in PBS and 0.1 mL of cells were incubated in 0.5 mL 5% sucrose on a microscope slide for 1 hour. Cells were lysed with 0.1 mL 0.05% Triton-X-100 for 10 minutes, and fixed with 0.8 mL of fixing solution (2% paraformaldehyde, 0.02% SDS in PBS) for 1 hour. The slides were then washed, blocked with 5% serum, 0.1% Tween in PBS and incubated with primary antibodies for 1 hour. Mouse anti-Sycp3 antibodies (Abcam) were used at a 1∶2000 dilution, rabbit anti-Sycp1 antibodies (Abcam) at 1∶250, rabbit anti-γH2AX antibodies (Upstate Biotechnology) at 1∶200, and mouse anti-Rad51 antibodies (Upstate Biotechnology) at 1∶125. Fluorescently labelled secondary antibodies were used at 1 µg/mL (Invitrogen), and DNA was stained with 2 µg/mL DAPI.
Chromosome spreads for metaphase I analysis were prepared as described in [24]. Briefly, testes were incubated for 20 minutes in 1% sodium citrate, minced with scissors, and the cells harvested by centrifugation. Cells were then washed and resuspended in fixing solution (3∶1 methanol∶glacial acetic acid), dropped onto slides, and the resulting chromosome spreads were stained with Giemsa solution. 100 metaphase I spreads were scored per animal, and two animals scored for each genotype.
A 460 bp fragment of the MMERVK10C endogenous retrovirus was amplified by RT-PCR from Tex19.1−/− mutant testes using primers listed in Supplementary Table S1 and cloned into pBluescript II SK+ (Stratagene). Sense and anti-sense digoxigenin-labelled riboprobes were generated using T3 and T7 RNA polymerase according to the supplier's instructions (Roche). In situ hybridisation on 6 µm wax sections of Bouin's-fixed testis tissue was performed essentially as described [25] using 100 ng/mL digoxigenin-labelled probe and a hybridisation temperature of 50°C. Bound probe was detected with alkaline phosphatase-conjugated anti-digoxigenin antibodies (Roche) and BCIP/NBT precipitating stain (Vector Labs), then sections counterstained with nuclear fast red according to manufacturer's instructions. The antisense MMERVK10C digoxigenin-labelled riboprobe was also used for non-radioactive Northern blotting of 1 µg testis RNA as described [22].
Spermatogenesis in the adult testis involves the differentiation of a small pool of spermatogonial stem cells into large numbers of mature sperm. Within the testis spermatogenesis takes place in the seminiferous tubules where the mitotic spermatogonia reside at the outermost edge of the tubule, and progressive stages of differentiation are found as layers of meiotic spermatocytes then haploid spermatids located more and more centrally towards the lumen of the tubule [26]. Published RT-PCR expression data for Tex19.1 in purified spermatogenic cell populations suggests that Tex19.1 expression is highest in mitotic spermatogonia, decreases as the spermatocytes progress through meiosis, and is present at low levels in round spermatids [27]. To establish the expression pattern of Tex19.1 protein in male testis during spermatogenesis we raised anti-peptide antibodies to Tex19.1. Immunohistochemistry on mouse testis shows strong cytoplasmic expression of Tex19.1 in spermatogonia that is downregulated as these cells differentiate and progress through meiosis (Figure 1A–C). We were able to detect cytoplasmic Tex19.1 protein in some meiotic spermatocytes (Figure 1C), but not in others (Figure 1B) suggesting that Tex19.1 protein expression is switched off as the germ cells proceed through meiosis. The expression of Tex19.1 protein in spermatogonia and spermatocytes is consistent with that observed for Tex19.1 mRNA by in situ hybridisation (Figure S1A–C).
Our finding that Tex19.1 is present in the cytoplasm of spermatogonia and early spermatocytes in the adult testis is not consistent with the published nuclear localisation of Tex19.1 protein in embryonic stem cells [17]. We confirmed that our anti-Tex19.1 antibody detects Tex19.1 and not a cross-reacting antigen by blocking the anti-Tex19.1 immunohistochemistry signal by competition with the immunising peptide, and by immunohistochemistry on Tex19.1−/− knockout testes (Figure S1D–F). We were also able to detect a predominantly cytoplasmic subcellular localisation of Tex19.1 by immunostaining germ cells isolated from 14.5 dpc embryonic testes (Figure S1G). Again the cytoplasmic anti-Tex19.1 immunostaining could be competed with the immunising peptide, and was absent in germ cells from Tex19.1−/− knockout embryos (Figure S1H, I). These data suggest that the anti-Tex19.1 antibody used in this present study specifically recognises endogenous Tex19.1 in germ cells, and that at least some Tex19.1 is present in the cytoplasm of spermatogonia and early spermatocytes in adult mouse testes.
To investigate whether the discrepancy between the cytoplasmic localisation of Tex19.1 in germ cells presented in this study and the nuclear localisation of Tex19.1 in embryonic stem cells described previously [17] is caused by the difference between the cell types studied we performed immunostaining for Tex19.1 on embryonic stem cells. In contrast to the previous report [17], we found that Tex19.1 is predominantly cytoplasmic in embryonic stem cells (Figure 1D–F).
To exclude the possibility that our anti-Tex19.1 antibody is unable to detect a nuclear population of Tex19.1 due to loss or masking of the epitope during the immunohistochemical procedures we biochemically fractionated 13 dpp prepubertal testes and embryonic stem cells into nuclear and cytoplasmic fractions. On Western blots the 42 kDa Tex19.1 band was barely detectable in the nuclear fraction, but was easily detectable in equivalent loadings of the whole cell lysate and cytoplasmic fractions (Figure 1G,H). The observed size of the anti-Tex19.1 band in the Western blots (42 kDa) correlates well with the predicted molecular weight of Tex19.1 (40.4 kDa). This 42 kDa band appears to be endogenous Tex19.1 as it is not present in testes from Tex19.1−/− knockout animals (Figure 2E). The biochemical fractionation of embryonic stem cells and testes therefore confirms the predominantly cytoplasmic subcellular localisation of Tex19.1 that we have observed by immunohistochemistry and immunostaining. Taken together the data presented here strongly suggests that Tex19.1 is a predominantly cytoplasmic protein in embryonic stem cells and germ cells.
Germ cells in many species possess specialised cytoplasmic structures termed nuage that are implicated in RNA metabolism. Although Tex19.1 appears to be a predominantly cytoplasmic germ cell protein, the subcellular localisation and cell-type distribution of Tex19.1 appears to be distinct from the nuage component Tdrd1 [28] (Figure S2). Thus Tex19.1 does not appear to be a novel component of nuage. As the subcellular localisation of Tex19.1 does not provide any major insight into what the cellular function of this protein might be, and the Tex19.1 protein sequence does not contain any functional domains to illuminate the potential biochemical function of this protein, we decided to take a genetic approach to analyse the function of Tex19.1 in the germline.
In order to investigate the function of Tex19.1 in germ cell development we generated Tex19.1−/− knockout mice. The Tex19.1 open reading frame was replaced with a neomycin-resistance cassette by homologous recombination in embryonic stem cells (Figure 2A), and the targeted deletion confirmed by Southern blotting (Figure 2B). The heterozygous Tex19.1+/− embryonic stem cells were used to generate chimaeric mice by blastocyst injection, and the Tex19.1− mutant allele bred to homozygosity (Figure 2C). Tex19.1−/− homozygous pups were born from heterozygous crosses at a sub-Mendelian frequency (72 wild-type, 131 heterozygous, 40 homozygous pups born from heterozygous matings, significant deviation from expected Mendelian 1∶2∶1 ratio, χ2-test p<0.01). The low rate of recovery of Tex19.1−/− homozygous animals at birth indicates that some Tex19.1−/− homozygous embryos are lost during embryonic development.
To confirm that the Tex19.1− mutant allele removes Tex19.1 mRNA and protein we performed RT-PCR on Tex19.1−/− testis cDNA with Tex19.1-specific primers (Figure 2D) and Western blotting on Tex19.1−/− testis protein extract with anti-Tex19.1 antibodies (Figure 2E). Both methods show that Tex19.1 is not expressed in the testes of Tex19.1−/− homozygous mice. We conclude that the Tex19.1− allele that we have produced is a null allele and that Tex19.1 function is ablated in the Tex19.1−/− homozygous mice.
The surviving Tex19.1−/− knockout mice are apparently healthy, with overtly normal morphology and behaviour. However both male and female Tex19.1−/− knockout mice have reduced fertility. Tex19.1−/− knockout females have a mean litter size of 5.0±2.2 SD, n = 9 compared to a litter size of 10.6±2.9, n = 23 for Tex19.1+/− heterozygous females (Student's t-test p<0.01). The reduced fertility in Tex19.1−/− homozygous females is consistent with expression of Tex19.1 in embryonic ovaries [17], and a detailed analysis of the cause of the subfertility in the Tex19.1−/− knockout female mice will be published elsewhere. Similarly, Tex19.1−/− knockout male mice are also severely subfertile when test-mated with wild-type female mice. Although one out of the eleven Tex19.1−/− knockout males tested for fertility was able to sire offspring, the remaining Tex19.1−/− knockout males were infertile. The sterile Tex19.1−/− knockout male mice were apparently able to mate with the wild-type females to produce a copulation plug, but these females did not give birth to any pups. Tex19.1−/− knockout male mice have smaller testes (Figure 3A, B), with the median testis weights of adult animals reduced from 111 mg in Tex19.1+/+ wild type and Tex19.1+/− heterozygous mice to 42.5 mg in Tex19.1−/− knockout littermates (Mann Whitney U-test, p<0.01). Furthermore, the median epididymal sperm count is reduced to 1.3×105 in Tex19.1−/− knockout mice from 1.3×107 in wild type and heterozygous littermates (Mann Whitney U-test, p<0.01, Figure 3C) suggesting that spermatogenesis is defective in the Tex19.1−/− knockout testes. We have not been able to detect any difference in testis weight or sperm count between Tex19.1+/+ wild-type and Tex19.1+/− heterozygous animals.
The extent of the spermatogenesis defect in the Tex19.1−/− knockout males varied between individual animals. When sperm was counted, a strong reduction was observed for most of the animals, but for the single fertile animal the sperm count was close to normal levels (Figure 3C). A similar variation in testis weight was also evident amongst Tex19.1−/− knockout animals (Figure 3B). This phenotypic variation was not influenced by the age of the mice at the time of analysis. Age-matched adult mice were analysed at 6 weeks, 3 months, 6 months and 9 months during the course of this study, and there appeared to be no correlation between the severity of the phenotype and the age at which the adult mice were examined. Rather phenotypic variation was observed in adult mice at all ages (R.O. and I.R.A., data not shown). The outbred component of the genetic background of these mice may contribute to this variation.
We investigated the spermatogenesis defect in Tex19.1−/− knockout mice further by examining the testis histology in these animals. We did not detect any overt differences in testis histology between Tex19.1+/+ wild-type and Tex19.1+/− heterozygous animals. However, Tex19.1−/− knockout testes have considerably narrower seminiferous tubules than their wild-type or heterozygous littermates due to a reduction in the number of post-meiotic germ cells (Figure 4A, B). This phenotype was also subject to some heterogeneity. In animals with a more severe phenotype all postmeiotic cell-types were missing and the most advanced meiotic cells were in pachytene stage (Figure 4C, D). In animals with a less severe phenotype a proportion of cells were able to complete meiosis and haploid cells could be detected, although often in comparatively low numbers (Figure 4C, E). Like the testis weight and sperm count phenotypes described in the previous section, there appeared to be no correlation between the severity of the testis histology phenotype and the age at which the adult mice were analysed (from 6 weeks to 9 months).
To test whether the reduction in the number of post-meiotic germ cells in Tex19.1−/− knockout testes arises from a decrease in the spermatogonial mitotic divisions or apoptosis of differentiating germ cells, we counted the number of B-type spermatogonia, early meiotic cells and apoptotic cells in testis sections from Tex19.1+/− heterozygous and Tex19.1−/− knockout animals. B-spermatogonia and early meiotic cells were identified by their location and histological appearance in the seminiferous tubules [26], and apoptotic cells were identified using the TUNEL assay to label fragmented chromatin. Whereas the number of B-type spermatogonia and early meiotic cells did not differ between Tex19.1+/− heterozygous and Tex19.1−/− knockout testes (R.O., data not shown), TUNEL staining showed an increase in the number of dying cells in adult Tex19.1−/− knockout testis (Figure S3A, B, N). In more severe Tex19.1−/− knockout seminiferous tubules, TUNEL-positive cells were found within or next to layers of meiotic germ cells (Figure S3C), but even at high magnification the nuclear morphology of these TUNEL-positive cells was not distinct enough to allow their developmental stage to be unambiguously identified (Figure S3H–J). In less severe Tex19.1−/− knockout seminiferous tubules, TUNEL-positive cells could also be found between the layers of meiotic germ cells and post-meiotic round spermatids (Figure S3D). At higher magnification, some of these TUNEL-positive cells could be identified as metaphase I spermatocytes (Figure S3E–G).
In order to further define the point during spermatogenesis when the Tex19.1−/− knockout cells are dying we examined the synchronous first wave of spermatogenesis that occurs in prepubertal mice. The first wave of spermatogenic germ cells initiates meiosis at around 10 dpp in the prepubertal testis, and progresses through the pachytene stage of meiosis from around 14 to 20 dpp to produce the first post-meiotic round spermatids around 21 dpp, and mature sperm at around 31 dpp [29],[30]. Analysis of apoptosis (Figure S3N) and testis histology (Figure S4) at various stages of prepubertal testis development revealed no overt differences in testis histology and no statistically significant increase in apoptosis at 16 dpp in Tex19.1−/− knockout testes. However by 19–22 dpp, a reduction in the number of meiotic and post-meiotic germ cells and an increase in the frequency of cell death are both evident in Tex19.1−/− knockout testes (Figures S3N, S4). In 22 dpp testes, clusters of TUNEL-positive cells can be seen within the layer of pachytene germ cells that line the lumen of the seminiferous tubule suggesting that at least some apoptosis is occurring at the pachytene stage of meiosis (Figure S3K–M). The high level of apoptosis in the Tex19.1−/− knockout testes increases by 29–31 dpp to the level seen in adult testes (Figure S3N). This data suggests that the reduction in the number of post-meiotic germ cells and increased levels of apoptosis seen in the adult Tex19.1−/− knockout testes is at least partly due to some Tex19.1−/− knockout germ cells initiating apoptosis during the pachytene stage of meiosis, and some Tex19.1−/− knockout germ cells initiating apoptosis during metaphase I.
Although the vast majority of the Tex19.1−/− null testes examined contained differentiating germ cells, two of thirty analysed knockout animals had an extremely severe phenotype with one testis that completely lacked germ cells. One of these agametic testes was isolated from a 31 dpp prepubertal mouse (Figure 4F) suggesting that this extreme phenotype is indicative of defects occuring during embryonic or early post-natal germ cell development rather than a progressive loss of spermatogonial stem cells in an ageing adult testis. However, as only a small number of testes exhibited this phenotype, we were not able to study this extreme phenotype further and instead focused on the meiotic phenotype evident in the vast majority of the Tex19.1−/− mutant testes.
We next attempted to determine the cause of the increased apoptosis in Tex19.1−/− null testes. Defects in homologous chromosome synapsis or homologous recombination during meiotic prophase can cause apoptosis in late pachytene spermatocytes [31],[32]. Therefore we used immunocytochemistry on meiotic chromosome spreads to analyse chromosome synapsis and homologous recombination in Tex19.1−/− knockout testes. In order to analyse chromosome synapsis, meiotic chromosome spreads were stained using Sycp3 as a marker for lateral elements of meiotic chromosomes and Sycp1 as a marker for synapsed homologous chromosomes [6]. In wild-type pachytene cells the autosomal chromosome axes stain completely for both markers, whereas the X and Y sex chromosomes remain largely asynapsed with only a small area of Sycp1 staining in the pseudo-autosomal region (Figure 5A). In contrast, about half the pachytene cells in Tex19.1−/− homozygotes have Sycp3-stained autosomal chromosomal axes that lack Sycp1 staining (Figure 5B–D, I). The asynapsed chromosomes in Tex19.1−/− knockout cells did not appear to be arranged in homologous pairs (Figure 5D). However it is not clear whether the asynapsed chromosomes have never paired in Tex19.1−/− knockout spermatocytes, or have paired but have subsequently fallen apart. In some of the incompletely synapsed Tex19.1−/− knockout cells, some chromosomes appeared to form chains linked by regions of apparent non-homologous synapsis (Figure 5C, asterisk). Incompletely synapsed pachytene cells comprise less than 1% of spreads from Tex19.1+/+ wild-type or Tex19.1+/− heterozygotous testes (Figure 5I). Thus Tex19.1−/− knockout animals exhibit defects in homologous chromosome synapsis during male meiosis.
During meiotic prophase, homologous recombination starts prior to homologous chromosome pairing and synapsis [33]. As progression of homologous recombination and chromosome synapsis are interdependent on each other [6],[7], we investigated whether the chromosome synapsis defect in Tex19.1−/− knockout spermatocytes was a consequence of an earlier defect in the initiation of homologous recombination. The appearance of DNA double strand breaks and the formation of early recombination foci during meiotic prophase can be detected by immunostaining for the phosphorylated histone γH2AX and the recombinase enzyme Rad51 respectively [33],[34]. γH2AX staining is normally present on chromatin during the leptotene and zygotene stages of early meiotic prophase. As synapsis proceeds during zygotene, the DNA double strand breaks are resolved, resulting in γH2AX staining disappearing from the autosomal chromosomes, but not the sex chromosomes. In normal Tex19.1+/+ wild-type pachytene cells, chromosome synapsis is complete and only the sex chromosomes stain for γH2AX (Figure 5E). However, the incompletely synapsed pachytene cells in Tex19.1−/− knockout testes, exhibit strong diffuse γH2AX staining (Figure 5F). This γH2AX staining is localised to the regions of the chromosome spreads that contain the unsynapsed chromosomes (Figure 5F). Similarly, immunostaining for the early recombination foci marker Rad51, which largely disappears from autosomal chromosomes as synapsis proceeds, suggests that Rad51 foci are formed in Tex19.1−/− knockout spermatocytes, but are not resolved or matured on the unsynapsed chromosomes (Figure 5G, H). Thus the formation of DNA double strand breaks and the assembly of early recombination foci both appear to be occurring in Tex19.1−/− knockout spermatocytes. This suggests that the defect in meiotic chromosome synapsis that we have observed in Tex19.1−/− spermatocytes does not appear to be a secondary consequence of impaired initiation of homologous recombination. Rather, the presence of DNA double strand breaks and early recombination foci in the unsynapsed regions of the incompletely synapsed Tex19.1−/− pachytene spermatocytes is consistent with impaired chromosome synapsis. Furthermore, the presence of DNA double strand breaks and early recombination foci in unsynapsed regions of incompletely synapsed Tex19.1−/− pachytene spermatocytes indicates that the unsynapsed chromosomes arise from a failure to initiate synapsis rather than premature desynapsis. The unsynapsed chromosomes in the incompletely synapsed pachytene Tex19.1−/− knockout cells are presumably sufficient to trigger apoptosis at the pachytene checkpoint [31],[32], and would account for the increased levels of cell death seen in pachytene stage meiotic germ cells in Tex19.1−/− knockout testes (Figure S3).
Although incompletely synapsed pachytene cells could explain the increased levels of cell death in the pachytene meiotic germ cells in Tex19.1−/− knockout testes, the presence of apoptotic metaphase I spermatocytes in these animals suggests that there may be an additional defect later in spermatogenesis to account for cell death at the metaphase I stage. Around half of the Tex19.1−/− knockout pachytene cells did not appear to have any overt defects in chromosome synapsis (Figure 5I), and would therefore presumably be able to progress to metaphase I and continue through spermatogenesis. To investigate whether there might be additional defects in chromosome behaviour at later stages of meiosis in Tex19.1−/− knockout spermatocytes, we prepared and analysed meiotic metaphase I chromosome spreads. During metaphase I of meiosis, homologous chromosomes are held together as bivalents by chiasmata (Figure 5J). 94% of the metaphase I spreads from Tex19.1+/− heterozygous testes contained only bivalent metaphase I chromosomes, 5% contained univalent sex chromosomes, and 1% contained univalent autosomes. However, only 34% of the metaphase I spreads from Tex19.1−/− knockout testes contained only bivalent metaphase I chromosomes, while 56% of the spreads contained univalent sex chromosomes, and 33% contained univalent autosomes (Figure 5K). 23% of the Tex19.1−/− knockout metaphase I spreads feature univalent autosomes and univalent sex chromosomes. Thus Tex19.1−/− knockout testes contain increased numbers of univalent chromosomes at meiotic metaphase I that could potentially trigger apoptosis at the metaphase I checkpoint [35],[36] and account for the apoptotic metaphase I cells seen in Tex19.1−/− knockout testes. Furthermore, the presence of univalent chromosomes in metaphase I spreads from Tex19.1−/− knockout testes is indicative of a defect in the formation or maintenance of chiasmata in post-pachytene spermatocytes.
Meiotic defects similar to those present in the Tex19.1−/− knockout testes have been observed in various different mouse mutants that carry defects in genes encoding components of meiotic chromosomes, the meiotic recombination machinery, or the synaptonemal complex [6],[7]. However, we have been unable to detect any Tex19.1 protein physically associated with meiotic chromosomes by immunostaining (R.O., data not shown), and our finding that Tex19.1 is predominantly localised to the cytoplasm rather than the nucleus suggests that Tex19.1 is unlikely to be a component of meiotic chromosomes or the synaptonemal complex. We therefore reasoned that the meiotic defects present in the Tex19.1−/− knockout testes are unlikely to be a direct effect of Tex19.1 on meiotic chromosome structure or function but rather may be an indirect consequence of changes in meiotic gene expression.
In order to detect changes in gene expression in the testis of Tex19.1−/− knockout mice, we performed microarray analysis using an Illumina MouseWG-6 v1.1 Whole Genome Gene Expression Beadchip containing 48,318 different probes. To exclude potential differences in transcript levels due to the loss of post-meiotic germ cells in the Tex19.1−/− testes we performed this analysis on testes from 16 dpp prepubertal mice during the first synchronous wave of spermatogenesis. At this stage of testis development, some germ cells are already in the pachytene stage of meiosis, but no obvious changes in cell composition were apparent between Tex19.1+/+ wild type and Tex19.1−/− knockout testes (Figure S4). RNAs from two different 16 dpp Tex19.1−/− knockout testes were compared with Tex19.1+/+ wild type or Tex19.1+/− heterozygous littermates, and transcripts that had consistent and greater than three fold changes in relative gene expression between the two groups of animals were identified. The Mouse Genome Database (http://www.informatics.jax.org) currently lists 97 mutations that are known to give rise to meiotic arrest during spermatogenesis [8]. These male meiotic arrest genes include genes that encode components of meiotic chromosomes, the meiotic recombination machinery and the synaptonemal complex such as Atm, Dmc1, γH2AX, Mlh1, Msh5, Rec8, Rad51, Smc1β, Spo11, Sycp1, Sycp2, Sycp3, Syce2 and Tex14. None of the male meiotic arrest genes listed in the Mouse Genome Database showed a consistent change in expression level in Tex19.1−/− knockout testes compared to littermate controls (I.R.A., data not shown). However, analysis of the microarray data suggested that the class II LTR-retrotransposon MMERVK10C [37] is upregulated by around four-fold in the testis RNA from each of the 16 dpp Tex19.1−/− knockout animals relative to their littermate controls (I.R.A., data not shown). The mouse genome contains around 16 approximately full-length copies of the MMERVK10C sequence in the genome, and a further 1200 fragments of the MMERVK10C endogenous retrovirus. Increased retrotransposon expression has been proposed to be responsible for impaired chromosome synapsis and meiotic defects during spermatogenesis in Dnmt3L, Miwi2 and Mili mutant mice [11]–[14]. As overexpression of the MMERVK10C retrotransposons could similarly be responsible for the meiotic defects seen in Tex19.1−/− mutant mice we sought to determine whether MMERVK10C expression is indeed upregulated in the testis in the absence of Tex19.1.
The levels of MMERVK10C expression in testis cDNA from two Tex19.1−/− knockout animals relative to their Tex19.1+/+ wild-type littermates were each tested by quantitative PCR (Figure 6A). The Sertoli cell marker Sdmg1 [22] was used to normalise cDNAs from different animals. Although there was no significant change in the expression of the ubiquitously expressed β-actin gene, or the germ cell marker Dazl [38], expression of the MMERVK10C endogenous retrovirus was increased by a factor of approximately four-fold in both Tex19.1−/− knockout animals (Student's t-test, p<0.01) (Figure 6A). Expression of LINE, SINE or IAP retrotransposons showed no significant change in the absence of Tex19.1 (Figure 6A).
To further validate the potential upregulation of MMERVK10C transcripts in the Tex19.1−/− knockout mice we performed Northern blots on testis RNA from the same two 16 dpp Tex19.1−/− knockout animals and their Tex19.1+/+ wild type littermates. Using a probe derived from the env gene of the MMERVK10C endogenous retrovirus we were able to detect a predominant 3.2 kb MMERVK10C env transcript in mouse testes, and some weaker MMERVK10C env transcripts at around 4.5 kb and 7.5 kb (Figure 6B). The Northern blot profile for MMERVK10C env transcripts is comparable to that of env-containing transcripts from HERV-K endogenous retroviruses in human teratocarcinoma cell lines [39]. Northern blotting confirmed that the predominant 3.2 kb MMERVK10C env transcript is consistently more abundant in testes from Tex19.1−/− knockout animals than in testes from their wild-type littermates at 16 dpp (Figure 6B).
In order to determine which cell types are accumulating MMERVK10C transcripts in the Tex19.1−/− knockout testes we performed in situ hybridisation on testis sections using a MMERVK10C env probe (Figure 6C–N). In Tex19.1+/+ wild-type and Tex19.1+/− heterozygous testes at 16 dpp, low levels of MMERVK10C env transcripts were present in some meiotic spermatocytes (Figure 6C,G). However, MMERVK10C transcripts were generally more abundant in Tex19.1−/− knockout testes than in testes from Tex19.1+/+ wild-type or Tex19.1+/− heterozygous littermates at 16 dpp (Figure 6D, H). The increased levels of MMERVK10C env transcript in 16 dpp Tex19.1−/− testes appeared to be largely due to the presence of strongly expressing cells located towards the centre of the tubules where meiotic spermatocytes are present (Figure 6K, L). Similarly in adult animals MMERVK10C env transcripts were upregulated in meiotic germ cells in the testes from adult Tex19.1−/− knockout animals relative to their heterozygous littermates (Figure 6E, F, I, J). A total of nine different Tex19.1−/− knockout animals at various ages were assayed for MMERVK10C expression in the testes by in situ hybridisation, and MMERVK10C expression in Tex19.1−/− knockout testes was consistently higher that in Tex19.1+/+ or Tex19.1+/− littermate controls. No in situ hybridisation signals were detected on testis sections using a sense MMERVK10C control probe (Figure 6M, N).
Taken together, the quantitative PCR, Northern blotting and in situ hybridisation data all suggest that transcripts from the MMERVK10C endogenous retrovirus are upregulated in the meiotic spermatocytes of Tex19.1−/− knockout testes.
The upregulation of retrotransposons in Dnmt3L, Mili and Miwi2 mutant mice is associated with defects in de novo DNA methylation of IAP and LINE elements in the male germline, which presumably allows increased transcription of these elements during spermatogenesis [11]–[14]. In order to investigate whether the upregulation of MMERVK10C retrotransposons in Tex19.1−/− knockout testis was caused by a similar mechanism, we investigated the DNA methylation status of CpG dinucleotides in MMERVK10C elements by bisulphite sequencing MMERVK10C elements from 16 dpp prepubertal Tex19.1−/− knockout testes. The MMERVK10C element includes a weak CpG island overlapping the LTR and 5′untranslated region (Figure S5). As promoters with weak CpG islands are good candidates for regulation by DNA methylation [40], we examined DNA methylation at CpG dinucleotides within this region. Sequence analysis of 30 independent clones from each of Tex19.1+/+ wild-type, Tex19.1+/− heterozygous and Tex19.1−/− homozygous 16 dpp testes showed that CpG dinucleotides in this region of the MMERVK10C element are predominantly methylated in the testis at this age (Figure S5). The MMERVK10C element was also methylated to a similarly high level in liver taken from the same animals as a somatic tissue control (Figure S5). The prepubertal testis is composed of approximately equal numbers of germ cells and somatic cells at 16 dpp [29],[30], therefore around half the clones analysed by bisulphite sequencing are likely to be derived from testicular germ cells and around half from testicular somatic cells. As all of the 16 dpp testis clones represented highly methylated DNA sequences (Figure S5), the MMERVK10C element appears to be highly methylated in both the germ cell and somatic cell compartments of Tex19.1+/+ wild-type, Tex19.1+/− heterozygous and Tex19.1−/− homozygous 16 dpp testes.
Although we have been unable to find any evidence that the methylation status of MMERVK10C elements in the testis changes in the absence of Tex19.1 (Figure S5), we cannot exclude the possibility that the absence of Tex19.1 causes reduced DNA methylation in a subset of MMERVK10C elements in the genome, or in a subset of germ cells in 16 dpp testes. If only a subset of germ cells have altered DNA methylation at MMERVK10C elements in Tex19.1−/− mutant testes then we estimate that this subset would need to represent less than 25% of the germ cell population to be below our detection limit in this assay (χ2-test, p<0.05). Nevertheless, our observations that loss of Tex19.1 causes the upregulation of MMERVK10C retrotransposon elements in the testis, but not IAP or LINE elements, combined with the absence of a detectable change in DNA methylation levels in MMERVK10C elements in Tex19.1−/− knockout testes, suggests that Tex19.1-mediated repression of retrotransposons may involve a mechanism that is distinct from Dnmt3L/Miwi2/Mili-mediated repression of retrotransposons. Thus we conclude that Tex19.1 is part of a novel genetic pathway that represses retrotransposons in the male germline.
This study describes the functional consequences of deleting the pluripotency-associated Tex19.1 gene in mice. Our data shows that loss of Tex19.1 causes impaired spermatogenesis and defects in chromosome synapsis during meiosis. Mutations in genes that are involved in various aspects of meiotic chromosome behaviour such as the initiation of recombination between homologous chromosomes, or the assembly of the synaptonemal complex, all typically cause defective chromosome synapsis during meiosis, and apoptosis in the male germline [6],[7]. However, although there is some similarity between these phenotypes and the Tex19.1 mutant phenotype, we have been unable to detect any localisation of Tex19.1 to meiotic chromosomes by immunostaining testis sections or testis chromosome spreads (R.O., data not shown). Indeed our data suggest that Tex19.1 is a predominantly cytoplasmic protein and is therefore unlikely to play a direct role in meiotic chromosome behaviour. Thus, although Tex19.1 mutant mice exhibit defects in chromosome pairing during meiosis, we do not believe that Tex19.1 is a component of meiotic chromosomes and favour the interpretation Tex19.1 is influencing meiotic chromosome behaviour indirectly.
Our finding that Tex19.1 is a predominantly cytoplasmic protein in germ cells and embryonic stem cells contradicts a previous study suggesting that Tex19.1 is a nuclear protein in embryonic stem cells [17]. The reason for the discrepancy between these studies is not yet clear. Kuntz et al. [17] raised monoclonal antibodies to Tex19.1 and observed nuclear staining with those antibodies in embryonic stem cells and pre-implantation embryos. The Tex19.1 peptide used by Kuntz et al. [17] to raise the monoclonal anti-Tex19.1 antibody is located C-terminally to the peptide that we have used to raise the anti-Tex19.1 antibodies in this study. Both peptides, and indeed the entire Tex19.1 open reading frame, lie within a single exon. In our study we have shown that Tex19.1 is predominantly cytoplasmic in embryonic stem cells by immunostaining and by Western blotting of subcellular fractions. We have also shown that Tex19.1 has a predominantly cytoplasmic localisation in germ cells by immunostaining germ cells isolated from embryonic testes, by immunohistochemistry on wax sections of adult testis and by Western blotting of subcellular fractions from prepubertal testes. Furthermore we have demonstrated the specificity of our antibody in the assays that we use by immunostaining and Western blotting on material from Tex19.1−/− knockout animals. As the cytoplasmic anti-Tex19.1 staining patterns that we present in this paper are lost in Tex19.1−/− knockout animals, at least some of the Tex19.1 protein that is present in germ cells and embryonic stem cells is cytoplasmic. However we cannot exclude the possibility that the two different antibodies raised in these two studies recognise mutually exclusive isoforms of Tex19.1 that have different subcellular localisations. Alternatively, the discrepancy between our study and the study by Kuntz et al. [17] could be caused by procedural differences, or by cross-reaction of anti-Tex19.1 antibodies with an unrelated antigen.
Tex19.1−/− null male mice showed some phenotypic variation between individuals ranging from completely agametic testes to fertility. This phenotypic variability may be partly due to the genetic heterogeneity in the outbred component of the genetic background used for this study. However, as some germ cells are more severely affected by the loss of Tex19.1 than other germ cells in the same animal, there is also some phenotypic variability in the absence of genetic variation. Furthermore, our finding that loss of Tex19.1 can impair spermatogenesis even in this heterogeneous genetic background suggests that mutations in the single human homologue, TEX19, could contribute to fertility problems in human populations. The human TEX19 gene contains two premature stop codons in the open reading frame that truncates the Tex19 protein from 351 residues in mouse to 164 residues in human [17]. The first premature stop codon in the human TEX19 gene is conserved in other primates suggesting that the C-terminal region of Tex19 is dispensable for function in primates [17]. The significance of this major difference in structure between human and mouse is at present unclear given our current level of understanding of the mechanisms underlying the phenotype in mouse.
The Tex19 genomic locus has undergone a duplication event in rodents to generate two closely related divergently transcribed genes [17]. The mutation that we have engineered removes the entire Tex19.1 open reading frame, but leaves Tex19.2 intact. Therefore Tex19.2 could potentially provide some functional redundancy with Tex19.1. Although Tex19.1 and Tex19.2 are reported to be expressed in testicular germ cells and testicular somatic cells respectively [17], there appears to be a moderate upregulation of Tex19.2 in Tex19.1−/− knockout testes as judged by quantitative RT-PCR (I.R.A., data not shown). It is not clear at present whether this upregulation of Tex19.2 occurs in the germ cells or somatic cells of the testis, but any upregulation of Tex19.2 that is occurring does not seem to be able to fully compensate for loss of Tex19.1. Nevertheless, deletion of the entire Tex19 locus may be required to rule out the possibility of some functional redundancy between these genes and may reveal additional functions for Tex19.1 in the germline.
This study demonstrates that Tex19.1 has a function in progression through meiosis in the male germline. Characterisation of the meiotic defect in Tex19.1−/− knockout spermatocytes indicates that homologous recombination is being initiated in the Tex19.1−/− knockout spermatocytes but that, for some chromosomes, synapsis does not occur. As homologous recombination and chromosome synapsis progress interdependently during meiosis, it is possible that the chromosome synapsis defect that we describe in Tex19.1−/− knockout spermatocytes is a secondary consequence of a defect in the progression of homologous recombination, or a secondary consequence of defects in the pairing between homologous chromosomes that normally precedes chromosome synapsis [7]. Further work is needed to dissect the molecular basis of the Tex19.1 chromosome synapsis defect in more detail, and to understand if and how the upregulation of MMERVK10C retrotransposons that we detect in Tex19.1−/− spermatocytes causes these defects in meiotic chromosome synapsis.
The Tex19.1 mutant phenotype bears some resemblance to the Dnmt3L, Miwi2 and Mili mutant phenotypes in that they all exhibit defects in meiotic chromosome synapsis and increased expression of retrotransposons in the germline [11]–[14]. However it is not yet clear whether there is a direct causal relationship between these two events. The increase in retrotransposon expression does not appear to be caused by defects in meiotic chromosome synapsis [11],[13], but it is not clear whether or how the increase in retrotransposon expression causes the defects in meiotic chromosome synapsis in any of these mutant mice. Increased transposition of mobile genetic elements could introduce quantitative, qualitative, or temporal changes in the DNA double strand breaks normally present during early meiotic prophase that could interfere with the homologous recombination events that normally precede and initiate chromosome pairing. Support for this model comes from the observation that mutating genes involved in piRNA function in flies activates the DNA damage signalling pathway [41],[42]. Alternatively, it is possible that repression of retrotransposons is important for the fidelity of homolog pairing and synapsis during meiosis, and that increased expression of these repetitive elements either interferes with homolog recognition and synapsis, or promotes pairing between non-homologous chromosomes. A third possibility is that proteins encoded by the MMERVK10C endogenous retrovirus mediate the defects in meiotic chromosome synapsis by interfering with host cell proteins involved in meiotic chromosome behaviour or regulation of the meiotic cell cycle. In this regard it is important to note that transgenic mice expressing the rec protein derived from the HERVK human endogenous retrovirus exhibit defects in spermatogenesis [43]. Lastly, there may not be a direct causal relationship between retrotransposon de-repression and chromosome asynapsis. Rather the Tex19.1, Dnmt3L, Miwi2 and Mili mutants may all cause defects in meiotic chromosome structure that lead to both retrotransposon de-repression and defective chromosome synapsis. Clearly further work is needed to clarify the molecular mechanism underlying the chromosome synapsis defect in the Tex19.1 mutant mice presented here, and in the Dnmt3L, Miwi2 and Mili mutant mice [11]–[13]. However, this study provides further evidence demonstrating a correlation between de-repression of retrotransposons and impaired chromosome synapsis during mouse meiosis.
Although there are gross similarities between the Tex19.1 mutant phenotype and the Dnmt3L, Miwi2 or Mili mutant phenotypes, there are also important differences. Dnmt3L, Miwi2 and Mili are all required to repress LINE and IAP retrotransposons in the germline, and these three genes appear to converge on DNA methylation and transcriptional repression of these sequences in the genome [11]–[14]. However, repression of LINE and IAP retrotransposons is not perturbed in Tex19.1−/− knockout testes suggesting that Tex19.1 is not involved in the transcriptional repression of LINE or IAP elements. Rather our data shows that transcripts from the MMERVK10C class of endogenous retroviruses accumulate in the germ cells in the absence of Tex19.1. These differences between the Tex19.1 mutant phenotype and the Dnmt3L, Miwi2 and Mili mutant phenotypes may reflect the existence of multiple mechanisms with different specificities to repress retrotransposons in the germline.
The Tex19.1 mutant phenotype is characterised by the accumulation of MMERVK10C retrotransposon transcripts, but the molecular basis for this phenotype is not yet clear. The upregulation of MMERVK10C transcripts could be caused by changes acting at any level of gene expression from the initiation of transcription to mRNA turnover. We have not been able to find any difference in the level of DNA methylation at MMERVK10C elements in Tex19.1 mutant testes. This provides further evidence that Tex19.1 belongs to a different genetic pathway than Miwi2, Mili and Dnmt3L for repression of retrotransposons in the germline. However, we cannot exclude the possibility that DNA methylation may be altered in a subset of MMERVK10C elements in a subset of germ cells in the Tex19.1 mutant testes, and that this subset of elements is responsible for the upregulation of MMERVK10C transcripts that we describe in the Tex19.1 mutant testes. An alternative model to explain the upregulation of MMERVK10C elements in Tex19.1 mutant testes is that Tex19.1 could be a transcriptional repressor of MMERVK10C elements. The nuclear localisation of Tex19.1 reported by Kuntz et al. [17] would be consistent with this type of mechanism operating. However, although we cannot exclude the possibility that some Tex19.1 acts in the nucleus in the germ cells in the adult testes, our finding that Tex19.1 is predominantly cytoplasmic in these cells would be more consistent with Tex19.1 acting to regulate gene expression at a post-transcriptional level. We are able to detect MMERVK10C transcripts in wild-type testes (Figure 6B,G) suggesting that some MMERVK10C transcripts must escape DNA methylation or transcriptional repression, and that post-transcriptional regulation of MMERVK10C mRNA may play a role in repressing the activity of this retrotransposon. The upregulation of MMERVK10C transcripts in Tex19.1 mutant testes does not appear to be the result of changes in RNA splicing as the MMERVK10C isoforms present in Tex19.1 mutant testes do not appear to be qualitatively different from those present in wild-type testes. However, the accumulation of MMERVK10C transcripts in Tex19.1 knockout testes would be consistent with Tex19.1 promoting degradation of MMERVK10C mRNA. Investigation into the biochemical function of Tex19.1 should provide a ready test of these models and generate some insight into the molecular mechanism of Tex19.1-dependent repression of MMERVK10C endogenous retroviruses.
Repression of retrotranposons in the mammalian germline requires mechanisms to distinguish retrotransposons from endogenous genes to allow repression to be targeted to the correct loci. piRNAs, a group of small RNAs that are physically associated with the piwi class of proteins, are abundant in male germ cells and some piRNAs have sequence homology to various classes of retrotransposon [14], [44]–[46]. The sequence homology between some piRNA molecules and retrotransposons is presumably used to target DNA methylation to retrotransposons rather than endogenous genes. Although there is good genetic evidence that the piwi class of proteins is involved in transcriptional repression of retrotransposons [12]–[14], there is also good biochemical evidence that piwi proteins and piRNAs are physically associated with the translational machinery in male germ cells [46],[47], suggesting a role in translation or mRNA turnover. Thus piRNA-mediated repression of retrotransposons may be working at multiple levels of gene expression in male germ cells. It will be informative to investigate whether the Tex19.1 pathway for repression of retrotransposons that we describe here also utilises piRNAs to target repression to MMERVK10C elements.
One of the interesting aspects of the Tex19.1 phenotype is that although the MMERVK10C subclass of retrotransposons is upregulated in Tex19.1 mutant testes, LINE, SINE and IAP retrotransposons are not. It is not clear how Tex19.1 determines specificity for the MMERVK10C element. Notably, IAP elements belong to the same subclass of endogenous retroviruses as MMERVK10C elements (class II LTR retrotransposons) but are not upregulated in Tex19.1 mutant testes. Sequences within the MMERVK10C promoter or transcript could be involved in targeting Tex19.1 activity to this element. Alternatively Tex19.1 may have the potential to regulate a wider range of retrotransposons than we have been able to identify here, but alternative mechanisms to repress retrotransposon expression during spermatogenesis, such as DNA methylation, may limit the phenotypic effects of losing Tex19.1 to a subset of its potential targets. Furthermore, as Tex19.1 expression is not restricted to spermatogenesis but also occurs in primordial germ cells, oocytes and pluripotent stem cells, it will be of interest to determine if Tex19.1 is involved in repressing MMERVK10C elements and other classes of retrotransposons in these cell types.
In addition to its role in the germline, Tex19.1 is also expressed in pluripotent cells. Like germ cells, pluripotent cells are viable targets for retrotransposon activity as any new transposition events could be propagated through successive generations. Therefore pluripotent cells presumably also need to modulate retrotransposon activity to ensure that the mutational load on the genome is not too high. Our finding that Tex19.1−/− homozygotes are born at a sub-Mendelian frequency is consistent with a role for Tex19.1 in pluripotent cells in early embryonic development. Further work is required to determine whether the loss of Tex19.1−/− homozygotes during embryogenesis is caused by defects in pluripotent cells, and whether pluripotent cells upregulate retrotransposon expression in Tex19.1−/− knockout embryos.
The ongoing battle between retrotransposons and the host genome has important consequences for evolution, and for genetic disease. Retrotransposons that can successfully evade genome defences in germ cells and pluripotent cells will be selected for during evolution, whereas germ cells and pluripotent cells are under selective pressure to keep the mutational load on the genome at sustainable levels. The striking differences in the relative abundance of different classes of retrotransposable elements between the mouse and human genomes suggest that this conflict is ongoing during mammalian evolution [9]. Although low levels of mutation and retrotransposition in the germline are required to generate the genetic variation essential for evolution, high levels of mutation or retrotransposition are deleterious to the survival of a species. In humans, endogenous retroviruses with intact coding sequences comprise a very small proportion of the genome [48], yet intact endogenous retroviral particles are found in human pluripotent stem cells, and in testicular germ cell tumours where the expression of endogenous retroviral proteins has been suggested to contribute to tumourigenesis [39],[43],[49]. Furthermore, a number of human genetic diseases are associated with de novo mutagenic retrotransposition events that disrupt the function of endogenous human genes [50],[51]. Our data suggests that Tex19.1 is part of a mechanism that protects the genome from the deleterious effects of retrotransposon activity in the germline, and thereby helps to maintain genomic stability through successive generations.
|
10.1371/journal.pcbi.1006652 | Multiscale computational model of Achilles tendon wound healing: Untangling the effects of repair and loading | Mechanical stimulation of the healing tendon is thought to regulate scar anisotropy and strength and is relatively easy to modulate through physical therapy. However, in vivo studies of various loading protocols in animal models have produced mixed results. To integrate and better understand the available data, we developed a multiscale model of rat Achilles tendon healing that incorporates the effect of changes in the mechanical environment on fibroblast behavior, collagen deposition, and scar formation. We modified an OpenSim model of the rat right hindlimb to estimate physiologic strains in the lateral/medial gastrocnemius and soleus musculo-tendon units during loading and unloading conditions. We used the tendon strains as inputs to a thermodynamic model of stress fiber dynamics that predicts fibroblast alignment, and to determine local collagen synthesis rates according to a response curve derived from in vitro studies. We then used an agent-based model (ABM) of scar formation to integrate these cell-level responses and predict tissue-level collagen alignment and content. We compared our model predictions to experimental data from ten different studies. We found that a single set of cellular response curves can explain features of observed tendon healing across a wide array of reported experiments in rats–including the paradoxical finding that repairing transected tendon reverses the effect of loading on alignment–without fitting model parameters to any data from those experiments. The key to these successful predictions was simulating the specific loading and surgical protocols to predict tissue-level strains, which then guided cellular behaviors according to response curves based on in vitro experiments. Our model results provide a potential explanation for the highly variable responses to mechanical loading reported in the tendon healing literature and may be useful in guiding the design of future experiments and interventions.
| Tendons and ligaments transmit force between muscles and bones throughout the body and are comprised of highly aligned collagen fibers that help bear high loads. The Achilles tendon is exposed to exceptionally high loads and is prone to rupture. When damaged Achilles tendons heal, they typically have reduced strength and stiffness, and while most believe that appropriate physical therapy can help improve these mechanical properties, both clinical and animal studies of mechanical loading following injury have produced highly variable and somewhat disappointing results. To help better understand the effects of mechanical loading on tendon healing and potentially guide future therapies, we developed a computational model of rat Achilles tendon healing and showed that we could predict the main effects of different mechanical loading and surgical repair conditions reported across a wide range of published studies. Our model offers potential explanations for some surprising findings of prior studies and for the high variability observed in those studies and may prove useful in designing future therapies or experiments to test new therapies.
| Many mechanically loaded tissues including skin, tendon, ligament, and heart respond to injury by forming a collagen-rich scar. Normally, tendons are comprised of highly aligned collagen fibers that help transmit forces between muscles and bones throughout the body and bear high loads. The Achilles tendon in particular can be exposed to loads up to 70 MPa, compared to 30 MPa in most other tendons [1]. These high loads often lead to injury, with Achilles tendon ruptures accounting for up to 45% of all tendon ruptures [2] and afflicting up to 2.5 million annually [3,4]. Many who suffer from an Achilles tendon rupture never regain complete function, especially because healing tendons form scar with reduced collagen fiber organization and stiffness compared to uninjured tendons [5].
While there seems to be general agreement that mechanical stimulation of the healing tendon, such as during physical therapy, influences scar mechanical properties [6], current treatments for patients with an Achilles tendon rupture have produced variable results [1,3,7–9]. To better understand the impact of loading, rat animal models of Achilles tendon rupture have been utilized so that scar tissue can be excised to determine quantitative biomechanical scar properties. Unfortunately, these studies have also led to a wide variety of results, with mechanical loading sometimes appearing to increase, but at other times appearing to decrease, tendon properties such as stiffness or rupture strength. One of the potential explanations for this variability could be due to differences in mechanics during healing, which could be altered through unloading (e.g. cast immobilization) or loading (e.g. free cage activity) of the tendon [10–14] as well as through the choice to repair the tendon with suture versus allowing natural healing without repair [12,15]. In this work, we developed a multiscale computational model of the healing rat Achilles tendon that integrates information about how local mechanics influences cellular alignment and collagen remodeling to predict the effects of various repair and loading protocols on tendon structure. We found that the multiscale model predicted the major observed trends in the evolution of tissue-level scar properties across a wide variety of published rat studies. Furthermore, the model simulations identified a potential mechanism underlying the apparently paradoxical finding that mechanical loading enhances collagen alignment in unrepaired Achilles tendons yet decreases it in repaired tendons.
Our multiscale computational model simulated mechanics and associated responses at multiple scales. At the organ level, we simulated various healing and loading conditions in a musculoskeletal model of the right rat hindlimb [16] implemented in OpenSim [17] to estimate associated strains in the tendon (Fig 1A). At the cellular level, these strains affected cellular behavior in two ways (Fig 1B and 1C). First, cyclic strains determined cell alignment according to a thermodynamic model of stress fiber dynamics developed and validated against in vitro experiments by our group [18] (Fig 1B). Second, mean strains modulated fibroblast collagen synthesis according to a relationship fitted to data from multiple published studies (Fig 1C) [19–22]. At the tissue level, we used an agent-based model (ABM) of wound healing similar to one published previously by our group [23] to integrate these cellular behaviors and predict the evolving collagen structure (Fig 1D).
We searched the literature for studies of transected Achilles tendons in rats, the most common animal model used to mimic an Achilles tendon rupture. These studies used a variety of time courses and mechanical loading protocols to treat the rats during the wound healing process. First, we focused on studies that tested the effects of natural healing of unrepaired, transected tendons and imposed unloading by either botulinum toxin (Botox) injection into the gastrocnemius muscle or tail suspension [10,24–26] or loading by allowing the rat to freely walk around its cage [10,13,27,28] (Fig 2A). Because no studies quantified both collagen alignment and total collagen mass, we then narrowed our search to include only studies that reported at least one of two quantitative measures that could act as surrogates for model comparisons. Based on the strong reported correlation between measured collagen alignment and intrinsic material properties of the tissue [6,14], we compared measured values of Young’s modulus (Ey in MPa) to the levels of collagen alignment predicted by our model. Similarly, based on previous studies showing that collagen concentration and tendon CSA rise in parallel [24,29,30], as well as the fact that tissue mass increases with tissue volume, we compared measured values of tendon cross-sectional area (CSA in mm2) to the total collagen content predicted by the model. Despite considerable variability among the reported values of these metrics, two clear trends were apparent in the data. First, Ey started near 0 MPa at 3 days and rose to about 20–40 MPa at 14 days in both unloaded and loaded conditions (Fig 2B and 2C). Second, CSA values in unloaded conditions remained at around 5 mm2 (Fig 2E), while loading increased CSA over time to around 10–20 mm2 at day 14 (Fig 2F).
Next, we used the model of the rat hindlimb implemented in OpenSim [16] to simulate these unloading and loading conditions and estimate strains in the healing region. We simulated unloading by fixing all joint ankles in plantar flexion, with minimum muscle activation and a 7mm gap distance between the tendon stumps [10,25,31], resulting in a predicted constant cell strain of E11 = 0.002 in the primary loading direction (Fig 3A). We simulated loading by prescribing joint angles and muscle activation corresponding to the rat gait cycle. Predicted strains oscillated between .009 and .043, reflecting the swing/stance phases of the gait cycle and yielding a cyclic strain amplitude of 0.034 and a mean strain of 0.030 (Fig 3A). Using our published model of stress fiber dynamics [18], both strain states produced cell alignment along the loading axis, with slightly lower predicted alignment for the loaded case (Fig 3B). Using the collagen synthesis curve we fitted to published data [32–35], mean strain from the loaded case was associated with about 2x greater collagen production than the mean strain from the unloaded case (Fig 3C).
Integrating these two cellular behaviors in the ABM component, the multiscale model predicted that the specific loading protocols we simulated should produce little difference in collagen alignment (Fig 2D) but a substantial difference in collagen content (Fig 2G). Trends in predicted alignment agreed with literature reports showing similar Ey for both groups at all time points (Fig 2B and 2C). Furthermore, differences in model-predicted collagen accumulation in the two loading states qualitatively matched reported differences of the tendon CSA in loaded (Fig 2F) compared to unloaded (Fig 2E) conditions.
In our next set of simulations, we explored how surgically repairing rat Achilles tendons would alter our predictions and the response to loading during healing (Fig 4A). We selected a set of studies that subjected experimental groups to either unloading with cast immobilization or loading with an exercise protocol, defined as 60 min/day treadmill exercise + free cage activity for the rest of the time, compared surgically repaired and unrepaired groups, and reported Young’s modulus or cross-sectional area (Fig 4B and 4D) [12,13,36]. For the unrepaired groups in this second set of studies, trends were similar to those shown in Fig 2, with loading enhancing CSA but not Ey. Similar to the unrepaired groups, loading of suture-repaired tendons increased tendon CSA compared to unloading (Fig 4D). However, loading in the suture-repaired groups surprisingly reduced Young’s modulus below the values observed in any other group by the 42-day final time point (Fig 4B).
Assuming that surgical repair eliminated the gap distance between the tendon stumps, our hindlimb simulations in OpenSim predicted higher strains for repaired conditions compared to unrepaired, with repaired-unloaded tendons experiencing a static strain of E11 = 0.012 and repaired-loaded tendons oscillating between 0.011 and 0.075 (cyclic strain of 0.064, mean strain of 0.049; Fig 5A). The much larger cyclic strains in the repaired-loaded group induced stress fiber disassembly along the loading axis in the cell alignment model, resulting in cells that were nearly randomly oriented (Fig 5B); the differences in predicted stress fiber orientation distributions for these four cases are shown in Fig 6. On the other hand, the large mean strains in the repaired-loaded group led to a higher rate of collagen synthesis in this group compared to the other conditions simulated (Fig 5C). Integrating these predictions using the ABM component, overall our multiscale model correctly matched the apparently paradoxical reports that repaired, loaded tendons have the lowest collagen alignment of any of these four conditions (Fig 4B and 4C) despite having the highest collagen content (Fig 4D and 4E). In other words, a single set of cellular response curves can explain features of observed tendon healing across a wide array of reported experiments in rats without fitting model parameters to any data from those experiments. Rather, the key to the predictions in our multiscale model is simulating the specific loading and surgical protocols to predict tissue-level strains, which then guide cellular behaviors according to response curves based on in vitro experiments.
Many of the studies we simulated here loaded healing tendons through unrestricted cage activity or through daily exercise added to normal cage activity. In these protocols, tendons are cyclically loaded in short bursts as the animals move about their cages, interspersed with short rest periods when they are standing still and longer rest periods when they sleep. We chose to model this situation by imposing cyclic loading at 1 Hz with a 1-hour on, 1-hour off protocol for 12 hours, followed by 12 hours of rest, and then repeating. To understand the effect of this choice on our results, we used the cell alignment model to simulate the effect of stretching with different protocols that all produced the same time-averaged frequency of 0.5Hz (Fig 7A). We compared stretching for repetitions of 6 hours at 1Hz followed by 6 hours of rest (0Hz), repetitions of 1 hour at 1Hz followed by 1 hour at rest, and continuous stretching at 0.5Hz. We simulated each repetition until steady state was reached, defined as the point when the difference in alignment at the end of two consecutive repetitions was less than 0.01 (Fig 7B). We found that the different stretching protocols produced different relationships between strain amplitude and predicted steady-state cell alignment, defined as the average order parameter over the last simulated cycle. At low strain amplitudes, all the protocols produced similar results, while at peak strains between 0.06 and 0.10 the three protocols resulted in predictions ranging from fairly strong alignment parallel to stretch (continuous stretching) to a modest degree of perpendicular alignment (1h on / 1h off and 6h on / 6h off). These differences arise from two features of the stress fiber model. First, stress fiber disassembly is triggered by high negative strain rates in the model, so much higher strains are required to influence alignment when loading is imposed at a lower frequency (Fig 7A). Second, because disassembly of stress fibers occurs on a much faster time scale than assembly, shorter durations of cyclic stretching can drive alignment down quickly, while much longer rest periods are needed to recover from each loading cycle (Fig 7B).
Our simulations of intermittent loading protocols raise the possibility that relatively minor differences in loading could have significant implications for fibroblast and collagen alignment. This could be an intriguing explanation for the surprising degree of variability we found in the literature among studies that employed apparently identical experimental protocols and outcome metrics (see Fig 2B, 2C and 2F), or even within groups in individual studies (see error bars in Fig 4B and 4D). To date, most published experiments have not tracked movement of rats with the detail required to simulate more realistic or even animal-specific loading protocols. Sams-Dodd observed that healthy rats traveled around 5700 cm within a 10 minute observation period in one of the only studies we could find that attempted to track travel distances [37]. Our simulations suggest that employing continuous movement tracking in future studies might provide additional insight into whether individual variability in activity and loading can explain some of the observed variability in tendon healing.
The most serious limitation of the modeling studies reported here is that we were only able to validate them through qualitative comparisons of model-predicted trends to experimentally measured surrogates. The most novel prediction of our model was the degree of collagen fiber alignment. Experimentally, collagen alignment can be directly measured from ultrasound [4,12] or polarized microscopy [38]. However, these measurements were reported in so few studies that we were forced to use a more commonly reported surrogate, the Young’s modulus (Ey), which has been shown to correlate with the degree of alignment. It is conceivable that other factors such as collagen density that might differ between the cases simulated here might have influenced Ey. Future studies may be able to draw on relationships between alignment and modulus such as those reported by Lake et al. [39] and Li et al. [40] to quantitatively estimate Ey. The comparison of total collagen content to tendon cross-sectional area (CSA) could also have limitations, since experimentally the rate of collagen production influences both collagen density (as assessed by biochemical assays or picrosirius red staining) and total CSA, but few studies provide data on both simultaneously. In addition, our collagen content predictions are in arbitrary units, since the studies we used to determine the effect of stretch on collagen synthesis reported relative changes rather than absolute synthesis rates.
The simplifications we made in the rat hindlimb model could also introduce some errors in the tendon-level strains we calculated. First, we included only the musculo-tendon units that directly comprise the Achilles tendon. Second, due to a lack of information in the literature about hip and knee motion during walking in the setting of rat Achilles tendon injury, we combined healthy hip and knee with injured ankle motion [41,42]. However, we expect errors introduced by this choice to be small, since the gastrocnemius and soleus muscles do not cross the hip, the soleus does not cross the knee, and the moment arms around the ankle are larger than at the knee for the gastrocnemius. Furthermore, as long as simulated strains from repaired tendons are higher than unrepaired and strains from loading are higher than from unloading, the overall trends predicted from our model should be robust to small changes in the exact magnitude of the predicted tendon strains.
Our agent-based model assumes that cells deposit collagen aligned with their own axis, and this assumption was critical to translating cell alignment predictions from the cytoskeletal model into tissue-level predictions of collagen structure. While the exact mechanisms by which fibroblasts deposit and orient collagen in vivo are still being debated, the general idea that collagen ends up locally aligned with the fibroblasts that deposit it remains strongly supported in the literature [43–45]. Furthermore, we have previously shown that agent-based models incorporating this same assumption correctly predict a range of scar structures observed following myocardial infarction in various animal models under different mechanical conditions [23]. We also considered and simulated several other alternative methods of determining collagen alignment in the model but found that none could predict all the observed trends apparent in the data reviewed here. For instance, several experiments have theorized that surrounding collagen fibers could “structurally constrain” the formation of new fibers in vivo [46–48]. Others have demonstrated strain-dependent modulation of collagen degradation that could influence overall alignment under uniaxial loading by selectively degrading fibers with certain orientations faster than others [49–51]. While all these effects could be present within the actual tendon, we found that strain-dependent cell alignment, deposition of collagen aligned with the cells, and strain-dependent changes in collagen synthesis rate were sufficient for capturing the major trends in the data as outlined above.
In this study, we used multiscale modeling to integrate information from the literature on fibroblast responses to stretch (alignment and collagen synthesis), scar formation following injury (collagen deposition and other features of the agent-based model), and musculoskeletal mechanics (rat hindlimb model implemented in OpenSim) to interpret apparently conflicting data from a range of experimental studies. We found that our computational model could reproduce several key features of observed tendon healing across a wide array of reported experiments in rats–including the paradoxical finding that repairing transected tendon reverses the effect of loading on alignment–without fitting model parameters to any data from those experiments. Rather, the key to the predictions in our multiscale model was simulating the specific loading and surgical protocols to predict tissue-level strains, which then guided cellular behaviors according to response curves based on in vitro experiments. These results suggest that the apparently conflicting data in the studies we reviewed may in fact reflect consistent biologic responses to local strains in the healing tendon, providing a new conceptual framework for interpreting existing data and devising potential therapies for Achilles tendon rupture.
We searched PubMed and Google Scholar for all papers that included the keywords “rat Achilles tendon rupture injury” in the title or abstract and used a scalpel to perform a full transection of both the Achilles tendon and the plantaris tendon, a tendon parallel to the Achilles that is proportionally larger in rats than in humans and can act as an “internal splint” [52]. From these, we next identified studies that specifically compared the effects of unloading from intramuscular injection of Botox, tail suspension, or cast immobilization against loading due to free cage activity or treadmill exercise. From these 30 identified papers, we then selected those that quantitatively measured the cross sectional area and/or the Young’s modulus of the healing scar, leaving us with 10 studies that met our search criteria [10–14,24–26,36,53]. We used cross sectional area (CSA) of the tendon as a surrogate measure for collagen production, based on previous studies showing that collagen concentration and tendon CSA rise in parallel [24,29,30], as well as the fact that tissue mass increases with tissue volume. Similarly, we used the Young’s modulus as a surrogate measure for collagen alignment based on the strong reported correlation between measured alignment and intrinsic material properties of the tissue [6,14], and the fact that Young’s modulus was commonly measured in the studies we found while collagen alignment was not.
We adapted a previously published rat right hindlimb model to conduct our simulations (Fig 1A) [16] (https://simtk.org/projects/rat_hlimb_model). Briefly, the model used anatomically accurate representations of the bones (spine, hip, femur, tibia, and foot), joints, and musculo-tendon units (each represented by one line segment from its origin to insertion) of the rat right hindlimb. Musculo-tendon units were represented as linear elements consisting of a muscle segment in series with a passive tendon segment. The muscle segment consisted of its own passive element in parallel with an active contractile element that generated force depending on a Hill-type model. The mechanical properties of muscle fiber and tendon were defined using fiber force-length, fiber force-velocity, and tendon force-strain curves determined by Millard et al. [54]. Muscle segments could be prescribed activation levels varying from 0 (rest) to 1 (full activation). Since muscles not attached to the Achilles tendon were irrelevant for this simulation, we simplified the model by only including the lateral and medial gastrocnemius and soleus muscles, the three musculo-tendon units that comprise the Achilles tendon. To predict Achilles tendon strains for the various cases simulated here, we prescribed experimentally measured joint angle profiles, muscle activation curves, and tendon mechanical properties as inputs, and obtained strain vs. time curves from forward dynamic simulations using custom MATLAB routines employed by two of the authors in a previous publication [55].
We used a computational model of stress fiber remodeling published previously by our group to estimate cell alignment behavior [18] (Fig 1B). The model represents the thermodynamics of stress fiber (SF) assembly and disassembly, capturing features such as the ability of tension to promote assembly by altering the free energy of bound actin subunits. Two specific features of the model are important for the predictions shown in this manuscript. First, on the time scale of individual loading cycles, large negative strain rates reduce stress fiber tension through the force-velocity behavior of myosin, promoting SF disassembly in the direction of stretch and net SF (and cell) orientation perpendicular to that stretch. On a longer time scale, the model assumes that cells can remodel the extracellular matrix and/or their attachments to the ECM to attain an equilibrium strain state that minimizes the sum of the energies associated with elastic stretch and the chemical potential of the bound and unbound actin subunits. This aspect of the model drives the response to mean boundary conditions, upon which cyclic responses are then superimposed. Here, we used the average and spread of predicted SF distributions as surrogate measures of cell alignment, the same approach we used in the original model validation against in vitro data. For a more complete description of the model and details on its validation, please see Chen et al. 2018 [18].
We adapted an agent based model (ABM) originally published by Rouillard and Holmes [23,62] for infarct healing to integrate cellular-level responses and predict the evolving tendon scar structure (Fig 1D). Table 1 lists all model parameters altered for these simulations, while Fig 9 shows a flowchart of a cell’s decision tree within the model. Rouillard and Holmes modeled fibroblasts as circular discs free to move in a square, two-dimensional space divided into 10-micron-square patches. Each patch contained information about the local collagen alignment and density. Fibroblasts could migrate, proliferate, undergo apoptosis, and remodel collagen. Fibroblast orientation guided fibroblast migration direction and deposition of collagen, and existing collagen fibers were degraded at a rate proportional to their local concentration. A local chemokine concentration gradient with a high concentration of chemokines within the wound area and a low chemokine concentration in the healthy tissue drove cell migration into the wound (Fig 10A). We made several modifications to this model to adapt it for Achilles tendon healing. We simulated healing of a rectangular wound area after complete transection or rupture (Fig 10B). The initial wound size, matrix structure, and collagen content within the wound depended on the healing condition (Table 1). An unrepaired transected tendon contained a low amount (0.1%) of randomly aligned collagen fibers, mimicking the randomly aligned provisional matrix in the wound area, while a suture-repaired tendon contained some aligned collagen (0.9%, alignment order parameter of 0.4), mimicking aligned collagen fibers from the healthy tendon stumps. Fibroblasts migrated into the wound space (Fig 10B, 10C and 10D) from the two opposing sides adjacent to healthy tissue. Rouillard determined cell alignment from a series of phenomenologic equations that represented the alignment response to stretch, contact guidance from surrounding collagen, and chemokine gradients, as well as their integration. Here, we replaced the original phenomenologic relationship governing stretch-induced alignment with the cell alignment predicted from the stress fiber model, and employed a more recent equation for integrating across the alignment cues published by Richardson et al. [63].
Based on previous studies showing that cells upregulate collagen production after exposure to both static [64] and cyclic stretch [32–35], cells synthesized collagen according to their mean strains, with higher strains corresponding to higher collagen synthesis rates. The mean strain was calculated by taking an average over one gait cycle period (1 second). In unloaded cases, this mean strain matched the static strain values (Fig 5). We determined the collagen synthesis amounts according to a sigmoidal curve fitted to data from four independent experiments in the literature (Fig 1C) [32–35]:
CollagenSynthesis(εm)=1.31+exp(−150(εm2−0.02))+0.6,
(2)
where εm is the mean strain felt by the cell. Cells deposited collagen aligned to their major axis of alignment. While the exact mechanisms by which fibroblasts deposit and orient collagen in vivo are still being debated, the general idea that collagen ends up locally aligned with the fibroblasts that deposit it remains strongly supported in the literature [43–45]. At each time point for which collagen content and orientation are reported, the local collagen content and orientation histograms from each collagen patch were averaged to determine a single area fraction, mean angle, and order parameter for the entire scar region.
|
10.1371/journal.pgen.1008084 | Meiotic gatekeeper STRA8 suppresses autophagy by repressing Nr1d1 expression during spermatogenesis in mice | The transition from mitotic to meiotic cell cycles is essential for haploid gamete formation and fertility. Stimulated by retinoic acid gene 8 (Stra8) is an essential gatekeeper of meiotic initiation in vertebrates; yet, the molecular role of STRA8 remains principally unknown. Here we demonstrate that STRA8 functions as a suppressor of autophagy during spermatogenesis in mice. Stra8-deficient germ cells fail to enter meiosis and present aberrant upregulation of autophagy-lysosome genes, commensurate with autophagy activation. Biochemical assays show that ectopic expression of STRA8 alone is sufficient to inhibit both autophagy induction and maturation. Studies also revealed that, Nr1d1, a nuclear hormone receptor gene, is upregulated in Stra8-deficient testes and that STRA8 binds to the Nr1d1 promoter, indicating that Nr1d1 is a direct target of STRA8 transcriptional repression. In addition, it was found that NR1D1 binds to the promoter of Ulk1, a gene essential for autophagy initiation, and that Nr1d1 is required for the upregulated Ulk1 expression in Stra8-deficient testes. Furthermore, both genetic deletion of Nr1d1 and pharmacologic inhibition of NR1D1 by its synthetic antagonist SR8278 exhibit rescuing effects on the meiotic initiation defects observed in Stra8-deficient male germ cells. Together, the data suggest a novel link between STRA8-mediated autophagy suppression and meiotic initiation.
| Meiotic initiation is a key feature of sexual reproduction that launches an intricate chromosomal program involving DNA double strand breaks (DSBs), homolog pairing, cohesion, synapsis, and recombination. Vertebrate gene Stra8 is an essential gatekeeper of meiotic initiation. However, the molecular role of STRA8 and its target genes remain elusive. Using mouse spermatogenesis as a model, we report that STRA8 suppresses autophagy by repressing the transcription of a nuclear hormone receptor gene Nr1d1, and in turn, silencing the expression of Ulk1, a gene essential for autophagy initiation. Given that autophagy is critical for protein and cellular organelle recycling and for preventing genomic instability, our study suggests that this newly demonstrated function of STRA8, as a suppressor of autophagy, may be an important mechanistic feature of its role in meiotic initiation.
| Meiosis is a fundamental process in sexual reproduction during which diploid cells halve their chromosome number by two rounds of cell divisions to generate haploid cells or gametes. In mammals, temporal regulation of meiosis in germ cells is sex-specific: meiosis in females begins during embryogenesis, whereas meiosis in males starts at puberty and persists throughout adulthood [1]. To enter meiosis, diploid cells must cease mitosis and then undergo one round of DNA replication, followed by the formation of DNA DSBs, meiotic chromosome pairing, cohesion, synapsis, and recombination. Although meiotic initiation has been extensively studied in some model organisms, such as yeast, flies, and worms, the mechanisms governing meiotic initiation in mammals remain elusive, largely because the molecular machinery controlling this process differs among species [2].
To date, the best characterized gatekeeper of meiotic initiation in vertebrates is stimulated by retinoic acid gene 8 (Stra8) [3]. STRA8 is thought to act as a basic helix-loop-helix (bHLH) transcription factor, based on its DNA binding and transcriptional activity [4]. Interestingly, Stra8 is expressed in a precise tissue-specific and developmental manner, whereby it is transitorily expressed only in premeiotic germ cells, of both sexes, shortly before their entry into meiosis [5, 6]. Functionally, Stra8 likely governs both meiotic initiation and early meiotic progression. In one study, Stra8-deficient germ cells (exons 2–7 deleted) of both sexes do not display molecular hallmarks of meiotic initiation (e.g., DSBs) and fail to enter meiosis in juvenile mouse testes and in developing ovaries [7, 8]. Another study that used a different Stra8-deficient mouse line, in which exons 2–4 were deleted on an F1 hybrid (C57BL/6 x 129) background showed that Stra8 functions instead in early meiotic prophase in spermatogenesis [9]. Nevertheless, Stra8-deficient germ cells of both sexes undergo meiotic arrest [7–9]. Thus, STRA8 is an essential regulator of meiosis that likely acts as a transcription factor, but currently there is little information on its molecular role and functional targets implicated in meiosis [10].
In this study, we report the unexpected finding that STRA8 acts as a suppressor of autophagy. Autophagy is a catabolic process involving self-digestion of protein and cellular organelles through lysosomes [11] and a major stress response pathway that promotes cellular survival by supplying nutrients and energy. In addition, autophagy plays a critical role in maintaining protein and cellular organelle quality control [12]. In recent decades, studies have identified genes that encode essential molecular factors of the autophagy pathway [11], including regulators of autophagy initiation (VPS34/PIK3C3, ULK1), nucleation (BECN1), elongation (ATG5-ATG12, ATG7, ATG16, and LC3B), and maturation through lysosome biogenesis (VPS18 and LAMP2). Moreover, TFEB has been identified as a master regulator of autophagy by inducing a broad spectrum of autophagy-lysosome genes [13]. Interestingly, lack of autophagy has been shown to instigate DNA damage events, including DNA DSBs, in somatic cells [14–16] (reviewed in ref. [17, 18]). The reported findings support a mechanism whereby STRA8 suppresses germ cell autophagy via direct transcriptional inhibition of a second transcription factor, NR1D1, which is needed for expression of the essential autophagy initiator ULK1. The data show that loss of Nr1d1 expression or inhibition of NR1D1 function by its synthetic antagonist SR8278 exhibited rescuing effects on the meiotic initiation block observed in Stra8-deficient male germ cells. Together, our results suggest that STRA8-mediated autophagy suppression is a mechanistic feature of its role in meiotic initiation.
To determine the molecular mechanism of STRA8-driven meiotic initiation, we used Stra8-deficient mice on a highly inbred C57BL/6 background [8]. We examined Stra8-deficient testes at 15–21 days postpartum (d.p.p.) by transmission electron microscopy. At this age, wild-type testes exhibited a normal presence of spermatogonia and meiotic spermatocytes (Fig 1A, panels a and b), whereas Stra8-deficient (Stra8-/-) testes exhibited a complete lack of meiotic spermatocytes despite the presence of spermatogonia (Fig 1B, panels a-c). Importantly, transmission electron microscopy revealed two novel characteristics in Stra8-deficient testes. First, the cellular integrity in the adluminal compartment of the seminiferous tubules appeared highly disrupted (Fig 1B, panel b), with the degenerative cells in this region having lost membrane integrity and erupted their cellular contents into the lumen (Fig 1B, panel d). Second, we frequently observed autophagosome structures (38 autophagosome structures observed in 200 germ cells from 2 Stra8-deficient testes), whereas in wild-type testes comparable autophagosome structures were not observed (0 autophagosome structures in 378 germ cells examined from 2 wild-type testes) (Fig 1A, panel b and c). Comparable autophagosome structures were not observed in Sertoli cells from both wild-type (Fig 1A, panel e) or Stra8-deficient testes (Fig 1B, panel c and e). This is consistent with a recent report that autophagy was not detected in spermatogonia, early spermatocytes and Sertoli cells in rat testes [19]. The autophagosome structures observed in Stra8-deficient germ cells were located in the cytoplasmic region of both non-degenerative (cells with intact membrane) (Fig 1B, panels f-i) and degenerative cells (Fig 1B, panel j). These double-membraned autophagosomes enclosed cellular organelles, including mitochondria (Fig 1B, panels f—i) and endoplasmic reticulum (Fig 1B, panel j). Together, our transmission electron microscopy study revealed aberrant autophagosome formation in Stra8-deficient germ cells.
During autophagy, autophagosomes serve as intermediate transport vesicles to target cellular components for degradation before their conversion into autolysosomes. Thus, aberrant appearance of autophagosomes in Stra8-deficient testes could result from either elevated autophagosome formation or impaired autophagosome turnover through fusion with lysosomes [20]. To distinguish between these possibilities, we examined autophagy activity in Stra8-deficient testes by using tandem fluorescent-tagged LC3 (RFP-GFP-LC3) by breeding RFP-GFP-LC3 transgenic allele into wild-type and Stra8-deficient backgrounds [21]. LC3 is a soluble protein and is distributed ubiquitously in cells. Upon autophagy activation, the cytosolic form of LC3 (LC3-I) is conjugated to phosphatidylethanolamine (PE) to form LC3-II, which is recruited to autophagosomal membranes, thereby serving as a well-characterized marker for autophagosomes [22]. In this reporter system, autophagosome vesicles (GFP-positive and RFP-positive puncta) can be distinguished from acidified autolysosome vesicles (GFP-negative and RFP-positive puncta) due to acidic quenching of the GFP signal, but not the RFP signal, after fusion with lysosomes. Whole-mount immunofluorescence imaging of seminiferous tubules from wild-type testes showed occasional merged GFP-positive and RFP-positive puncta, indicative of autophagosomes (Fig 2A). In contrast, seminiferous tubules from Stra8-deficient testes exhibited a significant increase in the number of both total vesicles (RFP-positive puncta) and autolysosome vesicles (GFP-negative and RFP-positive puncta) (Fig 2A). Direct fluorescence imaging of testicular cross sections confirmed a profound increase of autophagy in 98% of Stra8-deficient seminiferous tubules (Fig 2B). In contrast, 100% of the seminiferous tubules in wild-type testes exhibited diffuse GFP signal, suggesting cytoplasmic soluble LC3 and a lack of autophagosome formation (Fig 2B, lower left panels). To further confirm that the lack of autophagosomes in wild-type seminiferous tubules did not result from rapid turnover, wild-type and Stra8-deficient juvenile mice were treated with chloroquine, a weakly basic lysosomotropic agent that can block autophagosome fusion with lysosomes. Whereas chloroquine treatment resulted in an accumulation of autophagosome vesicles (GFP-positive and RFP-positive puncta) in Stra8-deficient testes, we did not observe an appreciable effect of autophagosome vesicle accumulation in wild-type testes, suggesting low autophagy activity (S1 Fig).
Autophagy is an essential intracellular degradation process. To evaluate autophagic degradation (flux) in wild-type and Stra8-deficient testes, we examined the protein level of p62 (or sequestosome 1, SQSTM1), a highly selective substrate for autophagic degradation [23]. The amount of p62 protein inversely correlate with autophagic flux activity: high levels of autophagy results in low p62 protein levels due to its degradation, while low levels of autophagy results in high p62 protein levels due to its accumulation. Wild-type testes contained seminiferous tubules with robust p62 protein accumulation at 21 d.p.p. (Fig 3A). In contrast, p62 protein level was almost completely lost in age-matched Stra8-deficient testes (Fig 3A). To determine if low p62 protein levels in Stra8-deficient testes were due to elevated autophagic degradation, possible changes due to Sqstm1 gene (encoding p62) expression and autophagosome degradation (by chloroquine treatment) were evaluated. Quantification of Sqstem1 mRNA showed comparable levels in age-matched wild-type and Stra8-deficient testes (Fig 3B) and whereas chloroquine showed no appreciable effect on p62 protein levels in wild-type testes (S2 Fig), chloroquine induced cytoplasmic p62 accumulation in germ cells of Stra8-deficient testes (Fig 3C). Together, these data suggest there is a rapid autophagic degradation of p62 in Stra8-deficient germ cells, reflective of high autophagic flux.
To help uncover the mechanism by which STRA8 influences autophagy, expression levels of 14 essential autophagy-lysosome genes were evaluated by quantitative RT-PCR (qRT-PCR). For these studies, juvenile testes at 10 d.p.p. were used to assure that the germ cell content is comparable between wild-type and Stra8-deficient testes and, thus, observed differences between mRNA levels are not due to differences in germ cell numbers (S3 Fig). Among these 14 genes, 6 genes, namely, Ulk1, Atg5, Map1lc3b, Vps18, Lamp2, and Tfeb, were significantly upregulated in Stra8-deficient testes (Fig 4). These genes encode essential factors for autophagosome formation (ULK1, ATG5, Map1lc3b), lysosome function (VPS18, LAMP2), as well as a master regulator of autophagy-lysosome genes (TFEB). Together, these data suggest that induction of autophagy in Stra8-deficient testes results from upregulation of specific autophagy-lysosome gene expression.
Our data in Stra8-deficient testes suggests that STRA8 may suppress autophagy by inhibiting autophagy-lysosome gene expressions. Currently, there are no available STRA8-expressing germ cell lines and, because Stra8 is transiently expressed on the verge of mitosis to meiosis transition, primary isolation and culture of Stra8-expressing cells could be challenging. Hence, to assess the role of STRA8 in autophagy suppression, STRA8 was ectopically and stably expressed in F9 embryonic carcinoma cells, a cell line regularly used for autophagy research [24]. Of note, autophagy machinery is present in every cell type and autophagy is an ongoing process in all cells; therefore, analysis of autophagy is often performed in cell lines [20]. To facilitate the identification of STRA8-expressing cells, STRA8 was tagged with GFP at its carboxyl terminus (S4 Fig). First, to test whether STRA8 suppresses autophagy induction, we used three commonly used inducers of autophagy, namely, amino acid starvation, rapamycin (mTOR inhibitor), and metformin (AMPK inducer) [25]. Autophagosome formation was detected by immunoblotting for LC3-II, which is a marker for autophagosomes [26]. It was found that, while LC3-II levels were significantly increased in control cells in all three conditions, there was no significant further increase of LC3-II levels in STRA8-expressing cells (Fig 5). These data suggest that STRA8 suppresses de novo autophagosome formation upon autophagy induction.
Although de novo autophagosome formation is impaired by STRA8 upon autophagy induction (Fig 5), we noted that there was a significant increase of LC3-II under basal condition (no autophagy induction) in STRA8-expressing cells, suggesting that STRA8 also inhibits autophagosome maturation, which results in autophagosome accumulation (upregulation of LC3-II) (Fig 6A). This result was confirmed at the cellular level by a significant increase of LC3 puncta (Fig 6B). Inhibition of autophagy flux frequently leads to autophagosome accumulation. Indeed, in our in vitro RFP-GFP-LC3 assay to monitor autophagy flux, STRA8 expression induced a significant accumulation of autophagosome vesicles (GFP-positive and RFP-positive puncta) that failed to mature into autolysosome vesicles (GFP-negative and RFP-positive puncta) (Fig 6C).
LC3-II and p62 are selectively degraded by autophagy. During chloroquine treatment, LC3-II and p62 accumulate due to inhibited autophagic flux, thereby serving as an indicator of autophagy flux activity. We show that the increment of p62 accumulation is significantly smaller in STRA8-expressing cells, suggesting that autophagy flux is being inhibited by STRA8. Consistent with the findings using in vitro RFP-GFP-LC3 assay, inhibition of autophagosome maturation by chloroquine resulted in a significant increase of LC3-II and p62 levels in control cells, but no significant further increase of LC3-II and p62 was observed in STRA8-expressing cells after chloroquine treatment (Fig 6B, 6D and 6E). Collectively, these data suggest that STRA8 also blocks autophagosome maturation under basal conditions.
To evaluate whether STRA8 influences autophagy-lysosome gene expression, the expression levels of the autophagy-lysosome genes that were upregulated in Stra8-deficient testes (Fig 4) were examined in these cells. We found that STRA8 expression alone was sufficient to cause a significant decrease in their expression levels, including Ulk1, Pik3C3, LC3B, Vps18, Lamp2, and Tfeb (Fig 6F). Together, these data suggest that STRA8 functions as a suppressor of autophagy by inhibiting autophagy-lysosome gene expression. Thus, loss of STRA8 function leads to the aberrant autophagy activation and upregulation of autophagy-lysosome gene expression in Stra8-deficient testes.
STRA8 contains a highly conserved bHLH domain that exhibits DNA binding activity [4]. To gain mechanistic insight into STRA8-mediated autophagy suppression, two STRA8 mutants in the bHLH domain were generated (S5A Fig): in the first mutant, point mutations were introduced in the first helix domain, which are known to disrupt the DNA binding activity of bHLH family transcription factors (mHelix) [27]; in the second mutant, point mutations were introduced in the basic domain, which disrupts the nuclear localization of STRA8 (mNLS) [4]. Both STRA8 mutants exhibited impaired nuclear localization (S5B Fig) and lost their ability to suppress autophagy activation (S5C Fig) as well as maturation (S5D Fig). These data suggest that the bHLH domain of STRA8 is critical for its autophagy suppression function.
To identify putative target gene(s) of STRA8 that could mediate its autophagy suppression function, we have performed an RNA-sequencing analysis in cells with transient ectopic expression of STRA8 under normal conditions. STRA8 upregulated 7 genes and downregulated 15 genes (≥ 2-fold change) (Fig 7A). Interestingly, none of the autophagy and lysosome genes upregulated in Stra8-deficient testes was detected under this condition, suggesting that STRA8 regulates autophagy through other target(s). Among the regulated genes by STRA8, we found that Nr1d1, a gene downregulated by STRA8, encodes a nuclear hormone receptor also known as Rev-erb-α. Nr1d1 is a critical circadian rhythm gene [28]. Recently, several studies have shown that NR1D1 acts as either an activator or an inhibitor of autophagy by regulating autophagy-lysosome gene expression, depending upon tissue context [29–31]. We therefore hypothesized that Nr1d1 could be a functional target of STRA8.
We confirmed that ectopic expression of STRA8 significantly reduced Nr1d1 levels in F9 cells and other cell types examined (S6A and S6B Fig), while the bHLH mutants of STRA8 failed to inhibit Nr1d1 expression (S6A Fig). Reciprocally, Stra8-deficiency induced a significant upregulation of Nr1d1 expression at mRNA level as detected by qRT-PCR analysis in testes (S6C Fig) and by in-situ hybridization (Fig 7B). To further evaluate if the induction of Nr1d1 mRNA is intrinsic to germ cells, we isolated the c-Kit-positive integrin α6-low differentiating spermatogonia, in which Stra8 is predominantly expressed, as well as the c-Kit-negative and integrin α6-high undifferentiated spermatogonia population, in which Stra8 is yet to be fully activated (S7 Fig) [32]. We found that Nr1d1 is more significantly upregulated in differentiating spermatogonia isolated from Stra8-deficient testes (Fig 7C). Moreover, we show that Stra8-deficient germ cells exhibited higher levels of NR1D1 expression at protein levels using immunofluorescence (Fig 7D).
To evaluate whether Nr1d1 is a direct genomic target of STRA8, the proximal region of the Nr1d1 promoter was examined for possible STRA8 binding sites. This identified a highly conserved canonical E-box (CAGCTG), the binding motif for members of the vertebrates bHLH protein family (Fig 7E). Chromatin immunoprecipitation (ChIP) detected robust STRA8 binding to this region of the NR1D1 promoter (Fig 7F), suggesting that STRA8 represses Nr1d1 transcription through E-box binding. Both bHLH mutants of STRA8 do not associate with the NR1D1 promoter at this E-box. Together, these data suggest that Nr1d1 is under direct transcriptional repression by STRA8.
To characterize whether STRA8 suppresses autophagy through NR1D1, we noted that past studies have shown that NR1D1 regulates autophagy through modulating the expression of Ulk1 gene [29–31], which encodes an essential autophagy initiator [33, 34]. Ulk1 mRNA is upregulated in Stra8-deficient testes, which we further confirmed using in-situ hybridization (Fig 8A). Moreover, we show that Ulk1 expression was significantly increased in c-Kit-positive integrin α6-low differentiating spermatogonia isolated from Stra8-deficient testes (Fig 8B), concomitant with upregulation of NR1D1 expression (Fig 7B–7D). Moreover, we examined a 2-kb region upstream of the transcription start site of mouse Ulk1 promoter and identified 3 RAR-related Orphan Receptor (ROR) DNA elements (ROREs), which consists of (A/G)GGTCA and could be putative NR1D1 binding sties [35, 36]. We therefore tested whether NR1D1 could directly regulate Ulk1 expression in mouse testis by a ChIP assay. Importantly, we found that NR1D1 binding affinity declined progressively from the distal to the proximal ROREs of the Ulk1 promoter (Fig 8C), suggesting that NR1D1 activates Ulk1 expression by engaging directly on the distal ROREs.
Transcriptional activation of Ulk1 is known to increase autophagy activity [37]. To test whether NR1D1 upregulation mediates the aberrant activation of Ulk1 transcription in Stra8-deficient testes, Ulk1 expression was evaluated in Stra8-/-;Nr1d1-/- double knockout mice. Testicular Ulk1 expression in double knockout mice was similar to that in wild-type, indicating that Nr1d1 is required for Ulk1 induction in Stra8-deficient testes (Fig 8D). Together, the results suggest that STRA8 suppresses autophagy by transcriptionally repressing Nr1d1 expression and, consequently, inhibiting the expression of essential autophagy initiation gene Ulk1.
To test whether repression of Nr1d1 activation by STRA8 is required to initiate meiosis, we evaluated whether NR1D1 inhibition could rescue Stra8-deficient testicular germ cells from meiotic initiation arrest. Stra8-deficient germ cells fail to enter meiosis during the first round of meiotic initiation in juvenile mouse testes [8]. Thus, we evaluated meiotic initiation of testicular germ cells into leptotene spermatocytes at 10 d.p.p.. In agreement with the previous report [8], we found that while spermatocytes at early meiotic prophase (leptotene) have appeared as a result of the first round of meiotic initiation in wild-type testes, these meiotic spermatocytes were absent in Stra8-deficient testes at this age (Fig 9A and 9D). Consistently, germ cells exhibiting nuclear distribution of SYCP3 (a synaptomeal complex protein) [38] together with foci of γ-H2AX (a hallmark of DNA DSBs) [39], two molecular characteristics of leptotene spermatocytes, were absent in Stra8-deficient testes (Fig 9B and 9E). Furthermore, gene expression analysis showed that testicular levels of Spo11, which encodes a topoisomerase essential for meiotic DSB formation [40, 41], Dmc1, which encodes a recombinase functioning in meiotic DSB repair [42, 43], and Sycp3, were significantly downregulated in Stra8-deficient mice (S8 Fig). Therefore, these missing hallmarks of meiotic initiation in Stra8-deficient testes at 10 d.p.p. provide a platform to evaluate the potential rescuing effects of NR1D1 inhibition on meiotic initiation.
Notably, juvenile testes from Stra8-/-;Nr1d1-/- mice at 10 d.p.p. contained germ cells with nuclear morphology resembling that of leptotene spermatocytes as observed in wild-type testes (Fig 9A; S9A Fig). Consistently, immunostaining revealed germ cells that exhibit nuclear distribution of SYCP3 together with foci of γ-H2AX staining in these testes (Fig 9B). Testicular levels of Spo11, Dmc1, and Sycp3 were significantly upregulated in Stra8-/-;Nr1d1-/- mice when compared to Stra8-/-;Nr1d1+/+ mice (Fig 9C). Moreover, Nr1d1 knockout alone showed no appreciable effect on meiotic initiation (Fig 9A; S9B and S9C Fig). Taken together, these results suggest that genetic loss of Nr1d1 exhibited rescuing effects on the meiotic initiation arrest in Stra8-deficient testicular germ cells.
To further examine the effect of NR1D1 inhibition on meiotic initiation arrest in Stra8-deficient testicular germ cells, we treated Stra8-deficient mice with SR8278, a synthetic NR1D1 antagonist [44]. Consistent with the results of genetic NR1D1 inhibition, we found that pharmacological inhibition recovered the appearance of germ cells with nuclear morphology as well as molecular hallmarks of leptotene spermatocytes in Stra8-deficient testes (Fig 9D and 9E). In addition, SR8278 treatment significantly stimulated testicular levels of Spo11, Dmc1, and Sycp3 in Stra8-deficient testes (Fig 9F). SR8278 treatment showed no appreciable effects on meiosis in wild-type testes (Fig 9D; S9D and S9E Fig). Taken together, the results from both genetic and pharmacological inhibition of NR1D1 suggest that aberrant upregulation of NR1D1 contributes to the meiotic initiation arrest of Stra8-deficient germ cells and that inhibition of Nr1d1 expression is an important feature of STRA8-directed meiotic initiation.
Despite dramatic sexual dimorphism in mammalian meiosis [45, 46], STRA8 appears to exhibit a comparable role in inducing both male and female meiosis [7, 8]. Thus, to examine whether STRA8 adopts similar mechanism of autophagy suppression in inducing meiosis in females, we investigated female meiosis, which occurs during embryonic day 13.5 (E13.5) to E16.5. Consistent with the observations in postnatal Stra8-deficient testes, autophagosome structures were frequently identified in germ cells of Stra8-deficient E14.5 ovaries (17 autophagosomes observed in 119 germ cells from 2 ovaries). In contrast, similar structures were not observed in age-matched wild-type fetal ovaries (158 germ cells from 2 wild-type fetal ovaries examined) (Fig 10A and 10B). Moreover, autophagy-lysosome genes as well as the STRA8 target gene, Nr1d1, were significantly upregulated in Stra8-deficient ovaries when compared to wild-type ovaries at the same developmental stage (Fig 10C). However, similar changes were not observed in embryonic testes at this age, because at this age meiosis is still inactive and STRA8 is absent until after birth (Fig 10D). Taken together, the results suggest that, similar to its role in male germ cell meiosis, STRA8 functions as a suppressor of autophagy in female meiosis.
Despite being an essential gatekeeper of meiotic initiation in mammalian germ cells, the molecular function of STRA8 has remained elusive to date. Here, by using mouse spermatogenesis as a model together with in vitro biochemical assays for autophagy, we report that STRA8 acts as a suppressor of autophagy by repressing Nr1d1 expression and consequently, inhibiting Ulk1 expression (Fig 11). Our data suggest a novel link between suppression of autophagy and meiotic initiation during mammalian germ cell development.
Stra8 was first reported as an essential gatekeeper of meiotic initiation in 2006 [7]. To our knowledge, the first study attempted to characterize the molecular function of STRA8 reports that STRA8 shuttles between nucleus and cytoplasm, but is mostly nuclear in freshly isolated germ cells [4]. This agrees with a study using immunostaining in PFA-fixed testicular sections, which localized STRA8 predominantly to the nucleus of meiosis-entering preleptotene spermatocytes [6]. In addition, protein-DNA cross-link studies showed STRA8 had DNA binding activity [4]. Moreover, STRA8 displays transcriptional activity when fused to a GAL4-DNA binding domain, [4]. Although these in vitro assays suggest STRA8 as a transcription factor, the targets of STRA8 and their molecular consequences have yet to be identified.
To characterize the meiotic gene program regulated by STRA8 in vivo, Soh and colleagues conducted RNA-sequencing analysis in wild-type and Stra8-deficient fetal ovaries at E14.5, when ovarian germ cells enter meiosis [47]. However, only meiotic genes were not selected for investigation by Soh and colleagues. Interestingly, examination of this earlier RNA-sequencing result revealed notable similarities to the data reported herein, namely, the autophagy-lysosome genes that exhibit significant upregulation in both Stra8-deficient testes at 10 d.p.p. and Stra8-deficient fetal ovaries at E14.5 in our study (Fig 4 and Fig 10C, respectively), including Ulk1, Atg5, Map1lc3b, Lamp2, Vps18, and Tfeb, also exhibited tendency of being upregulated in Stra8-deficient fetal ovaries at E14.5 in the RNA-sequencing result of Soh and colleagues. Moreover, except for Map1lc3a, the autophagy-lysosome genes that were not upregulated in Stra8-deficient testes at 10 d.p.p. in our study (Fig 4), i.e., Pik3c3, Atg7, Vps11, Uvrag, Lamp1, and Becn1, were also not upregulated in Stra8-deficient fetal ovaries at E14.5 in the RNA-sequencing result of Soh and colleagues. Thus, changes in autophagy-lysosome gene expression in Stra8-deficient fetal ovaries at E14.5 detected by Soh and colleagues using RNA-sequencing align with the data presented in the current study. These data together suggest that STRA8 inhibits the expression of a selective autophagy-lysosome gene program, thereby suggesting STRA8 functions as s suppressor of autophagy.
A more recent study by Shen and colleagues is focused on characterizing a potential role for STRA8 in preventing apoptosis of germ cells [48]. Generally speaking, autophagy is associated with a response to stress to prevent cells death; however, excessive autophagy leads to apoptosis [49]. Stra8-deficient germ cells exhibited profound activation of autophagy (Figs 1–4) which may represent a condition of uncontrolled autophagy activation that leads to their germ cell apoptosis [8].
To date, the roles of autophagy regulation in meiosis remains unclear. In budding yeast (S. cerevisiae), autophagy participates in the early phase of meiosis and is switched off upon meiotic division [50]. In fission yeast (S. pombe), however, autophagy is thought to be required for chromosome segregation during meiotic division [51]. Thus, these observations underscore the concept that mechanisms of meiotic initiation could differ in model organisms [2].
Our observation that autophagy is being actively suppressed by STRA8 during meiotic initiation in mouse spermatogenesis is in accordance with a recent study of autophagy in rat spermatogenesis, in which autophagy was found only to be activated during late meiotic spermatocytes but not in spermatogonia and early spermatocytes [19]. This is in line with the finding in yeast, in which autophagy is required for proper meiotic chromosome segregation [51]. Together, our findings suggest that activation of autophagy is specifically prevented by STRA8 during meiotic initiation in mammalian germ cell development. It should be noted that Stra8-deficient germ cells, when rescued by inhibiting NR1D1 antagonist SR8278, do not progress beyond leptotene stage (S10 Fig), suggesting that STRA8 has additional role(s) beyond the initiation of meiosis through NR1D1-independent mechanisms. For instance, based on our RNA-seq analysis, 15 out of 22 STRA8-regulated genes of are noncoding RNAs (Fig 7A). Given that non-coding RNAs play an essential role in meiosis prophase and homologous recombination [35], it is possible that STRA8 regulates this stage of meiosis through noncoding RNAs.
We observed that STRA8-deficiency resulted in a significantly autophagy-lysosome gene expression in testes; reciprocally, these autophagy-lysosome genes were downregulated in cultured cells with stable STRA8 expression. Based on these data and a past study that indicated STRA8 could bind to DNA and display transcriptional activity [4], we had expected that STRA8, like some other transcriptional regulators of autophagy, such as TFEB [13] or FXR [52], could potentially regulate a wide spectrum of autophagy-lysosome genes directly. However, no autophagy-lysosome genes were identified as direct targets of STRA8 by transient ectopic expression of STRA8 in our in vitro RNA-sequencing analysis under basal condition. Instead, Nr1d1 was identified as a target repressed by STRA8 transcriptionally. To date, three studies have reported a role of NR1D1 in either inducing or inhibiting autophagy, all involving transcriptional regulation of Ulk1 [29–31]. While Woldt and colleagues reported that NR1D1-deficiency in murine muscle leads to upregulation of autophagy [30], Chandra and colleagues reported that Nr1d1-knock down leads to reduced autophagy and downregulation of Ulk1 gene expression in human macrophage [29]. Interestingly, whether NR1D1 is an autophagy inducer or inhibitor remains unclear in the study reported by Huang and colleagues [31]. Thus, these studies point out a critical regulatory role of NR1D1 in autophagy through modulating autophagy-lysosome gene expression. Herein, the data show that genetic loss of NR1D1 prevented the upregulated Ulk1 expression normally observed in Stra8-knockout testes (Fig 8D), supporting a role for NR1D1 as an inducer of autophagy by activating autophagy-lysosome gene expression during germ cell development. It is currently unclear why there is a robust activation of autophagy that is precisely counteracted by STRA8-mediated suppression of autophagy during the transition to meiosis. One possibility is that the activation of autophagy is simply a response to cellular stress. A second, and more intriguing, possibility is that autophagy induction also plays an essential role in meiotic initiation and is an area that warrants further investigation.
Based primarily on models of somatic tumorigenesis, genetic disruption of critical autophagy factors can induce the formation of DNA DSBs upon metabolic stress [14, 15]. In germ cells, DSBs are required to initiate meiosis and permit the exchange of genetic information between maternal and paternal chromosomes through homologous recombination [53]. Stra8 activation is required for meiotic DSB formation [7, 8]. Thus, by characterizing STRA8 as a suppressor of autophagy, our work suggests autophagy suppression as a possible mechanism adopted by germ cells to form meiotic DSBs. The mechanisms underlying autophagy inhibition-induced DSB formation is currently unclear. Several studies show that autophagy-deficiency causes DNA damage through accumulation of autophagic substrate, p62 [54–56]. However, p62-deficient mice are reportedly to be fertile [57], suggesting that p62 accumulation alone is dispensable for meiosis. In addition, loss of autophagy has been shown to affect DNA damage repair machineries, such as HIPα (a molecule essential for chromosome condensation) [16] and Chk1 (a molecular regulator of DNA damage repair by homologous recombination) [58]. Thus, a role for autophagy suppression in DSB formation during meiotic initiation remains to be determined.
Mounting efforts have been directed to derive functional haploid gametes (sperm or oocytes) from spermatogonial stem cells, embryonic stem cells or induced pluripotent stem cells in cultures [59]. However, how to properly induce and sustain meiosis remains to be a major challenge in this field. Our study revealed that germ cells entering meiosis is exposed to an antagonistic pressure of simultaneous autophagy induction and STRA8-mediated autophagy suppression. Thus, manipulating autophagy pathway to mimic this in vivo condition may facilitate the induction of meiosis in cultures, thereby advancing the technology of in vitro haploid gamete production that may ultimately afford clinical utility in assisted reproduction technology.
All genetically modified mice were obtained from Jackson Laboratory: Stra8-deficient mice (Stock number: 023805), RFP-GFP-LC3 transgenic mice (Stock number: 027139), and Nr1d1-deficient mice (Stock number: 018447). For chloroquine treatment, mice were injected intraperitoneally with chloroquine dissolved in PBS at 100 mg/kg body weight daily. For SR8278 treatment, mice at 7 days of age were injected intraperitoneally with vehicle (5% DMSO, 10% Cremephol, and 85% PBS) or SR8278 at 100 mg/kg body weight daily for 3 days. Postnatal testes were dissected from male mice after euthanasia. For timed pregnancy, females at estrus stage were caged with males overnight, and the presence of vaginal plug was examined the following morning and the midday was defined as E0.5. Pregnant females were euthanized when embryos reached E14.5. Embryos were collected from uterine horns before gonad isolation under a binocular dissecting microscope.
All procedures and care of animals were carried out according to the Massachusetts General Hospital (MGH) and the University of Kansas Medical Center (KUMC) Institutional Animal Care and Use Committee (IACUC) under IACUC protocol number 2012N000015 and 2018–2461, respectively. Euthanasia was performed by CO2 inhalation followed by cervical dislocation (MGH) and decapitation (KUMC).
Tissue specimens were fixed in 2.5% glutaraldehyde in 0.1M sodium phosphate buffer (pH 7.4, Electron Microscopy Sciences, Hatfield, PA), then rinsed several times in 0.1M sodium cacodylate buffer. Specimens were post-fixed in 1.0% osmium tetroxide in cacodylate buffer for 1 hour at room temperature and rinsed several times in cacodylate buffer. Samples were then dehydrated through a graded series of ethanols to 100% and dehydrated briefly in 100% propylene oxide. Samples were pre-infiltrated 2 hours in a 2:1 mix of propylene oxide and Eponate resin (Ted Pella, Redding, CA), then transferred into a 1:1 mix of propylene oxide and Eponate resin and allowed to infiltrate overnight on a gentle rotator. The following day, specimens were infiltrated with fresh 100% Eponate resin for several hours, embedded in flat molds with fresh 100% Eponate and allowed to polymerize 24–48 hours at 60°C. Thin (70nm) sections were cut using a Leica EM UC7 ultramicrotome, collected onto formvar-coated grids, stained with uranyl acetate and Reynold's lead citrate and examined in a JEOL JEM 1011 at MGH or The JEOL JEM-1400 at KUMC transmission electron microscope at 80 kV. Images were collected using an AMT digital imaging system (Advanced Microscopy Techniques, Danvers, MA).
Human STRA8 expression plasmid was obtained from Origene (RC213536). To tag STRA8 with GFP, STRA8 cDNA was inserted into the pEGFP-N1 vector (Clontech). Mutant STRA8 plasmids were generated using Q5 mutagenesis kit (New England Biolabs). All plasmids were sequenced to confirm fidelity prior to use.
F9 embryonic carcinoma cells (ATCC CRL-1720; American Type Culture Collection) and 293T were cultured at 37°C in a humidified atmosphere of 5% CO2-95% air in DMEM containing 4.5 g/l glucose and supplemented with 10% FBS (Invitrogen), 2 mM L-glutamine, 50 μg/ml penicillin and 50 μg/ml streptomycin. F9 cells were cultured on 0.1% gelatin-coated tissue culture plates. Cells were transfected using Lipofectamine 2000 (Invitrogen) and then stably selected by G418 (Invitrogen) for 2 to 3 weeks, before STRA8-expressing cells (GFP-positive) were identified and isolated by flow cytometry. F9 cells expressing GFP alone were used as control. For LC3 staining, cells were fixed in ice-cold methanol for 2 minutes before immunostaining with antibody against LC3 (12741, Cell Signaling Technology). Images were captured by using a ZEISS LSM 800 confocal microscopy with Airyscan. To induce autophagy, control or Stra8-expressing cells were treated with rapamycin (0.1 mM), metformin (2 mM), or Earle's balanced salt solution (EBSS) for 2 hours before cell lysates were collected. To monitor the effects of blocking autophagosome maturation, control or STRA8-expressing cells were treated with chloroquine (20 μM) for 2 hours before cell lysates were collected. Hela cells carrying mRFP-GFP-LC3 were transfected by plasmid expressing empty pCMV6 vector or STRA8 (not tagged with GFP). Then, cells were fixed by 4% paraformaldehyde for 10 minutes at room temperature before observation under fluorescence microscope.
Testes were fixed in 4% paraformaldehyde, embedded in paraffin, and sectioned for analysis. Antibodies used include: p62 (ab56416, abcam; 1:2000 dilution for immunofluorescence and 1:10,000 dilution for immunohistochemistry), γ-H2AX 05–636, Millipore; 1:1,000 dilution), NR1D1 (sc-100910, Santa Cruz Biotechnology; 1:500 dilution), SYCP3 (sc-74569, Santa Cruz Biotechnology; 1:500 dilution). For immunofluorescence, detection was performed using Alexa fluor 546-conjugated goat anti-rabbit antibody and Alexa fluor 488-conjugated goat anti-mouse secondary antibodies. For immunohistochemistry, detection was performed using goat anti-mouse or goat anti-rabbit as secondary antibody for horseradish peroxidase-based DAB detection (DAKO). Images were captured using a Nikon ECLIPSE TE2000-S microscope and were analyzed by Image J software (National Institutes of Health).
Dissected seminiferous tubules from wild-type and Stra8-deficient testes carrying the RFP-GFP-LC3 reporter were fixed in 4% paraformaldehyde on ice for an hour. Tissue is washed with PBS to remove PFA. For whole-mount imaging, tissues of seminiferous tubules were mounted directly on glass slides in PBS for confocal microscopy by using a Nikon A1R microscope. For imaging on testicular cross sections, testes were cryosectioned. Then sections were washed with warm PBS to remove gelatin and were imaged under Nikon ECLIPSE TE2000-S microscope. No staining was performed under both conditions. Images from wild-type and Stra8-deficient samples mounted on the same slide were captured and processed in parallel using identical settings.
Total protein was isolated in RIPA buffer supplemented with 1 mM PMSF (Sigma) and protease inhibitor cocktail (Sigma P8340). Lysates were cleared by centrifugation at 14,000 X g for 10 min at 4°C, and protein concentrations in supernatants were determined (DC protein assay; BioRad). Equal amount of protein from each sample was mixed with LDS sample buffer (Invitrogen) plus sample reducing agent (Invitrogen), and denatured for 10 min at 70°C. Proteins were resolved in Bis-Tris gels (Thermo Fisher), and transferred to PDVF membranes. Blots were probed with antibodies against LC3A/B (12741, Cell Signaling Technology; 1:1,000 dilution), p62 (ab56416, Abcam; 1:20,000 dilution), GFP (sc-9996, Santa Cruz Biotechnology; 1:1,000 dilution), or pan-actin (MS-1295, Thermo Fischer; 1:1,000 dilution), washed and reacted with horseradish peroxidase-conjugated goat anti-rabbit or anti-mouse IgG (BioRad). Detection was performed with the Clarity ECL Western Blotting Substrate (BioRad).
293T cells were transfected with plasmid expressing GFP-tagged wild-type and mutant STRA8. 24 hours later, cell lysates were processed using the EZ-ChIP kit (Millipore, Temecula, CA) along with a rabbit polyclonal anti-GFP antibody (ab290; Abcam) for immunoprecipitation. Normal rabbit IgG was used as a negative control. Precipitated soluble chromatin was analyzed by PCR using primer sets to target E-box: forward, 5’-CCC TCC CCG GCT TCT CTC TCT CC-3’, reverse, 5’-GCA AAC CTT GCA AAC GTG AGG GC-3’; and to target exon 8: forward, 5’-CCG GAC CTG CGG ACC CTG AAC AA-3’, reverse, 5’-TCT GTA CAA GGG GGC AGC GGC AGA-3’.
To characterize NR1D1 binding to the Ulk1 promoter, testicular lysates were processed using the EZ-ChIP kit (Millipore, Temecula, CA) along with a rabbit polyclonal anti-NR1D1 antibody (#13418; Cell Signaling Technology) for immunoprecipitation. Normal rabbit IgG was used as a negative control. Precipitated soluble chromatin was analyzed by PCR using primer sets to target ROREs: primer set 1, forward, 5’-AAT GGG TAT GTG CGA CAA CA-3’, reverse, 5’-TGT CAT TTG GGG AGG GGT AT-3’; primer set 2, forward, 5’-TGC CAA GTT TGA CAA CCT GA-3’, reverse, 5’-CTG TAT GTG GGG ACG GAG AC-3’; primer set 3, forward, 5’-GCA CCT GCC TTT AAT TCC AA-3’, reverse, 5’-CGA CTG GTC TCG AAC TTG CT-3’.
MCF-7 cells were transiently transfected with control (pCMV6) or STRA8 plasmid. Cells were collected after 24 hours. Total RNA was isolated using the RNeasy Mini Kit (Qiagen). Following rRNA depletion using RiboZero kit (Epicentre/Illumina), RNA-Seq libraries were constructed using NEBNext Ultra Directional RNA library prep kit for Illumina (New England Biolabs) and sequenced on Illumina HiSeq2500 instrument, resulting in approximately 25 million reads per sample on average. STAR aligner was used to map sequencing reads to transcripts in hg19 reference genome. Read counts for individual transcripts were produced with HTSeq-count, followed by the estimation of expression values and detection of differentially expressed transcripts using EdgeR. Principle component analysis (PCA) was performed on the union of differentially expressed transcripts in all samples.
In situ hybridization (ISH) has been described previously [60]. Antisense Nr1d1 probe (825 bp) and Ulk1 probe (843 bp) were amplified from cDNA prepared from juvenile testes, and labeled with digoxigenin (Roche Diagnostics).
Testes were collected from age-matched wild-type and Stra8-deficient mice. Testicular cells were dissociated by two-step enzymatic digestion, followed by staining with PE-conjugated rat anti-human integrin α6 antibody (BD Pharmingen, clone GoH3) and APC-conjugated rat anti-mouse c-Kit antibody (BD Pharmingen, clone 2B8) as previously described [32]. 40,000–50,000 cells from each population were sorted directly into lysis buffer in NucleoSpin RNA XS kit (Takara) for subsequent RNA isolation.
Equal amount of total RNA from each sample was reverse transcribed by using SuperScript III from Invitrogen. quantitative RT-PCR was conducted by using SsoAdvanced Universal SYBR Green Supermix (BioRad). Primer sequences are listed below:
β-actin forward, 5’-CTG CCG CAT CCT CTT CCT C-3’
reverse, 5’-GCC ACA GGA TTC CAT ACC CA-3’
Mvh forward, 5’-GCT TCA TCA GAT ATT GGC GAG T-3’
reverse, 5’-GCT TGG AAA ACC CTC TGC TT-3’
Maplc3A forward, 5’-CAC ATC CTG GGT AGG TCC TG-3’
reverse, 5’-AAT GAC AAA CCC CAC AGA GC-3’
Maplc3B forward, 5’-CGG CTT CCT GTA CAT GGT TT-3’
reverse, 5’-AAC CAT TGG CTT TGT TGG AG-3’
Atg12 forward, 5’-TCC CCG GAA CGA GGA ACT C-3’
reverse, 5’-TTC GCT CCA CAG CCC ATT TC-3’
Atg5 forward, 5’-TGT GCT TCG AGA TGT GTG GTT-3’
reverse, 5’-GTC AAA TAG CTG ACT CTT GGC AA-3’
Atg7 forward, 5’-GTT CGC CCC CTT TAA TAG TGC-3’
reverse, 5’-TGA ACT CCA ACG TCA AGC GG-3’
Pik3c3 forward, 5’-CCT GGA CAT CAA CGT GCA G-3’
reverse, 5’-TGT CTC TTG GTA TAG CCC AGA AA-3’
Tfeb forward, 5’-CCA CCC CAG CCA TCA ACA C-3’
reverse, 5’-CAG ACA GAT ACT CCC GAA CCT T-3’
Ulk1 forward, 5’-AAG TTC GAG TTC TCT CGC AAG-3’
reverse, 5’-CGA TGT TTT CGT GCT TTA GTT CC-3’
UVRAG forward, 5’-ACA TCG CTG CTC GGA ACA TT-3’
reverse, 5’-CTC CAC GTC GGA TTC AAG GAA-3’
Vps11 forward, 5’-AAA AGA GAG ACG GTG GCA ATC-3’
reverse, 5’-AGC CCA GTA ACG GGA TAG TTG-3’
Vps18 forward, 5’-ACG AGG ACT CAT TGT CCC G-3’
reverse, 5’-CAT ACC CAG AAT GGG GGA TGC-3’
Lamp1 forward, 5’-CAG CAC TCT TTG AGG TGA AAA AC-3’
reverse, 5’-ACG ATC TGA GAA CCA TTC GCA-3’
Lamp2 forward, 5’-TGT ATT TGG CTA ATG GCT CAG C-3’
reverse, 5’-TAT GGG CAC AAG GAA GTT GTC-3’
Sqstem1 forward, 5’-AGG ATG GGG ACT TGG TTG C-3’
reverse, 5’-TCA CAG ATC ACA TTG GGG TGC-3’
Beclin1 forward, 5’-ATG GAG GGG TCT AAG GCG TC-3’
reverse, 5’-TCC TCT CCT GAG TTA GCC TCT-3’
Nr1d1 forward, 5’-ATG CCC ATG ACA AGT TAG GC-3’
reverse, 5’-CGG TGT GGA GTT GTA GCT GA-3’
All experiments were replicated at least three times independently. Different mice, tissues or cells were used during each experimental replicate. Animal assignment to each experimental group was made randomly. Quantitative data from the experimental replicates were pooled and are presented as the mean ± SEM or mean ± SD as indicated in the figure legend. Compiled data were analyzed by Student’s t-test.
|
10.1371/journal.pgen.1002289 | Bmp and Nodal Independently Regulate lefty1 Expression to Maintain Unilateral Nodal Activity during Left-Right Axis Specification in Zebrafish | In vertebrates, left-right (LR) axis specification is determined by a ciliated structure in the posterior region of the embryo. Fluid flow in this ciliated structure is responsible for the induction of unilateral left-sided Nodal activity in the lateral plate mesoderm, which in turn regulates organ laterality. Bmp signalling activity has been implied in repressing Nodal expression on the right side, however its mechanism of action has been controversial. In a forward genetic screen for mutations that affect LR patterning, we identified the zebrafish linkspoot (lin) mutant, characterized by cardiac laterality and mild dorsoventral patterning defects. Mapping of the lin mutation revealed an inactivating missense mutation in the Bmp receptor 1aa (bmpr1aa) gene. Embryos with a mutation in lin/bmpr1aa and a novel mutation in its paralogue, bmpr1ab, displayed a variety of dorsoventral and LR patterning defects with increasing severity corresponding with a decrease in bmpr1a dosage. In Bmpr1a-deficient embryos we observed bilateral expression of the Nodal-related gene, spaw, coupled with reduced expression of the Nodal-antagonist lefty1 in the midline. Using genetic models to induce or repress Bmp activity in combination with Nodal inhibition or activation, we found that Bmp and Nodal regulate lefty1 expression in the midline independently of each other. Furthermore, we observed that the regulation of lefty1 by Bmp signalling is required for its observed downregulation of Nodal activity in the LPM providing a novel explanation for this phenomenon. From these results we propose a two-step model in which Bmp regulates LR patterning. Prior to the onset of nodal flow and Nodal activation, Bmp is required to induce lefty1 expression in the midline. When nodal flow has been established and Nodal activity is apparent, both Nodal and Bmp independently are required for lefty1 expression to assure unilateral Nodal activation and correct LR patterning.
| Although vertebrates are bilaterally symmetric when observed from the outside, inside the body cavity the organs are positioned asymmetrically with respect to the left and right sides. Cases where all the organs are mirror imaged, known as situs inversus, are not associated with any medical defects. Severe medical problems occur however in infants with a partial organ reversal (situs ambigious or heterotaxia), which arises during embryonic development. Left-right asymmetry in the embryo is established by unilateral expression of Nodal, a member of the Tgf-ß superfamily of secreted growth factors, a role that has been conserved from human to snails. By performing a genetic screen in zebrafish for laterality mutants, we have identified the linkspoot mutant, which displayed partial defects in asymmetric left-right positioning of the internal organs. The gene disrupted in the linkspoot mutant encodes a receptor for bone morphogenetic proteins (Bmp), another member of the Tgf-ß superfamily of secreted growth factors. Further analysis of Bmp over-expression or knock-down models demonstrate that Bmp signalling is required for unilateral Nodal expression, through the initiation and maintenance of an embryonic midline barrier. Our results demonstrate a novel and important mechanism by which left-right asymmetry in the vertebrate embryo is established and regulated.
| In vertebrates the internal organs are positioned asymmetrically along the left-right (LR) axis. For example, in humans, the heart is positioned on the left side, as is the stomach whilst the liver is positioned on the right side. Within organs LR asymmetry also exists. For example, the two lungs appear identical however they are divided into lobes with 3 on the right lung and 2 on the left. Animals with situs inversus totalis (a LR reversal of all organs) have no pathological features [1] however severe medical problems occur in infants with a partial reversal in a subset of organs (situs ambigious or heterotaxia). These heterotaxic phenotypes occur during early embryonic development and can have both genetic as well as environmental causes [2], [3].
A ciliated organ at the posterior end of the embryo is required for LR-axis specification in the embryo. In this LR organ, the node in mouse or Kupffer's vesicle (KV) in zebrafish, cilia rotate and create a directional fluid flow from the right to left side of the embryo. This directional nodal flow induces a unilateral and asymmetric expression of Nodal in the left lateral plate mesoderm (LPM) directing organ laterality. Unilateral expression of Nodal is essential for correct LR-axis specification, a function that has been highly conserved from human to snails [2], [4], [5]. Although unilateral expression of Nodal is highly conserved and essential for LR–axis specification, there is still very little understanding of how this unilateral Nodal expression is initiated by nodal flow and maintained in the LPM.
Nodal is a member of the Tgf-ß superfamily of secreted growth factors. Nodal signaling is activated by the interaction of Nodal ligands with the type I and II Activin receptors and the Cripto coreceptor (reviewed by A.F. Schier [6]). Upon Nodal interaction with its receptor, intracellular Smad2 protein is phosphorylated, which after associating with Smad4 protein is translocated to the nucleus to activate transcription of downstream target genes. Extracellular antagonists such as Lefty and Cerberus can inhibit Nodal signalling either by direct interaction with Nodal or by competing with Nodal for binding to the receptor. The activity of Lefty proteins, Lefty1 and Lefty2, is controlled at the level of transcription. In most tissues Lefty expression is dependent on Nodal signalling [6]. During LR-axis formation in mouse embryos Lefty1 and Lefty2 have reciprocal expression patterns. While Lefty1 is expressed strongly in the presumptive floor plate and only weakly in the left LPM, Lefty2 is expressed strongly in the left LPM and only weakly in the presumptive floorplate [7]. During LR axis formation in zebrafish embryos lefty1 is expressed in the notochord. Only after LR patterning has been established are lefty1 and lefty2 expressed in the left cardiac field [8]. Nodal likely activates its own expression via a positive feedback loop while it also activates expression of its own antagonists Lefty1 and Lefty2. Genetic experiments in mouse demonstrated that Lefty1 is the more important antagonist and is essential for LR-axis formation [9]. It is believed that Lefty1 expression in the midline prevents Nodal from crossing the midline, blocking activation of Nodal signalling in the right LPM. Indeed loss of Lefty1 expression caused the ectopic expression of Nodal and other left-sided genes in the right LPM and resulted in various laterality defects. It has been suggested that Nodal and Lefty maintain the L/R asymmetry by a self-enhancement and lateral-inhibition (SELI) mechanism [10]. With the SELI model it is possible to explain how a small difference between two separated regions is converted into a robust difference through local activation and long-range inhibition [11].
Bmps have been implicated in LR patterning but data on their precise role has been contradictory [12]–[22]. This is partly due to Bmp ligands acting in opposite fashions, depending on the time and place of action during LR-axis specification [13], [16]. Bmp proteins are members of the Tgf-ß superfamily of growth factors. Extracellular antagonists of Bmp signalling are Noggin, Chordin and Follistatin. Upon interaction with their serine/threonine kinase type I and II Bmp receptors, Bmp ligands induce intracellular phosphorylation of Smad1, 5 or 8 proteins [23]. Mouse embryos deficient for the type I Bmp receptor Bmpr1a/Alk3 or Acvr1/Alk2 fail to form mesoderm, which has hampered the study of their role during LR-axis specification [24]–[26].
In the current work we describe the identification of the linkspoot (lin) mutant from a forward genetic screen for laterality mutants. A missense mutation in the bmpr1aa gene is responsible for the LR defect of lin mutant embryos. Due to a genome duplication event, there is a second gene encoding a Bmpr1a (bmpr1ab) in the zebrafish genome. By screening an ENU-mutagenized zebrafish library we identified a nonsense allele in the bmpr1ab gene. Genetic analysis reveals that a reduction in Bmpr1a activity results in left isomerism of the viscera, demonstrating an essential and early role in LR-axis specification. Previous genetic data has provided evidence that Bmp signalling is required to repress Nodal activation in the right LPM but various direct and indirect models have been proposed to explain this activity [12]–[22], [27]. Here we provide evidence that Bmp signalling via Bmpr1a inhibits Nodal activation in the right LPM indirectly by inducing lefty1 expression in the midline, offering a new model of the interactions between Nodal, Bmp and Lefty in induction and maintenance of LR asymmetry.
From an ENU-mutagenesis screen, we identified a unique mutant, linkspoot (linhu4087), that displayed a reduced ventral tail fin in combination with a heart-specific laterality defect (Figure 1A, 1B, 1E and 1F). At 30 hours post fertilization (hpf), 24.6% (n = 464) of the embryos derived from an incross of two lin heterozygous carriers displayed the small but noticeable reduction of the ventral tail fin (Figure 1A and 1E). Whilst the majority of lin mutant embryos with the ventral tail fin reduction had no other obvious morphological defects and survived to adulthood, 29% (33 out of 114 lin mutant embryos) showed cardiac defects resulting in cardiac failure and death at around 5 days post fertilization (dpf) (Figure 1F, 1G). Examination of the cardiac defect in lin mutant embryos revealed a midline positioning of the heart in contrast to a leftward positioning in wild-type siblings at 28 hpf. Furthermore at 48 hpf, when the heart in wild-type sibling embryos has completed looping toward the right, heart looping in these lin mutant embryos was incomplete (n = 6/8) (data not shown). Despite the aberrant heart looping in almost 30% of the lin mutant embryos, patterning of the myocardium and endocardium was grossly normal. Expression of tbx2b and has2 in the atrioventricular canal myocardium and endocardium, respectively, was comparable between lin mutant and sibling embryos (Figure S1). In addition, bmp4 expression was still restricted (although slightly expanded) to the venous pole, atrioventricular canal and arterial pole (Figure S1).
We observed that the laterality of the other visceral organs (direction of gut looping, positioning of the liver and pancreas) was unaffected in lin mutant embryos (30 out of 34) (Figure 1C, 1D, 1G and 1H). Since lin mutant embryos that did not display the cardiac defects described above survived up to adulthood we crossed homozygous lin mutant females with heterozygous lin carrier males. The resulting maternal and zygotic (MZ) lin mutant embryos displayed a reduction of ventral structures such as the tail fin and blood islands (Figure 1I). Such phenotypes have been associated with aberrant dorsoventral patterning of the embryo [28]. In addition, we observed in MZlin mutants, uncoordinated laterality defects in the viscera (Figure 1J–1L). Aberrant positioning of the heart and other viscera can be caused by defects in formation or function of the Kupffer's vesicle, resulting in disrupted LR patterning. We therefore examined cilia rotation in the KV, and found that both in Zlin and MZlin mutant embryos with a midline positioning of the heart, cilia rotation in the Kupffer's vesicle was unaffected (Figure S2 and Videos S1, S2, S3), suggesting that the laterality defect was not due to a disruption of Kupffer's vesicle function. Together, these results suggest that the affected gene product in lin mutants is required for dorsal-ventral and left-right axis specification.
To better understand the molecular nature of the lin mutant phenotype, we positionally cloned the gene that is disrupted in the lin mutant. Using bulk segregant analysis with SSLP markers we placed the lin mutation onto chromosome 13. Mapping of the lin locus using 570 mutant embryos resulted in the identification of a chromosomal region containing a zebrafish orthologue of the mammalian Bmpr1a/Alk3 gene, encoding a Bmp Type I receptor (Figure 2A). Since Bmp signalling is instructive for cardiac laterality as well as ventral tail fin formation [13], [29], [30], we sequenced the coding region of the bmpr1aa gene for mutations. We identified a base pair substitution (T > G) at position 1538 resulting in a leucine to arginine substitution at position 337 (L337R) in the kinase domain of the Bmpr1aa protein (Figure 2B). The T1538G polymorphism was invariably linked with the mutant phenotype (n = 570). No other non-synonymous substitutions were identified in the coding region of bmpr1aa that were linked with the mutant phenotype. Modelling of the corresponding region of human BMPR1A suggested that the L312R (corresponding to zebrafish L337R) substitution is incompatible with proper folding of this region and thereby likely destabilizes the entire kinase domain (Figure 2C, 2D).
To address the functional consequence of the L337R substitution, we introduced the lin mutation in the bmpr1aa gene and generated synthetic mRNA for injection into embryos. Surprisingly, injection of wild-type bmpr1aa mRNA into wild-type 1-cell stage embryos resulted in a loss of the ventral tail fin (Table 1). Injection of bmpr1aa L337R mRNA had a stronger inhibition of Bmp signalling since more of the injected embryos displayed a dorsalised phenotype, which was also stronger in its effect (Table 1). These results suggest that increasing wild-type Bmpr1a beyond physiological levels has a negative effect on Bmp signalling, possibly by titrating out other components of the signalling pathway. The dominant-negative effect is stronger for the Bmpr1aa L337R most likely because Bmpr1aa L337R is still able to form a receptor complex and interact with Bmp but it can no longer phosphorylate the receptor Smad protein due to its mutation in the kinase domain. To test this hypothesis we injected a lower dose of the wild-type bmpr1aa mRNA into embryos derived from an incross of two lin heterozygous carriers to determine whether we could rescue the tail fin defects of lin mutant embryos. Indeed we observed that injection of low levels of wild-type bmpr1aa was able to rescue the ventral tail fin defects in almost 50% of lin mutant embryos (Table 1). Consistent with our model that Bmpr1aa L337R has reduced signalling activity, we never observed a rescue of the tail fin defects of lin mutant embryos when we injected the bmpr1aa L337R mRNA. From these results we conclude that the gene that is disrupted in lin mutants encodes the Bmp receptor, Bmpr1aa, and that the lin mutation inactivates Bmpr1aa activity.
To further characterise the requirement for bmpr1a during zebrafish development, we analysed its expression. Interestingly, database searches revealed that due to a genome duplication event, a paralogue of bmpr1aa existed in the form of bmpr1ab/alk3b (exhibiting 80% identity at the protein level). We, therefore, simultaneously analysed the expression pattern of these two closely related genes. ISH analysis revealed that both bmpr1a paralogues are expressed from the 2-cell stage, indicating maternal deposition of the transcripts (Figure S3). Each paralogue was expressed in a ubiquitous fashion up until the 10-somite stage, however the signal for bmpr1aa was more intense compared to the signal of bmpr1ab suggesting different levels of expression. From 20-somites onwards, the expression of both paralogues became progressively restricted to anterior regions.
The similar expression patterns observed for bmpr1aa and bmpr1ab suggest comparable functions for the paralogues. To analyse this possibility further we screened a mutagenesis library for a bmpr1ab mutant. We identified a mutant harbouring a stop codon (TAC>TAA) in the second exon of the gene, truncating the protein 84 amino acids into the ligand-binding domain (Y84X) (Figure 2E). Although a DV patterning defect was reported upon morpholino knockdown of bmpr1ab [31], we observed no morphological phenotype in the majority of bmpr1ab zygotic mutants. Furthermore, maternal zygotic bmpr1ab mutants exhibited no observable phenotype (Figure 2F).
We next tested for possible redundancy between the two paralogues. By incrossing double heterozygous carriers for the two mutations (bmpr1aa+/−;bmpr1ab+/−), we observed a spectrum of dorsalised embryonic phenotypes, ranging from wild-type phenotypes to C4 dorsalisation in the most severe instances (categorisation according to Mullins et al., [28]) (Figure 2G). Genotyping revealed that the severity of the dorsalisation phenotype correlates with decreasing gene dosage of bmpr1aa and bmpr1ab, with double mutant embryos always exhibiting a C4 dorsalised phenotype. Importantly, this gene dosage effect was also observed on LR patterning, with 80% of embryos of genotype bmpr1aa−/−;bmpr1ab+/− presenting with a cardiac laterality defect (Figure 2H). Interestingly, loss of the bmpr1aa paralogue affected phenotypic severity more robustly than loss of the bmpr1ab. Unfortunately, we were unable to score the cardiac laterality phenotype of double mutant embryos as no cardiac field was detected in these embryos (Figure S4), consistent with previous observations that Bmp signalling is required for cardiac specification [32], [33]). These results demonstrate that the bmpr1a paralogues play partially redundant roles in both dorsoventral and LR patterning.
Since no role for Bmp1a in LR axis formation has been reported thus far, we further investigated how Bmp signalling via Bmpr1a regulates LR patterning. We analysed the expression pattern of marker genes whose expression is controlled by LR patterning in embryos derived from an incross of bmpr1aa+/−;bmpr1ab+/− parental fish. Expression analysis of the Nodal-related gene spaw revealed that embryos that retained at least one wild-type copy of bmpr1aa, displayed normal spaw expression (Figure 3A). However, embryos that had lost both wild-type copy of bmpr1aa and retained at least one wild-type copy of bmpr1ab displayed strong and bilateral expression of spaw in the entire LPM (Figure 3B). On the contrary, in embryos that had lost all wild-type copies of bmpr1aa and bmpr1ab we observed a reduction of spaw expression in the LPM by in situ hybridization (Figure 3C) and quantitative RT-PCR (Figure S5). The bmpr1aa/bmpr1ab double mutant embryos displayed a strong (C4) dorsalised phenotype resulting in a curling of the tail region. Although a Kupffer's vesicle was present in these embryos (data not shown), the structure of the tail is suspected to have physically intervened with the potential of the Kupffer's vesicle to activate and/or propagate spaw expression in the posterior LPM.
Similar disruptions to asymmetric gene expression were observed upon analysis of lefty1 expression in the cardiac field and pitx2 expression in the gut region. Expression of lefty1 was restricted to the left cardiac field in embryos that retained at least one wild-type copy of bmpr1aa (Figure 3D). Embryos that had lost both wild-type copy of bmpr1aa and retained at least one wild-type copy of bmpr1ab, however, displayed a clear bilateral expression of lefty1 in the cardiac field (Figure 3E). Since embryos without any wild-type bmpr1a gene lack the entire cardiac field, no lefty1 expression was observed in the cardiac region of these embryos (Figure 3F). Furthermore, pitx2 is expressed in the posterior LPM and its expression is regulated by Nodal activity; this expression was unaltered in embryos that still possessed at least one wild-type copy of bmpr1aa (Figure 3G). Consistent with the observed spaw and lefty1 expression, pitx2 expression was also bilateral in the LPM of embryos that had lost both wild-type copy of bmpr1aa and retained at least one wild-type copy of bmpr1ab and was compromised in embryos that had lost all 4 copies of the wild-type bmpr1a gene (Figure 3H, 3I).
These results suggest that during LR patterning Bmp signalling via Bmpr1a regulates Nodal activity. To address the interrelation between Bmp and Nodal signalling we tested the possibility that Nodal acts downstream of Bmp signalling during cardiac laterality. Therefore we attempted to rescue the Bmp-dependent cardiac laterality defect by implanting Nodal-soaked beads in the anterior LPM (ALPM), in order to induce ectopic Nodal signalling. To block Bmp signalling, Tg(hsp70l:nog3) embryos were heat-shocked at 16 hpf which resulted in a cardiac laterality defect in almost all embryos (6 out of 7; Figure 4A–4C). Interestingly, when a Nodal bead was placed in the right ALPM of non-heat-shocked embryos the heart tube was displaced from the left side towards the midline in approximately 50% of the embryos (Figure 4C, 4F). This effect of the Nodal bead was even stronger when the bead was placed in heat-shocked Tg(hsp70l:nog3) embryos. The cardiac tube in such embryos with reduced Bmp signalling was directed towards the right-sided bead in nearly 70% of cases (Figure 4D–4G). In a similar experiment using MZbmpr1aa mutant embryos we again observed that the cardiac tube was directed towards the Nodal bead in 75% of embryos examined (Figure 4H–4J). Together these results suggest that during generation of cardiac laterality Bmp and Bmpr1a act upstream of, or in parallel with, Nodal.
Our observation that spaw is ectopically expressed in the right LPM mesoderm in embryos that had lost 2–3 copies of their wild-type bmpr1a indicated that Bmpr1a is normally required to repress spaw expression in the right LPM. For this to be a direct effect of Bmp signalling it is expected that Bmp signalling is elevated in the right LPM, as recently reported studying mouse embryos [17]. Although we previously reported on elevated Bmp activity in the left anterior LPM before the cardiac tube is formed (22-somite stage)[34], we never observed enhanced Bmp activity in the right posterior LPM using an anti phospho-Smad1,5,8 antibody (data not shown). An alternative to the model in which Bmp activity directly regulates spaw expression in the right LPM is a model in which Bmp activity regulates spaw expression in an indirect manner. It is well established that Lefty1 in the midline is required to prevent Nodal protein produced in the left LPM from crossing the midline and inducing Nodal expression ectopically in the right LPM [9]. We, therefore, systematically analysed lefty1 expression in the midline of embryos with a gradual loss of Bmpr1a signalling. Doing so, we observed that embryos with 4 or 3 copies of the wild-type bmpr1a gene displayed normal and robust lefty1 expression in the embryonic midline (Figure 5A). Analysis of lefty1 expression in embryos that had lost 2 or 3 copies of the wild-type receptor gene, we observed an increase in the number of embryos with reduced lefty1 expression levels in the midline. Embryos that had lost all 4 copies of the wild-type bmpr1a gene consistently showed a near loss of all lefty1 expression (Figure 5A).
To address whether a disruption of fluid flow in Kupffer's vesicle might explain the reduced lefty1 expression we analyzed the lrrc50hu255h mutant, a loss-of-function allele of a conserved cilia protein that is required for cilia motility [35]. We observed that in the majority of lrrc50hu255h mutant embryos lefty1 was robustly expressed in the midline (Figure S6), suggesting that the observed reduction of lefty1 expression in bmpr1a mutant embryos was not a consequence of a disruption in Kupffer's vesicle function.
To test whether Bmp activity can regulate lefty1 expression in the midline we analysed lefty1 expression in embryos with increased or reduced Bmp activity. We manipulated levels of Bmp signalling by performing heat-shock experiments on embryos carrying the Tg(hsp70l:bmp2b) or Tg(hsp70l:nog3) transgenes, allowing temporally controlled upregulation or downregulation of Bmp signalling, respectively. The embryos were heat-shocked after gastrulation to prevent strong effects on dorsal-ventral patterning due to altered Bmp signalling levels and lefty1 expression was analysed at somitogenesis stages. Consistent with the data from the bmpr1a mutant analysis, we observed an upregulation of lefty1 expression in embryos with ectopic Bmp activity, intermediate levels of lefty1 in wild-type embryos and reduced lefty1 expression in embryos with reduced Bmp activity (Figure 5B). These results demonstrate that Bmp signalling is both required and sufficient for lefty1 expression in the midline.
Thus far it has been proposed that lefty1 expression in the midline is directly regulated by Nodal protein produced in the LPM [36]. Importantly, our observations demonstrate that lefty1 expression also requires Bmp signalling. Next, we wanted to address whether the regulation of lefty1 by Bmp is Nodal (in)-dependent. The suggestion that Nodal activity itself is not sufficient to induce lefty1 expression in the midline arises from our observation that upon reduction of Bmp signalling, lefty1 expression was reduced while spaw was still strongly expressed in the LPM. In addition, our observation that upon ectopic expression of bmp2b, lefty1 expression is induced while spaw expression is lost goes further to suggests that Bmp signalling can induce lefty1 expression in the absence of Nodal activity. We, therefore, wanted to address whether Bmp signalling regulates lefty1 expression in the midline independent from its regulation by Nodal activity. To investigate further the possibility that Bmp induces lefty1 expression in a Nodal-independent manner, we incubated embryos with the Nodal inhibitor SB431542 from tail-bud stage until the time point of analysis (18 hpf) [37]. As expected, we observed that in wild-type embryos treated with SB431542, spaw expression was compromised in the LPM, indicating the efficiency of the SB431542 treatment in blocking Nodal signalling (data not shown). When wild-type embryos or Tg(hsp70l:bmp2b) that were not subjected to heat-shock (both exhibiting wild-type Bmp levels) were treated with SB431542, we observed a loss of lefty1 expression from the midline. These results demonstrate that Nodal activity is indeed required for lefty1 expression at this stage, which is consistent with previous reports [5], [38]. However, when heat-shock induced Tg(hsp70l:bmp2b) embryos were directly treated with SB431542, lefty1 expression was induced in the midline. From these results we can conclude that in embryos with wild-type Bmp activity, Nodal is essential to drive robust expression of lefty1 in the midline. In addition, these results demonstrate that when Bmp signalling is ectopically activated, lefty1 expression is induced independently of Nodal. When comparing the level of lefty1 induction in heat-shocked Tg(hsp70l:bmp2b) embryos with or without the SB treatment we observed less ectopic lefty1 expression in the presence of the SB inhibitor (comparing Figure 6B and 6D). This result suggests a synergistic effect of Nodal and Bmp on lefty1 expression.
Thus far our results suggest that Nodal and Bmp regulate lefty1 expression in the midline independent from each other. To confirm such an independent regulation we tested whether Nodal can regulate lefty1 expression independent from Bmp signalling. To block Bmp activity Tg(hsp70l:nog3) embryos were heat-shocked at tail-bud stage (10 hpf), which resulted in reduced expression of lefty1 in the anterior midline at 18 hpf (Figure 6E, 6F). To induce Nodal in Bmp-depleted embryos, a Nodal bead was placed in the ALPM. As a consequence of Nodal bead implantation we observed restoration of the anterior lefty1 expression even in the absence of Bmp signalling (Figure 6G, 6H). These results demonstrate that Nodal can activate lefty1 expression independent from Bmp and confirm that Bmp and Nodal regulate lefty1 expression independent from each other. Together these results indicate that lefty1 expression is regulated by at least two parallel pathways involving Nodal and Bmp.
Finally, we wanted to address whether the observed effect of Bmp on spaw expression in the LPM is direct or indirect via its proposed role in regulating lefty1 expression. In Tg(hsp70l:bmp2b) embryos that were heat-shocked at the tail bud stage, we observed a strong down-regulation of spaw expression in the LPM (Figure 7A, 7B), which was coupled with ectopic lefty1 expression (Figure 5B). To test whether the upregulation of lefty1 in the midline was responsible for the downregulation of spaw expression in the LPM, we performed lefty1 knock-down by injecting embryos with a previously published morpholino that effectively targets lefty1 [39]. Interestingly, injection of the lefty1 MO in heat-shock induced Tg(hsp70l:bmp2b) embryos resulted in restoration of spaw expression in the left LPM, with ectopic expression also observed in the right LPM (Figure 7D), similar to non-heat-shocked embryos (Figure 7C). These results demonstrate that lefty1 expression in the midline is required for Bmp to repress spaw expression in the LPM and acts as an intermediary between Bmp signalling and spaw expression.
We describe here the identification of two novel zebrafish bmpr1a mutants; a bmpr1aa mutant allele from a forward genetic screen for laterality mutants and a bmpr1ab mutant allele by screening a mutagenized library. By generating and analyzing compound heterozygous and double mutant embryos for bmpr1aa and bmpr1ab, we observed a strong correlation between the number of wild-type bmpr1a gene copies being lost and the severity of the LR patterning defects observed. Most strikingly we observed a shift from the normal unilateral expression of the Nodal-related spaw gene in the left LPM to a bilateral spaw expression in both the left and the right LPM. This shift was accompanied by a reduction in the expression of lefty1 at the midline. This demonstrates that Bmp signalling regulates normal unilateral Nodal activation in the LPM, an observation supported by Nodal bead implantation in the LPM that restored cardiac laterality in Bmp-deficient embryos. Mechanistically our data suggests that there are two parallel pathways, a Bmp and a Nodal dependent pathway, to promote lefty1 expression in the midline and regulate LR patterning (see Figure 8 for proposed model). This model also explains the observation made in several animal models that ectopic Bmp signalling downregulates Nodal activation, suggesting that Bmp signalling is required on the right side to repress Nodal activation [13], [20], [22], [38]. Our data now demonstrates that, at least in zebrafish, this regulation of Nodal activity by Bmp is indirect and depends on the activation of lefty1 expression, as was demonstrated by knock-down of lefty1 in embryos with elevated Bmp signalling (Figure 7). Expanding the previous reaction-diffusion model of an agonist (Nodal) and antagonist (Lefty1), we can now include an additional level of regulation, in which Bmp induces Lefty1, which is required to establish unilateral Nodal activity in the LPM.
Lefty1 is essential for formation of the LR axis [9]. Loss of Lefty1 in mouse embryos results in a left-isomerism, whereby left-sided genes become expressed bilaterally. These described effects are very similar to those observed upon reducing Bmpr1a levels or Bmp signalling in the zebrafish embryo shown here. Others have reported that expression of Lefty1 in the midline is dependent on Nodal activity from the LPM in both zebrafish and mouse embryos [5], [36], [38]. Detailed analysis of the Lefty1 promoter region by Saijoh and colleagues identified a 1.2 kb upstream region of the Lefty1 gene that was sufficient to drive its midline expression [40]. In addition, it was reported that although Foxh1 binding sites are present in this upstream promoter region, these were not required to drive Lefty1 expression in the midline [36]. This suggests that, besides Nodal, additional factors are required for inducing midline Lefty1 expression. Indeed our data demonstrate that during zebrafish LR axis formation, Bmp signalling is required and sufficient to drive lefty1 expression in the midline. Firstly, we found that in mutants with reduced copies of the wild-type bmpr1a gene, lefty1 expression is gradually lost from the midline. Secondly, in transgenic embryos that ectopically express noggin3, a potent Bmp antagonist, lefty1 expression is diminished from the midline while Nodal signalling is still active (indicated by bilateral spaw expression). Thirdly, ectopic activation of the Bmp signalling pathway using a Tg(hsp70l:bmp2b) transgenic results in elevated and ectopic expression of lefty1 in the midline.
Our experiment using ectopic Bmp signalling in the absence of Nodal activity demonstrated that under conditions where Bmp signalling is sufficiently high, Nodal is not required to induce lefty1 in the midline. This might be important during early stages of LR axis formation. Based on the following observations, we hypothesize that at the initiation of LR axis formation, lefty1 expression in the midline is initiated by Bmp signalling independently of Spaw activity. Firstly, in zebrafish embryos lefty1 expression in the midline was observed at the 1–3 somite stage while spaw expression is initiated only at the 5-somite stage in the perinode region and at the 10-somite stage in the LPM ([5] and unpublished observations M. Verhoeven, E. Noël and J. Bakkers). Secondly, this initial lefty1 expression was unaffected by the injection of MOs that efficiently targeted spaw [5]. Thirdly, at these early somite stages expression of Bmp ligands is very strong in the tail bud region [41]. When blocking all Bmp signalling at this early stage in heat-shocked Tg(hsp70l:nog3) embryos, lefty1 expression was indeed not initiated in the midline. Together these results suggest that at the initiation of LR axis formation, lefty1 expression in the midline is initiated by Bmp while the maintenance of lefty1 expression in the midline requires both Nodal and Bmp (see model in Figure 8).
Although the zebrafish has been used extensively to identify new regulators by conducting forward genetic screens, there has been very limited success identifying novel mutants displaying LR patterning defects [42], [43]. This might be due to the variability and mixture of the phenotypes that can be observed (situs inversus, situs ambigious or situs solitus) as well as the natural occurrence of these phenotypes in the commonly used wild-type strains. Alternatively, an earlier and essential function of the gene product in embryo development masking any LR defects would hamper the identification of such LR genes. In addition, redundancy with paralogous genes often present in the zebrafish genome can mask the full loss-of-function phenotype. In lin mutant embryos, two copies of the wild-type bmp1aa gene are lost while the two wild-type bmpr1ab copies are still present. The lin/bmpr1aa mutant embryos displayed heart-specific laterality defects (although not fully penetrant) without displaying any gut laterality defects. Previously, we showed temporally distinct requirements for Bmp signalling functions during both LR axis formation and heart morphogenesis [13], [34]. The heart-specific laterality defect of lin/bmpr1aa mutant embryos (eg. loss of leftward cardiac jogging and rightward cardiac looping) is very similar to the cardiac laterality defect previously observed in the lost-a-fin/alk8 mutant. This suggests that during these processes Bmpr1a/Alk3 and Acvrl1/Alk8 play non-redundant functions similar to those described for these receptors during dorsoventral patterning [31]. These results also imply that either the regulation of heart laterality is more sensitive to reducing Bmp signalling activity than the digestive system or that this process is less compensated by wild-type maternal bmpr1aa RNA present in the oocyte. In agreement with the latter suggestion are the observations that bmpr1aa is maternally provided in the oocyte and that maternal zygotic (MZ)lin/bmpr1aa mutant embryos (from surviving lin/bmpr1aa homozygous females) displayed an increase in the strength of the LR patterning defects, including gut laterality defects.
To our knowledge, this is the first report describing the requirement for Bmpr1a in regulating LR axis formation. Mouse Bmpr1a mutant embryos do not form mesoderm at embryonic day 7.5 and subsequently die before embryonic day 9.5, preventing the study of LR axis formation in these mutants [26]. Interestingly, the closely related mouse Acvr1/Alk2 gene has been implicated in LR patterning [16]. Since the Acvr1 mutant mouse embryos also die early due to severe gastrulation defects, chimeric embryos were produced and analysed for LR patterning. Depending on the relative contribution of mutant cells to the chimeric embryos, a variety of laterality defects were described. In chimeric embryos with a relative high contribution of Acvr1 mutant cells, bilateral expression of Nodal and Pitx2 in the LPM was observed in combination with reduced expression of Lefty1 in the midline. The phenotypes described for the chimeric embryos with Acvr1 mutant cells corroborate our observations in the Bmpr1a compound heterozygous/mutant embryos, suggesting a conserved role for Bmp type I receptors during LR axis formation.
The Bmp signal that regulates Lefty1 expression in the midline does so independent of Smad1, one of the three Bmp-specific Smad proteins. Although Smad1 inactivation in mouse embryos resulted in the activation of Nodal expression in the right LPM, Lefty expression in the midline was unaffected in such embryos [15]. Alternatively, Smad5 could be responsible for transducing the Bmp signal. Embryos lacking Smad5 no longer express Lefty1 in the midline, which is accompanied by bilateral Nodal and Pitx2 expression in the LPM [12]. Several observations in mouse suggest that during LR axis specification, Bmp signalling can also repress Nodal activation in the right LPM more directly and independently from its regulation of Lefty1 in the midline. As mentioned above, Smad1-deficient embryos showed bilateral Nodal expression while Lefty expression in the midline was reported to be unaffected [15]. In a study by Mine and co-workers, elevated phospho-Smad1,5,8 levels in the right LPM compared with the left LPM of mouse embryos was reported [17]. In addition, an increase on the left side of phospho-Smad1,5,8 levels was observed in Chordin and Noggin double mutant embryos, combined with a loss of Nodal and Lefty1,2 expression. However in Chordin;Noggin double mutant embryos, perinodal Nodal was also reduced and defects in the morphology of the node and the density of cilia were described, suggesting an additional defect in the transduction of a signal from the node to the LPM in such embryos. This defect in communication between the node and the LPM most likely also explains why we observed a complete lack of spaw expression in the LPM of bmpr1aa;bmp1ab double mutant embryos. In zebrafish embryos, we did not observe a stronger phospho-Smad1,5,8 level in the right LPM compared to the left side during LR specification. However, at later stages we did observe the opposite in the anterior LPM where phospho-Smad1,5,8 levels were increased on the left side [34]. In addition, our observation that ectopic Bmp signalling in the Tg(hsp70l:bmp2b) embryos can no longer repress spaw activation in the LPM when Lefty1 is absent makes it very unlikely that such a direct repression of Bmp signalling on spaw expression exists in the zebrafish embryo. Together this indicates that the regulation of lefty1 by Nodal and Bmp during LR axis specification is conserved amongst various vertebrate species. However there are species-specific differences as to what other activities Bmp signalling has during this process. Possibly, differences in geometry or scale of the embryos and speed of their development might require additional regulatory mechanisms to maintain the crucial but very unstable unilateral Nodal activation during LR axis specification.
All zebrafish strains were maintained in the Hubrecht Institute using standard husbandry conditions. Animal experiments were approved by the Animal Experimentation Committee (DEC) of the Royal Netherlands Academy of Arts and Sciences. The bmpr1ahu4087 mutant was identified during a forward genetic screen performed at the Hubrecht institute. ENU mutagenesis was performed as previously described for the creation of the Hubrecht Institute target selected mutagenesis library [44]. F1 progeny of mutagenised males were outcrossed to create approximately 300 F2 families, which were then incrossed. F3 progeny were screened for cardiac laterality defects at 28–34 hpf. The bmpr1ahu4087 mutant can be identified using nested PCR with the following primers:
PCR1
Forward primer: AGCTCATCCGGAGAAGTATG
Reverse primer: TCCACTTCATTTGTGTCACTG
PCR2
Forward primer: TGTAAAACGACGGCCAGT ATATGTACCCAGCCCTGATG
Reverse primer: AGGAAACAGCTATGACCAT AGCTTCAGATTCAGATCAACAC
The bmpr1absa0028 mutant was identified from the mutagenesis library at the Sanger institute by screening finclip DNA using nested PCR with the following primers:
PCR1:
Forward primer: CCAGACTACATGCTTCATG
Reverse Primer: ATTGTGACAGGCCTACAATG
PCR2:
Forward primer: TGTAAAACGACGGCCAGT CAGAAGATGCCACAAACAAC
Reverse primer: AGGAAACAGCTATGACCATGGTCACACCGAGTAATTTCC
Products were then sequenced with M13F or M13R primers.
Published transgenic lines used were Tg(hsp70ll:nog3)fr14 and Tg(hsp70ll:bmp2b)fr13 [13].
Meiotic mapping of the linkspoot mutation was performed using standard simple sequence length polymorphisms. The primers used for SSLP can be found on www.ensembl.org.
The lefty1 morpholino was described previously [39].
The coding region of the bmpr1aa gene was cloned into pCS2+ by PCR amplification. The lin mutation was introduced in the pCS2+ bmpr1aa construct using the QuickChange kit (Stratagene). In vitro transcription was performed from Acc65I digested template using the SP6 mMessage mMachine kit for all injected mRNA (Ambion).
SB431542 (Sigma) was resuspended in DMSO to a concentration of 10 mM, and subsequently diluted to a working concentration of 150 µM in embryo medium. Control embryos were treated with an equal volume of DMSO. 30 embryos were treated per 5 ml of SB/DMSO solution.
In situ hybridization was carried out as previously described [45]. Embryos were cleared in MetOH and mounted in benzylbenzoate/benzylalcohol (2:1) before pictures were taken. Riboprobes were generated by transcription from a linearized template in the presence of 11-UTP.
Agarose beads (Affigel blue, BioRad) were rinsed twice in PBS and incubated for 1 hr at 37°C with 50 µg/ml recombinant mouse Nodal protein (R&D systems). Implants were performed as previously described [46].
|
10.1371/journal.pntd.0007363 | Development and validation of a multiplexed-tandem qPCR tool for diagnostics of human soil-transmitted helminth infections | Soil-transmitted helminths (STH) are a major cause of morbidity in tropical developing countries with a global infection prevalence of more than one billion people and disease burden of around 3.4 million disability adjusted life years. Infection prevalence directly correlates to inadequate sanitation, impoverished conditions and limited access to public health systems. Underestimation of infection prevalence using traditional microscopy-based diagnostic techniques is common, specifically in populations with access to benzimidazole mass treatment programs and a predominance of low intensity infections. In this study, we developed a multiplexed-tandem qPCR (MT-PCR) tool to identify and quantify STH eggs in stool samples. We have assessed this assay by measuring infection prevalence and intensity in field samples of two cohorts of participants from Timor-Leste and Cambodia, which were collected as part of earlier epidemiological studies. MT-PCR diagnostic parameters were compared to a previously published multiplexed qPCR for STH detection. The MT-PCR assay agreed strongly with qPCR data and showed a diagnostic specificity of 99.60–100.00% (sensitivity of 83.33–100.00%) compared to qPCR and kappa agreement exceeding 0.85 in all tests. In addition, the MT-PCR has the added advantage of distinguishing Ancylostoma spp. species, namely Ancylostoma duodenale and Ancylostoma ceylanicum. This semi-automated platform uses a standardized, manufactured reagent kit, shows excellent run-to-run consistency/repeatability and supports high-throughput detection and quantitation at a moderate cost.
| Soil-transmitted helminthiases are among the most prevalent and damaging neglected tropical diseases and have a significant global health impact. Accurate identification and quantitation of STH infection is a cornerstone of effective control. Direct observation and counting of eggs in faeces is the current gold-standard method for diagnosis of infection. This approach is time consuming and has poor sensitivity. As ongoing oral benzimidazole therapy across many endemic regions leads to a reduction in STH prevalence and intensity, these sensitivity limitations become an increasingly relevant issue, particularly with respect to monitoring treatment efficacy, identifying reductions in parasite transmission, and accurately quantifying infection burden in discrete populations in middle-income countries. PCR-based detection has long been proposed as an alternative approach to STH diagnosis and many protocols, including quantitative PCR-based methods, have been developed. However, these methods are largely bespoke and use non-standardized reagents that can greatly impact on the transferability and relative consistency of their performance. In the current study, we evaluate an automated, commercially-produced molecular diagnostic tool for validation of the major soil-transmitted helminths, including Ascaris lumbricoides, Trichuris trichiura, Necator americanus, Ancylostoma duodenale and Ancylostoma ceylanicum, and evaluate its performance in comparison to an established multiplexed qPCR using faecal samples from endemic settings.
| Soil-transmitted helminths (STH), including roundworms (Ascaris lumbricoides), whipworms (Trichuris trichiura) and hookworms (Necator americanus, Ancylostoma duodenale, Ancylostoma ceylanicum) represent a major cause of morbidity in tropical to sub-tropical developing and low-income countries [1]. In 2016, the Global Burden of Disease Study estimated that as many as 3.4 million disability adjusted life years (DALYs) are lost globally due to STH infections each year, of which 1.1, 0.5 and 1.8 million DALYs accounted for roundworm, whipworm and hookworm infections within all age groups in 2015 [2]. Total global infection prevalence is estimated to lie slightly above 1.9 billion infections depending on the diagnostic tool used, and as many as 5.3 billion people are at risk of infection worldwide [3, 4]. Infection risk directly correlates with inadequate sanitation, impoverished conditions, limited access to public health systems and population overcrowding [5]. Symptomology and severity of infection correlates with species of STH and intestinal burden, as well as host age, nutritional and health status [1]. Acute clinical symptoms are less prevalent, but include, for hookworm (Necator americanus and Ancylostoma spp.) and whipworm (Trichuris trichiura), anaemia and diarrhoea, and for roundworm (Ascaris lumbricoides), intestinal blockage and/or rupture, leading to ~135,000 deaths per year [6, 7]. Ultimately, the burden of disease caused by morbidity is far more significant than the impact caused by mortality, with long-term sequelae including malnutrition, stunting, wasting and decreased cognitive development [8].
Control of STH infection is dependent on oral anthelmintic therapy using benzimidazoles (BZ) [9]. Currently, the World Health Organization (WHO) recommends, regional mass drug administration (MDA) programs in endemic populations to deliver 400 mg single dose albendazole or 500 mg single dose mebendazole annually or biannually to reduce both infection prevalence and intensity [9]. These MDA programs follow the London Declaration on Neglected Tropical Disease (NTD) 2012 endorsement of the WHO goal to scale up global deworming in order to treat 75% of pre- and school-aged children at least once a year until 2020 in order to decrease STH burden in endemic areas [1]. Quantifying STH burden and the efficacy of BZ MDA programs is dependent on accurate and sensitive diagnosis and quantification of infection, particularly in regions where infection intensity is low but prevalence remains relatively high. Methods used include direct microscopy, formalin-ether concentration method, the McMaster egg counting technique, simple sodium nitrate flotation (SNF) including FLOTAC, the Kato-Katz thick smear (KKTS) and, more recently, PCR-based approaches [10]. The observation and enumeration of STH eggs in faecal samples is the current gold standard for diagnosis of STH infection, with the WHO-recommended, Kato-Katz thick smear the most widely used approach [11]. Advantages of the Kato-Katz method are its cost-effectiveness and its application in remote settings [11]. However, there are several drawbacks of this method, such as the requirement for immediate microscopic examination of multiple freshly collected faecal samples and technicians exhibiting parasitological expertise, a lack of standard protocols (e.g., for the amount of faecal matter examined, fixation method, calculation of eggs per gram faeces) and hence reproducibility, time consumption, labour intensity, the need for rapid assessment of faecal samples to avoid hookworm egg clearance and most importantly the underestimation of infection prevalence due to limited sensitivity [12]. The technique also requires large teams of technicians and equipment to be dispatched to remote communities with an additional requirement for electricity and running water, which is logistically challenging in remote areas. More recently, Inpankaew and colleagues (2014) have established the sodium nitrate flotation method, a direct faecal microscopy-based method used widely in veterinary settings in the past, of which a single application has a 6% higher sensitivity than KKTS (performed in quadruplicate over a two day period to reduce sensitivity limitations) for the detection of hookworm eggs in human stool at any given time point [13].
Copro-microscopic diagnostic methods in combination with regularly administered drug treatment are ideal in highly endemic regions; however, they are not representative in areas containing sub-populations with low prevalence and intensity infections [14]. Further, STH infections are overdispersed, with a majority of the infected harbouring moderate to light intensity infections (roundworm 1–49,999 epg, whipworm 1–9,999 epg, hookworm 1–3,999 epg) and only a minority suffering from high intensity infections (roundworm > 50,000 epg, whipworm > 10,000 epg, hookworm > 4,000 epg) [1, 15]. Consequently, even within highly endemic regions worldwide, the majority of infections may not be readily detectable by microscopy [16, 17]. Although the disease burden in regions of the world has decreased significantly over years of MDA control and increased socioeconomic development, STHs remain a major global human health issue [18]. It is assumed that with successful implementation of MDA treatment a larger decrease in infection intensity than infection prevalence can be observed [19], making the sensitivity limitations of the diagnostic tool a more significant issue than the cost associated with molecular diagnostic approaches [20], as a greater proportion of the population harbours low intensity infections that can potentially be missed by copro-microscopic diagnostic approaches [21]. While these low-grade STH infections may be of a lesser consequence in terms of disease burden, they are highly relevant in terms of any effort to interrupt infection transmission [22], seeing that adults in endemic countries represent infection reservoirs that inhibit the chance of an interruption of transmission cycle by school-based MDA [23]. One major limitation of MDA treatment is the inability to prevent re-infection once treatment has ceased, resulting in rebounding infection levels among targeted communities [24, 25]. Transmission of infection, with a particular focus on populations in low infection intensity settings, needs to be interrupted to stabilize control of STH infections [26], but is influenced by infection prevalence, infection intensity, human migration, regional demography, diagnostic tool application and drug efficacy [27]. Diagnostic tools contribute to reducing transmission by providing a knowledge basis for the specific, targeted and sustainable treatment of sub-populations at risk which serve as a transmission reservoir [28]. Notably, MDA treatment should be administered in combination with other intervention methods to reach satisfactory health outcomes [29]. There is a global need for an appropriate, rapid, cost-effective and sensitive tool for detection of STH infections in order to decrease burden of disease by identifying low-intensity infections and a subsequent targeted sustainable reduction in worm burden [30].
Real-time qPCR methods have recently been tested for STH diagnosis, targeting species-specific gene markers such as ITS-1 or ITS-2 [31–33]. However, these methods currently have limitations that make them not applicable in field settings such as the need for trained scientific personal [31]. Particularly challenging is the transfer of customized qPCR methods among laboratories with a requirement for optimization via significant molecular biological expertise. A reliable, automated diagnostic tool could have the potential to overcome issues related to reproducibility and create a standardized method of STH infection detection.
In the current study, we evaluate a multiplexed-tandem PCR (MT-PCR) based assay to differentiate, identify and quantify each major STH species in genomic DNA isolated directly from stool samples. The method is user-friendly, has high sensitivity and specificity and is produced as a standardized kit that is commercially available and readily transferrable to other laboratories. The method is semi-automated and requires little a priori expertise in molecular diagnostics or parasitology. Although unlikely to be cost-effective for routine diagnostics at the present time, the method provides a useful research tool for epidemiological studies of STHs in endemic regions, particularly in populations where prevalence and intensity of infection is highly variable and the limitations of direct egg counting by microscopic examination is impractical or insufficient.
Written informed consent for this prospective study was received from all study participants, or from parents or guardians for participants under the age of 18 years in Timor-Leste and Cambodia respectively. Ethics approval for the Cambodian based part of this study was provided by the National Ethics Committee for Health Research of the Ministry of Health in Cambodia (269NECHR, 27th of June 2016) as well as by the Human Research Ethics Committee of the University of Melbourne (1647208). Ethic approval for the Timor-Leste based part of this study has been received from the Human Research Ethics Committees at the Australian National University (2015/111) and the Timor-Leste Ministry of Health (2015/196).
In collaboration with an industry partner (AusDiagnostics Ptd. Ltd., Australia) we have used an established commercially available multiplexed molecular diagnostic platform, the Easy-Plex system, to develop an assay targeting human STH infections in faecal DNA samples [37]. Multiplexed-tandem polymerase chain reaction (MT-PCR) tests were developed targeting the β-tubulin 1 locus of Ascaris lumbricoides (Genbank accession number FJ501301.1), Trichuris trichiura (Genbank accession number AF034219.1), Necator americanus (Genbank accession number EF392851.1), Ancylostoma duodenale (Genbank accession number EF392850.1), and targeting the second internal transcribed spacer region of the nuclear ribosomal RNA gene (ITS-2) for Ancylostoma ceylanicum (Genbank accession number JN164660.1). The PCR primers for these assays are held in commercial confidence by AusDiagnostics Pty. Ltd.
The MT-PCR is a nested PCR method consisting of a primary PCR performed in multiplex on a Gene-Plex CAS1212 liquid handling robot (AusDiagnostics, Pty Ltd., Australia), followed by a secondary, tandem real-time qPCR (in which testing is conducted in single-plex in a tandem battery of reactions) run on a LightCycler 480 system (Roche, Switzerland). At the initiation of testing, DNA samples are placed on the Gene-Plex robot deck in a designated loading area and provided a unique sample ID (entered manually into a computer interphase linked to the Gene-Plex robot) which is carried over throughout the MT-PCR testing. Upon initiation of the amplification protocol, the robot distributes 5 μl of each sample DNA to a separate 20 μl reaction well in the “multiplex reaction strip tubes” provided by the STH MT-PCR kit manufacturer. No template controls (dH2O) are included with each MT-PCR run and carried over through both amplification phases. Each well of this strip contains lyophilized standardized quantities of each external primer pair for each STH species and an internal spike (consisting of 10,000 copies of a 120 bp, heterologous, synthetic oligonucleotide) to control for PCR inhibition and provide a quantification standard for each sample. The initial multiplex step is performed in a conventional thermocycler unit installed on the Gene-Plex robot deck. This amplification phase consists of 15 cycles with the following parameters: denaturation at 95°C for 10 seconds, annealing at 60°C for 30 seconds and extension at 72°C for 20 seconds with no initial denaturation or final extension phase.
Following the initial multiplex PCR, all first round product amplicons are diluted by the Gene-Plex robot (this is done to eliminate primary carry-over and PCR inhibition) and used as a template for the secondary tandem real-time PCR step. Tandem PCR is conducted on a 384-well plate in 20 μl reactions in a reaction master mix containing SYBR Green I/HRM dye. Each well of the 384-well plate contains lyophilized internal PCR primers for one target species or the internal spike control per reaction arranged in tandem array (i.e., multiplex amplicon from one sample is loaded onto a tandem array of six consecutive reaction wells, with each well containing one amplicon specific primer pair). Following loading of the tandem PCR plate, the plate is sealed using a self-adhesive heat-adherent film (MSB1001, BioRad, USA) and then loaded onto a LightCycler 480 system (Roche, Switzerland). The secondary amplification consisted of 30 cycles with the following profile: denaturation at 95°C for 10 seconds, annealing at 60°C for 15 seconds and extension at 72°C for 15 seconds with initial denaturation at 95°C for 10 minutes and no final extension phase.
Following the tandem real-time PCR phase, each amplicon is subjected to high-resolution melt-curve (HRM) analysis (from 72°C to 95°C). Following HRM, the melt profile for each amplicon for each sample is assessed for quality, purity (based on melt peak number, height and width) and specific identity (based on estimated melting temperature relative to control parameters determined using purified positive control DNAs for each targeted species during initial assay development by AusDiagnostics Pty. Ltd.). A positive or negative test call was determined based on these HRM results. Quantity of each test-positive amplicon was then calculated based on its amplification cycle threshold (Ct-value) relative to the cycle threshold of the sample-specific internal spike control. From the initial loading of the sample DNAs onto the Gene-Plex robot, all steps of the MT-PCR method are automated, including the determination of positive test results and their quantitation, which is provided at the end of the reaction in a tabular and graphical format by the software (MT Analysis Software, AusDiagnostics Pty. Ltd., Sydney) used to run the reaction protocol (MT Assay Setup Software for the multiplex-PCR phase, AusDiagnostics Pty. Ltd., Sydney; and LightCycler 480 Software Version 1.5 for the tandem-PCR phase, Roche, Switzerland). Upon initiation of the initial multiplex reaction protocol, the only additional operator input required prior to final test results is the application of a heat-sealing film between the multiplex and tandem-PCR phases and the transfer of the sealed tandem-PCR reaction plate from the Gene-Plex robot to the Lightcycler.
All samples tested in the current study were assessed previously for STH infections by multiplex qPCR [34, 35]. We evaluated both the quantitative and qualitative diagnostic performance of the MT-PCR method against the previous molecular test results. Run-to-run variation was assessed for the assay by testing all 462 (5 μl) faecal DNA samples from Timor-Leste in duplicate, with disagreements (n = 31; 6.7%) tested in triplicate. As this testing showed high run-to-run consistency between replicates, subsequent testing of samples from Cambodia (n = 302) has been performed in single replicate (5 μl). Diagnostic sensitivity of the MT-PCR was assessed using the previously published multiplex qPCR as the diagnostic gold standard; i.e., we defined a “true positive” as samples that are positive by both molecular diagnostic methods (multiplex qPCR and MT-PCR) and a “true negative” as samples that are negative by both methods. MT-PCRs that disagreed with the multiplex qPCR were defined as “false positive” or “false negative” respectively. These potentially “false positive” or “false negative” samples were restested via the previously published conventional qPCR with slight modifications. Briefly, we tested all disagreements using a single-plex qPCR approach for the respective gene targets (Ascaris lumbricoides n = 31, Ancylostoma ceylanicum n = 1, Trichuris trichiura n = 1 and Necator americanus n = 25). Tests were evaluated using 1x SensiFAST SYBR No-ROX mastermix (Bioline, UK), 2 μl template DNA and optimized primers as described elsewhere in a total reaction volume of 20 μl [32, 35]. The DNA quantification and melt-curve analysis for all tests was performed on the LightCycler 480 instrument (Roche, Switzerland) using the following conditions: 3 minutes at 95°C followed by 40 cycles of 9 seconds at 95°C and 30 seconds at 60°C with a final step-wise denaturation from 60°C to 97°C in 1.1°C/second increments. On retesting, we consistently found primer dimer formation at Ct values equal to 35 (n = 18) within the negative control of the N. americanus qPCR test after 35 cycles. For the purpose of verifying that sample degradation had not influenced MT-PCR testing, we considered any single-plex qPCR retests as negative if amplification was detected above Ct = 35.
Data and statistical analysis was performed using the Stata 12.1 (StataCorp LP, USA), Prism 7.0b (Graphpad Software Inc., USA), RStudio 1.1.463 (RStudio, Inc., USA) and Excel 15.4 software (Microsoft, USA). Power calculations for chi-squared tests were performed in RStudio 1.1.463 (RStudio, Inc., USA) using the pwr package 1.2–2 [38]. All tests used a significance level of p = 0.05 with a degree of freedom (df) of 2–1 (i.e., 1) (S3 Table). STH infection prevalence and intensity data were analysed in Excel 15.4 with visualization in Prism 7.0b. Interrater agreement values—described as Cohen’s κ—between both molecular diagnostic tests, 95% confidence intervals, significance of κ (using a p-value threshold for statistical significance of 0.05), sensitivity and specificity of MT-PCR have been determined using the Stata 12.1 software package. Cycle threshold values of MT-PCR vs qPCR were plotted against each other in a scatterplot and a simple linear regression analysis was performed using Prism 7.0b with R2 values confirmed through regression analysis in Stata 12.1.
This study evaluated a total of 764 faecal DNA samples (462 Timor-Leste and 302 Cambodia) that had been previously tested by qPCR examination [34, 35]. Each sample was tested by MT-PCR as described above, of which 20.3% were positive for Ascaris lumbricoides, 36.4% were positive for hookworm (32.3% were positive for Necator americanus, 4.1% Ancylostoma ceylanicum) and 1.4% were positive for Trichuris trichiura. No infections were detected for Ancylostoma duodenale. Infection prevalence in Timor-Leste with any soil-transmitted helminth was 42.9% with species-specifc infections of 33.5% for A. lumbricoides, 2.4% for T. trichiura, 10.4% N. americanus and 1.1% Ancylostoma ceylanicum. Hookworm infections dominated in the Cambodian cohort with prevalence of 65.9% for N. americanus and 8.6% for Ancylostoma ceylanicum infections. Overall infection prevalence with any STH species was generally higher in Cambodia (70.5%) than Timor-Leste (42.9%) (Fig 1).
Chi-square table values for all investigated species are as follows: A. lumbridcoides 276 true negative, 31 false negative, 0 false positive, 155 true positive (n = 462); N. americanus 494 true negative, 23 false negative, 2 false positive, 245 true positive (n = 764); T. trichuris 451 true negative, 0 false negative, 1 false positive, 10 true positive (n = 462); Ancylostoma spp. 732 true negative, 1 false negative, 0 false positive and 31 true positive (n = 764). Total diagnostic infection values for any STH species are 307 true negative, 46 false negative, 2 false positive and 409 true positive (n = 764) as shown in Table 1.
Interrater reliability values, kappa (κ), showed a very good agreement between MT-PCR and conventional qPCR (Table 2), as defined by Cohen [39], as follows: <0.2 poor agreement, 0.20–0.40 fair agreement, 0.40–0.60 moderate agreement, 0.60–0.80 good agreement and 0.80–1.00 very good agreement. Total amount of agreement between both molecular diagnostic methods for true positive and true negative validation shows a percentage of agreement for all tests > 93.29%. Diagnostic specificity (true negative rate) of the MT-PCR method relative to multiplex qPCR was above 99.60% for all tests (Table 2). Diagnostic sensitivity (true positive rate) relative to multiplex qPCR ranged from 83.33% (Ascaris) to 100.00% (Trichuris) with three tests above 91.42%. The quantitative capacity of the MT-PCR method was compared by linear correlation of MT-PCR vs qPCR considering samples that were infection positive by both diagnostic methods (Table 2, Fig 2). Quantitative capacity was visualized considering all samples independent of infection prevalence by the various diagnostic methods (Fig 2). MT-PCR gene copy number estimates ranged from 28–70,059,042 with a median of 590,956.3 for A. lumbricoides, 41–1145.5 with a median of 84.5 for T. trichura and 15–781,465 for hookworms, with species-specific medians of 4228 for N. americanus and 4940 for A. ceylanicum, providing a possible initial starting point of inferring infection intensity measure/intestinal burden from gene copy number counts.
We have evaluated the multiplex-tandem PCR as a semi-automated diagnostic tool suitable for human STHs, comparing its performance to a multiplexed qPCR assay. Although copro-microscopic analysis is overwhelmingly the predominant method for diagnosis of STH infections, there is no clear gold standard for determining “true positives” in STH diagnostics [22, 40]. Based on this, we adopted the approach of considering a “false positive” or “false negative” by MT-PCR based on previous multiplex qPCR results for this dataset [34, 35], followed by re-testing of the disagreeing samples using a single-plex qPCR with an established ITS-1 or ITS-2 gene marker (Genbank accession numbers AB571301.1, FM991956.1, AJ001599.1, EU344797.1) [35]. Using this approach, two “false positive” tests (n = 2) in the MT-PCR validation (n = 764) determined relative to qPCR were confirmed as true positives, yielding a total diagnostic specificity of above 99.78% for each species assay.
We note that all faecal samples tested in the current study are samples collected from field sites in Timor-Leste and Cambodia in 2016 and have been stored at -20°C since their original testing by qPCR. After several freeze-thaw cycles, it is possible that the samples were influenced by DNA degradation, which may influence their subsequent detection. To test this, we retested all samples that were “false negative” by MT-PCR (n = 55), using the conventional qPCR protocol they originally tested positive by. Per the above, 38 of these samples retested as positive with an average Ct value of 30.80 (S1 Table). The remaining 17 samples yielded faint positives with the majority of Ct values between 30 to 35 (S1 Table). This indicates that Ascaris lumbricoides MT-PCR has a slightly lower sensitivity than the conventional qPCR, but that sample degradation may have impacted on some of our test results, and requires further evaluation with fresh field samples. Using this approach, total diagnostic sensitivity of the MT-PCR ranged from 90.64 (Ascaris) to 100.00% (Trichuris).
Overall, our study supports recent efforts to develop qPCR as a diagnostic alternative to or complement of faecal microscopy [41]. Each MT-PCR assay had very good agreement (kappa > 0.85) and strong quantitative correlation (R2 > 0.7548) to a recently published multiplexed qPCR for these species [35, 36]. Limitations of the quantitative correlation include limited numbers of T. trichuris positive samples (n = 10) as well as no Ancylostoma spp. single infection positive samples, which are required to present a complete assay evaluation (Table 2). Both standard qPCR and the MT-PCR methods have clear advantages over faecal microscopy in terms of sensitivity. The limited sensitivity of faecal microscopy for STH detection, particularly hookworm, is well documented in the literature [42, 43]. This has increasingly been noted as a challenge to sustainable STH control in regions where prolonged oral MDA programs have resulted in a substantial reduction in worm burden such that the majority of infections now fall below the WHO definition for “light intensity” infections [1]. This reduction will clearly impact on the global STH disease burden and with the increased economic development in many STH endemic countries as well as the potential for the emergence of drug resistant helminths, the focus needs to begin to shift toward sustainable control and STH transmission interruption [44]. Programs to monitor transmission and to allow efficient control of STHs in populations where regional MDA programs are no longer sensible will require more sensitive diagnostic approaches than faecal microscopy can provide [21]. As shown in previous studies of parasitic nematode infections, nested qPCR diagnostics show a higher sensitivity measure compared to that of a standard qPCR while including a quantitative estimation missing in traditional PCR approaches [45].
The primary limitations associated with a shift to a molecular or PCR-based diagnostic for STH infections include (i) the complexity of PCR application in endemic settings, (ii) a limitation in the transferability or reproducibility of bespoke PCR techniques and a need for standardization for clinical applications, (iii) the lack of clear relatability (conversion of molecular diagnostic Ct values to epg counts) of PCR-based test results to WHO treatment/burden guidelines, which is based on faecal egg densities and found to be on average 4x lower than egg intensity counts by molecular diagnostic for Necator americanus [36], and (iv) the relative cost differences between microscopic and PCR-based detection. The MT-PCR method developed here is built around a user-friendly, largely automated robotic platform that requires minimal molecular biological expertise. All kit reagents are produced and standardized by a commercial entity accredited for clinical diagnostic assay production, supporting rapid and reliable transfer among laboratories. The MT-PCR method is quantitative and further evaluation of the correlation between copy numbers and eggs per gram will provide meaningful information on helminth intensity as well as assist in its translation to assess intensity-related morbidity. However, as found in other studies investigating the use of multiplexed qPCR for STH diagnosis [40], correlations between qPCR results and egg intensities are moderate. These results can be inferred via logistic regression, but lack of knowledge about target gene copy number per genome limits our ability to make conclusions about helminth burden (particularly for loci, such as the nuclear ribosomal RNA gene, that are of an unstable copy number) [46]. Establishing these threshold levels will require additional screening of well characterized control samples of known egg densities as well as further field evaluation including patient specific health records to assess burden.
Finally, regarding cost, recent estimates for multiplex qPCR-based testing of STH infected stools put the cost per sample at US$2.61 [32] as a single-well and qPCR mastermix are used per sample (run in duplicate). The same study estimated true costs per sample for microscopy at US$2.60, indicating that the cost-effectiveness of microscopy is likely over-estimated. At present, the MT-PCR method is not competitive with these costs, with testing costing approximately US$7.31 (AUS$10.17) per sample for the current assay configuration following DNA extraction (this calculation does not include cost of liquid handling robot or qPCR thermocycler; approximate cost of DNA extraction US$ 6.85 (AUSD$ 9.64)). However, our focus here was on assay evaluation not on maximizing the efficiency or economics of the system. Overall, we acknowledge the current limitation of the platform, considering all resources associated, to laboratory-based most likely non-endemic settings. We have addressed the issue of to date limited data availability of PCR based helminth diagnostics approaches which perform with an increased sensitivity compared to currently performed copro-microscopic approaches [22].
In summary, we find the MT-PCR method to be a rapid, semi-automated and user-friendly molecular diagnostic tool for STH infection that provides comparable performance to conventional multiplex qPCR and superior sensitivity to faecal microscopy.
|
10.1371/journal.pgen.1003252 | Antagonism Versus Cooperativity with TALE Cofactors at the Base of the Functional Diversification of Hox Protein Function | Extradenticle (Exd) and Homothorax (Hth) function as positive transcriptional cofactors of Hox proteins, helping them to bind specifically their direct targets. The posterior Hox protein Abdominal-B (Abd-B) does not require Exd/Hth to bind DNA; and, during embryogenesis, Abd-B represses hth and exd transcription. Here we show that this repression is necessary for Abd-B function, as maintained Exd/Hth expression results in transformations similar to those observed in loss-of-function Abd-B mutants. We characterize the cis regulatory module directly regulated by Abd-B in the empty spiracles gene and show that the Exd/Hth complex interferes with Abd-B binding to this enhancer. Our results suggest that this novel Exd/Hth function does not require the complex to bind DNA and may be mediated by direct Exd/Hth binding to the Abd-B homeodomain. Thus, in some instances, the main positive cofactor complex for anterior Hox proteins can act as a negative factor for the posterior Hox protein Abd-B. This antagonistic interaction uncovers an alternative way in which MEIS and PBC cofactors can modulate Abd-B like posterior Hox genes during development.
| Hox genes encode transcription factors necessary to achieve the morphological differences between anterior and posterior regions of the body. These genes have been functionally conserved during animal evolution, and similar classes can be recognized in vertebrates and invertebrates. To bind DNA and regulate many of their targets, Hox proteins interact with the MEIS and PBC transcriptional cofactors. However, this is not always the case for the most posteriorly expressed genes belonging to the Abdominal-B class. Here we show a new interaction between the Abd-B protein and these cofactors where, rather than cooperating with Abd-B, the cofactors antagonize its function. Given the conservation of the Hox proteins and their cofactors, this new mode of interaction may be also happening in other species, including vertebrates.
| In segmented animals the differential anterior-posterior morphology is achieved during development under the control of the Hox genes [1]. Hox genes encode a conserved family of transcription factors organized in clusters in most animals. Hox clusters originated before the divergence of protostomes and deuterostomes and as a result, orthologous Hox genes can be identified between vertebrates and invertebrates that are more similar to each other than to other Hox genes in the same species [2],[3].
The development of the unique organs present in a segment is controlled by the Hox protein expressed in that segment through the regulation of specific downstream targets. In Drosophila melanogaster the Abdominal-B (Abd-B) protein (orthologous to Hox9/13 in mammals) induces the formation of the posterior spiracles in the eighth abdominal segment (A8) through the transcriptional activation of empty spiracles (ems), cut (ct) and spalt (sal) among other genes [4], [5]. Similarly, expression of the Sex combs reduced protein (Scr, orthologous to Hox5) in the labial segment of the head induces the formation of the salivary glands through the activation of fork head, trachealess and huckebein [6]; while expression of Ultrabithorax (Ubx) and Abdominal-A (Abd-A, both orthologous to Hox6/8) in the abdominal segments prevent the development of thoracic structures by repressing Distalless and buttonhead in the abdomen [7], [8]. The specific in vivo regulation of precise targets by each Hox protein contrasts with the observation that Hox proteins bind very similar DNA sequences in vitro [9]. In Drosophila, anterior and central Hox proteins (Lab, Pb, Dfd, Scr, Antp, Ubx and Abd-A [henceforth collectively referred to as anterior-Hox for simplicity]) bind TAAT sites with only the posterior-Hox Abd-B protein binding the slightly different TTAT sites [10], [11]. This in vitro lack of Hox DNA binding specificity is resolved in vivo by the use of protein cofactors that increase Hox DNA affinity and extend the binding site therefore increasing specificity for downstream target genes [12].
In Drosophila, the best-studied Hox cofactors are the Extradenticle (Exd) and Homothorax (Hth) proteins (homologous to the Pbx and Meis proteins in vertebrates) [13]. In vitro studies show that the anterior-Hox proteins bind poorly to many of their targets in the absence of Exd and Hth [14]. The Hth, Exd and Hox proteins form a trimeric complex that binds DNA with higher affinity than any of the proteins separately [15]. Exd directs the formation of the trimeric complex by binding directly to both Hth and the Hox protein. Exd can bind to various domains in the anterior-Hox proteins including the YPWM domain (present in all anterior-Hox proteins but not in the Abd-B posterior Hox proteins), the UbdA domain (present only in Ubx and Abd-A) and possibly to other domains not yet characterized [9], [16]. Hth binds Exd directly through the Homothorax-Meis (HM) domain [15] but there is no evidence of Hth binding to Hox proteins directly.
Exd translocation to the nucleus requires its binding to Hth [17]. Accordingly, Exd remains in the cytoplasm of cells that do not express Hth, while Exd is nuclear in cells expressing Hth. Moreover, in hth mutants Exd localization is cytoplasmic. These observations suggested a model in which Hth binding to Exd allows the translocation of Exd to the nucleus where it can bind to the Hox proteins forming the trimeric complex that binds target genes [15]. The requirement of this complex for normal Hox-target activation explains the homeotic phenotypes observed in hth or exd mutants even though they express correct levels of Hox proteins [18].
In contrast to the Exd/Hth requirement for anterior-Hox protein function, there is no clear evidence pointing to Abd-B interacting with these cofactors although such evidence exists in vertebrates for Hoxa9 protein interaction with PBX. Abd-B has two functions: a morphogenetic function (m) required for the formation of segment specific structures, and a regulatory function (r) that represses the transcription of anterior-Hox genes [19]. These functions correlate with the existence of two protein isoforms, which differ by the inclusion of a 5′ exon [20], [21],[22],[23]. Mutations affecting the Abd-Bm isoform result in embryos where the posterior spiracles are almost absent and the A5–A8 denticle belts resemble that in A4 indicating that Abd-Bm performs most of the morphogenetic functions [4], [19], [24]. Mutations affecting the Abd-Br isoform have minor defects in A8 but result in the formation of a small A9 denticle belt anterior to the anal pads indicating that Abd-B represses the formation of an A9 segment. However, the r isoform has some morphogenetic activity as heat shock induction of both the Abd-B m and r isoforms can induce the formation of posterior spiracles when ectopically expressed [25], [26], [27].
Contrary to the anterior-Hox proteins, addition of the Exd cofactor does not increase Abd-B's binding affinity to DNA [14]. As a result, the case for Hth and Exd interaction with Abd-B has not been studied in detail.
Here we investigate the interaction of Exd/Hth and Abd-B and find that, surprisingly, these cofactors interfere with Abd-B function during embryogenesis. We show that the presence of Exd/Hth interferes with Abd-B binding to its direct target empty spiracles (ems). This interference does not require binding of Exd/Hth to DNA and is probably achieved by Exd/Hth binding to the Abd-B homeodomain. These results uncover a novel Exd/Hth complex function and explain why in Drosophila exd and hth transcription is repressed by Abd-B protein. This novel interaction extends our understanding on the capacity of PBX MEIS proteins to modulate Hox output.
Despite the importance of Exd and Hth for anterior-Hox function, there is not much evidence pointing to the Abd-B proteins interacting with these cofactors. In fact, during embryogenesis exd and hth are initially expressed homogeneously along the trunk epidermis until stage 11 (st11) when their transcription is downregulated in the posterior abdominal segments [28], [29], [30]. To study more in detail the expression of Hth in the A8 and A9 segments we double stained with Abd-B antibodies and observed that Hth expression is downregulated in the dorsal region of A8 and A9 (Figure 1A–1B). This downregulation depends on Abd-B function as dorsal levels of Hth are restored in Abd-B mutant embryos (Figure 1C). In this region it had been described that Abd-B excludes Exd protein from the nucleus [[28] and Figure 1D′ insets] probably because its effect on Hth expression.
To test if the downregulation of exd and hth observed in wild type embryos is required for the normal development of the A8 segment we artificially maintained their expression using the Gal4 system. Expression of Exd in the ectoderm using the arm-Gal4 or the 69B-Gal4 lines driving UAS-exd results in embryos with normal cuticles (Figure 2A) and the same is true for UAS-hth (not shown). In contrast, coexpression of Exd and Hth gives rise to larvae with abnormal posterior spiracles and a reduced A8 denticle belt (Figure 2B–2B′). Interestingly, in many embryos a small A9 denticle belt forms (Figure 2B″), a phenotype also observed in the Abd-BUab-1 and Abd-BUabX23-1 loss-of-function alleles [19]. As these phenotypes could be caused by abnormal Abd-B expression, we stained embryos expressing ectopically both cofactors with anti-Abd-B. Using several Gal4 lines we observed that Abd-B localization in cells ectopically expressing Exd/Hth is normal (Figure 1E–1E′) showing that the transformations caused in the posterior segments are not due to altered Abd-B expression.
To find out at what level of the Abd-B genetic cascade the spiracle defects are caused we analyzed the expression of the early Abd-B targets [5]. We observe that the expression of ct and sal is downregulated in these embryos (Figure 3A–3D) suggesting that overexpression of Exd/Hth interferes with the normal activation of Abd-B downstream targets. The ems gene is required for spiracle development and its expression in the posterior spiracles is regulated by an enhancer that depends on Abd-B function [25]. We observe that expression of the ems spiracle enhancer is also downregulated in embryos overexpressing Exd/Hth (Figure 3E–3F). Taken together, these results suggest that in the presence of the Exd/Hth complex Abd-B proteins are less efficient in the activation of their direct targets. These results indicate that although Exd/Hth are positive cofactors of anterior-Hox proteins, they may also have a previously unnoticed negative effect on Abd-B function.
To test if Exd/Hth expression affects the function of both Abd-B isoforms, we first studied how the phenotypes obtained after ectopically expressing Abd-Bm are affected by the simultaneous expression of Exd/Hth. As previously reported with other Gal4 drivers [31], ectopic expression in the ectoderm of UAS-Abd-Bm with arm-Gal4 causes the formation of ectopic posterior spiracles in all trunk segments (Figure 4A) and the same is true if Abd-Bm is coexpressed with two irrelevant UAS constructs (see materials and methods). In contrast, simultaneous expression of Hth and Exd with Abd-Bm severely reduces the length of the ectopic spiracles (Figure 4B) confirming that Abd-Bm cannot fully function in the presence of these Hox cofactors.
Ectopic expression of the Abd-Br isoform with arm-Gal4 does not induce ectopic spiracles (Figure 4C) despite the fact that antibody stainings indicate that the protein is expressed at high levels (Figure S1A–S1C). This is probably due to Abd-Br having an inefficient morphogenetic function, as using stronger Gal4 lines or increasing the expression levels of Abd-Br by performing the experiment at 29°C, a temperature favouring Gal4 activity, results in the formation of ectopic posterior spiracles (Figure S1D–S1E). This confirms previous experiments using heat shock inducible constructs that demonstrated that both the Abd-Bm and r isoforms perform the morphogenetic function albeit Abd-Br is less efficient [25], [26], [27].
To test if Abd-Br is also competed by Exd/Hth we took advantage of the weak morphogenetic capacity shown by Abd-Br at 25°C (Figure 4C) and studied how varying the levels of endogenous Exd or Hth affects its function. Ectopic expression of Abd-Br in a hth homozygous mutant background induces ectopic spiracles similar to what Abd-Bm does in a wild type background (Figure 4 compare 4D to 4A), indicating that endogenous Hth can partially block Abd-Br activity. This effect is dependent on Hth protein concentration, as in heterozygous hth/+ embryos, expression of Abd-Br at 25°C can induce small ectopic spiracles (Figure 4E). We also observe that the ectopic spiracles appear to be more complete in the A2–A7 segments where antibody stainings show there are lower levels of endogenous Hth protein. Similar to hth mutants, ectopic Abd-Br expression in exd zygotic mutants results in the formation of ectopic spiracles (Figure 4G–4H). These spiracles are smaller than those observed in a hth mutant background probably due to the maternal exd contribution. Taken together, the above results show that the function of both Abd-B isoforms is sensitive to the Exd/Hth protein levels.
Abd-B has been suggested to control directly ems transcription in the spiracle through an enhancer located in a 1.2 kb region upstream of the promoter but the Abd-B binding sites mediating this interaction have not been identified [25]. As we found that the ems spiracle enhancer is downregulated in embryos where Exd/Hth expression is maintained (Figure 3E–3F), we decided to confirm its direct regulation by Abd-B and study how Exd/Hth can affect its expression.
Subdivision of the 1.2 kb fragment shows that the central 0.35 kb element is responsible for spiracle expression (Figure S2A–S2D). The 0.35 kb element is regulated by Abd-B and behaves like the original 1.2 kb fragment, responding to ectopic Abd-B expression (Figure S2E) and losing its expression in Abd-B null mutants (Figure S2F). Further reduction of the 350 bp element from the 5′ or the 3′ end abolishes spiracle expression (Figure S2I–S2L). This 350 bp element contains six putative Abd-B binding sites (TTAT) five of which are conserved in twelve Drosophila species analyzed (Figure S3, red boxes). Chromatin Immunoprecipitation (ChIP) in S2 cells transfected with HA-tagged Abd-B shows that Abd-B can bind the ems posterior spiracle enhancer in vivo (Figure S2G).
Electrophoresis mobility shift assays (EMSA) confirm the binding of Abd-B to the 350 bp element (Figure 5B). To test if all putative Abd-B sites in the 350 fragment are bound by Abd-B with equal affinity we made six similar sized oligos covering the whole fragment (Figure 5A grey boxes) and tested their capacity to compete for Abd-B binding to the whole 0.35 fragment. At high concentration all oligos, except oligo three that does not contain predicted Abd-B binding sites, can compete for Abd-B binding (Figure 5C). However, at lower concentrations only oligo 4 and oligo 6 are able to compete efficiently (Figure 5C) indicating that these sites have higher affinity for Abd-B.
To confirm that Abd-B binds the oligos through the predicted sites, we mutated the TTAT sites in oligos 4 and 6, and analyzed the capacity of Abd-B to bind these oligos in EMSA. Mutation of both putative binding sites in oligo 4 abolishes Abd-B binding to it (Figure 5E compare lane 3 with 15); with mutation of site 4A (lane 7) having a stronger effect than mutation of site 4B (lane 11) when mutated independently. Similarly, mutation of the putative binding site in oligo 6 strongly decreases its ability to be bound by Abd-B in EMSA (Figure 5E compare lane 18 with 22). This confirms that Abd-B binds to the predicted sites, and shows that in vitro each site binds Abd-B with different affinities.
We next tested the capacity of mutant and wild-type cold oligos 4 and 6 to compete the ems0.35 fragment for Abd-B binding. As expected, even at the high concentration, oligos with mutant Abd-B sites cannot compete for binding (Figure 5D). Mutation of the only putative Abd-B binding site in oligo 6 almost abolishes its ability to compete (Figure 5D compare lanes 19–22 with 23–26). Mutation of both putative binding sites in oligo 4 almost abolishes Abd-B binding (Figure 5D lanes 3–6 compared with 15–18); again with site 4A (lanes 7–10) having more effect than site 4B (lanes 11–14) when mutated independently.
To test their in vivo requirement, we mutated single Abd-B sites in the ems0.35 enhancer. While mutation of site 1 or site 2 does not affect spiracle expression noticeably (Figure 6B and Figure S2H), single mutation of putative sites 4A, 4B or 6 slightly reduces expression (Figure 6C–6E). Simultaneous mutation in ems0.35 of sites 4A and 6 strongly reduces spiracle expression (Figure 6F) with only occasional spiracles having residual expression; while mutation of sites 4A and 4B completely abolishes spiracle expression in all embryos (Figure 6G). These results are consistent with Abd-B controlling the expression of the ems spiracle enhancer by binding to several sites in an additive manner. These experiments and the deletion series show that sites 4A, 4B and 6 are necessary but not sufficient for spiracle expression, as fragments D and E that do not affect these sites also lose spiracle expression (Figure S2J, S2L).
To understand how Exd/Hth compete Abd-B activation of ems we first analyzed the capacity of these cofactors to bind the ems spiracle enhancer. In EMSA experiments we could not detect Exd/Hth binding to any of the six ems oligos (Figure 7A lanes 5,10,15,20,25,30) in conditions where we could detect Abd-B binding to oligos 4 and 6 (Figure 7A lanes 18,28 asterisks).
We next analyzed the effect of adding Exd, Hth or Exd/Hth to oligo 6 where the Abd-B site overlaps a predicted Exd/Hth site (Figure S3). Separate addition of Exd or Hth has a small effect on Abd-B binding to the DNA, while adding simultaneously Exd/Hth decreases the affinity of Abd-B for oligo 6 in a concentration dependent manner (Figure 7B lanes 12–14). Interestingly, adding Exd/Hth to oligo 4 that does not contain any predicted Exd/Hth sites also interferes with Abd-B binding (Figure 7C lanes 12–14) as efficiently as with oligo 6 where the predicted Abd-B Exd/Hth binding sites overlap. These results suggest that Exd/Hth interference with Abd-B is not due to competition for occupancy of overlapping binding sites.
As Exd/Hth does not bind oligos 4 and 6 in vitro, it is possible that interference with Abd-B binding to DNA is not due to competition for DNA binding but due to direct binding of Abd-B to the Exd/Hth complex.
In the embryo, there are several naturally expressed Hth isoforms. Some isoforms contain the DNA binding homeodomain, while others lack the homeodomain but still include the HM domain [32]. To test if Hth proteins without the homeodomain are capable of competing Abd-B function in vivo, we studied the hth100-1 allele that only affects the homeodomain containing isoform [32]. In hth100-1 embryos, ectopic expression of UAS-AbdBr at 25°C does not form well-developed ectopic spiracles as those formed in hthP2 alleles (compare Figure 4F and 4D), indicating that the homeodomainless Hth isoform can compete Abd-Br morphogenetic function in vivo. However, in these embryos some small spiracle structures are formed not seen in a wild type background (compare Figure 4F and 4C), suggesting that although the homeodomain of Hth is not strictly necessary, the full isoform competes Abd-Br posterior spiracle morphogenetic function more efficiently.
Exd binds anterior Hox proteins trough several domains, among them the YPWM domain. Although Abd-B lacks this element, many Abd-B like proteins contain at a similar position with respect to the homeodomain a conserved tryptophan (W) amino acid [33]. To investigate the possibility that Abd-B and Exd/Hth interact through this amino acid we analyzed the capacity of Exd/Hth to interfere with an Abd-B protein where this tryptophan residue has been mutated to Alanine (Abd-B W*). As shown in Figure 8A (lanes 8–10), mutation of this tryptophan does not prevent Exd/Hth interference with Abd-Bm DNA binding.
To investigate if the competitive interaction requires a particular Abd-B protein domain, we analyzed if the competition occurs with both Abd-B isoforms. We observe that adding Exd/Hth interferes with both Abd-Bm and Abd-Br isoforms binding to oligo 4 (Figure 8A lanes 3–5 and 13–15). In these experiments we observe the formation of different size bands (Figure 8A lane 1 black arrows and lane 11 grey arrowheads). These bands are specific as they are supershifted by anti-AbdB (Figure 8A lanes 2 and 12). As the smaller Abd-Bm band in lane 1 coincides with the larger Abd-Br band in lane 11 we interpret these bands as being the result of in vitro translation from internal Abd-B methionines. The observation that Exd/Hth can compete with even the smallest Abd-Br fragment suggests that the interference is due to interaction with the C-terminal end of Abd-B where the homeodomain is located.
To find out if Exd or Hth interact directly with the Abd-B C-terminal region we performed GST pull-down experiments. We observed that Exd interacts with GST fused to Abd-B313, a C-terminal fragment including the last 181a.a (from 313 to 493) (Figure 8B, lane 3). We observed that Exd interaction with this fragment is reduced if we remove the homeodomain (AbdB313ΔHD, lane 4) and found that Exd can interact directly with the Abd-B HD fragment (lane 5). In contrast, GST fused to the Abd-B homeodomain has a weak interaction with Hth (Figure 8C lane 3). However, when cold Exd is added to the mixture higher levels of Hth are isolated (Figure 8C lane 5). These experiments show that in the absence of DNA, the Exd/Hth complex can bind the Abd-B homeodomain, providing a mechanistic explanation for the observed in vivo antagonistic effect between these proteins.
The evolution of Hox proteins was fundamental for the acquisition of morphological differences in the antero-posterior axis of animals. Comparison between all extant animals indicates that the difference between posterior (Abd-B like) and anterior-Hox genes occurred early in evolution. This happened before the Hox and ParaHox clusters diverged, in what has been called a protoHox cluster [34]. This early divergence has resulted in Abd-B having a different character to all other anterior-Hox proteins, with the most striking difference being Abd-B binding to a TTAT DNA core site [and TTAC with lower affinity [35]] while other Hox proteins bind to a TAAT core [10], [11]. The divergence is also reflected at the protein sequence level with all anterior-Hox proteins having a YPWM motif that is absent or highly reduced from Abd-B like proteins [3], [33]. At the functional level, a major difference is the use anterior-Hox proteins do of the Exd/Hth complex as a positive cofactor to increase target DNA-binding efficiency, while Abd-B does not require it [14], [18]. Here we have shown in vivo and in vitro a new relationship between Abd-B and Exd/Hth where these positive Hox cofactors can also have an antagonistic interaction with Abd-B protein function, providing an explanation to why Abd-B represses the transcription of exd and hth genes during development.
The ems gene had been suggested to be a direct target of Abd-B in the posterior spiracles [25]. Despite being one of the first putative Hox targets analyzed, the lack of direct mutational evidence has resulted in ems being excluded from most Hox-target compilations [1], [9]. Here we have trimmed down this element to 350 bp and demonstrated that the spiracle enhancer is directly activated by Abd-B. This enhancer contains several sites with different Abd-B affinities all of which conform to the TTAT core sequence. Mutation of single sites does not eliminate enhancer expression, while simultaneous mutation of two high affinity sites abolishes the in vivo enhancer function. This suggests that similarly to yellow and bric-a-brac, two confirmed Abd-B direct targets analyzed to date [35], [36], Abd-B sites act additively. Our observation that mutating the low affinity Abd-B binding site 1 has no effect on the ems spiracle enhancer expression, while the deletion of the element abolishes expression (Figure S2H, S2L) indicates the presence of binding sites for cofactor or collaborator proteins acting in concert with Abd-B to achieve intrasegmental specificity.
Although additional bona fide targets should be analyzed to test how general is Exd/Hth competition on Abd-B function, our results indicate that this may be widespread during embryogenesis. We have found that induction of Exd/Hth in the A8–A9 segments not only affects the posterior spiracles, but also perturbs ectodermal cuticular structures controlled by Abd-B as well as downregulates the expression of the Abd-B early spiracle targets analyzed [37]. Moreover, it was described that the ectopic expression of Hth in the Drosophila melanogaster male abdomen causes a lack of pigmentation [15]. As it has been found that Abd-B induces male abdominal pigmentation by activating transcription of the yellow gene in the A5 and A6 segments [36], the effect of Hth expression on male pigmentation could also be explained by Exd/Hth interfering during larval development with the activation of yellow by Abd-B. Similarly, in the accompanying paper, Graba and collaborators [38] show that Dll repression by Abd-B in the posterior abdominal segments is also competed by Hth activation. This differs from our results as we only observe effects when both Exd and Hth are expressed in vivo. The difference may be explained as due to certain targets being more sensitive than others to the competition. In fact, our in vitro experiments show that Hth can bind weakly to Abd-B, and that this binding is increased by the addition of Exd (Figure 8C). The effect on Abd-B function, rather than a competition for binding sites in each specific target, could be due to a blocking interaction of Exd/Hth on Abd-B a possibility that is suggested by the direct binding we observe between Abd-B and the Exd/Hth complex. This is also supported by our observation that binding of Abd-B to oligo 4 is competed by Exd/Hth despite the absence of putative binding sites for these cofactors on this element. The direct interaction of Exd/Hth with the Abd-B homeodomain offers a plausible explanation for the observed antagonistic effect that Exd/Hth causes in vivo and in vitro.
Despite the many instances where we show competition between Abd-B and Exd/Hth during embryogenesis, there is at least one important case where the competition does not seem to happen, and this is the regulation of hth and exd transcription itself. Maternal and zygotic Exd and Hth proteins are expressed homogeneously along the antero-posterior axis until extended germ band (st11) when posterior Hox proteins downregulate their expression in the posterior abdomen [28], [29], [30]. Thus, at least in this case, the presence of Exd/Hth is incapable of blocking the Abd-B repressive function on hth transcription on the dorsal side of A8 and A9. Why competition does not occur on hth downregulation during this stage of embryogenesis is unclear. A simple explanation could be that although Abd-B function is also competed by the presence of Exd/Hth, Abd-B's maintained expression will eventually overturn the blocking effect of the Exd/Hth protein therefore repressing exd and hth transcription. Alternatively, we cannot discard the existence of a dedicated factor expressed at this stage preventing the competition of Exd/Hth with Abd-B. The expression of such factor in some cells but not in others would explain why Abd-B represses Hth in only some but not all cells of A8 and A9. The existence of this additional factor could also explain the surprising observation that some cells in the Abd-B domain have nuclear Hth without corresponding nuclear Exd. Our results open up the possibility of the existence of a dual Hth/Exd interaction with Abd-B: the antagonistic interaction we uncover here and, a different one, where Abd-B may not be competed by Exd/Hth and in fact could be acting as a positive cofactor as it happens with more anterior Hox genes. This may be happening in the genital discs where both Exd/Hth and Abd-B are co-expressed [39].
No cofactors have yet been identified for the Abd-B protein. The finding that the main positive cofactor of the anterior-Hox proteins is a competitor for the posterior Hox proteins is interesting. It is well established that Abd-B represses anterior-Hox gene transcription [40]. The fact that it also represses the positive cofactors of anterior-Hox proteins reinforces the prevalence of Abd-B expression and function in posterior segments. Our finding that not only Exd/Hth reinforces anterior Hox function but also counteracts Abd-B function uncovers a complementary mechanism for the stabilization of the anterior vs posterior segment information, where any accidental ectopic Abd-B expression in anterior segments would be quickly dampened down by the presence of the Exd/Hth complex before it has had a significant transcriptional effect on the repression of anterior Hox genes or on hth and exd transcription.
Another important function could be in cells where Abd-B and anterior-Hox proteins are coexpressed. Although the negative cross-regulatory interactions between Hox genes in Drosophila results in most cells expressing either an anterior or a posterior Hox protein [40], in the central nervous system or the ventral ectoderm of the embryo there are well documented cases where both proteins are coexpressed. This is illustrated by the dMP2 and MP1 pioneer neurons in the central nerve cord [41], or by the A8 segment that requires both Abd-A and Abd-B function to shape the denticle belt [24]. It is easy to imagine that in cells where both anterior-Hox proteins and Abd-B are coexpressed, the levels of Exd/Hth complex present can modulate the transcriptional output favouring either the function of one or the other Hox protein. In addition Abd-B repression of exd and hth transcription would limit the targets Abd-A could activate to those bound with high affinity in the absence of the cofactors as it has been found for Ubx in the distal part of the appendage (haltere) [42].
An open question is to what extent a similar interference also happens in mammals where the Hox proteins have expanded to 39 orthologs and multiple MEIS and PBX proteins exist [2], [13], [43]. In vertebrates there is evidence of Pbx1 binding to posterior Abd-B like Hox proteins. HoxA9-Pbx1 crystal structure showed that the conserved W amino acid present in HoxA9 at a position homologous to the YPWM sequence interacts with Pbx-1 [44]. HoxB9 and HoxA10 that posses this conserved W increase their DNA binding affinity in the presence of Pbx-1 in a similar manner as what happens with anterior-Hox proteins [33]. In contrast, Pbx1 does not increase the affinity to DNA of HoxA11, HoxD12 and HoxD13, which lack this W amino acid [33]. In fact, observation of the published results suggest that some competition to DNA binding similar to what we observe with Abd-B in Drosophila may happen in vertebrates (see Figure 1A in [33]).
Several papers have reported detailed analysis of the molecular interaction of PBX/MEIS proteins with either HoxA9 or HoxA10. Similar to our findings in Drosophila, addition of increasing amounts of MEIS leads to a decrease of HoxA9 or Pbx/Hoxa9 binding to the DNA (see Figure 2 in [45]). Although in this work the authors observed the formation of a trimeric complex on DNA that we have failed to detect, binding of the trimeric complex to the promoter was unable to increase transcription [45].
More recently, it has been reported that during osteoblastogenesis Pbx1 negatively regulates HoxA10 mediated transcription [46]. Although both results coincide with our observations in Drosophila where Exd/Hth compete instead of collaborating with Abd-B, there is one case where PBX1a and MEIS1b interact with HoxA10 as positive cofactors in the transcriptional regulation of p21 [47]. Thus, although further experiments should be done in Drosophila and vertebrates to clarify if there is a dual function of Exd/Hth and Pbx/Meis on Abd-B like Hox proteins, we believe that the existing results are indicative of a novel antagonistic function that contrasts with their well known cooperative effect with anterior-Hox proteins. Our in vivo observations indicating the existence of antagonistic interactions and recent results showing that, in vitro, Hth/Exd interaction with Abd-B transforms the unique DNA binding specificity of Abd-B from TTAT to that of a more anterior Hox gene [48] show the enormous modulatory potential that these cofactors can have on the Abd-B like Hox protein output.
The following Gal4 driver and UAS lines were used: arm-Gal4, 69B-Gal4, prd-Gal4, nullo-Gal4, UAS-hth-gfp (isoform containing both the HM and homeodomain), UAS-exd, UAS-Abd-Bm, UAS-y, UAS-t. We used the hthP2 (null mutant affecting all isoforms), hth101-1 (mutant only affecting the homeodomain containing isoform), exdYO12, and the Abd-B loss-of-function allele UabX23-1 affecting only the Abd-Br function. The ems1.2-lacZ reporter line was a gift from Bill McGinnis [25].
To test the interference of Exd/Hth with Abd-Bm function we crossed homozygous arm-Gal4 males to UAS-exd; UAS-AbdBm; UAS-hth-GFP e/TM6B females. As a control we crossed the arm-Gal4 males to w; UAS-AbdBm; UAS-y, UAS-t/TM6B females. In both crosses we expect at least 50% of the embryos to have well developed ectopic spiracles due to the expression of UAS-Abd-Bm and absence of the two accompanying UAS constructs (either UAS-exd, UAS-hth in experimental or UAS-y, UAS-t in control embryos). We observed that in the cross generating arm-Gal4; UAS-exd; UAS-Abd-Bm; UAS-hth-GFP embryos, 54.7% of them had well developed spiracles and the rest formed small and medium spiracles as those shown in Figure 4B (n = 86). In the control cross generating arm-Gal4; UAS-Abd-Bm; UAS-y, UAS-t embryos, 85,4% had well developed ectopic spiracle formation (n = 76). These results indicate that coexpression of UAS-exd UAS-hth strongly reduces the effect of UAS-Abd-Bm expression while UAS-y UAS-t does not.
Anti-Exd and anti-Hth (Kindly donated by R. Mann and N. Azpiazu); anti-AbdB 1A2E and anti-ct 2B10 (Hybridoma Bank); and anti-ßGal mouse (Promega) primary antibodies were used. For sal in situ we used an antisense RNA probe.
The UAS-Abd-Br construct was made from an Abd-Bm cDNA cutting with appropriate enzymes to delete the first exon and the resulting fragment was cloned in UASp [This construct has already been donated for the experiments performed in [41]]. We also generated a mutant Abd-B where the conserved tryptophan at position 381 was substituted by Alanine (Abd-B*W) and subcloned into pCDNA3.
Fragments of the ems1.2 enhancer (Figure S2A) were subcloned into phs43-lacZ to create the following reporter genes ems0.9, ems0.26, ems0.35, ems0.3, emsFragA, emsFragD, emsFragE and emsFragF. In the ems0.35 enhancer we mutated the putative Abd-B sites 1, 2, 4A, 4B and 6 individually or in combination to create the single ems0.35mut1, ems0.35mut2, ems0.35mut4A, ems0.35mut4B, ems0.35mut6 or double mutant ems0.35mut4A4B and ems0.35mut4A6 reporter constructs. Site 1 TCATAAA was mutated to >TCTTCAA, site 2 ATAATGA>ATCCCGA, site 4A TCATAAA>TCGGGAA, site 4B TTTATTT>TTCCCTT and site 6 TCATAAA>TCGGGAA. Constructs were injected in D. melanogaster by Bestgene (USA) and the Drosophila Consolider-Ingenio 2007 transformation platform (Spain). Four to ten independent inserts were analyzed for each line.
DNA sequence analysis to identify conservation regions and DNA binding sites was performed with the JASPAR and the GENOMATIX programs.
Complementary oligonucleotides (Table S1) were synthesized (Sigma Aldrich). Radioactively labelled probes were generated by annealing and subsequent end filling with [α-32P]dCTP. The conditions used were similar to those described previously [49], [50]. Briefly, double-stranded, end-labelled DNA (50,000 cpm/binding reaction; 10 nM) was incubated with 2 µl of reticulocyte lysate reaction mixture containing each test protein or 2 µl of the lysate control and 50 mM NaCl, 5 mM EDTA, 0,5 mM DTT, 10 mM Tris-HCl (pH 7.8), 4% glycerol, 1 mM mgCl2, and 1 µg of poly dI-dC as nonspecific competitors, in a final reaction volume of 20 µl.
Experiments designed to detect DNA-protein complex formation were performed with a 30-min incubation at 4°C. Reaction mixtures were run on a 5% polyacrylamide gel to visualize complex formation by retardation of the 32P-labeled target DNA. In some experiments monoclonal anti-AbdB was incubated with aliquots of the reaction mixture for an additional 30 min.
The amount of Hth, Exd and Hth/Exd expressing protein lysate used in the experiments detecting Abd-B DNA binding interference, was 2×, 4× and 8× the quantity of protein lysate expressing Abd-B. In all cases the final amount of protein lysate was the same, using non-expressing lysate to equalize the final volume.
Gel electrophoresis was performed in 0.5× Tris-borate-EDTA buffer as described previously [51]. For each gel shift reaction, a control containing the reticulocyte lysate was used to detect possible DNA binding by endogenous lysate factors. Gel was dry at 80°C in vacuum, exposed to a phosphorimager screen and detected by a typhoon scanner.
ChIP was performed using transiently transfected Drosophila S2 cells [52]. 10×106 cells were seeded in 10 cm cell culture dish, and transfected one day later with either 5 µg pUASt-Abd-B-HA and 5 µg pAC-GAL4 plasmids or 5 µg empty pUASt and 5 µg pAC-GAL4 plasmids. 1/10 of cells were collected to monitor the protein expression by Western blot. The remaining cells were cross-linked, lysed and sheared to 350–1000 bp as described in [53]. Six microliters of anti-HA antibody (Abcam) was used per 100 µg sheared chromatin, and the immunoprecipitation was performed according to [54].
qRT-PCR was done using primers emsQPCR2for and emsQPCR2rev (Table S1) amplifying inside the ems0.35 enhancer sequence. The data are represented as recovered percentage from the input in AbdB-HA-transfected cells against GAL4-transfected cells.
Exd, Hth and Abd-B GST pull-down assays were performed [55] after cloning the Hth or Exd ORF in pCDNA3 (Invitrogen) and labeled in vitro with S35 by the TNT T7 Quick Coupled Transcription/Translation System (Promega).
From 1000 ml of bacterial culture expressing either GST (negative control), GST-Abd-B313 (a.a. 313 to 493), GST-Abd-B313ΔHD (lacking HD a.a. 384–445) or GST-AbdB-HD (a.a. 386–446), crude extracts were generated and mixed with 70 mg of glutathione agarose beads. After 5 hr of incubation at 4°C, the beads were washed three times in lysis buffer (50 mM Tris-Cl pH 8, 1 mM EDTA, 100 mM ClNa, PMSF 250 µM, DOC 0.1%, CaCl2 5 mM, lysozime 330 µg/ml, DNasaI 66 µg/ml, Triton X-100 1% and complete protease inhibitor 1× (Roche)), then 30 µg of beads-conjugated protein mixed with 300 µl of binding buffer (10 mM Tris-Cl pH 8, 5 mM EDTA, 0.5% DTT, 1 mM MgCl2, 150 mM ClNa, 0.1 mM PMSF and complete protease inhibitor 1× (Roche)), plus 30 µl of S35-labelled protein, and incubated for an additional 4 hr at 4°C. The beads were washed four times with binding buffer. A total of 40 µl of SDS loading buffer was added to the beads, which were boiled, spun, and half of supernatant loaded onto an 8% SDS-polyacrylamide gel. After electrophoresis, the gel was dried and detected by phosphorimager method.
|
10.1371/journal.ppat.1005211 | Fundamental Roles of the Golgi-Associated Toxoplasma Aspartyl Protease, ASP5, at the Host-Parasite Interface | Toxoplasma gondii possesses sets of dense granule proteins (GRAs) that either assemble at, or cross the parasitophorous vacuole membrane (PVM) and exhibit motifs resembling the HT/PEXEL previously identified in a repertoire of exported Plasmodium proteins. Within Plasmodium spp., cleavage of the HT/PEXEL motif by the endoplasmic reticulum-resident protease Plasmepsin V precedes trafficking to and export across the PVM of proteins involved in pathogenicity and host cell remodelling. Here, we have functionally characterized the T. gondii aspartyl protease 5 (ASP5), a Golgi-resident protease that is phylogenetically related to Plasmepsin V. We show that deletion of ASP5 causes a significant loss in parasite fitness in vitro and an altered virulence in vivo. Furthermore, we reveal that ASP5 is necessary for the cleavage of GRA16, GRA19 and GRA20 at the PEXEL-like motif. In the absence of ASP5, the intravacuolar nanotubular network disappears and several GRAs fail to localize to the PVM, while GRA16 and GRA24, both known to be targeted to the host cell nucleus, are retained within the vacuolar space. Additionally, hypermigration of dendritic cells and bradyzoite cyst wall formation are impaired, critically impacting on parasite dissemination and persistence. Overall, the absence of ASP5 dramatically compromises the parasite’s ability to modulate host signalling pathways and immune responses.
| The opportunistic pathogen Toxoplasma gondii infects a large range of nucleated cells where it replicates intracellularly within a parasitophorous vacuole (PV) surrounded by a membrane (PVM). Parasites constitutively secrete dense-granule proteins (GRAs) both into and beyond the PV which participate in remodelling of the PVM, recruitment of host organelles, neutralization of the host cellular defences, and subversion of host cell functioning. In addition, the GRAs critically contribute to cyst wall formation, a process that critically ensures parasite persistence and transmission. To act as effector molecules, some of the GRAs must be translocated across the PVM. Within the related apicomplexan parasite P. falciparum, a repertoire of proteins exported beyond the PVM contain a motif cleaved by a specific protease, Plasmepsin V. Examination of the repertoire of GRAs in T. gondii revealed that some proteins exhibit such export-like motifs suggestive of protease involvement. In this study, we have functionally characterized the related aspartyl protease 5 (TgASP5) in both virulent and persistent T. gondii strains, and have investigated the phenotypic consequences of its deletion in the context of overall parasite biology, its intracellular niche, the infected host cells and the murine model. Our findings revealed fundamental roles of TgASP5 at the host-parasite interface.
| The phylum Apicomplexa groups obligate protozoan parasites that are the causative agents of severe diseases in humans and animals such as malaria, toxoplasmosis, babesiosis and coccidiosis. The key process of invasion and subsequent multiplication within their host cells is maintained via secretion from three distinct phylum-specific organelles termed the micronemes, rhoptries and dense granules [1–3].
Plasmodium falciparum is the most notorious member of the Apicomplexa in terms of its impact upon human health [4]. During its intraerythrocytic stage development, P. falciparum modulates the infected red blood cell by exporting a large repertoire of proteins to impact notably on nutrient acquisition, rosetting and cytoadherence [5]. This host cell modulation is governed by the export of various effector proteins, many of which contain a plasmodium export element RxLxE/Q/D (PEXEL), whilst a smaller repertoire of exported proteins lack this motif and are thus termed PEXEL-negative exported proteins [6]. The protease underpinning this cleavage event, Plasmepsin V, is an integral membrane protein localized to the endoplasmic reticulum (ER) in an orientation such that the catalytic aspartyl protease domain faces the ER lumen [7]. Plasmepsin V cleaves the PEXEL motif after the leucine residue, ensuring secretion of the effectors into the host erythrocyte and subsequent parasite survival [8–10].
Toxoplasma gondii is amongst the most widely distributed parasites with nearly half of the human population chronically infected. Infection during pregnancy can result in severe neurological birth defects, whilst fatal cerebral toxoplasmosis can occur in association with immunosuppressive diseases and treatments. T. gondii follows a complex life cycle involving a haploid replicative stage, followed by chronic encystation in a broad range of intermediate hosts, and meiosis in the intestine of the definitive felid host [11]. Intermediate hosts are infected either by ingestion of either oocysts shed by felids, or by bradyzoites in tissue cysts within infected meat. The fast-replicating tachyzoites are responsible for the acute stage of infection and dissemination into all tissues, whereas the slow growing bradyzoite stage establishes chronic infection with resultant cysts predominantly found in the brain and striated muscle.
Following host cell invasion, T. gondii tachyzoites and bradyzoites are surrounded by a PVM that resists endo/lysosomal fusion [12] and secludes them from the host cell cytosol. Unlike rhoptry effector proteins that are secreted at the onset of invasion, dense granule proteins (GRAs) are secreted once the parasite resides within the PV. GRAs have been implicated in a variety of processes linked to the establishment of parasitism, including the formation of the membranous nanotubular network (MNN) produced within the PV at the posterior side of invading parasites [13, 14]. More recently, several GRAs have been demonstrated to cross the PVM and subvert host cellular functions. Specifically, GRA15 modulates host cell signalling pathways by NF-kB nuclear translocation and NF-kB–mediated transcription of cytokines and other effector molecules [15]. GRA16 reaches the host cell nucleus, where it positively modulates genes involved in cell cycle progression and the p53 tumour suppressor pathway [16]. GRA24 modulates the early immune response to infection by promoting host p38 MAPK activation [17]. The activity of these proteins within the host cell is suggestive of the presence of an export pathway similar to that of Plasmodium spp. In this context, the T. gondii genome was reported to contain genes exhibiting a signal peptide in combination with proximal sequences reminiscent of the PEXEL motif (PEXEL-like motif) [18]. Some of these genes encode novel dense granule proteins; GRA19, GRA20 and GRA21, that are cleaved at the PEXEL-like motif, however do not cross the PVM and are instead incorporated into the PV and PVM. Furthermore, the previously reported GRA3, GRA5, and GRA15 were also found to contain the exact PEXEL consensus motif RxLxD/E, whilst most of the other identified GRAs possess an N-terminal motif resembling the pattern RxLxD/E within approximately 140 residues of their predicted start methionine [18].
The presence of PEXEL-like motifs on several GRAs spoke for the existence of an aspartyl protease involved in cleavage and export of these effector proteins in T. gondii. Of the seven aspartic protease paralogues encoded by the T. gondii genome (ASPs), ASP5 and ASP7 are members of the evolutionarily distinguished group (D) of apicomplexan aspartic proteases comprising Plasmepsin V [19]. Since ASP7 is not expressed in tachyzoites and bradyzoites, ASP5 emerges as the most promising candidate to cleave T. gondii PEXEL-like motif containing proteins. Here we report the functional characterization of ASP5 in T. gondii type I and type II strains. We demonstrate that ASP5 is responsible for the cleavage of GRAs containing PEXEL-like motifs and is necessary for the export of GRAs beyond the PVM into the host cells. We also report the broader phenotypic consequences of the absence of ASP5: severely compromised parasite fitness, a block in the formation of the MNN, an inability to enhance dendritic cell (DC) migration, a dramatic remodelling of the host immune response, and an impairment in cyst wall formation without any impact on tachyzoite to bradyzoite conversion.
ASP5 has previously been described as a Golgi-resident protein when expressed as an epitope-tagged second copy [19]. To determine its role and importance in T. gondii, we first confirmed the localization of ASP5 by inserting a 3Ty-epitope tag at the carboxyl-terminus of the endogenous ASP5 locus in both type I (RHΔku80) and type II (Prugniaud, PRUΔku80) strains. Endogenous ASP5-3Ty co-localizes with the Golgi marker GRASP (Fig 1A) and shows two forms by western blot analyses that have not been previously reported. The 100 kDa band is in agreement with the predicted full length protein size (108 kDa), whereas a smaller form migrates with an apparent molecular weight of 55 kDa (short-ASP5). Markedly, short-ASP5 is not detectable when a cDNA copy of the gene is expressed in the parasites (Fig 1B and 1C). To determine if the short-ASP5 form identified by western blot is the result of a processing event, a pulse-chase experiment followed by co-immunoprecipitation (IP) was performed using anti-Ty antibodies on 35S-methionine metabolically labelled ASP5-3Ty expressing parasites. Even during the short pulse, short-ASP5 is readily detectable suggesting that it originates either from an alternative transcriptional initiation, splicing or translational start, but not from a processing maturation event (S1A Fig).
To functionally characterize ASP5, knockout mutants were generated in type I and II strains. In RHΔku80, two loxP sites were inserted on either side of the ASP5 coding sequence and the gene was excised by transient expression of Cre recombinase, followed by FACS sorting and cloning. The upstream loxP was inserted along with a KillerRed expression cassette while the downstream loxP was directly fused to a GFP without a promoter as described in S2A Fig. The excised parasites (RHΔku80Δasp5) were confirmed by genomic PCR and immunofluorescence analyses (Figs 1A and S2B). In parallel, the CRISPR/Cas9 approach was used to generate frame-shift knockout parasites in type I (RHΔasp5) and type II strains (PRUΔku80Δasp5 and ME49Δasp5, with insertion of a selection marker, HXGPRT and DHFR, respectively) (S3A–S3D Fig). In the parental lines RHΔku80 and PRUΔku80, ASP5 was C-terminally epitope-tagged at the endogenous locus prior to disruption of the gene (S3A–S3D Fig). The frame-shift in ASP5 induced by CRISPR/Cas9 editing was confirmed by genomic PCR and sequencing as indicated in S3B–S3D Fig.
Plaque assays were performed to assess the importance of ASP5 for parasite fitness over multiple lytic cycles (Fig 1E). RHΔku80Δasp5, RHΔasp5, and PRUΔku80Δasp5 parasites formed strikingly smaller plaques compared to the wild-type (wt) parasites and their respective parental non-excised lines, indicating a defect in one or more steps of the lytic cycle (Fig 1F). Morphologically, all of the organelles (inner membrane complex, mitochondrion, apicoplast, rhoptries and micronemes) appeared normal in the absence of ASP5 (S1E Fig). Both type I RHΔku80Δasp5 and RHΔasp5 parasites were functionally complemented with either a genomic or cDNA version of wt ASP5, or the cDNA coding for the mutated ASP5D/A where the aspartic residue in the first catalytic site (DTG) was converted to alanine (ASP5D/A). Complemented parasites expressing either ASP5 cDNA or gDNA were readily obtained in the absence of positive selection, whereas no transgenic parasites expressing ASP5D/A were obtained unless selection was applied. Although the level of ASP5 expression between endogenously tagged ASP5-3Ty (3 tags) and the complemented strain ASP5g-Ty (1 tag) cannot be directly compared, the full complementation of phenotype by plaque assay with a low level of ASP5g-Ty suggests that the protease is produced in excess in wild type parasites. Interestingly, complementation with ASP5 gDNA resulted in a complete reversion of the Δasp5 phenotype in plaque assay, whereas ASP5 cDNA led to only partial reversion, implying that short-ASP5 contributes to ASP5 functioning (Fig 1F). In contrast, ASP5D/A failed to complement the RHΔasp5 phenotype (S1B Fig). Table 1 recapitulates all the ASP5 modified parasites lines generated in this study.
The phenotypic consequences of Δasp5 were investigated for each step of the lytic cycle. Unexpectedly, intracellular growth assays revealed that all strains examined replicated at a similar rate suggesting that ASP5 does not impact on parasite growth (Fig 1D). To exclude that the rich culture media used here is actually masking a phenotype, we performed an intracellular growth assay in glucose depleted medium. While the control BCKDH mutant parasites were severely slowed, both RH and RHΔasp5 were not impacted under glucose starvation conditions (S1C Fig). Spontaneous egress is a none-synchronous event which cannot be assessed quantitatively, whereas the calcium ionophore A23187 is a strong inducer that tends to mask modest impairments in egress. We were only able to observe a significant defect in egress in both RHΔasp5 and PruΔku80Δasp5 parasites when using BIPPO (Fig 1G), a recently described potent inhibitor of phosphodiesterases that triggers egress in apicomplexan parasites [20]. Moreover, during the process of spontaneous egress, a significant fraction of Δasp5 parasites remained enclosed within a membranous structure, either PVM or host plasma membrane, which possibly delayed infection of new host cells (Figs 1H and S1D, S1 Movie).
To assess the role of ASP5 in trafficking of the GRAs to the PV and/or the PVM, we first examined the localization of the subset of GRAs for which specific antibodies were available (GRA1, 2, 3 and 7). GRA1 is expressed and secreted into the vacuolar space as a soluble protein that subsequently becomes peripherally associated with the MNN [21]. As shown in Fig 2A, the localization of GRA1, which possesses a putative PEXEL-like motif (RALNK), is not affected by the absence of ASP5. In contrast, GRA2 and GRA3 that are associated with the MNN in wt parasites [13], showed an altered staining pattern in Δasp5 parasites (Fig 2A). Upon strong fixation conditions adapted to visualize proteins accumulated in the vacuolar space, GRA2 is not aggregated and displays instead a punctate staining for around 80% of the PVs observed. Similarly, GRA7 and also GRA3 localization at the PVM was modified in the absence of ASP5, with no PVM staining observed in more than 70% of the vacuoles (Fig 2A). Given that several GRAs involved in MNN formation appeared perturbed in RHΔasp5 parasites, the morphology of the PV was examined by electron microscopy. Whilst the MNN in wt parasites is comprised of elongated nanotubules, a dramatic change of vacuolar space architecture was observed in the absence of ASP5. In contrast to RH parasites, the PV of RHΔasp5 parasites did not exhibit a typical MNN which is usually constituted of many long and intricate tubules. Instead, RHΔasp5 parasites contained vesicles and small tubules sparsely distrubuted throughout the vacuolar space (Figs 2B and S4). This indicates that parasites lacking ASP5 are unable to assemble an elaborated MNN. The PV lumen of RHΔasp5 parasites also appears different to that of parasites depleted in both GRA2 and GRA6 [13].
By electron microscopy, the PVM appeared intact and the host mitochondria and ER still appeared to be recruited at the periphery of the vacuole (Fig 2C). The outer membrane of the host mitochondria shows a close apposition to the PV membrane of RHΔasp5 parasites, with a mean distance of 12 ± 3 nm, as similarly documented for RH parasites (Fig 2C). Morphometric analyses were undertaken to quantify the extent of host mitochondria-PV membrane association in host cells 24 h p.i.: 26 and 18% of the PV membrane was associated with host mitochondria in RH and RHΔasp5 parasite-infected cells, respectively. This suggests that the mutant has the ability to recruit host mitochondria to its PV but to a lesser extent than wt parasites.
The recently described GRA19 and GRA20 were investigated here via expression of C-terminally HA-tagged second copies as previously described [18]. Both proteins are known to be processed within their PEXEL-like motif by an unidentified protease [18]. Transiently expressed GRA19-HA and GRA20-HA were modestly processed as previously observed in parental parasites, however this cleavage was abolished in the absence of ASP5 (Fig 3A). An R/A point mutant in the GRA19 PEXEL motif prevented processing as previously reported, and served here as a control. The absence of ASP5 did not alter the localization of either GRA19 or GRA20 in an obvious manner as documented by IFA (Fig 3B). Given that both processing and localization of several GRAs is affected by the absence of ASP5, we examined whether the overall secretion by dense granules was impaired. We developed a secretion assay whereby released GRAs were collected from the supernatant of extracellular parasites and referred to here as excretory secretory antigens (ESA) upon western blot analyses. Secretion of the microneme protein MIC2 was used here as a control for parasite viability and fitness. These assays revealed that secretion of GRA1, 2, 3 and 7 were comparable in RH and RHΔasp5 strain parasites (Fig 3C). Interestingly, GRA7 which was previously reported to be phosphorylated by an unidentified host cell kinase [22, 23] gave rise to a ladder of bands which appears to be extensively reduced or even abolished in the absence of ASP5 (Fig 3C). This suggests that GRA7 may be subtly miss-targeted in the absence of ASP5 and hence no longer accessible to the host kinase. Taken together, these results indicate that ASP5 is responsible for the cleavage of some PVM-enclosed GRAs and in its absence, these proteins are normally secreted by the dense granules yet are impacted in their final destination. This is likely to lead to defects in post-translational modifications (e.g. phosphorylation of GRA7) and altered protein activity given that the MNN is no longer formed.
Following parasite internalization and the concomitant PVM formation, two GRAs are known to cross the PVM and be exported into the host cell nucleus [16, 17]. To assess the fate of GRA16 and GRA24, second copies of GRA24 driven by a tubulin promoter and GRA16 driven by its endogenous promoter and fused to 3 Myc tags were expressed in type I parasites (Fig 4A and 4C). As previously reported, GRA16 and GRA24 show a dual localization in the PV as well as the host cell nucleus in RH parasites. In sharp contrast, in RHΔasp5 parasites both GRAs accumulate in the PV but fail to reach the host nucleus, even at a high MOI. Whereas ASP5 cDNA or gDNA complementation restored the host cell nuclear localization, the catalytically inactive ASP5D/A is not sufficient to promote GRA16 export (Fig 4A). In wild type parasites GRA16-3Myc showed two forms that presumably correspond to unprocessed and processed forms given the fact that the protein possesses a PEXEL-like motif starting at the arginine residue in position 63, corresponding after cleavage to a drop of ~5 kDa. Contrastingly, in RHΔasp5 parasites the unprocessed form of GRA16 (which migrated slightly slower) strongly accumulates while a residual level of the processed form was still detectable (Fig 4B). This might result from the action of a different protease, or may represent a degradation product. The small shift in the unprocessed band observed between RH and RHΔasp5 parasites cannot be explained and will require further investigation.
Scrutiny of the GRA24 sequence did not uncover the presence of such a motif, nor did the protein appear to undergo any detectable processing event by western blot analyses (Fig 4D).
These findings demonstrate that ASP5 contributes critically to the export of parasite effector proteins both with and without a PEXEL-like motif suggesting that ASP5 might additionally alter the function of protein(s) implicated in the translocation of effectors across the PVM.
Type I parasites lacking GRA16 and GRA24 exhibit no decrease in virulence in mice, however the deletion of these genes in type II strain parasites has been reported to show reduced virulence [16, 17]. In light of this, type I and type II Δasp5 parasites and their parental lines were assessed for virulence upon intraperitoneal (i.p.) inoculation into groups of susceptible female C57Bl/6 mice. Mice had to be sacrificed 7 days after infection with 5.101 RH parasites, whereas mice receiving the same inoculum of RHΔasp5 parasites survived for 13 days (Fig 5A). Despite this, upon inoculation of a larger number of parasites (5.103) no difference was observed between the two type I parasite lines (Fig 5A). This suggests that the delay observed during low dose infection could be explained by the reduced fitness observed in tissue culture. Mice infected with 106 type II parasites from both ME49 and ME49Δasp5 led to death of the animals over the acute phase of the infection (Fig 5B). In contrast, 80% of the mice infected with 105 ME49Δasp5 survived the infection at day 40 whereas the control ME49 parasite line succumbed to infection within 7 to 13 days (Fig 5B). The type II Δasp5 in vivo phenotype therefore correlates with previous observations made with GRA16 and GRA24 deficient parasites [16, 17]. Seroconversion was assessed for the four surviving mice, which all display a positive serological profile (S5 Fig). Cyst biogenesis in vivo remains unassessed and will be further investigated.
One of the key events in establishing a protective Th1 immune response against T. gondii is the ability of host immune cells to produce the pro-inflammatory cytokine interleukin 12 (IL-12), which in turn stimulates the production of interferon gamma (IFNγ) by natural killer (NK) cells, CD4+ and CD8+ T cells [15, 24]. IFNγ is the major pro-inflammatory cytokine driving multiple cellular defense mechanisms during both the acute and chronic phases of infection [25]. Importantly, the immunity related GTPases (IRG proteins) constitute a large family of interferon-inducible proteins that mediate early resistance to T. gondii infection in mice. Several studies have shown that IRGs, in particular Irga6 and Irgb6 are recruited to the nascent PVM, where they cause disruption of the vacuole and parasite death. While the ROP18 complex of type I parasites is able to phosphorylate IRG proteins, thereby preventing their oligomerization and loading onto the PVM, type II parasites are unable to block the action of IRG proteins due to the polymorphic nature of ROP5, which forms part of the ROP18 complex [26]. Recent studies have associated GRA7 to the ROP18 complex by acting as regulator for ROP18-specific inactivation of Irga6 [23, 27]. To determine whether vacuoles containing RHΔasp5 parasites failed to block IRG recruitment to the PVM, an IRG recruitment assay was performed for Irgb6. Our results indicate that RHΔasp5 parasites behave like RH parasites and remain non-susceptible to Irgb6 and Irga6 loading (Fig 6A, left panel). Conversely, Irgb6 was recruited to the PVM of PRUΔku80Δasp5 parasites as previously reported for PRUΔku80 (Fig 6A, right panel) [26]. Given these data and in spite of the impact of ASP5 on GRA7 phosphorylation (which forms part of the ROP18 complex), we propose that ASP5 activity is not essential for the activity of the ROP18 complex [23].
In light of the blockage in the export of GRA effectors, we interrogated the impact of Δasp5 on the macrophage response to infection. Differentiated bone marrow derived macrophages (BMDMs) were infected with type I and II strain parasites lacking ASP5, and IL-12p40 levels were measured 40 h pi (at the peak level of IL-12p40 secretion) by ELISA. As previously observed, PRU parasites induce significantly higher levels of IL-12p40 synthesis when compared to RH strain parasites (Fig 6B). Both type I and type II strain parasites showed a reproducible and significantly lower level of IL-12p40 produced by macrophages infected with Δasp5 parasites. Expectedly, RHΔasp5/asp5g complemented parasites rescued the IL-12p40 phenotype observed with RHΔasp5 parasites. This assay has also been carried out with RAW264.7 cells and with peritoneal exudate cells (PECs), and both assays produced results comparable to those obtained with BMDMs. Based on the current knowledge, both type II GRA15 and GRA24 promote IL-12 secretion in vitro and in vivo and it is therefore plausible that export and function of GRA15 is also blocked in the absence of ASP5 [15].
Chemokines are soluble mediators that are essential to contain parasite spreading and to control the infection. Previous studies have shown that T. gondii induces chemokine up-regulation in several cell types and specifically GRA6 [28], GRA24 [17] and GRA25 [29] are known to shape the immune response by regulating the expression of CXCL1, CXCL2, CCL2, CCL5 and CXCL10 [30]. Here, monolayers of pMEF cells were infected with type I and type II strain parasites and pelleted after 20 h. cDNA was synthesized from total RNA and the mRNA expression levels for each of these chemokines was measured by qPCR. These data reproducibly demonstrated a pronounced decrease in CXCL1, CXCL2, CCL2, CCL5 and CXCL10 expression reproducibly measured from cells infected with RHΔasp5 parasites when compared with RH and RHΔasp5/asp5g parasites (Fig 6C). The results obtained with PRU and PRUΔasp5 showed no difference for CXCL1, CXCL2 and CCL2, while a significant decrease in CCL5 and CXCL10 expression was observed.
To obtain a more global picture of the modification caused by the absence of ASP5, we performed a genome-wide expression profiling by RNA-sequencing of mouse BMDMs infected with parental or Δasp5 parasites from both type I and type II parasites (Fig 6D). We focused our analysis on genes that were modulated with more than twofold change when comparing each Δasp5 mutant with their respective parental strains. A broader, analysis of the highly modulated KEGG pathways in a type I context of infection revealed that a substantial number of both pro- (e.g., IL-1a and TNF) and anti-inflammatory (e.g., IL-10) cytokines, chemokines and their relative receptors (e.g., CXCL1, CXCL10) were significantly differentially-regulated upon Δasp5 parasite infection compared to parental lines (Fig 6D, left panel). Strikingly, in type II Δasp5 parasites, the parasite-induced transcriptional response is strongly reduced to a level resembling un-infected cells (Figs 6D, right panel, S6A and S6B). This result highlights that ASP5 might broadly impact type II-specific virulence factors and host cell effectors.
T. gondii tachyzoites can cross biological barriers [31, 32], however, once inside the host the precise mechanisms leading to systemic dissemination of the parasites remain unknown. T. gondii can exploit the migratory properties of dendritic cells (DC) to spread throughout the organism using the “Trojan horse” strategy. Specifically, upon infection by tachyzoites, DCs exhibit a hypermigratory phenotype [33, 34]. GRA5 has been described as one of the parasite effector molecules capable of increasing the migratory properties of DCs via CCR7 expression without DC activation [35]. To determine the potential impact of ASP5 in this process we first assessed the hypermotility phenotype and observed no significant differences between Δasp5 and the corresponding parental lines either in RH or PRU parasites (Fig 7A). We then conducted a transmigration assay, whereby we measured the ability of infected DCs to migrate in response to CCL19, a CCR7 ligand [34]. In this assay both RHΔasp5 and PRUΔku80Δasp5 parasites showed a considerable reduction in transmigration of infected DCs (Fig 7B).
Importantly, GRAs are anticipated to play key roles in other stages of the parasite and notably during cyst wall formation, a process that is central for parasite persistence and transmission [36]. To assess the role of ASP5 in cyst wall formation, we used the fluorescent Dolichos biflorus Agglutinin (DBA) lectin to detect the glycosylated protein CST1 (one of the few available markers of the T. gondii cyst wall) [37], following pH induced differentiation of PRUΔku80Δasp5 strain parasites in vitro. Stage conversion from tachyzoites to bradyzoites was measured by expression of SAG4 or BAG1. This conversion took place normally in PRUΔku80Δasp5 parasites and CST1 was also produced, glycosylated and targeted to the PV However, upon closer inspection of the IFA, it became obvious that cyst wall formation was already impaired just one week after induction of differentiation, and even more strikingly impaired two weeks later (Fig 8A and 8B).
Whilst intracellular, apicomplexan parasites reside within a specialised membranous niche (PVM), across which the parasite transports a plethora of effector molecules necessary for subversion and remodelling of host cell functions. Studies in P. falciparum have revealed that this process relies upon a PVM-resident translocation machinery (PTEX) that serves to facilitate export of parasite proteins across this membrane into the erythrocyte cytosol [38, 39]. Intimately associated with this process is Plasmepsin V, a protease known to cleave a specialised motif (PEXEL) in a wide repertoire of known exported proteins [40–42]. Similarly, T. gondii possesses components related to the PTEX translocon that are proposed to act as a molecular sieve at the PVM, allowing diffusion of small molecules across this membrane [43].
Here we report the characterisation of the Golgi-resident protease ASP5, which is responsible for the cleavage of PEXEL-like motif-containing proteins in T. gondii. Whilst Plasmepsin V appears to be primarily dedicated to cleavage of proteins destined to be exported beyond the PVM, deletion of ASP5 causes considerable pleiotropic effects by effecting both exported and PV/PVM-resident proteins (Fig 9). Accordingly and in contrast to P. falciparum, numerous PEXEL-like containing proteins remain within the PV and are not further exported. Deletion of ASP5, without affecting dense granule secretion, caused significant morphological aberrations of the PV, most notably being the defect in MNN formation. The role of this elaborated structure is still mysterious, although it is presumed to participate in parasite access to host cell nutrients. In this context, and rather unexpectedly, depletion of ASP5 does not appear to impose any restriction on intracellular parasite replication even in glucose depleted media. The molecular connection between ASP5 activity and MNN formation is not known, however such a phenotype was previously described when individual GRAs were knocked out [13]. Alternatively, the MNN could participate in the process of egress, which is unexpectedly affected in parasites lacking ASP5. It is known that some GRAs form high molecular weight complexes within the dense granules [44] and also exist as heteromeric complexes in the PV [44]. It is therefore conceivable that deletion of ASP5 could affect formation of these complexes and thus ASP5 will not only affect the activity of its direct substrates but also their interacting partners.
Through western blotting analyses we demonstrate that the vacuolar GRA19 and GRA20, together with the exported GRA16, are cleaved by ASP5 at the PEXEL-like motif. These proteins share a consensus motif “RRL” in their PEXEL-like sequence, which is likely part of a larger motif recognized and cleaved by ASP5. This slightly differs from the substrate preference of Plasmepsin V [45], and may suggest that the motif serves differing functions in T. gondii.
In wt parasites, multiple GRAs such as GRA16 and GRA24 are exported to the host cell nucleus. Upon ASP5 deletion however, there is a striking block in the export of GRA16 and GRA24 and they are absent in the host nucleus but still accumulated within the PV. Intriguingly, while GRA16 harbours a PEXEL-like motif and thus accumulates as an un-cleaved product in the absence of ASP5, GRA24 is apparently devoid of such a signal. Assuming that both GRA16 and GRA24 share the same translocation pathway to cross the PVM, the defect in GRA24 export suggests that a component of the export machinery is defective and hence possibly active only upon cleavage by ASP5. A recent study has demonstrated that ASP5 is directly responsible for the processing of GRA16 in a PEXEL-dependent manner at the site (RRLAE) [46]. In accordance with our findings, this study established that the mapped processing of GRA16 is actually not necessary for export of the protein to the host nucleus. Further dissection of GRA16 processing and identification of components of the translocation machinery and the potential implication of ASP5 in this process still await further investigation.
The aforementioned GRAs exported to the host cell nucleus play a significant role in reprogramming host cell gene expression and thus contribute to immunomodulation of the host. This response is significantly perturbed in parasites depleted of ASP5 as reflected by the targeted analysis of IL-12 and the chemokine levels, as well as by the global transcriptome analysis of BMDMs infected with either type I or type II parasites.
Previous studies in type II strain parasites reported specific host-transcriptome alterations when selected GRAs [15–17, 29] were knocked-out individually. While reproducing most of these findings our data significantly broaden the network of putative parasite-dependent host-regulated pathways. Importantly, the broader impact observed upon ASP5 deletion reflects the multiple GRAs affected simultaneously in this mutant. T. gondii also acts to subvert host cell functions through increasing the motility and migration of infected DCs. This parasite-induced hypermigration of infected DCs ensures rapid dissemination of the parasites from the intestinal site of entry to the rest of the body. This step is critical for the establishment of infection and persistence as it gives the parasite time to reach sanctuary organs prior to the onset of the immune response [47]. The mediators of DCs hypermigration are not fully characterized; however GRA5 has been reported to be associated with this phenomenon [35]. Parasites lacking ASP5 show a considerable loss of this capacity, which correlates with the down-regulation of CCR7 observed in the RNA seq data upon Δasp5 type I and type II parasite infection.
The establishment of chronic infection is underpinned by cyst formation. Not only do cysts ensure transmission from intermediate to definitive hosts, they are also the source of reactivation in situations of immunosuppression and thus a very important stage from a pathology view-point. Our data indicate that whilst ASP5 does not impact upon the capacity of the parasite to differentiate into bradyzoites, depletion of ASP5 severely compromises the parasites ability to build the cyst wall. The cyst wall provides a protective ‘shell’ within which the parasite is sheltered from the host cell environment, however exchanges across this barrier with the host cell are still anticipated to occur.
In conclusion, the role of the T. gondii PEXEL-like motif appears to be significantly broader than that reported to date for Plasmodium spp., wherein PEXEL cleavage gives rise to a newly exposed N-terminal sequence proposed to serve as a trafficking signal for proteins destined to be exported beyond the PVM. Given the substantial differences in host cell repertoire and accordingly, the host cell subversion requirements of these two apicomplexan parasites, it is not surprising that cleavage of this motif in T. gondii is likely implicated in a variety of functions. For example, the ASP5 PEXEL-like motif cleavage could be needed to elicit conformational changes needed for enzymatic activity, or alternatively, for sequential interaction/recognition by/with a host/parasite partner. In the context of protein targeting, given that T. gondii PEXEL-like motif cleavage occurs within the Golgi prior to trafficking to the DGs, it might specifically target a population of proteins to distinct secretory organelles. Concordantly, not all exported GRAs completely co-localize with canonical GRA-markers within the parasite [16, 17]. Whilst the data presented here have substantially contributed to the knowledge base surrounding not only ASP5 but also GRA functioning, many questions pertaining to the reasons behind this cleavage remain to be answered. Furthermore, we have not estimated the range of ASP5 substrates. Notably, the option that ASP5 mediated cleavage could expand to proteins from secretory organelles other than dense granules or following the default pathways for secretion, has not been comprehensively investigated. In spite of this, this study will serve as a solid platform upon which further investigations into the essential process of apicomplexan protein export can be completed.
E.coli XL-10 Gold chemo-competent bacteria were used for all recombinant DNA experiments. T. gondii tachyzoites parental and derivative strains were grown in confluent human foreskin fibroblasts (HFFs) maintained in Dulbecco’s Modified Eagle’s Medium (DMEM, Gibco) supplemented with 5% fetal calf serum (FCS), 2mM glutamine and 25 mg/ml gentamicin.
Genomic DNA (gDNA) from RH parasite was isolated with the Wizard SV genomic DNA purification system (Promega). Total cDNA was generated by RT-PCR using the Superscript II reverse transcriptase (Invitrogen). TgASP5 ToxoDB accession number: TGME49_242720. The C-terminal (Ct) of ASP5 was amplified with primers 4203–4204 on gDNA and cloned in pT8-TgMIC13-3Ty-HXGPRT [48] between ApaI and NsiI sites to give Ct-ASP5-3Ty-HXGPRT. This vector was then digested ApaI/PacI and cloned in pTub8-loxP-KillerRed-loxP-YFP-HXGPRT [49] to give Ct-ASP-LoxP-YFP-HXGPRT. The 5’ region of ASP5 was amplified with primers 4615–4616 and cloned in pTub8-loxP-KillerRed-loxP-YFP. This vector was digested PacI/SacII and the bleomycin selection cassette from pTub8-ARO-GFP-Ty-Ble [50], digested with the same sites and was inserted to give 5’ASP5-pTub8-loxP-KillerRed-Ble. Ct-ASP5-3Ty-HXGPRT was digested ApaI/NotI and ligated into p2854-DHFR-TS [51] to create Ct-ASP5-3Ty-DHFR. To create pTub8-ASP5c-Ty, cDNA was amplified with primers 1624–1592, digested MfeI/NsiI and cloned in pTub8-Ty [52] digested EcoRI/NsiI. To create pTub8-ASP5g-Ty, cosmid PSBL804 (D. Sibley, Toxodb.org) was digested BglII/EcoRI, the band corresponding to the genomic DNA of ASP5 was isolated and cloned at the same sites in pTub8-ASP5c-Ty. Importantly, the last intron is not present and the two last exons are fused. ASP5 catalytic residue D431 was mutated to A to create pTub8-ASP5c-D/A-Ty. The Q5 site directed mutagenesis kit (NEB) instructions was followed using primers 4795–4796 with the pTub8-ASP5c-Ty as template. TgGRA24 was amplified from gDNA using primers 4814–4815, and cloned in pTub8TgARO-Myc-Ble [50] between EcoRI and NsiI sites. This vector was digested with NotI and ligated with a PCR product using primer 4943–2642 with pTub5-CAT-Sag1 [52] as template to create the plasmid pTub8-GRA24-Myc-Ble-CAT. To construct the vector pLIC-PGRA16-GRA16-3-Myc, the promoter region and the coding sequence were amplified on gDNA using primers 5419–5421 and cloned into the pLIC-3Myc-dhfr vector using the LIC cloning method as described [53]. gRNA for CRISPR/Cas9 was generated with primers 4883–4969 on the pSAG1-CAS9gfp-U6gRNA [54] following the Q5 site directed mutagenesis kit (NEB) instructions. The HXGPRT and DHFR cassettes used to generate the CRISPR/Cas9 mediated KOs of ASP5 were amplified by KOD DNA polymerase (Novagen) using primers 5240–5241 and 5142–5143 respectively. PCR products were precipitated in sodium acetate and re-suspended in water prior to transfection.
T. gondii tachyzoites were transfected by electroporation as previously described [55]. Selection of transgenic parasites were performed with either mycophenolic acid and xanthine for HXGPRT selection [56], pyrimethamine for DHFR selection [51] or phleomycin for ble selection [21]. All stable expressing strains were cloned by limited dilution in 96-well plates and analyzed for the expression of the transgenes by IFA and for the genomic integration by PCR. The RHΔku80-DiCre (abbreviated RHΔku80) strain [49] was transfected with 40 μg of the plasmid Ct-ASP-LoxP-YFP-HXGPRT linearized AvrII. The resulting strain, RHΔku80asp5-3Ty, was transfected with 5’ASP5-pTub8-loxP-KillerRed-Ble linearized XhoI to create the RHΔku80loxPasp5-3Ty strain. 40 μg of pTub5-Cre [57] was transfected in this strain to obtain RHΔku80Δasp5 (S1 Fig). To generate RHΔasp5, 30 μg of pSAG1-CAS9gfp-U6gASP5 was transfected into RH parasites (S2 Fig). Transfected parasites where cloned by GFP+ FACS sorting 48 hr post-transfection. To generate PRUΔku80Δasp5 and ME49Δasp5, 30 μg of pSAG1-CAS9gfp-U6gASP5 and respectively 15 μg of KOD-amplified HXGPRT or DHFR selection cassette flanked by 25 nt homology regions were transfected into PRUΔku80asp5-3Ty or ME49Δasp5 (S2 Fig). Clones were obtained by limiting dilution after appropriate selection. 60 μg of pTub8-ASP5c-Ty, pTub8-ASP5g-Ty, were transfected in RHΔku80Δasp5 and in RHΔasp5. Parasites were passaged for several weeks to allow the Ty positive population to gradually increase. Parasites were cloned and Ty positive clones were selected. 60 μg of pTub8-ASP5c-D/A-Ty was transfected in RHΔasp5 and HXGPRT-selected. In absence of selection, no Ty positive population was observed even after 6 weeks post-transfection. Transient transfection of GRASP-GFP [58], pTub-GRA19-HA, pTub-GRA19-HA R124A, pTub-GRA20-HA [18] and pLIC-PGRA16-GRA16-3Myc was performed by using 40 μg of each plasmid as previously described [55].
A confluent monolayer of HFFs was infected with around 50 freshly egressed parasites for 7 to 8 days before the cells were fixed with PFA/GA. Plaques were visualized by staining with Crystal Violet (0.1%) as previously described [59].
RH, RHΔasp5, PRUΔku80 and PRUΔku80Δasp5 were allowed to grow on HFF for 24 hr prior to fixation with PFA/GA. IFA using α-GAP45 antibodies was performed and the number of parasites per vacuole was scored. For each condition, 200 vacuoles were counted. Data are mean value ± s.d. of three independent experiments. Medium depleted in glucose is DMEM 11966 supplemented with up to 6 mM glutamine and 25 μg/ml gentamicin [60].
Freshly egressed tachyzoites were added to a new monolayer of HFF, washed after 30 min and grown for 30 hr. The infected HFF were then incubated for 5 min at 37°C with DMEM containing either 3 μM of the Ca2+ ionophore A23187 (from Streptomyces chartreusensis, Calbiochem), 50 μM of BIPPO [20] or DMSO as control. Host cells were fixed with PFA/GA, and IFA using α-GAP45 antibodies was performed. 200 vacuoles were counted per strain and per condition, and the number of lysed vacuoles was scored. Data are mean value ± s.d. of three independent experiments.
Dense granule secretion assay was performed with freshly egressed parasite washed twice in intracellular buffer (IC; 5 mM NaCl, 142 mM KCl, 2 mM EGTA, 1 mM MgCl2, 5.6 mM glucose, 25 mM HEPES-KOH, pH 7.2) containing protease inhibitor cocktail. Parasite constitutive secretion was performed in similar IC buffer. After 1 hr at 37°C, a fraction was collected, which represent the total lysate, and the excreted secreted antigens (ESA) were collected upon sequential centrifugation. Following an initial centrifugation 5 min/4°C/1000 g, supernatants were transferred to a new tube and spun again 5 min/4°C/2000 g. The final supernatant was collected and analyzed by immunoblotting. Processing of the micronemal protein 2 (MIC2) upon secretion of the micronemes was used as control.
Antibodies described here were used for IFA and western blot analysis. The mAbs α-Ty tag BB2, α-Myc tag 9E10, α-HA (Covance Inc), α-ROP2-4 T3-4A7 [61], α-MIC2, α-GRA1, α-GRA2, α-GRA3 (J. F. Dubremetz), α-ACT1 [62] as well as the polyclonal Abs α-GAP45 [59], α-CAT [63], α-GRA7 (kindly provided by Prof. D.J.P. Ferguson) α-Hsp70 [64], α-Cpn60 [65] were used. Infected-HFF monolayers on coverslips were fixed with 4% paraformaldehyde (PFA)/0.05% glutaraldehyde (GA) for 10 min or for 30 min for GRAs PV and PVM localization prior to quenching in 0.1M glycine/PBS. Cells were then permeabilized with 0.2% Triton X-100/PBS (PBS/Triton) and blocked in the same buffer supplemented with 2% BSA (PBS/Triton/BSA). Cells were incubated with primary antibodies (Abs) diluted in PBS/Triton X-100/BSA for 1 hr followed by PBS/Triton washes (3 x 5 min). Cells were incubated with secondary Abs (Alexa488- or Alexa594-conjugated goat anti-mouse or goat anti-rabbit IgGs) in PBS/Triton X-100/BSA. Where appropriate, parasite and HFF nuclei staining was performed by incubating cells in DAPI (4’,6-diamidino-2-phenylindole; 50μg/ml in PBS) prior to final washing (3 x 5 min). Coverslips were mounted in Fluoromount G (Southern Biotech) on glass slides and stored at 4°C in the dark. Confocal images were collected with a Zeiss microscope (LSM700, objective apochromat 63x /1.4 oil) at the Bioimaging core facility of the Faculty of Medicine, University of Geneva. Stacks of sections were processed with ImageJ and projected using the maximum projection tool.
Crude extracts of T. gondii tachyzoites were subjected to SDS-PAGE Western blot analysis carried out using polyacrylamide gels under reducing conditions. Proteins were transferred to hybond ECL nitrocellulose. Primary and secondary antibodies (HRP conjugated, SIGMA) are diluted in PBS, 0.05% Tween20, 5% skimmed milk. Bound antibodies were visualized using the ECL system (Amersham).
Metabolic labelling of the tachyzoites was done with 50 mCi [35S]-labeled methionine/cysteine (Hartmann analytic GmbH) per ml for 4 h at 37°C followed by co-IP in RIPA buffer using α-Ty antibodies. A pulse of 7 min was followed by two chases of 15 and 60 min.
Freshly egressed parasites were allowed to invade HFF monolayers for 24 hr prior to fixation. Infected host cells were washed with 0.1M phosphate buffer pH 7.2 and were fixed with 2.5% glutaraldehyde in 0.1M phosphate buffer pH 7.2, post-fixed in osmium tetroxide, dehydrated in ethanol and treated with propylene oxide to embedding in Spurr’s epoxy resin. Thin sections were stained with uranyl actetate and lead citrate prior to examination using a Technai 20 electron microscope (FEI Company). Two independent sample preparation and multiple thin sections for each sample were examined.
HFF infected with wt or mutant parasites for 24 hr were fixed in 2.5% glutaraldehyde in 0.1 M sodium cacodylate buffer (pH 7.4) for 1 hs at room temperature, and processed as described [66] before examination with a Philips CM120 Electron Microscope (Eindhoven, the Netherlands) under 80 kV. Morphometric analysis to quantify the extent of association of host mitochondria with PV was performed as described [67].
Total RNA was extracted from infected primary mouse embryonic fibroblasts (pMEFs) using TRIzol reagent. cDNA was synthesized using Verso Reverse transcription (Thermo Fisher Scientific). Real-time PCR was performed using the Go-Taq real-time PCR system (Promega) and the CFX connect real-time PCR system (Biorad). The values were normalized to the amount of actin in each sample. The primer sets used are listed in [28]. Chemokine production was analyzed from three independent experiments.
Bone marrow derived macrophages (BMDMs) were obtained by flushing marrow from the hind tibias and femurs of C57Bl/6 mice. The cell suspension was passed through a nylon mesh and cultured in RPMI1640 medium supplemented with 10% FCS, 100 U/ml penicillin, 0.1 mg/ml streptomycin, and 15% L-cell conditioned medium at 37°C degrees in humidified 5% CO2. Non-adherent cells were passed the next day to 10-cm bacteriological petri dishes (4.106 cells/dish) and harvested for experiments six days later using a cell scraper.
BMDMs were seeded (5.105 per well) in 24 well plates containing coverslips and activated with 10 ng/ml of murine recombinant IFNγ for 24 h at 37°C, 5% CO2. The next day, activated cells were infected with freshly egressed and filtered T. gondii parasites (MOI = 1) for 1 h, then fixed in 4% PFA in PBS for 10 min, semi-permeabilized in 0.002% digitonin in PBS for 7 min at 4°C and blocked in 2% BSA/PBS for 30 min. IFA was carried out using the primary antibodies (Abs) mouse α-Irga6 (kindly provided by Prof. J.C. Howard), goat α-Irgb6 (Santa Cruz), mouse α-GRA1 and mouse α-GRA2. GRA1 and GRA2-containing positive vacuoles were analyzed for the presence of Irgb6 at the vacuole by counting 10 fields at a magnification of 100X. Data was analyzed and images were taken using a confocal laser microscope (FVI1200 IX-83; Olympus) and the software FLUOVIEW (Olympus) and are representative of three independent experiments.
BMDMs were seeded (105 per well) in 96-well plates and left to adhere over-night at 37°C, 5% CO2. Cells were then infected with freshly egressed and filtered T. gondii parasites (3.105 parasites per well) and culture supernatants were collected 40 h later and frozen at -20°C degrees. IL-12p40 levels were measured by ELISA according to manufacturer’s instructions in three independent experiments. The assay was also carried out using PECs and RAW264.7 cells.
BMDMs were plated to 80% confluence in RPMI with 10% FCS, 100 U/ml penicillin, 0.1 mg/ml streptomycin and infected with the different T. gondii strains at a MOI of 3. 18–20 hr post-infection, cells were rinsed with cold PBS, detached with trypsin and pelleted. Total RNA from samples in triplicates was extracted using a hybrid RNA extraction protocol with TRIzol (Life Technologies) and QIAGEN RNeasy Mini Kit. Sample pellets were lysed in TRIzol followed by the addition of chloroform to separate the aqueous layer and the organic layer. RNA from the upper aqueous phase was precipitated with 70% ethanol and isolated using the RNeasy column according to the manufacturer's instructions. Isolated RNA was subjected to single read 100 bp on a Illumina HiSeq 2500 at the Genomics platform at the University of Geneva, iGE3 (the Institute of Genetics and Genomics). The RNA samples were multiplexed across 3 sequencing lanes of the flow cell.
RNA Sequencing was performed on the Illumina HiSeq 2500 at the iGE3 genomics plateform of the University of Geneva (http://www.ige3.unige.ch/genomics-platform.php). The adapter sequences from the raw reads obtained from the RNA-Seq were trimmed using FASTX-Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/) (phred<20). The resulting reads, after quality control, were aligned to the latest mouse reference genome (GRCm38) using TopHat/Bowtie2 aligner and HTSeq-count was used to get the read counts of the genes [68–71]. Differential expression analysis was carried out using edgeR, a Bioconductor package in R (http://www.R-project.org). Normalized expression values from the count data were obtained from the normalization factors calculated by the TMM (trimmed mean of the M values) method. The heatmaps for the genes of interest were also generated in R using heatmap.2 in the gplots package. KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway analysis was performed using KOBAS 2.0 (KEGG Orthology Based Annotation System) [72, 73]. All the computations were performed at University of Geneva on the Baobab cluster.
BM-DCs were obtained as previously described [74]. Briefly, DCs were generated by culturing BM cells for 7–10 days in the presence of 20 ng/ml GMCSF in RPMI supplemented with 10% heat-incativated FCS, 50 mM 2-mercaptoethanol, 100 mM sodium pyruvate and 100 μM penicillin/streptomycin at 37°C in humidified 5% CO2.
DC (3–5.104) were challenged with freshly egressed tachyzoites at a MOI of 3, treated with LPS (100 ng/ml) or maintained in complete medium (non-infected). Cells were settled on gelatin-coated glass slides for 6–8 hr at 37°C. The cells were imaged every min for 45–60 min (Zeiss Cell Observer.Z1). Motility patterns were compiled using ImageJ (image stabilizer software and manual tracking plugins).
DCs were plated at a density of 1.106 cells/well and incubated with freshly egressed T. gondii tachyzoïtes (MOI 3) for 4 hr at 37°C and 5% CO2. DCs were then transferred into transwell filters (8 μm pore size; Corning) and incubated for 16 hr at 37°C and 5% CO2. Migrated DC were quantified in a hematocytometer.
In vitro tachyzoite to bradyzoite conversion was induced by exposing parasite cultures to pH 8.2 as described previously [75]. Briefly, 5.104 tachyzoites were allowed to infect HFF grown on glass coverslips inside 24-well plates. 24 hr post infection, bradyzoite differentiation was induced by replacing normal media with RPMI 1640 buffered with 50 mM HEPES to pH 8.2 and supplemented with 3% fetal bovine serum. Parasites were allowed to grow at 37°C in absence of CO2 for 4 days and alkaline media was changed daily. After 4 days of conversion, infected HFF were fixed with 3.7% formaldehyde, permeabilized with 0.5% Triton X-100 in phosphate buffered saline (PBS) for 20 min. After 1 hr incubation with 10% foetal calf serum (FCS) as blocking agent, the cells were stained for 1 hr with DBA conjugated with Alexa 594 (used at 10 μg/mL; Vector) and with α-BAG1 mAb (kindly provided by Prof. V.B. Carruthers) followed by Alexa Fluor 594 goat anti-rabbit IgG antibody 200 vacuoles were counted from 20 fields for each experiment to determine the positive/negative rate of DBA and BAG1 staining.
Mice were infected by intraperitoneal injection. The health of the mice was monitored daily until they presented severe symptoms of acute toxoplasmosis (bristled hair and complete prostration with incapacity to drink or eat) and were sacrificed on that day.
All animal experiments were conducted with the authorization Number (1026/3604/2, GE30/l3) according to the guidelines and regulations issues by the Swiss Federal Veterinary Office. No human samples were used in these experiments. Human foreskin fibroblasts (HFF) were obtained from ATCC.
|
10.1371/journal.pcbi.1000401 | SpaK/SpaR Two-component System Characterized by a Structure-driven
Domain-fusion Method and in Vitro Phosphorylation Studies | Here we introduce a quantitative structure-driven computational domain-fusion
method, which we used to predict the structures of proteins believed to be
involved in regulation of the subtilin pathway in Bacillus
subtilis, and used to predict a protein-protein complex formed by
interaction between the proteins. Homology modeling of SpaK and SpaR yielded
preliminary structural models based on a best template for SpaK comprising a
dimer of a histidine kinase, and for SpaR a response regulator protein. Our LGA
code was used to identify multi-domain proteins with structure homology to both
modeled structures, yielding a set of domain-fusion templates then used to model
a hypothetical SpaK/SpaR complex. The models were used to identify putative
functional residues and residues at the protein-protein interface, and
bioinformatics was used to compare functionally and structurally relevant
residues in corresponding positions among proteins with structural homology to
the templates. Models of the complex were evaluated in light of known properties
of the functional residues within two-component systems involving His-Asp
phosphorelays. Based on this analysis, a phosphotransferase complexed with a
beryllofluoride was selected as the optimal template for modeling a SpaK/SpaR
complex conformation. In vitro phosphorylation studies
performed using wild type and site-directed SpaK mutant proteins validated the
predictions derived from application of the structure-driven domain-fusion
method: SpaK was phosphorylated in the presence of 32P-ATP and the
phosphate moiety was subsequently transferred to SpaR, supporting the hypothesis
that SpaK and SpaR function as sensor and response regulator, respectively, in a
two-component signal transduction system, and furthermore suggesting that the
structure-driven domain-fusion approach correctly predicted a physical
interaction between SpaK and SpaR. Our domain-fusion algorithm leverages
quantitative structure information and provides a tool for generation of
hypotheses regarding protein function, which can then be tested using empirical
methods.
| Because proteins so frequently function in coordination with other proteins,
identification and characterization of the interactions among proteins are
essential for understanding how proteins work. Computational methods for
identification of protein-protein interactions have been limited by the degree
to which proteins are similar in sequence. However, methods that leverage
structure information can overcome this limitation of sequence-based methods;
the three-dimensional information provided by structure enables identification
of related proteins even when their sequences are dissimilar. In this work we
present a quantitative method for identification of protein interacting
partners, and we demonstrate its use in modeling the structure of a hypothetical
complex between two proteins that function in a bacterial signaling system. This
quantitative approach comprises a tool for generation of hypotheses regarding
protein function, which can then be tested using empirical methods, and provides
a basis for high-throughput prediction of protein-protein interactions, which
could be applied on a whole-genome scale.
| Because proteins so frequently function in coordination with other proteins,
identification and characterization of protein-protein complexes are essential
aspects of protein sequence annotation and function determination [1]. A
variety of empirical [2]–[4] and computational [5]–[14] methods for
identifying putative protein-protein interactions have been reported. Of particular
note is the Rosetta Stone approach for identifying interacting partners based on the
theory of gene fusion, whereby protein domains that are encoded separately in one
species may be homologous to domains that are “fused” in the
same open reading frame in another species [15]–[17]. Whereas
sequence-based domain fusion methods can be highly successful in identifying
putative functional relationships among proteins, the reliance on sequence homology
limits detection to protein sequences with adequate levels of sequence identity.
Another approach to identifying putative protein-protein interactions is described
by Lu and coworkers [18], whereby sequence-based searches against the PDB
database were performed in order to identify multi-domain structures having at least
one domain with good sequence identity to each putative interacting protein.
However, the sensitivity of this search method is also dependent on the levels of
sequence identity between the proteins of interest and the sequences of the domains
within the identified PDB domain-fusion template. Kundrotas and Alexov [6]
explored the use of structure-based comparisons in the identification of
multi-domain templates for homology modeling of complex structures. In this work, it
was determined that a structure-based protocol performed considerably better than
did a sequence-based protocol in recovering known protein-protein interacting
partners (86% recovery as opposed to 19%) in searches against
a database of known complexes, indicating that the structure-based method was more
sensitive in detecting remote homologs.
We describe the application of a quantitative structure-based comparison method to
the identification of putative protein-protein interactions, and show that this
approach increases sensitivity in detecting putative interactions at low
(<20%) levels of sequence identity, based on the general principle
that structure homology is more highly conserved in evolution than is sequence
homology [19]. Our approach, therefore, involves the generation of
a structure model, based on adequate (typically >30%) sequence
identity to a PDB domain, followed by structure-based homology searches against PDB
to identify multi-domain structures with adequate structure identity [20] to the
model of each putative interacting protein. Thus, we propose that our
structure-driven domain-fusion method can be used to identify domain-fusion
templates for modeling protein-protein interaction complexes, and that such searches
may prove to be more sensitive than sequence-based searches alone.
To explore this approach, we selected as the subject of our study a protein-protein
interaction that is representative of a common class of biological control systems,
known as the two-component signal transduction system [21]–[24]: the
interaction of SpaK and SpaR from Bacillus subtilis, which regulate
the biosynthesis of subtilin, an antimicrobial peptide lantiobiotic that inhibits
growth of a broad range of pathogenic Gram positive bacteria [25]–[27]. In this
study we introduce a structural bioinformatics methodology for identification of
putative protein-protein complexes, and we apply it to characterize the interactions
between SpaK and SpaR. We generate structure homology models of SpaK and SpaR, and
then use these models to identify multi-domain protein structures that have good
structure homology to the models. Using one of the so-identified domain-fusion
templates, we generate a model representing a hypothetical physical interaction
between SpaK and SpaR, which enables further analyses of residues involved in the
protein-protein interaction. In this way we extend the well-known sequence-based
domain-fusion method by leveraging structural data, and use it to generate
hypotheses regarding the interactions between the two proteins. We further report
the results of biochemical studies on wild type and mutant proteins that
characterize the interactions between SpaK and SpaR, and we assess the resulting
structural model of a putative SpaK/SpaR complex arising from our structure-driven
domain-fusion approach. Furthermore, our biochemical analyses confirm that SpaK
autophosphorylates and subsequently transfers a phosphoryl group to SpaR.
SpaK (gi: 6226707, Uniprot P33113) and SpaR (gi: 417799, Uniprot P33112) protein
sequences were input to the AS2TS protein structure modeling system ([28];
http://as2ts.llnl.gov/), which generated initial homology models
based on structures taken from the Protein Databank (PDB) (version released
December 11, 2007). Structural templates having global sequence homology to each
of SpaK and SpaR were further studied by examining domain-level homology.
As no suitable template for the N-terminal domain (218 residues) of SpaK was
identified, this domain was not modeled. Based on match length (227 residues),
e-value (4e-57), and sequence identity (28%), PDB entry 2c2a_A, a
sensor histidine kinase from Thermotoga maritima, was
identified as the primary template for modeling SpaK (Fig. 1). Additional templates identified by
AS2TS are shown in Supplemental Results Table S1. Two domains of SpaK (SpaK_d1:
residues 219–300 and SpaK_d2: 301–459) were modeled
separately, pending determination of relative conformation to be provided by
structure-driven domain-fusion analysis (see Results). Although identification
of a structure template with acceptable global sequence homology enables initial
model construction, there often remain sub-sequences in the protein of interest
that do not correspond to any portion of the template due to insertions or
deletions relative to that template. For this reason, and in order to construct
as complete a model as possible to confirm the fitness of the modeled complex,
the Local-Global Alignment (LGA) modeler gap-filling procedure (in-house
software) was used to construct necessary loops, gaps or insertions by
“grafting” in suitable regions from related structures in
PDB.
Similarly, SpaR was modeled as two separate domains, comprising residues SpaR_d1:
1–117 and SpaR_d2: 118–220. The N-terminal domain was
initially modeled based on the structural template 1mvo_A (crystal structure of
the PhoP receiver domain from Bacillus subtilis), which showed
the highest level of sequence identity (46%) to that domain (see
Supplemental Results Table S2). In order to complete the model,
the LGA gap-filling procedure was used to construct regions of missing
coordinates. PDB entry 2gwr_A, a response regulator protein from
Mycobacterium tuberculosis, was identified as the primary
template for homology modeling of the C-terminal domain of SpaR (match length
216, e-value 9e-58, sequence identity 30%). This template was also
used for the construction of the domain orientation (Fig. 2). Further refinement of the
constructed SpaK and SpaR models was performed based on the structure comparison
of modeled domains with other PDB templates that were structurally identified by
a PDB-search procedure using LGA and the PDB release of July 8, 2008. In all
created models the positioning of the sidechains for residues that were
identical in the template were copied to the models, and the coordinates for
missing side chain atoms were predicted using SCWRL [29].
The LGA software ([20], http://as2ts.llnl.gov/lga/) was used to perform structure homology
searches against the PDB database to identify all entries with detected
(LGA_S> = 35%) structural
similarity to any of the four modeled domains (see above) within the homology
models of SpaK and SpaR. We selected an LGS_S cutoff value of 35%
based on our observation that the number and quality of hits increased rapidly
at LGA_S< = 33% (data not
shown) and based on previous work [30] that determined
the minimal structure homology needed to assure quality of structure alignment.
Those entries with homology to both respective domains of SpaK and SpaR were
selected as putative domain-fusion templates for modeling a SpaK/SpaR complex
(Table 1). Reported in
Table 1 are the
sequence identities between SpaK or SpaR compared to each corresponding
domain-fusion template, whereby residue-residue correspondences were extracted
from the structure alignments between the models and the domain-fusion
templates. We do not report the PSI-BLAST calculated sequence identities, as
these are highly inaccurate and meaningless when calculated from sequence
alignments at low levels of sequence identity (i.e., below 10%).
The spaK and spaR genes were isolated from
Bacillus subtilus strain LH45, a subtilin-producing
derivative of strain 168 [31]. Synthetic oligonucleotide primers were used
to amplify spaR using methods described previously [32],[33]. Briefly, the
commercial vector pQE31 (obtained from Qiagen, Valencia, CA), was digested with
EcoRI and HindIII, and a fragment containing a truncated spaK
gene encoding the C-terminal half of SpaK was cloned into the multipurpose
cloning site of the QE31 vector to construct the pQE31-spaK
expression vector (Supplemental Fig. S1A). (Note that we succeeded in
expressing only the C-terminal residues of SpaK, as the full-length gene did not
yield an expression product.) The pQE31-spaR vector was
similarly constructed (details are shown in Supplemental Fig. S1B.
Vectors (MLD[pQE31-spaR] and
MLD[pQE31-spaK]) were transformed into
JM109. For expression of the histidine-tagged proteins, the expression plasmids
MLD[pQE31-spaK] and
MLD[pQE31-spaR] were transformed into
M15[pREP4] competent cells (Qiagen), and expressed according
to the manufacturer's protocol. Expressed His-tagged proteins were
purified using a Ni-NTA resin from Novagen to form slurries that were used to
pack a 1.6 cm column, and eluted proteins were dialyzed against a storage buffer
and stored in 50-ul aliquots at 80°C. A working stock was stored for
several weeks at 20°C. Protein concentrations were determined by Bio-Rad
protein assay using the manufacturer's protocol.
Mutant SpaK proteins were prepared by Ana-Gen Technologies (Palo Alto, CA) using
the Stratagene QuikChange Mutagenesis Kit. Synthetic forward and corresponding
reverse complement oligonucleotide primers were prepared for each of two
mutations introduced into SpaK (altered nucleotides are indicated in bold type):
at position H247 the histidine was changed to glutamine using forward primer
5′-GTGCTTTGGCACAAGAGATCAAGATTCCG-3′
and reverse primer: 5′-CGGAATCTTGATCTCTTGTGCCAAAGCAC-3′,
and at position G392 the glycine was changed to alanine using forward primer
5′-GTAAAAGACACGGCAAATGGATTTTCGG-3′
and reverse primer 5′-CCGAAAATCCATTTGCCGTGTCTTTTAC-3′.
Phosphorylation reactions were performed with each histidine-tagged SpaK wild
type and mutant protein in the absence and presence of histidine-tagged SpaR.
Upon addition of 32P-labeled ATP, reaction mixtures were incubated for 20
minutes at room temperature, after which the reactions were stopped by addition
of 5× phosphorylation sample buffer, then electrophoresed on a
12.5% SDS polyacrylamide gel. The gel was stained with Coomassie
blue, dried, and autoradiographed using Kodak X-OMAT AR film.
Phosphorimage analysis was performed to quantify incorporation and turnover of
phosphate in assays involving phosphorylation of 6xHis-SpaK. Four samples of
protein were incubated in the presence of 32P-labeled ATP, of which three were
followed by cold chase treatment with unlabeled 4 mM, 10 mM, or 50 mM ATP, using
reaction conditions described previously [34]. Samples were run
on a 12.5% SDS-PAGE gel and subjected to autoradiography (not shown)
and phosphorimaging. Image intensities of the radiolabeled-phosphorylated SpaK
gel bands were analyzed using the Molecular Dynamics Phosphorimager 400.
Thin-layer chromatography was performed using Polygram Cell 300 PEI cellulose
plates as described previously [35]. 6xHis-SpaK and 6xHis-SpaR were incubated
individually (SpaK) or in combination with 32P-labeled ATP in the absence or
presence of EDTA. One ul aliquots from each reaction were spotted onto TLC
plates, and chromatography was carried out in 0.75 M
KH2PO4, pH 3.75, after which the plate was dried and
autoradiographed.
The AS2TS protein structure modeling system [28] yielded over 30 and
over 140 PDB structures suitable as templates for modeling each of SpaK and
SpaR, respectively, from which were selected sets of the closest templates with
sequence identities ranging from 13% to 28% for SpaK and
24% to 46% for SpaR (see Supplemental Data Tables S1,
S2).
LGA-mediated structure homology searches against the PDB database using
constructed structural models of domains from SpaK (SpaK_d1, SpaK_d2) and SpaR
(SpaR_d1, SpaR_d2) yielded 6 domain-fusion templates with structural homology
(i.e., similarity based on structure alignment; [20]) ranging from
LGA_S = 37% to 95%, and
root mean square deviation (RMSD) calculated on superimposed C-alpha atoms
ranging from 1.11 to 2.96 (Table
1). Identification of domain-fusion templates suggested that SpaK and
SpaR interact forming an interface between domain 2 of SpaK and domain 1 of
SpaR. Sequence identities of SpaK and SpaR to corresponding template sequences
ranged from 4% to 25%, but in no instance was sequence
identity greater than 7% simultaneously to both SpaK_d2 and SpaR_d1.
Structural comparison of all identified domain fusion template structures showed
that they clustered into two distinct conformations, yielding the following
groups: (1) 1f51_AE and 2ftk_AE (Spo0F/Spo0B from B. subtilus),
and (2) 1th8_AB, 1thn_AB, 1tid_AB and 1til_AB (SpoIIAB/SpoIIAA from B.
stearothermophilus). PDB entry 2ftk was determined to be the
optimal domain-fusion template for modeling a SpaK/SpaR complex based on the
highest structure similarity to the corresponding two modeled domains: SpaK_d2
and SpaR_d1, and based on the expected intermolecular distance between the
putative functional residues H247 of SpaK and D51 of SpaR that were predicted as
active site residues (His and Asp) critical for exchanging a phosphoryl group
[36]. In order to form a covalent bond with the
phosphoryl group, the distances between atoms N of His and O of Asp were
expected to be in the range of about 5 Angstroms. The models created based on
templates 1f51 and 2ftk satisfied this requirement. 2ftk was also used to
complete the homology model of SpaK (Fig. 1) by providing relative positioning of
the central (SpaK_d1) and C-terminal (SpaK_d2) domains. The SpaK/SpaR complex
was modeled as a trimer, comprising a SpaK homo-dimer and a SpaR monomer, based
on the domain conformation between chains A and E from 2ftk (Fig. 3). The constructed model
of a SpaK/SpaR complex agreed with structural analysis of the Spo0F and Spo0B
interaction reported by Varughese and coworkers [37], who showed that
the geometry of Spo0F binding to Spo0B favors an associative mechanism for
phosphoryl transfer. In order to visualize the autophosphorylation of the
histidine kinase, and the subsequent phosphoryl transfer to Spo0F, they
generated in silico models representing these reaction steps,
proposing Spo0B as a model for the autokinase domain of KinA (histidine kinase,
consisting of an N-terminal sensor domain and a C-terminal autokinase domain).
The level of sequence identity between KinA and SpaK is about 27%,
and the KinA sensor domain comprises three PAS (Per-Arnt-Sim) domains that
correspond to the N-terminal part of SpaK (1–218; not modeled). The
autokinase domain corresponds to the modeled C-terminal part (219–459)
of SpaK, and consists of a phosphotransferase subdomain and an ATP binding
subdomain. In modeling SpaK we followed Varuguese and coauthors'
suggestion that the four-helix bundle of Spo0B is formed through the
dimerization of two helical hairpins from two monomers, and that it is a
prototype for the phosphotransferase domains of histidine kinases (see Fig. 1A). This concept is
supported by the high degree of structure similarity between the C-terminal
domain of Spo0B and the ATP binding domains of histidine kinases, as well as by
a report [38] of the crystal structure of the entire
cytoplasmic portion of a histidine kinase (a PDB structure, 2c2a), which we used
as a primary template for modeling individual domains of SpaK.
Inspection of the constructed SpaK/SpaR complex (Fig. 3A) allowed us to identify specific
residues putatively involved in the interaction between SpaK and SpaR or
believed to mediate transfer of phosphate from SpaK to SpaR (Fig. 3B). Specifically, we
identified the histidine residue at position H247 in SpaK that corresponds to
the histidine H30 that is phosphorylated in Spo0B (PDB entry 2ftk_A) (Table 2A), and we identified
3 aspartate residues in close proximity in SpaR (D8, D9, and D51), which we
presumed to be involved in transfer of a phosphoryl group bound to the H247
residue of SpaK, if SpaK and SpaR truly mediate a phosphorelay as postulated.
These residues corresponded to their equivalents (D10, D11, and D54) in Spo0F
(PDB entry 2ftk_E) (Table
2B). Three additional functional residues were identified, which
corresponded to functional residues that are highly conserved among response
regulator proteins [37]: T78, Y97, and K100 in SpaR, corresponding to
T82, H101, and K104, respectively, of Spo0F (Table 2B). Under global superposition, the
distances between corresponding functional residues were below 0.8 Angstroms and
the local RMSD(3) (root mean square deviation along the main-chain atoms
(N,CA,C,O) averaged over three residues: current and immediate neighbors along
peptide chain (local superposition); [20]) values were below
0.5 Angstrom, indicating significant structure similarity in corresponding
regions. The sites of phosphorylation, D51 of SpaR and H247 of SpaK, which
correspond to D54 of Spo0F and H30 of Spo0B, are shown in Figure 3.
In most histidine kinases the extracellular sensing domains are variable in
sequence, reflecting the wide range of environmental signals to which they
respond. Conversely, the cytoplasmic portions typically have a conserved
catalytic core comprising a set of characteristic sequence motifs known as the
H, N, G1, F and G2 boxes [39],[40] and can be dissected
into several distinct functional units [41],[42].
Corresponding functional units P1 through P5 were evident upon examination of
residues 219 through 459 of our modeled SpaK protein (Fig. 1B), which were determined to comprise
an N-terminal dimerization and histidine phosphotransfer domain (DHp; SpaK_d1)
and a C-terminal catalytic and ATP-binding domain (CA; SpaK_d2). P1 had a
conserved histidine residue (H247) belonging to the autophosphorylation site
known as the “H box”. Autophosphorylation was presumed to
occur from ATP in the active site of P4 (the kinase domain) to H247 of P1,
followed by transphosphorylation from H247 to an aspartate residue (D51) of
SpaR. P2 functional units have a specific domain for recognizing the response
regulator and assisting transfer of the phosphoryl group. P3 corresponds to the
linking domain, through which two SpaK subunits may form a dimer. P4 resembles
the ATP binding domain, which autophosphorylates the conserved histidine
residue. In histidine kinases most of the residues around the ATP binding site
of the P4 unit are conserved, especially those comprising the characteristic
sequence motifs (identified in Fig.
1B). In addition, the histidine kinase P4 unit has a loop-like lid
(ATP lid) between the F and G2 boxes (corresponding to the SpaK model, residues
409 to 417), which controls the closed-to-open conformational change of the
binding pocket. It is postulated that P5 acts as a regulative domain to modulate
the activity of autotransphophorylation, responding to signals from the external
environment [41].
To examine sequence homology in structure context between SpaK and various
histidine kinases in the 5 “box” regions, we used LGA to
globally align the SpaK homology model with all other histidine kinases from PDB
that have these structure motifs. Structures with corresponding
“box” regions included 2ftk_A, 1tid_A, 1b3q_A, and 2ch4_A.
In Table 3 are shown
structure-based alignments, including residue-residue correspondences, between
our SpaK model (based on 2c2a) and 2ftk_A in the H-box regions, and between SpaK
and 2ch4_A in the N-, G1-, F-, and G2-box regions. Calculated structural
alignments between our SpaK model and the PDB structures (including those not
shown) indicated significant structure conservation within these defined
sequence motifs. The residue-residue correspondences arising from the LGA
structure alignments were consistent with respect to highly conserved residues
identified by Stock and coworkers [21] and by Grebe and
Stock [43] (see bold-type residue-residue correspondences in
Table 3), even in the
more variable F-box regions. Within group HPK-3c, a small group of histidine
kinases into which Grebe and Stock [43] classified SpaK,
most histidine kinases have an F at the position corresponding to T404 in SpaK,
whereas SpaK T404 corresponds to a T in some proteins in group HPK 1a.
Furthermore, SpaK F407-Y408 has identity to the corresponding F-box FY in most
proteins in group HPK 1a. As group HPK 3c is closely related to group HPK 1a, it
is not surprising that there is ambiguity with respect to residue-residue
correspondences within the relatively variable F box among the proteins in these
two groups. Based on this ambiguity, we examined the alpha-carbon structure
alignment between the SpaK model and 2ch4_A to verify that the side chains of
the corresponding SpaK Y408 and 2ch4_A F491 were well aligned (not shown), which
further supported the residue-residue correspondence between these two residues.
Protein CheA (2ch4) is classified in group HPK 9, and as such the sequence
alignment also shows an F in the position corresponding to SpaK Y408.
To confirm whether SpaK undergoes auto-phosphorylation and subsequently transfers
a phosphate moiety to SpaR, each protein was tested individually and in
combination in the presence of radio-labeled ATP (Fig. 4). Combinations of 6xHis-SpaK and
6xHis-SpaR were created using 3 SpaK∶SpaR molar ratios of
4∶1, 4∶3, and 1∶2 shown in Fig. 4 A and B, lanes 3, 4, and 5,
respectively. Only SpaK was phosphorylated in isolation (Fig. 4B lanes 1, 2), indicating that SpaK
undergoes autophosphorylation. Phosphorylation of SpaR in the presence of SpaK
(Fig. 4B lanes
3–5) indicated that phosphate is transferred from SpaK to SpaR. This
transfer was incomplete at a molar ratio of SpaK∶SpaR of
4∶1, but reached completion at molar ratios of 4∶3 and
1∶2, indicating that transfer of phosphate from SpaK to SpaR reaches
saturation as SpaK approaches molar equivalence or reaches molar excess relative
to SpaR. These results imply that SpaR acts as a receptor for the phosphate
group that is transferred from SpaK.
Quantification of radio-labeled phosphate-bound 6xHis-SpaK was performed to
determine whether SpaK might exhibit phosphatase activity (Fig. 4C). Phosphor image analysis was used to
measure the incorporation of radio-labeled phosphate by 6xHis-SpaK (Fig. 4C, histogram 1). This
quantity served as baseline (100%) for comparison of 6xHis-SpaK
samples that had been incubated in radio-labeled Pi followed by cold-ATP chase
treatments (Fig. 4C,
histograms 2–4). Cold chase with lower concentrations of ATP (4 mM or
10 mM) reduced the level of radio-labeled SpaK to levels about one-third to
one-quarter that of the control, whereas a high concentration (50 mM) of
unlabeled ATP resulted in a decrease in the rate of phosphate turnover, thereby
reducing the level of radio-labeled SpaK only to about 70% that of
the control. The decrease in the turnover of radio-labeled Pi on SpaK at high
ATP concentration is suggestive of enzymatic inhibition of dephosphorylation (or
phosphatase activity) rather than simple hydrolysis.
Thin-layer chromatography was performed to further examine the possibility that
either SpaK or SpaR may exhibit phosphatase activity (Fig. 4D). Protein consisting of 6xHis-SpaK
alone (Fig. 4D, lane 2) or
6xHis-SpaK in combination with 6xHis-SpaR (lane 3) was phosphorylated in the
presence of radio-labeled ATP. In both cases, inorganic phosphate (Pi) was
detected, but slightly more Pi and considerably more radio-labeled protein were
detected when both proteins were present (compare Pi and Protein in lanes 2 and
3). The ATP-only control (lane 1) produced no detectable radio-labeled Pi,
indicating that simple hydrolysis of ATP was not occurring. Furthermore, when
phosphorylation was performed in the presence of EDTA, some phosphorylated
protein was observed, although no inorganic phosphate was detected (Fig. 4D lane 4). This result,
taken together with Fig. C, which suggested the presence of enzymatic
phosphatase activity, supports the claim that SpaK (and possibly also SpaR) may
possess enzymatic phosphatase activity.
Based on amino acid sequence alignment with other histidine kinases, the highly
conserved histidine at position H247 was presumed to be the site of possible
auto-phosphorylation, and a glycine located at position G392 in the C-terminal
end of SpaK was determined to correspond to the conserved DXG motif of the
nucleotide binding domain in related histidine kinases (Fig. 1A, Fig. 1B: H box and G1 box). In the
superfamily of phosphotransferases, the conserved residues that form a
corresponding motif (DXG in actin, GTG in hexokinase/glycerol kinase, and GNG in
acetate and propionate kinases) are observed to be present in binding to a- and
b-phosphate groups of the nucleotide [44]. Because several
histidine kinases are believed to exist as homo-dimers and it is believed that
phosphorylation occurs in trans, in which one monomer binds ATP in the
nucleotide-binding domain and then transfers the phosphoryl group to a histidine
located in the other monomer, we postulated that mutations at either of these
positions might reduce or abolish auto-phosphorylation of SpaK, but that
complementation between mutants might occur, effectively restoring function. We
used site-directed mutagenesis to construct two mutants (see Materials and Methods): one in which the
histidine at position H247 was changed to a glutamine (H247Q), and the other in
which the glycine at position G392 was changed to alanine (G392A). Locations of
mutated residues are shown in Fig.
1A. Phosphorylation studies of mutants H247Q and G392A revealed that
both mutations resulted in loss of phosphorylation when each mutant was tested
individually (Fig. 5 A, B;
lanes 4, 5) or when individually combined with SpaR (Fig. 5B; lanes 9, 10). However, when the
mutant proteins were combined, a detectable amount (approximately 25%
that of wild type) of auto-phosphorylation was observed (Fig. 5B, lane 6), suggesting that
complementation between the mutants had occurred, and supporting the hypothesis
that SpaK forms a homo-dimer. Furthermore, when H247Q and G392A together were
subjected to phosphorylation in the presence of wild type SpaR, the phosphoryl
moiety was transferred to SpaR (Fig. 5B, lane 12).
In this work we demonstrated a quantitative approach for modeling protein-protein
complexes using homology modeling followed by structure-based searches for
multi-domain template proteins. In a search for templates upon which to base the
model of a putative SpaK/SpaR complex, we used LGA, which applies two scoring
schemes: GDT (global distance test) and LCS (longest continuous segment). Based on a
previous study involving structure alignments between weakly homologous proteins
[30],
we applied a relatively stringent cutoff
(LGA_S> = 35%)—Pettitt
and coworkers [30] concluded that in order to assure the quality of
a structure alignment between two domains, the GDT_TS score (a component of
LGA's GDT) must exceed 25. In the current study we had observed a rapid
increase in the number of hits obtained using
LGA_S = 33% and below (not shown), and
therefore we selected LGA_S = 35% as a
conservative cutoff to assure confidence in selecting templates.
Although our approach can be used to identify domain-fusion protein structures that
imply a possible functional association between two proteins of interest, it does
not in itself provide sufficient information for modeling a physical interaction
between the proteins. Protein domains that have less than
30–40% sequence homology to a
“domain-fusion” template are likely to assume a similar
orientation [8],[45]–but at sequence identity levels below
this “interaction similarity twilight zone”, additional analysis
is needed to make a reasonable prediction regarding the relative orientation of the
interacting domains. In the current study, this additional analysis included
identification and inspection of putative functional residues coupled with
experimental analysis of mutant proteins. Thus, a protein-protein-complex model for
a SpaK/SpaR interaction was initially built based on a structure-driven
domain-fusion search method, followed by validation based on bioinformatic analysis
and experimentation.
Our modeling effort supported the hypothesis that SpaK and SpaR may function as a
histidine kinase sensor and a response regulator, respectively, in a two-component
system. Based on homology modeling and domain-fusion analysis, residues
corresponding to those believed to function in phosphorylation and subsequent
transfer of a phosphate moiety from sensor to response regulator in other
two-component systems were identified (Fig. 3, Tables 1,
2). Modeling of SpaK
enabled structure comparisons with related sensor proteins (2ftk_A, 1tid_A, 1b3q_A,
2ch4_A), identification of sequences corresponding to the 5 highly conserved regions
(“boxes”) that characterize class II two-component system
proteins [40],[41],[43] (Table 3), and mapping of these sequences to the homology model of SpaK
(Fig. 1B). Functional
residues and conserved sequence motifs of our modeled SpaK/SpaR complex matched well
with those of known sensor/response-regulator systems. Structure-based
residue-residue correspondences (Tables 2, 3)
agreed with sequence alignments used previously to classify histidine kinases [43],[46],[47], in
which SpaK was placed in group HPK 3c in an 11-group classification by Grebe and
Stock [43],
but was unclassified according to the 5-type classification of Kim and Forst [46].
Phosphorylation studies of SpaK and SpaR showed that SpaK auto-phosphorylates and
subsequently trans-phosphorylates SpaR (Fig. 4), confirming the hypothesis based on structure-driven
domain-fusion analysis that SpaK and SpaR are functionally related and physically
interact, and that the quaternary structure of the complex could enable transfer of
a phosphate moiety between the protein subunits. Phosphorylation and complementation
analyses using SpaK mutants suggested that residues H247 and G392 are important for
auto- and trans-phosphorylation and that SpaK likely forms a dimer in which ATP
binding and hydrolysis functions are split between the protomers (Fig. 5). Whereas both SpaK mutants
(H247Q and G392A) were deficient in auto-phosphorylation (Fig. 5, lanes 4,5), this function was apparently
restored when the mutants were combined (Fig. 5, lane 6), suggesting that complementation
had occurred between the mutants. Complementation between H247Q and G392A also
apparently restored trans-phosphorylation, as evidenced by phosphorylation of SpaR
in the presence of both mutants (Fig.
5, lane 12). In an equimolar mixture of mutants H247Q and G392A, one would
expect that approximately one-half of the resulting dimers would comprise a protomer
of each mutant. Furthermore, phosphorylation would occur from the H247Q mutant to
the G392A mutant, but not in the other direction, since G392A should not be able to
bind ATP. Therefore the levels of auto-phosphorylation or trans-phosphorylation
would not be expected to exceed one-half those of wild type SpaK. Also, although the
H247Q/G392A mixed dimer may have had restored function, it would be expected to have
functioned at less than the efficiency of a wild type SpaK dimer; since dimer
formations between non-productive forms would occur, one would expect
phosphorylation to proceed more slowly than in the wt. This is consistent with the
observation that phosphorylation of or by H247Q combined with G392A (lanes 6, 12)
occurred at levels considerably below those of wild type SpaK (lanes 3, 8).
In modeling the interaction between SpaK and SpaR we identified 6 suitable
domain-fusion templates (Table
1), which were structurally clustered into two groups (see Results), each
having a distinct conformation. Both groups displayed the same interaction pose with
respect to the domain-domain interaction. Although each of the identified
domain-fusion templates would have yielded a SpaK/SpaR complex model consistent with
the experimental data, the criteria for selecting 2ftk as the domain-fusion template
were based on combined structural identities between domains of 2ftk and the SpaK
and SpaR models, on the resulting distance between putative functional residues
involved in phosphate transfer (Fig.
3), and on the presence of a helical bundle domain, which enabled
construction of a complete model. Interestingly, the domain-domain conformation
between the helical bundle and the ATPase domains of 2c2a, used for modeling SpaK,
differed from that of the corresponding domains within 2ftk. This difference
suggests the possibility that a conformational change might take place when SpaK
interacts with SpaR. Furthermore, it should be noted that the phospho-transfer in
Spo0B-Spo0F (2ftk) occurs in the opposite direction (Asp to His) as that
demonstrated here in SpaK-SpaR (Figs.
4, 5). This is not
surprising, and does not diminish the value of 2ftk as a template for modeling a
SpaK/SpaR interaction, given the considerable mechanistic diversity observed among
structurally conserved domains comprising sensor/response-regulator systems [48].
Although structure modeling and experiments involving phophorylation studies strongly
suggest functional and physical interactions between SpaK and SpaR, we cannot be
entirely certain that our quaternary structure is correct with respect to domain
composition, conformation, or orientation, as the methodology is dependent on
existing structural data within PDB; it is possible that none of the domain-fusion
templates detected by our approach is truly representative of the physical
interaction between SpaK and SpaR, as homology modeling is, by definition, data
driven. Due to the low sequence homologies between SpaK and SpaR and the identified
domain-fusion templates, one could not conclude with any degree of certainty based
solely on template identification that the interaction pose modeled here is likely
to be correct [8]. However, combining bioinformatics analysis of known
functional motifs (sequence “boxes”) and putative interacting
residues with experimental evidence of function allows us to assert the value of the
homology model of a putative SpaK/SpaR protein-protein complex. Our approach detects
existing putative domain-fusion templates, which may suggest testable hypotheses
regarding quaternary structure and function; a structure-based approach for
identification of “Rosetta Stone” proteins greatly enhances
structure-function hypothesis generation by providing structural context for
putative functional residues. Additional bioinformatics analyses of a putative
protein-protein complex model, which may verify the correctness of the model,
include alignments of modified sequence profiles [7], for example, which
use quantitative methods applied at the domain-domain interface to evaluate the
likelihood of a stable interaction.
Although many two-component signal transduction systems have been identified by
sequence homology, we wish to point out that a purely sequence-based approach would
not have yielded the structural domain-fusion templates that were identified in this
study. The strength of our approach is in its ability to identify putative
domain-fusion templates based on structure homology searches in cases where sequence
identities between the proteins of interest and the putative domain-fusion templates
are low. Sequence identities of candidate domain-fusion templates to domains of SpaK
and SpaR ranged from 4% to 25%, but in no instance was
sequence identity greater than 7% simultaneously to both (Table 1). This point is
emphasized by the lack of sufficient sequence-based evidence for linking these
proteins using the standard domain-fusion approach: as of this writing, SpaK and
SpaR are not linked in this way, for example, in Prolinks [5], nor did we find them
linked by other sequence-based or empirical methods in DIP, BIND/BOND, MIPS, IntAct,
MPIDB, or InterPreTS [49]–[54]. Homology modeling of
SpaK and SpaR using a standard methodology [28] and subsequent
structure-based searches using a quantitative structure comparison algorithm [20] is what
enabled a more sensitive, structure-based homology search against PDB. In
conclusion, our method provides a basis upon which a high-throughput system for
identification of putative protein-protein interactions could be built on a
whole-genome scale.
|
10.1371/journal.pgen.1001378 | GWAS of Follicular Lymphoma Reveals Allelic Heterogeneity at 6p21.32
and Suggests Shared Genetic Susceptibility with Diffuse Large B-cell
Lymphoma | Non-Hodgkin lymphoma (NHL) represents a diverse group of hematological
malignancies, of which follicular lymphoma (FL) is a prevalent subtype. A
previous genome-wide association study has established a marker, rs10484561 in
the human leukocyte antigen (HLA) class II region on 6p21.32 associated with
increased FL risk. Here, in a three-stage genome-wide association study,
starting with a genome-wide scan of 379 FL cases and 791 controls followed by
validation in 1,049 cases and 5,790 controls, we identified a second independent
FL–associated locus on 6p21.32, rs2647012
(ORcombined = 0.64,
Pcombined = 2×10−21)
located 962 bp away from rs10484561 (r2<0.1 in controls). After
mutual adjustment, the associations at the two SNPs remained genome-wide
significant (rs2647012:ORadjusted = 0.70,
Padjusted = 4×10−12;
rs10484561:ORadjusted = 1.64,
Padjusted = 5×10−15).
Haplotype and coalescence analyses indicated that rs2647012 arose on an
evolutionarily distinct haplotype from that of rs10484561 and tags a novel
allele with an opposite (protective) effect on FL risk. Moreover, in a follow-up
analysis of the top 6 FL–associated SNPs in 4,449 cases of other NHL
subtypes, rs10484561 was associated with risk of diffuse large B-cell lymphoma
(ORcombined = 1.36,
Pcombined = 1.4×10−7).
Our results reveal the presence of allelic heterogeneity within the HLA class II
region influencing FL susceptibility and indicate a possible shared genetic
etiology with diffuse large B-cell lymphoma. These findings suggest that the HLA
class II region plays a complex yet important role in NHL.
| Earlier studies have established a marker rs10484561, in the HLA class II region
on 6p21.32, associated with increased follicular lymphoma (FL) risk. Here, in a
three-stage genome-wide association study of 1,428 FL cases and 6,581 controls,
we identified a second independent FL–associated marker on 6p21.32,
rs2647012, located 962 bp away from rs10484561. The associations at two SNPs
remained genome-wide significant after mutual adjustment. Haplotype and
coalescence analyses indicated that rs2647012 arose on an evolutionarily
distinct lineage from that of rs10484561 and tags a novel allele with an
opposite, protective effect on FL risk. Moreover, in an analysis of the top 6
FL–associated SNPs in 4,449 cases of other NHL subtypes, rs10484561 was
associated with risk of diffuse large B-cell lymphoma. Our results reveal the
presence of allelic heterogeneity at 6p21.32 in FL risk and suggest a shared
genetic etiology with the common diffuse large B-cell lymphoma subtype.
| Non-Hodgkin lymphoma (NHL) represents a diverse group of B- and T-cell malignancies
of lymphatic origin. The most common subtypes are of B-cell origin and are further
classified on the basis of their resemblance to normal stages of B-cell
differentiation [1]. Epidemiological studies indicate that these may have
different environmental and genetic risk factors, although some etiological factors
may also be shared [2]. Familial studies provide substantial evidence for a
genetic influence on susceptibility to the major mature B-cell neoplasms, including
diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL) and chronic
lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) [3], [4]. Recent genome-wide association
studies (GWAS) of the FL subtype of NHL identified associations with two variants
within the human leukocyte antigen (HLA) region, one at 6p21.33 (rs6457327) [5] and the other
at 6p21.32 (rs10484561) [6]. Additional true associations, particularly in the HLA
region, may have been missed because a limited number of samples were used in the
initial genome-wide screens, and the selection of a few top single nucleotide
polymorphisms (SNPs) for validation is further subject to chance. In this study, we
conducted a larger independent genome-wide scan of FL using 379 cases and 791
controls from the Scandinavian Lymphoma Etiology (SCALE) study of Sweden and
Denmark, which was used in the validation of the previous GWAS [6]. This scan was followed by two
stages of validation in European-ancestry cases of FL and other common B-cell NHL
subtypes and controls from the US, Canada and Australia (Table 1, Table S1, Table S2, Figure 1).
In total, 298,168 SNPs were analyzed in Stage 1 (λ = 1.028;
λ1000 = 1.055 [7]), in which we observed
suggestive associations (adjusted trend P-value<10−5) at 4q32.3,
6p21.32 and 10q25.3 (Table S3) with the strongest at rs2647012 (odds ratio
(OR) = 0.58, PPCAadjusted =
1.59x10−7) within the HLA class II region on 6p21.32. Sixteen
SNPs in close proximity to the HLA-DQ genes showed association with
adjusted P-values<10−4, including the previously reported
rs10484561 (Figure 2, Table S4) [6]. The previously
reported HLA class I associated SNP rs6457327 [5] was modestly associated with FL
risk (OR = 0.82, P = 0.03) in Stage 1, and
was not in linkage disequilibrium (LD; r2 = 0) with
any of the top 100 SNPs.
In Stage 2, we carried out an in silico validation of the top 40
SNPs from Stage 1 (Table S5) in 213 FL cases and 750 controls from the San Francisco Bay
Area, USA (Table 1), the study
that reported an association at 6p21.32 [6]. Among 38 out of 40 SNPs, seven
showed association (P<0.05) in Stage 2 (Table S5), six of which were located within the
6p21.32 region. We tested the independence of multiple association signals in
6p21.32 using a stepwise logistic regression analysis (entering SNPs based on a
criterion of likelihood ratio test p-value<0.05) and found that with rs2647012
(the top SNP within the region) forced in the model, only the addition of rs10484561
contributed significantly to the association with increased risk of FL. The OR for
this SNP, adjusted for rs2647012, was 1.43, P = 0.006 (Table S6).
After excluding previously identified and non-independent association signals, we
selected rs2647012, and an additional four top SNPs to be taken forward to a third
stage (Table
S7, S8), wherein these were genotyped in 836 FL cases and 3202 controls from
the Mayo Clinic (US) [8], National Cancer Institute-Surveillance, Epidemiology and
End Results (NCI-SEER, US) [9], Yale University (US) [10], New South Wales (NSW,
Australia) [11]
and British Columbia (BC, Canada) [12] studies. The association of rs2647012 with FL was
validated, showing consistent associations with similar ORs (no heterogeneity,
P = 0.32) across all independent studies and reaching
genome-wide significance in both the combined analysis of the validation samples
(P = 3×10−15) and the combined
analysis of all three stages (1428 FL cases, 4743 controls;
OR = 0.64,
P = 2×10−21) (Table.2, Figure 3). After adjustment for rs10484561, the
association at rs2647012 remained genome-wide significant with minimal change in
magnitude (ORadjusted = 0.70,
Padjusted = 4×10−12). The
LD between the two SNPs is low (r2<0.1 in the SCALE controls and
HapMap CEU [Utah residents with northern and western European ancestry]
samples release27). Taken together, our results suggest that the association at
rs2647012 is independent from rs10484561, and tags a different disease-predisposing
variant. We also found suggestive evidence for an association at rs6536942 on 4q32.3
(OR = 1.36,
P = 2×10−5) (Table 2, Figure S1A).
To fine-map the association signals in the HLA class II region, we imputed 10,639
SNPs within 600 kb surrounding the top SNP rs2647012 using data from the 1000
Genomes (1000G, 60 CEU subjects, August 2009) and HapMap projects (HapMapII release
22, CEU) in Stage 1. Among the imputed SNPs, 258 SNPs located in a strong LD block
of 236 kb (r2>0.8) showed stronger evidence of association than all
the genotyped SNPs within the region (Figure S2). Since a moderate discordance of
reference genotypes was observed between 1000 G and HapMapII, we analyzed only SNPs
showing a concordance of >95% in the two datasets and identified the
strongest association at rs9378212 (OR = 1.66,
P = 3.21×10−8), located 219 kb
upstream of rs2647012 (r2 = 0.56 in controls). We
subsequently confirmed the imputed genotypes by Taqman genotyping in 345 of the FL
case subjects used in Stage 1 and found a 99.4% concordance with the imputed
genotypes, demonstrating high confidence in the results of the imputation.
Next, we performed a haplotype analysis using rs2647012, rs10484561 and an additional
12 adjacent genotyped SNPs located within a block of minimal recombination. Out of
the eight haplotypes identified, three were neutral
(OR = 0.9–1.1), three increased risk (ORs>1.2;
strongest risk haplotype tagged by rs10484561) and two were protective (OR≤0.8;
both tagged by rs2647012) (Table S9), suggesting the presence of at least
two susceptibility alleles within the region. Coalescence analysis of the eight
haplotypes indicated that rs2647012 and rs10484561 arose on two distal branches of
the ancestral recombination graph [13] (Figure S3), which was also supported by the
analysis of median-joining network [14] using seven SNPs without any recombination (Figure 4). Further haplotype
analysis of the seven genotyped SNPs (Table S9) and the imputed SNP rs9378212 indicated
that the two alleles of rs9378212 tag the two different evolutionary lineages (Figure 4), each harboring either
rs2647012 or rs10484561. Thus, the associations at the two SNPs are likely due to
two distinct susceptibility variants, instead of a single risk allele, that arose
independently on different haplotype backgrounds.
The FL-associated SNP, rs10484561, was previously found to tag the extended haplotype
HLA-DQA1*0101-HLA-DQB1*0501-HLA-DRB1*0101
[6]. Here, to test
whether any HLA class II alleles may also be responsible for the observed
association at rs2647012, we imputed known HLA tag SNPs [15], [16] using data from the 1000G and
HapMapII European datasets. We confirmed the association of the
HLA-DRB1*0101-HLA-DQA1*0101-HLA-DQB1*0501 extended
haplotype, tagged by rs10484561. The association at rs2647012 remained significant
after adjustment for these three HLA alleles (OR = 0.64,
P = 8.11×10−6), suggesting that
these are not driving the association at rs2647012. Furthermore, rs2647012 was not
in strong LD (r2<0.8 in HapMap CEU or SCALE controls) with any other
known HLA tags [15], including those tagging FL-associated alleles previously
reported [17], [18]
(r2<0.39 with the six HLA-DRB1*13 tag SNPs
[rs2395173, rs2157051, rs4434496, rs6901541, rs424232, rs2050191] [17] and
r2<0.25 with the three HLA-B*0801 and
HLA-DRB*0301 tag SNPs [rs6457374, rs2844535,
rs2040410] [15]). Of the other 17 HLA class II alleles (∼39%
of all the class II alleles) that could be imputed, none showed significant
association or were found to be responsible for the association at rs2647012 (Table S10).
Detailed HLA allelotyping on large numbers of cases and controls is needed to
determine if particular HLA class II alleles are responsible for the observed
association at rs2647012.
To assess whether the FL-associated SNPs may be involved in the development of other
NHL subtypes, we genotyped the five SNPs selected for Stage 3 together with
rs10484561 in a total of 1592 DLBCL, 1075 CLL/SLL, 336 marginal zone lymphoma (MZL),
262 mantle cell lymphoma, 306 T-cell lymphoma and 878 rare or unspecified NHL cases
and 5220 controls from the SCALE2, SF2, BC, Mayo, NCI-SEER, Yale and NSW studies
(Table 1, Table S1, Figure 1). Among these SNPs,
rs10484561 showed evidence of association with DLBCL
(OR = 1.36,
P = 1.41×10−7) (Figure S1B) and
all NHL (OR = 1.23,
P = 6.81×10−7). ORs were consistent
across the seven studies. There was also a suggestive association for rs2647012 with
MZL (OR = 1.32,
P = 6.34×10−4) (Table.3), consistent across six
studies.
Finally, we investigated the possibility of additional susceptibility loci for FL
outside of the HLA region by performing a joint analysis of the top 41 to 1000
variants of our scan and the previously published GWAS of follicular lymphoma [6]. From this
combined analysis, we did not find any additional markers with a strong association
(P<10−6) with FL that were not in LD with our top 5 markers
taken forward to stage 3 (data not shown).
Through the identification of a second variant, rs2647012, that is independent of the
previously identified risk variant rs10484561 [6] within the 6p21.32 region, our
findings substantiate a major link between HLA class II loci and genetic
susceptibility to FL. In addition, our study revealed evidence that rs10484561 is
associated with DLBCL risk suggesting some shared biological mechanisms of
susceptibility between these two common NHL subtypes. The association of rs2647012
with FL risk was not detected in earlier GWAS studies [5], [6], and that of rs10484561 with
DLBCL risk previously reported was only marginal [6], perhaps because of the smaller
sample sizes in Stage 1. The number of FL cases scanned in this study was almost
double compared to the previous individual GWAS [6].
HLA class II molecules are expressed in antigen presenting cells such as
B-lymphocytes, and act to present exogenous antigens to CD4+ helper T-cells.
Efficiency of antigen presentation may influence lymphomagenesis through effects on
anti-tumor immunity or on immune response to infections that are directly or
indirectly oncogenic (e.g., through viral genome insertion or nonspecific chronic
antigenic stimulation) [19]. Allelic variants in coding regions may affect the
structure of the peptide binding groove of the class II molecules, leading to
differences in the efficiency of oncogenic peptide binding or T-cell recognition.
Coding sequence variation in the molecules encoded by the extended
HLA-DRB1*0101-HLA-DQA1*0101-HLA-DQB1*0501 haplotype
may be responsible for the association at rs10484561 [6].
Alternatively, variants in the regulatory sequences may influence the expression
level of the HLA molecules and consequently the efficiency of antigen presentation.
We note that rs2647012 is strongly associated with the average expression levels of
HLA-DRB4 (β = 0.78,
P = 3.4×10-22) and
HLA-DQA1 (β = -0.58,
P = 5.1×10−13) probes in
Epstein-Barr virus-transfected lymphoblastoid cell lines (mRNA by SNP browser) [20], and rs10484561
is also associated with the expression levels of HLA-DQA1 probes
(β = -0.884,
P = 1.6×10−10). We speculate that
this may be an alternative mechanism underlying the observed associations,
especially at rs2647012.
Interestingly, SNPs within the same LD block harboring rs2647012
(r2>0.7 in HapMap CEU) have previously been associated with rheumatoid
arthritis with the same direction of effect [21]. Since autoimmune disorders
such as rheumatoid arthritis and Sjögren syndrome are associated with increased
risk of NHL, in particular with DLBCL but also with FL [22], our finding may suggest a
molecular link between these diseases, although their associations within this
region of high LD could also be due to different causal variants.
Previously, large-scale candidate gene studies have pointed to susceptibility loci in
the HLA class III region mainly between the TNF variant
–308G->A (rs1800629) and risk of DLBCL [23], [24]. We provide
novel evidence of association of DLBCL with an independent HLA marker in the class
II region (rs10484561; r2 = 0), 1.1Mb away from
rs1800629, strongly suggesting that alleles in the HLA class II region may play an
important role in the pathogenesis of this subtype as well. The weaker association
of rs10484561 with DLBCL (OR 1.36) than with FL (OR 1.95) [6] could imply that the
DLBCL-association is confined to a subset of DLBCL tumors with specific
morphological or molecular features more closely related to FL, such as the germinal
center-like B-cell phenotype [25]. However, the observed effects could also be due to
modification of other concurrent DLBCL-specific susceptibility variants, or
rs10484561 could tag a more strongly associated marker in this region of high
LD.
Moreover, we found suggestive evidence of association at rs6536942 on 4q32.3, located
within an intron of the tolloid-like 1 (TLL1) gene, with FL risk.
However, larger studies are needed to validate this finding. Although the strongest
associations so far have been observed in the HLA region, and extended pooling of
available scan data failed to identify additional loci outside of HLA, we expect
that future larger meta-GWAS efforts will more robustly identify additional loci in
other regions.
In conclusion, our results strongly suggest that future genetic and functional work
focused on the HLA class II region will provide important insight into the disease
pathology of FL, DLBCL and other subtypes of NHL. In addition, further studies of
this region and potential interaction with environmental factors in NHL risk, and of
NHL prognosis are warranted.
The studies described in this manuscript have been approved by the ethics
committee of the respective institutions: Karolinska Institutet (Sweden),
Scientific Ethics Committee system (Denmark), University of California, Berkeley
(US), National Cancer Institute, National Institutes of Health (US), Mayo Clinic
(US), University of British Columbia (Canada), Yale University (US), University
of Sydney (Australia).
The SCALE study is a population-based study of the etiology of NHL carried out in
all of Denmark and Sweden during 1999 to 2002 [26]. NHL subtype diagnoses were
reviewed and reclassified according to the World Health Organization (WHO)
classification [1] as previously described [26]. For this GWAS (SCALE1) we
used DNA from 400 cases with follicular lymphoma (FL; 150 from Denmark and 250
from Sweden) and from 150 Danish controls, individually matched to the Danish FL
cases by sex and age at study inclusion. We also used material collected from
673 control subjects in a separate Swedish population-based case-control study
of rheumatoid arthritis (the Eira study) [21], [27]. The latter was conducted
during 1996 to 2005 among residents 18 to 70 years of age in the southern and
central parts of Sweden (including 90% of Swedish residents). Hence, the
population controls recruited in this study were considered to represent the
same study population as the Swedish component of the SCALE study with regard to
genetic variation. Genotyping completion rates were similar between cases and
controls; out of 400 cases and 823 controls genotyped, 379 cases (95%)
and 791 controls (96%) were included in the final analysis. Study
subjects used in Stages 2, 3 and validation in other NHL subtypes (Table 1, Table S1,
S2)
have been previously described [6], [8]–[12], and details are
available as supporting text (Text S1). For the SCALE2 NHL subtype
validation study, we used the rest of the lymphoma cases with blood samples
originally recruited in SCALE (n = 1869), Danish control
subjects not included in the GWAS (n = 556), a second set
of control subjects from the Eira study (n = 742) and a
third group of controls recruited in a national population-based case-control
study of breast cancer, the Cancer and Hormones Replacement in Sweden (CAHRES)
study [28]
(n = 720). The control subjects from this study were
randomly selected from the Swedish general population to match the expected age
distribution of the participating breast cancer cases (50 to 74 years).
Stage I genotyping of 317,503 single nucleotide polymorphisms (SNPs) was done on
the HumanHap300 (version 1.0) array. Validation genotyping was done using
Sequenom iPlex; SNPs in the human leukocyte antigen (HLA) region that failed
primer design for Sequenom assays were genotyped using Taqman (Applied
Biosystems).
The scan included 317,503 SNPs from the HumanHap300 (version 1.0) array. The
datasets were filtered on the basis of SNP genotyping call rates
(≥>95% completeness), sample completion rate (≥90%),
minor allele frequency (MAF; all subjects as well as cases and controls
separately ≥0.03) and non-deviation from Hardy-Weinberg equilibrium (HWE;
p<10−6). We also excluded SNPs with cluster plot
problems, and those on the X and Y chromosomes. Study subjects with gender
discrepancies and/or labelling errors were removed. We also removed individual
samples with evidence of cryptic family relationships (identified using
the–genome command in PLINK). To detect outliers in terms of population
stratification, we performed principal component (PC) analysis using the
EIGENSTRAT software (Figure S4). A subset of linkage
disequilibrium (LD) thinned SNPs was selected such that all pair-wise
associations had r2<0.2, and long-range regions of high LD,
reported to potentially confound genome scans, were removed [29]. Twenty-five
samples were removed as population outliers on the basis of their values on the
first three PCs. To adjust for possible stratification in our association
analyses we adjusted the regression analyses using the first three PCs; the
number of PCs used for adjustment was determined by plotting the eigenvalues and
locating the position of the “elbow” on the scree plot (Figure S5).
Wald tests, treating minor allele counts as continuous covariates were used to
test for association. The genomic inflation factor (λ) was calculated to be
1.0283 after adjusting for the first three PCs, suggesting the presence of
minimal stratification. Quantile-quantile plots for the associations before and
after adjustment are shown in Figure S6. Finally, we assessed associations
of age and sex with main genotypes among the control subjects to address the
possibility of confounding by these factors (Table S11).
As there was no evidence of associations of age or sex with genotypes among the
controls, we did not adjust for them in the final main effects analyses of
genotypes.
In Stage 2, similar quality control measures were applied as in Stage 1,
including genotyping call rate ≥95%, sample completion rate
≥90%, and MAF ≥0.05. We tested each validation study for
association using trend tests. For meta-analyses across studies and NHL
subtypes, we used the Cochran-Mantel-Haenszel method to calculate the combined
odds ratio and P-value, and χ2 tests for heterogeneity.
Multivariate logistic regression was used to test for independence of SNP
effects. For validation among other NHL subtypes, the control subjects were the
same as those in Stages 2 and 3 for validation in FL for all studies except
SCALE2. Only European-ancestry subjects were included, and the possibility of
population stratification affecting the results has been thoroughly explored and
found to be low in earlier investigations in the same populations [6], [8].
We used IMPUTEv1 for the imputation of SNPs from the 1000 Genomes pilot1 CEU data
(August 2009 release); and the HapMap Phase II release 22 CEU data. We set a
strict threshold for imputation, using only SNPs with confidence scores of
≥0.9, call rates ≥90%, non-deviation from Hardy-Weinberg
equilibrium P >0.001 and MAF >0.01. The imputation was done on the Stage 1
samples separately for each of the two reference datasets and SNPs showing a
discordance of >5% between the genotypes imputed with the two datasets
were excluded from further analysis. The data were then merged using HapMap II
as the master dataset to which additional imputed SNPs from the 1000 Genomes
dataset were added. HLA alleles were imputed by identifying tag SNPs [15] from the
genotyped and imputed SNP dataset. We used PLINK for haplotype imputation with
the tag SNPs and downstream association analyses. Only haplotypes with call
rates >90%, MAF>1% and probability thresholds >0.8 were
analyzed.
For coalescence analysis all 12 SNPs (genotyped in this study and within a region
of ∼177 Kb) adjacent to the two SNPs associated with the FL risk were used
to construct haplotypes. These were phased using the PHASE program [30] and
tested for association using PLINK. The ancestral haplotype was constructed from
the chimpanzee (PanTro2) allele whenever possible, and otherwise from the
macaque alleles. An ancestral recombination graph was constructed using the
program Beagle [13], [31] which allows recombination assuming an infinite site
mutation model. After identifying the first recombination event the haplotype
segment before the recombination spot was used to construct a median
–joining network using the Network program [14]. The alleles of the
imputed SNP rs9378212 were then phased on each haplotype segment using the PHASE
program.
The URLs for the data and analytic approaches presented herein are as
follows:
1000 Genomes http://1000genomes.org
HapMapII http://www.hapmap.org
IMPUTEv1 https://mathgen.stats.ox.ac.uk/impute/impute_v1.html
mRNA by SNP browser http://www.sph.umich.edu/csg/liang/asthma/
R script for recombination plot http://www.broadinstitute.org/science/projects/diabetes-genetics-initiative/plotting-genome-wide-association-results
|
10.1371/journal.pbio.0050150 | Setting the Tempo in Development: An Investigation of the Zebrafish Somite Clock Mechanism | The somites of the vertebrate embryo are clocked out sequentially from the presomitic mesoderm (PSM) at the tail end of the embryo. Formation of each somite corresponds to one cycle of oscillation of the somite segmentation clock—a system of genes whose expression switches on and off periodically in the cells of the PSM. We have previously proposed a simple mathematical model explaining how the oscillations, in zebrafish at least, may be generated by a delayed negative feedback loop in which the products of two Notch target genes, her1 and her7, directly inhibit their own transcription, as well as that of the gene for the Notch ligand DeltaC; Notch signalling via DeltaC keeps the oscillations of neighbouring cells in synchrony. Here we subject the model to quantitative tests. We show how to read temporal information from the spatial pattern of stripes of gene expression in the anterior PSM and in this way obtain values for the biosynthetic delays and molecular lifetimes on which the model critically depends. Using transgenic lines of zebrafish expressing her1 or her7 under heat-shock control, we confirm the regulatory relationships postulated by the model. From the timing of somite segmentation disturbances following a pulse of her7 misexpression, we deduce that although her7 continues to oscillate in the anterior half of the PSM, it governs the future somite segmentation behaviour of the cells only while they are in the posterior half. In general, the findings strongly support the mathematical model of how the somite clock works, but they do not exclude the possibility that other oscillator mechanisms may operate upstream from the her7/her1 oscillator or in parallel with it.
| Somites—the embryonic segments of the vertebrate body—are formed sequentially, with a spacing determined by a gene expression oscillator, the segmentation clock, operating in the cells at the tail end of the embryo. This system provides a rare opportunity to analyse how the timing of at least one set of developmental events is controlled. We previously proposed a mathematical model, showing how the oscillations could be generated by a delayed negative feedback loop, in which the products of two genes, her1 and her7, act as inhibitors of their own expression, and how Notch signalling between adjacent cells keeps their individual oscillations synchronised. Here we test and find support for this model in two ways. First, we show how to use the spatial pattern of gene expression to measure some of the temporal delays and molecular lifetimes that are critical for the occurrence of synchronised oscillations. Second, we use transgenic fish in which expression of her1 or her7 can be switched on at will by heat shock to probe the dynamics of the system and to analyse the logic of the control circuitry.
| How do genes set the tempo of embryonic development? This basic question is still largely unanswered. We lack quantitative information about the dynamics of gene expression in the embryo, and in most cases, we do not even know which genes govern timing, let alone how they do so. In this paper, we focus on one particular process, zebrafish somite formation, in which a better understanding may be attainable: previous work has identified specific genes as critical for the control of timing, and a detailed mathematical model has been proposed to explain how they could act as a timer. Our goal is to test this model through measurements of the dynamics of the real system.
The somites—the future segments of the vertebrate body axis—are laid down sequentially by a mechanism involving oscillating gene expression in the cells of the presomitic mesoderm (PSM) (see Figure 1A). Here, at the tail end of the embryo, the transcripts of certain genes undergo regular coordinated cycles of production and degradation [1–9]. This gene expression oscillator is called the somite segmentation clock, and in each of its cycles, one additional somite is formed. The cycling is rapid: in the zebrafish at 28 °C, the period is 30 min. A gradient of FGF8, with its high point in the tail bud where the Fgf8 gene is transcribed, and acting in opposition to retinoic acid released by anterior tissues [10–14], appears to define the extent of the PSM and thus the region within which cycling continues. As the cells in the PSM proliferate and the tail bud grows caudally, the maturation “wavefront” delimiting the region of high FGF8 concentration moves caudally also, causing the cells in the anterior region of the PSM to slow down and finally cease their oscillations as they emerge from the PSM and begin differentiation. The gradual slowing of the oscillation in the anterior part of the PSM is manifest there in a spatial pattern of travelling stripes of gene expression. Cells that are in different phases of the oscillation cycle as the maturation wavefront sweeps over them become arrested in different states, corresponding to expression of different genes and apparently defining which portion of a somite they will form. The succession of cells passing through the anterior part of the PSM to begin their differentiation as somites can thus be likened to magnetic tape passing the recording head in a tape recorder: the periodic somite pattern represents a spatial record of the temporal oscillation in the posterior PSM. It is this spatiotemporal map that makes the system particularly attractive for investigation of developmental timing.
What, then, is the mechanism of the intracellular oscillation, and how are the oscillations of adjacent cells coordinated? All of the mutations known to specifically disrupt the PSM oscillations in zebrafish lie in components of the Notch signalling pathway [15–18] (although in mammals and birds, there is evidence that the Wnt/beta-catenin and other pathways are also involved [1,3,4,9]). Previous work in the zebrafish has shown that a closely related pair of Notch target genes, her1 and her7, are strong candidates for a central role, along with the Notch ligand gene deltaC. her7, deltaC, and perhaps her1 are required for coordinated oscillation of other markers and for regular somite segmentation, and all three show oscillating expression in the PSM in synchrony with one another [16,19]. her1 and her7 both belong to the Hairy/E(spl) family of putative transcriptional repressor genes, several of which are regulated by, and regulators of, Notch signalling [16,19–22]. Notch signalling thus provides a communication mechanism to synchronize the her1/her7 oscillations of adjacent cells [6,23–25]. Moreover, her1 and her7 are thought to negatively regulate their own expression [19]. This suggests that a negative feedback loop based on direct autorepression of her1 and/or her7 could be the fundamental pacemaker mechanism of the somitogenesis oscillator (Figure 1B). But is this really the case, and if so, how is the period of oscillation specified? To answer these questions, intuition is not enough: one has to work out the underlying mathematics and analyse the system quantitatively.
We have shown by mathematical modelling [25,26] that the possibility of oscillation in such a simple negative feedback system depends critically on the delays involved in transcription and translation—that is, the time Tm that elapses from initiation of a transcript to its emergence into the cytosol as a mature mRNA, and the corresponding time Tp from initiation of translation to delivery of the functional protein molecule to its site of action. When these delays are taken into account, theory predicts that autoinhibition of her1 or her7 (or of both together) will give rise to robust oscillations, provided certain conditions are satisfied: in particular, the lifetimes of the Her1/Her7 protein and mRNA molecules must both be short compared with the sum of delays Tm + Tp. The predicted period of oscillation T, if oscillation occurs, is then given by the simple formula:
where τm and τp are the lifetimes of the mRNA and protein, respectively. This analysis explains the pattern of somite defects in mice in which the lifetime of the Her1/Her7 homolog Hes7 has been artificially lengthened [27].
The theory predicts, furthermore, that the oscillations in adjacent cells will be kept synchronized through communication via the Delta-Notch pathway (Figure 1C), but again only if certain conditions are met. Experimental evidence indicates that the Her protein regulates expression of the deltaC gene in parallel with that of the her gene, and that activation of Notch by DeltaC contributes to regulation of her gene expression [15,19]. In the corresponding mathematical model, synchronization then requires a relatively large delay, of the order of one oscillator cycle, in the Delta-Notch signalling pathway—a delay that includes the translational plus transcriptional delay for DeltaC. This delay is indeed expected to be long because of the time needed for delivery of DeltaC to the plasma membrane via the secretory pathway.
Taken all together, the experimental evidence suggests a simple mathematical model of how specific known genes may act as pacemakers and synchronizers of the somite segmentation clock. This model oscillator system will only work, and will only fit the observed oscillation period, if certain parameters have appropriate values. Previously, we inserted rough estimates, based on rather scanty data available from other systems, and argued that the theory was quantitatively, as well as qualitatively, plausible. Nevertheless, its validity rests on several unproven conjectures. These concern both the detailed logic of the regulatory interactions between her1, her7, and deltaC, and the actual values of the biosynthetic delays, molecular lifetimes, and other parameters for the cells in the zebrafish PSM.
We have therefore set out to check these conjectures experimentally, through artificial manipulation of the components and through measurement. To probe the dynamics of such rapid transcriptional oscillations, one needs a rapid way to switch expression of the individual genes on or off at will. For this purpose, we have generated transgenic lines of zebrafish containing an inducible cassette of either her1 or her7 under the control of a heat-shock–responsive promoter. We have used these and other methods to examine the following questions:
Do the oscillating genes her1, her7, and deltaC show transcriptional and translational delays of the magnitudes postulated by the theory? Do the gene products—the mRNAs and the proteins—have the postulated short half-lives? What is the regulatory circuitry linking the oscillating genes? Do the oscillations recover after resetting by overexpression of her1 or her7, and if so, how rapidly? Finally, when her1 or her7 is transiently overexpressed, what is the time course of resulting somite segmentation defects? In other words, how does expression of these oscillating genes affect somite segmentation, and at what point in the history of a given somite do they act?
The spatial wave pattern seen in a fixed specimen can be used to derive information about the temporal oscillations that generated that pattern in the living tissue. For this purpose, we need first of all to understand quantitatively how the spatial and temporal patterns are related.
At the posterior end of the PSM, oscillations of gene expression occur at a frequency ω0 of one cycle per 30 min, corresponding to the time taken to form one additional somite. At the anterior end of the PSM, oscillation is arrested. In between, the oscillation slows down gradually. This means that anterior PSM cells are retarded in oscillator phase relative to posterior PSM cells; the retardation is greater the further anterior the cells lie. As a result, one sees the different phases of the oscillator cycle mapped out in space along the length of the PSM, where they are manifest as stripes of expression of the oscillating genes [15,23,28,29]; the number of stripes reflects the number of cycles by which cells at the anterior end of the PSM are delayed relative to the cells at the posterior end. We can exploit this spatial graph of the oscillation cycle to make measurements of the timing of oscillator events. To do so, however, we need a precise statement of the relationship between position and oscillator phase along the length of the PSM. This depends on the rate at which the clock runs in the cells in the different positions, and on the way in which the cells move as the tail bud extends and additional somites form. For cells close to the midline, whose movement is oriented along the anteroposterior axis, the formula is as follows (see Materials and Methods for a derivation):
where ω(x) is the frequency of the oscillation (in cycles per unit time) in the cells at the given position x along the axis, T(x) = 1/ω(x) is their oscillation period, T0 = 1/ω0 is the somite cycle time (the period of the fundamental somite segmentation clock, i.e., 30 min for the zebrafish at 28 °C), u(x) is the velocity of the cells relative to the tail bud measured in somite lengths per somite cycle, S(x) is the local spatial wavelength (the distance from one peak to the next or one trough to the next) measured along the trajectory of the cells, and S0 is the length of a formed somite. As shown in Figure 2, we can use this formula to plot a graph of the oscillation period T(x) as a function of distance along the anteroposterior axis. From the graph, we see, for example, that halfway along the PSM, T(x) ~ 1.2 T0, while at a point anterior to this, three quarters of the way along the PSM, T(x) ~ 1.6 T0.
If two events of the oscillation cycle appear separated by a distance δx in a snapshot of the PSM, we can deduce that in the history of a given cell, as the cell moves through the PSM, the events are separated in phase by a fraction δx/S(x) of the local cycle, i.e., by a time interval
Similarly, if the concentration C of some molecule in the cells changes as a function of the cells' phase in the oscillation cycle, we can deduce its rate of change with time within a given cell (i.e., its material derivative, DC/Dt) from the rate of change from cell to cell with respect to position at a given instant:
We can use these relationships to deduce the timing and rate of processes in the intracellular cycle from snapshots of the spatial pattern of cells in different states in the PSM.
We used fluorescent in situ hybridisation (FISH) as described in Materials and Methods to analyse the expression of the cyclically expressed genes her1, her7, and deltaC (Figure 3). For each of these genes, two types of in situ hybridisation (ISH) signals were visible. Some cells showed diffuse labelling in their cytoplasm, corresponding to cytoplasmic mRNA molecules. Other cells showed intense dots in the nucleus, corresponding to transcripts in the course of synthesis; often such dots occurred as pairs, corresponding to the two gene copies present in G1 phase of the cell division cycle, but in some nuclei, only one dot was seen (either because other dots, though present, were not included in the optical section, or because of the stochastic nature of transcriptional control), and in some nuclei (presumably in G2 phase), there were three or four dots.
The bands of cells with nuclear dots lay slightly anterior to the bands of cells with cytoplasmic staining, and from this spatial shift, we can estimate the delay from the time when nascent transcripts first become detectable in the nucleus to the time when the mature transcripts first become detectable in the cytoplasm (Figure 4). This delay represents a lower bound to the transcriptional delay Tm as defined in the mathematical model: it leaves out of account the time Tinit that must elapse from disappearance of the inhibitory signal (free Her1/7 protein in the nucleus) to the appearance of detectable nascent transcripts. Tinit includes time required for bound inhibitory protein to dissociate from the DNA and time for the RNA polymerase to generate sufficient transcript to be recognised by the ISH probe. Adding Tinit (whose value we do not know) to the delays manifest in the ISH pattern, we arrive at the following estimates for the transcriptional delays: for her1, Tmher1 = 3.8 ± 1.0 min + Tinit-her1 (mean ± standard error of the mean [s.e.m.], n = 10); for her7, Tmher7 = 3.7 ± 1.4 min + Tinit-her7 (mean ± s.e.m., n = 7 ); and for deltaC, TmdeltaC = 8.4 ± 1.2 min + Tinit-deltaC (mean ± s.e.m., n = 7).
The numbers are probably more trustworthy for her1 and deltaC than for her7, for which the in situ staining intensity was rather weak. These values can be compared with those, based on completely different and much less-direct evidence, that we used previously in computing the behaviour of our mathematical model: there we assumed Tmher1 = 12 min, Tmher7 = 7.1 min, and TmdeltaC = 16 min (see [25], Figures 3 and 4). The new values would thus roughly agree with the old if Tinit were of the order of 3 to 8 min.
In a snapshot of the PSM, the steepness of the decline in the number of transcripts per cell as a function of distance going caudally from a peak of the transcript distribution gives a measure of the rate at which transcript concentrations decrease with time in a typical cell as it progresses past the peak of its oscillation cycle. The rate of transcript degradation in a cell must be at least as fast as the observed net rate of decrease in the number of transcripts it contains, and will be faster than this if new transcripts are being produced at the same time. Thus, by analysing the spatial pattern and using Equation 4, we can derive an upper bound to the transcript lifetime, τm. The relationship is as follows:
where m(x,t) is the concentration of transcripts in a cell at (x,t), and lifetime is defined in such a way that half-life = lifetime × ln2 ~ 0.69 × lifetime.
We can use our fixed ISH specimens to calculate this upper limit if we assume that the intensity of the ISH signal is proportional to the concentration of transcripts present; details of the analysis are given in Materials and Methods. From specimens stained with tyramide chemistry to give a FISH signal (as in Figure 3), we find τmher1 ≤ 6.6 ± 0.8 min (mean ± s.e.m., n = 12) ; τmher7 ≤ 8.1 ± 1.2 min (mean ± s.e.m., n = 8); and τmdeltaC ≤ 6.1 ± 0.6 min (mean ± s.e.m., n = 5).
Noise, background staining, and possible nonlinearity in the relationship between mRNA concentration and ISH signal all mean that these numbers are only rough estimates. As a partial check, we also made the same measurements from a set of specimens stained for her1 using NBT/BCIP chemistry (as in Figure 2). These yield τmher1 ≤ 5.6 ± 0.5 min (mean ± s.e.m., n = 12).
Given that the actual values of the lifetimes are likely to be shorter than the estimated upper bounds, the above measurements are consistent with the value of 4.3 min that we assigned to each of these lifetimes in computing the behaviour of our mathematical model.
Although cyclic transcription in the PSM has been well documented in zebrafish, frog, chick, and mouse [1–8], data concerning the dynamics of the corresponding protein products are extremely scarce. Only two or three instances of cyclic protein expression during somitogenesis have been described so far: in the chick PSM, Western blots against the Lunatic Fringe protein have been used to demonstrate its cyclic expression [30]; in the mouse PSM, immunohistochemistry with an anti-Hes7 antibody shows a stripy pattern, which has been taken as evidence for oscillations in Hes7 protein levels [31]; and the level of Notch activation has also been shown to oscillate in the mouse [32]. There is no published evidence as to whether levels of Delta protein oscillate; indeed, of the many studies reporting Delta1 expression during somitogenesis in chick and mouse, only one [33] has described it as oscillatory at the mRNA level. In the zebrafish, deltaC clearly shows oscillating transcription, and it is clear that DeltaC function is necessary for oscillator function [6,15,29]; but the question remains open whether oscillations of DeltaC protein levels are required, or indeed occur at all, in the zebrafish PSM. This is an important issue for our theoretical model, which supposes that oscillating levels of DeltaC protein provide the signals that keep neighbouring cells synchronized.
To address this question, we used a monoclonal antibody raised against the zebrafish DeltaC protein (zdc2; see Material and Methods). Immunostaining with this antibody revealed stripes of DeltaC protein in the anterior PSM, indicating that DeltaC protein levels do oscillate in this region at least (Figure 5). Levels of DeltaC protein in the posterior PSM were too low to detect confidently, and thus too low for us to see clear evidence of oscillations, but this should not be taken to imply that they were too low to be physiologically important or that they were actually non-oscillatory. Protein degradation in this region is rapid, levels of the mRNA are lower than in the anterior PSM, and we have found in other tissues that the amounts of DeltaC and DeltaD are frequently so small as to be barely detectable by immunofluorescence even where there is clear genetic evidence that they are functionally important ([34] and unpublished data). Presumably, a very small amount of Notch ligand is enough to activate Notch effectively. Although levels of DeltaC protein in the posterior PSM were thus too low for us to demonstrate oscillations clearly, in the anterior PSM there was no such difficulty. In this region, comparison of the DeltaC immunofluorescence pattern with the deltaC mRNA pattern allowed us to estimate the DeltaC translational delay (from the time of the beginning of synthesis of a DeltaC molecule to its arrival in intracellular vesicles, presumably via a journey to the cell surface [34–36]). As shown in Figure 5, the pattern of bands of DeltaC protein immunostaining is similar to the pattern of bands of deltaC mRNA, but shifted anteriorly by almost exactly one somite width. Because cells take one somite cycle time to move one somite width, this means that the cells that show the peak levels of mRNA (close to the anterior end of the PSM) display peak levels of the protein one somite cycle—i.e., 30 min—later.
In our mathematical model, we found that a DeltaC “translational delay” of the order of 20 min (on top of a transcriptional delay of 16 min) gave good synchronization of adjacent cells. This parameter in the model comprises the delay from initiation of DeltaC translation to delivery of mature DeltaC protein to its site of action at the surface of the signalling cell, plus the delay for signal transduction in the receiving cell (from activation of Notch at the cell surface to its arrival at its site of action in the nucleus). We do not know the size of the latter contribution, though it is likely to be small, because it involves only a protein cleavage event and transport from the cytoplasm into the cell nucleus. There is also some uncertainty as to the timing relationship between availability of DeltaC at the cell surface and presence of DeltaC in endocytic vesicles in the cytoplasm (which is where we chiefly detect the protein with our antibody) (discussed in [34] for DeltaD). With these provisos, the measured delay of 30 min from appearance of peak levels of deltaC mRNA to appearance of peak levels of DeltaC protein agrees reasonably well with the delay required in the mathematical model to enforce synchronisation of adjacent cells.
To test the dynamics of the control system and to check the regulatory interactions between its components, we generated stable fish lines carrying a transgene in which the zebrafish hsp70 heat-shock promoter was coupled to the cDNA sequence of either her1 or her7. In these fish, we could artificially trigger a pulse of expression of her1 or her7 and follow the consequences. The transgene was designed so that the resulting transcript would be similar to the native her1/7 mRNA, including the 3′ untranslated region (UTR), with the 5′ UTR being that of hsp70. We also inserted a sequence coding for a haemagglutinin (HA) tag in frame with the N-terminus of the Her protein, so that the transgene product would be easily detectable. By screening fish injected with the appropriate DNA constructs, we identified and isolated stable transgenic lines carrying either hsp70:HA-her1 or hsp70:HA-her7. At least two independent lines were isolated for each construct.
We first checked the expression of the transgenes following heat shock, using ISH and immunochemistry to detect their transcripts and the HA-tagged protein. Figure 6 shows observations of one of the hsp70:HA-her7 transgenic fish lines; the other lines behaved similarly. Without heat shock, the transgenic embryos showed (outside the PSM) only a barely detectable basal expression of transgenic her7, with expression of the endogenous her7 gene confined to the PSM in the normal way. After 7.5 min at 37 °C, her7 mRNA expression had clearly risen throughout the embryo; after 15 min, it was everywhere quite intense, matching the peak levels of endogenous her7 expression normally seen in the PSM; and after 20–30 min, it reached an even higher and maximal level, such that further stay at 37 °C produced no further increase. The exogenous Her7 protein was a little slower to appear, as judged by staining with anti-HA antibody (though the difference may be partly just a consequence of the different methods of detection): it was first detected after 20 min and reached a maximal level after about 40 min of heat shock.
As shown in Figure 6B and 6C for an hsp70:HA-her7 embryo heat shocked for 30 min, levels of transgene expression—both mRNA and protein—fell rapidly after the heat shock was ended and the embryo was left to recover at its normal temperature of 28 °C. The decline in mRNA levels appeared practically uniform over the whole embryo, implying that the transcripts have a similarly short lifetime in all tissues. Thus the normal pattern of variation of her7 transcript levels in space and time is most likely achieved through regulated transcription (as our model assumes) rather than regulated degradation.
Levels of the HA-tagged protein, though still high after 10 min of recovery, were much reduced after 30 min, and undetectably low after 1 h (Figure 6C). In principle, the decay rate could be estimated from measurements of the immunolabelling fluorescence intensity at different time points; in practice, this was difficult, because the estimate depends sensitively on the amount of stain to be counted as background, and because there was substantial variability from specimen to specimen. The results for the HA-tagged protein were consistent with a lifetime as short as proposed for native Her7 and Her1 in the original model (4.3 min), though they did not exclude a value two or three times longer (unpublished data). Findings for the hsp70:HA-her1 transgenics were similar (unpublished data).
Despite these uncertainties, it is clear that induction of the transgene is fully reversible, so that a half-hour heat shock gives rise to a brief pulse of expression of the tagged Her7 or Her1 protein, beginning within 20 min after the start of the heat shock and vanishing within an hour after the end of the shock.
We next examined how such a pulse of Her7 or Her1 overexpression would affect the expression of the endogenous oscillator genes (Figures 6, 7, and 8). Strikingly, induction of Her7 expression led within 40 min after the start of heat shock to the disappearance of all deltaC and her1 transcripts from the PSM, with the notable exception of its most-anterior boundary region, where a single narrow stripe of residual expression could usually be seen (Figure 6D). In addition, deltaC transcripts in the formed somites were unaffected. A longer heat shock (60 min) led to disappearance of her1 from the anterior PSM as well (see Figure 8, top row). Repression of these genes was never observed in non-transgenic heat-shocked embryos and cannot therefore be attributed to the effect of heat shock itself. These observations confirm that Her7 represses deltaC and her1 transcription in the PSM. The rapidity of the effect strongly suggests that the repression is direct, and not a secondary consequence of a change in the expression of some intervening gene. We can also infer that the repressive effect of Her7 is strong only in the posterior and intermediate PSM, becomes weak in the anterior PSM, and no longer operates in formed somites. This could be because the repressive activity of Her7 depends on some partner protein whose concentration falls off with distance from the tail end of the PSM; Her13.2 or a similar protein could be a candidate for this role [37].
We could not test so directly whether Her7 represses her7 transcription, because we did not have in situ probes that would distinguish between the closely similar transcripts of the native her7 and the hsp70:HA-her7 transgene. Nevertheless, using a 15-min heat shock, we observed that after 25 min of recovery at 28 °C, when heat-shock–induced transcripts from the transgene had largely disappeared, the overall level of her7 mRNA was uniformly low, and below normal endogenous levels in the PSM (Figure 6B). This implies that the endogenous her7 gene is down-regulated within 40 min after the onset of heat shock. Therefore, Her7 also represses its own expression in the PSM.
To test whether her7 overexpression might lead to a general inhibition of transcription in the PSM rather than a specific inhibition of cyclically expressed genes, we also examined the expression of deltaD, another member of the Delta gene family expressed in the PSM and required for oscillations, but not itself oscillatory, at least in the posterior PSM. Upon induction of her7 by a 40-min heat shock, deltaD transcript levels in the PSM were only slightly reduced (although with some blurring of the stripes in the anterior region); in contrast, expression of deltaD in proneural clusters in the neural tube was almost totally abolished (Figure 6D).
Using hsp70:HA-her1 fish, we tested whether the induction of her1 expression would have the same effect as for her7. The findings were similar, but not identical (Figure 7). Thus a 40-minute heat shock induced ubiquitous massive overexpression of Her1, and this reduced the PSM expression of deltaC and of her7; but it did not totally abolish the PSM expression of her7. Cyclic expression of her7 was still evident in these fish, at apparently normal frequency, but with reduced amplitude (Figure 7A and 7B). However, a longer heat shock, of 60 min, was enough to make expression of her7 undetectable (Figure 7C). Again, deltaC expression in formed somites was unaffected. Like her7, her1 appears to be inhibited by its own protein product, because the endogenous her1 expression pattern becomes invisible in the hsp70:HA-her1 fish in the aftermath of the heat shock (Figure 7D).
Several considerations suggest that her7 is more likely than her1 to be the key pacemaker of the transcriptional oscillations: (1) whereas morpholino knock-down of her1 expression disrupts formation of only the first few somite boundaries, knock-down of her7 disrupts all somite boundaries posterior to the first few [16,19,21,38]; and (2), as we have just shown, cyclic expression of her7 seems able to continue even in the face of strong forced overexpression of her1, whereas all expression of her1 seems to be lost in the face of similar forced overexpression of her7.
We thus predicted that a heat-shock–induced pulse of Her7 would reset the clock, and that disappearance of this exogenous Her7 would allow the expression cycle to start again in an altered phase. We therefore subjected batches of hsp70:HA-her7 embryos to a heat shock followed by an extended recovery period at 28 °C. By about 1 h after the end of heat shock, renewed expression of her7 in the PSM was already visible, but restoration of the normal pattern of stripes took much longer, of the order of 2 to 3 h (Figure 8).
Similar treatment of hsp70:HA-her1 embryos again gave similar, but not identical, results. Renewed expression of her1 mRNA was already visible 40 min after the end of heat shock, and then took more than 2 h to resolve into regular oscillations (unpublished data).
At least three factors may contribute to the slowness of recovery after heat shock. First, by transiently imposing a high level of Her7 or Her1 protein throughout the PSM, the treatment would be expected to erase the normal phase gradient responsible for the pattern of stripes; to re-establish this gradient, the anterior PSM must become populated with fresh cells that have the phase delay that results from following the normal trajectory from the posterior PSM, and this will require several hours. A second delaying factor may be the abnormal combination of regulatory molecules inside each cell at the start of the recovery period: concentrations of both mRNA and protein for her1, her7, and deltaC will all presumably be near zero, except for one member of the set of proteins—Her1 or Her7—which will be maximal. Such a condition never arises normally, and it may take some time to recover from this confused state. A third reason for the slowness of recovery lies in intercellular variation: neighbouring cells do not recover synchronously, presumably because of the variation in the heat-shock response from cell to cell and the essentially stochastic character of all gene regulation. Restoration of synchrony between desynchronised oscillators can be expected to be slow: a synchronising mechanism based on Delta-Notch signalling (or on any other short-range form of communication) will at first create microdomains within which cells are locally synchronised, but out of synchrony with the cells of the next microdomain, and this situation will then take a long time to resolve.
Whatever the detailed mechanics of recovery, it is clear that heat shock in the hsp70:HA-her7 transgenics rapidly and powerfully suppresses normal expression of her1, her7, and deltaC throughout almost the entire PSM. The normal pattern of somite segmentation is thought to be controlled by the cyclic expression of these genes. We therefore expected that the heat shock would disrupt segmentation of the next somites to emerge from the PSM. But we got a surprise.
Using hsp70:her1 and hsp70:her7 embryos, and heat shocking them at a variety of times between the zero- and 12-somite stages, we did indeed disrupt the pattern of segmentation; but the first four or five somites to form after the beginning of the heat shock were always normally spaced. The same delay from onset of heat shock to onset of somite disruption was seen in over 100 embryos, heat shocked for times ranging from 10 min to 90 min, whatever the stage at which the heat shock was begun.
The formation of four or five normal somites after the heat shock was followed by a disruption of segmentation in a region that varied in extent according to the duration of the heat shock and the age of the embryo when it was administered. Thus a 60-min heat shock starting at the five-somite stage (Figure 9) reproducibly caused a defect extending over five somite widths, from the level of somite 10 to that of somite 14 (n = 11 hsp70:her7 embryos scored), whereas the same heat shock starting at the 12-somite stage caused a defect extending over about three somite widths, from the level of somite 17 to that of somite 19 (n = 8); 60-min heat shocks at the intermediate stages (n = 14) gave intermediate effects. Shorter heat shocks caused defects that were less extensive and not always apparent; thus, out of 64 hsp70:her7 transgenic embryos heat shocked for 10, 20, or 30 min at the seven-somite stage, 27 showed defects, and these generally extended over only two somite widths, corresponding to the loss of one inter-somite boundary; but as always, the defects were delayed until four or five somites after the heat shock.
With all these different heat-shock regimes, somites that formed after the period of disruption appeared normal: the segmentation process eventually recovered. In sibling control embryos that we heat shocked, we saw no disturbance of segmentation, implying that the disrupted segmentation in the transgenic embryos was a specific effect of overexpression of Her1 or Her7. (Our findings here contrast with those of Roy et al. [39], who found that heat shock affected segmentation in wild-type embryos; but the segmentation defects they saw were much milder than in our transgenics, and their standard heat shock was more severe. In our hands, even a 40 °C 40-min heat shock did not cause any obvious segmentation defects in wild-type embryos, suggesting a difference between our wild-type strain and theirs.)
The five somites that form normally after a heat shock, before the segmentation defect is seen, consist of the cells that occupied the anterior half of the PSM at the time of the heat shock. (To be precise, according to our measurements at the five-somite to 12-somite stages, the distance from the posterior end of the notochord to the anterior boundary of the PSM equals 8.0 ± 1.3 [mean ± standard deviation (SD), n = 10] times the length of a formed somite; allowing for growth and cell movement, we calculate that the next five prospective somites occupy 46% of this region.) In the anterior PSM, evidently, the cells proceed with their segmentation programme regardless of whether her1, her7, and deltaC are oscillating or indeed expressed at all. But the occurrence of segmentation defects after the five-somite delay means that these genes do control the way the segmentation machinery of the anterior PSM is set going in the cells as they move into the anterior half of the PSM from the posterior half. And the eventual recovery of the segmentation pattern reflects the recovery of the oscillations in the posterior PSM.
Mathematical modelling has shown that a very simple delayed negative feedback loop, based on autorepression of her7 and/or her1, could account for the oscillations of gene expression that underlie somite segmentation. But the mathematics also shows that such a feedback loop can only generate oscillations if the kinetic parameters lie within appropriate ranges; if these conditions are satisfied, the period of oscillation should be primarily determined by the sum of the transcriptional and translational delays for the oscillator genes. In our original account of this model, we made rough estimates—or, in some cases, guesses—of the likely parameter values, mainly drawing on data from other gene expression systems. When we substituted these values into the model, it gave oscillations that matched the observations.
In this paper, we have shown how the beautiful spatiotemporal organization of the zebrafish somite segmentation system can be exploited so as to obtain direct measurements of several of the key parameters, including information about the delays. The measured values are close to the previous indirect estimates. We can substitute into the mathematical model (see the Supplementary Data in [25]) the new measured values of the transcriptional delays, the mRNA lifetimes (taking them to be equal to their measured upper bounds), and the translational delay for DeltaC, and re-compute the model's behaviour. To do this, we have to make an assumption about the value of Tinit, the time from disappearance of free inhibitory protein to the appearance of visible dots corresponding to nascent transcripts in the nucleus. If we assume Tinit = 0 min, we get only damped oscillations, with a period of approximately 29 min; if we assume Tinit = 3 min for each gene, we get correctly synchronized sustained oscillations with a period of 41 min, matching the period of 41 min computed using the old parameter estimates and reasonably close to the observed period of 30 min. Longer values of Tinit give synchronized sustained oscillations with a longer period. Shorter values of Tinit can give a similar outcome, but with a shorter oscillation period if other processes are also a little more rapid than the above estimates (but still within the margin of measurement error); for example, if we suppose that Tinit = 2 min, that the mRNA lifetimes are 20% (a minute or two) shorter than our measured upper bounds, and that the DeltaC translational delay is 25 min rather than 30 min, we get synchronized sustained oscillations with a period of 36 min. Values of a few minutes for Tinit are consistent with the evidence from other systems, as discussed in the previous paper [25].
There are, of course, some important parameters that remain to be measured directly, including the value of Tinit, the length of the translational delays for Her1 and Her7, and the protein lifetimes. Nevertheless, the measurements that we have been able to make significantly strengthen the case for believing that the mathematical model gives a true account of the oscillator mechanism of the real system. The transcriptional delays fit the proposal that autorepression of her7 (and/or her1) is responsible for the observed 30-min oscillation cycle. Likewise, the observed translational delay for DeltaC, of approximately 30 min, has the magnitude required if Delta-Notch signalling is to keep the oscillations in adjacent cells synchronized in the way that the model proposes.
The proposed oscillator mechanism depends not only on kinetic parameter values, but also on the logic of the gene control circuitry. Here, too, we made assumptions, and through our heat-shock experiments in transgenic fish, we have been able to check them. Previous work [19] had shown that reduced expression of her1 led to increased expression of her7, and vice versa, and that both her genes affected expression of deltaC; but the evidence that either of these her genes was autoinhibitory was inconclusive [38], and there was no proof that the inhibition was direct and not mediated through effects on intervening genes. Our experiments using heat shock to trigger a pulse of expression of her1 or her7 confirm all the postulated inhibitory actions, and the rapidity of the response to the heat shock strongly suggests that the inhibition is indeed direct in each case.
Lastly, our heat-shock experiments have allowed us to test the function of her1/7 oscillations in controlling the pattern of somite segmentation. In particular, we have been able to discover at what time in the history of a somite the products of these genes act. Even though her1 and her7 normally continue to be expressed in an oscillatory fashion in the anterior half of the PSM, they exert no influence in that region on the segmentation behaviour of the cells; their influence is exerted only at an earlier stage, while the cells are in the posterior half of the PSM or in the transition zone from posterior to anterior. The delay surprised us, but it fits well with previous observations showing that the tissue in the anterior part of the PSM is already determined with respect to various other tests. Its pattern of segmentation is unaffected by treatments that interfere with FGF signalling, including treatments that cause some disruption of her1 expression [10,12]; a block of this tissue maintains its pattern of segmentation even when rotated so as to reverse its anteroposterior axis relative to the rest of the embryo [10]; and in experiments on wild-type embryos in which heat shock was found to disrupt somite patterning, cells that lay in the anterior part of the PSM at the time of the heat shock were unaffected [39]. Our data, however, go beyond these previous observations in showing directly that the products of the oscillator genes her1 and her7 have no impact on the segmentation behaviour of cells in this region. Evidently, other patterning mechanisms, involving a variety of other dynamically expressed genes and cell–cell interactions [40–44], come into play in the anterior PSM. These in effect produce a delayed readout of the her1/7 clock cycle phase that the cells had on leaving the posterior half of the PSM. The question of how this system of readout genes is organised and influenced by the her1/7 clock is a subject for another paper (E. M. Özbudak and J Lewis, unpublished data).
Several of the key assumptions of our mathematical model have yet to be tested—in particular, assumptions as to the rates of transcription and translation, and the detailed functional dependence of the rate of transcription initiation on protein concentration. Moreover, measurements of the parameters on which the model depends are not enough by themselves to prove that the her1/7 oscillator is the fundamental pacemaker of the zebrafish somite segmentation clock: it remains possible, for example, that it could be only one of several loosely coupled oscillators operating in parallel [45].
In the mouse and the chick, other genes, not belonging to the Notch pathway, are indeed found to oscillate in the posterior PSM [1,3,4,9], along with homologs of the genes we have discussed here. Could it be, then, that the oscillations of her1 and her7 are directly driven by some other oscillator that operates independently of her1 and her7? The slow pattern of recovery of her1 and her7 oscillation that we have described after a heat-shock–induced pulse of her7 expression makes this unlikely. If her1 and her7 were merely slaves of another oscillator, their pattern of expression should recover promptly, under the command of the master oscillator, as soon as the artificial pulse of expression has disappeared; and we have seen that it does not do so, but takes about four times the normal somite cycle period to re-establish itself. The her1/7 oscillator may be subject to control by another oscillator; but if so, the relationship cannot be a simple one of master and obedient slave.
The findings of this paper all lend support to our model of the mechanism of the somite segmentation clock: the logic of the control circuitry, the magnitudes of the transcriptional and translational delays, and the lifetimes of the molecules, where we have been able to measure them, all appear to be as the model supposes. If the her7 or her1 autorepression loop is not the fundamental pacemaker of the observed oscillations, it seems probable that it is at least capable of generating oscillations of a similar character and tempo.
The example of the somite segmentation clock shows how it is possible to analyse and explain quantitatively how the timing of at least one developmental process is controlled. The timing mechanisms of other developmental processes await investigation.
The hsp70:HA-her1 and hsp70:HA-her7 DNA constructs consisted of 1.5 kilobases (kb) of the zebrafish hsp70 promoter and upstream regulatory regions [46], including the short 5′ UTR, cloned upstream of her1 (1.7 kb) or her7 (1 kb) full-length cDNAs starting at the initial ATG, and followed by the SV40 early polyadenylation signal sequence (220 base pairs). A stretch of 33 nucleotides (GCCTACCCTTACGACGTGCCTGACTACGCTAGC) was inserted after the initiation ATG, so as to provide an influenza HA tag (AYPYDVPDYA) at the amino terminus of the produced protein. For the hsp70:HA-her1 construct, an EF1alpha:GFP cassette [47] was added in reverse orientation to facilitate the identification of transgenic embryos by constitutive green fluorescent protein (GFP) expression.
The cassettes were cloned in pBluescript SK I-SceI [48] so as to be flanked by restriction sites for the homing endonuclease I-SceI. Approximately 2 nl of a solution of the resulting plasmids (15 μg/ml) were injected into freshly fertilised eggs together with I-SceI enzyme (250 μg/ml; New England Biolabs, http://www.neb.com) and 0.5% phenol red in 1× I-SceI digestion buffer (New England Biolabs).
Injected fish were raised to adulthood and screened for germline transmission, either by transgene-specific PCR on genomic DNA prepared from embryos that they spawned, or by GFP fluorescence in these progeny. Two independent lines were selected for each construct. For each line, adult transgenic fish were identified by PCR on DNA extracted from fin clips, and the transmission of the transgene was checked for Mendelian segregation (50% of the progeny of each fish inherited the transgene), thereby ensuring that integration had occurred only at a single site in the genome.
Embryos were kept at a temperature of 28 °C until the desired stage for heat shock. They were then transferred to pre-warmed E3 medium [49] in a 37 °C incubator for the desired length of time, then fixed immediately in ice-cold buffered 4% formaldehyde solution or returned to 28 °C for further development.
ISH was performed according to standard protocols. Digoxigenin-labelled RNA probes used were as previously described: her1 [22], her7 [16,19], deltaC [50], and deltaD [51]. For FISH, a peroxidase-conjugated anti-digoxigenin antibody was used (anti-DIG-POD, 1/50; Roche, http://www.roche.com), and peroxidase activity was detected using tyramide signal amplification with Alexa-488 (Molecular Probes, http://probes.invitrogen.com), FITC (PerkinElmer, http://www.perkinelmer.com), or Cy3 (Perkin Elmer) coupled to the tyramide, following manufacturer's instructions.
For dual FISH, the second RNA probe was labelled with fluorescein instead of digoxigenin, and detected using alkaline phosphatase-conjugated anti-fluorescein antibody (1/1,000; Roche) and Fast Red staining.
For co-staining of deltaC mRNA with DeltaC protein, we used a monoclonal antibody, zdc2, directed against the amino-terminal part of DeltaC (see below). Fixed embryos were quickly dehydrated and then rehydrated in a methanol series, then incubated with the zdc2 antibody (1/100) for 3 h, using 2 mg/ml bovine serum albumin as a blocking agent. Embryos were then incubated with a biotin-conjugated anti-mouse IgG antibody (1/200; Vector Laboratories, http://www.vectorlabs.com), before being washed and then refixed overnight in buffered 4% formaldehyde solution to stabilise the antigen–antibody complexes. Embryos were then processed for deltaC ISH as described above using Alexa-488–coupled tyramide to stain for anti-digoxigenin peroxidase activity, after which the peroxidase activity was destroyed by a 10-min incubation in 0.1 M glycine-HCl at pH 2.2. Biotin was then detected using biotin-streptavidin-HRP complexes (ABC kit; Vector Labs), with the HRP peroxidase activity being stained with Cy3-coupled tyramide (Perkin Elmer).
For HA tag immunodetection, we used a monoclonal rat anti-HA antibody (Roche 3F10, 1/500) in combination with Alexa-488–conjugated anti-Rat IgG secondary antibody (1/400; Molecular Probes). We used a standard immunohistochemistry protocol with short fixation in buffered 4% formaldehyde solution, permeabilisation with 1% Triton X-100, and 0.2% gelatine as a blocking agent.
All fluorescently stained specimens were counterstained with the far-red fluorescent nuclear marker TOPRO3 (Molecular Probes), flat mounted, and imaged as optical sections on a confocal microscope.
The extracellular region of zebrafish DeltaC was C-terminally fused to the rat CD4 tag and expressed as a soluble fusion protein by transient transfection of HEK293T cells. The protein was purified from tissue culture supernatant by immunoaffinity chromatography using the anti–CD4-tag monoclonal antibody OX-68 [52]. Hybridomas were generated by fusing splenocytes from immunized mice to the SP2/0 cell line. Hybridoma supernatants were screened by enzyme-linked immunosorbent assay (ELISA) to select antibodies that recognise epitopes that did not cross-react with other zebrafish Delta proteins and were also resistant to formalin fixation. Full details will be published elsewhere. The hybridoma was cloned, isotyped as a mouse IgG2a, and named zdc2.
Measurements of the geometry of the ISH patterns of her1, her7, and deltaC were made from confocal optical sections of flat-mounted embryos that were stained using tyramide signal amplification and a DNA counterstain as described above. First, images were warped using the Distort:Shear tool in Adobe Photoshop (Adobe Systems, http://www.adobe.com) to make the bands of gene expression in the anterior PSM run at right angles to the body axis. We applied the Threshold operation to the nuclear signal so that each part of the image was classified as nucleus or not nucleus. Using the Image:Calculations:Multiply tool in Photoshop, we then derived a pair of images, one showing just the ISH signal that lay within nuclei (corresponding to nascent transcripts), the other showing just the signal that did not lie within nuclei (corresponding to cytoplasmic mRNA). A graph of the mean signal intensity in the PSM for each image as a function of distance along the anteroposterior axis was obtained using the Analyze:Plot Profile tool of ImageJ. By comparing the two graphs, we determined the spatial interval from onset of the nuclear signal to onset of the cytoplasmic signal, for selected stripes in the anterior PSM (see Figure 4). We converted this to a time interval—an estimate of the transcriptional delay—using Equation 3 (see Results).
For estimation of mRNA lifetimes, we used the graph of the spatial distribution of cytoplasmic transcripts (derived as just described), smoothed the numerical data in Mathematica, subtracted the background (which we took to be the signal intensity in the minima of the graph), and from the smoothed data, measured (1/c)(dc/dx) in the region of steepest descent. We converted this to a lifetime as explained in Results. As a check on possible errors from nonlinearity of the staining, we performed a similar analysis on a set of ISH specimens stained with NBT/BCIP, but in this case, without discriminating between nuclear and cytoplasmic signals.
Once a steady state has been reached, in which somites are formed at a steady rate through a steady production process in the PSM, we can derive a simple relationship between the observed temporal and spatial oscillations of gene expression.
Let ϕ(x, t) denote the phase of the oscillation cycle for a cell at position x at time t. If v is the velocity of the cell, its phase at time t + dt will be ϕ(x + vdt, t + dt). Thus the rate of change of phase in this cell as it moves along its trajectory (the material derivative of ϕ) is
If we measure ϕ in cycles,
is simply the intracellular oscillation frequency in cycles per unit time; in other words,
where T is the current value of the period of oscillation in the given cell. We assume that the rate of cycling depends only on the position of the cell relative to the tail bud (as will be the case if, for example, the rate depends only on the concentration of FGF8). We choose our origin of coordinates to be in the tail bud. Then we can write
Cells reaching x at different times but having followed the same flowline since leaving the posterior PSM will differ in phase by an amount that simply reflects the difference in their time of exit from the posterior PSM. If cells in the posterior PSM all oscillate with period T0, it follows that
so that the spatial pattern in the PSM as a whole (posterior plus anterior) oscillates with period T0 (a snapshot of the PSM at time t looks the same as a snapshot at time t + T0, after one additional somite has emerged from the anterior end of the PSM). T0, the period of the fundamental oscillator in the posterior PSM, is thus equal to the time taken to form one extra somite, while the spatial stripes seen in the anterior PSM reflect the slowing of the oscillation in each cell as it moves out along its flowline. We can write
where ∂/∂xv denotes differentiation with respect to position along the flowline (parallel to v), and S(x) is the period of oscillation of the spatial pattern along this line. In a linear approximation, S(x) is simply the distance from peak to peak or trough to trough in the neighbourhood of x.
Putting all this together, we have
At the anterior end of the PSM, the velocity of the cells relative to the tail bud is just one somite length per somite cycle time, directed along the rostrocaudal axis, so that v = S0/T0, and in general, v = u S0/T0, where u(x) is the velocity measured in somite lengths per somite cycle time. Hence we find
Equivalently, we can write
where ω(x) = 1/T(x) is the frequency of the oscillation in a cell at x (in cycles per unit time) and ω0 = 1/T0 .
Note that at the anterior end of the PSM, u(x) = 1 and S(x) = S0, so that T(x) → ∞, reflecting the fact that the temporal oscillation has stopped, while at the posterior end of the PSM, in the neighbourhood of the tail bud, where u(x) = 0 and S(x) is large, T(x) → T0. For the estimates shown in the main text, which are based on measurements close to the central body axis, we make a linear interpolation for the value of u(x) as a function of rostrocaudal position in the intervening region: u(x) = x/L, where x is the distance along the rostrocaudal axis measured from the tail end of the notochord, and L is the length of the PSM measured from the tail end of the notochord to the most recently formed somite boundary. The results shown in the main text are in fact not very sensitive to the exact form of u(x).
We use these formulas to read the temporal course of events from the spatial pattern as seen in the anterior PSM, where temporal cycling is still in progress (though slowing down), and the peaks and troughs of the spatial pattern are clearly defined.
|
10.1371/journal.ppat.1000116 | Histidine-Rich Glycoprotein Protects from Systemic Candida Infection | Fungi, such as Candida spp., are commonly found on the skin and at mucosal surfaces. Yet, they rarely cause invasive infections in immunocompetent individuals, an observation reflecting the ability of our innate immune system to control potentially invasive microbes found at biological boundaries. Antimicrobial proteins and peptides are becoming increasingly recognized as important effectors of innate immunity. This is illustrated further by the present investigation, demonstrating a novel antifungal role of histidine-rich glycoprotein (HRG), an abundant and multimodular plasma protein. HRG bound to Candida cells, and induced breaks in the cell walls of the organisms. Correspondingly, HRG preferentially lysed ergosterol-containing liposomes but not cholesterol-containing ones, indicating a specificity for fungal versus other types of eukaryotic membranes. Both antifungal and membrane-rupturing activities of HRG were enhanced at low pH, and mapped to the histidine-rich region of the protein. Ex vivo, HRG-containing plasma as well as fibrin clots exerted antifungal effects. In vivo, Hrg−/− mice were susceptible to infection by C. albicans, in contrast to wild-type mice, which were highly resistant to infection. The results demonstrate a key and previously unknown antifungal role of HRG in innate immunity.
| It has been estimated that humans contain about 1 kg of microbes, an observation that reflects our coexistence with colonizing microbes such as bacteria and fungi. The fungal species Candida is present as a commensal at mucosal surfaces and on skin. Although it may cause life-threatening infections, such as sepsis, particularly in immunocompromised individuals, it seldom causes disease in normal individuals. In order to control our microbial flora, humans as well as virtually all life forms are armoured with various proteins and peptides that comprise integral parts of our innate immune system. Here we describe a new component in this system; histidine-rich glycoprotein (HRG), an abundant plasma protein. We show, using a combination of microbiological, biochemical, and biophysical methods, that HRG exerts a potent antifungal activity, which is mediated via a histidine-rich region of the protein, and targets ergosterol-rich membrane structures such as those of Candida. HRG killed Candida both in plasma as well as when incorporated into fibrin clots. In mouse infection models, HRG was protective against systemic infection by Candida, indicating a novel antifungal role of HRG in innate immunity.
| The innate immune system, based on antimicrobial peptides (AMP) and proteins, provides a first line of defence against invading microbes [1]–[3]. At present, over 880 different AMPs have been identified in eukaryotes (www.bbcm.univ.trieste.it/tossi/pag5.htm). During recent years it has become increasingly evident that many AMPs, such as defensins and cathelicidins, are multifunctional, also mediating chemotaxis, apoptosis, and angiogenesis [4]–[6]. Conversely, molecules previously not considered as AMPs, including proinflammatory and chemotactic chemokines [7], neuropeptides [8], peptide hormones [9],[10], the anaphylatoxin peptide C3a [11],[12], growth factors [13] and kininogen-derived peptides [14]–[17] have recently been found to exert antibacterial activities.
Histidine-rich glycoprotein (HRG) is a plasma protein which was first isolated in 1972 by Heimburger et al. [18],[19]. The protein is present in human plasma at 1.5–2 µM, but the local concentration when HRG is released from activated platelets is likely to be higher [20]–[22]. It is a type 3 cystatin family protein [23], along with α-2-HS-glycoprotein/fetuin-A, fetuin-B and kininogen, and is found in vertebrates as well as in some invertebrates. The structure contains two cystatin-like domains, a central histidine-rich region (HRR) with highly conserved GHHPH tandem repeats flanked by proline-rich regions, and a C-terminal region [20]. This modular structure of HRG facilitates multiple interactions, involving ligands such as heparin, plasminogen, fibrinogen, thrombospondin, heme, IgG, FcγR, and C1q. Due to its high content of histidine residues (∼13%), which are concentrated to the HRR, HRG can acquire a positive net charge either by incorporation of Zn2+, or by protonation of histidine residues at acidic conditions [20]. In this context it has been proposed that HRG acts as a pH and Zn2+ sensor, providing a mechanism for regulating the various activities of HRG [24]. HRG has recently been ascribed antiangiogenic [25] effects in vitro, as well as antitumor [26] effects in vivo. Recent studies on Hrg−/− mice furthermore suggest that HRG plays a role as both an anticoagulant and an antifibrinolytic modifier, and may regulate platelet function in vivo [22].
Previous work has also demonstrated that HRG exert direct antibacterial activities in vitro which are dependent on Zn2+and pH [27]. However, as many cationic proteins and peptide sequences display antimicrobial properties in vitro, the ultimate role(s) of HRG in innate immunity in vivo still remained unresolved. During the course of our studies, we observed that HRG had a significant activity against Candida. Candida, an eukaryote, is present as a commensal at mucosal surfaces and on skin. Although it may cause life-threatening sepsis in immunocompromised individuals it seldom causes invasive disease in immunologically normal individuals [28]. We therefore speculated that HRG could constitute a natural defence against Candida infections. In the present study we show, using a combination of microbiological, biochemical, and biophysical methods, that HRG exerts a potent antifungal activity particularly at low pH, which is mediated via its HRR, and targets ergosterol-rich membrane structures such as those of Candida. In mouse infection models, HRG protects against systemic infection by Candida, indicating a previously undisclosed antifungal role of HRG in innate immunity.
In order to assess possible antifungal effects of HRG, we tested the activity of the protein against various Candida isolates. HRG was shown to be antifungal against C. parapsilosis at normal pH (10 mM Tris, pH 7.4), and the activity was significantly increased in low pH buffer (10 mM MES, pH 5.5) (Figure 1A). It is well-known that activities of AMPs and antimicrobial proteins are dependent of the microenvironment. For example, various chemokines, defensins, LL-37 as well as heparin binding protein are partly, or completely, antagonized by high salt conditions or the presence of plasma proteins in vitro [27],[29],[30]. Therefore, the influence of salt was tested. The results showed that HRG partially retained antifungal activity at physiological Cl− levels (0.1 M) but only at low pH (Figure 1B). The antifungal activity against C. parapsilosis was both time- and dose-dependent (Figure 1C). In subsequent experiments various Candida strains (C. parapsilosis, C. albicans, C. glabrata and C. krusei) were incubated with HRG (at 3 µM) at neutral as well as low pH. Figure 1D demonstrates, in line with the above experiments, that HRG is particularly active at low pH. Thus, C. parapsilosis, C. albicans and C. krusei were all nearly completely killed by HRG at low pH, whereas C. glabrata exhibited a partial resistance at this concentration of HRG, the latter in analogy to C. glabrata displaying some resistance against histatin 5 [31]. Next, to investigate the binding of HRG to fungi, C. papapsilosis was incubated with HRG at low pH, washed, and analysed by immunoblotting. Since previous results indicated that the HRR of HRG, which binds heparin/heparan sulfate, mediates antibacterial effects [27], heparin was added for competition of binding to Candida. Figure 1E shows that HRG was able to bind to the fungal cells and that the binding was partially inhibited by an excess of heparin. This finding is compatible with the observation that heparin completely blocks the antifungal effect of HRG (Figure S1). As demonstrated by flow cytometry, HRG bound to C. parapsilosis at neutral pH, and the binding was significantly increased at pH 5.5 (Figure 1F), results compatible with the fungal killing assays (Figure 1A). In summary, therefore, the results demonstrate that the antifungal actions of HRG were pH-dependent and likely mediated via the heparin-binding region of the protein.
Many AMPs kill microbes by membrane lysis, while others may translocate through membranes and subsequently interact with intracellular targets, such as DNA and mitochondria, all eventually resulting in microbial killing [32],[33]. Considering the antifungal effects and the binding to Candida cells, it was of interest to further study the possible mode of action for HRG on Candida. Electron microscopy demonstrated that HRG caused membrane breaks in Candida cells and release of cytoplasmic components (Figure 2A), effects particularly noted at low pH, where significant extracellular material was detected. The effects were similar to those observed after treatment with the “classical” human AMP LL-37 (Figure 2A). These data suggest that HRG acts on fungal membranes, however they do not demonstrate the exact mechanistic events, as secondary metabolic effects on fungi also may trigger death and membrane destabilization. Therefore, the impermeant dye FITC was used to assess permeabilisation. The results showed that HRG indeed was able to permeabilise Candida membranes (Figure 2B). In line with previous antifungal and binding experiments (see Figure 1A and 1F), the permeabilisation was most apparent at low pH.
These results were further substantiated by the use of a liposome model to assess membrane permeabilisation. In correspondence with the effects of HRG on Candida, HRG caused liposome leakage. Compatible with the pH sensitivity observed for HRG, the molecule preferably disrupted ergosterol-containing liposomes at pH 6.0 when compared with pH 7.4 (Figure 2C, left panel). Notably, ergosterol-containing liposomes, mimicking fungal membranes, were more sensitive than cholesterol-containing ones, mimicking mammalian membranes (Figure 2C, right panel). These results are in agreement with numerous previous findings on the membrane-stabilizing effects of cholesterol [34], as well as the findings that ergosterol induce less membrane stability in phospholipids than cholesterol [35]. At lower pH, protonation of histidine groups (pKa for the isolated histidine group is approximately 6.5), effectively increases the net charge density of HRG, thus the observed effects are compatible with findings previously reported for histidine-containing consensus peptides and histidine-rich endogenous peptides [36],[37]. Also noteworthy is that HRG did not display any major conformational changes either at low pH, or in the presence of fungal mannan (Figure 2D) or ergosterol-containing phospholipid liposomes (not shown). Hence, large-scale conformational changes appear not to be critical for the antifungal action of HRG. Taken together, the combination of electron microscopy, FITC-studies, and liposome data demonstrates that HRG acts at least in part through membrane disruption, although it is possible that additional intracellular effects of HRG may also contribute to fungal death. It is also notable that the observed effects were most marked and consistent at low pH. At neutral pH, binding (Figure 1B), as well as permeabilization (Figure 2A and 2B) was less apparent and these observations reflected the diminished antifungal effects at pH 7.4 (Figure 1A and 1D).
In order to explore the structure-function relationships of epitopes of HRG, overlapping peptide sequences comprising 20mers (Figure 3A and Table S1) were synthesized and screened, at both neutral and acidic pH, for antifungal activities against C. parapsilosis as well as C. albicans. The experiments identified several antifungal regions. In particular peptides no. 20–24 and 26, spanning the HRR, displayed a significant antifungal activity against both Candida strains at low pH (Figure 3B). There was a clear correlation with net charge (at the respective pH) of the various peptide regions and their observed antifungal activity (Figure S2). Although intuitively apparent (Figure 3B), the analysis furthermore showed that peptides derived from the HRR were (with the exception of the K and R-rich peptide no. 27) characterized by an increase in net charge at low pH (Table S1 and Figure S2).
In order to further study the importance of the HRR we investigated the activity of recombinant HRG (rHRG) and a truncated version (rHRG1-240), lacking the HRR and C-terminal domain. In contrast to full-length rHRG, truncated rHRG (0.6 µM) displayed no activity at pH 5.5 against Candida (Figure 3C). Taken together, considering the well-known heparin binding capacity of HRR, its pH dependence, as well as the absence of antifungal activity of rHRG1-240, it was logical to focus on the HRR of HRG in the subsequent studies of antifungal activity.
The HRR contains 12 tandem repeats of five consensus sequences of amino acids, GHHPH [20], a motif highly conserved among various vertebrate species [27]. To examine the activity of this sequence motif further, a 20-mer peptide (GHHPH)4 [16],[27] was chosen for further studies. Similar to intact HRG, GHH20 was antifungal against C. parapsilosis and C. albicans, particularly at low pH (Figure 4A). As demonstrated by FACS analysis, Tetramethyl-6-Carboxyrhodamine (TAMRA)-labeled GHH20 peptide bound to C. parapsilosis, and in correspondence with the antifungal data, the binding was stronger at pH 5.5 when compared to neutral pH (Figure 4B). As illustrated by fluorescence microscopy, TAMRA-labeled GHH20 showed a significant binding to Candida at pH 5.5 (Figure 4C). As with the HRG holoprotein, heparin abolished the binding, reflecting the heparin-binding capacity of this region of the HRR [27]. Also in line with the above experiments on fungi, GHH20 preferably disrupted liposomes at pH 6.0, with no significant activity at pH 7.4 (Figure 4D). The GHH20 peptide caused liposome leakage within a few hundred seconds (not shown), which contrasted to the significantly slower HRG-induced liposome leakage (Figure 2B), likely a manifestation of the much higher molecular weight of the holoprotein. Again as with intact HRG, CD spectroscopy showed that GHH20 displayed no major conformational changes associated with the histidine protonation at low pH, nor on interaction with phospholipid liposomes or mannan (not shown). Taken together, the GHH20 peptide showed similar characteristics as the holoprotein HRG with respect to activity, binding, and membrane permeabilisation.
In order to investigate the functional relevance of the above in vitro activities, we first tested the role of HRG against fungi in relevant physiological “settings” ex vivo. Initial results showed that HRG was detected in blood fractions (plasma, serum) and in platelets, also in wound fluid from acute wounds, and chronic leg ulcers (Figure 5A). The latter wound type is characterized by unregulated and excessive proteinase activity leading to degradation of many plasma proteins [38],[39]. However, compared with plasma and serum HRG, the molecule was not fragmented in this chronic wound fluid fraction (Figure 5A). The protein was also detected in fibrin clots (Figure 5A) but not present in seminal plasma. It is of note that the molecule migrated aberrantly in the used gel systems; relative 55–60 kDa in 8% gels (Tris-Glycine) and 45–50 kDa in 16.5 gels (Tris-Tricine). Identical serum and plasma preparations of HRG were used in the two gel systems, and recombinant HRG showed the same anomalous migration (not shown). In addition to its presence in plasma and other biological fluids, HRG occurs at significant levels in, and binds avidly to, fibrin clots [40]. Coagulation was initiated in normal and HRG-deficient human plasma in the presence of FITC-labeled HRG (Figure 5B). FITC-labeled HRG bound to clots derived from HRG-deficient plasma, and notably, it appeared to be present at clot boundaries, suggesting that it may “coat” the clot surfaces. In clots from normal plasma, no staining was seen, indicative of an inhibition of binding of FITC-HRG by the excess of endogenous HRG (∼150 µg/ml). Clots, physiologically important “barriers”, formed during hemostasis and infection, could thus constitute a unique milieu with high levels of surface-immobilized HRG. Considering the above results we investigated whether the presence of HRG could reduce the growth of Candida in plasma. Firstly, the growth of C. parapsilosis was investigated in normal human plasma and in plasma depleted of HRG. The results showed that C. parapsilosis multiplied significantly faster in HRG-depleted human plasma (Figure 5C). Analogous results on fungal growth were observed using plasma from mice deficient in HRG (data not shown). It is of note that these results do not exclude the possibility that other antifungal mechanisms may be involved, such as those dependent of complement activation. Furthermore, although the total protein levels (as determined by the Bradford method) and contents (as assessed by SDS-PAGE on 8% gels, not shown) were the same in depleted plasma (51.0+/−1.2 g/l) when compared with control plasma (51.7+/−3.3 g/l), it cannot be excluded that additional changes of low abundance proteins, induced by passage over Ni-NTA agarose could affect Candida growth. Nevertheless, the observation that similar results were obtained with the mice plasmas points at HRG as the main factor responsible for the partial growth inhibition noted. Furthermore, as demonstrated in Figure 5D, fibrin clots derived from plasma of HRG deficient mice were significantly more prone to infection by C. parapsilosis than clots from wild-type mice, and similar results were obtained with human plasma depleted of HRG when compared with normal plasma (not shown). The observation that clots devoid of HRG showed detectable, although reduced, antifungal activity (Figure 5D) suggest the existence of other yet unidentified factors in clots also mediating fungal killing. Nevertheless, the results indicate that HRG contributes to antifungal activity under physiological conditions.
To investigate the role of HRG during Candida infection in vivo, we designed a mouse model of intraperitoneal infection with C. albicans. After infection, the body weight of the mice was followed for three days (Figure 6A). Hrg−/− mice showed a significantly increased weight loss at day 1 and 2 (p = 0.02) when compared with wild type mice, and the wild type mice regained their initial weight after three days. Blood samples were collected from the animals 2 days post infection, and the fungal load in blood was determined (Figure 6B). A significantly higher amount of Candida cells was detected in the blood of Hrg−/− mice when compared with wild type mice (p = 0.032), indicating that a systemic infection has developed in HRG-deficient mice. In a similar experiment, we determined the ability of the fungi to establish infection in target organs distant from the site of administration. The spleen and kidney were harvested 3 days after initiation of intraperitoneal infection and the fungal load was determined. The results showed significant differences between Hrg−/− mice and the wild type mice; one animal out of 10 in the control group showed fungal load in the spleens and kidneys compared with 8 out of 10 in the Hrg−/− group (p = 0.009) (Figure 6C). Histopathological examination of the kidney tissues from Hrg−/− mice showed dense neutrophil infiltrates and notably, Candida cells were visualised by PAS staining in the centre of these infiltrates (Figure 6D). These results show a striking protective role for HRG against invasive Candida infection in vivo.
The key findings in our study are the identification of an antifungal activity of HRG in vivo together with the characterization of possible epitopes of HRG mediating this effect, as well as mechanistic data on HRG targeting of Candida membranes. The results have implications for our understanding of novel antifungal properties of HRG, and demonstrate that HRG constitutes a previously undisclosed natural and antimicrobial defence system.
From a structural perspective, several lines of evidence indicate that the HRR is, at least to a significant extent, responsible for the HRG interaction with Candida membranes. Although the 3D structure of HRG has not yet been determined, modelling studies suggest that the HRR of HRG forms a polyproline (II) helical structure with numerous histidines. At physiological pH, HRG is net negatively charged (pI 6.45). However, due to its high content of histidine residues (∼13%), which are concentrated to the HRR, it can acquire a positive charge by protonation [20],[41], and this in turn likely facilitates the interactions between HRG and Candida. These results were substantiated by the finding that a region of HRG containing the motif sequence GHHPH, was antifungal, and that low pH enhanced this activity. The high conservation of this sequence among vertebrates likely reflects its importance for membrane interactions of HRG [27]. However, as evident in Figure 3B, there are also other antifungal regions in the protein, active irrespective of pH in the interval investigated, an observation compatible with the antifungal activity of HRG detected at neutral pH. It should be pointed out however, that the peptide data do not reflect the complex structure-activity relationships of the holoprotein. Although the CD experiments did not detect any major conformational changes upon interaction with liposomes or polysaccharides, it cannot be ruled out that conformational changes mediated by HRR interactions with intact fungal cells lead to the exposure of additional antimicrobial epitopes in the molecule. Nevertheless, a recombinant and truncated variant of HRG, lacking the histidine-rich and C-terminal domains, was not active against Candida, pointing to the HRR as an important, possibly the most important, effector of HRGs antifungal effects.
Many histidine-rich AMPs are known, among these the clavanins [36], histatins, and calprotectin [42]. We have previously shown that the antibacterial effects in vitro of various histidine-rich peptides, both consensus motifs and peptides derived from domain 5 of HMW kininogen [17] and from HRG [27] are enhanced at low pH or upon addition of Zn2+. Others have reported that the antimicrobial activity of clavanins were substantially increased in low pH as compared with neutral pH [36]. Furthermore, the antimicrobial effect of histatin 5 is enhanced at low pH [43], and histidine-rich variants of magainin, the LAH4-peptides, were recently shown to have increased antibacterial activity in low pH compared to neutral pH [37]. Taken together, the pH dependent activity of HRG is thus comparable to other histidine-rich proteins and peptides, and provides an additional link between pH sensitive AMPs and HRG. However, contrasting to histatins, which translocate through Candida membranes, bind mitochondria, and induce cell death by non-lytic ATP-release [44], HRG acts directly on fungal membranes.
Many AMPs are generated by proteolysis of larger, and non-antimicrobial holoproteins. For example, the cathelicidin LL-37 is released from hCAP18, and other AMPs are proteolytically generated from complement factor C3 and high molecular weight kininogen [3], [11], [12], [14]–[17]. Considering that intact HRG is antifungal, proteolysis of this molecule does not appear to be needed for activity. It is of note that like HRG, several antimicrobial proteins are antimicrobial per se, including bacterial permeability increasing protein, serprocodins such as proteinase 3, elastase and heparin binding protein, as well as lactoferrin [27],[45]. However, it is also described that antibacterial proteins, such as bacterial permeability increasing protein and lactoferrin, may give rise to peptides exerting antibacterial activities [46],[47]. Likewise, it has been shown that HRG may be degraded by plasmin [48], as well as in patients undergoing thrombolytic therapy [49] and bioactive fragments of HRG are involved in antiangiogenesis [26],[41]. Thus, although a major fragmentation of HRG was not observed in this work, e.g., in wound fluid and after binding to fibrin, it is likely that degradation of HRG may occur at sites of high proteolysis and plasmin activity. Indeed, the finding that the HRG-derived peptide GHH20, as well as numerous other other 20mer peptides were antifungal, and as particularly noted for HRR-derived peptides, exhibiting a similar pH dependence as HRG, exemplifies that the holoprotein is not a prerequisite for antifungal action. Clearly, such possibilities need to be addressed in future studies.
As previously mentioned, HRG is involved in various aspects of angiogenesis, coagulation, and fibrinolysis [20], reflecting its interactions with ligands such as heparin, plasminogen, fibrinogen, and thrombospondin. Additionally, it acts as an opsonin by bridging FcyRI receptors on macrophages to DNA on apoptotic cells, stimulating phagocytosis [50], and modulates the binding of IgG and immune complexes to FcγRI [50]. Considering these multiple roles, it is likely that HRG binding to microbial surfaces could induce additional “down-stream” effects, such as modulation of plasminogen activity and phagocytosis. The history of “classic” AMPs have shown that these molecules, initially believed to take part merely in direct microbial killing, have extended their roles into the ability to act as chemokines and to induce chemokine production leading to recruitment of leukocytes, promotion of wound healing, and an ability to modulate adaptive immunity [51]. Indeed, as interest in the in vivo functions of host defence peptides is increasing, it is important to consider the direct antimicrobial and immunomodulatory properties observed. Nevertheless, several findings in this study unequivocally demonstrate that HRG, like many AMPs, acts directly on microbes. Thus, in addition to the antifungal in vitro data, the enhanced fungal growth in HRG-deficient plasma, as well as the finding that Candida was detected at higher levels in blood of Hrg−/− animals, indicates a direct antifungal action of the molecule. It is also interesting to note that these HRG deficient animals have also been shown to be more susceptible to Streptococcus pyogenes infection (Shannon et al, unpublished results). However, considering both AMP and HRG multifunctionality in vitro as well as in vivo, it may be envisaged that additional actions, resulting in the observed antifungal effects, will likely be revealed. All of these effects may be dependent on binding of HRG to microbes and subsequent interactions with cells (e.g., neutrophils and macrophages) in different compartments (e.g., skin, internal organs, and blood). In this respect, the pH dependence of HRG is particularly interesting and relevant. It is well known that infection foci, including abscesses, are characterized by low pH levels reaching as low as pH 5, due to increased anaerobic metabolism and lactate production, as well as leukocyte mediated oxidative burst and subsequent acidification [52]. The capacity of HRG to kill Candida at these pH levels and the corresponding increase in salt-resistance at low pH suggest that HRG could target infection foci, resulting in a physiologically relevant concentration and localization of antifungal activity. As previously mentioned, HRG's opsonising activity could hypothetically lead to enhanced phagocytosis. Although it remains to be investigated, such localisation of antifungal activity to endosomal compartments, where acidification could result in enhanced HRG-mediated killing of phagocytosed fungi, could serve as an effective way of eliminating invading Candida cells at sites of tissue inflammation without releasing potentially toxic microbial components.
Again hypothetically, genetic deficiencies of HRG or acquired functional defects could provide interesting clues with respect to functional roles of HRG. In some patients, reduced levels of HRG are associated with a thrombophilic phenotype, indeed compatible with the phenotype observed in Hrg−/− mice, which had a shorter prothrombin time [22]. As these patients still have ∼20–50% of normal levels of HRG, the human phenotype of complete absence of HRG remains, however unknown. Although patients with low levels of HRG have not been reported to be more prone to infections, it must be remembered that examples from deficiencies of particular innate immune proteins, e.g., complement and mannose-binding lectin, illustrate that even homozygous deficiency and a complete absence of a particular innate immune molecule may give rise to surprisingly mild symptoms. For example, patients with mannose-binding lectin deficiencies are normally not at risk of developing infections unless compromised by immune suppression or severe disease [53]. In this context, it is particularly interesting that antibodies against HRG have been detected in patients with antiphospholipid syndrome [54], a disease associated with thrombodiathesis and systemic lupus erythematosus. Notably, the latter disease is associated with an increased risk for opportunistic infections, including Candida [55]. Taken together, and considering the role of HRG in innate immunity, it should be of interest to study potential associations between functional inactivation(s) or deficiencies of HRG as well as genetically determined differences, in relation to the occurrence of infections.
During the last three decades, research on innate immune molecules has demonstrated the significance of the innate immune system for prevention of invasion by microbes at biological boundaries. Previous studies have emphasized that various molecules, such as “classic” AMPs, complement factors, and cytokines, bridge between innate and adaptive immunity. The present work adds another significant component to this family of molecules, the plasma protein HRG.
The peptides GHH20 (GHHPHGHHPHGHHPHGHHPH) and histatin 5 (DSHAKRHHGYKRKFHEKHHSHRGPY) were synthesized by Innovagen AB (Lund, Sweden), and were of >95% purity. The purity and molecular weight was confirmed by MALDI-TOF MS analysis (Voyager, Applied Biosystems). 20-mer synthetic peptides (PEP-screen) spanning the sequence of HRG (Table 1) were obtained from Sigma-Genosys (St Louis, MO). Polyclonal rabbit antibodies against GHH20 and TAMRA-labeled GHH20 were from Innovagen AB (Lund, Sweden). HRG was FITC-labeled using the FluoroTag FITC Conjugation Kit (Sigma, St Louis, MO). Human serum and plasma were collected from healthy volunteers. Sterile wound fluids were obtained from surgical drainages after mastectomy. The use of human wound fluid was approved by the Ethics Committee at Lund University (LU 708-01). Seminal plasma was collected at the Fertility Center at Malmö University Hospital, Sweden. The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) accession number of human histidine-rich glycoprotein is NP_000403.
The fungi Candida parapsilosis BD 17837 and Candida albicans BD 1060 were clinical isolates. C. parapsilosis ATCC 90018, C. albicans ATCC 90028, Candida glabrata ATCC 90030, and Candida krusei ATCC 6258 isolates were from the American Type Culture Collection (ATCC, Rockville, MD).
Serum HRG was purified using nickel-nitrilotriacetic acid (Ni-NTA) agarose as described before [27]. The concentration of the protein was determined using the Bradford method [56].
Recombinant His-tagged HRGP and truncated version of HRG (HRG1-240), containing amino acids 1-240 was produced and purified as previously described [26],[27].
Plasma, serum, wound fluids, seminal plasma (1 µl), and platelets (fluid from 1×103 cells, disrupted by freeze thawing) were electrophoresed on 8% SDS-polyacrylamide (SDS-PAGE) gel or an 16.5% Tris-tricine gel and transferred to a nitrocellulose membrane (Hybond-C, GE Healthcare BioSciences, Little Chalfont, UK) [57]. The membrane was incubated in 3% skimmed milk in 10 mM Tris, 0.15 M NaCl, pH 7.4 for 1 h at room temperature, followed by incubation for 1 h with rabbit polyclonal antibodies against GHH20 (diluted 1:1000 in the same buffer). The membrane was washed 3 times, and incubated again for 1 h with horseradish peroxidase-conjugated secondary swine anti rabbit antibodies diluted 1:1000 (Dako, Carpinteria, CA). The image was developed using the ECL system (Amersham Biosciences).
Human plasma was subjected to a Ni-NTA agarose gel. The eluent (plasma completely depleted of HRG) was collected and used to form clots. Hrg−/− and C57BL/6 (wild type) mice [22] were used for preparation of fibrin clots from plasma of the respective animals. Plasma deficient of HRG and normal plasma were incubated with a total concentration of 10 mM Ca2+ in eppendorf tubes at 37°C over night. Clots were washed three times and then stored in 10 mM 2-Morpholinoethanesulfonic acid (MES), pH 5.5. Clots (∼0.04g) were used in viable count experiments. To investigate the localization of HRG in fibrin clots, human plasma and HRG-deficient plasma were incubated with 10 µl FITC-labeled HRG (0.4 mg/ml) and then processed as before in the presence of 10 mM Ca2+ over night. The clots were then washed in distilled water and mounted on slides using Dako mounting media (Dako).
C. parapsilosis, C. albicans, C. glabrata and C. krusei were grown to mid-logarithmic phase in Todd-Hewitt (TH) medium (Becton and Dickinson, Maryland, USA) at 27°C and washed in 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5. For dose-response experiments, purified HRG or GHH20 (0.03–6 µM) were incubated with 1×105 C. parapsilosis ATCC 90018 or C. albicans ATCC 90028 for 2 h at 37°C in 10 mM, Tris, pH 7.4 or in 10 mM MES-buffer, pH 5.5, plated on Sabouraud dextrose broth (Becton and Dickinson) agar, and incubated 48 hours at 27°C, whereafter the number of cfu was determined. In order to investigate the antifungal activity of HRG in presence of salt, 6 µM HRG were incubated with 1×105 C. parapsilosis ATCC 90018 for 2 h at 37°C in 10 mM, MES, pH 5.5 containing 0, 25, 50, 100 or 150 mM NaCl, plated and the number of cfu was determined. In kinetic experiments, 0.3 and 3 µM HRG were incubated with C. parapsilosis ATCC 90018 for 5, 15, 30, 60, or 120 minutes in 10 mM MES, pH 5.5, plated and the number of cfu was determined. For determination of the effect of HRG on various Candida strains, HRG (3 µM) was incubated with C. parapsilosis ATCC 90018 or BD 17837, C. albicans ATCC 90028 or BD 1060, C. glabrata ATCC 90030 or C. krusei ATCC 6258 in 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5, plated and number of cfu determined. Truncated and full length recombinant HRG, 0.6 µM rHRG, or rHRG1-240 were incubated with C. parapsilosis (1×105) for two hours and then plated and number of cfu determined. To investigate the in vitro antifungal activity of HRG, normal or HRG-deficient fibrin clots (∼0.04g) were incubated with C. parapsilosis ATCC 90018 for 2 h in 10 mM MES, pH 5.5, plated and number of cfu were determined. For inhibition studies, 0.3 µM HRGP were incubated with C. parapsilosis (1×105) in 10 mM MES, pH 5.5, in presence or absence of heparin (50 µg) for two hours and then plated and number of cfu was determined. In all experiments, 100% survival was defined as total survival of fungi in the same buffer and under the same conditions in absence of peptide, protein, or clots. The p-values were determined using Kruskall-Wallis one-way ANOVA analysis.
Radial diffusion assay (RDA) was performed essentially as described earlier [58]. C. parapsilosis ATCC 90018 and C. albicans ATCC 90028 were grown to midlogarithmic phase in TH-medium, and then washed with distilled water. 4×106 colony forming units was added to 5 ml of the underlay agarose gel (0.03% (w/v) trypticase soy broth (TSB), 1% (w/v) low electroendosmosis type agarose (Sigma), 0.02% (v/v) Tween 20 (Sigma). The buffers used in the underlay gels were 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5. The underlay gel was poured into an 85-mm Petri dish. After agarose solidification, wells of 4 mm in diameter were punched, and 6 µl of peptide solution was added to each well. Buffers were used as a negative control. Plates were incubated at 28°C for 3 h to allow diffusion of the peptides. The underlay gel was then covered with 5 ml of molten overlay. Antimicrobial activity of a peptide is visualized as a zone of clearance around each well after 18–24 h of incubation at 28°C. Peptides were tested in concentrations of 100 µM.
C. parapsilosis (1×105 cfu) were incubated with 0.6 µM HRG in 50 µl 10 mM MES, pH 5.5, with or without heparin (50 µg/ml) for 2 h at 37°C, centrifuged and the pellet was washed three times in 10 mM MES, pH 5.5. The pellet and the supernatant were resuspended in SDS sample buffer, electrophoresed (8% SDS-PAGE), and then transferred to a nitrocellulose membrane. Western blotting was performed as above.
C. parapsilosis ATCC 90018 fungi were grown in TH medium at 27°C to mid-logarithmic phase. The fungi were washed in 10 mM Tris, pH 7.4, and resuspended in the same buffer. C. parapsilosis (2×106/ ml) were incubated with 1 µl of TAMRA-labeled GHH20 (2 mg/ml) in 10 mM MES, pH 5.5, with or without heparin (50 µg/ml), left standing for 5 minutes on ice, and then washed twice in 10 mM Tris, pH 7.4. Fungi were fixed with 4% paraformaldehyde by incubation on ice for 15 minutes and in room temperature for 45 minutes. The fungi were then applied onto Poly-L-lysine coated cover glass and after an incubation time of 30 minutes, finally mounted on slides using Dako mounting media (Dako, Carpinteria, CA). In order to assess permeabilisation, C. albicans ATCC 90028 (2×106 cfu) were incubated with HRG or LL-37 (both at 10 µM) in 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5 for 30 minutes at 37°C. Samples were transferred to Poly-L-lysine coated cover glass and incubated for 45 minutes at 37°C, washed and 2 µg of FITC were added in a volume of 200 µl, and incubated for 30 minutes at 30°C, washed and then fixed as above. Samples were visualized using a Nikon Eclipse TE300 inverted fluorescence microscope equipped with a Hamamatsu C4742-95 cooled CCD camera, a Plan Apochromat 100X objective and a high N.A. oil condenser.
C. parapsilosis ATCC 90018 were grown in TH medium at 37°C to mid-logarithmic phase. The fungi were washed in 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5, and resuspended in the same buffer. HRG or LL-37 (10 µM) was incubated with C. parapsilosis (20×106 cfu) for two hours in a total volume of 10 µl in Tris buffer, pH 7.4 or in MES buffer, pH 5.5. Samples of C. parapsilosis fungi suspensions were adsorbed onto carbon-coated copper grids for 1 min, washed briefly on two drops of water, and negatively stained on two drops of 0.75 % uranyl formate. The grids were rendered hydrophilic by glow discharge at low pressure in air. Specimens were observed in a Jeol JEM 1230 electron microscope operated at 60 kV accelerating voltage. Images were recorded with a Gatan Multiscan 791 CCD camera.
C. parapsilosis ATCC 90018 were grown in TH medium at 27°C to mid-logarithmic phase. The fungi were washed in 10 mM Tris, pH 7.4 or 10 mM MES, pH 5.5 and resuspended in the same buffer. C. parapsilosis (5×107 in a total volume of 0.5 ml) were incubated with 10 µl of FITC-labeled HRG (0.4 mg/ml) or 10 µl TAMRA-labeled GHH20 (2 mg/ml) in 10 mM Tris, pH 7.4 or in 10 mM MES, pH 5.5, let stand for 5 minutes on ice and then washed in 10 mM Tris, pH 7.4. The cells were fixed with 4% paraformaldehyde by incubation on ice for 15 minutes and in room temperature for 45 minutes. Flow cytometry analysis was performed using a FACS-Calibur flow cytometry equipped with a 15 mW argon laser turned a 488 mm (Becton-Dickinson, Franklin Lakes, NJ). The fungal population was selected by gating with appropriate settings of forward scatter (FSC) and sideward scatter (SSC). The FL1 fluorescence channel (λem = 530 nm) was used to record the emitted fluorescence of FITC, and the FL3 fluorescence channel (λem = 585 nm) was used to record the emitted fluorescence of Texas red.
Dry lipid films were prepared by dissolving dioleoylphosphatidylcholine (1,2-dioleoyl-sn-Glycero-3-phoshocholine, >99% purity, Avanti Polar Lipids, Alabaster, AL) (60 mol%) and either ergosterol or cholesterol (both >99% purity, Sigma, St Louis, MO) (40 mol%), and then removing the solvent by evaporation under vacuum overnight. Subsequently, buffer (10 mM Tris, pH 7.4) was added together with 0.1 M carboxyfluorescein (CF) (Sigma, St Louis, MO). After hydration, the lipid mixture was subjected to eight freeze-thaw cycles consisting of freezing in liquid nitrogen and heating to 60°C. Unilamellar liposomes, of about Ø140 nm were generated by multiple extrusions through polycarbonate filters (pore size 100 nm) mounted in a LipoFast miniextruder (Avestin, Ottawa, Canada) at 22°C. Untrapped CF was then removed by two gel filtrations (Sephadex G-50) at 22°C, with Tris buffer as eluent. CF release was determined by monitoring the emitted fluorescence at 520 nm from liposome dispersions (10 mM lipid in 10 mM Tris). An absolute leakage scale was obtained by disrupting the liposomes at the end of the experiment through addition of 0.8 mM Triton X100 (Sigma, St Louis, MO), causing 100% release and dequenching of CF. Although calcein is frequently used for pH-dependent leakage studies, the high charge of this dye has been noted to influence its leakage behaviour in the presence of highly cationic peptides [59]. Instead, therefore, CF was used as a leakage marker at both pH 6.0 and 7.4, however, avoiding pH-dependent fluorescence effects through neutralization prior to probing the limiting leakage in case of pH 6.0 leakage. Throughout, a SPEX-fluorolog 1650 0.22-m double spectrometer (SPEX Industries, Edison, NJ) was used for the liposome leakage assay. Measurements were performed at 37°C.
The CD spectra of the peptides in solution were measured on a Jasco J-810 Spectropolarimeter (Jasco, U.K.). Measurements were performed at 37°C in a 10 mm quartz cuvet under stirring and the effect on protein/peptide secondary structure monitored in the range 200–260 nm. The background value, detected at 250 nm, was subtracted, and signals from the bulk solution were corrected for. The secondary structure was monitored at a concentration of 0.25 µM of HRG in buffer, in the presence of liposomes (lipid concentration 100 µM), and in the presence of mannan from Saccharomyces cerevisiae (0.02 wt%; Sigma-Aldrich, St. Luis, USA).
C. parapsilosis ATCC 90018 were grown in TH medium at 27°C to mid-logarithmic phase. The fungi were washed in 10 mM MES, pH 5.5 and resuspended in the same buffer. C. parapsilosis (2×107) cfu in a total volume of 10 µl was added to 50 µl of human normal plasma or HRG-depleted plasma (eluent from Ni-NTA agarose gel), and incubated for 0, 4, 8 or 18 hours at 27°C and then plated and number of cfu determined.
The original knockout mice 129/B6-HRGtm1wja1 were crossed with C57BL/6 mice (Taconic) for 14 generations to obtain uniform genetic background. These HRG-deficient mouse strain was called B6-HRGtm1wja1 following ILAR (Institute of Laboratory Animal Resources) rules. Wildtype C57BL/6 control mice and C57BL/6 Hrg−/− mice (8–12 weeks, 27+/−4g) were bred in the animal facility at Lund University. C57BL/6 Hrg−/−, lacks the translation start point of exon 1 of the Hrg gene [22]. Animals were housed under standard conditions of light and temperature and had free access to standard laboratory chow and water. In order to study Candida dissemination, C. albicans ATCC 90018 were grown to midlogarithmic phase, washed and diluted in PBS, pH 7.4. Two hundred and fifty µl containing 1×109 cfu was injected intraperitoneally into C57BL/6 or C57BL/6 Hrg−/− mice, divided into weight and sex matched groups. The animals were sacrificed 48 hours post infection, and blood was collected by cardiac puncture. The number of cfu was determined by viable count. In order to study fungal dissemination to target organs, the mice were infected as previously described and three days later the spleen and kidney were harvested on ice.
Representative animals were sacrificed three days post infection and the kidneys were removed into 4% formalin. The tissues were embedded in paraffin, sectioned and stained with Hematoxylin and eosin (H&E) and with Periodic acid-Schiff (PAS).
|
10.1371/journal.pcbi.1000053 | Efficient Olfactory Coding in the Pheromone Receptor Neuron of a Moth | The concept of coding efficiency holds that sensory neurons are adapted, through both evolutionary and developmental processes, to the statistical characteristics of their natural stimulus. Encouraged by the successful invocation of this principle to predict how neurons encode natural auditory and visual stimuli, we attempted its application to olfactory neurons. The pheromone receptor neuron of the male moth Antheraea polyphemus, for which quantitative properties of both the natural stimulus and the reception processes are available, was selected. We predicted several characteristics that the pheromone plume should possess under the hypothesis that the receptors perform optimally, i.e., transfer as much information on the stimulus per unit time as possible. Our results demonstrate that the statistical characteristics of the predicted stimulus, e.g., the probability distribution function of the stimulus concentration, the spectral density function of the stimulation course, and the intermittency, are in good agreement with those measured experimentally in the field. These results should stimulate further quantitative studies on the evolutionary adaptation of olfactory nervous systems to odorant plumes and on the plume characteristics that are most informative for the ‘sniffer’. Both aspects are relevant to the design of olfactory sensors for odour-tracking robots.
| Efficient coding is an overarching principle, well tested in visual and auditory neurobiology, which states that sensory neurons are adapted to the statistical characteristics of their natural stimulus - in brief, neurons best process those stimuli that occur most frequently. To assess its validity in olfaction, we examine the pheromone communication of moths, in which males locate their female mates by the pheromone they release. We determine the characteristics of the pheromone plume which are best detected by the male reception system. We show that they are in agreement with plume measurements in the field, so providing quantitative evidence that this system also obeys the efficient coding principle. Exploring the quantitative relationship between the properties of biological sensory systems and their natural environment should lead not only to a better understanding of neural functions and evolutionary processes, but also to improvements in the design of artificial sensory systems.
| According to the ‘efficient-coding hypothesis’ [1], the sensory neurons are adapted to the statistical properties of the signals to which they are exposed. Because not all signals are equally likely, sensory systems should best encode those signals that occur most frequently. This idea was first tested by Laughlin [2] in a pioneering study of first order interneurons in the insect compound eye, the large monopolar cells, which code for contrast fluctuations. He showed that the response function of these graded potential cells, measured by intracellular recording, approximates the cumulative probability distribution function of contrast levels measured in the natural fly's habitat with a photodiode.
The efficient coding hypothesis has been much studied in the visual system [2]–[7]; reviewed in [8] and to a lesser extent in the auditory system [9],[10]. However, it has been rarely discussed in the context of olfactory sensory neurons [11],[12].
With a nonlinear stimulus-response function, the neuron encodes differently an equal change in stimulus intensity depending on the actual concentration (Figure 1A). The key question is, how should a neuron weigh its input so as to transfer as much information as possible? Information theory [13],[14] provides the solution. In the simplest scenario (with no other constraints on the response range), the inputs should be encoded so that all responses are used with the same frequency [2]. The optimal stimulus statistics is given by the stimulus probability distribution (Figure 1B), which is obtained directly from the stimulus-response curve. This simple solution, however, does not hold in the case of olfaction because of the large differences in reaction time at different stimulus concentrations. This is a major difference with respect to Laughlin's approach, in which all response states were assumed to be equiprobable.
In this paper, we paralleled Laughlin's approach [2], adapting his method to suit the specificity of olfaction. We chose a well studied olfactory receptor neuron, the pheromone receptor neuron of male moths, to investigate its adaptation to the natural signal it processes, the sexual pheromone emitted by conspecific females. To our knowledge this neuron and its stimulus provide the only example in olfaction for which enough data are available on the odorant plume and the neuron transduction mechanisms to make a quantitative comparison possible between the predicted optimum signal and the natural signal.
Flying male moths rely on the detection of pheromone molecules released by immobile conspecific females for mating. The atmospheric turbulence causes strong mixing of the air and creates a wide spectrum of spatio-temporal variations in the pheromonal signal (Figure 2). The largest eddies are hundreds of metres in size and may take minutes to pass a fixed point, while the smallest spatial variations are less than a millimetre in size and last for milliseconds only [15],[16]. Due to inhomogeneous mixing, a very high concentration of pheromone can be found in a wide range of distances from the source, though their frequency decreases with distance [15]. Because of its complicated and inhomogeneous structure, the description of the plume must rely on statistical methods, notably the histogram of the fluctuations in pheromone concentration [15]–[19]. These fluctuations are essential for the insect to locate the source of the stimulus. Experiments in wind tunnels showed that moths would not fly upwind in a uniform cloud of pheromone [20]–[22]. Characteristics like the frequency and intensity of the intermittent stimulation play a key role in maintaining the proper direction of flight [23].
The goal of this paper is to present arguments specifying in which sense the perireception and reception processes occuring in pheromone olfactory receptor neurons (ORNs) can be considered as optimally adapted to their natural stimulus. Although, in the light of previous studies on similar sensory neurons, the ORN may be considered a priori as adapted to the pheromone plume, the exact nature of this adaptation and its proof are more challenging questions. Despite widespread agreement that environmental statistics must influence neural processing [24], precise quantification of the link proved difficult to obtain [8]. So, the main aim of this paper was to identify the specific characteristics to which the pheromone ORN is adapted and to provide quantitative evidence for their adaptation. We proceeded in two steps. First, using the statistical theory of information, we predicted the characteristics of the optimal pheromonal signal that the ORN is best capable of encoding based on the properties of the initial steps of signal transduction. Second, we compared these theoretically-derived properties with statistical characteristics most often determined in experimental measurements, i.e., the probability distribution function of the fluctuations in pheromone concentration, the spectral density function of the stimulation course and the intermittency of the odorant signal.
Pheromone components are detected by specialized ORNs located in the male antenna. We considered a specific ORN type of the moth Antheraea polyphemus detecting (E,Z)-6,11-hexadecadienyl acetate, the major component of the sexual pheromone in this species, for which a wealth of precise information is available (reviewed in [25]). The pheromone molecules are adsorbed on the cuticle, diffuse inside the sensory hair to the neuron membrane and are thought to be enzymatically deactivated [25] then degraded. The initial cell response is triggered by the binding of the pheromone molecules to the receptor molecules borne by the dendritic membrane and the ensuing receptor activation. A cascade of events follows, amplifying this initial response and finally leading to the generation of a train of action potentials conveyed to the brain. The pheromone concentration at each instant determines the ORN response. However the extreme temporal variability of pheromone concentration in plumes prevents a full description of stimulus-response relationships by direct electrophysiological measurements. For this reason we based our study on a model of perireception and reception processes describing how any stimulus (concentration of pheromone in the air) is transformed into the receptor response (concentration of activated receptors). This model, based on extensive biochemical, radiochemical and electrophysiological experiments, was developed by Kaissling and coworkers [25],[26]. It involves the following system of chemical reactions:(1)(2)(3)The network includes (1) the translocation of the ligand from the air (input pheromone signal Lair) to the hair lumen (L); (2) the reversible binding of L to receptor R and the reversible change of the complex RL to an activated state R* (output signal); (3) the reversible binding of L to a deactivating enzyme N and its deactivation to product P which is no longer able to interact with the receptor.
The concentrations of individual components in the network 1–3 are denoted by square brackets and the concentration values are functions of time. For simplicity we omit here the explicit dependence on the time variable t and adopt the following notation for the individual concentrations: Lair = [Lair](t), = [L](t), R = [R](t), RL = [RL](t), R* = [R*](t), N = [N](t), P = [P](t) and NL = [NL](t). The evolution of the system 1–3 in time given the external signal Lair is fully described by five first order ordinary differential Equations 4–8 and two conservation Equations 9 and 10:(4)(5)(6)(7)(8)(9)(10)Equations 9 and 10 follow from the fact that the total concentration of the receptor molecules, Rtot = R+RL+R*, as well as the total concentration of the deactivating enzyme, Ntot = N+NL, do not change over time. We assume that at t = 0 the concentrations L, RL, R*, NL and P are zero. The parameter values, derived from extensive experimental investigations, are given in Table 1.
The efficiency of information transfer in the system 1–3 depends critically on its stimulus-response relationship under single and repeated stimulus pulses. For transferring as much information as possible the response states must be optimally utilized. The actual amount of information transferred is limited by biological constraints. In the system studied, information transfer from Lair (stimulus) to R* (response) presents three main limitations.
First, it is limited by the finite number of receptor molecules per neuron which places an upper bound on the range of responses. Whatever the pheromone concentration (height of the step) the concentration of activated receptors cannot exceed at any time [26].
Second, temporal details in the stimulus course shorter than a certain lower limit Δt cannot be analyzed by the system. The smallest period of stimulation of the model studied here is 0.4 s [26],[27], in agreement with experimental measurements [28],[29]. With smaller periods, at higher frequencies, the amplitude of the oscillations of R* becomes too small to be effective. Therefore we set Δt = 0.4 s. Two successive pheromone pulses separated by a time shorter than Δt cannot be distinguished.
Third, information transfer in time is also limited by the response duration, which depends on the deactivation rate of the activated receptors. The time course of R* in response to stimulations of different heights Lair and limited duration (0.4 s) is shown in the inset of Figure 3A. The concentration of activated receptors rises at first, reaches RΔ* at the end of the stimulus pulse, i.e., RΔ* = R*(t = Δt), and finally decreases. We consider RΔ* as the “response” of the system and for the sake of simplicity in the following, we omit index Δ. The duration of the falling phase (receptor deactivation) gets progressively longer for higher pheromone concentrations. This deactivation takes typically much longer than the time resolution parameter Δt. The falling phase is often described by the half-fall time, τ(R*), which is the time required for R*(t) to decrease from R* to R*/2. The relationship between R* and τ(R*) is shown in Figure 3A. A unique value of R* corresponds to each value Lair, which defines the stimulus-response curve (Figure 3B). The fact that the deactivation of activated receptors is relatively slow suggests that the reception system cannot encode a long sequence of pheromone pulses in arbitrarily quick succession. This observation plays an important role in the definition of the optimal stimulus course.
In the simplest scenario (with no other constraints on the response range and stimulus-independent additive noise), the inputs should be encoded so that all responses are used with the same frequency [2],[30]. The optimal stimulus is thus described by its probability distribution function, which is obtained directly from the stimulus-response curve. Due to the large differences in reaction times at different stimulus concentrations, all response values R* from 0 to 0.24 µM cannot be considered as equally “usable” (the long falling phases decrease the efficacy of the information transfer). Therefore, the longer the half-fall time of a given response R* (i.e. the greater concentration R* is) the less frequent it must be. The particular form of the optimal response cumulative probability distribution function (CDF), FR(R*), which was determined by maximizing the information transferred and minimizing the average half-fall time (see Methods), is shown in Figure 3C. Then, based on the three factors mentioned (stimulus-response curve, Figure 3B; time resolution Δt = 0.4 s; and optimal response probability distribution, Figure 3C), an optimum stimulus course in time can be predicted as explained in the Methods section.
Examples of predicted temporal fluctuations in pheromone concentration are shown in Figure 4 at various time scales and compared to experimental observations. Even though the time resolution of the system studied here is only 0.4 s, it seems sufficient to capture the main bursts of pheromone (see the 10 s sample in Figure 4A). The comparison can be made more precise by describing statistically the heights and occurences in time of the pulses.
Concerning temporal aspects, the bursts of non-zero signal do not occur at periodic intervals but appear randomly. An important descriptor of the temporal structure is the intermittency [15],[16], which is the fraction of total time when the signal is present. The intermittency of the predicted optimal stimulus is 20%, which is in relatively good agreement with experimental data. It has been shown using various types of ion detectors [17],[19] as well as electroantennogram responses [17],[31], that the natural signal is always present less than 50% of the total time, and usually smaller values are found. The average intermittency values reported are 10–20% [15] and 10–40% [16],[17], depending on the experimental conditions, such as the detector size or the global meandering of the plume (see Discussion).
Concerning pulse height, the overall character of the predicted stimulus course is that pulses of high concentration are much rarer than those of low concentration. This feature of the predicted stimulus can be best quantified by the CDF, P(Lair), of the stimulus. The shape of the CDF is one of the most important properties for comparing theoretical predictions to experimental measurements because it describes the relative distribution of odorant concentrations throughout the plume. In fact, because measuring pheromone concentration in the field is not presently feasible [17], pheromone molecules must be replaced by measurable tracers. Relative quantities are valid for both pheromones and tracers (see Discussion). They are the only quantities known experimentally for pheromone plumes. So, although our model predicts them, we cannot compare values of Lair to actual measurements.
Given the definition of the optimal stimulus, function P(Lair) can be directly computed (see Methods). Figure 5 shows a comparison between experimentally measured (A) and predicted (B) concentration CDF. The optimal pheromone concentration CDF (Figure 5B, solid line) is not known in analytical form but it can be well approximated by an exponential CDF (Figure 5C, dashed line). The differences between the predicted and true exponential shape can be considered as non-significant, namely, very high values of Lair are predicted to be less frequent than in the exponential model. The exponential CDF is in agreement with experimental CDF (Figure 5A), [18],[19],[32],[33] and holds well especially for observations closer to the source (less than 100 m). Although the precise form of the CDF varies with distance from the plume centerline [19] and may be affected by the measurement technique, the shape is always highly skewed.
Other predicted relative quantities (peak-to-mean ratios, dimensionless concentrations Lair/〈Lair〉) were compared with their experimental counterparts. The results, summarized in Table 2, show that the predicted statistical properties of the stimulus are not contradicted by the experimental observations.
Spectral density functions of the concentration time course, which analyze the contribution of various frequencies to the overall stimulus course, characterize other properties of the plume which are independent on the nature of the odorant (pheromone or ion source) [19],[33]. Furthermore, spectral density function represents a point of view different from the concentration probability distribution.
Several spectral density functions, shown in Figure 6, were calculated from the predicted optimal pheromone stimulation (see Methods). The spectral shapes seem to be almost flat from 0.02 Hz to 0.2 Hz with a decreasing slope close to −2/3 above 0.2 Hz. The same slope −2/3, which is theoretically predicted by the inertial subrange theory [19], was reported in the spectral densities obtained from measurements close to the source (less than 100 m), in the range 0.1 Hz (or 0.5 Hz, depending on records) to 1 Hz [19],[33], although the precise range may depend on the technique of measurement.
The goals of this study were to determine to which extent early olfactory transduction in olfactory receptor neurons can be considered adapted (in the evolutionary sense) to odorant plumes and to specify the plume characteristics to which it is adapted. The formulation and resolution of this problem benefited from successful studies of efficient sensory coding undertaken in the field of vision and audition. However, transposition from these sensory modalities to olfaction is not straigthforward, which may explain in part why it has not been attempted earlier. Specificities of olfaction concern both the odorant plume and the sensory system.
In theory and in practice, the quantitative description of odor plumes and their spatiotemporal distribution is less straightforward than that of visual or auditory scenes. Contrary to light and sound, for which the physical description is essentially complete, the turbulent phenomena which underlie the plume characteristics are still an incompletely mastered domain of physics [34].
In Laughlin's classical experiment in vision a single time-independent variable, the contrast level, was measured [2] and directly compared with experimental data. In olfaction, however, the odorant concentration (an analogue to the contrast level) is essentially time dependent which results in a complex optimal stimulus course (Figure 4). Complexity and time dependence make a meaningful direct comparison between predictions and experimental records, but also between different experimental records, impossible. Instead, the comparison must rely on global, statistical descriptors [15],[17],[19],[33]. We identified 5 such descriptors of odor plumes, actually measured and usable in the present context (see Table 2), which summarize the present knowledge on odor plumes.
Moreover, there are no easy-to-use instruments to measure odor plumes in the field, comparable to luxmeters and microphones. For example, the absolute pheromone concentration cannot be easily known in field experiments [17]. This explains why no experimental values were given for this descriptor in Table 2. In practice, only ratios of concentrations are presented because they are independent of the dispersed molecules. The pheromone is often substituted by an ion or a passive tracer (polypropylene for example) whose concentration can be measured [15],[17],[19]. Because both pheromone and tracer compounds in the air are governed by the same physical laws, the relative (dimensionless) values are conserved, as confirmed by independent experiments with different sources [15]–[17],[33]. More generally, this limitation explains why we compared only relative quantities (i.e. shape of probability distributions, spectral density functions, peak-to-mean ratios, dimensionless concentrations Lair/〈Lair〉 and intermittency values). Other limitations of plume measurements are discussed below.
The essentially multidimensional and stochastic nature of the odor stimulus has a profound influence on the analysis of olfactory transduction system in its natural context, as undertaken here. Indeed to investigate the problems at hand, the kinetic responses of the system to a very large number of stimuli, varying in intensity, duration and temporal sequence must be known in order to simulate the diversity of stimuli encountered in a natural plume. This task is difficult, if not impossible, to manage in a purely experimental approach. However, this difficulty can be overcome with an exact dynamic model of the system because its response to the diverse conditions mentioned can be computed, provided it includes all initial steps from molecules in the air to the early neural response. This is the case of the perireception and reception stages of the moth pheromonal ORN and the reason why it was chosen in the present study. This choice brings about two questions, one about the validity of the model, the other on its position within a larger context.
The computational model employed has been thoroughly researched and improved over the last three decades [25], [35]–[37]. It describes perireceptor and receptor events in the ORN cell type sensitive to the main pheromone component of the saturniid moth Antheraea polyphemus. At the time of writing it represents the most completely researched computational models of its kind, agreeing with extensive experimental data from various authors and a wide range of experimental techniques. This model is the best description presently available for early events in any ORN and it summarizes in a nutshell a wealth of dispersed knowledge. This model is based on ordinary differential equations 4–8, following the law of mass action for chemical reactions, and is therefore purely deterministic. This approximation is acceptable when the concentrations of reactants are high enough above single-molecular levels, so that the stochastic fluctuations can be neglected. In this paper, the concentration of R* is always well above that corresponding to one activated receptor molecule per neuron (approximately 10−6.2 µM) because we do not investigate the effect of extremely small pheromone doses. Then, the response of the system can be considered as deterministic, in accordance with the efficient coding hypothesis [8].
The system studied here constitutes only a small part of the whole pheromonal system, although its role is absolutely essential and all other parts depend on it. First, in ORNs, post-receptor mechanisms modify the receptor signal, primarily by a large amplification factor and by sensory adaptation. Second, the ORN population includes cell types with different properties, e.g. the ORN type sensitive to the minor pheromone components can follow periodic pulses up to 10 Hz [29], a performance not yet accounted for in present models [27]. Third, in the brain antennal lobe, convergence of a large number of ORNs on a few projection neurons (PNs) provides another amplification and supports the ability of some PNs to follow periodic signals at 10 Hz or greater [38]. Evolutionary adaptation of an integrated ORN response is difficult to study at the present time because no complete model of the ORN from receptors to the generation of the receptor potential and the ensuing spike train, is yet available, at least with the required degree of precision. The same argument holds a fortiori for higher order processes. Notwithstanding, the study of the early sensory events is not as restrictive as it may seem because any incoming odor signal must be first transduced in the population of membrane receptors. No information can be extracted by the post-receptor transduction system which has not been encoded by the receptors in the first place. For this reason it is essential to investigate the nature of the adaptation of the initial events (pheromone interaction with receptors) to the pheromone signal.
Different response states of the pheromone reception system have different efficacies from the coding point of view: the “high” states, with large concentrations of activated receptors, take much more time to deactivate than the “low” states, so that for some time after its exposition to a large concentration of pheromone the system is “dazzled”. It means that in the optimal stimulus the low pheromone concentrations must be more frequent than the high ones. This is a difference with respect to the classical problem where the efficacy of all response states at transferring information is considered the same, as in the vision of contrasts for example. The problem to solve is to find the right balance between two conflicting demands: to use all response states (including the high ones) and to react rapidly (the short transient responses must be as frequent as possible), i.e. to maximimize the information transferred per time unit.
The solution to this optimization problem is provided by information theory as detailed in the Methods section. The optimal balance derives from Equation 19 which relates the average half-fall time and the maximum response entropy distribution. The key factor to consider in the optimization is the average half-fall time, which characterizes globally the “swiftness” of the system – smaller average half-fall time means faster stimulation rate. In other words, the average half-fall time characterizes the bias towards “low” response states. Simultaneously, the condition of maximum response entropy guarantees that the temporal dynamics of the system is as varied as possible and that during the course of stimulation every possible response state is used (with appropriate frequency). By taking into account only the average half-fall time, and not the precise sequence of its individual values, we therefore do not neglect or limit the temporal dynamics of receptor molecules activation. It is important to note, that the average half-fall time is not a free parameter of the problem; it is not set a priori: its optimal value follows from the optimization procedure (Equation 20). The resulting optimal response CDF is highly biased towards low response states, as expected (see Figure 3C).
The main achievement of the present investigation was to predict the characteristics of the stimulus optimally processed by the receptor system based on its biochemical characteristics and an information theoretic approach. The predicted optimal plume was shown to be close to the actual plumes for a series of characteristics, namely intermittency, peak/mean ratio and peak/standard deviation ratio of pheromone pulses, probability distribution of dimensionless pheromone concentration and spectral density function of pheromone concentration (Table 2, Figures 4–6). The correspondence between the predictions and measurements is very good for the last two characteristics (probability distributions) and fair for the first three (numerical values).
These differences in precision of the predictions may be interpreted by taking into account technical factors. Increasing the noise rejection threshold leads to a decrease of the measured intermittency [15],[19], while increasing the detector size or averaging the signal over longer time windows has the opposite effect [39]. So, for example, the small size of olfactory sensilla with respect to detectors may explain in part why in Figure 4B, the predicted intermittency seems lower than that in the corresponding experimental record sample, and also why the peak-to-mean ratio and peak-to-standard deviation ratio are relatively higher. The immobility of the measurement devices, in contrast with the active movements of the moths, is another significant factor. For example, long pauses (of the order of minutes) of zero signal are missing in the prediction but visible in the longest available field record (350 s, Figure 4C). They are caused simply by the plume being blown away from the immobile field detector. First, this loss of signal is clearly an extraneous effect, which cannot be included in our optimal signal predictions and therefore cannot be seen in our results. Second, the moth is not subjected to this extraneous effect, or at least not to the same extent, because, in case of signal loss, it actively seeks the pheromone plume, whereas the fixed detector must passively wait for its return. This difference of mobility may substantially affect the intermittency values, but does not affect the shape of probability distributions (see Methods), hence the better quality of the fits in the latter case. In conclusion, the results obtained suggest that the perireceptor and receptor system investigated here is evolutionary adapted to the pheromone plumes.
Even if one considers that the pheromone olfactory system must be a priori adapted to the average characteristics of the pheromone plumes, it does not logically follow that the system studied is itself necessarily well adapted. Indeed, it is conceivable that the global adaptation results mainly, not from perireception and reception processes but from other downhill intra- and intercellular processes involved in higher signal processing. The respective importance of the former and latter processes in global adaptation cannot be decided a priori. Therefore, the relatively close correspondence between predicted and observed plume characteristics presented here is not trivial. It suggests that the adaptation at the level of receptors is already substantial, and consequently that the global adaptation is not predominantly the result of post-receptor mechanisms involving amplification, sensory adaptation, convergence of different ORN types in the antennal lobes etc. The role of these mechanisms in the global adaptation of the animal remains to be established, as well as the relative importance of the various components of the olfactory system (receptor population, ORN as a whole, population of pheromonal ORNs in the antenna, projection neurons in the antennal lobes, etc.). The response characteristics of these other subsystems, e.g. their various temporal resolutions, will have also to be interpreted, maybe in relation with changing plume characteristics with distance to the source and other factors yet to be identified.
As mentioned in the Results section, information transfer in the pheromone reception system is limited by the finite response range, (), and by the deactivation rate of the activated receptors for each concentration value R*. This deactivation rate is described by the half-fall time τ(R*). The optimal performance of the system is thus reached by a trade-off between two conflicting demands: to employ full response range (maximum information) vs. to employ only the “fastest” responses (minimum average half-fall time). In other words we need to maximize the information transferred per average half-fall time. In the following we provide the mathematical framework that enabled us to find the probability distribution function over the response states R* that realizes this trade-off.
The optimal stimulus course in time was calculated as follows. First, at time t0 = 0 a random value p0 is drawn randomly from a uniform probability distribution function over the range [0,1]. The concentration corresponding to probability p0 is obtained by solving the equation(23)where FR(R*) is the optimal CDF given by formula 22 (Figure 3C). The predicted optimal concentration Lair,0 for a pheromone pulse of duration Δt = 0.4 s which corresponds to is obtained by solving the equation(24)where R*(Lair) is the stimulus-response function (Figure 3B). The value Lair,0 is plotted at t0 (Figure 4). Second, the concentration Lair,1 and time of appearance t1 of the next pulse are determined. Time t1 follows from the falling phase of activated receptors: optimality requires that no pheromone pulse appears before R* returns to its resting level. In practice it is considered that the resting level is reached when R* falls below 0.01 µM (less than 5% of the coding range). The concentration Lair,1 of the pulse at t1 is determined in the same way as for the pulse at t0 by drawing a new random number p1 from the uniform probability distribution function over [0,1]. The same process can be repeated as many times as needed to create an optimal pheromone pulse train of arbitrary length.
It is common in the literature on the statistical analysis of plumes [15],[18],[19] to define two types of mean concentrations. The total mean concentration, 〈Lair〉, describes the “true” mean concentration obtained from the whole record of concentration fluctuations in time, i.e., including the parts where no signal was available. On the other hand, the conditional mean concentration, 〈Lair〉cond, describes the mean concentration inside the plume, i.e., with zero concentrations excluded. The intermittency, γ, relates the two means as [19](25)(Analogously, the total variances and total standard deviations are calculated by taking into account also the parts where no signal is available [19].)
By combining Equations 23 and 24 we may symbolically express the optimal CDF of the stimulus, P(Lair), as(26)Though P(Lair) cannot be expressed in a closed form, it can be well approximated by the exponential CDF(27)where ξ = (5.24±0.01)×10−4 µM is the estimated value of 〈Lair〉cond by least-squares fitting of Fexp(Lair) to P(Lair).
In order to compare concentration probability distribution functions from different measurements meaningfully, authors [19] plot the CDF for a dimensionless concentration Lair/〈Lair〉. (In the Figure 5A C/〈C〉 is used, since the data plotted were obtained using a propylene source, not pheromone), see Figure 5A. The scale of such plots is affected by intermittency due to the presence of the total mean in the ratio. Furthermore, information about intermittency is included explicitly in the plots by letting the probability P(Lair = 0) of zero concentration be(28)Consequently the CDF P(Lair) must be renormalized [19]. Intermittency affects only the dimensionless scale, Lair/〈Lair〉, and the value of P(Lair = 0) but not the overall shape of CDF [19]. Therefore we can use formulas 25 and 28 to compare our predictions with experimentally measured data by correcting for different intermittency values.
The optimal stimulus course is represented by pulses of different pheromone concentrations, Lair, occurring in time intervals 0.4 s long. In order to calculate the spectral density function of such stimulation course we sample the time axis with step Δt = 0.4 s. Thus we obtain a series of pheromone concentrations at these time points, {Lair,j}, j = 1…n, where n should be even. The discrete Fourier transform, φk, of {Lair,j} is defined for k = 1,…,n values as [41](29)where i is the complex unit. The zero-frequency term is thus at position k = 1. The spectral density, S(f), of the complete time course of the stimulus can be calculated for a total of n/2+1 values of frequency f (given in Hz) [42](30)where m = 0, 1, 2,…, n/2−1, n/2 and f = m/(nΔt) are the frequency values. The function Π(f) is the Fourier transform of a pulse of unit height, 0.4 s long and starting at t = 0 [41],(31)where a = 2.5 and δ = −0.5. The function Π(f) appears in formula 30 because the whole stimulus course (such as shown in Figure 4, bottom panels) can be reconstructed by convolving the discrete series {Lair,j} with such a pulse of unit height in the time domain [41]. |
10.1371/journal.pcbi.1004615 | How Obstacles Perturb Population Fronts and Alter Their Genetic Structure | As populations spread into new territory, environmental heterogeneities can shape the population front and genetic composition. We focus here on the effects of an important building block of heterogeneous environments, isolated obstacles. With a combination of experiments, theory, and simulation, we show how isolated obstacles both create long-lived distortions of the front shape and amplify the effect of genetic drift. A system of bacteriophage T7 spreading on a spatially heterogeneous Escherichia coli lawn serves as an experimental model system to study population expansions. Using an inkjet printer, we create well-defined replicates of the lawn and quantitatively study the population expansion of phage T7. The transient perturbations of the population front found in the experiments are well described by a model in which the front moves with constant speed. Independent of the precise details of the expansion, we show that obstacles create a kink in the front that persists over large distances and is insensitive to the details of the obstacle’s shape. The small deviations between experimental findings and the predictions of the constant speed model can be understood with a more general reaction-diffusion model, which reduces to the constant speed model when the obstacle size is large compared to the front width. Using this framework, we demonstrate that frontier genotypes just grazing the side of an isolated obstacle increase in abundance, a phenomenon we call ‘geometry-enhanced genetic drift’, complementary to the founder effect associated with spatial bottlenecks. Bacterial range expansions around nutrient-poor barriers and stochastic simulations confirm this prediction. The effect of the obstacle on the genealogy of individuals at the front is characterized by simulations and rationalized using the constant speed model. Lastly, we consider the effect of two obstacles on front shape and genetic composition of the population illuminating the effects expected from complex environments with many obstacles.
| Geographical structure influences the dynamics of the expansion of populations into new habitats and the relative importance of the evolutionary forces of mutation, selection, genetic drift, and gene flow. While populations often spread and evolve in highly complex environments, simplified scenarios allow one to uncover the important factors determining a population front’s shape and a population’s genetic composition. Here, we follow this approach using a combination of experiments, theory, and simulations. Specifically, we use an inkjet printer to create well-defined bacterial patterns on which a population of bacteriophage expands and characterize the transient perturbations in the population front caused by individual obstacles. A theoretical understanding allows us to make predictions for more general obstacles than those investigated experimentally. We use stochastic simulations and experimental expansions of bacterial populations to show that the population front dynamics is closely linked to the fate of genotypes at different parts of the front. We anticipate that our findings will lead to understanding of how a wide class of environmental structures influences spreading populations and their genetic composition.
| Populations expand into new territory on all length and time scales. Examples include the migration of humans out of Africa [1], the recent invasion of cane toads in Australia [2], and the growth of colonies of microbes. Although populations often persist long after invading [3], events during their spread can have long-lasting effects on their genetic diversity [4, 5]. Considerable effort has been undertaken to understand the role of the invasion process on the evolutionary path of the population: The small population size at the edge of the advancing population wave amplifies genetic drift, reducing genetic diversity, which can culminate in the formation of monoclonal regions [4]. The fate of mutations—deleterious, neutral, or beneficial—occurring in the course of the expansion depends on the location of their appearance with respect to the edge of the wave [4, 6–10]. While the genetic consequences of such range expansions have been studied in the field [11, 12], the complexity of natural populations makes it difficult to draw general conclusions. Laboratory expansions of microbes have thus become a useful tool to illustrate, test, and inspire theoretical predictions [13–16].
The majority of theoretical and experimental work on range expansions has focused on homogeneous environments while habitats in nature are often spatially heterogeneous with regard to dispersal or population growth, the two processes that lead to the expansion. Incorporating environmental heterogeneity into models of spreading populations [3, 4, 17, 18] raises complex problems. Heterogeneity can affect any parameter that controls population dispersal or growth and there can be many spatial patterns of heterogeneity. Ecologists and population geneticists often focus on different consequences of environmental heterogeneity. Work in population dynamics and ecology typically concentrates on the effect of heterogeneity on invasibility and the speed and impact of an invasion in such environments [3, 17, 19–22] and is closely linked to the mathematical and physical aspects of front propagation [23, 24]. In contrast, population genetics studies usually assume a successful invasion and ask how environmental heterogeneities affect the population’s genetic composition [4]. Although heterogeneous carrying capacities [25], fragmented environments [26], single corridors or obstacles [8, 27], and environmental patterns found on earth [7, 8, 28] have been addressed from a theoretical perspective, a systematic understanding is still missing. In this work, we study the population dynamics and relate the dynamics of the population front to the consequences on the genetic composition of the spreading population, thereby linking the evolutionary and ecological consequences of range expansions.
We want to understand what happens when expanding populations confront environmental heterogeneities. For simplicity, assume that at each point the environment is a high quality habitat (large growth rate at population density well below carrying capacity) or a low quality habitat (very small or zero growth rate). What constitutes low quality habitats depends on the population: For a macroscopic expansion of a terrestrial animal or a plant, lakes and mountain ranges are examples of low quality habitat. For microbes, regions with poor nutrients may represent an obstacle to colony growth. If ρ is the fraction of the environment that allows growth, we can distinguish between two scenarios: For 0 < ρ ≪ 1/2, the ‘island scenario’, a largely inhospitable environment is interrupted by islands or oases of growth; in contrast, for 1/2 ≪ ρ < 1, the ‘lake scenario’, a largely hospitable environment is punctuated by obstacles that impede growth (Fig 1A). The island scenario, reminiscent of stepping stone models of population genetics [29], with a weak coupling between nearby islands by migration, is a situation where genetic drift can lead to genetic uniformity on individual islands due to founder effects [30]. Here, we address the lake scenario in the context of spatial expansions.
In addition to the fraction of the habitat not occupied by obstacles, ρ, we must also consider the number of obstacles, N, in the new habitat to be invaded. When N ≫ 1, i.e., when many (non-overlapping) obstacles are engulfed as the range expansion progresses, we expect that the principal effect of interest from a population dynamics perspective is the speed of the invasion and the roughness of the population front (top of Fig 1A). If, at the other extreme, the expansion domain only includes one obstacle with a size comparable to the size of the habitat invaded, the obstacle’s size, shape, and spatial arrangement are expected to greatly influence the shape of the front at the length scale of the habitat (bottom of Fig 1A).
We study the effect of isolated obstacles on the spread of populations. Using a combination of experiment, theory, and simulation, we characterize the obstacle’s effect in a regime of sizes where the shape of the front is well-described by a phenomenological model of expansion with constant speed. The constant speed model reveals general effects which hold independently of the mechanisms for population spread: The perturbation in the population front induced by the obstacle is determined by the obstacle’s width, but not by its precise shape. The front shape, induced by the obstacle, governs the effect on the genetic composition of the expanding population. Expanding past obstacles reduces genetic diversity and privileges genotypes that just miss an obstacle’s edges, an example of ‘geometry-enhanced genetic drift’, effects which are reflected in the genealogy of individuals at the front. In addition to the phenomenological model of front shape, we study a reaction-diffusion model, which enables us to compare experiments to a theoretical description in more detail and to understand the utility of the constant speed model in situations that extend beyond the experimental system studied here.
To derive these findings, we combine an analytical model, simulations, and experiments. While the experiments are the basis for theoretical work, they also allow us to test theoretical predictions. The analytical model provides the opportunity to make predictions for a variety of environments and length scales while simulations are used to explore regimes not accessible to analytical solutions. In addition to using established theoretical and experimental methods to study expanding populations, we present a new laboratory model system which allows us to quantitatively study population spread in heterogeneous environments: the expansion of bacterial viruses (bacteriophage) on a lawn of sensitive and resistant bacteria. Patches of resistant bacteria represent obstacles to the spread of the phage and can be generated using a printing technique, allowing us to quantitatively test predictions.
The growth dynamics of the phage system with phage and bacterial host differs from the one-species system with logistic growth, the FKPP equation [43], often used to study population expansions theoretically and used as the basis for our reaction-diffusion model. The dynamics of phage spread is governed by the density of phage, the density of bacteria, and the density of uninfected bacteria. However, at long times, the profiles of infected and uninfected bacteria are slaved to the profile of a traveling population wave of phage with a constant speed, and its motion mimics the dynamics of the simpler FKPP model. This similarity makes sense, because it is well-known that under broad conditions the solution to reaction-diffusion equations produces traveling waves with constant velocity whose speed is determined by linearization at the foot of the wave [43]. Hence, we expect that our phage system reflects well aspects of range expansions that depend on the biology at the leading edge of the front and believe that it offers the prospect of studying demographic and evolutionary processes in complex, yet well-defined environments.
To explore the effects of obstacles on the population front dynamics, we employed a microbial model system, bacteriophage T7 spreading on a lawn of E. coli cells. Phage T7, a virus of E. coli, infects bacterial cells and lyses them, releasing a large number of new phage particles which undergo passive dispersal and can infect nearby cells, a cycle of growth and replication that leads to an advancing population front. Phage T7 must kill the bacteria it infects [31] and its spread on a bacterial lawn is revealed by the growth of plaques (clearings in the lawn due to lysis of bacteria), a fast process easily visualized by bright-field or fluorescence microscopy. We produce a heterogeneous environment for phage spread by incorporating regions which do not support propagation of the population wave: While a wild-type bacterial region (marked by a constitutively expressed yellow fluorescent protein) corresponds to a region supporting phage production, a resistant bacterial patch, an obstacle (similarly marked by a red fluorescent protein (mCherry)), does not, see Fig 1B and Materials and Methods. A lawn with regions of susceptible and resistant bacteria represents a static, heterogeneous environment that the phage population travels across during its expansion and that can be easily visualized.
We designed an assay that allowed linear fronts of expanding phage populations to encounter obstacles of defined shape. We modified a method that used a consumer inkjet printer to print sugar solutions [32] to deposit bacteria in defined patterns on agar surfaces (such as was done using custom-made equipment [33]). The printer produces a field of bacteria on a rectangular (3.5 × 2 cm2) agar patch at sub-mm resolution (Fig 1C, Materials and Methods, S1 Protocol and S1 Fig). The printed founder cells grow into a lawn, which is inoculated with a linear front of phage T7 close to or at the region with resistant bacteria (Fig 1D). The phage population spreads on this heterogeneous lawn, with repeated cycles of infection and lysis of the susceptible bacteria leading to the loss of fluorescence and the expanding dark region. Fig 1D shows such a printed pattern at different stages of the phage invasion (see also S1 Video): A linear population wave of phage encounters the region of resistant bacteria, the obstacle. The front curves as it passes the widest part of the obstacle and the two curved regions move along the far side of the object until they unite with each other, giving rise to a kink that disappears with time as the front moves beyond the obstacle.
We used the difference between two consecutive images to detect the front of the plaque (Materials and Methods), and studied the front position as a function of time. We define the unperturbed front position d(t) as the position of the plaque’s edge at a horizontal distance of ±3 mm away from the obstacle center as displayed in Fig 1D. Fig 2A, displaying front position as a function of time, shows that the plaque grows at an approximately constant speed, but slows down slightly over time, presumably due to E. coli entering stationary phase [34]. The varying slope illustrates variation in front speed among replicates. Overall, the plaque extends with an approximately constant speed of 0.2 mm/h. The profile of the fluorescence signal in direction of the moving front is constant in time as shown in S2 Fig. Fig 2B shows the front shape over time for multiple replicates. The evolution of the front is very similar across replicates, despite small variations in front speed and initial conditions. While the perturbation of the front by the obstacle and the formation of a kink is intuitive at first, we aimed for a quantitative model which can describe front shape and make predictions which can be tested experimentally.
An arguably most minimal model assumes that the front moves with constant speed in direction normal to the front and ignores the microscopic details of how phage replicate inside bacteria and diffuse outside them. We dubbed this the ‘constant speed model’. Fig 2C–2G illustrates the dynamics of a front encountering a rhombus-shaped obstacle: (C) An initially linear front moves forward uniformly until the obstacle is encountered. (D) When it encounters the obstacle, the front stays linear, but is interrupted in the interval where it would overlap with the obstacle. (This is different from scenarios where a front of material encounters an obstacle and the obstacle “pushes” the material to the sides.) (E) Beyond the obstacle’s widest points, propagation with constant speed creates circular arcs in the shade of the obstacle that are connected to a linear front on either side of the obstacle. The circular elements span a region given by the obstacle’s width and encounter the obstacle at a 90° angle. (F) For a rhombus with height 2h and width 2w, the arcs from the two sides meet and a kink forms after the front has traveled a distance w 2 + h 2 beyond the point of maximum width. (G) The kink then heals due to the increasing radii of the circular segments, i.e., the size of the indent Δ decreases (Δ(d) ∼ 1/d for large “downstream” distances d, where d is the distance perpendicular to the front from the widest point of the obstacle to the unperturbed portion of the front, see below). Fig 2H shows that the height of the rhombus-shaped obstacle does not play a role in determining front shape and thus the size of the indent while the kink heals: For an obstacle which is taller (light red rhombus), the kink forms later and the circular arc where it forms is correspondingly shorter, but the shape of the downstream kink is independent of the obstacle height. Moreover, the circular shapes of the front on the far side of rhombus-shaped obstacles all fall onto the same master curve when plotted in units of w. A calculation shows that the indent size Δ as function of position d is indeed independent of h for rhombus-shaped obstacles and shows the same functional behavior if all lengths are expressed in units of w (see also S1 Appendix):
Δ ( d ) w = d w 1 - 1 - w 2 d 2 ≈ d ≫ w w 2 d . (1)
Counterintuitively, the width of the obstacle thus is a more important predictor of downstream front shape than obstacle height. For rhombus-shaped obstacles, obstacle height determines where the kink forms, but not the shape of the front after formation of the kink. Below, we will discuss more general obstacle shapes and the influence of obstacle size on the applicability of the constant speed model.
The constant speed model predicts the shape of the front at a given front position relative to the obstacle, this way allowing direct comparison with the experimentally determined front shapes in Fig 2B. While the constant speed model captures overall front shape and the transient character of the perturbation, the details of the predicted front (black line) differ from the experimental data (colored lines). The experimental profiles consistently lag behind the predicted front.
The constant speed model also predicts that the shape of the front, scaled by the obstacle’s width, is identical for all rhombus-shaped obstacles. To test this prediction, we repeated the experiment for three more obstacles, in total combining two different widths with two different heights. Fig 3A and 3B display front shapes and the indent sizes as measured for all four obstacle shapes. As predicted, the data collapse very well onto single lines if lengths are divided by w. (This is not the case for other scalings, see S3 Fig.)
Although the constant speed model successfully predicts how the experimentally determined front shape (colored lines) scales with the obstacle’s width and the front’s distance from the obstacle, the experimental profiles display a lag relative to the predicted front (black line) for all four different obstacles and, equivalently, have a larger indent size than predicted (Fig 3A and 3B). However, this quantitative disagreement does not affect the scaling behavior of Δ(d) (Fig 3B, see S4 Fig for the same data on linear scale).
The constant speed model captures the general features of the front dynamics observed in the phage experiment, but the deviation prompted us to study a more detailed model which considers more of the details of phage propagation. In addition to understanding the deviation, the more detailed model will shed light on the range of applicability of the constant speed model.
The dynamics of plaque growth on homogeneous lawns has attracted considerable interest in the past [35–39]. A reaction-diffusion model, which captures the phage-bacterial interaction, the phage life cycle, and focuses on bacteriophage T7, has been suggested by Yin and McCaskill [36]: phage bind bacteria to form infected cells, and these, with a rate constant, burst to release more phage. More complex successor models focusing on the delay between infection and release of progeny phage have been published [38]. We decided not to generalize these models to heterogeneous environments for two reasons: (i) The appropriate parameters are not known for our experiments and (ii) we aimed for a general description that allows us to reach conclusions that extend beyond the infection of E. coli by bacteriophage T7.
We therefore employed a coarse-grained reaction-diffusion model in which a species disperses by diffusion and replicates locally with logistic growth (the local reproduction rate increases linearly with population density, then decreases and reaches zero at the carrying capacity of the environment) except inside of obstacles. In the absence of obstacles, this model produces a traveling population wave with an exponentially shaped leading edge that moves at constant speed like the population wave resulting from the model by Yin and McCaskill [36] (see Materials and Methods for a brief discussion of the differences and similarities between our model and phage population spread). Mathematically, it is a version of the Fisher-Kolmogoroff-Petrovsky-Piscounoff equation (FKPP equation) [40–43], which captures the two processes underlying a range expansion, dispersal and growth. In our generalized version, the growth function depends on location to include the effect of obstacles. The time evolution of phage population density u(x, t) depending on location x and time t is given by:
∂ u ( x , t ) ∂ t = D eff ∂ 2 u ( x , t ) ∂ x 2 + k eff ( x ) u ( x , t ) K - u ( x , t ) , (2)
where the first term describes dispersal by diffusion with an effective diffusion coefficient Deff. The second term captures local logistic growth with reproductive rate keff(x) and constant carrying capacity K. By rescaling the phage density u(x, t), we can set K = 1 without loss of generality. In general, keff will depend on the bacterial density, the number of phage an infected bacterium releases (the burst size), the adsorption kinetics of the phage, the rate for lysis of infected host, etc. [36].
We used our data to estimate the values for the phage’s effective diffusion coefficient and effective reproductive rate on the lawn of susceptible bacteria, D ^ eff and k ^ eff, respectively. For biologically relevant initial conditions, an unimpeded, linear population front moves forward with front speed v = 2 D ^ eff k ^ eff and front width parameter ξ = D ^ eff / k ^ eff [43]. The front propagation is governed by the dynamics at the leading edge, a behavior we expect for the phage system (see Materials and Methods for a more detailed comparison to the phage system).
From Fig 2A we find that the plaque front extends with a speed of about 0.2 mm/h. With a rough estimate of the diffusion coefficient of 0.0144 mm2/h (Refs. [36, 44], Materials and Methods) we can determine an effective growth rate of k ^ eff = 0 . 7 / h for the phage in our experiments. We assume that the phage’s diffusion coefficient inside the obstacle remains the same, but that no growth is possible due to the lack of susceptible bacteria, thus allowing us to set keff(x) = 0 inside the obstacle and k eff ( x ) = k ^ eff otherwise. With the diffusion coefficient to be the same inside and outside the obstacle, individuals can diffuse into the obstacle, reminiscent of an absorbing boundary. We think this is the case in the experimental system as well, although it is possible that the effective diffusion coefficient differs slightly in the region with resistant bacteria from the region with susceptible bacteria.
We next numerically solved Eq 2 for the four different obstacles considered experimentally. Fig 3E displays two snapshots from the numerical solution of the wide obstacle (see S2 Video). To quantify front shape at the leading edge, we defined front position as the boundary at which the population density is larger than 5% of the carrying capacity (white line in Fig 3E, see Materials and Methods). Fig 3C displays the fronts. For the wide obstacle (and the three other obstacles, S5 Fig) we observe a lag of the front for the numerical solution (colored line) relative to the constant speed prediction (black line), in qualitative agreement with the experiments. This lag also manifests itself in an increased indent size (Fig 3D). To test sensitivity to the value of the diffusion coefficient, D ^ eff, we repeated the analysis for the wide obstacle with D ^ eff → 3 D ^ eff and D ^ eff → D ^ eff / 3 and found the lag to persist in both cases (S6 Fig, Materials and Methods). As expected, for decreasing D ^ eff the lag, relative to the constant speed prediction, becomes smaller. Taken together, the numerical solution of the reaction-diffusion model produces a lag similar to that seen in experiments of the phage model system (Fig 3) even though its parameters were not derived from the front’s shape.
Where does the lag originate from and under which circumstances is the constant speed model a good description? Both questions are closely related and can be explained by considering the relative importance of diffusion and movement of the front. While diffusion results in a mean distance traveled scaling with the square root of time, propagation of the front results in a position change of the edge of the front linear in time. In consequence, diffusion is the faster process at small length and time scales, while only propagation of the front leads to significant changes in population density at large length and time scales. The critical length dividing these two regimes is given by D ^ eff / v, the ratio of the diffusion coefficient D ^ eff to the speed of the advancing front v. Up to a prefactor, this ratio is given by the front width parameter ξ = D ^ eff / k ^ eff and is proportional to the width of the profile, perpendicular to the front, which is shown in Fig 3E and S2 Video [43]. Small kinks in the front will eventually be rounded and small bulges in the front smoothed out by diffusion. (We disregard possible number fluctuations at the frontier and associated possible front instabilities [45].) The process of invasion, however, plays the major role in determining front shape on length scales much larger than ξ, justifying the use of the constant speed model as an approximate, but intuitively useful model for understanding how populations spread around obstacles. For our experimental system, ξ ≈ 0.1 mm which is considerably, but not strikingly, smaller than the scale determining the shape of the obstacle (1 − 2 mm).
The simplicity of the reaction-diffusion model (Eq 2 only has two free parameters.) allows us to identify two mechanisms for the lag of the front relative to the constant speed model: a modified shape of the front close to the obstacle’s boundary (S7 Fig, panel A) and a slow-down of the front around the point of maximum width (S7 Fig, panel B); see also S1 Appendix.
First, phage particles diffuse into the obstacle, recognizable by the obstacle in Fig 3E turning yellow at its boundaries. The obstacle is therefore partially absorbing and the phage sink leads to a reduced population density close to the boundary. This flux into the obstacle does not lead to a slow-down of the overall front, since the population extends far to the sides of the obstacle. Instead, a lagging boundary layer arises whose width is of the order of the only length scale, the front width parameter ξ, and which moves at the same speed as the unperturbed front (S7 Fig, panel A). If the obstacle induces large perturbations to the front (predicted by the constant speed model), this boundary layer will not be an important component of overall front shape. If the induced perturbation is small, however, the boundary layer becomes an important constituent of overall front shape. Because our obstacles are only one order of magnitude larger than ξ, we expect the lagging boundary layer to contribute to overall front shape and thus to the observed lag. (We attribute the differing shapes of the front at the boundary layer between experiment (Fig 3A) and theory (Fig 3C) to the coarse-graining embodied in our model and differences in front detection.) This effect will be modified if diffusion into the obstacle is not possible.
Second, expansions of circular populations with radii smaller than ξ are significantly slowed down compared to linear population fronts or circular population fronts with radii much larger than ξ [43]. The constant speed model predicts that a circular segment arises with a radius initially smaller than ξ when the front passes around the point of maximum width (Fig 2H, S7 Fig, panel B). A temporary slow-down is therefore expected until the radius becomes significantly larger than ξ, leading to an apparent lag of the front close to the obstacle. In general, we expect a contribution to lagging of the front wherever the boundary of the obstacle is kinked or curved (i.e., many infinitely small kinks are present).
Both effects depend on details of the obstacle’s shape, but are tied to the length scale ξ. The perturbations predicted by the model of constant speed, however, are tied to the size of the obstacle: doubling the size of the obstacle leads to a doubling of the size of the perturbation due to the obstacle. Both effects should therefore lead to only small corrections to the front shape predicted by the constant speed model in the limit of large obstacles (large in all directions, linear size L ≫ ξ).
Since we expect the constant speed model to successfully predict the front shape for large obstacles, we can construct the front shapes for more general obstacle shapes and infer general properties of front shape that are independent of the shape of the obstacle (see below and S1 Appendix), which is not possible using experiments or numerical solutions alone.
While for rhombus-shaped obstacles the front shape is particularly simple (the front consists of two linear and two circular segments only, Fig 2H), we now consider general convex, mirror-symmetric obstacles. When the front encounters an obstacle (as when it first envelops the tip of a rhombus or the front half of a circle), the shape of the front remains planar. As the obstacle starts to decrease in width, each point along the boundary is the source of a circular segment contributing to the front (similar to Fig 2E) and the front thus encounters the obstacle at a 90° angle. Eventually, a kink or a “cusp” (a kink with infinite slope) forms on the far side (S8 Fig), which heals downstream from the obstacle. Note that when changing the size of the obstacle (without changing its shape) the front’s overall shape stays unchanged, but gets scaled by the same factor that the obstacle size increased or decreased.
In addition, as the kink heals downstream from the obstacle, we eventually recover a scaling result similar to Eq 1. In this respect, the front exhibits a universal behavior far away from the obstacle: the perturbation inherited by the front is determined by the obstacle’s width, but not by its precise shape. Some quantitative predictions of the constant speed model for isolated circular, elliptical and elongated tilted obstacles are found in S1 Appendix (S8 and S9 Figs). For objects that are not convex, we expect a similar overall behavior. An obstacle with a complicated shape still results in a kink which gradually heals as the front moves beyond the obstacle (S10 Fig, S3 Video).
We next examine how genetic composition of a population is shaped by obstacles it encounters, first predicting the obstacle’s effects based on the constant speed model followed by examining an experimental model system and simulations. As populations expand, genetic drift leads to the local reduction of genetic diversity and the formation of monoclonal sections at the front [4]. Thus we consider a population front that contains different neutral genotypes at different positions along the front encountering an obstacle.
Fig 4A displays a series of front shapes together with a simplified initial genotype distribution indicated by orange, green, cyan, blue, and red colors. In the spirit of the constant speed model, we focus on front shape dynamics alone. The front segment with the cyan genotype either cannot propagate within the obstacle or, in the case of bacteriophage T7, slows down dramatically since only diffusive motion is possible. Hence, this genotype is lost and does not contribute to the range expansion at later times. After passing the point of maximum width, the circular arcs of the front in the ‘shadow’ of the obstacle grow due to inflation and therefore genotypes marked in green and blue occupy a larger part of the front. As the kink heals, the green and blue genotypes occupy the part of the frontier that lies in the shadow of the obstacle. The abundance of these genotypes, which were a small fraction of the initial front, stays elevated even after the kink has healed. Note, however, that part of the increase in genotype abundance is transient since the arc length of the circular segments gets reduced during healing of the kink, although the radius still grows and the front thus locally experiences inflation. Fig 4A depicts a special symmetric initial condition of genotype frequencies that guarantees that genotypes benefitting from the inflation in the wake of the obstacle (green and blue genotypes) will meet precisely at the top of the obstacle. However, selectively neutral, grazing genotypes will meet at the top for quite general initial conditions, i.e., there is always a boundary that gets ‘pinned’ at the top of the obstacle.
The constant speed model argues that genotypes that fail to encounter the obstacle will be unperturbed, those whose segment of the front entirely collides with the obstacle will be eliminated, and those that graze the obstacle will be privileged because they will fill in the region downstream of the obstacle.
We tested this idea experimentally by using fluorescent proteins as labels for selectively neutral genotypes. Because we could not produce expansions with fluorescent phage, we used the expansion of three E. coli strains, which express different fluorescent proteins. Two of the strains have been characterized previously [13] and we constructed a third strain which behaves comparably for the purpose of the experiment. We created heterogeneous agar plates by adding a circular membrane with an impermeable region just below the top layer of agar. We then launched linear expansions of mixtures of the three marked strains and observed them as they grew past the circle that blocked access to nutrients (Materials and Methods, S11 Fig). Fig 4B displays the range expansion after approximately 1, 2, 6, and 10 days of growth (see S4 Video for additional time points and Materials and Methods for a description of replicate experiments).
Before the population meets the obstacle, genetic drift at the population front leads to separation into monoclonal regions of the three different colors [13, 47]. After formation of these sectors, their boundaries wander which results in a coarsening of sectors [13]. Abstracting from this effect, we observe that the sectors encountering the obstacle head-on are lost but the two that just graze the obstacle grow in its shadow, increasing the abundance of the corresponding genotypes, and meet at the top of the obstacle. These features are experimentally reproducible and verify the predictions of the constant speed model (Materials and Methods).
Next, we performed stochastic simulations, in which individuals reproduce on a lattice. In each step, a site along the front is randomly chosen and is copied onto one of the unoccupied neighbored sites thus propagating the front [5] (a variant of the Eden model [48, 49] extended here to track genotypes, see S12 Fig and Materials and Methods). Individuals never die, i.e., occupied sites never change. The obstacle is a set of lattice sites which cannot be occupied. Fig 4C shows an initially linear front encountering an obstacle. The obstacle leads to dynamics that are qualitatively similar to the bacterial range expansion described above (S5 Video). Fig 4C illustrates that individual genotypes can go extinct by two processes: the wiggling of sector boundaries caused by genetic drift [5, 13, 47] and collision with the obstacle (light green to light blue sectors, Fig 4C). The genotypes that graze the corners that define the obstacle’s width dominate the curved part of the front during the subsequent inflationary phase (green and purple sectors, Fig 4C) and meet at the top of the obstacle.
The founder effect of individuals near the point of maximum width also dominates the population’s genealogy downstream of the obstacle. Black lines in Fig 4C represent lineages of individuals at the front. As already evident from the labeling of genotypes with colors, none of the lineages pass through the area in front of the obstacle. In addition, most of the individuals at the curved part of the front originate from a small number of ancestors near the point of maximum width. Strikingly, none of these lineages pass through the point where the two populations meet behind the obstacle. Despite the expansion of the green and purple sectors right before they encounter each other, the parts of the population which meet at the top of the obstacle have no descendants at the front at late times. This effect arises because the two sectors encounter each other (almost) head-on just behind the obstacle (Fig 4C). Although this effect does not manifest itself in the sectoring pattern we deduced from the constant speed model (Fig 4A), it can be understood within the framework of the model: In Fig 2H, blue lines indicate the position of a virtual marker at the front coinciding with the overall shape of lineages behind the population front. This suggests that the constant speed model may also be used to predict the evolutionary dynamics of a spreading population in more complex environments.
In summary, we found that the constant speed model used to describe the front shape of an expanding population can be used to understand the effects of an obstacle on the diversity of neutral genotypes in an expanding population. These include the loss of genotypes encountering the obstacle head-on and a founder effect from individuals present at the point of maximum width. Since the obstacle does not affect fitness of individuals carrying specific genotypes, but in an intricate way increases random fluctuations, these effects are an example of ‘geometry-enhanced genetic drift’.
In the regime in which the constant speed model is valid, the effect of the obstacle on front shape is limited to a downstream region as wide as the obstacle and is transient due to healing of the kink (Fig 2H). If the habitat is much larger than a single obstacle, the overall front speed and shape is therefore not influenced by the presence of a single obstacle. What is the effect of many such obstacles introduced in Fig 1A? Insight can be gained by considering two obstacles which are offset and placed behind each other as displayed in Fig 5A. We focus on the population front between both obstacles arguing that in the presence of many obstacles the population encounters such pairs of obstacles subsequently. Instead of displaying the front at different time points, a blue arrow is used to indicate the path of an imaginary marker at the front which propagates with constant speed (compare to Fig 2H; the path of the marker can be derived from minimizing path length as explained in S1 Appendix on the ‘Analogy to geometrical optics’). The dashed gray arrow indicates the path of that marker in the absence of the second obstacle illustrating that the presence of the second obstacle lengthens the path and thus slows down the front between both obstacles. This effect is more readily visible in a regular pattern of rhombus-shaped obstacles (Fig 5B); the path of the virtual marker repeatedly changes direction, the speed of the front in normal direction is lower. To address the same scenario using the reaction-diffusion model, we extended our analysis of Eq 2 using the parameters employed to study the case of a single obstacle (Fig 3E). Fig 5C displays two snapshots of the numerical solution (S6 Video, Materials and Methods). Both obstacles transiently perturb the front, but not independently. Due to the first obstacle, the front reaches the right side of the second obstacle after it reaches the left side, resulting in the formation of a kink which is asymmetric and slightly shifted to the right. The front lags the unperturbed part of the front (dashed white line indicating front position at the boundary of the channel) in the wake of both obstacles, effectively resulting in a slow-down. As discussed above, for the obstacle size considered, the constant speed model is not a perfect description of front shape. The lag observed relative to the unperturbed front therefore originates from a combination of the geometrical slow-down in Fig 5A and a slow-down for reasons discussed above.
Extending our qualitative analysis of ‘geometry-enhanced genetic drift’ we repeated the stochastic simulation (Fig 4C) with two obstacles. In Fig 5D two snapshots of one realization of the simulation are displayed. Due to the stochastic nature of genetic drift, rigorously analyzing the effects of multiple obstacles on genetic diversity is not possible without detailed quantification. However, we observe two effects expected from our understanding of single obstacles. First, there is a sector boundary at the top of the second obstacle with two sectors encountering each other from opposite sides of the obstacle. Second, there is a lineage passing the first obstacle on the left and the second obstacle on the right and just grazes both obstacles (see Materials and Methods for a discussion of other instances of the simulation). These two observations illustrate two of the effects we expect many obstacles to have on genetic diversity. First, a subset of sector boundaries will be created or pinned by obstacles introducing an effective wandering of sector boundaries not arising from genetic drift at the front (the mechanism for wandering of sector boundaries in the absence of obstacles). Second, if lineages preferentially graze sides of obstacles, an effective description of the genealogy in a complex environment may be possible by considering a small subset of possible paths through the maze of many obstacles.
Organisms rarely spread across featureless habitats. Instead, they must find ways to survive and reproduce in the presence of environments that are heterogeneous in space and time. To investigate the effects of spatial heterogeneities on the dynamics and genetics of a spreading population, we combined experimental and theoretical approaches to understand the effect of single obstacles, of defined geometry, where organisms could not reproduce. When bacteriophage T7 encounters resistant E. coli the bacteriophage population front is perturbed in the wake of the obstacles by a sharp kink that slowly heals as the front moves on. A constant speed model gives an intuitive understanding of this perturbation, and a more detailed reaction-diffusion model rationalizes the deviation between experiment and the constant speed model’s predictions. In addition, the constant speed model explains that in a genetically diverse population, genotypes that run into the obstacle are eliminated and those that graze its sides increase in abundance, an example of ‘geometry-enhanced genetic drift’.
A mathematically rigorous analysis by Berestycki et al. predicted transient perturbations of planar waves encountering a single compact obstacle [50]. From a physical perspective, when the obstacle’s linear size L satisfies L ≫ ξ, where ξ characterizes the front width, considerable understanding of the perturbation is possible using a model based on front propagation locally and with constant speed. In this limit, the shape of the front can be found using a straightforward geometric construction that has an analogy in geometrical optics (S1 Appendix). Interestingly, in this regime, a linear front stays unperturbed while it envelops the obstacle, in contrast to a first intuition based on a front of fluid material encountering an obstacle such as lava flow encountering a barrier [51]. However, a front of forest fire resembles the situation of phage propagating on a lawn of bacteria; indeed, ideas very similar to the model of constant speed are used to predict forest fires [52]. Our analysis of the front predicted by the constant speed model shows that the width of the obstacle, and not its precise shape, determines the long-term dynamics of the perturbation caused by the obstacle.
The study of two obstacles placed behind each other and offset suggests an overall slow-down of the front in the presence of many obstacles. This effect is expected to depend on the density of obstacles. If obstacles are sparse, the healing of the perturbations implies that the front speed should be only marginally reduced compared to expansion in the absence of any obstacles. If obstacles are close enough to each other that the perturbation from the preceding obstacle has not healed much before the next obstacle is encountered, the perturbations will add up faster and an ensemble of obstacles will reduce front speed more. Obstacles regularly placed on a lattice are a special case: the existence of open channels, unobstructed by obstacles and much wider than the front width parameter, will allow the front to travel as fast as it would without obstacles; the remaining territory will then be explored in the wake of the front.
If the density of obstacles is so high that no free paths connect the different boundaries of the environment, the traveling wave cannot propagate around obstacles. When dispersal within obstacles is possible, the population can nevertheless expand via migration between regions with good growth conditions, which is essentially the island scenario depicted in Fig 1A. Invasion is not possible in a scenario where population spread is hindered by a connected set of impermeable obstacles (compare to the percolation threshold concept [53]).
When the size of the obstacle approaches the parameter that sets the width of the population front, the constant speed model breaks down. This regime can be understood by numerically solving a two-dimensional reaction-diffusion system (a generalized FKPP equation), which rationalizes the lag between the experimentally observed phage front and the constant speed model prediction, and bridges the gap to the regime where the length scale of the heterogeneities in the environment is much smaller than ξ and perturbations in front shape are therefore not expected.
Following these ideas will complement recent studies using reaction-diffusion models to study invasion in heterogeneous environments [20]. From the experimental side, extending the printing assay to environments with many obstacles or creating random environments by spotting a mixture of bacteria susceptible and resistant to phage onto an agar surface (S8 Video, S13 Fig, Materials and Methods) might shed light on this question in the future.
The models we used to describe the spread of phage populations were successful, even though they ignored the details of the bacteriophage life cycle. We found that for large obstacles the constant speed model is a good description for the front shape and expect in consequence the effects of ‘geometry-enhanced genetic drift’ to hold. How do these results apply to organisms whose spreading mechanism is very complex or even not well characterized? In general we expect that for other population waves than those considered here, similar considerations hold. Specifically, we expect a length scale to exist beyond which a constant speed approximation results in a good description of front shape. Thus, our findings based on the constant speed model such as universality of the shape of the population front and the genetic consequences should be applicable to population fronts with a differing underlying dynamics, including pushed fronts (fronts where the bulk of the wave and not the leading edge sets the dynamics [24]). Upon decreasing the obstacle size, we expect the constant speed model to break down and front shape and population spread to depend on the details of the biological system considered.
Similar considerations hold when the nature of the heterogeneities is changed. We here considered obstacles with vanishing growth rate. If, however, the obstacle was a region with reflecting boundary conditions, i.e., diffusive dispersal into the obstacle was not possible, we expect the behavior on large scales for large obstacles to be described by the constant speed model, while the behavior at small scales and near the boundary of the obstacle would be different.
When the obstacle strongly perturbs the shape of the population front, we predicted that these perturbations affect the fate of genotypes and lead to ‘geometry-enhanced genetic drift’. Analyzing the fate of lineages shows that the descendants of individuals trapped in front of the obstacle or born right behind it are lost in the long term. Our results are in qualitative agreement with a simulation study that demonstrated a decreased probability of survival of neutral (and deleterious) mutations occurring just in front of and right behind an obstacle [27]. Taken together, our results show that the long-term reproductive success of an individual depends on its position relative to the obstacle the population encounters as well as the random sampling that drives genetic drift, expanding the list of factors that contribute to ‘survival of the luckiest’ [54]. In addition to these effects, obstacles separating two genotypes for a considerable amount of time could also help preserve genetic diversity, similar to the mechanism of allopatric speciation. Theory and simulation, including a more detailed description of the evolutionary dynamics on top of the population dynamics is needed to disentangle these effects.
More work is also needed to understand the effects of many obstacles: Considering the effects of two obstacles placed behind each other and offset illustrates that obstacles can shape the genetic composition of a population by creating transition zones between two genotypes and constraining the spatial structure of lineages. More research is needed to illuminate how genetic drift at the population front and ‘geometry-enhanced genetic drift’ due to obstacles together shape the genetic makeup of a population.
Single obstacles could have pronounced effects on evolution beyond shaping the abundance of neutral genotypes. Because the small subpopulation that grazes the obstacle expands spectacularly, obstacles could make it easier for deleterious mutations to survive. This expansion protects deleterious mutations from extinction [55] and could establish a subpopulation which is large enough to survive for a considerable amount of time. This time span might be long enough for a second, beneficial mutation to occur, which has implications for the crossing of fitness valleys, similar to the effects due to genetic drift at the front [56]. Because the obstacle is not a population bottleneck, failure to acquire such a second mutation does not lead to a reduced fitness of the population in the long-term: genotypes that passed further away from the obstacle would eventually spread sideways and extinguish the deleterious allele if it is not rescued by a second beneficial mutation. A true spatial bottleneck would have fixed the deleterious allele and thus reduced the (absolute) fitness of the whole population. The rapid evolution of phage should allow such questions to be addressed experimentally in the future.
Higher organisms differ in two important aspects from the E. coli system and the stochastic simulation, they generally are diploid or polyploid and their population is dynamic behind the front. While the consequences of obstacles on diploid organisms undergoing recombination are an important area for future research, our results are relevant for the evolutionary dynamics of mitochondrial DNA carried by diploids. Gene flow behind the front will blur the sector boundaries which are frozen in both our experiments and the stochastic simulation. However, this diffusive blurring is slow (scaling as the square root of time since it is a diffusive process) while the front advances more rapidly (linearly with time). Hence, the boundaries remain well-defined for some distance behind the wave [57].
This study focused on a regime where the front dynamics can be described by a model of constant speed. In this regime, the results are insensitive to details of the expansions and details of the obstacle shape. Assuming that the population front is subdivided into monoclonal regions, an effect of ‘geometry-enhanced genetic drift’ can be described which is closely connected to the dynamics of front shape. We believe that these findings carry over to a wide variety of population expansions and beyond the neutral evolutionary dynamics considered here. Finally, although our analysis focused on single obstacles, we believe that our findings can be extended to natural environments, which typically display more complex heterogeneities. As a first step, by using the findings for isolated obstacles we expect to be able to describe observables such as the effective front speed and an ‘effective genetic drift’ in environments with obstacles such as those displayed in Fig 1A.
A semi-automated image analysis pipeline was used to extract front shapes, front positions, and indent sizes (such as in Figs 2A, 2B, 3A and 3B) from the fluorescence time-lapse information. First, the channel detecting YFP fluorescence was used to define a front right after the plaque boundary got established and the channel detecting mCherry was used to identify the three upper corners of the obstacle. This information was used to define a coordinate system with the obstacle’s center at the origin and the front extending in y-direction (referred to as ‘upper region’ in the following, e.g., Fig 1D). The image was cropped (5.3 mm in direction of front movement, 0.7 mm in direction opposite to front movement, and 3 mm to either side of the obstacle). After normalization using the upper, uninfected region, the difference between two consecutive frames (YFP channel) was used to identify the front. In the difference image the extending front manifests itself as a bright region whose upper boundary was identified using thresholding. The algorithm was tested manually since the front is easily detectable by eye, although the decay in fluorescence extends to about 1 mm (S2 Fig). A few frames were excluded from the analysis due to jumps in the front which could be detected automatically using a threshold for local slope of front shape. Finally, front position was determined from the curve of the front close to the boundaries of the cropped region. Indent size was derived as the distance between the most lagging part of the front and front position (after the kink has formed, i.e., curve of front was defined around the axis of bilateral symmetry). The corners of the obstacles identified were also used to identify the size of obstacles, which were slightly smaller than in the printing template. When displaying data for individual obstacles the median of all the obstacles included in the analysis was used; for collapse plots, the size of each single obstacle was used to rescale data. For analysis, front detection was limited to frames obtained within 22 h even if the experiment lasted longer. After this time front detection becomes more challenging most likely due to the bacterial lawn transitioning into stationary phase.
|
10.1371/journal.pcbi.1005915 | MPLasso: Inferring microbial association networks using prior microbial knowledge | Due to the recent advances in high-throughput sequencing technologies, it becomes possible to directly analyze microbial communities in human body and environment. To understand how microbial communities adapt, develop, and interact with the human body and the surrounding environment, one of the fundamental challenges is to infer the interactions among different microbes. However, due to the compositional and high-dimensional nature of microbial data, statistical inference cannot offer reliable results. Consequently, new approaches that can accurately and robustly estimate the associations (putative interactions) among microbes are needed to analyze such compositional and high-dimensional data. We propose a novel framework called Microbial Prior Lasso (MPLasso) which integrates graph learning algorithm with microbial co-occurrences and associations obtained from scientific literature by using automated text mining. We show that MPLasso outperforms existing models in terms of accuracy, microbial network recovery rate, and reproducibility. Furthermore, the association networks we obtain from the Human Microbiome Project datasets show credible results when compared against laboratory data.
| Microbial communities exhibit rich dynamics including the way they adapt, develop, and interact with the human body and the surrounding environment. The associations among microbes can provide a solid foundation to model the interplay between the (host) human body and the microbial populations. However, due to the unique properties of compositional and high-dimensional nature of microbial data, standard statistical methods are likely to produce spurious results. Although several existing methods can estimate the associations among microbes under the sparsity assumption, they still have major difficulties to infer the associations among microbes given such high-dimensional data. To enhance the model accuracy on inferring microbial associations, we propose to integrate multiple levels of biological information by mining the co-occurrence patterns and interactions directly from large amount of scientific literature. We first show that our proposed method can outperform existing methods in synthetic experiments. Next, we obtain credible inference results from Human Microbiome Project datasets when compared against laboratory data. By creating a more accurate microbial association network, scientists in this field will be able to better focus their efforts when experimentally verifying microbial associations by eliminating the need to perform exhaustive searches on all possible pairs of associations.
| Microbes play an important role both in environment and human life. However, the way microbes affect the human health remains largely unknown. Knowledge of the microbial interactions can provide a solid foundation to model the interplay between the (host) human body and the microbial populations; this can serve as a key step towards precision medicine [1]. Unfortunately, understanding microbes interactions is difficult, as most microbes cannot be easily cultivated in standard laboratory settings. However, the recent increase of quality and reduced costs of sequencing technologies (e.g., shotgun or PCR directed sequencing [2]) enable researchers to collect information from the entire genome of all microbes under different environment conditions. As a result, various datasets ranging from earth ecosystem to human microbiome have been made publicly available under the Human Microbiome Project [3] or the Earth Microbiome Project [4].
In this paper, we aim at analyzing the networks of associations (putative interactions) among the microbes of human microbiome in order to understand how microbes can affect the human health. To this end, there exist several challenges: First, the amount of sequenced data that corresponds to human microbiome available from public websites is scarce. To date, one of the largest metagenomic datasets of human niches is the NIH Human Microbiome Project (HMP) [3] which only provides a few hundreds of healthy individual samples (n) of various body sites, while the number of measured microbes (p) usually ranges from hundreds to thousands. As a consequence, the number of associations (p(p − 1)/2) is much greater than the number of samples (i.e., high-dimensional data). Another big challenge stems from the nature of the data itself. Sequencing data only provides the relative abundance of various species; this is because the sequencing results are a function of sequencing depth and the biological sample size [5]. Therefore, from a statistical standpoint, the relative taxon abundance falls into the class of compositional data [6]; this causes statistical methods such as Pearson or Spearman correlations (which work with absolute values) to generate spurious results.
To infer microbe associations for both compositional and high-dimensional data, several algorithms have been developed. A pioneering method called SparCC [7] applies log-ratio transform on compositional data and directly approximates the correlation among microbes based on sparsity assumption of microbial associations. However, SparCC does not consider the influence of errors in compositional data; this may reduce the correlation estimation accuracy. More precisely, SparCC approximates the basis variance (i.e., the variance of compositional data) under the assumption that average correlations are small. Second, the iterative procedure used to estimate the magnitude of correlations can exceed value 1; this may cause poor approximations if one tries to remedy the problem by setting up the threshold value to 1 or -1 for the estimated correlations; these series of approximations may reduce the correlation estimation accuracy quite significantly. SPIEC-EASI [8] calculates the covariance of the log-ratio transformed data to approximate the covariance of the absolute abundance of microbes; then, it uses either neighborhood selection (mb) [9] or graphical Lasso (gl) [10] to estimate the conditional dependencies among microbes. CCLasso [11] is similar to SPIEC-EASI which applies log-ratio transform on compositional data and imposes a l1 penalty on the inverse covariance matrix of the microbes and then solves it to obtain a sparse covariance matrix. However, it is not clear whether or not CCLasso can obtain a consistent estimator on the inferred microbial covariance without showing consistency analysis (see consistency analysis for graphical Lasso in S1 File section 8).
We note that although the above methods can estimate the covariance among microbes under the sparsity assumption, they still have major difficulties to infer the associations among microbes given such high-dimensional data. To solve the problem caused by high-dimensional data, we propose to integrate multiple levels of biological information to enhance the model accuracy on inferring microbial associations. Indeed, an increasing amount of scientific literature provides a large amount of data which can be mined not only for the co-occurrence of microbes, but also to predict microbes associations directly. For instance, pioneering work [12] considers automated analysis of the co-occurrence of bacterial species through statistical testing approaches (e.g., Fisher’s exact test). Recently, Lim et al. [13] incorporated machine learning techniques to automatically identify and extract microbial associations directly from the abstracts of scientific papers. Finally, Wang et al. [14] and Li et al. [15] use prior biological knowledge to reconstruct genes interaction networks.
To the best of knowledge, we are the first to consider experimentally verified biological knowledge as a priori information to derive microbial association networks. To this end, we transform the original problem of microbial associations estimation into a graph structure learning problem where nodes represent microbes and edges represent (pairwise) associations among microbes. With this new problem formulation, the graphical Lasso algorithm becomes suitable to infer the microbial association network. We also integrate the text mining results from the scientific literature as prior knowledge for inferring the microbes graph structure; the proposed algorithm Microbial Prior Lasso (MPLasso) turns out to be more accurate than other existing methods on inferring the microbial associations. The proposed MPLasso pipeline is shown in Fig 1.
We assess the performance of MPLasso in the presence of prior knowledge by first comparing it against other previously proposed methods (e.g., CCLasso, REBACCA [16], SparCC, SPIEC-EASI, and CCREPE [17]) through synthetic data generated from different graph structures (run time comparisons of existing methods are summarized in S1 File section 1 and S1 Table). We show that our proposed MPLasso outperforms all these methods in terms of area under the precision-recall curve (AUPR) and accuracy (ACC) of network associations prediction. Next, we evaluate the HMP datasets of two different sequencing techniques (shotgun and 16S ribosomal RNA (rRNA)) at five different body sites and compare the reproducibility of the estimated results. Taken together, our contributions are three fold:
In this paper, we consider high-throughput comparative metagenomic data obtained from the next-generation sequencing (NGS) platforms. More specifically, two types of gene sequencing data are considered: 16S rRNA and shotgun data. Shotgun data analyses are accomplished by unrestricted sequencing of the genome of all microorganisms present in a sample; on the contrary, the domain of 16S rRNA is restricted to bacteria and archaea. Data obtained from the human microbiome project (HMP) have a curated collection of sequence of microorganisms associated with the human body from both shotgun and 16S sequencing technologies.
For the 16S rRNA data, we consider the high-quality sequencing reads in 16S variable regions 3-5 (V35) of HMP healthy individuals from Phase one production study (May 1, 2010). The taxonomy classification of the 16S rRNA are performed using either mothur (HMMCP) [18] or QIIME (HMQCP) [19] pipelines. The resulting table for operational taxonomic units (OTUs) at each body site of the human samples can be obtained from http://hmpdacc.org/HMMCP/ and http://hmpdacc.org/HMQCP/. For the shotgun data (HMASM), we obtain data from http://hmpdacc.org/HMASM/ and use the trimmed sequences as inputs to the metaphlan2 [20] pipeline which can generate the OTU abundance for each sample.
The OTU table can be represented by a matrix D ∈ ℕn×p where ℕ represents the set of natural numbers. d i = [ d 1 i , d 2 i , … , d p i ] denotes the p-dimensional row vector of OTU counts from the ith sample (i = 1, …, n). To account for different sequencing depths for each sample, the raw count data (di) are typically transformed into relative abundances (x) by using log-ratio transform [6]. Statistical inference on the log-ratio transform of the compositional data (x) can be shown to be equivalent to the log-ratio transform on the unobserved absolute abundance (d) as: log ( x i x j ) = log ( d i / m d j / m ) = log ( d i d j ) . Here, we apply the centered log-ratio (clr) transform as follows:
c = clr ( x ) = [ log ( x 1 m ( x ) ) , log ( x 2 m ( x ) ) , … , log ( x p m ( x ) ) ] (1)
where m ( x ) = ( ∏ i = 1 p x i ) 1 p is the geometric mean of the composition vector x. The resulting vector c is constrained to be a zero sum vector.
The covariance matrix of the clr transform C = Cov[clr(c)] can be related to the covariance matrix of the log-transformed absolute abundances Γ = Cov[log D] via the relationship [6, 8] C = UΓU, where U = I p - 1 p J, where Ip is the p-dimensional identity matrix, and J is the p-dimensional all-ones vector. For the case where p > > 0, the finite sample estimator (C ^) serves as a good approximation of Γ ^; therefore, the finite sample estimator (C ^) serves as the basis on inferring the correlations among microbes. To account for the zero counts in samples, we add pseudo count to the original count data to avoid numerical issues when using the clr transform.
To infer the pairwise associations among microbes, we can transform the original inferring problem into a graph learning problem where each node represents an OTU (e.g., taxon) and each edge represents a pairwise association between microbes; the resulting graph is an undirected graph G = ( V , E ), where V and E represent the node and edge sets, respectively.
Suppose the observed data (d) are drawn from a multivariate normal distribution N(d|μ, Σ) with mean μ and covariance Σ. The inverse covariance matrix (precision matrix) Ω = Σ−1 encodes the conditional independence among nodes. More specifically, if the entry (i, j) of the precision matrix Ωi,j = 0, then node i and node j are conditionally independent (given the other nodes) and there is no edge among them (i.e., Ei,j = 0).
However, microbial data usually come with a finite amount of samples (n) but with high dimensionality (p); this makes the graph inferring problem intractable since the number of variables (p ( p - 1 ) 2) is greater than n. To solve this problem, an important assumption that needs to be made is to assume that the underlying (true) graph is reasonably sparse. One suitable algorithm to select the precision matrix under sparsity assumption is to utilize the graphical Lasso proposed previously [8, 10].
As shown in Fig 2, we propose to utilize the information obtained from the scientific literature in order to construct the prior matrix P ∈ ℝp×p, where each entry Pi,j ∈ [0, 1] represents the prior probability of associations between taxon i and taxon j. We can impose different amounts of penalties on the precision matrix; this is different from the standard formulation where the penalty (ρ) imposed on the precision matrix is the same. Therefore, by incorporating the prior information into the penalty matrix (P), the proposed MPLasso can be formulated as follows:
Ω ^ = arg max Ω { log det ( Ω ) - tr ( Ω C ^ ) - ρ | P ⊗ Ω | 1 } (2)
where C ^ is the empirical covariance of the microbial data, and Ω is the precision matrix of the estimated associations among microbes. Here det and tr denote the determinant and the trace of a matrix, respectively. |Ω|1 is the L1 norm, i.e., the sum of the absolute values of the elements of Ω and ⊗ represents the component-wise multiplication. When the value of Pi,j is large, this directly puts a heavy penalty and represents a weaker association between taxa and vice versa. This way, by imposing the prior information, we can accurately infer the associations among microbes.
We extract two types of data to be used as priors for our model. One type of data is from the microbial co-occurrence in literature that examines the number of abstracts where two taxa appear together and compares this to random chance. The second type of data is from the machine learning-based method that extracts the full details of the interaction, including the sign and direction of the interaction.
To acquire the prior knowledge (P) of microbial associations from reported experiments and published papers, we utilize the PubMed database (https://www.ncbi.nlm.nih.gov/pubmed/) that contains a massive amount of papers with abstracts. For the 16S rRNA data where the taxonomy level can only be achieved at the genus level, we adopt the statistical testing method (i.e., Fisher’s exact) [12] to identify the pairwise associations derived from the microbial co-occurrence in literature. On the other hand, for the shotgun data where the taxonomy level can be up to species level, we adopt both the microbial co-occurrence in literature and the machine-learning-based methods [13] to obtain such associations.
We modify the code available on https://github.com/CSB5/atminter that utilizes the Entrez search system to query all the possible combinations of taxon-taxon pairs from the data. More specifically, the query “taxon i AND taxon j” for genus (species) level are performed on PubMed database in order to obtain the number of papers that corresponds to this query term. Acquisitions of abstract’s content follow a similar way where the query term follows the format “species i AND species j” for each pair of species. Note that, all text-mining procedures are completely automated; that is, users only need to specify the species pairs and the tool will extract the information automatically (and comprehensively) from the PubMed database.
To select the optimal penalty parameter (ρ), we use the Bayesian information criterion (BIC) [23] which is a standard method for model selection. The BIC for Gaussian graphical models takes the form:
B I C = - 2 l n ( Ω ) + | E | log ( n ) (3)
where |E| is the number of edges in the association network, n is the sample size, and l n ( Ω ) = n 2 [ log ( det ( Ω ) ) - tr ( Ω C ^ ) ]. Based on (3), we choose ρ that minimizes BIC.
To show the effectiveness of our proposed model, we first compare our model against several state-of-the-art models: CCREPE, SparCC, REBACCA, CCLasso, SPIEC (mb) and SPIEC (gl). All these codes have been implemented using the R language. We set up p-value at 0.05 for CCREPE and the threshold of correlation for SparCC at 0.1 (see S1 File section 1 for precise simulation settings for each algorithm).
For MPLasso in real datasets, the true underlying network is only partially known and contains spurious information. To assess our algorithm performance with imperfect prior information, we consider prior information with different precision levels, where the precision level is defined as the number of true edges over the total number of edges in the prior information. The total number of edges in the prior network is set to be equal to the number of edges in the true underlying network. Therefore, a precision level of 0.1 indicates that 10% of the edges in the prior network are true edges, whereas the other 90% are spurious ones (see S1 File section 7 for details of introducing priors). We report the results we obtained for 0.5 precision level in the synthetic experiments while more results for different precision levels can be found in S1 File section 3 and S2 Fig.
Emboldened by the success of our proposed algorithm on synthetic data, we have applied MPLasso to infer the associations among microbes for HMP data. Acquisitions and preprocessing for both 16S rRNA and shotgun sequencing data are described in Material and methods section. We report the same three body sites (i.e., buccal mucosa, supragingival plague, and tongue dorsum) of each pipeline and filter out OTUs that appear in less than 10% of total samples—two more body sites (i.e., stool and anterior names) are reported in S1 File section 5 and S8 Fig. The total number of samples and OTUs are summarized in Table 2 and S8 Table.
We use the clr transformation in (1) and add pseudo count 0.1 to all the samples, then normalize the counts to get compositional data. However, there is no true correlation network of taxon-taxon associations in real data as opposed to synthetic data. To assess and compare the performance among different methods in real data experiments, we measure the reproducibility of the resulting networks. More specifically, we define the “gold standard” network as the one that uses the full dataset. The reproducibility is defined as the number of edges that had been correctly estimated when using only half of the samples in the full dataset compared to the “gold standard” network. We randomly select half of the samples in the full dataset of each body site and then average over 20 independent simulations. We compare the reproducibility of the MPLasso against SPIEC (gl) which has a better performance than other existing algorithms on synthetic datasets as well as CCLasso which has a better performance than other correlation based methods in [11].
The reproducibility results are summarized in Table 2. MPLasso has a better reproducibility over SPIEC (gl) and CCLasso; this implies that MPLasso is not only more robust, but also more accurate at inferring edges. We also consider reproducibility on different percentages of highly connected nodes in S9 Table. Only when we consider as little as only 25% of high degree nodes, CCLasso has a better performance (but even so for 2% only, on average).
We also summarize the statistics of the non-associated pairs found by the Fisher’s exact test, potential associated pairs, associated pairs found by MPLasso, and recovered associated pairs in S10 Table (see also S10 Fig and S1 File section 9). As shown, the known associations obtained from Fisher’s exact test is around 50% over all possible pairs of associations (i.e., around 50% prior information). The recovery rate of associated pairs of MPLasso is around 80%. For comparison, we also include the recovery rate of associated pairs for CCLasso and SPIEC (gl) algorithms in S11 Table. As we can see by comparing the S10 and S11 Tables, CCLasso tends to discover more edges than MPLasso and SPIEC (gl). Although CCLasso can obtain similar results on the recovery rate of associated pairs, it does not perform as well as MPLasso when considering the recovery rate (i.e., reproducibility) of both associated and non-associated pairs (see Table 2). In other words, CCLasso finds a greater amount of false positive taxa pairs when compared to MPLasso; this is evaluated through AUPR in synthetic experiments shown in Fig 3.
To compare the estimated association networks at each body site for different pipelines (i.e, HMASM, HMMCP and HMQCP), we select the “top players” (i.e., high degree nodes) and arrange them using a counterclockwise layout as shown in Fig 5. For the genus level data, since we only utilize the Fisher’s exact test (that only requires the information of the number of abstracts), we can use contents of published scientific literature to validate the inferred associations. In contrast, for the species level data, the machine learning-based approach has already used the contents of abstract to obtain the prior information; therefore, it is inappropriate to use any papers that appear in the PubMed search results to validate the inferred associations. To circumvent the potential circular validation, we only use the scientific literature that has not yet been used to create the prior information.
For the buccal mucosa (BucMuc), the association pair 〈Streptococcus mitis, Actinomyces naeslundii〉, which was found in HMASM (Fig 5(a)), has been shown to have associations [26]. Additionally, the associations are also detected at genus level data as shown in Fig 5(b). Note that the top degree nodes in HMMCP and HMQCP has 70% in common (i.e., belongs to same genus) which implies that the microbial composition of BucMuc is relatively robust.
For the supragingival plague (SupPla), the “top players” in species level data (Fig 5(d)) mainly come from two genera: Actinomyces and Prevotella which can be widely found in SupPla and also correspond well with the HMMCP dataset (Fig 5(e)). Similarly, the species level associations in tongue dorsum (TonDor) is dominated by Actinomyces as shown in Fig 5(g); this is because Actinomyces possess 10 different strains out of the total 103 taxa, yet this does not imply that all members of a particular genus group should be associated. Although not seen in Fig 5(h) and 5(i), genus Actinomyces is also a high degree node in the association network of the genus data.
One noticeable observation in the species level dataset (HMASM) is that the same genus belongs to the same community which means that edges are mostly found within OTUs of the same taxonomic group. This phenomenon is called assortativity and it has been widely observed in other microbial network studies [17]. However, this does not imply that all members of the same taxon should be ecologically associated. To quantify the similarity of high degree nodes that are found both in HMMCP and HMQCP datasets, we compute the correlation between node degrees at different body sites by utilizing the Spearman correlation method (see S1 File section 6). We found that TonDor has lower correlations (∼0.5) than other body sites (∼0.7); this can be directly observed from Fig 5(h) and 5(i) that have a few high degree genera in common.
Inferring associations (putative interactions) among microbes and understanding their influence on the human body is an important step towards precision medicine. Advancements of high-throughput sequencing techniques enable us to gather metagenomic sequence data from different environment and human niches. The available high-throughput experimental data, however, are compositional and high-dimensional in nature.
Existing microbial network inferring methods focus on inferring the compositional data and use the graph sparsity assumption to overcome problems caused by high-dimensional data. However, all of these approaches do not consider the information that can be obtained from the scientific literature to directly describe the associations among microbes or their co-occurrence. By integrating multiple levels of biological information into the statistical models, we have shown that one can dramatically increase the model accuracy and edges recovery rate. To the best of knowledge, this is the first work to propose this automated pipeline to infer the associations on microbial data, show its feasibility, and measure performance metrics on both synthetic and real datasets.
We have also shown that our proposed algorithm Microbial Prior Lasso (MPLasso) outperforms all other existing methods when using synthetic data with different graph structures which simulate different levels of sparsity. We have evaluated several combinations of sample sizes and number of taxa to demonstrate the applicability of our approach under different conditions and suggest rough guidelines for requisite sample size for the real data for the given assumption of the underlying graph structures.
Additionally, the use of prior information does not dominate the inferred results. Indeed, as summarized in S10 Table, the prior information obtained by the microbial co-occurrence in literature is only used to restrict the search space in order to infer associations that are more plausible (i.e., more likely to be associated) than other candidate pairs of associations. More specifically, we first calculate the probability of association among taxa. Next, if two taxa are not associated, we penalize the associations among these two taxa when solving MPLasso. Consequently, MPLasso can effectively select taxa that are highly associated with high statistical confidence. In this respect, prior information will not dominate the results, but rather improve the algorithm’s accuracy and robustness.
Our analyses on different levels of real HMP data show that MPLasso achieves better reproducibility than SPIEC (gl) and CCLasso; we have also found the assortativity at the species level data (HMASM) at different body sites. In other words, OTUs are more likely to interact with other phylogenetically related OTUs. Additionally, the detected genera at genus level (HMMCP and HMQCP) datasets show high correlations based on their node degrees (i.e., number of edges a node has to other nodes). Those high degree nodes (i.e., “top players”) have been found experimentally as being ubiquitous at each body site; this confirms that MPLasso can accurately detect the “top players” and even correctly infer the associations among them. The resulting microbial association network can suggest credible directions for experimentalists to validate the results without exhausting search for all possible associations.
Recent studies report that people affected by microbiome related diseases show different microbiome profiles when compared to healthy individuals. For example, results show that individuals affected by the inflammatory bowel disease (IBD) have (30-50)% percent less biodiversity of commensal bacteria (e.g., Firmicutes and Bacteroidetes) when compared to healthy individuals. Another example shows that individuals with Type 2 diabetes (T2D) exhibit significant changes in 190 microbial OTUs, with particularly high abundance of Enterobacteriaceae compared to healthy individuals [27]. Therefore, by creating a more accurate microbial association network, scientists working in this field will be able to accurately identify the relationship between microbiome related diseases (such as T2D) and groups of taxa based on the inferred network. This way, scientists can develop new drugs or use probiotics to directly target identified groups of taxa.
Finally, the estimated microbial association networks of the real datasets can be used to understand why and how various eco-systems evolve over time. Recent studies use association networks to fit dynamic models, e.g., differential equation-based model of gut microbiome evolution of mice [28]. These microbe associations represent the putative microbial interactions that provide partial information about the true interaction network. Therefore, by incorporating the association network as additional information, we may be able to infer the microbial interaction networks more accurately [29]. Overall, MPLasso shows promising results and outperforms state-of-the-art methods. In the present framework, our proposed MPLasso creates the inferred association network to provide additional partial information; this can be useful to reveal the underlying dynamics (i.e., interactions) of microbial communities. However, MPLasso was not tested on a dynamic model of microbial communities. Inferring the dynamics or interactions among microbial communities would require a new algorithm which is left as future work.
The MPLasso R package can be downloaded from here https://github.com/ChiehLo/MPLasso_RPackage
|
10.1371/journal.ppat.1006178 | A single mutation in the envelope protein modulates flavivirus antigenicity, stability, and pathogenesis | The structural flexibility or ‘breathing’ of the envelope (E) protein of flaviviruses allows virions to sample an ensemble of conformations at equilibrium. The molecular basis and functional consequences of virus conformational dynamics are poorly understood. Here, we identified a single mutation at residue 198 (T198F) of the West Nile virus (WNV) E protein domain I-II hinge that regulates virus breathing. The T198F mutation resulted in a ~70-fold increase in sensitivity to neutralization by a monoclonal antibody targeting a cryptic epitope in the fusion loop. Increased exposure of this otherwise poorly accessible fusion loop epitope was accompanied by reduced virus stability in solution at physiological temperatures. Introduction of a mutation at the analogous residue of dengue virus (DENV), but not Zika virus (ZIKV), E protein also increased accessibility of the cryptic fusion loop epitope and decreased virus stability in solution, suggesting that this residue modulates the structural ensembles sampled by distinct flaviviruses at equilibrium in a context dependent manner. Although the T198F mutation did not substantially impair WNV growth kinetics in vitro, studies in mice revealed attenuation of WNV T198F infection. Overall, our study provides insight into the molecular basis and the in vitro and in vivo consequences of flavivirus breathing.
| Flaviviruses include emerging pathogens such as WNV, DENV, and ZIKV that threaten global health. Despite causing significant morbidity, effective vaccines or therapeutic agents to protect humans against many flaviviruses are lacking. Because of the importance of antibodies in flavivirus immunity and vaccine protection, much effort is focused on understanding the factors that modulate antibody recognition of flaviviruses. Virus breathing, which allows viruses to sample different conformations at equilibrium, has the potential to transiently expose otherwise inaccessible antibody epitopes. Here, we report the identification a single mutation in the envelope protein that alters the exposure of a poorly accessible epitope and the stability of both WNV and DENV through changes in the ensemble of structures sampled by the virus. For WNV, this change attenuated infection and pathogenesis in mice, suggesting that virus conformational dynamics have relevant consequences in vivo.
| Flaviviruses are enveloped, positive-stranded RNA viruses typically transmitted to humans via infected ticks or mosquitoes. As many members of the flavivirus genus are emerging, they constitute a significant threat to global health. For example, approximately 390 million humans worldwide are infected annually with one of the four serotypes of dengue virus (DENV) [1]. West Nile virus (WNV) was introduced into North America in 1999 [2] and rapidly became the leading cause of arbovirus-related encephalitis in the United States [3]; Zika virus (ZIKV) emerged from Asia and Africa for the first time in 2007 and has since caused epidemics in French Polynesia [4], Oceania [5], and most recently, the Americas [6, 7]. Despite causing significant morbidity, licensed vaccines or therapeutic agents to protect humans against many flaviviruses are lacking. However, highly effective vaccines for some flaviviruses such as yellow fever virus, Japanese encephalitis virus, and tick-borne encephalitis virus are in use. The induction of a neutralizing antibody (NAb) response is a correlate of protection for these vaccines [8–12]. While a live-attenuated tetravalent DENV vaccine was recently licensed, its efficacy and durability varied by DENV serotype, pre-existing flavivirus immune status, and age of vaccine recipient [13–15]; the relationship between neutralization titer and protection for this vaccine is less clear. Because of the importance of antibodies for flavivirus immunity, a detailed understanding of flavivirus antigenic structure as well as the mechanisms of antibody-mediated neutralization is critical [16].
Assembled flavivirus particles are composed of three structural proteins: capsid (C), pre-membrane (prM), and envelope (E). The E protein, which consists of three structural domains (DI, DII, and DIII) connected to the viral membrane via a helical anchor, has critical roles in directing both the assembly of virions and their entry into cells. Flexible hinges between E protein domains enable conformational changes necessary for many steps of the viral life cycle, including fusion and maturation [17, 18]. Flaviviruses bud into the lumen of the endoplasmic reticulum as immature, non-infectious particles with a spiky surface composed of 60 icosahedrally arranged prM-E heterotrimers [19, 20]. During virus egress through the acidic environment of the trans-Golgi network, conformational changes in E expose a cleavage site within prM, which is recognized by host furin-like proteases. Cleavage of prM in the trans-Golgi network and release of the pr peptide in the extracellular environment give rise to mature and infectious virus particles covered by antiparallel E homodimers. NAbs can target epitopes in all three E protein structural domains and in quaternary structures composed of multiple domains within or across E dimers [21–31], and may block multiple steps in the virus entry pathway [32–35].
Cryo-electron microscopic (cryo-EM) reconstructions of DENV [36, 37], ZIKV [38, 39], and WNV [40] have provided detailed models for the structure of mature virions on which 90 E dimers lie flat against the viral membrane in a herringbone pattern. However, several lines of evidence suggest that many infectious flaviviruses exist as structures beyond those captured by high-resolution cryo-EM reconstructions [16]. Flaviviruses are released from infected cells as a heterogeneous population containing varying amounts of uncleaved prM due to incomplete maturation [41]. While the structures of these partially mature virions are not fully defined, they appear to contain regions that display mature- and immature-like arrangements of E proteins in varying proportions [42]. The presence of uncleaved prM on virions modulates the accessibility of many E protein epitopes recognized by NAbs [43–47]. Additionally, prM retention on virions allows for recognition by prM-reactive antibodies incapable of efficiently neutralizing infectivity. These prM-specific antibodies can enhance DENV infection and potentially contribute to severe clinical disease [48–52].
Flavivirus heterogeneity also arises from conformational flexibility of viral proteins that allows virions to sample an ensemble of conformations at equilibrium [53]. As with changes in virus maturation efficiency, virus conformational dynamics or ‘breathing’ has the potential to modulate antibody recognition and potency. Prolonged virus-antibody incubation reveals time-and temperature-dependent changes in antibody potency, the degree of which correlates generally with predictions of epitope accessibility on the mature virion; the neutralization potency of antibodies targeting cryptic epitopes can be enhanced more significantly compared to antibodies targeting highly accessible epitopes [54]. In this context, prolonged incubation of WNV and DENV in the absence of antibody does not irreversibly render them more sensitive to neutralization, suggesting that antibody binding may stabilize transiently sampled virion conformations [47].
The role of virus breathing in modulating antibody recognition is exemplified by the DENV1-specific human monoclonal antibody (mAb) E111, which neutralizes two DENV1 strains, Western Pacific (WP) and 16007, with a ~200-fold difference in potency [55]. Structural analyses revealed that E111 binds to an epitope in DIII that is predicted to be inaccessible based on existing cryo-EM models. The strain-dependent neutralization potency of E111 was neither explained by antibody binding affinity nor sequence variation within the epitope. Instead, a single residue in DII outside of the antibody footprint that differed between DENV1 WP and 16007 was responsible for modulating sensitivity to neutralization by E111 [56]. Thus, natural variation at this residue regulates the conformational dynamics of DENV1 in a way that affects exposure of the distal E111 epitope.
The determinants and functional consequences of a dynamic virion are poorly understood. In this study, we describe a single residue within the E protein DI-DII hinge that alters the neutralization sensitivity and stability of WNV and DENV virions through changes in conformational dynamics. Mutation at this residue in the WNV E protein attenuated infection and pathogenesis in mice, suggesting that changes in virus breathing have relevant consequences in vivo.
To identify epitopes targeted by NAbs in flavivirus-immune sera, we created a library of WNV reporter virus particles (RVPs) [57] containing mutations at solvent-accessible E protein residues for use in neutralization studies. We focused on residues within the DI-DII hinge region, which has been shown to be an important target for many potently neutralizing antibodies against both WNV and DENV [23, 24, 33]. Because virion maturation state and conformational dynamics may modulate epitope accessibility and neutralization sensitivity indirectly, we performed a series of control experiments, as described previously [26], to identify pleiotropic effects of mutations on the overall antigenic structure of virions. These experiments identified a mutation at E residue 198 (T198F, Fig 1A) that unexpectedly modulated sensitivity to neutralization by the DII-fusion loop (DII-FL)-reactive mAb, E60, despite being located distally from the predicted epitope [58]. RVPs incorporating T198F were 68-fold more sensitive to E60 neutralization than were wild type (WT) RVPs (average EC50 = 13.1 and 881 ng/ml for T198F and WT RVPs, respectively, p< 0.0001, Fig 1B and 1D). In contrast, there was little difference in sensitivity to neutralization by mAb E16 (average EC50 = 5.9 and 8.1 ng/ml for T198F and WT RVPs, respectively, Fig 1C and 1E), which binds an accessible epitope on the DIII lateral ridge (Fig 1A) [59–62].
Neutralization studies with antibodies that bind poorly accessible epitopes such as those in DII-FL [46, 58, 63] often reveal incomplete neutralization even at saturating antibody concentrations [45]. As flavivirus neutralization occurs when the number of antibodies bound to the virion exceeds a stoichiometric threshold [62], incomplete neutralization may reflect structural heterogeneity among a genetically homogeneous RVP population that limits epitope accessibility. Virions that are not bound by antibodies at a stoichiometry sufficient for neutralization remain infectious [62]. We noted that a significantly smaller fraction of T198F RVPs remained infectious at the highest E60 concentration tested (10 μg/ml) compared to WT (5.5% and 26.7% for T198F and WT RVPs, respectively, p = 0.001, Fig 1B). These results demonstrate that the T198F mutation increases the accessibility of a cryptic epitope in DII-FL.
The neutralizing activities of many mAbs, including those targeting DII-FL, can be modulated by the efficiency of virion maturation [45, 46, 64]. Specifically, virions that retain uncleaved prM often are more sensitive to neutralization by antibodies that bind epitopes predicted to be poorly accessible on the mature virion. For example, increasing the efficiency of prM cleavage (Fig 2A) reduced the sensitivity of WT WNV RVPs to neutralization by E60 relative to a standard WT RVP preparation (~70-fold reduction in EC50, p = 0.02, Fig 2B and 2C), as detailed previously [26, 45, 47]. To investigate whether T198F modulates the efficiency of prM cleavage during virion maturation, we analyzed the prM content of RVPs produced using standard conditions (Std-RVP) and in the presence of over-expressed human furin (Furin-RVP). Although three independent preparations of Std T198F RVPs contained an average of five times the level of uncleaved prM compared to that of WT RVPs prepared in parallel, furin over-expression in RVP producing cells resulted in efficient prM cleavage for both WT and T198F RVPs (Fig 2A), suggesting that the increased prM content of Std T198F RVPs was not due to an inability of the virion to adopt conformations in which the prM cleavage site was accessible during egress. Moreover, increased prM content was not sufficient to explain the greater sensitivity of T198F RVPs to neutralization by E60, as Furin T198F RVPs were more sensitive to neutralization by E60 than were Furin WT RVPs (148-fold reduction in EC50, p = 0.002, Fig 2B and 2C), despite undetectable levels of uncleaved prM (Fig 2A).
Consistent with prior studies [26, 45, 47], increasing the efficiency of prM cleavage increased the proportion of WT RVPs that were resistant to E60 neutralization at the highest mAb concentration tested (22% and 55% for Std and Furin WT RVPs, respectively, p = 0.03, Fig 2B and 2D). In contrast, we observed only a minimal difference between the proportion of Std and Furin T198F RVPs resistant to neutralization at saturating concentrations of E60 (5% and 11%, respectively, p = 0.12, Fig 2B and 2D). Additionally, in contrast to the 70-fold difference in EC50 between Std and Furin WT RVPs, there was a much smaller (7-fold) difference in EC50 between corresponding preparations of T198F RVPs (p = 0.09, Fig 2B and 2C). Altogether, these data suggest that the large increase in sensitivity of T198F RVPs to neutralization by E60 was not simply due to increased retention of uncleaved prM. Furthermore, the reduced impact of maturation state on E60 recognition of WNV T198F suggests that the E60 DII-FL epitope is distinctly displayed by this variant.
We investigated whether the increased neutralization sensitivity of WNV T198F was dependent on amino acid chemistry at this position. In addition to T198F, we produced Std WNV RVPs in which the threonine at position 198 was replaced with amino acids containing small (A), nucleophilic (C, S), hydrophobic (L, M), acidic (D), basic (K), or amide (N) side chains. Each of these variants resulted in infectious RVPs, with titers within 2-fold of the WT control produced in parallel (p = 0.47, S1 Fig). Compared to WT, all T198 RVP variants were similarly sensitive to neutralization by mAb E16 (<2-fold difference in EC50, Fig 3A and 3C). However, when tested against E60, neutralization sensitivity varied among RVP variants, although there was no clear correlation with amino acid chemistry. For example, T198F, T198M, T198K resulted in a similar (~50-fold) reduction in EC50 despite incorporating amino acids with distinct side chain characteristics (Fig 3B and 3D). Importantly, every T198 variant except for T198A and T198S resulted in a significant increase in sensitivity to neutralization by E60 to varying extents (~10 to 100-fold reduction in EC50 compared to WT, Fig 3B and 3D). Thus, the increased sensitivity of T198F RVPs to neutralization by E60 was not linked to a particular change in amino acid chemistry at WNV E residue 198.
We hypothesized that the increased accessibility of the E60 DII-FL epitope on T198F virions was due to changes in virus conformational dynamics or ‘breathing,’ which allows the transient display of poorly exposed epitopes [47, 54]. A potential consequence of virus breathing is a reduction in virus stability that can be inferred by the loss of infectivity over time (or ‘intrinsic decay’) at physiological temperatures, as has been established for picornaviruses [65, 66]. Among the ensemble of conformations sampled by a virion at equilibrium, a subset may result in irreversible structural changes that are incompatible with infectivity for a given cell type. We therefore investigated whether the T198F mutation altered the functional stability of WNV RVPs. Following incubation at 37°C for up to 72 hours, the infectivity of WT and T198F RVPs collected periodically was determined by titrating viruses on Raji-DCSIGN-R cells (Fig 4A). In agreement with previous findings, virion maturation state modulated the rate of intrinsic decay [47, 67]: the infectivity of Std RVPs decayed more rapidly compared to that of Furin RVPs for both WT and T198F. For Std RVPs, the T198F substitution resulted in a ~3-fold reduction in half-life relative to WT (average half-life of 5 versus 16 hours, respectively, p<0.0001, Fig 4B). Similarly, the infectivity of Furin T198F RVPs decayed at a rate that was approximately twice as fast as Furin WT RVPs (average half-life of 10 versus 22 hours, respectively, p = 0.001). These findings support the hypothesis that the T198F mutation alters the ensemble of conformations sampled by virions at equilibrium.
The threonine at residue 198 in WNV and the phenylalanine found at the corresponding residue in DENV (193) and ZIKV (198) is highly conserved (99.9 to 100%) among 1989, 2692, and 104 sequences of WNV, DENV, and ZIKV naturally occurring isolates, respectively, available in the Virus Variation database [68]. We introduced the reciprocal mutation at this residue (F193T) into the Western Pacific strain of DENV1 to investigate whether it similarly affected antigenic structure and dynamics of the virion. The infectivity of standard RVP preparations of DENV1 F193T was reduced by ~10-fold as compared to WT DENVI (p = 0.001, S1 Fig). While prM cleavage was much less efficient for DENV than WNV, as reported previously [69, 70] (Figs 2A and 5A), the F193T mutation in DENV1 did not alter maturation efficiency; three independent standard preparations of WT and F193T DENV1 RVPs contained a similar level of prM (1.1-fold difference, Fig 5A). We next investigated the effect of F193T on the neutralization of DENV1 by E60 [58]. Similar to results with WNV T198F, DENV1 F193T RVPs were more sensitive to neutralization than WT RVPs (average EC50 = 25 ng/ml and 200 ng/ml for F193T and WT RVPs respectively, p = 0.0006, Fig 5B and 5C). Moreover, the F193T mutation resulted in a 3-fold reduction in the half-life of infectivity of DENV1 RVPs (average half-life of 2.5 and 0.8 hours for WT and F193T, respectively, p<0.0001, Fig 5D and 5E). In contrast to our findings with WNV and DENV1, mutation at the analogous residue (F198) of ZIKV E protein had no effect on sensitivity to neutralization by E60 or virion stability in solution (Fig 5F and 5G). Together, these results suggest that a single residue in the DI-DII hinge of the E protein alters the exposure of a cryptic DII-FL epitope and the stability of flavivirus particles in solution in a context-dependent manner.
To explore further the possibility that the T198F mutation modulates WNV conformational dynamics, we performed kinetic neutralization assays, in which virus-antibody complexes were used to infect Raji-DC-SIGN-R cells immediately following a 1 h pre-incubation at room temperature, or further incubated at 37°C prior to addition of target cells (Fig 6). We studied only Furin RVP preparations containing little or no prM to eliminate the confounding effects of heterogeneity in virion maturation state on neutralization sensitivity (Fig 2B) [26, 45, 47]. As observed previously, increasing the virus-antibody incubation time for mAb E16, which binds an accessible epitope on DIII, resulted in only modest increases in neutralization potency against both WT and T198F RVPs [54]. In contrast, kinetic changes in neutralizing antibody potency were more pronounced with mAb E60. Increased sensitivity due to the T198F mutation was observed at all time points tested. For example, after 8 hours of incubation, we observed complete neutralization of Furin T198F RVPs, whereas a large proportion of Furin WT RVPs remained infectious at the highest E60 concentration tested. Even after 24 hours of incubation with E60, WT RVPs remained much less neutralized compared to T198F RVPs. These findings suggest that the T198F mutation alters accessibility of the E60 DII-FL epitope.
To extend our findings with RVPs capable of only a single round of infection, we investigated effect of the T198F mutation on standard preparations of fully infectious WNV encoding GFP [71]. As observed with RVPs (Figs 1 and 2), the T198F mutation reduced the efficiency of prM cleavage (Fig 7A), and resulted in increased sensitivity to neutralization by E60 (Fig 7B). Following prolonged incubation at 37°C, T198F reduced the infectious half-life of virions by ~2-fold relative to WT (average half-life of 2.8 and 6.5 hours, respectively, p = 0.002, Fig 7C), consistent with our observations with RVPs (Fig 4). As expected, raising the temperature at which viruses were incubated to 40°C reduced the infectious half-life of both WT and T198F RVPs (average half-life of 3.6 and 1.6 hours, respectively, Fig 7C) compared to incubation at 37°C; the 2-fold decrease in the half-life of T198F relative to WT viruses was maintained at 40°C (p<0.0001). Despite reduced stability in solution, T198F viruses nonetheless displayed similar growth kinetics as WT in Vero cells at 37°C and 40°C, and in mosquito (C6/36) cells at 28°C (Fig 7D), suggesting that the effect of T198F on virus stability in solution might be masked under conditions that allow efficient cell-cell spread of infection.
Because the T198F mutation did not impair WNV replication in vitro, we investigated its impact on pathogenesis in a well-established mouse model of infection [72]. We infected 5-week old C57BL/6J mice with either WT or T198F WNV and monitored survival for three weeks. In contrast to the high mortality rate (13 of 15) observed among mice infected with WT WNV, only 2 of 15 T198F-infected mice succumbed to infection (p<0.0001, Fig 8A). As expected, WT- and T198F-infected mice that died experienced rapid weight loss beginning at day 6 post-infection (Fig 8B). Weight loss was observed to a much lesser extent among mice that survived infection.
Type I interferon (IFN) and antibody responses have been shown to be critical for protection against lethal WNV infection [72–75]. We therefore investigated the outcome of T198F infection of mice treated with MAR1-5A3, a previously described blocking antibody against IFN-α/β receptor that prevents type I IFN-induced intracellular signaling in vitro and inhibits antiviral responses in mice [76, 77], and of congenic C57BL/6J mice that were genetically deficient in mature B cells and antibody (μMT strain). The attenuated phenotype of T198F in WT mice was not observed in MAR1-5A3-treated WT or μMT mice. Although T198F infection remained attenuated relative to WT infection (1/10 versus 7/10 deaths, respectively, p = 0.004, Fig 8C) in mice treated with GIR-208, an isotype control antibody targeting human IFNγ receptor, MAR1-5A3-treated mice were equally susceptible to lethal infection with WT or T198F virus (9/10 vs 8/10 deaths, respectively, p = 0.36, Fig 8C). Analogously, 8/8 μMT mice infected with either WT or T198F virus succumbed to lethal infection by day 12 and 13 post-infection, respectively (p = 0.08, Fig 8D). Thus, the T198F mutation attenuates WNV in mice in a manner that is dependent on type I IFN signaling or B cell responses.
To investigate whether the T198F virus was attenuated due to rapid clearance, we collected serum samples at days 2 and 4 following infection of WT mice with either WT or T198F WNV. At day 2 post-infection, serum viral load was 7-fold lower in T198F- compared to WT-infected mice (p = 0.009, Fig 8E); a similar reduction in serum T198F infectious titer was observed at this time point (p = 0.06, Fig 8F). However, by day 4 post-infection, T198F serum viremia reached WT levels (p = 0.82, Fig 8E). Moreover, there was no difference in the infectious titer of WT and T198F viruses harvested from spleens at 4 days post-infection (p = 0.79, Fig 8F). Despite similar levels of WT and T198F viremia by day 4 post-infection, viral burden in the brain of T198F-infected mice was severely reduced compared to WT-infected mice. By day 6, WT-infected mice had a median virus titer of 104 PFU/g in the brain, whereas no infectious virus was detectable in the brain of T198F-infected mice (p = 0.001, Fig 8G). Infectious virus became detectable in the brain of 2 of 8 T198F-infected in mice by day 8, although at levels that were over 10-fold lower than those found in WT-infected mice on the same day (p = 0.02, Fig 8G). Sequence analyses of viruses isolated from the brain of T198F-infected mice revealed no reversion to WT. These results suggest that WNV containing the T198F mutation is suppressed early in infection and is attenuated for neuroinvasion or neurovirulence.
As shown above, viremia in T198F-infected mice was reduced as early as day 2, but not by day 4 post-infection (Fig 8E and 8F). Because WNV-specific antibodies become detectable 4 days after infection [72, 73], we hypothesized that natural antibodies might accelerate the rate of T198F virus decay relative to WT. To test this hypothesis, we compared the stability of WT and T198F viruses incubated at 37°C in heat-inactivated serum samples obtained from naïve WT or μMT mice; viruses incubated in media were included as controls. We observed a ~2.3- and ~1.5-fold reduction in half-life of WT and T198F viruses in these serum samples, respectively, compared to incubation in media (Fig 9), suggesting that antibodies and other heat-resistant serum factors modulate the infectious half-life of WNV. Notably, the 2-fold reduction in half-life of T198F viruses compared to WT was observed following incubation in WT serum, μMT serum, and media (Fig 9D), suggesting that natural antibodies do not differentially modulate the rate of decay of WT and T198F WNV.
To investigate whether neutralizing antibodies play a role in the attenuation of WNV T198F, we pooled serum samples from five WT- or T198F-infected WT mice for use in neutralization studies with WT and T198F RVPs. We compared serum samples obtained at days 6 and 9 after infection to distinguish neutralizing activity mediated by IgM and IgG, which become detectable at 4 and 8 days after infection, respectively [72, 73]. At both days 6 and 9, there were minimal differences (maximum of 1.2 fold change in EC50) in the ability of sera from WT- and T198F-infected mice to neutralize WT or T198F RVPs (Fig 10A–10D).
To distinguish IgM- versus IgG-mediated neutralizing activity, pooled serum samples from infected mice were either treated with 2-mercaptoethanol (2-ME), which preferentially degrades IgM [72, 78], or were used for IgG purification. As expected, treatment with 2-ME resulted in a large reduction (74–196 fold, Fig 10A and 10B) in the neutralization potency of serum samples obtained from WT- and T198F-infected mice at day 6, during which IgM, but not IgG is present [72, 73]. At day 9, when IgG is present, 2-ME treatment resulted in a smaller reduction in serum neutralization potency (4–8 fold, Fig 10C and 10D). As observed with untreated sera, at both days 6 and 9, there were minimal differences (maximum of 1.8 fold change in EC50) in the ability of WT- or T198F-immune sera treated with 2-ME to neutralize WT and T198F RVPs. Although T198F RVPs were slightly more sensitive (2–4 fold) to neutralization by IgG purified from day 9 sera compared to WT RVPs, this was observed for both WT- and T198F-immune IgG (Fig 10E and 10F), suggesting that infection with WNV T198F did not uniquely elicit NAbs that preferentially neutralized T198F viruses.
Finally, to directly study the impact of the T198F mutation on immunogenicity, we immunized WT C57BL/6J mice with WT or T198F RVPs capable of only a single round of infection. Pooled sera from WT or T198F RVP-immunized mice at either day 10 or 21 displayed limited differences in their ability to neutralize WT and T198F RVPs (S2C–S2F Fig). Similar results were observed with individual serum samples at day 21 post-immunization; although T198F was neutralized with a 3–4 fold greater potency compared to WT RVPs, this was observed for sera obtained from both WT and T198F RVP immunization groups (S2G Fig). These results suggest that neither infection nor immunization with T198F elicited unique NAb responses in mouse polyclonal sera.
Our study demonstrates that a single residue in the E protein DI-DII hinge regulates conformational dynamics in distinct flaviviruses, with relevant consequences in vivo for WNV infection and pathogenesis. Although conformational flexibility has been described for different virus families [53], the first evidence of the dynamic properties of flaviviruses came from structural studies of mAb 1A1D-2, which is capable of neutralizing multiple DENV serotypes, yet binds to an epitope in the β-strand of DIII that is not predicted to be fully accessible on the mature virus particle [79]. Monoclonal antibody 1A1D-2 can bind to DENV particles at 37°C, but not at 4°C, suggesting that at an elevated temperature, this mAb trapped a conformation on which the DIII β-strand epitope was otherwise not accessible at lower temperatures [79]. Consistent with the conformational flexibility of flaviviruses, subsequent studies showed that exposure of DENV2 virions to physiological temperatures in the absence of antibody results in the formation of an expanded ‘bumpy’ structure, on which E protein dimers are more loosely arranged and are rotated outwards relative to their orientation on fully mature particles [80, 81]. However, this ‘bumpy’ structure was not observed for all DENV strains or serotypes [81, 82], suggesting that sequence variation contributes to the structural pathways sampled by virus breathing at equilibrium.
The molecular determinants that govern flavivirus breathing have not been defined. We recently demonstrated that natural variation at residue 204 in DII of the DENV1 E protein explained large genotypic differences in sensitivity to neutralization by a mAb targeting a cryptic epitope in DIII [56]. In our current study, we identified a single mutation in the E protein DI-DII hinge of WNV (T198F) and DENV1 (F193T) that increased sensitivity to neutralization by mAb E60, which targets a poorly accessible epitope that includes the DII-FL. Because flavivirus neutralization occurs once the number of antibodies bound to the virion exceeds a stoichiometric threshold, neutralization potency depends not only on antibody affinity, but also on epitope accessibility [62]. Based on this model, antibodies targeting poorly exposed epitopes may not achieve complete neutralization even at saturating concentrations [45, 62], as observed for E60 against WT WNV. We observed that the T198F mutation markedly reduced the proportion of neutralization resistant WNV virions at high E60 concentrations. Compared to WT, increased accessibility of this epitope on T198F virions was less dependent on maturation state, which has been shown to indirectly modulate epitope accessibility [45, 46, 64]. Following prolonged incubation with E60 for up to 24 hours, WT WNV virions remained much less sensitive to neutralization than T198F virions. These results demonstrate that the E60 DII-FL epitope is displayed uniquely on T198F virus particles. Notably, increased sensitivity to neutralization by E60 also was observed by introducing a mutation at the corresponding residue of DENV1, but not ZIKV, suggesting that the molecular mechanisms governing conformational flexibility and/or FL exposure may be distinct for ZIKV, in agreement with recent neutralization studies with mAbs [31, 83–85]. The structural basis for these functional data cannot be inferred directly from our studies. The phenylalanine at position 193 and 198 of DENV and ZIKV, respectively play a space filling role (S4 Fig), while for WNV, the analogous threonine at this position projects outwards and is solvent exposed. Thus, the local structural environment likely contributes to the effects of amino acid substitutions at this position. Moreover, while the structures of mature ZIKV and DENV particles share many similarities, a distinguishing feature is an extended loop surrounding the glycan at ZIKV E residue 154, which has been hypothesized to limit accessibility of the adjacent DII-FL on the neighboring E protein [38].
Time-dependent increases in E60 neutralization potency were still apparent for T198F virions, suggesting that while this mutation increased the overall accessibility of a cryptic DII-FL epitope, it did not result in a grossly open ‘ground-state’ conformation. In support of this hypothesis, neutralization studies with an expanded panel of mAbs revealed that T198F (and DENV1 F193T) did not uniformly confer large increases in the potency of antibodies targeting distinct epitopes throughout the E protein (S3 Fig). Indeed, T198F resulted in a relatively modest increase in the neutralization sensitivity of WNV to mAb E53, which targets residues within the nearby DII bc-loop in addition to those within DII-FL [46, 58]. Together, these results suggest that T198F alters the ensemble of conformations sampled by WNV to increase the accessibility of poorly accessible epitopes within DII-FL. Consistent with alterations in conformational dynamics, T198F also impacted the functional stability of WNV virions. Although the molecular basis for the loss of virus infectivity (intrinsic decay) following prolonged incubation is not understood [47, 65, 66], we hypothesize this could be a consequence of virus breathing. Among the ensemble of conformations sampled by a dynamic virus, a subset may lead to irreversible changes in the E protein that impair viral infectivity. The more rapid intrinsic decay of T198F virions suggests that this mutation alters the conformational landscape in a manner that more frequently leads to irreversible changes in the E protein that are incompatible with infectivity.
Although studies of antibody reactivity and intrinsic decay have provided clues into the dynamic properties of flaviviruses, the consequences of virus breathing for viral replication and pathogenesis remain poorly understood. Our finding that the T198F mutation in WNV reduced the efficiency of prM cleavage from virions prepared under standard conditions suggests that virus breathing may affect the accessibility of the prM cleavage site during Golgi transit, thus contributing to the heterogeneity in the maturation state of released virus particles [41]. While the corresponding mutation in DENV1 (F193T) increased both sensitivity to E60 neutralization and the rate of intrinsic decay, prM cleavage efficiency in the context of DENV1 was unaffected. We previously demonstrated that the rate of intrinsic decay differs between WNV and DENV [47, 86], suggesting that sequence variation and the presence of uncleaved prM may alter the structural pathways sampled by flaviviruses. The possibility that virus breathing affects prM cleavage efficiency, perhaps by modulating access to the furin cleavage site on prM, further adds to the complex interplay among the determinants of flavivirus structural heterogeneity. Indeed, the reduced efficiency of prM cleavage of both WT and F193T DENV1 compared to WT WNV may reflect differences in the structural flexibility of DENV and WNV.
We demonstrated that altered virus breathing impacts pathogenesis. The T198F mutation attenuated WNV pathogenesis in WT mice, but not in mice treated with a monoclonal antibody targeting the IFN-α/β receptor or in congenic C57BL/6J mice deficient in B cells and antibody, suggesting that T198F attenuation is dependent on type I IFN- or B cell-mediated immunity. Our finding highlights the role of both innate and adaptive immune responses in protection against WNV lethal infection. Specifically, type I IFN signaling has been shown to be important in priming and enhancing B cell responses, in addition to its established role in innate antiviral defense [87–89]. Prior studies have demonstrated that both neutralizing and non-neutralizing WNV-specific antibodies can protect against lethal infection [72–74, 90]. For weakly neutralizing antibodies targeting DII-FL, protection is dependent on non-neutralizing mechanisms [90]. Our data indicate that T198F attenuation is not likely due to increased susceptibility to NAbs, suggesting a possible role for antibody effector functions.
T198F viremia was reduced as early as day 2 post-infection, before WNV-specific antibodies become detectable [72, 73], suggesting that this early viral suppression also might be due to differential effects of innate immune responses. Additionally, we previously found that the presence of even low concentrations of WNV-specific antibody can decrease the infectious half-life of virions in vitro [54], perhaps by trapping conformations that are incompatible with infectivity. Although we demonstrated that natural antibodies did not differentially affect the rate of intrinsic decay of WT and T198F viruses in vitro, it is possible that in the presence of other immune factors, even low concentrations of WNV-specific and/or natural antibodies may facilitate rapid viral clearance in vivo to limit dissemination to vital organs, as has been shown for other viruses [91]. Indeed, the suppression of T198F viremia very early in infection, though transient, was sufficient to limit CNS dissemination, which typically occurs between days 4 and 5 after infection [72, 74]. Finally, the stoichiometric requirements of prM cleavage for the production of infectious virus particles have not been defined. Although increased prM retention on T198F virus particles did not significantly impair infectivity in vitro (Fig 7), it is possible that decreased maturation efficiency of T198F may result in lower infectivity of key target cells in vivo, thus contributing to its attenuation.
Beyond impacting pathogenicity, the in vivo consequences of virus conformational dynamics are unexplored. The vector competence and extrinsic incubation period (time from infected blood meal to transmission) for both WNV [92–95] and DENV [96–99] are temperature-dependent, which could correspond to changes in the extent of virus breathing. It is intriguing to consider that the reduced infectious half-life of T198F WNV and F193T DENV1 virions in solution may result in less efficient transmission to mosquitoes during an infected blood meal, especially from a febrile animal, given that the rate of intrinsic decay is accelerated at elevated temperatures (Fig 7C and [86]). Indeed, sequence analyses reveal that WNV E residue 198 and the analogous DENV E residue 193 are highly conserved in nature [68]. The impact of changes in conformational dynamics on virus attachment to target cells also is unexplored. In addition to increased susceptibility to immune clearance, changes in conformational flexibility also might impair T198F virus interaction with host attachment factors in the central nervous system or in the vessels lining the blood-brain barrier [100].
T198F was neutralized slightly more potently than WT RVPs by both WT- and T198F-immune mouse sera (Fig 10 and S2 Fig), suggesting that infection or vaccination with T198F did not skew the NAb response to preferentially neutralize T198F. Thus, although the T198F mutation impacts antigenicity as measured by changes in accessibility of a cryptic DII-FL epitope, its effects on immunogenicity are unclear. These results suggest, however, that antibodies targeting DII-FL do not significantly contribute to the overall neutralizing activity of polyclonal sera in mice. As the specificity of the polyclonal antibody repertoire elicited by flavivirus infection likely differs between mice and humans [22, 101–103], how changes in E protein conformational flexibility alter immunogenicity in humans remains to be determined. Recently, a number of potently neutralizing human monoclonal antibodies that target quaternary epitopes within or across flavivirus E protein dimers have been identified following natural infection or vaccination [23, 24, 27, 30, 31, 33, 34]. We speculate that the dynamic properties of E proteins have the potential to impact the exposure of these epitopes, and thus the induction of antibodies against them. The conformational flexibility of envelope proteins has been shown to modulate antibody recognition of HIV ([104]; among the different structures sampled by HIV envelope trimers at equilibrium, broad and potent NAbs preferentially target the highly ordered, ‘closed’ trimer conformation [104–106]. Based on these observations, current HIV immunogen design strategies to elicit broad and potent NAbs are focused on stably presenting the closed form of native envelope trimers [106–108]. Whether limiting conformational flexibility is a suitable strategy for flavivirus vaccine design awaits further studies.
Our ongoing studies aim to identify additional residues throughout the E protein that regulate conformational flexibility to facilitate studies on the impact of flavivirus breathing on immunogenicity and other aspects of flavivirus biology, including maturation, replication, and the pH threshold of fusion [109, 110]. We hypothesize that residues at the E protein hinge regions and dimer interface play critical roles in regulating virus breathing by virtue of their conformational flexibility [17, 18] and potential interactions that contribute to overall virion stability [111], respectively. The existence of antiviral compounds that inhibit virus breathing of selected picornaviruses suggests an important role for structural flexibility in the lifecycle of viruses [112–117]. Structural flexibility contributes to heterogeneity in the antigenic structure of virions by governing the exposure of cryptic epitopes that may be immunodominant [47, 54]. For example, antibodies that bind the conserved DII-FL are cross-reactive, poorly neutralizing antibodies with the potential to contribute to antibody dependent enhancement at high concentrations, which is especially relevant in the context of DENV infection [52, 118]. We have shown here for WNV and elsewhere for DENV [56] that exposure of cryptic epitopes can be modulated by amino acid substitutions at a distance. Thus, an improved understanding of the molecular determinants that regulate flavivirus breathing and the consequences of conformational dynamics on flavivirus biology has the potential to inform both the design of novel vaccines and identification of antiviral compounds.
HEK-293T (ATCC) and Vero (ATCC) cells were maintained in Dulbecco’s Modified Eagle medium (DMEM) containing 25 mM HEPES (Invitrogen) supplemented with 7% fetal bovine serum (FBS; Invitrogen) and 100 U/ml penicillin-streptomycin (P/S; Invitrogen). C6/36 (ATCC) cells were similarly cultured, except with the addition of 1X non-essential amino acids (Invitrogen). Raji-DC-SIGN-R cells (Raji B lymphoblast [ATCC] engineered to stably express DC-SIGN-R, Pierson lab [45, 54, 62, 119]) were cultured in RPMI 1640 medium containing Glutamax (Invitrogen) supplemented with 7% FBS and 100 U/ml P/S. HEK-293T, Vero, and Raji-DC-SIGN-R cells were maintained at 37°C in the presence of 7% CO2. C6/36 cells were maintained at 28°C in the presence of 7% CO2.
We used a previously described expression vector encoding the structural genes (C-prM-E) of the WNV NY99 strain [57] as a template for mutagenesis. Initially, threonine at residue 198 of the WNV E protein was replaced with phenylalanine by site-directed mutagenesis using the Pfu Ultra DNA polymerase system (Agilent Technologies). The reciprocal mutation (F193T) was introduced into a previously described expression vector encoding the structural genes (C-prM-E) of the DENV1 Western Pacific strain [86]. Mutation at the analogous residue of the ZIKV E protein (F198T) was introduced into a plasmid encoding the structural genes of the ZIKV strain H/PF/2013 [120]. This plasmid is described elsewhere [121]. Additional amino acid variants were introduced at position 198 of the WNV E protein using primers containing a degenerate codon (NNN). PCR cycling parameters were as follows: 1 cycle of 95°C for 1 min; 18 cycles of 95°C for 50 s, 60°C for 50 s, and 68°C for 9 min; and 1 cycle of 68°C for 7 min. PCR products were treated with DpnI (New England BioLabs) for 3 h at 37°C, prior to transformation into Stbl2 cells (Invitrogen) and propagation at 30°C. The entire C-prM-E region of each construct was sequenced to ensure that no additional mutations were present.
RVPs were produced by complementation of a GFP-expressing WNV sub-genomic replicon with plasmids encoding the structural genes of WNV, DENV, or ZIKV, as described previously [121, 122], with slight modifications. Briefly, HEK-293T cells were pre-plated in a low-glucose (1 g/liter) formulation of DMEM containing 25 mM HEPES (Invitrogen), 7% FBS, and 100 U/ml P/S, transfected with plasmids encoding the replicon and structural genes at a 1:3 ratio by mass using Lipofectamine 3000 (Invitrogen), and incubated at 37°C. For each microgram of DNA, 2 μl of Lipofectamine 3000 was used. Four hours post-transfection, cells were transferred to 30°C. Supernatant was harvested at 72–96 h post-transfection, passed through a 0.22 μm filter (Millipore), and stored at –80°C. To produce mature preparations of WNV RVPs containing low to undetectable prM, RVPs were produced as above by co-transfecting plasmids encoding the replicon, structural genes, and human furin at a 1:3:1 ratio. To detect prM in RVP preparations, a modified structural gene construct that encodes prM and E, with a V5 tag immediately downstream of the prM signal cleavage site [123] was used to complement a plasmid encoding capsid [57]. For immunization studies, 180 ml of transfection supernatant containing WT or T198F RVPs was passed through a 0.22 μm filter, layered on 20% sucrose (pH 7.4), and pelleted by ultracentrifugation at 32,000 rpm at 4°C for 5 h. The virus pellet was resuspended in 0.5 ml of PBS containing 1% BSA.
Infectious WNV encoding a GFP reporter gene was produced using a previously described molecular clone system in which a DNA fragment encoding WNV structural genes is ligated into a GFP-expressing WNV replicon plasmid (pWNV-GFP-backbone V3) and transfected directly into HEK-293T cells [71]. Briefly, 1 μg each of the backbone and structural gene plasmids was digested with BamHI and BssHII, and ligated with T4 DNA ligase (New England Biolabs) in a final volume of 40 μl at 16°C overnight. Next, the entire unpurified ligation mixture was transfected directly into HEK-293T cells using Lipofectamine 3000 (Invitrogen). Cells were incubated at 37°C in the presence of 7% CO2. Viral supernatant was harvested at 48 and 72 h post-transfection, filtered using a 0.22 μm filter (Millipore), and stored at –80°C. To detect prM in fully infectious virus preparations, a DNA fragment encoding WNV structural genes was modified to express a V5 tag immediately downstream of the prM signal cleavage site and was used for virus production as described above.
Clarified virus-containing supernatant was serially diluted 2-fold in a total volume of 100 μl and used to infect 5 x 104 Raji-DC-SIGN-R cells in an equal volume at 37°C. Cells were fixed in 1.8% paraformaldehyde at 48 h or 16 h following infection by RVPs or fully infectious viruses, respectively, and GFP-positive cells enumerated using flow cytometry. Virus titer was calculated using the linear portion of the virus-dose infectivity curve using the following formula: Infectious units (IU)/sample volume = (% GFP-positive cells) x (number of cells) x (dilution factor).
Viruses produced in HEK-293T cells using WNV-GFP-backbone V3 [71] as described above were used to inoculate Vero or C6/36 cells at a multiplicity of infection (MOI) of 0.05 for 2 h at the indicated temperatures, after which supernatant was collected to confirm the input virus titer. After washing twice with PBS to remove unbound virus, cells were further incubated at 37°C (Vero), 40°C (Vero), or 28°C (C6/36). At the indicated time points, virus supernatant was collected and clarified by centrifugation at 2000 rpm for 5 min. Virus titers were determined on Raji-DC-SIGN-R as described above.
RVP or fully infectious virus stocks were diluted to a level of infectivity that ensures antibody excess (~5 to 10%) and incubated with serial dilutions of mAbs or heat-inactivated (56°C for 1 h) sera for 1 h at room temperature before addition of Raji-DC-SIGN-R cells. To investigate the kinetics of neutralization, virus-antibody complexes were further incubated for additional lengths of time at 37°C as indicated prior to addition of Raji-DC-SIGN-R cells. All infections were performed in duplicate at 37°C. At 48 h (RVP) or 16 h (fully infectious virus) post-infection, infectivity was scored as a percentage of GFP-positive cells by flow cytometry. Antibody dose-response curves were analyzed using non-linear regression with a variable slope (GraphPad Prism v 6.0g, GraphPad Software Inc.) to calculate the concentration of antibody (EC50) required to inhibit infection by 50%, or the maximum inhibition of infectivity achieved at the highest antibody concentration tested (‘% Resistant’).
Serum samples were depleted of IgM by treatment with 0.1 M of 2-mercaptoethanol in 1X PBS for 1 h at 37°C, as described previously [72, 78]. Total IgG was purified from 50 μl sera pooled from WT-immune (n = 5) or T198F-immune (n = 5) five-week old WT C57BL/6J mice at day 9 post-infection using the Melon IgG purification kit (Thermo Scientific) in a final volume of 500 μl (1:10 dilution). Purified total IgG was quantified using a human IgG ELISA kit (Immunology Consultants Laboratory) for use in neutralization assays as described above.
Viruses were diluted to a similar level of infectivity as used in neutralization assays, allowed to equilibrate at the indicated temperature for 1 h (reference) and sampled periodically for the next 48–72 h. At each time point, aliquots were collected and stored at –80°C. All frozen samples were thawed simultaneously and used to infect Raji-DC-SIGN-R in triplicate to assess infectivity as described above. Infection was normalized to the level observed at the initial reference time point and fitted with a one-phase exponential decay curve (GraphPad Prism v 6.0g, GraphPad Software Inc.) to estimate the infectious half-life.
The level of prM in RVP preparations was determined by SDS-PAGE and Western blot analysis, as previously described [64, 123]. Briefly, RVPs were concentrated and partially purified by ultracentrifugation at 4°C (32,000 rpm for 5 h) through a 20% sucrose cushion, followed by re-suspension in TNE buffer (50 mM Tris, 140 mM NaCl, 5 mM EDTA, pH adjusted to 7.4) containing 1% Triton-X100. WNV and DENV1 E proteins were detected by a cross-reactive DII-FL reactive mouse monoclonal antibody, 4G2 (1 μg/ml). WNV prM-V5 was detected using a 1:5000 dilution of a mouse monoclonal antibody targeting V5 (Invitrogen), while DENV1 prM was detected using mouse monoclonal antibody, prM22 (0.5 μg/ml) [124]. IRDye 800CW goat-anti mouse IgG (LI-COR Biosciences) diluted 1:2500 was used as a secondary antibody. Protein bands were visualized and quantified using the Odyssey infrared imaging system (LI-COR Biosciences).
C57BL/6J mice were purchased from Jackson Laboratories (Bar Harbor, ME) and congenic μMT B cell-deficient were bred at Washington University under pathogen-free conditions. Five-week old WT C57BL/6J mice or eight-week old μMT mice were inoculated subcutaneously via footpad injection with 102 focus-forming units (FFU) of WNV NY99 WT or T198F, and monitored daily for survival. Where indicated, C57BL/6J mice were injected via an intraperitoneal route with 0.5 mg of a mouse monoclonal antibody targeting mouse IFN-α/β receptor (MAR1-5A3) [76] or an isotype control mouse antibody targeting human IFN-γ receptor 1 (GIR-208) one day prior to infection. Purified LPS-free monoclonal antibodies MAR1-5A3 and GIR-208 were purchased from Leinco Technologies. WT and T198F viral stocks were generated by in vitro transcription of an infectious two-plasmid cDNA clone as previously described [125]. The T198F mutation was introduced into plasmid pWN-AB, which consists of the 5'-UTR and structural genes, by site-directed mutagenesis as described above. For immunization studies, five-week old C57BL/6J mice were injected via an intraperitoneal route with 50 μl of WNV WT or T198F RVPs normalized by infectivity and relative E protein content as determined by antigen capture ELISA. Serum from immunized mice collected at days 10 and 21 were analyzed in neutralization studies.
High-binding 96-well plates (Corning) were coated with 3 μg/ml humanized mAb E16 in 100 μl coating buffer (100 mM BupH Carbonate Bicarbonate Buffer, Fisher) with pH adjusted to 9.6. Plates were washed six times with PBS containing 0.05% Tween 20 followed by incubation with 100 μl blocking buffer (PBS, 3% non-fat dry milk, and 0.05% Tween 20). RVPs were serially diluted 2-fold starting at a 1:100 dilution in 100 μl blocking buffer, added to plates, and incubated for 1 h at 37°C. Plates were washed again and were incubated with 100 μl of mouse mAb E16 diluted in blocking buffer (2 μg/ml) for 1 h at 37°C. Following washing, 100 μl of HRP-conjugated goat anti-mouse IgG (Thermo Scientific) diluted 1:10,000 in blocking buffer were added to plates and incubated for 1 h at 37°C. One-step Ultra TMB-ELISA (Thermo Scientific) substrate was added (100 μl/well) and incubated for six minutes at room temperature in the dark. The reaction was stopped by the addition of 100 μl 1N hydrocholoric acid (Fisher) and read on a plate reader (BioTek Synergy H1) at a wavelength of 450 nm.
On the indicated day post-infection, mice were sacrificed and organs collected following extensive perfusion with PBS. Organs were weighed, homogenized using a bead-beater apparatus, and titrated by plaque assay on BHK-21 cells [126]. Viral burden in serum samples was measured by plaque assay on Vero cells, and viral RNA from serum was isolated using the Viral RNA Mini Kit (Qiagen) and measured by quantitative fluorogenic reverse-transcription PCR as described previously [126].
Statistical analyses were performed using GraphPad Prism v 6.0g (GraphPad Software Inc.). For results of in vitro experiments, paired t-tests or a one-way ANOVA followed by Dunnett’s multiple comparisons test was used, for two or more comparisons, respectively. For survival analysis, Kaplan-Meier curves were plotted and analyzed by the log rank test. Mouse serum, spleen, and brain viral loads and titers were compared using the Mann-Whitney test.
Experiments were approved and performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocols were approved by the Institutional Animal Care and Use Committee at the Washington University School of Medicine (Assurance number A3381-01).
|
10.1371/journal.pgen.1007696 | Functional equivalence of germ plasm organizers | The proteins Oskar (Osk) in Drosophila and Bucky ball (Buc) in zebrafish act as germ plasm organizers. Both proteins recapitulate germ plasm activities but seem to be unique to their animal groups. Here, we discover that Osk and Buc show similar activities during germ cell specification. Drosophila Osk induces additional PGCs in zebrafish. Surprisingly, Osk and Buc do not show homologous protein motifs that would explain their related function. Nonetheless, we detect that both proteins contain stretches of intrinsically disordered regions (IDRs), which seem to be involved in protein aggregation. IDRs are known to rapidly change their sequence during evolution, which might obscure biochemical interaction motifs. Indeed, we show that Buc binds to the known Oskar interactors Vasa protein and nanos mRNA indicating conserved biochemical activities. These data provide a molecular framework for two proteins with unrelated sequence but with equivalent function to assemble a conserved core-complex nucleating germ plasm.
| Multicellular organisms use gametes for their propagation. Gametes are formed from germ cells, which are specified during embryogenesis in some animals by the inheritance of RNP granules known as germ plasm. Transplantation of germ plasm induces extra germ cells, whereas germ plasm ablation leads to the loss of gametes and sterility. Therefore, germ plasm is key for germ cell formation and reproduction. However, the molecular mechanisms of germ cell specification by germ plasm in the vertebrate embryo remain an unsolved question. Proteins, which assemble the germ plasm, are known as germ plasm organizers. Here, we show that the two germ plasm organizers Oskar from the fly and Bucky ball from the fish show similar functions by using a cross species approach. Both are intrinsically disordered proteins, which rapidly changed their sequence during evolution. Moreover, both proteins still interact with conserved components of the germ cell specification pathway. These data might provide a first example of two proteins with the same biological role, but distinct sequence.
| Living systems have the unique ability to reproduce copies of themselves. In animals, the reproductive cells or their precursors, the primordial germ cells (PGCs) are specified by two different modes during embryogenesis [reviewed in 1]. In the inductive mode, the embryo generates signals, which specify a subset of cells to differentiate into PGCs. This was initially described for mouse and axolotl, which seems to be the ancestral mode [reviewed in 2]. In the alternative, maternal-inheritance mode, the mother deposits a cytoplasmic determinant termed germ plasm into the oocyte [reviewed in 3, 4]. After fertilization, germ plasm is inherited by a subset of embryonic cells, which then differentiate into PGCs as shown for example in Drosophila, C. elegans, Xenopus, and zebrafish. Ablation and transplantation experiments demonstrated that germ plasm is necessary and sufficient for PGC specification [reviewed in 4, 5, 6].
Germ plasm activities can be triggered by a single Drosophila protein termed Oskar (Osk) [reviewed in 7]. Osk mutants fail to assemble germ plasm [8, 9], whereas mis-localization of Osk induces ectopic PGCs [10, 11]. Structural and biochemical studies revealed that Osk binds RNA and more recently, that it increases the helicase activity of its interaction partner Vasa [12–15]. Despite its potent activity as a germ plasm organizer, Osk homologs were not discovered outside of insect genomes [reviewed in 16].
In vertebrates, we identified the zebrafish bucky ball (buc) gene, which appears similar at the genetic level to Osk in Drosophila [17]. Buc mutants fail to assemble germ plasm, whereas its overexpression induces ectopic PGCs [17, 18]. Biochemical studies with Buc suggest that it acts as a scaffold bringing together RNA binding proteins like Hermes [19–24]. Interesting results with the frog homolog Velo1 showed that its N-terminal prion-like domain forms SDS-resistant granules and that these amyloid-like aggregates recruit RNA [25, 26]. Similar to Osk, Buc is a fast-evolving protein, which has not been found outside of vertebrate genomes [17, 27]. It is therefore not surprising that no sequence similarity between Osk and Buc was previously discovered [28]. Nonetheless, the striking overlap in their function in Drosophila or zebrafish was frequently highlighted [e.g. 29], but experiments directly addressing the functional conservation of Osk and Buc are not available.
Here we provide a biochemical basis for the functional equivalence of both germ plasm organizers. We show that overexpression of Drosophila Osk leads to the formation of ectopic PGCs in zebrafish. Although Buc and Osk share this unique activity, we did not detect conserved motifs in extensive sequence comparisons. However, we find that both germ plasm organizers share protein stretches of intrinsically disordered regions (IDRs). Upon overexpression, we observe that Osk and Buc formed protein aggregates similar to liquid-liquid phase separations or hydrogels as previously shown for other IDPs [30, 31]. Moreover, when we treated early zebrafish embryos with hydrogel disruptors, we observe the fragmentation of Buc aggregates. IDRs change their sequence rapidly during evolution, which obscures conserved interaction motifs [32]. We indeed discover that known biochemical interactors of Osk, i.e. Vasa protein and nanos mRNA, also interact with Buc. These data show that the functional equivalence of germ plasm organizers is based on similar biochemical interactions and could represent the first case of an unrelated protein pair with hidden evolutionary homology.
We first analyzed the functional equivalence of germ plasm proteins by analyzing their activity to reprogram somatic cells into PGCs. The germ cell induction assay exploits that the first somatic cells in zebrafish segregate from the germline at the 16-cell stage (Fig 1A) [17]. Injecting gfp-nos3’UTR reporter mRNA [33] into a middle blastomere containing endogenous germ plasm highlighted PGCs in 18 hours post fertilization (hpf) embryos (Fig 1B and 1J). By contrast, injection into a somatic cell (corner blastomere) leads to background activation of the PGC-reporter as previously published (Fig 1C and 1J) [17]. Co-injecting wild-type buc mRNA encoding amino acid 1–639 into somatic cells was sufficient to significantly increase PGC specification, but mutant buc mRNA coding for aa 1–361 was not different from negative controls (Fig 1D, 1E and 1J).
The Xenopus Velo1 protein was recently postulated to act as a functional homolog of Buc [25, 26]. To test this hypothesis experimentally, we overexpressed Xenopus Velo1 in zebrafish embryos and found ectopic germ cells (Fig 1F and 1J). To determine the specificity of the germ cell induction assay, we injected the zebrafish Piwi homolog Ziwi, which is also a germ plasm component [34]. However, Ziwi showed no activity confirming that not every germ plasm component is active in the germ cell induction assay (Fig 1G).
The germ plasm organizers Osk and Buc have a remarkable genetic similarity [reviewed in 6, 7, 35], but conserved sequences were not discovered [28]. To compare their function experimentally, we tested Osk in the germ cell induction assay. Fascinatingly, short Oskar (sOsk, aa 139–606), which is the active variant in Drosophila for specifying germ cells, also induced PGCs in zebrafish. By contrast, mutant sOsk (aa 139–254) showed no effect (Fig 1H–1J). The mutant controls are derived from the osk84 and the bucp43 alleles, which have identical RNA sequences to their wild-type counterparts besides a point mutation generating a premature STOP-codon [9, 17]. Thus, overexpression of buc or osk mRNA per se is not sufficient to induce germ cells. Injecting GFP-fusions of these mRNAs lead to fluorescent embryos suggesting that they are indeed translated into protein (S1A Fig). Quantification indicated a similar number of specified PGCs by Osk and Buc (S1B Fig). These results suggest that the germ plasm organizers sOsk, Buc and their homologs share the unique activity to specify germ cells.
Although these ectopic cells express the PGC reporter and migrate to the prospective gonad, it was recently shown that reprogrammed germ cells retain GFP fluorescence, while they differentiate into somatic tissue types of the three germ layers [36]. Similar to Buc [17], Osk induced an increased number of vasa mRNA positive cells at 3 hpf (S2 Fig). Moreover, we analyzed Vasa protein expression of cells expressing the PGC reporter in the 16-cell assay. Notably, GFP positive cells after Buc or Osk injection also expressed Vasa protein (Fig 2A–2C and 2G–2I). By contrast, embryos injected with mutant Buc (1–361) only stained the endogenous germ cells with Vasa, but ectopic GFP-positive germ cells were not detected (Fig 2D–2F). Taken together, we concluded that the cells specified by Buc and Osk differentiate into PGCs.
According to the sequence-structure-function paradigm, proteins with the same activity contain homologous sequence motifs to interact with similar binding partners [reviewed in 37, 38]. Conserved amino acid sequences were previously described for Xenopus Velo and zebrafish Buc, but not between Drosophila Osk and Buc [17, 28]. We therefore pursued a stepwise strategy for their direct, bioinformatic comparison starting with pairwise alignments (Fig 3A). We detected only 11.5% similarity between both proteins (Fig 3B; S1 Table). The long Osk (lOsk) isoform, which is inactive in germ cell induction in Drosophila [39], reduced similarity to Buc even further down to 10%. A comparison of zebrafish Buc with Drosophila Vasa as an unrelated sequence showed 18.5% similarity, while Vasa homologs in zebrafish and Drosophila were 59.4% similar (Fig 3B).
In previous studies, the alignment of orthologs from different species revealed conserved domains and thereby hidden similarities [40, 41]. We aligned the sequences of 14 vertebrate Buc orthologs discovering two conserved motifs (aa 24–84 and 114–128) within the previously described BUVE-sequence (aa 23–136) [17] and another novel motif in the center of Buc (aa 372–394) (Fig 3C). The same approach with Osk detected published motifs: the LOTUS-domain (aa 154–236 lOsk) [42, 43], the Lasp binding region (aa 290–369) [44], and the putative hydrolase homology sequence within the OSK domain (aa 452–490) [12, 14, 45] (Fig 3C). We then generated profile hidden Markov models (HMM) of both proteins, but to our surprise did not detect significant hits by searching the Drosophila genome for sequence similarities with the Buc-HMM. The Buc-HMM consensus sequence, however, showed 43% identity of aa 56–81 in Buc to the DAZ motif in zebrafish Dazl (Fig 3D; S3 Table) [46]. Searching with Osk-HMM identified Tdrd5 and -7 in zebrafish and Tejas in Drosophila, which all contain LOTUS-motifs, but no similarity to Buc (Fig 3D; S3 Table). Finally, comparing the HMM-models of sOsk and Buc to each other did also not discover conserved motifs (S3 Table). Taken together, our extensive bioinformatic analysis did not detect hidden sequence similarities between the two germ plasm organizers Osk and Buc and hence, could not explain their similar activity.
Intrinsically disordered proteins (IDPs) seem to be an exception to the conventional sequence-structure-function paradigm [reviewed in 47]. IDPs are defined by a disordered stretch of at least 30 residues [reviewed in 48]. Indeed, the Xenopus homolog Velo1 was shown to contain a low-complexity motif within the BUVE domain, which forms insoluble amyloids [25]. In addition, Osk and Buc were proposed to encode intrinsically disordered proteins [12, 49]. Similar to Osk and Buc, IDPs frequently evolve faster than structured proteins. Furthermore, IDPs can form liquid-liquid phase separations or hydrogels as found in RNP-granules or the germ plasm [30, 31, 48, 50, 51, reviewed in 52, 53, 54]. As the intrinsic disorder of Osk and Buc was previously not shown, we analyzed the intrinsic disorder prediction of Osk and Buc using the PONDR-VSL2 algorithm [55]. PONDR-VSL2 is a metapredictor, which conservatively combines the results of prediction algorithms. Both protein sequences displayed large disordered regions (Fig 4A and 4B). Interestingly, the previously identified prion-like domain in the N-terminus (aa 1–150) [25] appeared in this disorder prediction as the largest ordered sequence in Buc (Fig 4A). Prion-like domains and IDR are considered low complexity sequences suggesting that Buc almost entirely consists of low complexity sequences. We used zebrafish Vasa as a positive control for IDP prediction, which showed the known unstructured domain of about 200 aa at its N-terminus [56, 57], whereas Ziwi was largely structured (Fig 4C and 4D).
Osk does not display prion-like domains [25] but was recently shown to form aggregates in insect S2-cells supporting its prediction as an IDP [13]. To investigate, whether Buc forms similar protein aggregates, we transfected HEK293 cells with plasmids encoding fusions with monomeric GFP and eGFP. Buc and Osk formed protein aggregates, whereas the GFP control was uniformly distributed (Fig 4E–4G; S3A–S4C Figs). Moreover, when we cotransfected Buc-mGFP with Osk-Cherry, we found partially overlapping aggregates of Buc and Osk (Fig 4H–4J). These data indicate that Osk and Buc encode IDPs with a propensity to form cellular protein aggregates.
A short treatment with the aliphatic solvent 1,6-hexanediol dissolves hydrogels formed by IDPs as described for germline P-granules in the C. elegans ovary, but not amyloid-like aggregates like the Balbiani body in Xenopus oocytes [25, 58, 59]. To distinguish whether germ plasm in zebrafish forms amyloid-like aggregates or hydrogels, we treated ovaries of Buc-GFP transgenic females with hexanediol (HD). Using time lapse-microscopy we observed that a 30 min exposure to hexanediol did not disperse the Balbiani body (Fig 5A and 5B; S1–S4 Movies). Extending the treatment to 3 hrs or doubling the hexanediol concentration to 10% did also not dissolve the Balbiani body (S4A–S4D Fig). This result corroborates the amyloid-like character of Buc aggregates [25]. Nonetheless, we noted that some Buc-GFP granules drained off the Balbiani body leaving behind a perforated Buc-GFP scaffold (Fig 5B). Interestingly, 30 minutes after washing out hexanediol, the Balbiani body was restored similar to untreated oocytes (Fig 5C). Hexanediol did not affect oocyte microtubules or microfilaments (S4E–S4L Fig) in line with previous studies on Xenopus oocytes showing that none of these cytoskeletal elements seems to be required for Balbiani body integrity [60, 61]. The hexanediol experiments suggest that Buc condensates in the Balbiani body have a partially liquid and partially solid character during zebrafish oogenesis, which is consistent with a continuous hardening model of protein condensates [reviewed in 62].
The Xenopus germ plasm was proposed to acquire a more liquid character at the end of oogenesis [25]. Indeed, our own time-lapse imaging results with embryos from Buc-GFP transgenic mothers support the liquid behavior of embryonic germ plasm in zebrafish [63]. We therefore treated embryos with hexanediol and observed the integrity of germ plasm by time-lapse microscopy. To our surprise, the embryonic germ plasm never completely dissolved like shown for the C. elegans ovarian P-granules, but only fragmented (Fig 5D–5F). In contrast to the oocyte however, the germ plasm did not reaggregate after washout of the drug (Fig 5G and 5H). When we analyzed the surviving embryos at 3 hpf, the majority showed numerous fragments of Buc-GFP aggregates, whereas control-treated embryos showed no change (Fig 5I and 5K). 1,2,3-hexanetriol (HT) is chemically similar and frequently used as a control for the specificity of hexanediol [59]. Indeed, the more polar structure of hexanetriol disrupted Buc-GFP aggregates less efficiently than hexanediol (Fig 5J and 5K). These results support the hypothesis that zebrafish germ plasm forms an intracellular hydrogel, whose aggregation is probably mediated by intrinsically disordered regions (IDRs) of Buc.
We next addressed whether Buc aggregation is sufficient for germ cell specification. Buc(1–601)-GFP lacks 38 C-terminal amino acids thereby retaining most of the IDRs (Fig 4A). Buc(1–601) still forms protein aggregates in HEK293 cells (Fig 5L). Reducing Buc further to aa 1–361 still leads to protein aggregation compared to a GFP control (Fig 5M and 5N). The aggregation of wt Buc(aa1-639) and mutant Buc(aa1-601) was confirmed in zebrafish embryos (Fig 5O–5Q). However, Buc (aa1-601) injected embryos did not show ectopic PGCs (Fig 5R, 5S and 5U). Furthermore, the intrinsically disordered RNA-binding protein FUS [reviewed in 64] did not induce the PGC reporter (Fig 5T and 5U) suggesting that aggregation is not sufficient to specify germ cells and that other biochemical interactions are critical for germ cell specification.
The similar function of Osk and Buc postulates that they perform similar biochemical interactions, which then initiate the PGC-specification program. However, the fast sequence evolution of the IDRs in both proteins may obscure sequence similarities detectable by current alignment algorithms, which then bind to conserved interactors. Osk binds to Smaug, Valois, and Vasa protein [15, 65, 66]. To test whether these proteins are conserved in the Buc interactome, we immunoprecipitated Buc-eGFP from zebrafish embryos. To avoid non-specific interactions after overexpression, we used Buc-GFP transgenic fish, which express Buc under control of its own promotor [63]. We then identified binding partners by mass-spectrometry and searched the Buc interactome for the zebrafish homologs of Osk binding partners (Fig 6A). Interestingly, we found MACF1 highly enriched in the Buc interactome (S3 Table). Zebrafish mutants in macf1 and buc show defects in embryonic polarity and Balbiani body localization [67–70] supporting the specificity of the biochemical interaction. Another good indicator for the specificity of the pulldown was the detection of the germ plasm component Ziwi (piwil1) [34], which was not enriched in the Buc sample (S3 Table). This result indicates that we did not bring down the entire germ plasm during Buc pull-down. Among the zebrafish homologs of Osk binding partners, we focused on Vasa for further analysis, since its stronger enrichment suggested a greater probability to interact with Buc.
Exciting structural studies showed that Vasa interacts with the extended LOTUS domain of Osk [13]. More specifically, helix α2 (aa156-167) and α5 (aa226-234) in the LOTUS-extension of Osk are required for Vasa interaction. Interestingly, α5 encodes an IDR, which folds into a helix on interacting with Vasa. Since we could not detect these peptide sequences in Buc with bioinformatics, we verified biochemically that Buc interacts with Vasa during the period of germ cell specification. We pulled down Buc-GFP from embryonic extracts of transgenic embryos at 3 hpf (Fig 6B). As controls we used the H2A-GFP transgenic line, which is one of the few strains in zebrafish expressing a GFP-fusion under maternal control similar to Buc-GFP [71]. We detected Vasa in Western blots after Buc-GFP pulldowns, but not with H2A-GFP controls suggesting that Vasa interacts with Buc in vivo during PGC specification.
To further corroborate the interaction of Buc and Vasa in vivo, we used bimolecular fluorescence complementation (BiFC) in early zebrafish embryos [20, 72, 73]. BiFC takes advantage of a split Venus protein called VN (N-terminal) and VC (C-terminal), which then complement into a functional fluorescent protein, if they are brought in close proximity. Coinjecting Vasa-VN with VC-Venus or Buc-VC with VN-Venus fragments did not form a fluorescent protein confirming the specificity of BiFC (Fig 6C and 6D). By contrast, overexpression of Buc-VC with Vasa-VN formed fluorescent aggregates in zebrafish embryos supporting the hypothesis that Buc binds Vasa in vivo (Fig 6E and 6F).
Vasa protein was previously described to be ubiquitously expressed during the maternally controlled embryogenesis [74, 75], while Buc protein is confined to the four germ plasm spots [17, 63, 76]. To support their biochemical interaction, we determined whether endogenous Buc and Vasa protein expression overlap during germ cell specification. Labelling zebrafish embryos by antibody staining showed that Vasa is ubiquitous at the 16-cell stage and at 3 hpf as previously described (Fig 6G–6L). Buc localization overlaps with Vasa only in the germ plasm, which further supports the hypothesis that Buc and Vasa might interact in vivo.
Previous reports in chicken showed that Vasa overexpression reprograms embryonic stem cells to a germline fate [77]. Furthermore, Drosophila Osk enhances Vasa activity suggesting that Vasa performs a key activity during germline specification [13]. We therefore analyzed the role of Vasa in the zebrafish germ cell induction assay. Surprisingly, Vasa induced ectopic germ cells, whereas another Buc binding protein Hermes [19–21, 23] showed no activity (Fig 7A and 7B). This result suggests that Vasa performs a critical activity during germ cell specification.
As Osk activates Drosophila Vasa and Vasa triggers germ cell formation in zebrafish, we investigated, whether in vitro translated Osk-GFP binds to zebrafish Vasa. Indeed, Osk pulled down zebrafish Vasa whereas controls did not interact (Fig 7D) supporting the hypothesis that Osk and Buc share conserved interactions. The Buc(1–361) and Buc(1–601) mutants do not induce PGCs and we therefore analyzed its interaction with Vasa. To our surprise both mutant alleles bound Vasa like wt Buc, (Fig 7E), whereas a control protein (non-muscle myosin II) was not bound by Buc (S5 Fig). Although these results show that the interaction with Vasa is conserved among germ plasm organizer proteins, the data also indicate that the mutant Buc proteins lack another critical interaction.
Mutant Buc(1–361) binds to Vasa, but does not induce germ cells suggesting that full-length Buc performs additional interactions. Osk was recently shown to bind RNA e.g. nanos [12, 14] and many IDPs are RNA-binding proteins [reviewed in 78]. To address whether Buc interacts with zebrafish nanos3 mRNA [33], we coexpressed GFP-tagged versions in HEK293 cells. After immunoprecipitation of Buc, we detected zebrafish nanos3 by RT-PCR, but not a cotransfected competitor 3'UTR (SV40) or an abundant, endogenous control (18S rRNA) (Fig 8A). Similarly, Osk-GFP bound to zebrafish nanos3 mRNA. We then tested whether Buc(1–361) pulls-down nanos3-3'-UTR RNA. Indeed, this mutant Buc could not pull-down the RNA suggesting that it lacks a motif, necessary for RNA interaction (Fig 8B).
As Buc and Vasa interact with RNA, their interaction might be mediated indirectly via RNA. However, RNase treatment did not inhibit Buc-Vasa binding showing that the complex was held together by protein–protein interactions or was protected by RNA-bridging from nuclease activity (Fig 8C). These results discover two novel biochemical interactors of Buc i.e. Vasa protein and nanos mRNA, which are conserved with Drosophila Osk (Fig 8D).
Here we discover a conserved core complex, which is required for germ cell specification. This complex includes the conserved germline components Vasa protein [reviewed in 79, 80, 81] and nanos mRNA as well as a germ plasm organizer like Osk or Buc. These molecules are probably not the only components of the complex and might contain additional proteins or RNAs, since numerous, canonical germ plasm components are conserved in metazoan genomes [reviewed in 35, 82, 83]. For instance, while this manuscript was under revision, the Tudor protein Tdrd 6 was shown to interact with Buc in zebrafish [84]. This interesting study suggests that the Tdrd6 interaction controls the aggregation of Buc. Remarkably, Tudor as the founding member of this protein family was first discovered as a germ plasm component in Drosophila [85–87] thereby supporting the hypothesis of a conserved core of germ plasm components in metazoans.
In addition to a conserved interactome, Osk and Buc also share intrinsically disordered regions (IDRs), which probably form weak interactions to oligomerize (Fig 8D). Multimerization of intrinsically disordered proteins causes phase-transitions or biological condensates [reviewed in 88]. The hydrogel-disruptor hexanediol dissolves germ plasm in C. elegans, whereas we observed fragmentation in zebrafish suggesting that the liquid character of germ plasm varies in different species [50, 59, 89]. In Xenopus eggs however, the Buc homolog Velo forms amyloid aggregates, which are resistant to hexanediol [25, 58, 90]. The less liquid character of amyloids is also consistent with the initial description of the germ plasm harboring Balbiani body in spider oocytes, which shows a more solid state [91]. At the end of frog oogenesis however, it was reported that Velo does not form amyloid-like aggregates anymore, which is in line with a more liquid behavior of the germ plasm in the embryo [25, 63].
Our hexanediol experiments are similar to Mip6p aggregates in yeast [90]. A 30-min pulse of hexanediol treatment leads to the fragmentation of Mip6p aggregates and to reaggregation after wash-out of the drug. Interestingly, the reassembled Mip6p aggregates were inherited symmetrically during cell division, whereas Mip6p granules are inherited asymmetrically in untreated controls leading to more Mip6p positive cells. This fragmentation of Mip6p aggregates seem similar to the behavior of germ plasm in zebrafish, which eventually results in the increased number of Buc aggregates 2 hrs after hexanediol treatment.
Our data are consistent with a model, in which germ plasm organizers like Buc provide a scaffold, which nucleate a phase transition at a specific location in the embryo. This aggregation drives the recruitment of other germ plasm components and eventually germ cell specification (Fig 8D). Interestingly, RNAs might not only contribute to the specificity of different granules, but also seem to nucleate phase transitions by recruiting IDPs as shown in the fungus Ashbya gossypii [92]. This might explain why the IDRs of Buc are not sufficient for germline formation. Our results however suggest that phase-transition of germ plasm seems to have a rather permissive than an instructive role for germline formation. Although the liquid nature of germ plasm was described in different organisms, the purpose of forming these aggregates for germ cell development is still a matter of debate.
Vasa seems to be a central component for germline specification. It was already reported that Vasa overexpression in chicken embryonic stem cells induces germ cells [77]. In addition, a zebrafish vasa mutant does not maintain nanos3 mRNA expression and thereby loses its germline stem cells [93]. Interestingly, Osk activates Vasa helicase activity demonstrating that in Drosophila the germ plasm organizer has an instructive role in germline specification [13]. Buc might also regulate Vasa activity in zebrafish and not only act as a scaffold recruiting Vasa to the germ plasm. Our results showing germ cell induction after Vasa overexpression would be consistent with this model. However, Vasa is already expressed in somatic cells in the early zebrafish embryo, which raises the question, why its overexpression reprograms a corner blastomere to the germline. We speculate that overexpression of Vasa bypasses the requirement for an activator. Such an effect was also observed for intracellular signaling components. For instance, Smad proteins in the BMP pathway are active after overexpression, but their endogenous activation requires phosphorylation [94]. Similarly, overexpression of Vasa might therefore have sufficient activity to start germline specification. In this model, it is not the localization of Vasa protein, which marks the germline of a species, but its activity. We therefore speculate that the activity of Vasa would be a more reliable marker for the germline. It would be more precise to visualize the early germline by the localization of a germ plasm organizer as an activator like Oskar or by the downstream products of Vasa's helicase activity such as piRNA maturation [95–97]. This is especially interesting in species similar to zebrafish such as the sea urchin, where Vasa is ubiquitously present in the early embryo [98].
A fascinating finding of our study is that Osk and Buc share some biochemical interactions despite the absence of recognizable sequence homologies. These similarities are remarkable considering that vertebrates and dipterans split more than 500 million years ago [99]. Two alternative scenarios could explain this functional equivalence. Both proteins are analogous designs, which converged at recruiting a similar interactome during evolution. We cannot rule out this model, but it seems most plausible for somatic tissues, where the loss of an organ might not lead to an evolutionary dead-end. By contrast, tinkering with a germ plasm organizer during evolution would result in reduced or missing fertility, and eventually the extinction of the entire species. As the invention of novel proteins with identical functions is very unlikely [reviewed in 37, 100, 101], the convergence model for germ plasm organizer evolution becomes increasingly complex to explain.
We therefore favor the second scenario, in which Osk and Buc are homologs, which diverged from a common ancestor. They probably have unrelated sequences, because their role as intrinsically disordered scaffolds releases the constraints to maintain a defined protein structure as described for other IDPs [reviewed in 102]. This model is supported by the recent finding that the LOTUS domain is not sufficient to bind Vasa, but requires an intrinsically disordered extension (aa 226–234) of low evolutionary conservation [13]. In addition, germ cell-specification is a very early event during the evolution of multicellularity and hence, germ plasm organizer proteins have a long history of diverging during evolution [103]. The fast evolution of IDPs probably hides conserved motifs, which bind to a similar interactome such as Vasa and nanos mRNA [51, reviewed in 53, 54]. Indeed, a similar situation was previously described for the intrinsically disordered domains CID and NCBD [32].
This hypothesis would also predict that Osk and Buc have a similar structure, which would explain their conserved interactions. There are already known examples of protein pairs with structural similarity, which do not display a related amino acid sequence. For instance, Sumo and Ubiquitin show a sequence identity of 18%, but form almost identical structures [104]. Despite their similar structure, both have different biological roles [reviewed in 105]. Moreover, Hsc70 and Actin provide another example for structural similarity without sequence conservation [106]. Furthermore, biochemistry has isolated numerous analogous enzymes e.g. carbonic anhydrases from different organisms, which show identical biochemical activities without related sequences [reviewed in 107]. However, in none of these examples, the conservation of their biological role was investigated, i.e. whether the function of a protein is conserved in the other species like Osk in zebrafish. It will therefore be fascinating to learn how similar the structure of Buc is, compared to the known structure of Osk [12–14].
The functional equivalence of germ plasm organizers in the absence of sequence similarity might be more widespread including other species. For instance, C. elegans germ plasm or P-granules have a similar composition, since they also contain Vasa protein and nanos RNA [108, reviewed in 109, 110–112]. Although the identity of a germ plasm organizer protein in C. elegans is currently not clear, it has been speculated that MEG-3 or PGL proteins might act as P-granule nucleators similar to Osk [113–115]. MEG-3 binds RNA [113] and could therefore interact with the C. elegans homologs of nanos mRNA and Vasa protein similar to Osk and Buc. Furthermore, recent studies on PGL-3 show that it binds nanos 3 RNA albeit weakly [116]. Moreover, PGL proteins are not found in vertebrate or insect genomes and nucleate the formation of hydrogels [116]. PGL proteins genetically interact with Vasa homologs [112, 117], but the direct biochemical binding was not tested. These examples provide some candidates in C. elegans, whose conservation as germ plasm nucleators will not be revealed by sequence comparisons but need to be analyzed with functional and biochemical experiments.
However, Osk, Buc or germ plasm organizers in other species could only be termed true homologs, if the identity of a common ancestor is known. Without this information, the functional similarity of two proteins without sequence homology remains a fascinating, but unique case. Contrarily, Osk and Buc could also represent a widespread phenomenon. In a more global perspective, more protein-pairs could exist, which are currently termed "novel" or "species-specific" but show similar functions across distant species.
Fish were maintained as described [118] in accordance with regulations of the Georg-August University Goettingen, Germany. Zebrafish experiments were conducted according to EU directive 2010/63/EU and maintained according to the EuFishBioMed/Felasa recommendations (https://www.eufishbiomed.kit.edu/59.php). Experiments were approved by the Lower Saxony State Office for Consumer Protection and Food Safety (AZ14/1681).
Injections were performed into wild-type embryos (hybrid of *ABxTLF). 16-cell embryos were injected as previously described [17]. At least 20 embryos were sorted per injection and for biological replicates independent clutches of eggs were used. One blastomere was injected with 0.5 nl RNA-solution containing 100 pg/nl of PGC-reporter (GFP-nos-3´UTR) plus 100 pg/nl mRNA encoding a germ plasm component. Buc and short osk mRNAs contained their 5´ and 3´UTR sequences, respectively. Bucp43 and bucp106 were identical to wt mRNA except for a premature stop codon in 362 aa and 602 aa, respectively [17]. Short osk and osk084 mRNA were identical except for the premature stop codon in osk084 [9].
BiFC assays were performed with modifications as previously described [72]. Briefly, wild-type embryos (hybrid of *ABxTLF) were injected at the one-cell stage with the mRNAs encoding the VN- and VC-fusions (200 pg each). Embryos were imaged for fluorescence at the 3hpf stage with a LSM780 confocal microscope (Carl Zeiss Microscopy, Jena).
Dechorionated Buc-GFP transgenic embryos at the one cell stage were treated with 1,6-hexanediol or 1,2,3-hexanetriol (5% w/v in E3-medium) for 30 min, whereas control treated embryos were exposed to E3-medium. Embryos were incubated at 28.5°C in glass dishes for 30min and then transferred into fresh E3-medium in agarose coated dishes at 28.5°C until 3 hpf.
Sample preparation: Proteins were separated on denaturing 4–12% gradient SDS-PAGE (Invitrogen, Carlsbad/CA, U.S.A.). After Coomassie staining for visualization, each lane was cut into 23 equidistant slices irrespective of staining. For in-gel digestion, gel slices were washed with water, reduced with dithiothreitol (10 mM in 100 mM NH4HCO3, 50 min, 56°C) and alkylated with iodoacetamide (55 mM in 100 mM NH4HCO3, 20 min, RT, dark). In between, the gel slices were washed with acetonitrile for 15 min and dried in a Speedvac at 35°C. Gel slices were digested overnight at 37°C with porcine trypsin (12.5 ng/μl in 50 mM NH4HCO3, 5 mM CaCl2). Peptide extraction from the gel slices was performed with aqueous acetonitrile.
HEK-293 cells (0.2 x106/ well) were co-transfected with the indicated combinations of plasmids for protein and RNA expression (S4 Table). Cells were incubated for 48 hrs and screened for expression of GFP and Cherry fluorescence. Cells were then lysed in (0.5 ml) YSS buffer (50 mM Tris pH 8, 75 mM NaCl, 1 mM MgCl2, 100 mM sucrose, 1 mM DTT, 0.5% NP- 40, 1x complete protease inhibitor cocktail (Roche Mannheim)) and centrifuged for 10 min. (13,000 rpm, 4 ºC). 50 μl of the supernatant were kept aside as the input fraction and the rest was incubated with pre-blocked GFP nanotrap beads (Chromotek) for 3 hrs at 4°C. Beads were washed (YSS buffer) and the bound fraction was released from the beads in 5% SDS. RNA was isolated using phenol/chloroform/isoamylalcohol and precipitated in 0.3 M ammonium acetate/ 50% EtOH, washed with 70% EtOH and used for cDNA synthesis.
RNA was reverse transcribed for first strand synthesis using random hexamers and SuperScript II RTase (Thermo Fisher Scientific). cDNA was amplified using the primers described in S4 Table.
Plasmids used in this study are listed in S4 Table.
Protein sequences for Danio rerio Bucky ball, Drosophila melanogaster Oskar, and the respective orthologs were retrieved from the NCBI protein database. The vertebrate Buc sequences used for multiple alignments were: Danio rerio (gi|292610748), Oryzias latipes (gi|432930267), Tetraodon nigroviridis (gi|47225100), Takifugu rubripes (gi|410909482), Oncorhynchus mykiss (gi|642119256), Pimephales promelas (gi|73433600), Ictalurus punctatus (gi|311721748), Xenopus laevis (gi|148230857), Xenopus tropicalis (gi|301615136), Anolis carolinensis (gi|327275069), Gallus gallus (gi|118086206 / gi|513169732), Canis familiaris (gi|545522949 / gi|73976581). The Oskar insect sequences: Drosophila melanogaster (gi|45553317 / gi|24645205 / gi|317183309), D. sechellia (gi|195330556), D. simulans (gi|195572425), D. yakuba (gi|195499262), D. erecta (gi|194903569), D. ananassae (gi|194741640), D. pseudoobscura (gi|198454187), D. persimilis (gi|195152922), D. mojavensis (gi|195111098), D. virilis (gi|2498716 / gi|195389208), D. immigrans (gi|111663086 / gi|111663088), D. grimshawi (gi|195054868), D. willistoni (gi|195445335 / gi|195445337), Aedes aegypti (gi|83701126 / gi|157134733), Culex quinquefasciatus (gi|170041806), Anopheles gambiae (gi|118783859 / gi|333468779), Anopheles darlingi (gi|312371899), Acromyrmex echinatior (gi|332023144), Nasonia vitripennis (gi|302138022). Global/local pairwise alignments of Buc and Osk were performed using the EMBOSS tools Needle/Water (http://www.ebi.ac.uk/Tools/psa/) with default parameters. Multiple alignments of Buc/Osk and their respective orthologs were constructed with the T-COFFEE software version 8.69 using standard parameters [122]. Hidden Markov models (HMM) were built from the multiple alignments using the HMMER3 software in default configuration [123]. The HMMs were used to search the complete genomic protein sequence complement of Danio rerio and Drosophila melanogaster as obtained from the NCBI protein database. To detect potential distant relationship between the models, the HMMs were uploaded to the HHpred server [124]. The intrinsic disorder of proteins was predicted with PONDR-VSL2 in default configuration [125].
Error bars indicate the standard deviation of the average (at least three independent experiments). The statistical significance (P-value) of two groups of values was calculated using a two-tailed, two-sample unequal variance t-test with MS-Excel.
|
10.1371/journal.pntd.0004605 | Neutropenic Mice Provide Insight into the Role of Skin-Infiltrating Neutrophils in the Host Protective Immunity against Filarial Infective Larvae | Our knowledge and control of the pathogenesis induced by the filariae remain limited due to experimental obstacles presented by parasitic nematode biology and the lack of selective prophylactic or curative drugs. Here we thought to investigate the role of neutrophils in the host innate immune response to the infection caused by the Litomosoides sigmodontis murine model of human filariasis using mice harboring a gain-of-function mutation of the chemokine receptor CXCR4 and characterized by a profound blood neutropenia (Cxcr4+/1013). We provided manifold evidence emphasizing the major role of neutrophils in the control of the early stages of infection occurring in the skin. Firstly, we uncovered that the filarial parasitic success was dramatically decreased in Cxcr4+/1013 mice upon subcutaneous delivery of the infective stages of filariae (infective larvae, L3). This protection was linked to a larger number of neutrophils constitutively present in the skin of the mutant mice herein characterized as compared to wild type (wt) mice. Indeed, the parasitic success in Cxcr4+/1013 mice was normalized either upon depleting neutrophils, including the pool in the skin, or bypassing the skin via the intravenous infection of L3. Second, extending these observations to wt mice we found that subcutaneous delivery of L3 elicited an increase of neutrophils in the skin. Finally, living L3 larvae were able to promote in both wt and mutant mice, an oxidative burst response and the release of neutrophil extracellular traps (NET). This response of neutrophils, which is adapted to the large size of the L3 infective stages, likely directly contributes to the anti-parasitic strategies implemented by the host. Collectively, our results are demonstrating the contribution of neutrophils in early anti-filarial host responses through their capacity to undertake different anti-filarial strategies such as oxidative burst, degranulation and NETosis.
| Filariases are chronic debilitating diseases caused by parasitic nematodes affecting more than 150 million people worldwide. None of the current drugs are selective, neither able to eliminate the parasites nor to prevent new infections once the drug pressure has waned. Therefore, blocking the entry and the migration of the infective larvae (L3) could be an efficient way to control the infection. In the present study we investigated the early interaction between the host and the L. sigmodontis murine filariasis with a focus on the neutrophils in the innate host responses. We uncovered a key role of neutrophils in the control of infection provided by the CXCR4-gain-of-function mice (Cxcr4+/1013) that display a blood neutropenia as well as an accumulation of skin-infiltrating neutrophils. Overall, we reveal that in the early phase of filariasis, i.e. after L3 are delivered into the skin and before they reach their site for reproduction, neutrophils are critical elements of the host innate protective response arsenal. A better understanding of their indirect and/or effector role(s) may provide mechanistic clues to host factors implicated in parasitic nematode entry and potentially lead to the identification of new drug targets.
| Filarial nematodes constitute a large group of human pathogens (i.e. Onchocerca volvulus, Brugia malayi, Brugia timori, Wuchereria bancrofti, Loa loa, Mansonella spp.) infecting around 150 million people throughout the tropics with more than 1.5 billion at risk of infection. Filariases remain a health issue, as no selective and efficient treatments are able to prevent and eliminate filarial infections in the exposed and infected populations [1].
Early stages of infection are common to filarial nematodes with the infective stages of filariae (L3) being delivered in the skin of the host by blood feeding arthropods [2]. However later in their life cycle, depending on the filarial species, adult stages settle in their preferred tissues for maturation/reproduction/release of microfilariae, strongly suggesting different migration paths for the larvae. Consequently, adult filariae reside in connective tissues (i.e. dermis, subcutaneous tissue, aponeuroses, tendons), blood and lymphatic vessels, or coelomic cavities such as the pleural cavity in the case of the rodent filarial L. sigmodontis [3], which is the focus of our study.
L. sigmodontis is a widely used experimental model in which infective larvae migrate from the skin through the lymphatic system before ending up in the pleural cavity [4]. Almost 70–80% of the skin-inoculated L3 will not reach their maturation niches [4, 5], supportive of early defense mechanisms implemented in the skin by the mice. The innate immune response to nematodes can involve host cell populations such as eosinophils, neutrophils, mast cells and activated macrophages [6–9]. A clear consensus has emerged from the L. sigmodontis model regarding the role of eosinophils in the early protective immunity against L3, induced within two days in infected mice immunized with irradiated L3 [10–13]. Several lines of evidence from the various filarial models including L. sigmodontis also indicate that neutrophils might contribute to the host immunity, but at late stages of the infection notably through the control of the adult worms and blood-circulating microfilariae [14–16]. Although it has not been reported that an innate immune response dominated by neutrophils might kill incoming L3, neutrophils are recruited to sites of parasitic nematode entry in Nippostrongylus brasiliensis, Heligmosomoides polygyrus, Brugia pahangi [17–20], and L. sigmodontis infections [12]. Additionally, neutrophils were reported to contribute to macrophage-dependent resistance mechanisms against Strongyloides stercoralis and macrophages-dependent resolution of tissue damage induced by Nippostrongylus brasiliensis [21–23]. Moreover, along with their essential role in responses to microbial pathogens, neutrophil activation and recruitment could be attributed to the endobacteria Wolbachia harbored by some filarial parasites [24–26]. Interestingly, the resistance mechanisms of the host in early defense against L3 also target these symbionts [26, 27].
Neutrophil homeostasis, which is maintained by a balance between granulopoiesis in the bone marrow (BM) and migration between blood and tissues, is finely tuned by the chemokine system. The essential CXCL12/CXCR4 chemokine-chemokine receptor pair [28–30], which induces typical activation of the Gαi protein- and β-arrestin-dependent pathways [31–33] regulates with other chemokine receptors, hematopoiesis and the lymphoid and peripheral trafficking of neutrophil and lymphocyte subsets [33]. The CXCL12/CXCR4 axis is notably critical for the release of neutrophils from the BM [34] and lungs [35]. CXCR4 engagement also promotes the migration of neutrophils from inflamed skin into draining lymph nodes, a process thought to participate in the control of pathogens through initiation of immune responses or conversely in the spreading of infection [36–38].
We have previously shown that a CXCL12-dependent cell response is associated with the “resistant phenotype” of C57BL/6 mice to L. sigmodontis infection [39] raising the possibility that this chemokine participates in the host immune-mediated resistance mechanisms. Here we thought to examine this possibility in C57BL/6 mutant mice, which harbor an inherited heterozygous Cxcr4 mutation engendering a gain-of-CXCR4 function [Cxcr4+/1013], an anomaly linked to the rare immunodeficiency WHIM disorder [40, 41]. As a consequence, [Cxcr4+/1013] mice exhibit a peripheral leukopenia affecting blood lymphocytes and neutrophils as do patients. This leukopenia is transiently reversed in mice [41] and patients [42] upon inhibition of CXCR4 with the selective antagonist AMD3100 thus demonstrating the causal role of the gain-of-CXCR4 function. Interestingly, neutropenia of the mutant Cxcr4+/1013 mice occurs in the context of normal maturation of the granulocyte lineage [41] in support of the fact that transient normalization of blood counts upon CXCR4 blockade or patient’s infections [43] likely arises from disturbed neutrophil trafficking rather than from a production defect [41].
Our findings reveal a dramatic blockade of L. sigmodontis infection in mutant mice, which manifests early in the skin of mice inoculated with infective L3 and is related to a higher number of neutrophils in the skin of mutant mice as compared to their wt counterpart. The identified host-resistance mechanisms such as the increase of neutrophils infiltrating the skin, the oxidative burst response, and the release of NET promoted by L3 were extended to wt mice demonstrating that neutrophils are important contributors of the early host resistance mechanisms against the nematode.
Infective larvae were subcutaneously injected into wild type (wt) and Cxcr4+/1013 C57BL/6 mice and were recovered in the pleural cavity 20 days post-inoculation (p.i.), when the filarial load is still high and before it decreases, as observed in the resistant background of the mice (S1 Fig). The mean number of worms recovered from the wt mice (9.72 ± 0.65 SEM, Fig 1) was similar to the one previously reported for resistant C57BL/6 mice at this stage of the infection [44–47]. As compared, the mean filarial load dramatically dropped in the mutant mice to minute levels (3 ± 0.26 SEM) corresponding to a 70% decrease in the filarial parasites (Fig 1). An enumeration performed earlier at day 8 p.i., after arrival of most of the L3 larvae in wt mice pleural cavity (11 ± 0.5 SEM), indicated a similarly small number of worms recovered from the mutant mice (4.4 ± 0.7 SEM) (S2 Fig). These results excluded the possibility that a killing of L3 larvae in the pleural cavity of mutant mice may account for the decreased filarial recovery in these mice.
Upon arrival of the L3 in the pleural cavity, mice mount a cellular response directed against the worms, which reaches a peak 30 days p.i. when larvae are molting into young adults [39, 46, 48]. In this context, resistant C57BL/6 mice are characterized by having a higher increase of pleural exudate cells (PleCs) as compared to susceptible mice [39]. PleCs counts were increased upon infection in both type of mice but neither the levels nor the composition was significantly different between control wt and Cxcr4+/1013 mice (S3 Fig) at steady state and upon infection. These results are suggesting that the strongest resistance of mutant mice to filarial infection was not associated with any change in PleCs. We then investigated whether the leukocyte populations in the blood circulation were differentially affected in the mutant mice upon infection. Lymphocytes counts of wt and mutant animals were not significantly modified by 20 days post infection and the mutant mice remained lymphopenic (Fig 2A, left panel) throughout the infection up to 70 days p.i. In contrast, both eosinophil and neutrophils blood counts, also constitutively diminished in mutant mice, were almost normalized 20 days p.i. (Fig 2A, middle and right panels). Time course analysis of the circulating neutrophils counts in Cxcr4+/1013 mice throughout subcutaneous (SC) infection of L3 revealed an increase, which tended to normalize the circulating neutrophil counts from 20 days followed by a progressive decay back to initial levels at day 70 p.i. (Fig 2B). While a slight increase of blood neutrophil counts was observed in the first five days following injection (days 1–5), the levels remained significantly different between both wt and mutant mice (see S1 Table).
We further investigated further whether this late and sustained neutrophilia was related to infection as an indirect consequence of the SC delivery of L3 and/or directly via filarial immunomodulatory molecules present in whole body crude extracts (WBE). To do so we counted blood neutrophils in Cxcr4+/1013 mice throughout 20 days p.i. in four different experimental settings: inoculation of L3 (L3) versus L3 crude extracts (WBE) and SC inoculation compared to intravenous (IV) one. Neutrophils counts in Cxcr4+/1013 mice shown in Fig 2C were expressed as a ratio to wt mice counts, in order to illustrate the trend toward normalization (raw data are presented in S2 Table). Inoculation of L3 in Cxcr4+/1013 mice via the IV route (IV L3) was associated with an increase in circulating neutrophils, the levels of which tended toward a transient normalization between days 1 and 3 p.i. A similar response was observed upon IV inoculation of L3 crude extracts (IV WBE) indicating that this increase is independent upon the mobility of the larvae. As previously shown, SC injection with infective larvae (SC L3) promoted a long-lasting normalization of neutrophils counts at 20 days p.i., which persisted for at least 40 days (Fig 2B and 2C). SC inoculation of crude extracts (SC WBE) only promoted a transient normalization, which occurred much earlier at 5 days p.i. (Fig 2C). Altogether these results indicated that L3 either through immunomodulatory molecules (WBE) and/or their motility have, independently of the inoculation site, the potency to induce a transient neutrophilia in the mutant mice reaching neutrophils counts comparable to those in wt mice. However long-term normalization of blood neutrophil counts in the mutant mice over is strictly dependent upon the inoculation of live larvae into the skin.
We then compared the parasitic success upon IV and SC inoculation of Cxcr4+/1013 mice at 20 days p.i. Interestingly, we unraveled that upon IV inoculation filarial larvae also ended up in the pleural cavity. Strikingly, IV inoculation reversed the resistant phenotype of Cxcr4+/1013 mice observed upon SC inoculation (Fig 1) with a recovery of larvae in the pleural cavity reaching the levels obtained in the wt mice (Fig 3). Moreover this increase of the parasitic success was extended to wt mice, which displayed twice more filariae recovered in the pleural cavity upon IV inoculation as compared to SC inoculation (Fig 3). Collectively, these data support on one hand the concept that L3 migration from the skin to the pleural cavity would stand a blood stage. On the other hand, they strongly suggest that the mechanisms underlying the enhanced resistance of the Cxcr4+/1013 mice to infection are acting before larvae reach the bloodstream, and more generally bring evidence in support of the skin’s major role in the host defense against filariae.
We thought to compare the steady state amounts of eosinophils, macrophages, mast cells and neutrophils in the skin of wt or mutant mice to those reached in both mice upon 6 hours p.i. considering the potential role of these cells in parasite clearance at early stages of the infection. The constitutive levels of eosinophils were similar in the skin of control wt and Cxcr4+/1013 mice whatever the layer either the dermis/hypodermis or the subcutaneous loose connective tissue (SLCT). Additionally, upon infection, eosinophil levels were similarly increased (5-fold) in both layers of the mutant and wt mice (S4A Fig). The numbers of macrophages and mast cells, degranulated or not, were also found to be in the same range in Cxcr4+/1013 and wt mice, but not significantly increased upon infection in any of the two models (S4B and S4C Fig). In contrast, the number of resident neutrophils constitutively present in the dermis-hypodermis of Cxcr4+/1013 mice was significantly higher than in wt mice (12.4 ± 1 SEM vs 4.79 ± 0.6 SEM neutrophils per mm2 of skin respectively, p < 0.001) (Fig 4). While an increasing trend was also observed in SLC layer, it was not significant. Filarial infection promoted a significant recruitment of neutrophils in the dermis/hypodermis of both mice, resulting in comparable levels of neutrophils at 6 hours p.i. in Cxcr4+/1013 and wt mice (28 ± 3.7 SEM and 23.4 ± 2.7 SEM neutrophils per mm2 of skin respectively). Cxcl12, which is expressed in the dermal stroma [49, 50], was found at higher levels in mutant mice dermis as compared to their wt counterparts (S5 Fig). Additionally, Cxcl12 levels were markedly increased in the dermis of both Cxcr4+/1013 and wt mice 6 hours p.i. (S5 Fig). Thus the steady state levels of the chemokine and their increase after infection are mirroring the variation in dermal neutrophil numbers herein reported in mice strengthening the possible interplay between both processes. Significantly, this increase in skin neutrophil recruitment observed in the 6 hours following filarial infection was concomitant with an increase of blood neutrophil numbers in both wt and Cxcr4+/1013 mice, of 2.9 and 2.7 fold respectively (S6 Fig). This phenomenon observed both in mutant and wt mice is highly suggestive of a neutrophil release in the bloodstream triggered upon immune sensing of the filariae by skin-resident cells.
The possibility that the high number of neutrophils present in the skin of mutant mice contributed to the resistant phenotype of these mice was further investigated by depleting neutrophils prior to infection. To do so, mice neutrophils were selectively depleted by a classical method based on a single intraperitoneal injection of anti-Ly6G antibodies 6 hours before SC inoculation of filarial larvae. Either the anti-Ly6G 1A8 clone targeting more specifically neutrophils (i.e. ly6G+ cells) or the NIMP-R14 clone that can also deplete monocytes (i.e. ly6C+ cells) [51–53]. Depletion of circulating neutrophils was already effective at the time of infection (Fig 5A, D0). Then at 10 days p.i., blood neutrophil levels in wt or Cxcr4+/1013 mice recovered values in the same range than those of control mice (i.e. infected mice pre-injected with PBS). Importantly, depletion was also dramatic on skin-resident neutrophils as early as 6 hours after anti-Ly6G antibody injection leading to a 5 to 10 fold decrease compared with constitutive levels in wt and Cxcr4+/1013 mice, respectively (Fig 5B). During the time-span of the experiment, lymphocyte numbers remained unchanged between control and neutrophils-depleted wt or mutant mice. We then compared the number of worms recovered in the pleural cavity 20 days p.i. in control and neutrophils-depleted wt or mutant mice. Strikingly, neutrophil depletion with anti-Ly6G antibodies whatever the antibody used (1A8 or NIMP-R14 clones), dramatically increased the number of filariae recovered in the pleural cavity of Cxcr4+/1013 mice to levels similar to those reached in the wt mice (Fig 5C). In contrast, neutrophil depletion had no significant effect on the worm burden in wt mice although some increasing trend was observed upon treatment with the NIMP-R14 clone, which also affects ly6C+ monocytes (Fig 5C, right panel). These results indicated that the sole depletion of neutrophils, also affecting skin-resident ones, reverted the mutant mice to a wt phenotype, supporting a critical role for this population in the resistance of Cxcr4+/1013 mice to filarial infection.
Collectively, these data identified a critical role for skin neutrophils in the host protective mechanism against primary L3 infection. The resistant phenotype displayed by Cxcr4+/1013 mice might underlie, either that neutrophils must be in elevated numbers at the point of infection, and/or that mutant mice neutrophils have a heightened activation state or modified functions. From previous analyses of neutrophils derived from patients suffering from the WHIM syndrome [54] it is anticipated that Cxcr4+/1013neutrophils would be also prone to an enhanced Cxcl12-dependent chemotaxis [41]. Chemotaxis assays performed on BM-isolated neutrophils indeed confirmed that neutrophils derived from Cxcr4+/1013 mice were more sensitive to Cxcl12 than their wt counterparts (S7 Fig). This increased Cxcl12-induced chemotaxis was similarly displayed by neutrophils isolated from infected mutant mice 20 days p.i (S7 Fig). Moreover, neutrophils derived from wt and mutant mice, infected or not, were found to express equivalent levels of Cxcr4 and Cxcr2 receptors and displayed comparable chemotactic response to the Cxcr2 agonist Cxcl1 (S7 Fig). These results thus indicated that Cxcr4+/1013 mice derived neutrophils display in vitro a selective enhanced responsiveness to Cxcl12-induced chemotaxis as a consequence of the gain-of-CXCR4-function they harbor. We then investigated in vitro other neutrophil functions such as the capacity of these cells to produce reactive oxygen species (ROS) and to undergo NETosis. ROS production was assessed with a nitroblue tetrazolium (NBT) assay. Results indicated that both wt and mutant mice-derived neutrophils were able to reduce the colorless NBT to black deposits within the cells indicating that the production of the superoxide anion (O2-) was not altered in neutrophils derived from the Cxcr4+/1013 mice (Fig 6A). In contrast, quantification of the oxidative burst revealed that ROS production by neutrophils from both wt and Cxcr4+/1013 mice was increased upon exposure to L3 and significantly heightened in neutrophils derived from Cxcr4+/1013 mice as compared with wt mice ones (Fig 6A, right panel). This oxidative burst in response to L3 was associated with a significant decrease in the intracellular content of myeloperoxidase (MPO) in neutrophils from both wt and Cxcr4+/1013 mice (Fig 6B, left panel), which was mirrored by an increase of the neutrophil-released MPO (Fig 6B, right panel). Interestingly, neutrophils from Cxcr4+/1013 mice released higher levels of MPO to the medium, indicating a stronger susceptibility to the degranulation process induced by exposure to L3 (Fig 6B, right panel). Extracellular release of MPO, neutrophil elastase, histones and chromatin decorated with numerous active proteins are the signature of NETs formation. We therefore investigated the potential induction of NETosis upon neutrophils exposure to L3 larvae by quantifying both cell viability and extracellular DNA release using the cell-impermeable SYTOX dye (Fig 6C, left panel). Quantification indicated that upon 4 hours exposure with L3, neutrophils purified from Cxcr4+/1013 mice were significantly more engaged into a cell-death program than those from wt mice (Fig 6C, right panel). The presence of extracellular DNA in culture supernatants increased with the exposure time to L3 larvae (from 4 to 36 hours) and was more marked in cultures with Cxcr4+/1013 BM-derived neutrophils suggesting that these cells are more prone to release NETs (Fig 6D). Examination of immunofluorescence slides of skin from infected and control mice indeed revealed granular structures co-staining with MPO, neutrophil elastase and DAPI, which hallmark NETs structures [55], strongly supporting the existence of neutrophils undergoing NETosis in the skin of mutant mice infected with L3 larvae (S8 Fig).
We provide compelling evidence revealing the potential for skin neutrophils to contribute in the early host defense against primary L. sigmodontis infection by using the CXCR4-gain-of-function Cxcr4+/1013 neutropenic mice. These mice found to be highly resistant to primary infection allowed us to demonstrate that (i) mice harbor an elevated numbers of dermal neutrophils in steady state the depletion of which, prior to infection, ablates the resistant phenotype of the mutant mice, and that (ii) mutant mice-derived neutrophils produce increased amount of ROS mediators and NETs upon L3 larvae stimulation, as compared with wt mice-derived neutrophils. Further, an original setting of intravenous delivery resulted in a strong enhancement of the parasitic burden, which strikingly reached similar levels in both wt and mutant mice. This demonstrates that most of the incoming L3 larvae including in wt mice are destroyed in the skin upon typical SC infection thus strengthening evidence in favor of the essential role of the skin in the protective mechanism against primary L3 infection. Of note, the rate of larvae recovery in both mice upon IV delivery did not reach 100% success. One could argue that some L3 within the inoculums may be unviable, but it mainly suggests that the control of the filarial load does not only take place in the skin. Previous studies indeed revealed that among the 70% of larvae that are leaving the skin within the first day, only one third are reaching the pleural cavity [56] with some L3 being found in pulmonary alveoli and arteries [56] suggesting that the lung could act as a clearance organ [57]. Of importance and in line with this, the IV infection setting experimentally supported for the first time the Wenk’s early hypothesis that the infective larvae might pass through the cardiopulmonary blood system to reach the pleural cavity [4].
We found that mutant mice display a heightened steady state number of neutrophils in the skin. Although the mechanism of this increase remains to be determined, the enhanced expression of Cxcl12 in the dermis of mutant mice combined with the increased chemotaxis toward the chemokine displayed by the mutant mice-derived neutrophils likely contribute to this process. Numerous evidences support the hypothesis of a major contribution of this neutrophil resident pool in the protective mechanism of the mutant mice against primary L3 infection. First, the selective and transient depletion of neutrophils prior infection, which also affected neutrophils in the skin, resulted in normalization of the rate of larvae recovery in the pleural cavity of mutant mice; second, recruitment of neutrophils in the skin of wt and mutant mice 6 hours p.i. led to similar levels in both mice and was mirrored by a neutrophilia in both animals; third, infection via IV route bypassed the protective mechanism of the mutant mice and; fourth, mutant mice display normal steady state numbers of eosinophils, mast cells and macrophages in the skin, suggesting that these innate immune cells do not participate, at least quantitatively, in the resistant phenotype of the mutant mice. These results however do not rule out the potential qualitative contribution of these cells in the control of filarial infections notably in the setting up of adaptive immune response. The protective role of eosinophils was indeed described in mice vaccinated with irradiated L3 larvae [12, 13] and that of mast cells in filarial-infected CCL17 deficient mice [7]. The blood leukopenia affecting the mutant mice, which constitutes a finely tuned sensor of the leukocytosis induced by the infection, affected neutrophils and eosinophils that were found to transiently reach normal levels upon infection. Although this process is not related to the early protective mechanism of mutant mice it supports the general concept that leukocyte trafficking promoted by the infection might impact the adaptive immune response against filarial infection. Moreover it indicates the potency of filarial-antigens to induce leukocytosis as demonstrated upon injection of filarial crude extract. Such early neutrophilia (2 hours p.i) may be related to excretory-secretory (ES) proteins released by filarial nematodes and which originate from the oesophageal glands, the anterior sensory glands (amphids), the posterior sensory glands (phasmids), the secretory pore, the hypodermis through transcuticular secretion but also from exosome release [58]. Notably, we have recently reported that the excretory-secretory proteins from L3 differ quantitatively and qualitatively from the other stages of L. sigmodontis [59].
We investigated the possibility that the early protective mechanism of mutant mice might involve a higher activation of skin resident neutrophils as a consequence of the Cxcr4-gain of function. We indeed found an abnormally enhanced migration of the mutant-derived neutrophils toward CXCL12, thus extending our previous observation made in this mouse model [41]. Moreover, with regard to the large length of the L3 larvae (i.e. about 750 μm) [58, 60] we sought to investigate the potential of neutrophils in inducing NETs that can be released by in response to microbe size-sensing [61]. The NETs were reported to capture Gram-positive and Gram-negative bacteria, fungi, and viruses [62–64] as well as Apicomplexa parasites, Leishmania, Eimeria, Plasmodium, and Toxoplasma [65] and the Strongyloides stercoralis nematode [66]. NET formation or NETosis is a gradual process notably involving ROS generation, transport of MPO and the extracellular release of chromatin [67–69]. Both wt and mutant mice-derived neutrophils were able to generate ROS mediators upon L3 stimulation likely causing NETosis. Importantly, mutant mice-derived neutrophils display heightened levels of ROS content and increased release of MPO and extracellular DNA when cultured with L3 and were more prompt to death as compared to their wt counterpart. Whether the filarial excretory-secretory proteins and/or the bacterial content (i.e. Wolbachia) also contribute to activation of neutrophils, including triggering of NETosis in response to live L3 cannot be excluded. Further work is needed to investigate the mechanisms by which NETs contribute in L3 larval entrapment and killing, which can be indirect as suggested by recent work on the Strongyloides stercoralis nematode model [66].
In summary, our findings suggest that the higher responsiveness exhibited by Cxcr4+/1013 neutrophils may be critical in the mutant mice-protective antifilarial response thus accounting for the fact that neutrophil depletion in control wt mice does not affect parasitic burden. Hence, they emphasize the potential of skin resident neutrophils to contribute in the early host defense against filarial infections when they are sufficiently activated and present in significant numbers prior infection. Such increase of the proportion of more functionally active neutrophils maybe envisioned in a wild type host environment during inflammatory processes and the subsequent abnormal increase of aged neutrophils that represent an overly active subset with notably enhanced propensities to form NETs [70, 71] or in the context of co-infections. Indeed, the immune responses evoked by bacterial or viral infections are associated with changes in the local cytokine environment and increases in the numbers and the activation state of neutrophils that could have implications for the outcome of filarial infections in light of recent findings highlighting interplay between host immune responses against parasites and virus in the course of co-infection (reviewed in [72]). Finally, our results identifying an immuno-modulatory role for the CXCL12/CXCR4 pathway could have implications for the development of therapies that should be further studied in filarial nematodes-infected individuals.
All experimental procedures were carried out in strict accordance with the EU Directive 2010/63/UE and the relevant national legislation, namely the French “Décret no 2013–118, 1er février 2013, Ministère de l’Agriculture, de l’Agroalimentaire et de la Forêt”. National license number 75–1415 approved animal experiments: protocols were approved by the ethical committee of the Museum National d’Histoire Naturelle (Comité Cuvier, License: 68–002) and by the “Direction départementale de la cohésion sociale et de la protection des populations” (DDCSPP) (No. C75-05-15).
The filaria L. sigmodontis were maintained in our laboratory and infective third-stage larvae (L3) were recovered by dissection of the mite vector Ornithonyssus bacoti [73] as previously described [74, 75].
Wt and Cxcr4+/1013 mice were bred in our animal facilities on a 12-hours light/dark cycle. All data were obtained from 8–12-week-old mice. Mice were genotyped by PCR on genomic DNA as previously described [41] using specific oligonucleotide primers to distinguish the mutant and the endogenous Cxcr4 allele.
Mouse infections were carried out by SC inoculation of 40 infective L3 in 200 μL of RPMI 1640 (Eurobio, France) into the left lumbar area of mice. Only when indicated, were mouse infections carried out by an intravenous inoculation of 40 infective L3 in 50 μL of RPMI 1640 into the caudal vein. In some experiments where indicated, mice were injected with crude extracts of L. sigmodontis worms which were obtained from the homogenization and sonication of L3 recovered from infected mites; the L3 derived crude extract concentration was determined by a Bradford assay (Pierce) following the manufacturer’s instructions. After centrifugation, the supernatant was collected and the protein content was determined by the modified Bradford method (BCA Protein Assay kit, Pierce). Mice then received 10 μg of crude extract either subcutaneously in 200 μL RPMI 1640 or intravenously in 50 μL of RPMI 1640 as above. Mice were sacrificed at 6 hours, 20 and 80 days post-inoculation (p.i.) as indicated below.
Blood smears and blood cell counts were performed before and throughout the challenge at different time points. Blood smears were obtained from the tail vein, stained with May-Grünwald-Giemsa (VWR, France) and the percentages of the different leukocyte populations were determined for 200 cells. Total blood cell counts were determined from tail blood mixed with 1% acetic solution (1:5 vol) using a hemocytometer (KOVA Glasstic Slide).
At the indicated time p.i., mice were anaesthetized and sacrificed by terminal bleeding. Blood was allowed to clot for 30 min at room temperature then centrifuged and sera were collected and kept at -20°C until further use. The pleural cavity was washed with 10 mL of cold phosphate-buffered saline (PBS, EUROBIO, France), as previously described [12]. The infiltrating cells and the worms were collected from the pleural wash for further analysis. Pleural washes were then frozen at -20°C until further use. Worms were fixed in toto with 4% formaldehyde in cold PBS to avoid body shrinkage and the gender and development stages of the worms were analyzed by light microscopy. PleCs were centrifuged at 250 g for 8 min at 4°C, resuspended in 2 ml RPMI supplemented with 2% foetal calf serum (FCS, EUROBIO, France) and then counted in PBS/ 0.04% trypan blue (Sigma-Aldrich) using a haemocytometer (KOVA Glasstic Slide). Proportions of the different leukocyte populations were determined by flow cytometry using the following rat anti-mouse antibodies: anti-F4/80-APC (clone BM8), anti-SiglecF-PE (clone E50-2440), anti-Ly6G-FITC (clone RB6-8C5), anti-CD3-PE (clone 145-2C11), and anti B220-FITC (clone RA3-6B2). All antibodies were purchased from eBioscience except the anti-SiglecF-PE (BD Pharmingen) and used at a 1/40 dilution. Flow cytometry analysis was performed using a FACSVerse flow cytometer running the FACSuite software (BD Biosciences). Acquisition and analyses were performed as described in S2 Fig.
Two days prior to the challenge, mice were anesthetised and depilated by means of Veet hair removal cream on an area of flank skin in the left inguinal region over the inguinal lymph node. For the challenge, wt and Cxcr4+/1013 mice were then inoculated either with 40 infective larvae or RPMI as a control and sacrificed 6 hours p.i. Skin sections of 1 cm2 were taken from the inoculation site, fixed overnight in 4% paraformaldehyde (PFA, VWR, France) and embedded in paraffin. Paraffin sections (5 μm) were stained with hematoxylin-eosin (H&E, VWR, France) or toluidine blue (VWR, France) for the detection of eosinophils or mast cells respectively. A modified Hematoxylin Eosin (HE) staining, including an alkaline eosin solution-staining step (pH 8.4) for 20 seconds has been selectively chosen to minimize background tissue eosin staining. For immunohistochemistry analyses, i.e. macrophages and neutrophils visualization, deparaffinized slides were incubated with the anti-F4/80 (clone CI:A3-1, Abdbiotech) or anti-NIMP-R14 (clone NIMP-R14, Abcam) antibodies respectively in 3% PBS-BSA overnight at 4°C. A peroxidase-based system was used for detection. For immunofluorescence staining, i.e. NETs, deparaffinized slides were incubated with anti-MPO (3.3 μg/mL, AF3667 R&D, Lille, France) or anti-Elastase (ELA) (5 μg/mL, Ab68672, Abcam, Paris, France) antibodies followed by staining with goat anti-rabbit Alexa Fluor 488 or anti-rat Alexa Fluor 594 (20 μg/mL, A-11034 and A-11007, respectively, Life Technologies, Saint-Aubin, France). DNA was visualized upon DAPI counterstaining. Images were acquired using the digital slide scanner HPF-NanoZoomer RS2.0 (Hamamatsu) coupled to a high definition 3-CCD digital camera. In some cases, images were acquired using an Olympus DP72 camera coupled to an Olympus BX63 motorized microscope running the cellSens Dimension (v 1.9) software.
Neutrophils were depleted by a single intra-peritoneal injection of either 0.25 mg of anti-Ly6G clone NIMP-R14 (AdipoGen) or 0.5 mg of anti-Ly6G clone 1A8 (BioXCell) 6 hours prior to infection (performed as described above in the Material and Methods Section 2). The two antibodies have been used extensively to deplete neutrophils in mice [51–53] with a potentially better selectivity for neutrophils of the 1A8 clone with regard to other Gr-1+ cells [51]. The effect of the treatment was assessed on blood and skin-resident neutrophils. Circulating neutrophil depletion was confirmed by blood smears and differential blood cell counts from 6 hours to 20 days post-injection of the depleting antibody. The effect of the depletion on skin-resident neutrophils was assessed by immunohistological analyses (see above, Material and Methods section 5) using the anti-NIMP-R14 antibody on paraffin-embedded skin sections of control or injected mice 6 and 12 hours post-injection of the depleting antibody.
Chemotaxis assays were performed on BM leukocytes using a Transwell assay as previously described [34]. Leukocytes were isolated from the BM of wt and Cxcr4+/1013 control or infected (at 20 days p.i.) mice and resuspended in assay buffer (HBSS medium supplemented with 20mM HEPES and 0.5% bovine serum albumin). Leukocytes (3 x 106 cells) were then added to the upper chamber of transwell filters (Millipore, 3μm pore diameter) that were placed in 24-well cell culture plates containing 300μl assay buffer with or without the indicated chemokines. In some experiments cells were pre-incubated at 37°C with Cxcl12 10 μM and AMD3100 200 μM for 30 min before being placed in the upper transwell chamber to confirm the specificity of the Cxcl12-dependent migration upon Cxcr4. Chambers were then incubated for 60 min at 37°C with 5% CO2 and the cells that migrated to the bottom chamber were recovered and stained with anti-ly6G-FITC antibody for flow cytometry analysis. The number of neutrophils that migrated into the bottom chamber was determined by a flow cytometer (BD Biosciences) with relative cell counts obtained by acquiring events for a set time period of 30s. Chemotactic indexes were then calculated by dividing the number of neutrophils that were counted in the chemokine stimulated well by the number of neutrophil that were obtained in the non-stimulated well.
Cell surface expression of Cxcr2 and Cxcr4 was also assessed on BM leukocytes from naïve and infected mice at 20 d.p.i. Neutrophils were identified with an anti-ly6G-FITC antibody and stained with either anti-Cxcr2-APC (clone 242216, R&D) and anti-Cxcr4-PE (clone 247506, R&D) antibodies or their respective control isotypes. Samples were processed by a FACSVerse flow cytometer (BD Biosciences) and analyzed using FACS Suite software. BM leukocytes from naive wt and Cxcr4+/1013 control mice were subjected to a discontinuous 72–64% Percoll density gradient centrifugation in 15 mL Falcon tubes for 30 minutes at 4°C. Mature neutrophils were collected at the 72–64% interface (purity > 93%), washed three times in cold PBS then resuspended in PBS at the working concentration of 106 cells/mL. The following assays were performed: The production of superoxide anions (O2-) was investigated using a microscopic NBT assay. In brief, 105 mature neutrophils per mouse were incubated in an 8-well Lab-Tek Chamber Slide for 1 hour at 37°C allowing the cells to attach to the plate. The supernatant was then discarded and 100 μL of NBT (1 mg/mL) was added to each well. Cells were incubated for 1 hour at 37°C and slides were analyzed by optical microscopy (Olympus BX63 microscope). The slide was counterstained 3 min with a 10% Giemsa solution then water washed. Black NBT deposits within the cells revealed the production of O2-. ROS content was measured via a dichloro-dihydro-fluorescin diacetate (DCFH-DA) assay (Sigma-Aldrich). Briefly, 105 mature neutrophils were incubated with 125 μM DCFH-DA in a 96-well cell culture plate for 15 minutes at 37°C to allow its entry into cells where it get converted in DCFH. Cells were then left unstimulated or stimulated with 20 L3 for 15 minutes at 37°C. Neutrophil oxidative response changes DCFH to green fluorescent DCF. The reaction was stopped by placing cells at 4°C. L3 were removed from the wells and cells were stained with anti-Ly6G-PECy7 antibody (eBioscience, clone RB6-8C5, dilution 1/40). Samples were processed by a flow cytometer (BD Biosciences) and analyzed using FACS Suite software. ROS content was expressed as an activation ratio by dividing the FITC mean fluorescence intensity (MFI) of neutrophils from L3 stimulated conditions by that from unstimulated cells.
MPO content was measured in wt or Cxcr4+/1013 mature mouse neutrophils (105 cells) either unstimulated or stimulated with 20 L3 in a 96-well cell culture plate at 37°C for 1 hour. Cells were then permeabilized with Saponin-PBS (0.2% BSA + 0.05% saponin in PBS) and stained with anti-Ly6G-PECy7 (eBioscience, clone RB6-8C5, dilution 1/40) and anti-MPO-FITC (Hycult Biotech, clone 8F4, dilution 1/40). Samples were processed by a flow cytometer and analyzed using FACS Suite software. MPO content was expressed as MFI.
Neutrophil necrosis and NETosis were quantified by immunofluorescence light microscopy in wt or Cxcr4+/1013 mature mouse neutrophils (105 cells) either left unstimulated or stimulated with 20 L3 in 24-well culture plates at 37°C for 4 hour. SYTOX-Green (Thermo Fisher Scientific), which does not enter into live cells, was then added (dilution 1:15000) for the detection of either extracellular DNA indicative of NETs release or intracellular DNA of necrotic neutrophils. Viable cells appear as non-fluorescent cells whereas both necrotic and NETs releasing neutrophils become fluorescent. Necrotic neutrophils are characterized by compromised membrane integrity and pluri-lobed nuclei whereas NETs releasing neutrophils display a halo (DNA area > 500 μm2) corresponding to chromatin decondensation and NETs release [60, 76]. Images were acquired using an Olympus DP72 camera coupled to an Olympus BX63 motorized microscope running the cellSens Dimension (v 1.9) software. Cells were counted and the numbers of necrotic, NETs releasing and living cells were determined.
Extracellular DNA was also quantified: mouse neutrophils (1x105) were cultured with 10 infective larvae or 100 nM phorbolmyristate acetate (PMA) for 4, 24 and 36 h in vitro. The culture supernatants were collected and the extracellular DNA was quantified using the Quant-iT PicoGreen dsDNA Assay kit (Life Technologies), following manufacturer instructions. Samples were cultured with the PicoGreen reagent (1:1 dilution) for 5 min. The samples were measured with a spectrofluorometer at 480 nm excitation and 520 nm emission. A DNA standard curve was used to determine the concentration of free DNA from samples
Statistical analysis was carried out using GraphPad Prism v5. Sample size, normality (Shapiro-Wilk test) and homoscedasticity (Bartlett’s test) were tested prior to further analysis. Data from separate experiments were pooled where possible. Most of the time one-way ANOVA analyses followed by a Bonferroni post-hoc test were performed. T-tests were used when comparing filarial load and lymphocytes, eosinophils and neutrophils cell counts at 20 days p.i. between wt and Cxcr4+/1013 infected mice. One-way ANOVAs with repeated measures were used to compare neutrophils cell counts in the blood of infected mice during the course of the infection. Significance was defined as *: p < 0.05; **: p < 0.01 and ***: p < 0.001.
|
10.1371/journal.ppat.1006419 | Competition between influenza A virus subtypes through heterosubtypic immunity modulates re-infection and antibody dynamics in the mallard duck | Our overall hypothesis is that host population immunity directed at multiple antigens will influence the prevalence, diversity and evolution of influenza A virus (IAV) in avian populations where the vast subtype diversity is maintained. To investigate how initial infection influences the outcome of later infections with homologous or heterologous IAV subtypes and how viruses interact through host immune responses, we carried out experimental infections in mallard ducks (Anas platyrhynchos). Mallards were pre-challenged with an H3N8 low-pathogenic IAV and were divided into six groups. At five weeks post H3N8 inoculation, each group was challenged with a different IAV subtype (H4N5, H10N7, H6N2, H12N5) or the same H3N8. Two additional pre-challenged groups were inoculated with the homologous H3N8 virus at weeks 11 and 15 after pre-challenge to evaluate the duration of protection. The results showed that mallards were still resistant to re-infection after 15 weeks. There was a significant reduction in shedding for all pre-challenged groups compared to controls and the outcome of the heterologous challenges varied according to hemagglutinin (HA) phylogenetic relatedness between the viruses used. There was a boost in the H3 antibody titer after re-infection with H4N5, which is consistent with original antigenic sin or antigenic seniority and suggest a putative strategy of virus evasion. These results imply competition between related subtypes that could regulate IAV subtype population dynamics in nature. Collectively, we provide new insights into within-host IAV complex interactions as drivers of IAV antigenic diversity that could allow the circulation of multiple subtypes in wild ducks.
| Many features of pathogen diversification remain poorly explored although host immunity is recognized as a major driver of pathogen evolution. Influenza A viruses (IAVs) can infect many avian and mammalian hosts, but while few IAV subtypes circulate in human populations, subtype diversity is extensive in wild bird populations. How do these subtypes coexist in wild avian populations and do they compete within these natural host populations? Here we experimentally challenged mallard ducks with different IAVs to study how an initial infection with H3N8 determines the outcome of later infections (duration of infection and virus load) and antibody responses. There was complete protection to re-infection with the same H3N8 virus based on virus isolation. In addition, there was partial protection induced by H3N8 pre-challenge to other subtypes and development of heterosubtypic immunity indicated by significantly shorter infections and reduction in viral load compared to controls. This indicates that subtype dynamics in the host population are not independent. Amongst H3N8 pre-challenged groups, the highest protection was conferred to the H4N5 subtype which was most genetically related to H3N8. The H4N5 challenge also induced an increase in H3 antibody levels in that challenge group and evidence for original antigenic sin or antigenic seniority. Thus, previous infections with IAV can influence the outcome of subsequent challenge with different IAV subtypes. These results not only have relevance to understanding naturally occurring subtype diversity in wild avian populations but also in understanding potential outcomes associated with introduction of novel viruses such as highly pathogenic IAV H5 viruses in wild bird populations.
| Diversification is a common feature in pathogen populations and this often involves the evolution of antigenic variants [1]. Examples of this exist in various pathogen systems: viruses (influenza A virus, Dengue virus, Bluetongue virus, and rotaviruses), bacteria (Borrelia spp, Neisseria meningitis, and Pneumococcus) and protozoan parasites (Plasmodium spp. and trypanosomes). Antigenic variation within specific hemagglutinin (HA) and neuraminidase (NA) subtypes is well described with influenza A viruses (IAVs) as this is an important consideration in developing vaccines and vaccination strategies associated with seasonal influenza viruses in humans and IAV affecting domestic livestock and poultry. Though multiple IAV subtypes circulate in these host systems, antigenic interactions between these subtypes are equally important but less understood. Shared epitopes between different HA subtypes associated with the HA stalk have been described and these may be important target epitopes for universal IAV vaccines. Immunity to these shared epitopes also may provide partial protection in subsequent infections with heterologous IAV [2] and could potentially influence clinical outcome and regulate subtype diversity in host populations through competition [3].
IAV have the capacity to infect many different hosts, from birds to mammals; however the vast majority of influenza subtype diversity is found in wild birds, especially waterfowl, gull, and shorebird populations where low-pathogenic IAV (LPIAV) representing 16 HA and 9 NA subtypes are maintained [4]. In these wild bird populations, many HA/NA subtype combinations co-circulate and it is notable that their abundance and relative diversity appears to vary over time and space [5–7]. In addition, individual birds are often co-infected or sequentially infected with multiple IAV subtypes in a given season or year [8, 9] and naturally infected mallards (Anas platyrhynchos) in the wild exhibit patterns of re-infections indicating heterosubtypic cross-immunity [9]. Experimental infections have demonstrated that initial infection with a specific IAV induces immune responses and protection against infection with homologous strains [10]. Additionally, initial viral infections could induce a partial protection to heterologous subtypes in experimental settings [10–13] and several studies have shown that LPIAV pre-exposure protects against a lethal highly pathogenic IAV (HPIAV) challenge [11, 14, 15]. Despite these observations little is known about the mechanisms, extent (strength and specificity) and persistence of these immune responses and the outcomes of re-exposures that have critical significance to understanding the maintenance of IAV antigenic diversity in multi-strain/subtype-pathogen systems such as occur with IAV and ducks.
In this study, mallards were pre-challenged with a strain of H3N8 LPIAV, one of the most common subtypes in waterfowl, and pre-challenged ducks were subsequently re-challenged with the same strain (homologous challenge) or with different strains representing various subtypes (heterologous challenges). Groups of ducks in the homologous challenge were exposed at different time intervals to evaluate long-term antibody responses and potential protection as well as to further investigate the potential influence of age [16]. The different strains used in the re-challenge were chosen to represented a gradient in the degree of phylogenetic relatedness between the HA and belonged to subtypes commonly found in waterfowl populations.
The objective of the present study was to experimentally mimic re-exposures that commonly occur in mallards and other ducks in nature to determine the effects of subsequent challenge on susceptibility, duration and intensity of viral shedding and to characterize the humoral immune response.
Here, we studied the persistence of protection to homologous IAV challenge by initially infecting mallards with H3N8 and re-challenging three groups with the same virus at different time intervals. The initial H3N8 induced long-term protection against homologous re-infection for up to 15 weeks post-challenge, which is much longer than expected. None of the re-challenged birds, regardless of time interval, shed virus as determined by virus isolation and that was true for all age groups and time intervals. Although some RRT-PCR positives were detected in re-challenged individuals, the Ct-values were high and significantly higher than in control groups indicating a low number of RNA copies possibly associated with non-infective virus. The observed protection and detectable antibody response is inconsistent with results from previous studies that concluded that infection conferred no protection against re-infection and that detectable antibody responses in ducks were short-lived [19, 20]. It was considered that antibody responses may not be detectable or protective due to the truncated structure of some of the IgY forms. A possible explanation of this inconsistency is that these truncated antibodies neutralize IAV but lack Hemagglutinin Inhibitory (HI) activity and the proportion between these forms vary over time being the truncated from most prevalent later in the immune response [21]. Our results are in agreement with other studies in both ducks and gulls that reported partial to complete protection against re-infection depending on the viruses, time between infections, host species and age and detection method [10, 22, 23]. The long-term persistence of antibodies after natural infection, artificial challenge or vaccination has been reported in captive birds up to 6 to 9 months post-exposure [24–26] though it is not know which parameters (i.e. HI or MN titers) are correlated to protection. When examining patterns in H3N8 infections for the different age classes in the naïve controls, all individuals were susceptible and competent to infection. The duration of infection and viral load in controls varied according to age and birds of 5 and 9 weeks of age experienced the highest replication based on AUC and duration of infection. This is consistent with earlier findings and indicative of older birds being more resistant to infection and having shorter infections [16]. This age effect could influence subsequent transmission risk and subtype diversity in cases were subtypes display seasonal patterns.
Results from IAV heterologous challenge experiments of ducks, geese and gulls have reported varying levels of partial protection upon re-challenge (11, 29, 30, 35). Such protection has also been reported from field-based studies [9, 27] and from experimental work where birds were challenged with HPIAV and effects could be measured by morbidity/mortality responses[11, 14]. It has also been reported that the order in which IAV subtypes are used to challenge the host is important. For instance, H3 appears to be poorly immunogenic and less protective as a primary infection compared to H4 that induced a stronger protective response [13]. The variation in reported protection from previous studies may be the result of asymmetric responses associated with strain or subtype specific variation that may be dependent on the virus causing primary infections [12, 13, 28].
We observed a partial protection to re-challenge in the individuals subsequently infected with heterologous subtypes indicating development of heterosubtypic cross-immunity by H3N8. When assessing the outcome of re-challenge there was a significantly lower viral load and shorter duration of infection in CL samples for all the groups pre-challenged with H3N8 compared to the naïve controls. This shorter period of detectable viral shedding is in agreement with estimates from the field [29, 30] and was observed for all re-challenged viruses.
The extent of partial protection, as measured by a decrease in duration of detectable viral shedding following re-infection was influenced by the genetic relatedness of the HA. This suggests immune mediated competition through cross-reactive responses [31, 32] and this could be related to both acquired humoral and cell mediated mechanisms [33]. Broadly neutralizing antibodies have been defined across HA groups [34] and within HA group [35, 36] that target conserved epitopes in the stalk [37, 38]. HA stalk antibodies are boosted upon re-infection and it is therefore thought that cross-reactive anti-stalk as well as anti-NA antibodies [39] can diminish the severity of disease in re-infections. Studies in humans have found consistent patterns of cross-immunity within HA group with increased severity of the disease in cohorts exposed with an HA of the opposite group during childhood when studying age-specific mortality caused by 1918 H1N1 and by HPIAV H5 and H7 [40, 41]. It has been proposed that extinction of influenza strains in humans could be driven by population immunity by HA stalk and NA antibodies [39] through competitive exclusion between strains [3]. Moreover, initial infections with a specific virus increases the probability of later infections by viruses from a different HA clade and group in wild mallards [9]. Comparable processes are likely acting in the avian IAV system where lineage replacement with strains from different continents has been reported [42, 43]. Theory predicts that antigenic variants tend to organize as discrete non-overlapping strains [1] in populations where cross-reactivity between viruses induces competition (i.e. case of H13 and H16) contrasting to the situation where cross-reactivity induces enhancement or facilitation, like Dengue, and variants may be antigenically clustered [44]. Here, after the heterologous re-challenge mallards shed viruses even if they were able to clear infections rapidly. An implication of this is the potential for selective processes like viral escape and antigenic drift to act in the same way as leaky vaccines [45, 46] which in turn could drive antigenic evolution as observed for H5N1 HPIAV [47] or similarly for seasonal H3N2 in humans [48]. Indeed, the estimates of divergence for some HA subtype indicate that divergence is relatively recent [40].
Isolation success from RRT-PCR positive samples from secondary infections (re-challenge), as expected, was correlated with Ct-value, but was significantly lower compared to isolation results from primary infections in controls. This discrepancy has implications when interpreting field data and assigning infection states based on molecular diagnostics rather than virus isolation as uncertainty needs to be incorporated [49] and may explain why isolation rates from PCR positive samples often vary. This relationship also corroborates findings from the field where samples from adults have a lower isolation success than samples from juveniles [50] and where RRT-PCR positive samples used in experimental trials have not infected ducks [51]. It is therefore prudent to be cautious when using transformations of Ct-values (proxy for RNA copies) to EID50/ml or TCID50/ml by a standard curve based on a virus grown at optimal conditions such as cell culture or embryonated eggs as infectivity in hosts varies according to several parameters (age of host, previous exposure, specific virus and infection dose…).
Antibody levels (NP and H3 MN) after H3N8 infection decreased over time but most individuals remained positive for the duration of the experiment (15 weeks); all birds remained protected against homologous challenge. There was a significant boost in the antibody responses after homologous challenge for the long-term groups (11 and 15 weeks) but not for the group re-infected after 5 weeks where a rapid blocking of the infection may not have activated antibody recall. Additionally, the hyperimmune sera after homologous challenge did not cross-react with other IAV subtypes, except for one individual positive by H1 at a low titer of 20.
The heterologous re-infections resulted in a boost in the NP-antibody responses in all H3 pre-challenged groups. Interestingly most of the individuals had neutralizing antibodies against the HA antigens they had been exposed to. This also includes the H12N5 that replicated poorly but resulted in serological imprinting in three of the five individuals per group. Additionally no cross-reactivity to other subtypes was observed when tested by MN with a panel of HA (H1-H12 and H14-H15). H3-specific antibodies were detected by MN after H3N8 primary infection and persisted in the majority of individuals until the end of the experiment. There were interesting H3 patterns of antibody dynamics for the H3N8 X H4N5 group that showed a boost in H3 titer (Fig 7) consistent with original antigenic sin (OAS); H4 antibodies were also detected in this challenge group (Fig 8). This phenomenon of interference was first described after sequential influenza re-infections in humans [52, 53]. Older individuals can have a broader immunity through repeated exposure with highest titers to the strains individuals were exposed to early in life. These “senior strains”, are a singularity known as antigenic seniority or imprinting [40, 54]. Currently, we report OAS in avian hosts and between different HA subtypes (H3/H4 and possibly for H3/H10), however we expect that phenomenon could also arise between other IAV subtypes. Thus the order in which individuals are challenged with a specific virus could influence the future recognition of viruses and the specificity of the responses that ultimately shapes the outcome of later exposures in life. Since population immunity influences the emergence and spread of new strains and can influence vaccination success, it is critical to have a better understanding of these processes in different host species. We need to increase our understanding of cross-reactivity patterns and boost dynamics in re-infections to ultimately predict risks of IAV spread into different host populations in a context of non-naïve populations, for instance by using antibody landscapes [55].
The high degree of heterosubtypic immunity and subsequent competition found between common HA subtypes from ducks indicates that the transmission success and perpetuation of any subtype is dependent on the other viruses in the population and may explain the cyclic or chaotic nature in the prevalence of some subtypes. Partial immunity or complete immunity induced by pre-infection may reduce transmission potential in subsequent infections but also may promote the high degree of IAV antigenic diversity observed in wild avian populations. This competition may also result in subtype succession over time, like the predominance of H3 and H4 in fall migration [5–7] and spring blooming of other subtypes within Group 2 such as H7 or H10 [56]. With equal strength of HA immunity to all subtypes the present antigenic diversity found in wild birds would be unlikely. Surprisingly, the results from H13 and H16 experimental infections in black-headed gulls showed limited cross-immunity between subtypes and suggest independent cycles for these viruses [22].
However, for some strains or subtypes, cross-immunity may not be the only factor explaining their dynamics in the pathogen population. Pathogen fitness or success is usually measured by transmission risk that could be based on different transmission parameters such as the duration of infection and pathogen load [57] as well as how long viruses could remain infectious in the environment [58, 59]. We believe that virus fitness linked to host-specificity and functionality of IAV proteins [60] is also playing an important role as evidenced by the fact that the H12N5 IAV used in this study did not successfully replicate in mallards even though it was isolated from that host.
A possible interpretation of these results is that population immunity can reduce the probability of transmission and potential introduction success of exotic antigenic variants [61]. Previous studies have reported protection to HPIAV induced by pre-exposure to LPIAV strains in different bird species [11, 14]. In the context of H5N8 HPIAV clade 2.3.4.4 or other HPIAV the present results suggest that cross-immunity could also reduce viral shedding and contribute to stopping the spread of specific virus in wild ducks that have naturally been exposed to LPIAV [62–64].
We believe that the competitive processes described here and in other studies occur in nature; however, in natural host populations the complexity of the system increases due to the extensive subtype diversity of co-circulating viruses that these birds are continuously exposed to. Based on our results we propose that the extent of competition at individual level through host immunity could be determined by many different interacting parameters: the strains involved in infections, the exposure history (or boost responses like OAS) and likely time between exposure/s (assuming that immune memory decreases over time). These are many of the same factors that are considered in evaluating immune responses and protection against influenza in humans and domestic animals.
All LPIAV viruses used in these trials represented common North American subtypes and were originally isolated from wild mallards in Minnesota, USA [6]. Viruses included: A/Mallard/MN/AI07-4724/2007 (H3N8), A/Mallard/MN/AI11-4213/2011(H4N5), A/Mallard/MN/AI11-4982/2011(H6N2), A/Mallard/MN/AI11-4412/2011(H10N7) and A/Mallard/MN/AI11-3866/2011(H12N5). Virus stocks were propagated by a second passage in specific pathogen free (SPF) embryonated chicken eggs (ECE). Stocks were endpoint titrated using the Reed and Muench Method [65] in ECE to determine the median egg infectious dose (EID50/0.1 ml).
Ducks were inoculated with 0.1ml at an approximate dose/dilution of 106.0 EID50/0.1ml; the inoculum was split between intranasal and oropharyngeal routes. Previous work has shown that different routes of infection (intranasal, intratracheal, intraocular, intracloacal, or intra-ingluvial) result in 100% infection and similar shedding patterns [66]. Based on back titration, the titers of the inocula, expressed in EID50/ 0.1 ml, were 105.8 for H3N8 pre-challenge, 106.5 for H12N8 and H6N2, 106.4 for H3N8 (5 week challenge), 106.8 for H10N7, 106.0 for H3N8 (11 week challenge), 106.2 for H4N5 and 106.3 for H3N8 (15 week challenge).
The HA and NA segments of the isolates were amplified [67] and later sequenced using Sanger method. Geneious (version mac6_4_8_0_4) was used for sequence alignments and to calculate amino acid distances. Sequences are publicly available in GenBank (KX814369-KX814375).
All procedures were in accordance with the Animal Welfare Act and US regulations. All protocols for raising, infecting, and sampling mallards were approved by the Institutional Animal Care and Use Committee at the University of Georgia (UGA; AUP#: A2013 05-021-Y1-A1).
One-day-old mallards (Murray McMurray Hatchery, Webster City, IA, USA) were raised in captivity at the animal resources facilities of the College of Veterinary Medicine, UGA. Food and water were supplied ad libitum. All individuals were identified with bands and unique ID numbers. Individuals did not exhibit behavioral changes or overt symptoms of disease during the duration of the trial. All birds gained weight before and throughout the challenge studies. At four weeks of age, 40 ducks were moved to Biosafety Level (BSL) 2+ facilities and were pre-challenged with the H3N8 LPIAV.
The study design included re-challenges with homologous and heterologous viruses (S1 Table). Prior to the re-challenges, ducks were randomly divided into groups (five individuals per group, approximate ratio 1:1 females: males) and were moved into BSL 2+ poultry isolators at the Poultry Diagnostic Research Center, Athens, GA, USA that are intended to house up to five adult size mallards. After 2 days of acclimation, the re-challenge was conducted as previously described [12]. Each pre-challenged group had a control group (i.e. age matched naïve individuals). Ducks were humanely euthanized at the end of all challenge trials at 14 dpi following protocols approved at the UGA. For the homologous challenge, one group of five ducks was challenged with the same H3N8 at five weeks after the initial H3N8 pre-challenge and two additional groups were re-challenged with the H3N8 virus at weeks 11 and 15 post-H3N8 inoculation. Groups in the heterologous challenges were inoculated five weeks after initial H3N8 infection with subtypes representing different levels of relatedness between the key antigenic proteins HA and NA of the H3N8 used in the pre-challenge: H4N5, H10N7, H6N2 and H12N5.
Swabs were collected from the cloaca (CL) and oropharynx, (OP) and were placed in separate tubes containing 2 ml BHI transport media supplemented with antimicrobials [12]. Swab samples were collected on 0–8, 10, 12 and 14 and 21 (only after H3N8 pre-challenge) days post infection (dpi) and kept cold until transfer to the laboratory where they were stored at -80 C° until processing.
Virus isolations was performed on swabs from all birds during all challenges on 0, 2, 4, 6, 8, 10, 12, 14 dpi as previously described [68] using two 9 to 11 days-old SPF ECE. To insure that all birds were IAV negative before any bird movement or subsequent challenge, all birds were tested by virus isolation on 21 dpi after the H3N8 pre-challenge and on 0 dpi prior to all subsequent viral challenges.
Viral RNA from samples collected on 1 to 14 dpi were extracted using the MagMAX-96 AI/ND Viral RNA Isolation kit (Ambion, Austin, TX, USA) on the Thermo Electron KingFisher magnetic particle processor (Thermo Electron Corporation, Waltham, MA, USA) [68]. Negative (BHI media) and positive (diluted isolate) controls were included in each extraction. Molecular detection of the IAV Matrix gene by Real-time Reverse Transcriptase PCR (RRT-PCR) [17] was conducted with the Qiagen OneStep RT-PCR kit (Qiagen, Valencia, CA, USA) and the Cepheid SmartCycler System (Cepheid, Sunnyvale, CA, USA) [68]. Negative and positive controls were used on the extraction step and an IAV Matrix gene transcript (National Veterinary Services Laboratory, Ames, IA, USA) was included as positive control in the RRT-PCR. Ct–value stands for the cycle threshold when there is initial detection of the fluorescence signal at the beginning of the exponential phase of DNA duplication and is proportional to the initial number of RNA copies in a sample. Ct-values were therefore used as a proxy for viral load (S4 Fig). By using the same detection protocol and stock of the Matrix gene transcript (NVSL, Ames, IA, USA) it was previously established that Ct-values significantly correlate with the number of Matrix RNA gene copies when diluting the Matrix gene RNA transcript to generate a standard curve [57]. Samples with cycle (Ct) values lower than 40 were considered positive
Serum samples were collected from the right jugular vein and transferred into serum separator tubes (Becton, Dickinson and Company, Franklin Lakes, NJ, USA), centrifuged upon arrival to the lab and serum was stored at -20 C° until analysis. Serum samples were taken immediately prior to H3N8 pre-challenge and prior to all subsequent challenges as well as at 14 dpi and between challenges for the homosubtypic long-term groups. Serum samples were tested with a commercially available nucleoprotein (NP)-ELISA kit (bELISA; FlockChek AI MultiS-Screen antibody test kit; IDEXX Laboratories, Westbrook, ME, USA). Specific antibodies against the different HA subtypes used in the trial were detected using a virus microneutralization (MN) assay in Madin Darby Canine Kidney cells (MDCK; ATCC, Manassas, VA, USA) as described previously [61]. Sera were additionally tested using the same viruses used for inoculation in the challenge and with prototype strains for H1-H12 and H14-H15 [69] (S1 Appendix) to detect cross-reactivity. These antigens also were prepared in MDCK and tests were run using an antigen concentration of 100 TCID50 /25 μl.
All analyses were run on the R software [70] using the GAMM, nlme and lme4 packages. Model selection was done using AIC [18] corrected for small sample size (AICc) within the package AICcmodavg. To evaluate the viral shedding or load, Ct-values from the Matrix RRT-PCR runs that are proportional to the RNA copy numbers from original samples were used. First, we analyzed the variation in Ct-values in pre-challenged and naïve controls for the different treatment groups with different time periods between infections. The strategy was to use linear models and include individuals as random effect using mixed models (package GAMM) due to repeated sampling of the same individuals over the course of infection. The models that were evaluated included the factors: dpi, treatment group, both factors and the interaction. Ct-values from 1 dpi were not included in the analysis as they likely represented residual inoculum rather than true virus replication (at 1 dpi the Ct-values were close to 40). Models including the random effect of individual and the additive effect of dpi were tested but the increased complexity of the models was penalized based on the AICc and some of them had convergence problems. In the same way, the variation of Ct-values from all H3N8 control groups including age as factor was assessed. For the heterologous re-challenge, the variation in Ct-values from control and pre-challenged groups for each of the different viruses was evaluated as described before. The Area Under the Curve (AUC) was calculated by using the Ct-values after subtracting them from the cut-off value of 40 as previously done [22]. The total duration of infection or shedding was estimated by counting the days between inoculation and last positive virus isolations in CL swabs (which also includes cases of intermittent shedding). Birds that died or were euthanized before 14 dpi were not included in analysis. AUC and duration of infection between groups was compared using Krustal-Wallis test.
To study the correlation between AUC and duration of infection the Pearson correlation test was performed followed by a linear regression. To explore the relation between HA similarity and the degree of protection, we estimated the amino acid distance between the H3 and the different HA of re-challenges and the relative reduction in the duration of infection within groups. The amino acid distance was then correlated with the reduction in shedding per group where the mean and SE were estimated using a bootstrapping approach.
Next, to explore the influence of different variables on isolation success (binomial response: successful or unsuccessful) we used Generalized Linear Mixed Models (GLMM) as in [50]. The explanatory variables included in the models were: Ct-value, sample type (OP and CL swabs), dpi and treatment (pre-challenged or naïve individuals which means samples from a primary or secondary infection). Additionally, since individual birds were re-sampled and samples are not independent we added the individual as a random effect in the models. Interactions were not included to avoid convergence problems.
Last, MN antibody titers (as log2 transformed) at a single day of sampling were compared between groups using the Krustal-Wallis test and the paired t-test was used to compare values in two different days within groups. Samples with an MN titer lower than 20 were arbitrarily given a titer of log2 (2.5) for the model testing. The variation in antibody titers to H3 by MN was explored based on the different sampling occasions (dpi) and groups also using Generalized Additive Mixed Models in the package GAMM.
One duck died (12 dpi in group H6N2) and swabs were already IAV negative; necropsy showed no gross lesions caused by LPIAV. Another duck was euthanized due to difficulty walking which was associated with husbandry in captivity (6 dpi in group H3N8 X H6N8). These birds were not included in the analysis. The H3N8 x H12N5 group was not included in the analysis because of poor replication of this virus in the pre-exposed and naïve groups. All relevant data are available as Supporting Information files (Supplementary tables and S1 Dataset).
|
10.1371/journal.ppat.1003852 | The CD8-Derived Chemokine XCL1/Lymphotactin Is a Conformation-Dependent, Broad-Spectrum Inhibitor of HIV-1 | CD8+ T cells play a key role in the in vivo control of HIV-1 replication via their cytolytic activity as well as their ability to secrete non-lytic soluble suppressive factors. Although the chemokines that naturally bind CCR5 (CCL3/MIP-1α, CCL4/MIP- 1β, CCL5/RANTES) are major components of the CD8-derived anti-HIV activity, evidence indicates the existence of additional, still undefined, CD8-derived HIV-suppressive factors. Here, we report the characterization of a novel anti-HIV chemokine, XCL1/lymphotactin, a member of the C-chemokine family that is produced primarily by activated CD8+ T cells and behaves as a metamorphic protein, interconverting between two structurally distinct conformations (classic and alternative). We found that XCL1 inhibits a broad spectrum of HIV-1 isolates, irrespective of their coreceptor-usage phenotype. Experiments with stabilized variants of XCL1 demonstrated that HIV-1 inhibition requires access to the alternative, all-β conformation, which interacts with proteoglycans but does not bind/activate the specific XCR1 receptor, while the classic XCL1 conformation is inactive. HIV-1 inhibition by XCL1 was shown to occur at an early stage of infection, via blockade of viral attachment and entry into host cells. Analogous to the recently described anti-HIV effect of the CXC chemokine CXCL4/PF4, XCL1-mediated inhibition is associated with direct interaction of the chemokine with the HIV-1 envelope. These results may open new perspectives for understanding the mechanisms of HIV-1 control and reveal new molecular targets for the design of effective therapeutic and preventive strategies against HIV-1.
| Although HIV, the causative agent of AIDS, establishes a lifelong infection that cannot be eradicated even with effective treatment, the host immune system has the ability to contain its replication for many years in which the disease remains asymptomatic. Key players in HIV control are CD8+ T cells, specialized immune cells that can not only destroy infected cells, but also secrete soluble factors that suppress the virus without killing infected cells. CD8+ T cells produce multiple HIV-suppressive factors, including certain chemokines (soluble proteins that attract immune cells), which block the virus even before it can gain access to its target cells. In the present study, we characterize a new anti-HIV chemokine, XCL1 or lymphotactin, which is primarily produced by CD8+ T cells. A unique feature of XCL1 is that, unlike other antiviral chemokines, it has a very broad spectrum of activity against different variants of HIV-1 and directly binds the virus outer coat, rather than blocking specific receptors on the target cell. Also unique is that fact that XCL1 adopts two possible conformations, and only one of them is capable of HIV inhibition. These findings may open new avenues for the design of effective drugs or vaccines against HIV.
| The replication of HIV-1 is regulated in vivo by a complex network of cytokines and chemokines expressed by immune and inflammatory cells. Key players in the mechanisms of HIV-1 control are CD8+ T cells, which, in addition to their cytolytic activity, secrete soluble factors that suppress HIV-1 in a non-lytic fashion [1]–[5]. Following the initial observation of this latter phenomenon in 1986 by Walker and colleagues [4], subsequent studies demonstrated HIV-1 inhibition in co-cultures of CD8+∶CD4+ T cells separated by a semi-permeable membrane, as well as in cell-free supernatants from activated CD8+ T cells [1], [3], thus, ruling out the need for cell-to-cell contact. Moreover, the ability of CD8+ T cells to suppress HIV-1 replication was shown to correlate with the clinical stage of HIV-1 infection, suggesting a potential in vivo protective effect of this non-lytic CD8+ T cell activity [5]. Approximately 10 years after the initial description of soluble CD8+ T cell-derived inhibition of HIV replication, three chemokines of the CC (β) chemokine family (CCL3/MIP-1α, CCL4/MIP- 1β, CCL5/RANTES) were identified as major components of the soluble CD8+ T cell-derived anti-HIV activity [2]. These three chemokines act via a redundant mechanism of binding and downmodulating CCR5 to block entry of viruses with CCR5 coreceptor tropism. However, multiple lines of evidence indicate the existence of additional, still undefined, CD8-derived factors that can suppress HIV-1 infection. In particular, the observation that CD8+ T-cell culture supernatants can also inhibit CC-chemokine-resistant HIV-1 strains, such as those restricted to CXCR4 coreceptor usage [6], [7], substantiates a role for new, still uncharacterized anti-HIV factors produced by CD8+ T cells. Additionally, several reports have documented a suppressive effect of these factors at the transcriptional level [8]–[10], whereas CCR5-binding chemokines act at the level of viral entry/fusion.
In addition to CD8+ T cells, other cells of the immunohematological system can produce soluble HIV-suppressive factors, including CD4+ T cells, γ/δ T cells, NK cells, cells of the mononuclear phagocytic system, and platelets [11]–[13]. Recently, we identified a novel antiviral chemokine, CXCL4/PF4, which is mainly produced by megakaryocytes and platelets. CXCL4 was shown to inhibit a broad spectrum of HIV-1 isolates, irrespective of their coreceptor usage and genetic subtype; it acts at the level of viral entry via a novel mechanism mediated by direct interaction with the viral envelope [14].
In this study, we report the characterization of a novel anti-HIV-1 C-chemokine, XCL1, which exhibits a broad spectrum of activity against different biological variants of HIV-1. We present evidence that this chemokine blocks infection at an early step of the viral life cycle, namely, viral attachment and entry into host cells. Similar to our previous work with CXCL4/PF4, we found that XCL1 acts through an unconventional mechanism mediated by direct interaction with the HIV-1 envelope. Moreover, we investigated the correlation between the unique metamorphic properties of XCL1 [15] and its antiviral activity, showing that the alternatively-folded (all β-sheet) structure of XCL1 is the specific conformation responsible for HIV-1 blockade. These results offer insights into pathogenesis and provide new molecular targets for HIV-1 therapy and vaccine development.
As multiple lines of evidence indicated the existence of still uncharacterized HIV-suppressive factors produced by CD8+ T cells [16], we used a wide-platform cytokine array (RayBio), which evaluates in a semi-quantitative fashion 507 soluble factors, to screen culture supernatants from activated primary human CD8+ T cells. Among the top 25% most expressed proteins, we identified the C-chemokine XCL1/lymphotactin, in addition to other previously reported anti-HIV chemokines (Guzzo et al., in preparation). We focused our attention on XCL1 because it is produced preferentially by CD8+ T cells [17], [18]. Production of XCL1 by primary human CD8+ T cells was confirmed by ELISA in culture supernatants from CD8+ T cells activated ex vivo with either PHA, PMA plus ionomycin, or anti-CD3/CD28 antibodies (Figure 1). PMA plus ionomycin was the most potent stimulation for XCL1 production, followed by anti-CD3/CD28 antibodies, while PHA elicited the secretion of markedly lower levels.
To assess the ability of XCL1 to inhibit HIV-1 infection, acute infection assays were initially performed with recombinant human XCL1 (Peprotech) in primary human PBMC infected with laboratory-adapted viral strains. XCL1 potently inhibited HIV-1 infection irrespective of coreceptor specificity, as it was equally effective on strains specific for CXCR4 (X4; IIIB) and CCR5 (R5; BaL) (Figure 2A). To further examine the breadth of XCL1-mediated inhibition, we evaluated the ability of recombinant XCL1 to inhibit infection of primary PBMC with a panel of primary HIV-1 isolates with different coreceptor-usage specificity. As seen with the laboratory-adapted HIV-1 strains, XCL1 was equally effective on CCR5-specific and CXCR4-using primary isolates (Figure 2B). Of note, XCL1 did not reach 90% inhibition on two R5 isolates even at the highest dose used (1 µM). Treatment with XCL1 did not reduce cell viability, indicating that the HIV-inhibitory effect of XCL1 was not due to toxic or anti-metabolic effects on the cells (data not shown). The anti-HIV-1 activity of XCL1 was also confirmed in the engineered cell line TZM-bl [19]–[21], a HeLa cell derivative expressing CD4, CXCR4, and CCR5 (data not shown).
XCL1 is a metamorphic chemokine that interconverts in solution between two distinctly folded structures, namely the canonical chemokine fold (three-stranded anti-parallel β-sheet and C-terminal helix), which has been reported to bind and activate the specific XCL1 receptor (XCR1), and an alternative, four-stranded sheet that forms a dimeric β sandwich, which was reported to bind glycosaminoglycans (GAGs), but not XCR1 [15]. Thus, we tested two stabilized recombinant XCL1 variants produced in E. coli: CC3, a variant locked in the chemokine-folded structure [22], and W55D, a variant that preferentially adopts the alternatively-folded dimeric structure [15]. As a control, we also tested a full-length, wild type (WT) XCL1, which retains the ability to interconvert between protein folds. A dual-tropic primary HIV-1 isolate sensitive to XCL1-mediated inhibition, 92HT599, was used for acute infection assays to evaluate inhibition with XCL1 variants. As shown in Figure 3, we observed striking differences in the inhibition exhibited by the locked XCL1 variants, CC3 and W55D. Indeed, the W55D variant inhibited HIV-1 with similar potency, as did the WT chemokine, while the CC3 variant showed no appreciable inhibition at the doses used. We observed similar results in infection assays performed with both CXCR4-tropic (IIIB) and CCR5-tropic (BaL) HIV-1 isolates, with the all-β, alternatively-folded conformation (W55D) conferring inhibition, and the chemokine-folded conformation (CC3) showing minimal, if any, activity (Figure S1). These results suggested that the antiviral activity of XCL1 is dependent on the protein conformation and appears to be unrelated to activation of the XCR1 receptor, which is specifically bound and activated by the CC3 variant, but not by the W55D variant. In line with this observation, we were unable to detect XCR1 expression by flow cytometry in the HIV-1 target cells used in our study (data not shown).
Since we established that the anti-HIV activity of XCL1 is associated with XCL1 adopting the all-β (alternatively-folded) conformation, which does not bind and activate the XCR1 receptor but is capable of interacting with GAGs [23], we focused on the extracellular events in the HIV-1 replication cycle and assessed the ability of XCL1 to interfere with viral attachment and entry. To achieve a high level of consistency of these assays, experiments were performed using TZM-bl cells infected with the primary dual-tropic isolate, 92HT599, which is highly sensitive to XCL1-mediated inhibition. As seen in Figure 4A and B, the WT chemokine and the W55D variant effectively blocked viral attachment and entry, while the CC3 variant had no appreciable inhibitory effect, reflecting the pattern of inhibition seen in the infection assays (Figures 2 and 3). In parallel, we examined the inhibitory effects of two other anti-HIV chemokines, namely CXCL4/PF4 and CCL5/RANTES. In accordance with our previous work, CXCL4 effectively inhibited viral attachment and entry [14], whilst CCL5 had enhancing effects, as previously observed with other CXCR4-tropic HIV-1 strains [24]. Additional controls included T-20, a well-characterized fusion inhibitor [25], which showed no effect on viral attachment, but a significant reduction in viral entry, and an anti-CD4 monoclonal antibody (mAb), which showed only a slight reduction in attachment but a very marked inhibition of viral entry (Figure 4A,B).
Since we documented an inhibitory effect of XCL1 at the level of HIV-1 attachment and entry, we examined the ability of XCL1 to downmodulate the main HIV-1 cellular receptors, namely, CD4, CXCR4 and CCR5. Flow cytometry did not reveal any change in surface expression of these receptors after XCL1 treatment for 24 h (data not shown). Furthermore, we also investigated interactions between XCL1 and CD4 via binding assays with anti-CD4 antibodies targeting different domains of CD4 (D1, D2 and D3–4); we did not observe any modification in CD4 staining, suggesting that XCL1 does not interact with cell surface-expressed CD4 (data not shown).
In view of the data described above, we investigated the possibility that XCL1 may interact directly with HIV-1 virions, as we previously demonstrated for CXCL4/PF4 [14]. To investigate this hypothesis, we performed a virion capture assay by which immunomagnetic beads were armed with different XCL1 variants (WT, W55D, or CC3) as molecular “baits” to capture whole HIV-1 virions, as previously described [14]. Figure 5A shows that both WT and W55D efficiently captured HIV-1 virions. The specificity of this interaction was validated upon observation of reduced capture when XCL1-armed beads were pre-incubated with anti-XCL1 mAb or polyclonal antibody (pAb) prior to exposure to the virus. In agreement with our infection and attachment/entry assays, we did not observe any appreciable virion capture when the beads were armed with the CC3 XCL1 variant. Our data demonstrate that XCL1 can directly interact with HIV-1 virions, and that the all-β (alternatively-folded) XCL1 conformation (W55D) mediates this interaction, while the classic chemokine-folded conformation (CC3) does not. To support the relevance of these data to the antiviral activity of XCL1, we found that the same anti-XCL1 pAb that was used to block HIV-1 capture reversed the antiviral activity of XCL1 in PBMC infection experiments (Figure S2).
To demonstrate that XCL1 can interact directly with the external viral envelope glycoprotein, gp120, we performed co-immunoprecipitation studies with biotin-conjugated XCL1 WT and variants. As seen in Figure 5B, XCL1 WT and W55D were able to specifically co-immunoprecipitate gp120, while the CC3 variant did not. Importantly, the same anti-XCL1 pAb that prevented virion capture abrogated gp120 co-immunoprecipitation (Figure 5B). Taken together, these data support a mechanism of HIV-1 inhibition whereby XCL1 interacts with viral particles via direct binding to the external viral envelope glycoprotein, gp120. Furthermore, these data confirm the dependency of the anti-HIV-1 activity of XCL1 on the all-β (alternatively-folded) conformation.
As an additional test for specificity of the interaction between XCL1 and the HIV-1 envelope glycoprotein (gp120), we performed both virion-capture and infection assays using VSV-G pseudotyped virions, which contain an HIV-1 core surrounded by the VSV envelope. Figure 6A shows that XCL1 was unable to capture VSV-G pseudotyped virions, indicating the HIV-1 capture observed in Figure 5A required the presence of the HIV-1 envelope. To verify the relevance of these observations to the antiviral activity of XCL1, we performed acute infection assays with VSV-G pseudotyped virions in primary PBMC. We did not observe any inhibitory effect of XCL1, as evidenced by measuring both the absolute numbers of infected cells (Figure 6B), and the levels of reporter gene (GFP) expression within the gated population of infected cells (Figure S3). Additionally, we observed no inhibition of VSV-G pseudotyped virus attachment or entry in TZM-bl cells (data not shown). Altogether, these results further validate that the mechanism of XCL1 inhibition is via direct interaction with the HIV-1 envelope.
Since we demonstrated that the antiviral activity of XCL1 depends on the all-β conformation (W55D), previously shown to bind GAGs with high affinity [15], we examined the inhibitory activity of XCL1 in PBMC infected with X4- or R5-tropic HIV-1 following digestion of cell-surface GAGs with heparitinase. As seen in Figure 7, we observed that both WT and W55D XCL1 were equally effective at blocking HIV-1 infection in heparitinase-treated and -untreated cells, while in contrast the CC3 variant remained inactive in both conditions. The efficiency of GAG removal was evaluated by ELISA using two different anti-GAG mAbs (Figure S4). These data provide further evidence for an antiviral mechanism mediated by direct interaction of XCL1 with the viral envelope, irrespective of it's binding to GAGs and/or other structures expressed on the target cell surface.
In this study, we report the characterization of the CD8+ T cell-derived C-chemokine, XCL1, as a novel, broad-spectrum inhibitor of HIV-1 infection. XCL1 is primarily produced by activated CD8+ T cells and NK cells [17], and recruits T lymphocytes and dendritic cells via binding to and activation of a specific cellular receptor, XCR1 [26], [27]. A possible link between XCL1 and HIV-1 was previously suggested in two wide-screening studies of chemokines and chemokine receptors: the first reported low-level inhibition of HIV-1 replication by XCL1 [28], while the second identified a small subset of HIV-1 isolates that could use the XCL1 receptor, XCR1, as a coreceptor in cells transfected in vitro [29]. However, these data were not subsequently validated in further studies, nor were the potential underlying mechanisms investigated. Our work provides a thorough characterization of the anti-HIV-1 activity of XCL1, showing no apparent relationship between the antiviral action of XCL1 and the putative function of XCR1 as a coreceptor, although these findings do not exclude that XCR1 may serve as a minor HIV-1 coreceptor in specific cells or anatomical sites.
XCL1 is a unique metamorphic chemokine that can interconvert between two different conformational folds: the conserved chemokine fold (monomer), which was shown to bind to and activate XCR1, and an alternatively-folded (all β-sheet) dimeric conformation which does not activate XCR1, but instead binds to glycosaminoglycans (GAGs) with high affinity [15], [23]. Using XCL1 variants designed to predominantly fold into one of the two conformations, we found that only the alternatively-folded (all β-sheet) molecule elicited anti-HIV activity, while the chemokine locked in the classical, XCR1-interacting fold was inactive. In line with this observation, we were unable to detect XCR1 expression by flow cytometry in the HIV-1 target cells used in our study; in fact, the ability of CD4+ T cells to express XCR1 is controversial [26], [30]. At this stage, it is uncertain whether and to what extent the inherent tendency of alternatively-folded XCL1 to dimerize plays any role in HIV-1 inhibition, as a monomeric form of the alternatively-folded XCL1 is not available for testing. Regardless, since the alternatively-folded molecule binds cell-surface GAGs and not XCR1, our results led us to investigate the early events in the viral infectious cycle that take place at the target cell surface. Indeed, we documented XCL1-mediated blockade of HIV-1 at an early stage of infection, namely, viral attachment and entry. Moreover, we provide multiple lines of evidence that XCL1 inhibits HIV-1 through an unconventional mechanism mediated by direct interaction with the viral envelope, similar to that previously reported for the α-chemokine CXCL4 [14]. We showed that XCL1 efficiently captures infectious HIV-1 virions and binds to the external viral envelope glycoprotein, gp120, and that both of these interactions depend on the alternatively-folded XCL1 structure. Furthermore, the same polyclonal antibody that antagonized virion capture and gp120 binding by XCL1 also neutralized the antiviral activity of XCL1. In line with the proposed antiviral mechanism, we demonstrated that binding to cell-surface GAGs was not required for the antiviral activity of XCL1, despite the dependence on the alternative, GAG-binding conformation for HIV-1 inhibition. These findings indicate that gp120 is another selective target of the alternative XCL1 conformation in addition to GAGs.
The fact that the biologically active conformation of XCL1 against HIV-1 is the high-affinity GAG-binding structure raises several mechanistic considerations. Foremost, the amount, complexity and variability of the glycan shield that decorates the surface of gp120 most likely influences the ability of XCL1 to block HIV-1 infection, since nearly half the molecular mass of gp120 is comprised of N-linked and O-linked glycans, and changes to these carbohydrate moieties result in altered neutralization sensitivity [31]–[36]. Indeed, it is possible that XCL1 interacts with a negatively-charged domain on the surface of gp120 [37]. Future structure-function studies with mutagenized XCL1 will help delineate key domains of the chemokine that are responsible for HIV-1 inhibition.
Whether and to what extent endogenous XCL1 contributes to the mechanisms of virus control during the course of HIV-1 infection is presently unknown. Although we found that XCL1 is a broad-spectrum HIV-1 inhibitor, we observed some variability in sensitivity among HIV-1 isolates. Different degrees of sensitivity have been documented for a wide range of antiviral biomolecules, including neutralizing antibodies [38]–[42], in line with the remarkable variability of the HIV-1 envelope, which we identified as the primary target for XCL1 antiviral activity. Another unresolved question is the discrepancy (∼2-log) between the XCL1 concentrations required for HIV-1 blockade and the levels released by activated CD8+ T cells cultured in vitro. However, it is important to emphasize that our data were obtained with E. coli-produced recombinant XCL1, leading to a significant underestimation of the potency of this chemokine. In fact, the C-terminus of XCL1 is a 22-amino acid mucin-like domain containing a cluster of O-glycosylated serine and threonine residues, and previous work has demonstrated that mammalian cell-produced, fully glycosylated XCL1 exhibits approximately 2-log higher biological activity compared with non-glycosylated XCL1 produced in prokaryotic cells [43]. Currently, there are no commercial sources of mammalian cell-produced XCL1, and efforts are underway in our laboratory to produce glycosylated XCL1. In addition, it is conceivable that in vivo-activated CD8+ T cells may release larger amounts of XCL1 into the local microenvironment, particularly within secondary lymphoid tissues. We are currently investigating if CD8+ T cells derived from asymptomatic HIV-infected subjects produce higher concentrations of XCL1 than CD8+ T cells from uninfected subjects, in line with their reported higher production of crude antiviral factor activity [5].
The identification of the first HIV-suppressive chemokines (CCL5/RANTES, CCL3/MIP-1α and CCL4/MIP-1β) has led not only to novel insights into endogenous host defenses against HIV-1, but also to the definition of new molecular targets for antiviral drugs [44]–[47] and genetic markers of innate HIV-1 resistance [48], [49]. In a similar manner, this study could be a first step toward determining the potential physiological role of XCL1 in HIV-1 infection. Analysis of XCL1 expression in subjects that are naturally protected from HIV infection (exposed-uninfected) or from disease progression (long-term nonprogressors) may offer new insights on mechanisms of natural resistance to HIV. Furthermore, a precise identification of the XCL1-interactive surface on the viral envelope may lead to the development of novel HIV-1 entry inhibitors, as well as new molecular targets for vaccine design.
Recombinant human XCL1/lymphotactin was obtained from Peprotech (Rocky Hills, NJ); recombinant XCL1 WT and variants (CC3 and W55D) were cloned and produced by two of the authors (JF, BFV) at the Medical College of Wisconsin, Milwaukee, WI, as previously described [15]; and recombinant RANTES/CCL5 and CXCL4/PF4 were purchased from R&D Systems (Minneapolis, MN). Molar values were calculated based on the molecular weight of the monomeric chemokines. PBMC from healthy donors were activated with phytohemagglutinin (PHA; Sigma, St. Louis, MO) and recombinant human IL-2 (Roche Applied Science, Mannheim, Germany) in complete RPMI medium (Invitrogen, Carlsbad, CA), containing 10% fetal bovine serum (FBS, Hyclone, Thermo Scientific, Waltham, MA), glutamine at 2 mM, streptomycin at 50 µg/mL, and penicillin at 100 U/mL for 72 hr prior to HIV-1 infection. Cell surface glycosaminoglycan (heparin sulfate) digestion was performed by incubating PBMC (1×106 cells/mL) with heparitinase (Heparinase III, Sigma) at 2 U/mL for 2 hr at 37°C in recommended buffer (20 mM Tris-HCl pH 7.5 containing 1% FBS and 4 mM CaCl2). Digested PBMC were washed once in complete RPMI and then used in acute infection assays as described. TZM-bl cells (NIH AIDS Research and Reference Reagent Program, Germantown, MD) were maintained in DMEM (Invitrogen, Carlsbad, CA) containing 10% fetal bovine serum. CD8+ T cells were enriched via negative selection from PBMC with the EasySep enrichment kit (Stem Cell Technologies, Vancouver, Canada) and activated by either PHA (20 µg/mL, Sigma), PMA (0.05 µg/mL, Sigma) plus ionomycin (1 µg/mL, Sigma), or anti-CD3/CD28 antibody-loaded beads (T Cell Activation/Expansion Kit, Miltenyi Biotec, Auburn, CA), all in the presence of 50 U/mL of IL-2. After 3 days of activation, the cells were washed to remove stimuli, medium was replaced with complete RPMI supplemented with IL-2 at a cell density of 1×106 cells/mL, and the culture supernatants were harvested at day 5 and 7 post-stimulation.
CD8+ T-cell culture supernatants (after 3 days of activation and washing) were tested for XCL1 production using the Human XCL1/Lymphotactin DuoSet ELISA (R&D Systems). To confirm the efficiency of GAG removal following heparitinase digestion, a cell-based ELISA was performed according to a previously established protocol with some modifications [50]. Briefly, activated PBMC were washed in PBS and seeded at 5×104 cells/well in 50 µL of PBS, and dried by evaporation at room temperature overnight. Plates were washed once with PBS and immediately fixed in ice-cold 2% paraformaldehyde at 4°C for 20 min, followed by washing in PBS. The wells were blocked in 0.2% casein-PBS buffer for 1 hr at 37°C, washed once with PBS, and incubated with 10 µg/mL of anti-heparan sulfate mAbs, clone 10E4 (AMSBIO, Lake Forest, CA) and clone T320.11 (EMD Millipore, Temecula, CA) for 2 hr at RT. An anti-CD4 mAb (RPA-T4, BD Biosciences) was used as a control for the non-specific effects of digestion on cell-surface protein expression (2 hr at RT). After 3 washes with PBS, the wells were incubated with polyclonal HRP-conjugated anti-mouse antibodies (Thermo Fisher Scientific, Rockford, IL). After 3 washes with PBS, wells were incubated with substrate solution until color development and immediate incubation with stop solution (R&D Systems), followed by reading optical density at 450 nm. Background measurements obtained with secondary antibody alone were subtracted from all readings.
The ability of XCL1 to downmodulate cell surface expression of CD4, CXCR4, and CCR5 was investigated by flow cytometry. Briefly, CD4+ T cells were cultured in the presence/absence of XCL1 at 20 µg/mL for 24 hours. Cells were then washed and stained for receptor expression using anti-CD4, CXCR4, and CCR5 mAbs (BD Biosciences, San Jose, CA). To further determine possible interactions between XCL1 and CD4, we assessed the ability of XCL1 to interfere with binding of different anti-CD4 mAbs targeted to different domains of CD4. Six fluorochrome-conjugated antibodies were used: OKT4 (eBioscience, San Diego, CA), 13B8 (Beckman Coulter, Inc., Indianapolis, IN), VIT4 (Miltenyi Biotec), RPA-T4, Leu3A/SK3, and L200 (all 3 from BD Biosciences). Two unlabeled mouse mAbs were used, DB-81 [51] and 9H5A8 (Novus Biologicals, Littleton, CO), followed by subsequent anti-mouse-R-phycoerythrin staining (Sigma). Enriched CD4+ T cells were incubated with XCL1 at 20 µg/mL or with PBS for 30 minutes at 4°C, without washing, cells were then stained with the various anti-CD4 mAbs listed. All data were acquired on a BD FACS Canto flow cytometer (San Jose, CA) and analyzed with FlowJo software version 9.5.2 for Macintosh (TreeStar, San Carlos, CA).
The HIV-1 isolates used in this study included two laboratory strains [IIIB (X4) and BaL (R5)], the dual-tropic primary isolate, 92HT599 (X4R5), and a set of primary isolates derived in our laboratory and minimally passaged ex vivo (98USSG, 07USLD, 07USPC, 08USSE, 97IT6366, 08USKD), obtained by culture of PBMC from chronically infected individuals. Acute cell-free HIV-1 infection was performed by addition of the viral stocks (50–100 pg of p24 Gag antigen per well) to duplicate cultures of activated PBMCs (PHA+IL-2 for 72 h) in round-bottom 96-well plates seeded at 2×105 cells per well in RPMI+10% FBS+20 U/mL of IL-2, or to TZM-bl cells seeded in 24-well plates overnight at 5×104 per well in DMEM+10% FBS for infection. Infected cells were cultured in the presence/absence of XCL1 at doses ranging from 0.06–1.5 µM. The levels of HIV-1 replication were assessed by measuring the extracellular release of p24 Gag protein in cell-free culture supernatants taken daily between days 3 and 7 post-infection using a highly sensitive Alpha (Amplified Luminescent Proximity Homogeneous Assay) technology immunoassay (AlphaLISA HIV p24 Research Immunoassay Kit, PerkinElmer, Waltham, MA). On day 7 of infection, cells were harvested for viability testing via absolute counting by flow cytometry. Cell viability was determined by normalization of the total live-gated cell counts in XCL1-treated wells to the number of cells recovered from control wells (untreated with XCL1). To show the physiological relevance of our infection data, we performed an infection assay whereby HIV-1 IIIB was pre-incubated (prior to addition to target cells) with XCL1 alone (1 µM) or with XCL1 combined with anti-XCL1 pAb (10 µg/mL of the same pAb used to block HIV-1 capture by XCL1). To control for non-specific effects of the pAb we also included control wells (no XCL1 treatment) with pAb alone.
As a test for specificity of the interaction between XCL1 and the HIV-1 envelope glycoprotein (gp120), we performed infection assays with GFP-expressing VSV-G pseudotyped virus provided by Michael P. Marino (CBER/FDA, Bethesda, MD, USA). PBMC (5×104) were seeded in 96-well round bottom plates and infected in a 50 µL volume of pseudotyped virus (MOI of 10) in the presence or absence of XCL1 WT overnight, with each condition in quadruplicate wells. Wells were supplemented with an additional 50 µL of complete RPMI at 24 hr post-infection to yield a final well volume of 100 µL. Individual wells were harvested 48 hr post-infection for flow cytometry detection of GFP-positive (infected) cells. To supplement the data counting the absolute numbers of infected cells, mean fluorescence intensity was also determined to indicate the amount of virus infection within each GFP-positive event (infected cell).
The HIV-1 attachment and entry assays were performed on TZM-bl cells with the primary, dual-tropic HIV-1 isolate 92HT599. TZM-bl cells (106 per replicate; two replicates per treatment) were seeded in 12-well plates overnight to achieve a confluent cell monolayer. Without disturbing the monolayer, cells were washed with PBS to remove media, followed by pre-incubation for 15 min at room temperature with XCL1 diluted in PBS, and then exposed to 500 µL of the undiluted viral stock (124 ng/mL of p24) for 4 h (attachment) or 6 h (entry) at 37°C, in the continuous presence of XCL1. Two wells of untreated cells were incubated for 4 hr with virus at 4°C to determine the background signal level (trypsin-insensitive despite low-temperature conditions preventing virus entry). As specificity controls, replicate wells were pretreated with known inhibitors/inducers of viral attachment/entry prior to virus incubation: CXCL4/PF4 at 15 µg/mL (R&D Systems), peptide T-20 at 50 µg/mL (NIH AIDS Research and Reference Reagent Program, Germantown, MD), an anti-CD4 mAb at 20 µg/mL (azide-free RPA-T4, eBioscience, San Diego, CA), or CCL5/RANTES at 15 µg/mL (R&D Systems). After incubation, the cells were washed with PBS to remove unbound virus, without disturbing the cell monolayer. Entry assay wells were treated with pre-warmed bovine trypsin (Sigma) for 5 min at 37°C, followed by trypsin inactivation with cold DMEM medium containing 10% (vol/vol) FBS. Both trypsin-treated (entry assay) and untreated (attachment assay) cells were then washed two times with cold PBS, and lysed with 100 µL of 0.5% (wt/vol) Triton X-100 to quantify the amount of cell-associated p24 protein. The specific signal was calculated by subtracting the background p24 levels measured in wells incubated at 4°C (treated with trypsin) from the p24 levels measured in each test sample.
The virion capture assay was performed as previously described [14]. Briefly, immunomagnetic beads (4×104 per tube) covalently linked to a polyclonal antiserum to rabbit IgG (Invitrogen) were incubated with a polyclonal rabbit IgG antibody to human XCL1 (Peprotech), washed with PBS containing 0.05% (wt/vol) bovine casein and then loaded with recombinant human XCL1 (2.5 µg per reaction). After removing unbound XCL1 by repeated PBS washes, chemokine-armed beads were incubated with 0.5 mL of the viral stock (HIV-1 IIIB (X4); 20 ng of p24 Gag protein/test). To test the specificity of XCL1 interaction with the virus, the XCL1-armed immunomagnetic beads were pre-incubated with monoclonal (mAb) and polyclonal (pAb) anti-XCL1 antibodies (R&D Systems, 20 µg/mL) for 10 minutes at room temperature prior to virus addition. After incubation with virus for 1 h at room temperature, the beads were washed to remove unbound virus particles and treated with 0.5% Triton X-100 to lyse the captured virions. The amount of captured p24 Gag protein was quantified by AlphaLISA®.
As an additional measure for the exclusive interaction between XCL1 and the HIV-1 envelope glycoprotein (gp120), we assessed the ability of XCL1 to capture VSV-G pseudotyped virus. Anti-mouse immunomagnetic beads were armed with monoclonal anti-VSV-G antibody (KeraFAST Inc., Boston, MA) to show the efficiency of VSV-G pseudotyped virus capture in our experimental design. In parallel, anti-rabbit immunomagnetic beads armed with both anti-XCL1 pAb and subsequent XCL1 WT were tested for the ability to capture VSV-G-pseudotyped virus. For accurate comparison of capture between beads armed with anti-VSV-G mAb and beads armed with XCL1, equal amounts of VSV-G-pseudotyped virus was added to all capture reactions.
To evaluate the direct interaction between XCL1 and the gp120 external envelope glycoprotein, we performed co-immunoprecipitation experiments using purified YU2 gp120 protein. To assess this interaction, XCL1 WT, W55D and CC3 proteins were biotinylated using the LYNX Rapid Conjugation Kit (AbD Serotec, Kidlington, UK). In two conditions we assessed the specificity of XCL1-gp120 interactions via pre-incubation of biotinylated XCL1 WT with 5 µg of anti-XCL1 pAb (R&D Sytems) or goat IgG (R&D Systems), as a control, for 1 h in 100 µL of PBS. Following the presence or absence of antibody pre-incubation, a mixture of biotinylated XCL1 WT, W55D, or CC3 (2 µg) was incubated with gp120 (2 µg) in 100 µL of PBS+0.2% casein for 3 h at room temperature with constant rotation. Subsequent incubation with 50 µL of streptavidin-coated magnetic beads (Invitrogen) in 200 µL of RIPA buffer was performed for an additional 10 min incubation. The samples were washed three times with RIPA buffer, dissolved in SDS loading buffer, and loaded on 12% polyacrylamide gels and resolved by SDS gel electrophoresis. Protein was transferred onto nitrocellulose membranes and blotted with an anti-gp120 mAb (b24; a gift from George K. Lewis, University of Maryland, Baltimore, MD).
Anonymized samples of peripheral blood were obtained from healthy volunteer donors at the NIH Blood Bank under a protocol approved by the NIH IRB.
|
10.1371/journal.pcbi.1003515 | Viral Quasispecies Assembly via Maximal Clique Enumeration | Virus populations can display high genetic diversity within individual hosts. The intra-host collection of viral haplotypes, called viral quasispecies, is an important determinant of virulence, pathogenesis, and treatment outcome. We present HaploClique, a computational approach to reconstruct the structure of a viral quasispecies from next-generation sequencing data as obtained from bulk sequencing of mixed virus samples. We develop a statistical model for paired-end reads accounting for mutations, insertions, and deletions. Using an iterative maximal clique enumeration approach, read pairs are assembled into haplotypes of increasing length, eventually enabling global haplotype assembly. The performance of our quasispecies assembly method is assessed on simulated data for varying population characteristics and sequencing technology parameters. Owing to its paired-end handling, HaploClique compares favorably to state-of-the-art haplotype inference methods. It can reconstruct error-free full-length haplotypes from low coverage samples and detect large insertions and deletions at low frequencies. We applied HaploClique to sequencing data derived from a clinical hepatitis C virus population of an infected patient and discovered a novel deletion of length 357±167 bp that was validated by two independent long-read sequencing experiments. HaploClique is available at https://github.com/armintoepfer/haploclique. A summary of this paper appears in the proceedings of the RECOMB 2014 conference, April 2-5.
| Humans infected with a virus, such as the human immunodeficiency virus (HIV-1) or hepatitis C virus (HCV), host a population of billions of virus particles. Among these, there is an unknown number of genetically different strains, some of which can harbor drug resistance and immune escape mutations. It is of clinical importance to know the DNA sequences and abundances of these variants, as they can affect treatment outcome. Here, we present HaploClique, a computational approach to reconstruct these sequences and to predict large insertions and deletions from paired-end next-generation sequencing data. Using simulations, we demonstrate that HaploClique can reconstruct full-length HIV-1 variants from low-coverage samples. Using real-world clinical data, we predict a novel deletion of 357±167 bp in a HCV patient sample that has been validated by two independent long-read sequencing experiments.
| Genetic diversity is an important characteristic of evolving populations and it affects the chances of survival in changing environments. Assessing the genetic diversity of a population experimentally is generally labor-intensive and difficult. Populations of individual cells or viruses, however, can be analyzed efficiently using next-generation sequencing (NGS). Although single-cell approaches are still immature, direct NGS of mixed samples at deep coverage allows for probing populations in great detail. The challenges with this bulk sequencing approach are (i) to separate sequencing errors from genetic variation, (ii) to assemble the short NGS reads into an unknown number of different, unknown, longer haplotype sequences, and (iii) to estimate their frequency distribution.
Viruses such as human immunodeficiency virus (HIV-1) and hepatitis C virus (HCV) populate their hosts as swarms of related but genetically different mutant strains, each defined by its haplotype sequence. The structure of such a mutant cloud, which is often referred to as a viral quasispecies [1], is of clinical importance, because it has been shown to affect virulence [2] and pathogenesis [3]. In addition, low-frequency genetic variants may harbor resistance mutations that are capable of evolutionary escape from the selective pressure of host immune responses [4] and of medical interventions, such as anti-viral drug treatment [5]. NGS is currently introduced into clinical diagnostics, but the de facto standard procedure for assessing the quasispecies structure is simply based on single-nucleotide variant (SNV) calling. This approach allows only for estimating the per-site allele frequency spectrum of the virus population and it ignores patterns of co-occurrence among mutations. This limitation is critical, because epistatic interactions are abundant in RNA viruses [6]. Hence, one cannot predict viral phenotypes without knowing the underlying mix of haplotypes. Here, we address this challenge and present a computational approach for the viral quasispecies assembly problem.
The viral haplotype reconstruction problem is related to the human haplotype reconstruction problem, but it differs in several key aspects and faces different challenges. First, the number of unique haplotypes in a viral quasispecies is unknown unlike in the case of human diploid genomes. Second, viral populations typically exhibit more than two variants at each polymorphic locus and often all four different nucleotides. Hence, viral haplotypes cannot be described by binary sequences. Third, in a viral quasispecies, low-frequency variants are abundant and of clinical importance, yet they are difficult to distinguish from technical sequencing errors. Finally, RNA virus genomes are orders of magnitude shorter than the human genome, but exhibit more diversity within one host than the ∼0.1% diversity between the two parental human haplotypes [7].
Several methods for viral haplotype reconstruction have been developed in recent years, specialized for different NGS technologies, experimental designs, and quasispecies structures. In general, reconstruction can be performed either locally, in a genomic region that can be covered by the average read length, or globally, over longer regions such that overlapping reads are necessary for assembly. Local reconstruction means estimating the number of locally unique haplotype sequences and, at the same time, correcting sequencing errors. Probabilistic clustering [8]–[11] and k-mer statistics [12] have been proposed for this task. Global reconstruction is more challenging, as it requires computational solutions for assembling NGS reads, which has proven itself to be demanding even in settings without poly-ploidy [13].
For quasispecies assembly, approaches from different domains have been developed: (i) probabilistic mixture models [14], (ii) hidden Markov models [15], (iii) sampling schemes [16], (iv) combinatorial approaches based on analyzing the read overlap graph [8], [17]–[19], (v) coloring of overlap and conflict graphs by constraint programming [20], and (vi) exploiting the “identical by descent” information [21] in the HapCompass framework [22], originally designed for diploid single nucleotide polymorphism data.
The performance of global haplotype reconstruction depends on several factors, including the true underlying diversity of the population, the distribution of amplification and sequencing errors, the read length, and the distribution of the read coverage along the genome [23]–[25]. A major shortcoming of all existing methods is that they are unable to handle large insertions or deletions (indels). For example, large deletions can result from erroneous replication or, as observed recently in HIV-1, they may occur as alternative splice variants [26]. In the context of analyzing structural variation in the human genome, such as indels of varying sizes, the use of paired-end reads has been instrumental. For viral haplotype reconstruction, however, approaches that systematically exploit paired-end information are lacking.
In this paper, we present a new quasipecies assembly method for paired-end reads, called HaploClique, based on enumeration of maximal cliques (max-cliques) as a general approach to clustering NGS paired-end reads. Although, in general, the runtime of enumerating all max-cliques in a graph is exponential, it has recently been shown that the graphs induced by overlapping NGS reads can be handled efficiently [27], [28]. Here, we exploit this fact for the quasispecies assembly problem and develop a probabilistic model of sequence and structural similarity between reads.
Using max-clique enumeration for reference-based read assembly is orthogonal to combinatorial approaches for de novo assembly that rely on path finding in de Bruijn or similar graphs [29]–[31]. Instead of computing paths, we iteratively transform max-cliques into super-reads and then seek max-cliques of super-reads, thereby obtaining haplotype segments of increasing length. The haplotype segments can eventually be extended to global haplotypes if the degree of heterogeneity of the viral quasispecies is high enough. HaploClique is related to max-cut-driven approaches in human haplotype reconstruction [32], but the computational complexity of those approaches is prohibitive for virus populations of high and unknown ploidy. While HaploCliques enumerates all max-cliques, a max-cut approach seeks an optimal cut of the overlap graph.
HaploClique explicitly incorporates paired-end information for assembling viral haplotypes. We define the insert as the unsequenced fragment between the two ends of a paired-end read. We use linkage information among variant alleles in the distant pairs to identify reads that stem from the same haplotypes and generate error-corrected paired-end super-reads. Paired-end reads allow to bridge homogeneous, and hence ambiguous, genomic regions if the insert size is sufficiently large. They also increase the statistical power to distinguish local haplotypes from sequencing errors in homogeneous regions if the paired read is located in a more heterogeneous region. Employing our iterative clique enumeration procedure, we show that error-free full-length HIV-1 viral haplotypes can be reconstructed in a heterogeneous mix of five viral strains in silico from a data set with mean coverage of 600×. Furthermore, we demonstrate that, unlike existing methods, HaploClique can detect large indels in mixed virus populations in silico and in vivo. Finally, we apply HaploClique to a HCV Illumina paired-end NGS data set and predict a novel deletion of length bp that has been confirmed independently by two long-read NGS platforms.
We developed and implemented HaploClique, a computational viral quasispecies assembly method for paired-end NGS data. HaploClique defines a read alignment graph, in which each node corresponds to a single-end or paired-end alignment (Figure 1A). We draw edges between two nodes if the two corresponding alignments have sufficient overlap and are likely to stem from the same haplotype (Figure 1B). Each max-clique in this graph consists of a large number of reads from the presumed same haplotype segment. Thus, the consensus sequence of all reads in a max-clique is a prediction of a local haplotype sequence. We refer to such a consensus sequence as a super-read. This consensus sequence also serves to correct errors in the reads that participate in the super-read by replacing the sequence of the original reads with the consensus. This form of error correction benefits from phasing sequential variants through super-read construction. Paired-end reads are particularly helpful, as they allow to also phase distantly co-occurring variant alleles.
HaploClique proceeds by iterating (i) (super-)read alignment graph construction, (ii) max-clique enumeration, and (iii) super-read construction, until convergence. The lengths of the super-reads increase while iterating and convergence is established when super-reads have reached their maximum length. If the mixed sample is sufficiently heterogeneous, super-reads will eventually represent haplotypes of full length. Because we process paired-end reads and incorporate insert-size compatibility into our edge definition, we can also identify max-cliques that indicate larger insertions and deletions. These structural variations are recognized by too small or too large insert sizes among the alignments of the reads that participate in a max-clique. We analyzed HaploClique's performance on simulated data and demonstrate its use on in vivo HCV quasispecies sequencing data.
HaploClique integrates paired-end and base quality information for improved sequencing error correction and haplotype frequency estimation, which we assess first. Second, we evaluate HaploClique's behaviour when confronted with low heterogeneity among the different haplotype strains. Third, we demonstrate HaploClique's ability to detect large insertions and deletions in the quasispecies by making use of paired-end information. Fourth, we evaluate the quality of the local and global haplotypes that HaploClique predicts. Lastly, we compare HaploClique to state-of-the-art tools ShoRAH [33], PredictHaplo [14], and QuRe [16] in quasispecies reconstruction of a simulated five virus mix of well-known HIV-1 lab-strains.
In all of the following experiments, we simulated Illumina 2×250 bp paired-end reads using SimSeq [34] with fragment size 600 bp. To make the simulated data as realistic as possible, we estimated the required error profiles from an in-house MiSeq data set of a mixture of known HIV-1 strains. The average error-rate was 0.33% per base.
For the application of HaploClique to a clinical sample, HCV RNA was extracted from the plasma collected from a subject isolated 135 days post infection and the NS5 region RT-PCR amplified as previously described [35]. In this subject there was experimental evidence of antigen-specific CD8+ T cell responses targeting two epitopes in the NS5 region (K2629SKRTPMGF and W2820LGNIIMFA). The NS5B region encodes for the RNA-dependent RNA polymerase and is essential for the replication of the virus.
This amplicon was sequenced to a coverage of 80,000× on a MiSeq instrument using a 2×250 bp read kit. The resulting reads were aligned using BWA-MEM [36]. We found the insert size distribution to have a mean of 155 bp and a standard deviation of 167 as estimated by HaploClique. Despite this large standard deviation, HaploClique was able to discover a bp deletion. No other indels were reported by HaploClique.
In two independent sequencing runs of the same amplicon, once on a 454/Roche GS FLX+ system and once on a PacBio instrument, the presence of the deletion was confirmed. Both technologies yield longer reads than MiSeq that could successfully be aligned across the deletion breakpoint, allowing to determine breakpoint coordinates at base-pair resolution.
For the alignment of the longer reads, extreme affine gap costs have been used to find the deletion. In general, this leads to alignment artifacts in other regions, causing false positive haplotype calls. With a read length of 250 bp, we did not succeed to align reads across the large deletion.
Comparing coordinates, we found that the start positions predicted by HaploClique was 15 bp off the true position and the true length amounted to 444 bp. That is, the length difference between true and predicted deletion amounted to 87 bp, or 0.52 standard deviations.
We have presented HaploClique, a method for local haplotype reconstruction, structural variant detection of large insertions and deletions, and global haplotype assembly, which represents a principled approach to viral quasispecies assembly from NGS paired-end reads. HaploClique builds on a read alignment graph as underlying combinatorial model, where nodes correspond to single-end or paired-end alignments of reads. Edges are modeled in a probabilistic fashion.
They are based on sequence similarity of the read overlap by incorporating phred-style quality scores in combination with a position-wise prior for the non-overlapping parts of the reads, and on a criterion that measures insert size compatibility of the two alignments. While the sequence similarity criterion accounts for correct assembly of reads, the insert size criterion allows for detecting insertions and deletions in viral haplotypes that cannot be detected from single-end read alignments alone. We suggest a model that unifies sequencing error correction, clustering reads into haplotype groups, as well as assembling reads into longer fragments, all of which naturally emerge from the model.
In the read alignment graph, max-cliques represent maximal read sets that overlap and represent (locally) identical haplotype sequence. The advantage of the max-clique computation is twofold. First, it clusters reads, thereby separating reads stemming from different haplotypes.
Second, it enables sequencing error correction in a way that can make full use of co-occurrence, that is, statistical correlation of variant alleles within reach of the reads participating in a max-clique. In particular, the error correction exploits paired-end information if provided. The improved error correction is important, as it gives rise to improved frequency estimates and allows for distinguishing between haplotypes whose pairwise distance is below 1%.
HaploClique allows for reconstructing full-length global haplotypes using a read assembly procedure that is orthogonal to all existing assembly methods. In our iterative approach, we alternate between transforming max-cliques into super-reads, which form the nodes of a new alignment graph, and finding max-cliques in the new graph. We repeat this process until convergence, which is established when super-reads do not grow any longer.
HaploClique depends on three parameters to be adjusted manually: minimal read overlap, , a threshold for the probability that two overlapping reads stem from locally identical haplotypes, , and the minimal coverage to call the super-read sequence . In general, if one of the parameters is decreased, the number and size of cliques will increase. If and are too small, the purity of cliques will decrease, meaning reads from different but very similar haplotypes cluster. If is too large, cliques will grow slower and less frequent haplotypes may be missed. If is too large, reads are more likely to cluster not only if they stem from the same haplotype but also if they have technical errors in common; this leads to lower error correction efficiency. If is too small, there might not be enough statistical power to correct for sequencing errors. If is too large, the false negative rate will rise, as low-frequency haplotypes do not provide enough reads to form cliques. We used two different parameter sets for HaploClique. In the first iteration, local haplotype reconstruction with error correction is performed and we chose and . In practice, the results are insensitive to the parameter choice (Figure S2). For the following iterations, the quasispecies assembly, we assume that haplotypes are error-corrected and must match perfectly. We set to account only for the stochasticity of the Phred scores.
We evaluated HaploClique by extensive simulation studies. The simulated haplotypes were well-known and much analyzed HIV-1 virus strains. We kept coverage in the simulation study rather low, so as to evaluate our tool in the presence of only weak signals. We did this also in comparison with extant state-of-the-art tools. We demonstrated that our approach has superior error correction capabilities. This, in turn, yields accurate haplotype frequency estimates, even at the rather low coverage of 120× per haplotype. The tools we compared to were not able to provide similarly accurate frequency estimates. HaploClique proved to be insensitive to a coverage reduction of one order of magnitude with respect to prevalent sequencing experiments, which commonly operate at 5000× or higher.
Beyond improved frequency estimates, we also improve haplotype sequence reconstruction. In all experiments, more than 99% of the haplotype segments we predict perfectly matched true haplotype sequences. None of the other tools generated even only one such perfectly matching segment, possibly because they require much higher coverage. This improvement in terms of accuracy may be due to the probabilistic model that treats error correction and assembly within one unifying framework. Our simulations also indicated that the degree of heterogeneity required in order to reconstruct large enough haplotype segments can be lower than 1%.
We also ran HaploClique on a real, Illumina MiSeq dataset of coverage 80,000×, which was found to consist of two HCV strains one of which had a frequency of only approximately 3% and contained a deletion of size 444 bp, as conformed by independent 454/Roche and PacBio sequencing experiments. In the MiSeq dataset, the deletion in question could not be detected by state-of-the-art read alignment tools. HaploClique successfully predicts this deletion, despite the large standard deviation of the fragment size distribution ( 167 bp). These experiments document that our method can detect large deletions also in Illumina paired-end datasets that otherwise would be difficult to identify.
Despite these improvements over previous methods, there are limitations of this approach. For example, the runtime of HaploClique is exponential in the read coverage. This feature is critical in the first two iterations of the procedure, before the number of reads is decreased. We observed that, approximately, the runtime doubles for each additional 250 reads of coverage. The baseline runtime was ∼4 minutes for a data set with coverage 1000×, on a single 3 GHz core. To overcome this computational bottleneck, one may perform the first iterations of haplotype reconstruction on subsets of the data and then assemble the merged results. Another extension that may decrease the runtime is to employ improved clustering techniques [37].
In the future, we also plan to explore on human whole-genome data, including polyploid cancer genomes, to perform error correction of the paired-end reads by local haplotype reconstruction and to assemble diploid haplotypes. This problem is more challenging due to the larger genome size and smaller levels of diversity, but several ideas presented here and implemented in HaploClique may prove useful for this task.
HaploClique performs paired-end error correction, local and global haplotype reconstruction, and structural variant calling based on enumerating cliques in the read alignment graph. We first explain the graph construction and then how super-reads are built, how global haplotypes are iteratively constructed, and how haplotype abundancies are estimated.
Let be the set of all reads from a viral quasispecies sequencing experiment and the set of their alignments to a reference genome as computed by a read aligner. In this paper, we assume that each read can be uniquely mapped, which is a reasonable assumption for short, non-repetitive viral genomes. In our experiments, we use the Illumina MiSeq technology for sequencing and BWA-MEM [36] as a read aligner. However, HaploClique depends on the sequencing technology and read mapper only insofar that it expects reads to be equipped with quality scores and that the reads can be properly aligned.
We construct a graph where the read alignments are the vertices. An edge indicates that the two alignments and overlap sufficiently and that the corresponding reads are likely to originate from (locally) identical haplotypes. More precisely, we draw an edge between and if they satisfy two criteria based on sequence similarity and insert sizes, respectively. While the sequence-based criterion ensures that the reads and do not exhibit mutually contradictory sequences, the insert size-based criterion guarantees that and do not contradict each other in terms of their fragment sizes. We allow alignments and to be any combination of single- and paired-end reads, but the size-based criterion applies only if both alignment are based on paired-end reads.
Cliques are fully connected subgraphs of the read alignment graph . They indicate groups of reads that are all likely to stem from locally identical haplotypes. Hence max-cliques form maximal groups of reads that originate from locally identical haplotypes.
The algorithm proceeds by first sorting all nodes from left to right, in ascending order of the alignment coordinates such that starts left of . The algorithm then computes maximal cliques by processing all nodes in this order. Let be the induced subgraph of with vertices and let be all max-cliques in . For a node , let be all nodes that are connected to by an edge . If the rightmost coordinates of all alignments in a clique are smaller than the leftmost coordinate of , this clique cannot be further affected by any node — such cliques are maximal in and can be output if only nodes are left to be considered. After having processed all nodes , we declare all cliques that can still be affected by nodes to be active.
When processing node , we compute its neighborhood and add a new clique if intersecting with each active clique yields the empty set. Otherwise, for each active clique , we set if , and we add a new clique if . Among all new cliques to be added, we eliminate duplicates.
Max-clique enumeration is related to the problem of finding a minimum clique cover [38], where a minimal set of non-overlapping max-cliques is sought, whose vertices cover the graph.
Each max-clique is a set of reads with mutually compatible alignments. Therefore, we can construct consensus sequences for max-cliques, which we refer to as super-reads. The purpose of super-read construction is two-fold. First, super-reads represent haplotype segments. Second, super-reads can be used as input to further iterations of max-clique enumeration, with the goal of global haplotype reconstruction, which is discussed in the next section.
To construct super-reads, let be the alignments participating in a max-clique and let be the set of positions where at least of these alignments contain non-gap characters. We recall that denotes the probability that nucleotide gave rise to position in read , although might differ from due to sequencing errors.
We determine the nucleotide sequence of the super-read by means of a weighted position-wise majority vote. We set , for each position , where is defined to be zero when does not cover position . The parameter ensures that the super-reads have sufficient coverage of high quality. For later frequency estimation, we keep track of which original reads gave rise to which super-read.
For global haplotype reconstruction, or quasispecies assembly, we iterate the clique enumeration procedure. We align reads against the reference sequence, construct the read alignment graph, find max-cliques, and merge them into larger super-reads with updated phred scores that reflect corrected error profiles. In the next iteration, we use these super-reads as reads and restart the procedure until number and length of super-reads have converged. For the assembly step, we assume that reads have already been error-corrected in the first iteration and set . Reads have to match perfectly and only allows for stochasticity of the Phred scores. We start the iterations with a relative overlap of . Once the length and number of super-reads converged, we decrease by down to a minimum of .
We estimate haplotype abundance by counting the number of (original) reads that participate in the super-reads giving rise to the haplotypes. Original reads may participate in several super-reads and thereby contribute to abundance counts for several haplotypes. We resolve this issue by keeping track of the original read in each iteration, such that each read can be assigned to the final haplotypes after convergence. Reads contributing to several haplotypes abundances are then taken into account by weighting them accordingly.
The MiSeq raw read data set is available through the Sequence Read Archive under the BioProject accession number SRP034655. The MiSeq 2×250 bp error profiles for SimSeq [34] used in the simulations are available at https://github.com/armintoepfer/haploclique under data.
A summary of this paper appears in the proceedings of the RECOMB 2014 conference, April 2-5 [42].
|
10.1371/journal.pgen.1006714 | Microhomology-mediated end joining induces hypermutagenesis at breakpoint junctions | Microhomology (MH) flanking a DNA double-strand break (DSB) drives chromosomal rearrangements but its role in mutagenesis has not yet been analyzed. Here we determined the mutation frequency of a URA3 reporter gene placed at multiple locations distal to a DSB, which is flanked by different sizes (15-, 18-, or 203-bp) of direct repeat sequences for efficient repair in budding yeast. Induction of a DSB accumulates mutations in the reporter gene situated up to 14-kb distal to the 15-bp MH, but more modestly to those carrying 18- and 203-bp or no homology. Increased mutagenesis in MH-mediated end joining (MMEJ) appears coupled to its slower repair kinetics and the extensive resection occurring at flanking DNA. Chromosomal translocations via MMEJ also elevate mutagenesis of the flanking DNA sequences 7.1 kb distal to the breakpoint junction as compared to those without MH. The results suggest that MMEJ could destabilize genomes by triggering structural alterations and increasing mutation burden.
| Recurrent chromosome translocations juxtapose chromosomal fragments and alter expression of tumor suppressors or oncogenes at or near breakpoint junctions to develop distinct types of leukemias and childhood sarcomas. The prevalence of 2–20 bp of imperfect overlapping sequences (a.k.a. microhomology [MH]) at the breakpoint junctions suggests the type of repair events joining two chromosomal fragments and the formation of oncogenic chromosomal translocations. In this study, we discovered that MH-mediated end joining (MMEJ) operates with kinetics markedly slower than other repair options. The slower kinetics leads to extensive resection and drives hypermutagenesis at sequences flanking the break site. We also found that MH-mediated chromosomal translocations accumulate mutations at sequences up to several kilobases distal to the breakpoint junction as compared to those without MH. Our results revealed that MH contributes to genetic instability by facilitating chromosomal translocations and increasing mutational load at the sequences flanking the breakpoints.
| The presence of short stretches of overlapping sequence (microhomology, MH) is a frequent feature of pathogenic chromosomal translocation breakpoints in human cells and has been implicated in juxtaposing two DNA ends for the error-prone repair of DNA breaks in both yeast and vertebrates [1–3]. This so-called microhomology-mediated end joining (MMEJ) is genetically distinct from Ku-dependent classical end joining or homologous recombination and becomes a prominent repair option when conventional repair mechanisms become inactivated or unavailable. Accordingly, MMEJ is frequently regarded as a back-up to the canonical repair pathways although it is still operational in cells retaining other repair options and contributes to a wide range of cellular chromosome maintenance processes including telomere maintenance and programmed immune receptor gene rearrangements [4, 5].
MMEJ is a highly error prone pathway because it inevitably entails deletion of inter-MH sequences and one of the MHs. MMEJ is also prone to chromosomal rearrangements due in part to the loss of intra-chromosomal joining bias [6]. To initiate MMEJ, DNA ends should first be resected and the flanking MHs for annealing should form ssDNA [7–10]. DNA resection also triggers DNA damage-induced checkpoints and the association of the strand exchange protein (Rad51)-DNA complex with ssDNA to initiate the homology search during recombination [11–13]. Furthermore, the formation of ssDNA at DNA ends inhibits non-homologous end joining (NHEJ), committing cells to homologous recombination (HR) and the MMEJ pathway [14]. Enzymatically, DNA end resection in eukaryotic cells comprises two distinct stages: initial resection by the Mre11 complex and more extensive resection by Dna2/Blm (Sgs1 in yeast) and Exo1 [15–18]. MMEJ is thus deficient in mre11-deleted cells or those deleted for CtIP [7–9, 19–22], a protein associated with the Mre11 complex that regulates its nuclease activity. Furthermore, expression of hypomorphic rfa1 mutants, one of the three subunits in the replication protein A (RPA) ssDNA binding complex in yeast, elevates the MMEJ frequency almost 350-fold and induces gross chromosomal rearrangements with MHs at the breakpoint junctions [23]. Resection and the formation of ssDNA are thus key steps in MMEJ and likely dictate the types of repair outcomes and chromosomal integrity upon DNA breakage.
Interestingly, emerging evidence suggests that ssDNA also triggers elevated mutagenesis because cells ultimately need to fill-in the gaps formed during double strand break (DSB) repair and restore the DNA duplex by the actions of an error prone translesion polymerase [24–26]. DSB repair thus represents a significant source of mutagenesis and fuels genome instability in mitotic cells. Together these observations prompted us to consider if MMEJ could contribute to mutagenesis especially at the breakpoints of chromosomal translocations because ssDNA represents an obligate intermediate for the process. Indeed, breakpoint junctions of complex copy number variants often contain MH and are associated with a high frequency of mis-sense and in-del types of mutations at the flanking DNA likely due to error prone repair synthesis [27–29]. We surmise that some of these junctions and mutagenesis might arise by MMEJ.
To address if MMEJ is mutagenic, we set up a model MMEJ assay in yeast and placed a URA3 reporter gene at several locations distal to an HO recognition site (5.8-, 7.1-, 7.2-, 9.1-, 11.5-, 14.5-, and 20-kb from the break, see Fig 1A). A DSB generated by HO cleavage is then flanked by 15-, 18- or 203-bp of direct repeat sequences 51-bp distal to the HO recognition site to mediate inter-repeat recombination (Fig 1A). The strain also lacks HML and HMR, two silent templates for gene conversion, and expresses HO endonuclease from the GAL1/10 promoter integrated at the ade3 genomic locus. The entire open reading frame of the endogenous URA3 locus on chromosome V is also deleted to eliminate gene conversion between ura3 sequences. Upon addition of galactose to the culture medium, HO is expressed (S1 Fig), and the resulting DSB is repaired by Rad52-dependent, but Rad51-independent single strand annealing (SSA) or MH-mediated events via annealing of flanking direct repeats (Fig 1A).
We measured overall survival frequency and FOA-resistant (FOAR) survival frequency by fluctuation tests, which reflect DSB repair and repair-induced mutation frequency, respectively (S1 and S2 Tables)[30]. We also measured mutation frequency in the CAN1 gene located on the left arm of chromosome V as an internal control and used it to calculate the spontaneous mutation frequency intrinsic to cell proliferation and to determine the 95% confidence intervals (S3 Table).
We found that HO expression led to a nearly 10-fold reduction in survival in the strain with 15-bp repeats compared to that with 203-bp repeats, indicating that 15-bp repeats do not efficiently support DSB repair (S1 Table). HO expression led to an intermediate level of survival in the strain with 18-bp repeats as compared to those with 15- or 203-bp repeats. Induction of HO increased the mutation frequency of the reporter gene 5.1- to 931-fold in the strain with 15-bp repeats, 13.5-fold in the strain with 18-bp repeats, and 7.2-fold in 203-bp repeats (Fig 1, S2 Table). The highest mutation frequency was observed in the strain having the URA3 gene inserted closest (5.8 kb) to the 15-bp repeats. The location of URA3 did not have an impact on the survival frequency (near 8%) nor the frequency of can1 mutations (Fig 1, S1–S3 Tables). The symmetry of the mutagenesis profile at either side of the break suggests that the distance to the repeats is one of the key factors dictating the frequency of mutagenesis (Fig 1B). Overall, the HO-induced mutation frequency was 51 times higher in cells that employed 15-bp of MH for repair (compare FOAR frequency in the 7.1-kb telomere proximal location in 15-bp vs 203-bp repeat containing strains, Fig 1B), and the mutations were found at greater distal locations up to 14.5-kb from the break in MMEJ events compared to those in the 203-bp repeat strain (Fig 1B, S2 Table), suggesting that mutagenesis is inversely related to the repair frequency.
As a comparison, we also measured the mutation frequency before and after HO expression in strains lacking direct repeat sequences. The strains lacking repeats did not significantly (<4.1-fold) induce mutagenesis 7.1-kb distal to the DSB (Fig 1C, S2 Table).
To determine the types and spectra of mutations associated with DSB repair, ura3 genes were recovered from FOAR survivors and subjected to sequencing analysis (S2–S4 Figs, S4, S5 and S11–S16 Tables). Mutation spectra were analyzed in the ura3 reporter on either side of the break to rule out the effect of chromatin landscape on the mutagenesis profile. We found that mutations were scattered throughout the open reading frame of the URA3 gene but clustered to several hotspots with base substitutions/deletions at homo-polymeric runs. Surprisingly, we only detected two multiple mutants out of over 300 sequenced mutation events. The low frequency of widely spaced multiple mutations in the URA3 reporter gene at DSB likely attributed to the small size of reporter gene (0.8-kb). G to C transversion-type events were dramatically elevated (42.4% without HO expression vs 83.3%, 60.9% or 64.4% after HO expression, see S4 Table) among mutations in the reporter placed at the 7.1 kb telomere-proximal location but not at the 5.8 kb centromere-proximal location after HO expression. We also observed minor differences in the mutagenesis patterns in strains with 15- or 203-bp repeats; for instance, recombination between 203-bp repeats induced far fewer base substitution type mutations at adenine relative to 15-bp MH or no homology repair events (p = 0.0549; S2–S4 Figs, S4 Table). The results suggest that MH-mediated repair is a powerful source of mutagenesis even for sequences that are tens of kilobases away from the break site.
To elucidate the basis for elevated mutagenesis in MH-mediated repair, we determined the timing of repair product formation by polymerase chain reaction (PCR) using primers flanking the repeats (Fig 2A, red and black arrows). To restrict our measurements of repair kinetics to a single cell cycle, we treated cells with nocodazole either prior to (S6A and S6B Fig) or just after (Fig 2B, S6C Fig) HO expression and rendered cells arrested at the G2 phase of the cell cycle. The cell cycle profile was confirmed by flow cytometry (S5 Fig).
We discovered that 15- and 203-bp repeat mediated repair events operate with distinctly different temporal kinetics in both conditions regardless of the order of nocodazole treatment and HO expression: SSA products using 203-bp direct repeats emerged at 2–4 h post-HO expression whereas the MH-mediated repair products were initially detected at 2 h but slowly accumulate up to 6–8 h post-HO expression (Fig 2B and 2C). These results further support the inefficiency of MH-mediated repair events.
We surmise that the protracted MMEJ kinetics reflects the instability of MH annealing and therefore, is inherent to the MMEJ process. Indeed, the rate of repair is higher between longer MH (18-bp) repeats than for shorter ones (15-bp), and the reaction follows first order kinetics (Fig 2C). To further test this possibility, we monitored MMEJ frequency in strains carrying MH of different melting temperatures by incorporating one or more base mismatches within the 18-bp repeats, thereby reducing the stability of MH pairing (Fig 2D). We discovered that the frequency of MMEJ was proportional to the melting temperature of flanking MHs (Fig 2D and 2E), supporting the premise that the stability of MH dictates the MMEJ frequency and corresponds to a key parameter of successful repair by MMEJ. Interestingly, the position of the mismatch also impinged on the MH-mediated repair frequency such that the mismatches towards telomere-proximal or central locations more severely disrupt MH-mediated repair (Fig 2D and 2E).
Evidence suggests that the amount of resection is directly proportional to the time needed for the repair [31]. The slow kinetics of repair events using MH might be accompanied by extensive resection at flanking DNA sequences. We therefore measured the extent of resection in both SSA and MH-mediated repair events. To date, most resection assays measured the amount of ssDNA in donorless yeast cells that lack all efficient repair options except limited end joining events [15, 32, 33]. However, in cells where resection leads to successful repair, the amount of ssDNA corresponds to the sum of resection and repair synthesis, complicating the accurate measurement of the extent of resection. To determine the amount of end resection in MMEJ events, we instead measured the amount of new DNA synthesis because the resected DNA should ultimately be re-synthesized by repair synthesis (Fig 3A). To detect the amount of repair synthesis, we labeled newly synthesized DNA using a nucleoside analog, bromodeoxyuridine (BrdU) in strains expressing both a nucleoside kinase as well as an equilibrative nucleoside transporter [34].
The amount of new DNA synthesis (i.e. resection) was monitored by incubating nocodazole-arrested G2 cells carrying 18-bp MHs in medium containing BrdU, which incorporates into nascent DNA during repair synthesis upon HO expression. Genomic DNA isolated from cells at several time points post-HO expression was pulled down with anti-BrdU antibody and analyzed by qPCR using a series of primer sets that anneal to the regions flanking the DNA break (Fig 3A). We found that BrdU incorporation extended up to 7.8-kb distal to the nearest repeats (location C) in MH-mediated repair (Fig 3B). The results were in stark contrast to SSA wherein the incorporation was not detectable even at 3.9-kb from the proximal repeat (Fig 3B, green bars). BrdU incorporation was not detected in strains lacking an HO cleavage site or in those without repeats (Fig 3B, orange bars and black bars, respectively). To further examine the extent of resection (and re-synthesis) during SSA, we constructed a strain in which the 527-bp repeat is situated asymmetrically at 5-kb distal and 0.5-kb proximal to the break site (Fig 3B, blue bars). In this strain, at least 5-kb of resection should occur to expose the requisite homology if resection proceeds symmetrically. Indeed, we found that BrdU incorporation is detected strongly at 3-kb proximal (location B) and up to 4.5-kb from the break (location C), but steeply declined at a site 6.1-kb proximal to the break (location D)(Fig 3B), indicating that resection is halted within a narrow zone about 1–2 kb beyond the repeat sequence. The results also suggest that the BrdU profile faithfully reflects the extent of resection and that MH-mediated repair events are accompanied by extensive DNA synthesis flanking the break site commensurate with the slow repair kinetics.
Emerging evidence suggests that ssDNA engenders elevated spontaneous and UV-induced mutagenesis [24–26, 35, 36]. According to this finding, mutation frequency may be directly proportional to the amount of end resection at given chromatin locations [25, 26], which could explain elevated mutagenesis in MH-mediated repair. Indeed, we found that UV treatment led to a dramatic (83,814-fold) increase in FOAR (and thus Ura3-) frequency among survivors after HO expression when the URA3 gene was inserted at 7.1-kb distal to the break site, and a moderate increase (279-fold) at 14.5-kb distal to the break site in a strain carrying flanking 15-bp MH (Fig 4A, S6 Table). In contrast, UV irradiation increased the frequency of FOAR survivors when the URA3 gene was inserted 7.1-kb (268-fold) or 11.5-kb (276-fold) distal to the break in long repeat strains (Fig 4B, S6 Table). As predicted, strong strand bias toward base substitutions at pyrimidines of the unresected strand was detected in the mutation spectra of the reporter placed at either side of the break (pyrimidine:purine = 31:2 and 22:6 at 5.8-kb centromere-proximal and 7.1 kb telomere-proximal to the break site, respectively) after UV and HO induction (S7 and S8 Figs, S5, S8 and S17–21 Tables). The results were consistent with the BrdU incorporation profile obtained from the ChIP assay that showed resection and repair synthesis reached at least 7.8-kb from the break site in MH-mediated repair but not in SSA (see Fig 3B).
In yeast, resection proceeds by two distinct stages: short range resection by the Mre11 complex, and long-range resection by Exo1 or Sgs1/Dna2 [13, 15, 17]. To test if resection and the formation of ssDNA trigger elevated mutagenesis flanking a DSB, we deleted EXO1 or SGS1, two enzymes responsible for different resection pathways, and measured the mutation frequency of a URA3 reporter gene at 7.1-kb distal to the break site upon HO induction. Should resection underlie elevated mutagenesis, deletion of EXO1 or SGS1 should reduce the mutation frequency upon HO expression in a strain carrying 15-bp MH. Indeed, the mutation frequency was reduced to 22- to 11-fold in sgs1 or exo1 deletion cells, respectively (Fig 4C). This was true even without UV irradiation (Fig 1D). We also found that break-induced mutagenesis depends on the Rev1 and Rev3 error-prone polymerases [37], but only moderately on Pif1 [38] and Rad30, suggesting that bubble migration as seen in break-induced replication is not chiefly responsible for the elevated mutagenesis in MH-mediated repair (Figs 1D and 4C)[39–41].
Previously, we showed that flanking MH could trigger promiscuous end joining and chromosomal translocation if the repeats are placed in two different chromosomes [6]. To test if MH also induces hypermutagenesis at the regions flanking breakpoint junctions of chromosomal translocations, a yeast strain carrying two HO recognition sites, one at the MAT locus on chromosome III and the other at the ura3 locus on chromosome V, was engineered to have 17-bp MH 2-kb telomere proximal to both HO cleavage sites (Fig 5A). The strain also contains a galactose-inducible HO endonuclease gene and lacks the HML and HMR loci so that the addition of galactose to the culture medium will induce DSBs at both cleavage sites but not elsewhere. DSB repair in this strain can proceed by intra- or inter-chromosomal MH-mediated repair and NHEJ. To distinguish the types of repair events, we have placed hygromycin resistance (HPH) and TRP1 genes next to the HO cleavage site so MH-mediated repair will lead to hygromycin sensitivity and/or tryptophan auxotrophy. The formation of a chromosomal translocation was then determined by PCR across HO cleavage sites using primers that anneal to two different chromosomes (Fig 5A).
Upon inducing HO for only two hours, over 20% of cells survive, of which at least 90% repair one or both breaks by NHEJ (Fig 5B). The predominance of NHEJ events among survivors after 2 h of HO expression could be attributed to relatively faster kinetics of NHEJ compared to MH-mediated repair [10] and therefore, NHEJ likely acting before MMEJ for DSB repair. Faster NHEJ kinetics also limits the frequency of chromosomal translocations among survivors due to the inherent bias of NHEJ to intra-chromosomal joining [42]. In events where both ends are repaired by MH-mediated repair (i.e., hph−trp1–), no intra-chromosomal repair bias is found (S9C Fig). In contrast, most survivors from persistent HO expression arose from MH-mediated repair events (hph–), nearly half of which are chromosomal translocations, consistent with our previous observations (S9 Fig) [6].
To measure mutation frequency associated with MH-mediated chromosomal translocations, we inserted the URA3 reporter at 7.1-kb distal to the MAT cleavage site and measured the frequency of FOAR survivors after 2 h of HO expression. Induction of HO increased the FOAR frequency 55.6-fold and most (91.8%) FOAR survivors were hygromycin sensitive, indicating that MH-mediated repair events were significantly enriched among mutagenic repair (Fig 5C). The majority of mutations were few base pair indels or base substitutions at homopolymer runs as described in [25]. Notably, G to T transversions were increased 4-fold (p = 0.02) after DSB induction (S10 Fig, S9, S22 and S23 Tables). Strong enrichment (11-fold, p<0.001) of MH-mediated repair events (hph- events, confirmed by analyzing the repair junctions) among FOAR survivors further confirms the high mutagenicity of such events as compared to those repaired by NHEJ (hph+ events, confirmed by analyzing the repair junctions). Furthermore, among FOAR survivors, chromosomal translocation events increased by 2.3-fold, with total survivors increasing from 6.8% to 16.1%. Together, these results suggest that MH-mediated chromosomal translocation could induce hypermutagenesis at sequences flanking the breakpoints.
Repair of a DNA double strand break (DSB) is frequently associated with elevated mutagenesis due in part to mutagenic repair synthesis that reverts accompanying ssDNA back to duplex form [24–26, 35, 36]. Indeed, hypermutagenesis was reported in ectopic gene conversion and break-induced replication even if mutagenesis is not always associated with error-prone translesion polymerases [35, 39, 40, 43]. Certain trinucleotide repeats, short palindromes, and interstitial telomeric sequences also induce chromosomal fragility and mutagenesis to flanking DNA sequences [44–48], likely because they trigger the formation of DNA DSBs and mutagenic DNA repair synthesis. We now show that MH-mediated end joining (MMEJ) can be added to the list of pathways endowed with extremely high mutagenesis potential, even up to tens of kilobases from the break site, underscoring its genome destabilizing capacity. Importantly, many of these mutations share features of clustered mutagenesis at or near chromosomal translocation breakpoints in human cancer cells [24, 26], raising the possibility that MMEJ contributes to some of these mutations. Alternatively, hypermutagenesis at the breakpoint junctions of MMEJ events reflects that yeast repair synthesis relies on error-prone polymerases, whereas in vertebrate cells repair is achieved by higher fidelity polymerases.
We propose that hypermutagenesis in MMEJ is linked with its slow kinetics. This is an inherent feature of MMEJ due to its reliance on the annealing of short MHs (S11 Fig), which itself is thermodynamically unstable and also counteracted by the presence of RPA [3, 6, 23]. The differences between MMEJ, HR and NHEJ with respect to kinetics as well as cell cycle dependency might dictate the order and timing of repair pathway choice for DNA lesions in cells and thus repair outcomes and associated mutagenesis upon DNA damage. It may also explain why MMEJ is more prominent when other competing and faster acting pathways become depleted or deficient [19, 49]. Alternatively, or in addition to MH annealing, other constraints such as degradation of antagonizing factors or the late recruitment of MMEJ components to the break site could contribute to slow MMEJ kinetics. Indeed, emerging evidence in vertebrate cells indicates that MMEJ could be blocked by a proteasome inhibitor [50] and normally confined to unique sub-nuclear compartments [51]. Additional studies are necessary to determine the underlying basis of slow MMEJ kinetics.
To analyze MMEJ-induced mutagenesis, we employed an experimental strategy that entails replica-plating of the surviving colonies after several divisions on non-selective medium following HO expression (see Materials and Methods for additional details). The arrangement was necessary because MMEJ is a slow repair process and 5-FOA kills yeast cells rapidly and does not allow residual divisions needed to establish DSB induced mutagenesis. Acute cell killing by FOA medium might also explain why UV irradiation did not increase spontaneous mutagenesis in URA3 as compared to that measured CAN1 mutagenesis using canavanine containing medium for selection (see S2, S3, S6 and S7 Tables). Mutation frequencies without DSB induction were measured by direct plating onto FOA containing medium. Excessive killing of recently formed ura3 mutants after transfer to FOA containing medium could thus account for the apparent lack of mutagenesis upon UV irradiation. Nonetheless, our results fully establish that MMEJ is far more mutagenic than SSA and NHEJ, in which all mutation frequency measurements involved identical set-ups and the methodologies used.
The presence of MH in most pathogenic chromosome translocations and complex genome rearrangements highlights MH as a driver for genome destabilization via either variant end joining, HR, or template switch (TS) mechanisms [52]. Complex genome rearrangements (CGRs) and somatic rearrangements are also accompanied by dramatically high levels of mutagenesis of DNA sequences at or near breakpoints harboring MHs [27, 53]. Analysis of breakpoint junctions with single base-pair resolution from 95 tumor samples revealed that somatic rearrangements across all cancer cell types are frequently associated with hypermutagenesis up to 10-kb flanking the breakpoint junctions [29]. Most of these mutations are transversion types [28, 29]. These results raise an intriguing possibility that breakpoint mutagenesis could partially be attributed to MMEJ driven events. Specifically, mutations observed at locations far distal from the break site cannot readily be explained by current models of microhomology-mediated BIR or TS, yet are consistent with long-range mutagenesis in MMEJ [39, 52, 54].
Under experimental conditions, MMEJ and other repair events could sharply induce mutagenesis at DNA flanking DSBs. It raises the tantalizing possibility that break-induced mutagenesis could drive the progression of diseases and potentially dictate cellular responses to current treatment protocols. Mutations occurring at flanking DNA sequences could also offer a unique strategy to selectively target disease cells that harbor pathogenic chromosomal rearrangements using neighboring gene deficiency as additional biomarkers. However, mutagenesis might be confined to a small fraction of repair events and many of these mutations do not necessarily lead to gene deficiency. Nonetheless, it will be interesting to explore if MH at the breakpoint junctions impinges on the aggressiveness of diseases and/or the treatment outcomes and could thus be exploited to identify the best therapeutic approaches according to the types of repair events triggering chromosomal rearrangements.
All yeast strains (S10 Table) are derivatives of JKM139 or JKM179 and were made by amplification of the hygromycin B phosphotransferase (HPH) gene from pAG26 with 90-bp oligonucleotides, containing 20-bp of homology to HPH, various sizes of microhomology/homology sequence, and homology to the Z1 region of MATα/a on chromosome III. Briefly, the SS203 strain containing direct 203-bp repeats flanking the Z1 region of MATα on chromosome III was constructed by the Golden Gate technique using primers ssa1, ssa2, ssa3 and ssa4 [55]. For SS527 strain construction, 527-bp fragments encompassing MATα Z2 sequence and TAF2 3’ end sequence were amplified (527-F and 527-R) and fused with the HPH gene at the 3’ end by PCR, and integrated into the PHO87 gene locus by PCR-based gene targeting (primers to introduce homology for integration: Pho87-HYG-F and TAF2-3-R). Gene deletion mutants were constructed by a PCR-based technique using oligonucleotides flanked by terminal sequences homologous to the open reading frames of target genes. BrdU incorporating (BrdU-inc) strains were constructed by the one step integration method as described previously52.
Yeast cells grown in YEP-glycerol media for 16 h were serially diluted and plated onto YEP-dextrose and YEP-galactose plates. Galactose induces HO endonuclease expression. Short or pulsed HO expression was achieved by adding 2% (w/v) galactose to the logarithmically growing yeast cells in YEP-glycerol medium, and after the indicated time of incubation, aliquots of culture were removed and plated onto YEP-dextrose to inhibit further HO endonuclease expression. Survival frequency was calculated by dividing the number of colonies on YEP-galactose by the number of colonies on YEP-dextrose plates.
Logarithmically growing yeast cells were incubated in YEP-glycerol for 16 h and 2% (w/v) galactose was added to the culture 2.5 hours prior to or after nocodazole (15 μg/ml) induced G2 cell cycle arrest. At different time points (0–10 hours), aliquots of culture were harvested and genomic DNA was isolated using the MasterPure Yeast DNA Purification Kit (Epicentre Biotechnologies). The amount of repair product was determined by quantitative PCR using primers flanking the newly re-joined DNA and normalized by amplification of a control locus in the genome (YEN1 genomic locus). To eliminate uncut or NHEJ events, genomic DNA was digested with PsiI restriction enzyme prior to PCR analysis. The recognition sequence of PsiI is located in the inter-repeat DNA and is thus deleted in MMEJ products but not in NHEJ products.
A single colony of S. cerevisiae cells was inoculated in 2–3 mL YEP-dextrose and cultured for 12–24 h. One ml of cells was harvested, washed with YEP-glycerol, transferred to 200 ml YEP-glycerol, and cultured overnight. Nocodazole was added to the culture at a final concentration of 20 μg/ml, and cells were cultured for another 3 h (at this point, cells were examined under the microscope to ensure that >90% of cells are arrested at G2/M). A double strand break was induced by adding galactose at 2% final concentration, and BrdU was supplemented to the medium at 400 μg/ml. Cells were cultured for another 10- to 13-h (no repeats & 18-bp) or 4- to 6-h (203-bp & 527-bp), and then harvested and washed with 50 mM EDTA. Genomic DNA was isolated by standard glass bead-based DNA extraction. Isolated DNA was re-suspended in 200 μl TE supplemented with RNase A (100 ng/ml), incubated at 37°C for 1 h, and then sonicated to sheer the DNA to fragments ranging from 200 bp to 700 bp. DNA was separated by 1.2% agarose gel and fragments ranging from 200–700 bp were extracted using a gel purification kit (Qiagen). One μg of purified DNA fragments (20–50 μl), 10 μg ssDNA and 10 μl 10xPBS, supplemented with distilled H2O to a final volume of 100 μl, was mixed, pelleted in a microcentrifuge, and placed in a 100°C heat block for 10 min. The mixture was then supplemented with 400 μl PBST (PBS, 0.1% Triton X-100), and 1 μl anti-BrdU antibody, and incubated at 4°C with rotating for 2 hrs. Five μl of the reaction was taken as 1% input, and mixed with 200 μl elution buffer. IP reactions were supplemented with 30 μl Dyna magnetic protein G beads (Invitrogen), and incubated for another 2 h. DNA-antibody-protein G bead complexes were subjected to extensive washing as follows: 1) 1 ml lysis buffer (50 mM HEPES pH 7.5, 1 mM EDTA, 140 mM NaCl, 1% Triton X-100, 0.1% NaDoc) for 5 min, 3 times; 2) 1 ml high salt lysis buffer (50 mM HEPES pH 7.5, 1 mM EDTA, 500 mM NaCl, 1% Triton X-100, 0.1% NaDoc) for 5 min; 3) 1 ml washing buffer (100 mM Tris-HCl pH 8.0, 1 mM EDTA, 1% Triton X-100, 0.1% NaDoc) for 5 min; 4) TE (10 mM Tris-HCl, 1 mM EDTA) for 5 min. The supernatant was removed completely, and DNA-antibody complexes were eluted with 2 x 100 μl elution buffer (10 mM Tris-HCl, 1 mM EDTA, 1% SDS) by incubating the tube at 65°C for 15 min. Beads were precipitated by magnetic apparatus, DynaMag2, and the supernatant was transferred to a new tube. Eluted DNA-antibody complexes were supplemented with 10 μl glycogen, 25 μl 3 M NaAC (pH 5.6) and 500~750 μl ice cold ethanol and kept at -80°C for >2 hrs. DNA was precipitated by centrifugation at 13,000 rpm at 4 °C for 15 minutes. Precipitated DNA was resuspended in 300 μl distilled water and subjected to quantitative PCR analysis using a series of primer sets that anneal to the regions flanking the DNA break site.
Logarithmically growing yeast cells were incubated in YEP-Glycerol medium for 16 h and then diluted with fresh 2% (w/v) galactose (Gal) synthetic complete media to induce Gal-HO-endonuclease expression for generation of site-specific DSBs. After 4 h of growth in galactose 108 cells were spun down and plated onto 150 mm YEP-GAL plates. To test UV-induced mutagenesis, the YEP-GAL plates were subsequently irradiated with 20 J m-2 ultraviolet-C (UV-C) using a Stratalinker (Stratagene). UV-C treated and untreated cells on YEP-GAL plates were incubated at 30°C for 12 h and then replica plated onto media containing 1 mg/ml 5-fluoroorotic acid (5-FOA) and 60 mg/ml L-canavanine to select for ura3 and can1 mutants.
To measure the frequency of FOA-resistant colony formation, we used the replica plating of surviving colonies after short-term (12 h) growth on YEP-GAL instead of a more standard method that entails simply plating cells onto FOA-GAL plates. We opted for this strategy because MMEJ events proceed significantly slower than SSA or gene conversion events and such slow repair product formation could impinge on the rate of FOA resistant colony formation. Indeed, the standard plating method greatly underestimate (almost 89-fold lower) the FOA colony formation frequency in a strain with 15-bp repeats; in contrast, the values obtained from standard plating and replica plating methods are almost identical in SSA-induced mutagenesis. We concluded that measurement of mutation frequency by the standard plating method is not suitable for MMEJ-mediated mutagenesis analysis and far less accurate even if the replica plating method may lead to minor fluctuations. Importantly, the replica plating method used here is remarkably reproducible with only <20% fluctuation between different trials (3 independent trials).
In order to determine the level of induced mutations, we calculated “mutation frequency” with r/N (‘r’: the total number of mutants, ‘N’: the total number of cells plated). Since we scored mutation events that are induced by an HO break, and formed within a single cell cycle, the frequency should be more appropriate in this case. The assay measures the frequency of 5-FOA resistant colony formation per viable cells. For the statistical interpretation of the data, the web tool “FALCOR” was used to calculate confidence intervals about the median with the cumulative binomial distribution of the rank value of M [30]. Significance testing was done via the Mann–Whitney U test [56] using the FALCOR program. The binomial distribution function used to calculate 95% confidence intervals is: Pr (probability) = n!/k!(n-k)! x (0.5)n; n = number of cultures in the experiments, k = the rank value.
The ‘‘No-GAL”control cells (108) were plated on Media containing 5-fluoroorotic acid (5-FOA) and L-canavanine to measure the spontaneous mutation frequency. For an additional control, mutation frequency in a “no homology” strain was also measured. Continuous Gal-induced HO endonuclease expression led to only 0.1% survival in this strain; therefore, for accurate mutation frequency measurements, the “no homology” strains were treated with galactose in order to induce endonuclease for only 2 h. Otherwise, all strains were treated identically. For further analysis of type of mutation pattern, a single FOAR mutant was recovered from each culture to avoid scoring of redundant mutations arising from the same mutated parent, the URA3 reporter was amplified by polymerase chain reaction (PCR) using primers annealing upstream and downstream of the gene, and products were sent (Beckman Coulter) for single pass sequencing using multiple primer sets. The primer sequences and additional information are listed in the “Primer List” (Table 1).
Additionally, we performed the reconstruction experiments to illustrate the efficiency and the reproducibility of our mutagenesis detection method that involves replica plating rare mutant cells to FOA medium after 12 h of growth on YEP-GAL medium. Briefly, yeast cells with a wild-type URA3 gene and the mutated HO cleavage site at the MAT locus (URA+, FOA sensitive) were mixed with cells with mutations in ura3 and the HO site (URA-, FOA resistant) at two different ratios (105:1 and 10:1), plated onto YEP-GAL and FOA-GAL and grew them at 30°C for 12 h. The YEP-GAL plates were then replica-plated to FOA containing medium as described in our experimental protocol and incubated at 30°C for three more days. We scored the number of colonies on FOA plates and divided by the number of colonies grown on YEP-GAL plates. The median frequencies and the 95% confidence intervals were determined using the web tool “FALCOR”. We performed the experiments a total of three times to test the reproducibility of the mutagenesis measurement.
As shown in S24 Table, the mutation frequencies calculated by replica plating led to ~40% as compared to those analyzed by direct plating to FOA plates. The results suggest that the replica plating efficiency might correspond to approximately 50%. Most importantly, the mutation frequencies measured by the replica plating are remarkably constant and highly reproducible in three different tests with two different ratios of FOA+/- cell populations.
|
10.1371/journal.pcbi.1002745 | Emergence of the Mitochondrial Reticulum from Fission and Fusion Dynamics | Mitochondria form a dynamic tubular reticulum within eukaryotic cells. Currently, quantitative understanding of its morphological characteristics is largely absent, despite major progress in deciphering the molecular fission and fusion machineries shaping its structure. Here we address the principles of formation and the large-scale organization of the cell-wide network of mitochondria. On the basis of experimentally determined structural features we establish the tip-to-tip and tip-to-side fission and fusion events as dominant reactions in the motility of this organelle. Subsequently, we introduce a graph-based model of the chondriome able to encompass its inherent variability in a single framework. Using both mean-field deterministic and explicit stochastic mathematical methods we establish a relationship between the chondriome structural network characteristics and underlying kinetic rate parameters. The computational analysis indicates that mitochondrial networks exhibit a percolation threshold. Intrinsic morphological instability of the mitochondrial reticulum resulting from its vicinity to the percolation transition is proposed as a novel mechanism that can be utilized by cells for optimizing their functional competence via dynamic remodeling of the chondriome. The detailed size distribution of the network components predicted by the dynamic graph representation introduces a relationship between chondriome characteristics and cell function. It forms a basis for understanding the architecture of mitochondria as a cell-wide but inhomogeneous organelle. Analysis of the reticulum adaptive configuration offers a direct clarification for its impact on numerous physiological processes strongly dependent on mitochondrial dynamics and organization, such as efficiency of cellular metabolism, tissue differentiation and aging.
| Mitochondria control energy production, initiation of cell death and several other critical cellular processes. Most often, they form a constantly reshaping tubular reticulum spread over the cytosol. Despite extensive knowledge of mitochondrial physiology, accurate description of their large-scale architecture is missing, partly due to substantial variability of reticulum geometries found in different cell types, and a capability for fast radical changes. We address this shortcoming with a mathematical model representing the organelle as a cell-wide dynamical network subjected to opposing actions of fission and internal fusion – processes known experimentally but not yet accurately specified. This opens a way for the quantitative characterization of the large-scale organization by showing how particular types of the internal dynamics can shape the reticulum into the whole multitude of configurations observed in biological studies. Further analysis reveals that for a specific value of tip-to-side fission/fusion rates the network should undergo a radical reorganization. Because of the high morphological sensitivity to minute changes in fusion or fission rates close to the critical point, cells can quickly adapt the mitochondrial operation and structure to their actual needs at a low expenditure of energy.
| Whereas experimental biology provides important insights into numerous characteristics and protein complexes responsible for cellular physiology, understanding the functional properties of many organelles can best be achieved when cell-wide structural complexity is included. This is provided by mathematical models able to explore a wide range of spatial and temporal scales. In the present study we describe the spontaneous emergence of a cell-wide network of mitochondria as an effect of a small set of dynamical rules well characterized in biological studies.
Mitochondria are elongated intracellular organelles present in eukaryotic organisms ranging from yeasts to mammals. They are well known to produce the majority of cellular ATP, the universal form of energy required for most of cellular reactions and to be a critical checkpoint for the initiation of apoptosis. In the past few years mitochondria came into focus by ongoing discoveries of their central role in aging [1]–[3], ischemia [4], [5], development of cancer and common neurological and metabolic diseases [6]–[9] - processes dependent on complex interactions between mitochondrial subunits. In the cell they are present in variable numbers ranging from few to hundreds of entities, forming a dynamic partially interconnected reticular network which spreads over the whole cytosolic volume excluding the nucleus. The network architecture is rather diverse and flexible, is able to adjust itself on a time scale of minutes depending on the actual physiological condition, and is highly variable among different cell types. In the fully interconnected state, the network edges are approximately of cylindrical shape with a typical diameter a few hundred nm and highly varying lengths up to more than 10 µm [10]. The opposite extreme is the fully fragmented condition, where the mitochondria are roughly spherical vesicles with diameter similar to that of the aforementioned cylinders. In the majority of situations, a cell contains many networked clusters of varying sizes along with numerous individual mitochondria, thus representing an intermediate state between the two extremes (Fig. 1 A) [11], [12].
Quantitative description of this complex structure is currently not available, but it is thought to result from mitochondrial dynamics, governed by their constant intracellular motion along the cytoskeleton filaments and the ability to fuse and divide at varying positions at a time scale of minutes to hours [13]–[15]. The past years of experimental effort greatly enriched biochemical understanding of fusion and fission processes, which are accomplished via assembling the reaction-specific macromolecular complexes in mitochondrial membranes. Fission requires assembly of circular oligomers involving dynamin-like protein Drp1 attached to the outer membrane anchor Mff (in mammals) followed by GTP-dependent scission of the mitochondrial body perpendicular to the cylinder long axis. The fusion is performed by proteins (in mammals these include Mfn1, Mfn2 and Opa1) mediating tethering and subsequent connection of membranes surrounding the two input organelles and thereby generating a continuous body [16]–[19]. Despite variations in regulation and homology levels of the protein components, the above blueprint of the fusion/fission progression was found to be largely universal among different species [17]. Much less clear is what influence these elementary events have on the structural properties of the network as a whole and what their implications are regarding the functional efficiency of the cellular chondriome. Studies of other complex systems found that collective dynamics is often associated with phenomena not directly deducible from behavior of their isolated constituents [20], [21]. In relation to mitochondria, this distinction is potentiated by experimentally established interconnection between the organelle’s functionality and its reticulum configuration [22], [23 and Refs. herein].
Here, we report a whole-cell dynamical model of the chondriome able to capture both the recognizable variability of tissue-specific mitochondrial architectures and the network intrinsic flexibility in response to metabolic requirements, as observed experimentally [12], [24]. First, we examine the mitochondrial reticulum by image analysis of fluorescence microscopy in order to identify the key types of fission and fusion processes responsible for the reticulum connectedness. Then, the network structure and dynamics are recreated numerically using both mean-field ordinary differential equations and explicit stochastic agent methods. The evolutionary graph-based representation introduces well-definable concepts making possible accurate characterization of mitochondria in quantitative terms. Its in-depth analysis shows that (a) the two types of events - fusion and fission of mitochondrial bodies - are sufficient to explain the observed diverse organization of the mitochondrial reticulum, (b) the geometrical properties of the network are directly related and can be calculated from the frequencies of these two processes (and vice versa) based on a few assumptions well supported by experimental data, (c) the cellular mitochondrial reticulum should exhibit a percolation transition and (d) its plasticity can be well explained by the vicinity of its functional regime to the critical point.
The major characteristics determining collective properties of mitochondria are the ability of these elongated organelles to form tubular threads here referred to as segments and to further interconnect them into branched net-like structures extending over the volume of the cytosol within a cell. This continually changing reticular arrangement is clearly observable with fluorescence microscopy (Fig. 1 A). A comprehensive model of cellular chondriome must be based on a correct representation of essential dynamics of its components. Despite the overall agreement that the mitochondrial morphology is shaped by a delicate balance between fusion and fission [17], [25], different kinds of these processes are conceivable for the formation of three-dimensional tubular objects such as mitochondria. Which processes are actually active in living cells requires experimental elucidation.
Upon noting that the underlying dynamic behavior of mitochondria leaves a distinctive and stable footprint on the resulting network structure, the details of fission and fusion processes can be deduced from the morphological still image analysis. This is because different types of fission and fusion mechanisms (e.g. tip-to-tip in contrast to side-to-side fusion, etc.) do involve process-specific kinds of network nodes. Segmentation processing of confocal microscopy images was employed here to experimentally establish the types of fission and fusion events forming the mitochondrial reticulum based on its structural properties. Configuration of a spatial network is specified by the connectedness of its branches at node points, expressed as node degree k (number k of edges connecting the node to the rest of the network, Fig. 1 C) and does not explicitly involve branch lengths or other geometrical attributes [26], [27]. As a consequence, the mere presence or absence of nodes of degree k, rather than their relative abundance is sufficient in order to deduce a structure of the underlying dynamics and thus formulate a general model of the mitochondrial network and its evolution.
The main advantage of this procedure is circumvention of the problems related to explicit 4-dimensional reconstruction of the network, not sufficiently reliable yet (due to the limited resolution of the current microscopes in z-dimension, required to be less than the tubule radius, potentiated by the augmented noise resulting from fast scanning) [28]. With the structural aspects kept fixed by the experimental assessment, the consecutive mathematical model will explore all the possible geometrical configurations by varying the reaction rates. This standpoint mirrors the available evidence that organisms and tissues share fundamentally similar kinds of mitochondrial fission and fusion events, despite the differences in protein structure or compositional details of the corresponding molecular machineries [16].
The image processing algorithm identifies the network segments and their connection points (Materials and Methods, Fig. 1 B). Although some of the free mitochondrial ends (i.e. nodes of degree k = 1) in the pictures are artifacts arising from the apparent cutting a reticulum branch by confocal slicing, false higher degree nodes cannot be produced that way, leaving the branching points (k>2) unaffected. This enables the determination of the branching point types using single confocal sections per cell, therefore selecting the most advantageous focus position near the cover glass, where the reticulum structure is best resolvable (Fig. 1 A).
With a proper resolution, image voxel size has an effect on visible lengths of the mitochondrial tubules (and hence on apparent fraction of the bulk nodes k = 2) but not on the reticulum branching organization. Because the network structure is determined solely by the branching node types, quantitative image analysis is restricted to these only, without experimental evaluation of the bulk nodes contribution (inset at the bottom part of Fig. 1 B). On the other hand, in view of the importance of the nodes of degree two for the segment sizes and kinetics, these are included in the detailed model (see the next Section). There, relative contribution of all the node types will emerge dynamically from the fission/fusion activity.
Computational image analysis of mtGFP-harboring mitochondria in HeLa cells revealed an exclusive presence of apparent node degrees 1≤k≤4 (n = 7, see Materials and Methods for details). Complete lack of higher degree nodes indicates a very limited number of fission and fusion mechanisms involved in the network dynamics, in accordance with the small number of protein complex types known to perform these reactions [16], [24]. Relative abundance of nodes of different degrees is important, because it reveals geometrical constraints favoring the node formation of particular degree. So, the mechanisms capable of generating the branching node types k = 3 and 4 are dissimilar and can be enumerated explicitly. Assuming single node formation/disruption times much shorter than the characteristic time of fusion/fission kinetics, the degree 3 nodes result exclusively from interactions of a mitochondrial tip (k = 1) with a side surface (k = 2) (a), while the k = 4 nodes can only be created either by (b) fusion of a mitochondrial tip to an already existing branching site node (k = 3), or (c) by fusion of two bulk sites (k = 2) in organelles touching each other side-to-side. Thus, comparative abundance of nodes of degrees 3 and 4 reflects the relative impact of the above reaction types on mitochondrial structure. For example, the detection of a high fraction of crossing tubules (k = 4) would evidence for the significance of (b) or (c) for the network dynamics. Without the above assumption of fast elementary events, also interactions of three and four nodes would be conceivable, although experimental observations of such processes were not reported in the literature. Because of the expected small frequency of these higher-order interactions, the mathematical model utilizes events involving two nodes only: For example, generation of a k = 3 node from a triple of free tips would then involve a pair of tip-to-tip and tip-to-side fusions quickly following each other.
Because the thickness of confocal slices cannot be made less than approximately two times the typical mitochondrial diameter, some of the detected branching nodes represent false positive connections resulting from the overlay of two unrelated organelles along the microscope optical path. However, this artifact can be corrected for by estimating its probability from the known volume fraction of the cytosol occupied by mitochondrial bodies and their diameter.
In the examined micrographs, the occurrence of all apparent k = 4 nodes is 17.0±6.9% of the total branching points (Supplementary Material Table S1). The correction due to the aforementioned optical artifact reveals that at least 80% of those nodes with k = 4 result from the overlay of unrelated organelles. In contrast, for k = 3 the overlay contributes to 7% of the detected nodes, and for k<3 its effect is negligible. Hence, ∼96% of the actual branching points are of degree 3, while the fraction of k = 4 nodes is statistically insignificant. Although measurements performed here are focused at peripheral cellular regions with lower density of microtubules, the extreme dominance of the simplest branching type strongly indicates that the general ability of mitochondria to perform complex fusions of type (b) and (c) is extremely low. We conclude that the spatial network of mitochondria is essentially connected with branching points of degree 3 resulting from the tip-to-side fusion activity (a), consistent with reports using alternative methods [29]. Accordingly, the model discussed in the following is designed to exclusively exhibit nodes with 1≤k≤3.
We represent the mitochondrial reticulum with a graph (nodes linked by edges, Fig. 1 C) consisting of the following node types: free ends of mitochondrial segments (k = 1), bulk sites (k = 2) and branching points (k = 3). Network edges interconnecting the nodes define minimal (indivisible) constituents of the organelle and their length l introduces a spatial scale to the network. The graph formalism does not imply actual physical existence of discrete subunits of uniform size l inside the mitochondrial bodies, but incorporates in a formal way their divisibility potential. Physically, l corresponds to the average distance between the membrane-bound fusion or fission complexes projected on the mitochondrial body axis. Finite value of l reflects the fact that at any moment the chondriome contains finite number of the molecular machines potentially capable of the reactions. When fluorescently stained, such complexes are directly observable as discrete punctuate structures on the surface of mitochondrial membranes [30], [31]. Thus, the network edges reflect the maximal achievable fragmentation state of the chondriome, which disintegration into numerous vesicle-like tiny mitochondria is routinely observed experimentally and can be induced by strong downregulation of fusion or upregulation of fission [30], [32]. In the model, this potential for fission is provided by bulk nodes interconnecting the edges into linear threads and by branching nodes. On the cellular scale (>>l), the overall dynamics then leads to a structure (Fig. 1 C) similar to the real reticulum network as seen in Fig. 1 A. With the chondriome internally discretized by the edges and nodes, the graph-based formulation does not require any externally imposed coordinate system or lattice. As a whole, the network has L edges amounting to the total mitochondrial length in the cell. Because fission potential of the mitochondria is restricted by the finite number of the available molecular machineries, the discrete model reflects the biological situation better than continuous representation would do. In this organelle, the discretization parameter L representing the divisibility limit should be viewed as an important observable. For example, in the specific case of HeLa cell line, the total mitochondrial length as estimated by visual inspection corresponds roughly to 5000 µm. Because under the microscope mitochondria (diameter ≈0.2 µm) in the maximally fragmented configuration resemble spherical vesicles [30], [32], their diameter can be used as typical network edge length l = 0.2 µm, thus giving the value of the discretization parameter L = 3·104 (≈5000/0.2). Keeping L as a free model parameter provides simple means for exploring the chondriome sizes relevant for different cell types.
Applying the principle of minimal complexity sufficient for the creation of the above network structure, in general we postulate two fusion and two fission types [33] represented as reaction processes on nodes Xk (Fig. 1 D):
Thus, the biological fusion or fission processes correspond to the network node transformations (Fig. 1 D) governed by specific evolution rules (Eq. 1). The graph-based, non-spatial formulation of the reticulum has the advantage that it allows studying structural development imposed by fusion/fission dynamics while omitting the detailed consideration of the underlying regulatory pathways, which may be organism or tissue-specific. By not discriminating biochemical protein species responsible for the corresponding reactions, the processes of Eq. 1 are set to account for their cumulative phenomenological effect by reproducing the observed spatiotemporal dynamics of the reticular structure, rather than mechanochemical description of inter-protein interactions. Still, taking into consideration that in mitochondria only one type of fission molecular apparatus is found experimentally, the description can be further simplified by assigning equal probability of a fission event per network edge: b2 = (3/2)b1 ≡ (3/2)b. This connection of fission rates for bulk and branching nodes via the number of participating edges reflects the fact that in both cases the fission event occurs by scission across a tubular body of the organelle.
Both the network dynamics and its equilibrium configurations result from intensities (a[], b) of a few well-defined reactions, Eq. 1. The current study considers in detail a wide range of constant a[] and b, which is sufficient for the characterization and explanation of the experimentally observed reticulum variability [11], [12], [16], [18], [30], [32]. Notably, interpretation of the reaction propensities may be extended by turning them into explicit functions of time- and concentration-dependent protein-specific kinetic processes. This would allow examination of the upstream regulatory mechanisms. However, such an expansion would involve inclusion of extramitochondrial biochemical pathways, exceeding the scope of this report.
The modeling framework proposed here (Eq. 1) can be utilized as long as the assumptions of (a) the network topology (1≤k≤3), and (b) absence of correlations between reaction events are fulfilled. This makes the model well suited for examination of mitochondrial networks, but would require major modifications in order to be applied to other cellular spatial networks such as endoplasmic reticulum. The condition (b) implies spatial isotropy and homogeneity inside the cell volume. Some anisotropy could arise in the very periphery of widely spread cells where microtubules may become arranged in bundles or preferentially oriented towards the cellular distal edge. So, care should be taken when applying the model to such cells. On very short spatial and temporal scale comparable with the kinetics of single motor proteins, mitochondrial fission and fusion may become also temporally correlated due to the influence of cytoskeleton. For example, two tips of mitochondria created by a recent division event but still attached to the same microtubule would have higher immediate propensity for fusion. However, such a deviation can occur only if the density of the cytoskeleton is sufficiently low, e.g. in the very periphery of the cell. In the bulk of the cytosol, frequent transitions between the fibers resulting from the high density of the cytoskeletal mesh common for the eukaryotes are expected to average the effect of single filaments out, leading to a fast decay of such correlations. This supposition was checked and confirmed by control simulations using an extended model where the mitochondrial network elements were assigned spatial positions by connecting them to explicit mesh of cytoskeletal fibers (data not shown). For configurations and densities of microtubules typical for central cellular regions, the prevailing majority of mitochondrial segments were found in the immediate proximity to several differentially oriented fibers simultaneously, leading to their fast reorientation. Consequently, the well-mixed environment is sufficient for the present investigation, in which the extremely short time scales are not considered.
Accounting for limited diffusivity in crowded cytosolic environment was shown to be important for accurate representation of molecular chemical systems due to their impact on effective kinetics relative to dilute conditions [34]. Mitochondria are much larger objects and are driven actively by motor proteins, but their mobility can still be affected by hindrances and other distortions commonly modeled by variations in diffusion coefficients. This kind of influence is accounted for in the current graph-based representation too, although a different approach is utilized. The dynamics of mitochondria is governed here by node transformation rates, and thus the diffusion in physical space is not explicitly implemented. Instead, the actual effective values for fusion/fission rates are taken from experimental measurements performed in living cells [15], which overall account for all known and unknown factors affecting the dynamics, without discrimination between those internal and external to mitochondria. This corresponds to a well-mixed approximation and allows the model to reproduce the architectural build-up of the reticulum while avoiding ambivalences related to still scarce experimental data on complex motility patterns of mitochondria subject to multiple regulatory mechanisms. Importantly, the framework of dynamic quantitative graph introduced here can be utilized for an upgraded model where coordinates, velocities, and/or external forces are assigned explicitly to the network constituents.
In addition to fusion/fission dynamics, the mitochondrial reticulum is subject to ongoing biological renovation in the course of import/export of material accompanied by organelle degradation due to selective mitochondrial autophagy (mitophagy). Carried out by protein molecules quite uniformly distributed over the mitochondrial body, the import/export is not known to influence or to be sensitive to the network structure. The total mitochondrial mass varies synchronously with the cell mitotic cycle or as a response to signals changing gene expression patterns – both processes much slower than the fusion or fission [15], [18], [25]. Hence, value of L, the parameter accounting for the organelle size is assumed time-independent here. The mitophagy potentially could influence the size distribution by preferentially depleting very small clusters because it is limited to small (∼1 µm) organelles. However, its impact on the mitochondrial architecture is negligible under physiological conditions because of the differences between time scales characteristic for the autophagy and the fission/fusion: while the newly imported or synthesized mitochondrial material was experimentally found to persist there for up to a few weeks, the mitochondrial motility is sufficiently fast to adapt the network conformation on the time scale of minutes to hours [13]–[15], [35]–[37]. Hence, the reticulum will be considered here well-equilibrated from the material turnover point of view. Notably, the model formulation allows for an extension explicitly incorporating autophagy into alternative implementations specifically targeted for investigation of longer durations. This can be useful if properties exceeding the mitochondrial structural organization, such as quality control mechanisms, were to be examined in detail [2], [38].
For the following, a mitochondrial segment is defined as one or more network edges connected only through bulk sites (nodes of degree 2), and a cluster as a detached set of interconnected segments possibly containing also branching nodes.
When expressed in the numbers of nodes per cell xk, Eqs. 1 translate into a system of differential-algebraic equations governing the network dynamics:(2)
The terms on the right-hand side of the first two of Eq. 1 represent the node kinetics due to each of the fission and fusion processes. For example, consumption of mitochondrial tips (k = 1 nodes) in the course of tip-to-tip fusion -a1x1(x1-1) is a product of the total number of node pairs potentially capable of fusion x1(x1-1)/2 and the fusion propensity for a pair 2a1. The last of Eq. 2 reflects conservation of the mitochondrial mass in the cell expressed as total number of edges L. Technically, it restricts the system phase space x ≡ (x1, x2, x3) to a plane which position and orientation are uniquely specified by L and the network structure k ≡ (1,2,3) respectively. While changing the applied discretization L of the model merely shifts the phase plane parallel to itself, relative proportions of the nodes and thus the state of the reticulum as a whole are well defined by the parameterization parameters (a[], b). Alteration of the reticulum discretization L corresponds to a different expression level of fission/fusion complexes without changing their relative abundance, while the values of a[] and b reflect independent activities of each of the network transformation processes.
The steady state of the system is a function of the ratio of fusion and fission rates c1 ≡ a1/b and c2 ≡ a2/b rather than the rates themselves:(3)
A unique set of real non-negative solutions x describing the network state corresponds to each triple of parameters (c1, c2, L). Fig. 2 exemplifies the steady-state solutions (Eq.3) for a particular chondriome size corresponding to HeLa cell line (L = 3·104, see previous Section) and a wide range of (c1, c2). In the extreme of infinitely strong tip-to-side fusions c2→+∞, the network tends to a fully connected “crystal” x → (0, 0, (2/3)L) consisting of branching sites only. Because in this regime the reticulum has no other node types, its geometry is insensitive to c1. The opposite case (c2 → 0) allows for many different configurations specified by the ratio of tip-to-tip fusion and fission rates c1: the set of possible states is flanked here by a fully fragmented x → (2L, 0, 0), c1→0 and a single loop x → (0, L, 0), c1→+∞. Reticulum states corresponding to some of these extremes were experimentally induced by artificial manipulation of mitochondrial fission or fusion activities [30], [32]. However, physiologically the most relevant parameter range is fairly narrow (see below), with all xk being far from saturation (Fig. 2).
The deterministic description of the mitochondrial geometry (Eq. 2) establishes a well-defined analytical connection between the molecular biochemical parameters (a[·], b, L) and network-wide structural variables like the average numbers of nodes xk, mean segment length for non-loop segments and total number of segments (x1+3x3)/2 comprising the reticulum. These represent the simplest quantitative measures of the chondriome architecture. They can be determined experimentally from the analysis of high-resolution cell reconstructions recorded with sub-diffraction microscopy, which is currently being extensively developed and shall become available in biological laboratories [39]–[41]. This type of recordings could also go beyond the mean values and measure among others the probability distributions of mitochondrial segment lengths and cluster sizes (discussed below). In addition to a more complete characterization, the latter variables may prove critical for clarification of key features of this organelle by revealing potential effects of stochastic fluctuations on mitochondrial operation and homeostasis (see the Section “Percolation phase transition in the mitochondrial reticulum” below).
In order to obtain a more detailed insight into the expected mitochondrial reticulum architecture, an agent-based stochastic simulation [42] of the same system was developed. As was verified by investigation of the potential fusion points in a setting with explicit representation of the cytoskeleton (data not shown), for the long-term evolution a non-spatial approximation is justified inside the majority of the cytosolic volume. Hence, a stochastic model was developed, where a set of L reactant objects corresponding to the network edges is submitted to the processes of fission and fusion as above (Eq. 1), i.e. between random nodes without explicit positioning in space. Reaction events and timings are put under the operation of the Gillespie algorithm [43] where nodes participating in a particular event are chosen randomly with equal probability among nodes of the same type. In the absence of detailed biochemical data on fission and fusion rates, the simulation parameters were adjusted to reproduce the experimentally observed average frequencies of fusion and fission events ≈0.25 (cell·sec.)−1 [15]. Explicit representation of the cellular mitochondrial system within the agent model allows for a comprehensive insight into the network characteristics surpassing the mean-field approximation of the deterministic description above. Hence, the following discussion is focused on the stochastic properties of the reticulum expressed in terms of statistical distributions.
Alternatively to the node-based description, the discussed network can also be viewed as a set of four types of (partially) connected segments, discriminated by degrees of the two end nodes (Fig. 1 E). The network state is then characterized by segment numbers of each of the four types: separate open-end segments , separate loops , as well as surface and internal segments of branched clusters. Index i denotes the segment length measured in edges.
Using the agent-based model, we find that with a good accuracy the steady-state distributions of mitochondrial segment lengths (Fig. 3 A) can be expressed as a superposition of two qualitatively different, fast and slow decaying, terms with their relative strength being strongly dependent on c1 and less on c2. The two components can be examined analytically by considering a simplified network where tip-to-side fusions are switched off (i.e. c2 = 0). As the degree of nodes is restricted here to k = 1 and 2, such a network consists of (1-1) and (2-2) segment types only, and equations governing their dynamics can be formulated explicitly. The segment numbers of length i and are derived by taking into account all possible transitions between segment populations:(4)
Here the prime under the sum denotes omission of the term j = i. The consecutive terms on the right-hand side of the Eq. for correspond to (1) creation of the segment resulting from fission of longer segments, (2,3) disruption and creation due to fission of same-size open-end segments and loops respectively, (4) formation upon fusion of shorter segments, and (5,6) disruption resulting from fusion to different and same size segments, respectively, and (7) formation of loops. Expressions for the steady-state segment length distributions in this network reveal unambiguous distinction in length distributions between loops and open-ended segments (Fig. 3 B):(5)with being the total number of open-ended segments. While Eq. 5 takes into account finite-size effects, as anticipated in real cells, for an idealized infinite system one would expect a geometrical distribution of because of the uniform probability density for a segment production or disruption over bulk nodes. On the other hand, the sizes of loops are governed by power-law. As a result, among the small segments the number of loops strongly exceeds the amount of the open-ended segments, while for the long segments the proportions are reversed.
In the general network (c2>0), the tip-to-side fusion and fission generates branched clusters consisting of variable segment numbers. The possibility of transitions between them induces a mild deviation from Eq. 5. Still, for all segment types other than disconnected loops (2-2), the geometric distribution , where p is the probability of a segment end provides a good approximation to the agent-based result. The mean length of segments can be estimated from the number of nodes introduced in the deterministic model (Fig. 3 C, solid line) and is therefore related to fusion and fission rates through Eq. 3. In the equilibrated system, is essentially equal among all non-loop segment types or cluster sizes (Fig. 3 C). Importantly, due to the strong decrease of at lower i (cf. Eq. 5, Fig. 3 B), loop segments amount to only a few percent of the total mitochondrial mass (Fig. 3 C, crosses), hence the mean length can be viewed as a good network-wide measure of mitochondrial reticulum structure (Fig. 3 C, solid line).
The geometric law of segment lengths is a direct consequence of the balance between opposing actions of fusion and fission. The pattern results from the positional independence of fission sites along the body of a mitochondrial tube, accompanied by the equal chances of network nodes to become involved in a fusion event (i.e. with probability being independent from the lengths of the segments). This corresponds to the dynamic behavior of a well-mixed system, as expected upon a cell-wide equilibration of the chondriome resulting from the permanent motion of individual mitochondria along the cytoskeleton filaments [13], [14].
For general values of c1 and c2 the system contains numerous segment clusters of variable sizes. Because the lengths of segments in a cluster are geometrically distributed independent random variables, the sizes j of clusters (i.e. the total number of network edges within a cluster) with r segments (Fig. 4 B) in the agent-based system conform (Fig. 4 A, colored markers) to the negative binomial distribution (Fig. 4 A, dashed lines)(6)where p is the probability of a segment end. The integral size distribution of all clusters in the system(7)involves weighting by the number of r-segment clusters g(r), which has no known closed form expression for this network geometry. Yet, upon using the agent-simulated g(r) together with Eq. 6, one recovers n(j) numerically (Fig. 4 A, black solid line, grey stars).
Variations in cluster sizes n(j) expected in a single cell are illustrated in the scatter plot of cluster sizes vs. number of segments in the cluster, for a set of different tip-to-side fusion rates c2 (Fig. 4 C). The dependences can be viewed as linear, with their coefficients (slopes of the scatter) corresponding to well-defined mean segment lengths discussed above. Increase in c2 leads to the conversion of bulk sites into branching sites and the resultant gradual reduction of (see also Fig. 3 C) accompanied by a higher segment number. Much less intuitive are c2-dependent nonlinear changes in variance (spread) of the cluster sizes (Fig. 4 C) discussed in the following.
For the simplified dynamics considered above (c2 = 0, Eqs. 4,5), n(j,r) would be reduced to a mixture of geometric and power laws (see previous Section), because such a network consists from single-segment clusters only.
The presented reticulum model (Eq. 1) exhibits a percolation phase transition in parameter c2 (tip-to-side fusion/fission ratio). In the thermodynamic limit of infinitely large network, the transition over a percolation threshold would correspond to an abrupt formation of a giant (percolating) cluster. At critical value a qualitatively new structure arises - a global cluster created from otherwise unconnected chondriome components spreading over the remote intracellular regions (Fig. 5, blue stars and schemes). In the finite network of a real cell the abrupt change of the network order is smoothed in the vicinity of the critical point, but rapid emergence and then dominance of the giant cluster for c2> is manifested when the fractional size of the largest mitochondrial cluster is plotted as a function of c2 while keeping c1 constant. Continuous (second order) phase transitions are marked by a peak in size fluctuations or other susceptibility measures at the point of criticality [44]. In the modeled reticulum the peak is visible as a sharp change in variability of non-percolating cluster sizes, e.g. expressed as the average number of segments per cluster (Fig. 5, pink circles). In a scatter plot, the rapid rise in cluster size fluctuations near the transition point is undoubtedly noticeable as elevated spread in sizes of individual clusters comprising the cellular chondriome (Fig. 4 C, yellow and green versus red and blue).
In contrast, no phase transition is found in tip-to-tip rate c1, where the critical point should be expected only at c1 →+∞, similar to a one-dimensional percolation problem on the lattice [44]. This result can be intuitively understood: While increase in tip-to-tip fusion rate c1 merely raises the average segment length, the tip-to-side fusions (c2) interconnect the segments among themselves and thus promote the formation of a global networked structure. Although from a mathematical perspective the predicted criticality is not surprising given the network topology assumed in the model, its relevance for the structure of mitochondria may have long-standing implications for cellular organization and functionality.
The predicted phase transition implies two distinct classes of mitochondrial structures: a subcritical network consisting of a relatively uniform set of multiple disconnected mitochondria, as opposed to a supercritical one characterized by the presence of a dominant giant cluster accumulating the majority of the mitochondrial mass, accompanied by a few much smaller satellite mitochondria (Fig. 5 bottom part). On one hand, from the experimental perspective, because only a relatively narrow range of c2 values corresponds to the transitional regime, the great majority of cells under unselective conditions could be expected to be found far from the transition point . On the other hand, frequent observation of intermediate chondriomes would indicate an underlying regulatory or self-tuning mechanism selectively promoting quasi-critical configurations.
Both highly fragmented and highly interconnected structures were induced experimentally in several cell types with pharmacological treatment severely shifting the fission/fusion balance [30], [32]. However, quantitative experimental investigations of the chondriome's undisturbed geometrical organization are, to our knowledge, not yet available. Here, we preliminarily examined the relevance of the predicted percolation transition for chondriome structure by measuring relative sizes of disconnected mitochondrial clusters in confocal images analyzed in the first Section (see also Materials and Methods). We find that in a standard mammalian cell line (HeLa, Supplementary Material Table S1), the largest mitochondrial cluster comprises 35% of the total visible reticulum (Fig. 5, dashed line), a fraction far from both the high-end and low-end saturation echelons expected for single-phase configurations (Fig. 5, blue stars). This value is accompanied by high (st. dev. 20%) diversity among cells, pointing to an elevated level of fluctuations in the cluster sizes. This result is further supported by the examination of relative sizes of mitochondrial clusters belonging to the same cell (data not shown). Both parameters indicate that under the normal physiological conditions the mitochondrial network may operate near the percolation transition point. However, due to small sampling of this preliminary measurement, currently only a compatibility with this interpretation can be stated. A more rigorous validation of the percolation transition predicted here will be achievable, when precise tuning of fission and fusion frequencies along with three-dimensional reconstruction of mitochondrial networks in living cells becomes biologically possible.
Growing experimental evidence indicates the fundamental interdependence between mitochondrial metabolic activity and its network structure [22], [23 and Refs. herein]. Yet, until now, the theoretical analysis of this vital organelle was either focused on its biochemical and electrophysiological aspects [45]–[47], or was reducing its architecture to a set of linear objects [48]. The examples of other complex systems indicate, that the manifestation of collective properties of the mitochondrial net should be critically dependent on proper representation of its structure and dimensionality [20], [21], [44], [49], requiring a cell-wide multiscale formulation. The current work examined the chondriome organization experimentally and introduced its network-based mathematical representation. This led to a detailed insight into the organelle, emerging as a mesh of tubular segments interconnected into larger flexible clusters able to reach distant cellular regions.
We find that fusion and fission dynamics should lead to a branched reticulum of tubules which lengths are well approximated by a geometric law and which mean size in equilibrium is determined by relative rates of these processes (Eq. 3). The whole network consists of disconnected clusters of such tubules. The cluster sizes are well approximated with a superposition of negative binomial distributions (Eqs. 5, 6). Notably, the distribution shape is distinctly convex (Fig. 4 A), featuring numerous tiny clusters coexisting along with a few relatively large entities. This property is expected to promote the experimentally observed disposal of damaged mitochondria by cellular autophagosomes [2], [50]–[52]. One reason for this is that mitochondrial dysfunction was shown to inactivate the mitochondrial fusion machinery, consequently leading to smaller mitochondrial entities [53], [54]. In addition, prevalence of small clusters present in the network supports the formation of autophagosomal bodies, which are not able to engulf objects larger than a few µm in mammals [37], [55]. Moreover, smaller clusters are expected to exhibit high statistical variance in functional efficacy facilitating the determination of removal candidates based on the inner membrane potential gradient or similar markers. In this way, the viability of an organism could be optimized by maintaining homeostasis of high-quality mitochondrial material on the whole-cell level [38].
Gradual disruption of this quality control e.g. as a result of natural aging is known to coincide with simultaneous rearrangement of mitochondrial reticulum structure due to alteration of fission and fusion rates [2], [15], [51]. By experimentally manipulating mitochondrial dynamics, the network reorganization was found to be sufficient for the induction of a substantial slowing down of the aging process [1]. The cluster size distribution predicted here offers a quantitative explanation for these and similar observations related to the strong dependence of mitochondrial structure and function [23].
To what extent the ongoing network dynamics is able to smooth out the differences constantly arising from diverse functional activity in distant parts of the chondriome? On a shorter time scale (∼minutes), the ongoing fission prevents complete homogeneity of the mitochondria over the cell body. Inside the mitochondrial clusters, additional equilibration results from molecular diffusion along the organelle tubules [56]. The geometric (exponential) distribution of the segment lengths predicted here is characterized by a high variance, with very long tubules connected to multiple shorter ones. It would be interesting to check what biological implications this diversity has for the organelle performance. For example, taking into account that key mitochondrial proteins in the inner membrane tend to compartmentalize while their diffusion is very slow [56]–[59], the presence of very long segments could significantly hinder equilibration of compositional differences between reticulum branches.
The balanced fusion and fission dynamics as in Eq. 1 leads to a network capable of a phase transition, i.e. that possessing two qualitatively dissimilar organizational modes. The critical transitional region lays in a narrow range of tip-to-side fusion/fission rates (c2, Fig. 5) where the reticulum structure is able to rapidly change its configuration, i.e. is very susceptible to the proper balance of these opposing processes. An experimental examination of the chondriome structure in HeLa cells reveals that its geometry corresponds to the transitional regime, characterized by the maximal heterogeneity in sizes of the network subcomponents. If this result is confirmed for other cell types, this would imply that under normal physiological conditions the cell has to maintain the fission and fusion rates quite precisely matched, despite the necessary flexibility in other parameters. The tight positioning inside the narrow transitional region rather than in one of the numerous configurations away from the critical point would induce questions about factors responsible for such a specific arrangement. Their assessment exceeds the scope of the current study, whilst the details of the network dynamics and architecture examined here could serve as an important counterpart. Similar combinations of high susceptibility and phenotypic robustness were found in other complex adaptive systems [60]–[62].
Independently from the underlying factors, it would be, indeed, advantageous for a cell to operate in the vicinity of the critical point because here the mitochondrial reticulum can be reconfigured with minimal energetic and temporal cost. Examples for such transformations include the quick and radical mitochondrial fragmentation upon activation of the apoptotic cascade or in the course of mitosis, where it is essential for a proper partitioning of the organelle between the daughter cells [63], [64]. Furthermore, high susceptibility of the mitochondrial network in the critical regime to small changes of the branching parameter naturally generates a high diversity between cells. Well known experimentally, it was often discarded as inessential intercellular noise amid a homogeneous population. However, this view is now changing due to recent observations connecting cellular structural aspects to its functional and gene expression patterns [65]. For example, in the course of organogenesis, the narrowly positioned regime of elevated flexibility can facilitate the tissue-specific reticulum alteration. With such a mechanism in place, during the differentiation phase cells can attain chondriomes best suitable for specialized energetic needs, inducing the variability of mitochondrial geometries found in different tissues and cell types [11].
In conclusion, the proposed model explains the self-organization of the chondriome into a dynamic network and its operation as a cell-wide adjustable construct, which large-scale characteristics provide the necessary connection between microscopical biochemical parameters and qualitative features central for the functionality of living cells. The predicted size distributions of segments and clusters are easily interpreted in physiological terms and verifiable by corresponding experiments.
HeLa cells were grown in Dulbecco Minimum Essential Medium (Sigma), supplied with 10% FCS (PAA) and 1% Penicillin Streptomycin (Sigma). Effectene (Qiagen) was used for transient transfection in 3.5 cm Petri dishes (IBIDI) according to the manufacturer's instructions. mtGFP experiments were performed 24 h after transfection. 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES, 10 mM final concentration) was added to the medium 1 h before measurements. Live cell imaging was done at 37°C using Nikon Eclipse TE2000-E microscope. A small pinhole ensured that the thickness of confocal slices did not exceed 800 nm. In order to reduce the effects of perinuclear region irregularities, the system was focused to a fraction of the intracellular volume adjacent to the Petri dish bottom.
Raw confocal images of HeLa cells (Fig. 1 A) with fluorescently visualized mitochondria were subjected to digital analysis designed to determine the reticulum network structure. As the first step, the images were thresholded and skeletonized after contrast optimization. The threshold position was chosen for each image such that no mitochondrial signal was lost while the background was cut off. Subsequently, the binary maps representing spatial graphs of the reticulum resulting from the skeletonization operation (Fig. 1 B, main field) were treated with a segmentation algorithm designed to identify and statistically analyze mitochondrial linear segments and branching points interconnecting them into clusters. An input skeletonized image consists of white pixels (1) on the background (0) (Fig. 1 B), the former classified as nodes of the mitochondrial network. Two white pixels were considered forming a connected graph if they were found adjacent to each other by applying an 8-connectivity criterion [66]. Degrees k of the resulting network nodes were calculated as the number of neighbors adjacent to the corresponding white pixel and, if appropriate, corrected to account for oversampling. The segmentation algorithm scanned the map starting from one of its edges and proceeded line by line. Upon encountering a white pixel, the graph to which it belongs was followed using a depth-first search method [67] until the whole cluster of adjacent pixels was traversed (upper box in Fig. 1 B shows the segmented structure). These pixels were then excluded from further search and the procedure advanced until the whole image matrix was processed. The image processing algorithm requires no initial assumptions regarding possible distributions of the reticulum segment lengths, their clustering or extent of connectedness.
Image contrast adjustment, thresholding and skeletonization were done using ImageJ (US National Institutes of Health, Bethesda, MD). Segmentation analysis algorithms, statistical and visualization procedures, as well as ordinary differential equation numerical solutions were implemented in MATLAB (The MathWorks , Natick, MA). Stochastic agent-based model of the mitochondrial network was designed using Intel Corp. (Santa Clara, CA) C++ compiler and run under Linux v2.6 kernel. Random numbers were generated using VSL routines, part of Intel Corp. (Santa Clara, CA) Math Kernel Library. For non-commercial use, the computer files comprising the agent-based model can be obtained free of charge upon contacting one of the corresponding authors (the e-mail addresses are given on the first page).
|
10.1371/journal.pntd.0005815 | Increased level and interferon-γ production of circulating natural killer cells in patients with scrub typhus | Natural killer (NK) cells are essential immune cells against several pathogens. Not much is known regarding the roll of NK cells in Orientia tsutsugamushi infection. Thus, this study aims to determine the level, function, and clinical relevance of NK cells in patients with scrub typhus.
This study enrolled fifty-six scrub typhus patients and 56 health controls (HCs). The patients were divided into subgroups according to their disease severity. A flow cytometry measured NK cell level and function in peripheral blood. Circulating NK cell levels and CD69 expressions were significantly increased in scrub typhus patients. Increased NK cell levels reflected disease severity. In scrub typhus patients, tests showed their NK cells produced higher amounts of interferon (IFN)-γ after stimulation with interleukin (IL)-12 and IL-18 relative to those of HCs. Meanwhile, between scrub typhus patients and HCs, the cytotoxicity and degranulation of NK cells against K562 were comparable. CD69 expressions were recovered to the normal levels in the remission phase.
This study shows that circulating NK cells are activated and numerically increased, and they produced more IFN-γ in scrub typhus patients.
| Orientia tsutsugamushi is an obligate intracellular bacterium. It primarily invades endothelial cells, macrophages, monocytes, and dendritic cells. Plasma concentrations of interferon (IFN)-γ, several cytokines and chemokines, which are known to recruit natural killer (NK) cells and T cells, were found to be increased in scrub typhus patients. NK cells are known as essential immune cells against several pathogens. In murine models of Rickettsial infection, the clearance of bacteria was found to be significantly associated with NK cell activity. Not much is known regarding NK cells’ role in O. tsutsugamushi infection in humans. This study is very possibly the first to measure NK cells’ level and function of in scrub typhus patients, or to examine NK cell levels’ clinical relevance. This study’s results demonstrate that circulating NK cells are activated and numerically increased in scrub typhus patients. Notably, increased production IFN-γ by NK cells of scrub typhus patients suggests their contribution to enhancement of intracellular bacterial killing in infected antigen presenting cells. Moreover, disease severity corresponded to increased NK cell levels. These findings importantly suggest that NK cells play a role in protecting the host against O. tsutsugamushi infection.
| Orientia tsutsugamushi is an obligate intracellular bacterium that causes scrub typhus in humans. It is a mite-borne, endothelium-targeting intracellular bacterium. Scrub typhus is prevalent in Asia, Northern Australia, and the Indian subcontinent. Most patients may recover from scrub typhus without complications if provided with an early diagnosis and management [1]. However, some patients develop fatal complications with median mortality of 6.0% unless they are treated sufficiently early in the course of illness [2]. O. tsutsugamushi resides in the cytoplasm of host cells, which are mainly endothelial cells, macrophages, monocytes and dendritic cells [3,4,5,6]. Related studies have found elevated plasma concentrations of interferon (IFN)-γ, IFN-γ-inducing cytokines (e.g., interleukin [IL]-12, IL-15, IL-18, and tumor necrosis factor [TNF]-α), and chemokines induced by IFN-γ (e.g., IFN-γ-inducible protein 10 and monokine induced by IFN-γ). These are well known for recruiting natural killer (NK) cells and T cells in patients with scrub typhus [7,8]. Based on these findings, a combination of innate and adaptive immune responses likely contribute to host defense against O. tsutsugamushi.
Natural killer (NK) cells are essential effectors within our innate immunity system. They mediate the elimination of target cells directly or indirectly through secretion of effector molecules such as perforin/granzyme, cytokines (mainly IFN-γ), and chemokines [9,10]. NK cells were discovered in the mid-1970s as they showed the ability to lyse tumor cells without prior exposure [11,12]. It is now well established that NK cells are also effective against several viruses, fungi, parasites and some intracellular bacteria such as Salmonella, Listeria and Chlamydia [10,13,14]. NK cell-mediated cytotoxicity is a complex process that involves receptor-mediated binding and signaling, synapse formation, granule polarization, and granule release [15]. Infection by intracellular pathogens leads to a decreased expression of major histocompatibility complex (MHC) class I antigens in host cells. This decrease reduces the infected host cell’s ability to interact with NK cells’ inhibitory receptors. In turn, the infected cell’s becomes more susceptible to lysis by NK cells, which leads to the destruction of the intracellular pathogen [16]. Similarly, K562 cells (which lack the MHC complex required to inhibit NK activity) are easily killed by NK cells. For this reason, these cells are often used for detection of NK cytotoxicity [17].
In murine models of Rickettsial infection, the clearance of bacteria was found to be significantly associated with NK cell activity and mice with NK cell deficiency showed increased susceptibility to infection [18]. However, studies have yet to explore the role of NK cells in O. tsutsugamushi infection in humans. Accordingly, this study aims to examine the level and function of NK cells in patients with scrub typhus, as well as the clinical relevance of NK cell levels.
The study cohort comprised 56 patients with scrub typhus (30 women and 26 men; mean age ± SD, 66.8 ± 13.0 years) and 56 age- and sex-matched healthy controls (HCs; 28 women and 28 men; mean age ± SD, 62.7 ± 8.03 years). Patients’ diagnoses required detecting O. tsutsugamushi antibodies via a passive hemagglutination assay (PHA) using Genedia Tsutsu PHA II test kits (GreenCross SangA, Yongin, Korea). Patients were positive for infection if test results showed a titer of ≥ 1:80 in a single serum sample, or at least a four-fold rise in antibody titer at a follow-up examination. According to the number of dysfunctional organs involved, severity of scrub typhus was classified into the following 3 grades as previously described [19,20]: mild disease (no organ dysfunction); moderate disease (one organ dysfunction); and severe disease (≥ 2 organ dysfunctions). Organ dysfunction is: (1) renal dysfunction, creatinine ≥ 2.5 mg/dL; (2) hepatic dysfunction, total bilirubin ≥ 2.5 mg/dL; (3) pulmonary dysfunction, bilateral pulmonary infiltration on chest X-rays with moderate to severe hypoxia (PaO2/FiO2 < 300 mmHg or PaO2 < 60 mmHg or SpO2 < 90%); (4) cardiovascular dysfunction, systolic blood pressure < 80 mmHg despite fluid resuscitation; and (5) central nervous system dysfunction, significantly altered sensorium with Glasgow Coma Scale (GCS) ≤ 8/15. All healthy controls were recruited in the Jeollanam-do area, which was the same area as the areas where the patients have developed. Controls had no history of autoimmune disease, infectious disease, malignancy, chronic liver or renal disease, diabetes mellitus, immunosuppressive therapy, or fever within 72 hours prior to enrollment.
The Institutional Review Board of Chonnam National University Hospital approved this study’s protocol. All participants provided written informed consent in accordance with the Declaration of Helsinki.
This study used the following mAbs and reagents: Allophycocyanin (APC)-Cy7-conjugated anti-CD3, APC-conjugated anti-CD3, APC-conjugated anti-CD69, and fluorescein isothiocyanate (FITC)-conjugated anti-CD45, FITC-conjugated anti-CD56, FITC-conjugated anti-CD107a, FITC-conjugated anti-IFN-γ, phycoerythrin (PE)-conjugated anti-CD3, PE-conjugated anti-CD56, PE-conjugated anti-CD69, PE-Cy7-conjugated anti-TNF-α, PerCP-conjugated anti-CD3, PerCP-conjugated anti-CD45, FITC-conjugated mouse IgG isotype and PE-conjugated mouse IgG isotype control (all from Becton Dickinson, San Diego, CA). Cells were stained with combinations of appropriate mAb for 20 minutes at 4°C. Stained cells were analyzed on a Navios flow cytometer using Kaluza software (version 1.1; Beckman Coulter, Brea, CA).
Peripheral venous blood samples were collected in heparin-containing tubes, and PBMCs were isolated by density-gradient centrifugation using Ficoll-Paque Plus solution (Amersham Biosciences, Uppsala, Sweden). NK cells were phenotypically identified as CD3-CD56+ cells by flow cytometry, as previously described [21,22]. NK cells were isolated using CD56 MicroBeads (Miltenyi Biotec, Bergisch Gladbach, Germany). The purity of CD3-CD56+ cells was greater than 95% as analyzed by flow cytometry. For sorting of CD69+ and CD69- NK cells, PBMCs were stained with APC-conjugated anti-CD3, FITC-conjugated anti-CD56, PerCP-conjugated anti-CD45, and PE-conjugated anti-CD69 mAb, and were sorted to obtain CD45+CD3-CD56+CD69+ and CD45+CD3-CD56+CD69- NK cells using a FACS Aria I sorter (BD Biosciences, Mountain View, CA) at purities of > 98%.
IFN-γ and TNF-α expression in NK cells were detected by intracellular cytokine flow cytometry as previous described [20]. Briefly, freshly-isolated PBMCs (1 × 106/well) were incubated in 1 mL complete media. The media consisted of RPMI 1640, 2 mM L-glutamine, 100 units/mL of penicillin, and 100 μg/mL of streptomycin. It was supplemented with 10% fetal bovine serum (FBS; Gibco BRL, Grand Island, NY). The incubation period was 24 hours in the presence of IL-12 (50 ng/mL; Miltenyi Biotec, Bergisch Gladbach, Germany) and IL-18 (50 ng/mL; Medical and Biological Laboratories, Woburn, MA). For intracellular cytokine staining, we added 10 μL of brefeldin A (GolgiPlug; BD Biosciences, San Diego, CA). The final concentration of brefeldin A was 10 μg/mL. After incubation for an additional four hours, cells were stained with APC-Cy7-conjugated anti-CD3, PE-conjugated anti-CD56 and APC-conjugated anti-CD69 mAbs for 20 minutes at 4°C, fixed in 4% paraformaldehyde for 15 minutes at room temperature, and permeabilized with Perm/Wash solution (BD Biosciences) for 10 minutes. Cells were then stained with FITC-conjugated anti-IFN-γ and PE-Cy7-conjugated anti-TNF-α mAbs for 30 minutes at 4°C and analyzed by flow cytometry.
Isolated PBMCs or purified NK cells and K562 cells (CCL-243; ATCC) were used as effector and target cells, respectively. Cytotoxicities of PBMCs and NK cells were evaluated by flow cytometry at an effector-to-target (E:T) cell ratio of 20:1 and 4:1, respectively, as previously described [17,21]. Briefly, isolated PBMCs and purified NK cells were cocultured with K562 cells for 4 hours. Mixed effector and target cells were stained with FITC-conjugated anti-CD45 mAb for 20 minutes at 4°C, then washed once in phosphate buffered saline (PBS). They were afterwards resuspended in 0.5mL of PBS containing 20 μL of 1 μg/mL propidium iodide (Becton Dickinson), and then incubated for 15 minutes at room temperature. A flow cytometry determined the ratio of dead K562 cells.
Luminex assay was performed according to the manufacturer’s instructions. Briefly, 25 μL of each supernatant from Quantiferon tubes was thawed and analyzed undiluted to determine the concentrations of cytokines, including IFN-γ, IL-17, and TNF-α, by Luminex assay using Milliplex MAP human Cytokine/Chemokine panel (Millipore, Billerica, CA) on a Bio-Plex 200 system with Bio-Plex Manager Software (version 4.1.1; Bio-Rad, Hercules, CA). Plasma levels of IL-12p40 and IL-18 were measured using a commercially available ELISA kit (R&D Systems Inc, Minneapolis, MN) according to the manufacturer’s instructions.
Degranulation of NK cells in response to K562 cells was determined by flow cytometry as previously described [23,24,25]. Briefly, freshly isolated PBMCs were stained with FITC-conjugated anti-CD107a or isotype control mAb and then incubated with or without K562 cells at an E:T ratio of 20:1. After 1 hour, monensin (GolgiStop; BD Biosciences, San Diego, CA) and brefeldin A were added, and the cells were incubated for an additional 4 hours at 37°C in 5% CO2. After the incubation, the cells were stained with PerCP-conjugated anti-CD3 and PE-conjugated anti-CD56 mAb for 20 minutes at 4°C, fixed in 4% paraformaldehyde for 15 min at room temperature, and analyzed by flow cytometry.
All comparisons of percentages and absolute numbers of NK cells, expression levels of CD69 and CD107a in NK cells, and cytotoxicity were performed by analysis of covariance after adjusting for age and sex using Bonferroni correction for multiple comparisons. Expression levels of IFN-γ and TNF-α in NK cells were analyzed using unpaired t-test. The Mann-Whiney U test was used to compare plasma levels of cytokines in scrub typhus patients versus age- and sex-matched HCs. Linear regression analysis tested associations between NK cell levels and clinical or laboratory parameters. A Wilcoxon matched-pairs signed rank test compared changes in NK cell levels and activation according to disease activity. P values less than 0.05 were considered statistically significant. SPSS version 18.0 software (SPSS, Chicago, IL) performed the statistical analysis. GraphPad Prism version 5.03 software (GraphPad Software, San Diego, CA) performed graphic works.
The clinical and laboratory characteristics of scrub typhus patients are summarized in Table 1. A total of 56 scrub typhus patients were included in this study. According to disease severity criteria that is based on number of organ dysfunctions, 33 patients (58.9%) had mild disease; 14 patients (25%) had moderate disease; and 9 patients (16.1%) had severe disease.
Flow cytometry determined the percentages and absolute numbers of NK cells in the peripheral blood samples of 56 patients with scrub typhus and 56 HCs. NK cells were defined as CD3-CD56+ cells (Fig 1A). Percentages of circulating NK cells were significantly higher in scrub typhus patients than in HCs (median 36.3% versus 19.8% [p < 0.0005]) (Fig 1B). Absolute numbers of NK cells were calculated by multiplying NK cell fractions by total lymphocyte numbers (per microliter of peripheral blood). Patients with scrub typhus had significantly higher absolute numbers of NK cells than HCs (median 507.6 cells/μL versus 380.8 cells/μL [p < 0.05]) (Fig 1C).
Based on the relative expression of the surface marker CD56, NK cells can be subdivided into CD56bright and CD56dim cells, as they exhibit different phenotypical and functional characteristics [26]. Previous studies have revealed that NK cell percentages, especially the proportion of CD56dim NK cell subset, were increased in multiple organ failure syndrome after trauma and metastatic melanoma [27,28]. To determine whether increased NK cell numbers in scrub typhus were the consequence of the increased proportion of CD56dim NK cell subset, the ratios of CD56bright and CD56dim NK cell subsets were measured by flow cytometry. No significant difference was observed in the ratio of CD56bright/CD56dim NK cell subsets between scrub typhus patients and HCs (Fig 1D).
We used a regression analysis to investigate the correlation between NK cell percentages in the peripheral blood and other laboratory or clinical parameters (Table 2). The aim was to evaluate the clinical relevance of NK cell levels in 56 patients with scrub typhus. In the univariate linear regression analysis, circulating NK cell percentages positively correlated with age, leukocyte count, neutrophil count, and disease severity (p = 0.012, p = 0.001, p = 0.001, and p = 0.008, respectively). Meanwhile, circulating NK cell percentages inversely correlated with serum protein level and serum albumin level (p = 0.003 and p = 0.004, respectively). We observed no significant correlation between NK cell percentages and lymphocyte count, hemoglobin level, platelet count, total bilirubin level, aspartate aminotransferase level, lactate dehydrogenase level, C-reactive protein level, or the erythrocyte sedimentation rate (Table 2).
To determine whether NK cells were activated during infection by O. tsutsugamushi, we used flow cytometry to examine the expression of CD69 in circulating NK cells from 31 scrub typhus patients and 18 HCs. Percentages of CD69+ NK cells were significantly higher in scrub typhus patients than in HCs (median 14.9% versus 2.9% [p < 0.005]) (Fig 2A and 2B). Scrub typhus patients had significantly higher absolute numbers of CD69+ NK cells than HCs (median 122.7 cells/μL versus 8.2 cells/μL [p < 0.0001]) (Fig 2C).
A variety of cytokines can regulate NK cell functions, including activation, cytokine release, and cytotoxicity. Next, we measured plasma levels of IFN-γ, IL-17, and IFN-γ-inducing cytokines, such as IL-12, IL-18 and TNF-α, in 25 patients with scrub typhus and 15 age- and sex-matched HCs using Luminex assay and ELISA. Scrub typhus patients had significantly higher plasma IFN-γ levels than HCs (median 43.8 pg/mL versus 6.3 pg/mL [p < 0.001]) (Fig 3A). Plasma levels of all IFN-γ-inducing cytokines, including IL-12, IL-18, and TNF-α, were found to be significantly higher in scrub typhus patients than in HCs (medians: 236.4 pg/mL versus 33.1 pg/mL [p < 0.0001]; 2536 pg/mL versus 503 pg/mL [p < 0.0001]; and 63.8 pg/mL versus 8.9 pg/mL [p < 0.005], respectively) (Fig 3C, 3D and 3E). However, plasma IL-17 levels were comparable between scrub typhus patients and HCs (Fig 3B).
Based on our observation that IFN-γ-inducing cytokines were increased in scrub typhus patients, we next examined whether NK cells might be activated after stimulation with IL-12 and IL-18. We cultured PBMCs obtained from 3 HCs with IL-12 and IL-18 for 24 hours. We then determined the CD69+ cell levels in NK cells by flow cytometry. Percentages of CD69+ NK cells were significantly higher in IL-12- and IL-18-treated cultures than in untreated cultures (mean ± SEM 26.2 ± 4.6% versus 2.2 ± 0.6% [p < 0.05]) (Fig 4A and S1A Fig). Furthermore, to determine whether the production of IFN-γ by NK cells might be linked to the activation of NK cells, we measured the expression of IFN-γ in the CD69+ and CD69- NK cell populations of HCs in the presence of IL-12 and IL-18 for 24 hours at the single-cell level by intracellular cytokine flow cytometry. Percentages of IFN-γ+ NK cells were found to be significantly higher in CD69+ NK cell population than in CD69- NK cell population (mean ± SEM 25.5 ± 3.2% versus 12.6 ± 3.8% [p < 0.05]) (Fig 4B and S1B Fig).
NK cells are a critical component of the innate immune response because of their capacity to produce a variety of cytokines. Among the most prominent cytokines produced by NK cells are IFN-γ and TNF-α [29]. To investigate the expression of these cytokines in NK cells, we incubated PBMCs obtained from five patients with scrub typhus and 10 HCs for 24 hours in the presence of IL-12 and IL-18. We then examined the expressions of IFN-γ and TNF-α in the NK cell populations at the single-cell level by intracellular cytokine flow cytometry. Percentages of IFN-γ+ NK cells were higher in scrub typhus patients than in HCs (mean ± SEM 29.1 ± 12.0% versus 6.9 ± 1.6% [p < 0.05]) (Fig 4C and 4D). Similar results were obtained even in an experiment calculated by the absolute numbers of IFN-γ+ NK cells (mean ± SEM 125.3 ± 65.3 cells/μl versus 14.2 ± 5.2 cells/μl [p < 0.05]) (S2 Fig). However, all of the scrub typhus patients and HCs exhibited low levels of TNF-α+ NK cells, which were comparable between the two groups (S3 Fig).
To examine the cytotoxic effect of NK cells on K562 cells, we used PBMCs and purified NK cells obtained from 20 patients with scrub typhus and 30 HCs. Cytotoxicities of PBMCs and purified NK cells were evaluated by flow cytometry at an effector-to-target (E:T) cell ratio of 20:1 and 4:1, respectively. The cytotoxicities were found to be comparable between scrub typhus patients and HCs (Fig 5A and 5B). Upon stimulation with K562 cells, the CD107a expression in NK cells was also comparable between scrub typhus patients and HCs (Fig 5C). Based on our observation that the expression level of IFN-γ was higher in CD69+ NK cells than in CD69- NK cells, we hypothesized that CD69+ NK cells could have an enhanced capacity to kill K562 cells. Thus, we determined cytotoxicities of purified CD69- and CD69+ NK cell subsets in scrub typhus patients by flow cytometry. However, we observed no significant difference in cytotoxicity between CD69- and CD69+ NK cell subsets in scrub typhus patients (Fig 5D).
We observed that circulating NK cell levels and CD69 expressions increased in scrub typhus patients; thus, we sought to determine circulating NK cell levels and CD69 expressions in the active and remission phases of the illness. Active and remission phases were defined as the period from the onset of symptoms to the start of antibiotic therapy on admission and the resolution of all presenting symptoms of scrub typhus after antibiotic treatment, respectively. Eleven scrub typhus patients were available for follow-up examination. As shown in Fig 6A, no significant changes in NK cell levels were found according to disease activity. However, CD69 expression was found to be significantly reduced when the disease was in remission than when it was active (median 3.5% versus 15.1% [p < 0.005]) (Fig 6B).
This is the first study to measure the level and function of NK cells in scrub typhus patients and to examine the clinical relevance of NK cell levels. The present study showed that circulating NK cell levels, together with elevated expression levels of CD69 and IFN-γ, increased in scrub typhus patients. Increased percentages of circulating NK cells reflected disease severity. Moreover, the plasma levels of IFN-γ, IL-12, IL-18 and TNF-α were significantly higher in scrub typhus patients than in HCs. In particular, with stimulation of NK cells with IFN-γ-inducing cytokines (i.e., IL-12 and IL-18), CD69 expression in NK cells was found to be increased, and CD69+ NK cells produced more IFN-γ than CD69- NK cells. Elevated CD69 expression in the active phase was normalized in the remission phase. However, the cytotoxicity and degranulation ability of NK cells were comparable between scrub typhus patients and HCs. Taken together, these findings suggest that cytokine-mediated activated NK cells accelerate the production of IFN-γ in scrub typhus patients.
Our results showed that percentages and absolute numbers of total NK cells in peripheral blood increased in scrub typhus patients. This result is substantiated by previous studies that found increased circulating NK cell levels in various human acute viral infections [30,31,32], whereas NK cell levels in peripheral blood have been reported to be decreased in influenza or HBV infection [33,34]. However, little is known about the circulating NK cell levels in human bacterial infection as compared with those in viral infection. Giamarellos-Bourboulis et al. have reported increased circulating NK cell levels in Gram-negative severe sepsis in humans [35], whereas others have reported a decline in NK cells in severe sepsis or septic shock [36,37,38]. Interestingly, only one study reported no evidence of change in the frequency of total NK cells in scrub typhus patients [39]. These discrepant findings about circulating NK cell levels might be due to the differences in pathogen, severity, time point at which the sample was obtained, and cohort selection bias including age and sex studied. In particular, age and sex are well-known confounding factors affecting NK cell levels in humans [40,41]. In the present study, comparative analyses showed that NK cell levels were compensated by age and sex, and that NK cell levels were still significantly higher in scrub typhus patients. Furthermore, elevated NK cell numbers have been found to be a direct consequence of induced proliferation in humans infected with hantavirus [31]. When considered together, these findings suggest that the numerical increase in NK cells might be due to NK cells proliferating in response to O. tsutsugamushi infection. Further study would be needed to confirm whether increased NK cell numbers are due to proliferation or recruitment to the blood.
Our data showed that CD69 expression and IFN-γ production in circulating NK cells were increased in O. tsutsugamushi infection, consistent with previous studies using several intracellular bacterial pathogens, including Rickettsia conorii, Haemophilus ducreyi, Salmonella typhimurium, Mycobacterium tuberculosis, M. bovis BCG, and Listeria monocytogenes [18,29,42,43,44]. It has been relatively well established that NK cell activation is closely related to cellular crosstalk with APCs such as dendritic cells (DCs) and macrophages as well as endothelium stimulated or infected by bacterial pathogens [29]. Based on the previous mechanistic study on NK cell activation by Haemophilus ducreyi [42], we speculated that O. tsutsugamushi-infected APCs possibly promote activation and IFN-γ production of NK cells through secretion of IL-12 and IL-18. The notion is supported by our observation that plasma levels of IL-12 and IL-18 were increased in scrub typhus patients and that addition of IL-12 and IL-18 upregulated CD69 expression in NK cells. Further study is required to determine whether APCs and NK cells form conjugates during cocultures with O. tsutsugamushi. In addition, Kang et al. have reported that the induced production of IFN-γ in NK cells affected the production of inducible nitric oxide synthase (iNOS) in APCs [45], which might contribute to intracellular bacterial killing in Listeria infection. Collectively, these findings suggest that increased production of IFN-γ by activated NK cells via cellular crosstalk with O. tsutsugamushi-infected APCs contribute to intracellular bacterial killing in scrub typhus.
NK cells kill virally-infected or malignant-host cells by degranulation of granzymes and perforin [46,47]. In several previous studies, cytotoxicity of NK cells against bacteria-infected APCs was found to be increased [18,43]. Our data showed that cytotoxicity and degranulation of NK cells against malignant cells (e.g., K562) were comparable between scrub typhus patients and HCs, suggesting that O. tsutsugamushi infection does not enhance the cytotoxic effects of NK cells.
We observed a positive correlation between circulating NK cell levels and age, leukocyte count, neutrophil count, and disease severity. This positive correlation has also been described in Epstein-Barr virus infection [48], whereas a negative correlation has been reported in influenza [49]. As shown in S1 Table, the univariate linear regression analysis showed that the absolute numbers of total NK cells were also positively correlated with leukocyte count and lymphocyte count (p = 0.008 and p = 0.001, respectively), but lost its correlation with age, neutrophil count, total protein level, albumin level, and severity. However, little is known about the dynamics of circulating NK cells related to disease severity in bacterial infection as compared with viral infection. Interestingly, CD69+ NK cell levels were also found to be positively correlated with leukocyte count, neutrophil count, and disease severity (S2 Table and S4 Fig). Taken together, these findings indicate that the activated subset rather than total NK cells contribute more to the correlation between the absolute numbers of NK cells and severity or inflammatory parameters in scrub typhus. Considering the previous observation obtained using animal models that overzealous activation of NK cells was related to organ damage, regardless of the infection itself [50,51], there is a possibility that an increased level of NK cells might be a cause of severe scrub typhus infection. However, this notion was contradicted by our data which demonstrated that increased CD69+ NK cell levels in the acute phase of scrub typhus infection were recovered to the normal levels in the remission phase, implicating that NK cell activation and expansion might be a consequence of severe scrub typhus infection.
In this study, increase in NK cell number levels did not change over our observation time, which was a relatively short time interval. This result is consistent with data from our previous study on mucosal-associated invariant T cell levels. In that previous study, mucosal-associated invariant T cell levels also did not change over the same observed time period [20]. A long-term follow-up study could determine whether circulating NK cell levels recover to normal levels after scrub typhus infection. Such a follow-up study would be in line with other previous studies, which found that normalization of NK cell levels in acute viral infections took a relatively long time [48,49].
In summary, the present study demonstrates that circulating NK cells are activated and numerically increased, and they produced more IFN-γ in patients with scrub typhus. In addition, increased NK cell levels reflect disease severity.
|
10.1371/journal.ppat.1004332 | Primary Seronegative but Molecularly Evident Hepadnaviral Infection Engages Liver and Induces Hepatocarcinoma in the Woodchuck Model of Hepatitis B | Hepadnavirus at very low doses establishes in woodchucks asymptomatic, serologically undetectable but molecularly evident persistent infection. This primary occult infection (POI) preferentially engages the immune system and initiates virus-specific T cell response in the absence of antiviral antibody induction. The current study aimed to determine whether POI with time may culminate in serologically identifiable infection and hepatitis, and what are, if any, its pathological consequences. Juvenile woodchucks were intravenously injected with inocula containing 10 or 100 virions of woodchuck hepatitis virus (WHV) to induce POI and followed for life or up to 5.5 years thereafter. All 10 animals established molecularly detectable infection with virus DNA in serum (<100–200 copies/mL) and in circulating lymphoid cells, but serum WHV surface antigen and antibodies to WHV core antigen remained undetectable for life. By approximately 2.5–3.5 years post-infection, circulating virus transiently increased to 103 copies/mL and virus replication became detectable in the livers, but serological markers of infection and biochemical or histological evidence of hepatitis remained undetectable. Nonetheless, typical hepatocellular carcinoma (HCC) developed in 2/10 animals. WHV DNA integration into hepatic and lymphatic system genomes was identified in 9/10 animals. Virus recovered from the liver virus-negative or virus-positive phases of POI displayed the wild-type sequence and transmitted infection to healthy woodchucks causing hepatitis and HCC. In summary, for the first time, our data demonstrate that an asymptomatic hepadnaviral persistence initiated by very small amounts of otherwise pathogenic virus, advancing in the absence of traditional serological markers of infection and hepatitis, coincides with virus DNA integration into the host's hepatic and immune system genomes, retains liver pro-oncogenic potency and is capable of transmitting liver pathogenic infection. This emphasizes the role for primary occult hepatitis B virus infection in the development of seemingly cyptogenic HCC in seronegative but virus DNA reactive patients.
| Introduction of highly sensitive molecular assays for detection of hepatitis B virus (HBV) identified the existence of persistent occult HBV infection years after recovery from an episode of hepatitis B and in individuals exposed to HBV but without symptoms and classical markers of infection. Because HBV integrates into human DNA and is a potent human carcinogen, it is postulated that occult HBV infection can be a cause of hepatic cancer in many individuals in which the tumor origin remains currently unknown. A causative relation between occult HBV infection and hepatocarcinoma is highly challenging to investigate in humans since occult HBV persistence is rarely diagnosed with current clinical assays and cancer development takes 15–30 years. However, we have established excellent models of occult HBV infection in the eastern North American woodchucks which are naturally susceptible to a virus closely related to HBV and in which chronic infection advances to liver cancer. In the current study, exploring experimental primary occult infection in woodchucks, we proved that the silently progressing infection, which is not detectable by serological markers, can culminate in hepatocellular carcinoma and that the persisting virus remains infectious, and causes hepatitis and liver cancer when transmitted to virus-naïve hosts.
| It is estimated that 370 million people have serologically evident chronic hepatitis B virus (HBV) infection and over 2 billion have been exposed to this virus [1]. Chronic hepatitis B (CHB) frequently (20–25%) advances to cirrhosis and liver failure, while hepatocellular carcinoma (HCC) develops in approximately 5% of the cases [1]. HBV is considered to be primarily hepatotropic; however, it also infects cells of the immune system where it persists for decades even when hepatitis resolves [2]–[4]. These events are closely mimicked in the natural animal model of HBV infection, the eastern North American woodchuck infected with woodchuck hepatitis virus (WHV) [4], [5].
WHV invariably invades the immune system and persists there for life irrespective of whether infection is symptomatic and serologically evident, i.e., serum WHV surface antigen (WHsAg) and antibody to WHV core antigen (anti-WHc) positive, or asymptomatic and serologically silent, i.e., serum WHsAg and anti-WHc nonreactive [6]–[9]. Based on the findings in naturally and experimentally WHV-infected woodchucks, two forms of occult hepadnaviral persistence were uncovered. Secondary occult infection (SOI) continuing after resolution of acute hepatitis (AH) and apparent clearance of serum WHsAg is accompanied by lifelong persistence of anti-WHc and WHV DNA in serum, liver and immune system [5]–[7]. The liver in SOI can display moderate inflammation with periods of normal morphology; nonetheless HCC develops in up to 20% of animals [7]. This form of infection appears to commonly underlie reactivation of hepatitis B in immunocompromised patients and those on cytotoxic therapies [10] and development of HCC in individuals with past exposure to HBV [11], [12].
Another form of hepadnaviral persistence, primary occult infection (POI), was uncovered in offspring of woodchuck dams convalescent from AH and in animals inoculated with WHV doses ≤103 virions [6]–[9]. POI progresses in the absence of identifiable serum WHsAg, anti-WHc and antibodies to WHsAg (anti-WHs), and hepatitis, but WHV and its replication are detectable at low levels in the immune system and sporadically in the liver [6], [9], [13]. Virus-specific T cell, but not B cell, responses are induced and, unlike SOI, protective immunity is not established. Our recent study showed that repeated exposures to liver nonpathogenic WHV amounts, i.e., ≤103 virions, do not culminate in serologically detectable infection and hepatitis or generate immune protection [13]. This type of infection can be suspected in HBV DNA reactive individuals who are seronegative for HBV surface antigen (HBsAg) and antibodies to HBV core antigen (anti-HBc) [14].
The main objective of the current study was to identify lifelong liver pathological consequences of POI, given the known high oncogenic potency of WHV and HBV [4], [15] and the notion that occult HBV infection undetectable by clinical testing might be responsible for HCC of unknown etiology in some cases [11], [12]. We also aimed to identify characteristics of POI regarding virus transmissibility and pathogenic potency in virus-naïve hosts, the status of WHV DNA-host genome integration during POI, and compatibility between virus sequences occurring in the liver virus-negative and the liver virus-positive phases of POI.
All animal experiments and the animal maintenance protocols were performed in compliance with the Institutional Animal Care Committee at Memorial University, St. John's, Newfoundland and Labrador, Canada (protocol identification number 13–159-M) that follows the guidelines and is accredited by the Canadian Council on Animal Care in Science.
Infection experiments were carried out in 1–2 year old healthy woodchucks housed in the Woodchuck Viral Hepatitis Research Facility at Memorial University, St. John's, Newfoundland and Labrador, Canada. All the animals were captured from a pristine region of Northern Canada. Normal liver function and morphology was ascertained by testing serum biochemical markers of hepatic performance, including sorbitol dehydrogenase (SDH) and γ-glutamyl transferase (GGT), macroscopic inspection of the liver during laparotomy, and by histological examination of liver biopsy taken prior to initiation of the study. Prior exposure to WHV was excluded by negative testing for serum WHsAg and anti-WHc, and by the absence of WHV DNA as determined by highly sensitive polymerase chain reaction/nucleic acid hybridization assays (PCR/NAH) (sensitivity ≤10 copies or virus genome equivalents [vge]/mL or ≤10 vge/µg total DNA) [7]–[9]. WHV/tm3 inoculum (GenBank accession number AY334075 for 3 identical clones) induced serum WHsAg-positive hepatitis in >90% of woodchucks after intravenous (i.v.) administration of doses ≥103 DNase-digestion protected vge, i.e., virions [7]–[9]. WHV/tm5 inoculum (GenBank accession numbers KF874491-3 for 3 clones) was derived from a woodchuck with chronic hepatitis and HCC. WHV/tm5 whole genome sequencing showed 99.7% (3298/3308) and 99.58% (1668/1675) identity in the nucleotide (nt) and amino acid sequences, respectively, when compared to WHV/tm3. Prior to infection, WHV/tm5 was fractionated on cesium chloride gradient to separate virions from free WHsAg, essentially as reported [8]. The recovery of intact virions was ascertained by a DNase-digestion protection assay [2]. To induce POI, 3 animals were i.v. injected with 10 virions of WHV/tm5 and 7 others with 100 virions of WHV/tm3. In addition, 2 animals inoculated with 106 WHV/tm5 virions, 2 inoculated with 1010 WHV/tm3 virions, and 2 healthy woodchucks not exposed to WHV, all followed for duration of their lifespan, served as controls. The animals were bled biweekly until 16 weeks post-infection (w.p.i.) and then bimonthly. They were followed for life until senility (animals 1/F, 3/F, 7/M, 10/M and 12/F), HCC development (5/M, 9/M and 11/M), a WHV-unrelated severe health issue requiring termination of follow-up (2/F, 8/M and 14/F) or challenge with 1010 virions WHV/tm3 at 66 months post-infection (m.p.i.) (4/F, 6/M and 13/F). The last group of animals was observed for an additional 5.2 months, and bled biweekly until 14 weeks post-challenge (w.p.c.) and then monthly. Liver biopsies were obtained before WHV inoculation and at 6 w.p.i., 8 m.p.i. and then at approximately yearly intervals until autopsy. At autopsy, serum, peripheral blood mononuclear cells (PBMC), liver, bone marrow, spleen, lymph nodes and other organ samples were collected.
WHV inocula were prepared from the liver virus DNA-negative and the liver virus DNA-positive phases of POI by pooling 24 mL of serum and plasma from 6/M and 7/M, and from 8/M that was liver WHV DNA nonreactive for life. Pellets recovered by ultracentrifugation at 200,000× g for 20 hours at 4°C were suspended in 1.3 mL of sterile phosphate buffered saline, pH 7.4. One mL of suspension was i.v. injected into a virus-naïve woodchuck and 0.3 mL used for WHV quantification and sequencing. Thus, A/F animal was injected with 1370 vge from 6/M, C/F with 1460 vge from 7/M, and E/F with 1460 vge from 8/M, all obtained from the liver virus-negative phase of POI. Also, B/F was infected with 2070 vge from 6/M and D/F with 1460 vge from 7/M collected from the liver virus-positive phase of POI. As a control, F/F was injected with 1010 virions of WHV/tm3. Plasma and PBMC samples were collected weekly until 8 w.p.i., biweekly until 6 m.p.i, and then bi-monthly. Liver samples were obtained before inoculation, 6–7 w.p.i., 6 m.p.i., then yearly, and at autopsy.
PBMC and plasma were harvested from sodium EDTA-treated blood after density gradient centrifugation [7]–[9]. PBMC were cryopreserved, and serum and plasma samples stored at −20°C. Liver samples obtained at biopsy or autopsy were washed, snap frozen and stored at −80°C. For histological examination, liver samples were processed to paraffin, stained and hepatic inflammatory alterations enumerated [7], [8]. Liver neoplastic changes were assessed following morphological criteria reported before [7], [16].
WHsAg, anti-WHc were evaluated by enzyme-linked immunosorbent assays (ELISA) reported previously [7]–[9], [13], with sensitivities comparable to or greater than those of clinical assays for detection of equivalent HBV infection markers. The sensitivity of WHsAg ELISA was 3.25 ng/mL while anti-WHc were detectable up to end-point dilution of 1∶64,000. Serum SDH served as a biochemical measure of liver injury and serum GGT as an indicator of HCC development [7], [13].
DNA from 100–400 µL of serum or plasma and from PBMC, liver, bone marrow and lymph nodes was extracted by the proteinase K-phenol-chloroform method [6], [7]. WHV DNA was assessed by direct and, if negative, nested PCR/NAH using primers and conditions reported [7]–[9]. Each sample was tested with 3 primer sets specific for WHV core (C), envelope (S) and X genes [7]–[9]. For nested PCR/NAH detecting WHV covalently closed circular DNA (cccDNA) (sensitivity, ∼102 copies/mL), enzymatic treatment, primers and conditions previously established were applied [8], [13]. Detection of WHV cccDNA was verified by sequencing. WHV RNA was detected by reverse transcription-PCR (RT-PCR; sensitivity, <10 copies/mL) using RNA extracted with Trizol (Invitrogen Life Technologies, Burlington, Canada), treated with DNase (Sigma-Aldrich, Oakville, Canada), and reversely transcribed to cDNA [8], [13]. Each test RNA sample without reverse transcriptase added served as a DNA contamination control [8], [13]. In selected cases, WHV DNA was quantified by real-time PCR (sensitivity, 10–100 vge/mL) using DNA equivalent to 25 µL of plasma or 400 ng of total DNA from cells or tissues, and WHV C and X gene primers. For all assays testing WHV DNA or RNA presence, mock extractions and respective nucleic acid preparations from WHV-positive and WHV-negative woodchuck livers or PBMC were routinely included as controls [6]–[8]. NAH analysis of PCR products was always performed to verify the specificity of virus detection and the validity of controls [6]–[8].
Low levels of WHV DNA in POI made full virus genome amplifications unfeasible, therefore fragments amplified with C, S and X gene-specific primers and regions spanning WHV polymerase (P) gene between nucleotides (nt) 2948-407 and 1080–1755, X/preC region nt 1503–2122 and preS nt 2948-407 were sequenced (nt positions according to WHV/tm3 AY334075 in GenBank). These regions were selected because they were found to have the most variable sequence based on analysis of full-length WHV genomes using Sequencher v5 (Gene Codes Corporation, Ann Arbor, MI). Amplicons were cloned using the TOPO-TA system (Invitrogen). Ten clones per amplicon were sequenced bidirectionally [17]. The same variants found in at least 2 clones were reported.
Liver and bone marrow DNA of 10–20 kbp purified from agarose served as a template for inverse-PCR (invPCR), as reported [18]. To identify WHV X region-host genome junctions, DNA was digested with Nsi-I that cuts WHV at nt 1915 (nt positions according to WHV/tm3 AY334075 in GenBank) and the woodchuck's sequence at unknown sites. To detect WHV preS region-host genome junctions, DNA was treated with EcoR-I that cuts WHV/tm3 at nt 3308/1. Diluted digests were circularized with T4 DNA ligase and linearized with Sph-I (for X invPCR) or Pst-I (for preS invPCR). The possibility of self-ligated virus double-stranded DNA was excluded by Psi-I or Pml-I digestion. Primers were designed based on consensus sequence of WHV isolates identified in this laboratory (GenBank accession numbers: AY334075, AY6280 and GU734791). For the X region, direct and nested primer pairs were located at nt 1782–1808 and 1718–1737, and 1853–1876 and 1654–1673, respectively. For the preS region, direct primers were located at nt 3231–3253 and 3009–3028, and nested primers at 3202–3222 and 2964–2985. The bands carrying WHV sequences were identified by NAH. DNA was purified by excision from agarose and either directly sequenced bidirectionally or cloned and sequenced. Non-WHV sequences were analyzed with NCBI BLAST and Refseq (National Center for Biotechnology Information, Bethesda, MD). WHV sequences were mapped by aligning with the full-length WHV/tm3 using BioEdit (Ibis Biosciences, Carlsbad, CA).
WHV sequences derived from the liver WHV-negative and liver WHV-positive phases of POI reported in this study were submitted to GenBank under accession numbers KJ755421 for woodchuck 6/M and KJ755420 for 7/M. WHV sequences identified in plasma and spleen of 8/M animal with POI have GenBank accession numbers KJ755405, KJ755406, KJ755410, KJ755411, KJ755415 and KJ755416, while those in E/F woodchuck injected with plasma inoculum derived from 8/M animal have GenBank accession numbers KJ755407-KJ755409, KJ755412-KJ755414, and KJ755417-KJ755419. WHV genome-woodchuck DNA integration sites identified in livers and bone marrows of animals with POI which developed HCC have accession numbers KG817076-85, KG817088, KG817089, KG817091 and KG817092, and those found in livers, PBMC and lymphoid tissues in woodchucks with POI without HCC have accession numbers KG817074, KG817075, KG817086, KG817087, KG817090, and KG817093-99. The sequence of woodchuck HCC H19 gene fragment identified in this study has GenBank accession number KG8117082.
Animals inoculated with 10 or 100 virions of WHV/tm5 or WHV/tm3, respectively, showed no serological evidence of WHV infection for up to 5.5 years p.i., as revealed by undetectable serum WHsAg and anti-WHc (Figs. 1A and 1B). Nonetheless, WHV DNA was detected in serum/plasma and PBMC throughout the entire follow-up at levels of 100–200 vge/mL or <103 vge/µg cell DNA, respectively (Figs. 1A and 1B). In contrast, woodchucks injected with 106 or 1010 virions developed transient serum WHsAg positivity, anti-WHc for life, and biochemical (not shown) and histological evidence of self-limited AH (SLAH) (Fig. 1C). WHV cccDNA and/or WHV RNA were identified in PBMC (Fig. 2 and Fig. 3A) throughout the lifespan and in lymphoid organs at autopsy in woodchucks with POI (Fig. 3B), similarly as in animals with lifelong SOI continuing after SLAH and as reported [7]–[9]. Sequential plasma or serum, liver and PBMC samples from healthy WHV-naïve woodchucks serving as controls remained WHV DNA negative when tested by nested PCR/NAH, while the animals liver and PBMC samples were WHV RNA nonreactive by nested RT-PCR/NAH during the entire observation period (data not shown).
It was not until 32 to 40 m.p.i. that all animals injected with 10 or 100 virions became consistently liver WHV DNA reactive (Fig. 1) and showed evidence of hepatic WHV replication, i.e., detection of WHV cccDNA or WHV RNA or both (Fig. 2). However, 9/M showed transiently a low level of virus DNA in the liver at 6 w.p.i., while 1/F and 5/M were reactive from 26 m.p.i and 6 w.p.i. onwards, respectively (Figs. 1A and 1B). Liver samples from 2/F and 8/M were consistently WHV DNA negative, even up to autopsy performed at 22 and 34 m.p.i., respectively. Liver histology and serum SDH levels remained entirely normal during the whole follow-up, except minimal inflammatory lesions limited to a few portal areas found at 32 m.p.i. in 5/M and at autopsy in 1/F and 10/M (Fig. 1). Despite this infection pattern, typical multinodular HCC has developed in 5/M and 9/M at 55 m.p.i. (Fig. 1B). The diameter of tumor nodules ranged between 2–3 mm (numerous) to 1.5–2 cm (singular) and they were spread throughout the entire livers. Histological examination revealed foci of well-differentiated HCC with hepatocytes arranged in trabeculea (Fig. 4) and, occasionally, with regions of compact cancer tissue. The HCC appearance coincided with moderately elevated serum GGT levels (data not shown). Control woodchucks injected with liver pathogenic doses of WHV showed transiently elevated serum SDH (data not shown) and SLAH followed by SOI accompanied by persistent low-level WHV replication in both liver and PBMC, and intermittent minimal to mild liver inflammation (Fig. 1C), as reported [7], [9]. One of the woodchucks (11/F) inoculated with 106 virions of WHV/tm5 developed HCC at 70 m.p.i. (Fig. 1C). Healthy controls not exposed to WHV had normal serum SDH and GGT levels during follow-up. Their livers remained normal during their lifespan when inspected macroscopically during laparotomies and by histological examination of serial biopsies obtained at approximately yearly intervals (data not shown).
To determine whether virus persisting as POI retained its infective and pathogenic properties, WHV recovered by ultracentrifugation from pooled serum/plasma collected from the liver WHV-negative or the liver WHV-positive POI phases was administered at doses between 1370 and 2070 virions to virus-naïve woodchucks. All animals developed transiently serum WHsAg-positive infection from 57–84 d.p.i. lasting for up to 113 d.p.i (Fig. 5). The WHsAg appearance was delayed by 22–49 days when compared to F/F control injected with 1010 virions. Anti-WHc became detectable at 70–113 d.p.i. (at 57 d.p.i. in F/F) and persisted to the end of follow-up. In animals inoculated with WHV from the liver virus-negative phase of POI, hepatic WHV load at 7 w.p.i. ranged from 30 to 100 vge/µg DNA, whereas in those with WHV from the liver virus-positive phase between 2.5×103 and 1.1×106 vge/µg DNA (9.5×106 vge/µg DNA for F/F). However, subsequent liver biopsies showed comparable WHV DNA levels ranging between 2×102 and 2×103 vge/µg DNA. Similar WHV DNA loads were detected in sera (10-102 vge/mL) and PBMC (10-102 vge/µg DNA) from the beginning of infection regardless of the inoculum source, but these levels subsequently increased by 10–100-fold. All animals (n = 5) developed mild to minimal hepatitis that persisted through the observation period (Fig. 5). Interestingly, 3 of them, including two inoculated with WHV from the liver virus-negative POI phase, developed HCC within 4.5 to 35 m.p.i. accompanied by a variable degree of hepatitis (Fig. 5).
To recognize whether initiation of the liver virus-positive phase of POI might be related to the emergence of a specific WHV variant, 2060-bp of WHV sequences derived from the liver virus-negative and liver virus-positive phases of POI from 6/M and 7/M were compared. The results showed that the WHV sequence from the liver-virus negative phase of 6/M differed only by one non-synonymous mutation in the preC region when compared to that of the virus from the liver virus-positive phase (Table 1). When WHV sequences from the equivalent phases from 7/M were compared, WHV from the liver virus-negative period showed 5 non-synonymous mutations not encountered in the virus from the liver WHV-positive phase (Table 1). However, none of the mutations were compatible with that identified in the preC region of 6/M WHV sequence, suggesting that unlikely a unique hepatotropic variant initiated the liver virus-positive phase of POI.
To determine whether WHV derived from the liver-virus negative POI phase retained its sequence after administration to a virus-naïve host, WHV sequences in inoculum and spleen from 8/M, which remained liver virus-negative until autopsy (Fig. 1B), and WHV from plasma, PBMC and liver from E/F, which was injected with 8/M inoculum (Fig. 5), were compared to each other and to WHV/tm3. This analysis showed that 2060-bp of the WHV sequence from 8/M inoculum and spleen displayed very few point mutations when compared to WHV/tm3 (11/2060), 8 conferred amino acid changes and 6 occurrrd in both samples (Table 2). WHV sequences from serum and liver of E/F were highly compatible to that of 8/M inoculum, while E/F PBMC showed a number of non-synonymous variants which were unaccounted for in WHV/tm3 inoculum (n = 31), 8/M inoculum (n = 25) or E/F serum or liver (n = 21) (Table 2), suggesting that the virus after transmission propagated most actively in the lymphoid cells.
Multiple WHV DNA-host genome junctions were identified in animals with POI which developed HCC (5/M and 9/M) or not (1/F, 2/F, 7/M and 10/M) (Table 3). Among virus-host integrants detected in liver biopsy and autopsy samples from 5/M and 9/M, various host sequences were joined predominantly with WHV X gene and less often with the polymerase (P) gene, and preS region sequences (Table 3). None of the virus-host integration sites was identified more than once in the material investigated; however, particular junctions were frequently found in more than one clone (Table 3). Notably, in HCC tissue from 9/M, the 264-bp host sequence flanked by the virus X gene sequence showed 80% homology with mouse H19 cDNA (GenBank accession number AF214115.1). H19 is a tumor suppressor gene and its knockdown may play a role in HCC development [19]. Further to virus-host junctions, multiple virus DNA rearrangements were identified in liver samples from animals with POI-associated HCC, but less frequently in those without cancer (Table 3). Viral-host junctions were also detected in autopsy bone marrow, lymph node and PBMC samples in all 6 animals analyzed (Table 3).
4/F and 6/M with POI lasting for 5.5 years were challenged with 1010 virions of WHVtm3 to determine whether the animals might be protected from reinfection. Both woodchucks became serum WHsAg positive at 2 w.p.c. and remained positive until 18 w.p.c. (Fig. 6). Anti-WHc became detectable from 13–14 w.p.c. Serum SDH levels increased and peaked at 8–14 w.p.c., while liver histology displayed moderate to severe AH (Fig. 6). WHV DNA levels in serum and PBMC were similar to those detected in control animals over the course of SLAH (Fig. 1C). Additionally, samples collected from 4/F and 6/M at autopsy, when serum WHsAg was undetectable, displayed low levels of WHV DNA in serum, liver and lymphatic organs, implying existence of SOI. Thus, both 4/F and 6/M developed acute hepatitis despite being persistently infected with WHV at a low-level. This was in contrast to 13/F with established SOI, which was protected from challenge (Fig. 6), similarly as previously reported [7], [9], [13].
We uncovered that minute amounts of hepadnavirus establish infection that persists indefinitely in the woodchuck model of hepatitis B in the absence of conventional serological markers of infection and hepatitis, but is detectable molecularly when sensitive virus nucleic acid-specific amplification assays are applied. We also documented that this form of asymptomatically hepadnaviral carriage, designated previously as POI [5], [9], [13], has both pathogenic and epidemiological relevance since it can lead to the development of HCC and, under certain conditions, transmit infection and cause hepatitis and HCC in virus-naïve hosts. Another important finding, albeit expected, was that POI is associated with hepadnavirus DNA integration into the hepatic and immune system DNA, which likely underpins liver oncogenic potency of the virus persisting during the course of this asymptomatic form of hepadnaviral carriage.
The results from the current investigations also showed that during POI, WHV replication expands to the liver with time, but the level and/or type of cells infected appear to be inadequate to trigger hepatitis. In previous studies, woodchucks with experimental POI were followed for up to 25 m.p.i. without detection of WHV in the liver or evidence of HCC, while WHV replication was detectable in circulating and organ lymphoid cells [9], [13]. Also, offspring born to woodchuck dams with SOI, which acquired lymphatic system-restricted POI, did not show liver engagement and the development of HCC during the 42-month observation period [6]. Although there might be several factors contributing to the development of HCC during POI, the virus spreading to the liver and the extended period of POI follow-up appear to be critical.
WHV genome fragments from the liver virus-negative and the liver virus-positive POI phases showed essentially the same predicted amino acid sequences (Table 1), which also were highly compatible to that of wild-type WHV inocula used to induce POI in this study. We analyzed more than 62% of the total WHV sequence, including virus regions identified as having the highest sequence variability based on our preceding analysis of the complete WHV sequences reported in GenBank. We used this approach because the trace quantities of WHV found during POI and the inherently lower sensitivity of the extended PCR amplifying long WHV sequences made full virus genome amplification not feasible. We also identified that the WHV/tm3 sequence, as far as we were able to determine, was conserved in the animals injected with WHV prepared from the liver WHV-negative or the liver WHV-positive phases of POI, which developed WHV infection engaging both the liver and the lymphatic system. These findings imply that the dual tropism of WHV towards hepatocytes and immune cells is unlikely due to the existence of cell type-specific viral variants but is an intrinsic propensity of the naturally occurring virus. This is consistent with data from in vitro infection experiments in which the same wild-type WHV was serially passaged in cultured woodchuck hepatocytes and lymphoid cells [17]. This issue has not yet been investigated in HBV infection.
Our previous studies showed that serologically overt WHV infection coinciding with hepatitis is resultant from i.v. administration of WHV doses greater than 1×103 virions (liver pathogenic doses), while lower doses of the same wild-type virus (liver non-pathogenic doses) consistently induced POI in woodchucks [9], [13]. In the current study, concentration by ultracentrifugation of virus from animals with POI to levels above the previously identified liver pathogenic threshold was accomplished and, as documented, the recovered virus readily induced serologically overt infection and hepatitis upon transmission to virus-naïve animals (see Fig. 5). It appears that the inability of WHV to engage the liver during the initial phase of POI was related to the very low quantities of the produced virus which, however, can be temporally augmented to the level sufficient to invade the liver. In this regard, we detected a transient increase in plasma WHV load to approximately or above 1×103 vge/mL that preceded detection of WHV DNA and its replication intermediates in hepatic tissue. It can be assumed that this temporal increase in circulating WHV was adequate to engage the liver during later phase of POI, which prior to that was restricted to the lymphatic system. This appears to be consistent with identification of a 100 to 1000-fold greater affinity of synthetic analogues of WHV cell binding site for activated woodchuck lymphoid cells than woodchuck hepatocytes, suggesting that very low quantities of virus may preferentially invade the immune system [20], [21].
The mechanism of liver carcinogenesis in hepadnaviral infection is not well understood, but it is likely a multistep process in which persistent virus infection and virus genome integration into host DNA are among the principal contributors [22]. Random HBV DNA integration into the liver genome was found in up to 22% of patients with CHB and is a typical finding in HBV-related HCC (>80% patients) [22]–[24]. On the other hand, the status of HBV DNA integration into HCC DNA coinciding with occult HBV infection was only occasionally investigated and mainly in cirrhotic patients [11], [12]. Nonetheless, the data convincingly showed that HBV DNA integrates into both HCC and non-HCC liver DNA in serum HBsAg-negative patients, with or without detectable anti-HBc [11], [12]. In WHV-related HCC, virus DNA insertions were identified in tumors developing during chronic hepatitis and SOI continuing after SLAH [25]. WHV DNA integration was frequently found near the myc pro-oncogenes in HCC coinciding with chronic WHV hepatitis [26], [27]. We did not find this relation in woodchucks developing HCC during POI. However, this might become more apparent when a greater number of relevant cases are analyzed. About two-thirds of the virus-host genome junctions detected in this study encompassed the WHV X gene sequence (Table 3). This resembles the predisposition of HBV X gene to integrate into the host genome reported in serum HBsAg-negative patients with HCC [28]. In our study, HCC had developed in the absence of hepatitis and cirrhosis. In contrast to CHB, chronic WHV hepatitis never leads to cirrhosis and very rarely to fibrosis (<1%) [4]. However, the occurrence of HBV-related HCC in the absence of cirrhosis has been reported [28]. The present finding of the POI-associated HCC mimics the human disease situation where HBV-related HCC develops in the absence of apparent chronic liver disease and serological evidence of HBV infection. Notably, the finding of WHV DNA sequence insertions within bone marrow DNA in POI parallels HBV DNA and WHV DNA integration into lymphoid cells and lymphatic organ genomes in CHB and in woodchucks with chronic hepatitis and SOI [29], [30].
This study also revealed that POI during lifelong follow-up did not culminate in serologically apparent infection or hepatitis, and did not induce protective immunity. These findings add new dimensions to the previous investigations on POI [6], [9], [13]. Among others, our previous study showed that repeated i.v. injections (12 in total) with 100 WHV virions did not initiate serologically detectable infection or hepatitis, but molecularly evident POI was established and continued until challenge with a liver pathogenic dose (>103 virions) of the same virus inoculum [13]. In the current study (data not shown) and in the previous investigations [9], [13], [31], [32], WHV-specific T cell reactivity occurring in the absence of virus-specific antibody response did not protect from challenge with liver pathogenic doses of WHV (>103 virions). This is in marked contrast to WHV-specific T cell responses coinciding with virus-specific antibodies in SOI continuing after recovery from symptomatic WHV infection and hepatitis which yield total protection against challenge with even massive doses of WHV (>1010 virions) [7], [31]. The former may parallel a situation in unvaccinated individuals having repeated contacts with infected persons and intravenous drug users repetitively exposed to small amounts of HBV. Our data implies that these individuals would unlikely become serum HBsAg and anti-HBc reactive or immune to HBV infection, but the development of HCC in such persons cannot be excluded. Nonetheless, the current findings are in contrast to data indicating that one virion of HBV derived from a HBV transgenic mouse was able to induce serologically evident chronic hepatitis in chimpanzee [33]. Differences in the liver pathogenic potency between a single HBV isolate from a transgenic mice and intact, naturally occurring WHV might explain this discrepancy.
Although the existence of POI in humans has not yet been thoroughly investigated, the prevalence of HBV DNA-reactive infection seronegative for HBsAg and anti-HBc has been reported between 0.07 and 7.6% of subjects in different areas of HBV endemicity [12], [34]. It can be expected that HBV POI is much more frequent because the assays available for HBV DNA detection are approximately 10–100-fold less sensitive than these utilized in this study. Further, HBV-specific T cell responses in the absence of serum HBsAg and anti-HBc have been identified in HBV DNA-reactive patients, further supporting that this silent form of HBV infection naturally occurs [14]. It is of note that WHV-specific T cell responses were also examined in the current study and they persisted at borderline levels after a period of heightened reactivity lasting between 6 and 20 w.p.i. (data not shown).
In conclusion, this study revealed the oncogenic capacity and potential epidemiological significance of asymptomatic hepadnaviral carriage initiated by very small amounts of otherwise pathogenic virus that advances in the absence of traditional serological markers of infection and hepatitis. The data emphasize the role for primary occult HBV infection in the development of seemingly cyptogenic HCC in HBV seronegative patients.
|
10.1371/journal.pcbi.1003386 | Electrodiffusive Model for Astrocytic and Neuronal Ion Concentration Dynamics | The cable equation is a proper framework for modeling electrical neural signalling that takes place at a timescale at which the ionic concentrations vary little. However, in neural tissue there are also key dynamic processes that occur at longer timescales. For example, endured periods of intense neural signaling may cause the local extracellular K+-concentration to increase by several millimolars. The clearance of this excess K+ depends partly on diffusion in the extracellular space, partly on local uptake by astrocytes, and partly on intracellular transport (spatial buffering) within astrocytes. These processes, that take place at the time scale of seconds, demand a mathematical description able to account for the spatiotemporal variations in ion concentrations as well as the subsequent effects of these variations on the membrane potential. Here, we present a general electrodiffusive formalism for modeling of ion concentration dynamics in a one-dimensional geometry, including both the intra- and extracellular domains. Based on the Nernst-Planck equations, this formalism ensures that the membrane potential and ion concentrations are in consistency, it ensures global particle/charge conservation and it accounts for diffusion and concentration dependent variations in resistivity. We apply the formalism to a model of astrocytes exchanging ions with the extracellular space. The simulations show that K+-removal from high-concentration regions is driven by a local depolarization of the astrocyte membrane, which concertedly (i) increases the local astrocytic uptake of K+, (ii) suppresses extracellular transport of K+, (iii) increases axial transport of K+ within astrocytes, and (iv) facilitates astrocytic relase of K+ in regions where the extracellular concentration is low. Together, these mechanisms seem to provide a robust regulatory scheme for shielding the extracellular space from excess K+.
| When neurons generate electrical signals they release potassium ions (K+) into the extracellular space. During periods of intense neural activity, the local extracellular K+ may increase drastically. If it becomes too high, it can lead to neural dysfunction. Astrocytes (a kind of glial cells) are involved in preventing this from happening. Astrocytes can take up excess K+, transport it intracellularly, and release it in regions where the concentration is lower. This process is called spatial buffering, and a full mechanistic understanding of it is currently lacking. The aim of this work is twofold: First, we develop a formalism for modeling ion concentration dynamics in the intra- and extracellular space. The formalism is general, and could be used to simulate many cellular processes. It accounts for ion transports due to diffusion (along concentration gradients) as well as electrical migration (along voltage gradients). It extends previous, related formalisms, which have focused only on intracellular dynamics. Secondly, we apply the formalism to model how astrocytes exchange ions with the extracellular space. We conclude that the membrane mechanisms possessed by astrocytes seem optimal for shielding the extracellular space from excess K+, and provide a full mechanistic description of the spatial (K+) buffering process.
| The interaction between neurons and glial cells has been the topic of many recent studies within the field of neuroscience (see reviews in [1]–[3]). Astrocytes (a species of glial cells) play an important role in modulating excitatory and inhibitory synapses by removal, metabolism, and release of neurotransmitters [4], homeostatic maintenance of extracellular K+, H+, and glutamate [5], supply of energy substrates for neurons [6], and neuronal pathfinding during development and regeneration [7]. Astrocytic cells seem to have key roles in many central nervous system disorders, ranging from neuropathic pain and epilepsy to neurodegenerative diseases such as Alzheimers, schizophrenia and depression [8]. Computational models of neuron-glia interactions is a prerequisite for understanding the dysfunctional situations, and for assessing glial cells as a potential therapeutic target [9]. To give a few examples, such models have been used to simulate glial regulation of extracellular K+-concentration [10]–[13], and the relation between extracellular K+-dynamics and epileptic seizures [14]–[16] and spreading depression [17], [18].
Regulation of the extracellular K+-concentration is considered one of the key cellular functions of astrocytes [2]. During normal conditions, the extracellular K+-concentration () is typically maintained close to the baseline level (). However, when neurons fire action potentials, they expel K+ into the extracellular space. During periods of intense neural activity, the local extracellular K+-concentration may increase by several millimolars, and may interfere with neural activity [10], [19], [20]. Concentrations between 8 and 12 mM are often considered a limit to pathological conditions [3], [12], [21].
Orkand (1966) [22] discovered that astrocytes can funnel out excess K+ from high concentration regions by a process coined spatial buffering [12], [21], [22]. According to this concept, K+ is taken up by the glial cell from high-concentration sites, evoking a local depolarization of the glial membrane. K+ is then transported longitudinally inside the glial cell (and possibly through several glial cells connected by gap junctions into a glial syncytium [10], [23]), and eventually expelled into the ECS at more distal cites where is lower. However, it has also been argued that astrocytes may reduce by local uptake and temporal storage, not necessarily including transport over distances [19], [24]. Furthermore, diffusion through the ECS is also involved in transporting excess K+ out from high concentration regions. The relative importance of these different clearance mechanisms are under debate [25].
Electrical neural signalling is typically modeled using the cable equation, where dendrites and axons are represented as one-dimensional, possibly branching, electrical cables, and the transmembrane potential is the key dynamical variable [26], [27]. With the possible exception of the signalling molecule Ca2+ (see e.g., [28], [29]), ion concentrations are typically assumed to be constant. The effect of ionic diffusion (due to concentration gradients) on the net electrical currents is neglected in standard cable theory, and resistivities (which in reality depend on ion concentrations) are assumed to be constant. These are often good approximations, as concentrations of the main charge carriers (K+, Na+ and Cl−) in the extracellular- (ECS) or intracellular space (ICS) typically vary little at the short time-scale relevant for electrical neural activity ().
Glial function typically involves processes that take place at a longer time-scale (), at which significant variations in ionic concentrations may occur. For example, the process of spatial K+-buffering involves local uptake, a local depolarization of the astrocytic membrane, and longitudinal electrodiffusive transports through the intracellular- (ICS) and extracellular space (ECS) propelled both by voltage- and concentration gradients [30]. A mechanistic understanding of glial function thus requires a modelling scheme that in a consistent way can capture the intricate interplay between ion concentration dynamics and the dynamics of . Physically, is determined by the total electrical charge on the inside (or outside) of the membrane, which in turn is uniquely determined by the concentrations () of all ionic species that are present there [31]. In some heart cell models, ion concentrations have been reported to drift to unrealistic values in long-term simulations, while maintain realistic values [32]–[34]. Whether the relationship between and is consistent, is a general concern with models that explicitly depend on both. If applied to general problems, and in particular in long-term simulations, models that do not ensure an internally consistent relationship may give erroneous predictions.
Gardner-Medwin (1983) [10] proposed a pioneering computational model of the spatial buffering process, later re-analyzed by Chen and Nicholson (2000) [12]. In this model, spatial buffering was considered as an essentially one-dimensional transport process. The complex composition of the tissue (Fig. 1A) could then be simplified to a two-domain model as that illustrated in Fig. 1B [10], [12]. There, the ICS of all cells participating in the transport process (i.e. the astrocytes) have been represented as an equivalent cable (I-domain) which is coated by ECS (E-domain). The I-E system could be pictured phenomenologically as an representative single astrocyte, coated with the average proportion of available ECS per astrocyte. This geometrical simplification was motivated for one-dimensional transport phenomena through the glial syncytium [10], [12], but could in principle apply to any transport phenomena that justifies a geometrical simplification as that in Fig. 1. A limitation with these modelling studies [10], [12], and related modelling studies by Newman and coworkers [11], [21], is that was derived from standard cable theory, which neglects effect from diffusive currents on . The concern regarding a consistent relationship between and the ionic concentrations thus also applies to these models.
Qian and Sejnowski (1989) have previously developed a consistent, electrodiffusive scheme for modelling the dynamics on and ion concentrations [31]. Like the standard cable model, the electrodiffusive model assumes that transport phenomena are essentially one-dimensional. Unlike the standard cable model, the electrodiffusive model derived from the ion concentration dynamics, accounting for all ionic movements (membrane fluxes, longitudinal diffusion, and longitudinal electrical migration), as well as for the concentration-dependent variation of the intracellular resistivities. An important limitation with this previous electrodiffusive model is that it only includes intracellular dynamics, whereas the ECS was assumed to be isopotential and with constant ion concentrations [31]. This was a useful simplification for simulating a small intracellular compartment, such as a dendritic spine [31], but is not generally applicable to macroscopic transport mechanisms. In particular, it can not be applied for modelling the spatial buffering process, where ion concentration dynamics in the ECS plays a paramount role. In reality, the ECS comprises about 20% of the total neural tissue volume, while the remaining 80% is the ICS of various cells [12]. When a large number of cells participate in simultaneous ion exchange with the ECS, the impact on the ion concentrations in the ICS and ECS may be of the same order of magnitude.
The aim of this work is twofold: First, we generalize the electrodiffusive formalim [31] to a explicitly include the ECS. The result is a general mathematical framework for consistently modelling the dynamics of the membrane potential (), the intra- () and extracellular () ion concentrations for a set () of ionic species. We believe that this framework will be of general value for the field of neuroscience, as it can be applied to any system that justifies a geometrical description as that in Fig. 1B. Next, we apply the electrodiffusive formalism in a spatially explicit model of astrocytes exchanging ions with the ECS. We run simulations to investigate the efficiency of the spatial K+-buffering process, relative to that of local uptake/storage by astrocytes, and that of diffusion in the ECS alone. Unlike the previous models [10]–[12], [21], our astrocyte model is based on the prevailing view that Na+/K+/ATPase-pump is the main uptake mechanism for K+ [3]. Furthermore, as our model was based on a physically consistent electrodiffusive formalism, we arrive at a full mechanistic description of the buffering process, which quantitatively describes the intricate interplay between and the dynamics of ion concentrations.
This article is organized in the following way: The Model section contains two main parts. In the first part, we present the electrodiffusive formalism for computing the ion concentration dynamics in a system described by the geometry depicted in Fig. 1B. We consider this theoretical framework a key contribution of this work. However, the key concepts introduced in this part are summarized in Table 1, and with this in hand, the reader who is mainly interested the biological process of spatial K+-buffering by astrocytes may therefore skip to second part of the Model-section. There, the model for astrocytes exchanging ions with the ECS is presented. The Results section is devoted to simulations on the astrocyte model, and provides an improved biophysical insight in the electrodiffusive mechanisms utilized by astrocytes to spatially buffer K+. By comparing different versions of the model, we also assessed the importance of spatial buffering, relative to that of other clearance mechanisms such as local uptake/storage by astrocytes and diffusion through the ECS alone. Finally, in the Discussion section we address how our mathematical framework relates to previous electrodiffusive modeling frameworks. We also summarize the new insights that our simulations have given in the process of spatial K+-buffering by astrocytes.
In Fig. 1B, particles in I or E may move along the x-axis or across the membrane. In a segment of I, centered at x, and with volume , the particle concentration dynamics of an ion species is determined by:(1)where the transmembrane- (), the intracellular- () and the extracellular () flux densities of particle species , have units mol/(m2s). The first term on the left represents the ionic flux that enter this segment through the piece of the membrane with area . The negative sign follows from (by convention) being defined as positive in the direction from I to E. The second and third terms represent the ionic fluxes that enter(+)/leave(−) the section through the left/right boundaries, with cross section areas . If the net flux into the segment is nonzero, the ion concentration will build up over time, according to the right hand side of Eq. 1.
We divide Eq. 1 by , and take the limit , to obtain the continuity equation on differential form:(2)(3)We have also written up the continuity equation for the extracellular domain.
The axial flux densities are described by the generalized Nernst-Planck equation:(4)where is the valence of ion species , and the index n represents I or E. The first term on the right in Eq. 4 is the diffusive flux density (), driven by the concentration gradients, and the last term is the field flux density (), i.e., the flux density due to ionic migration in the electrical field. The effective diffusion constant is composed of the diffusion constant in dilute solutions and the tortuosity factor , which summarizes the hindrance imposed by the cellular structures [12], [35]. We use , where is the gas constant, the absolute temperature, and is Faraday's constant.
The formalism is general to the form of , which may include contribution from multiple membrane mechanisms, such as ion pumps, co-transporters and ion channels. It is sufficient to require that is known at any point in time given the voltage across the membrane, the ionic concentrations on either side of the membrane, and possibly some additional local information () reflecting the local state of the membrane:(5)
As boundary conditions, we shall apply the sealed-end condition, i.e., we assume that no fluxes enter or leave through the ends ( and ) of I or E:(6)
Equations 2–3, together with with Eqs. 4, 5 and 6, specify the system we want to solve. Before we derive the electrodiffusive formalism for this problem, we recall how the standard cable equation can be derived from the principles of particle conservation.
We here present a model of astrocytes exchanging ions with the ECS, as sketched in Fig. 3, and defined in further detail below. The astrocyte model was developed for macroscopic transport processes, involving a collection of astrocytes (possibly connected via gap junction into a syncytium) in a piece of tissue. For this problem, we used the geometrical simplification motivated in Fig. 1, i.e., we applied the geometry in Fig. 1B. We took the intracellular domain to represents a phenomenological “average” astrocyte (the cable, ), surrounded by a sheet of ECS (the coating, ). We used the empirical estimates that a fraction of neural tissue volume is ECS, while astrocytes take up a fraction of about of the total tissue volume [12]. The intracellular domain was therefore twice as voluminous as the intracellular.
Table 1 contains a list of definitions that are necessary for the reader to follow the remainder of the paper. The dynamics in the system was due to fluxes of ions crossing the membrane , or axial fluxes in the ECS or ICS due to diffusion () or migration in the electrical field (). We assumed that only the three main charge carriers (K+, Na+ and Cl−) contributed to electrodiffusive transport. For the diffusion constants (), we used values valid for electrodiffusion in diluted media [36], modified with the tortuosities () estimated in [12]. The same values have also been used in earlier, related studies [31], [37]. All relevant model parameters are listed in Table 2. The system input, and the astrocytic membrane mechanisms are defined in further details below.
An important contribution of this work was the general electrodiffusive formalism presented in the Model section. This formalism represents a framework for modeling the dynamics of the membrane potential (), the intra- () and extracellular () ion concentrations. The formalism is general to the choice of membrane mechanisms, and could be applied to model any transport process that justifies the geometrical simplification depicted in Fig. 1.
Here, we have applied the formalism to simulate spatial K+-buffering by astrocytes, using the specific implication to the atrocyte/ECS-model, also presented in the Model section. Our main objective has been to investigate the transport routes of K+ ions, from entering the system in the ECS of the input zone, to leaving the system at some point along the -axis. We remind the reader that a useful list of symbols and definitions can be found in Table 1.
We investigated the ion concentration dynamics in the astrocyte model (Fig. 3) in full detail. Fig. 4A–D shows the dynamics of selected variables in the input zone (at ). Fig. 4E–H shows how the same variables depend on at a time when the system was in SS. We explain this further below.
The input was applied from to in the input zone (). This is illustrated in Fig. 4A (solid line), which shows the flux density of K+ () entering the system in the input zone. We recall that the input was a cation exchange, so that there was an equal flux density of Na+ leaving the system (). For simplicity, was not included in the figure, but we keep in mind that whenever K+ entered/left the system, an equal amount of Na+ left/entered. The cation-exchange input thus caused an increase in and a decrease in in the input zone. This can be seen in Fig. 4B. The notation represents the deviations from baseline concentration (cf. Table 2).
As increased, the output from the system (being proportional to ) increased. Also this is illustrated in Fig. 4A (dashed line), which shows the flux density of K+ () leaving the system from a point in the input zone. We recall that also the output was a cation exchange, so that the efflux of K+ implied a corresponding influx Na+.
The input was given in the input zone, while the output occurred over the full axis, depending on the local value of . During a transient period, the constant input changed the ion concentrations in the system. The system reached steady state (SS) when became sufficiently high. Then, the total amount of K+ entering the system per second, and the total amount of K+ leaving the system per second, coincided (with the same being true for Na+). This is illustrated in Fig. 4E, which shows how the and are distributed over the -axis at a time , when the system was in SS. The areas under the curves for and were then equal. In the input zone, however, the output rate was about 1/3 of the input rate (Fig. 4A). This means that about 2/3 of the K+ that entered the system was transported in the positive -direction, and left the system from the decay zone. (We recall from Fig. 3 that the decay zone is defined as any part of the -axis outside the input-zone).
Fig. 4B–D shows how the local (at ) intracellular ion concentrations, the extracellular ion concentrations and changed from the input had been turned on until the system reached SS. For the present example it took 49 s from the constant input had been turned on until the slowest variable () reached 99% of its SS value. The other variables approached SS faster than this (e.g., 12 s for and 19 s for ). During SS, was about 7.7 mM, corresponding to a concentration (as the baseline concentration was ). Although the input was applied to the ECS of the input zone, the local intracellular K+-concentration had increased even more (). This reflects the astrocyte's propensity for local K+-uptake. The changes in ionic concentrations in the ECS and ICS coincided with a local depolarization of the astrocytic membrane, from the resting potential () to about , reflecting concentration dependent changes in the reversal potentials of the involved ionic species.
From here on, we focus on the SS-situation, i.e., on the activity of astrocytes during periods of on-going intense neural activity. For all system variables, the devition from the baseline (resting) conditions were generally biggest at the point , i.e., in the part of the input zone which is furthest away from the decay zone (Fig. 4E–H). The average value of , taken over the input zone () was approximately 10 mM (about 6.9 mM above the resting concentration). During the model calibration, the constant input rate () was tuned to obtain this value, which is on the threshold between functional and pathological conditions [3], [12], [21]. During SS, the gradients in ionic concentrations (Fig. 4F–G) and (Fig. 4H) were quite pronounced. We thus expect that both diffusive and electrical forces contribute to transporting ions through the system (from entering to leaving). This is explored further in the following section.
In addition to spatial buffering, K+ may also be buffered by diffusion through the ECS alone, or by local (space independent) storage by the glial cell, to be later released in the same region of the ECS [19], [24]. To investigate the relative importance these clearance mechanisms, we compared the 6 six model versions depicted in Fig. 8A, including one group of three spatially extended models (solid lines), and one group of three point models (dashed lines). Both groups included one model version with an active astrocyte, one model version where the astrocyte had been replaced by a corresponding increase in the ECS volume (the total ECS volume fraction increased to ), and one version where the original ECS volume fraction () was kept when the astrocyte was removed. The spatially extended model including the astrocyte, is the one we studied in the previous sections. The other models were reduced versions of this.
All model versions were exposed to the input signal described by Eq. 30, causing an increase in . The input was applied in the time window , which was sufficient for to reach its SS-value in all models. Fig. 8B shows the dynamics of the K+-concentration in the ECS at the point where the concentration was the highest (). In the spatially extended models, this occurred at , i.e. in the part of the input zone furthest away from the decay zone.
During SS, the net K+ efflux and influx from/to the system coincided. For the point models, having no spatial resolution, there was no distinction between the input zone and decay zone, as the input and output were injected to/subtracted from the same single compartment. The net output rate thus depended on in this single compartment. Therefore, all point models approached the same SS value (). For the spatially extended models was lower, as parts of the K+ could leave the system also outside the input zone. For these models, depended on how efficient they were in longitudinally transporting K+ out from input zone before (revisit Fig. 4 for more details).
To gain insight in the importance of local K+-uptake by astrocytes, relative to diffusion in the ECS, we compared the performance of the point model including the astrocyte (black, dashed line in Fig. 8B) to that of the spatially extended model including only the ECS (blue, full line). During the first few seconds after the stimulus had been turned on, the point model with the astrocyte (representing local uptake) was most efficient in terms of limiting . However, local uptake was limited by the storing capacitance of the astrocyte. After seconds with constant K+-influx to the system, the spatially extended model (representing diffusion through the ECS) performed better, as it could redistribute K+ over a larger spatial region. The astrocyte's ability to locally store excess K+ has been emphasized in previous investigations [19], [24]. Our simulations predicted that the local storage mechanism is mainly important in relatively short time spans after potassium release (a few seconds). A similar conclusion was also drawn from previous modelling studies [10], [12]. We here add an additional point to this discussion: The performance of the point model with extended ECS (dashed red lines) more or less coincided with that of the point model including the astrocyte (dashed black lines). In terms of local storage, the astrocyte (with its membrane being highly permeable to K+), essentially just acts to expand the local volume that the incoming flux of K+ enters into.
It has been argued that because K+-transport is aided by transmembrane processes as well as internal processes in the glial cell, K+ can be cleared more effectively by glia than would be possible by a much enlarged extracellular space [40]. To investigate this claim, we compared the three spatially extended models (solid lines). We found that the model including the astrocyte (black, solid line) was more successful in limiting than any of the other model versions. It was significantly more successful than diffusion in the ECS alone, even in the (rather hypothetical) system where the extracellular volume had been increased by a factor 3.
In conclusion: In terms of local storage, the astrocyte was not significantly more efficient than an increased enlarged extracellular space. In terms of spatial buffering, however, it was.
We presented a one-dimensional, electrodiffusive framework for modeling the dynamics of the membrane potential () and the ion concentrations of all included ion species in an intra- and extracellular domain (Fig. 2). The framework could have a broad range of applications within the field of computational neuroscience. In the current work, it was applied to simulate the role of astrocytes in K+-removal from high concentration regions.
The astrocyte/ECS-model provided a mechanistic understanding of how astrocytes may remove K+ from high-concentration regions. In summary, the model astrocyte responded to a local extracellular increase in by a local depolarization of the membrane. At the same time, this depolarization (i) increased astrocytic K+ uptake in the input zone, (ii) increased astrocytic K+-release outside the input zone, (iii) decreased axial K+ transport in the ECS, and (iv) increased axial K+ transport inside the astrocyte. Furthermore, by comparing different versions of the model, we predicted that (v) local storage of K+ by astrocytes may play an important role at a short time scale, while (vi) at a longer time scale, the ability to distribute K+ spatially will be crucial for maintaining a low extracellular K+-concentration. In this regard, we found (viii) that the astrocyte was more efficient spatial buffering mechanism than diffusion in an enlarged extracellular space.
The findings (i–ii) were due to well documented astrocytic membrane mechanisms that we implemented in the model. Uptake from high concentration regions was mediated by the K+/Na+-pump, while release into low-concentration regions of the ECS was mediated by the Kir-channel. This supports the dominant view of glial K+-buffering [2], [3], [25], [39], [41].
The findings (iii–iv) were due to an interplay between electrical and diffusive forces. When locally depolarized (in the region with high extracellular ), longitudinal voltage gradients arose, and ions in the ICS and ECS were exposed to an electrical force. In the ICS, the electrical force and diffusive force both drove K+-ions in the same direction (out from the high concentration regions). In the ECS, the electrical force acted in the opposite direction from the diffusive force, and reduced the net longitudinal transport through the ECS. Hence, in addition to being an efficient transport route for K+ out from high-concentration regions, the astrocyte actively reduces the extracellular K+-transport. This represents a (to our knowledge) novel mechanism that astrocytes may utilize to shield the extracellular space from excess K+. All these effects (i–iv) taken together turned the astrocyte into an efficient sluice for removing K+ from the input zone.
The findings (v–viii) shed light on the relative efficiency of spatial buffering and other K+-clearance mechanisms, such as local storage by astrocytes, or diffusion in the ECS alone. An interesting prediction was that, in terms of local storage, the astrocyte did not have a stronger effect on than an enlarged extracellular space would. In terms of longitudinal transports, however, the astrocyte performed much better (by spatial buffering) than diffusion in an enlarged extracellular space (Fig. 8). We do, however, wish to comment that these mechanisms are not mutually exclusive. In fact, an (initial) local accumulation of intracellular K+ is required for the astrocyte to initiate the spatial buffering process. It is this local accumulation that evokes the intracellular voltage- and concentration gradients that the astrocyte utilizes for intracellular K+-transport.
It is likely that the mechanisms responsible for spatial buffering vary between brain regions and between different species of glial cells [30]. Previous literature has suggested several mechanisms for spatial buffering apart from the ones that were included in our model. K+-uptake by Na+K/K+/Cl−-cotransporters and K+/Cl−-cotransporters are two candidate mechanisms that likely could affect the simulated results [38]. Furthermore, regions in the endfoot processes of astrocytes have been shown to have an extremely high K+-conductance compared to the membrane in general [42]. Such high concentration regions could improve the buffering process by transferring (siphoning) excess K+ into the vitreos humor or vasculature [43]. The buffering process may also be affected by water influx and swelling experienced by the astrocyte during the uptake process [13], [30], [38].
Rather than increasing the biological complexity, by e.g., including multiple candidate buffering mechanisms, we have in this study strived towards elucidating the fundamental physical processes involved in spatial K+-buffering. As our simulations demonstrated, K+-buffering is a highly complex process. It involves an intricate and sensitive interplay between and ionic concentrations, and between electrical and diffusive transports. We therefore highlight the importance of applying an electrodiffusive, physically consistent, modelling scheme which ensures a complete book-keeping of ion concentration dynamics and its effects on . In previous models of spatial buffering, was derived from standard cable theory [10]–[12], [21], where diffusive currents are assumed to have a negligible impact on , and where the resistivity is assumed to be constant (i.e., not dependent on ion concentration variations). During our simulations, intra- and extracellular resistivities changed by as much as 10% and 20%, respectively, and the diffusive current was about 25–30% of the field current in the ECS. The assumptions underlying standard cable theory are therefore poorly justified if applied to the spatial K+-buffering process.
The astrocyte/ECS-model was represented phenomenologically as a single astrocyte coated with the average proportion of available ECS per astrocyte (see Figs. 1 and 3). This geometrical representation is justified for macroscopic transport processes, when a large number of astrocytes perform the same function simultaneously [12]. For the current study, this was a reasonable assumption, as the input was a change in the ion-concentrations in the ECS, shared by all present astrocytes.
If we instead wanted to study a cell specific signal, such as e.g., the response of a single astrocyte to a transmembrane current injection, the geometrical representation in Fig. 1B would be less appropriate. Firstly, the notion of the ECS as a relatively thin coating following a single cell is only motivated at the macroscopic “average transport”-level. Secondly, if only a single cell was involved in a particular process, we would expect that . That is, a single active cell would have a significantly larger proportion of the ECS to its own disposal, compared to the macroscopic case, where all cells in a piece of tissue are active, and share the limited amount of available ECS. In single-cell models it is common to assume that conditions in E are constant, so that only I is modeled explicitly. In this limit, the electrodiffusive formalism reduces to the one-domain model presented previously by Qian and Sejnowski [31].
The framework presented here is essentially an expansion of the one-domain model by Qian and Sejnowski [31] to a two-domain model that includes both the ECS and ICS. Like the one-domain model, the framework ensures (i) a consistent relationship between and ionic concentrations . Unlike the one-domain model, the framework ensures (ii) global particle/charge conservation, and (iii) that the charges on either side of a piece of membrane must be equal in magnitude and opposite in sign (). The latter constraint is implicit when the the membrane is assumed to be a parallel plate capacitor, an assumption made in most models of excitable cells (see e.g., [26], [27], [31]). It is also related to the topic of electroneutrality.
Electroneutrality in electrodiffusive models of biological tissue has been the topic of many discussions [44]–[46]. It is relevant for how the electrical potential (v), occurring in the Nernst-Planck equation, is derived. Generally (at sufficiently course spatial resolutions so that the charge density can be assumed to be continuous), v obeys Poisson's equation:(43)where is the dielectric constant, and is the total charge density.
In biological tissue, the charge relaxation time is very small in any region except in the thin Debye layer () surrounding a bio-membrane. Any nonzero net charge density in the bulk solution will decay very rapidly () to zero [36]. Several models have simulated electrodiffusion by solving the Nernst-Planck equations in one or more dimensions, with Poisson's equation for (see e.g., [45], [47]–[50]). The advantage with such a procedure is that the Poisson-Nernst-Planck (PNP) equations can be implemented in a general way in three-dimensional space. The challenge is then to specify the appropriate boundary conditions for solving Eq. 43 in the vicinity of membranes. Generally, PNP-solvers apply a fine spatial resolution near the membrane, and simulation time steps smaller than the charge-relaxation time [48]. For these reasons, they tend to be extremely computationally demanding [51].
The formalism presented in this work belongs to a class of of one-dimensional models, including the cable equation and several electrodiffusive models [10], [12], [13], [17], [31], [52], [53], which bypasses the computationally heavy PNP-scheme. The physical interpretation of these models is as follows: Any net charge in a volume is implicitly assumed to be located in the thin Debye-layer surrounding the capacitive membrane, and is identical to the charge that determines . The remainder of the space (i.e., the bulk) will therefore be electroneural (). Note that any finite volume, enclosing a piece of membrane, will also be electroneutral. This follows from the charge symmetry condition (Eq. 18), constraining the charge on either side of the membrane to be equal in magnitude and opposite in sign. The charge symmetry condition and the electroneutrality condition are in this way closely related. In these electroneutral models, charge relaxation is implicit. This is a plausible assumption at time scales relevant for most biophysical processes. Accordingly, simulations may be run with time-steps ranging from 1 ms to 1 s, depending on the time course of the included membrane mechanisms.
To our knowledge, the formalism summarized in Fig. 2 is the first biodiffusive model where the intra- and extracellular voltage gradients have been derived from the charge symmetry condition. Eqs. 25 and 26 can be interpreted as summarizing all local and global electrical forces driving the system towards electroneutrality. A natural future ambition would be to generalize the electrodiffusive formalism to 2 or 3 spatial dimensions, so it can address the same 3-dimensional transport problems as PNP-solvers. The challenge will be to formulate the system as a grid of coupled constraints (electroneutrality in the bulk and Eq. 12 for across the membrane) for which the Nernst-Planck equations can be solved with time steps much longer than those involved in the charge relaxation process.
|
10.1371/journal.ppat.1000669 | A Novel System of Cytoskeletal Elements in the Human Pathogen Helicobacter pylori | Pathogenicity of the human pathogen Helicobacter pylori relies upon its capacity to adapt to a hostile environment and to escape from the host response. Therefore, cell shape, motility, and pH homeostasis of these bacteria are specifically adapted to the gastric mucus. We have found that the helical shape of H. pylori depends on coiled coil rich proteins (Ccrp), which form extended filamentous structures in vitro and in vivo, and are differentially required for the maintenance of cell morphology. We have developed an in vivo localization system for this pathogen. Consistent with a cytoskeleton-like structure, Ccrp proteins localized in a regular punctuate and static pattern within H. pylori cells. Ccrp genes show a high degree of sequence variation, which could be the reason for the morphological diversity between H. pylori strains. In contrast to other bacteria, the actin-like MreB protein is dispensable for viability in H. pylori, and does not affect cell shape, but cell length and chromosome segregation. In addition, mreB mutant cells displayed significantly reduced urease activity, and thus compromise a major pathogenicity factor of H. pylori. Our findings reveal that Ccrp proteins, but not MreB, affect cell morphology, while both cytoskeletal components affect the development of pathogenicity factors and/or cell cycle progression.
| The human pathogen Helicobacter pylori lives in the hostile environment of the human stomach. H. pylori possesses a spiral shape and high motility that enable the bacterium to swim through the stomach lumen and to come into close contact with epithelial cells. High urease activity in the bacterium counterbalances the low pH within the stomach, in order to persist within the viscous mucus layer. In this work, we analysed the molecular basis of the spiral structure of H. pylori. We demonstrate that the helical cell shape depends on so called coiled coil rich proteins (Ccrp), which form extended filamentous structures in vitro and in vivo, and are differentially required for the maintenance of proper cell morphology. In most bacteria analysed so far, the actin-like protein MreB affects cell morphology. Contrarily, H. pylori MreB is not involved in the maintenance of cell shape, but affects the progression of the cell cycle. Mutant cells were highly elongated, characteristic for a delay in cell division, and contained non-segregated chromosomes. The persistence of H. pylori in the hostile environment of the human stomach depends on the activity of urease. Interestingly, mreB mutant cells displayed significantly reduced urease activity, revealing a novel connection between the cytoskeletal element and an enzyme, and thus with pathogenicity. These experiments show that H. pylori has a novel type of system setting up helical cell shape, which has not yet been described for any bacterium. Our work will allow studying H. pylori cell cycle and pathogenicity at a new visual level.
| Helicobacter pylori is a Gram negative, highly motile, microaerophilic, spiral-shaped organism, which colonizes the stomachs of at least half of the world's population [1]. Infection of humans results in persistent gastritis, which can develop into peptic ulcer disease and adenocarcinoma [2],[3]. Motility is a key factor in the adaptation of infection, allowing for the penetration of the mucus and enabling the bacteria to colonize and to persist in the gastric lumen [4]. Both spiral shape and flagella contribute to the motility of this human pathogen. Whereas flagella of H. pylori have been studied intensively, our knowledge of the maintenance and establishment of spiral structure in H. pylori and in fact for any bacterium is marginal. In addition, nothing is known about any cytoskeletal protein in this pathogen.
Maintenance of cell morphology is highly important or essential for functioning and survival of most eukaryotic and prokaryotic cells. For many eukaryotic cells, it is also vital to be able to change the shape of the cell, and/or to be able to move via flexible extension/retraction of the cell membrane. Cytoskeletal elements actin and intermediate filaments are key elements of the eukaryotic cytoskeleton that controls cell morphology and cell rigidity. Due to its rapid polymerisation/depolymerization properties, actin is the driving force for motility involving membrane rearrangements, and is also involved in trafficking of vesicles and in cell division [5]. IF proteins, on the other hand, are characterized by extended coiled coil regions. The proteins are believed to be highly elongated and assemble into sheet structures based on extensive interactions between coiled coils [6]. IF like proteins provide mechanical strength to e.g. skin or blood vessel cells, and are involved in positioning of cellular organelles [7].
For most rod shaped bacteria analysed so far, the loss of genes affecting cell shape is lethal. Escherichia coli or Bacillus subtilis cells are unable to grow as round cells, into which they turn when gene products of rodA, mreB, or mreC are depleted. While RodA and MreC are membrane proteins, whose function is still unclear, MreB is an actin like protein that forms filaments in vitro, dependent on ATP [8]. In vivo, MreB forms filamentous helical structures underneath the cell membrane [9],[10]. In B. subtilis and in Caulobacter crescentus, these filaments are highly dynamic, and appear to move along the membrane with dynamics similar to those of eukaryotic actin [11],[12]. Movement of filaments is most likely based on ratchet-like extension of filaments at one end, and depolymerization (and thus shrinkage) at the other end. E. coli MreB and an MreB ortholog, Mbl, in B. subtilis, have been shown to interact with MreC [13],[14], which in turn appears to interact with enzymes that synthesize the extension of the murein sacculus [15]. Because the incorporation of new cell wall material occurs in a helical pattern [16], it has been proposed that the helical organization of MreB filaments in the cytosol may direct the helical localization of cell wall synthetic proteins within the periplasm/outside the cell. A disputed question is the effect MreB exerts on the segregation of duplicated chromosomes. Interfering with MreB levels or polymerization activity has been shown to strongly impair chromosome segregation in several organisms [10],[17], but arguments against a direct involvement of MreB in segregation have also been put forward [18].
The question of how bacterial cells can obtain a curved shape has only been investigated in the vibrio-shaped bacterium C. crescentus. CreS encodes for a coiled coil protein, crescentin, which has high similarity to IF proteins. Crescentin forms filamentous structures in vitro without the addition of any nucleotides. Deletion of creS leads to the generation of straight cells, and thus to loss of cell curvature, while the culture doubling time or any other obvious physiological aspect of the cell is not affected [19]. Crescentin localizes as a defined ribbon structure along the short side of the cells, suggesting that it forms a filamentous structure in vivo [19]. Recent evidence suggests that crescentin exerts its effect on cell curvature through mechanical control of cell growth [20].
In this work, we set out to analyse cytoskeletal elements in the human pathogen Helicobacter pylori. We have systematically inactivated genes encoding coiled coil-rich proteins, and for mreB. Surprisingly, deletion of mreB is not lethal, but affects a variety of cellular parameters, such as chromosome segregation, but not cell shape. Deletions of Ccrp (coiled coil rich proteins) genes have different effects on cell shape in different strains, from loss of helical shape to complete loss of a regular morphology. We have also established a system for the visualization of proteins in H. pylori, and show that Ccrp proteins have a specific pattern of localization, consistent with their function in cell shape maintenance.
To gain insight into the question of how H. pylori gains its helical cell shape, we searched for elements similar to known cytoskeletal or cell morphological elements. Chromosomes of all H. pylori strains analysed contain a gene with high similarity to mreB, followed by a mreC gene. Like in E. coli and B. subtilis, the MreC gene product is predicted to contain a single membrane span, and coiled coil regions. No mreD gene could be found in the genomes, but a rodA like gene, and several pbp genes (not shown). Interestingly, all strains contain two genes that have already been suggested to encode for IF-like proteins (HP0059 and HP1143 in strain 26695) [19], which are predicted to contain several extended heptad repeat regions, but also a so-called stutter, where coiled coil 4 is clearly discontinued for few amino acids [21]. However, HP0059 is almost entirely composed of heptad repeat regions, and lacks the characteristic N- and C-terminal domains of IF proteins, which are predicted to be globular. According to their predicted secondary structure we suggest to term this class of proteins as “coiled coil rich proteins” (Ccrp). We designate the H. pylori HP0059 or HP1143 gene products as Ccrp59 or Ccrp1143, respectively.
Because of the genetic (and morphological) variability of Helicobacter pylori, we generated constructs in several different strains, to obtain information on the general validity of gene deletions or localization patterns of fusion proteins. We focussed our work on the reference strain 26695 (moderately motile), on KE88-3887, a hyper-motile variant of strain 26695, and on the clinical isolates G27 and 1061, all of which are relatively well amenable for genetic analysis.
It should be noted that H. pylori strains have somewhat different morphologies. Strain 26695 is highly helical (Fig. 1A) (similar to KE88-3887, Fig. 2D), with an average of length of 3.0 +/− 0.5 µm (n = 72) and can be up to 4.0 µm in length, while cells of strain 1061 are much shorter with an average of length of 2.3 +/− 0.5 µm (n = 100) and their helical shape is less pronounced (compare Fig. 2A). Other strains of H. pylori also have varying degrees of cell curvature.
In order to study possible functions of genes predicted to encode for cell shape determinants, we inactivated genes HP0059 and HP1143 in H. pylori strains 26695, KE88-3887, G27 and 1061. To ensure expression of the downstream genes, all genes were disrupted by insertion of a cat gene driven by its own promoter but lacking a terminator. Growth analysis of all mutants revealed that inactivation of none of these genes showed any effect on the growth rate of H. pylori.
Interestingly, the inactivation of gene HP0059 resulted in the formation of 100% straight rods in strains 26695, KE88-3887 and G27 (Fig. 1B, Fig. 2E, and data not shown for G27), or 85% straight cells in strain 1061 (Fig. 2B), revealing a complete loss of the spiral shape in the absence of the HP0059 gene product. To rule out a possible downstream effect on gene HP0060, HP0060 was disrupted by introducing a pcat cassette. No change in growth or cell curvature could be detected compared with wild type cells (data not shown) showing that the loss of helical cell shape is due to the inactivation of gene HP0059. The same observation was made with strains 1061, KE88-3887 and G27 (data not shown). Inactivation of HP1143 had a mild effect on cell shape. Whereas 10 to 15% of 26695 wild type cells were straight (Fig. 1A), about 60% of HP1143 mutant cells were straight (Fig. 1C, or 52% for strain KE88-3887, Fig. 2F). These experiments show that the loss of genes HP0059 or of HP1143 affects cell curvature to different extents.
The deletion of HP1143 had an even more dramatic effect on cell shape in 1061 cells, about 70% of the cells were round, oval or irregularly shaped, while the remaining cells were straight or bulgy rod shaped (Fig. 2C). Single non-aggregated cells were basically undetectable. Thus, deletions of genes HP0059 and HP1143 have different effects on cell shape in different strain backgrounds. Absence of HP0059 generates loss of cell curvature in all strains tested, while lack of HP1143 results in a complete loss of cell shape in strain 1061.
H. pylori undergoes a transition from helical cells to coccoid cells upon prolonged starvation. We analysed whether ccrp mutant cells influence this morphological transition, whose mechanism is still poorly understood. Like wild type cells, all HP0059 or HP1143 mutant cells were coccoid 7 days after inocculation (i.e. 5 days into stationary phase), showing that the inactivation of Ccrp encoding genes does not influence the helical to coccoid transition.
To investigate if HP0059 and HP1143 are genetically linked, we generated a strain from the parent 26695, in which both genes are deleted. Interestingly, cells of the double mutant strain displayed a variety of cell shapes: while 65% of the cells were straight, 30% had an irregular curved shape, and 5% had a highly bent shape, such that the cell ends came together (Fig. 1D), which is never observed for wild type cells. For strain KE88-3887, the double deletion resulted in even more highly bent cells (suppl. Fig. S1). These findings show that the loss of both Ccrp encoding genes leads to a complete loss of regular cell shape in strains 26695 and KE88-3887, and exacerbates the phenotype of the single gene deletions. The double deletion of HP0059 and HP1143 in strain 1061 was similar to that of the HP1143 single gene deletion, in that most (>80%) double mutant cells were round, oval or irregularly shaped, while the remaining cells were straight or bulgy rod shaped (data not shown).
We wished to obtain insight into the biochemical properties of Ccrp proteins. Towards this end, we purified an N-terminally strep-tagged version of Ccrp59 to more than 95% purity (Fig. 3A). Ccrp59 could be purified in very low quantities as a soluble protein upon mild and short time (2 h) induction of the protein in E. coli cells, but appeared in inclusion bodies after prolonged induction. On SDS-PAGE, Ccrp59 migrated as monomer but also as a band that corresponded to a dimer (Fig. 3A), which is apparent in the Western blot in Fig. 3B. When subjected to centrifugation, a major proportion of Ccrp59 appeared in the pellet fraction (Fig. 3C), suggesting that it forms large assemblies. Over time (i.e. days to weeks), the amount of Ccrp59 in the pellet fraction increased (data not shown). When purified fractions were subjected to electron microscopy, it became clear that Ccrp59 forms extended filamentous structures in vitro, in the absence of any added cofactor (Fig. 4A), similar to IF-type proteins [21], and dissimilar to most other filament forming proteins. However, the fact that Ccrp59 can be purified as soluble protein clearly distinguishes it from IF proteins, which need to be refolded to be obtained in solution (compare Table 1 for dissimilarities and similarities). Filaments were generally straight, and present in bundles or sheets (Fig. 4A, grey triangles). The smallest observable filaments were 10 nm wide and about 50 nm long (Fig. 4A, black triangles); these may represent single Ccrp filaments. A closer investigation of the bundles showed that they consist of individual filaments (also 10 nm wide) that were positioned in parallel, and appeared to have staggered ends (hatched triangle). Bundles could frequently be observed to split into several single - or double filaments (Fig. 4A, white triangle), supporting the idea that Ccrp59 forms bundles or sheets of individual straight filaments. Most Ccrp59 bundles had a length of 120 to 160 nm, but bundles of more than 200 nm were also observed (Fig. 4A, red triangle). The diameter of individual filaments was in the range of 10 nm, similar to IF filaments from eukaryotic cells, while the larger filament bundles had a diameter of 30 to 60 nm. Interestingly, the length of Ccrp59 bundles as well as their width increased with time and with protein concentration. 5 to 7 days after purification, large bundles up to 950 nm length consisting of individual parallel filaments could be seen (Fig. 4A, lower panels), and were still observable after several weeks, showing that these structures are highly stable.
Next, we wished to know if Ccrp59 can form filaments in vivo in a heterologous system. Accordingly, we transfected D. melanogaster S2 Schneider cells (derived from macrophages) with a Ccrp59-YFP fusion. Straight filamentous structures of 1.7 µm±0.3 (n = 9) length, which were frequently branched, could be seen soon after induction of transcription of the fusion (Fig. 5A), albeit in few cells. These observations show that the Ccrp59-GFP fusion can form filaments in vivo. Interestingly, upon coinduction of wild type Ccrp59, straight and branched filamentous structures were observed in more cells (Fig. 5B), which measured 2.45 µm±0.43 (n = 42) µm on average. At later time points after induction, large aggregates filled the transfected cells, which appeared to consist of large bundles of filaments (Fig. 5C). Importantly, Ccrp59-GFP filaments were of uniform fluorescence over their full length, and were much longer than those seen in the EM. These data show that Ccrp59 forms extended filamentous structures in vitro as well as in vivo in a heterologous system. The observations also suggest that the intracellular concentration of Ccrp59 must be kept at a certain level to avoid crowding of the cytosol by uncontrolled polymerisation of Ccrp59.
In order to obtain information on Ccrp1143, the protein was purified as Strep-tagged versions analogous to Ccrp59, and was fully soluble (Table 1). When subjected to electron microscopy, Ccrp1143 was observed as single straight filaments of 50 to 60 nm length (Fig. 4B, left panel). Interestingly, a change of the purification condition from pH 8 to pH 7 resulted in the formation of structures, which were about 600 nm long and more than 20 nm wide, as well as of round or U-shaped structures of about 50 nm in diameter (Fig. 4B, right panels, green triangles). U-shaped structures were never observed for Ccrp59, suggesting that they are intrinsic to Ccrp1143. Therefore, similar to IF proteins [21], pH conditions considerably affect the ability of Ccrp1143 to form extended filaments (see Table 1). Like Ccrp59, Ccrp1143 filaments were still observed several weeks after purification, and thus highly stable structures.
These experiments show that both Ccrp proteins form filaments in vitro, which share several properties with IF proteins, but are dissimilar to IF type proteins because both can be purified as soluble proteins without the need for refolding.
We wished to obtain insight into the pattern of localization of cytoskeletal elements in live H. pylori cells. We adapted a system for the generation of GFP fusions for Bacillus subtilis cells to H. pylori, which allowed integration of the fusions at the original locus within the chromosome. This strategy was successful with strain 1061, and in some cases also with 26695, which does not easily take up and integrate plasmid DNA (in contrast to linear DNA).
Ccrp59 was visualized through the generation of a C-terminal GFP fusion that was integrated at the original locus, such that it was expressed under the native promoter, and was the sole source of Ccrp59 within the cells. Cells of strain 26695 expressing Ccrp59-GFP were helical like wild type cells and not straight (Fig. 6A), showing that the fusion was functional, even in the absence of wild type Ccrp59. Discrete Ccrp59-GFP foci could be detected within exponentially growing cells (Fig. 6A–E). Small cells contained 2 to 3 foci, while the number of foci increased with cell size. Foci were not of uniform fluorescence, but showed different intensities. Foci with high fluorescence intensity (indicated by grey triangles in Fig. 6A, C and D) were positioned at relatively regular intervals within the cells, with an average of 0.89 µm±0.2 (n = 62) between the foci, and were frequently interspersed with foci of low intensity. Imaging of different Z-planes within cells and ensuing 3D deconvolution suggested that some of the foci were connected with each other (Fig. 6C). Due to the low cell diameter of H. pylori (0.78 µm), and because of the weak fluorescence of Ccrp59-GFP (which allows capturing of only 4–5 Z-planes) it was not possible to clearly determine if the foci are arranged in a helical pattern (which some images suggest), or in which other pattern. However, the data are compatible with a helical localization of Ccrp59 filaments along the long axis of the cells (Fig. 6F).
Ccrp59-GFP also localized in a very similar arrangement in strain 1061 (Fig. 6D), suggesting that the observed localization reflects the true positioning of Ccrp59 in several if not all H. pylori strains.
Importantly, time lapse microscopy showed that Ccrp59-GFP signals were not moving through the cells, but were stationary positioned over a period of 4 minutes (Fig. 6E). We did not observe any movement of Ccrp59-GFP foci in any of the 120 cells analysed. Thus, Ccrp59-GFP foci are not freely diffusing elements, supporting the idea that they may constitute cytoskeletal elements that are statically localized along the length of the cells.
We have not been able to generate a functional Ccrp1143-GFP fusion. We created a strain derived from 1061 in which a complete Ccrp1143-GFP fusion was integrated into the original locus by single crossover, such that the fusion as well as the wild type gene HP1143 were present within the H. pylori chromosome. Between 40 to 60% of these cells showed abnormal cell shape (Fig. 6G), and contained one to two distinct Ccrp1143-GFP foci at random places within the cell (data not shown), suggesting that the Ccrp1143-GFP fusion is dominant negative. These data reinforce the idea that a loss of function of Ccrp1143 leads to aberrant cell morphology.
As mentioned above, H. pylori strains can have different degrees of helical cell curvature and different cell lengths. Occasionally, laboratory strains lose cell curvature altogether and become rod shaped. To investigate if differences in genes encoding for Ccrp proteins may be the basis for this phenomenon, we amplified the gene region of HP0058 up to the beginning of gene HP0060 from different strains and sequenced the PCR products. Interestingly the whole region differs in length from about 1500 bp in strain 26695, 1600 bp in strain J99, about 1000 bp in strain 1061 up to only 550 bp in strain SS1 (mouse adapted). This is in agreement with a previous study that showed that HP0059 is among the most divergent genes in H. pylori [22]. Analysing HP0059 (encoding Ccrp59) itself, the size of 855 bp, 984 bp, 750 bp or 500 bp in strains 26695, J99, HPAG1 and 1061, respectively, was determined. Gene jhp0050 (is similar to HP0059) in strain J99 is 663 bp long. Because strain KE88-3887 is a hyper-motile variant of strain 26695 both strains contain the same HP0059 sequence.
To our surprise, it was possible to generate a deletion of the mreB gene, through a replacement of the gene with a chloramphenicol acetyltransferase cassette. Therefore we inactivated the mreB gene in three different strains indicating that this result was not strain dependent. The generated mutant cells were able to grow, albeit at strongly reduced growth rate compared with wild type cells. To rule out an effect on the downstream mreC gene, we isolated total RNA from mreB mutants from different H. pylori strains and performed dot-blot hybridization with probes specific for the mreC gene, showing that mreC is expressed in ΔmreB cells like in wild type cells (Fig. 7A), ruling out a polar effect of the disruption of mreB. MreB mutant cells were obtained at a similar frequency compared with many deletions of non essential genes generated in our laboratory [23], strongly arguing against the generation of secondary suppressor mutations.
The growth rate of mreB mutant cells was severely decreased compared with wild type cells for strain 26695 (Fig. 7B) and for KE88-3887 and 1061 (data not shown). The growth curves of all wild type strains showed a lag phase of about 8 hours and an exponential increase in cell density until at least 25 h, whereas all mreB mutants (i.e. in the 3 different strains) displayed a highly decreased growth rate.
Interestingly, there was no change in cell morphology of mreB mutant cells other than cell elongation in comparison with wild type cells (Fig. 8, compare A with D for 26695, B with E for 1061 and C with G for KE88-3887). MreB deleted cells were still helical and had the same average cell diameter of 0.78 µm (n>100 cells) than that of wild type cells. Cell elongation can be easily seen in Fig. 8F. Interestingly, mreB mutant cells showed a strong defect in the segregation of chromosomes. In contrast to wild type cells of all strain, which contained one, two, or (rarely) three well defined nucleoids (Fig. 8A, B and C), mreB mutant cells contained brightly staining bilobed nucleoids and large DNA free cell regions (Fig. 8D, E and G). Similar to C. crescentus smc mutant cells that have a strong segregation defect [24], no anucleate mreB mutant H. pylori cells were observed. Fluorescence intensity of the bilobed nucleoids in mutant cells was similar to that of separated nucleoids in large wild type cells, while the length of the bilobed nucleoids was twice of that of single segregated nucleoids in wild type cells, showing that the nucleoids in mutant cells contain two largely duplicated chromosomes demonstrating a separation delay. MreB mutant cells from strain 1061 frequently contained a single extended non-segregated nucleoid in spite of the large cell size (Fig. 8E), showing that loss of MreB strongly affects chromosome segregation in H. pylori cells. MreB mutant cells from strain 1061 could reach more than 3 times the normal cell size (Fig. 8F). Blocking of chromosome segregation leads to a delay in cell division in bacteria such as E. coli or B. subtilis [25], suggesting that most likely, the elongation of cells lacking MreB is due to the defect in chromosome partitioning.
To obtain further insight into the function of MreB, we treated H. pylori cells with A22, which was reported to mediate disassembly of MreB filaments in vivo [26]. Addition of a low amount of A22 (10 µg/ml) to exponentially growing H. pylori cells led to the formation of abnormally shaped nucleoids (Fig. 8H), and the addition of 50 µg/ml A22 resulted in a similar phenotype than the mreB deletion: only 2 doubling times after addition of A22, nucleoids no longer separated into two (Fig. 8I), and growth proceeded at an extremely slow rate (Fig. 7B, open triangles). Resuspension of cells after A22 treatment into fresh growth medium fully restored growth of cells, showing that interfering with MreB function transiently and rapidly leads to disturbance of the cell cycle, but does not kill the cells under experimental conditions. The fact that A22 treatment resulted in a phenotype that closely resembles that of an mreB deletion reinforces the idea that the phenotype is not masked by a secondary mutation but is due to the inactivation of MreB activity.
The lack of MreB did not interfere with the formation of polar flagella (Fig. 8F, right panel). Our data suggest that MreB plays a direct or indirect role in the progression of the cell cycle, but not in cell shape determination.
We also examined the possible link between Ccrps and MreB, because the effect of the lack of MreB on cell shape could potentially be masked by the presence of Ccrps. Therefore, we treated HP0059 mutant cells with A22, and visualized the effect of inhibition of MreB on cell shape. While wild type cells of strain 26695 (Fig. 8J) or of strain 1061 (Fig. 8K) remained helical during addition of A22, mutant cells of strain 26695 (Fig. 8L) or of strain 1061 (Fig. 8M) remained straight and rod shaped like the non-treated cells, and displayed the same degree of growth retardation as the wild type cells. These data support the findings that cell shape is not affected by the loss of MreB activity, but is determined by Ccrp proteins.
The persistence of Helicobacter pylori in the hostile environment of the human stomach is ensured by the activity of urease. Urease catalyses the hydrolysis of urea into carbon dioxide and ammonia, which are buffering compounds essential to raises the pH in the microenvironment surrounding the cell [27] and to maintain the pH homeostasis in the bacterial cytoplasm [28]. Therefore, enzyme activity is essential for both early colonization events and for virulence [29],[30]. To test whether this major pathogenicity factor is affected by cytoskeletal elements, we determined urease activity in mreB mutant cells, and found a statistically significant (p<0.01) 2.5 fold decrease in activity in strain KE88-3887 (Fig. 9A), and a ∼6 fold decrease in strain 26695 (data not shown). Interestingly, addition of 1 µM NiCl2 restored urease activity up to wt level (data not shown). Western blot analysis showed that the urease level in the mreB mutant strain is similar to or even higher than that in the parental wild type strain (Fig. 9B). It is unclear how MreB exerts its effect on urease activity, but clearly, loss of this cytoskeletal element compromises H. pylori pathogenicity.
This report provides novel insight into the bacterial cytoskeleton and the function of cytoskeletal elements, and shows that the human pathogen H. pylori has a novel type of system for the establishment and maintenance of defined cell morphology. We show that two coiled coil rich proteins (Ccrp) are essential for the maintenance of proper cell shape in H. pylori, whereas the actin-like protein MreB is not involved in the generation of helical and/or rod cell morphology, like in many other bacteria. Deletion of gene HP0059 encoding for protein Ccrp59 resulted in the complete loss of helical cell curvature in strains 26695, KE88-3887, G27 and 1061. Loss of a second Ccrp protein, Ccrp1143, resulted in a mild reduction of cell curvature in strain 26695. However, in strain 1061, lack of Ccrp1143 resulted in a complete failure to maintain cell morphology, mutant cells were round or oval, or irregularly shaped. Thus, Ccrp proteins contribute differentially to cell morphology in different H. pylori strains, and are required for the maintenance of cell morphology in H. pylori. Intriguingly, in contrast to mreB and ftsZ, Ccrp encoding genes are highly variable both in terms of their length and in sequence between various H. pylori strains analysed in this study. It is thus plausible to propose that the great variety of cell shapes of H. pylori strains – from small bent cells to large and highly helical cells – stems from the nature of the Ccrp proteins. For example, Ccrp59 is much longer in strain 26695 than in 1061, which produces smaller and less helical cells than 26695. Thus, loss of Ccrp1143 in strain 26695 may be compensated for by Ccrp59, while Ccrp59 of strain 1061 may not be able to do so. Unfortunately, we do not have sophisticated genetic tools at present to test these intriguing ideas, and clearly, the situation is more complicated, because of the differential contribution of Ccrp proteins in different strains.
We show that both Ccrp proteins form extended filaments in vitro. Ccrp59 forms bundles of filaments in vitro, in the absence of any added cofactor, and is able to form extended filaments in macrophage cells, and thus in the absence of any cofactor from H. pylori. Ccrp59 bundles clearly consist of parallel stacks of filaments, which appear to be arranged in a staggered fashion, as is proposed to be the case for IF filaments from eukaryotic cells [6]. However, both Ccrp proteins identified in this study are initially soluble proteins when expressed in E. coli cells, and Ccrp59 lacks characteristic N and C terminal domains of IF proteins, showing that Ccrps are distinct from IF proteins. Possibly, Ccrp proteins are evolutionarily older versions of IF proteins, or even unrelated to IFs, and possibly a novel class of cytoskeletal elements in bacteria.
Towards a further analysis of Ccrps, we localized Ccrp59 in H. pylori cells. We found that a functional Ccrp59-GFP fusion forms distinct foci, whose position did not change over a time of several minutes, along the length of the cells. Thus, Ccrp59 is not a freely diffusing cytosolic protein, but remains at fixed positions, and may thus serve as a cytoskeletal structure that affects cell morphology. Due to the narrow width of H. pylori cells and the relatively weak fluorescence of Ccrp59-GFP, it was not possibly to unequivocally determine if the foci are arranged in a helical pattern. Because of the fact that Ccrp59 forms extended filaments, which can be longer than H. pylori cells, in vitro, and when expressed in S2 macrophages, we favour the idea that the foci consist of filaments that are connected and run along the long axis of the cells (Fig. 6F). Rigid Ccrp59 filaments could exert force onto the cell membrane and this way lead to a helical twist in the cell wall during synthesis. Alternatively, a helical growth pattern could also be achieved through a possible interaction of Ccrps and proteins that synthesize the cell wall, which would be positioned in a helical arrangement. MreB positions cell wall synthetic enzymes in B. subtilis [31], but H. pylori MreB clearly does not affect cell shape, so this function could be performed by Ccrps. Coiled coil rich/IF-like protein crescentin in C. crescentus forms a filamentous structure along the short axis of the cell, and likewise actin-like MamK and MamJ in magnetotactic bacteria, which align magnetosomes in a straight line along the short axis of the helical cells [32],[33]. On the other hand, the cytoskeletal element CfpA, found exclusively in spirochaetes, is part of filaments running along the long axis of the highly helical cells [34], which even persist and retain their helical path when the cells are gently lysed. Interestingly, CfpA is also predicted to contain a high degree of coiled coils. It will be important to determine the nature of the foci formed by Ccrp59-GFP within the cells, and to identify factors that interact with Ccrps in H. pylori, to find out how the proteins mediate the generation of helical curvature of the cells.
In C. crescentus, the IF-like protein crescentin is essential for the generation of cell curvature [19], while MreB is indispensable for the maintenance of rod shape, in striking contrast to H. pylori, where cell shape depends on two Ccrps, but not on MreB. Moreover, Ccrp59 is clearly different from crescentin, because crescentin forms individual long filaments [19], and not parallel bundles of filaments like Ccrp59. In S. coelicolour, the filament-forming coiled coil rich protein FilP affects cell rigidity, but not cell shape [35], while MreB is involved in differentiation (sporulation), but does not play a role during vegetative growth [36]. Our findings show that H. pylori employs a novel concept for the generation of complex cell shape and suggest that Ccrp proteins may set up complex cell shape in many other bacteria that contain MreB (which may serve different functions), and also in bacteria lacking mreB, such as Corynebacterium, which is rod shaped.
We also addressed the question of the function of MreB in H. pylori. MreB mutant H. pylori cells are viable, but grow much more slowly than wild type cells. Strikingly, mutant cells contained non-segregated but strongly fluorescent (and thus duplicated) chromosomes, and were highly elongated. Because a defect in chromosome segregation leads to a delay in cell division, cell elongation in mreB mutant cells most likely stems from the delay in cell cycle progression. Thus, in H. pylori, MreB affects chromosome segregation, but not cell shape, while in other bacteria, the observed defect in chromosome segregation may be due to an indirect effect caused by the loss of cell shape. Strikingly, mreB mutants contain considerably lower levels of urease activity, whereas the amount of urease is unchanged. At present, we have no clear indication as to how MreB might affect the activity of an enzyme. Possibly, MreB affects the activity of membrane proteins such as transporters, and the absence of MreB may thereby change intracellular levels of metals and ions. Indeed, a deletion of mreB in B. subtilis can be rescued by the addition of high concentrations of magnesium and sucrose [18], and urease activity in H. pylori mreB mutant cells can be rescued by an increase in the concentration of nickel, which is a limiting factor for the enzyme [37]. In any event, our findings severely alter the spectrum of cellular functions affects by MreB. Because high urease activity is a prerequisite for colonization and persistence of Helicobacter pylori in the hostile environment of the human stomach, we establish for the first time a connection, directly or indirectly, between the bacterial cytoskeleton and a pathogenicity factor.
To verify the different contributions of Ccrps and MreB in H. pylori, we added the MreB inhibitor A22 to HP0059 mutant cells. The addition of A22 resulted in slow growth in the mutant cells, which however retained their rod cell shape. Therefore, cytoskeletal elements in H. pylori strongly affect cell shape (Ccrps) and growth/pathogenicity (MreB), which emphasizes the potential to generate antibacterial chemicals by screening for compounds that affect the assembly of MreB and Ccrp proteins. The study of H. pylori at the level of cell biology and the investigation of its cytoskeleton has revealed a novel type of system for cell shape maintenance, and point to additional interesting features of its cell cycle that deserve further investigation.
Bacterial strains are listed in suppl. Table S1. H. pylori strains were routinely cultivated on Dent blood agar in a microaerobic atmosphere as described earlier [37]. Growth experiments were performed in Brucella broth with 5% fetal calf serum (BBF). Bacteria were precultured to an optical density at 600 nm (OD600) of approximately 1.0 in BBF and subsequently diluted 1∶150 in test media. Growth rate was assessed by optical density (OD600). All growth experiments were performed in triplicate and were repeated at least three times. E. coli strains were grown aerobically at 37°C in Luria-Bertani medium. When appropriate, growth media were supplemented with 50 µg/l ampicillin (Ap) or 20 µg/l chloramphenicol (Cm).
Restriction and modifying enzymes (New England Biolabs, USA) were used according to the manufacturer's instructions. Cloning was performed in E. coli according to standard protocols. Plasmids were isolated with a QIAprep Spin Miniprep Kit from Qiagen (Qiagen 27104). The chloramphenicol-acetyl-transferase gene catGC with (Pcat) promoter were amplified by PCR from plasmid pTnMax5 (suppl. Table S1) using primer CATS1 in combination with the primer CATAS1 (suppl. Table S2). The Pcat gene were fused to upstream and downstream DNA regions of mutagenized genes by using a modified version of the megaprimer PCR protocol [38],[39] as described earlier [23],[40],[41]. Marker exchange mutagenesis of H. pylori was performed by electroporation or natural transformation according to standard procedures [42]. H. pylori mutants carrying the Pcat gene inserted into the chromosome were selected by growth on Dent blood agar containing chloramphenicol (Cm) at concentrations of 20 mg/l. Correct insertion of cat and Pcat was verified by PCR analysis with appropriate primers listed in suppl. Table S2. All fluorescent tag vectors (see Protocol S1) were integrated into the H. pylori chromosome via single crossover integration, which was verified by PCR.
D. melanogaster S2 Schneider cells were grown in Schneider's Drosophila medium (Lonza Group Ltd.) supplemented with 5–10% fetal calf serum (FCS) at 25°C without addition of CO2. Cells were passaged every 2 to 3 days to maintain optimal growth. S2 cells were transfected using the cationic lipid Cellfectin (Invitrogen). The S2 cells were spread in a 6-well plate at 1×106 per well in 3 ml medium with 5% FCS. Supercoiled plasmids (0.3 µg of each plasmid) were complexed with lipid (10 µl Cellfectin reagent) in 200 µl serum-free medium. The complex was incubated at room temperature for 15 min, filled up with serum-free medium to 1 ml and was added to cells from which the growth medium had been removed (cells were washed once with serum-free medium). After 18 hrs, the supernatant was removed and replaced by 3 ml of medium containing 5% FCS. After further incubation for 24 hrs, the production of the proteins was induced by adding CuSO4 to a final concentration of 1 mM.
Recombinant versions of the H. pylori HP0059 and HP1143 proteins were produced in E. coli using the StreptagTM protein expression system from IBA (Göttingen, Germany) according to the manufacturer's instructions (http://www.iba-go.de). The coding sequences from H. pylori strain 26695 were amplified using the primer pairs listed in Table S2 and cloned via the BsaI restriction sites added as 5′-extensions (underlined) into plasmid pASKIBA-7TM (IBA-Göttingen). The plasmids were transferred to E. coli BL21 and expression was induced with 0.2 mg/l tetracycline. The bacteria were harvested by centrifugation and the recombinant proteins were purified to homogeneity on a Strep-TactinTM column according to the manufacturers' instructions. In case of HP1143 the coding sequence from H. pylori strain 26695 was amplified using the primer pairs listed in Table S2 with the Streptag sequence integrated and cloned via the NcoI and PstI restriction sites into plasmid pETDuet-1 (Novagene). Protein expression was performed according to the manufacturer's instructions. For protein purification at pH 7, buffer W (100 mM Tris/HCl pH 8, 150 mM NaCl, 1 mM EDTA) as well as the buffer E (100 mM Tris/HCl pH 8, 150 mM NaCl, 1 mM EDTA, 2.5 mM Desbiotin) were adjusted to pH 7.
Spin down assays were performed as follows: 20 µl of purified protein fractions in buffer W (usually 24 h after elution, with storage at 4°C) were centrifuged at 13000 rpm in a bench centrifuge, after removal of the supernatant, the pellet was resuspended in 20 µl buffer W. SDS sample buffer were added and equal volumes of supernatant and pellet were subjected to SDS PAGE analysis.
Urease activity was determined in fresh lysates by measuring ammonia production from hydrolysis of urea, as described previously [37],[43]. The concentration of ammonia in the samples was inferred from a standard NH4Cl concentration curve. Enzyme activity was expressed as µmol urea substrate hydrolysed min−1 (mg protein)−1.
Elution fractions of the Streptag purification and from gel filtration were applied to 200 mash copper grids and were negatively stained with 2% phosphotungstic acid pH 2.7 or with 1% uranyl acetate. Filaments were visualized under a Philipps/FEI CM10 (80000V) electron microscope equipped with a Bioscan Camera. Images were processed with Digital Micrograph (Gatan) software.
Fluorescence microscopy was performed on a Zeiss Axioimager microscope using a 100×Objective with A = 1.45. Cells were mounted on agarose gel pads on object slides. Images were acquired with a digital CoolSnap HQ CCD camera; signal intensities and cell length were measured using the Metamorph 6.3 program (Universal Imaging Corp., USA). 3D deconvolution was done using Autdeblur software. DNA was stained with 4′,6-diamidino- 2-phenylindole (DAPI; final concentration 0.2 ng/ml) and membranes were stained with FM4–64 (final concentration 1 nM). Filters used were: DAPI – ex360–370, dc400, em420–460, GFP – ex460–495, dc505, em510–550, FM4–64 ex480–550, dc570, em590. For acquisition of Z-stacks, 20 planes with a spacing of 0.2 µm were taken from bottom to top, or reverse. Signal intensities were measured in Metamorph program.
|
10.1371/journal.pgen.1005428 | Variability of Gene Expression Identifies Transcriptional Regulators of Early Human Embryonic Development | An analysis of gene expression variability can provide an insightful window into how regulatory control is distributed across the transcriptome. In a single cell analysis, the inter-cellular variability of gene expression measures the consistency of transcript copy numbers observed between cells in the same population. Application of these ideas to the study of early human embryonic development may reveal important insights into the transcriptional programs controlling this process, based on which components are most tightly regulated. Using a published single cell RNA-seq data set of human embryos collected at four-cell, eight-cell, morula and blastocyst stages, we identified genes with the most stable, invariant expression across all four developmental stages. Stably-expressed genes were found to be enriched for those sharing indispensable features, including essentiality, haploinsufficiency, and ubiquitous expression. The stable genes were less likely to be associated with loss-of-function variant genes or human recessive disease genes affected by a DNA copy number variant deletion, suggesting that stable genes have a functional impact on the regulation of some of the basic cellular processes. Genes with low expression variability at early stages of development are involved in regulation of DNA methylation, responses to hypoxia and telomerase activity, whereas by the blastocyst stage, low-variability genes are enriched for metabolic processes as well as telomerase signaling. Based on changes in expression variability, we identified a putative set of gene expression markers of morulae and blastocyst stages. Experimental validation of a blastocyst-expressed variability marker demonstrated that HDDC2 plays a role in the maintenance of pluripotency in human ES and iPS cells. Collectively our analyses identified new regulators involved in human embryonic development that would have otherwise been missed using methods that focus on assessment of the average expression levels; in doing so, we highlight the value of studying expression variability for single cell RNA-seq data.
| In order to function properly, cells express specific sets of genes that are regulated via complex transcriptional programs. During early stages of development, when an embryo consists of only a handful of cells, it is vital that these cells work together so that the embryo can develop into a healthy baby. Single cell studies allow us to understand how each cell contributes to ensuring proper regulation of the embryo, as well as identify the critical genes whose expression is important for development. While we understand that regulation of a gene occurs through the timing of when it is expressed and also the quantity of its expression, more recently we have come to recognize that the variability of a gene’s expression across single cells may also contribute to the viability of the organism. In this study, we analyzed the gene expression variability of human embryos at different developmental stages. We discovered distinctive patterns of variability across cells in the embryo; some genes had extremely stable expression, and others were variable but with increased homogeneity in expression at a particular developmental stage. We validated one of these stage-specific markers and found that it played a role in the maintenance of pluripotency of human pluripotent stem cells. Overall, these results can help unlock additional clues into understanding how embryonic development is regulated in humans.
| The regulatory program that ensures that a human embryo can develop successfully starting from a single cell zygote is one of the most fascinating examples of systems-level genetic control. During development, individual cells must quickly respond to internal and external signals while the number of cells that make up an embryo increases at a rapid rate. How an embryo is able to coordinate signals cooperatively across all cells, while subsets of cells undergo diverse fate transitions to specific lineages remains an open question. Inherent in an embryo’s regulatory program is the need to balance flexibility with robustness to ensure that development can continue in spite of perturbations that may occur. Studying how individual cells alter their transcriptomes as an embryo transitions through each developmental stage presents an opportunity to understand the core of the regulatory program, and specifically how robustness is maintained throughout development.
Single cell technology has revolutionized almost every arena of the biological sciences but perhaps none more significantly than developmental biology. Profiling the transcriptomes of individual cells provides a means to disentangle heterogeneous properties that can identify small numbers of distinct or rare cells amongst a population of cells that otherwise appear identical based on a handful of markers [1]. As early as the 2000s, studies have demonstrated the limitations of inferences derived from bulk cell approaches, where transcriptomes from multiple cells are combined to create an ensemble representation of a generalized single cell [2–4]. This ensemble-based model, referred to as an “averaged cell” by Levsky and Singer [5] is unable to capture the variability inherent in gene expression in cell populations and therefore provides only a marginal insight into transcriptional regulation. Applying single cell profiling to understand developmental processes has been invaluable for pinpointing specific genes that direct cell fate transitions towards distinct lineages [6]. The ability to resolve heterogeneity in gene expression amongst cell populations has been useful in revising and identifying more specific cell types; and in the process, we are only beginning to appreciate just how diverse the transcriptional states underlying each developmental stage can be [7].
Studying transcription at the single cell level has forced us to confront the fact that gene expression consists of both signal and noise [8–10]. While the recognition that transcription is a stochastic process predates single cell technology [11, 12], it is only recently that we have come to appreciate the insight that variability in gene expression can shed on understanding regulatory control [13–15]. In the context of single cells, inter-cellular variability represents a measure of consistency or dispersion of a specific signal amongst a cell population. As a byproduct, expression variability also inversely reflects our ability to predict or have confidence in the transcriptional state for a new cell. For instance, a gene with low expression variability has high generalizability for any future cells sampled at random and therefore may be valuable as a marker of that cell’s state. Conversely, a gene with high expression variability is one whose expression levels fluctuate widely across cells in the population, and this heterogeneity could be due to several factors that can provide information about its regulation, e.g. a difference in cell cycling, cell fate, or stochastic component of the regulatory program used by the cell [16–18] (see Fig 1A).
When characterizing a phenotype based on differential gene expression, the statistical methods that are typically used, such as a t-test or ANOVA, are based on deviations detected in average expression and assume constant levels of variability. While this is a sensible approach for identifying overall shifts, we acknowledge that studying dispersion or variability in expression is also important to understanding regulation in single cells [19, 20]. Single molecule studies performed at single cell resolution have demonstrated that gene expression is inherently stochastic, where even amongst isogenic cells, a gene’s expression level is not identically distributed [21]. Most of this stochasticity is because the production of a transcript requires the coordination of multiple components in a cell, some of which are present at very low concentrations. Gene expression can be modeled as a probabilistic event, and thus in a cell population, each cell contains a gene that is expressed in an all-or-nothing, binary manner according to a certain probability. This generates cell-to-cell heterogeneity in the population where some cells have genes whose expression is switched on while in other cells, expression is off [22]. Consequently, variability represents a very real component to understand gene expression, and by assuming constant variability and not studying this property directly, we may miss out on identifying key regulators in single cells [23, 24].
Recently, Yan et al. [25] used single cell RNA-sequencing (RNA-seq) to elucidate the transcriptional landscape of preimplantation human embryos from distinct stages of development (Fig 1B). This study used RNA-seq to comprehensively profile individual cells that were derived from the same human embryo at early stages of development. Major classes of genes with key temporal changes over the course of development were identified, including novel lncRNAs regulators, potentially new protein-coding transcripts, as well as dynamic patterns of alternative splicing. The application of single cell profiling provided new insights into transcriptional regulation of human embryos. It is worth highlighting, however, that the analytical methods used by Yan et al. [25] are based on averaging expression from the cell populations. While this is adequate for identifying average trends, their study fails to use the valuable opportunity presented in this data set to understand how variability in gene expression is distributed at the cell-to-cell level. For single cell RNA-seq data, an analysis of expression variability is useful in identifying genes that are invariantly or heterogeneously expressed in cell populations, which in turn may provide insight into regulation. Here, we use the data set generated by Yan et al. [25] to demonstrate the additional utility that analyzing variability of gene expression can bring to our understanding of transcriptional control in human embryonic development (see Fig 2 for an overview of the analysis performed). Our analysis approach is composed of the following key steps: (1) identification of stable genes across development, (2) detection of stage-specific variability markers (3) identification of control states based on different levels of expression variability.
Our results show that certain genes do have extremely stable inter-cellular expression during development, and that this stability appears to play an important role in regulating the cell. Stable genes are expressed across a range of levels, and those with low expression are enriched for cell type-specific processes whereas those with high expression are involved in fundamentally important cellular processes like maintenance, and metabolism. We also identified genes that could function as markers of stage, as these genes are noticeably expressed with increased homogeneity for cells of a particular developmental stage. Overall, our findings point to variability in gene expression as a regulatory feature of developmental processes in human embryos.
The embryos that were profiled in Yan et al. [25] were donated by women who had already undergone successful in vitro fertilization (IVF), delivered a healthy baby, and elected to donate the remaining high-quality embryos that had been cryopreserved. On average, the women were 30 years old, all had tubal-factor infertility, and all had partners with normal semen parameters. To be included in the profiling study, the embryos underwent extensive screening for good morphology and high quality. Quality assessments were performed using externally-derived, established scoring criteria for viability of the embryos [26], in conjunction with explicit definitions of embryonic stage (as outlined in the Materials and Methods section of Yan et al. [25]). To the best of our knowledge, we are confident that these embryos were viable.
We tested the data to ensure that an appropriate standard of data quality was met for studying gene expression variability. Stringent thresholds were used to discard genes that were expressed below a specific quality threshold in the majority of cells that were profiled (Materials and Methods, S1 Text). To measure inter-cellular expression variability, we adopted a statistic termed the SDC that represents the standard deviation (SD) of gene expression between cells of the same embryo that is averaged across multiple embryos (Materials and Methods, Fig 1A). The data were inspected to ensure that patterns of inter-cellular variability were not influenced by the number of cells contributing to each stage (see S1 Text). We also investigated the consistency of the inter-cellular variability measures between the embryos and observed that the transcriptomes of the embryos were relatively stable compared to each other (see Fig 3B). This observation lends support to treating the embryos as replicates to estimate the inter-cellular expression variability in our study (see S1 Text). We also evaluated the possibility that chromosomal copy number mosaicism could affect the variability analysis for those genes expressed on an aneuploid chromosome in a subset of cells. To check for chromosomal copy number effects on expression, we compared the normalized gene expression distribution for every chromosome (see S2 Text) and found no obvious aberrations in chromosome-wide expression levels [27].
The analysis of gene expression variability remains an area where guidelines for the best statistical practices are still maturing. For this reason, we investigated the utility of using two statistics, the SDC, based on the SD and the coefficient of variation (CV) for studying gene expression variability. The CV is a standardized measure of variability or dispersion that is calculated by taking the ratio of the SD and the average. It is a measure that has previously been used in studies of expression variability for microarray-based data [13, 14]. One advantage of the CV is that it addresses any potential correlation between the variability and average. Nevertheless, it can be more difficult to interpret what a CV value represents for an individual gene with respect to its average expression and expression variability. Another potential issue is that as a ratio of two statistics, the CV is also in theory subject to zero-inflation for genes that have very low levels of average expression. Such genes will be falsely assigned a higher level of variability as measured by the CV, independent of how dispersed the data for this gene really is.
A factor that often arises when studying expression variability is the assumption that the mean and variability of a gene are correlated. Based on our comparison of the CV and SD, the two statistics appear to handle the nature of this correlation quite differently (see S3 Text). We found that the CV had a higher negative correlation with the average expression than the SDC for all four stages. We also conducted simulation studies to test the ability of the two statistics to identify genes with three different levels of variability. We tested the SDC and CV on three different test cases and overall, the SDC appeared to be the better performing statistic with the ability to identify the correct number of variability states more decisively than the CV, and also in general, increased precision in classifying genes to their correct level of variability. We also repeated our analyses of expression variability to compare the results of our main findings when the CV statistic was used. We found that although there was an overlap in the stable genes identified, the two statistics still measure different properties of expression variability. We also noticed that within a specific developmental stage, genes with a low SDC and high average expression, were more likely to have a low CV value, however the converse was not true. Genes with a low CV had both average values and SDC values that spanned a wide spectrum (in some cases greater than the third quartile for SDC, and less than the first quartile for average expression). Therefore, for applications such as marker detection, using CV may be limited as it does not provide the kind of control that comes from setting specific filters on the SDC and average expression directly. Overall, our analyses, which are outlined in more detail in S3 Text, point to the SDC as being the more informative statistic to study gene expression variability for the Yan data set.
Focusing on genes whose expression remains invariant amongst all cells across the developmental stages may reveal the subset of core regulators that are integrally important to the developing embryo. Genes expressed at very precise levels in a cell population may correspond to widespread regulators that are ubiquitously switched on to maintain homeostasis, or reflect specialized processes that must be systematically turned off until the embryo is ready. We used Levene’s test to determine those genes with a non-significant change in expression across development (adjusted P-value > 0.05), and clustered the SDC values of these genes using a Normal mixture model to identify the gene clusters with low, medium and high variability. We identified 955 genes whose variability in gene expression varied the least across the four stages (see Materials and Methods, S1 Table, S1 Fig) and this group was designated as the group of stable genes. Inspection of the stable genes revealed that they could be classified into three distinct modes of absolute expression; corresponding to low, medium, and high levels of expression (Fig 4).
To understand how these different modes may be contributing to developmental regulation, we identified pathways and functions through Ingenuity Pathway Analysis (IPA) that were significantly enriched in each of these three groups. Overall, we observed that stable genes with low levels of expression were involved in pathways regulating specialized cell types (see S2 Table). For all four stages, we saw terms that related to disease processes affecting a specific cell type, e.g. melanoma and chronic leukemia. For specific stages, we also observed enrichment of tissue-specific diseases that affect different organs, such as the kidney (renal cancer, 4-cell and 8-cell and blastocyst), endometrium (endometrium carcinoma, 4-cell, and blastocyst), head and neck (8-cell), and colon (blastocyst) (see S2 Table).
For the stable genes with medium levels of expression, 59% were found to be common to all four stages, suggesting that they form a core set of housekeeping-type genes that are expressed in a non-stage specific nature (S3 Table, see S2 Fig for single cell profiles of some of the stable genes with medium expression). This gene set was enriched for fundamentally important pathways, including those controlling transcription (cleavage and polyadenylation of pre-mRNA), translation (EIF2 signaling, regulation of eIF4 and p70S6K signaling), protein ubiquitination, mTOR signaling and DNA damage (cell cycle: G2/M damage checkpoint regulation) (see S3A Table). The medium-expressed group of stable genes were also enriched for signaling pathways involving key growth factors and receptors (e.g. ERK5 signaling, estrogen signaling, and ephrin signaling). We observed enrichment for processes related to structural remodeling (e.g. remodeling of epithelial adherens junctions, epithelial adherens junction signaling, and regulation of actin-based motility by Rho). These are important processes for the growing embryo, as critical inter-cellular communication is known to be transmitted via adherens junctions. Enrichment in functional terms obtained from IPA support these themes as well, where enriched key terms include “initiation of translation of mRNA”, “protein and expression of RNA” (see S3B Table). Significant terms related to infection were also observed (“infection of embryonic cell lines”, “epithelial cell lines”, “viral infection”, “HIV infection”, “infection by HIV-1”, “infection by RNA virus”), and these terms are likely to reflect the massive degree of cellular proliferation occurring in the embryo.
The stable genes with high levels of average expression were among the smallest subset identified. These genes overlapped significantly with pathways already detected in the stable gene set with medium levels of average expression. Specifically, these genes were enriched for EIF2 signaling, regulation of eIF4 and p70S6K signaling, and mTOR signaling at all stages (S4 Table). For the blastocyst stage, significant enrichment of processes involving cell cycle (G1/S checkpoint regulation) and NADH repair was observed.
We used the list of stable genes that were expressed at the 4-cell stage to make inferences on which genes may be contributed by the maternal transcriptome. We identified genes that were highly expressed in the oocytes but lowly expressed in human embryonic stem cells (hESCs) and intersected this list of genes with the stable genes that had either medium or high average expression at the 4-cell stage embryos. We also identified genes that were likely to be part of the early zygote transcriptome by looking for stable genes that were either expressed at high or medium average levels at the 4-cell stage, but not highly expressed in the oocytes. We found 90 genes that were likely to be contributed by the maternal transcriptome (S5 Table). For the early zygote genes, we identified 626 genes that were active at the 4-cell stage, and 239 genes that were repressed or lowly-expressed (S6 and S7 Tables).
One of the genes likely to be contributed by the maternal transcriptome that we identified was DPPA3, a known maternal factor in the mouse that has a key role in embryonic development. DPPA3 was actually the only gene that was highly expressed at the 4-cell stage (the remaining 89 genes had medium expression at the 4-cell stage). We also showed that these stable, early zygotic genes were significantly enriched in genes that were known to be targets for the human pluripotency transcription factors SOX2 and NANOG (P-value < 0.05) but not OCT4 [28]. The analyses and results are further described in S4 Text.
While we hypothesize that many of the stably expressed genes serve an important regulatory function for human embryonic development, experimentally validating this hypothesis is a significant challenge. Instead, we can use the genome-wide catalogues that have been collected from studies of human disease genes, or genes that represent essential or robust elements of the genome to computationally infer the functional impact of the stable genes that we have identified, on the cell and embryonic development. We conducted a meta-analysis using seven gene catalogues that represent different aspects of essentiality or perturbation of the human genome. The catalogues that were used were the Genome-wide Association Study Catalog (GWAS Catalog) [29], the Online Mendelian Inheritance in Man (OMIM) Gene Map [30], a set of human orthologs of mouse essential genes [31], a set of the top 10% most ubiquitously expressed human genes [32], a set of human haploinsufficient genes [33], a set of human recessive disease genes that had a DNA copy number variant (CNV) deletion [34], and a set of loss-of-function genetic variants in human protein-coding genes [35]. We found that the stable genes were significantly associated with an enrichment of the essential genes and the top 10% of ubiquitously expressed genes (two-sided Fisher’s exact test P-value < 0.05, see S5 Text). The stable genes were marginally significant in overlap with the haploinsufficient genes and the OMIM Gene Map (P-value < 0.10). We also found a significant depletion of stable genes from the loss-of-function variants and the list of recessive disease genes affected by a CNV deletion (two-sided Fisher’s exact test P-value < 0.05). These results provided some evidence to support the claim that the stable genes had important functional consequences for regulation of the cell and by extension, we assume, the embryo. Stable genes were significantly enriched for genes that were essential or ubiquitously expressed, and less likely to be associated with loss-of-function variants or a known recessive disease gene affected by a CNV deletion.
Variability is a statistical property that reflects how the distribution of gene expression in all cells is shrinking or expanding. Genes that have changes in variability at different stages of development may therefore shed light on important stage-specific regulators of the embryo. To identify such genes, we applied the following criteria (1) a gene must have a statistically significant change in expression variability across the four stages based on Levene’s test [36], (2) the minimum level of variable expression for a specific stage (all other stages had a higher SDC), (3) an average expression level at a specific stage that was higher than all other stages (see Fig 5B–5D). To avoid confusion with markers that are found using average-based approaches that are typically employed, we refer to genes satisfying the three criteria listed, as variability markers.
For morulae and blastocyst stages, we identified eight and eleven stage-specific variability markers, respectively (Table 1, Fig 5C and 5D). These genes share a common theme in functioning to ensure the proper development of the embryo, and they mainly play a role in embryonic development, cell protection, or metabolism. For example, mesoderm development candidate 2 (MESDC2), a morulae variability marker, functions as a key mesenchymal chaperone protein for Wnt co-receptors [37] (see S3 Fig). For blastocysts, an example is epithelial cell adhesion molecule (EPCAM), which has a critical role in the formation of trophectoderm, the outer layer of the blastocyst that will eventually become the placental interface [38, 39]. Some of these genes act specifically to protect the cell from free radicals and other toxins, such as blastocyst variability markers peroxiredoxin 6 (PRDX6) [40–42] and glutathione S-transferase pi 1 (GSTP1) [43]. For single cell expression profiles of the variability marker genes, and functional information from the literature, see S3 and S4 Figs for the morula set, and S5 and S6 Figs for the blastocyst set.
When considering the functions that are served by the blastocyst marker genes, one theme was metabolism. In order to ascertain that this was not an artefact of the local environment that the human embryos were stored at, we looked at the expression patterns of the blastocyst marker genes in human induced pluripotent stem cell lines (iPSCs) and hESCs from three other studies [44–46] as well the hESCs that were profiled by Yan et al. [25]. For these variability marker genes, we saw high levels of expression relative to the global distribution of expression in all cell lines tested, providing some evidence to suggest that the patterns we observed for the blastocyst marker genes were not likely to be solely a product of the embryo’s environment (see S6 Text).
We also applied an ANOVA model to determine which variability markers could be identified using a standard approach (see S7 Text) and which were specific to our analysis based on gene expression variability. Some of the variability markers had statistically significant P-values from the ANOVA model, and these genes have been highlighted in bold and underlined in Table 1. The largest degree of overlap was observed for the blastocyst variability markers, where only two genes, LAMTOR1 and RPL17, were uniquely detected based on our variability-based approach.
To demonstrate the functional impact of the stage-specific variability markers, we validated one of the blastocyst variability markers, HDDC2, by shRNA-mediated knockdown in a human iPSC line. qPCR experiments confirmed that the knockdown of HDDC2 mRNA caused a significant decrease in the expression of key pluripotency markers DNMT3B and NANOG, two genes that are critical for embryonic development and maintenance of pluripotency in iPSCs (see Fig 6). We also tested the impact of the up-regulation of the HDDC2 locus on hESC differentiation using a hESC line with a stably-integrated inducible CRISPRa/Cas9-VP64 artificial transcriptional activator system. Over-expression of HDDC2 attenuates the drop in expression of the pluripotency marker NANOG and induction of the neuroepithelial marker PAX6 during the early stages of neural differentiation (see Fig 7). While these experiments cannot provide evidence for the effects of HDDC2 on cell viability or embryonic development, the results from the HDDC2 knockdown suggest that this gene plays a role in the maintenance of pluripotency. From the transcriptional effects caused by HDDC2 over-expression on early neural differentiation, we can infer that HDDC2 is able to either reinforce the persistence of the pluripotent phenotype, or is involved with specific interference of the neural differentiation process, or both.
In early developmental processes, evidence supports the role of gene expression variability as a necessary initiating step that precedes lineage commitment in progenitor cells [8, 47, 48]. As embryos transition from the 4-cell to the blastocyst stage, the global distribution of inter-cellular variability shifts from lower to higher values (Fig 3A). Based on the shape of the density distribution, we see an increase in the overall number of genes that are expressed more heterogeneously between cells as the embryo differentiates. This result likely reflects the diversification of transcriptomes observed in cells that are undergoing the necessary cell fate changes to become distinct lineages and specialized cell types [49] but which also contain inherent stochasticity [50].
The shape of the inter-cellular variability distribution demonstrates how at each developmental stage, the transcriptome is expressed with levels of precision that span a wide spectrum (Fig 3A). This is unsurprising given that during the 4-cell to the 8-cell stage, the embryo undergoes activation of the zygotic genome (ZGA) and a massive reprogramming of the transcriptome occurs [51, 52]. The stochastic nature of reprogramming has been observed in stem cells as well as the existence of alternative stem-cell states, and hence we expected to see higher levels of SDC [53]. Analyzing expression variability gives us a window into which components of the transcriptome are being used in cells of the developing embryo at different levels of precision. We are interested therefore in characterizing how precision changes as the embryo develops, and which genes are involved at either end of the regulatory control spectrum. We refer to these different levels of expression variability as control states. Mixture models were applied to identify within the data, the number of control states present at each stage and classify the subsets of genes observed under different levels of regulatory control (see Fig 8, Materials and Methods). The 4-cell stage had the largest number of control states, where each state could be interpreted as one of four distinct levels of variable expression (low, medium, high, very high) (Fig 8A). During development, we observed that these levels collapse down into simpler states, e.g. for the morula stage, there are three levels (see Fig 8C) and for the blastocyst stage, there are only two (Fig 8D).
We investigated whether genes with the same level of regulatory control were enriched for certain pathways or processes using IPA software. Much like the variability markers, we observed consistent changes in pathways that appear to protect and ensure adequate gene regulation in the developing embryo (S8A–S8D Table). For example, four pathways were enriched for low-variability (most stable, mixture 1) genes, EIF2 signaling, regulation of eIF4 and p70S6K, mTOR signaling, and protein ubiquitination. This overlap was significant for all four developmental stages, and represents pathways that control important functions that are required by every living cell (for an overview of the important enriched terms occurring per stage, see S7 Fig).
For the 4-cell stage only, the low-variability genes were enriched for the “DNA Methylation and Transcriptional Repression Signaling pathway” (S8A Table). The enriched genes (SIN3A, SAP18, SAP130, RBBP7, RBBP4, CHD4) were involved in regulation of histone deacetylation, and DNA methylation (DNMT1). Proper regulation of histone deacetylation is known to be a requirement for embryonic development [54]. DNMT1 is involved in maintaining single-stranded methylation of newly replicated DNA, and its expression is activated by cell cycle-dependent transcription factors in the S phase [55]. We also observed enrichment of a key developmental pathway “Ephrin B signaling” for the 4-cell low-variability group but at no other stage. Genes enriching this pathway have been directly implicated in fish and mammalian embryo development, including Rho-associated, coiled-coil containing protein kinase 2 (ROCK2), beta catenin (CTNNB1) and ras homolog family member A (RHOA) [56–58].
Pathways associated with telomere extension by telomerase, and telomere signaling were statistically significant amongst variably-expressed genes at the 4-cell and the blastocyst stages (mixtures 3 and 2 respectively, S8A and S8D Table). The importance of telomerase to preimplantation embryos has previously been established via its link to reproductive potential [59]. At the 4-cell stage, telomeres are thought to reset during genome activation, which may explain the higher number of telomerase-associated genes that have variable expression at this stage. The six genes that were annotated to the telomerase extension pathway by IPA (TERF2, TNKS, HNRNPA2B1, TERF2IP, TNKS2, POT1, XRCC5, RAD50, and NBN) have roles in double-strand DNA break repair and damage response, where they are responsible for protecting the telomere extension process. Telomere DNA elongation has been observed to occur between the 8-cell and the blastocyst stages in both animal [60, 61] and human embryo studies [62]. It is likely that this corresponds to the second enrichment of variable genes observed at the blastocyst stage. Although successful telomerase elongation is critical to the embryo, variability in expression of genes associated with this process may be due to the slightly different rates at which cells are using and completing this process.
The transition from the morula to the blastocyst features the appearance of the first cell type specification with the organization of the trophectoderm. Integrin binding and activation is an essential element of blastocyst implantation and trophoblast differentiation [63–65]. Integrins are a class of heterodimeric transmembrane cell surface receptors that participate in cell-cell interactions and regulate signals for cell adhesion, growth and survival. We found that genes in the IPA pathway “Integrin Signaling” first appeared in the gene clusters for the morula stage, and continued to be significantly enriched in the blastocyst stage, possibly indicating the variable process of cell differentiation within the morula before the visible segregation of the trophectoderm (S8C and S8D Table).
For the blastocyst stage, we find that low-variability genes were enriched for metabolic pathways, such as oxidative phosphorylation, oxidative stress, and glycolysis-related pathways (S8D Table). In fact, glycolysis and gluconeogenesis pathways are uniquely observed to occur for low-variability genes in the blastocyst stage. This may indicate the appearance of a more diverse regulatory program in blastocyst metabolism. Blastocysts are known to require a higher metabolic load due to the burst in developmental change that occurs at this stage. For example, the cytosolic form of malate dehydrogenase (MDH1), which is featured in both oxidative phosphorylation and gluconeogenesis pathways has been shown to be important for blastocyst development in mice [66]. On the other hand, by the blastocyst stage, because of the emergence of distinct lineages and cell types, cells across the embryo may begin to show signs of asynchronicity in cell cycling. We see this effect reflected in the statistically significant enrichment of cell-cycle related processes in variably-expressed genes at the blastocyst stage. Dependence on specific developmental and signal transduction pathways is also expected to be more heterogeneous at this stage owing to the distinct cell types that have different expression programs. Variably-expressed genes at the blastocyst stage were enriched for several signaling pathways, including ATM signaling, PI3K/AKT signaling, JAK/Stat signaling, ERK/MAPK signaling, and others (S8D Table).
While our analysis so far has concentrated on identifying the gene-specific measures that vary in a cell population (see Fig 9A), it is worthwhile to highlight those cells that are the most aberrant or consistent in the population, with respect to global gene expression. Sampling a fixed number of cells and analyzing patterns of variability allowed us to look at the overall heterogeneity of the cell population based on the global gene expression variability distributions. Inspection of the density plots obtained for each 4-cell sample revealed interesting properties that provide insight into the transcriptional noise occurring at a specific development stage (Fig 9B). Multiple draws of cells from the population can highlight how good of a representation any single cell sampled at random is likely to be. For situations where it is not possible to sample every cell in an organism or tissue, this kind of analysis gives us the ability to assess which parts of the transcriptome are more generalizable and stable (Fig 9B), and conversely, which cells may be unusual or aberrant.
For the 8-cell stage, we observed that all 4-cell combinations produce overlapping densities and suggest that most cells are relatively similar to each other in the embryo at this stage of development with respect to global variability (Fig 9C). The shape of the densities also indicate that most of the transcriptome is expressed with lower variability. For the morulae stage, it is clear that there are three dominant sub-clusters of 4-cell combinations and that this stage has the greatest degree of cellular diversity. For the blastocyst stage, this diversity appears to subside, and cells are more homogeneous with respect to their overall variability profile. The density for the blastocyst combinations suggest that although there is reduced cellular heterogeneity compared to the morulae stage, there are more genes expressed at higher levels of inter-cellular variability. These densities may reflect a transition signature of the morula cells. In blastocysts, there are at least two major tissue types (trophectoderm, inner cell mass) whereas the morulae are still in an undifferentiated state where cells are only starting to transition into two different types. The blastocyst, on the other hand, is where we see the first differentiation events occurring.
One limitation of the human embryo data set was that it featured relatively small numbers of embryos, and hence to test the robustness of our results, we looked to two separate studies that were performed in mouse as a source of data to validate our key findings (see S8 Text). The first study from Guo et al. [7] generated gene expression data for 48 genes in five to seven embryos per stage using the Fluidigm Biomark System 48.48 Dynamic Arrays. We selected the data from the 4-cell, 8-cell, 16-cell, 32-cell and 64-cell stages as these paralleled the most closely with the human developmental stages that were used in our analysis of the human embryos. Using the Guo data set we were able to verify the existence of a subset of stably expressed genes that had invariant expression across the development of the embryo, and these genes showed low, medium and high levels of average expression. We also detected sets of variability markers that displayed changes in average expression and expression variability that appeared to delineate a specific stage of development. Although only 48 genes were profiled in this data set, we saw that the distribution of expression variability for the total set of genes adopts wider values as the mouse embryos developed. Even with a higher number of mouse embryos, the distribution of inter-embryo variability remained relatively constant as we had observed with the human data.
In the second mouse study, Deng et al. [67] applied Smart-seq and Smart-seq2 single cell RNA-sequencing to profile the transcriptomes of cells taken from early stage embryos. Their experimental design was similar to that of the Yan human embryo data, but with a higher number of embryos and cells. We used the 4-cell, 8-cell, 16-cell and late blastocyst stages from their study to compare our results. Our analysis of the Deng data set also yielded results that aligned with those observed from our study of the human embryos. Moreover, because the number of genes in both data sets was more similar, we were able to compare the overlap between the lists of stable and variable genes that were identified in the mouse and human embryos. The overlap was statistically significant (two-sided Fisher’s exact test P-value < 10−23, odds ratio estimate 3.744), even after permutation tests (see S8 Text). We tested the gene lists from the human-mouse comparisons for enrichment of biological pathways and processes using the MSigDB database [68] (based on Hallmark gene set terms and C4 computational gene set terms) to further investigate the nature of both the overlap between the human and mouse stable genes, but also the species-specific differences. Genes that were stable in both human and mouse embryos, were enriched most strongly for MYC targets and housekeeping genes. We found that the discordant genes (either stable in mouse but variable in human, or vice versa) were enriched for more disease-specific terms and processes. Two terms that were unique to the list of genes variable in human but stable in mouse, reflected gene sets that regulate pluripotency (the Wong Embryonic Stem Cell Core [69] and the Mueller PluriNet protein-protein interaction network [70]). Based on these results, genes that are discordant with respect to expression variability may play a role in human-specific (or mouse-specific) diseases, or be required for species-specific embryonic development and maintenance of pluripotency.
We have taken a novel approach to uncovering how the transcriptome is regulated in human embryos. The use of technologies enabling analysis with a single cell resolution enables an embryo to be modeled as a defined population of cells, and by studying variability between these cells directly, we can identify genes whose expression is either homogeneously or heterogeneously distributed in the cell population. Our analyses have highlighted key developmental pathways that show different degrees of regulatory control, and we have demonstrated that analyses of expression variability can provide functional insights into stage-specific markers. It is important to emphasize that all our inferences on regulatory control of the transcriptome are based on data collected from human embryos. This has an impact on improving our understanding of transcriptional regulation in human developmental processes. A recent study has revealed that even for mice, a species that is often believed to be a good model of human biology, the differences in transcriptomes from humans are larger than expected [71]. In embryology, there are many technical, ethical and scientific limitations that understandably make the use of non-human embryos more feasible for—omic studies; however, it is useful to keep in mind that the results of our study provide valuable insights into how the transcriptome is controlled and used during the development of an actual human embryo.
The greater diversity of control states observed for the 4-cell stage is likely to be a result of the transcriptome being in a continued state of flux and reorganization following zygotic genome activation (ZGA). As the embryo transitions away from the 4-cell stage, plasticity in global gene expression gives way to the commencement of the more specific differentiation programs. This phenomenon may overlap with what other studies have described as “waves of transcriptional activation” [52]. We see genes in the blastocyst stage segregate to one of two different control states (see Fig 8D), and from this, we speculate that control of the transcriptome has polarized into low and high levels of precision, with fewer intermediately-variable genes observed than in earlier developmental stages (see Fig 8A and 8B).
The decreased complexity of regulatory control observed across development is also consistent with our model of how ZGA is influenced by the number of completed cell cycles. Our general understanding is that repressors and activators have opposing activity during ZGA that changes as cells of the embryo transition through successive cell cycles [72]. During early embryogenesis, the expression activity of maternal repressors is high at first but as these cells divide further their concentration is diluted in the embryo and therefore this activity eventually decreases. Activators, on the other hand, are initially present at low levels and may require successive cell cycles before attaining a threshold level for successful activation. The interplay between activators and repressors gives rise to the resulting levels of mRNA copies, and associated degree of heterogeneity observed in the embryo as it develops.
We identified genes with stable expression for the 4-cell to blastocyst stages, and these may form the core set of genes required for embryonic development to occur successfully. The stable genes were expressed at one of three distinct modes that spanned low, medium and high levels of expression, and this is interesting to note because it suggests that invariance of expression may be part of regulating genes required at all expression levels, and is not only confined to genes that are highly activated or need to be silenced. The assumption that variability diminishes inversely with the absolute level of transcription is not consistent with our observations. The first two modes of expression (low and medium) could be interpreted as genes that are turned off and on, respectively. Stable genes with low expression may be the result of a gene that has been switched off and their non-zero expression is the result of background signal observed in the RNA-seq experiment. Alternatively, these genes may be weakly expressed due to other mechanisms, e.g. from leaky transcription, where initiation of transcription occurs but not enough of the necessary components from the remaining machinery are available to follow through. The third expression mode (stable genes with high expression) may operate as accelerators or as higher level activators that support or function in roles related to the stable genes with medium expression. The fact that the pathways with significant enrichment for these genes overlap with those enriched for the medium-expressed (Fig 4), stable genes supports this interpretation.
Identifying genes based on their lack of variability has some parallel to developing catalogues of housekeeping genes, a concept that has been a fixture in molecular biology for nearly fifty years now [73]. While housekeeping genes represent a practical use in standardization and quality control, they have also garnered interest from a systems biology perspective because they represent a core, minimal set of genes necessary to support the cell. With growing technology, we continue to revise our definition of housekeeping genes; however, as far as we are aware, none of these approaches evaluate changes in variability directly. Our approach therefore has cross-over utility for the purpose of identifying potentially new housekeeping-type genes in single cells. Similarly, our approach may also have utility for defining stage-specific control genes that can be used to normalize or calibrate expression data of other genes with more heterogeneous expression.
We compared the stable genes that showed low, medium and high expression across all stages (see Fig 4 and S1 Table) with two sources of housekeeping genes, the top 10% of ubiquitously expressed genes collected by de Jonge et al. [32], and a list of housekeeping genes for RNA-seq data collected by Eisenberg et al. [73]. For both sets of housekeeping genes, we saw considerable overlap between the stable genes that had medium or high levels of expression compared to very few lowly-expressed stable genes (S9 and S10 Tables).
Across species, the property that divergence occurs at early and late, but not the middle part of development trajectories is a phenomenon referred to as the hourglass effect [74]. Other studies have shown that developmentally important genes demonstrate remarkable sequence similarity across a wide range of organisms, pointing to shared mechanisms that have been evolutionarily conserved. It has also been shown that gene expression control follows a similar pattern of conservation or stability, and experimental evidence to that effect has been most recently reported by Gerstein et al. [75]. Using transcriptomes of distant species generated by the ENCODE and modENCODE consortia, genes exhibiting developmental stage-specific expression showed diminished inter-species expression variability in the middle stages of development but were heterogeneous before and after. A recent paper by Liu et al. [76] also analyzed the same human embryo data set generated by Yan et al. [25] where they built co-expression gene modules and performed evolutionary conservation analysis on gene sequences as well as transcriptional regions upstream of each gene. They concluded that genes from stage-specific co-expressed modules have increased selective pressure at the 8-cell stage versus earlier stages (zygote to the 4-cell stage) and later (8-cell to the late blastocyst stage). In our study, the variability marker genes that were identified for the 8-cell and the morula stages also follow a pattern of diminished variability at these intermediate development stages (see Fig 5B–5D). This observation corresponds to Liu et al.’s discovery of increased selective pressure in co-expressed module genes at the 8-cell stage. Higher evolutionary pressure represents an increase of evolutionary constraint. Both independently derived observations point to an ‘hourglass’ in the preimplantation embryo, with lower evolutionary constraint and higher variability flanking a period of higher evolutionary constraint and lower variability.
The stage-specific variability markers we have identified show promise as potential determinants of embryo stage, however, they also shed light on understanding embryogenesis. It is likely that these markers are consistently expressed for all cells of the embryo because they provide or support critical functions for the developing embryo. This hypothesis could be readily tested in the context of model organisms like zebrafish and Drosophila. For human cells, these markers may be useful in future studies for identifying healthy embryos based on transcriptional states. Some of these applications may be in laboratory embryology where analysis of the least variable components of the transcriptome may serve as a scaffold to create predictive signatures of the developmental stage and potential of a biopsied embryo’s cells, especially for expression effects in genetically modified embryos. In conjunction with other existing criteria, information based on expression variability may eventually be used as a screen during routine day 3–5 biopsy for preimplantation embryos in IVF and other applications to improve pregnancy rates of women using these procedures.
We also see potential clinical applications in genetic diagnosis for reproductive potential for human embryo pre-implantation. Embryos that are otherwise chromosomally healthy are known to have a significant risk of implantation failure [77]. While there are many reasons underlying implantation failure, one significant cause may be due to innate errors in development due to disordered regulatory programs in the developing embryos that may result from genetic, epigenetic and environmental factors. As trophectoderm biopsies taken on day 5 (blastocysts) are regularly used as material for preimplantation genetic diagnosis (PGD), part of the biopsy material could be used to isolate cells and test the transcriptional regulatory state of the embryo in the trophectoderm for both expression level and SDC. This effect could be studied in a clinical trial through genetic fingerprinting analysis, determining the origin of a single birth following a double embryo transfer into the same patient. The expression state of cells in the implanting embryos versus the non-implanting embryos can be quantified and compared among a cohort, including the SDC of genes that represent important biomarkers for embryo reproductive potential.
There is value to understanding the variability of genes as a marker of cell state and regulatory potential [19]. Other variability markers may also unlock intriguing clues into how blastocysts are regulated (see S5 and S6 Figs). NDUFA12 helps to generate ATP for the cell via oxidative phosphorylation in mitochondria [78]. Mutations in this gene are often the cause of mitochondrial oxidative phosphorylation diseases, suggesting that NDUFA12 is critical for maintaining healthy human cells, and for blastocysts, NDUFA12 may play an important role in supplying these cells with sufficient energy [79]. LAMTOR1 is a membrane protein whose function is currently only known in the context of late endosomes and lysosomes, which are organelles that the cell relies upon to perform waste disposal. LAMTOR1 is anchored to the surface of these organelles, and through other intermediates can activate the mTORC1/MAPK signaling pathways that lead to cell growth and control of energy homeostasis [80]. At this point, it is difficult to say what impact LAMTOR1 has on blastocysts; however, a mouse study has shown that LAMTOR1 is responsible for proper epidermal development by regulation of lysosome-mediated catabolism [81]. It is possible that these processes may affect other tissues in the blastocysts more generally, and hence we see this gene being flagged as a variability marker.
Another marker of interest, ACTN4 is part of the α-actinin family, a set of related proteins that bind a wide range of molecules, including actin, to regulate a host of important processes in the cell [82]. α-actinin appears to operate in cell type-specific roles in the human body. Specifically, ACTN4 has been found to be ubiquitously expressed and distinct from the other α-actinins in that it interacts with transcription factors [83]. While its connection to blastocysts is also unclear at this point in time, ACTN4 has been detected in higher levels in cells with higher motility [84] and may be linked to metastatic processes in cancer [85, 86]. Motility in cells of the blastocyst may be important for blastocyst activation and trophectoderm motility [87] as previously observed in mouse embryos.
We observed that some of the variability markers that were identified using our approach overlapped with markers that can be found with standard approaches such as an ANOVA model. One example was EPCAM, a blastocyst variability marker identified also by ANOVA (see S5 and S6 Figs). EPCAM acts as a cell surface protein that is often used as a stem cell marker [88] because of its association with elevated levels in undifferentiated human embryonic stem cells [89]. Depending on the cell type that the expression profile was captured in, and other biological factors, there may be genes showing distinct changes in variability and/or average expression and both statistics are informative. Decreases in variability may be indicative of increased criticality, whereas increases in average expression may suggest a general trend of the cell population.
While transcription is fundamentally a stochastic phenomenon that affects all genes generically, the observation that some genes are more variably expressed than others, points to the existence of physical mechanisms underlying gene expression variability. These mechanisms can be grouped broadly into two major sources of variation. The first source refers to mechanisms embedded in the genome sequence, where the content of the promoter region and other regulatory sequences affects the degree of expression variability of a downstream gene. Nucleosome positioning, for example, is one such mechanism that can generate variable gene expression where a nucleosome typically occupies a region where transcription start sites and TATA elements exist [90]. Another example involves the interaction of long-range gene regulation where the expression of some genes require the interaction of other chromosomal segments. This mechanism can generate increased expression variability for genes that are regulated by this kind of nuclear architecture [91]. Other epigenetic mechanisms such as the presence of chromatin modifications have also been attributed to expression variability [92].
The second source of variation refers to the networks or interactions that a gene participates in that can influence its expression variability. We have seen through both experimental and simulation studies, that certain circuit formations, such as auto-regulatory motifs and feedback loops, can propagate signals to create different degrees of noise or variability [15, 16]. Pathways can be viewed as a collection of integrated circuits of genes, and the makeup of these circuits have been refined through evolutionary processes to become highly gene-specific and well-tuned. Therefore, depending on the placement of genes in their specific circuits, they may be more highly sensitive to exhibit fluctuations in their expression [93]. Related to this is the interaction with other molecules that have repressive or activator effects on a mRNA signal. For example, the interaction with microRNAs and shRNAs in the cell may deplete transcript levels of a gene however due to concentration, spatial regulation and other intracellular effects, this repression is likely to be observed non-uniformly in a cell population.
We used the RPKM (reads per kilobases per million reads) normalized RNA-seq data set generated by Yan et al. [25]. The data was downloaded from GEO using the accession number GSE36552. Whole-genome RNA-seq data from three embryos were available for the four embryo stages used in our study (4-cell, 8-cell, morula, blastocyst) except for morula where there were only two embryos. For morula and blastocyst, eight and ten cells were profiled respectively. In order to eliminate genes that may be affected by poor quality, genes with an RPKM-expression value ≥ 0.1 in at least 75% of cells from the same stage were retained. This resulted in 8105 genes that were used for further analysis.
To measure the inter-cellular expression variability of each gene g, we adopted the following statistic SDCg:
SDCg=1E∑j=1E1Nj∑i=1Nj(xij−x¯j)2
where xij is the expression level of gene g in the i-th cell of the j-th embryo for a total number of E embryos. For each embryo j, there are a total of Nj cells that have been profiled. x-j represents the average expression occurring in the j-th embryo. The SDC captures the SD of expression levels observed between cells belonging to the same embryo, that is then averaged across all embryos to give an overall measure of inter-cellular variability. We make the assumption that the embryos represent biological replicates, and therefore genes showing consistent inter-cellular variability between embryos are given higher weight by the SDC. This assumption is reasonable to make given that the embryos were screened extensively and only those passing stringent morphological criteria were included for RNA-sequencing (Fig 1).
As a contrast to the SDC, and as a means to evaluate inter-embryo variability of expression, we also computed the SDEg:
SDEg=1E∑j=1E[1Nj∑i=1Nj(xij−x¯j)2−1E∑j=1E1Nj∑i=1Nj(xij−x¯j)2]2.
All analysis and code is available upon email request.
For each stage, the SDC values were clustered using Normal mixture models to identify how many variability states were present. We used the Mclust function in the mclust R package (version 5.0.1) where the number of states was inferred by the function, and selected from possible values ranging from 1 to 9.
We used the Ingenuity Pathway Analysis database (Summer Release 2014) to identify significantly enriched pathways and processes for genes in different variability states at each stage. Significance was determined using a Fisher’s exact P-value that was adjusted for multiple testing correction using the Benjamini-Hochberg method [94]. Cut-offs were applied and are specified in the corresponding table legends.
We applied three simultaneous filters to identify stage-specific variability markers. To qualify as a marker of stage X, a gene had to have (1) the smallest SDC in stage X compared to all other stages; (2) a higher average expression in stage X compared to all other stages; (3) a statistically significant change in variability as detected by Levene’s test (adjusted P-values < 0.05).
For the knockdown experiments, a transgene-free human iPS cell line (clone C11 [95]) was transduced with lentiviral particles delivering pLKO.1-Puro family vectors constitutively expressing shRNAs against either control (EGFP) or HDDC2 (x2) mRNAs (see S11–S13 Tables). Puromycin selection (2μg/mL in the culture medium) was applied at days 4–6 after transduction, after which RNA was harvested for qPCR analysis.
For inducible up-regulation experiments, a human ES cell line inducibly expressing Cas9-VP64, an artificial transcriptional activator, was transduced with lentiviruses allowing for a puromycin-selectable expression of chimeric gRNAs targeting either unrelated gene or HDDC2. 2 days after the start of puromycin selection commencing at day 4 after transduction, induction of gene activation was started by addition of doxycycline at 1μg/mL. 2 days after that neuronal differentiation was started using a standard dual Smad inhibition protocol (on days 0 and 2 of differentiation, KOSR medium without FGF but supplemented with 5μM SB431542, 10 5μM Dorsomorphin and doxycycline at 2μg/ml). RNA was harvested for qPCR analysis on day 3.
We first identified genes that failed to meet statistical significance for Levene’s test (adjusted P-value > 0.05). These genes were clustered according to their SDC values into three groups that corresponded to low, medium and high levels of variability. We used the hclust function to perform agglomerative hierarchical clustering. Functional enrichment analysis was performed on the low variability group using IPA software. The set of stable genes were taken to be the cluster with the lowest level of variability (Cluster 1, S1 Fig).
All possible 4-cell combinations were elucidated for each stage, and the variability profile constructed for each combination.
|
10.1371/journal.pbio.1000612 | Mechanism of Neuroprotective Mitochondrial Remodeling by
PKA/AKAP1 | Mitochondrial shape is determined by fission and fusion reactions catalyzed by
large GTPases of the dynamin family, mutation of which can cause neurological
dysfunction. While fission-inducing protein phosphatases have been identified,
the identity of opposing kinase signaling complexes has remained elusive. We
report here that in both neurons and non-neuronal cells, cAMP elevation and
expression of an outer-mitochondrial membrane (OMM) targeted form of the protein
kinase A (PKA) catalytic subunit reshapes mitochondria into an interconnected
network. Conversely, OMM-targeting of the PKA inhibitor PKI promotes
mitochondrial fragmentation upstream of neuronal death. RNAi and overexpression
approaches identify mitochondria-localized A kinase anchoring protein 1 (AKAP1)
as a neuroprotective and mitochondria-stabilizing factor in vitro and in vivo.
According to epistasis studies with phosphorylation site-mutant dynamin-related
protein 1 (Drp1), inhibition of the mitochondrial fission enzyme through a
conserved PKA site is the principal mechanism by which cAMP and PKA/AKAP1
promote both mitochondrial elongation and neuronal survival. Phenocopied by a
mutation that slows GTP hydrolysis, Drp1 phosphorylation inhibits the
disassembly step of its catalytic cycle, accumulating large, slowly recycling
Drp1 oligomers at the OMM. Unopposed fusion then promotes formation of a
mitochondrial reticulum, which protects neurons from diverse insults.
| Mitochondria, the cellular powerhouse, are highly dynamic organelles shaped by
opposing fission and fusion events. Research over the past decade has identified
many components of the mitochondrial fission/fusion machinery and led to the
discovery that mutations in genes coding for these proteins can cause human
neurological diseases. While it is well established that mitochondrial shape
changes are intimately involved in cellular responses to environmental
stressors, we know very little about the mechanisms by which cells dynamically
adjust mitochondrial form and function. In this report, we show that the
scaffold protein AKAP1 brings the cAMP-dependent protein kinase PKA to the outer
mitochondrial membrane to protect neurons from injury. The PKA/AKAP1 complex
functions by inhibiting Drp1, an enzyme that mechanically constricts and
eventually severs mitochondria. Whereas active, dephosphorylated Drp1 rapidly
cycles between cytosol and mitochondria, phosphorylated Drp1 builds up in
inactive mitochondrial complexes, allowing mitochondria to fuse into a
neuroprotective reticulum. Our results suggest that altering the balance of
kinase and phosphatase activities at the outer mitochondrial membrane may
provide the basis for novel neuroprotective therapies.
| Opposing fission and fusion events determine the shape and interconnectivity of
mitochondria to regulate various aspects of their function, including ATP
production, Ca++ buffering, free radical homeostasis,
mitochondrial DNA inheritance, and organelle quality control. In addition,
fragmentation of neuronal mitochondria is necessary for their transport to and
proper development and function of synapses. Moreover, mitochondrial fission is an
early step in the mitochondrial apoptosis pathway, and inhibiting fission can block
or delay apoptosis in a variety of cell types, including neurons [1]–[3].
Fission and fusion processes are catalyzed by large GTPases of the dynamin
superfamily. Mitochondrial fission requires dynamin-related protein 1 (Drp1), which,
similar to the “pinchase” dynamin, is thought to mechanically constrict
and eventually sever mitochondria. Normally a largely cytosolic protein, Drp1 is
recruited to the outer mitochondrial membrane (OMM) by a poorly characterized
multiprotein complex that includes the transmembrane proteins Fis1 and Mff [4]–[6].
Mitochondrial fusion is carried out by the concerted action of OMM-anchored GTPases
(mitofusin-1 and -2 in vertebrates) and optic atrophy 1 (Opa1), a GTPase localized
to the intermembrane space [4]. A properly controlled fission/fusion balance appears to
be particularly critical in neurons, since mutations in mitochondrial fission/fusion
enzymes are responsible for common neurological disorders in humans [7]–[10]. All
mitochondria-restructuring enzymes are essential for mammalian development, as mice
that lack Drp1, Opa1, or either of the two mitofusins die during early embryogenesis
[11]–[14].
Our understanding of the signaling events that regulate this group of organelle
shaping GTPases is limited. Drp1, in particular, is subject to complex
posttranslational modification by ubiquitylation, sumoylation, nitrosylation, and
phosphorylation [15],[16]. Highly conserved among metazoan Drp1 orthologs, the two
characterized serine phosphorylation sites are located 20 amino acids apart,
bordering the C-terminal GTPase effector domain (GED). Since the numbering differs
between Drp1 orthologs and splice variants, we will refer to these sites by the
kinase first shown to target them, rather than their sequence number.
SerCDK (Ser616 in human, Ser635 in rat splice variant 1) is
phosphorylated by the cyclin-dependant kinase 1/cyclin B complex, leading to
fragmentation of the mitochondrial network during mitosis. Phosphorylation of
SerPKA (Ser637 in human, Ser656 in rat splice variant 1) is mediated
by both PKA and Ca2+/calmodulin-dependent protein kinase I (CaMKI).
Three laboratories, including ours, found that SerPKA phosphorylation by
PKA promotes mitochondrial elongation presumably through Drp1 inhibition [17]–[19], whereas a
fourth group reported the opposite effect upon targeting of the same site by CaMKI
[20].
We previously showed that Bβ2, a neuron-specific and OMM-targeted regulatory
subunit of protein phosphatase 2A (PP2A), sensitizes neurons to various insults by
promoting mitochondrial fission [21]. Here, we identify outer mitochondrial PKA as the
opposing fusion and survival promoting kinase. Evidence for a model is presented in
which the multifunctional scaffold protein AKAP1 recruits PKA to the OMM to
phosphorylate Drp1 at SerPKA. Phosphorylation traps Drp1 in large, slowly
recycling complexes, allowing mitochondria to fuse into a neuroprotective
reticulum.
In an effort to identify signal transduction pathways that alter the
mitochondrial fission/fusion balance, we investigated the effect of elevating
cAMP levels in PC12 pheochromocytoma cells and primary hippocampal neurons.
Mitochondria were visualized by transfection with mitochondria-targeted GFP or
by live-staining with the fluorescent dye MitoTracker Red. Application of the
adenylate cyclase agonist forskolin either alone or in combination with the
phosphodiesterase type IV inhibitor rolipram resulted in the conversion of
mostly punctiform or short and stubby mitochondria to highly elongated and
interconnected organelles (Figure
1A–D). Mitochondrial shape changes were quantified by image
analysis [22]
as well as by blinded comparison to a set of reference images (length score 0 to
4; Figure
S1A; [21]). Both methods yielded comparable elongation measures
that correlated on a cell-by-cell basis (Figure S1B, C, and D). Mitochondrial
elongation was readily apparent within 30 min of forskolin addition and was
insensitive to inhibition of protein synthesis by cycloheximide, which together
suggests non-genomic actions of cAMP (Figure 1B). Mitochondrial elongation mediated
by forskolin or the membrane permeable cAMP analog cpt-cAMP lasted for at least
20 h in PC12 cells and in the soma of hippocampal neurons (Figure 1C,D).
To implicate PKA in the morphogenetic effects of cAMP, we established clonal PC12
cell lines that inducibly express the PKA inhibitory peptide PKI fused to GFP
and the outer mitochondrial anchor domain of MAS70p (omPKI). Cells with
inducible expression of an OMM-directed form of the protein phosphatase 1 (PP1)
modulator inhibitor-2 (omInh2) were analyzed for comparison (Figure 2A). Induction of omPKI
by doxycycline resulted in mitochondrial fragmentation, measured both by
subjective scoring and image analysis. Conversely, mitochondria decorated with
the PP1 inhibitor were significantly elongated compared to mitochondria of
uninduced cells (Figure
2B–D).
Given that mitochondrial fragmentation is associated with apoptotic and
non-apoptotic cell death [2],[3], we examined cell lines expressing
mitochondria-targeted PP1 and PKA inhibitors for their susceptibility to
apoptotic stressors. Inhibition of neither outer mitochondrial PP1 nor PKA
affected survival or growth of PC12 cells under basal conditions (Figure 2F and unpublished
data). In contrast, induction of omPKI was associated with increased sensitivity
to several apoptosis inducers, including H2O2 and
staurosporine (Figure 2E,F).
These results indicate that PKA activity at the mitochondrial surface opposes
fragmentation of this organelle and increases resistance to apoptosis
inducers.
AKAPs recruit the PKA holoenzyme to specific subcellular sites and substrates,
which is critical for the physiological actions of the kinase [23]. Of the
three AKAPs that have been localized to mitochondria, the OMM-anchored AKAP1
(also known as D-AKAP1, AKAP121, AKAP149) has previously been shown to enhance
survival in PC12 cells [24]. We expressed AKAP1-GFP in hippocampal neurons and
scored apoptosis either under basal conditions or 2 d after treatment with
rotenone, an inhibitor of complex 1 of the electron transport chain that is
commonly used as a chemical model of Parkinson's disease. Wild-type AKAP1,
but not a point mutant that cannot bind the PKA holoenzyme (AKAP1ΔPKA
= I310P, L316P; Figure S2A), had a potent neuroprotective
effect under both conditions (Figure 3A). Conversely, knockdown of endogenous AKAP1 with two
different shRNAs (Figure S2B,C) dramatically amplified both
basal and rotenone-induced neuronal death (Figure 3A). Additionally, as we observed in
stressed PC12 cells, inhibiting outer mitochondrial PKA by expressing omPKI
resulted in increased basal apoptosis in hippocampal neurons. Thus, AKAP1
enhances neuronal survival by recruiting PKA to the OMM. Our finding that AKAP1
and outer mitochondrial PKA are critical for neuronal viability contrasts with
the mild phenotypes seen in unconditional AKAP1 knockout mice [25]. It is
possible that other mitochondrial AKAPs may be able to compensate if AKAP1 is
never expressed.
To investigate whether the pro-survival effects of PKA/AKAP1 are associated with
changes in mitochondrial morphology, TMRM-stained mitochondria of hippocampal
neurons were imaged 2–3 d after transfection. As previously reported in
other cell types [26], AKAP1-GFP colocalized perfectly with mitochondria in
hippocampal neurons (Figure
3B). Compared to OMM-targeted GFP or the AKAP1 mutant that cannot
recruit PKA, overexpression of wild-type AKAP1 increased mitochondrial length
scores in the somal mitochondria (Figure 3B,C). AKAP1 expression also increased mitochondrial form
factor in dendrites of hippocampal neurons (Figure
S3A,B). Targeting the PKA catalytic subunit directly to the OMM
(omPKA) resulted in even more striking mitochondrial fusion, often culminating
in perinuclear aggregation of mitochondria into a single mass. Conversely,
either expression of OMM-tethered PKI or silencing of AKAP1 with two shRNAs
induced significant somatic mitochondrial fragmentation as compared to
OMM-targeted GFP and nonsense shRNA controls (Figure 3C). AKAP1 knockdown also induced
mitochondrial fragmentation in dendrites of hippocampal neurons (Figure
S3A,B). To rule out off-target effects, we silenced the endogenous
protein in PC12 cells by co-expression of rat AKAP1-directed shRNAs with or
without plasmids encoding human AKAP1, which diverges at both shRNA target
sites. Human AKAP1 reversed the mitochondrial fragmentation induced by rat
AKAP1-directed shRNAs (Figure S4A,B), indicating that loss of
mitochondrial interconnectivity is due to specific silencing of the endogenous
protein.
Lentivirus expressing AKAP1-GFP was stereotaxically injected into the right
striatum and hippocampus of rats; the left hemisphere received injections with a
control lentivirus encoding mitochondrial (m)GFP (targeted via the N terminus of
cytochrome oxidase VIII). After 7–14 d, mitochondrial shape was assessed
by automated morphometry of confocal z-stacks from perfusion-fixed brain
sections (Figure 3D).
Compared to mGFP, lentiviral delivery of AKAP1 into the hippocampus and striatum
induced robust mitochondrial fusion as indicated by two independent metrics,
form factor, which measures elongation of individual mitochondria, and
cumulative area:perimeter ratio, which measures mitochondrial interconnectivity
(Figure 3E,F). Thus,
neuroprotection by PKA/AKAP1 is associated with mitochondrial elongation both in
vitro and in vivo.
We carried out epistasis experiments in hippocampal neurons to determine whether
PKA-mediated changes in mitochondrial shape and neuronal survival are causally
related. Initially, we showed that overexpression of the antiapoptotic Bcl2
protein prevented apoptosis induced by inhibiting PKA with omPKI but had no
effect on omPKI-induced mitochondrial fragmentation (Figure 4A,B). Similarly, Bcl2 rescued neurons
from apoptosis due to AKAP1 silencing without restoring normal mitochondrial
morphology (Figure S5). Therefore, mitochondrial fragmentation due to PKA
inhibition occurs either independently of or upstream of apoptosis.
To distinguish between these possibilities, we interfered with mitochondrial
fission by dominant-negative (K38A mutant) Drp1 or Fis1 knockdown. Blocking the
mitochondrial fission machinery fully restored viability of neurons transfected
with omPKI, demonstrating that PKA inhibition kills by inducing mitochondrial
fission. In contrast, AKAP1 silencing with two shRNAs compromised neuronal
viability in a manner that could not be reversed by inhibiting mitochondrial
fission (Figure
S5). Our data thus point to an essential function of AKAP1 beyond
localizing PKA to maintain mitochondrial integrity. In this regard, the AKAP1
gene gives rise to multiple splice variants, which have been localized to the
endoplasmatic reticulum and nuclear lamina, in addition to mitochondria. Also,
AKAP1 interacts not only with PKA but also with various other signaling enzymes
and mRNA [26].
The data thus far suggest a model in which the cAMP signaling through the outer
mitochondrial PKA/AKAP1 complex opposes the mitochondria fragmenting activity of
several protein phosphatases including PP2A [21], PP2B/calcineurin [18],[18], and likely
PP1 (Figure 2C,D).
Kinase/phosphates-regulated mitochondrial shape changes control neuronal
vulnerability to a variety of challenges (Figure 4C).
PKA/AKAP1 could maintain elongated mitochondria by inhibiting fission or
promoting fusion reactions. The mitochondrial fission enzyme Drp1 was previously
shown to be phosphorylated by PKA and phospho-mimetic substitution of the
targeted Ser residue causes mitochondrial elongation [18],[19]. We therefore examined the
mobility of Drp1 by fluorescence recovery after photobleaching (FRAP), replacing
endogenous Drp1 in PC12 cells with the GFP-tagged protein by a single-plasmid
RNAi/expression strategy (Figure S2D) [18]. Similar to the endogenous
fission enzyme, GFP-Drp1 was localized diffusely in the cytosol, as well as in
punctate structures on mitochondria (Figure 5A). Fluorescence recovery of
mitochondrial GFP-Drp1 could be well approximated by double exponential fits
(R2>0.99) and was largely independent of the
size of the bleached area, indicating that Drp1 mobility follows reaction-
rather than diffusion-limited kinetics [27]. Drp1 fluorescence
recovery was also shown to be unaffected by prior fragmentation of mitochondria
via protonophore treatment (Figure S6), indicating that Drp1 turnover
determines mitochondrial shape and not vice versa.
Silencing of AKAP1 increased the cytosolic pool at the expense of the
mitochondrial pool of GFP-Drp1 and accelerated fluorescence recovery of
mitochondrial GFP-Drp1. Activation of PKA by forskolin/rolipram treatment had
the opposite effect, increasing mitochondrial localization of GFP-Drp1 and
slowing its turnover (Figure
5B,C). Faster Drp1 recovery was reflected in an increase in the
plateau of the double exponential fit (mobile fraction) and a decrease in the
50% recovery time constant (t1/2); therefore, we expressed
Drp1 turnover as the ratio of the two curve-fitting parameters (Figure 5D).
In support of the imaging experiments, subcellular fractionation of GFP-Drp1
expressed in COS cells showed that 80% of the fission protein is
cytosolic, while 20% localizes to a heavy membrane fraction that includes
mitochondria (Figure 5E,F).
Co-expression of OMM-targeted PKA catalytic subunit induced phosphorylation of
Drp1 SerPKA mainly in the mitochondrial fraction as detected with a
phosphorylation state-specific antibody. Moreover, omPKA expression resulted in
a robust relocalization of Drp1 towards mitochondria (43%, Figure 5E,F). Together, these
data implicate Drp1 as a target of PKA/AKAP1 and suggest that phosphorylation
may slow the catalytic cycle of Drp1 by trapping the fission protein at the
mitochondrial membrane.
To explore the role of Drp1 phosphorylation in mitochondrial remodeling by PKA,
we replaced endogenous Drp1 in PC12 cells with mutant Drp1 that cannot be
phosphorylated by PKA (SerPKAAla,
SerPKA = Ser656 in rat splice variant 1) and
analyzed mitochondrial morphology at basal or forskolin/rolipram-stimulated cAMP
levels. Mutation of SerPKA in Drp1 rendered cells largely refractory
to mitochondrial elongation by cAMP (Figure 6A,B). To confirm the critical
importance of Drp1 SerPKA in the mitochondria-shaping effects of PKA
in non-neuronal cells, HeLa cells expressing either wild-type or
SPKAA-mutant Drp1 were also transfected with the catalytic subunit of
PKA, either with or without an OMM targeting sequence. Compared to control
transfected cells, both forms of PKA resulted in significant mitochondrial
elongation, with mitochondrial PKA being more effective than cytosolic PKA. The
non-phosphorylatable Drp1 mutant attenuated mitochondrial remodeling by either
form of PKA to a similar extent (Figure 6C,D).
We then explored the role of basal PKA activity in the maintenance of
mitochondrial shape. Incubation of HeLa cells with the PKA inhibitor H89
resulted in a protracted loss of phosphate from Drp1 SerPKA but did
not affect Drp1 phosphorylation at the neighboring SerCDK (Figure 6E). Paralleling the
time course of Drp1 SerPKA dephosphorylation, mitochondria in
wild-type GFP-Drp1-expressing cells underwent fragmentation upon endogenous PKA
inhibition. HeLa cells expressing pseudophosphorylated Drp1
(SerPKAAsp mutant) on the other hand were completely refractory to
H89-induced mitochondrial fission (Figure 6F,G).
To examine the influence of AKAP1 on Drp1 phosphorylation, COS cells were
co-transfected with Drp1 and either empty vector, wild-type AKAP1, or the
PKA-binding deficient AKAP1 mutant and stimulated with forskolin/rolipram.
Especially at lower forskolin/rolipram concentrations, recruitment of the PKA
holoenzyme to mitochondria via wild-type AKAP1 significantly enhanced Drp1
phosphorylation at SerPKA (Figure 7A,B). Essentially identical results
were obtained with neuronal PC12 cells (Figure S7). Rationalizing these findings, we
hypothesize that SerPKA of mitochondrial Drp1 is more exposed to
phosphorylation by PKA; alternatively, SerPKA-phosphorylated,
mitochondrial Drp1 may be relatively protected from cytosolic protein
phosphatases, such as calcineurin.
Next, we investigated the role of Drp1 phosphorylation in AKAP1-mediated
mitochondrial fusion. For this set of experiments, HeLa cells expressing empty
vector, wild-type, or PKA binding-deficient AKAP1 in addition to wild-type or
SPKAA-mutant Drp1 were treated with forskolin/rolipram to
activate PKA. As in neuronal cells, AKAP1 promoted mitochondrial network
formation that depended on recruitment of PKA.
Furthermore, PKA/AKAP1-induced mitochondrial elongation was nearly abolished when
endogenous Drp1 was replaced with the Ser→Ala mutant (Figure 7C,D).
We confirmed an epistatic relationship between AKAP1 and Drp1 SerPKA
in primary hippocampal neurons. Substitution of endogenous with
Ser→Ala-mutant Drp1 resulted in mitochondrial fragmentation, whereas the
Ser→Asp mutant caused mitochondrial elongation as assessed in dendrites
(form factor: Figure S8B; length: Figure S8C). Co-expression of wild-type AKAP1
lengthened dendritic mitochondria in wild-type GFP-Drp1 substituted neurons, but
not in neurons expressing either SerPKA mutant. Neurons were
unresponsive to AKAP1 ΔPKA transfection (Figure S8).
These experiments indicate that PKA/AKAP1 inhibits mitochondrial fission by
phosphorylating Drp1 at SerPKA.
AKAP1 has previously been reported to promote phosphorylation of the proapoptotic
Bcl2 family member Bad in PC12 cells, and Bad phosphorylation and cytosolic
sequestration was proposed to underlie the neuroprotective effect of PKA/AKAP1
[24].
To examine whether Drp1 may instead be the critical prosurvival substrate of
PKA/AKAP1, we challenged transiently transfected PC12 cells with the classical
apoptosis inducer staurosporine, a broad-spectrum kinase inhibitor previously
shown to dephosphorylate Drp1 at SerPKA
[18]. Compared
to AKAP1 ΔPKA or scrambled shRNA, overexpression of wild-type AKAP1
significantly attenuated apoptosis induced by staurosporine (0.5 µM for 24
h) when endogenous Drp1 was replaced with wild-type GFP-Drp1. Replacement with
Ser→Ala mutant Drp1, instead, negated the antiapoptotic effect of AKAP1
(Figure 7E,F).
Conversely, AKAP1 knockdown in wild-type GFP-Drp1 expressing cells moderately
increased the percentage with apoptotic nuclei compared to scrambled shRNA.
Pseudophosphorylated (S→D mutant) Drp1 overcame the proapoptotic effect of
AKAP1 shRNA, reducing staurosporine-induced apoptosis to levels comparable to
AKAP1 overexpression (Figure
7E,F). These results provide strong evidence that AKAP1 acts through
Drp1 (and not Bad) to inhibit the intrinsic apoptosis cascade in PC12 cells.
FRAP and subcellular fractionation experiments so far suggest that AKAP1 recruits
PKA to phosphorylate Drp1 at the OMM, which inhibits the membrane scission
activity of Drp1 by trapping the protein in slow-turnover complexes (Figure 5). We returned to FRAP
experiments in HeLa cells to establish a requirement for PKA binding to AKAP1
and SerPKA phosphorylation of Drp1 in this process. Compared to
AKAP1ΔPKA, co-expression of wild-type AKAP1 significantly slowed
fluorescence recovery of mitochondrial GFP-Drp1 (>3-fold increase in
t1/2; Figure
8A–C, Video S1). Furthermore, as seen in
forskolin/rolipram-stimulated PC12 cells, wild-type AKAP1 expression resulted in
a shift of GFP-Drp1 localization from cytosol to mitochondria (Figure 8A). In striking
contrast, PKA site-mutant (SPKAA) GFP-Drp1 was completely insensitive
to PKA/AKAP1 recruitment, displaying rapid recovery comparable to wild-type Drp1
cotransfected with AKAP1ΔPKA (Figure 8A,C).
Cytosolic Drp1 is a tetramer, which upon translocation to the OMM is thought to
oligomerize into spiral or ring-shaped superstructures. In analogy to dynamin,
GTP hydrolysis may trigger disassembly of Drp1 oligomers and concomitant
membrane fission [28]. In order to provide evidence that outer
mitochondrial PKA promotes SerPKA-dependent oligomerization of Drp1,
we performed crosslinking experiments with dithiobismaleimidoethane (DTME), a
membrane-permeant, reversible, and Cys-reactive crosslinker. Transfected COS
cells were crosslinked in vivo, and cell lysates were subjected to
ultracentrifugation (300,000 x g) to assess Drp1 assembly into large,
sedimentable complexes. Compared to co-expression of OMM-targeted GFP,
OMM-targeted PKA enhanced oligomerization of wild-type but not
SerPKA-mutant Drp1 (Figure 8D), providing a biochemical correlate for the slowly
recycling, mitochondrial Drp1 pool we observed in our FRAP experiments.
How does PKA/AKAP1-mediated phosphorylation stabilize Drp1 oligomers at the OMM?
SerPKA phosphorylation was previously shown to inhibit intrinsic
GTP hydrolysis by Drp1 [19]. Dnm1, the yeast ortholog of Drp1, assembles into
fission-competent complexes in its GTP-bound state, but assembly also stimulates
GTP hydrolysis, effectively limiting the size of Dnm1 oligomers at the OMM [28].
Therefore, an attractively simple hypothesis compatible with our observations is
that PKA/AKAP1 phosphorylation of Drp1 attenuates GTP hydrolysis, thereby
allowing Drp1 oligomers to grow beyond a size that is compatible with membrane
remodeling.
This hypothesis predicts that GTPase-impaired Drp1 mutants should behave
similarly to phospho-SerPKA Drp1. We initially analyzed the widely
used dominant-negative K38A mutant of Drp1, which affects a critical Lys residue
in the nucleotide binding fold. However, Drp1 K38A was found to localize
exclusively in large spherical aggregates that only occasionally overlapped with
mitochondria and that did not exchange with one another as detectable by FRAP
analysis (unpublished data). Aiming for a milder phenotype, we targeted Thr55 in
the Drp1 GTPase domain (Figure
9A) based on a mutagenesis study of the related dynamin-2. Ser61, the
corresponding residue in dynamin-2, lies within the nucleotide binding pocket
but does not directly coordinate GTP [29]. Mutation of Ser61 to Asp
was shown to significantly lower the maximal rate of GTP hydrolysis
(Kcat) without affecting the Km or apparent affinity
of dynamin-2 for GTP [30].
To assess the effect of the T55D mutation on mitochondrial morphology, HeLa cells
were transfected with different ratios of wild-type and T55D-mutant GFP-Drp1
plasmid (replacing endogenous Drp1 by RNAi from the same plasmid). Drp1 T55D
expression resulted in a dose-dependent increase in mitochondrial length,
demonstrating that the mutation impairs Drp1's fission activity (Figure 9B,C). Remarkably,
GFP-Drp1 T55D also displayed a more pronounced, punctate mitochondrial
localization than wild-type Drp1 (Figure 9B). Quantitative analysis confirmed that GFP-Drp1
colocalization with mitochondria (Manders' coefficient) was positively
correlated with relative expression levels of Drp1 T55D, as well as with
mitochondrial length (Figure
9C). Further FRAP analysis demonstrated that the T55D substitution
dramatically attenuates mitochondrial Drp1 dynamics, again in a dose-dependent
manner, slowing turnover by up to 10-fold (Figure 9D,E; Video S2).
Similar to phosphorylation at SerPKA, the T55D mutation also
increased the propensity of Drp1 to form sedimentable oligomeric structures
after intact cell crosslinking (Figure 9F).
Therefore, inhibition of GTP hydrolysis (either by mutation of the GTPase domain
or by phosphorylation of SerPKA by PKA/AKAP1) leads to a stable
accumulation of Drp1 at mitochondria, likely by interfering with a disassembly
step that is required for mitochondrial fission.
In an effort to provide direct evidence that PKA phosphorylation slows the
turnover of mitochondrial Drp1 complexes, we imaged HeLa cells expressing
GFP-Drp1 and mitochondrial dsRed2 at high magnification (630×). Wild-type
GFP-Drp1 foci were mostly found to abut mitochondria, seemingly randomly
translocating along their length and pivoting around the organelle (Figure 10A, Video S3).
Mitochondrial fission events were observed only at sites of GFP-Drp1
accumulation, and GFP-Drp1 punctae frequently divided to segregate with the new
mitochondrial ends. When PKA was co-expressed, GFP-Drp1 punctae appeared larger
and less dynamic, and the frequency of mitochondrial fragmentation events
decreased. Foci formed by GTPase-impaired, T55D mutant Drp1 displayed a similar
albeit even more accentuated behavior, which correlated with the near absence of
fission events during the 1 h recording period (Figure 10A, Video
S3).
Movies were subjected to automated particle tracking analysis [31],
extracting average lifetimes of mitochondria-associated GFP-Drp1 punctae at
37°C. Wild-type Drp1 punctae could be tracked for an average of 3.5 min,
whereas co-expression of PKA or the T55D mutation increased the persistence of
GFP-Drp1 punctae to 5.4 and 6.7 min, respectively (Figure 10B).
This report establishes a role for outer mitochondrial PKA and, in particular, the
PKA/AKAP1 complex in the maintenance of mitochondrial integrity and the protection
from neuronal injury. The importance of cAMP/PKA signaling in cell survival is well
documented [26]. Phosphorylation and inactivation of Bad, a pro-apoptotic
Bcl2-family protein, has been put forward as one of the critical survival promoting
substrates of mitochondria-localized PKA [32],[33]. While AKAP1 expression in PC12
cells was shown to increase Bad phosphorylation at multiple sites [24], our results
indicate that Bad phosphorylation does not significantly contribute to the
anti-apoptotic function of AKAP1. Instead, we present evidence that PKA targeting
via AKAP1 opposes cell death mainly by gating Drp1-dependent mitochondrial fission.
Specifically, inhibition of endogenous PKA via OMM-targeted PKI leads to
mitochondrial fragmentation and sensitizes neurons to pro-apoptotic stimuli, both of
which are reversed by blocking the mitochondrial fission machinery. Conversely,
recruiting PKA to mitochondria via expression of AKAP1 resulted in mitochondrial
elongation in cell culture and in vivo and protected hippocampal neurons from
rotenone toxicity. Both mitochondrial elongation and survival enhancement by AKAP1
required PKA anchoring and SerPKA-phosphorylatable Drp1, indicating a
critical role for the PKA-Drp1 axis.
However, our results do not rule out the possibility that PKA may cause mitochondrial
elongation by promoting mitochondrial fusion events in addition to inhibiting
mitochondrial division. For instance, a recent report showed that forskolin can
stimulate mitochondrial fusion in a cell-free assay [34].
AKAP1 is a large, multifunctional adaptor protein with several splice variants and a
highly conserved N-terminal transmembrane domain that acts as a mitochondrial
targeting sequence [35]. Besides localizing the PKA holoenzyme, AKAP1 also
interacts with the tyrosine phosphatase PTPD1, and through PTPD1 with the tyrosine
kinase Src, as well as two Ser/Thr phosphatases, PP1 and PP2B [36]–[38]. The C-terminus of AKAP1
contains Tudor and KH RNA binding domains, which were suggested to localize
nucleus-derived mRNAs encoding mitochondrial proteins close to their destination
[39].
A recent study described nuclear aggregation of mitochondria upon overexpressing an
N-terminal fragment of AKAP1 containing the PKA binding domain. The authors did not,
however, investigate whether this effect was PKA dependent and instead attributed
the phenotype to a lack of RNA binding to the missing C terminus [37]. In the present
study, we consistently observed elongation but rarely nuclear aggregation of
mitochondria, regardless of whether full-length or C-terminally truncated AKAP1
(residues 1–524) was expressed. High levels of AKAP1 overexpression did
sometimes induce nuclear aggregation of mitochondria, which may therefore be
secondary to exaggerated fusion of the organelle. Mitochondrial remodeling depended
on an intact PKA binding domain and was phenocopied and reversed by direct OMM
tethering of PKA and PKI, respectively. Thus, PKA targeting is both necessary and
sufficient for AKAP1-dependent regulation of mitochondrial morphogenesis.
Both the PKA and the PTPD1/Src interaction domains are important for maintenance of
mitochondrial membrane potential by AKAP1 [40]. Hence, either mitochondrial
recruitment of PTPD1/Src or an as yet undefined structural role of AKAP1 may explain
why mitochondrial fission inhibitors rescue hippocampal neurons from omPKI
expression but not from AKAP1 knockdown.
Consistent with an essential role for AKAP1 in neuronal survival, a recent report
demonstrated that ischemia induces expression of the E3 ubiquitin ligase Seven
In-Absentia Homolog 2 (Siah2), which targets AKAP1 for rapid proteasomal degradation
[41]. Our
study predicts that the hypoxia-induced loss of PKA anchoring at the OMM leads to
disinhibition of Drp1 and contributes to the massive mitochondrial fragmentation
that is a hallmark of ischemic brain injury [42].
We have identified a conserved PKA phosphorylation site in the GTPase effector domain
of Drp1 as the principal mediator of PKA/AKAP1-induced mitochondrial remodeling.
AKAP1-mediated redistribution of PKA was shown to augment Drp1 SerPKA
phosphorylation, and mitochondria of cells expressing
SerPKAAla-substituted Drp1 were unresponsive to cAMP and PKA/AKAP1. As to
a mechanism, Drp1 activation via AKAP1 silencing was associated with accelerated
cycling of the fission enzyme between cytosolic and mitochondrial pools. Conversely,
Drp1 inhibition via cAMP or PKA recruitment or overexpression resulted in the
accumulation of stable Drp1 oligomers at mitochondria and in an extension of the
lifetime of mitochondrial Drp1 foci. Because SerPKA phosphorylation
decreased the Kcat of GTP hydrolysis and because a mutation that
stabilizes the GTP bound form of Drp1 mimicked the effects of PKA phosphorylation on
localization and dynamics of the fission enzyme, modulation of Drp1's GTP cycle
emerges as a probable mechanism for the mitochondria-stabilizing and neuroprotective
actions of PKA/AKAP1.
A previous FRAP study demonstrated that the apoptosis inducer staurosporine causes
accumulation of slowly recycling mitochondrial YFP-Drp1 complexes, which colocalize
with Bax and Bak. Since the arrest of Drp1 cycling occurs after mitochondria have
fragmented but before they release cytochrome C, this phenomenon may be related to
the proapoptotic christae remodeling activity of Drp1 [43]. Given that the pan-kinase
inhibitor staurosporine actually inhibits Drp1 phosphorylation at SerPKA
[18],
mitochondrial accumulation of Drp1 during apoptosis likely occurs by a mechanism
distinct from the one reported here, such as Drp1 sumoylation ([43],[44], but see [45]).
Seemingly at odds with our findings, another study previously suggested that
calcineurin-mediated dephosphorylation of Drp1 at SerPKA promotes
translocation of the fission enzyme to mitochondria, a conclusion largely based on
overexpression of phosphorylation site-mutant Drp1 [17]. Confirming and extending
the findings of that report, we found that pseudophosphorylated
(SPKAD-mutant) GFP-Drp1 partitions mostly with the cytosolic fraction
(>90%), oligomerizes less readily than wild-type Drp1, and only
infrequently forms mitochondrial punctae (unpublished data), which is essentially
opposite to the phenotype of Drp1 phosphorylated by PKA. Because Asp substitution of
Drp1 SerPKA at most incompletely reproduces the inhibitory effect of
SerPKA phosphorylation on in vitro GTP hydrolysis [17]–[19], we propose that
the supposedly phosphomimetic substitution of SerPKA with an acidic
residue locks Drp1 into a partially inhibited state, arresting the enzyme at a
different stage of its subcellular translocation cycle.
Enhanced colocalization or cofractionation of Drp1 with mitochondria has previously
been interpreted as evidence for Drp1 activation (e.g. [17],[20],[46]). Our data and those of Zunino et
al. [44] argue
that Drp1 regulation is more complex, in that mitochondria-associated pools of Drp1
may sometimes be inactive. Similar considerations apply to higher order oligomeric
assembly of Drp1, which is clearly required for its function as a mechanoenzyme
[28]. The
crosslinking and particle tracking data presented here indicate that excessive
oligomerization of Drp1 into particles unable to constrict and sever mitochondria
occurs as a consequence of SerPKA phosphorylation or mutation of the
GTPase domain.
In support of a model in which PKA/AKAP1 fuses mitochondria by accumulating Drp1 in
inactive superstructures (Figure
10C), recent studies on the mechanism of action of dynamin suggest that
the endocytosis motor assembles into relatively short oligomers (3 to 4 rungs of a
spiral) before GTP hydrolysis-driven disassembly leads to membrane destabilization
and scission. In contrast, disassembly of larger dynamin oligomers (assembled in the
absence of GTP) does not effectively mediate membrane scission [47],[48].
Outer mitochondrial PKA-induced super-oligomerization of Drp1 could inhibit cell
death by several, non-mutually exclusive mechanisms. For instance, mitochondrial
networks resulting from unopposed fusion can sustain higher metabolic activity [49], are relatively
resistant to Bax insertion and cytochrome C release [50], and may also be more effective
at sequestering cytotoxic calcium and reactive oxygen species [51],[52]. More directly, PKA-mediated
depletion of the cytosolic Drp1 pool could potentially interfere with pathological
Drp1 activation by sumoylation [43] and nitrosylation [15] and compete with Drp1 recruitment
into Bax/Bak positive foci during apoptosis [53]. The interplay between
multi-site phosphorylation and other posttranslational modifications of Drp1 in the
regulation of mitochondrial homeostasis and cell death is undoubtedly complex and
will require further attention.
The following antibodies were used: rabbit anti-GFP (ab290, Abcam), mouse
IgG1 anti-MTCO2 (cytochrome oxidase subunit II, Neomarkers),
rabbit anti-ERK (Santa Cruz), mouse anti-phospho-SerPKA Drp1 [22], rabbit
anti-phospho-SerCDK Drp1 (Ser616, Cell Signaling), rabbit
anti-LacZ and mouse IgG2a anti-V5 epitope tag (Invitrogen), mouse
anti-neurofilament (2H3, Developmental Studies Hybridoma Bank, Iowa City), and
mouse anti-MAP2B (BD Transduction Laboratories). For immunofluorescence
staining, Alexa fluorophore-coupled secondary antibodies were purchased from
Invitrogen. Infrared fluorophore-coupled secondary antibodies for quantitative
immunoblot analysis were purchased from LI-COR Biosciences (Lincoln, NE).
The core domain of rat AKAP1 that is present in all published splice variants
(residues 1–524) was isolated by reverse transcriptase PCR and fused to
the N terminus of EGFP. The PKA-binding deficient mutant I310P, L316P was
generated by mutagenesis according to the QuikChange protocol (Stratagene).
N-terminally GFP tagged PKA, PKI [54], and PfARP32–239
[55] were
modified by the addition of N-terminal OMM targeting sequences: hexokinase
I1–30 for PKA and MAS70p1–29 for PKI. For
the V5-tagged AKAP1 constructs, the AKAP1 cDNA was excised from the GFP
constructs with BglII and SalI and ligated into the pcDNA3.1HisV5 vector
digested with BamHI and XhoI. AKAP1 and Fis1 were silenced by H1-promoter-driven
expression of shRNAs [56] (pSUPER plasmid [57] or lentivirus [58]).
19 b target sites in the mRNAs were (numbering relative to translation start
site in rat mRNAs): AKAP1/#1: 780–788, AKAP1/#4: 740–758, Fis1/#3:
315–333, Fis1/#4: 421–439. The nonspecific control shRNA had a
similar base composition but was randomized for no more than 14 consecutive
matches to any mammalian mRNA. The Fis1 and control shRNAs were described
previously [21]. Drp1 expression plasmids encoded rat splice variant 1
with an N-terminal EGFP tag under CMV promoter control, as well as Drp1-directed
shRNA driven by the H1 promoter [18]. GFP-Drp1 was rendered RNAi resistant and coding
mutations were incorporated by site-directed mutagenesis [18]. All Drp1 transfection
experiments in this article involved concomitant silencing of the endogenous
protein.
For analysis of Drp1 phosphorylation, COS cells were cotransfected with
GFP-tagged AKAP1 and Drp1 at 2∶3 plasmid mass ratios using Lipofectamine
2000. After 24 h, cells were stimulated with various concentrations of
forskolin/rolipram for 45 min, lysed in SDS sample buffer containing 2 mM EDTA
and 1 µM microcystin, and sonicated with a probe tip to shear DNA. For
AKAP1 immunoprecipitation, PC12 cells were transfected with AKAP1 WT- or
AKAP1ΔPKA-GFP. After 48 h, AKAP1 was immunoprecipitated with GFP antibodies
and protein A-agarose beads essentially as described [59]. Protein samples were
resolved on 8% or 10% polyacrylamide gels and transferred to
nitrocellulose membrane, followed by antibody detection using a LI-COR Odyssey
infrared fluorescence scanner. Band intensities were quantified using the ImageJ
gel analysis macro set, normalizing to loading controls in the same lane.
COS cells were transfected with triple HA-tagged Drp1 and either omGFP or omPKA
expressing plasmids at 3∶1 mass ratio and cultured for 24 h. After a wash
with PBS, cells were incubated with HBSS containing up to 0.5 mM
dithiobismaleimidoethane (DTME, from freshly prepared 50× stocks in DMSO)
for 5 min at 37°C. The HBSS/DTME was removed and cells were lysed in buffer
containing 60 mM Tris pH 6.8 and 2% SDS. Insoluble debris was removed by
centrifugation (10 min at 14,000 x g) and cleared lysates (300 µl) were
layered onto 1 ml of a 300 mM sucrose cushion and subjected to
ultracentrifugation (30 min at 300,000 x g). Pellets were dissolved in SDS
sample buffer containing 5% β-mercaptoethanol to cleave the
crosslinker and analyzed by SDS-PAGE alongside an aliquot of the lysate prior to
ultracentrifugation.
Hippocampal neurons from E18 rat embryos [60] were cultured in Neurobasal
medium with B27 supplement (Gibco) and transduced with lentivirus or transfected
using LipofectAmine 2000 (0.15%, 2 µg/ml DNA) at 10–21 d in
vitro. For survival assays based on counting transfected neurons or apoptotic
nuclei, GFP- or β-galactosidase-expressing plasmids were cotransfected.
After 3 to 5 d, cultures were challenged with rotenone (400 nM continuous) and
fixed with 3.7% paraformaldehyde and processed for immunofluorescence
staining for the transfection marker (αGFP and αLacZ) and neurofilament
protein as a neuronal marker (2H3). To score survival, transfection
marker-positive neurons with intact processes were counted in quadruplicate
wells of a 24-well plate. Apoptosis was quantified as the percentage of
transfected neurons with condensed, irregular, or fragmented nuclei (labeled
with 1 µg/ml Hoechst 33342). All survival assays based on counting neurons
were performed blind to the experimental conditions.
HeLa cells, PC12 cells, and hippocampal neurons cultured on, respectively,
untreated, collagen-, and poly-D-lysine-coated, chambered No. 1 cover glasses
(20 mm2 chamber, Nalge Nunc) were infected with lentivirus or
transfected using LipofectAmine 2000 as above. Hippocampal cultures were
infected or transfected at 8 to 10 d in vitro and analyzed 3 to 5 d later. For
live cell imaging 24 to 96 h post-transfection, cells were incubated with 100 nM
TMRM for 30 min at 37°C to visualize mitochondria, and images through the
midplane of the soma were captured using a Zeiss LSM 510 laser-scanning confocal
microscope. For analysis of fixed cells, cultures were subjected to
immunofluorescence staining with antibodies to cytochrome oxidase II (HeLa cells
only) and GFP [22], and images were captured with a Leica
epifluorescence microscope.
Mitochondrial morphology was scored by reference image-based and software-based
methods. For the former, coded images were assigned scores from 0 to 4 by
comparison to a set of reference images with increasing degrees of mitochondrial
elongation and clustering (Figure S1A). For automated morphometry,
images were processed using ImageJ software and plugins, involving either
“rolling ball” background subtraction or deblurring by 2-D
deconvolution with a computed point spread function. Using a custom-written
ImageJ macro, processed images were converted to binary (black and white) images
by auto-thresholding, and mitochondrial particles were analyzed for length,
width, area (a), and perimeter (p) [22]. Metrics that reliably
reported the effects of manipulating components of the mitochondrial
fission/fusion machinery (e.g. wild-type or dominant-negative Drp1 expression,
Fis1 or Drp1 RNAi) included form factor (p2/(4π*a)) and
cumulative area:perimeter ratio (Σa/Σp). The form factor is reported as
an average of all particles in a region-of-interest (ROI), has a minimum value
of 1 (for perfect circles), and captures well the transition from punctiform to
elongated, complex shaped, but still isolated mitochondria. The cumulative
area:perimeter ratio is computed as the summed particle area in a ROI divided by
the summed particle perimeter (including the perimeter of enclosed spaces or
“holes”). This metric is particularly effective at detecting the
transition from elongated, isolated mitochondria to a reticular network of
interconnected mitochondria. Colocalization of GFP-Drp1 with mitochondria (Figure 9C) was quantified as
the Manders' coefficient using the JaCoP plug-in for ImageJ [61].
All animal procedures were approved and carried out according to the
Institutional Animal Care and Use Committee (IACUC) at the University of Iowa.
Adult male Lewis rats were anesthetized with 91/9.1 mg/kg of ketamine/xylazine
and placed in a stereotaxic apparatus. Concentrated stocks (∼108
particles/ml) of lentivirus (feline immunodeficiency virus), prepared by the
University of Iowa Viral Vector Core, were bilaterally injected into the
striatum and hippocampus using a 10 µl Hamilton syringe with a 30°
beveled 30 gauge needle. Virus expressing mitochondria-targeted GFP (via
residues 1–31 of cytochrome oxidase subunit VIII) and AKAP1-GFP were
delivered into the left and right hemisphere, respectively. The following
coordinates were used with the incisor bar at 3.3 (from bregma): Striatum,
AP = 0, ML = ±3.5,
DV = −4.5; Hippocampus,
AP = −4.0, ML = ±2.0,
DV = −3.5. Animals were sacrificed 7–14 d later
and transcardially perfused with 4% paraformaldehyde. 40 µm thick
coronal cryostat sections were processed for indirect immunofluorescence with
antibodies to GFP and MAP2B and counterstained for nuclei with TOPRO-3. Ten to
22 confocal z-sections of infected neurons, 0.8 µm apart, were captured
and analyzed for mitochondrial shape by digital morphometry as described
above.
Hela and PC12 cells were grown on collagen-coated, chambered No. 1 cover glasses
(Nunc, Thermo Fisher), transfected 24 to 48 h, and stained with MitoTracker Deep
Red 30 min prior to analysis. Cells with non-mitochondrial GFP-Drp1 aggregates,
a sign of overexpression, were excluded from the analysis. Using the 488 nm
laser line of a Zeiss LSM 510 confocal microscope, an approx. 5×5 µm
region of mitochondria-associated GFP-Drp1 was bleached and fluorescence
recovery was tracked by capturing images every 5 s for 5 min. Image stacks were
analyzed, and recovery curves normalized to pre-bleach intensity and corrected
for acquisition bleach [27] were approximated by single and double exponential
recovery equations using a custom-written ImageJ macro, yielding 50%
recovery time constant (t1/2) and mobile fraction (mFx). Drp1
dynamics followed double exponential recovery kinetics with
R2 values of generally greater than 0.99.
Control experiments, in which bleach recovery of cytosolic and mitochondrial
Drp1 was tracked separately by masking the GFP channel with the MitoTracker
channel, showed that mitochondrial masking does not significantly affect the
results (unpublished data).
HeLa cells cotransfected with GFP-Drp1, dsRed2/mito, and either empty vector or
PKA catalytic subunit at 15∶4∶1 plasmid ratios were subjected to
time lapse imaging using the 63× objective of a motorized Leica AF6000
epifluorescence microscope under temperature (37°C) and CO2
(5%) control. Frames were captured at 30 s intervals for 60 to 90 min,
contrast-enhanced (contrast-limited adaptive histogram equalization
[CLAHE] plugin for ImageJ), and analyzed with the Particle Tracker
plugin for ImageJ [31]. Text files containing particle trajectories were
parsed with a custom-written macro to obtain mean particle lifetimes and
lifetime histograms.
COS cells expressing GFP-Drp1 with or without omPKA were permeabilized in 0.5
mg/ml digitonin, 100 mM KCl, 20 mM HEPES pH 7.4, 1 mM EDTA, 1 mM EGTA, 1 mM
benzamidine, 5 µg/ml leupeptin, 1 mM dithiotreitol (DTT), and 1 mM
phenylmethylsulfonyl fluoride (PMSF). Cytosolic and heavy membrane fractions
were prepared by centrifugation (10 min, 20,000 x g, 4°C).
Data were analyzed by Student's t test (two-tailed) for
single comparisons and by one-way analysis of variance (ANOVA) followed by
pairwise Bonferroni post hoc tests for multiple comparisons.
|
Subsets and Splits