Abstract
The study of metabolomics has improved our knowledge of the biology behind type 2 diabetes and its related metabolic physiology. We aimed to investigate markers of adipose tissue morphology, as well as insulin and glucose metabolism in 53 non-obese male individuals. The participants underwent extensive clinical, biochemical and magnetic resonance imaging phenotyping, and we also investigated non-targeted serum metabolites. We used a multi-modal machine learning approach to evaluate which serum metabolomic compounds predicted markers of glucose and insulin metabolism, adipose tissue morphology and distribution. Fasting glucose was associated with metabolites of intracellular insulin action and beta-cell dysfunction, namely cysteine-s-sulphate and n-acetylgarginine, whereas fasting insulin was predicted by myristoleoylcarnitine, propionylcarnitine and other metabolites of beta-oxidation of fatty acids. OGTT-glucose levels at 30 min were predicted by 7-Hoca, a microbiota derived metabolite, as well as eugenol, a fatty acid. Both insulin clamp and HOMA-IR were predicted by metabolites involved in beta-oxidation of fatty acids and biodegradation of triacylglycerol, namely tartrate and 3-phosphoglycerate, as well as pyruvate, xanthine and liver fat. OGTT glucose area under curve (AUC) and OGTT insulin AUC, was associated with bile acid metabolites, subcutaneous adipocyte cell size, liver fat and fatty chain acids and derivates, such as isovalerylcarnitine. Finally, subcutaneous adipocyte size was associated with long chain fatty acids, markers of sphingolipid metabolism, increasing liver fat and dopamine-sulfate 1. Ectopic liver fat was predicted by methylmalonate, adipocyte cell size, glutathione derived metabolites and fatty chain acids. Ectopic heart fat was predicted visceral fat, gamma-glutamyl tyrosine and 2-acetamidophenol sulfate. Adipocyte cell size, age, alpha-tocopherol and blood pressure were associated with visceral fat. We identified several biomarkers associated with adipose tissue pathophysiology and insulin and glucose metabolism using a multi-modal machine learning approach. Our approach demonstrated the relative importance of serum metabolites and they outperformed traditional clinical and biochemical variables for most endpoints.