Dietary metabolic signatures and cardiometabolic risk

Abstract

Aims

Observational studies of diet in cardiometabolic-cardiovascular disease (CM-CVD) focus on self-reported consumption of food or dietary pattern, with limited information on individual metabolic responses to dietary intake linked to CM-CVD. Here, machine learning approaches were used to identify individual metabolic patterns related to diet and relation to long-term CM-CVD in early adulthood.

Methods and results

In 2259 White and Black adults (age 32.1 ± 3.6 years, 45% women, 44% Black) in the Coronary Artery Risk Development in Young Adults (CARDIA) study, multivariate models were employed to identify metabolite signatures of food group and composite dietary intake across 17 food groups, 2 nutrient groups, and healthy eating index-2015 (HEI2015) diet quality score. A broad array of metabolites associated with diet were uncovered, reflecting food-related components/catabolites (e.g. fish and long-chain unsaturated triacylglycerols), interactions with host features (microbiome), or pathways broadly implicated in CM-CVD (e.g. ceramide/sphingomyelin lipid metabolism). To integrate diet with metabolism, penalized machine learning models were used to define a metabolite signature linked to a putative CM-CVD-adverse diet (e.g. high in red/processed meat, refined grains), which was subsequently associated with long-term diabetes and CVD risk numerically more strongly than HEI2015 in CARDIA [e.g. diabetes: standardized hazard ratio (HR): 1.62, 95% confidence interval (CI): 1.32–1.97, P < 0.0001; CVD: HR: 1.55, 95% CI: 1.12–2.14, P = 0.008], with associations replicated for diabetes (P < 0.0001) in the Framingham Heart Study.

Conclusion

Metabolic signatures of diet are associated with long-term CM-CVD independent of lifestyle and traditional risk factors. Metabolomics improves precision to identify adverse consequences and pathways of diet-related CM-CVD.

Structured Graphical Abstract

Structured Graphical Abstract.

This figure summarizes the general approach of this work. The bottom portion of this figure (on the CCA analysis) was inspired by Wang et al.⁷⁸ The readers are referred to this excellent work for further exposition on the utility of CCA in high-dimensional data analysis.

See the editorial comment for this article ‘Improving precision in estimating diet-disease relationships with metabolomics’, by A. Mente et al., https://doi.org/10.1093/eurheartj/ehac616.

Introduction

Since foundational observations by Keys nearly four decades ago on the association between dietary intake and cardiometabolic-cardiovascular disease (CM-CVD), interventional and observational studies have aimed to understand the clinical impact of dietary intake on CVD outcomes, giving rise to different policy and clinical recommendations regarding which diets may be ‘CVD-optimal’.¹ Inference from the population to an individual-level remains an important limitation to translation of observational nutritional research. Indeed, the individual-level metabolic response to diet—tied to weight gain, diabetes, and CM-CVD—relies on a broad array of features beyond macronutrient intake, including metagenomic composition and host-related lifestyle factors.² Understanding an individual’s dietary intake via history still has limitations, as the diet history does not provide insight into the potential metabolic effect of diet on the individual due to biological heterogeneity in the metabolic response to diet.³ These issues have limited potential insights into the extent of and mechanisms by which dietary intake is a critical, modifiable metabolic risk factor for CM-CVD.

Metabolite profiles of dietary intake have recently been advanced as objective measures of the metabolic responses of the host to genetic and lifestyle/environmental factors. However, precise mechanisms by which an individual’s diet influences metabolism and downstream CM-CVD are multi-factorial (e.g. microbiome, host factors) and not well understood. Recent efforts suggest selected food and beverage exposures and their nutrient composition are associated with dynamic changes in human metabolism^4,5 that may influence downstream CVD risk in older adults.⁶ Nevertheless, integrating broad metabolite profiles alongside composite dietary patterns that reflect common patterns of intake may more precisely identify heterogeneity that underlies metabolic response to diet and CVD in young, racially diverse populations at high lifetime risk. Here, we examine >4000 individuals across two major U.S.-based cohort studies across age, sex, and race to derive and validate multi-metabolite signatures of dietary pattern against CM-CVD. Our ultimate goal was to link variations in self-reported dietary pattern, metabolites, and clinical outcomes as an important step in deciphering the relations between diet, metabolism, and clinical CM-CVD.

Methods

The general approach of our research method is shown in the Structured Graphical Abstract.

Study cohorts and metabolite profiling

The Coronary Artery Risk Development in Young Adults (CARDIA) is a prospective cohort study of 5115 White and Black participants at baseline in 1985–1986 (age 18–30 years, four field centres: Birmingham, AL; Chicago, IL; Minneapolis, MN; Oakland, CA). Our analytic sample includes 2259 individuals with metabolite profiling at Year 7 (fasting at least 8 h) and completed diet history from Year 7 and complete covariates (see below). Assessment of standard cardiovascular risk factors and physical activity has been previously described.^7–10 For validation purposes, we studied 2006 participants of the Framingham Heart Study (FHS) Second Generation (Offspring) cohort who underwent fasting metabolite profiling at their 5th examination cycle. Metabolite profiling was performed at the Broad Institute of Harvard-MIT (Cambridge, MA; methods previously published^11–13) and included (in CARDIA) amino acids and related polar metabolites and sugars (‘HILIC-positive’ and ‘HILIC-negative’), lipids (‘C8-positive’), and free fatty acids, lipid-derived mediators and bile acids (‘C18-negative’). FHS metabolites were performed on an earlier version of the Broad platform (with a subset of the modern metabolomics platform¹¹). All individuals provided written informed consent, and CARDIA and FHS studies were approved by the Institutional Review Board at each participating institution.

Assessment of diet in CARDIA

Dietary intake at Year 7 was assessed by in-person interviews by trained and certified staff members using the CARDIA diet history questionnaire.^14,15 Information obtained included brand name (if known), usual frequency consumed (per day/week/month), and portion sizes estimated by use of food models. Nutrient and food groups were output from the Nutrition Data System for Research programme.^14,16 For this analysis, 17 food and beverage groups were studied: red meats, processed meats, dairy, candy, sugar-sweetened beverages, refined grains, whole grains, fish and seafood, chicken, nuts, legumes, fruits, fruit juices, vegetables, dark green vegetables, eggs, and diet beverages. Animal fat and plant fat were also studied. A healthy eating index (HEI2015) score was created as a measure of dietary quality.¹⁷ Dietary exposures were continuous in analysis.

Clinical outcomes

We investigated incident diabetes and CVD in FHS and CARDIA. We defined diabetes as per previous work, and those methods are reproduced here with minimal alteration for scientific rigor.¹³ In CARDIA, incident diabetes was defined as the composite of self-reported diabetes, fasting glucose ≥7.0 mmol/l or 2-h glucose of ≥11.1 mmol/l on an oral glucose tolerance test (if performed). Age of onset of diabetes was taken as self-reported age of onset of diabetes or, if not reported, age at first examination with self-reported diabetes or with plasma glucose measurements consistent with diabetes. Time to diabetes or censor was measured as the difference in age from the Year 7 study visit to censoring (loss to follow-up) or to the onset of diabetes. In FHS, diabetes was defined as medication use for diabetes, fasting glucose ≥7.0 mmol/l, random glucose ≥11.1 mmol/l or haemoglobin A1c ≥6.5% (where measured), with survival time as time to event or censor. The definition for CVD in CARDIA included fatal/non-fatal myocardial infarction, acute coronary syndrome, stroke, heart failure, carotid or peripheral artery disease. For FHS, CVD included fatal or non-fatal myocardial infarction, angina, coronary insufficiency, heart failure; stroke; transient ischaemic attack; claudication; or cardiovascular death. Time to CVD or censor was measured as the time from the metabolite measures to censoring in both cohort studies. Models for diabetes excluded individuals with prevalent diabetes at the time of metabolite measures, and models for CVD excluded individuals with CVD at baseline.

Statistical methods

Association between metabolites and food groups in CARDIA

We began our analysis in CARDIA. Metabolites with less than 10% missingness across subjects were imputed as half the lowest detected value, log transformed, and standardized. Metabolites that had greater than or equal to 10% missingness or coefficient of variation >10% were excluded. Dietary intake and covariates were transformed with hyperbolic arcsine and standardized. We estimated linear regressions for each dietary intake (predictor) and metabolite level (as outcome) pairing separately, using Benjamini-Hochberg false discovery rate (FDR) to control for Type 1 error across the metabolome, separately for each combination of food group and adjustment model. Regressions were performed with three sets of adjustments: (Model 1) estimated total caloric intake; (Model 2) estimated total caloric intake, age, sex, and race; (Model 3) estimated total caloric intake, age, sex, race, annual household income, education, physical activity, and smoking (never, former, or active).

Quantifying predictive accuracy and prognostic association of diet-related metabolites in CARDIA

Next, we sought to address (i) how addition of metabolites to total caloric intake would improve prediction of self-reported dietary intake and (ii) how metabolite-based scores built on dietary intake (where dietary intake is the dependent variable in elastic nets, described in the following context) would perform in association with clinical CM-CVD outcomes, relative to self-reported dietary intake itself.

First, we constructed elastic nets for each of the 19 food groups as a function of the entire metabolome with adjustment for total caloric intake forced into models. To estimate the increase in explanatory information from the metabolome for each food group, we compared the R² of a base model for food intake (outcome) based on total caloric intake (predictor) to the R² of an elastic net for food intake (outcome) that included total caloric intake (forced into model) and the metabolome (penalized). Elastic net hyperparameters were optimized by cross-validation. Resulting metabolite-based predictions of food intake were included in Cox models for CM-CVD outcomes in CARDIA and compared with self-reported intake. Cox regressions for incident diabetes were adjusted for age, sex, race, body mass index (BMI), fasting glucose, and parental history of diabetes; regressions for CVD were adjusted for age, sex, race, BMI, total cholesterol, high-density lipoprotein cholesterol (HDL-C), systolic blood pressure, use of antihypertensive medication, self-reported diabetes and smoking (as never, former, and active).

Identifying composite metabolite signatures of common dietary intake patterns in young adults

One of the limitations of traditional multivariable regression in nutritional research is its reliance on specifying one food group (as the dependent variable) as a function of many metabolites, proteins, or other biomarkers of interest. A common solution is to pick an a priori ‘ideal’ food pattern score (e.g. Mediterranean diet, HEI) or ‘reduce’ the dietary space to independent components (e.g. principal components of diet) for regression against biomarkers. These approaches pre-specify food patterns not necessarily representative of common intake patterns across the United States¹⁸ (e.g. Mediterranean), resulting in limited power and generalizability at the individual level. In addition, methods such as principal components analysis would not necessarily allow us to jointly optimize relations between the metabolome and diet. To address this limitation, we used regularized sparse canonical correlation analysis (CCA) to identify a parsimonious set of metabolites maximally related to variation in self-reported food intake in CARDIA (Structured Graphical Abstract).

CCA is a method of identifying pairs of latent dimensions among two domains of data (e.g. across food groups and metabolites). The canonical variates within a domain (e.g. food groups or metabolites) are all independent of each other. However, the first canonical variate from the diet domain will be maximally correlated to the first canonical variate from the metabolite domain and likewise for each subsequent pair of canonical variates across domains. Unlike principal components analysis, this approach allows diet and metabolome to ‘supervise’ their relation to each other (e.g. dietary variate 1 ∼ metabolomic variate 1; dietary variate 2 ∼ metabolomic variate 2, and so on), maximizing the relation between dietary intake and the metabolome. The resulting canonical variates are described by weightings that can be used to generate summary scores for each CARDIA participant for each diet or metabolomic variate. Metabolomic variate-based scores can be then used in downstream association analyses as a metabolic signature of the dietary intake described by the corresponding dietary variate.

In traditional CCA, each metabolomic canonical variate would include all metabolites measured. However, this is likely to be less externally translatable, and not all metabolites are necessary to optimally encode a given pattern of food intake. To identify a parsimonious set of metabolites that jointly capture maximal variation in diet in the population, we utilized iterated LASSO regression within the CCA to optimize the metabolite domain, a method previously developed to identify correlated components in audio and video streams.¹⁹ We sequentially optimized 19 canonical variates (one for each diet variable) for sparsity in the metabolite domain (more ‘sparsity’ = fewer metabolites included in any given variate) to identify an optimum number of metabolites (called ‘cardinality’). We used repeated cross-validation to identify the optimal cardinality between 1 and 562 metabolites to maximize the canonical correlation for each pair of canonical variates. Given several cardinalities may have similar cross-validated canonical correlations, we selected the smallest cardinality with a cross-validated canonical correlation within 1% of the maximal cross-validated canonical correlation we identified. Because this optimization requires repeated penalized regression across multiple cross-validation folds and numerically intensive matrix operations, it was executed in parallel across hundreds of compute nodes in the Great Lakes High Performance Supercomputing Cluster (Advanced Research Computing Group, University of Michigan, Ann Arbor) using slurm and OpenMPI to distribute tasks. This approach gave us the optimal cardinalities (e.g. numbers of metabolites) for each of the 19 canonical variates in the metabolite domain to include in the final regularized CCA.

We carried forward the first metabolomic CCA variate for further survival analysis against CM-CVD, given its high correlation with the corresponding dietary variate (canonical correlation = 0.77) and the interpretability of the weightings of different food groups in the corresponding dietary variate. Cox regressions for diabetes and CVD were adjusted as mentioned previously. In addition, models for the HEI and for the HEI plus the first metabolomic CCA score were also estimated.

Relation of metabolic signatures of common dietary intake patterns in the FHS

We identified the subset of metabolites measured in CARDIA also measured in FHS using the Human Metabolome Database (http://hmdb.ca). Metabolite levels which were not detected in <10% were imputed as half the lowest detected level, log transformed and standardized. Validation was performed for the top metabolite canonical variate from CARDIA in the FHS. We built a ‘reduced’ score in CARDIA by fitting the first metabolomic variate score (dependent variable) as a function of 24 metabolites in the CCA in CARDIA that were also measured in FHS (independent variables) in linear regression; the coefficients from this ‘reduced’ score were used in conjunction with FHS metabolite levels to allow application of the CCA-based score to FHS. The correlation between the initial ‘full’ and the ‘reduced’ metabolomic variate scores in CARDIA was good (r = 0.92). Regressions for incident diabetes were performed adjusted for age, sex, BMI, fasting glucose, and parental history of diabetes; regressions for incident CVD were performed adjusted for age, sex, BMI, total cholesterol, HDL-C, use of lipid lowering medications, systolic blood pressure, use of blood pressure medications, treated diabetes, and current smoking.

All analyses were performed in R (R Foundation for Statistical Computing, Vienna, Austria) and a two-tailed P < 0.05 (with FDR adjustment as noted) considered statistically significant. All models were fit on complete cases.

Results

Clinical characteristics of the study samples

The clinical characteristics of the CARDIA study sample are shown in Table 1. The study sample was on average in young adulthood (age 32.1 ± 3.6 years) with a BMI 25.7 ± 4.9 kg/m² with 45% women and 44% Black Americans. The cross-correlation among dietary intake measures is shown in SupplementaryFigure S1. The characteristics of the FHS cohort (older, more prevalent cardiometabolic risk) in this study are shown in SupplementaryTable S1.

Table 1.

Characteristics of the CARDIA analytic sample (N = 2259)

Characteristics
Age (years)	32.1 (3.6)
Women	1008 (45)
White	1267 (56)
BMI (kg/m2) (N = 2246)	25.7 (4.9)
Waist circumference (cm) (N = 2238)	82.0 (11.8)
Total cholesterol (mmol/l) (N = 2258)	4.56 (0.85)
High-density lipoprotein cholesterol (mmol/l) (N = 2258)	1.40 (0.35)
Diabetes (self-reported) (N = 2257)	41 (2)
Parental history of diabetes (N = 2109)	391 (19)
Systolic blood pressure (mmHg)	108 (11)
Diastolic blood pressure (mmHg)	69 (9)
Reported use of hypertension medication (N = 2258)	20 (1)
Smoking
ȃNever	1379 (61)
ȃFormer	362 (16)
ȃCurrent	518 (23)
Glucose (mmol/l) (N = 2251)	5.0 (0.5)
Total physical activity by CARDIA questionnaire (arbitrary units)	375.6 (281.5)
Years of education	14.9 (2.5)

Values are n (%) or mean (standard deviation) at Year 7.

Associations between the circulating metabolome and food groups and pathways of cardiometabolic risk

Our first hypothesis was that survey-based dietary intake would be associated with a specific set of circulating metabolites implicated in CM-CVD (Structured Graphical Abstract; Supplementary Data File 1). Several metabolites associated with dietary exposure reflected food-related components or catabolism (Table 2), including red and processed meat (multiple plasmalogens, phospholipids,⁵ hydroxyproline,²⁴ choline/carnitine metabolites), fish [long-chain unsaturated triacylglycerols: C60:12, C58:9, C58:10; long-chain unsaturated fatty acid-containing fats, as well as ω-3 omega fatty acids (DHA, EPA)], citrus or juice consumption (proline-betaine, arginine),^5,57 white meat (3-methylhistidine⁵⁸), diet beverages (saccharin), and plant-based intake (trigonelline, a niacin metabolite linked to favourable metabolic effects⁵⁹). In addition, several diet-associated metabolites potentially represented metabolic interaction with host features linked to cardiometabolic risk (e.g. liver metabolism, gut microbiome), including hippurate^30,31 (related to gut microbial diversity, lower visceral adiposity, a high fruit/whole grain diet) and trimethylamine-N-oxide [TMAO]⁶⁰ (related to red meat consumption and CVD, modifiable with plant-based diets⁶¹). Importantly, we identified associations between diet and metabolites both known and those not previously widely described in dietary epidemiology that—while not necessarily directly food-derived—have been implicated in cardiometabolic disease^38,49 [e.g. dimethylguanidinovaleric acid (DMGV), ceramide/sphingomyelin lipid metabolism].

Table 2.

Selected metabolites and their relation to food groups. The number of arrows indicate the strength and direction of association from regressions associating metabolite levels (response) to food group intake (predictor), adjusted for total caloric intake, age, sex, race, income, education, physical activity, and smoking.

Metabolite	Red meat	Fish and sea food	Chicken	Eggs	Animal fat	Refined grains	Dairy	Diet drinks	Sugary drinks	Dark green vegetables	Fruit	Whole grains	Fruit juice	Nuts	Clinical–pathologic associations
Choline/carnitine metabolism
Butyrobetaine	↑↑	↓	↑	↓	↑	↑↑	↑↑	—	↓	↓↓	↓↓	↓↓	↓	↓↓	Diet-related metabolites; cholines found in red meat, eggs; selected choline metabolites associated with diabetes and CVD (in different directions)20,21
Carnitine	↑↑	↑	↑↑	↑	↓	↓	↑	↑	↓	↓↓	↓↓	↓	—	↓↓
Choline	↑	↓	↑	↑↑	↑	↑	↑	↑	↓↓	↑↑	↓	↑	↓	↓
Trimethylamine-N-oxide	↑↑	↑↑	—	↑	↓	↓	↑	—	↓↓	↓↓	↓↓	↓↓	↓	↓
Betaine	↓↓	↓	—	↑	↓	↓	↓↓	↓	↓	↑↑	↑↑	↑↑	↑	↑
Alpha-glycerophosphocholine	↑↑	↑↑	↓	↑↑	↓	—	↓	↑	↑	↓↓	↓	↓↓	↓	↓↓
Glutamate/glutamine metabolism
Glutamate	↑↑	↑	↑	↑↑	↑	—	↑	↑↑	↓	↓↓	↓	↓↓	↓	↓	Associated with diabetes risk,22 glutamine associated with fat inflammation23
Glutamine	↓↓	↓↓	↓↓	↓	↑	↑	↓	↓↓	↑	↑	↑	↑↑	↑	↓
Hydroxyproline	↑↑	↑	↑↑	↑↑	↑↑	↑↑	↓↓	↑	↑	↓↓	↓↓	↓↓	↓↓	↓↓	Diet-related metabolite, derived from meat consumption; higher circulating level related to greater meat consumption;24,25 higher concentrations associated with CVD
Asparagine	↓↓	↑	↓↓	↓	↓	↓	↑↑	↓	↓	↑↑	↑↑	↑↑	↑	↑	Intestinal barrier function; Lower levels related to diabetes;26 may have a role in maintaining intestinal epithelial barrier protection (via intestinal TLR4)27
Cinnamoylglycine	↓↓	↓↓	↑	↓	↓	↓↓	↑↑	↑	↓↓	↑↑	↑↑	↑↑	↓	↑↑	Gut microbial product; associated with lower BMI,28 increased microbial diversity, lower T2D risk, and decreased visceral fat29
Hippurate	↓↓	↓↓	↓↓	↑	↓	↓↓	↑↑	↓	↓↓	↑↑	↑↑	↑↑	↑↑	↑↑	Gut microbial product; higher levels related to lower metabolic syndrome30 and visceral adiposity;31 may be related to greater fruit/whole grain consumption
Branched-Chain Amino Acids
Isoleucine	↑↑	↓	↑	↑↑	↓	↑	↑	↑	—	↓↓	↓	↓↓	↓	↓	Essential amino acids; greater insulin resistance or diabetes risk;11 BCAA feeding may decrease insulin resistance (through mTOR-dependent mechanisms);32 BCAA catabolic defect may be related to diabetes;33 associated with decreased dietary quality (lower alternative HEI34)
Valine	↑↑	↑	↑	↑	↓	↓	↑↑	↑↑	↑	↓	↓	↓↓	↓	↓↓
Leucine	↑↑	↑	↑	↑↑	↑	—	↑	↑↑	↓	↓	—	↓↓	↓	↓
Trigonelline	↓↓	↓	↓↓	↑	↓↓	↓	↑↑	↑	↓↓	↑↑	↑↑	↑↑	↓	↓	Diet-related metabolite (select plants); reduces inflammation, oxidative stress35,36
DMGV	↑	↓	↓↓	↓	↓	↑↑	↑	↑↑	↑↑	↓	↓	—	↑↑	—	Diet-related metabolite (higher levels with sugary beverages; lower levels with vegetables37); related to hepatic steatosis, diabetes;38 decreased after acute exercise39
Glycine	↓↓	↓	↓↓	↓	↓↓	↑	↑	↓	↓	↑↑	↑↑	↑↑	—	↑↑	Related to dysglycemia, adiposity, oxidative stress, incretin secretion,40 and lower risk of diabetes41
Taurine	↑	↑↑	↑	↓	↑	↓	↓	↑	—	↑	↑↑	↑	↑↑	↓	Diet-related metabolite (found in meats, fish, dairy42); beneficial effects on oxidative stress and blood pressure43
Polyunsaturated fatty acids
DHA	↓↓	↑↑	↑	↑↑	↓↓	↓↓	↓↓	↓↓	↑	↑↑	↑↑	—	↑	↓↓	Diet-related metabolite (ω-3: oily fish, plants; ω-6: nuts, vegetable oils); EPA related to reduced CVD; linoleic acid related to lower CVD risk44
EPA	↓↓	↑↑	↑	↓	↓↓	↓↓	↓	↓	↓	↑↑	↑↑	↓	↓	↓↓
Alpha-linolenic acid	↓↓	↑	↑	↓	↓↓	↓↓	↓	↑↑	↓	↑↑	↑	↑↑	↑	—
Linoleic acid	↓↓	—	↑	↓	↓↓	↓↓	↓↓	↑↑	↓↓	↑	↑↑	↑↑	↑	↑↑
Tri- and diacylglycerols
Saturated TAG	↑↑	↓	↓↓	↑	↑	↑	↑↑	—	↑↑	↓↓	↓↓	↓↓	↑	↓↓	Diet-related metabolites; longer chain unsaturated TAGs related to fish intake;5 increased intracellular DAG may increase cellular insulin resistance;45 greater degree of unsaturation related to higher dietary quality (higher alternative HEI34)
Unsaturated TAG	↓↓	↑↑	↑↑	↑↑	↓↓	↓↓	↓↓	↓	↑	↑↑	↑↑	↓	↑	↓↓
DAG	↑↑	↓	↑	↑	↓↓	—	↓	↑	↑↑	↓	↓	↓↓	↑	↓
Sphingolipids and ceramides
Sphingosine 1-phosphate	↑↑	↑	↑↑	↑	↓	↑↑	↓↓	↑↑	↑	↓	↓↓	↓↓	↓↓	↓↓	Sphingosine 1-phosphate46 and ceramides may be related to CVD;47 ceramides implicated in insulin resistance and cardiometabolic risk;48 sphingolipids associated with CVD risk,49 and sphingomyelins associated with decreased dietary quality (lower alternative HEI34)
Sphingomyelins	↑↑	↑↑	↑↑	↑	↑↑	↑	↑	↑↑	—	↓↓	↓↓	↓↓	↓	↓↓
Ceramides	↑↑	↑	↑↑	↑↑	↓↓	↑	—	↑	↓	↓	↓	↓↓	↓	↓↓
Bile acids
Glycocholic acid	↑	—	↓	↓	↓	—	↓	—	↑	↓	↓↓	↓↓	↑↑	↑	Complex regulation, with influence of microbiome; glycocholic acid, taurocholic acid, glycochenodeoxycholic acid higher in individuals with diabetes50
Glycocholate	↓	↓	↓↓	↓↓	↓	↑	↓	↑	↓	—	↓	↑	↑	↑↑
Taurocholic acid	↑↑	↑	↓	—	↑	↓	↑↑	↑	↑	—	↓	↓↓	↑	—
Glycochenodeoxycholic acid	↑	↑	↓	↓↓	↓	↑↑	↓↓	↓	↑↑	↓	↓↓	↓↓	↑	↑
Plasmalogens	↑↑	↑↑	↑↑	↑↑	↑↑	↓	↓↓	↓	↓	↓↓	↓↓	↓↓	↓↓	↓↓	Diet-related metabolite (higher levels associated with red meat intake5)
Phytanic acid	↓↓	↓↓	↓↓	↓	↑	↓↓	↑↑	↑	↓	↑↑	↑	↑↑	↑↑	↑	Diet-related metabolite (fatty acid in dairy51); may have effects on PPAR activity52 and UCP-1 expression,51 impacting fatty acid oxidation and overall energy balance (protective against insulin resistance52)
Saccharin	↑↑	↓	↑↑	↓↓	↓↓	↑↑	↑↑	↑↑	↓↓	↓↓	↓↓	—	↓	↓	Additive associated with diet drinks
Pantothenate	↓↓	↑	↑	↑↑	↓↓	↓↓	↑↑	↑↑	↓	↑↑	↑↑	↑↑	↑↑	↑	Vitamin B5; found in whole grains, eggs and other foods53,54
5-acetylamino-6-amino-3-methyluracil	↑	↓	↓	↓	↑↑	↑↑	↑↑	↑↑	↓	↓	↓↓	↓↓	↓↓	↓	Caffeine metabolite55
Pipecolic acid	↓↓	↑↑	↓	↓	↓	↓↓	↓↓	↓↓	↓	↑↑	↑↑	↑↑	—	↓↓	Associated with bean consumption56
Proline-betaine	↓↓	↑	↓↓	↓↓	↓↓	↓↓	↑↑	↓↓	↑↑	↑↑	↑↑	↑↑	↑↑	↑↑	Related to citrus consumption57
3-methylhistidine	↑↑	↑↑	↑↑	↑↑	↓	↓↓	↑	↑	↑	↓	↑	↓↓	↓	↓↓	Previously associated with white meat (chicken) intake, with plasma levels falling within days after consumption58
Cotinine	↑	↑	↓	↑	↓	↑	↓	↓	↑↑	↓	↓↓	↓↓	↑	↓	Nicotine metabolite (smoking)

Metabolic signatures for individual food groups in young adults and long-term CM-CVD risk

Metabolites improved prediction (by R²) of survey-based dietary measures beyond total caloric intake for nearly all food groups (Figure 1). Furthermore, metabolome-wide signatures of self-reported intake were strongly associated with long-term incident diabetes and CVD in CARDIA [Figure 2; 234 incident diabetes events with median 23.0 years (IQR 18.0–23.0 years) follow-up; 107 incident CVD events with median 24.9 years (IQR: 24.8–25.1) follow-up]. Although the 95% confidence intervals (CIs) showed overlap, the point estimate for association for incident diabetes and CVD was higher for the metabolic signatures relative to self-reported dietary intake.

Figure 1.

Precision of the metabolome to identify dietary intake. The proportion of variance explained (R²) by total caloric intake (TCI; black circles) and metabolite levels plus total caloric intake (red circles). As noted in text, this figure represents the results of elastic nets where dietary intake is the dependent variable and the metabolome (with or without TCI) is the independent variable.

Figure 2.

Metabolite signatures of dietary intake are more consistently associated with long-term cardiometabolic and cardiovascular disease relative to survey-based dietary intake. Standardized hazard ratios for diabetes (left) and CVD (right) for metabolite scores (red/lower dots, based on elastic net regression of dietary intake as a function of the metabolome) and survey-based dietary intake (black/upper dots) are displayed. Metabolite signatures generally exhibited larger effect sizes than the survey estimates for CM-CVD outcomes. Models presented here are fully adjusted (as specified in the Methods for diabetes and CVD).

Identifying a composite metabolite-based score of common dietary pattern in young adulthood and its relation to long-term CM-CVD risk

We next used penalized CCA to identify metabolite patterns maximally related to dietary variation in CARDIA (Structured Graphical Abstract). LASSO regularized CCA identified a first metabolomic canonical variate with 50 metabolites that was maximally correlated with food intake patterns (weightings in Supplemental Data File 2). Penalized CCA weightings for the first canonical variate alongside relations between metabolites and food group from linear models are shown in Figure 3. The first canonical variate for diet was weighted positively on red and processed meat, eggs, chicken, and sugary drinks, with negative weights for vegetables, dairy, whole and refined grains, and fruit. Metabolite weightings for the first canonical variate were used to calculate scores (as described in Methods). Of note, scores were related to BMI (r = 0.25, P < 0.0001), and they were higher in Blacks and men. In survival analysis in CARDIA for diabetes and CVD, the fully adjusted effect size for the CCA-based metabolite score [diabetes: standardized hazard ratio (HR) = 1.62 (95% CI: 1.32–1.97), P < 0.0001; CVD: HR = 1.55 (95% CI: 1.12–2.14), P = 0.008] was greater than that of HEI for both outcomes in separate models and in models with both together (Figure 3, with overlap in the 95% CIs for both). In FHS, we observed significant variability in the reduced CCA score by sex (higher in males: score difference 0.54, P < 0.0001), and BMI (r = 0.17, P < 0.0001). In our FHS sample, there were 283 incident cases of diabetes with a median follow-up of 18.0 years (IQR: 12.0–20.0 years) and 518 incident cases of CVD with a median follow-up of 21.6 years (IQR: 13.5–23.4 years). The CCA-based metabolite score was associated with incident diabetes after full adjustment (standardized HR: 1.38, 95% CI: 1.20–1.59, P < 0.0001) with a trend for CVD (HR: 1.09, 95% CI: 0.99–1.20, P = 0.08).

Figure 3.

Penalized canonical correlation and survival analysis for metabolite signatures of common dietary patterns in CARDIA and FHS. Panel (A) shows a heat map of regression coefficients relating metabolite levels (outcome) to dietary exposures (independent variables), adjusted for total caloric intake, for metabolites selected by the penalized CCA. Weightings for the first canonical variate are shown as marginal bar plots. Panel (B) shows forest plots of standardized hazard ratios for the first metabolomic CCA score for incident diabetes and CVD from Cox models in CARDIA. In general, the CCA-based metabolite score has a larger effect size than HEI for both outcomes. Panel (C) shows corresponding hazard ratios from FHS.

Discussion

Despite physiologically plausible associations between intake of select foods or patterns and cardiometabolic outcomes,^62,63 limitations of survey-based dietary instruments (e.g. recall and ascertainment bias, interrelation among food groups) and inter-individual differences in metabolism³ have prompted a call for increased rigour in studies of nutrition in CM-CVD. In response, several contemporary studies have investigated association between circulating or urinary metabolic biomarkers with survey-based dietary intake or in the context of dietary challenge.^4,5,64–73 The results of the present investigation directly address the relevance of the metabolome in dietary assessments in CM-CVD. First, we observed a broad array of metabolites associated with survey-derived dietary intake, reflecting food-related components or catabolites, interactions with host features (e.g. microbiome), or pathways broadly implicated in CM-CVD. While metabolite patterns described a significant proportion of the variation in self-reported dietary intake, specific food-related metabolite patterns (e.g. red meat, processed meats, vegetables) were more strongly related to long-term incident diabetes and CVD than survey-based dietary intake itself. Finally, using penalized CCA, we identified a parsimonious set of metabolites linked to a diet pattern classically unfavourable for CM-CVD (e.g. high in red/processed meat, sugary drinks; low in vegetables, grains, nuts, fruits). Metabolite signatures reflecting this pattern were associated with long-term clinical CM-CVD beyond a healthy eating index in CARDIA, and these associations replicated in FHS for diabetes and CVD (though non-significant for CVD after full adjustment). Collectively, these findings suggest the importance of metabolic correlates of diet in improving precision for relevant outcomes in cardiometabolic disease (beyond the dietary exposures themselves), framing efforts to develop functional metabolic biomarkers of CM-CVD rooted in acute and chronic diet exposures.

Until recently, the question of whether ‘precision’ molecular measures of dietary intake can serve as biomarkers reflecting the impact of diet on CVD has not been widely explored. Contemporary approaches have linked metabolites to composite measures of dietary quality, testing association of identified metabolites (either singly or in composite scores) with CVD.^6,34 In a recent seminal study in the PREDIMED randomized trial and large observational studies (Nurses’ Health Studies I and II and Health Professionals Follow-up Study), Li et al.⁶ derived a multi-metabolite score based on a Mediterranean diet pattern that was associated with CVD, proposing this score as a measure of adherence and response to a Mediterranean dietary pattern in future studies. Nevertheless, a unique strength and novelty of the current approach is its applicability to free-living dietary patterns: instead of building metabolite associations on an a priori ‘ideal’ diet not common in high-risk populations,¹⁸ we used methods (canonical correlation) to match common variation in dietary intake across young White and Black adults in the cohort with a composite, multi-metabolite profile, which was in turn related to CM-CVD outcomes across two large cohorts with nearly two decades of follow-up.

Beyond the innovations in methods and population studied, we identified a broad metabolome associated with dietary intake and plausible mechanisms of CM-CVD observed in model systems, human physiologic studies, and prior cohort epidemiology (Table 2). In addition to associations with plausible food composition-related metabolites (e.g. saccharin in diet beverages), we identified metabolites reflecting host–diet interactions (those linked to host microbiome or intestinal function, e.g. hippurate, asparagine, cinnamoylglycine) or those broadly related to insulin resistance or inflammation (e.g. lipid derivatives, including sphingomyelin/ceramide pathway metabolites; see Table 2 for references). Similar to Li et al. where the multi-metabolite Mediterranean diet score was not tightly associated with the survey-based diet itself (R² ≈ 9%), metabolites explain a moderate fraction of variation in self-reported diet intake in CARDIA (Figure 1). Nevertheless, individual and composite (CCA) metabolite-based dietary intake scores were significantly associated with long-term CM-CVD, even in cases where the parent, survey-based dietary measure was not. While we cannot specifically pinpoint the reasons for this divergence using observational data, it may reflect limited precision in quantifying diet by survey, the ability of the metabolome to capture clinical risk beyond dietary exposures, or underlying complementary information provided by both the dietary intake history and the metabolome. Deconvoluting metabolic signatures for individual dietary exposures is likely to be difficult based on our results, due to (i) lack of specificity of common metabolites for a single food group, (ii) reliance on common metabolites and/or those with structural annotation for association (e.g. relatively rare metabolites may reflect specific foods), (iii) dynamic relations between food ingestion and acute changes in circulating or urinary metabolite excursions, and (iv) the influence of bioavailability of nutrients in food and other genetic or environmental factors (e.g. smoking, exercise) that may influence metabolism. Diet remains a complex exposure, with advantages of observational studies (with comprehensive questionnaires of diet) complemented by controlled, acute and sustained dietary challenge needed to sharpen precision for individual foods or macronutrients.^4,74

The results in CARDIA extend findings in older adults with prevalent cardiometabolic disease (e.g. dyslipidemia, hypertension, obesity) to a younger biracial cohort, demonstrating universal importance for diet-derived metabolite signatures in CVD and diabetes. Nevertheless, several important limitations merit mention. The present assessment is insensitive to metabolites that change transiently during feeding. Nevertheless, select biomarkers of diet intake that rise after acute and chronic exposure to certain foods linked to CM-CVD (e.g. red meat with TMAO⁷⁵) were still associated with the self-reported measure in CARDIA, consistent with our goal to capture metabolic correlates of chronic dietary intake. Importantly, CARDIA and FHS were different populations, with a younger, more geographically and racially diverse population in CARDIA. Despite recognized differences in dietary pattern by geography and emerging recognition of important differences potentially related to social determinants of health,^18,76,77 the replication of major metabolite-based signals from CARDIA to FHS suggests potential for a wide generalizability across time, race, and geography. We cannot surmise the origin of diet-associated metabolites in this study, and residual confounding (as with any nutritional epidemiologic study) may be an important limitation. The finding by our group and others of select non-dietary metabolites (e.g. cotinine) in signatures of lower ‘quality’ diet and the moderate predictive ability (R²) of metabolite-based scores for individual food groups or composite patterns⁶ have led to the proposal for prospective, controlled feeding studies to understand the metabolic response to diet.⁷⁴ In addition, integration of other axes of metabolic function (e.g. metagenomics, proteomics, transcriptional changes) across target tissues of interest (e.g. liver, adipose, muscle, urine) is likely to provide more comprehensive assessment of the metabolic impact of diet and is the subject of ongoing precision medicine efforts.⁷⁴ The mitigation of effect size for CVD after adjustment for risk factors in FHS also raises the intriguing possibility that ‘intermediate’ traits (e.g. diabetes, hypertension, hyperlipidaemia) may mediate the associations between diet-associated metabolic signatures and CVD; larger studies (including studies that focus on distinct components of the CVD endpoint separately) are needed to confirm this hypothesis. While linear regression assumptions are not explicitly tested (e.g. non-linearity, heteroskedasticity), the idea of non-linearity is an important aspect of metabolic responses to diet, with dramatically larger sample sizes required to confidently address this. Finally, we recognize that the association of HEI and the CCA-based metabolite scores with CVD or diabetes have overlapping confidence bounds. Nevertheless, with advancing techniques for metabolite quantification, it is likely that parsimonious metabolite-based scores may serve as more widely, rapidly deployable ‘surrogate’ measures of nutrition’s impact on cardiometabolic health, a concept of intense interest. Ultimately, extending these observations to prospective studies of diet in controlled settings (including interventional studies during specific dietary exposure) will identify not only food-based metabolic biomarkers linked to CVD but will also allow interrogation of axes of metabolism implicated in cardiometabolic health revealed through dietary challenge.

In conclusion, in a sample of young White and Black adults, we identified specific metabolites associated with dietary food groups and patterns, demonstrating that metabolite signatures of food groups are more strongly linked to diabetes and CVD than the survey-defined food groups themselves. Identified metabolites not only reflected constituents of specific consumed foods, but also in vivo processing (e.g. gut microbial metabolism). Using unbiased approaches, we generate metabolite-directed patterns of food consumption in the community, defining metabolic measures of dietary intake strongly associated with CVD and diabetes across in young adults over two decades of follow-up. Further underscoring efforts to improve nutrition through the use of innovative integrative methods⁷⁸ toward improved cardiometabolic health.

Supplementary material

Supplementary material is available at European Heart Journal online.

Funding

This work was supported by grants from the National Institutes of Health (R01 HL136685 to R.V.S. and V.L.M.; K23-HL138260, R01-HL156975 to M.N., K01-HL127159 to K.M.) and the American Heart Association. The Coronary Artery Risk Development in Young Adults Study (CARDIA) is conducted and supported by the National Heart, Lung, and Blood Institute (NHLBI) in collaboration with the University of Alabama at Birmingham (HHSN268201800005I and HHSN268201800007I), Northwestern University (HHSN268201800003I), University of Minnesota (HHSN268201800006I), and Kaiser Foundation Research Institute (HHSN268201800004I). The Framingham Heart Study (FHS) acknowledges the support of Contracts NO1-HC-25195, HHSN268201500001I and 75N92019D00031 from the National Heart, Lung and Blood Institute and NIH grant R01DK080739 for this research. V.L.M. is supported by the Melvyn Rubenfire Professorship in Preventive Cardiology. M.N. is also supported by a Career Investment Award from the Department of Medicine, Boston University School of Medicine. R.S.V. is supported in part by the Evans Medical Foundation and the Jay and Louis Coffman Endowment from the Department of Medicine, Boston University School of Medicine. This manuscript has been reviewed by CARDIA for scientific content. The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services.

Data availability

The data underlying this article can be obtained through the CARDIA (https://www.cardia.dopm.uab.edu/) or Framingham Heart Study (FHS) studies (http://www.framinghamheartstudy.org) directly.