| name | esmfold-structure-prediction |
| description | Run the legacy public ESMFold v1 single-sequence structure model reproducibly for proteins or simple multimers. Use for the 2022 ESM-2-based model; use the separate ESMFold2 skill for Biohub's 2026 diffusion/all-atom model. |
| license | MIT |
ESMFold v1 Structure Prediction
This skill covers the legacy public esmfold_v1 model. Use
$cx-esmfold2-structure-prediction for the 2026 Biohub diffusion/all-atom model.
For a concrete input, follow $cx-modeling-problem-execution through execution.
Gate
Ask once before installing a pinned fair-esm/repository revision, downloading
weights, or using GPU compute. Prefer local inference; remote ESM Atlas folding
requires separate approval because it transmits the sequence. Follow
$cx-compute-environment.
Workflow
- Validate sequence alphabet, chain separators, construct, and length. Do not
silently truncate; plan chunking only when scientifically defensible.
- Pin repository/package,
esmfold_v1 checkpoint, weight checksum, recycles,
chunk size, batching, seed, device, and precision.
- Smoke-test one short sequence, then run into
artifacts/<run-id>/esmfold/. Retain FASTA, PDB outputs, logs, timings,
parameters, and failed sequences.
- Report pLDDT, chain/interface uncertainty, clashes, missing confidence
information, and consistency across settings. Compare against AlphaFold DB or
an MSA-based method when the claim depends on structural detail.
- Record with
$science-provenance; review with $science-review.
Boundaries
- Single-sequence speed does not imply equal accuracy to MSA/template methods.
- pLDDT is confidence, not experimental accuracy for a specific unseen target.
