ESM2 Embeddings and ESMFold
Usage Notes
- Mean pooling: pool over all residues → single vector per protein (sequence-level tasks)
- Per-residue: one vector per AA → residue-level tasks (variant scoring, annotation)
- Layer selection: last/second-to-last for most tasks; intermediate layers can improve structural tasks
- ESMFold vs AF2: ESMFold ~60× faster (no MSA), but less accurate for proteins with many homologs — use for rapid screening, AF2 for final analysis
[Description truncada. Veja o README completo no GitHub.]