Integrative Transcriptomic Analysis of Breast Cancer Subtypes Using Consensus Gene Expression Modeling and Biological Pathway Decoding

Received: 17 December 2025 | Revised: 21 January 2026 and 13 March 2026 | Accepted: 15 March 2026 | Online: 23 May 2026

Corresponding author: Sumendra Yogarayan

Abstract

Breast cancer is among the three leading causes of cancer-related mortality in women, highlighting the importance of accurate molecular subtyping for treatment using personalized medicine. Despite major advances in transcriptomics, the analysis of gene expression data remains challenging due to high dimensionality and biological variability. This study proposes a reproducible computational framework for robust diagnosis of breast cancer subtypes based on gene expression profiles, using a subset of the GSE45827 comprising 120 tumor and normal tissue samples representing six molecular subtypes. The proposed pipeline incorporated rigorous preprocessing, consensus-based feature selection, comprehensive benchmarking across classical machine learning, ensemble learning, and deep learning models, feature selection combined with Shapley Additive Explanations (SHAP)-based importance analysis, Boruta, Tabular Network (TabNet) attention masks, and stability selection to identify biologically relevant and reproducible biomarkers. To minimize information leakage, nested cross-validation was employed throughout model development, while external validation was conducted using the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) cohort to evaluate generalizability across independent datasets. Among the evaluated approaches, the stacking ensemble classifier achieved the best overall performance, reaching mean accuracy and macro-F1 scores of 96.7% and 96.8%, respectively, on the GSE45827 dataset, and 94.2% and 94.0% on the METABRIC cohort. These results surpassed those obtained using TabNet, Autoencoder + Logistic Regression (AE+LR) pipelines, and Prediction Analysis of Microarray 50 (PAM50) baseline models. Moreover, the interpretability-based analysis identified i) several biologically significant genes, including Breast Cancer 1 (BRCA1), Tumor Protein p53 (TP53), and Phosphatidylinositol-4,5-Bisphosphate 3-Kinase Catalytic Subunit Alpha (PIK3CA), as key contributors to subtype discrimination, and ii) the pathways in which these genes are involved, such as the Phosphatidylinositol 3-kinase - Ak strain transforming (PI3K-Akt) signaling and Deoxyribonucleic Acid (DNA) repair. Overall, the proposed framework provides a reproducible, interpretable, and high-performing methodology for reliable feature identification in gene expression data.