Leakage-controlled benchmarking of multi-omics patient-graph construction for pan-cancer tumor-type classification...

Pan-cancer multi-omics analysis requires models that integrate complementary molecular signals while preserving biologically meaningful relationships among patients. This study presents a leakage-controlled benchmarking framework for patient-graph learning in pan-cancer classification and prognosis analysis, focusing on how graph construction affects downstream performance. The benchmark explicitly separates fold-specific graph formation from downstream prediction. Using the TCGA Pan-Cancer cohort of 8204 primary tumors across 31 cancer types, RNA expression and copy-number variation data were used to compare early feature fusion, lightweight similarity network fusion (SNF-lite), and fused k-nearest neighbor similarity graphs under a common GATv2 encoder family with a matched attention-head search space and inner-validation selection procedure. A strict 5 × 3 nested cross-validation protocol ensured that imputation, gene selection, feature scaling, similarity computation, and neighbor search were fitted on training folds only. At G^'=2000, graph-level fusion approaches achieved about 0.92 accuracy and 0.89 Macro-F1, outperforming early fusion at about 0.89 accuracy and 0.84 Macro-F1. Fused kNN graphs also showed higher neighborhood label purity than SNF-lite despite similar predictive performance. A weighted topology audit showed that local label agreement alone did not determine graph utility. Gene and omics ablations showed that RNA carried the dominant subtype-discriminative signal, while CNV and mutation contributed weaker but complementary information. A Cox auxiliary objective retained classification performance when used alone and enabled out-of-fold prognostic stratification. These findings show that patient-graph construction is a key design choice in pan-cancer multi-omics learning and that leakage-controlled evaluation is essential for reliable and biologically informative benchmarking in computational oncology.