AI-Augmented Physics-Based Docking for Antibody-Antigen Complex Prediction

Abstract

Predicting the structure of antibody-antigen complexes is a challenging task with significant implications for the design of better antibody therapeutics. However, the levels of success have remained dauntingly low, particularly when high standards for model quality are required, a necessity for efficient antibody design. Artificial intelligence (AI) has significantly impacted the landscape of structure prediction for antibodies, both alone and in complex with their antigens. We utilized AI-guided antibody modeling tools to generate ensembles displaying diversity in the complementarity-determining region (CDR) and integrated those into our previously published AlphaFold2-rescored docking pipeline, a strategy called AI-augmented physics-based docking. We highlight that the quality of the ensemble is crucial for docking performance, that including too many models can be detrimental and that prioritization of models is essential for achieving good performance. In this study, we also compare docking performance with AlphaFold, the new benchmark in the field. We distinguish between two types of success tailored to specific downstream applications: 1) criteria sufficient for epitope mapping, where gross quality is adequate and can complement experimental techniques, and 2) criteria for producing higher-quality models suitable for engineering purposes. Our results robustly demonstrate the advantages of AI-augmented docking over AlphaFold2, further accentuated when higher standards in quality are imposed. Docking performance is noticeably lower than the one of AlphaFold3 in both epitope mapping and antibody design. While we observe a strong dependence on CDR-H3 length for physics-based tools on their ability to successfully predict, this helps define an applicability range where physics-based docking can be competitive to AlphaFold3.