TRACE-Net: A Physics-Inspired Geometric Deep Learning Framework for Antibody–Antigen Docking Ranking

Elucidating antibody–antigen interactions at atomic resolution is fundamental for therapeutic antibody design. However, accurately discriminating near-native conformations from massive decoy pools remains a critical bottleneck for current high-accuracy structure prediction, as existing scoring functions often fail to capture complex interfacial interactions and monomeric deformation. To address this challenge, we present TRACE-Net, a physics-inspired geometric deep learning framework. TRACE-Net introduces three innovative modules: Context-Aware Edge Feature Recalibration (CEER), Internal Deformation-Aware Gated Fusion (IDGF), and Top-Weighted Combined Ranking Loss (TWCRL). Trained on a comprehensive data set of 1442 complexes, the model effectively reduces geometric noise and quantifies spatial divergence as a structural indicator of localized geometric deformation forced during flexible interface adaptation. Experimental results demonstrate that TRACE-Net effectively mitigates the severe ranking inversion of near-native decoys, yielding a competitive improvement in the Top-1 success rate, particularly on challenging targets. Overall, this framework offers a sensitive and informative scoring strategy for assessing predicted structures and advancing computational immunology.