Identification of a Minimal, Generalizable Gene Expression Signature for Breast Cancer Subtyping Through Cross-Cohort Machine Learning Integration

Breast cancer is the most commonly diagnosed cancer among women and remains a leading cause of cancer-related mortality. Accurate classification into molecular subtypes—Luminal A, Luminal B, HER2-enriched, and Basal-like—is important because these subtypes differ in biology, prognosis, and treatment response. However, many gene-expression classifiers are developed using a single dataset or expression platform, limiting generalizability across patient populations and technologies. This study aimed to identify a smaller, biologically meaningful gene-expression signature that preserves molecular subtype information across independent cohorts and expression technologies, providing a computational foundation for future reduced-gene molecular assays.

Gene-expression data from The Cancer Genome Atlas (TCGA; RNA sequencing) and GEO dataset GSE81538 (microarray) were analyzed. Differentially expressed genes were identified using the Kruskal-Wallis test with false discovery rate correction, resulting in 761 genes shared between cohorts. Logistic regression, support vector machine, and random forest models were trained using this shared gene set and evaluated through reciprocal cross-cohort validation. Genes were ranked by model-derived importance, and multiple reduced signature sizes were evaluated. Intersecting the top 100 genes from each cohort produced a final 34-gene signature.

The 34-gene signature retained substantial predictive performance while reducing more than 18,000 measured genes to a compact panel. Reciprocal validation achieved macro-F1 scores of 0.816 and 0.761 across cohorts and platforms. Principal component analysis demonstrated subtype-related separation, while STRING protein–protein interaction analysis showed significant network enrichment (p < 1 × 10⁻¹⁶), supporting biological connectivity among selected genes.

These findings demonstrate the feasibility of cross-cohort and cross-platform feature selection for identifying compact molecular signatures. The 34-gene panel represents a candidate research signature requiring additional analytical, experimental, and clinical validation before potential clinical use.