Motif-Based Graph Neural Networks for Predicting Quantum Molecular Properties

Comprehensive exploration of chemical compound space (e.g., geometric, energetic, electronic, and thermodynamic properties) is often required in designing new drugs and materials. However, traditional quantum modelling methods (e.g., density-functional theory) for calculating molecule properties suffer from high computational complexity as they scale combinatorially with molecular size. Recently, various graph neural network models have been developed for fast estimation of molecule properties. However, their interpretability has yet to be improved. We have developed a motif-based graph neural network model that offers better interpretability and is able to better predict quantum properties of chemical molecules. Our approach automatically learns a set of motifs from molecules, which are represented as attributed relational graphs, and utilizes the learned motifs to produce context-aware embeddings of atoms, which summarize local molecular structures (e.g., functional groups).