Biohub and ESMFold2: esm 3.4.1.post1

Released support and access

The released SDK includes ESMFold2; a floating GitHub installation is unnecessary. Use the isolated environment in SKILL.md. Current Biohub ESMC6B and ESMFold2 weights are public with MIT model-card labels; ESMC also links third-party license notices. Availability of weights does not imply a small memory footprint or access to every hosted model.

ESMFold2 combines ESMC representations with all-atom diffusion prediction. It returns a molecular complex, unlike ESM3's ESMProtein structure track. Use result.complex.to_mmcif() for mixed polymers/ligands and preserve chain IDs, modifications and chemical components. A predicted static structure is not a binding-energy, activity or dynamics assay.

Hosted example

Illustrative; requires an authorized key and sends the input to Biohub.

import os
from esm.sdk import esmfold2_client
from esm.sdk.api import ESMProteinError, FoldingConfig
from esm.utils.structure.input_builder import ProteinInput, StructurePredictionInput

input_data = StructurePredictionInput(
    sequences=[ProteinInput(id="A", sequence="MPRTKEINDAGLIVHSPQWFYK")]
)
config = FoldingConfig(num_loops=20, num_sampling_steps=100, include_pae=True)
with esmfold2_client(model="esmfold2-fast-2026-05", url="https://biohub.ai",
                     token=os.environ["ESM_API_KEY"], request_timeout=300) as client:
    result = client.fold_all_atom(input_data, config=config)
if isinstance(result, ESMProteinError):
    raise result
with open("predicted.cif", "w") as handle:
    handle.write(result.complex.to_mmcif())

fold_all_atom posts /api/v1/fold_all_atom; settings are passed by keyword because the second positional argument of the synchronous method is model_name, not config. The released deserializer returns one MolecularComplexResult, despite an annotation allowing a list.

The response exposes plddt, ptm, iptm (mapped from interface_ptm), pae, and optional embeddings. Some values may be absent; inspect their shape and units for the actual result before summarizing. Chain/atom arrays may require mapping to polymer residues; do not assume every complex token is one amino acid. Keep low-confidence regions and alignment coverage visible. No numerical accuracy or confidence calibration was measured in this refresh.

MSA and configuration limits

• esmfold2-fast-2026-05 is the documented single-sequence hosted example and ignores supplied MSAs, with an SDK warning. Current constants also name esmfold2-2026-05; confirm account availability before using another model.
• ProteinInput.msa accepts an SDK MSA object or None, not a raw A3M string. Keep the query sequence, row alignment and chain assignment consistent. The client rejects an over-limit MSA for the MSA-enabled model; inspect the released ESMFOLD2_MAX_MSA_SEQS and current service limits rather than inventing a depth.
• FoldingConfig defaults to 20 loops and 100 sampling steps. Lower values trade computation for quality; they are not validated accuracy-equivalent settings. lm_mask_pct=None resolves to 0.1 for the fast model, 0.0 for the full model.
• Hosted include_distogram=True is rejected by the released client. Although include_pair_chains_iptm exists in FoldingConfig, the released fold_all_atom serializer does not send it. Do not promise pair-chain IPTM output through that path.
• FoldingConfig has no seed field. Record settings and returned identifiers; a local seed is not transmitted by this hosted API.

Local ESMFold2

The following is an official-API illustration, not an executed pretrained run:

from esm.models.esmfold2 import EsmFold2Model, ESMFold2InputBuilder

model = EsmFold2Model.from_pretrained("biohub/ESMFold2", device="cuda").eval()
result = ESMFold2InputBuilder().fold(
    model, input_data, num_loops=20, num_sampling_steps=100,
    num_diffusion_samples=1, seed=0,
)
with open("local_prediction.cif", "w") as handle:
    handle.write(result.complex.to_mmcif())

This needs large model weights and suitable GPU resources. Do not download them for a signature check. The repository also documents longer-sequence Fold-CP and an independent Transformers implementation; those deployment paths were not executed here and require their own dependency/parallelism review.

Complex input conventions

ProteinInput, DNAInput, RNAInput, and LigandInput belong in StructurePredictionInput.sequences. IDs must preserve chain identity. LigandInput(ccd=["SAH"], id="L") uses a list of CCD codes; a bare string is rejected. Use a CCD list or SMILES deliberately rather than providing ambiguous competing representations. Modification.position, covalent-bond residue indices and atom indices are 0-based; they do not share the 1-based ESM3 FunctionAnnotation convention. Validate molecule chemistry and all residue/atom mappings before inference. Prediction availability does not validate them.

Sources: official README, ESMFold2 card, folding SDK, input types, API configs.