Metabolomic profiling of glucose homeostasis in African Americans: the Insulin Resistance Atherosclerosis Family Study (IRAS-FS)

Introduction African Americans are at increased risk for type 2 diabetes. Objectives This work aimed to examine metabolomic signature of glucose homeostasis in African Americans. Methods We used an untargeted liquid chromatography-mass spectrometry metabolomic approach to comprehensively profile 727 plasma metabolites among 571 African Americans from the Insulin Resistance Atherosclerosis Family Study (IRAS-FS) and investigate the associations between these metabolites and both the dynamic (SI, insulin sensitivity; AIR, acute insulin response; DI, disposition index; and SG, glucose effectiveness) and basal (HOMA-IR and HOMA-B) measures of glucose homeostasis using univariate and regularized regression models. We also compared the results with our previous findings in the IRAS-FS Mexican Americans. Results We confirmed increased plasma metabolite levels of branched-chain amino acids and their metabolic derivatives, 2-aminoadipate, 2-hydroxybutyrate, glutamate, arginine and its metabolic derivatives, carbohydrate metabolites, and medium- and long-chain fatty acids were associated with insulin resistance, while increased plasma metabolite levels in the glycine, serine and threonine metabolic pathway were associated with insulin sensitivity. We also observed a differential ancestral effect of glutamate on glucose homeostasis with significantly stronger effects observed in African Americans than those previously observed in Mexican Americans. Conclusion We extended the observations that metabolites are useful biomarkers in the identification of prediabetes in individuals at risk of type 2 diabetes in African Americans. We revealed, for the first time, differential ancestral effect of certain metabolites (i.e., glutamate) on glucose homeostasis traits. Our study highlights the need for additional comprehensive metabolomic studies in well-characterized multiethnic cohorts. Supplementary Information The online version contains supplementary material available at 10.1007/s11306-023-01984-1.


Introduction
Metabolomics, the comprehensive profiling of small molecule metabolites in biological systems, has proven to be a powerful tool for understanding biochemical pathways and disease mechanisms (Newgard, 2017). High-throughput metabolomic profiling technologies using mass spectrometry and nuclear magnetic resonance can identify thousands of metabolites simultaneously. These technological advances have led to the successful application of metabolomics to identify new biomarkers for cardiometabolic diseases and improve our mechanistic understanding of these diseases and related traits (Newgard, 2017).
African Americans are at increased risk of obesity, type 2 diabetes and other cardiometabolic diseases. However, few large-scale metabolomic studies have examined glucose homeostasis in this population, particularly for the sophisticated dynamic measures. In this study, we used an untargeted liquid chromatography-mass spectrometry (MS) metabolomic approach to examine plasma metabolites among 571 African Americans from the Insulin Resistance Atherosclerosis Family Study (IRAS-FS) and investigate the associations between these metabolites and both the basal and dynamic measures of glucose homeostasis traits. We also compared our results in African Americans with our previous results in IRAS-FS Mexican Americans (Palmer et al., 2018).

Study participants
Study participants were recruited from the Insulin Resistance Atherosclerosis Family Study (IRAS-FS) (Henkin et al., 2003;Palmer et al., 2018). This study included 42 African American families from Los Angeles, CA. These families were recruited based on large family size, irrespective of disease status. A total of 572 African Americans without diabetes were examined for association between metabolites and glucose homeostasis traits. A similar metabolomics study in Mexican Americans from the IRAS-FS has been previously reported (Palmer et al., 2018). The study protocol was approved by the institutional review board of the clinical and analysis sites. All participants provided their written informed consent.

Glucose homeostasis measurements
Insulin sensitivity index (S I ) and glucose effectiveness (S G ) were calculated from the frequently sampled intravenous glucose tolerance test (FSIGT) using mathematical modeling method (MINMOD version 3.0; Harms Software, CA) (Pacini & Bergman, 1986). MINMOD uses FSIGT glucose and insulin data to fit two mathematical models for glucose disappearance and insulin kinetics, respectively.
Glucose disappearance is calculated as: Insulin kinetics is calculated as: and I(t) represent the time courses of glucose and insulin in plasma following glucose injection. G b and I b are basal values. X(t) describes the insulin effect on net glucose disappearance and p 1 , p 2 , p 3 , G 0 , n, γ, h and I 0 are parameters. Parameter p 1 represents S G which is the effect of glucose per se at basal insulin to normalize glucose concentration within the extracellular glucose pool. The ratio between p 3 and p 2 represents S I which is the insulin dependent increase in the net glucose disappearance rate. Acute insulin response to glucose (AIR) was calculated as the increase in insulin concentrations at 2 to 8 min above the basal (fasting) insulin level after a bolus glucose injection at 0 to 1 min. AIR is a measure of pancreatic function and insulin release. Disposition index (DI) was calculated as AIR multiplied by S I , which reflects the ability of the body to compensate for changes in either β-cell function or peripheral insulin action. Basal measures included the updated homeostatic model assessment of insulin resistance (HOMA-IR) and beta cell function (HOMA-B) (Levy et al., 1998).

Metabolite measurements
Metabolite profiling was performed on stored (at -80 °C) fasting plasma samples collected at the 1999-2002 baseline survey. Metabolite detection and quantification was conducted by Metabolon, Inc. (Durham, North Carolina) using untargeted liquid chromatography-mass spectroscopy (MS) (DiscoveryHD4 panel). Samples were prepared using the automated MicroLab STAR system (Hamilton Company, Salt Lake City, UT). A methanol extraction was used to remove protein, dissociate small molecules bound to protein or trapped in the precipitated protein matrix, and to recover chemically diverse metabolites. The resulting extract was divided into five fractions: two for analysis by two separate reverse phase/ultra-performance liquid chromatography-MS/ MS methods with positive ion mode electrospray ionization (ESI), one for analysis by reverse phase/ultra-performance liquid chromatography-MS/MS with negative ion mode ESI, one for analysis by hydrophilic interaction liquid chromatography/ultra-performance liquid chromatography-MS/MS with negative ion mode ESI and one sample was reserved for backup. All methods utilized a Waters ACQUITY ultra-performance liquid chromatography (UPLC) and a Thermo Scientific Q-Exactive high resolution/accurate mass spectrometer interfaced with a heated electrospray ionization source and Orbitrap mass analyzer operated at 35,000 mass resolution. Raw data were extracted, peak-identified and quality control processed using Metabolon's hardware and software. Compounds were identified by comparison to library entries of purified standards or recurrent unknown entities. Peaks (3) dI dt = −n ⋅ I(t) + ⋅ (G(t) − h) ⋅ t, I(0) = I 0 were quantified using area under the curve. Several types of controls were analyzed in addition to experimental samples: a technical replicate, pooled matrix sample generated from a small volume of each experimental sample; process blanks, extracted water samples; and QC standards, a cocktail of QC standards chosen not to interfere with the measurement of endogenous compounds were spiked into every analyzed sample, allowed instrument performance monitoring and aided chromatographic alignment. This panel identified and provided relative quantification of known chemical compounds among amino acid, carbohydrate, energy, lipid, nucleotide, and peptide super pathways. In addition to individual named biochemicals; super-and sub-pathways were annotated based on a combination of pathway and chemical structure similarities to serve as a guide for interpretation. Prior to return, data were block corrected for a run day, normalized by batch, and volume extracted. Missing data for metabolites were imputed to the minimum value for the respective metabolite. Each metabolite in original scale was rescaled to set the median equal to one. Since the metabolite distributions were highly skewed, metabolite values were winsorized at 1% and 99% to reduce the impact of outliers on parameter estimation. Among 572 African Americans, one individual had 132 metabolite outliers (± 4 SD) and was excluded from downstream analyses. In addition, six metabolites with > 50% missing values were excluded. A total of 727 out of 733 metabolites passed the quality control and were retained for the final analyses (Supplementary Table 1).

Univariate metabolite analysis
Phenotype variables were transformed to approximate normal distributions as follows: S I , HOMA-IR and HOMA-B were natural log transformed; AIR and DI were square-root transformed but retained the mathematical sign of the original value; S G was not transformed. General linear mixed model was used to test for the association between each metabolite and glucose homeostasis traits after adjusting for age, sex, BMI and pedigree structure in IRAS-FS. The restricted maximum likelihood method in PROC MIXED in SAS was used to fit for each glucose homeostasis trait. The proportion of variance (R 2 adj ) explained by significantly associated metabolites from univariate metabolite models for a given glucose homeostasis trait was estimated using residual values of mixed model fitting after conditioning on covariates. To control for multiple tests, the Bonferroni-corrected P value of 6.8 × 10 -5 (0.05/727 annotated metabolites) was used as the threshold for statistical significance.

Metabolite selection model
The metabolite levels within each metabolic pathway (Supplementary Table 1) are potentially correlated with each other. We used the scaled elastic net regularized regression method to disentangle the linear dependency of correlated metabolites in the prediction of glucose homeostasis traits. In the scaled elastic net regression model, the LASSO penalty function (L 1 ) and ridge penalty function (L 2 ) shrink the estimates of the regression coefficients towards zero relative to the maximum likelihood estimates. This shrinkage prevents overfitting arose from either collinearity of the metabolites or high-dimensionality of data when p > > n (where p is the number of metabolites, and n is the number of individuals). In the procedure of regularization with an elastic net, the coefficient of ridge regression is estimated, then a LASSO algorithm is performed on the ridge regression coefficient to shrink the coefficient. The ridge regression penalty function (L 2 ) penalizes the square of the regression coefficients for the predictors, shrinking coefficients from correlated predictors proportionally toward zero. The LASSO penalty (L 1 ) imposes a penalty on the absolute value of the coefficients, shrinking coefficients by a constant factor, and can select a subset of predictors by shrinking coefficients for the least predictive predictors exactly to zero. The scaled elastic net uses these penalties to minimize the objective function in Eq. 4.
where n is the number of observations (i = 1, 2, …, n), y * i is the n × 1 vector of residual values of glucose homeostasis traits after adjusting for age, sex, BMI and pedigree structure with the mixed model, p (j = 1, 2, …, p) is the number of metabolites, β 0 and β = (β 1 , β 2 ,….., β p ) T are the fitting parameters for intercept (known as bias) and metabolites, X ij is the n x p matrix standardized metabolites, 0 ≤ α ≤ 1 is the tuning parameter that controls the balance between the LASSO (α = 1) and ridge (α = 0) penalties, λ ≥ 0 is positive regularization (or penalty) parameter with the degree of shrinkage increasing as λ increases for a given α value (Goeman et al. 2014). The adjusted R 2 given in Eq.
(2) was used as the stopping criteria.
where i is equal to 1 if there is an intercept and 0 otherwise, n is the number of observations used to fit the model, and p is the number of parameters needed to be estimated in the model.
For the internal validation of the scaled elastic net model, 2000 bootstrapping were conducted by using the model-average approach. As the initial round of model average approach using bootstrapping contains many effects, we used the refit option to obtain a more n − p parsimonious model for better generalization. Thus, for each data sample in the refit, a least squares model was fit with no effect selection (only intercept in the model).
After 2000 bootstrapping of which 90% of each sample is used as training data, the number of non-zero, non-zero percentage and estimated quantiles (25%, 50% and 75%) information for intercept and parameter estimates were used. Metabolites that were selected for more than 20% of samples after 2000 bootstrapping were reported for the metabolite selection model. Before running model-average, a log scale grid search algorithm was used to specify the optimal ridge regression parameter, L 2 for the scaled elastic net. The predicted values ŷ (i) for average model i are given by where X is the design matrix of the data to be scored. Averages are obtained as: j (SAS Inst., 2021). All statistical analyses were performed using SAS (SAS Inst, 2016) and the statistical R package, version 3.3.3 (http:// www.r-proje ct. org/).

Metabolite effect size comparisons for glucose homeostasis traits between African Americans and Mexican Americans
Identical analytical protocols were implemented for data processing in both African Americans and Mexican Americans in IRAS-FS, i.e., metabolite data quality control, outcome transformation and covariate adjustments. The difference in metabolite effect sizes between African Americans and Mexican Americans were tested using the Z statistic. To control for multiple tests, the Bonferroni corrected P value of 6.8 × 10 -5 (0.05/727 annotated metabolites) was used as the threshold for statistical significance. The metabolites that showed significant associations with glucose homeostasis traits in both populations were further characterized.
Participants had a mean S I of 1.64 ± 1.17 × 10 -4 min/μU/ mL, mean AIR of 1013 ± 830 μU/mL/min, and a resulting mean DI of 1436 ± 1269. In comparison with IRAS-FS Mexican Americans (Supplementary Table 2) (Palmer et al., 2018), African Americans were more insulin resistant based on the dynamic glucose homeostasis measures (i.e., reduced S I ); however, African Americans had reduced levels of both HOMA-IR and HOMA-B when compared with Mexican Americans (Supplementary Table 2). Most pairwise correlations (except between DI and HOMA-B) among the six glucose homeostasis traits in African Americans were significant with absolute correlation coefficients ranging from 0.16 to 0.84. AIR was more correlated with DI and S G (r S = 0.62 and 0.30, respectively) in IRAS-FS African Americans than in IRAS-FS Mexican Americans (r S = 0.50 and 0.20, respectively); however, AIR is less correlated with S I and HOMA-B (r S = -0.16 and 0.18, respectively) in IRAS-FS African Americans than in IRAS-FS Mexican Americans (r S = -0.27 and 0.35, respectively; Supplementary Table 3).

Associated metabolites in univariate models
The associations between 727 metabolites and glucose homeostasis traits were assessed using the general linear mixed model with adjustment for age, sex, BMI and pedigree structure. There were 79, 2, 58, 40, 45 and 19 metabolites significantly associated with S I , AIR, DI, S G, HOMA-IR and HOMA-B, respectively (P < 6.8 × 10 -5 , See Table 2 and  Supplementary Tables 4a-f for each glucose homeostasis trait).
Of the 58 metabolites associated with DI, two metabolites (glycine and 1,2-dilinoleoyl-GPC (18:2/18:2)) were positively associated with DI but the remaining 56 metabolites were negatively associated with DI (Table 2 and Supplementary Table 4c). Twelve amino acid and 37 lipid metabolites were predominantly associated with DI. Seven of the 24 BCAA metabolites analyzed were negatively associated with DI (7.53 × 10 -7 < P < 5.82 × 10 -5 ). For the lipid group, 13 metabolites in long chain fatty acid, 5 metabolites in diacylglycerol, 6 metabolites in phospholipid, 6 metabolites in polyunsaturated and 3 metabolites in sphingolipid sub-pathways were associated with DI (Supplementary Table 4c). As observed for AIR, glucose was most negatively associated with DI (β = -51.03, P = 1.32 × 10 -16 ) (Supplementary Table 4c). Given that DI is calculated as the product of S I and AIR, a remarkable number of metabolites overlapped between S I and DI, i.e., 34 of the 58 metabolites associated with DI were shared with S I . Most of these shared metabolites were less significantly associated with DI than S I . Three metabolites in the long chain fatty acid sub-pathway (margarate (17:0), palmitate (16:0), and stearate (18:0)) are, however, more significantly negatively associated with DI than S I . Moreover, 10 of the 24 metabolites that were unique to DI were from the long chain fatty acid sub-pathway (Supplementary Table 4c).
The associated metabolites for S G were mainly clustered into two groups: amino acid (4 metabolites) and lipid (35 metabolites) ( Table 2 and Supplementary Table 4d). All 4 amino acid metabolites were negatively associated with S G (1.30 × 10 -5 < P < 6.29 × 10 -5 ). Except for three metabolites (propionylglycine, 1-linoleoyl-2-linolenoyl-GPC (18:2/18:3) and 1,2-dilinoleoyl-GPC (18:2/18:2)) in the lipid group, all other metabolites (including 10 fatty acid derivatives, 12 long chain fatty acids, 3 phospholipids, 5 polyunsaturated fatty acids, 2 sphingolipid metabolites, etc.) were negatively associated with S G (Supplementary Table 4d). Metabolite associations for S I , DI and S G generally showed a similar pattern: the metabolites from amino acid and lipid groups were predominantly negatively associated with these three dynamic measures of glucose homeostasis traits. Sixteen metabolites overlapped between S I and S G with 11 metabolites from the lipid super-pathway group, whereas 26 metabolites overlapped between DI and S G with 22 metabolites from the lipid super-pathway group.

Associated metabolites in regularized regression models
Overall, a smaller number of metabolites were selected by regularized regression models when compared with univariate models due to correlation among metabolites (Supplementary Table 4a-f and 5). The largest difference in the number of selected metabolites between the two models was observed for S I (Supplementary Table 4a). The pairwise correlation heat map for metabolites associated with S I in univariate models shows that metabolites within each of the sub-pathways (BCAA, diacylglycerol, long-chain fatty acid, lysolipid, phospholipid, sphingolipid, and gamma-glutamyl amino acid metabolisms) were positively correlated with each other (Supplementary Fig. 1). All the 15 metabolites associated with S I in the regularized model overlapped those selected by univariate models (Supplementary Table 4a) and the metabolite glycine accounted for 11.4% of S I residual variance. Two (glucose and 1-docosapentaenoyl-GPC (22:5n3)) of the 7 metabolites associated with AIR in the regularized model overlapped those associated in the univariate model (Supplementary Table 4b) and glucose accounted for 9.4% of AIR residual variance. Ten of the 12 metabolites associated with DI in the regularized model overlapped those associated in the univariate model (Supplementary  Table 4c) and the top three metabolites (glucose, palmitate and leucine) accounted for 19.2% of DI residual variance. Four of the 6 metabolites associated with S G in the regularized model overlapped those associated in the univariate model (Supplementary Table 4d), and the metabolite palmitate accounted for 6.0% of S G residual variance. Ten of the 11 metabolites associated with HOMA-IR in the regularized model overlapped those associated in the univariate model (Supplementary

Comparison of metabolite associations with glucose homeostasis traits between African Americans and Mexican Americans
African Americans were more insulin resistant (low S I ) but had better first phase insulin response (AIR) than Mexican Americans in the IRAS-FS (Supplementary Table 2). We explored whether there were any metabolites that were differentially associated with these six glucose homeostasis traits between these two populations. For most of the metabolites analyzed, no significant differences in associations with these glucose homeostasis traits were observed (Supplementary Table 6). However, the effect sizes of glutamate on S I was significantly different for S I (β = − 0.35 and -0.12 in African Americans and Mexican Americans, respectively; P = 4.5 × 10 -6 , Table 3 and Supplementary  Table 6). Glutamate was also differentially associated with DI, S G and HOMA-IR at nominal significance with larger effects observed in African Americans than in Mexican Americans (Table 3 and Supplementary Table 6).

Discussion
We performed a comprehensive assessment of the signature of metabolomics associated with dynamic and basal measures of glucose homeostasis traits in 571 non-diabetic African Americans. We applied both univariate and regularized regression models to identify metabolites that were significantly associated with six glucose homeostasis traits (S I , AIR, DI, S G , HOMA-IR and HOMA-B). The univariate models identified multiple significantly correlated metabolites within the same sub-pathway; however, the regularized regression models identified the most important metabolites within each sub-pathway. Overall, the metabolites involved in amino acid metabolism (BCAA metabolism; glycine, serine and threonine metabolism; lysine metabolism; methionine, cysteine, SAM and taurine metabolism; glutamate metabolism; urea cycle, arginine and proline metabolism), carbohydrate metabolism (glycolysis, gluconeogenesis, and pyruvate metabolism) and lipid metabolism (diacylglycerol and phospholipid metabolism) were associated with both basal and dynamic measures of glucose homeostasis traits. In addition, we observed that glutamate affects glucose homeostasis traits more strongly in African Americans than in Mexican Americans.
Circulating levels of glycine were lower in obese (Felig et al., 1969;Okekunle et al., 2017;Takashina et al., 2016;Yan et al., 2012) and insulin-resistant individuals (Ejaz et al., 2016;Gall et al., 2010;Newgard et al., 2009;Takashina et al., 2016) as compared to healthy individuals in populations of European and East Asian ancestries. Plasma glycine levels were also negatively associated with fasting glucose levels and type 2 diabetes risk (Chen et al., 2022;Ferrannini et al., 2013;Floegel et al., 2013;Guasch-Ferre et al., 2016;Newgard et al., 2009;Palmer et al., 2015;Svingen et al., 2016;Vangipurapu et al., 2019;Walford et al., 2016;Wang-Sattler et al., 2012). In concordance with the previous findings, we observed that the metabolites (glycine, N-acetylglycine or serine; Supplementary Table 4a-f) in the glycine, serine and threonine metabolism sub-pathway were positively associated with dynamic measures of S I and DI but negatively associated with basal measures of HOMA-IR and HOMA-B in IRAS-FS African Americans. The metabolite glycine also explained the largest residual variance (11.4%) of S I in IRAS-FS African Americans. Glycine plasma concentration is tightly regulated by glucagon, which is a major regulator of hepatic glycine metabolism. Elevated plasma glucagon levels could drive increased glycine degradation in insulin resistant states (Alves et al., 2019). In addition, impaired hepatic BCAA metabolism in obesity was shown to contribute to the decrease in glycine circulating concentration; however, the underlying cellular mechanisms still remain to be fully elucidated (Alves et al., 2019).
Lysine and its metabolite 2-aminoadipate were associated with increased type 2 diabetes risk in populations of both European and Asian ancestries (Chen et al., 2019;Takashina et al., 2016;Wang et al., 2013). The metabolite 2-aminoadipate was negatively associated with dynamic measures of S I and DI but positively associated with basal measures of HOMA-IR in IRAS-FS African Americans (Supplementary  Table 4a-f). The in vitro studies suggest that 2-aminoadipate had an effect on insulin secretion in pancreatic β-cells and isolated islets  and the positive associations between 2-aminoadipate levels and type 2 diabetes risk in human may be due to the development of insulin resistance secondary to chronic hyperinsulinemia (Newgard, 2017;Wang et al., 2013).
The metabolite 2-hydroxybutyrate in the sub-pathway of methionine, cysteine, SAM and taurine metabolism was an early biomarker of insulin resistance and glucose intolerance in a nondiabetic population of European ancestry (Gall et al., 2010) and its levels was elevated in type 2 diabetes patients compared with healthy controls in populations of Asian ancestry (Li et al., 2009). Consistent with our previous findings in IRAS-FS Mexican Americans (Palmer et al., 2018), this metabolite was negatively associated with dynamic measures of S I , DI and S G in IRAS-FS African Americans (Supplementary Table 4a-f). The metabolite 2-hydroxybutyrate is converted from alpha-ketobutyrate when the NADH/NAD + ratio is elevated in high lipid oxidations of insulin resistant states (Gall et al., 2010). The NMD supplementation is recently proven to be effective in recovering insulin sensitivity in prediabetic women (Yoshino et al., 2021).
Plasma glutamate levels were differentially associated with glucose homeostasis traits between African Americans and Mexican Americans in IRAS-FS with larger effects (negatively associated with dynamic measures of S I , DI and S G but positively associated with measures of HOMA-IR) observed in African Americans than in Mexican Americans. Oral glutamate supplementation was previously shown to impair insulin sensitivity in a short-term dietary intervention (Chevassus et al., 2002). Increased circulating levels of glutamate were previously associated with insulin resistance (Chen et al., 2022;Cheng et al., 2012;Newgard et al., 2009;Vangipurapu et al., 2019) and type 2 diabetes risk in populations of European (Chen et al., 2019;Cheng et al., 2012;Ferrannini et al., 2013;Liu et al., 2019;Stancakova et al., 2012;Vangipurapu et al., 2019), Asian (Chen et al., 2019;Takashina et al., 2016) and African (Chen et al., 2022). ancestries. Glutamate stimulated glucagon release from pancreatic α-cells (Adrover et al., 2015) and increased transamination of pyruvate to alanine, a strong promoter of gluconeogenesis ). These mechanisms may partially explain the increased type 2 diabetes risk associated with circulating glutamate levels. Populations of African ancestry are generally more susceptible to insulin resistance than populations of other ancestries (Meigs et al., 2014); however, the underlying physiological mechanisms of increased glutamate sensitivity on glucose homeostasis traits in African Americans merits further investigations.
The amino acid sub-pathway of urea cycle, arginine and proline metabolism was consistently enriched for dynamic measures of S I and basal measures of HOMA-IR and HOMA-B in IRAS-FS African Americans (Supplementary  Table 4a-f). The metabolite 2-oxoarginine in this pathway was negatively associated with S I , but most significantly positively associated with measures of HOMA-IR and HOMA-B levels. It also explained the largest residual variance of HOMA-IR (8.1%) and HOMA-B (5.9%). Similarly, the metabolite arginine in this pathway was negatively associated with S I but positively associated with HOMA-IR and HOMA-B levels (Supplementary Table 4a-f). Increased plasma levels of 2-oxoarginine and arginine were positively associated with HOMA-B measures in IRAS-FS Mexican Americans (Palmer et al., 2018), and increased circulating arginine levels were associated with increased type 2 diabetes risk in populations of European (Guasch-Ferre et al., 2016) and African (Chen et al., 2022) ancestries. The metabolite 2-oxoarginine is a guanidino compound metabolite of arginine catabolism. Arginine promoted insulin secretion (Sener et al., 2000). However, the underlying mechanisms of arginine metabolites regulating glucose homeostasis traits remains unclear.
The carbohydrate sub-pathway of glycolysis, gluconeogenesis, and pyruvate metabolism was consistently enriched for dynamic measures of S I , AIR, DI and S G , and basal measures of HOMA-IR in IRAS-FS African Americans (Supplementary Table 4a-f). Several metabolites (glucose, fructose, mannose, lactate, pyruvate, mannose or ribonate) were negatively associated with S I , AIR, DI, S G , but positively associated with HOMA-IR, consistent with reported higher levels of these metabolites in individuals with type 2 diabetes than control subjects of European ancestry (Drogan et al., 2015;Floegel et al., 2013;Guasch-Ferre et al., 2016;Padberg et al. 2014;Wurtz et al., 2012aWurtz et al., , 2012b. The metabolite glucose explained the largest residual variance of AIR (9.4%) and DI (10.9%) and confirmed its important role of regulating these glucose homeostasis traits.
Circulating medium-and long-chain fatty acids were elevated in prediabetes and type 2 diabetes patients in populations of European and Asian ancestries (Gall et al., 2010;Li et al., 2009;Menni et al., 2013). Plasma levels of long-chain fatty acids were negatively associated with dynamic measures of S I , DI and S G in the IRAS-FS Mexican Americans (Palmer et al., 2018). We observed a similar pattern of increased levels of palmitate (C16), margarate (C17) and stearate (C18) associated with reduced S I in IRAS-FS African Americans (Supplementary Table 4a-f). In addition, more long-chain fatty acids species (n = 13, Supplementary Table 4c) were inversely associated with DI in IRAS-FS African Americans. The metabolite palmitate also explained a substantial portion of residual variance of DI (5.8%) and S G (6.0%). The increased plasma medium-and long-chain fatty acid levels may be biomarkers of increased adipose fatty acid releases and/or reduced fatty acid oxidations of insulin resistant states.
Our study has the following strengths. It included the largest sample size of African Americans in a semi-untargeted metabolomic study with 727 "known" plasma metabolites assayed. While most studies have fasting glucose and insulin measures only (Chen et al., 2022;Rebholz et al., 2018;Yu et al., 2019), we analyzed both basal and dynamic glucose homeostasis, of which dynamic measures could be considered physiologically more proximal to pathogenic components of type 2 diabetes and provide discrete insights into the pathogenesis of type 2 diabetes (Palmer et al., 2018). The comparison of metabolomic profiles between African Americans and Mexican Americans in IRAS-FS allows identification of differentially associated metabolites, while other studies only reported associations adjusted for ancestry (Chen et al., 2022;Rebholz et al., 2018). We used regularized models to identify key metabolites associated with these glucose homeostasis traits, which could partially overcome the bias derived from high correlation of metabolites in the same metabolic sub-pathway; and finally, substantial proportions of residual variance in these glucose homeostasis traits (ranging from 16 to 41% in univariate models) have been explained by these metabolites even after adjustment for the effect of BMI (Supplementary Table 5). However, the current study is observational in nature, which precludes assessment of causality between metabolites and these glucose homeostasis traits, thus unable to determine which participants may develop diabetes in the future.
In summary, we confirmed in a representative population of African ancestry that increased plasma metabolite levels of BCAA and their metabolites, 2-aminoadipate, 2-hydroxybutyrate, glutamate, arginine and its metabolites, carbohydrate metabolites, and median-and longchain fatty acids are associated with insulin resistance, while increased plasma metabolite levels in the glycine, serine and threonine metabolic sub-pathway were associated with insulin sensitivity. We also observed a differential ancestral effect of glutamate on glucose homeostasis with much stronger effect observed in African Americans than those observed in Mexican Americans. Overall, this study suggests that these metabolites may be useful biomarkers in the identification of prediabetes individuals at risk of type 2 diabetes in African Americans, which warrant further studies. Our findings also extend the scientific literature on the role of these metabolites in the etiology of insulin resistance and impaired insulin secretion and highlight the need for additional comprehensive metabolomic studies in well-characterized multiethnic cohorts.