Comparing Voxel-Based and Graph-Based Convolutional Neural Network Models to Quantitatively Predict Binding Affinity Globally Across Structures and Ligands

Abstract

After identifying and validating initial hits, drug discovery progresses through hit-to-lead and lead-optimization stages. Quantitatively predicting binding affinity accurately could greatly accelerate the drive to higher potency, opening a bigger window for addressing other drug endpoints. A global pKi prediction model across targets and ligands can bridge the gap between the initial hits and more detailed models created after target-specific data becomes available. Recently in the virtual screening hit identification arena, voxel-based convolutional neural networks have been rivaled by graph convolution network architectures. Here we direct improved versions of both architectures to the difficult task of regression over global pKi prediction. To access performance across unseen chemical series, we use a broad matched-molecular pair based benchmark. To date, our wider voxel-based architectures are edged out by our latest graph convolutional networks. We expand the study by comparing how target-based and ligand-based splits frame the prediction difficulty. Finally, we ask whether including less quantitative but more numerous data points like pKi inequalities can improve our predictions of quantitative pKi.