Metabolite Signatures of Metabolic Risk Factors and their Longitudinal Changes

This study tested metabolite associations with risk factors cross-sectionally and with risk factor changes over time to uncover mechanistic links between metabolomics dysregulation and metabolic risk.

Abstract

Context:

Metabolic dysregulation underlies key metabolic risk factors—obesity, dyslipidemia, and dysglycemia.

Objective:

To uncover mechanistic links between metabolomic dysregulation and metabolic risk by testing metabolite associations with risk factors cross-sectionally and with risk factor changes over time.

Design:

Cross-sectional—discovery samples (n = 650; age, 36–69 years) from the Framingham Heart Study (FHS) and replication samples (n = 670; age, 61–76 years) from the BioImage Study, both following a factorial design sampled from high vs low strata of body mass index, lipids, and glucose. Longitudinal—FHS participants (n = 554) with 5–7 years of follow-up for risk factor changes.

Setting:

Observational studies.

Participants:

Cross-sectional samples with or without obesity, dysglycemia, and dyslipidemia, excluding prevalent cardiovascular disease and diabetes or dyslipidemia treatment. Age- and sex-matched by group.

Interventions:

None.

Main Outcome Measure(s):

Gas chromatography-mass spectrometry detected 119 plasma metabolites. Cross-sectional associations with obesity, dyslipidemia, and dysglycemia were tested in discovery, with external replication of 37 metabolites. Single- and multi-metabolite markers were tested for association with longitudinal changes in risk factors.

Results:

Cross-sectional metabolite associations were identified with obesity (n = 26), dyslipidemia (n = 21), and dysglycemia (n = 11) in discovery. Glutamic acid, lactic acid, and sitosterol associated with all three risk factors in meta-analysis (P < 4.5 × 10⁻⁴). Metabolites associated with longitudinal risk factor changes were enriched for bioactive lipids. Multi-metabolite panels explained 2.5–15.3% of longitudinal changes in metabolic traits.

Conclusions:

Cross-sectional results implicated dysregulated glutamate cycling and amino acid metabolism in metabolic risk. Certain bioactive lipids were associated with risk factors cross-sectionally and over time, suggesting their upstream role in risk factor progression. Functional studies are needed to validate findings and facilitate translation into treatments or preventive measures.

Cardiovascular disease (CVD) is the foremost cause of death in the developed world and is emerging as a leading cause of death in the developing world (1, 2). The simultaneous presence of multiple CVD risk factors is used to diagnose metabolic syndrome (3–5). Metabolic syndrome is associated with cardiovascular morbidity and mortality, and current estimates place its prevalence in U.S. adults at nearly one-third (6, 7). Thanks to advances in mass spectrometry, population-based studies are now able to identify metabolite associations with metabolic risk factors in plasma and other biosamples (8–12). We hypothesized that metabolic pathway disturbances underlie the major components of metabolic syndrome. We aimed to identify plasma metabolite signatures of obesity, dyslipidemia, and dysglycemia cross-sectionally and of serial changes in the same metabolic traits.

To that end, we designed an efficient 2×2×2 factorial study to evaluate plasma metabolites as biomarkers of obesity, dyslipidemia, and dysglycemia. Metabolomic profiling was conducted using gas chromatography-mass spectrometry (GC-MS) on plasma samples from 650 Framingham Heart Study (FHS) participants. Independent external replication was performed on biosamples from 670 participants in the BioImage Study. We also evaluated associations of our complete metabolite set with longitudinal changes in body mass index (BMI), lipids (triglyceride [TG] and high-density lipoprotein-cholesterol [HDL-C]), and fasting glucose during 5–7 years of follow-up in FHS participants.

Subjects and Methods

Study samples

The FHS is an observational study that began recruitment of three generations of participants in 1948, 1971, and 2002, respectively (13–15). In discovery, we sampled attendees at Offspring Cohort Examination 8 (2005–2008; n = 3021) and Third Generation Cohort Examination 1 (2002–2005; n = 4095). Exclusion criteria included off-site examination (n = 107), no blood specimen (n = 126), lack of unrestricted consent (n = 335), prevalent CVD at baseline (n = 216), pharmacological treatment for diabetes (n = 257) or dyslipidemia (n = 1201), missing BMI data (n = 4), smoking at baseline examination (n = 772), or age < 50 or > 79 years for the Offspring cohort (n = 123) or < 25 or > 59 years for the Third Generation cohort (n = 183). The eligible sample was 3792 individuals. Replication samples were selected from individuals enrolled in the BioImage Study, an ongoing prospective observational study of subjects free of prevalent CVD (ClinicalTrials.gov Identifier NCT00738725) (16). In June 2009, enrollment was complete for 6822 subjects from the metropolitan areas of Chicago, Illinois, and Fort Lauderdale, Florida. Study participants included men ages 55–80 and women ages 60–80 years; exclusion criteria included self-reported use of lipid- or glucose-lowering medications. Nonfasting plasma and serum samples from FHS and BioImage were collected according to standard operating procedures, aliquoted immediately after centrifugation, and stored at −80°C with less than three freeze-thaw cycles before assay.

Clinical measures

Metabolic syndrome risk factor definitions were similar to those of the National Cholesterol Education Program's Adult Treatment Panel III report (5). Obesity was defined as BMI ≥ 30 kg/m². Obesity was selected as a stratifying variable rather than waist circumference because it was available in both discovery and replication, and the correlation between BMI and waist circumference was extremely high (in pooled Offspring and Third Generation cohort participants, r = 0.92 [n = 7001], adjusted for age, sex, and cohort). Dyslipidemia was defined as plasma TG ≥ 150 mg/dL or low plasma HDL-C (<40 mg/dL in men or <50 mg/dL in women). Dysglycemia was defined as fasting blood glucose ≥ 100 mg/dL. Plasma aspartate transaminase (AST) and alanine transaminase (ALT) levels were measured in FHS and BioImage blood samples using the kinetic method (Beckman Liquid-Stat Reagent Kit).

Study design

In discovery and replication, we implemented an efficient 2×2×2 factorial design by sampling for elevated vs nonelevated levels of BMI, lipids, and glucose (Table 1). Individuals were stratified to high vs normal BMI, dyslipidemia vs normal lipids, and high vs normal fasting glucose. This stratification produced eight groups of equal size, each with a unique combination of presence and absence of the three metabolic risk factors, which permitted analyses of individual risk factor axes (ie, BMI, dysglycemia, or dyslipidemia) using data from all participants. The 2×2×2 factorial design ensured that when selecting any of the three metabolic risk factors as a discriminatory factor for analysis, the distribution of the other two factors remained balanced across the eight groups. We aimed to sample 80–84 individuals for each group, while balancing age and sex (and cohort in FHS). The final sample size for the FHS was 650 participants; for BioImage, the sample size was 670. The longitudinal sample included members of the discovery sample who attended the follow-up examination (FHS Offspring Cohort Examination 9 [2011–2013] or Third Generation Cohort Examination 2 [2008–2010]). Participants without follow-up data for the metabolic traits of interest (BMI, TG, HDL-C, and glucose) were excluded. The final sample size for the longitudinal analysis was 554 participants. For individuals who reported treatment for dyslipidemia (n = 133) or diabetes (n = 30) at follow-up, values were imputed empirically to adjust for treatment effect (described below).

Table 1.

Study Design and Clinical Characteristics of Discovery and Replication Samples

	Group 1	Group 2	Group 3	Group 4	Group 5	Group 6	Group 7	Group 8
Obesitya	-	-	-	-	+	+	+	+
Dysglycemiab	-	+	-	+	-	+	-	+
Dyslipidemiac	-	-	+	+	-	-	+	+
Framingham Heart Study (discovery)
Sample size, n	84	83	84	84	82	84	74	83
Age, y	52 ± 14	54 ± 13	52 ± 14	54 ± 15	53 ± 12	54 ± 11	50 ± 14	55 ± 13
Women, %	50	49	50	50	50	50	57	49
BMI, kg/m2	24.5 ± 2.8	25.0 ± 2.7	25.9 ± 2.7	26.6 ± 2.3	32.4 ± 2.3	35.9 ± 5.1	35.4 ± 4.8	34.9 ± 4.9
Glucose, mg/dL	92 ± 5	107 ± 15	92 ± 5	107 ± 7	93 ± 4	109 ± 7	93 ± 5	122 ± 37
HDL-C, mg/dL	65 ± 18	62 ± 15	45 ± 11	44 ± 14	61 ± 13	58 ± 13	42 ± 10	42 ± 10
TG, mg/dL	81 ± 26	84 ± 26	167 ± 115	175 ± 91	92 ± 28	95 ± 29	165 ± 77	202 ± 148
BioImage Study (replication)
Sample size, n	83	84	83	84	84	84	84	84
Age, y	69 ± 6	69 ± 6	69 ± 6	70 ± 6	67 ± 6	68 ± 6	68 ± 6	68 ± 6
Women, %	49	50	51	50	50	50	50	50
BMI, kg/m2	25.6 ± 2.6	25.9 ± 2.6	25.8 ± 2.7	26.9 ± 2.3	33.6 ± 3.5	35.4 ± 4.8	33.8 ± 3.3	34.1 ± 3.8
Glucose, mg/dL	89 ± 8	110 ± 13	90 ± 9	114 ± 19	90 ± 8	112 ± 12	92 ± 6	116 ± 19
HDL-C, mg/dL	63 ± 14	64 ± 15	54 ± 17	50 ± 15	59 ± 13	59 ± 14	46 ± 11	48 ± 13
TG, mg/dL	104 ± 27	106 ± 35	221 ± 119	238 ± 118	105 ± 33	110 ± 29	224 ± 92	223 ± 107

The + and − signs represent the presence and absence, respectively, of obesity, dysglycemia, and dyslipidemia in each of the eight metabolic risk factor groups.

^aObesity was defined as BMI ≥ 30 kg/m².

^bDysglycemia was defined as a fasting blood glucose level ≥ 100 mg/dL.

^cDyslipidemia was defined as either TG level ≥ 150 mg/dL or HDL-C level < 40 mg/dL in men or < 50 mg/dL in women.

Metabolomic profiling

GC-MS profiling of plasma samples was conducted at the TNO (Zeist, The Netherlands) with randomized cohorts and blinded analysis. A list of analyte targets was created from a preliminary profiling experiment before target peaks were measured in experimental samples. For GC-MS analysis, 100-μL samples were extracted with methanol and dried under a stream of nitrogen. For two-step derivatization, oximation was completed with ethoxamine in pyridine, and then trimethylsilylation was completed by adding 100 μL of N-methyl-N-(trimethylsilyl) trifluoroacetamide (MSTFA). To each sample, a cocktail of internal standards was added for normalization: d3-leucine, ¹³C5-ribose, d3-glutamic acid, d5-phenylalanine, d4-alanine, d4-cholic acid, d4-citric acid, trifluoroacetylanthracene, and dicyclohexyl phthalate. The samples were injected into an Agilent 7890 N gas chromatograph connected to an Agilent 5975 mass spectrometer (for more detail, see Pellis et al, Ref. 17). The analytical runs were arranged into 27 randomized batches, each containing 24 primary samples and three quality control (QC) samples. QC samples were generated by pooling plasma specimens from 15 healthy donors (seven males, eight females) following standard collection protocols. Injections were made in duplicate for experimental and QC samples. ChemStation software (version E02.00.493; Agilent Technologies) was used for data processing (18). A target table was constructed using an in-house library containing the mass spectra and retention times of > 600 reference metabolites (authentic standards), > 100 annotated metabolites (spectral match with the National Institute of Standards and Technology library or spectra of silylated compounds published in literature), and > 200 unknown metabolites found in serum or plasma in previous studies. Compounds were detected either underivatized, oximated and silylated, or silylated. Some lipid compounds, eg, sphingomyelins, were detected as a thermal degradation product by losing the phosphocholine group in the injector (19). Metabolites (known or unknown) specific for this study were added to the target table by searching the first three batches of samples for peaks present but not accounted for by the preselected target table. Peaks present in blanks and chemicals were excluded from the target table based on analysis of blank solvents and blank derivatizations. When one metabolite yielded ≥ two derivatization products, the most reliable was selected (based on specificity of ions, lack of coelution, and intensity), and the others were excluded. A total of 175 targets were detected in this study and quantified in all samples by reconstructing an ion chromatogram of a specific mass from the mass spectrum of the target. The quantification for all targets was manually checked by visual control, and peak integration was corrected as needed. Components detected in the GC-MS runs were quantified through peak areas normalized to an internal standard peak area. The internal standard was selected individually for each component: the standard yielding the smallest variability observed in QC samples. To compensate for different batch medians and intra-batch linear trends, batch and trend corrections were applied as needed. These methods are described in detail by van der Kloet et al (20). Normalized peak areas representing relative concentrations of the detected metabolites were used for statistical data analysis without being converted into absolute concentrations.

Statistical methods

Statistical analyses were performed using SAS software version 9.2 for Windows (SAS Institute Inc). We calculated rank normal scores of each metabolite using the Blom method (21). In discovery and replication, general linear mixed models were used to identify metabolites that differed across all eight metabolic groups or with individual risk factors (SAS procedure GLIMMIX). Metabolites were treated as outcome variables, and metabolic group was a categorical predictor. Covariates included age, sex, sex*age, cohort, and batch effect, with additional adjustment in FHS for cohort, cohort*sex, and cohort*age. The effect of each main risk factor was studied using models that adjusted for the other two risk factors. Analogous methods were applied in replication, except cohort adjustments were not necessary. False discovery rate (FDR) was used to select metabolites for replication (SAS procedure MULTTEST). Metabolites for replication were chosen as follows: 1) FDR ≤ 0.01 in the eight-group comparison; and 2) top 10 metabolites from each main effect analysis, prioritized by discovery FDR. Metabolites that were highly correlated with a previously selected metabolite were excluded (r ≥ 0.7). Meta-analyses were carried out using a fixed-effect approach with inverse-variance weighting (22). Specifically, the estimated combined effect was the inverse variance weighted average of those in discovery and replication. Associations of AST and ALT with metabolites were conducted using general linear mixed models adjusting for age, sex, alcohol consumption (average drinks per week), and metabolic risk factors (obesity, dyslipidemia, and dysglycemia), with additional adjustment in FHS for cohort, cohort*sex, and cohort*age. Plasma levels of ALT and AST were natural log-transformed and standardized, and outlier values were excluded.

In our longitudinal study, we used general linear models (SAS procedure GLM). Each metabolite was treated as the predictor, and change from baseline for each continuous metabolic trait (BMI, TG, HDL-C, and glucose) was the outcome. Covariates included age, sex, cohort, and baseline values of the metabolic traits. For participants on lipid or diabetes medications at follow-up, values were imputed empirically. Specifically, HDL-C was reduced by 10%, TG was elevated by 20%, and glucose was increased by 10% or to 126 mg/dL (whichever was higher), respectively. As a sensitivity analysis, metabolite associations were retested with treated individuals excluded. Multi-metabolite associations with longitudinal change were assessed using stepwise regression with forward selection. For each trait, candidate lipids were considered to be those with FDR < 0.25 in the single-marker analysis, and a P value threshold corresponding to FDR < 0.25 was used for stepwise selection [at step i, P < (0.25 × i)/119]. Aggregate partial r² was calculated for multi-metabolite panels after forward selection was complete. Using this approach, highly correlated markers could not enter the model.

The Bonferroni method was used to correct for multiple testing (P < .05/number of measured metabolites). For discovery and longitudinal analyses, P < 4.2 × 10⁻⁴ was used (0.05/119). The replication threshold was P < 1.4 × 10⁻³ (0.05/37). For meta-analysis, a more stringent cutoff was implemented by including a term for the number of metabolic risk factors evaluated (n = 3): P < .05/(37 × 3) = 4.5 × 10⁻⁴.

Results

Study sample characteristics

Clinical characteristics of study participants in discovery and replication are shown in Table 1. Mean age and percentage of women were constant across groups. BioImage participants were slightly older than FHS participants (16). BMI, lipids, and glucose levels differed systematically across metabolic groups per study design. Supplemental Table 1 reports clinical characteristics of the longitudinal samples, and the distribution of changes from baseline for each metabolic trait is provided in Supplemental Figure 1, A–D. The mean changes from baseline to follow-up were 0.6 ± 2.6 kg/m² for BMI, −0.5 ± 84.6 mg/dL for TG, 3.6 ± 10.7 mg/dL for HDL-C, and 0.5 ± 20.4 mg/dL for glucose.

Metabolite associations in discovery

In discovery, 149 metabolites were detected by GC-MS. Of these, 26 were unknown, one was a duplicate, and two were exogenous compounds; these were excluded before analysis. The median coefficient of variation in analytical replicates was 9.3% (25th percentile = 6.4%; 75th percentile = 10.9%). One analyte, cysteine, was excluded due to poor measurement quality. The final metabolite set consisted of 119 metabolites, consisting of alcohols (4.2%), amino acids (23.5%), lipids (38.7%), organic acids (13.4%), sterols (2.5%), sugars (9.2%), and others (8.4%). See Supplemental Table 2 for an annotated list of these metabolites.

We identified cross-sectional associations of metabolites from multiple classes with obesity (n = 26), dyslipidemia (n = 22), and dysglycemia (n = 11) at P < 4.2 × 10⁻⁴ (Tables 2–4). The top three metabolites associated with obesity were glutamic acid (P = 2.1 × 10⁻¹⁸), sitosterol (P = 1.6 × 10⁻¹⁰), and uric acid (P = 3.0 × 10⁻¹⁰) (Table 2). The strongest markers for dyslipidemia were sphingomyelins (SMs), including SM (d18:1/24:0) and SM (d18:1/24:1) (P = 1.4 × 10⁻²⁷), SM (d18:2/24:0) and SM (d18:2/24:1) (P = 4.5 × 10⁻²⁰), and SM (d18:1/16:0) (P = 7.4 × 10⁻¹¹) (Table 3). For dysglycemia, the top three metabolites were glucose (P = 1.4 × 10⁻⁴²), fructose (P = 2.9 × 10⁻⁷), and 2-hydroxybutanoic acid (P = 5.5 × 10⁻⁷) (Table 4). Glutamic acid and lactic acid were positively associated with all three risk factors. Several metabolites were associated with multiple risk factors, including direct associations of branched chain amino acids (BCAAs) and inverse associations of sitosterol and campesterol with obesity and dyslipidemia (Supplemental Table 3). Forty metabolites differed across the eight metabolic groups at P < 4.2 × 10⁻⁴ (Supplemental Table 2).

Table 2.

Cross-Sectional Metabolite Associations With Obesity in Discovery, Replication, and Meta-Analysis

Metabolite	Discovery (FHS)			Replication (BioImage)			Meta-Analysis
	βa	SE	P Valueb	βa	SE	P Valuec	βa	SE	P Valued
Glutamic acid	0.62	0.07	2.1 × 10−18	0.36	0.07	2.3 × 10−7	0.49	0.05	5.3 × 10−24
Mannose	0.43	0.08	1.5 × 10−8	0.51	0.07	2.5 × 10−12	0.47	0.05	7.0 × 10−20
Sitosterol	−0.48	0.07	1.6 × 10−10	−0.34	0.07	2.9 × 10−6	−0.41	0.05	2.2 × 10−15
Uric acid	0.43	0.07	3.0 × 10−10	0.31	0.07	1.8 × 10−5	0.38	0.05	2.4 × 10−14
Glycerol	0.36	0.06	1.1 × 10−8	0.25	0.07	2.2 × 10−4	0.31	0.05	1.3 × 10−11
Glycine	−0.34	0.07	7.7 × 10−7	−0.30	0.07	2.9 × 10−5	−0.32	0.05	7.4 × 10−11
2-Ketoglutaric acid	0.37	0.07	5.7 × 10−7	0.30	0.07	8.2 × 10−5	0.33	0.05	1.8 × 10−10
Asparagine	−0.28	0.07	4.0 × 10−5	−0.33	0.07	3.7 × 10−6	−0.30	0.05	5.3 × 10−10
Indole-3-propionic acid	−0.33	0.08	1.8 × 10−5	−0.25	0.08	1.1 × 10−3	−0.29	0.05	8.0 × 10−8
Inositol	−0.30	0.07	2.2 × 10−5	−0.23	0.07	1.9 × 10−3	−0.27	0.05	1.5 × 10−7
SM (d18:2/16:0)	0.30	0.07	4.8 × 10−6	0.15	0.07	0.02	0.23	0.05	8.8 × 10−7
4-Hydroxyglutamate semialdehyde	−0.21	0.07	4.6 × 10−3	−0.29	0.07	7.2 × 10−5	−0.25	0.05	1.3 × 10−6
2-Hydroxybutanoic acid	0.31	0.08	3.2 × 10−5	0.16	0.08	0.04	0.24	0.05	7.8 × 10−6
LPA 16:0	−0.31	0.07	2.7 × 10−5	−0.14	0.07	0.05	−0.23	0.05	1.3 × 10−5
Leucine	0.30	0.07	1.5 × 10−5	0.13	0.07	0.09	0.22	0.05	1.3 × 10−5
Phenylalanine	0.32	0.07	1.9 × 10−5	0.14	0.08	0.08	0.23	0.05	1.5 × 10−5
Lactic acid	0.34	0.07	6.3 × 10−6	0.09	0.07	0.21	0.22	0.05	4.0 × 10−5
Aminomalonic acid	−0.29	0.07	4.6 × 10−5	−0.10	0.07	0.12	−0.19	0.05	8.6 × 10−5
Sn-Glycerol-3-phosphate	−0.25	0.07	6.0 × 10−4	−0.15	0.07	0.04	−0.20	0.05	9.2 × 10−5
Campesterol	−0.42	0.08	3.3 × 10−8	-	-	-	-	-	-
Valine	0.35	0.07	1.2 × 10−6	-	-	-	-	-	-
LPC 18:2	−0.34	0.07	1.7 × 10−6	-	-	-	-	-	-
SM (d18:2/18:0)	0.32	0.07	7.5 × 10−6	-	-	-	-	-	-
Pyruvic acid	0.34	0.08	8.7 × 10−6	-	-	-	-	-	-
Creatinine	0.32	0.08	3.0 × 10−5	-	-	-	-	-	-
SM (d16:1/18:0)	0.29	0.07	1.1 × 10−4	-	-	-	-	-	-
Isoleucine	0.25	0.07	2.5 × 10−4	-	-	-	-	-	-
Tyrosine	0.27	0.07	3.1 × 10−4	-	-	-	-	-	-

Dashes represent the absence of replication and meta-analysis data for a metabolite. See Supplemental Table 2 for complete nomenclature.

^aEstimated β coefficients represent the mean differences in standardized metabolite measures between participants with and without obesity.

^bThe discovery P value threshold was P < 4.2 × 10⁻⁴ (0.05/119).

^cThe replication P value threshold was P < 1.4 × 10⁻³ (0.05/37).

^dThe meta-analysis P value threshold was P < 4.5 × 10⁻⁴ [0.05/(37 × 3)]. All significant markers in meta-analysis are shown in order of ascending P value.

Table 3.

Cross-Sectional Metabolite Associations With Dyslipidemia in Discovery, Replication, and Meta-Analysis

Metabolite	Discovery (FHS)			Replication (BioImage)			Meta-Analysis
	βa	SE	P Valueb	βa	SE	P Valuec	βa	SE	P Valued
SM (d18:1/24:0) and SM (d18:1/24:1)	−0.73	0.06	1.4 × 10−27	−0.44	0.07	4.2 × 10−11	−0.59	0.05	6.2 × 10−38
Glutamic acid	0.45	0.07	6.9 × 10−11	0.50	0.07	6.9 × 10−13	0.48	0.05	5.3 × 10−23
Myo-inositol 1,2-cyclic phosphate	0.33	0.07	9.0 × 10−6	0.53	0.07	2.7 × 10−15	0.44	0.05	1.6 × 10−19
SM (d18:1/17:0)	−0.46	0.07	4.9 × 10−10	−0.33	0.07	4.0 × 10−6	−0.40	0.05	9.2 × 10−15
SM (d18:0/16:0)	−0.45	0.07	1.9 × 10−10	−0.24	0.07	1.1 × 10−3	−0.35	0.05	3.5 × 10−12
Leucine	0.29	0.07	2.8 × 10−5	0.41	0.07	2.6 × 10−8	0.35	0.05	4.1 × 10−12
Alanine	0.29	0.08	1.9 × 10−4	0.44	0.07	4.3 × 10−9	0.36	0.05	5.7 × 10−12
Lactic acid	0.37	0.07	6.9 × 10−7	0.33	0.07	8.2 × 10−6	0.35	0.05	1.7 × 10−11
Erythronic acid	0.22	0.06	5.2 × 10−4	0.43	0.07	2.0 × 10−9	0.32	0.05	2.4 × 10−11
Sitosterol	−0.27	0.07	2.9 × 10−4	−0.37	0.07	3.9 × 10−7	−0.32	0.05	5.1 × 10−10
Maltose	0.26	0.07	5.3 × 10−4	0.29	0.08	1.7 × 10−4	0.27	0.05	2.9 × 10−7
Serine	−0.31	0.07	2.8 × 10−5	−0.22	0.07	3.6 × 10−3	−0.26	0.05	4.4 × 10−7
Uric acid	0.24	0.07	4.5 × 10−4	0.22	0.07	2.3 × 10−3	0.23	0.05	3.0 × 10−6
SM (d18:2/16:0)	−0.30	0.07	7.8 × 10−6	−0.12	0.06	0.07	−0.21	0.05	7.5 × 10−6
SM (d16:1/24:1)	−0.46	0.07	4.6 × 10−10	0.00	0.07	0.98	−0.23	0.05	8.5 × 10−6
3-Methyl-2-oxovaleric acid	0.20	0.07	1.6 × 10−3	0.23	0.07	1.6 × 10−5	0.21	0.05	1.6 × 10−5
SM (d18:2/24:0) and SM (d18:2/24:1)	−0.67	0.07	4.5 × 10−20	-	-	-	-	-	-
SM (d18:1/16:0)	−0.49	0.07	7.4 × 10−11	-	-	-	-	-	-
SM (d18:2/20:0)	−0.35	0.07	1.5 × 10−7	-	-	-	-	-	-
LPC 18:2	−0.34	0.07	1.1 × 10−6	-	-	-	-	-	-
Isoleucine	0.33	0.07	2.1 × 10−6	-	-	-	-	-	-
SM (d17:1/16:0)	−0.31	0.07	1.5 × 10−5	-	-	-	-	-	-
Campesterol	−0.29	0.07	1.3 × 10−4	-	-	-	-	-	-
SM (d18:1/23:0)	−0.28	0.07	1.4 × 10−4	-	-	-	-	-	-
Valine	0.27	0.07	2.2 × 10−4	-	-	-	-	-	-
Quinic acid	−0.27	0.08	4.1 × 10−4	-	-	-	-	-	-

Dashes represent the absence of replication and meta-analysis data for a metabolite. See Supplemental Table 2 for complete nomenclature.

^aEstimated β coefficients represent the mean differences in standardized metabolite measures between participants with and without dyslipidemia.

^bThe discovery P value threshold was P < 4.2 × 10⁻⁴ (0.05/119).

^cThe replication P value threshold was P < 1.4 × 10⁻³ (0.05/37).

^dThe meta-analysis P value threshold was P < 4.5 × 10⁻⁴ [0.05/(37 × 3)]. All significant markers in meta-analysis are shown in order of ascending P value.

Table 4.

Cross-Sectional Metabolite Associations With Dysglycemia in Discovery, Replication, and Meta-Analysis

Metabolite	Discovery (FHS)			Replication (BioImage)			Meta-Analysis
	βa	SE	P Valueb	βa	SE	P Valuec	βa	SE	P Valued
Glucose	0.95	0.06	1.4 × 10−42	0.95	0.06	9.0 × 10−47	0.95	0.04	2.1 × 10−102
Mannose	0.25	0.08	8.5 × 10−4	0.41	0.07	1.5 × 10−8	0.33	0.05	9.6 × 10−11
Lactic acid	0.31	0.07	4.3 × 10−5	0.37	0.07	9.2 × 10−7	0.34	0.05	1.4 × 10−10
Glutamic acid	0.30	0.07	1.1 × 10−5	0.31	0.07	8.2 × 10−6	0.31	0.05	2.7 × 10−10
2-Hydroxybutanoic acid	0.38	0.08	5.5 × 10−7	0.23	0.08	3.2 × 10−3	0.31	0.05	1.3 × 10−8
3-Methyl-2-oxovaleric acid	0.24	0.07	4.8 × 10−4	0.30	0.07	4.6 × 10−5	0.27	0.05	7.6 × 10−8
Alanine	0.26	0.08	8.8 × 10−4	0.27	0.07	2.5 × 10−4	0.26	0.05	6.6 × 10−7
Fructose	0.39	0.07	2.9 × 10−7	0.09	0.08	0.25	0.24	0.05	6.4 × 10−6
Glycine	−0.23	0.07	7.9 × 10−4	−0.20	0.07	4.4 × 10−3	−0.22	0.05	1.0 × 10−5
Aminomalonic acid	−0.35	0.07	1.1 × 10−6	−0.09	0.07	0.18	−0.21	0.05	1.6 × 10−5
2-Ketoglutaric acid	0.24	0.07	1.4 × 10−3	0.21	0.07	5.4 × 10−3	0.22	0.05	2.2 × 10−5
SM (d18:1/24:0) and SM (d18:1/24:1)	−0.22	0.06	5.5 × 10−4	−0.16	0.07	0.02	−0.19	0.05	2.9 × 10−5
Malic acid	0.30	0.07	3.5 × 10−5	0.13	0.07	0.09	0.22	0.05	3.3 × 10−5
4-hydroxyglutamate semialdehyde	−0.28	0.07	1.1 × 10−4	−0.14	0.07	0.05	−0.21	0.05	3.3 × 10−5
SM (d18:1/17:0)	−0.26	0.07	4.2 × 10−4	−0.16	0.07	0.03	−0.21	0.05	4.4 × 10−5
Sitosterol	−0.16	0.07	0.03	−0.24	0.07	9.1 × 10−4	−0.20	0.05	8.6 × 10−5
Malic acid	0.30	0.07	3.5 × 10−5	-	-	-	-	-	-
Pyruvic acid	0.30	0.08	8.9 × 10−5	-	-	-	-	-	-
C18:0 FA (stearic acid)	0.28	0.08	3.2 × 10−4	-	-	-	-	-	-
C16:0 FA (palmitic acid)	0.26	0.07	4.0 × 10−4	-	-	-	-	-	-

Abbreviation: FA, fatty acid. Dashes represent the absence of replication and meta-analysis data for a metabolite. See Supplemental Table 2 for complete nomenclature.

^aEstimated β coefficients represent the mean differences in standardized metabolite measures between participants with and without dysglycemia.

^bThe discovery P value threshold was P < 4.2 × 10⁻⁴ (0.05/119).

^cThe replication P value threshold was P < 1.4 × 10⁻³ (0.05/37).

^dThe meta-analysis P value threshold was P < 4.5 × 10⁻⁴ [0.05/(37 × 3)]. All significant markers in meta-analysis are shown in order of ascending P value.

Metabolite associations in replication and meta-analysis

Thirty-eight metabolites met the replication criterion, of which 37 were successfully measured in BioImage samples. Replication results for metabolite associations with obesity (n = 10), dyslipidemia (n = 12), and dysglycemia (n = 7) at P < 1.4 × 10⁻³ are provided alongside their corresponding discovery results in Tables 2–4. We compared estimated effect sizes between discovery and replication phases for each metabolite and found high degrees of similarity (Supplemental Figure 2, A–C). All metabolites in replication showed consistent directionality of associations. Correlations between metabolites in both discovery and replication phases are shown in Supplemental Figures 3 and 4. In meta-analysis of results from discovery and replication, 32 markers were associated with one or more risk factors, 16 were associated with two or more risk factors, and three (glutamic acid, lactic acid, and sitosterol) were associated with all three risk factors (Tables 2–4).

Transaminase associations

Because the liver transaminases ALT and AST are robustly associated with metabolic syndrome and regulate several of our identified metabolites (23–26), we examined metabolite associations with ALT and AST. Glutamic acid and 2-ketoglutaric acid were significantly associated with ALT and AST in discovery and replication (Supplemental Table 4).

Metabolite associations with longitudinal changes in risk factors

We identified metabolites associated with serial change in BMI (n = 1), TG (n = 5), HDL-C (n = 2), and glucose (n = 1) at P < 4.2 × 10⁻⁴ (Table 5). All metabolites associated with longitudinal changes in risk factors were also associated with the same metabolic risk factor cross-sectionally (P < 4.2 × 10⁻⁴ in discovery), with consistent directionality of association. After exclusion of treated individuals, the associations of two longitudinal markers were attenuated, but the effect sizes and directionalities were unchanged (Supplemental Table 5).

Table 5.

Metabolite Associations With Longitudinal Changes in Metabolic Traits

Change in Metabolic Trait	Metabolite	Estimated βa	SE	P Valueb	Partial R2
BMI	LPA 16:0	−0.48	0.13	2.8 × 10−4	0.025
TG	SM (d18:1/16:0)	−21.38	4.94	1.8 × 10−5	0.035
	SM (d18:1/17:0)	−17.61	4.10	2.1 × 10−5	0.035
	Quinic acid	−13.74	3.48	9.1 × 10−5	0.030
	Sitosterol	−14.46	3.69	1.0 × 10−4	0.029
	SM (d18:1/24:0) and SM (d18:1/24:1)	−17.83	4.92	3.2 × 10−4	0.025
HDL-C	SM (d18:2/20:0)	2.59	0.67	1.2 × 10−4	0.029
	SM (d18:1/24:0) and SM (d18:1/24:1)	2.62	0.68	1.4 × 10−4	0.028
Glucose	Glucose	5.31	1.05	5.2 × 10−7	0.048

All markers of longitudinal change were also associated with the corresponding risk factor in cross-sectional analyses. See Supplemental Table 2 for complete nomenclature.

^aEstimated β coefficients represent the mean differences in longitudinal change (native units) in the stated metabolic trait per 1 SD difference in metabolite measures.

^bThe P value threshold for significance (P < 4.2 × 10⁻⁴) was determined by the Bonferroni method (0.05/119 metabolites).

Multi-metabolite panels were associated with longitudinal change in each metabolic trait (Table 6). For serial change in BMI, the sole marker was lysophosphatidic acid (LPA) 16:0, which explained 2.5% of change in BMI vs 2.1% explained by baseline BMI. Six metabolites were selected in the TG model: SM (d18:1/16:0), quinic acid, mannose, uric acid, phosphate, and 1-methylhistidine. In sum, these markers explained 15.3% of change in TG; baseline TG explained 23.2% of change. The metabolite panel for HDL-C consisted of SM (d18:2/20:0) and SM (d18:1/22:0), which explained nearly twice as much change in HDL-C as did baseline HDL-C (6.2 vs 3.8%). 1,2-Diglyceride and 3-methyl-2-oxovaleric acid together accounted for 4.4% of change in glucose, relative to 21.6% explained by baseline glucose.

Table 6.

Multi-Metabolite Associations With Longitudinal Changes in Metabolic Traits

Change in Metabolic Trait	Parameter	Estimated βa	SE	P Valueb	Partial R2
BMI	Baseline BMI	−0.07	0.02	8 × 10−4	0.021
	Baseline covariatesc				0.070
	LPA 16:0	−0.48	0.13	3 × 10−4	0.025
	Multi-metabolite panel				0.025
TG	Baseline TG	−0.57	0.05	<1 × 10−4	0.232
	Baseline covariatesc				0.399
	SM (d18:1/16:0)	−25.25	5.18	<1 × 10−4	0.046
	Quinic acid	−12.02	3.34	<4 × 10−4	0.025
	Mannose	11.33	3.58	2 × 10−3	0.020
	Uric acid	16.53	4.10	<1 × 10−4	0.032
	Phosphate	−10.71	3.80	5 × 10−3	0.016
	1-Methylhistidine	−9.45	3.41	6 × 10−3	0.015
	Multi-metabolite panel				0.153
HDL-C	Baseline HDL-C	−0.16	0.04	<1 × 10−4	0.038
	Baseline covariatesc				0.127
	SM (d18:2/20:0)	3.11	0.68	<1 × 10−4	0.040
	SM (d18:1/22:0)	−1.96	0.56	6 × 10−4	0.023
	Multi-metabolite panel				0.062
Glucose	Baseline glucose	−0.52	0.04	<1 × 10−4	0.216
	Baseline covariatesc				0.275
	1,2-Diglyceride	−3.22	0.92	6 × 10−4	0.023
	3-Methyl-2-oxovaleric acid	3.05	0.92	1 × 10−3	0.021
	Multi-metabolite panel				0.044

Metabolites also associated cross-sectionally with the corresponding metabolic risk factor are in bold. See Supplemental Table 2 for complete nomenclature.

^aEstimated β coefficients represent the mean differences in longitudinal change (native units) in the stated metabolic trait per 1 SD difference in metabolite measures.

^bA dynamic P value threshold for entrance into the model was defined based on FDR < 0.25 [at step i, P < (0.25 × i)/119].

^cBaseline covariates were as follows: age; sex; cohort; batch; and baseline values for total cholesterol, HDL-C, TG, glucose, and BMI.

Discussion

We measured circulating levels of 119 metabolites and identified metabolomic signatures of key metabolic risk factors at baseline and of their longitudinal changes during 5–7 years of follow-up. Many top cross-sectional findings in discovery were replicated in an independent external cohort, and meta-analysis identified three metabolites (glutamic acid, lactic acid, and sitosterol) associated with all three risk factors cross-sectionally. Our cross-sectional results identified risk factor associations with glutamic acid and other major amino acids—particularly branched chain and aromatic amino acids—that may contribute to metabolic risk through their involvement in transaminase-mediated pathways. Bioactive lipids also emerged as important markers of metabolic risk factors cross-sectionally and of their changes over time. Multi-metabolite analyses demonstrated the predictive power of the top lipid markers, suggesting an upstream role for lipid dysregulation in metabolic risk factor progression.

Amino acids

The top cross-sectional amino acid biomarker was glutamic acid (glutamate), which was directly associated with all three risk factors in discovery and replication. Glutamic acid is one of the most abundant amino acids and a key player in amino acid and carbohydrate metabolism. Glutamic acid has been linked to metabolic dysfunction, with increasing plasma levels identified in obesity, insulin resistance, and progression to type 2 diabetes in humans (8, 9, 27). The BCAAs—leucine, valine, and isoleucine—were directly associated with obesity and dyslipidemia in discovery. Aromatic amino acids phenylalanine and tyrosine were directly associated with obesity in discovery, whereas glycine and asparagine were inversely associated. Elevated levels of BCAAs and the same aromatic amino acids (as well as decreased levels of glycine) have been identified in recent cross-sectional studies as biomarkers of obesity, early insulin resistance, and risk of diabetes in humans (8, 10, 12, 27–29). A number of previous intervention studies, however, found that higher dietary BCAA intake reduces body fat and body weight in both humans and animal models, indicating that cross-sectional associations are not necessarily predictive of the results of intervention studies (30–33). We also examined levels of liver transaminases ALT and AST in association with our metabolites, and found strong associations of ALT and AST with glutamic acid and 2-ketoglutaric acid (its cognate α-keto acid) in discovery and replication. Transaminase-mediated reactions are key to amino acid catabolism and to the regulation of insulin secretion by pancreatic β-cells (34). We focused on ALT and AST because increased levels have been linked to obesity, diabetes, and CVD in large-scale epidemiological studies (23–26). Taken together, our results implicate dysregulated glutamate cycling via transaminase-mediated reactions, likely during metabolism of BCAAs and other amino acids, in the metabolic dysregulation underlying metabolic syndrome.

Lipids

One of the top cross-sectional lipid biomarkers was the lysophosphatidylcholine (LPC) species 1-linoleoyl-sn-glycero-3-phosphocholine (LPC 18:2). LPC 18:2 showed inverse associations with obesity and dyslipidemia in discovery. LPCs are low-abundance lipid derivatives involved in signaling processes; they are theorized to act through G protein-coupled receptors to trigger glucose-dependent insulin secretion and activate glucose uptake by adipocytes (35, 36). LPC 18:2 has recently emerged as a biomarker of metabolic dysfunction, showing robust inverse associations with insulin resistance, dysglycemia, type 2 diabetes, and obesity (12, 37, 38). LPC 18:2 was also reported as a predictive biomarker of incident coronary heart disease (39). These findings suggest that reduced levels of LPC 18:2 may contribute to the metabolic dysfunction that promotes CVD. LPA, a lipid derivative of LPCs, was our top longitudinal marker for BMI. LPA 16:0 (1-palmitoyl-L-α-lysophosphatidic acid) was inversely associated with BMI at baseline and with change in BMI in single marker and multi-metabolite analyses, explaining 2.5% of the longitudinal change in BMI. LPA 16:0 is the most common circulating form of LPA, a lipid mediator with roles in essential cellular processes. One signaling pathway in which LPA is generated involves the action of autotaxin (ATX), which is highly expressed in adipocytes (40). ATX is elevated in the visceral fat of obese and insulin-resistant individuals (41, 42). ATX-knockout mice exhibit enhanced adiposity and reduced plasma LPA, suggesting that disruption of ATX/LPA signaling in adipose tissue contributes to obesity and that lowered plasma LPA is a biomarker of this process (43). The detailed mechanism by which ATX mediates LPA signaling in humans, however, remains unclear, and other LPA-producing pathways that are independent of ATX also exist (44, 45). Complicating this picture further, some studies have reported direct associations between LPA and obesity in mice (46).

Our results point to sphingomyelins as biomarkers of dyslipidemia. Tightly linked to sterol metabolism, sphingomyelins are bioactive sphingolipids that regulate cell growth and immune functions (47, 48).

Ten sphingomyelin species were inversely associated with dyslipidemia in discovery, and five of these were inversely associated with longitudinal change in TG and/or directly associated with change in HDL-C. In multi-metabolite analyses, top sphingomyelin biomarkers explained 4.6 and 4.0% of longitudinal change in TG and HDL-C, respectively.

Two phytosterols, sitosterol and campesterol, were inversely associated with obesity and dysglycemia in discovery and replication, and sitosterol showed an inverse relation to TG in longitudinal analysis. These natural plant sterols, structurally similar to cholesterol, lower cholesterol absorption by displacing cholesterol from micelles in the intestinal tract, among other mechanisms (49). Increased intake of plant sterols is associated with cholesterol-lowering activity (50, 51), and reduced levels of plasma phytosterols may serve as a proxy for altered cholesterol homeostasis (52).

Organic acids

Lactic acid (lactate) was our top organic acid biomarker of all three risk factors in discovery. Lactic acid is an indicator of oxidative capacity, and hyperlactatemia is broadly characteristic of acute illness. In the clinical setting, lactic acid levels serve as a prognostic biomarker of a number of metabolic disorders (53). For example, increased plasma lactic acid has been found in the context of insulin resistance, diabetic ketoacidosis, and type 2 diabetes (54–56). Lactic acid has also been linked to incident diabetes in a large case-control study, suggesting a causal role of oxidative stress in metabolic disease (57). Particularly in the pediatric population, elevated lactate is associated with inborn errors of metabolism that affect key metabolic pathways involving pyruvate dehydrogenase, gluconeogenesis, and the tricarboxylic acid cycle (58). Multiple organic acids were directly associated with dysglycemia, including 2-hydroxybutanoic acid, malic acid, and pyruvic acid. Quinic acid was inversely associated with dyslipidemia cross-sectionally and with change in TG in single and multi-metabolite analyses. Quinic acid is a naturally occurring plant cyclic polyol found as chlorogenic acid in common dietary constituents, including tea, coffee, blueberries, apples, grapes, and cereals (59–61). A functional role for quinic acid as a prometabolite that enhances antioxidant production and DNA repair enhancement has been described (62). Although human studies on chlorogenic acid have been inconsistent, this compound has been shown to improve lipid metabolism, BMI, and glycemic traits in mice (63–65).

Strengths and limitations

Although our discovery and replication studies were modest in size (n = 650 and 670, respectively), the cost-effective factorial design of independent discovery and replication phases enabled the identification of a robust set of metabolites associated with metabolic risk factors. Sample selection and statistical analysis for the replication phase were highly consistent with discovery, allowing us to validate many top markers. The use of an independent external, comparably sized cohort for replication provided superior power to identify robust signals. The availability of longitudinal data enabled us to identify metabolites associated with risk factors at baseline and their serial changes and to estimate their predictive power in multi-metabolite panels. It should be noted, however, that baseline covariates including age, sex, and baseline metabolic risk factor values explained a greater proportion of longitudinal changes in metabolic risk factors than multi-metabolite panels alone.

Our study is limited by its use of predominantly middle-aged individuals of European ancestry in discovery and longitudinal analysis. We cannot generalize these findings to other ethnic groups or younger ages. The replication sample, however, was multiethnic (69% Caucasian, 13% Hispanic, 12% African American, 4% Asian, 2% other), and results were highly consistent with discovery. We were also limited by the number of detectable metabolites; newer platforms offer more comprehensive metabolomic screening. Of note, a recent population-based study showed that body weight change is associated with various lipoprotein classes, but our results lacked data on associations between lipoproteins and metabolic traits (66). In replication, restriction to uncorrelated metabolites limited our ability to analyze complete pathways.

Conclusion

We employed a factorial study design to identify cross-sectional metabolomic signatures of obesity, dyslipidemia, and dysglycemia, with independent external replication of top findings. Using follow-up data from the discovery sample, we identified metabolites associated with longitudinal changes in metabolic traits. Amino acid biomarkers of obesity and dysglycemia revealed a glutamic acid-enhanced signature of metabolic risk that may represent a systemic metabolic derangement involving BCAA metabolism and transaminase-mediated pathways. Organic acid intermediates were most highly associated with dyslipidemia and may represent shifts in cellular metabolism in response to oxidative stress. Longitudinal results highlighted alterations in bioactive lipids, with lysophospholipid derivatives as biomarkers of BMI and its longitudinal change and sphingomyelins associated with dyslipidemia and serial changes in TG and HDL-C. Our multi-metabolite analyses demonstrated a number of longitudinal associations because multi-metabolite panels explained a substantial proportion of longitudinal change in each metabolic trait compared with the corresponding baseline trait value. Our findings shed light on the profound breadth and depth of metabolomic derangements associated with metabolic syndrome risk factors. Follow-up studies guided by these results may help in the targeting of therapies to treat metabolic risk factors or prevent their emergence.