Benchmarking protein pocket-aware molecular generative tools

Abstract

Protein pocket-aware molecular generative models have been an emerging approach to aid structure-based drug design to develop small molecules that can fit and bind into the protein target of interest. However, there is a lack of validation of these models on clinically relevant targets. In this FYP, we had evaluated five generative models (Pocket2Mol, PocketFlow, Lingo3DMol, DiffSBDD, and PMDM) to generate de novo molecules against 5-hydroxytryptamine receptor 2C G protein-coupled receptor (GPCR), dopamine receptor D2 GPCR, cyclin-dependent kinase 2, and glycogen synthase kinase 3β. We found that while most generated ligands were structurally unique and valid, they did not have a valid ligand conformation with invalid bond length, bond angle, and high intramolecular steric clashes. Furthermore, many had varying physicochemical properties which were non-drug-like. Despite these limitations, DiffSBDD, Lingo3DMol, and PMDM were able to generate very few unique ligands that retained important interactions with GPCRs and kinases with favourable physicochemical properties.