Molecule Meets Protein Pocket 3D-Aware Molecular Optimization for Protein Targets

TMLR Paper7107 Authors

Abstract: Lead optimization, refining drug candidates to improve binding to protein targets, is a key challenge in drug discovery. We introduce a 3D-aware generative framework that performs fragment-level molecular optimization conditioned on the geometry of the protein's binding pocket. Our model represents the molecule-protein complex as a sparse 3D graph and applies grouped vector attention to learn spatial interactions. It decomposes the molecule into a stable scaffold and generates new fragments using a Variational Autoencoder (VAE) and a SMILES-based transformer guided by local pocket structure. To handle the imbalance in fragment sizes, we incorporate a focal loss. On the CrossDock2020 benchmark, our method outperforms prior approaches in generating diverse, novel, and chemically valid candidates with improved Vina scores-while generalizing to unseen proteins.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Wenbing_Huang1

Submission Number: 7107