Reproducibility package for "A Systematic Evaluation of Feature Representations and Transfer Learning for Cross-Platform Breast Cancer Subtyping"

This record provides the materials supporting the study “A Systematic Evaluation of Feature Representations and Transfer Learning for Cross-Platform Breast Cancer Subtyping.” The study evaluated gene-level, pathway-level, PAM50, and a prespecified nine-module biological representation under strict cohort-disjoint transfer using 3,897 breast cancer samples from TCGA-BRCA, METABRIC, GSE20685, and GSE25066. The benchmark examined the effects of cohort composition, assay platform, preprocessing strategy, feature representation, classifier choice, and label provenance on cross-cohort subtype classification. Particular emphasis was placed on potential circularity introduced when subtype labels were completed using biological marker modules that overlapped conceptually or at the gene level with the evaluated module representation. The package therefore includes module-independent and reported-only label analyses, marker-exclusion analyses, and a repeated disjoint-marker audit in which genes used for external label completion were separated from genes used as predictive module features. The frozen disjoint-marker design comprised 3,840 intended configurations: 10 marker splits; 2 split directions; 16 cohort-disjoint transfer specifications; 2 preprocessing modes; 2 model classes; and 3 representation families. After exclusion of an unintended partial checkpoint repeat, 3,768 of 3,840 intended configurations were retained, corresponding to 98.1% completion. These included 1,256 matched module–gene comparisons and 1,256 matched module–PAM50 comparisons.