Integrative Machine Learning and Network Analysis of Skeletal Muscle Transcriptomes Identifies Candidate Pioglitazone-Responsive Biomarkers in Polycystic Ovary Syndrome

TL;DR

This study uses machine learning and network analysis to find genes that may predict how PCOS patients respond to pioglitazone treatment.

Contribution

The study introduces an integrative approach combining machine learning and network analysis to identify novel biomarkers for pioglitazone response in PCOS.

Findings

1459 differentially expressed genes were identified, including immune and fibrotic signaling genes like IFNB1 and DNMT3A.

Machine learning models effectively distinguished PCOS from controls using a compact gene panel.

ITK, WT1, and BRD1-associated loci were highlighted as key regulatory hubs in the co-expression network.

Abstract

Background/Objectives: Polycystic ovary syndrome (PCOS) is a common endocrine–metabolic disorder in which skeletal muscle insulin resistance contributes substantially to cardiometabolic risk. Pioglitazone improves insulin sensitivity in women with PCOS, yet the underlying transcriptional changes and their potential as treatment-response biomarkers remain incompletely defined. We aimed to reanalyse skeletal muscle gene expression from pioglitazone-treated PCOS patients using modern machine learning and network approaches to identify candidate biomarkers and regulatory hubs that may support precision therapy. Methods: Public microarray data (GSE8157) from skeletal muscle of obese women with PCOS and healthy controls were reprocessed. Differentially expressed genes (DEGs) were identified and submitted to Ingenuity Pathway Analysis to infer canonical pathways, upstream regulators, and disease functions. Four supervised machine learning algorithms (logistic regression, random forest, support vector machines, and gradient boosting) were trained using multi-step feature selection and 3-fold stratified cross-validation to provide superior Exploratory Gene Analysis. Gene co-expression networks were constructed from the most informative genes to characterize network topology and hub genes. A simulated multi-omics framework combined selected transcripts with representative clinical variables to explore the potential of integrated signatures. Results: We identified 1459 DEGs in PCOS skeletal muscle following pioglitazone, highlighting immune and fibrotic signalling, interferon and epigenetic regulators (including IFNB1 and DNMT3A), and pathways linked to mitochondrial function and extracellular matrix remodelling. Within this dataset, all four machine learning models showed excellent cross-validated discrimination between PCOS and controls, based on a compact gene panel. Random forest feature importance scoring and network centrality consistently prioritized ITK, WT1, BRD1-linked loci and several long non-coding RNAs as key nodes in the co-expression network. Simulated integration of these transcripts with clinical features further stabilized discovery performance, supporting the feasibility of multi-omics biomarker signatures. Conclusions: Reanalysis of skeletal muscle transcriptomes from pioglitazone-treated women with PCOS using integrative machine learning and network methods revealed a focused set of candidate genes and regulatory hubs that robustly separate PCOS from controls in this dataset. These findings generate testable hypotheses about the immunometabolism and epigenetic mechanisms of pioglitazone action and nominate ITK, WT1, BRD1-associated loci and related network genes as promising biomarkers for future validation in larger, independent PCOS cohorts.

Figures