De Novo Molecular Generation of Molecules with Consistent Synthetic Strategy

Abstract

The discovery of new pharmaceutical compounds is essential for developing novel medicines, a process that is both costly and time-consuming. In this context, generative Artificial Intelli gence (AI) methods offer a valuable, cost-effective approach for generating large amounts of de novo pharmaceutical candidates. However, these drug candidates must still be synthesized for use in in vitro and in vivo studies, necessitating the creation of efficiently synthesizable drug candidates. In this thesis, we explore two AI models for drug design developed by As traZeneca: REINVENT and AiZynthFinder. REINVENT is a generative AI model designed for creating small molecules fine-tuned with Reinforcement Learning (RL), whereas AiZyn thFinder is a framework for retrosynthesis. We explore the combination of these models through two novel reward functions and demonstrate that REINVENT, through querying of AiZynthFinder, can be conditioned to generate molecules that share an identical synthetic route. This is done in two ways, by provision of a template reaction, or by letting REINVENT independently find identical routes, allowing the generated molecules to be synthesized in parallel.