A deep learning model for structure-based bioactivity optimization and its application in the bioactivity optimization of a SARS-CoV-2 main protease inhibitor

Abstract

Bioactivity optimization is a crucial and technical task in the early stages of drug discovery, traditionally carried out through iterative substituent optimization, a process that is often both time-consuming and expensive. To address this challenge, we present Pocket-StrMod, a deep-learning model tailored for structure-based bioactivity optimization. Pocket-StrMod employs an autoregressive flow-based architecture, optimizing molecules within a specific protein binding pocket while explicitly incorporating chemical expertise. It synchronously optimizes all substituents by generating atoms and covalent bonds at designated sites within a molecular scaffold nestled inside a protein pocket. We applied this model to optimize the bioactivity of Hit1, an inhibitor of the SARS-CoV-2 main protease (M^pro) with initially poor bioactivity (IC₅₀ : 34.56 μM). Following two rounds of optimization, six compounds were selected for synthesis and bioactivity testing. This led to the discovery of C5, a potent compound with an IC₅₀ value of 33.6 nM, marking a remarkable 1028-fold improvement over Hit1. Furthermore, C5 demonstrated promising in vitro antiviral activity against SARS-CoV-2. Collectively, these findings underscore the great potential of deep learning in facilitating rapid and cost-effective bioactivity optimization in the early phases of drug development.