| name | esm2-protein-embeddings |
| description | Extract reproducible ESM-2 protein embeddings, residue representations, likelihoods, or zero-shot mutation scores. Use when a frozen protein language model is needed as a feature extractor or baseline. |
| license | MIT |
ESM-2 Protein Embeddings
Gate
Ask once before installing a pinned ESM/Transformers revision, downloading model
weights, or using GPU compute. Keep sequences local unless remote inference is
explicitly approved. Follow $cx-compute-environment.
Workflow
- Define the prediction unit, exact sequence/construct, downstream task, and
inference-time information. Deduplicate and cluster sequences before splitting.
- Pin model/checkpoint revision, tokenizer, representation layer, max length,
truncation/chunk overlap, pooling, special-token handling, batch size, device,
precision, and code revision.
- Smoke-test one sequence alone and in a mixed batch; require invariant output
within numerical tolerance. Extract into
artifacts/<run-id>/esm2/ with
sequence hashes, embeddings/scores, dimensions, masks, logs, and failures.
- For supervised work, fit all transforms and heads on training data only.
Compare nearest-homolog, composition, k-mer/profile, and simple linear
baselines; report performance versus nearest-train identity.
- Record with
$science-provenance; review with $science-review.
Boundaries
- Pin the pretraining/model release and acknowledge database-cutoff leakage.
- Embedding similarity or mutation likelihood is not functional validation.