Normalization to Specific Gravity Prior to Analysis Improves

Oct 6, 2014 - Out of these, 19 samples were further selected based on simultaneously high or non/low consumption of six specific foods (coffee, red wi...
0 downloads 16 Views 1MB Size
Article pubs.acs.org/ac

Normalization to Specific Gravity Prior to Analysis Improves Information Recovery from High Resolution Mass Spectrometry Metabolomic Profiles of Human Urine William M. B. Edmands,† Pietro Ferrari, and Augustin Scalbert* Biomarkers Group, Nutrition and Metabolism Section, International Agency for Research on Cancer (IARC), 150 cours Albert Thomas, 69372 Lyon Cedex 08, France S Supporting Information *

ABSTRACT: Extraction of meaningful biological information from urinary metabolomic profiles obtained by liquidchromatography coupled to mass spectrometry (MS) necessitates the control of unwanted sources of variability associated with large differences in urine sample concentrations. Different methods of normalization either before analysis (preacquisition normalization) through dilution of urine samples to the lowest specific gravity measured by refractometry, or after analysis (postacquisition normalization) to urine volume, specific gravity and median fold change are compared for their capacity to recover lead metabolites for a potential future use as dietary biomarkers. Twenty-four urine samples of 19 subjects from the European Prospective Investigation into Cancer and nutrition (EPIC) cohort were selected based on their high and low/nonconsumption of six polyphenol-rich foods as assessed with a 24 h dietary recall. MS features selected on the basis of minimum discriminant selection criteria were related to each dietary item by means of orthogonal partial leastsquares discriminant analysis models. Normalization methods ranked in the following decreasing order when comparing the number of total discriminant MS features recovered to that obtained in the absence of normalization: preacquisition normalization to specific gravity (4.2-fold), postacquisition normalization to specific gravity (2.3-fold), postacquisition median fold change normalization (1.8-fold increase), postacquisition normalization to urinary volume (0.79-fold). A preventative preacquisition normalization based on urine specific gravity was found to be superior to all curative postacquisition normalization methods tested for discovery of MS features discriminant of dietary intake in these urinary metabolomic datasets.

M

conversion to a homoscedastic noise structure (i.e., uniform variance across the intensity scale).7,8 Other sources of unwanted variation are specific to the sample analyzed. Unlike biological fluids such as peripheral blood and cerebral spinal fluid that are homeostatically regulated, urine volume and solute concentrations vary greatly according to hormonal, physiological, dietary and behavioral factors.9−11 Variations in volume as high as 15-fold are commonly observed12 and this may introduce a major bias when analyzing metabolomic data unless a proper normalization method is applied. A typical urine normalization method uses creatinine, a breakdown product of muscle proteins constitutively excreted in urine. However, this single point normalization method proved to be largely unsuitable, as creatinine excretion varies greatly according to muscle mass, physical activity and renal impairment.13,14

etabolomic investigations utilize observations made by mass spectrometry (MS) or nuclear magnetic resonance (NMR) spectroscopy and multivariate chemometric methods to identify often subtle modifications of metabolic homeostasis due to environmental exposures or disease etiologies.1 Untargeted high resolution MS coupled to ultrahigh performance liquid chromatography (UHPLC) has become the most widely used technique for acquisition of metabolomic data. It allows the simultaneous detection of thousands of low molecular weight metabolites with a high degree of sensitivity, resolution, reproducibility and a broad dynamic range.2−6 Various sources of unwanted variations may still affect the quality of the data. A major source of variation in metabolomic datasets is multiplicative analytical noise, which increases according to relative MS signal intensity. More intense mass spectral signals are prone to incorporate a greater proportion of analytical noise compared to lower intensity signals. This multiplicative noise structure can readily dominate subsequent chemometric modeling. A generalized log transformation is crucial to ameliorating this dominant unwanted variance by © 2014 American Chemical Society

Received: August 25, 2014 Accepted: October 6, 2014 Published: October 6, 2014 10925

dx.doi.org/10.1021/ac503190m | Anal. Chem. 2014, 86, 10925−10931

Analytical Chemistry

Article

on their richness in polyphenols, major antioxidants of the diet.20

A number of other normalization methods have been proposed for urine analysis and can be categorized into postacquisition curative8 and preacquisition preventative15−17 methods. Preacquisition methods differ from postacquisition methods as urine concentrations are normalized by dilution or reconstitution before analysis. Veselkov et al. compared four methods for postacquisition normalization and multiplicative noise correction for urine.8 Median fold change normalization (MFCN) or probabilistic quotient normalization were found to be optimal for correction of systematic variations of urine dilution. Warrack et al. also compared four methods of postacquisition normalization to total urinary volume, creatinine, osmolality and mass spectral total usable signal (MSTUS; the total intensity of reproducible peaks common to all samples).12 They found that normalization to osmolality or MSTUS provided the best discrimination between groups of rats fed two doses of a toxic agent. More recently, Chen et al. also showed that several methods of postacquisition normalization applied to serially diluted urine samples failed to correct for variations in urine concentrations, due to signal saturation or ion suppression in most concentrated samples and default of detection of some metabolites in most diluted samples.17 In contrast to postacquisition normalization methods, Mattarucchi et al. investigated the use of MSTUS as a corrective factor to dilute and normalize urine concentrations before their analysis by MS.16 This preacquisition normalization method improved the predictive power of orthogonal partial least-squares discriminant analysis (OPLS-DA) models. However, no comparison was made with other normalization methods. Furthermore, a limitation of this preacquisition method of normalization is that it requires preliminary analysis of the urine samples to measure MSTUS. This appears impractical when large series of samples are analyzed, as commonly done in epidemiological studies. Chen et al. also recently showed that preacquisition normalization to creatinine level improved grouping of rat urine samples in PCA analyses.17 In this work, we compare the effects of preacquisition normalization to specific gravity (as a surrogate of total solute concentration, preSGN) and three methods of postacquisition normalization to urinary volume (VolN), MFCN and specific gravity normalization (SGN) in a study involving free-living individuals of the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort and aiming at the identification of novel dietary biomarkers. Specific gravity measured by refractometry was found to be an excellent alternative to the independent measurement of MSTUS.18

Table 1. Reported Dietary Intakes of Six Polyphenol-Rich Foods in 19 Subjects from the EPIC Cohort Selected for the Studya high consumers

coffee red wine tea citrus fruits apple and pear chocolate-containing products a

non/low consumers

n

mean intake (g/24 h)

n

mean intake (g/24 h)

10 8 5 6 10 5

985 366 1132 220 400 163

9 8 5 6 9 5

45 0 0 0 0 0

n, Number of (non) consumers.

Top and bottom categories of food intake 24-HDR measurements were considered as discriminatory classes (Yvariables) by partitioning the data into both “high-consumer” and “non/low-consumer” to facilitate discriminant MS feature identification. The mean 24 h urinary volume was 1859 ± 811 mL and the mean specific gravity was 1.018 ± 0.008. In addition, first-pass urine samples collected from three subjects in the morning following a polyphenol-rich dinner were used as positive controls. The polyphenol-rich meal included a cup of coffee, a cup of green-tea, a glass of red wine, an orange, an apple and two squares (50 g) of dark chocolate. Urine samples collected in the same conditions after a polyphenol-free meal were used as negative controls. These samples provided a means to filter potential discriminant MS features identified by chemometric modeling of the data obtained on the cohort sample subset. Preacquisition Specific Gravity Measurement and Sample Preparation. For un-normalized analyses (NoN), urine samples (40 μL) were diluted 8-fold with LC-MS grade water (Optima from Fisher, France). Samples were vortexed and centrifuged (13000g, 10 min). The supernatant (320 μL) was transferred to an LC-MS grade high recovery glass vial (Agilent Technologies Inc., France). A quality control (QC) sample was created by pooling aliquots (100 μL) of all the 19 urine samples. For prenormalized analyses, urine samples were allowed to thaw and were centrifuged (13000g, 10 min). An aliquot of the supernatant (100 μL) was placed upon the lens aperture of a digital refractometer (Euromex clinical digital refractometer RD.5712, NL) previously calibrated with LC-MS grade water, to measure specific gravity. Once all specific gravity measurements had been recorded, a dilution scheme was defined to bring all urine samples to the lowest urinary density measured (d = 1.006; for example, a sample with a density of 1.042 was diluted 7-fold). All samples were then further diluted 8-fold as previously described. Samples were ordered for pipetting according to increasing specific gravity to reduce potential experimental error. Urine samples were diluted with LC-MS grade water and centrifuged (13000g, 10 min). It was checked that all samples were accurately diluted to the lowest specific gravity (1.006). The supernatant (350 μL) was transferred to a high-recovery glass vial and randomized for LC-MS analyses. LC-MS Data Acquisition. Data was acquired using an Infinity 1290 binary ultrahigh performance liquid chromatog-



EXPERIMENTAL SECTION Solvents. UHPLC grade water and methanol (Fisher, Optima brand for LC-MS, France) and formic acid (SigmaAldrich, France) were used for sample preparation and chromatography. Study Design. Urine samples were selected from a subset of participants from the EPIC calibration study where both a 24 h urine collection and 24 h dietary recall (24-HDR) taken on the same day were available (n = 480).19 Completeness of the 24 h urine collection was monitored using p-amino benzoic acid (PABA). Out of these, 19 samples were further selected based on simultaneously high or non/low consumption of six specific foods (coffee, red wine, tea, orange, apple and chocolatecontaining products) of the subjects as reported in the 24 h dietary recalls (Table 1). These six foods were selected based 10926

dx.doi.org/10.1021/ac503190m | Anal. Chem. 2014, 86, 10925−10931

Analytical Chemistry

Article

2, mzwid = 0.015, minfrac = 0.5. The fillpeaks function was used to integrate missing peak intensities in samples following grouping using raw data from the corresponding peak group area. Consistent mobile phase background signals were removed by considering only mass spectral features >1.5-fold higher in the QCs compared to triplicate injected blanks. Reproducible mass spectral features (CV 0.3 was considered an acceptable fit of each OPLS-DA model. OPLS-DA models failing to meet this minimum, returning a negative Q2 or no model were therefore excluded from further analysis (see Table S1 of the Supporting Information for a summary of OPLS-DA modeling results). Features were selected above minimum filtration thresholds of 0.6 OPLS-DA loadings coefficient and 0.025 magnitude of modeled covariance. The variable influence on projection (VIP) plot of each mass spectral variable (>1.5) and its 99% confidence interval (>0.1 after subtraction of the 7fold jack-knifed confidence interval) were used to further filter the significant OPLS-DA loadings coefficients. A final filtration of discriminant MS features was made using data from the subjects consuming a polyphenol-rich dinner composed of the same six foods. MS features with an average >1.5 fold increased intensity following consumption of the polyphenol-rich foods were selected. This filtration was used to limit potential false positives in OPLS-DA models built with the samples of the cohort. The total number of discriminant MS features were reduced by 15.4%, 13.5% and 11.5% for the VolN, MFCN and SGN normalization methods, respectively, and provided further reassurance in the mass spectral features identified. A script was created in the R language to perform data preprocessing including generalized log transformation, postacquisition normalization methods and reproducible feature filtration. The preprocessed data was then exported for multivariate data analysis in the SIMCA-P+ (v. 12.0.1.0, Umetrics AB, Umea, Sweden) software. The results of OPLSDA and VIP analyses in SIMCA-P+ were saved as .txt files for each model. The text files for each OPLS-DA model and the

raphy (UHPLC) system (Agilent Technologies Inc., Santa Clara, CA, USA) coupled to a 6550 “ion funnel” quadrupoletime-of-flight (Q-TOF) mass spectrometer (Agilent Technologies Inc., Santa Clara, CA, USA). Two reference masses, tetrafluoroacetic acid fragment ion [M − H]− at m/z 112.985 587 and HP-921 [hexakis-(1H,1H,3H-tetrafluoropentoxy)phosphazene] [M + FA − H]− at m/z 966.000725, were simultaneously sprayed into the source by a separate nebulizer using the Agilent “Dual AJS” (Agilent Jet Stream) electrospray source. This dual-nebulizer provides a stable lock mass at both high and low masses with a low flow of reference mass solution from a dedicated nebulizer providing a typical mass resolving power of 27 000 ± 500 (m/z 966). Mass spectral source parameters were the following: fragmentor voltage, 175 V; sheath gas temperature, 400 °C; sheath gas flow, 12 L·min−1; nebulizer pressure, 45 psig; capillary voltage, 3500 V; nozzle voltage, 300 V; drying gas temperature, 290 °C; drying gas flow, 12 L·min−1; scan rate, 1.67 spectra·s−1. We assessed the stability and reproducibility of the UHPLCQTOF system prior to sample acquisition by 70 injections of the pooled QC urine sample to test for chromatographic drift and signal attenuation over the course of long-term constant operation of the system in negative mode. This test demonstrated the sensitivity (10 827 features detected including 8610 (approximately 80%) with CV 0.1) following subtraction of the 99% confidence interval. The number of discriminant MS features matching the threshold filtration criteria were identified for every model (Table 2). Postacquisition normalization to the urinary volume Table 2. Number of MS Features in Urine of Free-Living Subjects from the EPIC Cohort Discriminating Consumers and Non/Low Consumers of Six Polyphenol-Rich Foods, Using Different Postacquisition and Preacquisition Normalization Methodsa number of discriminant MS features

coffee NoN MFCN VolN SGN PreSGN

75 60 77 90 169

red wine 83 120 88

tea

citrus fruits

chocolatecontaining products

apple and pear

18 21 1 9 86

2 7 3 4 71

11 15

2

Discriminant features were identified and filtered as described in the Experimental Section. Abbreviations: NoN, un-normalized; MFCN, median fold change normalized; VolN, post-acquisition normalized to urinary volume; SGN, postacquisition specific gravity normalized; PreSGN, preacquisition specific gravity normalized. a

resulted in poor subsequent modeling of the data and negative values of the Q2 cross validation parameters for two of the six food categories (red wine and chocolate-containing products) and a Q2 close to zero (0.064: apple and pears). A negative Q2 was also calculated for one of the six food categories (chocolate-containing products) when using postacquisition median fold change and a poor Q2 (0.176: apple and pears). A summary of the OPLS-DA models can be found in the Supporting Information (Table S1). Normalization methods ranked in the following decreasing order when comparing the number of total discriminant MS features recovered to that obtained without normalization: preacquisition normalization to specific gravity (4.2-fold), postacquisition normalization to specific gravity (2.3-fold), postacquisition median fold change normalization (1.8-fold), postacquisition normalization to urinary volume (0.79-fold). When the postacquisition normalization methods were compared in terms of potential for discriminant MS feature identification, normalization based on specific gravity as a surrogate measurement of total solute concentration gave a larger number of discriminant MS features and appeared preferable to both of the other methods of postacquisition normalization (Table 2). The nature of mass spectral features characteristic of dietary intakes also varied according to the normalization methods. Significant overlap in significant features can be seen between the postacquisition normalization methods applied to the same dataset (Figure 3A). Of these features, 12% were shared among all normalization methods, and 50% were shared between the MFCN and SGN datasets. On the basis of these results, the use of urinary volume as a postacquisition normalization, even for 24 h urine collection, should be avoided and MFCN and SGN normalization preferred. Preacquisition normalization by dilution of urine according to their respective specific gravity provided the largest number of potentially discriminant MS features with 1.4-fold more MS

Figure 2. OPLS-DA S-plots showing potential of different normalization methods for identification of discriminant MS features of coffee intake within urine samples. (A) Un-normalized data (R2X, 0.49; R2Y, 0.82; Q2, 0.37; 19 observations; 13 932 variables). (B) Postacquisition median fold change normalization (R2X, 0.23; R2Y, 0.813; Q2, 0.35; 19 observations; 13 932 variables). (C) Preacquisition normalization by sample dilution to a common specific gravity (R2X, 0.35; R2Y, 1; Q2, 0.75; 19 observations; 11 822 variables). All data were generalized logtransformed and pareto-scaled. The S-plot shows the reliability of the modeled correlation (p(corr)[1]) on the x-axis against the modeled covariation or variable magnitude (p[1]). The gray boxes highlight highly correlated mass spectral features in the loading with lower risk of spurious assignment (>0.6 OPLS loadings coefficient and >0.025 covariance). 10929

dx.doi.org/10.1021/ac503190m | Anal. Chem. 2014, 86, 10925−10931

Analytical Chemistry

Article

series of samples. Preacquisition normalization to creatinine was recently proposed.17 However, creatinine level is also known to vary according to muscle mass and any factor that may affect muscle mass such as age, sex or BMI.28 We establish here the feasibility and benefit of preacquisition normalization to specific gravity on recovery of biological information. Normalization was achieved through dilution of urine samples to the lowest specific gravity value found among samples before mass spectrometry analysis. Recently normalization by differential sample volume injection was also proposed and this may further simplify sample preparation.17 Preacquisition normalization to specific gravity resulted in a significant increase in the number of discriminant MS features detected (Figure 3B) that can be explained by a leveling-off of signal intensities over all urine samples. Large differences in solute concentrations would otherwise result in large differences between samples in the detection of peaks with intensities close to the noise.15,16 Variations in total solute concentration between urine samples may also result in differential charge competition and ion suppression effects in the electrospray source of the mass spectrometer. Charge competition in biological matrixes such as urine is not only a factor of linear changes in all metabolite concentrations as in a differential dilution of a technical replicate but is also very much dependent on the complex combinations of solutes and their concentrations.29−32 It is thus erroneous to make the assumption that all analytes in a complex matrix will necessarily have a linear relationship with their concentration in solution when measured by electrospray mass spectrometry, as assumed with postacquisition normalization methods such as MFCN. For these reasons, preacquisition normalization is also expected to improve the comparability of urine samples.

Figure 3. Area-proportional Venn diagrams showing the number of MS features discriminant of dietary intake identified in urine by OPLSDA modeling using different normalization methods. (A) Overlap in the nature of discriminant MS features identified using three postacquisition normalization methods: normalization to urine volume (VolN; yellow circle), median fold change normalization (MFCN; red circle) and postacquisition normalization to urine specific gravity (SGN; green circle). (B) Area-proportional Venn diagrams showing the number of discriminant MS features using the best identified postacquisition method (SGN; green circle) against the preacquisition to specific gravity normalization method (PreSGN, blue circle). Also included is the breakdown of discriminant MS features of the specific food intakes: (i) coffee, (ii) red wine, (iii) tea, (iv) citrus fruits, (v) chocolate-containing products, (vi) apple and pear. Data were glog transformed and pareto-scaled. Areas of the circles are proportional to number of unique features matching the discriminant MS feature identification criteria.



features when compared to SGN normalization (Figure 3B). The two independently acquired datasets demonstrated a 30% overlap between discriminant MS features identified by both normalization methods. Notable improvements in discriminant MS feature recovery were also seen for all food categories considered individually with 2.0-, 2.6- and 2.7-fold increases in the number of discriminant MS features for respectively, coffee, tea and citrus fruits. An exception was seen for red wine where preacquisition SGN normalization did not improve discriminant MS feature recovery. It was also negligible for chocolatecontaining products and apple and pear. Different normalization methods have been proposed to analyze the urine metabolome and take into account the large variations in solute concentrations. Some authors used MSTUS and osmolality to normalize signal intensities after data acquisition.16 Both methods showed an improved discrimination of urine samples when compared to normalization over urine volume.12 In the present work, we also found that urine MSTUS was poorly correlated to urine volume (r = 0.55) but strongly correlated with specific gravity (r = 0.96) easily measured at limited cost with a refractometer. SGN provided the highest number of discriminant MS features when compared to MFCN or VolN (respectively, 1.3- and 2.9-fold increases; Figure 3A). Significantly, we found preSGN rather than postacquisition methods improved potential discriminant MS feature recovery (Figure 3B). MSTUS assessed by mass spectrometry has been used in a previous metabolomic study to normalize urinary total solute concentration before metabolomic analysis.16 However, this implies two successive mass spectrometry analyses and makes this approach largely impractical for studies with large

CONCLUSIONS This study demonstrates the advantages of a preacquisition normalization strategy for analyzing the urinary metabolome in order to maximize biomarker recovery and limit the loss of information that may result from the large variations in urine concentration. Methods of postacquisition normalization as used by most researchers have the potential to alter considerably the interpretation and outcome of metabolomic investigations on urine. Preacquisition normalization efficiently mitigates the effects of systematic sources of unwanted variance prior to chemometric modeling. This study shows its greatest potential for subsequent identification of biomarkers of dietary intake. However, this normalization approach has potentially broader implications for improving the outcome of urinary metabolomic investigations. This study highlights the importance of the choice of an adequate normalization strategy as part of the experimental design of metabolomic studies and its potential effect on biological inference.



ASSOCIATED CONTENT

S Supporting Information *

Additional information as noted in the text. This material is available free of charge via the Internet at http://pubs.acs.org.



AUTHOR INFORMATION

Corresponding Author

*A.S. E-mail: [email protected]. Tel.: +33 (0)4 72 73 80 95. Website: www.iarc.fr. 10930

dx.doi.org/10.1021/ac503190m | Anal. Chem. 2014, 86, 10925−10931

Analytical Chemistry

Article

Present Address

(22) Smith, C. A.; Want, E. J.; O’Maille, G.; Abagyan, R.; Siuzdak, G. Anal. Chem. 2006, 78, 779−787. (23) Prince, J. T.; Marcotte, E. M. Anal. Chem. 2006, 78, 6140−6152. (24) Dieterle, F.; Ross, A.; Schlotterbeck, G.; Senn, H. Anal. Chem. 2006, 78, 4281−4290. (25) Swinton, J. Vennerable: Venn and Euler area-proportional diagrams, R package version 2.0/r75, 2009. (26) Wold, S.; Antti, H.; Lindgren, F.; Ö hman, J. Chemom. Intell. Lab. Syst. 1998, 44, 175−185. (27) Trygg, J.; Wold, S. J. Chemom. 2002, 16, 119−128. (28) Barr, D. B.; Wilder, L. C.; Caudill, S. P.; Gonzalez, A. J.; Needham, L. L.; Pirkle, J. L. Environ. Health Perspect. 2005, 113, 192− 200. (29) Tang, K.; Page, J. S.; Smith, R. D. J. Am. Soc. Mass Spectrom. 2004, 15, 1416−1423. (30) Kebarle, P.; Verkerk, U. H. Mass Spectrom. Rev. 2009, 28, 898− 917. (31) Beach, D. G.; Gabryelski, W. Anal. Chem. 2013, 85, 2127−2134. (32) Enke, C. G. Anal. Chem. 1997, 69, 4885−4893.



School of Public Health, University of California, Berkeley, USA Author Contributions

The paper was written through contributions of all authors. All authors have given approval to the final version of the paper. Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS This work was supported by the European Union (NutriTech FP7-KBBE-2011-5 Grant #289511, EUROCAN FP7-KBBE2010.2.4.1-2 Grant #260791).



REFERENCES



(1) Dunn, W. B.; Broadhurst, D. I.; Atherton, H. J.; Goodacre, R.; Griffin, J. L. Chem. Soc. Rev. 2011, 40, 387−426. (2) Want, E. J.; Wilson, I. D.; Gika, H.; Theodoridis, G.; Plumb, R. S.; Shockcor, J.; Holmes, E.; Nicholson, J. K. Nat. Protoc. 2010, 5, 1005− 1018. (3) Gika, H. G.; Theodoridis, G. A.; Wingate, J. E.; Wilson, I. D. J. Proteome Res. 2007, 6, 3291−3303. (4) Gika, H. G.; Theodoridis, G. A.; Earll, M.; Wilson, I. D. Bioanalysis 2012, 4, 2239−2247. (5) Wong, M. C.; Lee, W. T.; Wong, J. S.; Frost, G.; Lodge, J. J. Chromatogr. B: Anal. Technol. Biomed. Life Sci. 2008, 871, 341−348. (6) Benton, H. P.; Want, E.; Keun, H. C.; Amberg, A.; Plumb, R. S.; Goldfain-Blanc, F.; Walther, B.; Reily, M. D.; Lindon, J. C.; Holmes, E.; Nicholson, J. K.; Ebbels, T. M. D. Anal. Chem. 2012, 84, 2424− 2432. (7) Durbin, B.; Hardin, J.; Hawkins, D.; Rocke, D. Bioinformatics 2002, 18, S105−S110. (8) Veselkov, K. A.; Vingara, L. K.; Masson, P.; Robinette, S. L.; Want, E.; Li, J. V.; Barton, R. H.; Boursier-Neyret, C.; Walther, B.; Ebbels, T. M.; Pelczer, I.; Holmes, E.; Lindon, J. C.; Nicholson, J. K. Anal. Chem. 2011, 83, 5864−5872. (9) Berliner, R. W.; Levinsky, N. G.; Davidson, D. G.; Eden, M. Am. J. Med. 1958, 24, 730−744. (10) Wilson, I. D.; Plumb, R.; Granger, J.; Major, H.; Williams, R.; Lenz, E. M. J. Chromatogr. B: Anal. Technol. Biomed. Life Sci. 2005, 817, 67−76. (11) Walsh, M. C.; Brennan, L.; Malthouse, J. P. G.; Roche, H. M.; Gibney, M. J. Am. J. Clin. Nutr. 2006, 84, 531−539. (12) Warrack, B. M.; Hnatyshyn, S.; Ott, K. H.; Reily, M. D.; Sanders, M.; Zhang, H. Y.; Drexler, D. M. J. Chromatogr. B: Anal. Technol. Biomed. Life Sci. 2009, 877, 547−552. (13) Saude, E.; Adamko, D.; Rowe, B.; Marrie, T.; Sykes, B. Metabolomics 2007, 3, 439−451. (14) Boudonck, K. J.; Rose, D. J.; Karoly, E. D.; Lee, D. P.; Lawton, K. A.; Lapinskas, P. J. Bioanalysis 2009, 1, 1645−1663. (15) Mattarucchi, E.; Guillou, C. Anal. Chem. 2011, 83, 9719−9720. (16) Mattarucchi, E.; Guillou, C. Biomed. Chromatogr. 2012, 26, 512−517. (17) Chen, Y.; Shen, G.; Zhang, R.; He, J.; Zhang, Y.; Xu, J.; Yang, W.; Chen, X.; Song, Y.; Abliz, Z. Anal. Chem. 2013, 85, 7659−7665. (18) Jacob, C. C.; Dervilly-Pinel, G.; Biancotto, G.; Le Bizec, B. Metabolomics 2014, 10, 627−637. (19) Slimani, N.; Bingham, S.; Runswick, S.; Ferrari, P.; Day, N. E.; Welch, A. A.; Key, T. J.; Miller, A. B.; Boeing, H.; Sieri, S.; Veglia, F.; Palli, D.; Panico, S.; Tumino, R.; Bueno-de-Mesquita, B.; Ocke, M. C.; Clavel-Chapelon, F.; Trichopoulou, A.; van Staveren, W. A.; Riboli, E. Cancer Epidemiol., Biomarkers Prev. 2003, 12, 784−795. (20) Perez-Jimenez, J.; Neveu, V.; Vos, F.; Scalbert, A. J. Agric. Food Chem. 2010, 58, 4959−4969. (21) Kessner, D.; Chambers, M.; Burke, R.; Agus, D.; Mallick, P. Bioinformatics 2008, 24, 2534−2536.

NOTE ADDED AFTER ASAP PUBLICATION This paper was published ASAP on October 17, 2014, with errors in the TOC figure and in Figure 3. The corrected version was reposted on October 21, 2014.

10931

dx.doi.org/10.1021/ac503190m | Anal. Chem. 2014, 86, 10925−10931