Decipher Fundamental Atomic Interactions to Unify Generative Molecular Docking and Design

Atomic interactions are fundamental to molecular structures and functions. We constructed PocketXMol, an all-atom AI model, to learn these interactions for general pocket-interacting molecular generative tasks. PocketXMol unified distinct molecular tasks under a single computational framework without re-quiring fine-tuning. It was evaluated on 11 typical tasks, covering docking and design of small molecules and peptides, and compared against 49 baselines using 45 metrics. PocketXMol outperformed state-of-the-art methods in 9 tasks and was competitive in the remaining ones. We successfully adopted PocketXMol to design novel small molecules to inhibit caspase-9 with efficacy comparable to commercial pan-caspase inhibitors. We also adopted PocketXMol to design PD-L1-binding peptides and demonstrated its success rate 50,000 times higher than random library screening. Three peptides underwent rigorous cellular validation, exhibiting effective binding to the cell membrane. Their potential for molecular probing and therapy was further confirmed through injections in tumor mouse models and the ligand inhibition assay.