Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design
Researchers have developed a new contrastive geometric learning model called ConGLUDe, which unifies structure-based and ligand-based approaches for computational drug design. This model uses a geometric protein encoder and a fast ligand encoder to align ligands with protein representations and potential binding sites. ConGLUDe can perform ligand-conditioned pocket prediction, virtual screening, and target fishing, achieving competitive performance in zero-shot virtual screening and state-of-the-art results in ligand-conditioned pocket selection and target fishing. AI
IMPACT This unified approach to drug design could accelerate the discovery of new therapeutics by improving virtual screening and target identification.