Background The PAM50 classifier predicts breast cancer prognosis but requires 50 genes and specialised platforms. We derived CorePAM, the smallest data-driven PAM50 subset maintaining non-inferior prognostic performance relative to a full 50-gene Cox elastic-net model, without pre-specifying gene count. Methods Cox elastic-net regression (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\alpha = 0.5\)</EquationSource></InlineEquation>) with deterministic 10-fold cross-validation was applied in the SCAN-B cohort (N = 3069; GSE96058). Gene selection followed a pre-specified non-inferiority margin (<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(\Delta \)</EquationSource></InlineEquation>C-index = 0.010). External validation used four independent cohorts: TCGA-BRCA (N = 1072; RNA-seq), METABRIC (N = 1978; microarray; disease-specific survival), GSE20685 (N = 327; microarray), and GSE1456 (N = 159; microarray). Incremental value over a clinical model (CORE-A: age and ER status) was assessed by bootstrap <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(\Delta \)</EquationSource></InlineEquation>C-index. Secondary analyses evaluated pathologic complete response (pCR) prediction in four neoadjuvant cohorts (N = 697) plus I-SPY2 (N = 986). Results CorePAM comprises 24 genes with an out-of-fold C-index of 0.670 (gap vs 50-gene maximum: 0.009). The score was independently associated with survival in all validation cohorts: TCGA-BRCA (HR = 1.20), METABRIC DSS (HR = 1.41), GSE20685 (HR = 1.40), GSE1456 (HR = 1.71); all <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(p < 0.02\)</EquationSource></InlineEquation>. Random-effects meta-analysis (K = 4) yielded pooled HR = 1.37 (95% CI 1.24−1.52; <InlineEquation ID="IEq5"><EquationSource Format="TEX">\(I^2\)</EquationSource></InlineEquation> = 38.2%; <InlineEquation ID="IEq6"><EquationSource Format="TEX">\(p = 1.6 \times 10^{-9}\)</EquationSource></InlineEquation>). CorePAM provided incremental value beyond CORE-A and remained significant after adjustment for T-stage and nodal status. For pCR, pooled OR 1.69 (95% CI 1.39−2.05; <InlineEquation ID="IEq7"><EquationSource Format="TEX">\(I^2\)</EquationSource></InlineEquation> = 0%; <InlineEquation ID="IEq8"><EquationSource Format="TEX">\(p = 1.9 \times 10^{-7}\)</EquationSource></InlineEquation>). Conclusions CorePAM—a 24-gene PAM50-derived score—met the pre-specified <InlineEquation ID="IEq9"><EquationSource Format="TEX">\(\Delta \)</EquationSource></InlineEquation>C-index non-inferiority criterion relative to a 50-gene Cox elastic-net comparator in derivation and retained prognostic discrimination across four external RNA-seq and microarray cohorts, remained significant after anatomical staging adjustment, and was associated with pCR in secondary neoadjuvant analyses. This reduction from 50 to 24 genes may simplify future assay development, although platform-specific analytical validation is required before clinical deployment.

Corepam: a 24-gene PAM50-derived expression score with cross-platform external validation for breast cancer prognosis

期刊论文
出版模式：

开放获取
发表日期：

2026年5月21日

查看全文

Rafael de Negreiros Botan,
João Batista de Sousa

摘要

Background

The PAM50 classifier predicts breast cancer prognosis but requires 50 genes and specialised platforms. We derived CorePAM, the smallest data-driven PAM50 subset maintaining non-inferior prognostic performance relative to a full 50-gene Cox elastic-net model, without pre-specifying gene count.

Methods

Cox elastic-net regression (\(\alpha = 0.5\)) with deterministic 10-fold cross-validation was applied in the SCAN-B cohort (N = 3069; GSE96058). Gene selection followed a pre-specified non-inferiority margin (\(\Delta \)C-index = 0.010). External validation used four independent cohorts: TCGA-BRCA (N = 1072; RNA-seq), METABRIC (N = 1978; microarray; disease-specific survival), GSE20685 (N = 327; microarray), and GSE1456 (N = 159; microarray). Incremental value over a clinical model (CORE-A: age and ER status) was assessed by bootstrap \(\Delta \)C-index. Secondary analyses evaluated pathologic complete response (pCR) prediction in four neoadjuvant cohorts (N = 697) plus I-SPY2 (N = 986).

Results

CorePAM comprises 24 genes with an out-of-fold C-index of 0.670 (gap vs 50-gene maximum: 0.009). The score was independently associated with survival in all validation cohorts: TCGA-BRCA (HR = 1.20), METABRIC DSS (HR = 1.41), GSE20685 (HR = 1.40), GSE1456 (HR = 1.71); all \(p < 0.02\). Random-effects meta-analysis (K = 4) yielded pooled HR = 1.37 (95% CI 1.24−1.52; \(I^2\) = 38.2%; \(p = 1.6 \times 10^{-9}\)). CorePAM provided incremental value beyond CORE-A and remained significant after adjustment for T-stage and nodal status. For pCR, pooled OR 1.69 (95% CI 1.39−2.05; \(I^2\) = 0%; \(p = 1.9 \times 10^{-7}\)).

Conclusions

CorePAM—a 24-gene PAM50-derived score—met the pre-specified \(\Delta \)C-index non-inferiority criterion relative to a 50-gene Cox elastic-net comparator in derivation and retained prognostic discrimination across four external RNA-seq and microarray cohorts, remained significant after anatomical staging adjustment, and was associated with pCR in secondary neoadjuvant analyses. This reduction from 50 to 24 genes may simplify future assay development, although platform-specific analytical validation is required before clinical deployment.