<p>Accurate classification of breast cancer subtypes is critical for prognosis and personalized treatment strategies. The PAM50 gene signature, consisting of 50 genes, serves as a widely accepted molecular standard for subtype classification. In this study, we propose a robust stacked ensemble learning framework that integrates heterogeneous base learners, Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and eXtreme Gradient Boosting (XGBoost), to classify PAM50 subtypes from high-dimensional gene expression data. The outputs of the base models are concatenated into meta-features, which are then passed to a Linear Support Vector Machine (SVM) meta-classifier for the final prediction of the stacked ensemble framework. Extensive evaluation on the METABRIC microarray data (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$n=1756$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>=</mo> <mn>1756</mn> </mrow> </math></EquationSource> </InlineEquation>) and TCGA-BRCA RNA-Seq (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$n=1148$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>=</mo> <mn>1148</mn> </mrow> </math></EquationSource> </InlineEquation>) data demonstrates that our stacked ensemble model achieves highly competitive performance with 98.31% accuracy on the METABRIC test set. On the external TCGA-BRCA cohort, it generalized with 95.88% accuracy, surpassing several recently reported approaches. Our framework achieves competitive performance and exceeds several previously reported approaches, while acknowledging that cross-study differences in preprocessing and evaluation settings may affect direct comparability. Further ablation studies and interpretability analyses using SHapley Additive exPlanations (SHAP) confirm the complementary strengths of the individual learners, with GRU contributing most significantly by providing 51.7% of the overall meta-classifier weight. This ensemble framework not only improves classification performance across heterogeneous transcriptomic datasets but also offers transparent decision-making, paving the way for future multi-omics integration in breast cancer subtype analysis.</p>

A stacked ensemble meta-classifier framework for PAM50 breast cancer subtype prediction from gene expression signatures

期刊论文
发表日期：

2026年3月7日

查看全文

Ekarsi Lodh,
Tapan Chowdhury,
Asim Shaw,
Debosmita Bedajna,
Saptarshi Bera,
Shalini Majumder,
Manashi De

摘要

Accurate classification of breast cancer subtypes is critical for prognosis and personalized treatment strategies. The PAM50 gene signature, consisting of 50 genes, serves as a widely accepted molecular standard for subtype classification. In this study, we propose a robust stacked ensemble learning framework that integrates heterogeneous base learners, Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and eXtreme Gradient Boosting (XGBoost), to classify PAM50 subtypes from high-dimensional gene expression data. The outputs of the base models are concatenated into meta-features, which are then passed to a Linear Support Vector Machine (SVM) meta-classifier for the final prediction of the stacked ensemble framework. Extensive evaluation on the METABRIC microarray data ( $n=1756$ $n = 1756$ ) and TCGA-BRCA RNA-Seq ( $n=1148$ $n = 1148$ ) data demonstrates that our stacked ensemble model achieves highly competitive performance with 98.31% accuracy on the METABRIC test set. On the external TCGA-BRCA cohort, it generalized with 95.88% accuracy, surpassing several recently reported approaches. Our framework achieves competitive performance and exceeds several previously reported approaches, while acknowledging that cross-study differences in preprocessing and evaluation settings may affect direct comparability. Further ablation studies and interpretability analyses using SHapley Additive exPlanations (SHAP) confirm the complementary strengths of the individual learners, with GRU contributing most significantly by providing 51.7% of the overall meta-classifier weight. This ensemble framework not only improves classification performance across heterogeneous transcriptomic datasets but also offers transparent decision-making, paving the way for future multi-omics integration in breast cancer subtype analysis.