ESMFold

Predict¶

Predict 3D structure for one or more protein chains (separated by colons) using ESMFold

from biolmai import BioLM
response = BioLM(
    entity="esmfold",
    action="predict",
    params={},
    items=[
      {
        "sequence": "GAMEDTQVAW"
      },
      {
        "sequence": "MKTIIALSYIFCLVFADYKDDDD:VLLPAGKQ"
      }
    ]
)
print(response)

curl -X POST https://biolm.ai/api/v3/esmfold/predict/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "sequence": "GAMEDTQVAW"
    },
    {
      "sequence": "MKTIIALSYIFCLVFADYKDDDD:VLLPAGKQ"
    }
  ]
}'

import requests

url = "https://biolm.ai/api/v3/esmfold/predict/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "items": [
        {
          "sequence": "GAMEDTQVAW"
        },
        {
          "sequence": "MKTIIALSYIFCLVFADYKDDDD:VLLPAGKQ"
        }
      ]
    }

response = requests.post(url, headers=headers, json=payload)
print(response.json())

library(httr)

url <- "https://biolm.ai/api/v3/esmfold/predict/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  items = list(
    list(
      sequence = "GAMEDTQVAW"
    ),
    list(
      sequence = "MKTIIALSYIFCLVFADYKDDDD:VLLPAGKQ"
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))

POST /api/v3/esmfold/predict/¶

Predict endpoint for ESMFold.

Request Headers:

• Content-Type – application/json

• Authorization – Token YOUR_API_KEY

Request

• params (object, optional) — Configuration parameters:

  • batch_size (int, default: 2) — Maximum number of items processed per request

  • max_sequence_len (int, default: 768) — Maximum number of amino acid residues per chain in a sequence

  • max_n_multimers (int, default: 4) — Maximum number of chains allowed in a multimer sequence

• items (array of objects, min: 1, max: 2) — Input sequences:

  • sequence (string, min length: 1, max length: 771, required) — Protein sequence using the extended amino acid alphabet with “:” as chain separator, allowing up to 3 non-consecutive “:” characters

Example request:

http Copy

POST /api/v3/esmfold/predict/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "items": [
    {
      "sequence": "GAMEDTQVAW"
    },
    {
      "sequence": "MKTIIALSYIFCLVFADYKDDDD:VLLPAGKQ"
    }
  ]
}

Status Codes:

• 200 OK – Successful response

• 400 Bad Request – Invalid input

• 500 Internal Server Error – Internal server error

Response

• results (array of objects) — One result per input item, in the order requested:

  • pdb (string) — Predicted protein structure in PDB format, including coordinates and standard PDB records

  • mean_plddt (float) — Mean predicted Local Distance Difference Test (pLDDT) confidence score over all residues

  • ptm (float) — Predicted Template Modeling (pTM) score estimating global fold accuracy

Example response:

http Copy

HTTP/1.1 200 OK
Content-Type: application/json

      {
  "results": [
    {
      "pdb": "PARENT N/A\nATOM      1  N   GLY A   1       2.704  13.046  14.934  1.00 67.45           N  \nATOM      2  CA  GLY A   1       1.649  12.125  15.325  1.00 67.60           C  \nATOM      3  C   GLY A   1 ... (truncated for documentation)",
      "mean_plddt": 64.428955078125,
      "ptm": 0.012728862464427948
    },
    {
      "pdb": "PARENT N/A\nATOM      1  N   MET A   1       1.328 -16.623 -16.398  1.00 42.52           N  \nATOM      2  CA  MET A   1       2.126 -17.137 -15.289  1.00 44.35           C  \nATOM      3  C   MET A   1 ... (truncated for documentation)",
      "mean_plddt": 47.16063690185547,
      "ptm": 0.10539554059505463
    }
  ]
}

Performance¶

• ESMFold runs inference on NVIDIA A10G GPUs with 4 vCPUs and 16 GB RAM, using mixed-precision computation, chunked attention, and optional CPU offloading to efficiently handle single- or multi-chain inputs up to 768 residues total.

• Inference speed is tuned for rapid turnaround: for a 384-residue monomer, network forward time is around 14 seconds on a single GPU, typically ~6× faster than a single AlphaFold2 model and >60× faster than AlphaFold2 on short sequences (<200 residues) when MSA search is considered.

• Structural accuracy is competitive with other single-sequence and MSA-based predictors: on CASP14 targets, ESMFold reaches mean LDDT ≈ 0.68 (vs. ≈ 0.85 for AlphaFold2 and ≈ 0.81 for RoseTTAFold). On the CAMEO benchmark, mean TM-scores are ≈ 0.90 (easy), 0.79 (medium), and 0.45 (hard), slightly below AlphaFold2 but close to RoseTTAFold on easy and medium targets.

• Confidence estimates (mean pLDDT and pTM) are well calibrated: mean pLDDT strongly correlates with realized accuracy (LDDT/TM-score), enabling downstream pipelines to filter or prioritize BioLM ESMFold predictions by quality when trading off against slower, more accurate models such as AlphaFold2 or against nanobody-specialized models such as NanobodyBuilder.

Applications¶

• Rapid single- and few-chain structure prediction in protein engineering pipelines, enabling teams to assess structural viability of designed variants in seconds without MSA generation; accelerates iterative design for therapeutics or industrial biocatalysts within the API limits of up to 4 chains and 768 residues total.

• High-throughput structural screening of protein variant or metagenomic libraries by ranking candidates using mean pLDDT and pTM scores from the API response; useful for prioritizing stable folds and well-packed cores for downstream experimental validation; less suitable for detailed modeling of large complexes or interfaces beyond four chains.

• Structural annotation of orphan or low-homology proteins in discovery programs, providing predicted 3D coordinates (PDB output) and confidence metrics directly from sequence; supports target selection and domain boundary assessment when experimental structures or deep MSAs are unavailable; accuracy decreases on very hard, novel folds, so predictions should be combined with additional evidence.

• Computational ranking of protein design candidates, combining ESMFold mean pLDDT and pTM with other design scores to focus wet-lab screening on structurally plausible designs; reduces cost by deprioritizing clearly misfolded or disordered variants; not recommended as the sole criterion for functional optimization, as activity and specificity still require additional modeling or experiments.

Limitations¶

• Maximum Sequence Length: The sequence input (a single chain or multiple chains separated by :) must not exceed 768 amino acids in total. Longer proteins must be truncated or split into separate API calls.

• Batch Size: The items list in ESMFoldPredictRequest supports at most batch_size = 2 sequences per request. Larger sets of sequences must be split across multiple requests.

• Multimeric Input: Multimer predictions are supported for up to 4 chains per sequence, separated by : characters. Complexes with more than 4 chains must be decomposed into smaller sub-complexes or modeled chain-wise.

• Protein Complexes: ESMFold was trained only on single-chain structures. Multimer predictions are therefore out-of-distribution and typically less accurate than specialized multimer models (for example AlphaFold-Multimer), especially for detailed interface geometry.

• Accuracy vs. Speed: ESMFold is much faster than AlphaFold2 but usually less accurate on difficult targets (e.g. low-MSA-depth CASP/CAMEO-style benchmarks). For final ranking of a small number of critical designs, higher-accuracy MSA-based models are often preferable.

• Confidence Estimation (pLDDT / pTM): The response fields mean_plddt and ptm correlate with model accuracy, but low-confidence predictions (mean_plddt < 70) or highly novel/orphan sequences should be treated cautiously and ideally cross-checked with additional models or experimental data.

How We Use It¶

ESMFold enables rapid, high-throughput prediction of protein structures directly from sequence data, accelerating protein design and engineering workflows by removing the need for multiple sequence alignment searches. Within BioLM-driven design cycles, ESMFold predictions feed standardized, API-level structural features (coordinates, pLDDT, pTM) into downstream thermodynamic, biophysical, and sequence-based models, allowing teams to quickly prioritize variants for synthesis and testing in enzyme optimization, antibody maturation, and targeted protein modification campaigns.

• Supports iterative protein engineering by quickly assessing structural impact of sequence changes across single chains and small multimers (up to 4 chains in one request).

• Integrates with BioLM scoring and filtering models to focus experimental resources on candidates with favorable predicted structure, stability, and developability profiles.

Related¶

• AlphaFold2 – Complementary MSA-based structure predictor; useful for benchmarking accuracy and confidence estimates against ESMFold’s single-sequence predictions.

• ESM-2 3B – Underlying protein language model whose frozen embeddings ESMFold uses to predict 3D structures directly from sequence.

• RoseTTAFold – Alternative deep learning-based structure prediction model; useful for cross-validating ESMFold results on challenging targets.

• ESM-IF1 – Inverse folding model that can be paired with ESMFold to design sequences compatible with predicted backbones and to evaluate alternative sequences for a given structure.

References¶

• Lin, Z., et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123–1130.