Article pubs.acs.org/JAFC
Metabolomics Investigation To Shed Light on Cheese as a Possible Piece in the French Paradox Puzzle Hong Zheng,† Christian C. Yde,† Morten R. Clausen,† Mette Kristensen,§ Janne Lorenzen,§ Arne Astrup,§ and Hanne C. Bertram*,† †
Department of Food Science, Aarhus University, Kirstinebjergvej 10, DK-5792 Aarslev, Denmark Department of Nutrition, Exercise and Sports, University of Copenhagen, DK-1958 Frederiksberg C, Denmark
§
S Supporting Information *
ABSTRACT: An NMR-based metabolomics approach was used to investigate the differentiation between subjects consuming cheese or milk and to elucidate the potential link to an effect on blood cholesterol level. Fifteen healthy young men participated in a full crossover study during which they consumed three isocaloric diets with similar fat contents that were either (i) high in milk, (ii) high in cheese with equal amounts of dairy calcium, or (iii) a control diet for 14 days. Urine and feces samples were collected and analyzed by NMR-based metabolomics. Cheese and milk consumption decreased urinary choline and TMAO levels and increased fecal excretion of acetate, propionate, and lipid. Compared with milk intake, cheese consumption significantly reduced urinary citrate, creatine, and creatinine levels and significantly increased the microbiota-related metabolites butyrate, hippurate, and malonate. Correlation analyses indicated that microbial and lipid metabolism could be involved in the dairyinduced effects on blood cholesterol level. KEYWORDS: SCFA, cholesterol, TMAO, feces, urine, metabolome, fermented dairy products
■
French paradox.8 Astrup suggested that the relationship between dairy saturated fatty acids and CVD probably needs to be revisited.9 A recent study by Soerensen et al.10 has also shown that lower increases in total and LDL cholesterol levels were observed after intake of cheese (0.41 ± 0.15 and 0.47 ± 0.12 mmol/L, respectively) and milk (0.57 ± 0.13 and 0.53 ± 0.11 mmol/L, respectively) relative to the control group (0.89 ± 0.12 and 0.84 ± 0.11 mmol/L, respectively).10 A potential explanation is that a high calcium content in dairy products increases the excretion of fat.10 However, the underlying metabolic mechanism of this phenomenon remains unclear. Metabolomics has emerged as a promising tool to identify dietary markers11,12 and also to elucidate the metabolism of specific dietary components. Consequently, several studies have reported the use of metabolomics for studying the metabolic response to intake of polyphenol-rich diets and beverages such as green tea,13 wine,14 almonds,15 and fruits and vegetables.16 Intriguingly, studies have also demonstrated that metabolomics can be an efficient tool to elucidate endogenous effects associated with specific dietary compounds. Thus, recently using metabolomics we showed that whey protein had an effect on body weight regulation in mice.17,18 This finding could be ascribed to endogenous effects on the TCA cycle17 and altered energy metabolism in several tissues.18 Consequently, we hypothesize that metabolomics could also be a useful tool to elucidate the metabolic effects associated with intake of different dairy products.
INTRODUCTION Dietary saturated fatty acids have been linked to increased lowdensity lipoprotein (LDL) cholesterol level, thereby increasing the risk of cardiovascular disease (CVD).1 However, there seems to be an exception known as the French paradox, which describes the observation of low coronary heart disease (CHD) mortality despite high intake of saturated fatty acids in France; hence, this paradox has attracted much attention. Intriguingly, Artaud-Wild et al.2 found that the mortality caused by CHD in France was around 5 times lower than that in Finland (198 versus 1031 per 100,000 men aged 55−64 years), but inhabitants in both countries consumed very similar amounts of saturated fatty acids (24 versus 26 cholesterol-saturated fat index). This observation may be explained by the higher saturated fatty acid consumption from plant-based foods in France compared with Finland, for example, vegetable oil intake of 8 versus 1 g/1000 kcal and vegetable intake of 81 versus 19 g/1000 kcal. However, a meta-analysis strongly suggested that dairy foods, despite their relatively high content of saturated fatty acids, may possess some unique properties in relation to CVD and CHD.3 In fact, different types of dairy products may influence the risk of CVD differently. Fermented dairy products have been proposed as functional foods with cholesterollowering effect and thereby protect against CVD compared with nonfermented dairy products.4 Consequently, cheese as a fermented dairy product has been reported to reduce LDL cholesterol level when compared with butter of equal fat content, for example, 40 g of fat as butter or matured cheddar cheese5 and 13% of energy from butter or hard Samsø cheese.6 Moreover, because France is the second highest cheese consumption country in the world,7 Petyaev and Bashmakov even hypothesized that cheese could be responsible for the © XXXX American Chemical Society
Received: December 4, 2014 Revised: February 23, 2015 Accepted: March 2, 2015
A
DOI: 10.1021/jf505878a J. Agric. Food Chem. XXXX, XXX, XXX−XXX
Article
Journal of Agricultural and Food Chemistry
region between 0.6 and 9.4 ppm excluding the water signals from 4.7 to 5.0 ppm for fecal WSE sample, and the region between 0.4 and 8.1 ppm excluding the chloroform signals from 7.1 to 7.3 ppm for fecal LE sample were subdivided into 0.01 ppm spectral regions (bins) and integrated. Finally, the binned data were normalized to the total signal amplitude of each NMR spectrum to minimize the dilution effect for urine but not for feces. Multivariate Data and Statistical Analyses. Principal component analysis (PCA) is an unsupervised model that provides visualization of a multivariate data set and thereby enables identification of clusters of different diet groups in the present study. In addition, orthogonal partial least-squares discriminant analysis (OPLS-DA) as a supervised model was used to compare the difference between any two diet groups. The advantage of OPLSDA is to separate the classification property in the predictive component P[1] and the within-class variation in the orthogonal component O[1], which can achieve better separations and thereby make data interpretation easier.20 PCA and OPLS-DA were carried out on Pareto-scaled data by using SIMCA 13.0 software (Umetrics, Umeå, Sweden). Pareto scaling keeps the structure of the original spectra and thereby reduces the impact of low-intensity, noisy signals. For OPLS-DA, a leave-one-out cross-validation (LOOCV) method was performed, where R2Y and Q2 were calculated as the explained variance in the Y matrix and the predictive capability of the model, respectively. Values of these two parameters close to 1.0 represent an excellent model. Furthermore, the significance of the models was tested by CV-ANOVA in the SIMCA software.21 The significance of variables for the separation between two groups in OPLS-DA was assessed by the variable importance in the projection (VIP) method, and those variables with VIP scores above 2.0 are considered important and used for further analysis. The NMR signals (VIP > 2.0) were carefully inspected for the quality of spectral alignment and assigned on the basis of our previous paper for urine12 and reported data for feces22 as well as HMDB,23 as shown in Table S2 in the Supporting Information. In addition, 2D 1 H−1H total correlation (TOCSY) and 2D 13C−1H heteronuclear single-quantum coherence (HSQC) experiments were performed on representative samples of urine and feces. All NMR spectra were processed by the Processor module in Chenomx NMR Suite software (version 7.7, Edmonton, Canada) including phase, baseline, and shim correction and saved as the CNX file format. The identified metabolite peak was integrated and quantified using the Profiler module in the software. Each metabolite peak was checked manually to ensure a good fitting, and the data were then output to an Excel file for univariate analysis. Prior to univariate analysis, all data were evaluated for normal distribution based on the Anderson−Darling test (P value < 0.05) in SAS software (PROC UNIVARIATE procedure, SAS 9.2, SAS Institute Inc., Cary, NC, USA). When non-normality was observed, data were log-transformed and assessed for normal distribution once again. Analysis of variance (ANOVA) was carried out by using a linear mixed effects model (PROC MIXED procedure, SAS 9.2)12 to evaluate the effect of diet, period, and their interaction for each quantified metabolite. The mixed model included the fixed effects of diet, period, and their interaction, whereas the intercept of model and individuals was used as random effect. Least-squares mean procedure was performed to calculate means and standard errors. Statistical significance was assessed using Student’s t test with Bonferroni correction for multiple comparisons, and a Bonferroni-adjusted P value 10 h/week), milk allergy, lactose intolerance, participation in other studies, or inability to follow the intervention. The intervention periods were separated by a washout period of at least 14 days. During each intervention, 5 days of feces samples (days 10−14) and 24 h urine samples (day 14) were collected and frozen at −80 °C. The detailed study design has previously been described.10 The trial was approved by the Municipal Ethical Committee of Copenhagen (H-1-2011-004) and registered in the database (NCT01317251, http://www.clinicaltrial.gov). NMR Measurements. NMR spectra were recorded on a Bruker Avance 600 spectrometer equipped with a 5 mm TXI probe (Bruker BioSpin, Rheinstetten, Germany) operating at a frequency of 600.13 MHz. The main acquisition parameters included 32K data points, a relaxation delay of 2 s, a spectral width of 7288.63 Hz, and an acquisition time per scan of 2.25 s. For urine, samples were thawed, vortexed, and centrifuged at 10000g for 5 min, and 500 μL of supernatant was transferred to a 5 mm NMR tube and mixed with 100 μL of a 0.75 M phosphate buffer solution containing 0.5% sodium trimethlysilyl propionate-d4 (TSP) in D2O. A standard Bruker “zgpr” pulse program with water presaturation was used to acquire 1H NMR spectra at 25 °C. For fecal extracts, approximately 40 mg of lyophilized fecal powder was weighed into a 1.5 mL Eppendorf tube, and then 300 μL of ice-cold methanol was first added and then mixture vortexed thoroughly and placed on ice for 10 min. Then, ice-cold chloroform and water were consecutively used for extraction acording to the same procedure with methanol. The mixture last was put in a 4 °C refrigerator overnight for separation. On the next day, the mixture was centrifuged at 1400g for 30 min at 4 °C, and the upper methanol− water layer as water-soluble extract (WSE) and the lower chloroform layer as lipid extract (LE) were transferred to fresh 1.5 mL Eppendorf tubes, respectively, and then dried by using vacuum centrifugation for around 3 h. The dried sample was stored at −80 °C until NMR analysis. The dried WSE was redissolved in 550 μL of D2O, 25 μL of H2O, and 25 μL of D2O containing 0.05% TSP, and the dried LE was redissolved in 575 μL of CHCl3-d and 25 μL of CHCl3-d containing 0.05% tetramethylsilane (TMS). Then, the mixture was vortexed thoroughly and transferred to a 5 mm NMR tube immediately. Both the water-soluble and lipid samples were analyzed by using a standard “zgpr” pulse program at 25 °C. NMR Data Preprocessing. NMR data were processed in the Topspin 3.0 software (Bruker BioSpin) as described in our previous paper.12 The 1H spectra of urine and feces were referenced to TSP and TMS signals at 0 ppm, respectively. The “icoshift” procedure was applied to align NMR spectra in MATLAB (R2012a, The Mathworks Inc., Natick, MA, USA).19 The spectral region from 0.0 to 10.0 ppm without the residual water signals from 4.7 to 5.0 ppm for urine, the B
DOI: 10.1021/jf505878a J. Agric. Food Chem. XXXX, XXX, XXX−XXX
Article
Journal of Agricultural and Food Chemistry coefficients (R) were calculated to explore the relationship between each metabolite and total or LDL cholesterol in SAS 9.2 (PROC CORR procedure, SAS 9.2) and considered statistically significant when a P value was