Hydrophilic Interaction Chromatography for Mass Spectrometric

This was to ensure that the HPLC and MS systems had time to equilibrate and .... The remaining data formed the training set with response variables as...
8 downloads 0 Views 366KB Size
Download PDF
Anal. Chem. 2007, 79, 8911-8918

Hydrophilic Interaction Chromatography for Mass Spectrometric Metabonomic Studies of Urine Simon Cubbon, Timothy Bradbury, Julie Wilson,† and Jane Thomas-Oates*

Department of Chemistry, University of York, Heslington, York, YO10 5DD, UK

High-performance liquid chromatography (LC) coupled to mass spectrometry (MS) is increasingly being used for urinary metabonomic studies. Most studies utilize reversedphase separation techniques, which are not suited to retaining highly polar analytes. Metabonomic studies should encompass a representative “fingerprint” that contains the largest amount of information possible. In this work, we have analyzed human urine samples with LC-MS, comparing traditional reversed-phase separation with hydrophilic interaction chromatography (HILIC), using both positive and negative electrospray ionization modes. The resulting data were analyzed using principal components analysis and partial least-squares-discriminant analysis. Discriminant models were developed for the response variables gender, diurnal variation, and age and were evaluated using external test sets to classify their predictive ability. The developed models using both positive and negative ionization mode data for reversed-phase and HILIC separations were very comparable, indicating that HILIC is a suitable method for increasing the fingerprint coverage for LC-MS metabonomic studies. Metabonomics is defined as “the quantitative measurement of the dynamic multiparametric response of a living system to pathophysiological stimuli or genetic modification”1 and comprises a suite of “omics” technologies. The field has seen substantial growth in recent years, due perhaps to its success within the pharmaceutical industry, where metabonomics is now used for the identification of potential biological markers of disease, efficacy, and toxicity.2-7 A “true” metabonomic study should involve a comprehensive analysis with no preselection of analytes, in order to obtain as much information as possible. Proton nuclear magnetic resonance * To whom correspondence should be addressed. E-mail: [email protected]. Tel: +44 (0) 1904 434459. Fax: +44 (0) 1904 432516. † York Structural Biology Laboratory, Department of Biology. (1) Nicholson, J. K.; Lindon, J. C.; Holmes, E. Xenobiotica 1999, 29, 11811189. (2) Drexler, D. M.; Feyen, J. H. M.; Sanders, M. Drug Discovery Today Technol. 2004, 1, 17-23. (3) Lindon, J. C.; Holmes, E.; Bollard, M. E.; Stanley, E. G.; Nicholson, J. K. Biomarkers 2004, 9, 1-31. (4) Robertson, D. G.; Reily, M. D.; Baker, J. D. J. Proteome Res. 2007, 6, 526539. (5) Walgren, J. L.; Thompson, D. C. Toxicol. Lett. 2004, 377-385. (6) Wilson, I. D.; Plumb, R.; Granger, J.; Major, H.; Williams, R.; Lenz, E. M. J. Chromatogr., B 2005, 67-76. (7) Nicholson, J. K.; Conelly, J. C.; Holmes, E. E. Nat. Rev. Drug Discovery 2002, 1, 153-161. 10.1021/ac071008v CCC: $37.00 Published on Web 10/31/2007

© 2007 American Chemical Society

(1H NMR) spectroscopy is a highly reproducible technique and has been used extensively for metabonomic fingerprinting.7 1H NMR is nondiscriminatory and provides information about all the metabolites in a biological sample above the limit of detection. However, sensitivity can be a problem in NMR, and metabolites present in low concentrations may not be detected. Mass spectrometry (MS) on the other hand is a highly sensitive but selective technique. Used in conjunction with high-performance liquid chromatography (HPLC), which provides separation of the components, mass spectrometry allows detection and quantification of low-level metabolites.8 A comprehensive LC-MS study should utilize both positive and negative ionization modes with a chromatographic method that allows the retention and separation of as many components as possible. However, many LC-MS studies only use reversedphase (RP) chromatography, which instantly discriminates against highly polar analytes. Although only nonpolar and mildly polar analytes are retained, RP-LC-MS is still the most widely used metabonomic MS platform.4,6,9-15 As urine is predominantly aqueous, a significant proportion of the content is likely to be highly polar and would typically be unretained on RP systems and thus not contribute to the data obtained. Hydrophilic interaction chromatography (HILIC) is analogous to normal-phase (NP) chromatography in that it utilizes a polar stationary phase, allowing the retention of polar analytes.16 However, unlike NP, HILIC allows the use of aqueous solvents, making this separation technique compatible with ESI-MS. In direct contrast to RP-LC, gradient elution HILIC begins with a low-polarity organic solvent and elutes polar analytes by increasing the polar aqueous content. Compounds are retained by partitioning into a water rich layer, which is partially immobilized on the stationary phase. MS-compatible buffers are typically used to reduce any undesirable electrostatic interactions between the analytes and stationary phase.16 It is because of HILIC’s compat(8) Dettmer, K.; Aronov, P. A.; Hammock, B. D. Mass Spectrom. Rev. 2007, 26, 51-78. (9) Hodson, M. P.; Dear, G. J.; Roberts, A. D.; Haylock, C. L.; Ball, R. J.; Plumb, R. S.; Stumpf, C. L.; Griffin, J. L.; Haselden, J. N. Anal. Biochem. 2007, 362, 182. (10) Lu, G.; Wang, J.; Zhao, X.; Kong, H.; Xu, G. Chin. J. Chromatgr. 2006, 24, 109-113. (11) Lutz, U.; Lutz, R. W.; Lutz, W. K. Anal. Chem. 2006, 78, 4564-4571. (12) Plumb, R. S.; Granger, J. H.; Stumpf, C. L.; Johnson, K. A.; Smith, B. W.; Gaulitz, S.; Wilson, I. D.; Castro-Perez, J. Analyst 2005, 844-849. (13) Tang, H.; Wang, Y. Prog. Biochem. Biophys. 2006, 33, 401-417. (14) Williams, R. E.; Lenz, E. M.; Lowden, J. S.; Rantalainen, M.; Wilson, I. D. Mol. Biosyst. 2005, 1, 166-175. (15) Sumner, L. W. Biotechnol. Agric. For. 2006, 57, 21-32. (16) Hemstro ¨m, P.; Irgum, K. J. Sep. Sci. 2006, 29, 1784-1821.

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007 8911

ibility with MS and the ability to retain polar compounds16,17 that we have chosen to assess its suitability for application in LCMS-based metabonomic studies. In this work, we have analyzed urine from healthy human volunteers comparing a traditional RP-LC-MS approach with HILIC-LC-MS, using both positive and negative electrospray ionization modes. LC-MS analysis of biological samples creates vast quantities of data requiring multivariate statistical methods for interpretation. Both principal components analysis (PCA) and partial least-squares-discriminant analysis (PLS-DA) have been utilized to visualize the information-rich data.18 Here, PLS-DA was used to discriminate between sex, time of collection, and age in order to compare the performance of HILIC-LC-MS with the traditional RP-LC-MS approach. MATERIALS AND EXPERIMENTS Sample Collection. Urine was collected from 19 male and 11 female healthy volunteers, who had no restrictions placed upon their diet or lifestyle. The volunteers were aged from 22 to 61. Each volunteer was asked to provide two samples of urine, one being the first void of the day and the second, any void after 15:00 h. All samples were frozen at -80 °C within 2 h of collection. Previous studies have shown that both the pH and metabolite levels can change upon storage of urine.19-21 All samples were thus stored at -80 °C for a period of at least one week to allow any degradation of urinary components to be consistent across the whole sample cohort, as described by Saude and Sykes.21 Sample Preparation. Samples were thawed at room temperature and aliquoted into 1.5-mL microcentrifuge vials (Sarstedt) with 1 mL from each sample being taken to create a sample pool to provide a reference sample. The aliquoted samples were then refrozen at -80 °C, prior to any further sample preparation and analysis. Reversed-Phase Analysis. For reversed-phase analysis, the aliquoted samples were defrosted at room temperature and centrifuged at 10186g for 8 min. The supernatant was then collected and passed through a 0.45-µm syringe filter (VWR International) into sample vials fitted with 250-µL deactivated glass inserts (Agilent Technologies), ready for analysis. Hydrophilic Interaction Chromatography Analysis. For HILIC analysis, the aliquoted samples were defrosted at room temperature before being mixed in a 1:1 ratio with MeCN (Fisher Scientific). The addition of MeCN is necessary to avoid the creation of a pure water “plug”, which would greatly affect the chromatography. After the addition of MeCN, the samples were centrifuged at 10186g for 8 min before being filtered through a 0.45-µm syringe filter (VWR International) into sample vials fitted with 200-µL deactivated glass inserts (Agilent Technologies), ready for analysis. Sample Analysis. Before any samples were analyzed by either RP or HILIC, four identical aliquots of the pooled urine reference sample were analyzed. This was to ensure that the HPLC and (17) Idborg, H.; Zamani, L.; Edlund, P.-O.; Schuppe-Koistinen, I.; Jacobsson, S. P. J. Chromatogr., B 2005, 828, 9. (18) Wold, S.; Sjo ¨stro ¨m, M.; Eriksson, L. Chemom. Intell. Lab. Syst. 2001, 58, 109-130. (19) Fura, A.; Harper, T. W.; Zhang, H.; Fung, L.; Shyu, W. C. J. Pharm. Biomed. Anal. 2003, 513-522. (20) LeBeau, M. A.; Miller, M. L.; Levine, B. Forensic Sci. Int. 2001, 161-167. (21) Saude, E. J.; Sykes, B. D. Metabolomics 2007, 3, 19-27.

8912 Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

Table 1. LC Parameters for RP and HILIC Separationsa reversed phase

a

HILIC

time, min

%B

time, min

%A

0.0 9.0 21.0 24.0 27.0 30.0

5 20 95 95 5 5

0.0 15.0 19.0 19.5 30.0

5 95 95 5 5

Conditions: 20-µL injection, 600 µL min-1 flow rate.

Table 2. MS Parameters Used for Positive and Negative Modes of Ionization parameter

value

capillary voltage nitrogen nebulizing gas flow nitrogen drying gas flow drying gas temperature scan range

(2,500 V 3.3 L min-1 6.0 L min-1 300 °C m/z 40-1000

MS systems had time to equilibrate and were performing satisfactorily. All subsequent samples were analyzed in random order to control for analytical variation with time. A pooled reference sample or a blank was analyzed after every 10 samples to identify sample carryover and to check for reproducibility. Reversed-Phase Liquid Chromatography. For RP analysis, a Chromolith RP18e (100 × 4.6 mm) (Merck) was used along with a guard column (5 × 4.6 mm) (Merck) on an Agilent 1100 LC (Agilent Technologies). Mobile phase A was 0.1% formic acid v/v (Fisher Scientific), while mobile phase B was MeCN modified by the addition of 0.1% (v/v) formic acid. The gradient started with 5% mobile phase B, increasing to 20% B at 9 min, and then to 95% B at 21 min. The mobile phase was held isocratic for 3 min before returning to the starting conditions within 3 min (total run time was 30 min). The injection volume was 20 µL, and the column was eluted at a flow rate of 600 µL min-1 (summarized in Table 1). Hydrophilic Interaction Liquid Chromatography. For HILIC analysis, a ZIC-HILIC (3.5 µm, 100 × 4.6 mm) (SeQuant) column was used along with a guard column (20 × 2.1 mm) (SeQuant) on an Agilent 1100 LC (Agilent Technologies). Mobile phase A consisted of 5 mM ammonium acetate modified with 0.1% (v/v) formic acid (pH 4), while mobile phase B consisted of MeCN modified by the addition of 0.1% (v/v) formic acid. The gradient started with 5% mobile phase A increasing linearly to 95% over a period of 15 min. The mobile phase was held isocratic for 4 min before returning to the starting conditions within 30 s. The mobile phase was kept at 5% A for the remaining time to allow equilibration (total run time was 30 min). The injection volume was 20 µL, and the column eluted at a flow rate of 600 µL min-1 (summarized in Table 1). Electrospray Ionization Mass Spectrometry. An Applied Biosystems QSTAR pulsar i TOF mass spectrometer equipped with a TurboIonSpray source was used. The LC outlet from an Agilent 1100 series HPLC was directly coupled with no splitting. The capillary voltage was held at (2500 V depending upon the ionization mode; N2 nebulizing gas, 3.3 L min-1; and N2 drying

Figure 1. Typical UV254 chromatograms obtained on separation of the same urine sample using (a) reversed-phase and (c) hydrophilic interaction chromatography (inset are extracted ion chromatograms (m/z 114.07 ) [M + H]+) of creatinine; peak areas equal 132 450 and 103 300 counts for RP and HILIC, respectively). Corresponding positive mode total ion chromatograms for (b) reversed phase and (d) HILIC are shown, normalized to the most intense peak.

gas, 6.0 L min-1 at 300 °C. The MS was operated in full scan mode with m/z range of 40-1000 (summarized in Table 2). Data Extraction. Raw data were exported using the metabolomics export script (Applied Biosystems). Peaks files were created prior to generating a 3D data matrix of m/z and tR versus intensity for each sample analyzed. The following settings were used: tR tolerance, 0.5 min; LC peak width (min/max), 0.1/10 min; intensity threshold, 10 counts/s for positive ionization mode and 1 count/s for negative ionization mode; mass accuracy, 200 ppm; maximum peak number, 5000. The data were exported as a text file, ready for import into Excel (Microsoft) for further data manipulation. Data Normalization. Human biological samples exhibit a large amount of variation caused by physiological factors such as sex, state of health, age, diet, stress, or diurnal cycles among others.6,12,22-24 The extracted data thus require normalization to account for the variation and to give each sample equal importance in analysis by multivariate statistics. Despite this being an important step, many metabolomics/ metabonomics papers fail to identify or describe any data (22) Granger, J. H.; Plumb, R. S.; Stumpf, C. L.; Wilson, I. D.; Castro-Perez, J.; Major, H., Montreal, Canada, June 8-12, 2003. (23) Antti, H.; Ebbels, T. M. D.; Keun, H. C.; Bollard, M. E.; Beckonert, O.; Lindon, J. C.; Nicholson, J. K.; Holmes, E. Chemom. Intell. Lab. Syst. 2004, 139-149. (24) Lenz, E. M.; Wilson, I. D. J. Proteome Res. 2007, 6, 443-458.

preprocessing.17,25-27 The most common data normalization step involves scaling each sample according to the creatinine content.28-31 However, as creatinine levels may be perturbed by illness or disease, this normalization technique may not be ideal.30,32 A more recently adopted technique involves scaling to the total ion count of each sample. This scaling method is typically used for NMR data23,28 but has recently been applied to LC-MS data.8,30,33 We compared the two approaches and found that both methods give equivalent results. In this study, we have chosen to (25) Idborg, H.; Zamani, L.; Edlund, P.; Schuppe-Koistinen, I.; Jacobsson, S. P. J. Chromatogr., B 2005, 14-20. (26) Pan, Z.; Gu, H.; Talaty, N.; Chen, H.; Shanaiah, N.; Hainline, B.; Cooks, R.; Raftery, D. Anal. Bioanal. Chem. 2007, 387, 539-549. (27) Yang, J.; Xu, G.; Zheng, Y.; Kong, H.; Wang, C.; Zhao, X.; Pang, T. J. Chromatogr., A 2005, 214-221. (28) Constantinou, M. A.; Papakonstantinou, E.; Spraul, M.; Sevastiadou, S.; Costalos, C.; Koupparis, M. A.; Shulpis, K.; Tsantili-Kakoulidou, A.; Mikros, E. Anal. Chim. Acta 2005, 542, 169-177. (29) Felitsyn, N. M.; Henderson, G. N.; James, M. O.; Stacpoole, P. W. Clin. Chim. Acta 2004, 219-230. (30) Kind, T.; Tolstikov, V.; Fiehn, O.; Weiss, R. H. Anal. Biochem. 2007, 363, 185-195. (31) Schneider, U.; Schober, E. A.; Streich, N. A.; Breusch, S. J. Clin. Chim. Acta 2002, 81-88. (32) Heavner, D. L.; Morgan, W. T.; Sears, S. B.; Richardson, J. D.; Byrd, G. D.; Ogden, M. W. J. Pharm. Biomed. Anal. 2006, 40, 928. (33) Koulman, A.; Tapper, B. A.; Fraser, K.; Cao, M.; Lane, G. A.; Rasmussen, S. Rapid Commun. Mass Spectrom. 2007, 21, 421-428.

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

8913

Figure 2. PLS-DA scores plots for a gender response variable (9 male training set, [ female training set, b male external test set, and 2 female external test set) analyzed in positive mode ESI-MS: (a) reversed-phase training set data. (b) HILIC training set data. (c) Reversed phase with external test set overlaid. (d) HILIC with external test set overlaid.

use the total ion count for normalization rather than creatinine levels in what now appears to be becoming common practice. Statistical Analysis. The data were analyzed by both PCA and PLS, using SIMCA+ 11.5 (Umetrics). PCA is a data reduction technique that allows the data to be visualized in a few dimensions while retaining the features that contribute most to the variance. PCA does not require any prior knowledge of class and was used here to detect any inherent trends within the data and to identify any potential outliers that could affect subsequent discriminant analysis. Each of the four data cohorts, i.e., those obtained using the two separation methods in each of the ionization modes, was subsequently analyzed by PLS-DA. As a supervised technique, overfitting the data can be a problem with PLS, and therefore, 25-30% of the data were withheld to create an independent test set for each data cohort. The remaining data formed the training set with response variables assigned to sex, time of collection, or age. PLS models were developed using venetian blind internal cross-validation (CV) for evaluation, and the variable importance for the projection (VIP) values were used to identify and eliminate nondiscriminatory variables. When a suitable number of variables with satisfactory internal CV were found, the external test sets were imported and analyzed using the developed model. These test sets were never used during the model building process and as such are a true independent test of a PLS model’s predictive power. RESULTS AND DISCUSSION Separation using the RP and HILIC columns was optimized, to allow comparisons to be made under optimal conditions for 8914

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

each column. Figure 1 depicts typical UV254 chromatograms obtained on analysis of the same urine sample and illustrates the differing peak profiles of UV-absorbing compounds using RP and HILIC separation techniques. The inset extracted ion chromatograms (XICs) for the endogenous urinary metabolite creatinine (m/z 114.07 for [M + H]+) show retention times of 2.1 min on the RP system and 8.3 min on the HILIC column. Both XIC peaks have comparable baseline peak widths and are Gaussian-shaped peaks. The total ion chromatograms (TICs) illustrate the urinary compounds detected by MS using the two LC separation methods. Reversed-phase separation appears to yield more peaks in the TIC than HILIC; this was also reflected when the raw data were extracted, as more variables could be generated from RP-LC-MS data than for HILIC-LC-MS. However, as is described below, HILIC-LC-MS produced robust statistical models (using positive and negative ionization) despite fewer variables being extracted than from the RP data set. Similarly, although negative ionization mode data generated fewer variables than positive mode, robust and useful models were again developed. PCA was used to analyze all of the four training data sets (( RP and ( HILIC) in an unbiased manner to identify any points outside the 95% confidence limits, which also had a large distance from the model (DModX in SIMCA+). The resulting training set was then assigned response variables based upon gender, diurnal variation (a.m. and p.m. samples), and age, for analysis by PLSDA. Within SIMCA+, the explained variation, in terms of R2 and 2 Q values, gives an indication of the fit of the model and its predictive ability. Internal CV was used to determine an optimal

Figure 3. PLS-DA scores plots for the response variables diurnal variation (a, b) where ([ a.m. training set, 9 p.m. training set, 2 a.m. external test set, b p.m. external test set), and age (c, d) where ([ 21-30 age group training set, b 31-40 age group training set, 9 41-61 age group training set, ] 21-30 age group external test set, O 31-40 external age group test set, and 0 41-61 external age group test set) analyzed in positive mode ESI-MS with both the training and test set data shown. (a) Reversed phase colored according to diurnal variation. (b) HILIC colored according to diurnal variation. (c) Reversed phase colored according to age. (d) HILIC colored according to age.

balance between fit and predictive ability and to determine the number of components in the model. VIP scores were used to assess the discriminatory power of each variable and identify unimportant variables that did not add predictive power to the model. After these variables were removed, the model was rebuilt and the process repeated. When a satisfactory model was achieved, the external test set was imported to assess the predictive power and, therefore, the robustness, of the developed PLS-DA model. Positive Ionization. The majority of LC-MS-based metabonomic studies utilize only the positive ionization mode. We have used both positive and negative ionization, in order to assess whether this can increase the coverage of urinary components. The PLS-DA plots show clear separation in terms of gender for both separation methods using positive ionization ESI-MS (Figure 2). It is clear that the results from HILIC separation (Figure 2b) are comparable to, if not better than, to those from a more traditional RP separation (Figure 2a) approach. This is confirmed by the classification results on the independent test set, where classification rates of 94 and 100% were achieved for RP and HILIC, respectively (the lower classification for RP-LC-MS corresponds to one fewer external test set sample being incorrectly predicted). Panels c and d in Figure 2 show the PLSDA scores plots with the external test set data overlaid. Both models for PLS-DA gender discrimination only utilized one latent variable. No gain in classification rates was observed by using two latent variables.

For further comparison of the separation techniques, PLS-DA was used to discriminate between diurnal variation and age. Figure 3 shows scores plots with diurnal variation and age as the discriminating factors. Both RP (Figure 3a) and HILIC-MS data (Figure 3b) show good clustering for discrimination by diurnal variation, although, as might be expected, the plots show some overlap (Figure 3a and b). This is probably caused by the fact that some volunteers may not have donated the first void of the day or may have donated a sample earlier than the stipulated 15:00 h collection time for the afternoon sample. Despite the overlap, classification rates on the external test sets were 94% for RP and 87% for HILIC. HILIC’s lower classification rate was due to the fact that one of the samples in the external test set (a p.m. donation, highlighted by a red circle) lies well outside the 95% confidence limit (Figure 3b) and was subsequently classed as a false positive. The PLS-DA model for RP separation gave the highest classification when one latent variable was used, whereas the HILIC separation model gave higher classification results with two latent variables. For discrimination by age, the donors were split into three arbitrary classes: ages 21-30, 31-40, and 41-61 years. A general trend with increasing age could be seen along the first principal component for both RP and HILIC, although the groups merge at age boundaries, as expected. This trend along the first principal component explains why, for both RP and HILIC separation models, only one latent variable was required to give the highest classification rates. The samples corresponding to the 21-30 age group are more tightly clustered those of the 31-40 and 41-61 Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

8915

Figure 4. PLS-DA scores plots for negative mode ESI-MS with the response variables: gender (a, b) where (9 male training set, [ female training set, b male external test set, and 2 female external test set), diurnal variation (c, d) where ([ a.m. training set, 9 p.m. training set, 2 a.m. external test set, b p.m. external test set), and age (e, f) where ([ 21-30 age group training set, b 31-40 age group training set, 9 41-61 age group training set, ] 21-30 age group external test set, O 31-40 external age group test set, and 0 41-61 external age group test set). (a) Reversed-phase data, colored according to gender. (b) HILIC data, colored according to gender. (c) Reversed-phase data, colored according to diurnal variation. (d) HILIC data, colored according to diurnal variation. (e) Reversed-phase data, colored according to age. (f) HILIC data, colored according to age.

age groups, which are less well defined. Classification results on the independent test sets were the poorest in this study, with classification rates of 71 and 86% for RP (Figure 3c) and HILIC (Figure 3d), respectively. PLS-DA models for both RP and HILIC data were able to predict the younger age groups (21-30) reasonably well, but the models were unable to accurately predict the older groups (31-40 and 41-61). Samples from the external test sets are overlaid onto the RP (Figure 3c) and HILIC (Figure 3d) scores plots and illustrate the poor prediction of class for the older age groups. Negative Ionization. We also investigated negative mode ESIMS to determine whether further information could be obtained over that produced just using positive mode, in order to provide as comprehensive a MS fingerprint as possible. The negative mode 8916

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

ESI-MS data from both RP and HILIC separation techniques were analyzed by PCA, as with the positive mode data, and again PLSDA was carried out on the data with response variables assigned to gender, diurnal variation, and age. PLS-DA scores plots for separation based on gender for RP (Figure 4a) and HILIC (Figure 4b) show good clustering for the two groups. The results for both positive (Figure 2a) and negative ionization (Figure 4a) are comparable for RP separation. However, the defined clusters seen with the positive ionization HILIC-ESIMS data (Figure 2b) are lost when the polarity of ionization is changed (Figure 4b). For both RP and HILIC PLS-DA models based upon gender, the classification rate on the independent test set was 88%, with both models having the same classification rate when either one or two latent variables were used. This value is lower than those achieved for positive ionization data from both

Table 3. Comparison of External Test Set Classification Results for Reversed-Phase and HILIC Separation Techniques Used in Positive and Negative Mode ESI-MS Studiesa separation method ionization mode positive negative

Table 4. Comparison of the m/z and tR Values for the Top Five Variables for Each Separation Method and Both Modes of Ionization, As Determined by the Variable Importance for the Projection (VIP) Values for Discrimination by Gender separation method

Y variable

reversed phase

HILIC

gender diurnal age gender diurnal age

94 94 71 88 81 64

100 87 86 88 81 64

a The value indicates the percentage of correct classification results.

separation techniques, where classification rates of 94 and 100% were obtained for RP and HILIC, respectively. Separation based upon diurnal variation for RP (Figure 4c) and HILIC (Figure 4d) ESI-MS data in negative mode shows clustering comparable to that in positive mode (Figure 3a and b). However, classification of diurnal variation was worse with negative ionization data, giving 81% correct classification of the independent test sets for both separation methods. This can be compared to the classification rates of 94 and 87% achieved for RP and HILIC, respectively, using positive mode ESI-MS. Both PLS-DA models did not increase the classification rate when two latent variables were used over one, as the majority of the variation can be accounted for using the first principal component. The final PLS-DA model corresponds to age. As with positive ionization, both RP (Figure 3e) and HILIC (Figure 3f) negative mode ESI-MS data show overlap of the age groups, with the general trend of increasing age along the first principal component. With negative ionization data, the age group 31-40 appears to form a tighter cluster than was observed for positive ESI-MS data (Figure 2c and d), although the oldest age group (41-61) again shows poor clustering. Even with the tighter clustering of the middle age group (31-40), classification of the external test set for both separation methods using negative ESI-MS data was poor at 64%. For RP separation, the use of two latent variables increased the classification rates, but for HILIC separation, only one latent variable was required to gain the maximum classification rate. Table 3 summarizes the classification results on the independent test set. The two different chromatographic column chemistries allow comparable classification rates, showing that HILIC is a suitable separation technique to be employed for the analysis of human urine in metabonomic studies. It is clear that, for gender, diurnal variation, and age, the classification rates obtained using positive mode ESI-MS are higher than for negative mode, suggesting that a positive ionization approach is able to generate more robust models. However, when a metabonomic study is undertaken, a comprehensive picture of the components present in the sample should be sought and both positive and negative ionization considered. Table 4 presents the five variables that gave the highest VIP values generated by the PLS-DA models for discrimination by gender for the two modes of ionization and for the two separation methods. It is clear that the variables used to construct the models

RP

HILIC

ionization method

rank (VIP)

m/z

tR (min)

m/z

tR (min)

positive

1 2 3 4 5 1 2 3 4 5

171.10 90.05 815.25 217.13 197.06 101.03 145.01 260.99 495.19 265.10

6.92 2.70 11.37 6.92 8.90 3.48 3.47 10.28 15.45 11.27

206.04 181.00 105.02 192.11 428.01 184.09 186.98 175.02 263.08 107.04

2.55 6.85 2.73 7.70 2.63 2.50 2.92 5.98 2.87 2.97

negative

from the positive and negative ionization mode data are all totally different, as both the m/z and retention times (tR) are different. This shows that different compounds are contributing to the developed models, highlighting the fact that both modes of ionization are important in increasing the urinary metabolome coverage. Further to this, comparing the variables for RP-LC-MS and HILIC-LC-MS also reveals that different compounds are contributing to the developed PLS-DA models. This is entirely consistent with our initial expectation that different compounds would be retained on the different columns and again reinforces the need to utilize complementary separation methods when as much information as possible is sought from urinary samples. CONCLUSION In this study, we have shown HILIC to be a separation technique that is able to provide statistical results comparable to those of RP for the LC-ESI-MS metabonomic study of human urine samples. The use of independent test data shows that the developed PLS-DA models are as robust for HILIC as for RP chromatography. From the results presented above, we can recommend the following procedures for metabonomic studies that utilize LC-ESI-MS: (1) HILIC should be considered as a separation method complementary to RP as polar compounds that were unretained using RP may now be retained and thus contribute to the data set giving a more comprehensive fingerprint. (2) Negative ionization mode should also be utilized in order to provide the most comprehensive metabonomic fingerprint possible from urine samples. While ESI is the most commonly used ionization method for urinary LC-MS studies, it should be noted that it favors ionization of some classes of compounds over others (e.g., hydrophobic compounds are favored). The use of atmospheric pressure chemical ionization, which is suited for the analysis of polar and ionic small molecules, is thus likely to complement ESI for HILICLC-MS and thus may well offer further advantages if used alongside ESI. With respect to the amounts of data that are typically generated from metabonomic studies, e.g., NMR, LC-MS (positive and Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

8917

negative mode, RP and HILIC), we also recommend that the data are analyzed both separately and as a whole by concatenation or hierarchical analysis25,26,34,35 to obtain a “global” view of the data, along with the identification of the most important variables for each discriminative model developed; this is now the subject of investigation in our groups.

ACKNOWLEDGMENT This work is supported by an EPSRC (DTA) grant and Smith & Nephew. J.T.-O. gratefully acknowledges Thermo Finnigan’s support and funding from the Analytical Chemistry Trust Fund, the RSC Analytical Division, and EPSRC.

(34) Forshed, J.; Idborg, H.; Jacobsson, S. P. Chemom. Intell. Lab. Syst. 2007, 85, 102. (35) Forshed, J.; Stolt, R.; Idborg, H.; Jacobsson, S. P. Chemom. Intell. Lab. Syst. 2007, 85, 179.

Received for review May 17, 2007. Accepted September 17, 2007.

8918

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

AC071008V