Serum Metabolites Associated with Healthy Diets in African Americans and European Americans

ABSTRACT

Background

High diet quality is associated with a lower risk of chronic diseases. Metabolomics can be used to identify objective biomarkers of diet quality.

Objectives

We used metabolomics to identify serum metabolites associated with 4 diet indices and the components within these indices in 2 samples from African Americans and European Americans.

Methods

We studied cross-sectional associations between known metabolites and Healthy Eating Index (HEI)-2015, Alternative Healthy Eating Index (AHEI)-2010, the Dietary Approaches to Stop Hypertension Trial (DASH) diet, alternate Mediterranean diet (aMED), and their components using untargeted metabolomics in 2 samples (n₁ = 1,806, n₂ = 2,056) of the Atherosclerosis Risk in Communities study (aged 45–64 y at baseline). Dietary intakes were assessed using an FFQ. We used multivariable linear regression models to examine associations between diet indices and serum metabolites in each sample, adjusting for participant characteristics. Metabolites significantly associated with diet indices were meta-analyzed across 2 samples. C-statistics were calculated to examine if these candidate biomarkers improved prediction of individuals in the highest compared with lowest quintile of diet scores beyond participant characteristics.

Results

Seventeen unique metabolites (HEI: n = 6; AHEI: n = 5; DASH: n = 14; aMED: n = 2) were significantly associated with higher diet scores after Bonferroni correction in sample 1 and sample 2. Six of 17 significant metabolites [glycerate, N-methylproline, stachydrine, threonate, pyridoxate, 3-(4-hydroxyphenyl)lactate)] were associated with ≥1 dietary pattern. Candidate biomarkers of HEI, AHEI, and DASH distinguished individuals with highest compared with lowest quintile of diet scores beyond participant characteristics in samples 1 and 2 (P value for difference in C-statistics <0.02 for all 3 diet indices). Candidate biomarkers of aMED did not improve C-statistics beyond participant characteristics (P value = 0.930).

Conclusions

A considerable overlap of metabolites associated with HEI, AHEI, DASH, and aMED reflects the similar food components and similar metabolic pathways involved in the metabolism of healthy diets in African Americans and European Americans.

Introduction

Healthy dietary patterns are recommended for health promotion and the prevention of chronic diseases. Higher diet quality, often assessed using the Dietary Guidelines for Americans, Alternative Healthy Eating Index (AHEI), the Dietary Approaches to Stop Hypertension Trial (DASH) diet score, and the Mediterranean diet score, has been associated with a lower risk of cardiovascular diseases, type 2 diabetes, kidney disease, and all-cause mortality (1–5). At an individual level, identifying those with greater compared with lower adherence to these diet quality indices and characterizing the types of dietary patterns each person is following are of significant interest because clinicians and policymakers can use this evidence to provide appropriate interventions. However, characterizing dietary patterns and assessing quality of diets are based on self-reported data, which are affected by systematic bias (i.e., recall bias, social desirability bias) (6).

Objective biomarkers of dietary patterns that do not have the systematic errors of self-reported data and are relatively easier to collect than 24-h urinary biomarkers (e.g., urea nitrogen for protein intake) have the potential to improve the assessment of dietary intake. Recently, studies have used metabolomics to identify novel biomarkers of dietary intake through characterization of small molecules in biological samples. These studies have detected biomarkers of certain food components (e.g., fruits, dairy) and beverages (e.g., red wine) (7–10). There is emerging evidence on metabolomic markers of dietary patterns using data from feeding studies and observational studies (11–1,9). However, prior studies of metabolomics of dietary patterns had certain limitations, including a relatively smaller sample size (n < 1500), focus on specific populations (i.e., postmenopausal women, Finnish male smokers), assessment of only 1 dietary pattern, or use of the targeted metabolomics approach, thereby limiting the metabolites that can be detected. Furthermore, only a few studies have examined if these metabolites are replicated in a different data set (11, 13). Thus, more comprehensive analyses and replication of previous findings, particularly in a general population, are needed to inform clinical practice and advance our knowledge of candidate biomarkers of healthy diets.

The current study aimed to identify metabolites associated with 4 diet quality indices [Healthy Eating Index (HEI)–2015, AHEI-2010, the DASH diet, alternate Mediterranean diet (aMED)] using untargeted metabolomics and to detect a combination of metabolites that are representative of healthy dietary patterns. A unique aspect of our study is that we used 2 different samples within a large prospective cohort study to assess if significant metabolites can be replicated.

Methods

Study design

We used baseline (visit 1, 1987–1989) data from the Atherosclerosis Risk in Communities Study (ARIC), a community-based cohort of 15,792 middle-aged adults (aged 45–64 y at enrollment). Mostly black and white men and women were recruited into ARIC from 4 sites in the United States: Washington County, Maryland; Forsyth County, North Carolina; Minneapolis, Minnesota; and Jackson, Mississippi (20). All participants provided informed consent, and the study protocol was approved by the Institutional Review Boards at all study sites. Procedures were followed in accordance with institutional committees on human experimentation.

Dietary assessment

At baseline, trained interviewers administered a modified 66-item semiquantitative Willett FFQ to assess participants’ usual food and beverage intakes in the past year (21). Visual aids were provided to help participants estimate portion sizes. Participants reported the frequency with which foods or beverages of a defined serving size were consumed. The FFQ had 9 choices ranging from “almost never” to “>6 per day.” Participants were not advised or counseled to follow a specific dietary pattern, and we calculated adherence to 4 dietary indices (HEI-2015, AHEI-2010, DASH, aMED) based on the responses from the FFQ. The 4 dietary indices are often used as a measure of diet quality, and details on the calculation of the indices and the components within the indices are published in our previous articles (4, 5).

Briefly, HEI-2015 assesses alignment to the Dietary Guidelines for Americans 2015–2020 and emphasizes higher intakes of whole grains, total fruits, whole fruits (without fruit juices), total vegetables, greens and beans, total protein, and seafood and plant protein and lower intakes of refined grains, saturated fats, and added sugars (22). AHEI is similar to HEI but was developed based on reported associations between foods and nutrients and chronic diseases (1). AHEI emphasizes higher intakes of whole grains, whole fruits, total vegetables, nuts and legumes, and healthy fats (i.e., omega-3 fatty acids and PUFA); moderate alcohol consumption; and lower intakes of refined grains, sugar-sweetened beverages, trans fats, saturated fats, and sodium. The DASH diet score was developed based on the DASH trial and emphasizes higher intakes of total fruits, total vegetables, whole grains, nuts and legumes, and low-fat dairy and lower intakes of red and processed meats, sugar-sweetened beverages, and sodium (2). aMED is a modified version of the traditional Mediterranean diet score and emphasizes higher intakes of whole grains, total fruits, total vegetables, fish, nuts and legumes, and MUFA; moderate alcohol consumption; and lower intakes of red and processed meat (23).

In sample 1, Spearman's correlations between the diet scores ranged from 0.48 for aMED and DASH to 0.77 for DASH and HEI (Supplemental Table 1). In sample 2, correlations ranged from 0.54 for aMED and DASH to 0.73 for DASH and HEI.

Metabolomic profiling

Metabolomic profiling was conducted in 2 phases in 2010 and 2014 using fasting (≥8 h) serum specimens that had been collected and stored at –80°C since baseline (10). The first sample (2010) was a random sample of African American ARIC participants from a single study center (Jackson, MS), and the second sample (2014) comprised African American and European American participants with sequencing data from all 4 ARIC study centers.

Untargeted metabolomic profiling was conducted by Metabolon using a GC-MS and ultra-HPLC system (ACQUITY liquid chromatographer and a ThermoFisher Scientific Q-Exactive high-resolution MS) with an electrospray ionization and Thermo Scientific Orbitrap MS analyzer (24, 25). Details on assay procedures have been published previously (24, 25). Briefly, an automated liquid handling robot (Hamilton Labstar; Hamilton Robotics) was used for sample extraction. At Metabolon, a data normalization step was performed to correct variation resulting from instrument interday tuning differences. Each compound was corrected in run-day blocks by registering the medians to a value of 1 and normalizing each data point proportionately (24). Using helium as carrier gas and a temperature ramp from 60° to 340°C in a 17.5-min period, samples were separated on a ThermoScientific 5% diphenyl/95% dimethyl polysiloxane fused silica column (20 m × 0.18 mm ID; 0.18-μm film thickness) for GC-MS. Using water, methanol, 0.1% formic acid, and 0.05% perfluoropentanoic acid, samples were gradient eluted from 2.1 × 100 mm Waters BEH C18 1.7-μm column for ultra-HPLC. After raw data were extracted, peaks were identified using Metabolon's in-house software and subsequently matched to a chemical library that had >5000 commercially available, purified standard compounds.

We restricted our analyses to known metabolites. All named metabolites have been confirmed using reference standards and were either tier 1 or tier 2 identification, with the exception of metabolites with a footnote in the tables that denote a lower level of certainty of their identity. Metabolites were identified as tier 1 when a minimum of 2 orthogonal measurements (e.g., accurate mass, retention index, fragmentation pattern) were compared to a reference standard (26, 27). If a reference standard was not available but evidence on physicochemical properties or spectral similarities was available, metabolites were identified as tier 2 (26, 27).

Metabolites were excluded if >80% of values were missing. For the rest of the metabolites with missing values, the lowest detectable value for that specific metabolite was imputed. Then, metabolites were rescaled to a median of 1 by dividing by the sample-specific median, and log transformed (log_e). Metabolites with a variance <0.01 (on a log scale) or without a variance were excluded. Outliers that were 5 SDs above or below the mean were capped at 5 SDs from the mean (28). After these preprocessing steps, there were 374 metabolites in sample 1 and 759 metabolites in sample 2. Fewer metabolites were available in sample 1 because an older platform was used to generate metabolomics data. With the improvement in the platform, more metabolites were captured in sample 2. In sample 1, 27 unknown metabolites were retroactively named, which were included in the analyses. We focused on 374 metabolites that were available in both sample 1 and sample 2.

Covariates

At baseline, participants self-reported sociodemographic information (age, sex, race/ethnicity, and education) and health behaviors (smoking status, duration and frequency of physical activity). BMI (in kg/m²) was derived from measured height and weight. Serum creatinine was used to calculate estimated glomerular filtration rate (eGFR) using the 2009 Chronic Kidney Disease Epidemiology Collaboration equation (29).

We analyzed age, physical activity, total energy intake, and clinical factors (BMI, eGFR) as continuous variables. Sex (male, female), a combined variable for race and study center in sample 2 (blacks in Maryland, blacks in North Carolina, whites in North Carolina, whites in Minnesota, blacks in Mississippi), education (<12 y, 12 y, >12 y), and smoking status (never, former, current smoker) were analyzed as categorical variables.

Statistical analysis

To create the analytic sample for this cross-sectional study, we restricted the study population to participants for whom metabolomic profiling was conducted and dietary intake data were available. We excluded participants with implausible dietary intake data (women: <500 or >3500 kcal/d; men: <700 or >4500 kcal/d), missing dietary intake information, or missing covariate information. The final analytic sample was 1806 participants in sample 1 and 2056 participants in sample 2 (Supplemental Figure 1).

We examined characteristics of the study population according to quintiles of all diet indices using proportions for categorical variables and means ± SDs for continuous variables.

Metabolomics data generated from the same set of serum specimens were used to study associations with diet indices. For 374 metabolites that were available in both sample 1 and sample 2, we used data from sample 1 as a discovery data set and data from sample 2 as a replication data set. For the primary analysis, we used multivariable linear regression models to assess the association between 1 unit increase in each diet score and individual metabolites by sample, adjusting for participant characteristics. We used linear regression models to examine the full range of diet scores. We verified that linear regression models were appropriate by assessing the assumption of linearity for 5% of randomly selected metabolites in each sample. Due to the large number of statistical tests, we used the most conservative method (Bonferroni-adjusted P values) to account for multiple testing (30). In sample 1, we used a threshold of 3.34 × 10⁻⁵ (0.05/4 dietary patterns/374 metabolites). In sample 2, we restricted the analysis to associations that were statistically significant in sample 1 and used a threshold of 1.35 × 10⁻³ (0.05/37 associations) to account for multiple testing. Then, we meta-analyzed significant metabolites in sample 1 and sample 2 using fixed-effects models (31). Metabolites (n = 385) that were available only in sample 2 were analyzed using the same set of covariates and a Bonferroni threshold of 3.25 × 10⁻⁵ (0.05/4 dietary patterns/385 metabolites) but were not meta-analyzed.

Then, we repeated this analysis by replacing overall diet indices with individual components within each diet index to study the associations between food components and metabolites. In the analysis of food components, we used a Bonferroni-adjusted threshold of 3.26 × 10⁻⁶ (0.05/41 food components/374 metabolites) for sample 1. In sample 2, we focused on associations that were statistically significant in sample 1 and used a threshold of 6.76 × 10⁻⁶ (0.05/74 associations). For metabolites that were only available in sample 2 (n = 385), we used a Bonferroni threshold of 3.17 × 10⁻⁶ (0.05/41 food components/385 metabolites).

Next, we used Spearman's correlation coefficients to depict the interrelation between significant metabolites (32). Last, we calculated C-statistics to assess whether the addition of significant metabolites improved the prediction of those in the highest compared with lowest quintiles of each dietary pattern beyond participant characteristics (age, sex, race–center, total energy intake, education, smoking status, physical activity, BMI, eGFR). We used extreme quintiles for the calculation of C-statistics to assess the ability of a panel of metabolites to classify participants with the highest compared with lowest level of adherence to diet scores. All analyses were conducted using Stata software version 15.0 (StataCorp) and R software version 3.6.2 (R Foundation for Statistical Computing).

Results

Baseline characteristics

Participants with the highest diet quality scores were more likely to be older, a high school graduate, more physically active, and were less likely to be current smokers compared with participants with the lowest diet quality scores (Table 1). They also had higher BMI, except that individuals with the highest aMED scores had slightly lower BMI. Total energy intake was higher for those with the higher AHEI and aMED scores, and it was lower for those with the higher HEI and DASH scores. Characteristics of the study population were similar in the 2 samples except for race/ethnicity. All of the participants were black in sample 1; those in the highest quintiles of HEI and aMED were more likely to be black and those in the highest quintiles of AHEI and DASH were less likely to be black in sample 2.

TABLE 1.

Selected baseline characteristics of African Americans and European Americans in ARIC¹

	HEI-2015		AHEI-2010		DASH		aMED
	Quintile 1	Quintile 5	Quintile 1	Quintile 5	Quintile 1	Quintile 5	Quintile 1	Quintile 5
Sample 1 (n = 1825)
Sample size, n	362	361	362	361	362	361	362	361
Median score (range)	58.7 (45.8–68.5)	79.9 (72.6–92.6)	33.3 (21.2–40.6)	61.4 (53.3–84.3)	16 (11–18)	29 (27–34)	2 (0–3)	6 (6–8)
Women, %	48.3	77.0	61.6	66.2	46.9	78.3	68.2	62.0
Age, y	52.8 ± 5.9	53.5 ± 5.6	52.4 ± 5.8	52.8 ± 5.6	52.6 ± 5.9	53.5 ± 5.6	52.8 ± 5.9	52.6 ± 5.7
Black, %	100	100	100	100	100	100	100	100
High school graduate, %	46.6	71.4	54.1	70.0	43.9	68.7	56.3	72.3
Physical activity index	2.0 ± 0.6	2.3 ± 0.8	2.0 ± 0.6	2.3 ± 0.8	2.1 ± 0.6	2.4 ± 0.7	2.0 ± 0.7	2.3 ± 0.8
Smoking status
Never smoker, %	36.7	58.4	50.0	50.1	41.7	58.8	53.3	52.6
Former smoker, %	21.2	26.3	19.9	25.5	20.4	26.4	18.2	21.6
Current smoker, %	41.9	15.2	30.1	24.4	37.9	14.8	28.4	25.7
BMI, kg/m2	29.1 ± 5.8	30.2 ± 5.9	29.5 ± 5.6	29.8 ± 6.1	28.8 ± 5.8	30.8 ± 6.1	30.1 ± 6.0	29.6 ± 6.1
eGFR, mL·min−1·1.73 m−2	111 ± 20.2	113 ± 18.2	111 ± 20.7	112 ± 16.9	113 ± 18.3	112 ± 18.5	110 ± 22.6	112 ± 17.9
Total energy intake, kcal/d	1682 ± 663	1376 ± 527	1360 ± 508	1870 ± 692	1733 ± 566	1443 ± 538	1312 ± 472	1822 ± 678
Sample 2 (n = 2056)
Sample size, n	412	411	412	411	412	411	412	411
Median score (range)	59.0 (42.5–72.9)	80.9 (73.8–81.9)	35.4 (20.4–44.3)	66.9 (57.4–83.3)	18 (12–20)	30 (28–35)	2 (0–3)	7 (6–8)
Women, %	43.7	65.6	50.7	61.8	38.3	67.8	57.0	53.2
Age, y	53.9 ± 5.6	55.3 ± 5.9	54.1 ± 5.6	54.6 ± 5.9	53.7 ± 5.6	55.3 ± 5.9	53.7 ± 5.5	54.5 ± 5.9
Black, %	21.7	25.5	28.2	20.2	38.4	17.0	22.5	24.2
High school graduate, %	70.4	82.7	71.3	82.7	65.7	72.3	72.1	81.0
Physical activity index	2.3 ± 0.8	2.6 ± 0.8	2.3 ± 0.7	2.7 ± 0.8	2.2 ± 0.7	2.7 ± 0.8	2.3 ± 0.8	2.6 ± 0.8
Smoking status
Never smoker, %	29.1	43.6	36.4	41.2	31.3	43.8	39.8	41.1
Former smoker, %	31.3	38.9	31.8	36.0	28.8	37.2	30.1	34.3
Current smoker, %	39.6	17.5	31.8	22.9	39.8	18.9	30.1	24.6
BMI, kg/m2	27.4 ± 5.2	28.0 ± 5.9	27.8 ± 5.4	28.0 ± 5.8	27.9 ± 5.2	27.8 ± 5.9	28.0 ± 5.4	27.3 ± 5.5
eGFR, mL·min−1·1.73 m−2	99.2 ± 16.5	100 ± 17.6	100 ± 16.4	101 ± 16.5	102 ± 17.0	99.7 ± 16.3	101 ± 16.3	101 ± 16.0
Total energy intake, kcal/d	1776 ± 686	1450 ± 487	1463 ± 559	1896 ± 593	1823 ± 622	1534 ± 496	1400 ± 509	1872 ± 590

¹Values are means ± SDs for continuous variables and percentages for categorical variables. Metabolomic profiling was conducted in 2 phases in 2010 and 2014. The first sample (2010) was a random sample of African American ARIC participants from a single study center (Jackson, MS), and the second sample (2014) comprised African Americans and European American participants with sequencing data from all 4 ARIC study centers. AHEI-2010, Alternative Healthy Eating Index–2010; aMED, alternate Mediterranean diet; ARIC, Atherosclerotic Risk in Communities Study; DASH, Dietary Approaches to Stop Hypertension; eGFR, estimated glomerular filtration rate; HEI-2015, Healthy Eating Index–2015.

Metabolites associated with high diet quality

In sample 1, a total of 37 metabolites (HEI: n = 10; AHEI: n = 9; DASH: n = 16; aMED: n = 2) were significantly associated with higher diet scores after Bonferroni correction (Table 2). At the Bonferroni threshold of 1.35 × 10⁻³ (0.05/37 associations), 27 of these 37 metabolites (HEI: n = 6; AHEI: n = 5; DASH: n = 14; aMED: n = 2) replicated in sample 2. Of these 27 replicated metabolites, 17 metabolites were unique, and the remaining metabolites were associated with multiple dietary patterns. Of these 17 unique metabolites in both sample 1 and sample 2, the most common metabolite category was lipids (n = 5, 29%), followed by amino acids (n = 4, 24%), cofactors and vitamins (n = 4, 24%), xenobiotics (n = 2, 12%), and carbohydrates (n = 2, 12%). Higher HEI and DASH scores were associated with a wide range of types of compounds (metabolic pathways), including amino acids, carbohydrates, cofactors and vitamins, lipids, and xenobiotics.

TABLE 2.

Significant metabolites associated with diet quality scores in sample 1 and sample 2 in African Americans and European Americans in ARIC ¹

		Sample 1		Sample 2		Sample 1 and Sample 22
Metabolite	Superpathways		P value		P value		P value
HEI-2015 (n = 10)
N-methylproline	Amino acid	0.03283	4.92 × 10−6*	0.01693	1.09 × 10−4*	0.02123	1.23 × 10−8
3-(4-Hydroxyphenyl)lactate	Amino acid	–0.00735	4.53 × 10−6*	–0.00432	2.34 × 10−3	–0.00564	9.51 × 10−8
Proline	Amino acid	–0.00482	2.20 × 10−5*	–0.00214	4.16 × 10−2	–0.00338	1.16 × 10−5
Glycerate	Carbohydrate	0.02546	2.71 × 10−21*	0.02090	1.53 × 10−24*	0.02257	<1.00 × 10−40
Pyridoxate	Cofactors and vitamins	0.01508	3.86 × 10−7*	0.01116	1.05 × 10−4*	0.01306	2.34 × 10−10
Threonate	Cofactors and vitamins	0.03417	1.62 × 10−17*	0.02216	5.33 × 10−22*	0.02513	3.57 × 10−37
Pantothenate	Cofactors and vitamins	0.01120	8.78 × 10−7*	0.00647	2.72 × 10−3	0.00871	2.50 × 10−8
Deoxycarnitine	Lipid	–0.00498	2.64 × 10−5*	–0.00458	4.17 × 10−6*	–0.00475	4.27 × 10−10
Stachydrine	Xenobiotics	0.04731	3.19 × 10−11*	0.05497	1.84 × 10−18*	0.05164	2.00 × 10−28
Paraxanthine	Xenobiotics	–0.02521	6.23 × 10−6*	–0.02187	1.45 × 10−3	–0.02388	3.20 × 10−8
AHEI-2010 (n = 9)
3-(4-Hydroxyphenyl)lactate	Amino acid	–0.00865	1.97 × 10−5*	–0.00569	1.13 × 10−3*	–0.00695	1.40 × 10−7
3-Methyl-2-oxovalerate	Amino acid	–0.04895	5.69 × 10−7*	–0.00236	8.68 × 10−1	–0.00862	2.13 × 10−1
Proline	Amino acid	–0.00732	4.09 × 10−7*	–0.00156	2.28 × 10−1	–0.00414	1.76 × 10−5
Glycerate	Carbohydrate	0.01631	2.41 × 10−6*	0.01063	2.89 × 10−5*	0.01262	6.44 × 10−10
Pyridoxate	Cofactors and vitamins	0.20531	5.9 × 910−8*	0.01490	2.45 × 10−5*	0.01752	1.01 × 10−11
Pantothenate	Cofactors and vitamins	0.01584	5.04 × 10−8*	0.00845	1.38 × 10−3	0.01180	1.42 × 10−9
CMPF	Lipid	0.02421	3.20 × 10−5*	0.02611	1.20 × 10−6*	0.02524	1.49 × 10−10
DHA (22:6n3)	Lipid	0.00815	2.08 × 10−5*	0.01123	1.09 × 10−9*	0.00975	1.66 × 10−13
Bradykinin	Peptide	–0.04959	1.85 × 10−7*	–0.00277	8.36 × 10−1	–0.03391	1.14 × 10−5
DASH (n = 16)
3-(4-Hydroxyphenyl)lactate	Amino acid	–0.01003	1.15 × 10−5*	–0.01173	1.31 × 10−7*	–0.01091	6.74 × 10−12
Indolepropionate	Amino acid	0.02680	7.04 × 10−7*	0.03211	2.49 × 10−6*	0.02885	8.28 × 10−12
N-acetylornithine	Amino acid	0.02131	4.01 × 10−9*	0.02084	1.45 × 10−9*	0.02107	2.27 × 10−17
N-methylproline	Amino acid	0.05151	6.31 × 10−7*	0.03894	1.34 × 10−8*	0.04277	5.64 × 10−14
Threonine	Amino acid	–0.00795	2.69 × 10−5*	–0.00291	7.73 × 10−2	–0.00509	4.14 × 10−5
Glycerate	Carbohydrate	0.04346	1.25 × 10−29*	0.03655	3.69 × 10−30*	0.03939	<1.00 × 10−40
Erythronate3	Carbohydrate	0.01718	9.72 × 10−8*	0.01381	2.18 × 10−16*	0.01453	9.57 × 10−23
Threonate	Cofactors and vitamins	0.04938	8.54 × 10−18*	0.03915	1.49 × 10−27*	0.04202	<1.00 × 10−40
Pyridoxate	Cofactors and vitamins	0.03463	3.35 × 10−16*	0.02879	1.65 × 10−10*	0.03190	2.45 × 10−25
Pantothenate	Cofactors and vitamins	0.02146	4.95 × 10−11*	0.01883	2.43 × 10−8*	0.02019	5.31 × 10−18
γ-Tocopherol	Cofactors and vitamins	–0.04949	3.81 × 10−7*	–0.03455	8.40 × 10−4*	–0.04248	1.92 × 10−9
Linoleate (18:2n–6)	Lipid	0.00778	5.26 × 10−6*	0.00649	1.11 × 10−4*	0.00712	2.47 × 10−9
Myo-inositol	Lipid	0.01089	1.51 × 10−8*	0.01219	3.56 × 10−15*	0.01168	1.96 × 10−22
Bradykinin	Peptide	–0.04988	3.53 × 10−6*	0.00532	7.55 × 10−1	–0.03423	1.63 × 10−4
Catechol sulfate	Xenobiotics	0.02705	2.76 × 10−10*	0.03899	2.09 × 10−14*	0.03200	9.75 × 10−23
Stachydrine	Xenobiotics	0.07985	6.52 × 10−15*	0.09139	1.51 × 10−20*	0.08587	2.45 × 10−34
aMED (n = 2)
Glycerate	Carbohydrate	0.02084	8.43 × 10−8*	0.01782	9.38 × 10−10*	0.01891	3.77 × 10−16
Stachydrine	Xenobiotics	0.04378	1.99 × 10−5*	0.05125	9.27 × 10−9*	0.04804	8.11 × 10−13

¹ (expressed as 1 unit higher in each diet score) was calculated from multivariable linear regression models adjusting for age, sex, race–center (sample 2 only), total energy intake, education, smoking status, physical activity index, BMI, and eGFR based on creatinine. *Metabolites were considered significant at the Bonferroni threshold in sample 1 (0.05/4 dietary patterns/374 metabolites = 3.34 × 10⁻⁵) and in sample 2 (0.05/37 associations = 1.35 × 10⁻³). AHEI-2010, Alternative Healthy Eating Index-2010; aMED, alternate Mediterranean diet; ARIC, Atherosclerotic Risk in Communities Study; CMPF, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate; DASH, Dietary Approaches to Stop Hypertension; HEI-2015, Healthy Eating Index–2015.

²Significant metabolites associated with diet quality scores were meta-analyzed.

³The metabolite has not been officially confirmed based on a standard.

Glycerate was associated with all 4 diet quality indices, and 6 of 17 unique metabolites were associated with >1 diet index (Figure 1, Table 3). The DASH diet had the highest number of unique metabolites (n = 8), 4 of which were amino acids, cofactors and vitamins, and lipids. In both samples, the other dietary patterns (aside from DASH) had a similar number of unique metabolites (range, 0–2), all of which were lipids. A heatmap of significant metabolites showed that those in the carbohydrate and xenobiotic categories were highly correlated with each other in both samples (shown in red clusters in Figure 2).

FIGURE 1.

Significant metabolites overlapping among healthy dietary patterns in 2 subsamples of African Americans and European Americans in the Atherosclerosis Risk in Communities Study. The total number of significant metabolites in sample 1 and sample 2 was 17. AHEI, Alternative Healthy Eating Index; aMED, alternate Mediterranean diet; DASH, Dietary Approaches to Stop Hypertension; HEI, Healthy Eating Index.

TABLE 3.

Significant metabolites associated with food components within each dietary pattern in sample 1 and sample 2 in African Americans and European Americans in ARIC¹

Metabolite	Superpathways	Sample 1Significant food components	Sample 2Significant food components	Other associated diets
HEI-2015 (n = 6)
N-methylproline	Amino acid	Total fruits (+)	Total fruits (+), whole fruits (+), dairy (+)	DASH
Glycerate	Carbohydrate	Total fruits (+), whole fruits (+), whole grains (+), dairy (+)	Total fruits (+), whole fruits (+), dairy (+), refined grains (+)	AHEI, DASH, aMED
Pyridoxate	Cofactors and vitamins	—	Total fruits (+), whole fruits (+), dairy (+)	AHEI
Threonate	Cofactors and vitamins	Total fruits (+), whole fruits (+)	Dairy (+), refined grains (+)	DASH
Deoxycarnitine	Lipid	—	Total fruits (–), total vegetables (–), dairy (–)	—
Stachydrine	Xenobiotics	Total fruits (+)	Total fruits (+), whole fruits (+), whole grains (+)	DASH, aMED
AHEI-2010 (n = 5)
3-(4-Hydroxyphenyl)lactate	Amino acid	—	Fruits (–)	DASH
Glycerate	Carbohydrate	Fruits (+), whole grains (+)	Fruits (+), SSB (–)	HEI, DASH, aMED
Pyridoxate	Cofactors and vitamins	—	Trans fat (+)	HEI
CMPF	Lipid	ω-3 fatty acids (+)	ω-3 fatty acids (+), trans fat (+)	—
DHA (22:6n3)	Lipid	ω-3 fatty acids (+), trans fat (+)	ω-3 fatty acids (+), trans fat (+)	—
DASH (n = 14)
3-(4-Hydroxyphenyl)lactate	Amino acid	—	Total fruits (–), dairy (–)	AHEI
Indolepropionate	Amino acid	Total fruits (+)	Total fruits (+)	—
N-acetylornithine	Amino acid	—	Total fruits (+), total vegetables (+), sodium (–), SSB (+)	—
N-methylproline	Amino acid	Total fruits (+)	Total fruits (+), dairy (+)	HEI
Glycerate	Carbohydrate	Total fruits (+), whole grains (+), red meat (+)	Total fruits (+), total vegetables (+), whole grains (+), dairy (+), red meat (+)	HEI, AHEI, aMED
Erythronate2	Carbohydrate	Total fruits (+)	Total fruits (+), nuts and legumes (+), dairy (+)	—
Threonate	Cofactors and vitamins	Total fruits (+)	Total fruits (+), total vegetables (+), dairy (+), red meat (+)	HEI
Pyridoxate	Cofactors and vitamins	SSB (+)	Total fruits (+), dairy (+), SSB (+)	HEI
Pantothenate	Cofactors and vitamins	—	Dairy (+)	—
γ-Tocopherol	Cofactors and vitamins	Total fruits (–)	Total fruits (–)	—
Linoleate (18:2n–6)	Lipid	—	Total fruits (+), total vegetables (+), SSB (+)	—
Myo-inositol	Lipid	—	Total fruits (+), total vegetables (+), SSB (+)	—
Catechol sulfate	Xenobiotics	Total fruits (+)	Total fruits (+), total vegetables (+), nuts and legumes (+), dairy (+)	—
Stachydrine	Xenobiotics	Total fruits (+)	Total fruits (+), whole grains (+), dairy (+)	HEI, aMED
aMED (n = 2)
Glycerate	Carbohydrate	Fruits (+), whole grains (+)	Fruits (+), whole grains (+)	HEI, AHEI, DASH
Stachydrine	Xenobiotics	Fruits (+)	Fruits (+)	HEI, DASH

¹A  + sign indicates that the level of the metabolite was higher among those with higher scores of the specific food component, and a – sign indicates that the level of the metabolite was lower among those with higher scores of the specific food component. Higher scores indicate higher diet quality for all components. Metabolites were considered significant at the Bonferroni threshold in sample 1 (0.05/41 food components/374 metabolites = 3.26 × 10⁻⁶) and in sample 2 (0.05/74 associations = 6.76 × 10⁻⁴). aMED, alternate Mediterranean diet; AHEI-2010, Alternative Healthy Eating Index–2010; ARIC, Atherosclerotic Risk in Communities Study; CMPF, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate; DASH, Dietary Approaches to Stop Hypertension; HEI-2015, Healthy Eating Index–2015; SSB, sugar-sweetened beverage.

²The metabolite has not been officially confirmed based on a standard.

FIGURE 2.

Spearman correlation coefficient for 17 serum metabolites significantly associated with healthy dietary patterns in sample 1 and sample 2, respectively, in African Americans and European Americans in the Atherosclerotic Risk in Communities Study. CMPF, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate. *The metabolite has not been officially confirmed based on a standard.

For metabolites available only in sample 2, 25 unique metabolites were significantly associated with high diet quality scores (HEI: n = 14; AHEI: n = 8; DASH: n = 20; aMED: n = 14) (Supplemental Table 2). More xenobiotic and lipid metabolites were associated with aMED in sample 2. Two xenobiotics (4-allylphenol sulfate, 2-aminophenol sulfate) and 2 lipids (N-stearoyltaurine, oleoyl sphingomyelin) were associated with all 4 diet quality scores. The DASH diet had the highest number of unique metabolites (n = 6), such as α-carboxyethyl hydroxychroman (CEHC) sulfate.

Metabolites associated with food components of high-quality diets

In both samples, many of the metabolites were positively associated with the fruit component within these diet quality indices (n₁ = 8; n₂ = 14) (Table 3). In both samples, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate (CMPF) and DHA (22:6n3) were positively associated with ω-3 fatty acids, and DHA (22:6n3) was negatively associated with trans fat (Table 3). Pantothenate was not associated with any food components in sample 1, but it was associated with higher low-fat dairy consumption in the DASH diet in sample 2 (Table 3).

Prediction of healthy dietary patterns with influential metabolites and participant characteristics

Compared with the models with only participant characteristics, C-statistics to predict those in the highest compared with lowest quintile of each diet quality score, except for aMED in sample 2, improved significantly when a panel of significant metabolites were added to the models (P < 0.02 for all models; Table 4). The DASH diet had the greatest improvement in C-statistics in sample 1 (difference in C-statistics = 0.076) and sample 2 (difference in C-statistics = 0.059), with the addition of 14 metabolites to the model. Other dietary patterns had a similar magnitude of increase with the additional significant metabolites (range of differences in C-statistics: 0.018–0.043). However, for aMED in sample 2, the C-statistic did not improve significantly with the addition of 2 metabolites to the model (difference in C-statistics = –0.001, P = 0.930).

TABLE 4.

C-statistics and difference in C-statistics for prediction of higher diet quality in African Americans and European Americans in ARIC using significant metabolites replicated across 2 samples compared with multivariable models that include only participant characteristics¹

Dietary pattern	C-statistics for participant characteristics	C-statistics for participant characteristics and metabolites2	Difference in C-statistics3	P value4
Sample 1
HEI-2015 (n = 6)	0.805	0.841	0.036 (0.017, 0.055)	<0.001
AHEI-2010 (n = 5)	0.799	0.818	0.018 (0.003, 0.033)	0.015
DASH (n = 14)	0.815	0.891	0.076 (0.052, 0.100)	<0.001
aMED (n = 2)	0.781	0.795	0.014 (0.002, 0.026)	0.019
Sample 2
HEI-2015 (n = 6)	0.802	0.845	0.043 (0.024, 0.061)	<0.001
AHEI-2010 (n = 5)	0.795	0.822	0.026 (0.011, 0.041)	0.001
DASH (n = 14)	0.834	0.894	0.059 (0.040, 0.079)	<0.001
aMED (n = 2)	0.791	0.789	–0.001 (–0.014, 0.013)	0.930

¹Participant characteristics include age, sex, race–center (only in sample 2), total energy intake, education, smoking status, physical activity index, BMI, and eGFR based on creatinine. aMED, alternate Mediterranean diet; AHEI-2010, alternative Healthy Eating Index–2010; ARIC, Atherosclerotic Risk in Communities Study; DASH, Dietary Approaches to Stop Hypertension; eGFR, estimated glomerular filtration rate; HEI-2015, Healthy Eating Index–2015.

²Metabolites for each dietary pattern are presented in Table 2. n indicates the number of significant metabolites associated with the respective dietary pattern.

³95% CIs in parentheses.

⁴ P value for the comparison of the C-statistics which include metabolites and participant characteristics relative to C-statistics which include only participant characteristics for the respective dietary pattern.

Discussion

In this sample of African Americans and European Americans, we identified 17 unique metabolites associated with high diet quality scores in 2 different samples using an untargeted metabolomics approach. These metabolites represented a wide range of compounds, including amino acids, carbohydrates, cofactors and vitamins, lipids, and xenobiotics. Several predictive metabolites that represent greater alignment to the high diet quality scores were consistent across multiple diet indices. Some of these predictive metabolites were uniquely associated with individual components of a dietary pattern, suggesting that they may be representative of specific dietary patterns. Metabolites that were consistently associated with high diet scores in sample 1 and sample 2 significantly improved prediction of high diet quality beyond participant characteristics, except for aMED.

Our data showed a considerable overlap of metabolites of high-quality diets. Six significant metabolites were associated with <1 dietary pattern [glycerate, N-methylproline, stachydrine, threonate, pyridoxate, 3-(4-hydroxyphenyl)lactate] and replicated across 2 different samples, which may reflect similar food components within diet indices, such as fruit intake. Similar to our study, prior studies found that several metabolites overlapped across multiple healthy dietary patterns (12, 13). However, there were slight differences in the set of metabolites that represented each dietary pattern. This suggests that adherence to a generally high-quality diet can be identified by the overlapping metabolites. Metabolites unique to each dietary pattern in our study may reflect the specificity of the dietary patterns, if replicated in other populations.

Using a discovery and replication approach, we found 17 metabolites with the strongest evidence for a link with high-quality diets. CMPF was unique to AHEI and was associated with higher intake of ω-3 fatty acid consumption, a unique component of AHEI. CMPF was associated with fish and seafood in African Americans in ARIC (10), fish and ω-3 fatty acids in postmenopausal women (13), fish and seafood in participants of a colorectal adenoma case–control study (33), and seafood and plant protein in Finnish male smokers (12). CMPF is reported to be formed from fish, vegetable, and fruit intake. Results from our study and previous studies suggest that CMPF may be a specific biomarker of AHEI and ω-3 fatty acids (13, 34).

Deoxycarnitine, a lipid that was negatively associated with HEI and total fruit, total vegetables, and dairy consumption, was unique to this diet index. Similar to our study, deoxycarnitine was negatively correlated with grapefruit consumption in a case–control study of colorectal adenoma (33). Deoxycarnitine is a precursor of l-carnitine and is thought to be derived from consumption of animal foods (35). In a short-term feeding trial, participants who followed a Western dietary pattern that was high in refined grains and animal foods for 2 wk had an elevated level of deoxycarnitine (36). These results show that deoxycarnitine is likely an indicator of poor diet quality, and the association observed in our study and a previous study may be due to correlated intakes of foods (i.e., individuals with higher intake of fruits may have lower intakes of animal foods).

The DASH diet had the largest number of significant metabolites. Some of these associations were consistent with those reported in previous studies. Stachydrine, a significant metabolite associated with the DASH diet, HEI, and aMED, was identified as the second most influential metabolite that distinguished the DASH diet from the control diet in a metabolomics study of the original DASH trial (11). N-methylproline, N-acetylornithine, and catechol sulfate, which were significantly associated with the DASH diet in both samples in our study, were also identified as candidate biomarkers in the previous analysis of the DASH trial. γ-Tocopherol and threonate, which were associated with a higher DASH diet score in our study, were also among the metabolites that predicted highest compared with lowest adherence to the DASH diet in the Cancer Prevention Study–II Nutrition cohort composed of postmenopausal women (13). However, α-CEHC, a metabolite that was unique to the DASH diet in sample 2, was not associated with the DASH diet in these 2 previous studies. α-CEHC is an α-tocopherol metabolite that results from a series of ω- and β-oxidations, and it has been associated with long-term vitamin E supplementation (37). It has been suggested to be a biomarker of vitamin E status (38). In future studies, it may be useful to assess if α-CEHC is replicated as a strong candidate biomarker of the DASH diet.

In our study, only 2 metabolites (stachydrine and glycerate) were significantly associated with the Mediterranean diet, and the addition of these 2 metabolites did not improve prediction of those in the highest compared with lowest quintile of this diet index. Several prior studies in Britain and Spain examined metabolites associated with the Mediterranean dietary pattern, but these studies focused on a smaller set of specific metabolites (acylcarnitines, sphingolipids, phospholipids, and amines) or examined changes in plasma metabolites in response to Mediterranean diet interventions, making the comparison of the findings less straightforward (14, 39–41). Furthermore, the composition of a Mediterranean-style diet likely differs in Mediterranean countries compared with non-Mediterranean countries considering the differences in dietary and cultural habits. This may explain why only a few significant metabolites were found in sample 1 in our study. Nevertheless, our findings of 2 metabolites (stachydrine and 2-aminophenol sulfate) were consistent with those of a metabolomics study conducted in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study in the United States, which also used the aMED score (12). However, our findings differed from those of a study of postmenopausal women, which reported that many important metabolites that distinguished extreme quintiles of the Mediterranean diet were in the lipid category and were all associated with fish consumption (13). Given that this previous study was composed exclusively of women, dietary patterns may be different than those observed in our study population. The limited number of metabolites associated with the aMED score suggests the need for further studies in the US study populations.

A few additional metabolites deserve mention. DHA (22:6n3) was associated with AHEI in both samples, and it was unique to AHEI. This metabolite was reported to be associated with a higher AHEI score and ω-3 fatty acid intake and lower trans fat consumption in a prior study, consistent with our findings (13). DHA is derived from fish oil; thus, these associations are biologically plausible. Levels of 4-allyphenol sulfate, also known as chavicol, were higher in individuals with higher diet scores in sample 2, and this metabolite was positively associated with the fruit component within all dietary patterns. 4-Allyphenol sulfate is a food component that has been detected in herbs, pineapples, and gingers (34, 42). In an untargeted metabolomic analysis of the DASH–Sodium trial, serum levels of 4-allyphenol sulfate were higher when participants switched from a high-sodium intervention to the medium-sodium level and from medium to low sodium (42). It was hypothesized that seasoning added to the lower sodium version of the DASH diet may have resulted in a higher level of 4-allyphenol sulfate (42). Given the consistent associations with fruit intake in our study, it may be a key marker of a high diet quality or additional seasoning, which requires further investigation.

Our study has a number of strengths, including a relatively large sample size and the use of predefined diet indices to define healthy diets that are easily reproducible in other populations. Our findings also have broader generalizability relative to previous studies that were conducted in special populations, such as male smokers and postmenopausal women (12, 13). Furthermore, to our knowledge, this is the first study to examine untargeted metabolomics data in relation to healthy dietary patterns in 2 different samples. However, several limitations should be taken into account in interpreting the results. Although the FFQ was modified from a validated questionnaire to include foods frequently consumed in the ARIC study population, it may not have covered all foods (21). Metabolomic profiling was conducted using serum specimens collected >20 y ago. Degradation of metabolites may have occurred, although it would be expected to be nondifferential by the degree of adherence to healthy dietary patterns. Due to differences in the timing of metabolomic profiling, fewer metabolites were quantified in sample 1 compared with sample 2. However, we were able to verify that many significant metabolites were replicated in sample 2. Last, despite efforts to adjust for confounders rigorously, there is still a possibility of residual confounding as with any observational study.

In conclusion, we detected a set of metabolites associated with healthy dietary patterns in 2 subsamples of African Americans and European Americans. This set of significant metabolites for HEI, AHEI, and DASH improved the prediction of adherence to the respective diet index beyond participant characteristics. Metabolites that were common across multiple healthy dietary patterns are representative of high diet quality and could indicate important metabolic pathways through which healthy diets impact chronic disease risk. Future studies should examine if these metabolites are replicated in different study populations. Furthermore, data from clinical trials that isolate differences in diet from other confounders such as genetic and environmental factors are warranted to have greater certainty that these candidate biomarkers represent high-quality diets.

Notes

The Atherosclerosis Risk in Communities Study was supported by the National Heart, Lung, and Blood Institute, NIH, Department of Health and Human Services (HHSN268201700001I, HHSN268201700002I, HHSN268201700003I, HHSN268201700004I, and HHSN268201700005I). CMR was supported by a Mentored Research Scientist Development Award from the National Institute of Diabetes and Digestive and Kidney Diseases (K01 DK107782) and a grant from the National Heart, Lung, and Blood Institute (R21 HL143089). The funding agencies had no role in study design, data collection, analysis, drafting of the manuscript, and the decision to submit this manuscript for publication.

Author disclosures: The authors report no conflicts of interest.

Supplemental Tables 1 and 2 and Supplemental Figure 1 are available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/jn/.

Abbreviations used: AHEI, Alternative Healthy Eating Index; aMED, alternate Mediterranean diet; ARIC, Atherosclerosis Risk in Communities Study; CEHC, carboxyethyl hydroxychroman; CMPF, 3-carboxy-4-methyl-5-propyl-2-furanpropanoate; DASH, Dietary Approaches to Stop Hypertension Trial; eGFR, estimated glomerular filtration rate; HEI, Healthy Eating Index.