esmc-protein-modeling

name	esmc-protein-modeling
description	Run released Biohub ESMC protein language models for representations, masked likelihoods, sequence scoring, mutation analysis, or sparse-autoencoder features. Use as the current ESMC alternative to ESM-2 when model scale, revision, and inference boundaries are explicit.
license	MIT

ESMC Protein Modeling

For a concrete dataset, use $cx-modeling-problem-execution to continue through execution and downstream evaluation.

Gate and workflow

1. Ask once before installing a pinned Biohub esm/Transformers revision, downloading ESMC/SAE weights, using GPU compute, Hugging Face authentication, or sending sequences to Biohub. Never expose tokens or private sequences.
2. Define the downstream task and split before extraction. Pin model/revision, tokenizer, layer/SAE, pooling, length/chunking, masks, precision, and device.
3. Smoke-test one sequence alone and in a mixed batch. Require padding/batch invariance within tolerance, then write results to artifacts/<run-id>/esmc/ with hashes, embeddings/logits/features, masks, configs, environment, logs, and failures.
4. Fit transforms and prediction heads on training data only. Compare ESM-2, nearest-homolog, composition/profile, and simple linear baselines; report by nearest-train identity and applicability domain.
5. Record with $science-provenance; review with $science-review.

Boundaries

• Likelihood, an embedding, or an SAE interpretation is not measured function.
• Record the pretraining/database cutoff and audit homolog or retrieval leakage.