Existing AI models for molecular structure prediction and design largely rely on task-specific algorithms with no unified molecular representation, which limits transferability across tasks and molecule types. On 18 February 2026, a joint team led by Associate Professor Jianzhu Ma of AIR published in Cell a unified model of 3D molecular generation via atomic interactions, named PocketXMol. Using atom prompts as task specifications, PocketXMol handles structure prediction and de novo design for both small molecules and peptides without task-specific fine-tuning, reaching state-of-the-art performance on 11 of 13 computational tasks; designed caspase-9 inhibitors performed comparably to a commercial drug, and generated PD-L1-binding peptides showed high affinity and in vivo tumour targeting.
https://air.tsinghua.edu.cn/info/1007/2524.htm