Reinforcement Learning for Molecule Design: A Taxonomy of Algorithms, Benchmarks, and Road to Synthesizable Leads.

Abstract

Reinforcement Learning (RL) has been in the forefront in traversing the vast chemical space for novel drug discovery. In this survey, we discuss the landscape of applying RL in two related paradigms: de novo drug design, the process of generating novel drugs from scratch and drug structure optimization, an approach to enhance the drug-likeness of an existing drug, with a case study for the latter.We describe the field along a taxonomy spanning value-based, policy-gradient, reward-proportional, and search-based methods, highlighting how molecular representation methods have evolved from pioneering RNN-based approaches to the latest transformer-based language models. Language models are treated as a cross-cutting layer across these families, most recently as agents that acton a molecule through chemical tools. We critically assess the frequently used datasets for training the models and benchmarks to evaluate the performance of those models across various drug generation and structure optimization tasks, arguing that strong benchmark performance does not always translate into real-world utility. We also showcase the recurrent challenges in modeling drugdesign algorithms and possible mitigation, with a final discussion on the overall developmental landscape in this field with future prospects.

Abstract as supplied by the indexing source. © the authors / publisher.