Machine Learning Classification of Breast Cancer Molecular Subtypes and Clinical Phenotypes from GEO Expression Cohorts, with a Leakage-Controlled External Validation Protocol on GSE1456

Preservation copy of an article published in International Journal of Oncology Research. Read the full article: https://ioro.org/ijor/article/604478159518/604478159518.

Background. Breast cancer comprises biologically distinct intrinsic subtypes with different outcomes. Public GEO cohorts allow reproducible computational classification, but small test sets, heterogeneous labels and information leakage often inflate reported performance. This study classified breast cancer subtypes and clinical phenotypes from expression data and tested whether subtype classification transfers to an independent cohort. Methods. Four Affymetrix cohorts (GSE20685, GSE45827, GSE2034, GSE7390) were processed through a common pipeline: probe-to-gene annotation, quality control, differential-expression (DEG) feature selection on the training split only, and Logistic Regression, SVM and Random Forest classification (80/20 stratified hold-out), each against its own available label. A model trained on GSE45827 with nested cross-validation (CV) was then tested once on GSE1456 as an independent external cohort. Results. Hold-out accuracy was 90.9–92.4% for six molecular clusters (GSE20685), 96.2–100% for four subtypes (GSE45827), 70.7–79.3% for bone relapse (GSE2034) and 80.0–90.0% for ER status (GSE7390); in GSE2034 no model clearly exceeded the 75.9% majority-class baseline. On GSE1456 (102 labelled tumours), the SVM selected by nested CV (97.4% internal accuracy) reached only 48.0% accuracy (bootstrap 95% CI 38–58; majority baseline 38.2%). Basal (recall 0.88) and HER2 (0.80) tumours partly transferred; luminal A (0.21) and luminal B (0.30) did not. Conclusion. Within a cohort, expression-based classification is accurate for subtype and ER labels but not for relapse. Subtype classification did not transfer reliably to an independent cohort: external accuracy (48.0%) was far below the internal estimate (97.4%).