Dairy Intake in 2 American Adult Cohorts Associates with Novel and Known Targeted and Nontargeted Circulating Metabolites

ABSTRACT

Background

The role of dairy in health can be elucidated by investigating circulating metabolites associated with intake.

Objectives

We sought to identify metabolites associated with quantity and type of dairy intake in the Framingham Heart Study Offspring and Third Generation (Gen3) cohorts.

Methods

Dairy intake (total dairy, milk, cheese, yogurt, and cream/butter) was analyzed in relation to targeted (Offspring, n = 2205, 55.1 ± 9.8 y, 52% female, 217 signals; Gen3, n = 866, 40.5 ± 8.8 y, 54.9% female, 79 signals) and nontargeted metabolites (Gen3, ∼7031 signals) in a 2-step analysis including orthogonal projections to latent structures with discriminant analysis (OPLS-DA) in discovery subsets to identify metabolites distinguishing between high and low intake; and linear regression in confirmation subsets to assess putative associations, subsequently tested in the total samples. Previously reported associations were also investigated.

Results

OPLS-DA in the Offspring targeted discovery subset resulted in a variable importance in projection (VIP) >1 of 65, 60, 58, 66, and 60 metabolites for total dairy, milk, cream/butter, cheese, and yogurt, respectively, of which 5, 3, 1, 6, and 4 metabolites, respectively, remained after confirmation. In the Gen3 targeted discovery subset, OPLS-DA resulted in a VIP >1 of 17, 15, 13, 7, and 6 metabolites for total dairy, milk, cream/butter, cheese, and yogurt, respectively. In the Gen3 nontargeted discovery subset, OPLS-DA resulted in a VIP >2 of 203, 503, 78, 186, and 206 metabolites, respectively. Combining targeted and nontargeted results in Gen3, significant associations of 7 (6 unannotated), 2, 12 (11 unannotated), 0, and 61 (all unannotated) metabolites, respectively, remained. Candidate identities of unannotated signals included fatty acids and food flavorings. Results supported relations previously reported for C14:0 sphingomyelin, and marginal associations for deoxycholates.

Conclusions

Dairy in 2 American adult cohorts associated with numerous circulating metabolites. Reports about diet-metabolite relations and confirmation of previous findings might be limited by specificity of dietary intake and breadth of measured metabolites.

Introduction

Metabolomics, the study of small molecule metabolites, has been used to distinguish between differing dietary patterns and to identify biomarkers of dietary intake, including dairy (1). To date, most observational studies of metabolomics and dairy intake have used only a limited number of metabolites or targeted metabolomics data (2–6), thus possibly biasing results by circumscribing the possible set of metabolites to most typically several hundred known signals. In contrast, nontargeted metabolomics approaches in dietary studies offer a unique opportunity to characterize more comprehensive metabolic activity of individuals in response to habitual consumption patterns (7, 8).

Although trials have sought to comprehensively evaluate the metabolomic effects of dairy products or their constituents on the serum (or other tissue/fluid) metabolome, the dairy products have been consumed in relative isolation (e.g., a specific cheese alone), rather than assessed alongside a variety of dairy products within the totality of dietary intake (9–18). Although that approach is useful for identifying biomarkers of intake specific to particular dairy products, it is less informative at the population level, where individuals might consume many different types of dairy on any given day. In addition, systemic byproducts of dairy metabolism depend on a number of factors that can limit immediate relevance of trial data to population health. Alongside the underlying disease/health status of the population studied, such factors can include the matrix in which dairy is consumed (e.g., cheese compared with milk), alongside fat and protein content (19, 20) and fermentation (13).

In the present study, we analyzed targeted and nontargeted circulating metabolomic and dietary data from the Framingham Heart Study Offspring and Third Generation (Gen3) cohorts, to identify metabolomic signatures related to the quantity and type of dairy intake. In particular, the nontargeted data, comprised of >7000 signals, enabled a relatively agnostic search for metabolites associated with dairy consumption. Investigating metabolomic signals of dairy intake could help elucidate the complex and conflicting role of dairy in human health; differentiating between signals associated with various types of dairy in environments of heterogeneous consumption could help point to specific beneficial or harmful effects and pathways of certain products; and, in the context of population-based studies with varied diets, it could facilitate the validity of existing and future diet history methodologies.

Methods

Participants

The National Heart, Lung, and Blood Institute's (NHLBI's) Framingham Heart Study (FHS) Offspring and Gen3 cohorts are community-based, longitudinal studies of cardiovascular disease that began in 1971 (21) and 2002 (22), respectively. In the fifth examination (1991–1995; mean age 55 y) of the Offspring cohort, and the first examination (2002–2005; mean age 40 y) of the Gen3 cohort, 3799 and 4095 participants, respectively, underwent a standard medical examination, consisting of laboratory and anthropometric assessments, as well as dietary intake assessment. Of these, 3712 Offspring participants’ and 4078 Gen3 participants’ data were available from the Database of Genotypes and Phenotypes (dbGaP)/National Center for Biotechnology Information, National Library of Medicine (NCBI/NLM) website (23, 24). Plasma samples from a subset of participants from each cohort underwent metabolomics analyses, conducted by the Broad Institute of the Massachusetts Institute of Technology and Harvard University Metabolomics Platform (Cambridge, MA). We included participants with ≥1 metabolite value (Offspring n = 2468, Gen3 n = 987). We excluded those with missing or invalid dietary data (Offspring n = 228, Gen3 n = 111), who were nonfasting (Offspring n = 19, Gen3 n = 8), or missing covariates (Offspring n = 16, Gen3 n = 2). Thus, ≤2205 Offspring and ≤866 Gen3 participants were included in the primary analysis, although the final number varied per metabolite (Supplemental Figures 1and 2).

The original data collection protocols were approved by the Institutional Review Board at Boston University Medical Center, and written informed consent was obtained from all participants. The present study protocol was reviewed by the Tufts University Health Sciences Institutional Review Board. All data presented in the current publication are based on the use of study data obtained with permission (controlled access) downloaded from dbGaP, under study accession phs000007.v28.p10 (25–29).

Dietary data

The Harvard semiquantitative 126-item FFQ (30) was used to assess dietary intake at Offspring examination 5 and Gen3 examination 1. The FFQ included a list of foods for which participants were asked to report frequency of consumption of standard serving sizes of each food item over the previous year. Assessment of dairy intake included 11 questions on products made from milk, specifically whole milk, low-fat/skim milk, cream, sour cream, sherbet/ice milk, ice cream, yogurt, cottage/ricotta cheese, cream cheese, other cheese (e.g., American, cheddar), and butter. Possible responses of frequency of consumption ranged from never/<1 time/mo to ≥6 times/d. Total dairy (servings per week) was calculated as the sum of each of the relevant individual line items that corresponded to the USDA “MyPlate” definition of dairy as “foods made from milk that retain their calcium content” including milk, sherbet/ice milk, ice cream, yogurt, cottage/ricotta cheese, and other cheese (31). Total fluid milk was estimated as the sum of full-fat and skim/low-fat milk; total cheese as the sum of cottage/ricotta cheese and other cheese. We also assessed foods made from milk that do not retain calcium, including cream, sour cream, cream cheese, and butter, which were collapsed into a “cream/butter” category. Conversion of standard serving sizes to grams is as follows: skim milk: 245 g; whole milk: 245 g; cream: 15 g; sour cream: 12 g; sherbet/ice milk: 96 g; ice cream: 66 g; yogurt: 227 g; cottage/ricotta cheese: 105 g; cream cheese: 28 g; other cheese: 28 g; butter: 5 g. Invalid FFQs were defined as those that estimated daily caloric intake as <600 kcal/d, or ≥4000 kcal/d for women, ≥4200 kcal/d for men, or those that had ≥12 blank items (30). The relative validity of the FFQ for dairy intake has been previously reported, and shows reasonable correlation with estimates from dietary records (e.g., highest for yogurt, r = 0.94–0.97; lowest for cheese, r = 0.38–0.57) (32). Dairy intake was adjusted for total energy intake using the residual method (33, 34). Quartile categories of intake were generated for total dairy, as well as for milk, cheese, yogurt, and cream/butter. Other dietary factors derived from the FFQ included intakes of fruit, vegetables, meat, fish, and nuts (all servings per week).

Metabolite data

Details on the analyte protocols have been previously described (35–37) and are available at dbGaP (25–29), briefly summarized here. In the Offspring, metabolite profiling in positive-ion mode using the hydrophilic interaction liquid chromatography (HILIC) and lipid LC-MS methods was conducted on plasma from a subset of participants (35, 37) who attended examination 5 (n = 1057) using LC/MS (Shimadzu Nexera X2 U-HPLC; Shimadzu Corp.) coupled to an Exactive Plus orbitrap mass spectrometer (Thermo Fisher Scientific). Metabolite profiling in the negative-ion mode using the HILIC LC-MS methods was conducted using a 4000 QTRAP triple-quadrupole mass spectrometer (AB SCIEX) equipped with an HTS PAL autosampler (Leap Technologies) and 1200 Series binary HPLC pump (Agilent). Among the 217 targeted analytes were 54 amino acids, biogenic amines, and other polar plasma metabolites, as well as 104 fatty acids including cholesterol esters (CEs), diacylglycerols, lysophosphatidylcholines, lysophosphatidylethanolamines, phosphatidylcholines (PCs), sphingomyelins (SMs), and triacylglycerols (TAGs) (Supplemental Table 1). Lipids were denoted by headgroup, total acyl carbon content, and total acyl double bond content.

In Gen3, both targeted and nontargeted analyses were performed on plasma samples from ≤998 participants of the first examination using similar methodology to that described above. Targeted metabolomics was used to identify 75 negatively charged polar metabolites using LC/MS-MS with HILIC amide chromatography (29) (Supplemental Table 1). For supervised, targeted extraction of the data, the following software was used: MultiQuant software (v1.1; Applied Biosystems/Sciex) for negative-ion mode, and TraceFinder software (Thermo Fisher Scientific) for positive-ion mode. Progenesis QI (Nonlinear Dynamics) was used for automated peak integration, and metabolite peaks were manually reviewed for quality of integration and compared against a known standard to confirm identity (28). For all metabolite data, unitless values represent the plasma relative abundance normalized to pooled plasma and then an internal standard. Nontargeted analyses included ≤7030 metabolites per participant, also analyzed by LC/MS, of which 115 were known, and the remaining unknown (labeled X-1118, X-2288, etc., identified by m/z ratio).

Clinical and lifestyle covariates

Clinical and lifestyle factors were assessed by in-person interview and examination in both cohorts, and included age (years), sex (male/female), current smoking status (yes/no), BMI (kg/m²), and history of cardiovascular disease and diabetes (both yes/no).

Statistical analysis

A flowchart of analyses is provided in Supplemental Figure 3. We natural log–transformed metabolite values to reduce skewness. To reduce signal-to-noise in the metabolite data given the nonrandomized nature of the exposure, we residual-adjusted log-transformed metabolite values for age, sex, BMI, smoking status, and presence of and/or treatment for diabetes or cardiovascular disease, analysis batch, and analysis plate (as applicable), to remove the variability in metabolite values associated with these characteristics. We then randomly subset the data in each cohort into discovery and confirmation sets in a 2:1 distribution. In the discovery set, the adjusted log-transformed metabolite values were used in orthogonal projections to latent structures discriminant analysis (OPLS-DA) with Pareto scaling (38) (SIMCA software with Omics Skin, v15.0.2; Umetrics, now Sartorius Stedim Data Analytics, AB), to identify the metabolites that could discriminate between participants in the extreme quartile categories of total dairy, milk, cheese, yogurt, and cream/butter intake. Metabolite values missing in >20% of the sample, or participants who were missing >20% of metabolite values were excluded. Values missing at <20% were interpolated by the software using a least-squares fit over successive iterations. We then identified metabolite values with variable importance in projection (VIP) scores >1 (targeted) and >2 (nontargeted), the latter VIP score selected to prospectively reduce the number of important metabolites (of potentially >7000) carried into further testing (39). VIP scores rank each of the individual features (metabolites) according to their contribution to the ability to discriminate different classes (extreme quartile categories of intake) within the data, thus indicating the most important variables in a given model (39).

Next, the log-transformed (unadjusted) metabolites were regressed on dairy intake in the confirmation subsets (SAS v9.4; SAS Institute). Regressions were adjusted for age (years), sex (male/female), BMI (kg/m²), history of diabetes (yes/no), history of heart disease (yes/no), total energy intake (kilocalories per day), batch, and plate, as applicable (model 1). Model 2 was further adjusted for other dairy types as applicable (e.g., in analyses of milk, the model included yogurt, cream/butter, and cheese, all in servings per week) and other dietary covariates including intake of fruit, vegetables, meat, fish, and nuts (all servings per week).

Those metabolites that remained statistically significant at a nominal α < 0.05 in model 2 in the confirmation subset regressions were carried into analyses in the full samples to maximize sample size. In the full samples, linear models were used to generate least-squares adjusted means of log-transformed metabolite values and assess trends across quartile categories of intake. The model was adjusted as for model 2, above. Although metabolite lists were not identical between the cohorts, when a significantly associated metabolite in one cohort was also measured/available in the other cohort, we sought to replicate the association in the other cohort as well.

In the case of significant unknown/unannotated signals from the nontargeted data set (i.e., Gen3), we searched the Human Metabolome Database (v4.0) (40) and/or METLIN (41) by m/z ratio with positive-ion mode, initially with primary adduct M + H, molecular weight tolerance ±5 ppm, to putatively annotate the unannotated signals with ≤5 candidate molecular identities based on hypothesized food or metabolic origins (42). In the case of no initial hit, we expanded the search to include molecular weight tolerance ±10 ppm and other adducts. No independent laboratory testing was conducted to confirm putative identities.

Finally, we assessed relations of metabolites previously reported in the literature to be associated with dairy (2, 4, 6, 8–18, 43–48) (Supplemental Table 2) where these reports overlapped with annotated metabolites available in the present data sets. These 36 metabolites were analyzed in the full samples using linear regression methods as described above.

Results

Participant characteristics of Offspring and Gen3 cohorts are provided in Table 1. Offspring participants were older, with correspondingly higher rates of type 2 diabetes and cardiovascular disease than Gen3. Offspring participants had lower overall energy intake and less total dairy intake, specifically lower cheese, cream/butter, and yogurt intake, than Gen3 participants.

TABLE 1.

Characteristics of participants in Framingham Heart Study Offspring and Third Generation (Gen3) cohorts¹

Characteristic	Offspring (n = 2205)	Gen3 (n = 866)
Age, y	55.1 ± 9.8	40.5 ± 8.8
Female, n (%)	1152 (52.2)	475 (54.9)
BMI, kg/m2	27.4 ± 4.9	26.6 ± 5.3
Current smoking, n (%)	396 (18.0)	131 (15.1)
History of T2D, n (%)	179 (8.1)	48 (5.5)
History of CVD, n (%)	743 (33.7)	8 (0.9)
Energy, kcal/d	1861 ± 615	2073 ± 672
Total dairy, servings/wk	10.2 ± 7.8	12.8 ± 9.0
Milk, servings/wk	5.4 ± 6.3	5.7 ± 7.3
Cheese, servings/wk	2.6 ± 3.1	4.5 ± 4.6
Cream/butter, servings/wk	5.2 ± 9.3	7.6 ± 9.0
Yogurt, servings/wk	0.8 ± 1.8	1.3 ± 2.1
Fish, servings/wk	2.3 ± 1.9	2.2 ± 1.9
Fruit, servings/wk	14.8 ± 10.6	14.5 ± 12.5
Vegetables, servings/wk	21.6 ± 13.7	24.3 ± 16.3
Meat, servings/wk	9.4 ± 5.6	11.1 ± 6.7
Nuts, servings/wk	2.2 ± 3.6	3.0 ± 4.2

¹Data are presented as mean ± SD for continuous variables or n (%) for categorical variables, as indicated. Conversion of standard serving sizes to grams is as follows: dairy (skim milk, 245 g; whole milk, 245 g; cream, 15 g; sour cream, 12 g; sherbet/ice milk, 96 g; ice cream, 66 g; yogurt, 227 g; cottage/ricotta cheese, 105 g; cream cheese, 28 g; other cheese, 28 g; butter, 5 g); fruit (raisins, 28 g; prunes, 117 g; bananas, 114 g; cantaloupe, 134 g; watermelon, 241 g; apples/pears, 138 g; apple juice, 186 g; oranges, 131 g; orange juice, 186 g; grapefruit, 120 g; grapefruit juice, 185 g; other fruit juice, 126 g; strawberries, 75 g; blueberries, 73 g; peaches, 128 g); vegetables (tomatoes, 123 g; tomato juice, 152 g; tomato sauce, 125 g; red chilli sauce, 16 g; string beans, 68 g; broccoli, 78 g; cabbage/coleslaw, 75 g; cauliflower, 62 g; Brussels sprouts, 78 g; carrots, raw, 36 g; carrots, cooked, 78 g; corn, 82 g; peas/Lima beans, 80 g; mixed vegetables, 91 g; beans/lentils, 131 g; winter squash, 102 g; summer squash, 90 g; yams/sweet potatoes, 100 g; spinach, cooked, 90 g; spinach, raw, 56 g; iceberg/head lettuce, 56 g; romaine/leaf lettuce, 56 g; celery, 20 g; beets, 85 g; alfalfa sprouts, 17 g; garlic, 0.05 g); meat (eggs, 50 g; chicken, 140 g; bacon, 13 g; hot dogs, 45 g; processed meats, 27 g; liver, 98 g; hamburger, sandwich, or casserole meat, 85 g; meat as main dish, 140 g); fish (canned tuna, 98 g; dark/other fish, 112 g; shrimp/shellfish, 85 g); nuts (peanut butter, 16 g; peanuts/tree nuts, 28 g). CVD, cardiovascular disease; T2D, type 2 diabetes.

In the Offspring discovery subset, OPLS-DA of lowest compared with highest quartile categories with 208 targeted, adjusted metabolites resulted in a VIP >1 of 65 metabolites for total dairy, 60 for milk, 58 for cream/butter, 66 for cheese, and 60 for yogurt. These were carried into multivariate adjusted regressions in the Offspring confirmation subset, resulting in nominally significant associations of 11 metabolites for total dairy, 4 for milk, 2 for cream/butter, 7 for cheese, and 5 for yogurt, after full adjustment (model 2) (Supplemental Table 3). Carrying these metabolites into the full Offspring sample (discovery and confirmation subsets combined), least-squares adjusted means and P-trend across quartiles of intake resulted in nominally significant associations of 5 metabolites with total dairy, 3 with milk, 1 with cream/butter, 6 with cheese, and 4 with yogurt (Table 2). Of these several remained significant after Bonferroni correction, including cis/trans-hydroxyproline, pantothenate, and uridine for total dairy and milk, suggesting the associations in total dairy were driven by milk intake; C54:4 TAG for cream/butter; C46:0 TAG, C54:3 TAG, C54:4 TAG, C54:5 TAG, C54:6 TAG for cheese; and C20:5 CE for yogurt (Table 2). Of these metabolites, 5 (hydroxyproline, niacinamide, pantothenate, taurodeoxycholate, and uridine) were also available in the Gen3 data, and were assessed against intake in the Gen3 sample. Pantothenate and uridine were statistically associated with total dairy intake in Gen3 consistent with the Offspring. However, associations of hydroxyproline with total dairy, or hydroxyproline, pantothenate, and uridine with milk, were nonsignificant in the Gen3 sample (Table 2, Supplemental Table 4).

TABLE 2.

Least-squares adjusted means ± SE of log-transformed circulating metabolite values across increasing quartiles of dairy intake in the Framingham Heart Study Offspring cohort, total sample (n = 2205)¹

		Quartile category of dairy intake
Metabolite		Q1	Q2	Q3	Q4	P-trend	Sig.2	In Gen33
Number in quartile category		551	551	552	551
Total dairy (quartile median, servings/wk)	n	2.5	6.5	10.4	20.0
C18:1 LPE	1818	0.402 ± 0.026	0.368 ± 0.024	0.365 ± 0.024	0.343 ± 0.025	0.05	N	n/a
C18:2 LPE	1818	0.043 ± 0.026	0.039 ± 0.025	0.026 ± 0.024	−0.013 ± 0.026	0.03	N	n/a
C54:3 TAG	1818	0.326 ± 0.021	0.321 ± 0.019	0.318 ± 0.019	0.302 ± 0.020	0.26	N	n/a
C56:3 TAG	1818	0.443 ± 0.031	0.440 ± 0.029	0.443 ± 0.029	0.408 ± 0.030	0.25	N	n/a
cis/trans-Hydroxyproline	2202	0.241 ± 0.023	0.195 ± 0.022	0.190 ± 0.021	0.143 ± 0.022	0.0002	Y	P = 0.44
Gentisate	1816	−0.553 ± 0.066	−0.574 ± 0.062	−0.618 ± 0.061	−0.612 ± 0.064	0.40	N	n/a
Glycodeoxycholates	1816	0.092 ± 0.056	0.020 ± 0.053	0.024 ± 0.052	−0.070 ± 0.055	0.01	N	n/a
Pantothenate	1816	0.040 ± 0.034	0.075 ± 0.032	0.125 ± 0.031	0.193 ± 0.033	<0.0001	Y	P = 0.0484
Taurodeoxycholates	1809	0.027 ± 0.067	−0.015 ± 0.063	−0.029 ± 0.062	−0.071 ± 0.065	0.18	N	P = 0.87
Uridine	1816	−0.250 ± 0.017	−0.214 ± 0.016	−0.198 ± 0.016	−0.140 ± 0.016	<0.0001	Y	P = 0.0014
Niacinamide	2202	0.195 ± 0.028	0.234 ± 0.026	0.184 ± 0.026	0.257 ± 0.028	0.07	N	P = 0.49
Milk (quartile median, servings/wk)		0.0	3.0	5.5	17.5
C46:0 TAG	1559	−0.127 ± 0.043	−0.122 ± 0.043	−0.078 ± 0.042	−0.079 ± 0.043	0.34	N	n/a
cis/trans-Hydroxyproline	2202	0.229 ± 0.022	0.186 ± 0.022	0.196 ± 0.022	0.154 ± 0.022	0.006	Y	P = 0.75
Pantothenate	1816	0.046 ± 0.033	0.059 ± 0.033	0.163 ± 0.032	0.165 ± 0.033	0.001	Y	P = 0.98
Uridine	1816	−0.246 ± 0.016	−0.222 ± 0.016	−0.196 ± 0.016	−0.148 ± 0.016	<0.0001	Y	P = 0.10
Cream/butter (quartile median, servings/wk)		0.0	0.9	3.0	13.4
C54:4 TAG	1818	0.137 ± 0.017	0.110 ± 0.017	0.095 ± 0.016	0.043 ± 0.017	<0.0001	Y	n/a
Pyridoxate	1814	0.110 ± 0.047	0.041 ± 0.045	0.170 ± 0.045	0.029 ± 0.046	0.11	N	n/a
Cheese (quartile median, servings/wk)		0.5	1.0	3.0	5.5
C14:0 LPC	1818	0.071 ± 0.027	0.103 ± 0.026	0.128 ± 0.026	0.149 ± 0.027	0.01	N	n/a
C44:1 TAG	1818	−0.521 ± 0.075	−0.445 ± 0.074	−0.402 ± 0.073	−0.294 ± 0.077	0.008	N	n/a
C46:0 TAG	1559	−0.144 ± 0.043	−0.151 ± 0.042	−0.098 ± 0.042	−0.004 ± 0.044	0.001	Y	n/a
C54:3 TAG	1818	0.341 ± 0.020	0.315 ± 0.019	0.323 ± 0.019	0.297 ± 0.020	0.10	N	n/a
C54:4 TAG	1818	0.123 ± 0.017	0.104 ± 0.017	0.096 ± 0.016	0.067 ± 0.017	0.004	Y	n/a
C54:5 TAG	1818	0.017 ± 0.022	−0.019 ± 0.021	−0.024 ± 0.021	−0.059 ± 0.022	0.005	Y	n/a
C54:6 TAG	1818	0.017 ± 0.022	−0.019 ± 0.021	−0.024 ± 0.021	−0.059 ± 0.022	0.005	Y	n/a
Yogurt (quartile median, servings/wk)		0.0	0.0	0.5	3.0
C20:5 CE	1818	−0.420 ± 0.043	−0.376 ± 0.039	−0.354 ± 0.040	−0.263 ± 0.040	0.001	Y	n/a
C20:5 LPC	1818	−0.099 ± 0.036	−0.099 ± 0.033	−0.050 ± 0.034	−0.017 ± 0.034	0.03	N	n/a
C58:11 TAG	1818	0.324 ± 0.036	0.315 ± 0.033	0.345 ± 0.034	0.392 ± 0.034	0.03	N	n/a
Pyridoxate	1814	0.022 ± 0.049	0.062 ± 0.045	0.111 ± 0.047	0.168 ± 0.046	0.01	N	n/a
Salicylurate	1814	0.804 ± 0.116	0.673 ± 0.106	0.735 ± 0.109	0.859 ± 0.109	0.17	N	n/a

¹Metabolite values are log-transformed. Model adjusted for age, sex, BMI, smoking status, history of diabetes, history of cardiovascular disease, intake of other dairy (except in total dairy; i.e., for milk, the model included cheese, cream/butter, and yogurt), and intake of fish, fruit, vegetables, meat, and nuts, as well as batch or plate (as applicable). Conversion of standard serving sizes to grams for dairy is as follows: skim milk, 245 g; whole milk, 245 g; cream, 15 g; sour cream, 12 g; sherbet/ice milk, 96 g; ice cream, 66 g; yogurt, 227 g; cottage/ricotta cheese, 105 g; cream cheese, 28 g; other cheese, 28 g; butter, 5 g. CE, cholesterol ester; Gen3, Framingham Heart Study Third Generation cohort; LPC, lysophosphatidylcholine; LPE, lysophosphatidylethanolamine; n/a, not available; Q, quartile; TAG, triacylglycerol.

²Significant after Bonferroni correction: P < 0.05/11 (0.0045) for total dairy; <0.05/4 (0.0125) for milk; <0.05/2 (0.025) for cream/butter; <0.05/7 (0.007) for cheese; <0.05/5 (0.01) for yogurt.

³Tested in identical models in Gen3 total sample, where metabolite value was measured and annotated. See also Supplemental Table 4.

⁴Association in Gen3 was directionally consistent with that in the Offspring. See also Supplemental Table 4.

In the Gen3 targeted data, discovery subset, OPLS-DA of lowest compared with highest quartile categories with 57 known, adjusted metabolites resulted in a VIP >1 of 17 metabolites for total dairy, 15 for milk, 13 for cream/butter, 7 for cheese, and 6 for yogurt. As in the Offspring, these metabolites were carried into the confirmation subset of Gen3, and regressed on dairy intake. These multivariate regressions resulted in nominally significant associations of 3 metabolites for total dairy, 5 for milk, 3 for cream/butter, 0 for cheese, and 0 for yogurt, after full adjustment (model 2) (Supplemental Table 5). In the Gen3 nontargeted data, discovery subset, OPLS-DA comparing the highest with the lowest quartile of intake, the VIP >2 signals included 203 metabolites for total dairy, 503 for milk, 78 for cream/butter, 186 for cheese, and 206 for yogurt. As in targeted analyses described above, these were carried forward to regression models in the confirmation subset, resulting in 11 significant metabolites for total dairy, 3 for milk, 11 for cream/butter, 1 for cheese, and 69 for yogurt in the fully adjusted model (Supplemental Table 6). Finally, combining the discovery and confirmation subsets and the targeted and nontargeted results in the full Gen3 sample, least squares-adjusted means and P-trend across quartiles of intake resulted in significant associations of 7 metabolites (6 unannotated) with total dairy, 2 with milk, 12 (11 unannotated) with cream/butter, 0 with cheese, and 61 with yogurt (all unannotated) (Table 3). Of the annotated signals, 7 (α-hydroxybutyrate, hippurate, oxalate, pantothenate, quinolinate, kynurenic acid, and taurine) were also available in the Offspring data. Neither oxalate with total dairy, nor α-hydroxybutyrate with milk were replicated in the Offspring; however, the association of pantothenate with cream/butter was statistically significant and directionally consistent in the Offspring (Table 3, Supplemental Table 7).

TABLE 3.

Least-squares adjusted means (SE) of log-transformed circulating metabolite values across increasing quartiles of dairy intake in the Framingham Heart Study Third Generation cohort, total sample (n = 866)¹

			Quartile category of dairy intake
Metabolite	Unannotated m/z2	n	Q1	Q2	Q3	Q4	P-trend	Sig.3	In Offspring4
Number per quartile category			216	217	217	216
Total dairy (quartile median, servings/wk)			4.5	8.5	12.5	23.4
Hippurate		865	−0.73 ± 0.19	−0.86 ± 0.18	−0.77 ± 0.19	−0.83 ± 0.19	0.48	N	P = 0.19
Kynurenic acid		865	−0.23 ± 0.10	−0.24 ± 0.10	−0.22 ± 0.10	−0.28 ± 0.10	0.25	N	P = 0.18
Oxalate		860	−0.52 ± 0.10	−0.50 ± 0.10	−0.45 ± 0.10	−0.40 ± 0.10	0.004	Y	P = 0.58
X1840	431.3395	864	8.50 ± 0.24	8.22 ± 0.24	8.31 ± 0.24	8.03 ± 0.24	0.0003	Y	n/a
X1852	496.4205	855	10.10 ± 0.28	10.08 ± 0.28	10.00 ± 0.28	10.14 ± 0.28	0.68	N	n/a
X203	244.1905	846	8.85 ± 0.16	8.83 ± 0.16	8.83 ± 0.16	8.90 ± 0.16	0.70	N	n/a
X2204	431.3388	863	9.34 ± 0.27	9.21 ± 0.27	9.35 ± 0.27	9.18 ± 0.27	0.23	N	n/a
X291	214.1787	833	5.80 ± 0.25	5.79 ± 0.24	5.85 ± 0.25	5.77 ± 0.25	0.65	N	n/a
X3341	197.0669	838	10.67 ± 0.31	10.58 ± 0.31	10.64 ± 0.31	10.10 ± 0.32	0.0002	Y	n/a
X3517	213.0981	814	10.37 ± 0.35	10.24 ± 0.34	10.38 ± 0.35	9.62 ± 0.35	0.00002	Y	n/a
X3726	207.1742	828	7.47 ± 0.22	7.39 ± 0.22	7.45 ± 0.22	7.44 ± 0.22	0.79	N	n/a
X4778	96.0449	841	9.93 ± 0.42	9.78 ± 0.41	9.67 ± 0.42	9.27 ± 0.42	0.0006	Y	n/a
X6385	94.0657	849	8.63 ± 0.47	8.49 ± 0.47	8.31 ± 0.47	7.95 ± 0.48	0.006	N	n/a
X942	402.2634	862	10.33 ± 0.13	10.50 ± 0.13	10.62 ± 0.13	10.74 ± 0.13	<0.0001	Y	n/a
Milk (quartile median, servings/wk)			0.0	3.0	5.5	17.5
α-Hydroxybutyrate		865	0.00 ± 0.15	0.06 ± 0.15	0.01 ± 0.15	−0.15 ± 0.15	0.04	N	P = 0.99
Hippurate		865	−0.87 ± 0.18	−0.77 ± 0.18	−0.68 ± 0.18	−0.89 ± 0.19	0.40	N	P = 0.47
Kynurenic acid		865	−0.23 ± 0.10	−0.24 ± 0.10	−0.24 ± 0.10	−0.28 ± 0.10	0.19	N	P = 0.21
Maleate		865	−0.14 ± 0.11	−0.11 ± 0.10	−0.25 ± 0.11	−0.24 ± 0.11	0.03	N	n/a
Quinolinate		865	−0.28 ± 0.07	−0.23 ± 0.07	−0.25 ± 0.07	−0.27 ± 0.07	0.75	N	P = 0.70
X1866	308.2218	861	9.21 ± 0.14	9.06 ± 0.13	9.13 ± 0.14	9.20 ± 0.14	0.25	N	n/a
X2920	427.3531	863	10.02 ± 0.16	9.96 ± 0.16	9.89 ± 0.16	9.88 ± 0.17	0.22	N	n/a
X3878	184.071	751	5.23 ± 0.45	5.20 ± 0.45	5.53 ± 0.46	5.10 ± 0.47	0.82	N	n/a
Cream/butter (quartile median, servings/wk)			0.5	2.9	7.0	18.0
Indole-3-carboxyl		865	−0.12 ± 0.16	−0.20 ± 0.16	−0.22 ± 0.17	−0.20 ± 0.16	0.38	N	n/a
Pantothenate		865	−0.15 ± 0.11	−0.20 ± 0.11	−0.26 ± 0.11	−0.29 ± 0.11	0.01	N	P = 0.035
Taurine		865	0.03 ± 0.23	0.14 ± 0.23	0.04 ± 0.23	0.00 ± 0.23	0.47	N	P = 0.21
X5942	510.3926	864	11.63 ± 0.10	11.70 ± 0.10	11.70 ± 0.10	11.82 ± 0.10	0.0002	Y	n/a
X1427	445.2948	766	8.16 ± 0.45	7.63 ± 0.46	7.74 ± 0.46	7.40 ± 0.46	0.001	Y	n/a
X250	96.045	704	7.79 ± 0.45	7.76 ± 0.46	7.60 ± 0.46	8.38 ± 0.45	0.001	Y	n/a
X2979	210.1488	864	10.28 ± 0.19	10.22 ± 0.19	10.41 ± 0.19	10.86 ± 0.19	<0.0001	Y	n/a
X3632	144.1019	864	9.46 ± 0.11	9.45 ± 0.11	9.47 ± 0.12	9.72 ± 0.11	<0.0001	Y	n/a
X4249	221.0643	842	10.17 ± 0.26	10.20 ± 0.27	10.58 ± 0.27	10.96 ± 0.26	<0.0001	Y	n/a
X4259	181.0719	834	9.12 ± 0.27	9.19 ± 0.28	9.45 ± 0.28	9.93 ± 0.27	<0.0001	Y	n/a
X4778	96.0449	841	9.30 ± 0.40	9.17 ± 0.40	9.64 ± 0.41	10.53 ± 0.40	<0.0001	Y	n/a
X5033	797.622	864	12.51 ± 0.04	12.55 ± 0.04	12.55 ± 0.04	12.60 ± 0.04	<0.0001	Y	n/a
X5146	692.5573	864	11.86 ± 0.09	11.99 ± 0.09	12.02 ± 0.09	12.11 ± 0.09	<0.0001	Y	n/a
X6385	94.0657	849	7.95 ± 0.45	7.71 ± 0.46	8.15 ± 0.46	9.44 ± 0.45	<0.0001	Y	n/a
Cheese (quartile median, servings/wk)			0.9	3.0	3.5	7.5
X2459	165.1257	850	6.29 ± 0.41	6.36 ± 0.41	6.36 ± 0.41	6.55 ± 0.41	0.18	N	n/a
Yogurt (quartile median, servings/wk)			0.0	0.0	1.0	3.0
X678	624.5547	865	11.72 ± 0.09	11.69 ± 0.09	11.62 ± 0.09	11.52 ± 0.09	0.00001	Y	n/a
X580	625.5589	865	10.59 ± 0.12	10.56 ± 0.12	10.48 ± 0.12	10.36 ± 0.12	0.00002	Y	n/a
X597	598.5389	864	10.63 ± 0.12	10.58 ± 0.12	10.49 ± 0.12	10.39 ± 0.12	0.00003	Y	n/a
X596	596.524	865	11.60 ± 0.10	11.55 ± 0.10	11.50 ± 0.10	11.39 ± 0.10	0.00004	Y	n/a
X671	581.511	865	12.32 ± 0.11	12.24 ± 0.11	12.19 ± 0.12	12.08 ± 0.12	0.00004	Y	n/a
X608	570.5088	864	10.37 ± 0.12	10.35 ± 0.12	10.28 ± 0.12	10.15 ± 0.12	0.00007	Y	n/a
X843	565.4456	865	11.56 ± 0.11	11.45 ± 0.12	11.43 ± 0.12	11.31 ± 0.12	0.0001	Y	n/a
X717	235.2418	865	10.01 ± 0.12	9.98 ± 0.12	9.91 ± 0.12	9.81 ± 0.12	0.0003	Y	n/a
X634	603.496	862	9.66 ± 0.16	9.55 ± 0.17	9.48 ± 0.17	9.34 ± 0.17	0.0005	Y	n/a
X889	197.19	865	9.99 ± 0.11	9.91 ± 0.11	9.87 ± 0.11	9.79 ± 0.11	0.0005	Y	n/a
X938	425.3626	865	11.18 ± 0.10	11.13 ± 0.10	11.06 ± 0.10	11.01 ± 0.10	0.0007	Y	n/a
X749	409.3675	864	10.08 ± 0.12	10.02 ± 0.12	10.04 ± 0.12	9.89 ± 0.12	0.0007	Y	n/a
X795	225.1848	865	12.64 ± 0.09	12.57 ± 0.09	12.51 ± 0.09	12.46 ± 0.09	0.0007	Y	n/a
X633	535.472	864	10.23 ± 0.14	10.15 ± 0.15	10.13 ± 0.15	9.98 ± 0.15	0.0007	N	n/a
X705	419.388	865	10.63 ± 0.11	10.58 ± 0.11	10.55 ± 0.11	10.44 ± 0.11	0.0007	N	n/a
X741	614.5347	865	12.56 ± 0.12	12.51 ± 0.12	12.44 ± 0.12	12.36 ± 0.12	0.0008	N	n/a
X695	235.2055	865	11.11 ± 0.09	11.06 ± 0.09	11.02 ± 0.09	10.95 ± 0.09	0.0008	N	n/a
X1030	553.4108	865	10.95 ± 0.12	10.87 ± 0.13	10.83 ± 0.13	10.73 ± 0.13	0.0008	N	n/a
X904	243.1954	864	10.00 ± 0.13	9.92 ± 0.13	9.88 ± 0.13	9.77 ± 0.13	0.001	N	n/a
X673	563.5005	865	11.29 ± 0.11	11.24 ± 0.11	11.17 ± 0.11	11.10 ± 0.11	0.001	N	n/a
X1063	345.2784	863	9.20 ± 0.12	9.15 ± 0.12	9.10 ± 0.12	9.00 ± 0.12	0.001	N	n/a
X899	393.3726	865	10.57 ± 0.12	10.49 ± 0.12	10.49 ± 0.12	10.37 ± 0.12	0.001	N	n/a
X802	533.4567	865	12.66 ± 0.09	12.60 ± 0.09	12.56 ± 0.09	12.50 ± 0.09	0.001	N	n/a
X1074	524.3946	865	10.93 ± 0.12	10.90 ± 0.12	10.85 ± 0.12	10.75 ± 0.12	0.002	N	n/a
X650	309.2786	862	8.59 ± 0.16	8.54 ± 0.16	8.48 ± 0.16	8.34 ± 0.16	0.002	N	n/a
X712	567.4624	865	11.22 ± 0.10	11.14 ± 0.10	11.10 ± 0.10	11.04 ± 0.10	0.002	N	n/a
X698	255.2317	865	10.70 ± 0.09	10.65 ± 0.09	10.62 ± 0.09	10.56 ± 0.09	0.002	N	n/a
X709	551.467	865	12.75 ± 0.10	12.67 ± 0.10	12.62 ± 0.10	12.58 ± 0.10	0.002	N	n/a
X1069	526.4104	865	12.19 ± 0.09	12.14 ± 0.09	12.11 ± 0.09	12.04 ± 0.09	0.003	N	n/a
X855	541.4474	865	11.74 ± 0.10	11.66 ± 0.10	11.63 ± 0.10	11.56 ± 0.10	0.003	N	n/a
X477	680.58	853	9.73 ± 0.31	9.43 ± 0.31	9.53 ± 0.32	9.17 ± 0.32	0.003	N	n/a
X1012	321.2786	865	9.86 ± 0.13	9.78 ± 0.13	9.74 ± 0.13	9.65 ± 0.13	0.003	N	n/a
X701	237.2209	865	10.14 ± 0.10	10.11 ± 0.10	10.08 ± 0.10	10.00 ± 0.10	0.004	N	n/a
X807	453.3566	865	12.00 ± 0.08	11.93 ± 0.08	11.93 ± 0.08	11.87 ± 0.08	0.004	N	n/a
X3096	347.2214	830	9.43 ± 0.49	9.24 ± 0.50	9.61 ± 0.50	9.99 ± 0.51	0.005	N	n/a
X811	377.3411	865	10.76 ± 0.09	10.71 ± 0.09	10.70 ± 0.09	10.63 ± 0.09	0.005	N	n/a
X431	786.6593	830	9.90 ± 0.33	9.62 ± 0.34	9.72 ± 0.34	9.34 ± 0.34	0.005	N	n/a
X516	654.5658	859	9.96 ± 0.29	9.70 ± 0.29	9.76 ± 0.30	9.50 ± 0.30	0.006	N	n/a
X732	153.1273	865	11.57 ± 0.09	11.53 ± 0.09	11.48 ± 0.09	11.44 ± 0.09	0.006	N	n/a
X1156	549.377	864	11.24 ± 0.12	11.17 ± 0.12	11.12 ± 0.13	11.06 ± 0.13	0.006	N	n/a
X454	714.6187	865	11.72 ± 0.16	11.85 ± 0.16	11.90 ± 0.16	11.96 ± 0.16	0.007	N	n/a
X1005	363.289	864	9.82 ± 0.11	9.74 ± 0.11	9.75 ± 0.11	9.65 ± 0.11	0.008	N	n/a
X1054	486.4156	865	10.92 ± 0.11	10.89 ± 0.12	10.80 ± 0.12	10.77 ± 0.12	0.008	N	n/a
X3616	433.3165	843	8.43 ± 0.43	8.29 ± 0.43	8.35 ± 0.44	7.93 ± 0.44	0.009	N	n/a
X1427	445.2948	766	7.58 ± 0.45	7.56 ± 0.46	8.01 ± 0.46	8.11 ± 0.46	0.009	N	n/a
X1194	547.3607	863	10.80 ± 0.13	10.73 ± 0.13	10.67 ± 0.13	10.62 ± 0.13	0.01	N	n/a
X1017	485.3844	865	11.75 ± 0.09	11.67 ± 0.09	11.64 ± 0.09	11.61 ± 0.09	0.01	N	n/a
X690	517.4612	860	9.76 ± 0.14	9.71 ± 0.15	9.66 ± 0.15	9.59 ± 0.15	0.01	N	n/a
X1028	479.3357	865	12.01 ± 0.08	11.97 ± 0.08	11.94 ± 0.08	11.90 ± 0.08	0.01	N	n/a
X579	584.5247	862	10.51 ± 0.22	10.47 ± 0.23	10.36 ± 0.23	10.24 ± 0.23	0.01	N	n/a
X956	517.461	856	9.24 ± 0.18	9.24 ± 0.18	9.18 ± 0.18	9.06 ± 0.18	0.01	N	n/a
X1137	551.3942	865	10.84 ± 0.14	10.76 ± 0.14	10.72 ± 0.14	10.65 ± 0.14	0.01	N	n/a
X1135	263.1641	864	9.89 ± 0.11	9.89 ± 0.12	9.81 ± 0.12	9.77 ± 0.12	0.02	N	n/a
X573	589.4816	864	10.36 ± 0.18	10.33 ± 0.18	10.35 ± 0.18	10.17 ± 0.18	0.02	N	n/a
X1146	267.1953	865	11.16 ± 0.08	11.10 ± 0.08	11.07 ± 0.08	11.05 ± 0.08	0.02	N	n/a
X1174	545.3449	861	10.08 ± 0.13	9.99 ± 0.13	9.99 ± 0.14	9.91 ± 0.14	0.02	N	n/a
X414	760.6424	797	10.04 ± 0.37	9.75 ± 0.37	9.85 ± 0.38	9.54 ± 0.38	0.03	N	n/a
X1181	517.3137	792	8.29 ± 0.22	8.24 ± 0.23	8.08 ± 0.23	8.07 ± 0.23	0.03	N	n/a
X1255	521.345	864	10.80 ± 0.12	10.74 ± 0.12	10.71 ± 0.13	10.66 ± 0.13	0.03	N	n/a
X886	498.3789	865	12.07 ± 0.09	12.01 ± 0.09	11.98 ± 0.09	11.96 ± 0.09	0.04	N	n/a
X1131	695.5724	791	9.17 ± 0.30	9.49 ± 0.30	9.47 ± 0.30	9.55 ± 0.30	0.04	N	n/a
X3995	605.55	772	9.86 ± 0.27	9.82 ± 0.27	9.82 ± 0.27	9.62 ± 0.27	0.046	N	n/a
X1222	523.3613	860	10.46 ± 0.14	10.40 ± 0.14	10.38 ± 0.14	10.32 ± 0.14	0.05	N	n/a
X1133	249.1848	865	12.67 ± 0.09	12.62 ± 0.09	12.56 ± 0.09	12.56 ± 0.09	0.05	N	n/a
X440	734.6276	828	10.05 ± 0.42	9.76 ± 0.42	9.93 ± 0.42	9.57 ± 0.43	0.05	N	n/a
X1158	475.3034	864	10.37 ± 0.13	10.32 ± 0.13	10.26 ± 0.13	10.24 ± 0.14	0.07	N	n/a
X1271	519.3294	852	9.86 ± 0.16	9.75 ± 0.16	9.74 ± 0.16	9.68 ± 0.16	0.08	N	n/a
X423	774.6596	799	9.04 ± 0.41	8.69 ± 0.42	8.64 ± 0.42	8.58 ± 0.42	0.21	N	n/a
X1272	231.1741	858	8.87 ± 0.16	8.81 ± 0.16	8.69 ± 0.16	8.75 ± 0.16	0.26	N	n/a

¹Metabolite values are log-transformed. Model adjusted for age, sex, BMI, smoking status, history of diabetes, history of cardiovascular disease, intake of other dairy (except in total dairy; i.e., for milk, the model included cheese, cream/butter, and yogurt), and intake of fish, fruit, vegetables, meat, and nuts, as well as batch or plate (as applicable). Conversion of standard serving sizes to grams for dairy is as follows: skim milk, 245 g; whole milk, 245 g; cream, 15 g; sour cream, 12 g; sherbet/ice milk, 96 g; ice cream, 66 g; yogurt, 227 g; cottage/ricotta cheese, 105 g; cream cheese, 28 g; other cheese, 28 g; butter, 5 g. Gen3, Framingham Heart Study Third Generation cohort; n/a, not available; Q, quartile.

²Unannotated/unknown metabolites were searched by m/z ratio in the Human Metabolome Database and/or METLIN (41). The m/z ratio and putative identities are provided in Supplemental Table 8. Identities have not been verified.

³Significant after Bonferroni correction: P < 0.05/14 (0.004) for total dairy; <0.05/8 (0.006) for milk; <0.05/14 (0.004) for cream/butter; <0.05/69 (0.0007) for yogurt.

⁴Tested in identical models in the Offspring total sample, where metabolite value was measured and annotated.

⁵Association in the Offspring was directionally consistent with that in Gen3.

Candidate identities of unannotated signals in Gen3 were putatively annotated based on biological plausibility, using the m/z ratio searched via the Human Metabolome Database and METLIN. Putative identities included primarily fatty acids (e.g., di- and triacylglycerols) as well as food flavorings and additives, particularly for yogurt, which had yielded the greatest number of unannotated signals (Supplemental Table 8).

Finally, we assessed relations of 36 metabolites previously reported in the literature (Supplemental Table 2) where the annotated equivalents were available in either or both of the Framingham cohorts. The strongest relations (all P-trend < 0.0001) were observed for C14:0 SM, which was higher with higher intake of total dairy and all products in the Offspring (unavailable in Gen3). Marginal associations (uncorrected for multiple testing) were observed for deoxycholates and total dairy and milk in both cohorts. Threonine trended toward associations with all dairy except yogurt in both cohorts. Proline trended toward associations with total dairy and milk in the Offspring but not in Gen3; aspartate with yogurt in Gen3 but not in the Offspring; and leucine with total dairy and milk in Gen3 but not in the Offspring. In addition, in Gen3, indole-3-lactate was marginally associated with total dairy and milk, and putrescine with cheese and yogurt (both unavailable in the Offspring) (Supplemental Table 9).

Discussion

In this study, we evaluated ≤7000 serum metabolite features in ≤3071 participants of the Framingham Heart Study cohorts. We used a data reduction method designed to discriminate between high and low intake, coupled with attempted confirmation via linear regression, to identify the most distinguishing markers of dairy/dairy-type intake. We observed 92 unannotated metabolites, many biologically plausible, as well as several long-chain fatty acids, pantothenate, and uridine, to be associated with various dairy types. We were able to tentatively replicate several associations in the opposite cohort. In addition, we tested 36 dairy-metabolite associations previously reported in the literature, but were able to confirm relatively few. To our knowledge, this is the largest observational study to examine targeted and nontargeted metabolomics data in relation to habitual dairy product consumption, as well as specific dairy types.

Most observational studies of dairy and metabolomics to date have either focused on targeted markers, notably fatty acids, or used relatively small metabolite sets relative to thousands of measurable circulating metabolites (2, 4–6, 8–10, 43, 49, 50). Several studies on food groups (including dairy) have emerged from a series of nested case-control studies in the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial. One such study involved 502 participants, investigating 412 known and 231 unknown metabolites against 36 food groups, and reported associations of butter with methylpalmitate (15 or 2), pentadecanoate (15:0), and 10-undecenoate (11:1n–1), and milk with homostachydrine (2). A second nested case-control study, in female participants, evaluated correlations of 617 known metabolites with 55 food groups, reporting associations of cheese with myristoyl SM (SM d18:1/14:0) and 10-undecenoate (11:1n–1); ice cream with 10-undecenoate (11:1n–1); milk with galactonate, lactose, phenylacetylglycine, homostachydrine, hydantoin-5-propionic acid, orotate, and 2-aminoheptanoate; yogurt with δ-tocopherol and γ-tocopherol; and butter with 10-undecenoate (11:1n–1), caprate (10:0), and 3-hydroxyoctanoate (4). Although these metabolites were not universally available in our data sets, C14:0 SM and several other SMs in the Offspring were associated with dairy intake. In Gen3, similar to homostachydrine, several signals plausibly related to metabolites of coffee and cocoa associated with cream/butter and total dairy intake, as were plausibly short- and medium-chain fatty acids with yogurt, although these would need verification.

The largest observational study prior to the present one, conducted in the Cancer Prevention Study-II Nutrition Cohort, used untargeted metabolomics (1186 serum metabolites) to evaluate 91 food groups, including dairy foods, in 1369 postmenopausal women (6). Wang and colleagues (6) reported partialled correlations between milk and galactonate, 2,8-quinolinediol sulfate, phenylacetylglycine, as well as 2 unannotated signals; between yogurt and 1 unannotated signal; and between butter and caprate (10:0), caprylate (8:0), 10-undecenoate (11:1n–1), SM (d18:1/25:0, d19:0/24:1, d20:1/23:0, d19:1/24:0), SM (d17:1/16:0, d18:1/15:0, d16:1/17:0), and 2 unannotated signals. In the present study, we did not assess correlations, but used a data reduction method designed to discriminate between high and low intake to identify the most distinguishing markers of dairy intake. In this, we observed 92 unannotated metabolites, many biologically plausible, to be associated with various dairy types, and were also able to test several of the associations reported by Wang et al. in 1 or both cohorts. However, of their results, only the SM associations were confirmed.

Where larger-scale nontargeted metabolomics has been more widely used, it has been in controlled feeding trials. Several high-quality but short-term feeding trials have investigated nontargeted circulating (serum) metabolites in response to dairy or dairy protein consumption in generally healthy populations (9, 10, 13). A recent study comparing nontargeted serum metabolomic effects (2302 unique signals) of 2 wk of 400 g/d of either nonfermented milk or a probiotic yogurt in 14 healthy young men, found both postprandial and 2-wk differences between the products (13). Postprandial and/or 2-wk changes were reported for amino acids (proline, lysine, threonine, phenylalanine, asparagine, tyrosine, tryptophan, citrulline, taurine), bile acids [gluconic acid, δ-gluconolactone, (cheno)deoxycholic acid, glycoursodeoxycholic acid, glycocholic acid, tauroursodeoxycholic acid, 3β-hydroxy-5-cholenoic acid], and indole derivatives (indole-3-lactic acid, indole-3-acetaldehyde, indole-3-acetic acid, 3-indole propionic acid). The authors noted that over one-third of the changes observed in the fasting serum after 2 wk of intake were readily observed in the postprandial tests, indicating rapid and sustained metabolomic changes as result of acute and habitual consumption. Of the results reported by Pimentel et al. (13), we could not confirm those for the amino acids, although there were suggestive trends, notably for threonine. In addition, there were compelling trends for the deoxycholates for total dairy and milk (but not other dairy types), as well as indole-3-lactate.

Although circulating fatty acids—C14:0, C15:0, C17:0, and trans-C16:1n–7, and potentially trans-C18:1n–7 and conjugated linoleic acid—have often been suggested as primary biomarkers of both total dairy and dairy fat intake in general populations (50, 51), they are not as consistent as might be desired of biomarkers, and also can associate with other components of habitual diet, such as fish (9, 10). In the present study, these lipids, where available, were not among the top hits discriminating high from low dairy intake in serum. Several long-chain fatty acids measured in the Offspring, such as TAGs with cheese (C46:0, C54:4, C54:5, and C54:6 TAGs) and butter (C54:4 TAG), and C20:5 CE with yogurt, were among the discriminatory metabolites withstanding full adjustment for other dairy and food intake and correction for multiple testing, suggesting they could be good biomarkers for select dairy types. In the Gen3 nontargeted data, several short- and medium-chain fatty acids could have emerged among the unannotated signals, including C14, C15, and C17 di- and triacylglycerols, in association with yogurt intake. However, none of these metabolites were assessed in both cohorts, and thus these associations could not be verified.

The challenges associated with replicating previously reported dairy-metabolite associations are many. For example, trials often use strictly controlled dairy intake, allowing for fine discrimination between consumption and nonconsumption of various dairy products, or dairy entirely. In free-living observational cohorts where few individuals are nonconsumers, there can be too much metabolic “noise” from other dairy types, including those in mixed dishes, to definitively identify a signature of any single given dairy product. In addition to our inability (due to unavailability) or failure to confirm previously reported associations, we were only able to replicate a few associations in both cohorts in the present study: pantothenate and uridine. Many associations did not replicate. Differences in age and generation, time frame (i.e., early 1990s in the Offspring compared with early 2000s for Gen3), and changes in food trends and product availability, could contribute to these contrasting results, as well as to nonconfirmation of previous observational literature. For example, yogurt intake was associated with aspartate in Gen3, but not in the Offspring, which might reflect the plausibly more common consumption of artificially sweetened yogurts in the early 2000s, compared with a decade earlier, as well as generational shifts in taste preferences. In addition, serum sample degradation and small differences in analytic protocols could have further contributed to inconsistent results between the 2 cohorts. The factors contributing to our failure to replicate either between our own cohorts, or the work of published literature, provide relevant lessons for future work in confirming biomarkers of complex dairy intake, especially in population-based studies using biobanked specimens.

We were surprised not to observe a greater number of associations for cheese intake among either the targeted, or notably, the nontargeted metabolite set, given the fermented qualities of many cheeses and their putative impacts on the gut microbiome (and gut metabolites). This could be attributable to the broad characterizations and groupings of cheese defined by the FFQ, as well as the definitions we used, which might not be sufficiently specific, in addition to the use of cheese in many mixed dishes, which are more difficult to assess with the FFQ. Interestingly, an earlier study from the PLCO author group reported no associations of milk-based products with metabolites in either serum or urine (3). At least 1 other observational study of food groups and serum metabolomics (356 known signals) also reported no associations with dairy, despite confirming other diet-metabolite relations (5). As in our study, lack of associations might be attributable to the heterogeneity of dairy (or cheese) as a whole food group, or limited numbers of metabolite targets might preclude relevant findings.

In contrast, we were pleased with the 68 unannotated signals that emerged unique to yogurt—13 of them significant after correction—which in part seemed to reflect the broad range of flavorings and food additives obvious to anyone browsing the supermarket. We could have been able to observe so many signals because of the specificity (at the time) of yogurt as a dairy type, as well as the stark contrasts in consumption in our populations; half of each cohort were nonconsumers of yogurt, in contrast to other dairy types, which had broader ranges of consumption.

Continuing investigations of dairy (and other food) biomarkers and confirmation of the present results in cross-sectional and longitudinal studies, as well as in feeding studies, could help elucidate the complex and conflicting role of dairy in human health, particularly in heterogeneous and changing foodscapes. Our findings on aspartate and yogurt, for example, could be time- and generation-specific, whereas others, such as specific long-chain fatty acids as well as SMs, could prove to be more consistent biomarkers across time and evolving foodscapes. Both time-/generation-specific and consistent biomarkers could have a role in demonstrations of validity of existing and future diet history methodologies. Whether the presently identified metabolites are further validated as direct biomarkers of dairy, or as endogenous effects of eating dairy, they could be used to investigate metabolite-disease relations, thereby leading toward a more nuanced understanding of the pathways of dairy consumption in health.

Strengths and limitations

We benefited from 2 well-characterized cohorts with both targeted and nontargeted metabolomics data. Because targeted metabolomics data sets introduce bias in possible discovery of important metabolite markers, our utilization of the nontargeted data set expands the discovery horizon to thousands of markers. We included randomly selected discovery and confirmation sets in both cohorts, to ensure robust findings and minimize the possibility of type II error within these populations. However, although our dietary data, derived from FFQ, are commonplace in large cohort studies, there is known measurement error that could have attenuated our results. In addition, relations between habitual diet, such as that assessed by FFQ, and metabolites, might not be as salient as biomarkers as when diet is assessed more proximally (e.g., postprandially), although the need for biomarkers of habitual diet remains. In addition, the serum samples were stored for ≤2 decades, most notably in the Offspring, prior to metabolomics analyses, which might have impacted both sample stability and also quantity and quality of metabolite data. Relatively little crossover existed between annotated or targeted metabolites between the 2 cohorts, and as such there was little opportunity to replicate all findings across cohorts. In addition, given the cross-sectional approach and the single measurement of metabolites, we were unable to assess variation in metabolites within individuals, which might be greater than the variation between individuals, and we were unable to account for this using the present data. Finally, metabolite values presented herein are not directly translatable for clinical use or for determining clinical thresholds, thus limiting their potential relevance in translational work.

Conclusions

Dairy intake was associated with numerous circulating metabolites, including long-chain fatty acids, pantothenate, uridine, and SMs in 2 cohorts of American adults. Reports about diet-metabolite relations might be limited by the specificity of dietary intake and the breadth of measured metabolites. Challenges with replicating findings of other studies, or between samples and populations, are manifold given the expansion and evolution of metabolomics platforms and approaches, as well as shifts and trends in dietary intake.

Notes

This research was supported by Danone North America (AH and PFJ) and the USDA Agricultural Research Service (ARS), agreement no. #58-1950-4-003 (AH and PFJ). The Framingham Heart Study is conducted and supported by the National Heart, Lung, and Blood Institute (NHLBI) in collaboration with Boston University (contract nos. N01-HC-25195, HHSN268201500001I, and 75N92019D00031). Funding support for the Framingham Food Frequency Questionnaire dataset was provided by ARS contract #533k06-5-10, ARS agreement #58-1950-9-001, #58-1950-4-401, and #58-1950-7-707. Funding support for the Framingham Central Metabolomics—HILIC Installment 1 and 2, Metabolomics—HILIC Installment 1, 2, and 3, and Metabolomics—Lipid Platform Installment 1 and 2 datasets was provided by NIH grant R01 DK081572. Funding support for the Framingham Negatively Charged Polar Metabolomics—Amide Installment 1 and the Framingham Targeted and Nontargeted Metabolomics—HILIC Installment 1 datasets was provided by Massachusetts General Hospital departmental funding.

Author disclosures: The contributions of AH and PFJ to this research were partially supported by Danone North America. PFJ serves on the Danone North America Essential Dairy and Plant-Based Advisory Board. CD reports no conflict of interest.

This manuscript was not prepared in collaboration with investigators of the Framingham Heart Study and does not necessarily reflect the opinions or views of the Framingham Heart Study, Boston University, or NHLBI. The views expressed in this article are of those of the authors. The sponsors had no role in the conduct of the study; collection, management, analysis, or interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.

Supplemental Figures 1–3 and Supplemental Tables 1–9 are available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/jn/.

Abbreviations used: CE, cholesterol ester; dbGaP, Database of Genotypes and Phenotypes; FHS, Framingham Heart Study; Gen3, Framingham Heart Study Third Generation Cohort; NHLBI, National Heart, Lung, and Blood Institute; OPLS-DA, orthogonal projections to latent structures discriminant analysis; PLCO, Prostate, Lung, Colorectal, and Ovarian (Cancer Screening Trial); SM, sphingomyelin; TAG, triacylglycerol; VIP, variable importance in projection.