Correlation of plasma metabolites with glucose and lipid fluxes in human insulin resistance

Summary Objective Insulin resistance develops prior to the onset of overt type 2 diabetes, making its early detection vital. Direct accurate evaluation is currently only possible with complex examinations like the stable isotope‐based hyperinsulinemic euglycemic clamp (HIEC). Metabolomic profiling enables the detection of thousands of plasma metabolites, providing a tool to identify novel biomarkers in human obesity. Design Liquid chromatography mass spectrometry–based untargeted plasma metabolomics was applied in 60 participants with obesity with a large range of peripheral insulin sensitivity as determined via a two‐step HIEC with stable isotopes [6,6‐2H2]glucose and [1,1,2,3,3‐2H5]glycerol. This additionally enabled measuring insulin‐regulated lipolysis, which combined with metabolomics, to the knowledge of this research group, has not been reported on before. Results Several plasma metabolites were identified that significantly correlated with glucose and lipid fluxes, led by plasma (gamma‐glutamyl)citrulline, followed by betaine, beta‐cryptoxanthin, fructosyllysine, octanylcarnitine, sphingomyelin (d18:0/18:0, d19:0/17:0) and thyroxine. Subsequent machine learning analysis showed that a panel of these metabolites derived from a number of metabolic pathways may be used to predict insulin resistance, dominated by non‐essential amino acid citrulline and its metabolite gamma‐glutamylcitrulline. Conclusion This approach revealed a number of plasma metabolites that correlated reasonably well with glycemic and lipolytic flux parameters, measured using gold standard techniques. These metabolites may be used to predict the rate of glucose disposal in humans with obesity to a similar extend as HOMA, thus providing potential novel biomarkers for insulin resistance.


| INTRODUCTION
Obesity is often accompanied by metabolic disorders such as dyslipidemia and insulin resistance, both part of the metabolic syndrome, which in turn is a major risk factor for type 2 diabetes (T2DM), cardiovascular pathology, non-alcoholic fatty liver disease and different types of cancer. 1,2 Insulin resistance develops prior to the beginning of overt T2DM, making its early detection of vital clinical importance.
However, direct accurate evaluation of insulin resistance in relation to fasting insulin and (compensatory) hyperinsulinemia is currently only possible using complex, invasive and time-consuming examinations such as the two-step stable isotope based hyperinsulinemic euglycemic clamp (HIEC), a method which is regarded as the gold standard. This method also allows for distinguishing between hepatic and peripheral insulin sensitivity (Rd).
An interesting developing research field in this respect is untargeted metabolic profiling, which enables the detection of in principle thousands of plasma metabolites. 3 These metabolites are products that reflect levels of cellular (dys)function. As they are influenced by both environmental (dietary) and biological (genetic) factors, plasma metabolites may provide insight into the balance of genotype and phenotype of T2DM. Moreover, due to the unbiased nature that characterizes metabolomic platforms, they can provide a tool to unveil novel underlying mechanisms of insulin resistance, metabolic syndrome and its dire consequences in humans with obesity.
Branched chain amino acids and other plasma metabolites such as glycerol, α-hydroxybutyrate and mannitol 4 are increased in patients with T2DM and might serve as potential novel biomarkers for insulin resistance. [5][6][7][8] However, there is a lack of studies in humans with metabolic syndrome that correlate novel metabolites with the gold standard measurement of glucose fluxes in the HIEC and, to the knowledge of this research group, no studies are available that combine both parameters. Consequently, this results in a lack of novel biomarkers able to predict aberrant glycemic control in patients with insulin resistance in an earlier phase. 9 A recent systematic review evaluated metabolite markers identified using high-throughput metabolomics techniques in patients cohorts with prediabetes and T2DM; however, in these patients, insulin resistance was determined via simple or indirect tests such as the oral glucose test, but never by HIEC. 8 Alternative methods have been developed such as the Homeostatic Model Assessment (HOMA) and Quantitative insulin sensitivity check index (QUICKI), which calculate insulin resistance fairly accurately using fasting insulin and glucose concentrations. Nevertheless, as surrogate indirect methods, they have their limitations compared with the direct clamp method and are still unable to detect early stages of aberrant glucose metabolism and insulin resistance. 10 Thus, in order to determine plasma metabolites, which are able to predict insulin resistance in an earlier stage and explore their interaction with glucose and lipid fluxes in the fasting state, liquid chromatography mass spectrometry (LC-MS)-based untargeted plasma metabolomics (Metabolon) was applied in 60 participants with obesity with an extensive range of Rd as determined by a two-step HIEC with stable isotopes [6,6-2 H 2 ]glucose and [1,1,2,3,3-2 H 5 ]glycerol, which also allowed for the measurement of insulin regulated lipolysis (Ra).
Other metabolic parameters determined were hepatic insulin sensitivity, as measured via suppression of endogenous glucose production (EGP), and resting energy expenditure (REE), measured by indirect calorimetry, as well as dietary intake. This study shows that a panel of plasma metabolites could be used to predict (peripheral) insulin resistance as a substitute for the invasive laborious HIEC. Indeed, several metabolites were identified that significantly correlated with glucose and lipid fluxes, led by plasma citrulline and gamma-glutamylcitrulline, followed by plasma betaine, beta-cryptoxanthin, fructosyllysine, octanylcarnitine, sphingomyelin (d18:0/18:0, d19:0/17:0) and thyroxine.

| Study participants
Male (n=30) and female participants (n=30) with obesity, defined as body mass index (BMI) ≥ 30 kg m −2 , were recruited via local advertisements. Participants were excluded in case of a primary lipid disorder, childhood onset obesity (due to higher risk of developing T2DM also based on more genetic predisposition), use of exogenous insulin, all medical and psychiatric conditions except for obesity related diseases, coagulation disorders, uncontrolled hypertension, renal insufficiency, substance abuse (nicotine or drugs, alcohol >2 units day −1 ), pregnancy and breastfeeding. All participants provided written informed consent, and all study procedures were approved by the Academic Medical Center IRB and conducted in accordance with the Declaration of Helsinki.

| Study design
Following informed consent and screening, participants visited the clinical research unit after an overnight fast. Participants were instructed to record food intake and dietary habits online (mijn. voedingscentrum.nl/nl/eetmeter) the week before the study visit.
During the study day, blood samples were drawn followed by the two-step HIEC during which glucose and lipid fluxes were determined as well as REE, as described below.

| Two-step HIEC and REE
REE was measured in all participants during the basal state of the HIEC by indirect calorimetry. During 20 minutes, oxygen consumption and CO 2 production were measured continuously using a ventilated hoodsystem (Vmax Encore 29; SensorMedics, Anaheim, CA).
REE was then calculated from oxygen consumption and carbon dioxide production. 11 To measure insulin resistance, a two-step HIEC was performed. 12 After an overnight, fast participants visited the hospital where they received two catheters in the peripheral veins of both arms. One catheter was used to infuse the [ Insulin infusion was increased after 2 hours of insulin infusion (t=2 h) to 60 mUÁm −2 Áminute −1 . At t=0, 2 and 4 hours, five blood samples were taken to assess glucose and glycerol enrichments.
Rates of disposal (Rd) of glucose and rates of appearance (Ra) of glycerol were calculated using the modified forms of the Steele equations for (non)steady state measurements as described previously. 13

| Biochemistry
Fasting glucose (Hitachi), insulin (Diagnostic products) and C-reactive protein (CRP, Roche, Switzerland) were determined in fasted plasma samples. Total cholesterol, high-density lipoprotein cholesterol (HDLc) and triglycerides (TG) were determined in EDTA-containing plasma using commercially available enzymatic assays (Randox, Antrim, UK and DiaSys). All analyses were performed using a Selectra autoanalyser (Sopachem, The Netherlands). Low-density lipoprotein cholesterol (LDLc) was calculated using the Friedewald formula. Insulin was determined on an Immulite 2000 system (Diagnostic Products, Los Angeles, CA, USA). As described previously, plasma analyses were performed, 14

| Metabolomic profiling
As mentioned above, EDTA plasma samples were taken from participants before the clamp in the fasting state. Plasma metabolite untargeted profiling analysis was carried out on plasma by Metabolon (Durham, NC), using ultrahigh-performance liquid chromatography coupled to tandem mass spectrometry, as previously described. 17 This method allowed for the study of 934 annotated plasma metabolites.
Raw data were normalized to account for interday measurement differences. Then, each biochemical was rescaled to set the median equal to 1. Missing values, generally due to the sample measurement falling below the limit of detection, were then imputed with the minimum observed value.

| Statistical analysis
Univariate (Spearman's rank) analyses were performed to determine which metabolites correlated significantly (P-value of <.05) with Rd, EGP suppression, REE and glycerol suppression of the included participants. Subsequent multivariate linear regression was used to correct for age and BMI.

| Machine learning to discern relative importance of plasma metabolites with Rd
In the multivariate analysis, the model was a gradient boosting regressor, 18

| Glucose and lipid fluxes
Insulin sensitivity was determined in each subject via a two-step HIEC. The use of [6,6-2 H 2 ]glucose infusion enabled the assessment of the ability of insulin to suppress EGP (EGP suppression, a marker of hepatic insulin sensitivity) and whole-body glucose rate of disposal

| Combining plasma metabolites to predict peripheral insulin sensitivity (Rd)
The 17 metabolites found to correlate significantly with Rd originate from a number of different metabolic pathways. To investigate whether a reliable prediction of Rd could be made from a combination of these metabolites, a Gradient Boosting Machine Learning model was used. 18 The average R 2 achieved by the model across the 20 performed shuffles was 0.24. Figure 3 shows the top 10 metabolites that were found to be most important in the Rd predictive model, headed by plasma (gamma-glutamyl)citrulline levels, followed by HOMA and fasting insulin levels.

| DISCUSSION
In this exploratory cross-sectional study (untargeted), metabolic profiling was applied using fasting plasma samples of 60 participants with obesity with an extensive range of peripheral insulin sensitivity, revealing several metabolites that significantly correlated with glucose and lipid fluxes as determined by stable isotope-based two-step HIEC.
Subsequent machine learning analysis showed that a combination of these plasma metabolites may potentially be used to predict peripheral insulin resistance and sensitivity (Rd) with a reasonable accuracy.
Machine learning analysis on EGP and REE was not possible due to the low number of correlating parameters. However, plasma gammaglutamylcitrulline emerged as the most important component in the  The medium-chain acylcarnitines, decanoylcarnitine and octanoylcarnitine were both negatively associated with insulin sensitivity, and especially octanoylcarnitine was found to be predictive of Rd (Figure 3). There is growing evidence of plasma acylcarnitines as potential biomarkers for insulin resistance. 27 In study participants with T2DM, octanoylcarnitine and decanoylcarnitine were both increased compared with controls. 28 A substrate overload of fatty acids may result in incomplete long-chain fatty acid β-oxidation, allowing for accumulation of acylcarnitines which via pro-inflammatory NFkBassociated pathways impair insulin action, thus promoting insulin resistance. 27 In participants with obesity and T2DM, in whom the degree of insulin resistance was measured via a HIEC similar to this study, plasma acylcarnitines were found to be significantly elevated, 29 supporting the current observations. which is in accordance with previous data in patients with TDM2. 31 In line, a positive significant correlation was observed in this study between plasma betaine and peripheral, hepatic and adipose tissue insulin sensitivity. Betaine, also known as trimethylglycine, is an amino acid, which is derived in the human body via dietary intake or via mitochondrial oxidation of choline in kidney and liver (driving TMAO production), 32,33 and low betaine is significantly associated with key criteria of metabolic syndrome. 34,35 Moreover, in the current study, the significant positive correlation was confirmed between plasma beta-cryptoxanthin (β-CRX, a carotenoid) and peripheral and hepatic insulin sensitivity. 36,37 In obese animal models, β-CRX supplementation was inversely related to the risk of insulin resistance and liver dysfunction, possibly mediated by inhibiting inflammatory gene expression via the suppression of macrophages and of the signalling of TNFα and gut-derived endotoxins (LPS). [38][39][40] In humans, serum β-CRX was likewise inversely related to insulin resistance 41 and decreased amount of visceral fat. 42 Finally, the negative correlation between fructosyllysine and both Rd and REE is in agreement with other studies where the increase of this metabolite has been associated with obesity and insulin resistance in humans. 43 Fructosyllysine is a glyco-amino acid, also known as an early glycation product, which is directly toxic to tissue and is a precursor of advanced glycation end product formation. 44,45 In conclusion, unravelling the interaction between plasma metabolites with glucose and lipid fluxes in persons with insulin resistance is essential for our understanding of its pathophysiology and uncovering yet unknown biomarkers and novel treatments. By combining targeted plasma metabolomics with gold standard stable isotopebased glucose and lipid fluxes in participants with obesity with an extensive range of Rd, some well-known and some less known indicators are provided for insulin resistance and a panel of these metabolites is shown that may possibly be used to predict Rd to a similar extend as fasting insulin and HOMA. Thus, plasma levels of (gamma-glutamyl)citrulline may serve as a viable candidate for early clinical diagnosis of insulin resistance (eg, in combination with HOMA) when a stable isotope-based HIEC is not available.