| name | esmfold2-structure-prediction |
| description | Run Biohub ESMFold2 locally from the released Hugging Face weights or through the approved Biohub Platform. Use for all-atom prediction of proteins, complexes, DNA, modifications, and ligands with optional MSA input and diffusion sampling; distinct from the legacy ESMFold v1 model. |
| license | MIT |
ESMFold2 Structure Prediction
For a concrete input, follow $cx-modeling-problem-execution and continue through
execution rather than returning setup instructions.
Gate and setup
Ask once before installing a pinned Biohub esm commit/Transformers revision,
downloading biohub/ESMFold2 and ESMC weights, using GPU compute, or sending
inputs to https://biohub.ai. Keep private sequences local unless remote transfer
is explicitly approved. Resolve an immutable commit from the authoritative
https://github.com/Biohub/esm repository and immutable Hugging Face revisions
for biohub/ESMFold2 and its ESMC dependency before approval; never install main.
Workflow
- Define protein/DNA entities, ligand CCDs, modifications, stoichiometry, MSA
source, cofactors, metal ions, and expected complex. Validate IDs and
residue/atom mappings. For ATP or another metal-dependent ligand, do not imply
Mg-bound chemistry when Mg is absent; ask or explicitly model/report the
chosen CCD and cofactor state.
- Record ESMFold2/ESMC weight revisions and checksums, input builder version,
loops, diffusion steps/samples, seed, device, precision, and remote model ID
when applicable.
- Smoke-test a short official example. Run the real input into
artifacts/<run-id>/esmfold2/; retain input objects, optional MSA, mmCIF,
pLDDT/pTM/ipTM, logs, environment, timings, and all sampled structures.
- Validate the installed release's actual output schema before naming metrics.
Compare samples and inspect interface confidence, geometry, clashes, ligand
state, and MSA dependence. Route structure rendering to
$cx-pymol-visualize and geometry/contact annotation to
$cx-plip-interaction-analysis using a prespecified pocket cutoff and atom map.
Use an orthogonal predictor for decision-critical claims.
- Record with
$science-provenance; review with $science-review.
Boundaries
- ESMFold2 is not legacy ESMFold/
esmfold_v1; record the exact model ID.
- Confidence and lab validation reported for the model family do not validate a
new target, complex, binder, or affinity prediction.
- If the pinned local release cannot represent the requested entity or chemistry,
stop that path and route to
$cx-protenix-structure-prediction,
$cx-openfold3-structure-prediction, or $cx-boltz-structure-prediction
according to modality and license; do not silently drop the unsupported entity.
