README.md

Antibody Inverse Folding Benchmark

Dataset

512 heavy + 512 light antibody variable domain structures sampled from the test split of the structure AntiFold [1] and AbMPNN [2] structure dataset.

Selection procedure

1. For every sequence in the test split, fold the native sequence with IgFold [3] and compute CDR3 RMSD against the reference PDB (framework-aligned, Cα only).
2. Keep only structures where CDR3 RMSD < 1 Å — this ensures IgFold can faithfully represent the structure, eliminating evaluation bias from IgFold's own limitations.
3. Randomly sample 512 heavy and 512 light chains from the filtered set (seed 42).

Files

if_bench_combined.csv   — task table (1024 rows)
pdbs/
  oas/                     — OAS-derived structures
  sabdab/                  — SAbDab-derived structures

Table columns

Column	Description
chain_type	heavy or light
pdb_local	path to PDB relative to this directory, e.g. pdbs/oas/di9q.pdb
chain_id	chain identifier in the PDB file
global_index	index in the original dataset
residue_start	start index for variable domain extraction (inclusive); NaN = use full chain
residue_end	end index for variable domain extraction (exclusive); NaN = use full chain
native_seq	native variable domain amino acid sequence

Metrics

Each model is run once per structure. Metrics are computed as follows.

AAR (Amino Acid Recovery)

Fraction of positions where the generated sequence matches the native sequence, reported per IMGT region: FR1, CDR1, FR2, CDR2, FR3, CDR3, and globally.

Global RMSD and CDR3 RMSD

For each generated sequence:

1. Fold with IgFold → Cα coordinates of the generated structure.
2. Extract Cα from the reference PDB (variable domain chain).
3. IMGT-number both sequences → find shared positions.
4. Global RMSD: Kabsch-align on all shared Cα → RMSD over all shared positions.
5. CDR3 RMSD: Kabsch-align on framework Cα only → RMSD over CDR3 positions (IMGT 105–117).

Links

[1] Magnus Haraldson Høie, Alissa M Hummer, Tobias H Olsen, Broncio Aguilar-Sanjuan, Morten Nielsen, Charlotte M Deane, AntiFold: improved structure-based antibody design using inverse folding, Bioinformatics Advances, Volume 5, Issue 1, 2025, vbae202, https://doi.org/10.1093/bioadv/vbae202 [2] Frédéric A Dreyer, Daniel Cutting, Constantin Schneider, Henry Kenlay, Charlotte M Deane, Inverse folding for antibody sequence design using deep learning, arXiv preprint, 2023, arXiv:2310.19513, https://arxiv.org/abs/2310.19513 [3] Ruffolo, J.A., Chu, LS., Mahajan, S.P. et al. Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies. Nat Commun 14, 2389 (2023). https://doi.org/10.1038/s41467-023-38063-x