ESM2 Protein Language Model

Prerequisites

How to run

First time? See Installation Guide to set up Modal and biomodals.

Option 1: Modal

cd biomodals
modal run modal_esm2_predict_masked.py \
  --input-faa sequences.fasta \
  --out-dir embeddings/

GPU: A10G (24GB) | Timeout: 300s default

Option 2: Python API (recommended)

import torch
import esm

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

# Process sequences
data = [("seq1", "MKTAYIAKQRQISFVK...")]
batch_labels, batch_strs, batch_tokens = batch_converter(data)

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

# Get embeddings
embeddings = results["representations"][33]

Key parameters

ESM2 Models

Output format

embeddings/
├── embeddings.npy       # (N, 1280) array
├── pll_scores.csv       # PLL for each sequence
└── metadata.json        # Sequence info

Sample output

Successful run

$ modal run modal_esm2_predict_masked.py --input-faa designs.fasta
[INFO] Loading ESM2-650M model...
[INFO] Processing 100 sequences...
[INFO] Computing pseudo-log-likelihood...

embeddings/pll_scores.csv:
sequence_id,pll,pll_normalized,length
design_0,-0.82,0.15,78
design_1,-0.95,0.08,85
design_2,-1.23,-0.12,72
...

Summary:
  Mean PLL: -0.91
  Sequences with PLL > 0: 42/100 (42%)

What good output looks like:

• PLL_normalized: > 0.0 (more natural-like)
• Embeddings shape: (N, 1280) for 650M model
• Higher PLL = more natural sequence

Decision tree

Should I use ESM2?
│
├─ What do you need?
│  ├─ Sequence plausibility score → ESM2 PLL ✓
│  ├─ Embeddings for clustering → ESM2 ✓
│  ├─ Variant effect prediction → ESM2 ✓
│  └─ Structure prediction → Use ESMFold
│
├─ What model size?
│  ├─ Fast screening → esm2_t12_35M
│  ├─ Standard use → esm2_t33_650M ✓
│  └─ Best quality → esm2_t36_3B
│
└─ Use case?
   ├─ QC filtering → PLL > 0.0 threshold
   ├─ Diversity analysis → Mean-pooled embeddings
   └─ Mutation scanning → Per-position log-odds

PLL interpretation

Typical performance

Throughput: ~100-200 sequences/minute with 650M model.

Verify

wc -l embeddings/pll_scores.csv  # Should match input + 1 (header)

Troubleshooting

OOM errors: Use smaller model or batch sequences Slow processing: Use esm2_t12_35M for speed Low PLL scores: May indicate unusual/designed sequences

Error interpretation

RuntimeError: CUDA out of memory

KeyError: representation

ValueError: sequence

Next: Structure prediction with

chai

boltz

protein-qc