Hierarchical graph neural network with subgraph perturbations for key gene cluster discovery in cancer staging

Abstract

Analyzing highly individual-specific genomic data to understand genetic interactions in cancer development is still challenging, with significant implications for the discovery of individual biomarkers as well as personalized medicine. With the rapid development of deep learning, graph neural networks (GNNs) have been employed to analyze a wide range of biomolecular networks. However, many neural networks are limited to black box models, which are only capable of making predictions, and they are often challenged to provide reliable biological and clinical insights. In this research, for sample-specific networks, a novel end-to-end hierarchical graph neural network with interpretable modules is proposed, which learns structural features at multiple scales and incorporates a soft mask layer in extracting subgraphs that contribute to classification. The perturbations caused by the input graphs' deductions are used to evaluate key gene clusters, and the samples are then grouped into classes to produce both sample- and stage-level explanations. Experiments on four gene expression datasets from The Cancer Genome Atlas (TCGA) show that the proposed model not only rivals the advanced GNN methods in cancer staging but also identifies key gene clusters that have a great impact on classification confidence, providing potential targets for personalized medicine.