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

Area of Science:

• Computational chemistry and structural biology
• Drug discovery and medicinal chemistry
• Machine learning in bioinformatics

Background:

• Accurate prediction of small-molecule ligand conformations to protein targets is crucial for drug discovery.
• Existing deep learning docking frameworks often neglect essential local geometric features, limiting pocket identification precision and binding pose reliability.
• Current pocket prediction methods lack adaptability due to reliance on external tools or fixed thresholds.

Purpose of the Study:

• To introduce CGDock, an end-to-end protein-ligand docking framework utilizing curvature-aware geometric flows for enhanced discrete structural representation learning.
• To improve the accuracy of binding pose predictions by integrating local geometric features and employing a ligand-guided adaptive pocket prediction module.
• To develop a unified architecture that simplifies the docking workflow and accommodates protein structural heterogeneity.

Main Methods:

• Integration of discrete Ricci curvature into molecular graph representations to strengthen local structural feature encoding for proteins and ligands.
• Implementation of a ligand-guided adaptive pocket prediction module for estimating ligand-specific binding regions.
• A unified architecture for iterative geometric optimization, refining protein-ligand conformation after pocket identification.

Main Results:

• CGDock demonstrates competitive performance on the PDBind v2020 dataset for protein-ligand docking.
• The framework successfully integrates local geometric features, improving pocket identification and binding pose prediction accuracy.
• The curvature-aware geometric flow operator proves effective as a plug-and-play geometric descriptor.

Conclusions:

• CGDock offers a significant advancement in protein-ligand docking by incorporating curvature-aware geometric flows and adaptive pocket prediction.
• The framework provides accurate binding pose predictions and simplifies the drug discovery workflow.
• The developed geometric descriptor has potential for broader applications in protein-ligand interaction modeling.