This paper demonstrates the use of hybrid chemical language models (CLMs) for de novo drug design, leveraging both molecular structure and bioactivity information. A generative CLM created a virtual library of PI3Kγ ligands, which was then refined using a CLM-based classifier trained on patented structures and PI3Kγ inhibitors. Several generated molecules were commercially available and tested, revealing a new sub-micromolar PI3Kγ ligand. Synthesis and testing of top-ranked molecules confirmed their potent activity, demonstrating the method's potential for scaffold hopping and hit-to-lead optimization. The most potent compounds inhibited PI3K-dependent Akt phosphorylation in a medulloblastoma cell model.

Michael Moret, Irene Pachon Angona, Leandro Cotos, Shen Yan, Kenneth Atz, Cyrill Brunner, Martin Baumgartner, Francesca Grisoni, Gisbert Schneider