De Novo molecular design
The de novo molecular design problem involves generating novel molecular structures or focused molecular libraries with desirable properties. It solves a so-called inverse design problem. The field of de novo molecular design has benefited tremendously from recent advances in ML. Within a very short time, numerous exciting approaches have been suggested. Notably, methods like recurrent neural networks (RNN), generative adversarial networks (GANs), and autoencoders were adapted to problems of rational design of organic and inorganic materials, synthesis planning, and device optimization.
We develop artificial intelligence (AI) method that enables the design of chemical libraries with the desired physicochemical and biological properties or both. We proposed a method called ReLeaSE for generating chemical compounds and focused chemical libraries with desired physical, chemical, and/or bioactivity properties that are based on deep reinforcement learning (RL). The general workflow for the ReLeaSE method includes generative (G) and predictive (P) neural networks. In this system, the generative model G is used to produce novel chemically feasible molecules, that is, it plays the role of an agent, whereas the predictive model P plays the role of a critic. P estimates the agent’s behavior by assigning a numerical reward (or penalty) to every generated molecule. The reward is a flexible function of the numerical property/activity of a generated molecule, and the generative model is trained to maximize the expected reward.