Invited Talk
in
Expo Workshop: AI for Science: Foundation Models and Agentic Systems for Closed-Loop Discovery Mon, Jul 6, 2026 • 12:40 AM – 1:00 AM PDT

Teaching Foundation Models to Read, See, and Assemble Molecules

Zhewei Wei

Abstract

Can a foundation model learn to assemble a molecule? This talk begins with MotifAgent, a multi-agent framework where molecular motifs collaborate to learn valid connection rules. I then step back to the structural foundations that make assembly possible: MolBasic teaches language models to read molecular graphs from SMILES, while MolSight teaches vision-language models to see chemical topology in molecular images. Together, these works argue for a structure-grounded path toward molecular foundation models that can read, see, and assemble molecules.

Speaker

Zhewei Wei

Zhewei Wei is a Professor at Renmin University of China, where he serves as Vice Dean of the Gaoling School of Artificial Intelligence. His research focuses on fundamental algorithms and models for artificial intelligence. He received his BSc from Peking University in 2008 and his PhD from the Hong Kong University of Science and Technology in 2012. He has received the PODS Test-of-Time Award and a Best Research Paper nomination at VLDB 2024, along with a Youth Outstanding Paper Nomination at the World Artificial Intelligence Conference (WAIC) 2023. He serves as an Associate Editor of IEEE TPAMI and has been an Area Chair for ICML, NeurIPS, and ICLR.