| ## data set specific | |
| dataset: GraphMutationDataset | |
| data_file_train: /share/vault/Users/gz2294/Data/DMS/ClinVar.HGMD.PrimateAI.syn/training.csv | |
| data_file_train_ddp_prefix: /share/vault/Users/gz2294/Data/DMS/ClinVar.HGMD.PrimateAI.syn/training | |
| data_file_test: /share/vault/Users/gz2294/Data/DMS/ClinVar.HGMD.PrimateAI.syn/testing.csv | |
| data_type: ClinVar | |
| loop: true # add self loop or not | |
| node_embedding_type: esm # esm, one-hot, one-hot-idx, or aa-5dim | |
| graph_type: af2 # af2 or 1d-neighbor | |
| add_plddt: true # add plddt or not | |
| add_conservation: true # add conservation or not | |
| add_position: true # add positional embeddings or not | |
| add_sidechain: true # add side chain or not | |
| use_cb: true | |
| loaded_msa: false | |
| add_msa: true # add msa or not | |
| add_dssp: true # add dssp or not | |
| alt_type: concat # concat or alt | |
| computed_graph: true | |
| max_len: 251 | |
| radius: 50 # radius for KNN graph, larger than curoff_upper | |
| ## model specific | |
| load_model: null | |
| model_class: PreMode_Star_CON | |
| model: equivariant-transformer-star2-softmax | |
| neighbor_embedding: true | |
| cutoff_lower: 0.0 # graph related | |
| cutoff_upper: 36.0 # graph related | |
| max_num_neighbors: 36 # graph related | |
| x_in_channels: 1313 # x input size, only used if different from x_channels, 1280 + 1 + 20 + 12 | |
| alt_projector: 2593 # alt input size, 1280 + 1 + 20 + 12 + 1280 | |
| x_in_embedding_type: Linear_gelu # x input embedding type, only used if x_in_channels is not None | |
| x_channels: 512 # x embedding size | |
| x_hidden_channels: 512 # x hidden size | |
| vec_in_channels: 35 # vector embedding size | |
| vec_channels: 32 # vector hidden size | |
| vec_hidden_channels: 512 # vector hidden size, must be equal to x_channels (why? go to model page) | |
| distance_influence: both | |
| share_kv: false | |
| num_heads: 16 # number of attention heads | |
| num_layers: 2 | |
| num_edge_attr: 444 # 1, from msa_contacts | |
| num_nodes: 1 | |
| num_rbf: 32 # number of radial basis functions, use a small size for quicker training | |
| rbf_type: expnormunlim | |
| trainable_rbf: true | |
| num_workers: 10 | |
| output_model: EquivariantBinaryClassificationStarPoolScalar | |
| reduce_op: mean | |
| output_dim: 1 | |
| activation: silu | |
| attn_activation: silu | |
| # aggr: mean # has to be mean because different protein sizes, removed and set to default (note previous default was add) | |
| drop_out: 0.6 | |
| ## training specific | |
| trainer_fn: PreMode_trainer | |
| seed: 0 | |
| lr: 1e-4 # important | |
| lr_factor: 0.8 # important | |
| weight_decay: 0.0 | |
| lr_min: 1e-6 # important | |
| lr_patience: 2 # important | |
| num_steps_update: 8 # important, how many steps before updating the model, use large number for large batch size | |
| lr_warmup_steps: 4000 # important | |
| batch_size: 8 | |
| ngpus: 4 | |
| num_epochs: 20 | |
| loss_fn: weighted_loss_pretrain | |
| data_split_fn: "" | |
| y_weight: 1.0 | |
| contrastive_loss_fn: null | |
| reset_train_dataloader_each_epoch: true | |
| test_size: null | |
| train_size: 0.95 | |
| val_size: 0.05 | |
| ## log specific | |
| num_save_epochs: 1 | |
| num_save_batches: 2000 # save every 1000 batches, this also control the validation frequency | |
| log_dir: /share/vault/Users/gz2294/PreMode.final/CHPs.v4.noPretrain.retrain.seed.0/ | |