Atom-level generative foundation model for molecular interaction with pockets

Abstract

Understanding molecular interactions is essential to structural biology and drug discovery. Despite the progress of AI models in revealing and exploiting the interaction mechanisms for various applications, they are predominantly tailored to specific tasks without fully exploiting the underlying transferability across molecular data and tasks. Here, we present PocketXMol, an atom-level generative foundation model to decipher fundamental atomic interactions for general protein-pocket-interacting molecular tasks. It adopts a novel unified generative framework with an innovative task prompt mechanism and an exclusive atom-level representation, making it applicable to diverse tasks covering structure prediction and design of small molecules and peptides, without requiring fine-tuning. PocketXMol was compared to 55 baseline models across 13 typical tasks, achieving state-of-the-art performance in 11 tasks and remaining competitive in the others. We successfully utilized PocketXMol to design novel small molecules that inhibit caspase-9 with efficacy comparable to that of commercial pan-caspase inhibitors. Furthermore, we employed PocketXMol to design PD-L1-binding peptides, demonstrating a success rate substantially higher than random library screening. Three representative peptides underwent further experiments, which validated the cellular specificity and confirmed their potential for molecular probing and therapeutics. PocketXMol presents a powerful and versatile tool with promising prospects for future applications and will have a profound impact on AI-aided drug discovery.