Predicting Basal-Like Breast Cancer Using Pathway-Targeted Genomic Features and SMOTE-Enhanced Ensemble Models
Year: 2026
Type: conference-paper
Abstract: Predicting complex disease phenotypes often requires evaluating gene networks rather than isolated variants. This study asks whether pathway-level mutation profiles can predict the Basal-like breast cancer subtype in a real clinical cohort. Using The Cancer Genome Atlas (TCGA) PanCancer Atlas, we built a binary mutation matrix for 690 patients across the mTOR and TGF-β pathways and evaluated four classifier families. In a naive baseline, every model’s recall on the Basal-like class was low (11.4–31.8%) despite deceptively high overall accuracy. Addressing the cohort’s 17.2% class imbalance through feature filtering and in-fold SMOTE raised recall to 55.8–67.7% across all four models and produced a consistent Matthews Correlation Coefficient (MCC ≈ 0.30–0.34). SHapley Additive exPlanations (SHAP) identified PIK3CA as the dominant predictive driver, acting as a negative predictor of Basal-like status, consistent with its known low mutation frequency in this subtype, a finding cross-checked (though not independently validated) through exact real-variant mapping onto Chaos Game Representation (CGR) images of the surrounding genomic sequence. Six candidate expression genes were then selected independently inside every cross-validation training fold, avoiding the leakage a single cohort-wide selection would introduce, and integrating their mRNA expression with mutation features achieved the highest overall predictive performance (Random Forest: 85.7% accuracy, MCC 0.523, AUC 0.894). This real but honestly bounded signal motivates further study of a low-cost, pathway-targeted sequencing panel as an exploratory adjunct to expression-based subtyping.
Institution: Biruni University
Language: English
Subfield: Cancer Research
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7223667478