Phys-AbGAT: A Physics-Informed Multi-Task Graph Attention Network for Robust Antibody-Antigen Binding Affinity Prediction.

Abstract

Accurate prediction of antibody-antigen binding affinity is essential for therapeutic antibody design. Existing methods often model the dissociation constant (KD) and Gibbs free energy (ΔG) independently, overlooking their thermodynamic relationship and complex paratope-epitope spatial interactions. We introduce Phys-AbGAT, a physics-informed multi-task graph attention network that represents the antibody-antigen interface as a spatial graph. It integrates ESM2 evolutionary embeddings with radial basis functions and employs a thermodynamically motivated auxiliary loss to encourage coupling between log10⁡(KD) and ΔG. On an independent 42-complex benchmark, Phys-AbGAT achieved Pearson correlation coefficients of 0.569 for log10⁡(KD) and 0.573 for ΔG. In matched-set comparisons, Phys-AbGAT achieved numerically lower RMSE and higher PCC than four baseline methods. The differences remained significant after Holm correction for MVSF-AB and PPA-Pred2, but not for AREA-AFFINITY or CSM-AB. Higher attention scores were assigned to several aromatic paratope residues, providing model-level hypotheses for structural inspection. These results support Phys-AbGAT as an interpretable multi-task geometric framework for joint prediction of two affinity-related quantities. Code and data are available at https://github.com/cliffgao/Phys-AbGAT  . Supplementary data are available at Bioinformatics online.