MultiSAAI: Sequence-informed Antibody-Antigen Interactions Prediction using Multi-scale Deep Learning

Abstract

MotivationAntibodies possess high specificity toward target antigens, making them critical for therapeutic applications. However, experimental screening of antibody-antigen interactions is labor-intensive and costly. Existing structure-based computational methods often face significant challenges in modeling the dynamic conformational flexibility that governs antibody-antigen interactions, which limits their ability to accurately identify binding. Therefore, there is an urgent need for accurate and efficient sequence-based prediction method to accelerate antibody discovery and reduce experimental burden. ResultsWe propose MultiSAAI, a sequence-informed framework that models antibody-antigen interactions by explicitly accounting for the distinct roles of antibody heavy and light chains in antigen binding. MultiSAAI integrates language model embeddings, physicochemical properties, geometric constraints, and residue substitutability to characterize antibody-antigen interactions across multiple scales, employing a multi-scale network architecture that simultaneously evaluates global residue-pair compatibility and local amino acid fitness at the binding interface. Furthermore, the incorporation of site-specific information and biologically grounded binding principles allows the model to more closely reflect the actual mechanisms of interactions. Benchmark results demonstrate that MultiSAAI achieves AUROC scores of 0.757 on the generic antibody-antigen interactions dataset and 0.946 on the SARS-CoV-2 dataset, outperforming existing methods such as A2binder and AbAgIntPre. Finally, large-scale preliminary antibody screening further validates the potential of MultiSAAI for high-throughput therapeutic antibody discovery.