Graph Neural Networks for Network-Based Biomarker Discovery in Complex Diseases

Network-based biomarker discovery in complex diseases faces significant challenges due to the intricate interplay of molecular interactions and multi-omics data heterogeneity. This paper introduces a comprehensive graph neural network (GNN) framework that integrates protein-protein interaction networks with multi-omics datasets to identify disease-specific biomarkers. Our approach employs graph convolutional networks (GCN) with attention mechanisms and variational graph autoencoders for learning hierarchical representations of biological networks. We implement a multi-scale architecture combining node-level, subgraph-level, and graph-level features to capture local and global network topology. The framework achieves 94.7% accuracy in cancer type classification across 33 cancer types using TCGA datasets, 89.4% accuracy in Alzheimer’s disease biomarker identification from ROSMAP cohorts, and 93.5% precision in cardiovascular disease gene prediction. Validation on five independent datasets including TCGA, GTEx, STRING, DisGeNET, and ADNI demonstrates superior performance compared to traditional machine learning methods, with 15-40% improvement in F1-scores. The discovered biomarkers show strong biological relevance with pathway enrichment p-values below 10-15. Our framework enables interpretable biomarker discovery through attention weight analysis and graph explain-ability methods, facilitating translation to clinical applications.

Ҳали таржима қилинмаган