A metabolome and microbiome wide association study of healthy eating index points to the mechanisms linking dietary pattern and metabolic status

Abstract

Background:

Healthy eating index (HEI), a measure of diet-quality, associates with metabolic health outcomes, however the molecular basis is unclear. We conducted a multi-omic study to examine whether HEI associates with the circulatory and gut metabolome and investigated the gut microbiome-HEI interaction on circulating and gut metabolites.

Methods:

Through a cross-sectional study, we evaluated diet-quality in heathy individuals (the ABO Glycoproteomics in Platelets and Endothelial Cells [ABO] Study, N=73), metabolites (measured at Metabolon Inc.) in plasma (n=800) and gut (n=767) and the gut-microbiome at enterotype and microbial taxa (n=296) levels. Pathway analysis was conducted using Metaboanalyst 4.0. We performed multi-variable linear regression to explore both the HEI-metabolites and HEI-microbiome associations and how metabolites were affected by the HEI-microbiome interaction. In the Fish oils and Adipose Inflammation Reduction (FAIR) Study (N=25), analyses on HEI and plasma metabolites were replicated. Estimates of findings from both studies were pooled in random-effects meta-analysis.

Results:

The HEI-2015 was associated with 74 plasma and 73 gut metabolites (mostly lipids) and with 47 metabolites in the meta-analysis of the ABO and FAIR Studies. Compared to Enterotype-1 participants, those with Enterotype-2 had higher diet-quality (p=0.01). We also identified 9 microbial genera associated with HEI, and 35 plasma and 40 gut metabolites linked to the HEI-gut microbiome interaction. Pathways involved in the metabolism of polar lipids, amino acids and caffeine strongly associated with diet-quality. However, the HEI-microbiome interaction not only influenced the pathways involved in the metabolism of branch-chain amino acids, it also affected upstream pathways including nucleotide metabolism and amino acids biosynthesis.

Conclusions:

Our multi-omic analysis demonstrated that changes in metabolism, measured by either circulatory/gut metabolites or metabolic pathways, are influenced by not only diet-quality but also gut microbiome alterations shaped by the quality of diet consumed. Future work is needed to explore the causality in the interplay between HEI and gut-microbiome composition in metabolism.

Introduction

Diet quality represents an assessment of overall diet, and is influenced by the quality of individual food and nutrients [1, 2]. Healthy eating index (HEI), a measure of diet quality (https://www.fns.usda.gov/resource/healthy-eating-index-hei), has been used in several studies. Adherence to higher diet quality score is associated with a lower risk of chronic diseases and mortality [3, 4]. Exploring mechanisms underlying diet pattern-disease relationships may identify candidate nutritional biomarkers with potential clinical relevance. However, biomarkers and metabolic pathways influenced by diet quality and involved in the development of several diseases are far from being explored.

The most recent approach used in mechanistic studies is high-throughput techniques defined as - omics methods. Metabolomics studies have identified both 1) nutritional biomarkers which include amino acids, a range of lipids, acylcarnitines, and a range of phenolic compounds, and 2) metabolic pathways such as lysolipid and food and plant xenobiotic pathways, that were associated with either higher or lower diet quality [5-7]. Furthermore, the microbiome composition, an environmental factor which plays a major role in health status, is influenced by dietary intake and host metabolism [8]. The diversity of the gut microbiome has been related to both the circulatory and fecal metabolome [9]. In addition, the relative contributions of the interaction between genetic variants and dietary pattern to molding metabolic and microbiome environment are of particular interest.

Although a number of studies have investigated diet quality relationships using single -omics approach, to our knowledge, no study investigated the relationship between diet quality and gut metabolome and no study has applied multi-omic methods to identify mechanisms driving key roles in the relationship between eating pattern and metabolism. Also, the interplay between gut microbiome and dietary quality on metabolites has not been previously evaluated. When these techniques are simultaneously assessed in an integrative network they could contribute to unveiling the interrelationships of different biomolecules involved in several pathways [10, 11]. Thus, to address this critical gap, we used integrative approaches in the current study to identify circulatory and gut metabolome and gut microbiome associations with dietary pattern as defined by HEI in healthy volunteers. These HEI-related metabolites were evaluated 1) to determine metabolic pathways related to the HEI score and 2) to investigate whether they are influenced by the interaction of gut enterotype and diet quality. The study was designed to be hypothesis generating, allowing for identification of novel relationships between HEI and multi-omic measurements of the metabolome and microbiome. Findings of this study might shed novel insight into the molecular and metabolic basis for the link between food patterns and metabolic health outcomes.

Methods

Study design and population-main study

In a cross-sectional study design, we analyzed data from the ABO Glycoproteomics in Platelets and Endothelial Cells (ABO) Study in which 150 healthy, lean and overweight non-pregnant or non-lactating women and men aged 18–50 y were recruited from the University of Pennsylvania between 2012 and 2014, as previously described [8, 12]. Subjects were excluded if they 1) had known illness or history of organ transplant 2) current use of prescription medication (excluding oral contraceptives) and 3) used tobacco or had history of using tobacco products within the last 30 days. Eligible participants were asked to stop using supplements and over-the-counter medications for at least two weeks, and to fast at least 12 hours prior to their scheduled visit. In addition, they were provided a stool collection kit (Commode Specimen Collection System, Fisher Scientific, Pittsburgh, PA, United States) to collect a stool sample within the 24 h prior to the study visit. Fasting blood samples (N=150) and stool samples (N=136) were collected from the participants at the study visit (supplementary Figure 1). Blood samples were processed into plasma and blood cells and stored at −80 °C. Stool samples were stored at 4°C following collection, and aliquots were obtained and frozen at −80°C within 36 h of sample collection for subsequent analysis. Dietary assessment was conducted in all participants, and metabolomics profiling was completed in a subset of participants as described below (supplementary Figure 1).

Study population-comparison study

In order to explore generalizability and reproducibility of our findings[13], we examined associations of HEI with plasma metabolites in a comparison sample, the Fish oils and Adipose Inflammation Reduction (FAIR) study, which used identical approaches to collect dietary and metabolomics data compared to the ABO study. This allowed for independent validation to improve the power of our study to detect biologically meaningful relationships between diet quality and metabolites. In the FAIR study 29 non-pregnant/non-lactating women and men aged 18-50 y, with BMI >=30 kg/m² (obese) were recruited from the Philadelphia region. Subjects were excluded if they 1) had known illness or history of organ transplant 2) current use of prescription medication (excluding oral contraceptives) and 3) used tobacco or had history of using tobacco products within the last 30 days. The FAIR study was a randomized controlled trial with fish oil intervention registered as NCT02010359 number and conducted at the University of Pennsylvania between 2014-2017. For the current study, only baseline dietary and metabolomics data of 25 participants were included in the analysis (supplementary Figure 1).

A written consent form was read and signed by each participant in both studies (ABO and FAIR) and the protocol of these studies was reviewed and approved by the Institutional Review Boards of the University of Pennsylvania and Vanderbilt University. All assessments were identical for the ABO and FAIR study.

Dietary intake assessment

Habitual dietary intake over the year preceding the study visit was assessed in both the ABO and the FAIR Studies using food frequency questionnaires (FFQ) from the National Cancer Institute’s Diet History Questionnaire (DHQ II) which was composed of 134 food items [14]. Data derived from the FFQ was converted into Nutrient intake values of 191 dietary variables using Diet*Calc version 1.5.1.

HEI, a measure of diet quality (food adequacy and food moderation), was derived from DHQ II Diet*Calc output (https://epi.grants.cancer.gov/hei) using the SAS University Edition (https://www.sas.com/en_us/software/university-edition.html). The latest version of HEI (HEI-2015) included 13 components. The score for seven of the 13 components ranges from 0 to 10 points, while for the other components, it ranges from 0 to 5. The overall HEI score, which was used in the current study, ranges from 0 to 100 points. Among these components, higher intakes of refined grain, sodium, added sugars and saturated fats (items of food moderation) are assigned a lower score, whereas higher intakes of total fruits, whole fruits, greens and beans, total vegetables, total protein foods, seafood and plant proteins, whole grains, dairy and polyunsaturated and monounsaturated fatty acids [PUFA and MUFA] (items of food adequacy) are assigned a higher score (https://epi.grants.cancer.gov/hei/developing.html#f1b).

Metabolomics assessment

Metabolomics profiling of plasma and stool samples of a subset of ABO study participants (75 plasma samples and 75 stool samples from the same individuals, analyzed in a single batch) and 25 plasma samples in the FAIR study were conducted using multiple mass spectrometry techniques at Metabolon (Metabolon Inc., Morrisville, NC, United States), including multiple QC steps and comparison to Metabolon’s large well-validated metabolomics compound library[15,16]. In the ABO Study 812 and 767 metabolites were identified and quantified in plasma and stool samples, respectively, while 787 plasma metabolites identified in the FAIR study were included in our analysis. Of these metabolites, 12 in plasma and 3 in stool in the ABO study and 5 in the FAIR study had the same value for all the participants and therefore they were excluded from the analysis. There were 640 plasma metabolites which were identified in both the ABO and FAIR studies. Metabolite values were standardized by Metabolon by setting the median across all samples equal to 1 (to account for batch effects from multi-day runs) and natural log-transformed (to account for a deviation from normality). Metabolite levels below the limit of detection were imputed with the lowest measured level for that particular metabolite.

Microbiome Profiling

The protocol for the microbiome measurements was explained in detail earlier [8]. Briefly, this process included three steps: sample processing, DNA extraction and DNA sequencing. Sample processing: Participants were provided a stool collection kit (Commode Specimen Collection System, Fisher Scientific, Pittsburgh, PA, United States) to collect a stool sample within 24 h prior to their study visit. Aliquots were made in 36 h from the time of sample collection and stored at −80°C. DNA extraction: DNA isolation was performed with the use of the PSP Spin Stool DNA Plus Kit (Stratec, Germany). Barcoded primers (Eurofins Genomics, Louisville, KY, United States) were used to amplify the 16S rRNA gene section. DNA sequencing: After being cleaned (MinElute PCR Purification kit, Qiagen, Germantown, MD, United States), DNA libraries were pooled and subsequently sequenced on the MiSeq platform, 300 bp paired-end reads, at an average depth of 158,000 reads/sample (Illumina Inc., San Diego, CA, United States) in two batches; at the University of Pennsylvania Next-Generation Sequencing Center (UPenn NGSC, N = 107) and the Vanderbilt University Technologies for Advanced Genomics (VANTAGE) Core (N = 29).

Statistical analysis

In the ABO and FAIR Studies, associations between the HEI (total score, ranging from 0-100) and individual plasma /gut metabolites were examined using multivariable linear regression adjusted for sex, age, and BMI. We tested the distribution of the HEI score by the Wilk- Shapiro test and observed that this variable was normally distributed (the Wilk- Shapiro test p-value = 0.1938). Both uncorrected and False Discovery rate (FDR)-corrected p-values were evaluated and all associations with a nominal p<0.05 are reported in these analyses. For plasma metabolites, the results from the ABO and FAIR studies were meta-analyzed using multivariable-adjusted effect estimate (β) with an accompanying measure of uncertainty (SE) derived from our analysis in the main and replication studies. In both fixed-effects and random-effects models, two popular statistical models for meta-analysis [17], summary estimates with the calculation of corresponding 95% CIs of the individual studies were obtained. Statistical heterogeneity among the ABO and FAIR studies was explored by the Cochran Q test and I² statistics, with the values of 25%, 50% and 75% regarded as low, moderate and high heterogeneity, respectively (http://handbook.cochrane.org). Heterogeneity was considered statistically significant at P ≤ 0.05. We used the random-effects values (since the random-effects models incorporate both within- and between-study variations) in a forest plot to evaluate the associations across significant metabolites (p< 0.05).

As described in our previous work [8], the Partitioning Around Medoids (PAM) method was used to identify two gut enterotypes in the ABO study, in order to group individuals based on broad characterization of their gut microbiome composition. The proportion of Ruminococcaceae was significantly higher in Enterotype 2 compared to Enterotype 1. Age and race did not differ between these two enterotypes, while Enterotype 2 had higher number of women and lower BMI [8]. We ran a logistic regression model (adjusting for age, sex and BMI) to investigate the association of HEI and gut enterotype. To explore whether gut microbe composition is related to the diet quality index, we did multivariable regression analysis (adjusting for age, sex and BMI) which assessed the association of individual gut microbial taxa (n=296) separately with HEI. Significant associations were plotted to screen real relationships.

In a second analysis only in the ABO Study, we used a linear regression approach to assess the interaction of gut enterotype (the two enterotypes identified by the PAM method) and HEI on plasma/gut metabolites. These models were fitted by regressing the interaction of gut enterotype (used as a dummy variable) and the HEI score on metabolites. More precisely, the regression models were conducted separately for each metabolite, and included main effects (enterotype, HEI score, and covariates age, sex, and BMI) and the interaction term of enterotype (2 vs. 1) * HEI score (0-100). Significant associations from the interaction analyses are shown in Supplementary Table 2 and 3. Associations with a nominal p<0.05 are reported.

We used Metaboanalyst 4.0 (https://www.metaboanalyst.ca) to perform pathway and enrichment analyses. Significant metabolites from our analyses were mapped to known human metabolic pathways in the KEGG and SMPDB libraries, and compared with all known compounds in those pathways, to identify top enriched pathway selected by a threshold p< 0.05. The data analysis was implemented in statistical framework R 3.1.0 (www.rproject.org) [18].

Results

Characteristics of the ABO and FAIR populations are provided in Table 1. The HEI score ranged between 38.19 and 87.64 in the ABO Study and between 47.76 and 82.24 in the FAIR Study. Median (interquartile range [IQR]) age of participants in the ABO Study was 28 (24-34) y and 60% of them were white. In the FAIR Study, participants had mean (SD) age of 37 (9), and 60% were white. Most of the participants in both studies were female (58% in the ABO and 68% in the FAIR Studies). Compared to participants the FAIR study, subjects ABO had relatively higher mean HEI (67.5±10.9 in the ABO and 65.7±7.9 in the FAIR Studies). Consistent with the design of the studies, BMI was higher in the FAIR study, which recruited only obese individuals, compared with the ABO study which recruited healthy individuals irrespective of body mass.

Table 1-.

Characteristics^a of the main and comparison study populations.

Characteristic	ABO Study N=73	FAIR Study N=25
Age (y)	28 (24-34)	36.64 (8.77)
Sex, %
Female	42 (57.53)	17 (68)
Body mass index (kg/m2)	23.58 (20.9-29.90)	33.80 (32.70-38.30)
Healthy eating index	67.51 (10.85)	65.67 (7.94)
Race b , %
White	44 (60.27)	15 (60)
Black	19 (26.03)	8 (32)
Asian	7 (9.59)	n/a
Other	3 (4.11)	2 (8)

^aMean (SD)/Median (IQRs) or n (%); Medians (IQRs) were reported for variables which did not have a normal distribution.

^bRace was not available in the FAIR Study

Plasma and gut metabolites were related to diet quality in healthy individuals

In the ABO Study plasma dataset, 74 metabolites were significantly associated with the HEI score (Table 2) out of 800 metabolites that were tested. Among these metabolites, top associations belonged to sphingomyelins (different analogues based on different fatty acids on their side chains), amino acids derivatives and fatty acids. Also, the metabolites of urea/ornithine cycle showed positive associations with the HEI score. In the gut dataset, 73 metabolites were significantly associated with the HEI score (Table 3) out of 767 metabolites that were examined. Around 20% of these metabolites were diacylglycerols which were positively associated with the dietary score. Carnitine and fatty acid derivatives and amino acids were inversely related to the score.

Table 2-.

Multivariable linear regression analyses for the association^a of the Healthy Eating Index with plasma metabolites^b

Metabolite	β	SE	R2c	P-valued
glutamate, gamma-methyl ester	−0.007	0.002	0.209	0.0005
sphingomyelin (d18:1/22:2, d18:2/22:1, d16:1/24:2)	−0.010	0.003	0.337	0.0006
1-lignoceroyl-GPC (24:0)	0.020	0.006	0.283	0.0006
sphingomyelin (d18:2/23:1)	−0.012	0.003	0.275	0.0007
sphingomyelin (d18:0/18:0, d19:0/17:0)	−0.019	0.006	0.212	0.0026
3-hydroxypyridine sulfate	0.041	0.013	0.293	0.0035
stearoyl sphingomyelin (d18:1/18:0)	−0.008	0.003	0.203	0.0039
N-delta-acetylornithine	0.025	0.008	0.188	0.0045
oxalate (ethanedioate)	0.009	0.003	0.331	0.0050
glycosyl-N-stearoyl-sphingosine (d18:1/18:0)	−0.010	0.004	0.227	0.0056
palmitoylcarnitine (C16)	−0.008	0.003	0.277	0.0062
margaroylcarnitine*	−0.012	0.004	0.236	0.0080
sphingomyelin (d18:1/19:0, d19:1/18:0)	−0.009	0.003	0.213	0.0081
trigonelline (N'-methylnicotinate)	0.037	0.014	0.136	0.0103
tryptophan betaine	0.043	0.017	0.140	0.0105
N-acetylglutamine	−0.014	0.005	0.127	0.0108
sphingomyelin (d18:2/23:0, d18:1/23:1, d17:1/24:1)	−0.007	0.003	0.186	0.0109
delta-tocopherol	−0.021	0.008	0.125	0.0113
15-methylpalmitate	−0.014	0.005	0.124	0.0115
methyl glucopyranoside (alpha + beta)	0.030	0.011	0.253	0.0115
docosapentaenoate (n6 DPA; 22:5n6)	−0.010	0.004	0.270	0.0125
sphingomyelin (d18:1/17:0, d17:1/18:0, d19:1/16:0)	−0.006	0.002	0.139	0.0127
lactosyl-N-palmitoyl-sphingosine (d18:1/16:0)	−0.006	0.002	0.147	0.0132
N-methylproline	0.033	0.013	0.095	0.0134
Argininate	0.015	0.006	0.245	0.0140
N-palmitoylglycine	−0.009	0.004	0.131	0.0149
1-palmitoyl-2-linoleoyl-GPI (16:0/18:2)	−0.025	0.010	0.109	0.0157
1-palmitoyl-2-arachidonoyl-GPI (16:0/20:4)	−0.010	0.004	0.133	0.0161
tricosanoyl sphingomyelin (d18:1/23:0)	−0.007	0.003	0.132	0.0165
sphingomyelin (d18:2/24:2)	−0.007	0.003	0.341	0.0166
4-ethylphenylsulfate	0.041	0.017	0.182	0.0179
S-methylcysteine sulfoxide	0.023	0.009	0.185	0.0199
glycosyl ceramide (d18:2/24:1, d18:1/24:2)	−0.008	0.003	0.171	0.0220
ceramide (d18:1/17:0, d17:1/18:0)	−0.011	0.005	0.220	0.0232
sphingomyelin (d18:1/18:1, d18:2/18:0)	−0.005	0.002	0.348	0.0248
1-oleoyl-2-linoleoyl-GPC (18:1/18:2)	0.008	0.004	0.132	0.0249
1-(1-enyl-stearoyl)-2-arachidonoyl-GPE (P-18:0/20:4)	−0.010	0.004	0.090	0.0254
sphingomyelin (d18:1/15:0, d16:1/17:0)	−0.005	0.002	0.085	0.0255
sphingomyelin (d18:2/21:0, d16:2/23:0)	−0.007	0.003	0.351	0.0259
10-heptadecenoate (17:1n7)	−0.013	0.006	0.171	0.0263
1-(1-enyl-palmitoyl)-2-palmitoyl-GPC (P-16:0/16:0)	−0.006	0.003	0.160	0.0266
Phenylpyruvate	0.008	0.004	0.239	0.0268
Glycerate	0.006	0.003	0.274	0.0269
4-vinylphenol sulfate	0.030	0.013	0.201	0.0273
Tartarate	0.014	0.006	0.225	0.0280
sphingomyelin (d18:1/21:0, d17:1/22:0, d16:1/23:0)	−0.006	0.003	0.108	0.0281
2-oxindole-3-acetate	0.019	0.008	0.254	0.0293
1-stearoyl-2-arachidonoyl-GPI (18:0/20:4)	−0.005	0.002	0.131	0.0298
Quinate	0.047	0.021	0.124	0.0298
1,2-dipalmitoyl-GPC (16:0/16:0)	−0.003	0.001	0.232	0.0298
octadecanedioate	−0.012	0.005	0.101	0.0304
N-stearoyl-sphingosine (d18:1/18:0)	−0.010	0.004	0.287	0.0317
sphingomyelin (d18:2/24:1, d18:1/24:2)	−0.005	0.002	0.168	0.0320
catechol sulfate	0.021	0.010	0.105	0.0353
Thyroxine	−0.005	0.002	0.182	0.0364
L-urobilin	−0.030	0.014	0.081	0.0372
4-acetamidobutanoate	0.015	0.007	0.154	0.0377
sphingomyelin (d18:1/20:2, d18:2/20:1, d16:1/22:2)	−0.009	0.004	0.226	0.0379
O-acetylhomoserine	−0.015	0.007	0.237	0.0379
N-acetyl-cadaverine	−0.021	0.010	0.250	0.0389
7-alpha-hydroxy-3-oxo-4-cholestenoate (7-Hoca)	0.006	0.003	0.272	0.0395
1-palmitoyl-2-arachidonoyl-GPC (16:0/20:4n6)	−0.004	0.002	0.080	0.0404
4-vinylguaiacol sulfate	0.023	0.011	0.085	0.0414
2-oxoarginine	0.014	0.007	0.243	0.0424
isobutyrylcarnitine (C4)	0.012	0.006	0.155	0.0425
O-sulfo-L-tyrosine	−0.005	0.002	0.302	0.0443
palmitate (16:0)	−0.007	0.004	0.189	0.0448
stearoyl ethanolamide	−0.006	0.003	0.256	0.0449
tartronate (hydroxymalonate)	0.008	0.004	0.381	0.0454
N-methyltaurine	0.026	0.013	0.086	0.0455
13-methylmyristate	−0.012	0.006	0.158	0.0468
1-methylnicotinamide	−0.010	0.005	0.155	0.0469
N-stearoyltaurine	−0.010	0.005	0.129	0.0487
phenol sulfate	0.014	0.007	0.116	0.0488

^aAdjusted for age (years), sex, body mass index at blood draw (in kg/m²)

^bMetabolites were log-transformed. The reported results belong to 74 significant associations out of 800 metabolites that were tested.

^cCompared to the models that did not include HEI, in models with HEI, the R² increased for all 74 significant metabolites.

^dAll associations with a nominal p<0.05 are shown. No metabolite was significant after correction for multiple comparisons (using the False Discovery Rate method).

Abbreviations: SE=Standard Error

Table 3-.

Multivariable linear regression analyses for the association^a of the Healthy Eating Index with gut metabolites^b

Metabolite	β	SE	R2	P-valuec
diglycerol	−0.048	0.012	0.197	0.0001
oleoyl-oleoyl-glycerol (18:1/18:1)	0.087	0.022	0.256	0.0002
2-hydroxystearate	−0.051	0.014	0.175	0.0003
mead acid (20:3n9)	−0.078	0.021	0.216	0.0004
behenoylcarnitine (C22)	−0.043	0.013	0.176	0.0014
oleoyl-linoleoyl-glycerol (18:1/18:2)	0.060	0.018	0.212	0.0016
dihomo-linoleate (20:2n6)	−0.029	0.009	0.180	0.0021
arachidoylcarnitine (C20)	−0.045	0.015	0.190	0.0029
N-palmitoylglycine	−0.038	0.012	0.224	0.0031
carnosine	−0.060	0.020	0.198	0.0032
2-hydroxydecanoate	0.019	0.006	0.164	0.0037
arachidonate (20:4n6)	−0.029	0.010	0.163	0.0048
linoleoyl-linolenoyl-glycerol (18:2/18:3)	0.055	0.019	0.213	0.0050
dimethylmalonic acid	−0.016	0.006	0.127	0.0051
linoleoyl-linoleoyl-glycerol (18:2/18:2)	0.050	0.017	0.215	0.0053
13-methylmyristate	−0.027	0.009	0.142	0.0055
secoisolariciresinol	0.037	0.013	0.115	0.0070
histamine	−0.070	0.025	0.154	0.0073
docosapentaenoate (n6 DPA; 22:5n6)	−0.051	0.019	0.120	0.0082
2-hydroxypalmitate	−0.047	0.018	0.117	0.0090
palmitoyl-linoleoyl-glycerol (16:0/18:2)	0.043	0.016	0.228	0.0096
palmitoyl-oleoyl-glycerol (16:0/18:1)	0.036	0.014	0.174	0.0104
nonadecanoate (19:0)	−0.025	0.009	0.095	0.0106
1-oleoylglycerol (18:1)	0.041	0.016	0.121	0.0125
cinnamate	0.023	0.009	0.143	0.0148
2-hydroxy-3-methylvalerate	−0.032	0.013	0.143	0.0151
4-hydroxycinnamate	0.039	0.016	0.102	0.0155
docosahexaenoate (DHA; 22:6n3)	−0.031	0.013	0.096	0.0166
margarate (17:0)	−0.029	0.012	0.096	0.0171
diacylglycerol (16:1/18:2, 16:0/18:3)	0.032	0.013	0.243	0.0172
eicosenoylcarnitine (C20:1)	−0.024	0.010	0.244	0.0181
3-hydroxyphenylacetate	0.035	0.015	0.282	0.0185
thymidine 5'-monophosphate	−0.016	0.006	0.186	0.0187
oleate/vaccenate (18:1)	0.016	0.007	0.160	0.0195
alpha-hydroxyisovalerate	−0.028	0.012	0.145	0.0204
delta-tocopherol	−0.016	0.007	0.134	0.0204
2-hydroxymyristate	−0.046	0.020	0.109	0.0211
2-piperidinone	−0.031	0.013	0.182	0.0211
glutamyl-meso-diaminopimelate	−0.032	0.014	0.108	0.0220
1,11-undecanedicarboxylate	0.016	0.007	0.084	0.0226
glycerol	0.025	0.011	0.105	0.0271
glycine	−0.017	0.007	0.116	0.0274
10-hydroxystearate	0.025	0.011	0.145	0.0277
oleoyl-linolenoyl-glycerol (18:1/18:3)	0.048	0.021	0.143	0.0278
glycylisoleucine	−0.014	0.006	0.108	0.0278
1-linoleoyl-2-linolenoyl-GPC (18:2/18:3)	0.030	0.013	0.072	0.0279
docosapentaenoate (n3 DPA; 22:5n3)	−0.033	0.015	0.128	0.0281
epiandrosterone suliate	0.038	0.017	0.072	0.0305
palmitoyl-linolenoyl-glycerol (16:0/18:3)	0.037	0.017	0.237	0.0306
N-butyroyl-sphingosine (d18:1/4:0)	−0.036	0.016	0.187	0.0307
15-methylpalmitate	−0.029	0.013	0.077	0.0309
gamma-glutamyl-epsilon-lysine	−0.029	0.013	0.172	0.0315
alanine	−0.011	0.005	0.107	0.0329
eicosapentaenoate (EPA; 20:5n3)	−0.028	0.013	0.100	0.0333
methionine	−0.012	0.005	0.083	0.0353
3-hydroxyphenylacetate suliate	0.022	0.010	0.126	0.0359
alpha-hydroxyisocaproate	−0.024	0.011	0.113	0.0362
17-methylstearate	−0.021	0.010	0.070	0.0367
cysteine s-sulfate	−0.020	0.009	0.189	0.0369
alpha-CEHC suliate	0.027	0.013	0.089	0.0385
N-acetyltryptophan	−0.011	0.005	0.170	0.0391
N-stearoyl-sphingosine (d18:1/18:0)	−0.027	0.013	0.168	0.0392
phenethylamine	0.025	0.012	0.135	0.0393
ceramide (d18:2/24:1, d18:1/24:2)	−0.027	0.013	0.103	0.0407
palmitoleoyl-linoleoyl-glycerol (16:1/18:2)	0.025	0.012	0.145	0.0414
linoleate (18:2n6)	0.020	0.010	0.143	0.0419
C-glycosyltryptophan	−0.019	0.009	0.085	0.0424
ceramide (d18:1/20:0, d16:1/22:0, d20:1/18:0)	−0.025	0.012	0.115	0.0427
pentadecanoate (15:0)	−0.024	0.012	0.065	0.0431
tryptamine	0.041	0.020	0.070	0.0435
N-methylalanine	−0.017	0.008	0.100	0.0446
xanthine	−0.016	0.008	0.210	0.0469
nicotinate ribonucleoside	−0.023	0.011	0.061	0.0473

^aAdjusted for age (years), sex, body mass index at blood draw (in kg/m²)

^bMetabolites were log-transformed. The reported results belong to 73 significant associations out of 767 metabolites that were tested.

^cAll associations with a nominal p<0.05 are shown. No metabolite was significant after correction for multiple comparisons (using the False Discovery Rate method).

Abbreviations: SE=Standard Error

Among the significant plasma and gut metabolites, six overlapped between the two tissues. These metabolites included delta-tocopherol, 15-methylpalmitate, docosapentaenoate (n6 DPA; 22:5n6), N-palmitoylglycine, N-stearoyl-sphingosine (d18:1/18:0), and 13-methylmyristate. These metabolites had the same direction of association, almost identical magnitude and similar R-squared between the two datasets.

Plasma metabolites also associate with HEI in a high-risk population

In order to assess the generalizability of the HEI-metabolite associations, we analyzed associations between plasma metabolites and HEI in a comparison dataset (FAIR Study). In contrast to the ABO study, which recruited healthy lean or overweight individuals, the FAIR study recruited obese individuals. These individuals did not have any overt disease but would be considered high-risk for future cardiometabolic disease. There were 29 metabolites significantly associated with HEI (supplementary Table 1) out of 787 metabolites that were examined. The relatively smaller number of associations compared with the ABO Study is likely due to the reduced power given the smaller sample size in the FAIR Study. Three metabolites that replicated between the ABO and FAIR studies were sphingomyelin (d18:1/17:0, d17:1/18:0, d19:1/16:0) [with B=−0.006, P=0.01 in ABO and B=−0.004, P=0.008 in FAIR], 1,2-dipalmitoyl-GPC (16:0/16:0) [with B=−0.003, P=0.03 in ABO and B=−0.0001, P=0.03 in FAIR], and 4-acetamidobutanoate [with B=0.015, P=0.04 in ABO and B=−0.01, P=0.03 in FAIR]. In order to increase power for detection of novel metabolite-HEI associations, we conducted a meta-analysis of the associations between HEI and 640 common plasma metabolites in the ABO and FAIR studies. Figure 1 shows the pooled estimates of random effects meta-analyses for these two studies. Based on random effects p-values, 47 metabolites out of 640 metabolites showed significant associations with the HEI score. As shown in Figure 1, of these metabolites, 12 including phenylpyruvate, oxalate, isobutyrylcarnitine (C4), ergothioneine, N-delta-acetylornithine, N-methylproline, methyl glucopyranoside (alpha + beta), trigonelline (N'-methylnicotinate), 4-ethylphenylsulfate, stachydrine, thymol sulfate, and quinate were positively associated with HEI. Metabolites that were inversely associated with HEI were mostly lipids. These may represent metabolites that associate with HEI across the obesity spectrum from lean to obese.

Figure 1-.

Pooled estimates of the association between healthy eating index (HEI) score and metabolites from ABO and FAIR studies. Overall estimates obtained from forest plots and random-effects meta-analysis of ABO and FAIR studies evaluating the associations of HEI with plasma metabolites. The estimate on the X-axis represent multivariable-adjusted effect estimate (β) and were derived from linear regression models adjusting for age, sex and BMI in each study. The reported results belong to 47 significant associations out of 640 metabolites (identified in both the ABO and FAIR Studies) that were tested. Closed squares and horizontal bars represent the overall estimates and 95% CIs. I^2 represents statistical heterogeneity among the ABO and FAIR Studies, with the values of 25%, 50% and 75% regarded as low, moderate and high heterogeneity, respectively.

Four metabolic pathways were identified to be influenced by diet quality

Interestingly, pathway topology analysis showed that three pathways (glycerophospholipid metabolism, glycine, serine and threonine metabolism, and caffeine metabolism) for plasma metabolites and two pathways (histidine metabolism, and caffeine metabolism) for gut metabolites were significantly (raw P-value < 0.05; pathway impact values ≥ 0.02) associated with the HEI score (Figure 2). Furthermore, caffeine metabolism overlapped between plasma and gut metabolites in relation to the diet score. This association appeared to be independent of actual caffeine intake, as caffeine was not included in the HEI and caffeine intake was not correlated to HEI (r= −0.055, p = 0.51) in our data).

Figure 2-.

Metabolomic pathway analysis, generated by MetaboAnalyst software, for plasma metabolites (A) and gut metabolites (B). Circles indicate the matched pathways. The color of each circle changes according to the p-values (more significant changes of metabolites in the pathway are shown by darker colors), while the pathway impact score is depicted by the size of the circle. The most impacted pathways with the highest statistical significance scores are annotated for clarity. In plasma the raw P-values were 0.006, 0.01, and 0.02 for glycerophospholipid metabolism, glycine, serine and threonine metabolism, and caffeine metabolism, respectively. In stool, the raw P-values for the two significant pathways, histidine metabolism, and caffeine metabolism, were 0.03 and 0.04, respectively.

Gut microbiome is related to diet quality

We hypothesized the adherence to a healthy diet would associate with differences in gut microbiome composition. We previously identified two enterotypes within our dataset [8]. There was a significant difference in HEI by gut enterotype before (mean±SD= 63.98±11.69 in enterotype 1 compared to 69.01±10.45 in enterotype 2, p=0.01 (result of independent t-test)) and after adjusting for age, sex and BMI (Odds ratio= 0.036, p=0.046 (result of multivariable logistic regression)). We also analyzed associations between HEI and 296 individual gut microbial taxa, adjusting for age, sex and BMI. After removing associations driven by taxa that were only present in a small number of individuals, we identified significant associations between HEI and 9 microbial genera shown in Table 4. However, these associations did not reach statistical significance after FDR adjustment for multiple testing. As Table 4 shows Higher HEI was associated with reduced abundance of Parabacteroides, Alistipes, certain Bacteroides, Incertae Sedis and Firmicutes taxa and increased abu vo9bndance of Prevotella 7, certain Bacteroides and Blautia. These data suggest that diet quality may affect gut microbiome composition, but these results need to be replicated in independent samples.

Table 4-.

Association ^a of healthy eating index with gut microbial taxa ^b

Bacterial taxon phylum	Bacterial taxon genus	β	SE	R2	P-value c
Bacteroidetes	Parabacteroides	−2.63E-05	9.80E-06	0.084	0.008
Bacteroidetes	Prevotella_7	1.96E-04	8.01E-05	0.066	0.016
Bacteroidetes	Alistipes	−2.80E-05	1.17E-05	0.095	0.018
Bacteroidetes	Alistipes	−2.77E-05	1.19E-05	0.066	0.021
Bacteroidetes	Bacteroides	−3.14E-05	1.42E-05	0.077	0.029
Bacteroidetes	Bacteroides	5.30E-05	2.43E-05	0.061	0.031
Firmicutes	Incertae_Sedis	−1.03E-04	4.97E-05	0.076	0.041
Firmicutes	Blautia	2.46E-05	1.19E-05	0.055	0.041
Firmicutes	-	−2.11E-05	1.06E-05	0.039	0.049

^aAssociations were performed using multivariable linear regression adjusted for age (years), sex, body mass index at blood draw (in kg/m²)

^bThe reported results belong to 9 significant associations out of 296 gut microbial taxa that were tested.

^cAll associations with a nominal p<0.05 are shown. No result was significant after correction for multiple comparisons (using the False Discovery Rate method).

Abbreviations: SE=Standard Error

The relationship between HEI and metabolites is influenced by gut enterotype interactions

Gut microbiome variation may influence the relationship between diet and metabolism. We ran a gut enterotype-dietary pattern interaction analysis using linear regression models to identify interactions between the two gut microbial enterotypes identified in our analysis and HEI on plasma and gut metabolites. These analyses revealed 35 plasma metabolites with significant microbiome-HEI interactions (Supplementary Table 2) out of 800 metabolites that were examined. More specifically, individuals who were categorized as Enterotype 2 and had higher diet quality were more likely to have lower levels of some of the lipid compounds including ceramides. However, this group of participants were more likely to have higher levels of both N-acetyl or N-acyl derivative of amino acids (N-acetylglutamine, N-linoleoylglycine and N-stearoyltaurine) and PUFAs which mostly included omega-3 and omega-6 ones. The most significant associations for this interaction analysis belonged to a glycerophospholipid known as 1-palmitoyl-2-linoleoyl-GPC (16:0/18:2) (β= 0.008, p=0.0009).

We repeated the HEI-enterotype interaction analyses for gut metabolites, and identified significant interactions between enterotype and HEI on 40 gut metabolites (Supplementary Table 3) out of 767 metabolites that were examined. Among gut metabolome, those involved in the synthesis of DNA, including metabolites involved in the synthesis of ribose (sedoheptulose-7-phosphate) and the nucleotide thymidine 5'-monophosphate, and both primary and secondary bile acids (chenodeoxycholate, glycolithocholate sulfate) were higher in the Enterotype 2 participants with higher diet quality. However, fatty acid derivatives and amino acids (AAs) including branch chain AAs (leucine and isoleucine) were lower in this group. The most significant association was identified between the Enterotype 2 and higher HEI with thymidine 5'-monophosphate (β=0.041, p=0.0007).

The glycine and serine metabolism pathway was enriched due to the interaction between gut enterotype and HEI

Metabolites that were significantly influenced by the microbiome-diet quality interaction were further investigated by metabolite enrichment analysis using Metaboanalyst 4.0 (Figure 3). For plasma metabolites, our findings revealed that the glycine and serine metabolism pathway was significantly enriched (p=0.044) for the gut enterotype-diet quality interaction. This top enriched pathway included four metabolites (adenosine 5'-monophosphate (AMP), S adenosylhomocysteine (SAH), alpha-ketoglutarate and 3-phosphoglycerate) identified as significant in our interaction analysis (Supplementary Figure 2).

Figure 3-.

Metabolite sets enrichment analysis (MSEA) overview reached through plotting log of p-values which are obtained from pathway enrichment analysis on Y-axis and pathway impact values which are obtained from pathway topology analysis on X-axis.

For gut metabolites, there were three pathways that were significantly influenced by the interaction. These pathways included aminoacyl-tRNA biosynthesis (p=9.7x10⁻⁶), valine, leucine and isoleucine biosynthesis (p=0.008) and pyrimidine metabolism (p=0.03).

Discussion

The findings of this multi-omic study of diet quality highlighted circulatory and gut metabolites in healthy adults in the ABO Study and circulatory metabolites in an obese population in the FAIR Study that were associated with the HEI diet quality score. The meta-analysis of findings from the ABO and FAIR Studies identified 47 circulatory metabolites related to diet quality. Our data also highlighted microbial genera associated with the HEI score. Moreover, we also showed a gut enterotype-dependent association of diet quality with circulatory and gut metabolites suggesting that the quality of diet could influence systematic and gut metabolisms in a gut microbiome-dependent manner.

Findings from our meta-analysis suggest that amino acids derivatives (N-acetylglutamine, N-methylproline and O-acetylhomoserine) and lipids (phospholipids, sphingomyelins) circulatory metabolites are mostly altered with changes in diet quality. We have replicated previous findings on circulatory metabolites associations with the HEI score [5, 19, 20]. The metabolites that have been previously identified in male participants and are consistent with our findings included N-δ-acetylornithine, a metabolite that was known to be involved in arginine synthesis [5], ergothioneine, a sulfur compound derivative of betaine and histidine [19], and a homologue of stachydrine (homostachydrine) [5] [in our study we identified stachydrine as a HEI biomarker]. Similarly, results in postmenopausal women also suggest that alterations in the HEI score would directly change the levels of ergothioneine [20]. In contrast to these two studies, our research not only identify essential fatty acids as the blood biomarkers of HEI but revealed that sphingosines were inversely related to this diet quality index. These differences in findings could be attributed to the specific metabolite fractions that have been measured in a single study but not analyzed in other studies. Our comparison did not include studies in which other indices of diet quality including alternate healthy eating index (AHEI) and dietary approaches to stop hypertension (DASH) were used since metabolomic biomarkers of each measure may reflect its food components and therefore are specific to that particular diet index.

These identified biomarkers may be involved in mechanisms explaining how adherence to an unhealthy dietary pattern would increase risk for several diseases. Overload of polar lipids (sphingosine and phospholipids) may lead to a vicious cycle by imposing mitochondrial dysfunction and a subsequent decrease in fatty acid oxidation which would itself cause more lipid overload. Changes in mitochondrial functions affect metabolism and health and therefore contribute to several diseases including metabolic syndrome and cardiovascular disorders [21]. In addition, our data showed that metabolites involved in fatty acid and amino acid break down (oxalate, isobutyrylcarnitine (C4) [22]) may have direct associations with the HEI index. Therefore, a healthy dietary pattern may contribute to lower disease risk by increasing the utilization of metabolites such as fatty acids and amino acids. Furthermore, alteration of lipid profiles might lead to changes in triglyceride lipolysis and glucose metabolism, a phenomenon observed in prediabetic patients [23].

We observed that the gut microbiome, both at enterotype and microbial taxa levels, was linked to the HEI score. Likewise, the gut microbiome has been shown to be associated with long-term diet and short-term diet interventions. More precisely, it was suggested that animal fat intake was associated with Bacteroides in a previous work [24]. Consistently, we found that higher HEI representing lower intake of saturated fats (one of the components of the index) was associated with reduced abundance of Bacteroides. Moreover, dietary fiber intake was associated with abundance of Prevotella [25, 26], the finding which is in harmony with our result indicating that having higher diet quality, gained through higher intake of dietary components rich in fiber (whole grain, total vegetable, whole fruits, green and beans) was associated with increased abundance of a Prevotella taxon. While previous literature focused on specific food groups rather that a dietary pattern, these findings suggest that the quality of diet might alter the gut microbiome composition [27].

To address whether HEI-gut microbiome interaction would influence circulatory and gut metabolome, we first investigated the relationship between gut metabolites and the HEI score. Our results on gut metabolites suggest positive associations of oleate and linoleate and lipid molecules containing these two fatty acids with higher diet quality. However other lipids including fatty acids and phospholipids, amino acids derivatives and peptides (glycylisoleucine and gamma-glutamyl-epsilon-lysine) showed inverse associations with HEI. In separate analysis models, we identified metabolites (circulatory and gut metabolome) that varied by diet quality and gut microbiome interaction. Gut metabolites that were influenced by having higher diet quality in the Enterotype 2 participants included those involved in the DNA synthesis and bile acids derivatives (which were directly affected) and fatty acids and branch chain amino acids (which were inversely affected). Of circulatory metabolites, our study showed that healthy PUFAs and acyl and acetyl derivatives of amino acids were directly, and polar lipids were inversely influenced in this group of participants.

The findings described above along with our data indicating that the Enterotype 2 participants had lower BMI [8] might shed light on mechanisms underlying how the interaction between diet quality and gut microbiome composition would be linked to states of health and diseases [28, 29]. Studies have revealed that long-chain essential PUFAs may be predictive of inflammatory and cardiovascular diseases, cancer, and other chronic disorders [30]. While fecal short chain fatty acids were reported to have positive effects on health [31], our study documented for the first time that long chain saturated gut free fatty acids and amino acids (including branch chain amino acids) were inversely influenced by the interaction and therefore could be elevated in metabolic diseases. Although the mechanisms driving these observations are unclear, it might be through the same mechanisms (oxidative stress, inflammation, mitochondrial dysfunction, and impaired metabolite catabolism) in which circulatory free fatty acids and amino acids play important role in development of diseases [32, 33].

Our findings also suggest pathways specific to the relationships between HEI and circulatory/gut metabolites and between the diet quality-gut microbiome interaction and circulatory/gut metabolites. Glycerophospholipid metabolism, a complex pathway involved in the metabolism of polar lipids and diacylglycerols (DAGs) [34] was found to explain why a large number of sphingomyelins and other polar lipids were identified as biomarkers of diet quality in our study. Moreover, assessing circulatory metabolites, we identified enriched pathways (glycine, serine and threonine metabolism) demonstrating that pathways involved in amino acid metabolism could be influenced by either diet quality or the interaction of diet quality and gut microbiome. For the gut metabolome, however, our data suggests that amino acid-involved pathways were altered at two levels: 1) upstream level where aminoacyl-tRNA biosynthesis was mediated by diet quality and microbiome interaction and 2) downstream level where amino acid-related pathways including histidine metabolism and valine, leucine and isoleucine biosynthesis were affected by diet quality and the interaction, respectively. Caffeine metabolism was also altered by diet quality but as the caffeine intake was not correlated with diet quality in our data, any alteration in this pathway could not be attributed to the caffeine intake. It was demonstrated that the activity of CYP1A2, an enzyme involved in the clearance of caffeine, varies by race, sex, genetics and diseases [35, 36]. Therefore, we hypothesize that diet quality might change caffeine metabolism through affecting CYP1A2 activity. This remains to be further explored. Finally, pyrimidine metabolism which is characterized as a pathway generating nucleotides in the process of normal cell proliferation [37], was one of the other pathways affected by the interaction. Considering this finding, we could speculate that the interaction between diet and the gut microbiome affects pathways of fundamental importance to cellular homeostasis. In sum, lipids, amino acids, caffeine and nucleotide metabolism were altered by diet quality which could be dependent or independent of differences in gut microbiome composition.

Our work was the first study relating diet quality with gut metabolome and microbiome in a multi-omic framework that has allowed us to assess how healthy/unhealthy diet-induced gut microbiome composition would affect body metabolic status. By sampling from a healthy population, we aimed to avoid potential confounding by metabolic and inflammatory complications that could be present in a disease setting, however, we validated the results at the circulatory metabolite level in a sample of a high-risk population (obese adults) to enhance the generalizability of our findings. The result of our meta-analysis of these two samples was consistent with other major published works [5, 20] in which diet quality was similarly assessed with HEI but in a subgroup of population with variable health status. Our study is distinctly strengthened by collecting large number of metabolites (both at circulatory and gut levels) and gut microbial taxa in addition to dietary data which represented a long-term habitual dietary intake [38]. However, our study has several limitations that should be acknowledged. Our sample size was modest, meaning we were limited in statistical power, and associations did not remain significant after FDR adjustment. While we were able to address this by measuring metabolites in an independent sample to validate the plasma-metabolite-HEI relationships, we did not have a validation sample for gut metabolites and microbiota. We consider these analyses preliminary and hypothesis-generating, and present our unadjusted findings for informational purposes, and to spur further research in this area [39]. Future studies will be required to validate these findings. Although diet data was collected by a reliable assessment tool (FFQ) used in most epidemiological studies, our findings are limited by measurement error as it is a self-reported data. Also, in our cross-sectional study design, we were unable to determine causation between the exposure and outcome, an issue that did not allow us to infer whether gut microbiome alterations were because of changes in metabolism or vice versa. In addition, given the complexity of these relationships, other unmeasured confounders such as environmental exposures and menstrual cycle may exist. However, adjustment for race did not change the results and we did not observe significant differences by sex.

In conclusion, we found that HEI, as a predefined measure of diet quality, was associated with several circulatory and gut metabolites. We also identified that diet quality might alter the metabolites through changing gut microbiome composition. This phenomenon might be driven by modulating several metabolic pathways susceptible to the diet quality or gut microbial taxa. Findings of this study 1) provide molecular and metabolic mechanisms for the studies investigating the link between food patterns and metabolic health outcomes and 2) can be used in future preventive and therapeutic studies exploring novel targets for the intervention. Further research on longitudinal data is needed to replicate our findings, find the causality in the observed relationships and determine whether the associations between diet quality and metabolic health outcome are medicated by gut microbiome composition.

Supplementary Material

1730063_Sup_material

supplementary Figure 1- Participant flow chart

Supplementary Figure 2- Scatterplot highlighting significant gut microbiome-HEI interaction on plasma levels of adenosine 5'-monophosphate (AMP) [A], S adenosylhomocysteine (SAH) [B], alpha-ketoglutarate [C] and 3-phosphoglycerate [D]. Lines represent linear regression relationships.

Supplementary Table 1- Multivariable linear regression analyses for the association of the Healthy Eating Index with plasma metabolites in the Fish oils and Adipose Inflammation Reduction (FAIR) study (N=25)

Supplementary Table 2- Multivariable linear regression analyses for the association of the interaction between Healthy Eating Index and gut enterotype with plasma metabolites

Supplementary Table 3- Multivariable linear regression analyses for the association of the interaction between Healthy Eating Index and gut enterotype with gut metabolites

Abbreviations:

AA

Amino acid

AHEI

Alternate healthy eating index

AMP

Adenosine 5'-monophosphate

DAG

diacylglycerol

DASH

Dietary approaches to stop hypertension

DHQ

Diet History Questionnaire

FAIR

Fish oils and Adipose Inflammation Reduction

FFQ

food frequency questionnaires

HEI

Healthy eating index

IQR

Interquartile range

MUFA

Monounsaturated

PAM

Partitioning Around Medoids

PUFA

Polyunsaturated fatty acid

SAH

Adenosylhomocysteine

UPenn NGSC

University of Pennsylvania Next-Generation Sequencing Center

VANTAGE

Vanderbilt University Technologies for Advanced Genomics