1380
Chem. Res. Toxicol. 2002, 15, 1380-1386
Analytical Reproducibility in 1H NMR-Based Metabonomic Urinalysis Hector C. Keun,*,† Timothy M. D. Ebbels,† Henrik Antti,† Mary E. Bollard,† Olaf Beckonert,† Go¨tz Schlotterbeck,‡ Hans Senn,‡ Urs Niederhauser,‡ Elaine Holmes,† John C. Lindon,† and Jeremy K. Nicholson† Biological Chemistry, Biomedical Sciences, Faculty of Medicine, Imperial College of Science, Technology and Medicine, London, SW7 2AZ, U.K., and Pharma Preclinical Research Basel, F. Hoffmann-La Roche AG, CH-4070-Basel, Switzerland Received July 10, 2002
Metabonomic analysis of biofluids and tissues utilizing high-resolution NMR spectroscopy and chemometric techniques has proven valuable in characterizing the biochemical response to toxicity for many xenobiotics. To assess the analytical reproducibility of metabonomic protocols, sample preparation and NMR data acquisition were performed at two sites (one using a 500 MHz and the other using a 600 MHz system) using two identical (split) sets of urine samples from an 8-day acute study of hydrazine toxicity in the rat. Despite the difference in spectrometer operating frequency, both datasets were extremely similar when analyzed using principal components analysis (PCA) and gave near-identical descriptions of the metabolic responses to hydrazine treatment. The main consistent difference between the datasets was related to the efficiency of water resonance suppression in the spectra. In a 4-PC model of both datasets combined, describing all systematic dose- and time-related variation (88% of the total variation), differences between the two datasets accounted for only 3% of the total modeled variance compared to ca. 15% for normal physiological (pre-dose) variation. Furthermore, 95% correlation (r2) between sites, with an analytical error comparable to normal physiological variation in concentration (4-8%). The excellent analytical reproducibility and robustness of metabonomic techniques demonstrated here are highly competitive compared to the best proteomic analyses and are in significant contrast to genomic microarray platforms, both of which are complementary techniques for predictive and mechanistic toxicology. These results have implications for the quantitative interpretation of metabonomic data, and the establishment of quality control criteria for both regulatory agencies and for integrating data obtained at different sites.
Introduction In recent years, the science of toxicology has begun to explore the potential of novel “-omics” technologies, namely, genomics, proteomics, and metabonomics, which respectively can characterize in a highly parallel fashion the response of living systems to chemical exposure in terms of gene expression, protein expression, or metabolic regulation (1-3). These technologies offer rapid, mechanistic information, are often noninvasive or minimally invasive, and are to some degree quantitative. Thus, they facilitate incorporation of toxicological data at earlier stages of drug development, with potential savings of many millions of dollars. While these approaches utilize different analytical techniques and generate varying biochemical data, they provide complementary informa* To whom correspondence should be addressed. Tel: 44-(0)20-75943142. Fax: 44-(0)20-7594-3226. Email:
[email protected]. † Imperial College of Science, Technology and Medicine. ‡ F. Hoffmann-La Roche AG.
tion and face common challenges that must be addressed for their successful application to toxicity assessment (2). As the use of -omics technologies evolves from essentially qualitative measurements, it becomes ever more crucial to assess the reliability of data generated from these new technologies. Reproducibility (4) and robustness are clearly important for any ‘real world’ implementation, but will also influence answers to fundamental questions, such as whether or not signature profiles of chemicals and other stressors can be confidently defined. For any analytical technique, high reproducibility can increase quantitative accuracy and sensitivity and, by decreasing the number of replicates necessary for a given task, can also increase sample throughput. Ultimately, in systems biology approaches, this translates into the use of fewer experimental animals. The generation of databases for pooling of such data from different studies and their interpretation, particularly for regulatory agencies in the case of toxicological applications, will require
10.1021/tx0255774 CCC: $22.00 © 2002 American Chemical Society Published on Web 10/17/2002
Analytical Reproducibility of Metabonomic Studies
clear guidance and quality control on analytical reproducibility. Metabonomics represents a systems approach for measuring time-related biochemical responses in key intermediary biochemical pathways as a result of physiological, pathological, or interventional genetic events, and to-date this has been primarily achieved through the use of 1H NMR1 spectroscopy on biofluids such as urine or plasma (5-7). In addition, intact tissue samples can be analyzed directly using magic-angle-spinning 1H NMR spectroscopy (8-10). The application of 1H NMR spectroscopy to the study of biofluids and tissues has thus been demonstrated as an efficient method for studying the effects of drug toxicity, for clinical diagnosis, and for investigating gene function (11-14). High-field 1H NMR spectra of biofluids typically contain several thousand resolvable lines, potentially providing structural and quantitative information on hundreds of compounds from a single, nondestructive measurement in a few minutes. The manual analysis of even a small number of such spectra is a laborious and complex task, and metabonomic platforms utilize a wide range of statistical data reduction, multivariate analysis (15,16), and modeling techniques to facilitate automated NMR pattern recognition (NMR-PR) (17-22). Since the variables generated from NMR spectra are numerous and highly correlated, multivariate methods such as PCA are a prerequisite for a correct interpretation of the variation occurring in the data. PCA can be described as a multivariate bilinear decomposition method compressing a data matrix built up by N observations (samples), arranged in rows, and K variables, arranged in columns, down to a few principal components (PCs) describing the systematic variation in the data. A PC is a linear combination of the original variables describing the maximum systematic variation in the original data. Each PC consists of two vectors, a score vector (t), explaining the between-observation variation, and a loading vector (p), explaining the between-variable variation. Following the orthogonality constraint of the PCA method, each calculated PC is orthogonal, i.e., not correlated, to all the other PCs building up the model. The variation not explained by the calculated PCs is referred to as the residual and should ideally contain the random variation in the data. The number of PCs to be calculated can vary with the application and can be decided by a number of different methods. Most common is the crossvalidation method, but size of eigenvalue and visual interpretation of explained variation are also efficient tools. Previous reports have commented on the similarity of models generated by NMR-based metabonomic methods, not only when identical studies were repeated across several sites (22) but also despite differences in toxicological protocols (11). The impact of NMR parameters has also been assessed (23). However, to-date little quantitative information has been available. Here we present quantitative estimates for the influence of both sample preparation and NMR spectrometer type on variability in metabonomic data. Urine samples from a single study of hydrazine toxicity in the rat were divided and inde1 Abbreviations: NMR, nuclear magnetic resonance; PR, pattern recognition; PCA, principal components analysis; PLS-DA, partial least-squares discriminant analysis; TSP, sodium trimethylsilyl[2,2,3,32H ]propionate; GC, gas chromatography; MS, mass spectrometry. 4
Chem. Res. Toxicol., Vol. 15, No. 11, 2002 1381
pendently run at two different sites, each using a different spectrometer observation frequency (500 and 600 MHz). Our results suggest that metabonomic urinalysis is an extremely precise analytical tool. The number and detection of outliers is also demonstrated, together with an assessment of normal physiological variation that in our study appeared several times larger than the analytical variance. This evidently high analytical reproducibility of metabonomics appeared very favorable when compared to the interlaboratory transferability of other novel platforms for pathophysiological investigation.
Experimental Procedures Animal Studies. Male 8-10 week old Sprague-Dawley rats were housed in metabolism cages and urine samples collected daily from each animal. A standard diet, Purina chow 5002, was given to all animals, and free access to food and water was permitted throughout the study. A temperature of 21 ( 2 °C and a relative humidity of 55 ( 10% were maintained with fluorescent lighting between 06.00 and 18.00 ( 1 h. A dose of hydrazine dihydrochloride in saline (10 mL/kg) was administered p.o. at zero hours. Animals were randomly assigned to dose groups (vehicle only, 30 mg/kg, and 90 mg/kg) with 8 animals per group. Each dose group was split into two groups, one euthanased at 48 h post-dose (end of day 2) and the other euthanased at 168 h post-dose (end of day 7) for histopathological evaluation. Urine samples were collected over ice at between 24 h pre-dose and 16 h pre-dose, and then at 0, 8, 24, 48, 72, 96, 120, 144, and 168 h post-dose (a total of 180 samples). On collection, urine samples were centrifuged at ∼1200g for 10 min, to remove any solid debris, and were subsequently split into two aliquots, with one sent to each test site. Samples were stored at -40 °C pending 1H NMR spectroscopic analysis. This study was carried out in accordance with relevant national legislation and was subject to appropriate local review (Animal Welfare Act of 1978 and the Animal Protection Orders of 1981 and 1991, Swiss Federal Veterinary Office). NMR Spectroscopy. Three hundred microliters of 0.2 M sodium phosphate buffer (pH 7.4) containing 1 mM TSP (sodium trimethylsilyl[2,2,3,3-2H4]propionate) and 50 µL of D2O were added to 600 µL of each urine sample. All samples were centrifuged at 1200 rpm for 10 min to remove any solid debris. 1H NMR spectra were measured at 600 MHz (at Imperial College, IC) and at 500 MHz (at Roche, R) at 300 K using a flow-injection system (Bruker Biospin, Karlsruhe, Germany). Sample preparation was conducted independently at both sites. The water resonance was suppressed by using the first increment of a NOESY pulse sequence with irradiation during a 1 s relaxation delay and also during the 100 ms mixing time. For each sample, 64 transients were collected into 64 K data points using a spectral width of 20.036 ppm. The total acquisition time was ∼4 min per sample. Prior to Fourier transformation, an exponential line-broadening factor of 0.3 Hz was applied to each free induction decay. The IC spectra were phased automatically using an in-house routine written in MATLAB (The MathWorks, Natick, MA), before being baseline corrected and referenced to TSP (δ 0.0) using XWIN NMR (Bruker Biospin, Karlsruhe, Germany). The Roche spectra were manually phased and baseline corrected, but otherwise processed similarly. Data Reduction and Pattern Recognition. Each NMR spectrum was reduced to 245 variables, calculated by integrating regions of equal width (0.04 ppm) corresponding to the δ range 0.2-10.0 using AMIX (version 2.5.9, Bruker Analytik, Karlsruhe, Germany). This simplifies statistical analysis of the data, and reduces the impact of small variations in chemical shift due to, for example, variation in pH or ionic strength. Several spectral regions were excluded from the analysis. First the regions δ 4.50-5.98 were deleted to remove any spurious effects of variability in the suppression of the water resonance and any
1382
Chem. Res. Toxicol., Vol. 15, No. 11, 2002
Keun et al.
Figure 1. Mean score trajectories of PC analysis of urinary NMR spectral data for each dose group showing progression of metabolic effects of hydrazine treatment. Time of sampling is indicated. High dose data also show standard deviations across replicates (n ) 8 up to 48 h; n ) 4 post 48 h). NMR data from one site only are shown. cross-relaxation effects (mediated by chemical exchange of protons) on the urea signal. Further regions, corresponding to metabolites of hydrazine, 1,4,5,6-tetrahydro-6-oxo-3-pyridazine carboxylic acid (δ 2.74-2.94 and δ 2.46-2.66), and acetyl/ diacetylhydrazine (δ 2.14-1.90), were removed to minimize variation not due to changes in endogenous metabolite concentrations. While pharmacokinetic information and some endogenous profile information are removed by these exclusions, the resulting analysis is thus focused on reliably profiling metabolic change within the living system. The regions around the remaining citrate resonance (δ 2.66-2.74) were merged by summing into a single integrated region (labeled 2.7C) to reduce further the effects of pH-dependent chemical shifts. Finally, to take account of large variations in urine concentration, all spectra thus reduced were then normalized to a constant integrated intensity of 100 units. Multivariate analysis was performed using SIMCA-P software (version 8.0, Umetrics AB, Umeå, Sweden). These data were mean-centered, in which the mean of each variable was subtracted from each column of the dataset. This is equivalent to subtracting the mean spectrum from each other spectrum in the dataset and performing subsequent analyses on the meansubtracted spectra. Multivariate variance was calculated as the variance in the scores for each component, and the contribution to this modeled variance from differences between the datasets was assumed to be equivalent to the total variance of the score differences between the two spectra (an IC spectrum and a Roche spectrum) of the same sample. PLS can be described as the regression extension of PCA. Instead of describing the maximum variation in the measured data (X), which is the case for PCA, PLS attempts to derive latent variables, analogous to PCs, which maximize the covariation between the measured data (X) and the response variable (Y) regressed against. PLS discriminant analysis (PLSDA) applies the PLS algorithm to classification, using a Y matrix that represents an orthogonal unit vector for each class. PLSDA was applied to the combined IC and Roche datasets to examine if there were systematic differences that could reliably distinguish between them, and to characterize further these spectral differences. All cross-validation (PCA and PLS-DA) was iterated 7 times with every seventh sample removed from the analysis.
Results The progression of metabolic response in time, or metabolic trajectory, on exposure to hydrazine is illustrated in Figure 1, using NMR data obtained from one laboratory. The high dose group (90 mg/kg) exhibits a response of greater duration and magnitude than the low dose group (30 mg/kg). The low dose trajectory coincides with the control trajectory after 48 h, i.e., appears to
Figure 2. Score (t) scatter plots for PC1 vs PC2 (top) and PC3 vs PC4 (bottom) from a PCA model of both datasets combined. Triangles and circles represent IC and Roche datasets, respectively. The ellipse denotes the 95% significance limit. An asterisk indicates a putative outlier.
recover fully, while the high dose trajectory although approaching control space does not appear to recover completely. Histopathological examination of the animals confirmed a dose-related incidence and severity of principally midzonal fatty degeneration in the liver. All these observations were consistent with previous metabonomic studies of hydrazine toxicity in the rat (24). The results of PCA analysis on the data from both laboratories combined are shown in Figure 2, and model statistics for the first four components are listed in Table 1. No discernible time-dependence could be observed for scores from higher components. This approach immediately illustrates that the overall distribution of both datasets, resulting from the time-course of metabolic events after hydrazine treatment, is strikingly similar. The differences between the two datasets are clearly small by comparison, and do not affect the result or interpretation of the pattern recognition protocol utilized here. Table 1 shows that the variation between the duplicate datasets (3% after 4 PCs) is at least 5 times smaller than the variation within either the control group (26%) or the pre-dose (17%) samples. Figure 2 also indicates another useful feature of PCA, in that one sample (rat 8 at 48 h) is somewhat removed from the main cluster as visualized by the scores. This sample would normally be classified as a strong outlier, i.e., a
Analytical Reproducibility of Metabonomic Studies
Chem. Res. Toxicol., Vol. 15, No. 11, 2002 1383
Table 1. Summary Statistics for PCA of the Combined Datasetsa PCA
PC1
PC2
PC3
r2X (cum) q2 (cum) % of overall var t(Roche-IC) t(Roche-IC) (cum) control control (cum) pre-dose pre-dose (cum)
0.62 0.57
0.74 0.63
0.82 0.72
PC4 0.88 0.79
1.43 1.43 15.10 15.10 8.95 8.95
15.59 3.45 53.47 20.58 41.21 13.56
1.70 3.27 12.89 19.80 7.50 12.94
1.98 3.19 120.31 25.97 82.55 17.22
Table 2. Summary Statistics for PLS-DA of the Combined Datasetsa PLS-DA
PC1
PC2
PC3
PC4
PC5
r2X (cum) r2Y (cum) q2 (cum)
0.04 0.68 0.62
0.63 0.71 0.66
0.74 0.76 0.71
0.80 0.82 0.77
0.83 0.87 0.83
a Explained variation, r2; cross-validated explained variation, q2. Y is the classification matrix; X is the reduced variable matrix.
a Explained variation, r2; cross-validated explained variation, q2. Multivariate variance is calculated from the scores in each component separately. Cumulative (cum) variance is the sum of current and earlier components. t(Roche-IC), difference in scores between the datasets; control, control sample scores only; pre-dose, pre-dose sample scores only.
Figure 4. Overlay of the mean reduced spectrum from all data and the mean difference spectrum.
closely resemble the average differences between the two datasets (Figure 4). Overall, a 5-component model obtained the most robust fit to the Y classification matrix, with a cross-validated correlation of ∼83% (Table 2).
Figure 3. PLS-DA analysis of the combined datasets. Top: Positive and negative scores in the first X component (t[1]) can classify samples with 90% accuracy. Triangles and circles represent IC and Roche datasets, respectively. The ellipse denotes the 95% significance limit. Bottom: the loadings of each variable for this component closely resemble the average difference spectrum. Note that only one citrate resonance is labeled (*) due to overlap with hydrazine metabolites.
The average spectral difference between the two duplicate datasets (Figure 4) and the PLS-DA model loadings (Figure 3) together characterize the systematic variation introduced by metabonomic sample processing and analysis. Neighboring regions with an equal magnitude of change but with the opposite sign are indicative of small shifts in NMR resonance position, and appear to account for many of the differences seen, for example, those of creatine and hippurate. These predominantly occur as a result of small differences in sample pH due to imperfect buffering of the urine. The largest difference corresponds to the allantoin amide resonance (δ 6.04), which varies according to the amount of cross-saturation from suppression of the water resonance as a result of solvent chemical exchange. For the same level of presaturation power, spectra acquired at 600 MHz would be expected to exhibit lower water suppression than at 500 MHz; hence, the allantoin resonance is larger in the IC spectra on average. The presence of a slight increase in lactate may result from the effects of microbial anaerobic metabolism during sample processing.
sample greatly exceeding normal limits in score space and thus possessing undue influence on the model (16), and would be excluded from the analysis, but is included here to illustrate the appearance of a significant difference between the duplicate spectra. The exclusion of this one sample from multivariate analysis improved the precision and robustness of the models produced (data not shown). Even though the differences between the datasets are small, they are consistent enough that it is possible to separate the datasets using a PLS-DA model, as illustrated in Figure 3 and Table 2. The first component accounts for just 4% of the total variation in the spectral data, but can classify the spectra into the two datasets with significant success. The loadings for this component
Figure 5 shows that the score differences between spectra from the same sample (∆t ) tRoche - tIC) are small and very evenly distributed with at least two notable exceptions, one already identified as rat 8 at 48 h and the other as rat 14 at 72 h, both denoted by arrows. Both samples produced spectra in the IC dataset that would normally be classified as outliers, either by PCA analysis of all the data (Figure 2) or during PCA analysis of each dose group individually (Figure 6). It is interesting to note that the corresponding spectra in the Roche dataset, generated from the same urine sample, were not outlying. This indicates that such outliers may not reflect an unusual metabolic state, but originate from sporadic differences in the spectral quality. The most common cause of poor spectral quality is inadequate water signal
1384
Chem. Res. Toxicol., Vol. 15, No. 11, 2002
Keun et al.
Figure 5. Score differences between datasets from the combined dataset PCA model, showing variation in dataset differences across samples. The first two components only are shown. Two extreme outliers are identified.
Figure 7. Interlaboratory correlation of metabonomic data. Each set of reduced variables (normalized peak intensities) was mean-centered and scaled to unit variance. Citrate, δ 2.74-2.66; taurine, δ 3.46-3.42; hippurate, δ 7.86-7.82. Table 3. Interlaboratory Correlation of Metabonomic Dataa metabolite
r2
% error
% cv (pre-dose)
r2 (unnormalized)
citrate taurine hippurate
0.99 0.94 0.93
8.4 7.2 4.2
11.7 21.8 12.5
0.38 0.45 0.32
a Citrate, δ 2.74-2.66; taurine, δ 3.46-3.42; hippurate, δ 7.867.82. No outliers were excluded from this analysis. % error was estimated by the rms difference between IC and Roche/mean variable value. cv, coefficient of variation.
8%. Again the normal physiological variation, as estimated by the coefficient of variation of the pre-dose measurements, appears several times greater. Correlations from un-normalized data in Table 3 also demonstrate that the normalization procedure is important in achieving this remarkable degree of reproducibility.
Discussion
Figure 6. Score scatter plots illustrating outlier detection. Top: PC1 vs PC2 from a PCA model of control animal (IC data only). Bottom: PC3 and PC4 from a PCA model of low-dose group animals (IC data only). Ellipses signify 95% normal limits.
suppression for dilute samples, the effect of which is exaggerated by the normalization step of data processing. In this study, less than 3% of samples produced appreciable score differences (∆t, Figure 5). The high degree of interlaboratory correlation in terms of the normalized peak intensity, i.e., relative concentrations, of three major urinary constituents, hippurate, taurine, and citrate, is shown in Figure 7. The few outlying samples clearly account for the majority of nonequivalence between the two datasets. Univariate statistics in Table 3 show that for these three metabolites analytical precision is high, with a percentage error (an estimation of the coefficient of variance) of between 4 and
The results presented here quantitatively assess the precision of NMR-derived metabonomic data by duplicating NMR sample preparation and spectral acquisition at two sites. Examination of the data by established PR protocols illustrates that, with the exception of easily detected outlying spectra, both datasets were nearidentical in their information content. It was also demonstrated that any analytical variation was many times smaller than the detected normal physiological variation in urine composition. Further to these observations, the data were precise enough to characterize small consistent differences between the datasets acquired at each site, originating from pH differences and from the different spectrometer fields at which the analyses were conducted. The superior dispersion of a NMR spectrometer operating at 600 MHz compared to 500 MHz is not a critical factor due to the reduction of the data into integral regions. The importance of interlaboratory comparison when assessing the reliability of an analytical technique has
Analytical Reproducibility of Metabonomic Studies
recently been highlighted by reports of marked variability in gene expression data with sample handling and microarray platform (25). In a study investigating cisplatin toxicity, the overall correlation between two laboratories for >1.5-fold gene changes was poor (r2 ∼ 0.61), with one laboratory reporting around twice the number of gene changes and some genes were even reported to have opposite regulation (25). Furthermore, as part of the same study, a comparison of three micro-array platforms produced only ∼10% agreement in identified >1.5-fold gene changes. These interlaboratory inconsistencies could originate from more fundamental problems of reproducibility in gene-expression studies. The relationship between signal intensities and transcript concentration is proportional only within a limited range, and is significantly affected by the presence of normal cRNA background (26). Even within the experimentally determined ‘linear’ range, duplication of hybridization for the same RNA sample on two micro-arrays of the same platform resulted in ∼10% of genes exhibiting spurious >2-fold differences (26). Another study has reported that of 7129 genes over 99% showed significant (p < 0.05) variability in probe effectiveness using 20 probe-pairs per gene (27). The results of these irregularities can be considered as outliers, which anecdotal evidence suggests can account for ∼15% of genes in a typical micro-array study (28). By contrast, previous work on metabolic profiling has repeatedly demonstrated high analytical repeatability. For example, Fiehn et al. (29) have shown that GC-MS analysis of plant extracts produced an analytical error for several metabolites of ∼8% compared to a biological variance estimated at 26-56%. NMR-based methods have the advantage over MS and micro-array approaches in that the signal-to-noise ratio can be improved by merely extending the data acquisition time. In addition, MS signal intensity is often affected by differential ionization. In an NMR study of urinary profiles from control (n ) 20) and diabetic patients, Messana et al. (30) have described normal variation in relative citrate and hippurate concentrations as low as ∼4% and 6%, respectively, comparable to the level of precision of NMR-based metabonomic data described in the current study. Earlier work on the effect of NMR parameters such as spectrometer field on metabonomic analysis suggested that baseline differences, although small, could produce some systematic variation (23). The application of baseline correction in the present work appears to counteract efficiently this possibility, regardless of whether automated (IC) or manual (Roche) correction is used. Furthermore, it would appear that the automated phasing routine used by IC is at least as successful as manual correction. The NMR spectroscopy protocol used allowed a considerable amount of baseline to be sampled, making both phase and baseline correction a straightforward procedure. These findings provide compelling evidence that NMR spectroscopy combined with PR is a robust and precise approach to metabonomics. The reproducibility of the NMR data indicates the reliability of the data acquisition and processing protocols used in this work. It implies that, in principle, large-scale, multi-laboratory in vivo metabonomic studies will have a high degree of tolerance to unavoidable differences in environment and sample handling, without sacrificing the diagnostic power to identify the source and effects of any such confounding
Chem. Res. Toxicol., Vol. 15, No. 11, 2002 1385
variation. The analyses reported here were at least as reproducible as the most precise functional genomic methods at the protein level (31), and clearly more precise than differential analysis of expression using cDNA/ mRNA micro-arrays (28, 32).
Acknowledgment. The authors (H.C.K., T.M.D.E., M.E.B., H.A., O.B.) acknowledge the members of the Consortium for Metabonomic Toxicology (Eli Lilly, Bristol Meyers Squibb, Roche, Novo Nordisk, Pfizer, Pharmacia) for financial support.
References (1) Burchiel, S. W., Knall, C. M., Davis, J. W., Paules, R. S., Boggs, S. E., and Afshari, C. A. (2001) Analysis of genetic and epigenetic mechanisms of toxicity: potential roles of toxicogenomics and proteomics in toxicology. Toxicol. Sci. 59, 193-195. (2) Tennant, R. W. (2002) The National Center for Toxicogenomics: using new technologies to inform mechanistic toxicology. Environ. Health Perspect. 110, A8-10. (3) Aardema, M. J., and MacGregor, J. T. (2002) Toxicology and genetic toxicology in the new era of “toxicogenomics”: impact of “-omics” technologies. Mutat. Res. 499, 13-25. (4) Thompson, M. (2000) Towards a unified model of errors in analytical measurement. Analyst 125, 2020-2025. (5) Nicholson, J. K., Lindon, J. C., and Holmes, E. (1999) ‘Metabonomics’: understanding the metabolic responses of living systems to pathophysiological stimuli via multivariate statistical analysis of biological NMR spectroscopic data. Xenobiotica 29, 1181-1189. (6) Shockcor, J. P., and Holmes, E. (2002) Metabonomic applications in toxicity screening and disease diagnosis. Curr. Top. Med. Chem. 2, 35-51. (7) Nicholson, J. K., Connelly, J., Lindon, J. C., and Holmes, E. (2002) Metabonomics: a platform for understanding drug toxicity and gene function. Nat. Rev. Drug Discuss. 1, 153-161. (8) Garrod, S., Humpfer, E., Spraul, M., Connor, S. C., Polley, S., Connelly, J., Lindon, J. C., Nicholson, J. K., and Holmes, E. (1999) High-resolution magic angle spinning 1H NMR spectroscopic studies on intact rat renal cortex and medulla. Magn. Reson. Med. 41, 1108-1118. (9) Bollard, M. E., Garrod, S., Holmes, E., Lindon, J. C., Humpfer, E., Spraul, M., and Nicholson, J. K. (2000) High-resolution 1H and 1H-13C magic angle spinning NMR spectroscopy of rat liver. Magn. Reson. Med. 44, 201-207. (10) Griffin, J. L., Walker, L., Shore, R. F., and Nicholson, J. K. (2001) High-resolution magic angle spinning 1H NMR spectroscopy studies on the renal biochemistry in the bank vole (Clethrionomys glareolus) and the effects of arsenic (As3+) toxicity. Xenobiotica 31, 377-385. (11) Robertson, D. G., Reily, M. D., Sigler, R. E., Wells, D. F., Paterson, D. A., and Braden, T. K. (2000) Metabonomics: evaluation of nuclear magnetic resonance (NMR) and pattern recognition technology for rapid in vivo screening of liver and kidney toxicants. Toxicol. Sci. 57, 326-337. (12) Gavaghan, C. L., Holmes, E., Lenz, E., Wilson, I. D., and Nicholson, J. K. (2000) An NMR-based metabonomic approach to investigate the biochemical consequences of genetic strain differences: application to the C57BL10J and Alpk:ApfCD mouse. FEBS Lett. 484, 169-174. (13) Griffin, J. L., Williams, H. J., Sang, E., and Nicholson, J. K. (2001) Abnormal lipid profile of dystrophic cardiac tissue as demonstrated by one- and two-dimensional magic-angle spinning 1H NMR spectroscopy. Magn. Reson. Med. 46, 249-255. (14) Waters, N. J., Holmes, E., Williams, A., Waterfield, C. J., Farrant, R. D., and Nicholson, J. K. (2001) NMR and pattern recognition studies on the time-related metabolic effects of alpha-naphthylisothiocyanate on liver, urine, and plasma in the rat: an integrative metabonomic approach. Chem. Res. Toxicol. 14, 1401-1412. (15) Eriksson, L., Johansson, E., Kettaneh-Wold, H., and Wold, S. (1999) Introduction to Multi- and Megavariate Analysis using Projection Methods (PCA & PLS). UMETRICS Inc., Box 7960, SE90719 Umeå, Sweden, pp 267-296. (16) Jurs, P. C. (1986) Pattern recognition used to investigate multivariate data in analytical chemistry. Science 232, 1219-1224. (17) Holmes, E., Nicholson, J. K., Nicholls, A. W., Lindon, J. C., Connor, S. C., Polley, S., and Connelly, J. (1998) The identification of novel biomarkers of renal toxicity using automatic data reduction techniques and PCA of proton NMR spectra of urine. Chemom. Intell. Lab. Syst. 44, 245-255.
1386
Chem. Res. Toxicol., Vol. 15, No. 11, 2002
(18) Anthony, M. L., Sweatman, B. C., Beddell, C. R., Lindon, J. C., and Nicholson, J. K. (1994) Pattern recognition classification of the site of nephrotoxicity based on metabolic data derived from proton nuclear magnetic resonance spectra of urine. Mol. Pharmacol. 46, 199-211. (19) Holmes, E., Nicholls, A. W., Lindon, J. C., Ramos, S., Spraul, M., Neidig, P., Connor, S. C., Connelly, J., Damment, S. J., Haselden, J., and Nicholson, J. K. (1998) Development of a model for classification of toxin-induced lesions using 1H NMR spectroscopy of urine combined with pattern recognition. NMR Biomed. 11, 235-244. (20) Holmes, E., Nicholls, A. W., Lindon, J. C., Connor, S. C., Connelly, J. C., Haselden, J. N., Damment, S. J., Spraul, M., Neidig, P., and Nicholson, J. K. (2000) Chemometric models for toxicity classification based on NMR spectra of biofluids. Chem. Res. Toxicol. 13, 471-478. (21) Holmes, E., Nicholson, J. K., and Tranter, G. (2001) Metabonomic characterization of genetic variations in toxicological and metabolic responses using probabilistic neural networks. Chem. Res. Toxicol. 14, 182-191. (22) Antti, H., Bollard, M. E., Ebbels, T. M. D., Keun, H., Lindon, J. C., Nicholson, J. K., and Holmes, E. (2002) Batch statistical processing of 1H NMR-derived urinary spectral data. J. Chemometr. 16, 1-9. (23) Potts, B. C. M., Deese, A. J., Stevens, G. J., Reily, M. D., Robertson, D. G., and Theiss, J. (2001) NMR of biofluids and pattern recognition: assessing the impact of NMR parameters on the principal component analysis of urine from rat and mouse. J. Pharm. Biomed. Anal. 26, 463-476. (24) Nicholls, A. W., Holmes, E., Lindon, J. C., Shockcor, J. P., Farrant, R. D., Haselden, J. N., Damment, S. J., Waterfield, C. J., and Nicholson, J. K. (2002) Metabonomic investigations into hydrazine
Keun et al. toxicity in the rat. Chem. Res. Toxicol. 14, 975-87. (25) Wild, S. (2001) Recent data from the ILSI/HESI genomics nephrotoxicity subcommittee. Proceedings of Southern California Chapter of the Society of Toxicology Fall 2001 Meeting. http:// www.toxicology.org/memberservices/regionalchapter/scal_new/ files/p2001f.doc. (26) Chudin, E., Walker, R., Kosaka, A., Wu, S. X., Rabert, D., Chang, T. K., and Kreder, D. E. (2002) Assessment of the relationship between signal intensities and transcript concentration for Affymetrix GeneChip arrays. Genome Biol. 3, RESEARCH0005. (27) Chu, T. M., Weir, B., and Wolfinger, R. (2002) A systematic statistical linear modeling approach to oligonucleotide array experiments. Math. Biosci. 176, 35-51. (28) Nadon, R., and Shoemaker, J. (2002) Statistical issues with microarrays: processing and analysis. Trends Genet. 18, 265271. (29) Fiehn, O., Kopka, J., Dormann, P., Altmann, T., Trethewey, R. N., and Willmitzer, L. (2000) Metabolite profiling for plant functional genomics. Nat. Biotechnol. 18, 1157-1161. (30) Messana, I., Forni, F., Ferrari, F., Rossi, C., Giardina, B., and Zuppi, C. (1998) Proton nuclear magnetic resonance spectral profiles of urine in type II diabetic patients. Clin. Chem. 44, 15291534. (31) Gygi, S. P., Rist, B., Gerber, S. A., Turecek, F., Gelb, M. H., and Aebersold, R. (1999) Quantitative analysis of complex protein mixtures using isotope-coded affinity tags. Nat. Biotechnol. 17, 994-999. (32) Vingron, M., and Hoheisel, J. (1999) Computational aspects of expression data. J. Mol. Med. 77, 3-7.
TX0255774
