Abstract
Gaining structural insights into antibody-antigen interactions is essential for understanding immune recognition and therapeutics design. Accurately modeling these complexes remains challenging for both physics-based approaches and AI-based, co-folding methods such as AlphaFold3. These methods not only struggle to generate near-native conformations, but, more critically, often fail to rank those correctly, revealing fundamental limitations for antibody-antigen modeling. We present DeepRank-Ab, a geometric deep learning-based scoring function tailored to antibody-antigen interfaces, together with a rigorously curated benchmark of ~2.3 million decoys from 1,442 complexes, providing the diversity required for robust training and unbiased evaluation. We systematically assessed graph representations, structural and energetic features, and sampling strategies. Our analysis identified that atom-level representations coupled with Voronoi-based surface decomposition and antibody-specific features are the most effective formulation for accurate scoring. Across multiple independent test sets, DeepRank-Ab consistently outperforms all evaluated methods, including AlphaFold3, HADDOCK and state-of-the-art scoring functions. It increases AlphaFold3 Top1 success rate by 35.5% and improves the mean Top1 DockQ by more than a factor of two. DeepRank-Ab generalizes beyond its training distribution, achieving 100% Top5 success rate on external antibody-antigen CAPRI targets, surpassing all tested methods. These results establish DeepRank-Ab as effective scoring method substantially improving identification of near-native antibody-antigen conformations.