Xingang Peng, Ruihan Guo, Fenglin Guo, Ziyi Wang, Ziyi Wang, Jiayu Sun, Jiaqi Guan, Yinjun Jia, Yan Xu, Yanwen Huang, Muhan Zhang, Jian Peng, Xinquan Wang, Chuanhui Han, Zihua Wang, Zihua Wang, Jianzhu Ma
We present PocketXMol, an atom-level model that unifies generative tasks related to protein pocket interactions. Using atomic prompts as task specifications, PocketXMol supports various molecular tasks, including structure prediction and de novo design of small molecules and peptides, without task-specific fine-tuning. PocketXMol achieved strong performance on 11 of 13 computational benchmarks and remained competitive on the remaining two, outperforming 55 baseline models. We applied PocketXMol to design caspase-9-inhibiting small molecules, achieving efficacy comparable with commercial pan-caspase inhibitors. We also adopted PocketXMol to generate PD-L1-binding peptides, resulting in a success rate that largely exceeds library screening. Three representative peptides underwent further experiments, which validated their cellular specificity and confirmed their potential for molecular probing and therapeutics. PocketXMol provides a general platform for AI-aided drug discovery and enables a wide range of future applications.