P07-28
Validation of the reproducibility of hit-to-candidate using ChemTS
Tomoki YONEZAWA *1, Masateru OHTA2, Shoichi ISHIDA3, Kei TERAYAMA3, Teruki HONMA2, Kazuyoshi IKEDA1, 2
1Faculty of Pharmacy, Keio University
2R-CCS, RIKEN
3Graduate School of Medical Life Science, Yokohama City University
( * E-mail: yonezawa-tm@pha.keio.ac.jp )
Various structure generation methods have been developed, and their application to drug design is expected to enhance the efficiency of the drug development process. In this study, we have attempted to generate the chemical structures of marketed drugs and clinical candidates using AI to verify the applicability of structure generation to drug discovery. To make this verification practical, we aimed to reproduce the structure development process from hit compounds to approved drugs or clinical candidates.
Structure development from hit compounds often involves multi-objective optimization to improve not only the primary activity against the target, but also physical properties such as solubility and membrane permeability, as well as pharmacokinetic properties. Among structure generation AIs, ChemTS generates structures while performing reinforcement learning to improve the target parameters. We used a predictive model of physicochemical and ADME properties, and set the predicted value as the reward function. This enables the search for structures with improved predicted values. We combined the main activity evaluation by docking or 3D shape similarity with the predictive model of physicochemical and ADME properties such as solubility, membrane permeability, and metabolic stability, and used it as a reward to generate structures using ChemTS.
To find applicable examples of structure development that could be evaluated using the predictive model, we employed the following approach. A review paper published in the Journal of Medicinal Chemistry provided examples of successful structure development of approved drugs and clinical candidates from hit compounds. Hit-drug and hit-candidate pairs were extracted from the paper, and related activity information were also obtained from ChEMBL. When generating the structure of DORAVIRINE, an approved HIV reverse transcriptase inhibitor, we successfully reproduced its structure with docking, membrane permeability, solubility, and metabolic stability as rewards. In a subsequent verification test with TEPOTINIB, a c-Met inhibitor, we attempted to generate its structure with docking, membrane permeability, and metabolic stability as rewards. We successfully generated similar structures of highly active compounds other than TEPOTINIB. Based on these verification results, we will discuss the strengths and weaknesses of structure generation using ChemTS.