The data obtained and analyzed in this study as well as previously published articles on the 53 subjects is to our knowledge unique in its extensive nature, combining clinical, biochemical, radiological and untargeted serum metabolomics data for comprehensive phenotypic metabolic characterization, as well as enhancing our knowledge of adipocyte biology and insulin- and glucose metabolism. In this study, our primary objective was to examine markers of insulin and glucose metabolism, whilst considering complete untargeted serum metabolomics. Moreover, we presented k-means and hierarchical cluster analyses in an attempt to identify unique metabolic phenotypes, considering our high-dimensional dataset. Predictive machine learning models were constructed in a stepwise fashion with additional pre-processing techniques to reduce number of predictors for each outcome. Our approach demonstrated that relative importance of serum metabolites outperformed traditional clinical and biochemical variables for most endpoints.
Predictive machine learning models based on oral glucose- and insulin tolerance tests, highlighted several metabolites as the most important predictors for glucose and insulin metabolism. Fasting glucose was associated with a known biomarker of obesity, namely cysteine-s-sulphate, which is involved in intracellular insulin action (13) and n-acetylgarginine, which has been suggested to modulate glucose homeostasis, insulin sensitivity and promote lipolysis, through arginine-nitric oxide modulation of intracellular AMPK and PI3K(14). In addition, cysteine is involved in gluthathione synthesis, which is known for its relation to beta-cell dysfunction.
Fasting insulin was predicted by the metabolite eugenol, which has been shown to lower blood glucose and blood lipids, as well as lower markers of inflammation (15), in addition to other metabolites of beta-oxidation of fatty acids. According to animal models, eugenol facilitates insulin sensitivity and stimulates glucose uptake via skeletal muscle tissue and activation of the GLUT4-AMPK signaling pathway. Increased relative importance of propionylcarnitine, a fatty ester lipid molecule, indicates that dysregulated fatty acid metabolism and lipid metabolism in the beta-oxidation of long-chain fatty acids might cause lipid accumulation in tissues, supporting the role as an important metabolite for fasting insulin levels. Carnitine is essential for cellular energy since it transports long-chain fatty acids into the mitochondria for beta-oxidation, as well as transporting toxic compounds out of this cellular organelle to prevent their accumulation. Body weight also demonstrated high relative importance in both scaled and unscaled models, while serum bilirubin was nearly statistically significant.
Glucose levels at 30 minutes were predicted by 7-Hoca and taurochenodeoxycholate, microbiota derived metabolites, as well as, fatty-chain acids. Type 2 diabetes is associated with alteration in bile acid signaling. Studies in animal models reveal that taurochenodeoxycholate increases glucose-induced insulin secretion, mimicking the effect of electrical stimulation of beta-cells and enhancing cytosolic Ca2+ concentration.
Both insulin clamp and HOMA-IR, were predicted by metabolites involved in beta-oxidation of fatty acids and biodegradation of triacylglycerol. Tartrate is considered a xenobiotic metabolite that is related to BMI, insulin resistance and adiponectin, while 3-phosphoglycerate is a significant intermediate in glycolysis as well as a non-ATP product of PGK1, which is critical for constructing serine and secreting insulin (16). According to unscaled predictions models, an important predictor was the metabolite 5-hydroxyhexanoate, a medium chain fatty acid and strong predictor of insulin clamp and its parent compound hexanoic acid, which has been shown to promote insulin induced phosphorylation of the AKT-mTOR signaling pathway and balances lipid metabolism in a hepatocellular model (17). According to predictions models with unscaled predictors, metabolites of glycolysis, and metabolites of the tricarboxyl acid (TCA) cycle in the mitochondria of cells (particularly pyruvate, which is central in converting fatty acids and glucose to acetyl CoA, and through the TCA cycle), predicted HOMA-IR (6). Its regulatory enzymes have been discussed as a target for new antidiabetic pharmacological agents (18). In addition, a substantial part of pyruvate enters mitochondria beta cells and TCA cycle via conversion to oxaloacetate, which correlated with glucose-induced insulin secretion. According to the scaled models, the artificial sweetener acelsulfame and methyl-4-hydroxybenzoate-sulfate, as well as, tartronate-hydroxymalonate, which is involved in fatty acid biosynthesis and mitochondrial energy production, proved to be important predictors. Acelsulfame has previously been associated with increasing BMI and glucose intolerance.
Information derived from OGTT was used to calculate an area under curve value (AUC) for both glucose and insulin measures. These newly constructed endpoint variables were associated with several examined metabolites. According to unscaled prediction models, AUC for glucose was associated with bile acid metabolites (Taurochenodeoxycholate), which we discussed previously and is linked to metabolic dysregulation. In scaled prediction models, nonadecanoyl-gpc and glutamate, were almost statistically significant.
AUC insulin was predicted by subcutaneous adipocyte size as well as a metabolite of sphingolipid metabolism, a compound involved both in intracellular signaling and cell membrane turnover. Sphingolipids have previously been shown to be associated with insulin resistance, possibly via downstream insulin signaling alterations (6). In addition to this, the scaled prediction models identified serum-bilirubin and propionylcarnitine, as important predictors for AUC insulin.
Adipocyte hypertrophy has been extensively studied as a mediator in the development of insulin resistance and hyperinsulinemia and our results are in line with previous results (3).
Finally, subcutaneous adipocyte size was found to be associated with several long chain fatty acids, metabolites involved in fatty acid biosynthesis, markers of beta-oxidation of fatty acids, as well as markers of sphingolipid metabolism. Specifically, an offspring compound of ethyl benzoate was an important predictor. Benzoates have been shown to directly affect both insulin and glucagon secretion (19). The scaled prediction models also identified that liver fat and dopamine-sulfate 1 and methyl-4-hydroxybenzoate-sulfate, were important predictors for adipocyte cell size. Previous research has suggested a regulatory role for peripheral dopamine-sulfate in adipose tissue.
Clustering analyses identified three unique phenotypic groups, where levels of insulin resistance, defined by insulin clamps, differed significantly between the groups. At a tendency level, amount of visceral liver fat also differed but failed to reach statistical significance. We found several markers of amino acid metabolism that predict visceral adipose tissue, a finding that is in line with previous research as amino acid metabolites have been shown to predict insulin resistance (20). We also found a bile acid metabolite, as well as a glycolysis metabolite to predict visceral liver fat, two cellular processes we have mentioned previously to be associated with insulin resistance.
In our previous research, we observed that in these subjects both visceral and subcutaneous fat area by MRS evaluation were predicted by metabolites of fatty acid oxidation. Lipid oxidation metabolites also predicted liver lipid accumulation, and cardiac lipid storage was predicted by a metabolite of branched chain amino acid (BCAA) turnover (9). BCAA have previously been linked to IGT and overt type 2 diabetes and our findings are in line with these results (6, 21). Our findings in this study are thus an addition to previous findings. Ectopic lipid accumulation in liver was predicted by amount of subcutaneous adipose tissue, liver transaminases (s-ALAT), methylmalonate mma (??) and lipid metabolites. According to scaled models, predictors for visceral fat were subcutaneous adipocyte cell size and ectopic liver fat and insulin clamp. However, linear regression shows that only adipocyte cell size, age and alpha-tocopherol, were associated with visceral fat. Our data are not able to distinguish whether visceral fat accumulation precedes ectopic fat storage in the liver. In general, cross-validation for the machine learning model for ectopic adipose tissue surrounding the heart tissue was poor, nevertheless age, diastolic blood pressure, 2-acetamidophenol-sulfate, gamma-glutamyltyrosine and visceral fat, were the best predictors.
A major strength of this study is the extensive examination of subjects using clinical and biochemical variables, imaging data and untargeted metabolomics. Some limitations of our study should be considered. The relatively small number of subjects included in our study complicates our ability to cross-validate and generalize our machine learning models. Validation models on test dataset were impracticable in some cases due to size of the cohort. We believe that a trade-off between a lesser regression-mean squared error (RMSE) value and R2 is satisfactory in this dataset to signify the superior model for each endpoint.