A systematic evaluation of feature representations and transfer learning for cross-platform breast cancer subtyping

Cross-platform breast cancer subtyping remains challenging because performance can change across independent cohorts, assay platforms, preprocessing strategies, and label definitions. We benchmarked gene-level, pathway-level, PAM50, and a prespecified nine-module biological representation under strict cohort-disjoint transfer using 3897 patients from TCGA-BRCA, METABRIC, GSE20685, and GSE25066. Because label provenance can couple ground truth to evaluated features, inference was separated across reported-only labels, an externally-completed-label-excluded set treated as a partial procedural control, and a repeated disjoint-marker audit. Under reported-only labels, mean macro-AUROC was 0.949 for modules, 0.971 for genes, 0.973 for PAM50, and 0.961 for pathway rank scores; paired module-minus-gene and module-minus-PAM50 effects were − 0.022 and − 0.024. The frozen disjoint-marker audit retained 3768 of 3840 intended configurations (98.1%), comprising 1256 configurations per representation, and yielded module-minus-gene and module-minus-PAM50 differences of − 0.036 and − 0.039. In the externally-completed-label-excluded set, removing ESR1, PGR, and ERBB2 from the module definitions reduced module macro-AUROC from 0.936 to 0.928 (difference − 0.0087; 95% transfer-specification bootstrap CI − 0.0121 to − 0.0055), while depleted-module differences versus genes and PAM50 had confidence intervals spanning zero. After restricting every representation to nine ranked features, mean top-5 Jaccard was 0.774 for modules versus 0.229 for genes, 0.454 for PAM50, and 0.389 for pathways; corresponding RBO@9 values were 0.903, 0.281, 0.580, and 0.446. These dimension-matched metrics support attribution concentration and cross-transfer stability, not validated clinical interpretability. Overall, representation rankings depended on label provenance and marker overlap; biologically structured modules remained compact and attribution-stable but did not provide universally superior discrimination.