Does the Inclusion of Genomic Covariates Improve Survival Prediction in Colon Adenocarcinoma Patients? A Reassessment Based on Penalized Cox Regression Models

We critically reassess the claim that genome-wide RNA-seq features yield strong prognostic biomarkers in cancer survival analysis, focusing on colon adenocarcinoma (COAD) from The Cancer Genome Atlas (TCGA). The aim is to evaluate whether genomic information provides additional predictive value beyond standard clinical covariates for survival prognosis. Motivated by known pitfalls in observational omics studies, including dichotomization, confounding, multiple testing, and instability, we benchmark progressively richer genomic modeling strategies against a clinical-only baseline. We consider penalized Cox regression models for joint modeling of gene expression, specifically lasso, elastic net and overlapped group lasso with penalties. Across all approaches, genomic predictors do not provide measurable improvements in predictive performance over standard clinical covariates in the TCGA-COAD dataset. Penalized models tend to shrink genomic coefficients toward zero, and pathway-based models retain only the clinical covariate group. These findings suggest that readily available clinical variables capture most of the predictive signal in this context, while the additional contribution of genomic data appears limited when evaluated within a careful modeling and validation framework.