Identifying biomarkers of dietary patterns by using metabolomics^,,

Abstract

Background: Healthy dietary patterns that conform to national dietary guidelines are related to lower chronic disease incidence and longer life span. However, the precise mechanisms involved are unclear. Identifying biomarkers of dietary patterns may provide tools to validate diet quality measurement and determine underlying metabolic pathways influenced by diet quality.

Objective: The objective of this study was to examine the correlation of 4 diet quality indexes [the Healthy Eating Index (HEI) 2010, the Alternate Mediterranean Diet Score (aMED), the WHO Healthy Diet Indicator (HDI), and the Baltic Sea Diet (BSD)] with serum metabolites.

Design: We evaluated dietary patterns and metabolites in male Finnish smokers (n = 1336) from 5 nested case-control studies within the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study cohort. Participants completed a validated food-frequency questionnaire and provided a fasting serum sample before study randomization (1985–1988). Metabolites were measured with the use of mass spectrometry. We analyzed cross-sectional partial correlations of 1316 metabolites with 4 diet quality indexes, adjusting for age, body mass index, smoking, energy intake, education, and physical activity. We pooled estimates across studies with the use of fixed-effects meta-analysis with Bonferroni correction for multiple comparisons, and conducted metabolic pathway analyses.

Results: The HEI-2010, aMED, HDI, and BSD were associated with 23, 46, 23, and 33 metabolites, respectively (17, 21, 11, and 10 metabolites, respectively, were chemically identified; r-range: −0.30 to 0.20; P = 6 × 10⁻¹⁵ to 8 × 10⁻⁶). Food-based diet indexes (HEI-2010, aMED, and BSD) were associated with metabolites correlated with most components used to score adherence (e.g., fruit, vegetables, whole grains, fish, and unsaturated fat). HDI correlated with metabolites related to polyunsaturated fat and fiber components, but not other macro- or micronutrients (e.g., percentages of protein and cholesterol). The lysolipid and food and plant xenobiotic pathways were most strongly associated with diet quality.

Conclusions: Diet quality, measured by healthy diet indexes, is associated with serum metabolites, with the specific metabolite profile of each diet index related to the diet components used to score adherence. This trial was registered at clinicaltrials.gov as NCT00342992.

See corresponding editorial on page 293.

INTRODUCTION

Consuming a diet that conforms to national dietary guidelines promotes well-being and prevents chronic disease throughout the life span. Research has mostly focused on individual food or nutrient associations with disease, yet overall diet quality is influenced not only by the quality of these distinct factors but their combinations, quantity, and interrelations (1–3). Diets can be summarized by dietary patterns, which represent the complexity of overall diet exposure (4–8). The most consistent evidence for dietary pattern–disease associations has emerged from cohort studies applying predefined diet indexes based on national dietary guidelines or evidence of health benefit from the literature (9).

Systematic reviews and large meta-analyses have convincingly demonstrated that higher diet quality, as described with the use of predefined indexes such as the Healthy Eating Index (HEI)¹³, is associated with lower risk of overall mortality (5, 9, 10) and lower incidence of cardiovascular disease (CVD) and type 2 diabetes (10, 11). Better compliance with regional dietary patterns defined as Mediterranean and Nordic or Baltic diets have also shown associations with reduced mortality, CVD, type 2 diabetes, and cancer incidence (12–16), as well as improved cardiometabolic risk factors (17). The WHO Healthy Diet Indicator (HDI), based on recommended nutrient cutoffs, is associated with improved life expectancy (18). In many instances, dietary pattern–cancer associations have been found that were not evident when investigating single foods or nutrients (19–24). These findings and the need for additional dietary pattern research were emphasized by the 2015 US Dietary Guidelines Advisory Committee (1). Furthermore, the Interagency Committee on Human Nutrition Research has identified nutritional biomarkers as a substantial research gap necessary to aid in understanding eating patterns and human metabolic processes (25).

Our objective was to determine the associations between 4 dietary patterns and serum metabolite peak intensity with the use of metabolomics, an emerging method that measures small intermediates and products of metabolism (which we collectively refer to as “metabolites”) in biospecimens (26, 27). Metabolites include parent compounds and their metabolites with molecular weight <1000 Da, including amino acids, monosaccharides, small lipids, cofactors and vitamins, energy cycle intermediates, nucleotides, and exogenous xenobiotics (28). A nutritional metabolomics approach can be used to identify metabolites that could serve as candidate nutritional biomarkers if they are replicated in future studies and conform to a classical measurement model. Furthermore, metabolomics can highlight the metabolites and metabolic pathways influenced by diet, which could be further assessed in prospective studies of disease outcomes with the use of mediation analysis to provide insights into the mechanisms that may be driving dietary pattern–disease relations. Although a number of small studies have evaluated individual food item associations with blood metabolite concentrations, to our knowledge, very few have evaluated dietary patterns, and, of these, most have used data-driven approaches with small samples (3, 29–31). Our 2 purposes were as follows: 1) to identify metabolites correlated with 4 dietary patterns and their components [with the HEI-2010, the Alternate Mediterranean Diet Score (aMED), the HDI, and the Baltic Sea Diet (BSD)], and 2) to gain insights into the potential biological mechanisms influenced by diet quality. To achieve these aims, we conducted cross-sectional partial correlation analyses with the use of baseline dietary data and metabolites measured in fasting serum from 5 nested case-control studies within the large, prospective Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study.

METHODS

Study design and population

The ATBC Study was a randomized, double-blind, placebo-controlled, 2 × 2 factorial, primary prevention trial that tested the effects of α-tocopherol and β-carotene supplementation on the incidence of lung and other cancers. Details of the trial have been published previously (32) (Supplemental Figure 1). Between 1985 and 1988, 29,133 healthy male smokers (≥5 cigarettes/d) aged 50–69 y at random assignment who resided in southwestern Finland were identified with the use of a central population register and enrolled. Although the trial ended in 1993, participants have been followed for outcomes for >20 y. Men were excluded from the trial if, at enrollment, they had a history of malignancy other than nonmelanoma skin cancer or carcinoma in situ, severe angina on exertion, chronic renal insufficiency, liver cirrhosis, alcoholism, or other medical conditions that could limit long-term participation. All participants provided written informed consent. The trial and follow-up study were approved by institutional review boards at the US National Cancer Institute and Finnish National Public Health Institute (32, 33). All procedures were in accordance with the Helsinki Declaration of 1975 as revised in 1983.

Demographic characteristics and medical history, smoking status, diet, and physical activity (leisure and occupational) data were collected via a questionnaire during the prerandom assignment baseline visit. Height and weight were measured by trained study personnel (32). Dietary intake for the 12 mo before the initial study visit was measured with the use of a self-administered validated food-frequency questionnaire (FFQ) (203 foods and 73 mixed dishes) based on the Finnish diet and accompanied by a color portion-size picture booklet (34). Trained personnel checked responses at a study center visit. Nutrients were assigned with the use of the National Food Composition Database and software from the Finnish National Institute for Health and Welfare (34). Dietary data for participants who had an implausible caloric intake (<1000 or >5000 kcal/d) and/or incomplete data were excluded.

Nested case-control studies

Metabolite and valid dietary data were available for 1336 participants from nested case-control studies as follows: 1) esophageal cancer, n = 64 controls and 69 cases; 2) pancreatic cancer, n = 114 controls and 115 cases; 3) pancreatic and lung cancer, n = 146 controls and 175 cases; 4) pancreatic cancer, n = 326 controls and 140 cases; and 5) prostate cancer, n = 187 controls only (35). All participants were free of cancer at the time of diet assessment and blood collection.

Metabolite measurement and quality control

Fasting blood samples were collected from participants at their prerandom assignment baseline visit. Serum samples were prepared according to established protocols (36) and stored at −70°C (32). Frozen, once-thawed, and single-aliquot serum samples were sent to Metabolon (Metabolon, Inc.) between 2013 and 2015. Sample extracts were spiked with quality control (QC) standards were added to sample extracts and analyzed with the use of liquid chromatography and mass spectrometry, ultrahigh-performance liquid chromatography and tandem mass spectrometry, and gas chromatography and mass spectrometry with pooled QC technical replicates. Peaks were identified by comparison with Metabolon’s library of pure standards of known chemical structure based on retention time and index, m/z, and chromatographic data.

Matched case and control samples were handled in the same standard manner within each nested case-control set and were analyzed consecutively. Each batch included 10% blinded replicate QC samples. The QC samples were pooled serum samples taken from 2 separate populations of male smokers. The median (IQR) intraclass correlation coefficients over all metabolites measured in each of the 5 nested studies were as follows: 1) 0.89 (0.57–0.98), 2) 0.92 (0.62–0.99), 3) 0.91 (0.50–0.98), 4) 0.90 (0.61–0.98), and 5) 0.87 (0.55–0.97).

Overall, 994–1220 serum metabolites were measured with partial overlap between studies; 626–722 were chemically identified. Metabolite peak intensity was normalized according to run day (metabolite value and median run-day value). A minimum of nonmissing values were assigned for metabolite values below the limit of detection (LOD). Metabolites for which ≥90% of participants fell below the LOD were excluded (n = 32–89/study set) (Supplemental Table 1). The median (IQR) for the percentage of participants with metabolite values below the LOD (i.e., missing) per metabolite for each of the 5 nested studies was as follows: 1) 0% (0–0.2%), 2) 0% (0–3%), 3) 0% (0–1.7%), 4) 0.2% (0–3.8%), and 5) 0.7% (0.5–6.4%). Data were available for 1602 metabolites across the 5 nested studies (881 of which were identifiable; Supplemental Table 2).

Dietary patterns

Diet quality was assessed with the use of 4 predefined diet quality indexes. A summary of diet index scoring algorithms can be found in Table 1 (detailed summary in Supplemental Tables 3–6), as follows: 1) The HEI-2010 was developed by the USDA and the National Cancer Institute as an index of overall diet quality based on the 2010 US Dietary Guidelines for Americans (8, 37, 38); 2) The HDI is based on the international WHO dietary guidelines for the prevention of chronic diseases (43, 44); 3) The aMED was developed on the basis of key findings from epidemiologic studies in Europe from the 1960s that investigated mortality risk factors, and the applied score includes modifications for a non-Greek population (39–42); and 4) The BSD was modeled on previous analyses of Nordic-style dietary patterns based on the BSD Pyramid created by the Finnish Heart Association, Finnish Diabetes Association, and University of Eastern Finland (45). The aMED score excluded legume intake because of low consumption in the study population (peas were included in the starchy vegetables group). The aMED and BSD components were energy-adjusted with the use of the density method. The HEI-2010 standards for all components are density-based, and components within the HDI are scored as a percentage of energy.

TABLE 1.

Summary of scoring systems for predefined diet quality indexes measured in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study¹

	Dietary pattern
	HEI-20102		aMED3		HDI4		BSD5
Food or nutrient	Score	Criteria	Score	Criteria	Score	Criteria	Score	Criteria
Fruit
Total fruit	5	≥0.8 cup eq/1000 kcal	1	≥ median (sv/d)		—		—
Whole fruit (excludes juice)	5	≥0.4 cup eq/1000 kcal		—		—		—
Apples, pears, and berries		—		—		—	0–3	Lowest to highest quartile (g/d)
Vegetables
Total vegetables (no potato)	5	≥1.1 cup eq/1000 kcal	1	≥ median (sv/d)		—		—
Greens and beans	5	≥0.2 cup eq/1000 kcal		—		—		—
Fruits plus vegetables (no potato)		—		—	1	≥400 g		—
Cabbages, roots, tomatoes, lettuce, and cucumber		—		—		—	0–3	Lowest to highest quartile (g/d)
Grains
Whole grains	10	≥1.5 oz eq/1000 kcal	1	≥ median (sv/d)		—		—
Refined grains	10	≤1.8 oz eq/1000 kcal		—		—		—
	0	≥4.3 oz eq/1000 kcal		—		—		—
Rye, barley, and oats		—		—		—	0–3	Lowest to highest quartile (g/d)
Dairy
Total dairy	10	≥1.3 cup eq/1000 kcal		—		—
Low-fat milk							0–3	Lowest to highest quartile (g/d)
Protein
Total protein foods	5	≥2.5 oz eq/1000 kcal		—		—		—
Fish and seafood	5	≥0.8 oz eq seafood/plant proteins/1000 kcal	1	≥ median (sv/d)		—	0–3	Lowest to highest quartile (g/d)
Nuts and seeds		—	1	≥ median (sv/d)		—		—
Red and processed meats		—	1	< median (sv/d)		—	3–0	Lowest to highest quartile (g/d)
Protein		—		—	1	10–15 %E		—
Sugars
Free sugars		—		—	1	<10 %E		—
Fats and oils
Unsaturated fatty acids	10	MP:S ≥2.5	1	≥ median M:S	1	P 6–10 %E	0–3	P:ST, lowest to highest quartile
	0	MP:S ≤1.2		—	0	P <6 %E or >10 %E		—
		—		—	1	S <10 %E		—
		—		—	0	S ≥10 %E		—
Total fat, %							3–0	Lowest to highest quartile
Beverages
Ethanol		—	1	5–25 g/d		—	1	<20 g/d
Nutrients
Sodium	10	≤1.1 g/1000 kcal		—		—		—
	0	≤2.0 g/1000 kcal		—		—		—
Cholesterol		—		—	1	<300 mg		—
Empty calories (from alcohol, solid fats, and added sugars)	20	≤19 %E		—		—		—
	0	≤50 %E		—		—		—
Dietary fiber		—		—	1	>25 g		—
Maximum score	100		8		7		25

¹One ounce = 28.35 g (dry) or 29.57 g (liquid); 1 cup = 8 oz based on the USDA Measurement Conversion Tables (www.ars.usda.gov). aMED, Alternate Mediterranean Diet; BSD, Baltic Sea Diet; eq, equivalent; HDI, Healthy Diet Indicator; HEI-2010, Healthy Eating Index 2010; M:S, ratio of monounsaturated fat to saturated fat; MP:S, ratio of monounsaturated and polyunsaturated fat to saturated fat; oz, ounce; P, polyunsaturated fat; P:ST, ratio of polyunsaturated fat to saturated and trans fat; S, saturated fat; sv, serving; %E, percentage of energy.

²The HEI-2010 has 12 components: total fruit, whole fruit, total vegetables, greens and beans, whole grains, refined grains, dairy, total protein, fish and seafood, fatty acid ratio, sodium, and empty calories. Each component was scored from 0 (no intake) to 20, with a maximum score of 100. Intake between the minimum and maximum is scored proportionately (above or below recommended cutoffs for serving size equivalents per 1000 kcal/d) (8, 37, 38).

³The aMED has 8 components: vegetables, fruits, nuts, whole grains, red and processed meat, fish, alcohol, and MP:S. Participants were assigned a score of 1 if intake was greater than or equal to the median number of servings per day for vegetables, legumes, fruit, nuts, whole grains, fish, and the M:S and less than the median for red and processed meat and for ethanol consumption 5–25 g/d (men); otherwise a score of 0 was assigned. The total score sums the dichotomous variables (39–42).

⁴The HDI has 7 components: specific cutoffs for percentage of energy as SFAs, PUFAs, protein, and free sugars; grams of dietary fiber, fruits, and vegetables (excluding potatoes); and milligrams of cholesterol. Sodium was excluded because there was no measure for salt added at the table. Each food group or nutrient guideline was assigned a score of 1 if intake met the guideline and 0 otherwise. The overall score was the sum of dichotomous variables (range: 0–7) (43, 44).

⁵The BSD has 9 components: fruit (apples, pears, and berries), vegetables (cabbage, roots, tomatoes, lettuce, and cucumber), fish, whole grains (barley, oats, and rye), low-fat milk, processed meat, restricting fat percentage, ratio of polyunsaturated to saturated and trans fat, and limiting alcohol. For fruit, vegetables, whole grains, low-fat or fat-free milk, fish, and fat ratio, quartile 1 scored 0, quartile 2 scored 1, quartile 3 scored 2, and quartile 4 scored 3. Scoring was the opposite for meat products and total fat. Alcohol intake scored 1 if intake was ≤20 g/d and 0 if intake was >20 g/d (45).

The HEI-2010 and aMED were chosen because they were well characterized in the Dietary Patterns Methods Project, which investigated their association with cancer and CVD mortality (46, 47). The Dietary Patterns Methods Project used standardized and consistent scoring algorithms (48) for all diet quality indexes across 3 large cohorts with different FFQs. Previous methods used to define dietary patterns have been inconsistent in the literature, making it difficult to compare results across studies (49). The HDI was selected because scoring is based on recommended nutrient cutoffs; therefore, it represented a contrast to food group–based indexes. The BSD was used to represent a healthy regional diet, given this study’s Finnish population.

To allow for comparison with previous literature for the HEI-2010 and aMED, we created component variables based on the USDA’s Food Patterns Equivalents Database (FPED), which provides cup or ounce equivalents of foods in the database. To obtain FPED values for the food line items from the ATBC Study FFQ, each food in the Finnish FFQ was linked to an identical or comparable food in the FPED based on similar food group or nutrient profiles, and FPED values were added (50). The FFQ response data were then reanalyzed to calculate the intake of FPED food groups for each ATBC Study participant.

Statistics

We assessed differences in baseline characteristics between cases and controls and across nested studies with the use of Wilcoxon’s Signed Rank or Kruskal-Wallis tests for continuous variables and chi-square or Fisher’s exact tests for categorical variables. We calculated the partial Spearman correlation coefficient between each diet index score (continuous or ordinal) and each fasting serum metabolite (continuous, ln-transformed) separately for 5 nested case-control studies. Analyses were adjusted for predefined covariates (51–53), including age at blood draw (years), BMI (kg/m²), number of years smoked, total number of cigarettes smoked per day, caloric intake (kilocalories per day), education (less than elementary or at least elementary), leisure-time physical activity (light to moderate or heavy), occupational physical activity (nonworking, sedentary, light to moderate, or heavy), and case status (case or control). We then obtained a summary estimate for each index-metabolite association measured in ≥2 studies (n = 1316 metabolites) with the use of fixed-effects meta-analysis (54). We repeated these analyses for diet index components. We considered associations to be statistically significant if the probability was <9.5 × 10⁻⁶ for the diet index or, for diet-index components, a more stringent threshold of <1.1 × 10⁻⁶ based on Bonferroni correction. Study heterogeneity was assessed with the use of a Cochrane Q P value (P < 0.01 threshold for significance). In sensitivity analysis, we controlled for known smoking metabolite cotinine to explore residual confounding by smoking (55). We also compared diet index–metabolite associations between cases and controls with the use of Fisher’s r to z transformation and the Wald test for homogeneity.

We then evaluated whether diet quality indexes were associated with metabolic pathways. Metabolites were divided into 46 subpathways with ≥5 measured metabolites per pathway with the use of KEGG PATHWAY metabolism subgroups for Homo sapiens (56) (Supplemental Table 7). First, within each study, we obtained a P value while assessing the relation between each diet quality index and each metabolite by linear regression. We combined the P values across studies by fixed-effects meta-analysis. We obtained a single measure of significance for each metabolic pathway by combining the P values of all metabolites within the pathway by Fisher’s method. We retrieved an overall measure of significance with the use of a parametric bootstrap with 10,000 bootstrap samples (57). First, we performed linear regression of each diet index on covariates (age, BMI, years or number cigarettes smoked, daily caloric intake, education, occupational or leisure physical activity, and case status) to get estimates of the null multivariate models. Then we generated a bootstrap sample from estimated null models and recalculated measures of significance for each sample. Finally, we obtained the empirical P values for each pathway as the proportion of bootstrap samples with measures of significance more extreme than that calculated from the original data. Partial correlation analyses were conducted in SAS version 9.3, and metabolic pathway analyses were conducted in R version 3.1.2.

RESULTS

Participant characteristics

The characteristics of our study population are described in Table 2. The median (IQR) age at blood collection was 57 y (54–62 y). Most participants were overweight (median BMI 26), were heavy smokers (median 20 cigarettes/d and smoked for 37 y), had less than an eighth grade education, and had either moderate or heavy levels of leisure-time and or work-related physical activity. Cases had baseline characteristics that were similar to controls, except for smoking intensity (P > 0.05). Across the 5 nested studies, there were differences in age at blood draw, smoking duration, and leisure-time physical activity level (P < 0.05). The range of number of cigarettes smoked per day was not normally distributed. Because of extreme values, smoking intensity also differed by case-control status and study.

TABLE 2.

Baseline characteristics of participants in 5 case-control studies nested within the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study¹

Characteristic	Total (n = 1336)	Controls (n = 837)	Cases (n = 499)	P2	P3
Age at blood draw, y	57 (54–62)	58 (54–62)	57 (53–61)	0.19	<0.0001
BMI, kg/m2	26 (24–28)	26 (24–28)	26 (24–28)	0.31	0.93
Smoked regularly, y	37 (32–42)	37 (31–42)	37 (32–42)	0.89	0.02
Cigarettes smoked/d,4 n	20 (15–25)	20 (15–25)	20 (15–25)	0.034	0.02
Daily energy intake, kcal/d	2610 (2152–3094)	2622 (2179–3094)	2593 (2135–3102)	0.38	0.23
Education
Less than elementary	1053 (79)	671 (80)	382 (77)	0.12	0.87
At least elementary	283 (21)	166 (20)	117 (23)
Leisure-time physical activity
Light	510 (38)	317 (38)	193 (39)	0.77	0.007
Moderate or heavy	826 (62)	520 (62)	306 (61)
Occupational physical activity
Nonworking	575 (43)	358 (43)	217 (43)	0.43	0.07
Sedentary	204 (15)	119 (14)	85 (17)
Light	235 (18)	154 (18)	81 (16)
Moderate or heavy	322 (24)	206 (25)	116 (23)

¹Values are medians (IQRs) or n (%). Cases and controls obtained from 5 nested studies are the following: 1) esophageal cancer, n = 64 controls and 69 cases, 2) pancreatic cancer, n = 114 controls and 115 cases, 3) pancreatic and lung cancer, n = 146 controls and 175 cases, 4) pancreatic cancer, n = 326 controls and 140 cases, and 5) prostate cancer, n = 187 controls only.

²Comparison of cases and controls. Wilcoxon’s Signed Rank test for continuous variables and chi-square or Fisher’s exact test for categorical variables.

³Comparison across 5 nested studies. Kruskal-Wallis test for continuous variables and chi-square or Fisher’s exact test for categorical variables comparing differences across the 5 nested case-control studies.

⁴Mean number of cigarettes smoked per day was 19.5 for controls and 20.5 for cases. Data were not normally distributed.

Diet scores

For the diet quality indexes, a higher score signified better compliance with the dietary pattern or a healthier diet. The median (IQR) HEI-2010 score was 54 (50–59) out of a total score of 100, the median (IQR) aMED score was 3 (2–4) out of 8, the median (IQR) HDI score was 2 (1–3) out of 7, and the BSD score was 13 (10–16) out of 25 points (Table 3). There was no heterogeneity of the dietary quality scores across the 5 nested case-control studies (all P > 0.05).

TABLE 3.

Dietary pattern scores and intake of diet index components in men in case-control studies nested within the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study¹

Dietary pattern and components	Component value	P2
HEI-2010
Total score (max 100)	54.1 (47.9, 59.0)	0.50
Total vegetables, cup Eq/d	2.1 (1.6, 2.7)
Greens and beans, cup Eq/d	0.0 (0.0, 0.0)
Total fruit, cup Eq/d	0.8 (0.4, 1.3)
Whole fruit, cup Eq/d	0.7 (0.4, 1.2)
Whole grains, oz Eq/d	5.1 (2.6, 7.7)
Dairy, cup Eq/d	3.5 (2.4, 4.5)
Total protein foods, oz Eq/d	6.4 (4.9, 8.2)
Seafood and plant protein, oz Eq/d	0.9 (0.5, 1.4)
Fatty acid ratio	0.9 ± 0.3
Sodium (not at the table), g/d	4.8 (4.0, 5.8)
Refined grains, oz Eq/d	5.6 (3.7, 8.3)
Calories from solid fats and added sugars, kcal/d	852 (649, 1063)
aMED
Total score (max 8)	3 (2, 4)	0.99
Total vegetables (no potatoes), cup Eq/d	0.7 (0.5, 1.1)
Total fruit, cup Eq/d	0.7 (0.4, 1.2)
Nuts, oz Eq/d	0.0 (0.0, 0.0)
Fish, oz Eq/d	0.8 (0.5, 1.3)
Whole grains, oz Eq/d	4.9 (2.7, 7.3)
M:S	0.6 (0.5, 0.7)
Alcohol, g/d	10.4 (2.3, 24.6)
Red and processed meat, oz Eq/d	3.8 (3.0, 4.9)
HDI
Total score (max 7)	2 (1, 3)	0.60
Saturated fat, %E	17.5 (14.3, 20.7)
Polyunsaturated fat, %E	3.2 (2.5, 4.7)
Cholesterol, mg/d	536.8 (412.0, 684.2)
Protein, %E	14.4 (13.2, 15.7)
Fiber, g/d	24.7 (18.5, 31.7)
Fruits and vegetables (no potatoes), g/d	191.4 (119.5, 282.2)
Added sugars, %E	7.0 (5.0, 10.0)
BSD
Total score (max 25)	13 (10, 16)	0.99
Fruit (apples, pears, and berries), g/d	50.4 (25.0, 86.8)
Vegetables (cabbage, roots, lettuce, tomatoes, cucumber, and legumes), g/d	43.0 (25.3, 72.7)
Fish (salmon and freshwater fish), g/d	31.2 (19.2, 48.3)
Whole grains (rye, barley, and oat products), g/d	125.2 (74.8, 188.4)
Low-fat milk, g/d	240.1 (129.4, 469.5)
Red and processed meat, g/d	126.1 (98.0, 160.6)
Fat, %E	40.5 (37.0, 44.0)
P:ST	0.17 (0.12, 0.30)
Alcohol, g/d	10.4 (2.3, 24.6)

¹Values are medians (Q1, Q3) or means ± SDs, n = 1336. One oz = 28.35 g (dry) or 29.57 g (liquid); 1 cup = 8 oz based on the USDA Measurement Conversion Tables (www.ars.usda.gov). Study 1 included 64 controls and 69 cases, study 2 included 114 controls and 115 cases, study 3 included 146 controls and 175 cases, study 4 included 326 controls and 140 cases, and study 5 included 187 controls only. aMed, Alternate Mediterranean Score; BSD, Baltic Sea Diet; Eq, equivalent; HDI, Healthy Diet Indicator; HEI-2010, Healthy Eating Index 2010; max, maximum; M:S, ratio of monounsaturated to saturated fat; oz, ounce; P:ST, ratio of polyunsaturated fat to saturated and trans fat; Q1, 25th percentile; Q3, 75th percentile; %E, percentage of energy.

²For differences across nested studies calculated with the use of the Kruskal-Wallis test.

Correlation between dietary index scores, components, and metabolites

The HEI-2010 was associated with 23 serum metabolites, including 17 with an identifiable chemical structure (r = −0.16 to 0.20; P = 8.5 × 10⁻¹³ to 8.3 × 10⁻⁶) (Table 4). Of the identifiable metabolites, 3 were amino acids, 2 were cofactors or vitamins, 9 were lipids, and 3 were exogenous xenobiotics. The aMED dietary pattern was associated with 46 metabolites, with 21 being identifiable (r = −0.30 to 0.24; P = 5.0 × 10⁻⁶ to 6.0 × 10⁻¹⁵), including 4 amino acids, 1 carbohydrate, 2 cofactors or vitamins, 11 lipids, and 3 xenobiotics. The HDI score was associated with 23 metabolites, with 11 being identifiable (r = 0.12–0.20; P = 6.8 × 10⁻⁶ to 1.5 × 10⁻¹¹). Three were amino acids, 2 were cofactors or vitamins, 4 were lipids, and 2 were xenobiotics. The BSD was correlated with 33 metabolites, with 10 being identifiable (r = −0.16 to 0.19; P value = 5.4 × 10⁻⁶ to 1.5 × 10⁻¹¹). These included 2 amino acids, 1 carbohydrate, 3 cofactors or vitamins, and 4 lipids.

TABLE 4.

Fasting baseline serum metabolites associated with dietary patterns in men in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study (Bonferroni-corrected)¹

Metabolite	Super pathway	r	P2	Associated components (direction of correlation)	Replication across diet indexes
HEI-2010
X-12544		0.20	8.51 × 10−13	Whole grain (+); limit refined grains (+)	HEI-2010, HDI
X-17145		0.18	7.11 × 10−8	Whole fruit (+); whole grain (+)	HEI-2010, aMED, BSD
X-19448		0.17	1.65 × 10−6	Whole grain (+)	HEI-2010, HDI
X-12627		0.17	2.34 × 10−6	Seafood and plant protein (+)	HEI-2010
X-14603		−0.16	4.94 × 10−6	Limit SOFAAS (+)	HEI-2010
Homostachydrine	XEN	0.15	1.99 × 10−8	Whole grain (+); total protein (−)	HEI-2010, HDI
Pyroglutamine	AA	−0.15	5.66 × 10−8	Total and seafood and plant protein (−); limit SOFAAS (−)	HEI-2010
Linoleate (18:2n–6)	LIP	0.15	6.67 × 10−8	Dairy (−); MP:S (+); limit SOFAAS (+)	HEI-2010, aMED, HDI, BSD
Docosapentaenoate (n–3 DPA; 22:5n–3)	LIP	0.15	8.48 × 10−8	Seafood (+); limit refined grains (+)	HEI-2010
Ergothioneine	XEN	0.14	1.40 × 10−7	Vegetables (+); dairy (−); total and seafood and plant protein (+); limit sodium (−); limit SOFAAS (+)	HEI-2010, aMED
Docosahexaenoate (22:6n–3)	LIP	0.14	1.58 × 10−7	Total and seafood and plant protein (+)	HEI-2010, aMED
X-17323		−0.14	1.58 × 10−7	Total protein (−); MP:S (−); limit SOFAAS (−)	HEI-2010, aMED
2-Aminophenol sulfate (X-12253)	XEN	0.14	2.88 × 10−7	Whole grain (+)	HEI-2010, HDI
Dihomo-linoleate (20:2n–6)	LIP	0.14	5.15 × 10−7	MP:S (+); limit SOFAAS (+)	HEI-2010
Stearidonate (18:4n–3)	LIP	0.14	5.97 × 10−7	Seafood and plant protein (+); limit SOFAAS (+)	HEI-2010
1-Linoleoylglycerophosphoinositol	LIP	0.13	9.38 × 10−7	MP:S (+); limit SOFAAS (+)	HEI-2010, HDI
Myo-inositol	LIP	0.13	1.77 × 10−6	None	HEI-2010
Pyridoxate	C/V	0.13	1.96 × 10−6	None	HEI-2010
Adrenate (22:4n–6)	LIP	0.13	3.36 × 10−6	None	HEI-2010
3-Hydroxyisobutyrate	AA	0.12	5.88 × 10−6	None	HEI-2010
Threonate	C/V	0.12	6.84 × 10−6	Vegetables (+), total or whole fruit (+)	HEI-2010, aMED, BSD
N-δ-acetylornithine	AA	0.12	7.63 × 10−6	Vegetables (+)	HEI-2010
Deoxycarnitine	LIP	−0.12	8.34 × 10−6	None	HEI-2010
aMED
1-Myristoleoylglycerophosphocholine (14:1)	LIP	−0.30	1.39 × 10−8	M:S (+)	aMED
X-02269 (X-11469)		0.24	1.24 × 10−11	Fish and seafood (+)	aMED, BSD
Scyllo-inositol	LIP	0.24	2.73 × 10−8	None	aMED
X-17138		0.21	6.00 × 10−15	M:S (+)	aMED, HDI, BSD
X-17145		0.21	1.01 × 10−9	None	HEI-2010, aMED, BSD
Mead acid (20:3n–9)	LIP	−0.21	1.22 × 10−6	M:S (−)	aMED
Phytanate	XEN	−0.21	1.51 × 10−6	M:S (−)	aMED
γ-CEHC	C/V	0.20	2.03 × 10−13	M:S (+), vegetables (+)	aMED, HDI, BSD
X-16944		0.20	3.51 × 10−13	M:S (+)	aMED, HDI, BSD
cis-4-Decenoyl carnitine	LIP	0.20	1.30 × 10−12	M:S (+)	aMED, HDI, BSD
X-11261		0.20	1.54 × 10−9	M:S (+)	aMED, BSD
X-11521		0.19	5.08 × 10−12	M:S (+)	aMED, BSD
X-18921		0.18	2.31 × 10−11	M:S (+)	aMED, HDI, BSD
X-11478		0.18	2.61 × 10−11	M:S (+)	aMED, BSD
X-14939		0.18	1.03 × 10−10	M:S (+)	aMED, BSD
3-Carboxy-4-methyl-5-propyl-2-furanpropanoate	LIP	0.17	1.83 × 10−10	Fish and seafood (+)	aMED
X-12056		0.17	4.03 × 10−10	Vegetables (+)	aMED
X-13435		0.17	1.17 × 10−6	M:S (+)	aMED, HDI, BSD
Stachydrine	XEN	0.16	6.01 × 10−9	Fruit (+)	aMED
X-11305		0.16	1.26 × 10−6	M:S (+)	aMED
Indolebutyrate	AA	−0.16	1.39 × 10−6	M:S (−)	aMED
X-11315		0.15	1.58 × 10−8	Fruit (+)	aMED, BSD
Tryptophan betaine	AA	0.15	1.93 × 10−8	Nuts (+)	aMED
X-18249		−0.15	2.79 × 10−8	M:S (−)	aMED
3-Hydroxy-2-ethylpropionate	AA	−0.15	5.92 × 10−8	None	aMED, BSD
X-15486		0.15	6.88 × 10−8	M:S (+)	aMED
Linoleate (18:2n–6)	LIP	0.15	7.53 × 10−8	M:S (+)	HEI-2010, aMED, HDI, BSD
Linolenate (α or γ 18:3n–3 or 18:3n–6)	LIP	0.15	1.12 × 10−7	M:S (+)	aMED, HDI
X-17323		−0.14	1.12 × 10−7	M:S (−)	HEI-2010, aMED
Chiro-inositol	LIP	0.14	2.04 × 10−7	None	aMED
X-11308		0.14	2.50 × 10−7	Fruit (+)	aMED, HDI, BSD
X-17653		0.14	2.86 × 10−7	M:S (+)	aMED, HDI, BSD
Threitol	CHO	0.14	3.56 × 10−7	Fruit (+)	aMED, BSD
Threonate	C/V	0.14	6.53 × 10−7	Vegetables (+); fruit (+)	HEI-2010, aMED, BSD
1-Linoleoylglycerol (1-monolinolein)	LIP	0.14	7.18 × 10−7	M:S (+)	aMED
X-12225		0.14	3.65 × 10−6	None	aMED
DHA (22:6n–3)	LIP	0.13	8.20 × 10−7	Fish and seafood (+)	HEI-2010, aMED
X-16935		0.13	9.39 × 10−7	M:S (+)	aMED, HDI, BSD
Ergothioneine	XEN	0.13	1.02 × 10−6	Vegetables (+); fish and seafood (+)	HEI-2010, aMED
X-11378		0.13	1.10 × 10−6	M:S (+)	aMED, HDI, BSD
X-17654		0.13	1.10 × 10−6	M:S (+)	aMED, HDI, BSD
X-13835		0.13	1.60 × 10−6	Fish and seafood (+)	aMED
X-18914		−0.13	3.21 × 10−6	None	aMED, BSD
Methyl palmitate (15 or 2)	LIP	−0.13	3.71 × 10−6	None	aMED
N-methylproline or N-methyl proline	AA	0.13	5.01 × 10−6	Fruit (+)	aMED
HDI
X-19448		0.20	1.38 × 10−8	Fiber (+)	HEI-2010, HDI
cis-4-Decenoyl carnitine	LIP	0.19	1.48 × 10−11	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
Linoleate (18:2n–6)	LIP	0.18	2.51 × 10−11	Polyunsaturated fat 6–10% (+)	HEI-2010, aMED, HDI, BSD
X-16935		0.18	6.30 × 10−11	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
X-11378		0.17	2.66 × 10−10	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
X-11308		0.17	2.69 × 10−10	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
2-Aminophenol sulfate (X-12253)	XEN	0.16	3.92 × 10−9	Fiber (+)	HEI-2010, HDI
Homostachydrine	XEN	0.16	1.17 × 10−8	Fiber (+)	HEI-2010, HDI
γ-CEHC	C/V	0.16	1.54 × 10−8	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
X-17653		0.16	1.63 × 10−8	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
X-17654		0.15	2.61 × 10−8	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
X-18921		0.15	4.14 × 10−8	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
S-methylcysteine	AA	0.15	5.32 × 10−8	None	HDI
X-12544		0.15	5.73 × 10−8	Fiber (+)	HEI-2010, HDI
X-16944		0.15	9.83 × 10−8	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
γ-CEHC glucuronide	C/V	0.15	3.12 × 10−7	Polyunsaturated fat 6–10% (+)	HDI
X-13435		0.15	8.53 × 10−6	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
1-Linoleoylglycerophosphoinositol	LIP	0.13	1.30 × 10−6	Polyunsaturated fat 6–10% (+)	HEI-2010, HDI
4-Guanidinobutanoate	AA	0.13	1.79 × 10−6	None	HDI
Linolenate (α or γ 18:3n–3 or 18:3n–6)	LIP	0.13	3.09 × 10−6	Polyunsaturated fat 6–10% (+)	aMED, HDI
X-17138		0.13	4.05 × 10−6	Polyunsaturated fat 6–10% (+)	aMED, HDI, BSD
Betaine	AA	0.13	4.94 × 10−6	None	HDI
X-12306		0.12	6.81 × 10−6	None	HDI
BSD
cis-4-Decenoyl carnitine	LIP	0.19	1.50 × 10−11	P:ST (+)	aMED, HDI, BSD
X-17138		0.18	2.66 × 10−11	Vegetables (+); P:ST (+)	aMED, HDI, BSD
X-11549		0.18	3.57 × 10−8	P:ST (+)	BSD
X-17145		0.18	2.73 × 10−7	None	HEI-2010, aMED, BSD
γ-CEHC	C/V	0.17	1.45 × 10−9	P:ST (+)	aMED, HDI, BSD
X-02269 (X-11469)		0.17	7.22 × 10−7	Fish (+)	aMED, BSD
X-11521		0.16	2.88 × 10−9	Vegetables (+); P:ST (+)	aMED, BSD
Indolepropionate	AA	0.16	6.06 × 10−9	None	BSD
X-16944		0.16	1.73 × 10−8	P:ST (+)	aMED, HDI, BSD
X-11261		0.16	1.71 × 10−6	P:ST (+)	aMED, BSD
X-12637		−0.16	3.55 × 10−6	None	BSD
X-13435		0.16	4.29 × 10−6	P:ST (+)	aMED, HDI, BSD
X-12636		−0.16	4.57 × 10−6	P:ST (−)	BSD
1-Palmitoleoylglycerophosphoinositol	LIP	−0.16	5.58 × 10−6	P:ST (−)	BSD
X-18914		−0.15	2.46 × 10−8	P:ST (−)	aMED, BSD
X-16935		0.15	7.95 × 10−8	P:ST (+)	aMED, HDI, BSD
X-17653		0.15	8.45 × 10−8	P:ST (+)	aMED, HDI, BSD
X-11378		0.15	1.13 × 10−7	P:ST (+)	aMED, HDI, BSD
X-09789		0.14	2.41 × 10−7	Whole grains (+)	BSD
X-17654		0.14	3.22 × 10−7	P:ST (+)	aMED, HDI, BSD
X-11478		0.14	3.64 × 10−7	P:ST (+)	aMED, BSD
X-18921		0.14	6.58 × 10−7	P:ST (+)	aMED, HDI, BSD
Linoleate (18:2n–6)	LIP	0.14	6.83 × 10−7	P:ST (+); reduce total fat percentage (+)	HEI-2010, aMED, HDI, BSD
Threonate	C/V	0.14	6.91 × 10−7	Fruit (+); vegetables (+)	BSD
α-Tocopherol	C/V	0.14	7.55 × 10−7	P:ST (+)	BSD
3-Hydroxy-2-ethylpropionate	AA	−0.14	8.35 × 10−7	P:ST (−)	aMED, BSD
Threitol	CHO	0.13	1.01 × 10−6	Fruit (+); reduce total fat percentage (+)	aMED, BSD
10-Undecenoate (11:1n–1)	LIP	−0.13	1.16 × 10−6	P:ST (−)	BSD
X-14939		0.13	2.97 × 10−6	P:ST (+)	aMED, BSD
X-11438		−0.13	3.49 × 10−6	P:ST (−)	BSD
X-11308		0.13	5.04 × 10−6	P:ST (+)	aMED, HDI, BSD
X-11315		0.13	5.41 × 10−6	Vegetables (+)	aMED, BSD

¹n = 1336. Metabolites beginning with “X-“ represent metabolites with unknown chemical identity. Metabolite information can be found in the Human Metabolome Database (www.hmdb.ca). Metabolic super- and subpathways, mass, retention index, mass spectrometry platform, and other identifying information can be found in the Supplemental Materials. AA, amino acid; aMED, Alternate Mediterranean Diet Score; BSD, Baltic Sea Diet; CHO, carbohydrate; C/V, cofactors/vitamins; HDI, Healthy Diet Indicator; HEI-2010, Healthy Eating Index 2010; LIP, lipid; M:S, ratio of monounsaturated fat to saturated fat; MP:S, ratio of monounsaturated and polyunsaturated fat to saturated fat; P:ST, ratio of polyunsaturated fat to saturated and trans fat; SOFAAS, solid fats and added sugars; XEN, xenobiotics.

²Fixed-effects meta-analysis summary estimate from partial Spearman correlation, controlling for age at blood draw (years), BMI (kg/m²), number of years smoked regularly, number of cigarettes smoked per day, daily caloric intake, education (less than elementary or at least elementary), leisure-time physical activity (light to moderate or heavy), and occupational physical activity (nonworking, sedentary, light to moderate, or heavy). Statistically significant with Bonferroni correction for multiple statistical testing at P < 0.05/(1316 metabolites × 4 dietary patterns) = 9.5 × 10⁻⁶.

There was limited evidence of heterogeneity of diet index–metabolite correlations across studies (Cochran’s Q, P > 0.01), with the exception of unknown X-18249 (aMED Q, P = 0.0001), unknown X-12544, and betaine (HDI Q, P = 0.006 and 0.007, respectively). Several Bonferroni-significant metabolites replicated across ≥2 diet index scores, e.g., metabolites related to whole grains and fiber (2-aminophenol sulfate and homostachydrine), fruit (threonate and threitol), ratio of unsaturated fatty acid to SFA [linoleate (18:2n–6), 1-linoleoglycerophosphoinositol, γ-carboxyethyl-hydroxychroman, cis-4-decenoyl carnitine, and linolenate (α or γ 18:3n–3 or 18:3n–6)], and seafood (ergothioneine and DHA).

We examined correlations between HEI-2010 diet index score–specific components and metabolites (results for identified and named metabolites associated with HEI-2010 components are provided in Table 5; detailed results for all indexes are provided in Supplemental Tables 8–11). Overall, the components used to score the diet indexes that were based on foods (e.g., fruit, dairy, and whole grains) were more strongly correlated with metabolites than were macro- or micronutrients (e.g., “limit sodium”), with the exception of percentage of polyunsaturated fats. The total HEI-2010 score was positively correlated with metabolites associated with most diet index components, except for greens and beans and sodium intake. Fifty percent of aMED-related metabolites were positively correlated with the ratio of unsaturated to saturated fat component; the remaining metabolites were positively correlated with fish, fruit, vegetable, and nut and seed intake, but not whole grains or limiting red and processed meat and ethanol. Sixty-five percent of HDI-related metabolites were positively correlated with the polyunsaturated fat component. Other metabolites were associated with fiber. There were no metabolites correlated with the other HDI components (saturated fat <10% energy, cholesterol <300 mg/d, protein 10–15% energy, fruit and vegetables >400 g/d and free sugars <10% energy). The BSD was associated with metabolites primarily correlated with the ratio of polyunsaturated to saturated and trans fat component, but also with fruit, vegetables, whole grains, fish, and reducing fat percentage.

TABLE 5.

Fasting baseline serum metabolites associated with the components used to score the HEI-2010 (chemically identified metabolites; Bonferroni correction and r > |0.15|) in men in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study¹

HEI-2010 components and metabolites	Superpathway	Subpathway	r	P2
Total vegetables
Ergothioneine	Xenobiotics	Food component or plant	0.19	7.09 × 10−12
Oxalate (ethanedioate)	Carb; C/V	Glyoxylate/dicarboxylate; ascorbate/aldarate	0.16	1.62 × 10−7
N-δ-acetylornithine	Amino acid	Urea cycle; arginine/proline	0.15	2.60 × 10−8
Threonate	C/V	Ascorbate/aldarate	0.15	8.98 × 10−8
Total fruit
Methyl-β-glucopyranoside	Carb	Fructose, mannose, galactose, starch, sucrose	0.26	4.27 × 10−10
Oxalate (ethanedioate)	Carb; C/V	Glyoxylate/dicarboxylate; ascorbate/aldarate	0.24	7.44 × 10−15
Scyllo-inositol	Lipid	Inositol	0.22	2.46 × 10−7
Stachydrine	Xenobiotics	Food component or plant	0.22	3.33 × 10−16
Threonate	C/V	Ascorbate/aldarate	0.22	7.77 × 10−16
Glycerate	Carb	Glycolysis, gluconeogenesis, pyruvate	0.19	5.48 × 10−12
N-methylproline or N-methyl proline	Amino acid	Urea cycle; arginine/proline	0.18	9.74 × 10−10
Threitol	Carb	Pentose	0.17	1.13 × 10−9
Xylonate	Carb	Pentose	0.17	2.00 × 10−9
Ascorbate (vitamin C)	C/V	Ascorbate/aldarate	0.15	2.99 × 10−7
Whole fruit
Oxalate (ethanedioate)	Carb; C/V	Glyoxylate/dicarboxylate; ascorbate/aldarate	0.21	5.50 × 10−12
Threonate	C/V	Ascorbate/aldarate	0.20	9.05 × 10−13
Stachydrine	Xenobiotics	Food component or plant	0.19	5.35 × 10−12
Threitol	Carb	Pentose	0.18	1.57 × 10−10
Glycerate	Carb	Glycolysis, gluconeogenesis, pyruvate	0.17	1.46 × 10−9
N-methylproline or N-methyl proline	Amino acid	Urea cycle; arginine/proline	0.16	1.50 × 10−7
Xylonate	Carb	Pentose	0.16	3.42 × 10−9
Whole grain
Homostachydrine	Xenobiotics	Food component or plant	0.31	<1.00 × 10−16
2-Aminophenol sulfate	Xenobiotics	Chemical	0.21	9.35 × 10−14
Dairy
Galactonate	Carb	Fructose, mannose, galactose, starch, sucrose	0.29	4.73 × 10−11
Myristoyl sphingomyelin	Lipid	Sphingolipid	0.25	1.18 × 10−8
2-Aminoheptanoate or 2-aminoheptanoic acid	Lipid	MCFA; FA, amino	−0.22	7.59 × 10−11
cis-4-Decenoyl carnitine	Lipid	Carnitine	−0.19	5.18 × 10−12
Ergothioneine	Xenobiotics	Food component or plant	−0.18	2.78 × 10−11
Acisoga	Amino acid	Polyamine	−0.17	6.46 × 10−10
3-Carboxy-4-methyl-5-propyl-2-furanpropanoate	Lipid	FA, dicarboxylate	−0.16	2.12 × 10−9
Riboflavin (vitamin B-2)	C/V	Riboflavin	0.16	1.49 × 10−8
S-methylcysteine	Amino acid	Cysteine, methionine, sam, taurine	−0.16	2.22 × 10−9
Protein foods
Ergothioneine	Xenobiotics	Food component or plant	0.22	8.88 × 10−16
N-acetyl-3-methylhistidine	Amino acid	Histidine	0.17	1.41 × 10−9
trans-4-Hydroxyproline	Amino acid	Urea cycle; arginine/proline	0.17	2.34 × 10−10
3-Methylhistidine	Amino acid	Histidine	0.17	8.13 × 10−10
3-Carboxy-4-methyl-5-propyl-2-furanpropanoate	Lipid	FA, dicarboxylate	0.16	4.15 × 10−9
DHA (22:6n–3)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.15	9.50 × 10−8
Pyroglutamine	Amino acid	Glutamate	−0.15	2.42 × 10−8
Seafood and plant protein foods
3-Carboxy-4-methyl-5-propyl-2-furanpropanoate	Lipid	FA, dicarboxylate	0.34	<1.00 × 10−16
Palmitoyl-oleoyl-glycerophosphoglycerol (2)	Lipid	Lysolipid	0.32	1.08 × 10−12
DHA (22:6n–3)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.30	<1.00 × 10−16
1-Docosahexaenoylglycerol (1-monodocosahexaenoin)	Lipid	Monoacylglycerol	0.30	3.78 × 10−11
EPA (20:5n–3)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.28	<1.00 × 10−16
Stearidonate (18:4n–3)	Lipid	LCFA; PUFA (n–3 and n–6)	0.19	1.77 × 10−11
Docosapentaenoate (22:5n–3)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.18	2.11 × 10−11
1-Docosahexaenoylglycerophosphocholine (22:6n–3)	Lipid	Lysolipid	0.18	1.54 × 10−7
N-acetyl-3-methylhistidine	Amino acid	Histidine	0.16	1.34 × 10−8
3-Methylhistidine	Amino acid	Histidine	0.16	3.22 × 10−9
Creatine	Amino acid	Creatine	0.15	1.38 × 10−7
Ergothioneine	Xenobiotics	Food component or plant	0.15	2.80 × 10−8
Pyroglutamine	Amino acid	Glutamate	−0.15	3.53 × 10−8
Ratio of unsaturated to saturated fat
cis-4-Decenoyl carnitine	Lipid	Carnitine	0.41	<1.00 × 10−16
γ-CEHC	C/V	Tocopherol	0.38	<1.00 × 10−16
γ-CEHC glucuronide	C/V	Tocopherol	0.37	<1.00 × 10−16
Linoleate (18:2n–6)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.34	<1.00 × 10−16
1-Myristoleoylglycerophosphocholine (14:1)	Lipid	Lysolipid	−0.34	9.14 × 10−11
Linoleoylcarnitine	Lipid	Carnitine	0.32	1.09 × 10−12
Phytanate	Xenobiotics	Food component or plant	−0.31	2.26 × 10−13
Methyl palmitate (15 or 2)	Lipid	FA, branched	−0.28	<1.00 × 10−16
γ-Tocopherol	C/V	Tocopherol	0.27	<1.00 × 10−16
1-Linoleoylglycerophosphoinositol	Lipid	Lysolipid	0.27	<1.00 × 10−16
Mead acid (20:3n–9)	Lipid	LCFA; PUFA (n–3 and n–6)	−0.26	8.51 × 10−10
Myristoyl sphingomyelin	Lipid	Sphingolipid	−0.26	4.87 × 10−9
1-Linoleoylglycerol (1-monolinolein)	Lipid	Monoacylglycerol	0.24	<1.00 × 10−16
Linolenate (α or γ 18:3n–3 or 18:3n–6)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.23	1.11 × 10−16
Palmitoyl-linoleoyl-glycerophosphocholine (1)	Lipid	Lysolipid	0.23	1.87 × 10−7
Palmitoyl-linoleoyl-glycerophosphoinositol (1)	Lipid	Lysolipid	0.23	2.17 × 10−7
Pentadecanoate (15:0)	Lipid	LCFA	−0.23	1.11 × 10−16
1-Palmitoleoylglycerophosphoethanolamine	Lipid	Lysolipid	−0.23	2.33 × 10−11
1-Palmitoleoylglycerophosphoinositol	Lipid	Lysolipid	−0.23	9.81 × 10−11
Palmitoyl-linoleoyl-glycerophosphocholine (2)	Lipid	Lysolipid	0.22	7.75 × 10−7
17-Methylstearate	Lipid	FA, branched	−0.22	3.33 × 10−15
1-Pentadecanoylglycerol (1-monopentadecanoin)	Lipid	Monoacylglycerol	−0.21	8.92 × 10−8
Caprate (10:0)	Lipid	MCFA	−0.20	1.39 × 10−13
γ-Muricholate (hyocholate)	Lipid	Secondary bile acid	0.20	1.61 × 10−8
Myristoleate (14:1n–5)	Lipid	LCFA	−0.20	6.57 × 10−13
13-Methylmyristic acid	Lipid	FA, branched	−0.20	6.96 × 10−12
1-Pentadecanoylglycerophosphocholine (15:0)	Lipid	Lysolipid	−0.20	1.99 × 10−8
3-Hydroxy-2-ethylpropionate	Amino acid	Valine, leucine/isoleucine	−0.20	4.53 × 10−13
10-Undecenoate (11:1n–1)	Lipid	MCFA	−0.19	9.28 × 10−12
2-Linoleoylglycerophosphocholine	Lipid	Lysolipid	0.19	6.28 × 10−8
Dihomo-linoleate (20:2n–6)	Lipid	LCFA; PUFA (n–3 and n–6)	0.18	2.53 × 10−11
Myristate (14:0)	Lipid	LCFA	−0.18	1.16 × 10−10
Campesterol	Lipid	Sterol or steroid	0.17	7.91 × 10−10
10-Nonadecenoate (19:1n–9)	Lipid	LCFA	−0.17	2.49 × 10−10
α-CEHC sulfate (X-12435)	C/V	Tocopherol	0.16	6.18 × 10−7
Docosadienoate (22:2n–6)	Lipid	LCFA; PUFA (n–3 and n–6)	0.16	9.88 × 10−9
1-Linoleoylglycerophosphocholine (18:2n–6)	Lipid	Lysolipid	0.16	1.06 × 10−8
1-Myristoylglycerol (1-monomyristin)	Lipid	Monoacylglycerol	−0.16	7.18 × 10−9
2-Linoleoylglycerol (2-monolinolein)	Lipid	Monoacylglycerol	0.16	7.09 × 10−9
10-Heptadecenoate (17:1n–7)	Lipid	LCFA	−0.15	7.21 × 10−8
Limit sodium
1-Oleoylglycerophosphoinositol	Lipid	Lysolipid	0.15	2.74 × 10−8
Limit refined grains
4-Androsten-3β,17β-diol disulfate (1)	Lipid	Sterol or steroid	0.22	3.33 × 10−16
1-Palmitoleoylglycerophosphoinositol	Lipid	Lysolipid	0.19	6.15 × 10−8
Ethyl glucuronide	Xenobiotics	Detoxification; chemical	0.18	8.35 × 10−11
α-Hydroxyisovalerate	Amino acid	Valine, leucine/isoleucine	0.17	3.30 × 10−10
Docosapentaenoate (22:5n–3)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.15	9.38 × 10−8
γ-CEHC	C/V	Tocopherol	−0.15	3.58 × 10−8
5α-Androstan-3β,17β-diol disulfate	Lipid	Sterol or steroid	0.15	3.39 × 10−8
Limit solid fats and added sugars
1-Myristoleoylglycerophosphocholine (14:1)	Lipid	Lysolipid	−0.29	4.17 × 10−8
cis-4-Decenoyl carnitine	Lipid	Carnitine	0.27	<1.00 × 10−16
Linoleoylcarnitine	Lipid	Carnitine	0.26	7.96 × 10−9
Phytanate	Xenobiotics	Food component or plant	−0.26	1.53 × 10−9
Linoleate (18:2n–6)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.25	<1.00 × 10−16
Myristoyl sphingomyelin	Lipid	Sphingolipid	−0.25	2.58 × 10−8
γ-CEHC	C/V	Tocopherol	0.21	1.85 × 10−14
3-Hydroxy-2-ethylpropionate	Amino acid	Valine, leucine/isoleucine	−0.20	6.55 × 10−13
Dihomo-linoleate (20:2n–6)	Lipid	LCFA; PUFA (n–3 and n–6)	0.19	1.51 × 10−12
γ-CEHC glucuronide	C/V	Tocopherol	0.19	1.70 × 10−10
Linolenate (α or γ 18:3n–3 or 18:3n–6)	Lipid	Essential FA; PUFA (n–3 and n–6)	0.19	1.15 × 10−11
1-Linoleoylglycerol (1-monolinolein)	Lipid	Monoacylglycerol	0.19	9.38 × 10−12
1-Palmitoleoylglycerophosphoethanolamine	Lipid	Lysolipid	−0.19	9.61 × 10−8
Methyl palmitate (15 or 2)	Lipid	FA, branched	−0.18	2.06 × 10−11
1-Linoleoylglycerophosphoinositol	Lipid	Lysolipid	0.18	2.20 × 10−11
1-Pentadecanoylglycerophosphocholine (15:0)	Lipid	Lysolipid	−0.18	1.99 × 10−7
Pentadecanoate (15:0)	Lipid	LCFA	−0.17	1.45 × 10−9
13-Methylmyristic acid	Lipid	FA, branched	−0.17	7.81 × 10−9
Caprate (10:0)	Lipid	MCFA	−0.15	2.88 × 10−8
Docosadienoate (22:2n–6)	Lipid	LCFA; PUFA (n–3 and n–6)	0.15	7.65 × 10−8
Ergothioneine	Xenobiotics	Food component or plant	0.15	3.43 × 10−8
17-Methylstearate	Lipid	FA, branched	−0.15	4.08 × 10−8

¹n = 1336. Full data (including unidentified metabolites) and metabolite identifying information can be found in the Human Metabolome Database (www.hmdb.ca). Metabolic super- and subpathways, mass, retention index, mass spectrometry platform, and other identifying information can be found in the Supplemental Materials. Carb, carbohydrate; CEHC, carboxyethyl-hydrochroman; C/V, cofactors/vitamins; FA, fatty acid; HEI-2010, Healthy Eating Index 2010; LCFA, long-chain fatty acid; MCFA, medium-chain fatty acid; SAM, S-adenosyl methionine.

²Fixed-effects meta-analysis summary estimate from partial Spearman correlation, controlling for age at blood draw (years), BMI, number of years smoked regularly, number of cigarettes smoked per day, daily caloric intake, education (less than elementary or at least elementary), leisure-time physical activity (light to moderate or heavy), and occupational physical activity (nonworking, sedentary, light to moderate, or heavy). Statistically significant with Bonferroni correction for multiple statistical testing at P < 0.05/(1316 metabolites × 36 components) = 1.1 × 10^–6).

The majority of metabolites that were positively associated with diet quality indexes had positive correlations with whole grain and fiber, fruit, vegetable, fish, polyunsaturated fat, reducing saturated fat, and limiting empty calories components. If a metabolite was inversely associated with a particular diet index score, it was generally also inversely associated with the diet index components. There were several metabolites associated with each diet index score that were not related to any specific component at the Bonferroni-corrected α-level. Results did not differ with control for cotinine; thus, we present results without cotinine adjustment. There were no differences in diet index–metabolite associations between cases and controls.

Pathway analysis

There were 7 metabolic subpathways, each containing ≥5 measured metabolites, representing the combination of the P values of all individual metabolite associations (either positive or inverse) within a specific metabolic pathway, that were associated with higher or more optimal HEI-2010 scores; 9 pathways were associated with higher aMED scores, 5 with HDI scores, and 7 with BSD scores (all P < 2.7 × 10⁻⁴) (Figure 1). The lysolipid pathway contained the largest number of metabolites associated with diet quality (n = 26–29 across diet indexes), and was associated with the aMED, HDI, and BSD scores. This pathway contained 79 measured metabolites overall. The food component and plant metabolic subpathway was associated with all 4 diet indexes. There were also unique pathways associated with individual dietary pattern scores: HEI-2010 with chemical metabolism, aMED with dicarboxylate metabolism, and BSD with benzoate metabolism.

FIGURE 1.

Metabolic subpathways and proportion of total measured metabolites within each pathway associated with diet quality indexes in 1336 men in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study. All pathways were statistically significantly associated with diet indexes at a Bonferroni significance level (P < 0.05/(46 × 4) = 2.7 × 10⁻⁴). Pathways are defined according to KEGG PATHWAY metabolism subpathways for Homo sapiens (http://www.genome.jp/kegg/pathway.html). Analyses were conducted with the use of multivariate linear regression with fixed-effects meta-analysis (Fisher’s method). The total number of metabolites measured by the Metabolon platform within each metabolic subpathway from highest to lowest were as follows: lysolipid (79); fatty acid, dicarboxylate (27); food component or plant (25); tocopherol (25); tryptophan (25); benzoate (18); chemical (16); carnitine (15); urea cycle (14); histidine (12); drug (11); pentose (10); essential PUFAs (7); long-chain PUFAs (7).

DISCUSSION

Our findings highlight circulating metabolites and metabolic pathways associated with consuming a high quality diet, and identify a suite of potential candidate biomarkers of dietary patterns that could serve to augment the measurement of diet by traditional error-prone questionnaires (3, 58). Some metabolites were positively associated with multiple diet quality indexes, including those correlated with fruit, vegetables, whole grains, fish, and unsaturated fat, although each index had a unique profile.

We replicated previous findings for metabolite associations with fruit, dairy, meat, fish and seafood, and nut intake (29, 59–63) (Supplemental Table 12). Several metabolites have previously been identified as dietary biomarkers. For example stachydrine (otherwise known as proline betaine) is a validated citrus biomarker (64), and 2-aminophenol sulfate has been identified as a rye biomarker in feeding studies (65, 66). Observations that, to our knowledge, are potentially novel for individual food component–metabolite associations include N-δ-acetylornithine (involved in arginine synthesis) as a marker of vegetable intake, oxalic acid (high in brassica vegetables) and fruit and vegetables, homostachydrine and whole grains, and myristoyl-sphingomyelin and galactonate as markers of dairy intake (Supplemental Table 13).

Each diet index had a metabolite profile that reflected the underlying components used to score adherence. The food-based diet indexes (HEI-2010, aMED, and BSD) were associated with metabolites related to most underlying diet components. However, most macro- and micronutrients were not as strongly correlated with circulating metabolites as absolute intake levels of whole foods. For example, the nutrient-based HDI metabolite profile was not associated with most macronutrient criteria (percentage of saturated fat, sugars, or protein) or restricting absolute sodium and cholesterol. These differences could be attributable to stronger correlations between food exposures and metabolites, or differences in self-reported food compared with nutrient measurement error properties. Urinary (as opposed to serum) biomarkers might better characterize intake of some of these nutrients (i.e., sucrose and fructose for sugars, and nitrogen for protein and sodium) (60, 67–69).

The magnitudes of diet index–metabolite correlations were modest, but similar to previously reported findings (60). Habitual dietary intake may be less strongly associated with circulating metabolites than recent intake (e.g., feeding studies). However, biomarkers identified in epidemiologic studies (e.g., cohorts or case-control) tend to be more sensitive and robust, because they are captured despite metabolite degradation over time during sample storage, they reflect regular intake or have longer half-lives, and they also tend to be more specific (70). The strength of dietary pattern–metabolite associations may also be influenced by component categorization. For example, fruit-related metabolites were associated with the HEI-2010, aMED, and BSD, but not the HDI, which combined fruit and vegetables into a single component. In the context of dietary biomarker discovery, our magnitudes of correlation are similar to those from the Observing Protein and Energy Nutrition Study, which found raw correlations between protein intake measured by gold-standard urinary nitrogen and protein self-reported by FFQ of 0.33 and 0.22 for men and women, respectively (69). Urinary nitrogen has been used to calibrate for dietary measurement error (71).

Pathway analysis provides greater power to detect important metabolic pathway associations with diet quality. The strongest associations of metabolic pathways with diet quality included lysolipid and food and plant xenobiotic. The lysolipid pathway, which included the largest number of measured metabolites overall, includes lysophospholipids involved in cell signaling, energy metabolism, and membrane integrity and stability. Metabolites within this pathway were predominantly associated with the ratio of unsaturated to saturated fat. Food and plant xenobiotic metabolites (25 measured in total) represent exogenous food constituents involved in many biological functions, such as inflammation; cell proliferation, oxidation, and signaling; energy metabolism; cell membrane integrity; neuroprotection; fatty acid transport; and nutrient metabolism (72). Food and plant xenobiotic metabolites were associated with most major food groups (e.g., vegetables, fruit, seafood and protein, and whole grains). Importantly, pathway associations with diet quality that were not observed when looking at single metabolite–diet index associations became evident. For example, pentose metabolism, but not individual metabolites involved in pentose metabolism (e.g., fruit-related metabolites threitol and xylonate), and histidine metabolism, but not individual histidine metabolites (e.g., protein and seafood–related 3-methylhistidine) were statistically significantly associated with HEI-2010.

Identifying metabolite profiles of healthy eating patterns could help to elucidate the mechanisms that drive the numerous health benefits observed with consuming high-quality diets when applied in a prospective context. Furthermore, metabolites associated with diet index components could be developed as dietary biomarkers of those components through replication in other large cohorts, validation in feedings studies, and evaluation of the food metabolome of diet exposures of interest. The dietary biomarker classification (i.e., recovery, concentration, replacement, or predictive biomarkers) would depend on the metabolite properties (29, 67, 73). Previous epidemiologic studies have identified metabolite profiles of predominantly data-driven (3, 74–76), or geography-defined dietary patterns (77, 78). Some predefined dietary patterns have been explored for their effects on metabolite concentrations. We previously identified 5 serum metabolites associated with the HEI-2010 from a nested case-control study within the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial. These included an inverse association with γ-tocopherol (vitamin E) and positive associations with threonate (a vitamin C metabolite) and pyridoxate (vitamin B-6) (60). Threonate was associated with HEI-2010 in the current analysis. We did not observe an inverse association with γ-tocopherol, which could reflect population differences in dietary intake. For example, these 2 populations had a very different intake of vitamin E supplements at baseline (79).

A Mediterranean diet was compared with a low fat diet for effects on the urinary metabolome in the Prevención con Dieta Mediterránea trial; metabolites were measured with the use of nuclear magnetic resonance spectroscopy. Metabolites associated with the Mediterranean diet included carbohydrates, amino acids, lipids [e.g., cis-9-octadecenoic (oleic) and 1,8-octanedioate (octanedioic) acids], and microbial cometabolites (80). We did not observe the same associations, possibly because of minimal overlap in metabolites identifiable by nuclear magnetic resonance compared with mass spectrometry. Olive oil consumption (high in oleic acid) in Finland was also extremely low in 1985. Another study randomly assigned participants to a 6-mo Mediterranean diet or control. Plasma metabolites measured by liquid chromatography and mass spectrometry that discriminated the Mediterranean diet included glycerophospholipids and lysophosphatidylcholines, similar to our observation of a major role for lipid species (81). The New Nordic Diet was discriminated from the Average Danish Diet by urine metabolites related to fish (trimethylamine N-oxide), polyphenol-containing foods (e.g., hippurate), fruits and vegetables (3,4,5,6-tetrahydrohippurate), and wheat (arbutin) in a parallel intervention study (82).

Our study has a number of strengths that build on previous metabolomics studies, including a larger sample size, greater number of identified metabolites, strict control for multiple comparisons, and use of fasting blood samples. We compared predefined diet indexes, which are easily reproduced and translated into public health messages, in contrast to data-driven approaches (83). Our results, based on habitual dietary intake, represent real-world conditions, as opposed to data obtained from feeding studies. Our FFQ included >270 food items and was designed specifically for Finnish participants; however, because all self-reported dietary intake data contain measurement error, the diet-metabolite associations we observed are likely attenuated (34). The homogenous Finnish male smoking population may limit generalizability, although some diet-metabolite associations we observed are similar to those of other populations that include women and nonsmokers (60). Finally, we cannot exclude the possibility of residual confounding because of unmeasured confounders related to lifestyle habits.

In conclusion, we found that diet quality indexes representing healthy dietary patterns are associated with serum metabolite concentrations. Several metabolites are commonly associated with different definitions of a healthy dietary pattern. However, the overall metabolite profiles were related to the diet index components, reflecting different underlying constructs. These findings may inform the definition and construction of future diet quality indexes that will have beneficial metabolic impact. A replication of our findings in diverse populations and quantification of the dose-response of metabolite concentrations to dietary patterns in feeding studies will inform their application to reduce measurement error in dietary assessment. Our findings provide support for hypothesis generation, e.g., which diet quality–related metabolic intermediates may drive observed associations between dietary patterns, such as polyunsaturated fats, and disease.