Mixed meal tolerance testing highlights in diabetes altered branched-chain ketoacid metabolism and pathways associated with all-cause mortality

Abstract

Background

Elevated BCAA levels are strongly associated with diabetes, but how diabetes affects BCAA, branched-chain ketoacids (BCKAs), and the broader metabolome after a meal is not well known.

Objective

To compare quantitative BCAA and BCKA levels in a multiracial cohort with and without diabetes after a mixed meal tolerance test (MMTT) as well as to explore the kinetics of additional metabolites and their associations with mortality in self-identified African Americans.

Methods

We administered an MMTT to 11 participants without obesity or diabetes and 13 participants with diabetes (treated with metformin only) and measured the levels of BCKAs, BCAAs, and 194 other metabolites at 8 time points across 5 h. We used mixed models for repeated measurements to compare between group metabolite differences at each timepoint with adjustment for baseline. We then evaluated the association of top metabolites with different kinetics with all-cause mortality in the Jackson Heart Study (JHS) (N = 2441).

Results

BCAA levels, after adjustment for baseline, were similar at all timepoints between groups, but adjusted BCKA kinetics were different between groups for α-ketoisocaproate (P = 0.022) and α-ketoisovalerate (P = 0.021), most notably diverging at 120 min post-MMTT. An additional 20 metabolites had significantly different kinetics across timepoints between groups, and 9 of these metabolites—including several acylcarnitines—were significantly associated with mortality in JHS, irrespective of diabetes status. The highest quartile of a composite metabolite risk score was associated with higher mortality (HR:1.57; 1.20, 2.05, P = 0.00094) than the lowest quartile.

Conclusions

BCKA levels remained elevated after an MMTT among participants with diabetes, suggesting that BCKA catabolism may be a key dysregulated process in the interaction of BCAA and diabetes. Metabolites with different kinetics after an MMTT may be markers of dysmetabolism and associated with increased mortality in self-identified African Americans.

Introduction

Diabetes is a highly prevalent and morbid disease, and the identification of at-risk patients could lead to early implementation of lifestyle or pharmaceutical interventions to reduce its incidence [[1], [2], [3]]. Previous research revealed that fasting levels of BCAAs, i.e., leucine, isoleucine, and valine, are associated with insulin resistance and can predict incident diabetes over a decade before disease onset [[4], [5], [6], [7], [8]]. In addition, their immediate breakdown products, branched-chain ketoacids (BCKAs), are also elevated in patients with type 2 diabetes, although the association with disease is less robust than that with BCAAs [5, 9]. Subsequent large-scale analysis showed that incident diabetes, BCAA levels, and BCKA levels are associated with variants in the PPM1K gene, which encodes the activator of the branched-chain α-ketoacid dehydrogenase (BCKDH) complex [10]. This suggests that BCAAs and BCKAs—both upstream of the BCKDH complex—may not only be the markers of disease but may also contribute at least in part to disease pathogenesis. Small molecule inhibitors of branched-chain ketoacid dehydrogenase kinase, which in turn inhibits the BCKDH complex, are under development for metabolic diseases [11, 12].

Although differences in fasting BCAA and BCKA levels have been shown between those with and without diabetes, much less is known about their changes in response to physiologically relevant stimuli, such as a meal. Previous studies have examined the differences in metabolic response after a mixed meal tolerance test (MMTT) between healthy controls and those with diabetes [[13], [14], [15], [16]], but they are limited by small number of metabolites interrogated or lack of measurements at frequent time intervals to establish robust kinetics. Therefore, the first goal of this study was to compare and quantify the differences in BCKA and BCAA levels after an MMTT in multiracial cohort with and without diabetes. We then performed an exploratory analysis of how diabetes might impact broader metabolomic changes after an MMTT.

Given that we and others have previously identified groups of metabolites whose steady state levels presage the onset of cardiometabolic disease outcomes [6, [17], [18], [19]], metabolites with different kinetics after an MMTT perturbation raise the possibility that they may reflect impaired nutrient handling and dysmetabolism and consequently also portend poorer long-term outcomes. Thus, we sought to characterize these metabolites by testing their fasting levels’ associations with mortality in the Jackson Heart Study (JHS), a large population-based cohort of self-identified African Americans. Several previous studies of metabolites as markers of mortality have focused almost exclusively on European and American Caucasian populations [[20], [21], [22], [23], [24], [25]], and with African Americans at much higher risk of developing diabetes than Caucasians [2], it is of epidemiological importance to verify whether metabolites reflecting dysmetabolism have relevance in this understudied population.

Methods

MMTT participants

We recruited adults with and without clinically diagnosed type 2 diabetes from outpatient clinics at the Beth Israel Deaconess Medical Center, Boston, MA between April 2019 and January 2020. We limited participants with diabetes to those treated with a stable dose of metformin to minimize the effects of medications on metabolomic profiles. The control group consisted of participants without obesity or other significant comorbid conditions and was matched to participants with diabetes by age and sex (see Supplemental Methods in the Supplemental Materials for full inclusion and exclusion criteria). Participants fasted overnight and received a mixed meal consisting of a nutritional drink and a protein bar (combined 450 kcal; 58% carbohydrate, 25% protein, and 17% fat) consumed within 15 min (see Supplemental Table 1 for full nutritional information). Participants took all their regular medications except for metformin, which was held before the MMTT. We collected peripheral blood samples 30 min before the meal, immediately before the meal, and at 15, 30, 60, 90, 120, 180, 240, and 300 min after the meal. We also collected finger stick glucose readings at each timepoint to monitor for hyper- and hypoglycemia.

We estimated that a sample size of 44 participants (22 per group), using a conservative SD estimate of 0.3 on the log scale based on previous literature [26], would have an 80% power to detect 25% increased BCAA levels in the diabetes group relative to the control group with a one-sided P < 0.05. Although the study was prospectively designed for 1-sided hypothesis that BCAA and BCKA levels would increase more in those with diabetes after an MMTT than in controls, we present statistical significance below based on 2-sided P < 0.05. After approximately 50% recruitment, the sponsor stopped the study early because of slow recruitment and having detected significant differences in the changes from baseline BCKA levels between groups. Pfizer Inc. funded the study. The Institutional Review Board at the Beth Israel Deaconess Medical Center approved the study, and all participants provided written informed consent.

BCAAs, BCKAs, insulin, and c-peptide quantification

We thawed frozen plasma samples at ambient temperature and vortexed for 30 seconds and added a 30 μL sample aliquot to a 96-well polypropylene plate (Analytical Sales & Services, Inc., Flanders, NJ) containing 270 μL of cold methanol prepared with deuterium labeled internal standards. We vortexed samples at 1500 × g for 5 min, placed them at -20min°C for 2 h, centrifuged them at ∼1643 × g at 4°C for 20 min, transferred 200 μL of the supernatant to a new 96-well polypropylene plate, dried them under N₂, and reconstituted them in 100 μL of 0.2% acetic acid. Last, we vortexed samples at 1500 × g for 5 min and placed them in the autosampler for analysis.

We injected a 5 μL sample onto a Waters Acquity HSS T3 1.8 μm 2.1 × 100 mm column equipped with a Waters Acquity HSS T3 1.8 μm Vanguard Pre-Column (Waters Corporation) using a Shimadzu Nexera chromatography system (Shimadzu Corporation) coupled to a SCIEX 5500 mass spectrometer (SCIEX Corporation). Mobile phase A consisted of a 0.2% acetic acid solution in water. Mobile phase B was acetonitrile. Isomer and analyte separation was achieved using a linear gradient of B from 0 to 85% over 7.5 min, flow rate 0.35 mL/minute, and column temperature of 40 ^oC for total run time of 12 min. Source conditions (heated capillary temperature, gas 1, gas 2, and curtain gas) were set at 500°C , 50°C , 50°C , and 30°C , respectively. We set declustering potential at -50 for the BCAAs and -60 for the BCKAs and the CE at -8 for all analytes. We substituted surrogate ¹³C stable isotope labeled analytes for endogenous analytes to generate standard curves. We assayed analytes using negative ion mode at -4.5 kV. We used the following multiple reaction monitoring and parent to parent transitions. Endogenous BCAAs and BCKAs were: leucine/isoleucine 129.9>129.9, valine 116.0>116.0, α-ketoisocaproate/α-keto-β-methylvalerate 128.9>128.9, and α-ketoisovalerate 114.9>114.9. Surrogate ¹³C labeled BCAAs and BCKAs were: ¹³C₆ leucine/isoleucine 135.9>135.9, ¹³C₅ valine 121.9>121.9, ¹³C₆ α-ketoisocaproate/α-keto-β-methylvalerate 134.9>134.9, and ¹³C₅ α-ketoisovalerate 119.9>119.9. Deuterium labeled internal standard BCAAs and BCKAs were: D₁₀ leucine/isoleucine 140.0>140.0, D₈ valine 124.0>124.0, D₃ α-ketoisocaproate 131.9>131.9, D₈ α-keto-β-methylvalerate 137.0>137.0, and ¹³C₄D₄ α-ketoisovalerate 122.9>122.9. We performed data acquisition and peak integration with SCIEX Analyst software version 1.6.2 and used Watson LIMS 7.5 for regression and quantitation.

We measured insulin and c-peptide levels at each timepoint using commercial ELISA assays [Human Insulin Quantikine ELISA kit (Cat# DINS00) and Human Insulin C-peptide Quantikine ELISA kit (Cat# DICP00), R&D Systems, Minneapolis, MN USA]. All the measurement procedures were performed exactly according to the manufacturer’s protocols.

Extended metabolite profiling

After collecting samples in EDTA anticoagulated tubes, we immediately centrifuged them at 1000 × g for 10 min at 4°C and transferred the plasma into 400 uL aliquots for storage at -80°C . We extracted metabolites by adding 10 uL or 30 uL of stored plasma to 90 uL or 70 uL of acetonitrile/methanol solution (3:1, vol/vol) containing isotope labeled internal standards for positive mode MS analysis or negative mode MS analysis respectively. We assayed amino acids, biogenic amines, nucleotides, neurotransmitters, vitamins, and other polar metabolites by liquid LC-MS in positive ion mode and assayed TCA cycle intermediates, sugars, and bile acids by LC-MS in negative ion mode as previously described [27, 28].

In brief, we loaded 10 uL of reconstituted sample onto either a 150 mm × 2.1 mm Atlantis hydrophilic interaction liquid chromatography (HILIC) column (Waters Corporation) for positive mode or loaded 5 uL on a 100 mm × 2.1 mm, 3.5 μm XBridge amide column (Waters Corporation) for negative mode using a HTS PAL autosampler (Leap Technologies, Carrboro, NC) or Agilent 1290 Infinity autosampler. We separated metabolites using an Agilent 1200 Series HPLC system (Agilent Technologies, Santa Clara, CA) coupled to a 4000-QTRAP mass spectrometer (AB SCIEX, Foster City, CA) in positive mode and an Agilent 1290 infinity HPLC binary pump system (Agilent Technologies) coupled to a 6490-QQQ mass spectrometer (Agilent Technologies) in negative mode. We used the softwares MultiQuant v2.1 (AB SCIEX) and MassHunter (Agilent Technologies) for automated peak integration of positive and negative modes respectively, and we manually assessed metabolite peaks for quality control of peak integration.

JHS participants and metabolite profiling

The design and methods of JHS have been described [29]. In brief, JHS is a large, community-based, observational study that recruited 5306 self-identified African-American participants from rural and urban areas of the Jackson, MS metropolitan area. Participants underwent a baseline examination between 2000 and 2004, and subsequent follow-up occurred annually by telephone interview and review of medical records and death certificates until the end of 2018. Diabetes status was classified by the 2010 American Diabetes Association definition, self-reported diabetes, or use of medications for diabetes within 2 wk before clinic visit [30]. Metabolite profiling was performed on baseline plasma samples from 2751 participants and measured similarly to the extended metabolites described above and as previously described [19].

Statistical analyses

The primary objective was to compare the kinetics and to quantify BCAA and BCKA levels in participants with diabetes compared with controls after an MMTT, and the secondary objective was to compare the same changes for insulin and c-peptide. In exploratory analyses, we examined the kinetics of an additional 194 targeted metabolites. We defined baseline values as the average of -30 min and 0 minute measurements and calculated relative fold change for each timepoint by dividing by the baseline value. For exploratory metabolite LC-MS peak areas, which are relatively quantified, we natural log transformed the values and performed normalization for each metabolite by subtracting batch medians and adding the global median. We then subtracted each timepoint’s value by the log transformed baseline value. We did not impute missing metabolite data from the MMTT, which was caused by inability to obtain blood samples at certain timepoints. In JHS, metabolite data were natural log transformed and scaled to mean 0 and SD 1 by run batch to allow for comparison across batches and metabolites. N-oleoyl glycine was the only metabolite with missingness (0.3%), and because its missingness was due to values below the limit of detection, we imputed missing values to half of the minimally detected value in each batch. Imputation was also a necessary step for the regularized regression algorithm used below.

For comparisons of metabolite kinetics, we applied a mixed model for repeated measurements (MMRM) that included log baseline, group, timepoint (as categorical variable), log baseline interaction with timepoint, and group interaction with timepoint as fixed effects and participants as random effects. We tested for overall interaction between group and timepoints using analysis of variance. For differences at each timepoint, we compared the least squares estimates for group means and their mean differences at each timepoint, which controls for baseline, timepoint, and baseline interaction with timepoint effects. We compared metabolite kinetics by individual metabolites and aggregated by metabolite class as reported on the University of California San Diego Metabolomics Workbench website (www.metabolomicsworkbench.org).

For the primary and secondary endpoints, we also calculated the peak increase from baseline and area under the effect curve (AUEC), which was calculated using the linear trapezoid method on untransformed data in any participant with the first, last, and ≥6 values available. For comparison of peak and AUEC, we used analysis of covariance (ANCOVA) with log baseline as a covariate and group as a factor. Full details of the prespecified statistical analysis plan are included in the Supplemental Materials. Because metabolites in the exploratory analysis were not quantified by absolute levels, AUEC was not calculated for them.

To further characterize metabolites that had different kinetics elicited by the perturbational MMTT, we identified metabolites that had significant interaction effect between timepoint and diabetes status from the MMRM analysis and no significant difference at baseline between the MMTT groups, which we established using a 2-sampled t-test with significance defined as 2-sided P < 0.05. To our knowledge, no large population-based cohort study has yet to incorporate an MMTT in its assessment, so we sought to examine whether fasting levels of these metabolites had any prognostic significance. Specifically, we performed Cox proportional hazards regression on the association of fasting (i.e., last ate at least 8 h ago) metabolites of interest in JHS with all-cause mortality. We adjusted for age, sex, and diabetes status in the first model and additionally adjusted for BMI, estimated glomerular filtration rate (eGFR) calculated based on the CKD-EPI equation [31], total cholesterol, HDL, current smoking status, systolic blood pressure (SBP), hypertension medication use, and CVD history in the full model. In both models, we added interaction terms of metabolites with diabetes to test for differential effects among those with and without diabetes. We additionally tested for interaction by sex.

Last, we constructed a metabolite risk score by combining these metabolites in a Cox regression regularized by ridge penalties with 10-fold cross validation and assessed the score’s association with all-cause mortality after adjustment for the variables in the full model above. We bootstrapped the CI for the metabolite score’s improvement in the Harrell’s concordance index over the full clinical model alone by randomly resampling the cohort 10,000 times. To establish reference points for the metabolite risk score’s strength of association with all-cause mortality, we compared its association with those of diabetes status and HbA1c (both assessed contemporaneously with metabolites) with mortality; metabolite risk score and HbA1c were categorized by quartiles in Cox regressions for ease of interpretation. For sensitivity analysis, we assessed the metabolite risk score’s performance after excluding each metabolite.

For the comparisons of BCAAs, BCKAs, insulin, and c-peptide, we defined significance with a two-sided P value < 0.05. In the exploratory analysis of comparing metabolome-wide kinetics between groups post-MMTT and of metabolites’ associations with all-cause mortality in JHS, we corrected for multiple hypothesis testing using a false discovery rate (FDR) < 0.05 across the number of analytes tested at each step. All statistical analyses were performed in R version 4.2.2 (www.r-project.org), and ridge regression was performed using the glmnet package version 4.1-4.

Results

Primary and secondary endpoints

We approached 133 individuals who met the eligibility criteria. Of the 41 individuals who consented to participate, 17 did not undergo the study because of loss to follow-up, health status changes resulting in loss of eligibility, or withdrawal of consent (Supplemental Figure 1). Ultimately, 13 participants with diabetes and 11 without diabetes underwent the MMTT study, and Table 1 summarizes their baseline characteristics. Overall, participants with diabetes were more likely to self-identify as African Americans and had high BMI and low total and low-density lipoprotein cholesterol levels. In the diabetes group, blood samples were unable to be obtained for 2 participants at 30 min, 1 at 90 min, 1 at 180 min, 1 at 240 min, and 2 at 300 min and in the control group, 1 at 120 min and 2 at 300 min.

TABLE 1.

Baseline characteristics of mixed meal tolerance test participants¹

	Diabetes group (N = 13)	Control group (N = 11)	Overall (N = 24)	P value
Age	57 ± 7	58 ± 9	57 ± 8	0.799
Female	9 (69%)	8 (73%)	17 (71%)	1
Race				0.050
African American	8 (62%)	2 (18%)	10 (42%)
Asian	1 (8%)	1 (9%)	2 (8%)
White	4 (31%)	8 (73%)	12 (50%)
BMI, kg/m2	29.6 ± 3.0	25.0 ± 3.4	27.5 ± 3.9	0.002
Fasting glucose, mg/dL	142 ± 80	92 ± 12	116 ± 60	0.079
Hemoglobin A1c, %	6.9 ± 0.6	5.4 ± 0.4	6.3 ± 0.9	<0.001
Creatinine, mg/dL	0.81 ± 0.12	0.84 ± 0.16	0.82 ± 0.14	0.654
Total cholesterol, mg/dL	169 ± 36	199 ± 35	184 ± 38	0.064
Triglyceride, mg/dL	137 ± 75	127 ± 113	132 ± 94	0.811
LDL, mg/dL	89 ± 24	113 ± 27	101 ± 28	0.042
HDL, mg/dL	52 ± 9	61 ± 19	56 ± 15	0.177
Statin use	8 (73%)	8 (73%)	16 (73%)	1

Numbers are expressed as mean ± SD and n (%). Race was self-identified. Hemoglobin A1c data were missing for 2 control participants. Creatinine, total cholesterol, triglyceride, LDL, HDL, and statin use data were missing for 2 participants with diabetes. Fasting glucose was missing for 3 participants with diabetes. P values were calculated based on two-sided two-sample t-tests for continuous variables and two-sided Fisher’s exact tests for categorical variables.

¹HDL, high-density lipoprotein cholesterol; LDL, low-density lipoprotein cholesterol.

Figure 1 shows the kinetics of BCAA, BCKA, insulin, and c-peptide levels for the raw values as well as the least squares estimates from the MMRM model with 95% CIs at each timepoint. At baseline, participants with diabetes had high levels of BCAA, BCKA, insulin, and c-peptide, as expected. These absolute differences persisted for many of these analytes after the MMTT. After adjustment for baseline differences via the MMRM model, there were no significant differences in the percent changes between the groups for BCAA levels after MMTT; however, we did observe significant overall interaction effects between diabetes status and time for α-ketoisocaproate (P = 0.022) and α-ketoisovalerate (P = 0.021). When looking at specific timepoints, there were higher levels of α-ketoisocaproate at 90 min (ratio of percent change, 1.21, P = 0.033) and at 120 min α-ketoisocaproate (ratio 1.25, P = 0.013) for participants with diabetes, and there were higher levels of α-ketoisovalerate at 120 min (ratio 1.21, P = 0.036). There was also a higher relative fold change for insulin at 120 min (ratio 1.85, P = 0.049). Thus, although the excursions of BCAA were similar after considering baseline differences, BCKA metabolite changes were different in individuals with diabetes. Last, we found that with adjustment for baseline values in the ANCOVA models, there were no significant differences in the peak change from baseline or the AUEC between participants with and without diabetes for BCAAs, BCKAs, insulin, or c-peptide (Supplemental Table 2).

FIGURE 1.

Change over time of estimated least squares means (LSM) from mixed model for repeated measures (MMRM) analysis for raw value and fold change of branched chain amino acids, their respective branched chain ketoacids, insulin, and c-peptide after mixed meal tolerance test by diabetes status. MMRM model for raw values included group, timepoint (as categorical variable), and group interaction with timepoint as fixed effects and participants as random effects. MMRM model included log baseline, group, timepoint (as categorical variable), log baseline interaction with timepoint, and group interaction with timepoint as fixed effects and participants as random effects. Overall interaction between group and timepoint was tested with analysis of variance. Significant P-interaction values of raw value comparisons were: 0.00049 for α-ketoisocaproate, 0.00026 for α-keto-β-methylvalerate, 0.0044 for α-ketoisovalerate, 2.1 × 10^-5 for insulin, and 8.7 × 10^-10 for c-peptide. Significant P-interaction values of fold change comparisons were: 0.022 for α-ketoisocaproate, 0.021 for α-ketoisovalerate, 1.3 × 10^-5 for insulin, and 2.2 × 10^-7 for c-peptide. For analytes with significant overall interaction effect, asterisks indicate significant differences at a specific timepoint with a two-sided p-value < 0.05.

Exploratory metabolites

To capture a sense of the broader metabolic changes before and after a meal challenge, we examined 194 metabolites in the exploratory analysis. At baseline, 16 metabolites had significantly different levels between the 2 groups (Figure 2). Even in this small cohort, many findings were consistent with previous studies, such as an inverse association for citrulline and positive associations for aminoadipic acid, alanine, and dimethylguanidino valeric acid with diabetes ([32], [33], [34]). Forty-six metabolites had a nominally significant (P < 0.05) overall interaction between group and timepoint from the MMRM analysis (Supplemental Table 3). Figure 3 shows the least squares means graphs for the 20 metabolites that reached FDR significance for overall group and timepoint interaction. When metabolites were grouped by molecular class (See Supplemental Table 3 for individual metabolite’s class), FDR significant interaction effects between group and timepoint emerged for saturated fatty acids, unsaturated fatty acids, acylcarnitines, hydroxy fatty acids, oxo fatty acids, quinoline carboxylic acids, N-substituted nicotinamides, and amino acids (Figure 4). For both saturated and unsaturated fatty acids, participants with diabetes had slower relative decline and recovery of levels after the MMTT compared with controls, and the levels did not return to baseline by 5 h after the MMTT. By contrast, acylcarnitine levels fell for both groups after the MMTT but the decline was less rapid for participants with diabetes. Participants with diabetes did not have a delayed increase in hydroxy and oxo fatty acids (acetoacetic acid) that was seen for those in the control group. Last, quinoline carboxylic acids increased and returned to baseline in those with diabetes, whereas those without diabetes had a more sustained elevation in levels across timepoints.

FIGURE 2.

Volcano plot of differences in baseline fasting metabolites between participants with diabetes and controls. Labeled metabolites have two-sided p-value < 0.05 from a two-sample t-test. CAR 3:0, propionylcarnitine; CAR DC3:0, malonylcarnitine; DMGV, dimethylguanidino valeric acid.

FIGURE 3.

Change over time of estimated least squares means (LSM) from mixed model for repeated measures (MMRM) analysis for relative fold change of metabolites with significant overall group and timepoint interaction effects (all false discovery rate < 0.05). MMRM model included log baseline, group, timepoint (as categorical variable), log baseline interaction with timepoint, and group interaction with timepoint as fixed effects and participants as random effects. Asterisks indicate significant differences between groups at a specific timepoint with a two-sided p-value < 0.05. Overall group and timepoint P-interaction values are: 1.3 × 10^-14 for 3-hydroxybuyric acid, 9.6 × 10^-14 for palmitic acid, 2.7 × 10^-12 for stearic acid, 4.3 × 10^-11 for CAR 2:0, 5.1 × 10^-10 for linoleic acid, 2.8 × 10^-9 for CAR 6:0, 1.9 × 10^-8 for CAR 12:0, 4.2 × 10^-8 for CAR 8:0, 1.3 × 10^-6 for N-oleoyl glycine, 7.0 × 10^-6 for CAR DC3:0, 9.2 × 10^-6 for butyric acid, 9.9 × 10^-6 for acetoacetic acid, 2.0 × 10^-5 for docosahexaenoic acid (DHA), 6.5 × 10^-5 for 14,15-epoxyeicosatrienoic acid (EpETrE), 7.4 × 10^-5 for 2-hydroxybutyric acid, 0.00025 for CAR:10, 0.0011 for arachidonoyl-ethanolamide (EA), 0.0018 for N-acetylglutamic acid, 0.0023 for xanthurenic acid, and 0.003 for kynurenic acid. CAR indicates acylcarnitine with number preceding colon indicates length of carbon chain and number following colon indicates number of double bonds. CAR DC3:0 denotes malonylcarnitine.

FIGURE 4.

Change over time of estimated least squares means (LSM) from mixed model for repeated measures (MMRM) analysis for relative fold change of metabolites grouped by class. MMRM model included log baseline, group, timepoint (as categorical variable), log baseline interaction with timepoint, and group interaction with timepoint as fixed effects and participants as random effects. Asterisks indicate significant differences between groups at a specific timepoint with a two-sided p-value < 0.05. Overall group and timepoint P-interaction values are: 2.2 × 10^-25 for saturated fatty acids (FA), 1.1 × 10^-15 for unsaturated FA, 4.5 × 10^-9 for acylcarnitines, 7.1 × 10-6 for hydroxy FA, 1.0 × 10-5 for oxo FA, 1.4 × 10-5 for quinoline carboxylic acids, 0.00019 for N-substituted nicotinamides, and 0.0067 for amino acids.

Metabolite profiles and all-cause mortality in a population-based cohort

From above, we identified 20 metabolites that had significantly different trajectories across timepoints between those with and without diabetes, of which 12 metabolites were assayed in our JHS cohort. We included 2441 JHS participants who fasted and had no missing covariates for model adjustment (Supplemental Figure 1); Table 2 shows their baseline characteristics. Given that the focus of this analysis was to characterize novel metabolites identified via their differences elicited by the earlier MMTT study, we excluded N-acetylglutamic acid and 2-hydroxybuyric acid, which showed a baseline difference between the MMTT groups. From the univariate analysis of these 10 MMTT-elicited metabolites, 9 metabolites reached FDR significance in the fully adjusted model accounting for diabetes status, age, sex, BMI, eGFR, total cholesterol, HDL cholesterol, smoking, SBP, hypertension medication use, and CVD history (Table 3). C6:0 carnitine showed the strongest association with incident all-cause mortality (HR 1.18 per SD increase in metabolite level, 95% CI: 1.08, 1.29, FDR adjusted P = 0.00267). None of these 10 metabolites demonstrated an interaction effect with diabetes (Supplemental Table 4 shows results stratified by diabetes status) or sex (Supplemental Table 5 shows results stratified by sex). Last, as a sensitivity analysis, the addition of treatment with diabetes medications as a covariate to the full model did not change the significance of associations (Supplemental Table 6).

TABLE 2.

Baseline characteristics of Jackson Heart Study participants²

	Diabetes (N = 550)	No Diabetes (N = 1891)	Overall (N = 2441)	P value
Age (y)	60 ± 11	54 ± 13	56 ± 13	<0.001
Female	362 (66%)	1155 (61%)	1517 (62%)	0.044
Body mass index, kg/m2	34 ± 7	31 ± 7	32 ± 7	<0.001
Estimated glomerular filtration rate, mL/min/1.73 m2	91 ± 25	94 ± 21	94 ± 22	0.003
Fasting glucose, mg/dL	140 ± 59	91 ± 9	102 ± 35	<0.001
Hemoglobin A1c, %	7.6 ± 1.8	5.5 ± 0.5	6.0 ± 1.3	<0.001
Total cholesterol, mg/dL	202 ± 46	200 ± 39	200 ± 41	0.208
HDL, mg/dL	50 ± 13	53 ± 15	52 ± 15	<0.001
Current smoker	57 (10%)	238 (13%)	295 (12%)	0.159
Systolic blood pressure, mmHg	131 ± 16	127 ± 17	128 ± 17	<0.001
Hypertension medication use	417 (76%)	877 (46%)	1294 (53%)	<0.001
History of cardiovascular disease	90 (16%)	134 (7%)	224 (9%)	<0.001
Follow-up time, y	14.1 (10.3, 16.0)	14.9 (12.5, 16.0)	14.6 (12.0, 16.0)
Deaths	220 (40%)	364 (19%)	584 (24%)

Numbers are expressed as mean ± SD and n (%) except for follow-up time, which is expressed as median (IQR). Fasting glucose was missing for one participant without diabetes. P values were calculated based on two-sided two-sample t-tests for continuous variables and two-sided Fisher’s exact tests for categorical variables.

²HDL, high-density lipoprotein cholesterol.

TABLE 3.

Association of metabolites with all-cause mortality in the Jackson Heart Study

Metabolite	N	Events	Model 1				Model 2
			HR	95% CI:	FDR p	P-interaction	HR	95% CI	FDR p	P-interaction
CAR 6:0	2441	584	1.283	1.182, 1.393	2.51E-8	0.708	1.178	1.079, 1.287	0.00267	0.774
3-Hydroxybutyric acid	2441	584	1.147	1.051, 1.251	0.00284	0.820	1.143	1.047, 1.247	0.0134	0.774
CAR 2:0	2441	584	1.217	1.117, 1.326	1.82E-5	0.820	1.132	1.037, 1.235	0.0155	0.774
CAR 12:0	2441	584	1.219	1.120, 1.326	1.44E-5	0.820	1.132	1.036, 1.236	0.0155	0.774
Kynurenic acid	2441	584	1.235	1.138, 1.341	2.44E-6	0.577	1.149	1.029, 1.283	0.0278	0.750
CAR 10:0	2441	584	1.199	1.104, 1.302	3.10E-5	0.708	1.109	1.016, 1.209	0.0334	0.774
CAR 8:0	2441	584	1.184	1.092, 1.282	6.33E-5	0.708	1.101	1.011, 1.198	0.0374	0.774
Xanthurenic acid	2441	584	0.927	0.848, 1.013	0.102	0.577	0.910	0.836, 0.991	0.0374	0.750
N-Oleoyl glycine	2441	584	1.146	1.036, 1.268	0.0105	0.577	1.113	1.005, 1.232	0.0449	0.372
Acetoacetic acid	2441	584	1.061	0.977, 1.152	0.161	0.688	1.061	0.977, 1.152	0.162	0.750

HRs are expressed as per SD increase in log transformed metabolite level. HRs were calculated from Cox proportional hazards regression models. Model 1 covariates included metabolite, diabetes status, age, and sex. Model 2 covariates included metabolite, diabetes status, age, sex, body mass index, estimated glomerular filtration rate, total cholesterol, high density lipoprotein cholesterol, current smoking status, systolic blood pressure, hypertension medication use, and cardiovascular disease history. Interaction effect was tested with the addition of metabolite by diabetes status interaction term to each of the above models. P values were adjusted based on false discovery rate of 0.05. CAR indicates acylcarnitine and number preceding colon indicates length of carbon chain and number following colon indicates number of double bonds.

Given that uncorrelated and partially correlated metabolites can provide added predictive information, we aggregated the metabolites into a score via ridge regression. The fully adjusted HR was 2.57 (95% CI: 1.71, 3.86; P = 5.8 × 10^-16) with diabetes status included as a covariate and 2.56 (95% CI: 1.70, 3.85; P = 7.1 × 10^-16) without, and the metabolite score remained significant irrespective of which metabolite was left out (Supplemental Table 7). The addition of the full metabolite score improved the Harrell’s concordance index from 0.784 (95% CI: 0.766, 0.802) to 0.789 (95% CI: 0.771, 0.807) over the clinical covariates listed above, which includes diabetes status (bootstrapped Δ 95% CI: 0.0016, 0.010; P = 0.0002). There was no significant interaction effect of the metabolite score with diabetes (P-interaction = 0.283). Individuals with the highest quartile metabolite score had a HR of 1.57 (95% CI: 1.20, 2.05, P = 0.00094) for mortality compared with those with the lowest quartile. In comparison, diabetes status was associated with a HR of 1.75 (95% CI: 1.47, 2.08) for mortality, and those with the highest quartile HbA1c had a HR of 1.33 (95% CI: 1.04, 1.69) compared with those with the lowest quartile HbA1c. When adjusted for the same clinical covariates (as above without diabetes), the metabolite risk score had a concordance index of 0.780 (95% CI: 0.762, 0.798) for mortality, compared with 0.784 (95% CI: 0.766, 0.802) for diabetes and 0.781 (95% CI: 0.761, 0.801) for HbA1c.

Discussion

In this study, we sought to first compare and quantify the kinetics of BCAAs and BCKAs after an MMTT and to then explore kinetics for 194 additional metabolites in a small multiracial cohort with and without diabetes. After identifying 20 metabolites that had different postprandial kinetics, we examined the association of all-cause mortality with 10 of these metabolites that were measured after fasting in JHS to characterize their epidemiological significance and to illustrate how an MMTT may highlight important metabolic pathways.

We quantified for the first time the kinetics of BCKA changes after an MMTT, and our findings related to the upstream BCAAs are consistent with previous literature [26, 35, 36]. First, we observed higher baseline values of BCAAs and BCKAs in participants with diabetes in our MMTT study, but after adjustment for baseline differences, BCAA levels were similar between groups across all timepoints. In contrast, 2 BCKAs, α-ketoisocaproate and α-ketoisovalerate, remained elevated in participants with diabetes at 120 min before returning to a similar baseline, indicating a relatively slow metabolism of BCKA in diabetes.

From our exploratory analysis, we observed that diabetes affects many metabolite classes’ responses to a meal, notably FFAs, acylcarnitines, and quinoline carboxylic acids. At baseline, participants in the diabetes group had elevated saturated and unsaturated FAs, which decreased postprandially as a result of increased uptake of triglyceride rich lipoprotein particles into adipose tissue and a decreased release of FFAs from adipose tissue [37]. Consistent with previous observations, the levels of palmitic acid and linoleic acid, 2 of the most abundant saturated and unsaturated fatty acids respectively, had a reduced rate of decrease and recovery after the MMTT in the diabetes group, possibly reflecting poorer FFA uptake and utilization in skeletal muscle [26, 38]. In further support of impaired fatty acid metabolism, a delayed kinetics pattern was observed for 3-hydrobutyric acid, acetoacetic acid, and medium-chain acylcarnitines. Of note, all participants with diabetes were regularly consuming metformin, which has been shown to decrease plasma FFA [39]. Although participants withheld metformin the day of MMTT, its chronic use may have resulted in lower postprandial FFA and a smaller between group difference. Additionally, regular metformin use affected levels of ornithine, citrulline, and certain plasma phosphatidylcholines in metabolomic studies of individuals with diabetes [40, 41], but neither ornithine nor citrulline had notable postprandial kinetic differences in our MMTT study.

Tryptophan is an essential amino acid that is metabolized via the kynurenine pathway into downstream metabolites that include kynurenic acid and xanthurenic acid, both of which we observed to have differential kinetics after MMTT and high levels in those without diabetes. Observational studies have shown that high levels of kynurenic acid is associated with prevalent diabetes [42], insulin resistance [43], heart failure [44], and myocardial infarction [45], and high levels of xanthurenic acid is associated with prevalent diabetes [42] and low cardiovascular and all-cause mortality [46]. However, how mechanistically kynurenine pathway metabolites relate to cardiometabolic diseases remains uncertain with reports finding that kynurenic acid improves energy expenditure and browning of adipose tissue [47], and xanthurenic acid induces pancreatic dysfunction [48]. Our findings show that in contrast to fasting, late after a meal challenge, xanthurenic acid and kynurenic acid are relatively more suppressed in those with diabetes than normal individuals, a finding which will require further studies to elaborate its significance.

Previous metabolomics studies of MMTT have been mostly performed in healthy individuals [13], and a few have compared comprehensive metabolic changes between those with and without type 2 diabetes [14, 15]. Our study is consistent with a previous report of differences in the time course responses of FFA and amino acids after a meal challenge [14], and we extend the literature to report novel findings for metabolites such as N-oleoyl glycine, xanthurenic acid, kynurenic acid, and arachidonic acid derivatives. Another report confirms our findings for differential postprandial response in acylcarnitines, but metabolites were only measured at baseline and 150 min after meal [15]. Here, we are able to provide more detailed kinetics on acylcarnitines over the course of 5 h.

From the MMTT perturbation, we uncovered differences among group kinetics of several metabolites that were not readily apparent when comparing their fasting levels. These differences highlight metabolites that might be markers of impaired nutrient handling and dysmetabolism, which may translate to differential risks for adverse outcomes. In the absence of a large cohort study that administered MMTTs, we examined whether fasting levels of these particular metabolites would act as static markers of impaired nutrient metabolism and tested their associations with all-cause mortality in JHS. Nearly all of the metabolites that were highlighted by the MMTT study and tested in JHS were associated with mortality even after adjustment for several cardiometabolic risk factors and irrespective of diabetes status. The latter finding suggests that although diabetes may alter metabolism, the markers of this change may also reflect pathophysiology that is not unique to diabetes.

Of the metabolites found to be associated with all-cause mortality in JHS, 3-hydroxybutyric acid [21, 22, 25] and xanthurenic acid [46] have been previously associated with all-cause mortality. However, most of these cohorts included primarily Caucasian participants [21, 22, 46], and one included a small number of female African Americans [25]. Thus, our study provides important confirmation of these findings in the largest African-American cohort to date. We further extend the literature with previously unreported associations for medium-chain acylcarnitines and N-oleoyl glycine. It is important to note that rather than arriving at these candidate markers of mortality through metabolome-wide association studies, we identified them through their differential postprandial kinetics in diabetes, providing additional insights into a mechanism by which their levels are modulated in humans. Furthermore, a composite score of these MMTT discovered metabolites’ fasting levels was associated with mortality risk independent of diabetes status. Future population-based studies are needed to elucidate how the postprandial metabolome reflects metabolic health, and our description of postprandial metabolite kinetics will serve as important background for the design of such studies.

Our study has several limitations. First, although our meal’s macronutrient composition matches recommended intake [49], its total calories is less than that of other published MMTT [13], and a larger caloric load may elicit different responses. Second, because we could only match MMTT groups by age and sex, differences may be confounded by other factors. For example, there were numerically more African Americans in the MMTT diabetes group, which was not adjusted for in the MMRM model because of convergence issues. In addition, we could not adjust for severity or duration of diabetes in our analyses, but we tried to select a relatively homogenous diabetes group in the MMTT study. Our exploratory metabolite kinetics analyses are for hypothesis generation and require replication, but some of the findings are consistent with previous observations as noted above. Although the assayed metabolites span a breadth of metabolic pathways, there may be notable findings in the unmeasured metabolome, and the MMTT study's sample size may be underpowered to detect some smaller yet biologically relevant between group differences. Of the notable metabolites with different kinetics, we had only a subset of them available in JHS and tested the association of mortality with their fasting—not postprandial—levels. JHS enrolled exclusively self-identified African-American participants,; therefore, our novel metabolite associations with mortality will need replication. We chose to extend our MMTT study’s findings in JHS because African Americans have a high burden of diabetes and are an important population to study, but we do not believe our findings are unique to African Americans. Here, we have used race as a surrogate for social factors, but how ancestry related genetic variations determine fasting and postprandial metabolites deserves further investigation.

In conclusion, we found that in contrast to BCAAs, BCKAs are elevated after an MMTT among individuals with diabetes and that other metabolites, including FFAs, acylcarnitines, and products of tryptophan metabolism, also have different postprandial kinetics. Fasting levels of metabolites with different kinetics, such as medium-chain acylcarnitines, are associated with increased mortality in self-identified African Americans and warrant further investigation as markers of dysmetabolism.

Author disclosures

MW, MB, PL, KPW, BT, AK, and RJRF were employees of Pfizer Inc. at the time the study was conducted. All other authors have no conflicts of interest.

Funding

Pfizer, Inc. sponsored the research, participated in the study design, and had no restrictions on publication. MYM is supported by NIH 5T32HL007208. MYM, JGW, and REG are supported by R01DK081572. The Jackson Heart Study is supported and conducted in collaboration with Jackson State University (HHSN268201800013I), Tougaloo College (HHSN268201800014I), the Mississippi State Department of Health (HHSN268201800015I, HHSN26800001) and the University of Mississippi Medical Center (HHSN268201800010I, HHSN268201800011I, and HHSN268201800012I) contracts from the National Heart, Lung, and Blood Institute and the National Institute for Minority Health and Health Disparities.

Disclaimer

The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services.

Data Availability

Primary data described in the manuscript are available upon request to the correspondent author. Data from the Jackson Heart Study are available via requests through the database of Genotypes and Phenotypes sponsored by the National Institutes of Health.

Appendix A. Supplementary data

The following are the Supplementary data to this article: