Deep Reinforcement Learning for Optimized Drug Design

Authors

Abstract

In de novo drug design, computational strategies are used to create new molecules with bespoke properties that have a good affinity towards the desired biological target. In this study, we present a model for molecular de novo drug design that utilizes a Reinforcement Learning (RL) framework. This is a policy-based approach that captures the syntax of molecular representation in terms of SMILES string and generates potential new compounds with certain specified desirable properties. The proposed RL approach consists of two neural networks that act as a generator of new compounds and predictor of the property in question, respectively. The property may come from a physicochemical or biological point of view, and a QSAR model must be established to allow for regression between each generated structure and the desired property. First, the generative model is trained to generate valid molecules. Then, the policy gradient method of RL is applied to make the model produce fine-tuned molecules. We performed statistical experiments demonstrating the efficiency of the proposed strategy in a single task regime where each endpoint of interest is independently optimized. Nevertheless, this approach can be expanded simultaneously to allow multi-objective optimization of several target properties, which is the need for drug discovery where the drug molecule should be optimized with respect to the properties of potency, selectivity and pharmacokinetic properties to avoid unwanted side effects and minimize toxicity. Our future studies will address this issue.

Keywords

Drug Discovery, Deep Algorithm, Reinforcement Learning

Subject

Generating Drug Compounds with Deep Reinforcement Algorithm

Related Project

D4 - Deep Drug Discovery and Deployment (PI: Bernardete Ribeiro; co-PI: Joel P. Arrais)

Conference

Bioinformatics Open Days, February 2020

PDF File