An optimization-based framework for gene expression-based diagnostic classification of Wilms tumor

Abstract In this study, eSRA is a novel optimization algorithm proposed to enhance machine learning model efficiency in large-scale problems with numerous dimensions and intricate patterns. For comparison purposes and to assess eSRA’s efficiency, it was compared with several popular methods for gene expression analysis of Wilms tumors (WT) using real-life medical datasets for gene expression classification, combining microarray (GSE66405, GSE73209, GSE2712, GSE11024) and RNA sequencing (TARGET-WT) with more than 260 samples of tumor and normal samples taken together from the kidneys. For this analysis, eSRA was used for gene set identification and hyperparameter tuning of the XGBoost classification algorithm, and eSRA was used alone. Experimental outcomes confirmed eSRA’s top rank for accuracy (0.9769), F 1 score (0.985), and ROC AUC (0.998) with superior performances compared with traditional and hybrid models. The model’s robustness and scalability were confirmed with unaltered performances regardless of population sizes and dimensions. SHAP value-based interpretation concluded genes like CALB1, CLDN8, NUP155, ALDH4A1, CYB5A, and GATA2 account for primary contributions to classifications related to tumors. Functional enrichment confirmed that these genes are involved in cell proliferation, embryonic kidney development, and vascular remodeling, processes central to WT biology. Overall, this work shows that eSRA is a powerful, efficient optimizer that enhances machine learning performance and provides biologically meaningful insights, making it a promising tool for biomedical data analysis and cancer genomics research.