Data Preprocessing
COMPLETE4-stage pipeline: UniProtKB acquisition β ESM-2 embedding extraction β GO annotation processing β MMseqs2 cluster-based splitting. All 2,620 proteins processed.
Downloads HIV-1 (taxid 11676) and M. tuberculosis (taxid 83332) protein sequences from UniProtKB/Swiss-Prot. Filters to reviewed entries only. Fetches Gene Ontology annotations (MF, BP, CC namespaces).
Loads facebook/esm2_t33_650M_UR50D from HuggingFace. Extracts per-residue embeddings (1280-dim) for all 2,620 proteins. Saves to HDF5 cache. Sequences >1,011 residues are truncated to model max length (1022 tokens).
Removes proteins with zero GO annotations (219 removed β 2,620 retained). Builds GO vocabulary of 354 most-frequent terms (min 5 proteins per term). Computes inverse-frequency class weights for imbalanced multilabel loss.
Clusters all proteins at 30% sequence identity to prevent data leakage. Cluster-level stratified split: train 1,833 / val 262 / test 525. Ensures no homologous sequences appear in both train and test.
| Field | Value | Notes |
|---|---|---|
| Total proteins | 2,620 | After removing proteins with 0 GO annotations |
| HIV-1 proteins | ~847 | UniProtKB Swiss-Prot, taxid 11676 |
| M. tuberculosis proteins | ~1,773 | UniProtKB Swiss-Prot, taxid 83332 |
| Proteins with PDB structure | 841 (32%) | Has CΞ± coordinates for physics loss |
| GO vocabulary size | 354 | Terms with β₯5 annotated proteins |
| Mean GO terms/protein | 4.69 | Sparse multilabel annotation |
| Split method | MMseqs2 30% | Cluster-level to prevent data leakage |
| Sequence length range | 30 β 1,011 aa | Median 280 aa, mean 318 aa |
| Embedding dimension | 1,280 | ESM-2 650M per-residue |