File size: 5,303 Bytes
4205816
 
6e61de9
 
4205816
6e61de9
 
 
 
 
 
 
 
 
 
 
 
 
 
4205816
 
0c42f71
 
 
 
4a7f2cd
0c42f71
4a7f2cd
0c42f71
6e61de9
4205816
0c42f71
 
 
 
6e61de9
0c42f71
 
 
 
6e61de9
 
 
 
 
 
 
 
 
 
0c42f71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e61de9
0c42f71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e61de9
 
0c42f71
 
6e61de9
 
 
 
 
0c42f71
 
 
6e61de9
 
 
 
 
0c42f71
6e61de9
 
 
 
 
 
 
 
 
 
 
 
 
0c42f71
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77d6ca3
 
 
 
 
 
 
0c42f71
 
 
 
 
6e61de9
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
---
license: cc-by-4.0
task_categories:
  - other
tags:
  - antibody-design
  - protein-structure
  - benchmark
  - biology
  - structural-biology
language:
  - en
size_categories:
  - 1K<n<10K
configs:
  - config_name: metadata
    data_files:
      - split: train
        path: "metadata/final_summary.csv"
---

# 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](https://openreview.net/forum?id=PyZvVIJbSy) (ICLR 2026 GEM Workshop)

**Code**: [github.com/mansoorbaloch/chimera-bench](https://github.com/mansoorbaloch/chimera-bench)

## Dataset Summary

| Property | Value |
|----------|-------|
| Complexes | 2,922 |
| PDB structures | 2,721 |
| Pre-computed features | 2,941 `.pt` files |
| Splits | 3 (epitope-group, antigen-fold, temporal) |
| Numbering schemes | IMGT, Chothia |
| Contact cutoff | 4.5 A |
| Resolution cutoff | 4.0 A |
| Baselines evaluated | 11 methods, 6 paradigms |

## Download

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

```bash
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

| Split | Train | Val | Test | Generalization Axis |
|-------|------:|----:|-----:|---------------------|
| epitope_group | 2,338 | 292 | 292 | Unseen epitope patterns |
| antigen_fold | 2,338 | 292 | 292 | Unseen antigen folds |
| temporal | 2,337 | 292 | 293 | Prospective (by deposition date) |

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

## Evaluation Metrics

| Group | Metrics |
|-------|---------|
| Sequence quality | AAR, CAAR, PPL |
| Structural accuracy | RMSD (Kabsch-aligned CA), TM-score |
| Binding interface | Fnat, iRMSD, DockQ |
| Epitope specificity | EpiF1 (precision, recall, F1) |
| Designability | n_liabilities (NG, DG, DS, DD, NS, NT, M motifs) |

## Quick Start

```python
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

```bibtex
@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.