Comparative and Integrative Mapping of Obesity and Cardiometabolic Risk Across Populations Using Lipidomics and Lipoproteomics

Noviani S Minaee

Cardiometabolic diseases (CMDs, including obesity, type 2 diabetes, and cardiovascular disease, are complex, multifactorial conditions and among the leading causes of global mortality. Although Body Mass Index (BMI) is widely used as a surrogate measure for adiposity, it often fails to capture the underlying metabolic heterogeneity contributing to disease risk. Omics technologies such as lipidomics and lipoproteomics offer promising avenues for uncovering molecular mechanisms and improving stratification. However, their high dimensionality, multicollinearity, and redundancy pose major analytical challenges for feature selection, model generalisability, and biological interpretability. This thesis investigates a range of statistical and machine learning approaches to address these challenges using data from two population cohorts: Australian (n=1806) and Saudi Arabian (n=289). Metabolic profiling included 112 lipoprotein parameters via 1H NMR spectroscopy and 895 lipid species via liquid chromatography–mass spectrometry (LC-MS). Obesity was associated with marked shifts in lipoprotein subfractions, notably a redistribution from larger, buoyant LDL1–3 to smaller, denser LDL4–6 particles and a parallel shift in HDL subfractions. These patterns were consistent across populations and were not detectable using standard lipid measures. Several particle number ratios, such as LDL1–3:LDL4–6 (AUROC = 0.74) and HDL1:HDL4 (AUROC = 0.71), demonstrated strong discriminatory power for obesity status. To model BMI, eight feature selection methods: Lasso, Ridge, Elastic Net, Random Forest, Support Vector Machine (SVM), Orthogonal PLS-R (OPLS-R), Hierarchical PLS-R (HPLS-R), and Selection of Features and Estimation of the Model via Classification and Regression Techniques (SELECT), were systematically benchmarked across 10 repeated train–test splits and five feature set sizes (1, 20, 50, 100, 200). Lasso showed the best individual model performance (mean RMSE=0.749, R²Y=0.455, Q²Y =0.419), but its selected features lacked biochemical diversity, being dominated by triacylglycerides (TAGs) and phospholipids. In contrast, a consensus-based approach that pooled the top 20 features from each method demonstrated improved generalisation (RMSE=0.799, R²Y=0.381, Q²Y =0.362) and included a broader mix of lipid classes such as sphingomyelins, phosphatidylcholines, and structurally informative TAG variants. These results demonstrate that combining complementary methods can enhance both model accuracy and interpretability. Moreover, the inclusion of biologically diverse lipid species reveals metabolic signals missed by single-method approaches. This work provides a robust analytical framework for omics-based modeling of cardiometabolic risk and offers insight into the complex lipid remodelling that underlies obesity and its related disorders.

Comparative and Integrative Mapping of Obesity and Cardiometabolic Risk Across Populations Using Lipidomics and Lipoproteomics

Files and links (1)

Abstract

Details

Metrics