Yinjun Jia

Keywords molecular interaction generative model unified AI model molecular structure prediction peptide design drug design molecular docking protein pocket atom-level interaction foundation model Introduction Artificial intelligence (AI) has revolutionized molecular structure prediction1,2,3,4 and design,5,6,7,8,9 yet current models typically rely on specialized algorithms tailored to specific tasks.

Editor’s summary Despite progress in drug discovery, approximately 90% of druggable disease targets still lack small-molecule therapies. Although virtual screening can accelerate hit identification, traditional methods such as molecular docking remain too slow for genome-scale applications. Jia et al. introduce DrugCLIP, a contrastive learning framework that embeds protein pockets and small molecules into a shared latent space, enabling virtual screening up to 10 million times faster than docking.