Multi-Level Attention Graph Neural Network for Clinically Interpretable Pathway-Level Biomarkers Discovery
Precision medicine, regarded as the future of healthcare, is gaining increasing attention these years. As an essential part of precision medicine, clinical omics have been successfully applied in disease diagnosis and prognosis using machine learning techniques. However, existing methods mainly make predictions based on gene-level individual features or their random combinations, none of the previous work has considered the activation of signaling pathways. Therefore, the model interpretability and accuracy are limited, and reasonable signaling pathways are yet to be discovered. In this paper, we propose a novel multi-level attention graph neural network (MLA-GNN), which applies weighted correlation network analysis (WGCNA) to format the omic data of each patient into graph-structured data, and then constructs multi-level graph features, and fuses them through a well-designed multi-level graph feature fully fusion (MGFFF) module to conduct multi-task prediction. Moreover, a novel full-gradient graph saliency mechanism is developed to make the MLA-GNN interpretable. MLA-GNN achieves state-of-the-art performance on transcriptomic data from TCGA-LGG/TCGA-GBM and proteomic data from COVID-19/non-COVID-19 patient sera. More importantly, the proposed model's decision can be interpreted in the signaling pathway level and is consistent with the clinical understanding.