medsci-agent: Skill for OpenCode

ESM Protein Language Models

Overview

ESM (Evolutionary Scale Modeling) is Meta AI's family of protein language models trained on millions of protein sequences. ESM-2 provides per-residue and per-sequence embeddings. ESMFold predicts 3D structure from sequence alone.

Key Models

• ESM-2: Embedding model (8M to 15B params). Use esm2_t33_650M_UR50D as default.
• ESMFold: Single-sequence structure prediction (no MSA needed).

Usage Patterns

import torch, esm

model, alphabet = esm.pretrained.esm2_t33_650M_UR50D()
batch_converter = alphabet.get_batch_converter()
model.eval()

data = [("protein1", "MKTLLILAVL")]
batch_labels, batch_strs, batch_tokens = batch_converter(data)

with torch.no_grad():
    results = model(batch_tokens, repr_layers=[33], return_contacts=True)

embeddings = results["representations"][33]  # (batch, seq_len, 1280)
contact_map = results["contacts"]            # predicted contacts

ESMFold Structure Prediction

model = esm.pretrained.esmfold_v1()
model.eval()
with torch.no_grad():
    output = model.infer_pdb("MKTLLILAVL")
# output is a PDB-format string

Key Details

• Input sequences use standard single-letter amino acid codes.
• Maximum sequence length ~1024 residues for ESMFold; ESM-2 handles longer.
• Embeddings from final layer are most informative for downstream tasks.
• Contact prediction uses attention maps; symmetric and valid for i-j where |i-j| >= 6.
• Install: pip install fair-esm.