GRABS: a content-addressed antibody-antigen binding-affinity dataset with predicted structures and a leak-free benchmark
Machine-learning prediction of antibody-antigen binding affinity depends as much on the training data as on the model, yet public affinity measurements are scattered across databases with incompatible formats, mix genuine dissociation constants with censored or predicted values, and concentrate most of their label mass on a few heavily mutated targets. We present GRABS (Geometric Resource for Antibody Binding Studies), a curated, content-addressed, structure-ready dataset that consolidates nine public sources, of which the seven with redistributable licenses yield 4,380 canonical antibody-antigen records} (median-collapsed from 4,761 raw measurements), each with an experimentally measured dissociation constant (K d ), paired heavy and light variable regions (VH, VL), and an antigen sequence, covering 2,015 distinct antigen sequences. Every structure, graph, and feature is addressed by a content key, md5(VH+VL+antigen)[:10], and verified non-circularly: each graph's node features are decoded back to sequence and re-hashed, matching the content key for 4,282/4,282 (100%) structure-backed records. This integrity is load-bearing: an alignment ablation shows that a previously-reported strong cross-target result (R 2 ≈ 0.415) was a misalignment-mediated source/batch leak that content-key regeneration dissolves. On the resulting leak-free, antigen-cluster-grouped benchmark, cross-target ranking of variants of a held-out antigen collapses to the noise floor (within-held-out-antigen Spearman ρ ≈ 0 for every model family: sequence, structure, and folding-confidence), with structure giving no advantage, a rigorously leakage-controlled quantification of the limited reach of current methods. GRABS ships model-ready artifacts (Boltz-2 predicted complexes, per-residue ESM-2 embeddings, geometric graphs, feature vectors) with antigen-stratified splits, and is intended for leakage-free benchmarking and within-target affinity ranking (measure-then-rank), not zero-shot prediction for antigens absent from training.
Authors
- João Meidânis (ORCID: https://orcid.org/0000-0001-7878-4990)
- Priscila Pini Zenatti (ORCID: https://orcid.org/0000-0002-8662-7458)
- Luciano de Melo Martins (ORCID: https://orcid.org/0009-0006-1192-0808)
Institutions
- Universidade Estadual de Campinas (UNICAMP) (BR)
Publication Details
- Journal
- bioRxiv (Cold Spring Harbor Laboratory)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.64898/2026.11.17.767263
- Primary Topic
- vaccines and immunoinformatics approaches
- Type
- preprint