A deterministic, open reconstruction of the 21-gene breast-cancer recurrence score: external discrimination validation across nine public cohorts

Background. The 21-gene recurrence score (RS) is an established prognostic and predictive tool in early, hormone-receptor-positive (HR+), HER2-negative breast cancer. Independent, transparent re-implementations of the published RS methodology are valuable for reproducibility, education and as engineering substrates for downstream decision-support, but their analytical fidelity and discrimination must be demonstrated before use. We describe OncoBreast21, a deterministic, content-hash-stamped implementation of the published Paik (2004) 21-gene algorithm, and its external validation. Methods. OncoBreast21 v1.8.0 (deterministic; no machine learning in the score) was evaluated on nine independent public cohorts (METABRIC, TCGA-BRCA, and seven GEO series). Cohort expression (microarray or RNA-seq) was bridged to the engine's reference scale by a genefu-style quantile rescaling that emits a synthetic pre-image for the unmodified engine normalisation; the production scoring function was never altered. An independent implementation of the published Paik formula served as an internal control. Discrimination was assessed by Harrell's C-index (bootstrap 95% CI, 400 resamples), Cox proportional-hazards models (continuous per +10 points, by TAILORx band, and multivariable), multivariate log-rank across bands, and concordance with the reference (Spearman rho, Pearson r, Bland-Altman, Cohen's kappa). Reporting follows TRIPOD and REMARK. Results. Eight cohorts carried a survival endpoint (5,488 in-indication patients; 976 events). Across all eight, the engine's C-index was statistically indistinguishable from the independent reference implementation (absolute difference <= 0.004): METABRIC 0.615 vs 0.616; GSE7390 0.678 vs 0.681; GSE17705 0.665 vs 0.667; GSE12093 0.633 vs 0.628; GSE6532 0.591 vs 0.593; GSE11121 0.610 vs 0.609; SCAN-B 0.570 vs 0.572; TCGA-BRCA 0.544 vs 0.548. Rank concordance with the reference was near-perfect (Spearman rho 0.997-0.999). The score added prognostic information beyond nodal status, grade and age on five independent cohorts (e.g., METABRIC multivariable hazard ratio 1.095 per +10 points, p = 1.2x10^-8). The TCGA-BRCA null was reproduced by the reference formula on the same patients, indicating an underpowered endpoint rather than an engine defect. Conclusions. OncoBreast21 faithfully reproduces the published 21-gene methodology and its discrimination across diverse platforms, cohorts and endpoints. Discrimination is validated; because cohort scores are platform-relative reconstructions, absolute-scale calibration and band cut-points are reported separately and are the subject of an ongoing concordance study. The implementation is investigational / research-use-only and is not interchangeable with any cleared clinical assay.