Explainable graph neural networks for drug-target interaction prediction: a gnnexplainer-augmented framework with molecular graph representation

Abstract

Drug-target interactions (DTIs) are important systems to predict, as they define the set of drug candidate molecules that are capable of binding efficiently to disease relevant protein target. The traditional computational methods use feature engineering on molecular fingerprint or similarity measures based on sequence level, which have limited ability in modeling the rich topological structure of the molecular graph. To solve the problems listed above, we propose a new explainable graph neural network framework, GNN-XAI, that unites multi-layer Graph Convolutional Networks (GCNs) with the GNNExplainer attribution method to merely ensure high-accuracy DTI prediction while providing biochemically interpretable explanations of binding site interactions. Molecules are drawn as attributed graphs, with the nodes drawn and their attributes painted, and the edges drawn and their types colored. Evaluation of the framework was conducted on BindingDB and ChEMBL-29 benchmark sets using AUC-ROC, AUPR and F1-score results of 0.961, 0.948 and 93.2 respectively were achieved. The plausibility of the biological fidelity of the derived subgraph attributions was verified by comparing to experimental crystallographic binding data in the GNNExplainer. Benchmark ablation experiments and generalization across datasets prove the robustness and clinical benefits of the proposed approach.