Abstract
Stratification of breast cancer (BC) into molecular subtypes by multigene expression assays is of demonstrated clinical utility. In principle, global RNA-sequencing (RNA-seq) should enable reconstructing existing transcriptional classifications of BC samples. Yet, it is not clear whether adaptation to RNA-seq of classifiers originally developed using PCR or microarrays, or reconstruction through machine learning (ML) is preferable. Hence, we focused on robustness and portability of PAM50, a nearest-centroid classifier developed on microarray data to identify five BC "intrinsic subtypes". We found that standard PAM50 is profoundly affected by the composition of the sample cohort used for reference construction, and we propose a strategy, named AWCA, to mitigate this issue, improving classification robustness, with over 90% of concordance, and prognostic ability; we also show that AWCA-based PAM50 can even be applied as single-sample method. Furthermore, we explored five supervised learners to build robust, single-sample intrinsic subtype callers via RNA-seq. From our ML-based survey, regularized multiclass logistic regression (mLR) displayed the best performance, further increased by ad-hoc gene selection on the global transcriptome. On external test sets, mLR classifications reached 90% concordance with PAM50-based calls, without need of reference sample; mLR proven robustness and prognostic ability make it an equally valuable single-sample method to strengthen BC subtyping.