Graph neural networks with molecular segmentation for property prediction and structure–property relationship discovery

Abstract

Graph neural networks (GNNs) have been widely used for predicting properties and discovering structure–property relationships in chemistry and drug discovery. However, current GNNs treat all atoms as independent of each other, neglecting the crucial role of functional groups in molecules. In this work, graph neural networks based on molecular segmentation are proposed. An unsupervised segmentation method is proposed to partition molecular graph data into multiple functional group-based clusters. Segmentation message passing neural network is proposed to learn function groups, which generate embeddings in and between molecule clusters. To explain structure–property relationships, we propose a new explainer to identify substructures that are more compatible with the chemical principles analysis. Our approach attempts to train and explain GNNs more rationally by segmenting molecules. Experimental results show that our approach achieves more efficient and rational performance prediction and explanation on chemical and drug datasets.