Forecasting the molecular interactions: A hypergraph-based neural network for molecular relational learning

Abstract

Molecular relational learning refers to forecasting the interaction performance between pairs of molecules. Graph neural networks (GNNs) enables the effective modeling of chemical molecules by transferring them as abstract structures of graph, thus considering node-level interactions instead of atom’s. However, the abilities of modeling and predicting of classical GNN methods is restricted by edges’ redundancy and insufficient process of interaction modeling in a pair of molecular graphs. Therefore, based on the novel hypergraph convolutional neural network and bi-view cross interaction message passing and updating, we propose a hypergraph-based neural network method named IE-HGNN (Internal-External Bi-view Hypergraph Neural Network). We evaluate our method on two real-world molecular datasets, i.e., LEP and ZhangDDI, demonstrating IE-HGNN achieves excellent performance for both macro and micro molecules, while maintaining high training efficiency. We believe that this study introduces a novel approach for exploring the field of molecular relational learning and interaction prediction.