Integrating Medicinal Chemist Expertise with Deep Learning for Automated Molecular Optimization

J Med Chem. 2026 Feb 12;69(3):3609-3638. doi: 10.1021/acs.jmedchem.5c03746. Epub 2026 Jan 30.

Abstract

Successful compound optimization heavily relies on medicinal chemist expertise. In this work, we curated nearly 9000 molecular optimization strategies from the medicinal chemistry literature. Driven by expert knowledge, we constructed the MolOpt framework based on graph deep learning to expand these structural optimization strategies. Leveraging both expert-derived strategies and MolOpt, we developed AutoOptimizer, an automatic platform used for molecular optimization. To demonstrate the platform's practical application, we conducted case studies on fibroblast growth factor receptor 4 (FGFR4) and hematopoietic progenitor kinase 1 (HPK1). Experimental validation identified M8 and M9, which exhibited IC50 values of 17.6 and 46.5 nM against FGFR4 and HPK1, respectively, representing a 77.6-fold and 51.6-fold improvement over starting molecules. To our knowledge, this represents the first deep learning-generated molecular optimization strategy database grounded in the expertise of medicinal chemists. We anticipate that AutoOptimizer will provide valuable insights and accelerate lead optimization, thereby advancing drug discovery efforts.

MeSH terms

  • Chemistry, Pharmaceutical*
  • Deep Learning*
  • Drug Discovery / methods
  • Humans
  • Protein Kinase Inhibitors* / chemistry
  • Protein Kinase Inhibitors* / pharmacology
  • Protein Serine-Threonine Kinases / antagonists & inhibitors
  • Protein Serine-Threonine Kinases / metabolism
  • Structure-Activity Relationship

Substances

  • Protein Kinase Inhibitors
  • Protein Serine-Threonine Kinases