X-GCN: An Explainable and Uncertainty-Aware Graph Convolutional Framework for Multi-Omics Disease Risk Prediction

Abstract

To provide accurate and helpful insights, healthcare early illness risk prediction requires the integration of diverse multi-omics data. Unlike existing models (MOGONet, ExplainMix, CNNs), X-GCN presents an integrated explainable graph-based framework that combines graph convolution with hierarchical attention in a unified multi-omics supra-graph, enabling uncertainty-aware disease risk prediction and biologically grounded explanations. Multi-omics data such as transcriptomics, proteomics, genomics, and epigenomics are represented by X-GCN as graph data, where nodes replace biological entities and edges replace molecular interactions. X-GCN focuses on critical biomarkers and ensures prediction transparency through hierarchical attention techniques.X-GCN surpasses complex models such as MOGONet (88.7% accuracy, 0.912 AUC) and ExplainMix (90.2% accuracy, 0.920 AUC) with tests on The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) datasets, recording 92.4% accuracy, 0.935 AUC, and 0.91 F1-score. X-GCN also decreases model uncertainty by 18% and identifies experimentally validated biomarkers for cancer and cardiovascular disease prediction. X-GCN provides an interpretable and uncertainty-aware computational framework for multi-omics disease risk prediction, serving as a foundation for biomarker discovery and future clinical validation.