Plasma Metabolites Related to the Consumption of Different Types of Dairy Products and Their Association with New‐Onset Type 2 Diabetes: Analyses in the Fenland and EPIC‐Norfolk Studies, United Kingdom

Scope To identify metabolites associated with habitual dairy consumption and investigate their associations with type 2 diabetes (T2D) risk. Methods and results Metabolomics assays were conducted in the Fenland (n = 10,281) and EPIC‐Norfolk (n = 1,440) studies. Using 82 metabolites assessed in both studies, we developed metabolite scores to classify self‐reported consumption of milk, yogurt, cheese, butter, and total dairy (Fenland Study‐discovery set; n = 6035). Internal and external validity of the scores was evaluated (Fenland‐validation set, n = 4246; EPIC‐Norfolk, n = 1440). The study assessed associations between each metabolite score and T2D incidence in EPIC‐Norfolk (n = 641 cases; 16,350 person‐years). The scores classified low and high consumers for all dairy types with internal validity, and milk, butter, and total dairy with external validity. The scores were further associated with lower incident T2D: hazard ratios (95% confidence interval) per standard deviation: milk 0.71 (0.65, 0.77); butter 0.62 (0.57, 0.68); total dairy 0.66 (0.60, 0.72). These associations persisted after adjustment for known dairy‐fat biomarkers. Conclusion Metabolite scores identified habitual consumers of milk, butter, and total dairy products, and were associated with lower T2D risk. These findings hold promise for identifying objective indicators of the physiological response to dairy consumption.


Metabolite and fatty acid samples processing Sample derivation and internal validation setsthe Fenland study
Blood samples were collected into lithium heparin coated tubes and spun to separate the plasma.Samples were stored in -80ºC in freezers until analyzed.Before the analysis, the samples were removed from the freezer, placed on roller mixers at 30 rpm to thaw for 15-20 minutes and centrifuged for 1 minute at 2000 rpm.Calibration standards and internal standards were also briefly centrifuged and quality control (QC) samples were centrifuged for 5 minutes at 2750 rpm.All vials were then shaken for 15 minutes at 1200 rpm and vortexed and then transferred to 96-well plates with rows A-H and columns 1-12.Derivatization agent (50 µl) was added to all the wells.The internal standards (10 µl) were added to all wells for quantification apart from A1, which belonged to the blank (deionized water with phosphate buffer saline).Positions B2-D2 contained the zero samples (blank samples with internal standards).Positions E1-C2 contained the calibration samples, positions D2-F2 contained the commercial QC samples, positions G2 and H12 contained the pooled QC samples and positions H3-G12 contained the study samples.The sample processing was done on a Hamilton STAR liquid handling station (Hamilton Robotics Ltd, Birmingham, UK).Targeted metabolomic profiling of 10,684 blood plasma samples was performed with a commercial kit (Absolute IDQ p180 kit, BIOCRATES Life Sciences AG, Innsbruck, Austria).Flow injection analysis (FIA) MS (AB SCIEX 5500 Qtrap mass spectrometer, Sciex Ltd, Warrington, UK), isocratic with methanol, in positive ionization mode was performed for lipids and acylcarnitines.FIA-MS in negative ionisation mode was performed for hexose.Ultraperformance liquid chromatography MS (UPLC-MS; Waters ltd, Manchester, UK coupled to ABSciex 5500 Qtrap mass spectrometer, Sciex Ltd, Warrington, UK) was performed for amino acids and biogenic amines with a 5-minute gradient elution starting with 100% water and increasing to 95% acetonitrile and 0.2% formic acid over a Waters Acquity UPLC BEH C18 column (2.1 x 50 mm, 1.7 µm).Amino acids and biogenic amines were quantified and expressed in an absolute scale, whereas lipids and hexose were semi-quantified and expressed in a relative scale.
The raw metabolomics data were processed with the MetIDQ software (provided with the kit).Then the data went through a check for outliers, non-detects, peak picking, normalization with the use of the QC samples (metabolite values were divided by the ratio of the mean value of the QC samples within a plate to the mean value of the samples across all the plates) and finally batch correction.During batch correction, zero values were replaced with numbers randomly selected from a uniform distribution between 0.1 and a minimum observed within batch.

External validation setthe EPIC Norfolk study
From sample collection to sample preparation to be shipped to Metabolon, Inc., blood samples were stored in the gas phase of liquid nitrogen (-175ºC).From sample preparation to shipment to Metabolon, samples were stored in -70ºC.Upon receipt of the samples by Metabolon, they were stored in -80ºC until one day before the analysis and overnight under nitrogen the night before the analysis.Deproteinisation was done with methanol, under vigorous shaking for 2 minutes (Glen Mills GenoGrinder 2,000).The samples were then centrifuged and placed on a TurboVap (Zymark) to remove the organic solvent.The order of the study samples on the plates was randomized and QC samples (two commercial QC standards and 4 pooled QC samples) were evenly distributed across the study samples.A blank sample and internal standards were also used.Untargeted metabolomic profiling was performed in 1,503 blood citrated plasma samples (DiscoveryHD4 ® platform, Metabolon, Inc.; Waters ACQUITY UPLC, Thermo Scientific Q-Exactive spectrometer with a heated electrospray ionization (HESI-II) source and Orbitrap mass analyzer with 35,000 mass resolution).Samples were divided into five fractions of which two were used for reverse phase UPLC MS/MS with positive ion mode ESI (mobile phases with 0.1% formic acid in water and 0.1% formic acid in methanol), one was used for reverse phase UPLC MS/MS with negative ion mode ESI (mobile phases with 6.5 mM ammonium bicarbonate in water, pH=8 and 6.5 mM ammonium bicarbonate in 95% methanol and 5% water), one was used for hydrophilic interaction liquid chromatography (HILIC) UPLC MS/ MS with negative ion mode ESI (mobile phases with 10mM ammonium formate in 15% water, 5% methanol, 80% acetonitrile and 10mM ammonium formate in 50% water and 50% acetonitrile) and one was kept as a backup.The UPLC column used was a Water BEH C18 (2.1 x 100 mm, 1.7 µm) at 40ºC and the HILIC column was a Waters BEH Amide (2.1 x 50 mm, 1.7 µm) at 40ºC.Gradient elution with water, methanol, 0.05% pentafluoropropionic anhydrate and 0.1% formic acid was used in all the columns.Metabolite identification was based on the retention index, accurate mass matched to the Metabolon internal library of metabolites within a range of ±10 ppm and the standards used.The raw metabolomics data were processed and went through normalization by ``block correction'' i.e. setting median to 1 and adjusting each data point proportionately and QC checks with in-house software from Metabolon, Inc.

Metabolite annotation
For all the lipids, ``Cx:y'' indicates x carbon atoms and y double bonds.For the acylcarnitines, x denotes the number of carbon atoms of the carboxylic acid, which is esterified with carnitine.So carnitine is denoted as C0.For PCs there is the additional indication of ``aa'' when it contains two fatty acids bound through an ester linkage or ``ae'', when one of the two fatty acids were bound through an ether bond.So a PC might be denoted as PC aa Cx:y or as PC ae Cx:y.For LPCs, ``a'' denotes the presence of a fatty acid, so then this becomes LPC aa Cx:y or LPC ae Cx:y.SMs are denoted as SM Cx:y or SM-OH Cx:y if it is a hydroxy-SM.

Fatty acids assay
Phospholipid fractions were isolated from total plasma lipids with solid-phase extraction.
Then free fatty acids were extracted with hydrolysis and derivatization and methylated to form fatty acid methyl-esters.These esters were then used in gas chromatography (7890N; Agilent Technologies) with a 30 m capillary column with a diameter of 0.25 mm.Sample was processed with sequential multipurpose sampler systems (Gerstel GmbH & Co. KG, Mulheim an der Ruhr, Germany).One blank sample and two QC samples were used in all the batches: one from human plasma pooled from the study samples and one from horse plasma.
Standards of all fatty acids were used for the gas chromatography calibration.Study and QC samples were stored in -80ºC before analysis.Fatty acids were identified by comparing their retention times with those of the standards.

Approximate metabolite matching between the targeted and untargeted metabolomics
For the external validation, the metabolites between the targeted and untargeted metabolomics platforms were matched.Exact matching was done for 21 out of 21 amino acids, 10 out of 13 biogenic amines and 15 out of 40 acylcarnitines (Supplemental Table 1).
For phosphatidylcholines (PCs), lyso-phosphatidylcholines (LPCs) and sphingomyelins (SMs), the FIA-MS/MS did not allow for the identification of the fatty acids in these lipids.
Thus, the lipids reported from the targeted metabolomics platform are the sum of all their isobaric (same weight within a range of ±0.5 Da) and isomeric (same number of atoms, but different structure) compounds.Likewise for hexose sugars.The untargeted metabolomics platform identified specific isobaric/isomeric compounds for some lipids, while for others, it also identified the sum of compounds, but with a higher precision.Lipids were matched approximately based on the number of carbon atoms, the number of double bonds and the molecular mass.After this process, approximate matching was possible for 9 out of 14 LPCs, 17 out of 74 PCs, 8 out of 11 SMs, and hexose (Supplemental Table 1).

Supplemental
CI: confidence interval; EPIC: European Investigation into Cancer and Nutrition; HR: hazard ratio; OCSFA: odd-chain saturated fatty acid; T2D: type 2 diabetes Supplemental Figures Supplemental Figure 1.Receiver operating characteristic (ROC) curves and corresponding AUCs for prediction models of metabolite scores reflecting (A) milk, (B) yogurt, (C) cheese, (D) butter and (E) total dairy consumption in the internal validation set (n=4,246).Socio-demographic, lifestyle model: age, sex, test site, smoking status, physical activity, lipid lowering drugs, hormone-replacement therapy and BMI; Top metabolite score: metabolites with absolute values of elastic net coefficients >mean+2×SD.AUC: area under the curve

Table 1 .
Metabolite matching between the targeted metabolomics platform in the Fenland study and the untargeted platform in the EPIC Norfolk diabetes casecohort study

Table 2 .
Covariates used in different models for the different analyses in the Fenland and EPIC-

Table 3 .
Likelihood ratio (LR) tests and net reclassification improvement (NRI) for the models with or without the fatty acids a and the metabolite score in the fatty acid set of the Fenland study for internal validation All models include age, sex, test site, smoking status, physical activity, lipid lowering drugs, hormonereplacement therapy and BMI b) Metabolites approximately matched between the derivation (targeted) set and the external validation (untargeted) set.Selection of metabolites was based on the top metabolites defined as metabolites with absolute values of the elastic net coefficients >mean+2×SD Supplemental a) Fatty acids: C15:0, C17:0, C16:1n7t in the Fenland Study and C15:0, C17:0 in the EPIC-Norfolk Study.b) Socio-demographic and lifestyle factors: age, sex, test site, smoking status, physical activity, lipid lowering drugs, hormone-replacement therapy, and BMI c) Selection of metabolites was based on the top metabolites defined as metabolites with absolute values of elastic net coefficients > mean+2×SD or >mean+SD LR: likelihood ratio; NRI: net reclassification improvement

Table 5 .
Hazard ratios of T2D per 1 SD of metabolite scores not significantly associated with dairy consumption after 16,360 person-years of follow-up (641 cases of T2D) in the incident diabetes casecohort study within the EPIC Norfolk study

.
Classification ability of the metabolite scores for consumption of different dairy products: a secondary external validation analysis of the metabolite scores derived from the Fenland Study and tested in the EPIC-Norfolk Study, United Kingdom a a) All the statistics were adjusted for socio-demographic and lifestyle factors including age, sex, test site, smoking status, physical activity, lipid lowering drugs, hormone-replacement therapy, and BMI b) The set of metabolites used in the analysis to derive the metabolite scores in the Fenland Study using the different cut-off point of the mean+1SD for the absolute values of the elastic net coefficients c) " No" refers to the model including covariates only; and " Yes" , the model including covariates and the metabolite score d) p values for the metabolite score for the classification of consumers and non-consumers AUC: area under the curve; EPIC: European Investigation into Cancer and Nutrition; NRI: net reclassification improvement Supplemental

Table 7 .
Associations of metabolite scores including selected top metabolites classifying each of total and types of dairy consumption with other dairy products and food groups in the EPIC Norfolk case-cohort study

Table 8 .
Hazard ratios of T2D per 1 SD of metabolite scores classifying dairy consumption in the incident diabetes case-cohort study within the EPIC Norfolk study a Strength of associations of metabolites, socio-demographic and lifestyle covariates with consumption levels of total and types of dairy products in the derivation set of the Fenland study with the 82 metabolites matched with those in the diabetes case-cohort of the EPIC Norfolk study.Colors show the relative strength of associations within each column.Intensity of red color indicates strength of positive associations and intensity of blue color indicates strength of negative associations.† Strengths of associations are expressed as the beta coefficients of elastic net penalized logistic regression and ordinary logistic regression.‡ Strengths of associations are expressed as the beta coefficients of elastic net penalized linear regression and ordinary linear regression.EPIC: European Prospective Investigation into Cancer and Nutrition Strength of associations of metabolites, socio-demographic and lifestyle covariates with consumption levels of total and types of dairy products in the derivation set of the Fenland study with the 82 metabolites matched with those in the diabetes case-cohort of the EPIC Norfolk study and the OCSFAs.Colors show the relative strength of associations within each column.Intensity of red color indicates strength of positive associations and intensity of blue color indicates strength of negative associations.† Strengths of associations are expressed as the beta coefficients of elastic net penalized logistic regression and ordinary logistic regression.‡ Strengths of associations are expressed as the beta coefficients of elastic net penalized linear regression and ordinary linear regression.EPIC: European Investigation into Cancer and Nutrition; OCSFA: odd-chain saturated fatty acid