A COMPUTATIONAL FRAMEWORK FOR IDENTIFYING GENE SIGNATURES USING HUMAN ADIPOSE AND SKELETAL MUSCLE TRANSCRIPTOMES

Abstract

Obesity is a multi-cellular metabolic syndrome characterised by dysregulated adipose tissue growth, mitochondrial bioenergetics and disrupted glucose and lipid homeostasis. Although transcriptome-scale profiling provides a thorough understanding of these changes, the accuracy of conventional statistical methods is limited by the nonlinear and high dimensional structure of gene expression data. To address these challenges, an integrated machine-learning and explainable artificial intelligence framework was developed to discover transcriptomic predictors of obesity in human skeletal-muscle and adipose tissue. Ensemble classification models i.e. Random Forest and Extreme Gradient Boosting were applied to normalised gene expression matrices derived from publicly available datasets(GSE22309 and GSE103766) with systematic hyperparameter optimization. SHapley Additive exPlanations( SHAP) promote model interpretability, enabling quantitative measure of each gene contributions to classification outcomes. Key transcripts such as BDH1, MRPS31, RARS2, STOML2, AKAP10 and UCP3 were aligned with mitochondrial metabolic pathways involving ketone oxidation, ribosomes function, cristae stability and proton leak modulation. Network analysis identified NUCB2, SRSF1 and ACTG2 as the key topological nodes while Enrichment analysis using Enricher revealed significant alterations in oxidative phosphorylation interactions, reactive oxygen and nitrogen mediated signalling pathways. Overall this study shows the significance of an xAI enabled framework for identifying biologically significant transcriptomic signatures related to obesity.