INTEGRATED SYSTEMS BIOLOGY AND MACHINE LEARNING FRAMEWORK FOR THE IDENTIFICATION OF MITOCHONDRIAL AND SIGNALING SIGNATURES IN ISCHEMIC AND DILATED CARDIOMYOPATHY

Abstract

Cardiovascular diseases (CVDs), predominantly manifested as ischemic cardiomyopathy (ICM) and dilated cardiomyopathy (DCM), constitute the leading cause of morbidity and mortality worldwide, driven by a complex architecture of genetic, metabolic, and environmental alterations. Traditional clinical diagnostic modalities, while effective for hemodynamic assessment, often fail to capture the granular molecular heterogeneity required for early intervention and personalized therapeutic strategies. The advent of high-throughput transcriptomics has provided a systems-level view of these pathologies, yet the translation of high-dimensional genomic data into robust clinical biomarkers remains hindered by the "curse of dimensionality" and the noise inherent in biological datasets. This study presents a comprehensive, integrated bioinformatics and machine learning pipeline to identify a compact, biologically potent gene signature for CVD classification. Drawing conceptual inspiration from the "Moana" classification framework—which prioritizes biological structure over raw feature variance—we employed a pathway-guided feature selection strategy on RNA-sequencing data (GSE55296) and microarray validation datasets (GSE57338). Functional enrichment analysis using KEGG and Reactome databases revealed a profound dysregulation of mitochondrial metabolism, oxidative phosphorylation, and cardiomyopathy-related pathways. Subsequent protein–protein interaction (PPI) network analysis, refined through the intersection of multiple topological centrality algorithms (Degree, Maximal Clique Centrality, and Betweenness), distilled the transcriptome into a core signature of six genes: COX4I1, NDUFA5, NDUFV2, NDUFS3, NDUFS4, and AKT1. These features were utilized to train and evaluate supervised machine learning models. The Random Forest classifier demonstrated superior stability and performance, achieving a mean cross-validation accuracy of approximately 69.5%, significantly outperforming a Deep Neural Network (DNN) in this limited-sample regime. The findings underscore the critical role of mitochondrial Complex I and IV dysfunction in heart failure and demonstrate the efficacy of integrating systems biology with ensemble learning to develop interpretable, high-precision molecular diagnostics.