PNNL

Abstract

Protein-ligand interactions (PLI) are fundamental to biochemical research. We build upon the recently introduced distance-aware, deep-learning approach to integrate knowledge representation and reasoning for PLI prediction by incorporating atomic level information into a graph-based embedding. We use 3D structural information in two distinct graph neural network (GNN) architectures to predict the activity of a protein-ligand pair and then modify the network to accommodate regression problems. GNNF is the base implementation which employs distinct featurization to enhance domain-awareness, while GNNP is a novel implementation which can predict with no prior knowledge of the intermolecular interactions. We achieved top performance with 0.979 test accuracy for GNNF and 0.958 for GNNP. These models are further adapted for regression tasks to predict experimental binding affinities and pIC50. Using PDBbind2016 data, we achieve a Pearson r of 0.66 and 0.65 on experimental affinity and 0.50 and 0.51 on pIC50 with GNNF and GNNP, respectively, outperforming the existing models. Most importantly, the GNNP model can be used to screen a large ligand library and predict biophysical properties against a given protein target.