GEN-SPARC – a pathway-aware framework for reliable and interpretable biomarker discovery in multi-cancer transcriptomics

High-dimensional transcriptomic studies frequently struggle with the imbalance between the vast number of measured genes and the limited number of patient samples, leading to instability, overfitting, and unreliable biomarkers. To address this enduring challenge, we propose GEN-SPARC (Gene Selection and Pathway-Aware Regularized Classifier), a novel framework that integrates stability learning, knockoff-based false discovery control, and biologically informed regularization within a unified architecture. GEN-SPARC begins by deriving robust gene importance estimates through bootstrap ensemble modeling and SHAP-based stability aggregation, then applies the Model-X knockoff principle to rigorously control the false discovery rate, ensuring that only statistically reliable genes are retained. These selected features are further organized into co-expression modules via partial-correlation graph learning, enabling the model to capture higher-order biological structure. A pathway-regularized sparse optimization step then leverages these modules to perform classification in a manner that is both interpretable and resistant to overfitting. Applied to eight major cancer families—breast, colorectal, brain, renal, prostate, liver, lung, and leukemia—on CuMiDa microarray datasets, GEN-SPARC consistently demonstrates strong internal validation performance, higher AUROC and precision–recall scores, and markedly lower feature volatility compared to standard lasso, elastic-net, and ensemble approaches. By coupling knockoff-driven statistical rigor with pathway-level interpretability, GEN-SPARC offers a reproducible, interpretable, and biologically meaningful framework for robust biomarker discovery in small-sample, high-dimensional cancer genomics.