讲座预告｜mCLM: 一个融合功能且易于合成的模块化化学语言模型

清华大学基础模型研究中心将于7月7日下午13:00举办系列讲座第24期，伊利诺伊大学厄巴纳-香槟分校（UIUC）教授：Heng Ji (季姮)将以"mCLM: A Function-Infused and Synthesis-Friendly Modular Chemical Language Model (mCLM: 一个融合功能且易于合成的模块化化学语言模型)" 为题作报告。

本次讲座对公众开放，欢迎校内外人士参加并共同探讨基础模型领域的前沿话题，校外人士请自行解决入校报备事宜。

报告时间：2025年7月7日（星期一） 13:00-14:00

报告地点：清华大学自强科技楼1号楼312B会议室

主 讲 人： Heng Ji（季姮）

主 持 人： 李涓子（清华大学计算机系教授）

报告题目 

mCLM: A Function-Infused and Synthesis-Friendly 

Modular Chemical Language Model 

(mCLM: 一个融合功能且易于合成的模块化化学语言模型)

报告摘要

Everything in our wonderful world is composed of molecules. Recent advances in block  chemistry involve the manual design of drugs and materials by decomposing molecules into graph substructures—i.e., functional modules—and reassembling them into new molecules with desired functions. However, the process of discovering and manufacturing functional molecules has remained highly artisanal, slow, and expensive. Most importantly, there are many instances of known commercial drugs or materials that have well-documented functional limitations that have remained unaddressed. In this talk I will present our recent efforts at teaching computers to speak two complementary languages: one that represents molecular subgraph structures indicative of specific functions, and another that describes these functions in natural language. Unlike existing approaches that add such knowledge as a post hoc step, we developed a function- and synthesis-aware modular chemical language model (mCLM). Inspired by bilingual speakers who frequently “code-switch”, we propose a novel neural encoder that integrates molecular structure and natural language. mCLM incorporates both function- and synthesis-related knowledge into the small molecule tokenization process a priori. In experiments on 430 FDA-approved drugs, we find mCLM capable of significantly improving 5 out of 6 chemical functions critical to determining drug potentials. More importantly, mCLM can reason on multiple functions and improve the FDA-rejected drugs (“fallen angels”) over multiple iterations to greatly improve their shortcomings.