Development of protein-ligand scoring functions using 3D Convolutional Neural Networks

Molecular docking and virtual screening are indispensable computational tools in modern drug design. A crucial component of these methodologies is the scoring function (SF), employed to estimate the binding affinity between a small molecule (ligand) and a biological target of pharmaceutical interest (usually a protein). In recent years, machine learning techniques have gained prominence in SF development, surpassing traditional physics-based approaches. In this work, we present the development of SFs based on Convolutional Neural Networks using DockTDeep – a computer program developed by our research group. The preliminary results are competitive with other methods from the scientific literature, achieving R = 0.721 in CASF-2016 and R = 0.718 in CASF-2013 while keeping a simple molecular representation that requires no physics descriptors manually crafted by specialists. The results obtained in this work show the potential of this type of SF for virtual screening applications.