END-TO-END INTERPRETABLE GRAPH LEARNING FOR PATIENT CLASSIFICATION

Genes don't operate in a vacuum - they operate in the form of complex networks. Traditional gene expression data analysis often includes the analysis of co-expression patterns to understand these interactions; however, most machine learning methodologies don't properly account for context-dependent relationships between input features. Here, we propose a novel latent graph learning framework, titled Learnable Graph Interaction Module (LGIM), that employs a differentiable graph module to learn interactions between genes. We test our model on seven TCGA cancer datasets, where it either outperforms or performs comparably to the baseline models while learning meaningful gene representations. Conducting an interpretability analysis on the learned gene interaction graph for breast cancer, we notice that the extracted nodes and edges of higher importance correspond to being more predictive, and to known protein-protein interactions respectively. Furthermore, the clusters in the learned graph corroborate with relevant biological pathways.