Graph Neural Network-Based Blood Transcriptomic Signature for Metabolic Phenotype Stratification and Prognosis Prediction in Obesity

Abstract

Background: The rising prevalence of obesity presents a global health crisis, yet significant heterogeneity exists within the obese population. Differentiating between metabolically healthy obesity (MHO) and metabolically unhealthy obesity (MUO) is crucial for personalized risk assessment and treatment, but current clinical methods are often insufficient. This study addresses the need for a robust, non-invasive tool to stratify these distinct metabolic phenotypes. Methods: We developed a novel machine learning framework based on a Graph Neural Network (GNN) to analyze blood transcriptomic data. The model leverages the expression patterns of key genes involved in lipid metabolism and inflammation to construct a patient-specific gene co-expression network. We trained and validated the model using transcriptomic and clinical data from multiple independent cohorts of individuals with obesity, sourced from low- and middle-income countries (LMICs). Findings: Our GNN-based model achieved high accuracy in distinguishing between MHO and MUO individuals, with an area under the receiver operating characteristic curve (AUC) of 0.92 (95% CI 0.89-0.95) in the primary validation cohort. The model successfully identified distinct gene expression signatures and network topologies associated with the MUO phenotype, highlighting dysregulated inflammatory signaling and lipid metabolism pathways. Furthermore, the model provided a risk score that was significantly associated with the prospective development of cardiometabolic complications. Interpretation: This study presents a powerful, interpretable, and blood-based transcriptomic signature for the stratification and prognostic assessment of obese individuals. The GNN framework offers a significant advancement over traditional machine learning approaches by capturing complex biological interactions. This tool has the potential to facilitate early and accurate identification of high-risk obese individuals, enabling targeted interventions and advancing precision medicine in diverse populations.