SPOTLIGHT: structure-based prediction and optimization tool for ligand generation on hard-to-drug targets – combining deep reinforcement learning with physics-based de novo drug design

Abstract

We present SPOTLIGHT, a proof-of-concept for a method capable of designing a diverse set of novel drug molecules through a rules-based approach. The model constructs molecules atom-by-atom directly at the active site of a given target protein. SPOTLIGHT does not rely on generation cycles and docking/scoring to optimize its molecules and requires no a priori information about known ligands as the molecule construction is purely based on classical interactions. We patch the model with deep Reinforcement Learning (RL) using a Graph Convolution Policy Network (GCPN) to tune molecule-level properties directly during the generation phase. Our method has shown promising results when applied to the ATP binding pocket of the well-studied HSP90 protein. We show that our model upholds diversity while successfully producing strong binders to the protein. Given the stochasticity at each step, we do not expect it to reproduce known ligands exactly. However, we show how it uses significant fragments of known ligands as substructures while also providing an alternate way for tuning between similarity and novelty.