首页|LeadFlow aligns generative priors with real-world hit-to-lead optimization patterns for structure-based molecular optimization

LeadFlow aligns generative priors with real-world hit-to-lead optimization patterns for structure-based molecular optimization

Runze Zhang Jia Li Hao Zhou Mingliang Wang Xiaomin Luo Duanhua Cao Mingyue Zheng Qi Huang Beijing Chen Xia Sheng Xinyu Jiang Lehan Zhang Xingyou Wang Yaoyu Zheng Zhehuan Fan Dan Teng Jie Yu Keyue Qiu Mingan Chen

来源：

中国科学院科技论文预发布平台

LeadFlow aligns generative priors with real-world hit-to-lead optimization patterns for structure-based molecular optimization

Runze Zhang ¹Jia Li ²Hao Zhou ³Mingliang Wang ⁴Xiaomin Luo ⁵Duanhua Cao ⁶Mingyue Zheng ⁷Qi Huang ⁸Beijing Chen ⁸Xia Sheng ⁵Xinyu Jiang ⁵Lehan Zhang ⁵Xingyou Wang ⁵Yaoyu Zheng ⁹Zhehuan Fan ⁵Dan Teng ¹⁰Jie Yu ¹¹Keyue Qiu ³Mingan Chen¹²

作者信息

1. Science for Life Laboratory, Department of Cell and Molecular Biology, Uppsala University, BMC, Uppsala, Sweden
2. The State Key Laboratory of Chemical Biology, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China;Zhongshan Institute for Drug Discovery, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Zhongshan 528400, China
3. Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China
4. University of Chinese Academy of Sciences, Beijing 100049, China;Department of Medicinal Chemistry, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China;Zhongshan Institute for Drug Discovery, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Zhongshan 528400, China
5. University of Chinese Academy of Sciences, Beijing 100049, China;Drug Discovery and Design Center, State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China
6. State Key Laboratory of Cardiovascular Diseases and Medical Innovation Center, Medical Artificial Intelligence Innovation Center, Shanghai East Hospital, Frontier Science Center for Stem Cell Research, School of Life Sciences and Technology, Tongji University, Shanghai, 200092, China
7. School of Information Science and Technology, ShanghaiTech University, Shanghai, 201210, China;University of Chinese Academy of Sciences, Beijing 100049, China;Drug Discovery and Design Center, State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China;School of Pharmaceutical Science and Technology, Hangzhou Institute for Advanced Study, UCAS, Hangzhou, 330106, China
8. Zhongshan Institute for Drug Discovery, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Zhongshan 528400, China
9. University of Chinese Academy of Sciences, Beijing 100049, China;School of Pharmaceutical Science and Technology, Hangzhou Institute for Advanced Study, UCAS, Hangzhou, 330106, China
10. Drug Discovery and Design Center, State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China
11. Lingang Laboratory, Shanghai 200031, China;School of Information Science and Technology, ShanghaiTech University, Shanghai, 201210, China;Drug Discovery and Design Center, State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai 201203, China
12. iHuman institute, ShanghaiTech University, Shanghai, 201210, China
折叠

Abstract

Generative models for structure-based molecular optimization (SBMO) need supervision that reflects real-world hit-to-lead optimization. Here we introduce LeadFlow, which combines maximum common substructure (MCS) supervision from assay-consistent chemical series, Bayesian Flow Networks for joint generation of continuous 3D coordinates and discrete molecular graphs, and experimental activity-based Direct Preference Optimization (DPO). On MolGenBench, LeadFlow recovered reference active scaffolds in approximately 400 of 600 series and active molecules in approximately 120 series, with active-recovery metrics improving by 1.33- to 2-fold over the best baseline. Scaffold-definition ablations supported improved active chemotype coverage and fused-ring elaboration. Experimental activity-based DPO enriched the generation of higher-activity molecules relative to the pretrained model and a docking-score-based counterpart. In a prospective ClpP agonist optimization campaign, generation followed by prioritization yielded five synthesized compounds with lower mean EC values than the starting compound A1. DC-ClpP-03 reached 33.4 nM, an approximately 180-fold improvement. These results support aligning training supervision and preference objectives with real-world hit-to-lead optimization as a practical strategy for SBMO.

关键词

Structure-based molecule optimization/Hit-to-lead optimization/Molecular generative models/Bayesian flow networks/Preference optimization/ClpP

Key words

Structure-based molecule optimization/Hit-to-lead optimization/Molecular generative models/Bayesian flow networks/Preference optimization/ClpP

引用本文复制引用

Runze Zhang,Jia Li,Hao Zhou,Mingliang Wang,Xiaomin Luo,Duanhua Cao,Mingyue Zheng,Qi Huang,Beijing Chen,Xia Sheng,Xinyu Jiang,Lehan Zhang,Xingyou Wang,Yaoyu Zheng,Zhehuan Fan,Dan Teng,Jie Yu,Keyue Qiu,Mingan Chen.LeadFlow aligns generative priors with real-world hit-to-lead optimization patterns for structure-based molecular optimization[EB/OL].(2026-09-23)[2026-09-26].https://chinaxiv.org/abs/202609.00304.

学科分类

化学

首发时间： 2026-09-23

下载量：0

点击量：14

段落导航