GTCExplainer: Interpretable Graph Convolutional Networks for Molecular Activity Prediction

ABSTRACT With the growing application of Graph Convolutional Networks (GCNs) across various domains, particularly in the field of bioinformatics, the demand for their interpretability has become more urgent. In bioinformatics, many critical tasks, including drug discovery and protein–ligand interaction analysis, rely heavily on molecular activity prediction, where functional substructures play a decisive role in shaping molecular activities. However, existing explanation methods rely on the features of nodes and edges generated by the GCN model, while overlooking the critical structural information within the graph. On the other hand, how to ensure the compactness of the explanation methods also needs to be solved. To address these issues, we propose GTCExplainer in this paper. The Graph Transformer is introduced to capture more comprehensive graph data information, and a hierarchical edge selector is designed to select suitable edges for the explanation. In addition, a substructure reward mechanism is proposed to generate rewards that ensure the explanation's compactness while including key structures closely related to the molecular activity prediction of the GCN model. Experimental results on multiple biomolecular datasets show that our proposed GTCExplainer can effectively provide explanations for the GCN model on molecular activity prediction.