CGDock: Curvature-Aware Geometric Flow Framework for Protein-Ligand Docking

Accurately predicting the conformation of small-molecule ligands to protein targets is a critical step in drug discovery and therapeutic development. Although deep learning has enabled substantial advances in molecular docking, many existing frameworks still overlook essential local geometric features when representing proteins and ligands. This limitation compromises the precision of pocket identification and diminishes the reliability of the predicted binding poses. Additionally, numerous pocket prediction modules rely on external tools or fixed thresholds to delineate candidate binding regions, which lack adaptability and constrain the overall efficiency. To address this limitation, we present CGDock, an end-to-end protein-ligand docking framework that is based on curvature-aware geometric flows for discrete structural representation learning. CGDock integrates discrete Ricci curvature into molecular graph representations, which strengthens the encoding of local structural features for both proteins and ligands. A ligand-guided adaptive pocket prediction module is incorporated to estimate the relevant binding region for each ligand, enabling the framework to accommodate structural heterogeneity across proteins. Building on a unified architecture, CGDock first identifies the ligand-specific pocket and subsequently refines the protein-ligand conformation through an iterative geometric optimization. This design simplifies the docking workflow and delivers accurate binding pose predictions. Comprehensive experiments indicate that CGDock demonstrates competitive performance on the PDBind v2020 data set. Furthermore, the curvature-aware geometric flow operator serves as a plug-and-play geometric descriptor for local structural characterization and can be readily extended to broader protein-ligand interaction modeling frameworks.