A Pilot Reproducibility and Robustness Assessment of a TCGA Gene-Expression Cancer Classifier

Preprint — not yet peer reviewed. This pilot computational study evaluates a TCGA gene-expression cancer classifier across six cancer types: BRCA, COAD, READ, LUAD, LUSC, and OV. The analysis focuses on patient-grouped cross-validation, training-only preprocessing, class-specific performance, tissue-source-site confounding, uncertainty, selective prediction, and the severe confusion between COAD and READ. The primary patient-grouped 1,000-tree Random Forest achieved 93.43% accuracy and 80.52% balanced accuracy, but READ recall was only 0.95%; 104 of 105 READ samples were predicted as COAD. An exploratory 100-gene linear SVM improved READ recall to 60.00% in the binary COAD–READ task, although it was not externally validated. The findings show why overall accuracy alone can conceal serious class-specific failures. External validation was not performed and is planned for future work. The author remains responsible for the analysis, code, figures, references, and interpretation. AI assistance was used for mentoring, code-drafting support, debugging suggestions, literature-search organization, and document editing.