Pathway-based multi-omics survival modeling: evaluating weak-modality contributions and robustness across cancer types

Abstract

Multi-omics survival prediction must distinguish whether a modality enters a model from whether it improves prediction. This thesis evaluated the contributions of copy number variation (CNV) and somatic mutation (MUT) beyond clinical variables and ribonucleic acid (RNA) expression, together with the effects of pathway-based modeling and missing-modality training. Data from The Cancer Genome Atlas (TCGA) covered liver hepatocellular carcinoma (LIHC), breast invasive carcinoma (BRCA), and lower-grade glioma (LGG). Complete-case cohorts comprised 346 and 345 patients for LIHC progression-free interval (PFI) and overall survival (OS), respectively; 778 patients for each BRCA endpoint; and 503 patients for LGG OS. Gene set variation analysis (GSVA) mapped RNA expression to Hallmark pathways, each combined with seven CNV and four MUT summaries. Pathway-grouped composite minimax concave penalty (cMCP) Cox models were compared with Elastic Net using identical features and five frozen train–test splits. A gated residual model tested additional CNV/MUT information beyond a clinical-plus-RNA baseline. Paired bootstrap and permutation controls supported evaluation. Matched ensembles trained with and without modality dropout were tested under synthetic and natural missingness. Elastic Net assigned zero coefficients to at least one molecular block in every cancer–endpoint combination. Relative to same-feature Elastic Net, cMCP improved the mean concordance index (C-index) by 0.0374 for BRCA PFI and 0.0658 for BRCA OS, with both paired confidence intervals above zero. These discrimination gains did not consistently reduce absolute-risk prediction error. Conditional CNV/MUT increment was supported only for LGG OS, with a paired C-index increase of 0.0164 and a 95% confidence interval of [0.0051, 0.0300]; residual corrections were zero in all LIHC and BRCA splits. Modality dropout improved mean discrimination under synthetic missingness for LIHC and LGG but reduced it for BRCA, while natural-incompleteness results varied by endpoint. Exploratory graph ablations did not support a biological-topology advantage, and mediation analysis yielded 11 indirect statistical associations without establishing causality. The results separate penalization benefit, conditional modality information, and robustness to missing inputs. BRCA supported structured penalization, whereas LGG supported conditional weak-modality increment. These internally validated findings define endpoint-specific benefit boundaries without establishing universal model superiority or clinical utility.