Protein-Ligand Binding Predictions with SO(3)-Equivariant Neural Networks

Abstract

Virtual high-throughput screening of protein-ligand binding affinity is a valuable tool in computational drug design. Recently, structure-based machine-learning techniques have emerged as a powerful strategy for accurately predicting binding affinity. Simultaneously, SO(3)-equivariant neural networks can accurately predict physics-based properties of small molecules. The introduction of translation-and rotation-equivariant features allows the network to learn and exploit local geometry with fewer parameters than a full network. Here, we present recent advancements using an SO(3)-equivariant neural network to learn structure-based binding affinity predictions. Our network generates rotationally-equivariant features for both the ligand and receptors, based upon Tensor Field Networks with an equivariant attention mechanism. The ligand and receptor features are then combined in a single rotationally-equivariant interaction network. We apply our network to a variety of metrics against open-source screening benchmarks, and compare our performance with other strategies such as graph-based and voxel-based neural networks.