Multi-Property De Novo Drug Design Using Multi-Objective Deep Reinforcement Learning

Drug design has been an expensive and slow but vital process for saving lives, especially in the presence of world pandemics such as COVID-19. To make the drug design process more cost-effective and efficient, researchers have been exploring the aid of computers and even artificial intelligence. In this work, we developed a deep reinforcement learning (RL) based multi-objective method to enable multi-property de novo drug design. The goal of such an algorithm is to make RL-based drug design processes more applicable in real life compared to single-property-driven design approaches. We have studied two multi-objective methods in our experiments: the weighted sum method and the lexicographic-order method. The experimental results showed that compared to the single-objective method, both methods can effectively improve the efficiency of the drug design process. We ran the multi-obj ective algorithms to increase Janus kinase 2 (Jak2) inhibition and hydrophobicity (LogP) drug-likeness and found that the lexicographic-order method (the most optimal method in our experiments) can effectively increase the J ak2 inhibition by 26.9% and LogP drug-likeness by 10.0% simultaneously compared to the baseline (unbiased drug design) while 54.5% of the generated compounds are chemically valid, which far outperforms previously proposed methods in literature such as the single-objective approaches. We also found that since the weighted-sum method is sensitive to the selection of different weights, the lexicographic method is more reliable. The proposed method can be easily extended to account for more properties. Overall, it will greatly reduce the time frame and costs of drug design.