chimera-bench

CHIMERA-Bench v1.0

A unified benchmark for epitope-specific antibody CDR sequence-structure co-design.

Paper: CHIMERA-Bench: A Benchmark Dataset for Epitope-Specific Antibody Design (ICLR 2026 GEM Workshop)

Code: github.com/mansoorbaloch/chimera-bench

Dataset Summary

Download

This dataset contains binary PyTorch files (.pt) and PDB structures that require manual download. Use the HuggingFace CLI:

export CHIMERA_DATA_ROOT=/path/to/chimera-bench-v1.0
huggingface-cli download mansoorbaloch/chimera-bench --repo-type dataset --local-dir $CHIMERA_DATA_ROOT

Directory Structure

chimera-bench-v1.0/
  README.md
  metadata/
    final_summary.csv           # 2,922 complexes with 32 columns
    excluded_complexes.csv      # 59 excluded complexes with reasons
    antibody_sequences.fasta    # VH+VL sequences for all complexes
    contamination_audit.json    # PLM training data overlap analysis
  splits/
    epitope_group.json          # Primary split (2338/292/292)
    antigen_fold.json           # Fold-based generalization (2338/292/292)
    temporal.json               # Prospective evaluation (2337/292/293)
  complex_features/             # Per-complex PyTorch tensors (2,941 files)
    {complex_id}.pt
  structures/                   # PDB structure files (2,721 files)
    {pdb}.pdb

Complex Features Format

Each .pt file is a Python dict with:

Sequences

• complex_id: str -- unique identifier ({pdb}{Hchain}{Lchain}_{Agchain})
• heavy_sequence, light_sequence, antigen_sequence: str -- one-letter AA

Coordinates

• heavy_atom14_coords: float32 (N_h, 14, 3) -- 14-atom representation
• heavy_atom14_mask: bool (N_h, 14) -- valid atom flags
• heavy_ca_coords: float32 (N_h, 3) -- CA-only coordinates
• Same for light_* and antigen_*

Annotations

• epitope_residues: list of (chain, resid, resname) tuples
• paratope_residues: list of (chain, resid, resname) tuples
• contact_pairs: list of (ab_chain, ab_resid, ab_resname, ag_chain, ag_resid, ag_resname, distance)

Numbering

• numbering: dict with imgt and chothia sub-dicts, each containing heavy and light lists of (resnum, icode, aa) tuples
• cdr_masks: dict with imgt and chothia sub-dicts, each containing heavy and light int lists (-1=framework; heavy: 0=H1, 1=H2, 2=H3; light: 3=L1, 4=L2, 5=L3)

Surface Features

• ag_surface_points: float32 (128, 3) -- sampled antigen surface points
• ag_surface_normals: float32 (128, 3)
• ag_surface_curvatures: float32 (128, 2) -- mean and Gaussian curvature
• ag_surface_chemical_feats: float32 (128, 6) -- hydropathy, charge, H-bond donor/acceptor, aromaticity, polarity
• Same for heavy_surface_* and light_surface_*

Splits

Each split JSON has keys train, val, test mapping to lists of complex_id strings.

Evaluation Metrics

Quick Start

import torch, json, pandas as pd

# Load metadata
summary = pd.read_csv("metadata/final_summary.csv")

# Load a split
with open("splits/epitope_group.json") as f:
    split = json.load(f)
print(f"Train: {len(split['train'])}, Val: {len(split['val'])}, Test: {len(split['test'])}")

# Load a complex
feat = torch.load(f"complex_features/{split['test'][0]}.pt", weights_only=False)
print(feat['complex_id'], feat['heavy_sequence'][:20], "...")
print(f"Epitope residues: {len(feat['epitope_residues'])}")
print(f"CDR-H3 (IMGT): positions where cdr_masks['imgt']['heavy'] == 2")

Citation

@inproceedings{
ahmed2026chimerabench,
title={{CHIMERA}-Bench: A Benchmark Dataset for Epitope-Specific Antibody Design},
author={Mansoor Ahmed and Nadeem Taj and Imdad Ullah Khan and Hemanth Venkateswara and Murray Patterson},
booktitle={ICLR 2026 Workshop on Generative and Experimental Perspectives for Biomolecular Design},
year={2026},
url={https://openreview.net/forum?id=PyZvVIJbSy}
}

License

Data: CC-BY 4.0. Code: MIT.