COX-GAT: A Leakage-Controlled Graph Attention Network For Pathway Dysregulation Scoring And Multi-Omics Prognostic Biomarker Discovery in Clear Cell Renal Cell Carcinoma
Keywords:
Clear cell renal cell carcinoma; graph attention network; data leakage; pathway dysregulation; external validationAbstract
Clear cell renal cell carcinoma (ccRCC) shows marked prognostic heterogeneity beyond conventional staging, yet many multi-omics graph neural network studies risk information leakage and few are validated on independent cohorts. We introduce Cox-GAT, a leakage-audited graph attention framework that integrates transcriptomic, copy-number and clinical data on a protein–protein interaction graph and is trained under a Cox partial-likelihood objective. Edge weights, feature normalization and risk thresholds were derived from training data only and verified by automated checks; the pooled Kaplan–Meier stratification and the pathway ranking are exceptions, using the pooled out-of-fold median risk score. Across 15 patient-level cross-validation folds on TCGA-KIRC, Cox-GAT achieved a mean concordance index of 0.7697 with significant stratification in every fold. Pathway scoring of learned embeddings identified 31 prognostic Reactome pathways, led by cholesterol biosynthesis, SUMOylation and the TCA cycle, and gradient attribution nominated candidate prognostic genes including ANGPTL4, CYP51A1, ALDOB and AXIN2. Six of the ten top pathways and three of the four highlighted genes were consistent with prior renal cancer literature. Without refitting, the models generalized to an independent CPTAC-3 cohort (ensemble C-index 0.707). Under identical folds, Cox-GAT showed a modest advantage over DeepSurv, GCN-surv and GraphSAGE-surv. The largest ablation losses followed removal of the clinical covariates or the residual connection and random rewiring of the PPI graph, whereas the real graph exceeded no graph by only 0.0065 (untested). Gene selection used the full cohort and attribution a single fold. Cox-GAT offers a reproducible, externally validated multi-omics survival framework for ccRCC.

