New AI Model Enhances Protein-Ligand Binding Prediction

This research introduces RicciBind, a novel geometric representation framework designed to significantly improve the prediction of protein-ligand binding affinity, a critical step in drug discovery. Existing machine learning methods often struggle to effectively model both localized geometric arrangements and broader, coordinated interactions between molecules. RicciBind addresses this by integrating two key components. First, it uses Ricci curvature to capture the tightness of local interactions within molecular structures, thereby creating curvature-aware hierarchical representations of atomic interactions. Second, it employs an optimal transport-based cluster matching mechanism to align protein and ligand clusters across different domains, ensuring globally consistent correspondences and revealing higher-order interaction patterns. By combining curvature-guided structural encoding with optimal transport-driven alignment, RicciBind effectively models complex interaction semantics. Extensive experiments demonstrate its superior predictive performance and generalization across various benchmarks and virtual screening tasks, also confirming the vital role of Ricci curvature in enhancing molecular interaction representations.