Tumor stratification remains a significant challenge due to the complex molecular landscape of cancer, as intertumoral heterogeneity can lead to different clinical outcomes and treatment responses. In this study, we propose an unsupervised Graph Neural Network (GNN)-based framework, applicable across omic modalities, to stratify cancer patients and uncover mechanisms underlying tumor heterogeneity. Our approach integrates network propagation via Random Walk with Restart (RWR) and permutation testing to construct patient-specific networks, which are then embedded using InfoGraph to obtain patient-level representations. To ensure robustness, we evaluate dimensionality reduction methods (PCA, t-SNE, UMAP) and clustering strategies (HDBSCAN, DBSCAN, affinity propagation), selecting the optimal configuration based on stability metrics (silhouette score, AMI, ARI). We then identify cluster-specific molecular patterns. In a case study of Glioblastoma Multiforme (GBM), we analyzed mutation profiles from 356 TCGA patients to construct patient-specific networks. We identified 11 clusters, four showing significant survival differences (p = 0.0018 and p = 0.036). Transcriptomic analysis revealed 221 differentially expressed genes (|logFC| > 2, p < 0.05) between identified clusters. To investigate underlying mechanisms, we applied a three-layer Graph Isomorphism Network (GIN) with PGExplainer, identifying cluster-specific subgraph motifs and pathways. One cluster was enriched in ion and membrane transport processes, while another showed enrichment in immune and signaling pathways. Overall, this framework combines unsupervised GNN-based representation learning, network propagation, and explainable AI to stratify tumors and assign biological meaning to clusters, providing insights into intertumoral heterogeneity.

Co-authors: Latif Hatipoglu, Nurcan Tuncbag