Theory and Computation of PDE Team

Guoqing Liu

2026年11月05日 15:00 至 16:00

Online

Zoom 518 868 7656 (BIMSA)

Making a new molecule starts with a plan: working backwards from the target to simple, available building blocks. This task, known as retrosynthesis, is essential to drug discovery, and it is a hard problem for AI. The space of possible answers is enormous, and the reactions that matter most are often the rarest in the data. This talk presents RetroChimera, recently published in Nature. Instead of building one ever-larger model, we combine two models that think differently: one edits the molecule step by step, while the other writes the answer from scratch. Each excels on different kinds of reactions, and a learned ensemble lets them cover each other's blind spots. RetroChimera handles both common and rare reactions, halves the error rate of strong baselines and transfers to real data from two pharmaceutical companies. In blinded tests, expert chemists preferred its predictions over reactions published in the literature. A predicted reaction, however, is not yet an experiment. I will close with QFANG, a reasoning language model that turns a reaction into step-by-step laboratory instructions, moving AI from planning chemical synthesis towards carrying it out.

Guoqing Liu is a Senior Researcher at Microsoft Research AI for Science in Cambridge, UK. His research interests include reinforcement learning, large language models, and AI for Science, with a focus on post-training language models for scientific reasoning and on AI systems that move chemical synthesis from route planning towards laboratory execution. He has published extensively as a first or corresponding author in leading scientific journals and machine learning conferences, including Nature, Nature Machine Intelligence, ICML, NeurIPS, and ICLR, with over 2,000 citations on Google Scholar. Previously, he was a researcher at Microsoft Research Asia, focusing on reinforcement learning. He received his PhD from the University of Science and Technology of China through a joint program with Microsoft Research Asia.