DeepRank-Ab: a scoring function for antibody-antigen complexes based on geometric deep learning

Abstract Gaining structural insights into the interactions between antibodies and their corresponding antigens is essential for understanding immune recognition and for guiding therapeutic antibody design. However, accurately modelling these complexes remains a significant challenge for both physics-based docking approaches and AI-based, co-folding methods such as AlphaFold3. These methods not only struggle to generate near native conformations, but, more critically, they often fail to score and rank such conformations correctly, revealing fundamental limitations when applied to antibody-antigen systems. To address these limitations, we present here DeepRank-Ab, a geometric deep learning-based scoring function tailored to the unique characteristics of antibody-antigen interfaces. Its development was enabled by a rigorously curated benchmark comprising more than 2.3 million decoys generated from 1,442 complexes, providing the diversity required for robust training and unbiased evaluation. Leveraging this resource, we systematically assessed multiple levels of graph representation, structural and energetic feature sets, and sampling strategies. Building on our previous DeepRank-GNN-esm work, our analysis identified that atom-level graph representation coupled with Voronoi-based surface decomposition and antibody-specific features is the most effective formulation for accurate scoring. Across multiple independent test sets, including models from unbound unbound docking and structures generated by AlphaFold, DeepRank-Ab consistently outperforms all evaluated methods, including AF3, HADDOCK and state of the art scoring functions such as FTDMP. It increases the AF3 Top 1 success rate by 35.5% and improves the mean Top 1 DockQ by more than a factor of two. DeepRank-Ab further generalizes robustly beyond its training distribution, achieving 100% Top 5 success rate on external antibody-antigen CAPRI targets, surpassing all tested methods. Together, these results demonstrate that DeepRank-Ab is a highly effective scoring method that substantially improves the identification of near-native antibody-antigen conformations. Competing Interest Statement The authors have declared no competing interest. Footnotes ↵‡ These authors jointly supervised this work Added a missing author. Added a possible explanation for the good performance on antibody-peptide complexes. Added link to published dataset.