<![CDATA[High-dimensional genomic datasets (p>n) pose persistent challenges for predictive modeling and biomarker-oriented feature selection due to multicollinearity and instability of selected feature sets under resampling. Although Elastic Net is widely used to address correlated predictors via combined L1/L2 regularization, the practical role of the L1/L2 mixing ratio (α) is often treated as a secondary tuning choice driven primarily by predictive accuracy. This study investigates how varying α shapes the trade-off among selection stability, solution sparsity, and predictive performance along the Elastic Net regularization path. Experiments were conducted using the publicly available METABRIC breast cancer cohort (n = 1,964) with 21,113 gene expression features and a binary overall survival status outcome. Logistic regression with Elastic Net penalty was fitted across a grid of α values, with the regularization strength (λ) selected by cross-validation. Feature selection stability was evaluated under repeated resampling using the Jaccard index, Dice coefficient, and Adjusted Rand Index (ARI), while sparsity was summarized by the average number of non-zero coefficients; predictive performance was assessed using AUC, accuracy, and F1-score. Results show a monotonic decline in stability as α increases: α = 0.2 yields the highest stability (Jaccard 0.324, Dice 0.487, ARI 0.434), whereas LASSO (α = 1.0) produces the lowest stability (Jaccard 0.278, Dice 0.431, ARI 0.400). In contrast, predictive performance varies only marginally across α (AUC 0.696–0.704; accuracy 0.666–0.671; F1-score 0.738–0.742), while sparsity changes substantially (average selected features 110–204). Coefficient path analyses further illustrate abrupt shrinkage under LASSO versus smoother, group-preserving shrinkage under Elastic Net, consistent with improved reproducibility under lower-to-moderate α. Frequency-of-selection analysis highlights genes repeatedly selected across resampling, supporting interpretability of stable configurations without claiming causal biomarker validity. Overall, the findings demonstrate that α is a substantive modeling choice that materially affects stability and sparsity even when accuracy is similar, motivating stability-aware tuning for high-dimensional genomic prediction and reproducible feature discovery.]]>
