High-dimensional gene expression datasets present unique challenges for cancer biomarker discovery and classification. Here, we propose a deep learning framework incorporating principal component analysis (PCA) for dimensionality reduction and a custom optimizer, PCAGroupAdam, which scales gradients by each principal component's explained variance. The framework was evaluated across six gene expression cohorts — five original datasets plus an independent validation cohort (N = 327). PCAGroupAdam achieved classification performance statistically equally strong to Adam and RMSprop on accuracy, F1, and ROC-AUC, with a significant loss advantage over vanilla SGD (p < 0.0001), consistent with the general benefit of adaptive optimization. Beyond performance, the framework's variance-aware structure enables a direct route from predictions to gene-level interpretation: SHAP analysis identified biologically relevant genes — including AGR2, TSPAN8 — linked to cancer progression, validated through functional annotation (GO/KEGG) and STRING protein–protein interaction analysis. An uncharacterised long non-coding RNA, BC006965, was also found associated with breast cancer, nominating it as a candidate for experimental follow-up. We report these findings under a corrected, leakage-free evaluation protocol and highlight the methodological safeguards — class-balanced training, full multi-trial reporting, and external validation — needed for reliable optimizer comparisons in small-cohort transcriptomic studies.