Phenotyping Tea Consumers by Nutrikinetic Analysis of Polyphenolic End-Metabolites Ewoud J. J. van Velzen,†,§ Johan A. Westerhuis,*,† John P. M. van Duynhoven,§ Ferdi A. van Dorsten,§ Christian H. Gru ¨ n,§ Doris M. Jacobs,§ Guus S. M. J. E. Duchateau,§ Danie¨l J. Vis,† and Age K. Smilde† Biosystems Data Analysis, Swammerdam Institute for Life Sciences, Universiteit van Amsterdam, Nieuwe Achtergracht 166, 1018 WV Amsterdam, The Netherlands, and Unilever Research and Development, Olivier van Noortlaan 120, 3133 AT Vlaardingen, The Netherlands Received December 15, 2008
An integration of metabolomics and pharmacokinetics (or nutrikinetics) is introduced as a concept to describe a human study population with different metabolic phenotypes following a nutritional intervention. The approach facilitates an unbiased analysis of the time-response of body fluid metabolites from crossover designed intervention trials without prior knowledge of the underlying metabolic pathways. The method is explained for the case of a human intervention study in which the nutrikinetic analysis of polyphenol-rich black tea consumption was performed in urine over a period of 48 h. First, multilevel PLS-DA analysis was applied to the urinary 1H NMR profiles to select the most differentiating biomarkers between the verum and placebo samples. Then, a one-compartment nutrikinetic model with first-order excretion, a lag time, and a baseline function was fitted to the time courses of these selected biomarkers. The nutrikinetic model used here fully exploits the crossover structure in the data by fitting the data from both the treatment period and the placebo period simultaneously. To demonstrate the procedure, a selected set of urinary biomarkers was used in the model fitting. These metabolites include hippuric acid, 4-hydroxyhippuric acid and 1,3-dihydroxyphenyl-2-O-sulfate and derived from microbial fermentation of polyphenols in the gut. Variations in urinary excretion betweenand within the subjects were observed, and used to provide a phenotypic description of the test population. Keywords: Crossover • Gut • Metabolizers • Metabonomics • Microbiota • Multilevel Data Analysis • NMR • Nutrition • Pharmacokinetics • Phenotyping • PLS-DA • Polyphenols • Urine
Introduction From nutritional interventions studies, it is known that the physiological and metabolic responses to a dietary treatment can strongly vary among humans.1-4 Genotypic diversity, lifestyle and gut microbial variations are important and known factors contributing to these intrinsic variations.2,5-13 Assessment of interindividual differences is particularly important for nutritionists to anticipate the effectiveness and dosing of the dietary intake, and to provide a better understanding of the underlying mechanism of action. Hence, large response differences between subjects are more likely to demonstrate variable efficacy than a nutritional intervention with more consistent responses in the test population.14 Until now, the response differences that exist between subjects in a test population have not often been addressed in nutritional studies. Recently, this topic has come more into focus with the advent of nutrigenomics15-17 and metabolomics.1,6,18 These disciplines use transcriptomic15 and metabolic profiling techniques1 to differentiate between study subjects. In drug * To whom correspondence should be addressed. Phone: +31 20 525 6546. Fax: +31 20 525 6971. E-mail:
[email protected]. † Universiteit van Amsterdam. § Unilever Vlaardingen. 10.1021/pr801071p CCC: $40.75
2009 American Chemical Society
discovery studies, however, segmentation is explicitly based on identifiable and standardized pharmacokinetic measures to anticipate for potentially new drug targets.14,19,20 The advantage of using pharmacokinetics (PK) is that a quantitative description is obtained by which drugs are absorbed, distributed, metabolized and eliminated (ADME).21 In analogy, we now propose a similar strategy for assessing the metabolic phenotypes in nutrition-based intervention studies. The PK properties are particularly of interest in this strategy because they provide direct insight into the diversity and effectiveness of the indigenous gut microbiota of the subjects in the study population. The use of PK parameters to distinguish between response phenotypes is not trivial in nutritional studies. The impact of dietary interventions on the endogenous and exogenous metabolite levels in blood, urine or faeces of healthy subjects is typically small in relation to the background diet and may be highly variable across the study population.3,22,23 This is different from drug discovery studies since drugs by nature should have a strong metabolic impact. As a result, the xenobiotic metabolites can be more easily distinguished from baseline biofluid metabolites.24 Also detailed knowledge about the exact composition of a food product (e.g., phytochemicals), Journal of Proteome Research 2009, 8, 3317–3330 3317 Published on Web 04/20/2009
research articles the active ingredient(s), their metabolism and mechanism of action is often lacking. This hampers targeted analysis of theoretically expected metabolites, and distinguishes it clearly from PK studies. Because of these fundamental differences between pharmacological and nutritional studies, we prefer to use the expression nutrikinetics instead of pharmacokinetics. As the exact dietary constituents and their ADME properties may not be exactly known in advance, an initial assessment of the most important metabolic changes following a nutritional intervention is needed prior to the nutrikinetic analysis. For an unbiased biomarker selection, a metabolomics approach is most suitable since it does not require any prior knowledge of the underlying metabolic pathways.25,26 Typically, the biofluid metabolic profiles from spectroscopic and/or chromatographic experiments are evaluated in metabolomics.27,28 Recently, we applied metabolomics in a small explorative kinetic experiment to discern the most distinguishing urinary biomarkers after a single-bolus oral intake of decaffeinated black tea solids.29 The excretion characteristics across the selected biomarkers and the individuals were assessed, and the results clearly demonstrated the potential benefit of metabolomics in the discovery of unknown dietary metabolites in humans. In this previous study, however, we did not use a nutrikinetic model to estimate the kinetic properties of the detected biomarkers. Instead of using a nutrikinetic model, a model-free approach was used for parameter estimation. The procedure is therefore less suitable for use in the current phenotyping study which requires uniform nutrikinetic model parameters for distinguishing metabolic phenotypes. In the current study, the integrated approach of metabolomics and nutrikinetics was demonstrated for the case of a placebo-controlled, full crossover designed nutritional intervention study with black tea polyphenols. The metabolomics approach identified increased urinary excretion of several gutmediated metabolites of tea flavonoids. The nutrikinetic properties of three representative phenolic biomarkers were evaluated and subsequently allowed us to describe metabolic phenotypes within this study population. The nutrikinetic parameters were estimated by fitting a customized onecompartment model with first-order kinetics, a lag time, and a baseline function whereby the crossover design of the study was taken into account. This clearly distinguishes our new method from the model fitting procedure in a common onecompartmental PK analysis in which the experimental design is usually not considered.30
Experimental Section Study Protocol. The study had a randomized, placebo controlled, double blind, full crossover design (Figure 1A) in which 20 healthy nonsmoking male volunteers participated. The subjects were 18-40 years of age and their Body Mass Index (BMI) was between 19 and 29 kg/m2. During each of the intervention periods, the subjects visited Mediscis (Berchem Antwerp, Belgium) where they maintained on a low-polyphenol diet for 4 days (Supplementary Document 1 in Supporting Information). Between the interventions, a 10-day wash-out period was included during which the subjects were free to consume their normal diet without any restrictions. The volunteers were also requested to follow a similar dietary and lifestyle pattern for the duration of the study. On the morning of the third day, the volunteers consumed a capsule containing 2500 mg of dried black tea extract powder, a capsule containing a red grape extract, or a placebo (sucrose) with an adequate 3318
Journal of Proteome Research • Vol. 8, No. 7, 2009
van Velzen et al. amount of water (>200 mL). The tea extract was prepared from a spray dried aqueous extract of Lipton Yellow Label (code LYL640, U.S. blend) and contained 800 mg of polyphenols, expressed as gallic acid equivalents.31 The intervention study actually included a third intervention with grape polyphenols which was crossed over with the black tea and the placebo interventions. However, in this paper, we only consider the black tea and the placebo interventions. The protocol was approved by the Commissie voor Medische Ethiek-ZNA Middelheim, Antwerpen, Belgium, and conducted in accordance with the ICH-GCP guidelines for Good Clinical Practice (ICH GCP, 1996). Urine Collection. Urine samples were collected after spontaneous urination, at nonequidistant time points, during 48 h after administering a black tea or placebo capsule. As illustrated in Figure 1B, the time points of the urine sampling are different for all subjects in both intervention periods (Supplementary Table 1 in Supporting Information). The weight of all urine samples produced was measured and a small volume of concentrated hydrochloric acid was added to adjust the pH between 3 and 4. From each acidified urine sample, an aliquot of approximately 10 mL was stored at -20 °C before analysis. Preparation of Samples. Urine samples were thawed at room temperature. To 450 µL of each urine, 200 µL of phosphate buffer solution (0.6 M Na2HPO4/NaH2PO4, pH 6.5) and 50 µL of deuterium oxide (D2O) were added. The phosphate buffer solution contained 0.05 mg · mL-1 3-(Trimethylsilyl)propionic acid-d4 sodium salt (TSP, 98% atom% D, CAS 2449321-8) as an internal standard. After homogenization, the samples were centrifuged at 10 000 rpm for 5 min. From the clear supernatant, 650 µL was then transferred into a 5 mm NMR tube. Preparation of QC Standards. The analytical procedure was controlled by 7 working standards. Solutions of sodium hippurate hydrate (g99%, Sigma, CAS 532-94-5) and a blank were prepared in the phosphate buffer and used as quality control (QC) standards. The concentrations were 0.87, 1.19, 1.53, 2.23, 3.64, and 3.75 mg · mL-1. Similar to the preparation of the urine samples, 50 µL of D2O/TSP was added to the QC standards and transferred into a 5 mm NMR tube (650 µL). The QC acceptance interval of the estimated hippurate concentrations was set to (10% (w/v). NMR Data Acquisition. One-dimensional 1H NMR spectra were acquired on a Bruker Avance 600 MHz NMR spectrometer operating at a proton frequency of 600.13 MHz. A 5 mm triple resonance (TXI) probe was used and the temperature was kept constant at 300 K. The spectra were acquired with presaturation of the water resonance using a noesy1dpr pulse sequence RD90°-F1-90°-Fmix-90°-FID (Bruker Biospin, Germany). Here, F1 is a 4 µs delay time, and Fmix is the mixing time (150 ms). The free induction decays (FIDs) were collected into 64 K points (128 scans) with a spectral width of 9000 Hz, an acquisition time of 3 s and a relaxation delay of 3 s. The spectra were phase- and baseline corrected using Topspin 1.3.4 software (Bruker, Rheinstetten, Germany). NMR Data Processing. An exponential window function was applied to the FID with a line-broadening factor of 0.5 Hz prior to the Fourier transformation. The Fourier transformed NMR spectra were manually phase- and baseline corrected, calibrated and normalized to the methyl resonance of TSP at δ 0.0 ppm. The molar quantity (xint m ) of each NMR signal was then calculated according to eq 1:
Phenotyping Tea Consumers by Nutrikinetic Analysis int xijm )
int ixijm int isijm
×
int uim
ps vs × × × cTSP [µmol] pxj vx F
(1)
int where xijm is the molar quantity represented by signal j in the NMR spectrum of subject i after intervention int (int ) placebo int or black tea) and at collection time point m (µmol); ixijm is the int intensity of the NMR signal; isijm is the integral of the (methyl) TSP resonance at δ 0.0 ppm; ps is the number of protons of
research articles methyl TSP () 9); pxj is the number of protons associated with the NMR signal j; vs is the volume of TSP solution in the NMR tube () 200 µL); vx is the urine volume in the NMR tube () int is the mass of the urine fraction of subject i 450 µL); uim collected at time point m after intervention int; F is the urine density () 1.0 g · mL-1) and cTSP is the TSP concentration () 0.29 µmol · mL-1).29 The resulting NMR spectra were subdivided in discrete regions (‘buckets’) of equal width (δ 0.00225 ppm) from which
Figure 1. Schematic representation of (A) the crossover designed, placebo-controlled intervention study in which 20 subjects were administered a single bolus dose of black tea. The experiment consisted of a baseline (2 days) and two intervention periods (placebo and black tea, 2 days each) with a wash-out period of 10 days in between. A third intervention which involved a grape extract is ignored here. (B) Urine collection scheme. The time points of the spontaneous urine voids were different between the subjects, and between the placebo period (npla) and the treatment period (ntea). Journal of Proteome Research • Vol. 8, No. 7, 2009 3319
research articles the integrated areas were determined using an in-house written matlab routine (version 2008b, The MathWorks). When this small bucket-size was used, hardly any fine-structure and multiplicity was lost in the data. Because bucketing could not completely compensate for line broadening effects and positional shifts (due to differences in pH, ion strength, etc.), also Correlation Optimized Warping was applied on the bucketed data.32,33 This segment-wise optimization was, however, limited to small segment lengths of 0.05 ppm to constrain positional shifts and large misalignments in the NMR spectrum. Analysis of Treatment Effect. Multilevel PLS-DA22 was performed on the NMR urinary profiles to explore (a) whether the black tea intervention led to systematic metabolic changes, and (b) which metabolites were the most differentiating biomarkers between the verum and placebo samples. In the current crossover designed study, a multilevel approach was particularly useful because it separated the between-subject variation from the within-subject variation prior to the multivariate data analysis (PLS-DA). Separation of these confounded sources of variation is advantageous in nutritional metabolomics studies because dietary treatments effects are typically small and often overwhelmed by intrinsic variations that exists between subjects.3 In this multilevel analysis, pooled 48 h urines were used. For each subject, and for each intervention, pooled urines were prepared by adding aliquots of the urine collections together. The masses of these aliquots were proportional with the masses of the urine collections. After collecting the NMR spectra, first the variation splitting property of the multilevel data analysis was used to separate the between-subject variation from the within-subject variation. Then, the within-subject data was autoscaled and regressed against the intervention class labels (-1 for placebo samples, +1 for verum samples) using Partial Least Squares Discriminant Analysis (PLS-DA). By means of a 5-fold cross-model validation (CMV) scheme,34 the prediction error of the multilevel PLS-DA model was determined and expressed in terms of DQ2.35 To obtain stable class predictions, and stable biomarker selections, the average result of 20 CMVs was calculated. To validate whether the prediction error of the multilevel PLS-DA model was not obtained by chance, a comparison was made with permutated data.22,36-39 In this permutation test, the prediction errors from 2000 randomly permutated data sets were collected and represented as the H0-distribution of no-effect. The treatment effect was considered statistically significant if the p-value obtained from this permutation test was e0.05. For the multilevel data analysis and the permutation testing, in-house written Matlab routines were used.22 Biomarker Selection. To select the most discriminative NMR biomarkers for black tea consumption from the multilevel model, the rank products (RPs) were calculated.22,39,40 In the RP, the NMR signals obtained from the pooled 48 h urines were ranked according to their absolute size in the multilevel PLSDA regression coefficient. In the currently applied multilevel analysis, the resulting RP is the product of 20 CMV rounds, whereby the NMR signals with the lowest RP-values have the strongest discriminative power (biomarker signals). To determine which biomarker signals can be considered significant, the RP values of the multilevel model were compared with the RP values obtained from 2000 permutations. To limit the chance of selecting false positive biomarkers, only signals with a RP-value below the 10% significance limit of the permutations were taken into consideration for nutrikinetic analysis. 3320
Journal of Proteome Research • Vol. 8, No. 7, 2009
van Velzen et al. Calculation of Nutrikinetic Parameters. In the nutrikinetic assessment, three kinetic parameters were estimated. The first parameter was the net cumulative urinary excretion (output) of the biomarkers over a period of 48 h after the tea intervention c (net xtea in mol). The second parameter was the rate constant of the excretion (ke in h-1), whereby a first-order kinetic model was assumed for the metabolite excretions.30,41 The third parameter was the lag time τ (h) and is defined as the time difference between the start of the intervention and the time point for a biomarker to appear in the urine. Characteristics of the Nutrikinetic Model. The metabolic impact of dietary interventions in humans is often small, subtle, and difficult to discern from a background diet. Furthermore, the metabolic effects can be highly variable within a study population.3,22 It is therefore that nutrikinetic studies requires different demands on the underlying compartmental models for parametrization as compared to drug response (PK) studies. An important aspect in the analysis of nutrikinetic data is to provide good estimations of the metabolic baseline levels. These baseline levels can be relatively large for urinary metabolites (e.g., hippuric acid), and may complicate the kinetic modeling considerably. The nutrikinetic model anticipates on this by fitting the baseline level in the placebo period simultaneously with the model fitting in the treatment period. In this procedure, the crossover design is fully exploited in the parametrization, even though the data from both intervention periods are imbalanced. In a common one-compartmental PK model, the fitting procedures are not integrated. Also the crossover design in the data is not considered, as well as the interdependency of the model parameters in the fitting procedure. These main considerations limit the (direct) use of a common PK model in nutrikinetic experiments. Indices Used in Nutrikinetic Models. In the definition of the nutrikinetic equations, the following indices were used: i ) 1,..., I for the number of subjects and j ) 1,..., J for the number of NMR signals. There are two intervention occasions, that is, placebo and tea, and npla ) 1,..., Npla represents the time point of a urine collection in the placebo period and ntea) 1,..., Ntea represents the time point of a urine collection in the tea intervention period (which is different for each subject). The equations described in the remainder of the paper lack the indices i and j for reasons of simplicity. Cumulative Data. The cumulative NMR data collected in the two intervention periods were used to estimate the nutrikinetic quantities of the urinary excretion for each biomarker j and for each subject i. We will now demonstrate this for the placebo period. Given a Npla data vector, xpla, the cumulative output vector, cxpla, is again a Npla vector which has elements according to eq 2. npla
c pla xnpla
)
∑x
pla [ ] m mol
(2)
m)1
where xpla m is the excreted amount of biomarker j, for subject i, pla in the placebo intervention at time point m (mol) and cxnpla is the cumulative excreted amount of the same biomarker up to, and including, time point npla (mol). In a similar way the total excreted amount cxnteatea in the tea intervention period was calculated. The urine fraction concentrations (Figure 2B) were converted to total cumulative output data and plotted against the collection time points. From this, the cumulative excretion plot
Phenotyping Tea Consumers by Nutrikinetic Analysis
research articles
Figure 2. Aromatic region of (A) the normalized urinary 1H NMR spectra (δ 6-9 ppm) of volunteer 7, acquired at 7 time points after black tea consumption. From the NMR signal at δ 7.78 ppm (hippuric acid), (B) the concentration curve, and (C) the cumulative output curve are illustrated.
was constructed for fitting. An example is given in Figure 2 where the NMR urinary excretion profiles of subject 7 are illustrated. From this subject, the urines were collected at 7 time points after the tea consumption. In Figure 2A, a part of the acquired NMR spectra is shown (aromatic region between δ 6.0 and δ 9.0 ppm). The NMR signal at δ 7.78 ppm is used here as a representative biomarker for the cumulative urinary excretion of hippuric acid. As shown in Figure 2C, the cumulative urinary output nicely describes a first-order kinetic model. Determination of Kinetic Parameters. In the placebo period, a constant urinary excretion was assumed for all metabolites. In the cumulative data set, these baseline levels describe a first-order linear function. The slope and the offset of this function can be estimated with least-squares regression. Therefore, the estimated cumulative excreted amount of a metabolite j at time point npla of a subject i, collected in the placebo period, can be calculated according to eq 3. xnpla ˆ pla + βˆ tnpla + enpla pla ) R pla [mol]
c
c pla xˆnpla
)R ˆ pla + βˆ tnpla [mol]
(3)
Where tnpla (h) represents the time difference between the placebo intervention and the urine sampling at time point npla; c pla xˆnpla is the estimated cumulative output at time point npla (mol); βˆ is the estimated slope associated with the base level of
metabolite j (mol · h-1); R ˆ pla is the estimated offset associated pla with the base level at t ) 0 (mol) and enpla is the model residual fraction at time point npla (mol). After tea consumption (treatment), the metabolites were compensated for their baseline levels. In the current approach, we assumed similar baseline levels as detected in the placebo period since the volunteers followed a controlled dietary and lifestyle pattern for the duration of the study. As a result, the slope of the basal levels (βˆ ) in both the placebo and the treatment periods was considered equal for each metabolite. The initial baseline level for any metabolite at t ) 0 in the placebo period and the treatment period was however strongly influenced by the early morning urine sample and varied between the intervention periods. An important factor that influenced the composition and sample volume of the early morning urine sample was the time difference with the former urination. Therefore, the intercept value of the linear offset term in the treatment period (R ˆ tea) could not be considered equal pla to the placebo period (R ˆ ). In the treatment period, the metabolic baseline levels may change due to the tea consumption. This net metabolic effect can be approximated by a first-order kinetic model.30,41,42 As shown in eq 4, parametrization of this model was achieved by fitting a composite function through the cumulative intensity values cxnteatea for each metabolite j, at each time point ntea, for each subject i. Journal of Proteome Research • Vol. 8, No. 7, 2009 3321
research articles
{ {
van Velzen et al. -kˆe(tntea-τˆ)
c tea xntea c tea xntea
tea )R ˆ tea + βˆ tntea + cxˆmax (1 - e )R ˆ tea + βˆ tntea, tntea < τˆ [mol]
c tea xˆntea c tea xˆntea
ˆ tea tea )R ˆ tea + βˆ tntea + cxˆmax (1 - e-ke(tn -τˆ)), tntea g τˆ tea )R ˆ + βˆ tntea, tntea < τˆ [mol]
) + entea ˆ tea, tntea g τ
[mol]
[mol]
(4)
where tntea (h) represents the time difference between the tea tea is intervention and the urine sampling at time point ntea; cxˆntea the estimated cumulative intensity value at time point ntea tea (mol); cxˆmax is the estimated maximum output of metabolite j tea (mol); R ˆ is the estimated offset associated with the base level in the treatment period at t ) 0 (mol); kˆe is the estimated firstorder rate constant (h-1); τˆ is the estimated lag time (h) and enteatea is the model residual fraction at time point ntea (mol). In the simultaneous fitting of the placebo function (eq 3) and the composite function (eq 4), βˆ is the common model parameter for estimating the basal metabolic levels in both intervention periods. tea The model parameters (ke, τ, β, cxmax , Rpla, Rtea) can be found by solving the least-squares minimalization criterion given in eq 5: min(|epla | 2 + |etea | 2)
(5)
where epla is a Npla data vector containing the model residuals of the linear placebo function (eq 3) and etea is a Ntea data vector containing the model residuals of the composite function (eq 4). With the use of the optimized model parameters, the net cumulative urinary excretion (netcxˆtea in mol) of each metabolite j (48 h after the tea intervention) was derived from the nutrikinetic model (eq 4) according to eq 6: c tea ˆ net x
ˆ
tea ) cxˆmax (1 - e-ke(48-τˆ))
[mol]
(6)
Error Estimation of the Fitted Parameters. The jackknife sampling scheme was used to estimate the error of the fitted parameters. The basic idea behind the jackknife estimator is to systematically recompute the error estimate leaving out one sample at a time. In the current study, the replacement of the (time-point) samples from the placebo period and the treatment period was done pairwise. This pairwise approach was applied because the fit parameters were simultaneously calculated in two dependent models (eq 3 and 4). In the sampling scheme used here, all possible sample pairs (1 placebo sample + 1 treatment sample) were left out of the data set once. From the subsets of the remaining samples, multiple estimations for tea the parameters (ke, τ, β, cxmax , Rpla, Rtea) were calculated. The resamplings were used to evaluate each of the estimators (θ) and the jackknife estimate of standard deviation (sjack). In the estimation of sjack, a bias correction factor (Npla × Ntea - 2)/ (Npla × Ntea) for each metabolite j and per subject i was taken into account.43 The estimated (mean) fit parameters for each subject given in the remainder of the paper were therefore represented as θˆ ( ˆsjack. Specificity of the Nutrikinetic Analysis. To test whether the cumulative output of the biomarkers indeed derive from the black tea intake and not from diuretic activity,44 a comparison was made against the urinary excretion of creatinine, lactate, betaine and oxaloacetate. These endogenous metabolites were selected because they have not shown elevated urinary levels 3322
Journal of Proteome Research • Vol. 8, No. 7, 2009
in a previous human intervention study with black tea.45 We therefore assumed a constant urinary excretion of these metabolites for the duration of the sampling period. In this test, a positive result was considered if the black tea treatment did not lead to a net endogenous effect after fitting the nutrikinetic model. The proportion of the actual positives was used as a statistical measure to express the specificity of the selected biomarkers.46,47 The false positive test results on the other hand represent the type I error in the analysis. Metabolic Phenotyping. The estimated nutrikinetic quantities for the selected metabolites were used to describe the phenotypes in the test population. To put the low- and medium-abundance nutrikinetic quantities for all metabolites on a similar scale as the high-abundance ones, the data columns were first normalized to their maximum values. As a result, variations in the nutrikinetic parameters became independent on their absolute quantities, keeping the data structure intact.48 Dependent on the estimated (normalized) net cumulative output levels of the biomarkers after 48 h (netcxˆtea), differences between strong metabolizer phenotypes and poor metabolizer phenotypes could be assessed. Strong metabolizers were designated to subjects that were able to extensively metabolize polyphenols, while those subjects with a rather deficient metabolism were termed poor metabolizer.49 In a similar way, a distinction of slow- and fast metabolizer phenotypes (based on kˆe), and slow- and fast responder phenotypes (based on τˆ) was made.
Results Subjects in Study. All participating subjects (20) successfully completed the study. Analytical Test Procedure. The estimated hippurate levels in the 7 QC samples indicated that the analytical procedure (sample handling, NMR acquisition and quantification) performed within the 10% QC acceptance interval. The relative differences between the estimated and the expected concentration levels varied between -5.4% and +7.5%, whereas the mean relative difference was 1.4%. Effect of Tea Consumption. To investigate whether the consumption of black tea leads to systematic and significant metabolic changes in the urinary NMR spectra, multilevel PLSDA was performed in combination with cross model validation (CMV) and permutation testing. CMV of the multilevel PLSDA model showed that on average 7.9 out of 40 samples were misclassified (prediction error ) 19.5%) with a DQ2 value of 0.33. In Figure 3A, the DQ2 value of the multilevel model is compared with the permutations (represented as a H0 distribution of no-effect). The p-value obtained from this permutation test (p < 0.0005) indicated that the classification model (and the selected biomarkers) can be considered statistically significant (R ) 0.05). Also, based on other classification criteria, that is, Q2, Area under the ROC curve (AUROC) and the Number of Misclassifications, a significant treatment effect was observed (Supplementary Figure 1 in Supporting Information). Identification of NMR Signals. From the rank product (RP, Figure 3B,C), a small list of discriminative urinary biomarkers associated with tea intake could be identified (Supplementary Table 2 in Supporting Information). Only 256 biomarker signals were considered significance limit (p < 0.1, Figure 4D). The majority of the biomarker signals had chemical shift values between δ 6.0 ppm and δ 8.0 ppm, and could be assigned to aromatic metabolites (Figure 4). Three of these signals (δ 7.78 ppm, d; δ 7.59 ppm, t and δ 7.50 ppm, t) derived from the
Phenotyping Tea Consumers by Nutrikinetic Analysis
research articles
Figure 3. DQ2 prediction errors of (A) the multilevel PLS-DA model (red) and 2000 permutations (blue) determined after Cross Model Validation. A significant effect was observed (p < 0.0005). The (B) associated rank product values (RP1/20) and (C) the insert, demonstrated that approximately 5% of all NMR signals were important in the discrimination between the placebo and the treatment groups. The p-values of the NMR signals shows that approximately (D) 256 out of 4400 were significantly on the 10% level (dashed line, R ) 0.10). Journal of Proteome Research • Vol. 8, No. 7, 2009 3323
research articles
van Velzen et al.
Figure 4. Spectral comparison between (A) the rank product values (RP1/20) and (B) a representative urinary 1H NMR spectrum (subject 7) acquired after black tea consumption. The dotted lines represent the assignments of the most distinctive biomarkers in the classification between the tea and placebo group. H ) hippuric acid; 4H ) 4-hydroxyhippuric acid; DHPS ) 1,3-dihydroxyphenyl-2-O-sulfate; U ) unknown.
aromatic-ring protons of hippuric acid.22,29,45 Also, the 4-hydroxy derivate of hippuric acid (4-hydroxyhippuric acid) was one of the most important contributors to the RP. The isolated aromatic signal of 4-hydroxyhippuric acid resonating at δ 7.71 ppm (d) was used for quantification. A third NMR signal at δ 6.55 ppm (d) could be assigned to 1,3-dihydroxyphenyl-2-Osulfate. This molecule has been previously identified in human urine after black tea consumption.29 Some other distinctive aromatic resonance patterns with low rank products were observed. Particularly in the region δ 6.9-7.0 ppm, several biomarker signals could be recognized that were significantly elevated after the tea intake. The low signal strengths, however, hampered an unambiguous assignment of these resonances. Nevertheless, some tentative assignments could be made in combination with a supporting GC-MS study that specifically aimed at the analysis of polyphenol metabolites in biofluids28 (Supplementary Table 2 in Supporting Information). Nutrikinetic Analysis of the Excretion. The time-course of the urinary excretion of hippuric acid, 4-hydroxyhippuric acid and 1,3-dihydroxyphenyl-2-O-sulfate was followed over 48 h after the intervention. Figure 5 shows the excretion of the 1,3dihydroxyphenl-2-O-sulfate in the treatment period following black tea intake in comparison with the placebo measurements. 3324
Journal of Proteome Research • Vol. 8, No. 7, 2009
In Figure 5A, the cumulative NMR signals are shown for a single subject (subject 10), whereas in Figure 5B, the excretion curves of all subjects (s1-s20) are illustrated. Figure 5B also demonstrates the large variability in the excretion characteristics of this metabolite among the subjects. Interestingly, almost similar excretion profiles of two subjects (s2 and s12) were recorded after the tea and the placebo interventions. This observation suggests that polyphenol metabolism, bioavailability and ADME were substantially different from the other subjects in the study population. This may ultimately influence the effectiveness of the tea treatment. Determination of Nutrikinetic Parameters. The composite nutrikinetic model (eq 4) in combination with a linear baseline function (eq 3) was used to fit the urinary excretion of the identified biomarkers in all 20 subjects (Supplementary Figures 2-4 in Supporting Information). An example is given in Figure 6A where the measured and fitted amount of hippuric acid for subject 10 is plotted after placebo and black tea consumption. On the basis of the first-order exponential model in Figure 6B, the net molar output of hippuric acid (netcxˆtea) was estimated at 1316 ( 43 µmol (θˆ ( ˆsjack) after black tea intake. The associated rate constant (kˆe) and the delay time (τˆ) were 0.14 ( 0.02 h-1 (θˆ ( ˆsjack) and 8 ( 1 h ( θˆ ( ˆsjack) respectively, suggesting a delayed and slow excretion of this biomarker. In Figure 6C, the
research articles
Phenotyping Tea Consumers by Nutrikinetic Analysis
Thus, the output levels of the 3 investigated metabolites tended to be low in the poor metabolizer phenotypes, whereas the strong metabolizer phenotypes exhibited increased output levels. This general tendency was also observed when the phenotypic differentiation was based upon the (normalized) first-order rate constant, kˆe (Figure 7B), and the (normalized) delay time, τˆ (Figure 7C). A second trend was observed when the relationship between the normalized values of netcxˆtea and kˆe was investigated. As shown in Figure 7D, the cumulative output of the metabolites had the tendency to be inversely related to the excretion rate. Thus, the metabolites with relative high urinary output levels were generally characterized by relative low excretion rates (and vice versa). This relationship was particularly evident for 4-hydroxyhippuric acid, for which c a reciprocal trend between net xˆtea and kˆe was observed.
Discussion
Figure 5. The cumulative urinary excretion of 1,3-dihydroxyphenyl-2-O-sulfate during 48 h after black tea intervention (black) and placebo intervention (red). The nutrikinetics is demonstrated (A) for subject 10 over 8 time-points by means of the aromatic NMR signal at δ 6.55 ppm, and (B) for all subjects (s1-s20) by means of the cumulative excretion curves.
residual error of the predicted output is shown. The residuals do not show a time-dependent error structure and therefore suggested that the nutrikinetic model used here was a likely match to the data. The estimated nutrikinetic quantities for the selected biomarkers, and for all subjects, are given in Table 1. Specificity of Nutrikinetic Analysis. To test whether the output of the selected biomarkers was a direct effect of the tea intake, a similar nutrikinetic analysis was performed on some endogenous metabolites that were assumed to have a constant urinary excretion for the duration of the sampling periods (creatinine, lactate, betaine and oxaloacetate). After nutrikinetic analysis, in 80-90% of all 20 subjects, no significant endogenous effects were observed. The proportion of these true positives was used to express the specificity of the selected biomarkers (80-90%), whereas the false positive error rate (10-20%) represented the type I error in the analysis. Phenotyping Based on Nutrikinetic Properties. On the basis of the urinary clearance of hippuric acid, 4-hydroxyhippuric acid and 1,3-dihydroxyphenyl-2-O-sulfate, a differentiation between the strong- and poor metabolizers in the test group was made. The differentiation is illustrated in Figure 7A where the subjects are displayed as points and representing the (normalized) cumulative output of the urinary end-metabolites after 48 h. A uniform (continuous) distribution of metabolizer phenotypes was observed rather than a bimodal distribution of two metabolizer subpopulations. Furthermore, the metabolites had a tendency to follow a similar excretion characteristic.
The concept of integrating metabolomics and nutrikinetics for describing metabolic phenotypes was demonstrated for the case of representative human nutritional intervention on a small subset of urinary metabolites. The nutrikinetic parameters used for the phenotyping were estimated by facilitating a one-compartmental nutrikinetic analysis with first-order kinetics, a lag time and a baseline function whereby the crossover design in the data was fully exploited. The jackknife error estimate of the nutrikinetic parameters (netcxˆtea, kˆe, τˆ) and the error structure of the fitted output curves suggested that the model introduced here was a likely match to the data. Among the urinary metabolites that were significantly elevated after the intake of black tea polyphenols, three gut microbial metabolites were selected for the nutrikinetic analysis, that is, hippuric acid, 4-hydroxyhippuric acid and 1,3dihydroxyphenyl-2-O-sulfate. These phenolic compounds are known metabolites from colonic microbial fermentation of dietary polyphenols.22,29,45,50-52 The urinary excretion for these metabolites was generally slow among the subjects (kˆe< 1 h-1) and delayed (τˆ ∼ 8-10 h). This suggesed that the excretion was mainly rate-determined by the microbial degradation in the gut, and not so much by phase II metabolism in the liver, small intestine or kidney. With the currently applied dose of black tea, the maximum level of polyphenols in the circulation would be limited to approximately 0.16 µg/mL (i.e., 800 mg in 5 L of blood), and it has been shown that at such relatively low levels glycine conjugation42 and sulfatation53-55 usually occur at much faster rates (ke > 1 h-1). After oral dosing of relatively simple phenolic acids, like benzoic acid,42 resveratrol,53 benzophenone,54 or ferulic acid,56 the urinary excretion of the phase II metabolites typically starts within a few hours after the administration (
