ESMFold2

Biohub ESMFold2 all-atom structure prediction using the ESMC-6B language model.

How SubSeq Runs It

Jobs run /ref/bin/esmfold2_subseq.py in an A100 profile.
The wrapper is mounted from the ESMFold2 ref tree, while model weights and Hugging Face cache are loaded from the A100-local /localref mirror.
The model root is forced to /localref/models; --model-preset=fast is the default for normal FASTA jobs unless a preset is specified.
Network access is disabled for jobs, and output is forced to /outputs.
Each FASTA file is folded as one complex. Multiple records in one file become multiple protein chains.

Input

Upload one or more FASTA files under /inputs. A single-chain protein can be one record; a protein complex can be one FASTA file with multiple records.

>A
MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG

For richer ESMFold2 inputs, upload JSON with sequences entries for proteins, DNA, RNA, ligands, modifications, MSAs, pockets, distograms, or covalent bonds.

{
  "id": "protein_ligand",
  "sequences": [
    {"type": "protein", "id": "A", "sequence": "MKTAYIAKQRQISFVKSHFSRQDILDL"},
    {"type": "ligand", "id": "L", "smiles": "CCO"}
  ]
}

Example Arguments

-I=/inputs
--model-preset=fast
--num-sampling-steps=32
--num-diffusion-samples=1

Use -i=/inputs/example.fasta to fold one FASTA file explicitly, or -j=/inputs/complex.json for structured JSON.

--model-preset=full
--num-sampling-steps=50
--num-loops=3
--chunk-size=64
--esmc-precision=bf16

Use fast for typical single-sequence FASTA folding. Choose full for richer structured JSON inputs, especially when using MSAs, ligands, nucleic acids, modifications, pockets, distograms, or covalent bonds.

Arguments

Inputs

-I, --input-dir: fold every FASTA file in a directory. Default: /inputs.
-i, --input: fold one FASTA file as one complex.
-j, --input-json: fold one structured JSON file.
-J, --json-dir: fold every JSON file in a directory.
--sequence: fold one direct single-chain protein sequence.
--complex-id: set the output stem and complex id for --sequence or single FASTA jobs.

Model

--model-preset=full|fast: choose the full ESMFold2 model or ESMFold2-Fast. Default: fast.
--esmc-precision=bf16|fp32|fp8: precision for the ESMC language model. Default: bf16.
--chunk-size=<int>: structure-module chunk size. Default: 64; use 0 to disable chunking.
--tf32, --no-tf32: enable or disable TF32 matmul on NVIDIA GPUs. Default: enabled.

Sampling

--num-loops=<int>: number of recycling/refinement loops. Default: 3.
--num-sampling-steps=<int>: diffusion sampling steps. Default: 32.
--num-diffusion-samples=<int>: number of samples to generate per input. Default: 1.
--seed=<int>: random seed for reproducible sampling.
--noise-scale=<float>: override ESMFold2's diffusion noise scale.
--step-scale=<float>: override ESMFold2's diffusion step scale.
--max-inference-sigma=<float>: cap the maximum inference sigma used by the sampler.
--early-exit: enable ESMFold2's early-exit mode.

32 steps is the normal fast preset for routine FASTA jobs. Use 50 for full, MSA-backed jobs, or quality checks. Try 100-200 only as an expensive retry when confidence is borderline; prefer adding MSA/context before increasing steps. More diffusion samples are useful for ranking/diversity, not as a reliable fix for missing context.

Validation

--dry-run: parse and validate inputs without loading the model or folding.

Server-Owned Paths

--model-root is forced to /localref/models.
--output-dir is forced to /outputs.
--model-id and --allow-download are not user-controlled in SubSeq jobs.

Outputs

The wrapper writes mmCIF structures and JSON confidence summaries, for example /outputs/example.cif and /outputs/example.json.

The JSON summary includes plddt_mean, ptm, iptm, and the output CIF path. It does not include Boltz-specific fields such as confidence_score or complex_ipde.

Submit

Queue a run from New Job -> ESMFold2.