Target Specific Drug Design with Deep Reinforcement Learning
Theodore Beck Sternlieb · 2022
In this paper we present a deep learning approach to diversifying early leads in drug discovery.By learning a generative model over a space of know druglike small molecules we are able to take know leads and explore the molecular space around them to find chemically similar molecules.We do this by treating a small molecule as a graph where nodes represent di↵erent atoms and edges, bonds.Initially, we fit an autoregressive model P ✓ (•) over a dataset of small molecules which learns to build molecular graph atom by atom.We then bias P ✓ (•) using reinforcement learning to place higher probability densities on portions of the molecular space which score well on a set of reward functions.We choose reward functions which value druglike molecules as well as a reward for molecules which have a high binding a nity to a specific protein of interest.Finally after training the model, we sample molecules x ⇠ P ✓ (•|M g ✓ g x) where M g , an early lead molecule, is a subgraph of x, to generate novel drug candidates.