Deep learning of protein-ligand interactions – remembering the actors

One of the intensely investigated applications of deep learning in drug design is the prediction of compound potency (affinity) based upon three-dimensional structures of protein–ligand complexes. Consistently accurate ligand binding affinity predictions would represent a milestone event for the field and put structure-based ligand design on a new level. For this purpose, convolutional neural networks (CNNs) with voxel representations of ligand binding sites as well as graph neural networks (GNNs) including message passing neural networks (MPNNs) are applied. GNNs/MPNNs learn directly from molecular graphs. In general, MPNNs are becoming increasingly popular for representation learning in chemistry. For affinity predictions using GNNs/MPNNs, protein–ligand complex structures are translated into interaction graphs.