Machine Learning Modeling for Identifying Cancer Subtypes From Gene Expressions

Cancer is a genetically heterogeneous disease, where diverse gene expression patterns complicate diagnosis and treatment. This study presents a machine learning framework using K ‐Means clustering, recursive subclustering, and statistical feature selection to explore candidate subtype–like transcriptomic structure from ribonucleic acid (RNA) gene expression data. Leveraging a subset of The Cancer Genome Atlas (TCGA), comprising 10,095 samples across 19 cancer types and over 13,000 genes, we implement a recursive clustering process that refines mixed clusters into more homogeneous groups and examines whether coherent subtype–like structure emerges. Feature importance in the dataset samples is analyzed using logistic regression, tree‐based methods, and principal component analysis (PCA) to identify genes that contribute most strongly to the separation of the resulting groups. We additionally benchmark the proposed pipeline against alternative clustering methods (agglomerative Ward, spectral, and a spectral‐seeded K ‐Means hybrid) and quantify robustness via seed‐stability and stratified hold‐out experiments. After recursive refinement, the dominant subclusters reach approximately 86%–90% dominant‐label purity against the original TCGA cancer labels, providing an internal consistency check rather than a supervised accuracy measure; single‐pass partition metrics are lower and are reported separately. Rather than a diagnostic tool, the framework provides a reproducible pipeline for hypothesis generation and subsequent biological validation of transcriptomic structure within and across cancer types.