A Mechanism-Aware Dual Attention Deep Model for Molecular-Protein Binding Affinity Prediction with Enhanced Generalizability and Interpretability

Accurate prediction of molecular-protein binding affinity (MPBA) is paramount in drug discovery, yet current computational models often lack generalizability, interpretability, and explicit mechanistic understanding. To address these limitations, we introduce BindMecNet (Binding Mechanism Network), a novel mechanism-aware, multi-scale deep model. BindMec-Nets core innovation lies in its Interfacial Interaction Prediction Module, which explicitly predicts an atom-residue level interaction map, serving as a crucial mechanistic inductive bias. This map guides a Mechanism-Aware Dual Attention Interaction Module, ensuring that information exchange between protein and ligand representations is focused on genuinely interacting regions, fostering a deeper understanding of binding mechanisms. We employ a robust two-stage training strategy: initial pre-training on PDBBind with a multi-task loss (affinity and interaction prediction), followed by fine-tuning on challenging generalization datasets using predicted protein and ligand structures. Our comprehensive evaluations on unseen protein families using predicted binding conformations demonstrate BindMecNets superior performance, significantly outperforming state-of-the-art deep learning baselines. Ablation studies confirm that both the PDBBind pre-training and the explicit mechanistic inductive bias are critical for achieving this enhanced generalizability and accuracy. Furthermore, BindMecNets predicted interaction maps offer valuable, interpretable insights into binding hotspots, paving the way for more rational and mechanism-driven drug design.