Immune Builder API

The Immune Builder API provides deep learning-based structure prediction for immune receptor proteins including antibodies, nanobodies, and T-cell receptors (TCRs). This API is based on the ImmuneBuilder project, which uses state-of-the-art deep learning models to predict protein structures with high accuracy.

Overview

ImmuneBuilder generates structures with state-of-the-art accuracy while being much faster than AlphaFold2. The API supports three main modes:

  • Antibody mode: Predicts antibody structures using ABodyBuilder3
  • Nanobody mode: Predicts nanobody structures using NanoBodyBuilder2
  • TCR mode: Predicts T-cell receptor structures using TCRBuilder2+

Command Line Interface

Examples

Predict antibody structure

lev engine submit immune-builder \
  --mode antibody \
  --heavy-chain heavy.fasta \
  --light-chain light.fasta

Predict antibody structure with specific model type

lev engine submit immune-builder \
  --mode antibody \
  --heavy-chain heavy.fasta \
  --light-chain light.fasta \
  --model-type language

Predict TCR structure

lev engine submit immune-builder \
  --mode tcr \
  --alpha-fasta alpha.fasta \
  --beta-fasta beta.fasta

Predict nanobody structure

lev engine submit immune-builder \
  --mode nanobody \
  --heavy-chain heavy.fasta

Flags

  • --mode (str) (Required)
    • The modeling mode to use
    • Options:
      • antibody: Predict antibody structures (requires heavy and light chain FASTA files)
      • tcr: Predict T-cell receptor structures (requires alpha and beta chain FASTA files)
      • nanobody: Predict nanobody structures (requires heavy chain FASTA file only)
  • --heavy-chain (str) (Conditional)
    • Path to the heavy chain FASTA file
    • Required for antibody and nanobody modes
  • --light-chain (str) (Conditional)
    • Path to the light chain FASTA file
    • Required for antibody mode only
  • --alpha-fasta (str) (Conditional)
    • Path to the alpha chain FASTA file
    • Required for tcr mode only
  • --beta-fasta (str) (Conditional)
    • Path to the beta chain FASTA file
    • Required for tcr mode only
  • --model-type (str) (Default: plddt)
    • Model type for antibody mode
    • Options:
      • plddt: Use pLDDT-based model
      • language: Use language model-based approach
    • Only applicable for antibody mode

Python Interface

Examples

Predict antibody structure:

from engine import EngineClient

client = EngineClient()
client.authorize()

job_id = client.submit_immune_builder(
    mode="antibody",
    heavy_fasta="heavy.fasta",
    light_fasta="light.fasta"
)

Predict antibody structure with specific model type:

job_id = client.submit_immune_builder(
    mode="antibody",
    heavy_fasta="heavy.fasta",
    light_fasta="light.fasta",
    model_type="language"
)

Predict TCR structure:

job_id = client.submit_immune_builder(
    mode="tcr",
    alpha_fasta="alpha.fasta",
    beta_fasta="beta.fasta"
)

Predict nanobody structure:

job_id = client.submit_immune_builder(
    mode="nanobody",
    heavy_fasta="heavy.fasta"
)

Outputs

  • ImmuneBuilder_output.pdb (PDB file)
    • The predicted 3D structure in PDB format

Performance and Accuracy

Based on the ImmuneBuilder repository, the models achieve the following performance:

  • ABodyBuilder2: CDR-H3 loops with RMSD of 2.81Å (0.09Å improvement over AlphaFold-Multimer)
  • ABodyBuilder3: CDR-H3 loops with RMSD of 2.42Å (0.12Å improvement over AbodyBuilder2 on a different dataset)
  • NanoBodyBuilder2: CDR-H3 loops with RMSD of 2.89Å (0.55Å improvement over AlphaFold2)
  • TCRBuilder2+: State-of-the-art TCR structure prediction

All models are significantly faster than AlphaFold2 while maintaining high accuracy.

Processing Time

Typical processing times vary by mode and sequence length, but are generally much faster than AlphaFold2.

References

Updated: