SPARC‐PATH: A Computational Intelligence Framework for Interpretable and Platform‐Agnostic Gene Selection in Lung Cancer Diagnosis

Accurate and interpretable lung cancer diagnosis from gene expression data remains challenging due to the high dimensionality of transcriptomic profiles, platform‐specific biases, and a lack of biologically grounded model explanations. To address these limitations, we propose SPARC‐PATH, a deep learning framework that integrates cross‐platform gene selection with patient‐specific sparse pathway attention for interpretable lung adenocarcinoma classification. SPARC‐PATH introduces a cross‐domain adaptive feature selection module that learns to identify a shared set of discriminative genes across RNA‐Seq and microarray datasets by minimizing both classification error and distributional divergence. These genes are then mapped to biological pathways using enrichment analysis, and per‐sample pathway activity scores are computed to reflect individualized disease mechanisms. A sparse attention‐based Transformer assigns dynamic importance weights to pathways for each patient, enabling personalized interpretability. Trained on TCGA‐LUAD and validated on three independent GEO microarray cohorts (GSE19188, GSE37745, and GSE31210), SPARC‐PATH achieves consistent performance across all cohorts (AUC up to 97.3%) and reveals pathway activation profiles consistent with known lung cancer mechanisms. Within the scope of the RNA‐Seq‐to‐microarray transfer task studied here, the framework offers a biologically grounded and interpretable approach to transcriptome‐based cancer diagnosis and provides a foundation for future validation across broader transcriptomic platforms and ancestrally diverse populations.