Deep Learning-Ready Voxel Representation of Protein-Ligand Complexes from an Enhanced PBDbind v.2020 Dataset

Abstract

A critical aspect of successful deep learning (DL) modelling in computer-aided drug discovery (CADD) is the representation of biomolecular data. Voxel grid representations have emerged as a straightforward method for depicting 3D molecular structures of protein-ligand complexes. Proper structural preparation of these complexes is also crucial, particularly in models where the orientation of hydrogen atoms and the accurate assignment of protonation/tautomeric states are vital. The PDBbind, a widely used dataset, can be improved in this regard. This work presents an enhanced version of the PDBbind v.2020 refined set concerning structural preparation, a voxel representation of these structures suitable for DL model training and a diverse set of docking-generated poses that could be used to develop new scoring functions for pose prediction. We also introduce DockTGrid, a software library developed to generate these voxel representations, which can be adapted to create new molecular features. With this work, we aim to provide the CADD community with high-quality, accessible resources to facilitate the development of DL models for drug discovery.