Molecular inverse design using fragment-wise reinforcement learning

Abstract

Strategies to design novel molecules and materials with desired properties are the ultimate goal of chemical research. Reinforcement learning is a machine learning method that learns optimal action for each state using only the reward function without prior knowledge and is drawing attention as a molecular inverse design strategy that goes beyond human intuition. This study proposes a reinforcement learning model that designs molecules using the actions of adhering two rings at the fragment level. In the task of generating a molecule with a designated partition coefficient (logP), it showed higher accuracy than previous models that design molecules at the atomic level, and overcome the problem of creating unstable molecules. In addition, even in the partition coefficient optimization process that took into account the synthetic accessibility score, it showed higher accuracy than the atom-wise model and generated a chemically more stable molecule containing a ring. Our molecular design strategy is expected to be useful in industrial fields that utilize the characteristics of cyclic compounds, such as new drug development or organic light-emitting diodes.