New GNN Predicts Protein-Ligand Binding Affinity with Curvature Information

Accurate prediction of protein-ligand binding affinity is a crucial step in structure-based drug discovery. While recent geometric deep learning methods have shown promise by representing molecular complexes as 3D graphs, most approaches primarily focus on static interaction geometry from a single bound conformation, often overlooking molecular flexibility and the conformational changes induced by binding. To address this limitation, a new graph neural network called CPES (Curvature-Informed Potential Energy Surface) has been developed. CPES integrates physics-informed curvature representations to model the dynamic conformational flexibility of molecules. It derives curvature spectral descriptors from the Hessian of the potential energy surface, capturing local principal curvatures. The model uses spectral cross-attention to compare unbound and bound states, thereby capturing binding-induced dynamic changes. This dynamic information is then fused with static structural features learned through geometry-aware message passing and hierarchical representations. Extensive evaluations demonstrate that CPES achieves improved predictive performance and offers enhanced physical interpretability for binding affinity prediction.