ABodyBuilder2 is an antibody-specific deep learning model for rapid, high-accuracy prediction of Fv structures, including all six CDR loops, from paired heavy (H) and light (L) chain sequences. It achieves a mean CDR-H3 backbone RMSD of 2.81 Å on a benchmark of 34 antibodies and produces full all-atom models without requiring MSAs, templates, or external sequence databases. The API returns PDB-format structures suitable for antibody screening, affinity maturation, and structure-guided engineering workflows.

Predict

Predict properties or scores for input sequences

python 
from biolmai import BioLM
response = BioLM(
    entity="abodybuilder2",
    action="predict",
    params={},
    items=[
      {
        "H": "EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
        "L": "DIQMTQSPSSLSASVGDRVTITCRASQDISNYLNWYQQKPGKAPKLLIYHTSRLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQGNTLPYTFGQGTKVEIK"
      },
      {
        "H": "QVQLQQSGPELEKPGASVKMSCKASGYTFTSYNMHWVKQTPGRGLEWIGAIYPGNGDTSYNQKFKGKATLTADKSSSTAYMQLSSLASEDSAVYYCARSNYYGSSYFDYWGQGTTLTVSS",
        "L": "QIVLSQSPAILSASPGEKVTMTCRASSSVNYMDWYQQKPGSSPKPWIYAPSNLASGVPVRFSGSGSGTSYSLTISRVEAEDAATYYCQQWSFNPPTFGAGTKLEIK"
      }
    ]
)
print(response)
bash 
curl -X POST https://biolm.ai/api/v3/abodybuilder2/predict/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "H": "EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
      "L": "DIQMTQSPSSLSASVGDRVTITCRASQDISNYLNWYQQKPGKAPKLLIYHTSRLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQGNTLPYTFGQGTKVEIK"
    },
    {
      "H": "QVQLQQSGPELEKPGASVKMSCKASGYTFTSYNMHWVKQTPGRGLEWIGAIYPGNGDTSYNQKFKGKATLTADKSSSTAYMQLSSLASEDSAVYYCARSNYYGSSYFDYWGQGTTLTVSS",
      "L": "QIVLSQSPAILSASPGEKVTMTCRASSSVNYMDWYQQKPGSSPKPWIYAPSNLASGVPVRFSGSGSGTSYSLTISRVEAEDAATYYCQQWSFNPPTFGAGTKLEIK"
    }
  ]
}'
python 
import requests

url = "https://biolm.ai/api/v3/abodybuilder2/predict/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "items": [
        {
          "H": "EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
          "L": "DIQMTQSPSSLSASVGDRVTITCRASQDISNYLNWYQQKPGKAPKLLIYHTSRLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQGNTLPYTFGQGTKVEIK"
        },
        {
          "H": "QVQLQQSGPELEKPGASVKMSCKASGYTFTSYNMHWVKQTPGRGLEWIGAIYPGNGDTSYNQKFKGKATLTADKSSSTAYMQLSSLASEDSAVYYCARSNYYGSSYFDYWGQGTTLTVSS",
          "L": "QIVLSQSPAILSASPGEKVTMTCRASSSVNYMDWYQQKPGSSPKPWIYAPSNLASGVPVRFSGSGSGTSYSLTISRVEAEDAATYYCQQWSFNPPTFGAGTKLEIK"
        }
      ]
    }

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

url <- "https://biolm.ai/api/v3/abodybuilder2/predict/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  items = list(
    list(
      H = "EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
      L = "DIQMTQSPSSLSASVGDRVTITCRASQDISNYLNWYQQKPGKAPKLLIYHTSRLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQGNTLPYTFGQGTKVEIK"
    ),
    list(
      H = "QVQLQQSGPELEKPGASVKMSCKASGYTFTSYNMHWVKQTPGRGLEWIGAIYPGNGDTSYNQKFKGKATLTADKSSSTAYMQLSSLASEDSAVYYCARSNYYGSSYFDYWGQGTTLTVSS",
      L = "QIVLSQSPAILSASPGEKVTMTCRASSSVNYMDWYQQKPGSSPKPWIYAPSNLASGVPVRFSGSGSGTSYSLTISRVEAEDAATYYCQQWSFNPPTFGAGTKLEIK"
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))
POST /api/v3/abodybuilder2/predict/

Predict endpoint for ABodyBuilder2.

Request Headers:

Request

  • params (object, optional) — Configuration parameters:

    • plddt (bool, default: False) — Whether to include pLDDT scores in the response

    • seed (int, optional, default: 42) — Random seed for prediction consistency

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

    • H (string, min length: 1, max length: 2048, required) — Heavy chain amino acid sequence

    • L (string, min length: 1, max length: 2048, required) — Light chain amino acid sequence

Example request:

http 
POST /api/v3/abodybuilder2/predict/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "items": [
    {
      "H": "EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
      "L": "DIQMTQSPSSLSASVGDRVTITCRASQDISNYLNWYQQKPGKAPKLLIYHTSRLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQGNTLPYTFGQGTKVEIK"
    },
    {
      "H": "QVQLQQSGPELEKPGASVKMSCKASGYTFTSYNMHWVKQTPGRGLEWIGAIYPGNGDTSYNQKFKGKATLTADKSSSTAYMQLSSLASEDSAVYYCARSNYYGSSYFDYWGQGTTLTVSS",
      "L": "QIVLSQSPAILSASPGEKVTMTCRASSSVNYMDWYQQKPGSSPKPWIYAPSNLASGVPVRFSGSGSGTSYSLTISRVEAEDAATYYCQQWSFNPPTFGAGTKLEIK"
    }
  ]
}
Status Codes:

Response

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

    • pdb (string) — Predicted antibody structure in standard PDB format.

    • plddt (array of arrays of floats, optional) — Predicted Local Distance Difference Test (pLDDT) scores per residue:

      • Outer array length: 2 (chains: heavy chain “H”, light chain “L”)

      • Inner array length: equal to the number of residues in the corresponding chain

      • Values range: 0.0–100.0 (higher values indicate higher local structure confidence)

Example response:

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

      {
  "results": [
    {
      "pdb": "REMARK  ANTIBODY STRUCTURE MODELLED USING ABODYBUILDER2                         \nREMARK  STRUCTURE REFINED USING OPENMM 8.2, 2025-06-18                          \nATOM      1  N   GLU H   1     -17.169... (truncated for documentation)"
    },
    {
      "pdb": "REMARK  ANTIBODY STRUCTURE MODELLED USING ABODYBUILDER2                         \nREMARK  STRUCTURE REFINED USING OPENMM 8.2, 2025-06-18                          \nATOM      1  N   GLN H   1     -17.778... (truncated for documentation)"
    }
  ]
}

Performance

  • Runs on CPU-optimized instances (2 vCPUs, 8 GB RAM) without requiring a GPU, enabling cost-efficient large-scale inference for antibody structures

  • Predictive accuracy (backbone RMSD, Å) on the 34-antibody benchmark: - CDR-H3: 2.81 (vs 2.90 for AlphaFold-Multimer; ~10% lower error than ABlooper, IgFold, EquiFold) - Framework regions: 0.54–0.57 (within typical X-ray experimental error)

  • Chemical surface and stereochemistry: - All-atom outputs with accurate side-chain torsions (χ1/χ2) and exposed/buried classification comparable to AlphaFold-Multimer - Zero peptide-bond, cis-bond, D-amino acid, or heavy-atom clash violations in benchmarked models after refinement, unlike EquiFold and some other antibody models

  • Compared to other BioLM structure predictors: - More accurate and substantially faster than AlphaFold2 / AlphaFold-Multimer for antibodies, especially on CDR-H3 - More accurate for antibodies than generalist models such as ESMFold, while being far more suitable for high-throughput antibody-focused workloads

Applications

  • Accurate prediction of antibody variable region (VH/VL) structures for therapeutic design, providing all-atom Fv models that can be used for downstream docking, epitope mapping, and structure-based affinity maturation workflows in antibody discovery pipelines.

  • Rapid structural modeling of large antibody panels (up to 8 VH/VL pairs per API call, each chain ≤2048 residues), enabling high-throughput triage of candidates from display libraries or immune repertoire sequencing by linking sequence variants to 3D CDR conformations and paratope geometry.

  • Nanobody and TCR structure prediction via the broader ImmuneBuilder API, allowing teams to use the same interface for single-domain antibodies and TCRs when appropriate, and to compare structural features across immune modalities in multi-format biologics programs.

  • Use of model-provided residue-level error estimates to filter or down‑weight uncertain regions (especially CDR-H3), helping focus experimental characterization on designs where predicted loop conformations and VH–VL orientations are likely to be reliable.

  • Limitations include reduced accuracy for highly unusual or very long CDR-H3 loops and lack of explicit antigen context; predicted structures should be combined with experimental data and additional modeling for final drug design and developability assessment.

Limitations

  • Maximum Sequence Length: Each antibody heavy (H) and light (L) chain sequence must not exceed 2048 amino acids.

  • Batch Size: The items array in ImmuneBuilderABodyBuilder2PredictRequest supports a maximum of 8 antibody sequence pairs per request.

  • Antibody-specific only: ABodyBuilder2 is trained for paired antibody variable domains (VH/VL). It is not suitable for general protein structure prediction or for other immune receptors such as nanobodies or T-cell receptors; use nanobodybuilder2 or tcrbuilder2 instead.

  • CDR-H3 edge cases: Although ABodyBuilder2 achieves state-of-the-art average accuracy for CDR-H3, predictions for unusually long or rare CDR-H3 loops (e.g. beyond ~22 residues) may be less reliable than for typical lengths.

  • No evolutionary data: The model does not use multiple sequence alignments or evolutionary information, so in settings where evolutionary couplings are critical, AlphaFold-Multimer or similar MSA-based methods may give more accurate results.

  • Output type: The API returns only an all-atom structure in pdb format per item. It does not expose per-residue error estimates, embeddings, attention maps, or other internal model features for downstream analysis.

How We Use It

ABodyBuilder2 enables rapid, sequence-to-structure modeling of antibody variable regions (VH/VL) through a standardized API, making accurate Fv and CDR, especially CDR-H3, structures available early in antibody discovery and optimization. These structures integrate with BioLM sequence embeddings, developability predictors, and docking/affinity models to prioritize designs, focus library construction, and inform multi-round in vitro campaigns while scaling to large next-generation sequencing repertoires.

  • Supports high-throughput structural screening of candidate antibodies to accelerate lead selection and maturation.

  • Provides consistent 3D inputs for downstream property prediction (e.g., stability, liability, epitope/paratope analyses) and ranking.

References

  • Abanades, B., Wong, W. K., Boyles, F., Georges, G., Bujotzek, A., & Deane, C. M. (2023). ImmuneBuilder: Deep-Learning models for predicting the structures of immune proteins. Communications Biology, 6, 575. https://doi.org/10.1038/s42003-023-04927-7