Incorporating targeted protein structure in deep learning methods for molecule generation in computational drug design

Abstract

Traditional drug discovery suffers from high costs and low productivity, with compounds frequently failing due to insufficient efficacy or off-target binding. Structure-based approaches aim to address these challenges by directly incorporating protein target information during molecule design, potentially reducing late-stage failures. In this review, we focus on current deep learning methods for structure-based drug discovery. We discuss the range of approaches used to encode and utilise protein structural information, from early shape-based approaches to more recent co-folding models that predict protein and ligand structures as a single task. We aim to provide insight into how deep learning approaches that incorporate structural information can be used to design molecules with enhanced binding potential while maintaining chemical and physical plausibility and offer suggestions as to the future directions of the field.