MHC-PLM / ensemble.json
zmactep's picture
Release MHC-PLM model weights and metadata
f5551d6 verified
Raw History Blame Contribute Delete
4.06 kB
{
"schema_version": 1,
"name": "MHC-PLM core-selected ensemble",
"repo_id": "zmactep/MHC-PLM",
"architecture": "ESM-C plus peptide-to-MHC cross-attention and explicit 9-mer MIL",
"selection_scope": "core-disjoint-fold1",
"default_members": [
"stage3",
"600m",
"stage4",
"replica_s3333"
],
"best_single_member": "stage4",
"aggregation": "arithmetic mean of post-sigmoid member scores",
"inference_seed": 2026,
"input_limits": {
"max_peptide_length": 64,
"max_context_length": 16,
"max_mhc_length": 64,
"max_alleles": 14
},
"score_semantics": {
"ba": "normalized affinity in [0,1], higher is stronger; approximate IC50 nM = 50000^(1-score)",
"el": "presentation score in [0,1]; not a calibrated clinical probability"
},
"model_defaults": {
"backbone": "hf",
"backbone_dtype": "bf16",
"attention_implementation": "sdpa",
"interaction_dim": 256,
"interaction_heads": 8,
"interaction_layers": 2,
"core_length": 9,
"allow_reverse": true,
"dropout": 0.1,
"lora": true,
"lora_rank": 16,
"lora_alpha": 32,
"lora_dropout": 0.05,
"gradient_checkpointing": false,
"backbone_chunk_size": 256,
"pooling_temperature": 0.15
},
"members": [
{
"name": "stage3",
"description": "ESMC-300M main trajectory, stage 3",
"weight_path": "weights/stage3.safetensors",
"weight_sha256": "2ef7668d28ac2c451e210624b2f7dd46418790e9feb8c6d830c8ab1f5a640892",
"weight_size_bytes": 12349296,
"weight_partial": true,
"model": {
"model_name": "biohub/ESMC-300M",
"model_revision": "a59b831785f907e96e6a246b1d142bfb76df31ee"
}
},
{
"name": "600m",
"description": "independent ESMC-600M trajectory",
"weight_path": "weights/600m.safetensors",
"weight_sha256": "7e6dda37c80ae11dbe44954f442f0490e16bdb3c1f23aabec2867fed198dedda",
"weight_size_bytes": 13360248,
"weight_partial": true,
"model": {
"model_name": "biohub/ESMC-600M",
"model_revision": "a7e82012c83126b9eedb055fea9fa84b6c02f094"
}
},
{
"name": "stage4",
"description": "ESMC-300M low-learning-rate stage 4; best individual member",
"weight_path": "weights/stage4.safetensors",
"weight_sha256": "8a2efbdebfe10d2a723be25e2e342a5209b546f2ea3fed1dcd73bb599b499493",
"weight_size_bytes": 12349296,
"weight_partial": true,
"model": {
"model_name": "biohub/ESMC-300M",
"model_revision": "a59b831785f907e96e6a246b1d142bfb76df31ee"
}
},
{
"name": "replica_s3333",
"description": "independent ESMC-300M seed trajectory",
"weight_path": "weights/replica_s3333.safetensors",
"weight_sha256": "aedbaf249908fad38fd61643639042ce5a2030da70ac1c3d5cc586cdbe478237",
"weight_size_bytes": 12349296,
"weight_partial": true,
"model": {
"model_name": "biohub/ESMC-300M",
"model_revision": "a59b831785f907e96e6a246b1d142bfb76df31ee"
}
}
],
"metrics": {
"core_disjoint": {
"composite": 0.866341447431238,
"ba_examples": 12911,
"ba_auroc_500nm": 0.8795781333432586,
"ba_average_precision_500nm": 0.8432509682864474,
"ba_pearson_r": 0.7494314418958933,
"ba_spearman_r": 0.7513894503510444,
"ba_mse": 0.027961074615603626,
"el_examples": 572664,
"el_positive_fraction": 0.14436039279193315,
"el_auroc": 0.9633044656226264,
"el_average_precision": 0.8730517488631739
},
"full_fold": {
"composite": 0.8572583300612281,
"ba_examples": 21794,
"ba_auroc_500nm": 0.8881501373829886,
"ba_average_precision_500nm": 0.8529772425709462,
"ba_pearson_r": 0.7654025002504102,
"ba_spearman_r": 0.7655603005570821,
"ba_mse": 0.028142128247225746,
"el_examples": 1516156,
"el_positive_fraction": 0.08164859025060746,
"el_auroc": 0.9644079925097755,
"el_average_precision": 0.8110726901017381
}
}
}