Chem. Res. Toxicol. 2000, 13, 471-478
471
Chemometric Models for Toxicity Classification Based on NMR Spectra of Biofluids Elaine Holmes,*,† Andrew W. Nicholls,† John C. Lindon,† Susan C. Connor,‡ John C. Connelly,‡ John N. Haselden,§ Stephen J. P. Damment,§ Manfred Spraul,| Peter Neidig,| and Jeremy K. Nicholson† Biological Chemistry, Biomedical Sciences Division, Imperial College of Science, Technology and Medicine, Sir Alexander Fleming Building, South Kensington, London SW7 2AZ, U.K., Departments of Safety Assessment and Analytical Science, SmithKline Beecham Pharmaceuticals, The Frythe, Welwyn, Herts AL6 9AR, U.K., Preclinical Safety Sciences-Toxicology, Glaxo Wellcome Research and Development, Park Road, Ware, Herts SG12 0DP, U.K., and Applications Department, Bruker Analytische Messtechnik GmbH, D-76287 Rheinstetten, Germany Received December 20, 1999 1H
NMR spectroscopic and pattern recognition (PR)-based methods were used to investigate the biochemical variability in urine obtained from control rats and from rats treated with a hydrazine (a model hepatotoxin) or HgCl2 (a model renal cortical toxin). The 600 MHz 1H NMR spectra of urine samples obtained from vehicle- or toxin-treated Han-Wistar (HW) and SpragueDawley (SD) rats were acquired, and principal components analysis (PCA) and soft independent modeling of class analogy (SIMCA) analysis were used to investigate the 1H NMR spectral data. Variation and strain differences in the biochemical composition of control urine samples were assessed. Control urine 1H NMR spectra obtained from the two rat strains appeared visually similar. However, chemometric analysis of the control urine spectra indicated that HW rat urine contained relatively higher concentrations of lactate, acetate, and taurine and lower concentrations of hippurate than SD rat urine. Having established the extent of biochemical variation in the two populations of control rats, PCA was used to evaluate the metabolic effects of hydrazine and HgCl2 toxicity. Urinary biomarkers of each class of toxicity were elucidated from the PC loadings and included organic acids, amino acids, and sugars in the case of mercury, while levels of taurine, β-alanine, creatine, and 2-aminoadipate were elevated after hydrazine treatment. SIMCA analysis of the data was used to build predictive models (from a training set of 416 samples) for the classification of toxicity type and strain of rat, and the models were tested using an independent set of urine samples (n ) 124). Using models constructed from the first three PCs, 98% of the test samples were correctly classified as originating from control, hydrazine-treated, or HgCl2-treated rats. Furthermore, this method was sensitive enough to predict the correct strain of the control samples for 79% of the data, based upon the class of best fit. Incorporation of these chemometric methods into automated NMR-based metabonomics analysis will enable on-line toxicological assessment of biofluids and will provide a tool for probing the mechanistic basis of organ toxicity.
Introduction There is a strong current demand for improved tools for screening novel therapeutic agents in toxicological studies. High-resolution 1H NMR spectroscopy of biofluids, cells, and tissues coupled with appropriate chemometric and pattern recognition (PR)1 methods offers a novel and robust approach for in vivo toxicological screening of drugs and provides a means of investigating their mechanistic toxicological effects (1-6). 1H NMR spectroscopic analysis allows simultaneous detection (and in principle quantitation) of hundreds of low-MW species within a biological matrix, resulting in the generation of †
Imperial College of Science, Technology and Medicine. SmithKline Beecham Pharmaceuticals. Glaxo Wellcome Research and Development. | Bruker Analytische Messtechnik GmbH. 1 Abbreviations: DMG, dimethylglycine; FT, Fourier transform; FID, free induction decay; HW, Han-Wistar; NAGs, N-acetylglycoprotein fragments; NMR, nuclear magnetic resonance; 2-OG, 2-oxoglutarate; PCA, principal components analysis; RD, relaxation delay; SIMCA, soft independent modeling of class analogy; SD, SpragueDawley; TMAO, trimethylamine N-oxide. ‡ §
an endogenous metabolic profile that is altered characteristically in response to changes in physiological status, toxic insult, or disease processes (1). The biochemical consequences and the mechanisms of toxicity of a range of compounds have been studied through noninvasive 1H NMR spectroscopic measurements on biofluids (7, 8). However, at high proton resonance frequencies (e.g., 600 MHz), the resulting one-dimensional (1D) 1H NMR spectra often contain several thousand signals, many of which are overlapping. The power and efficiency of diagnostic 1H NMR spectroscopy of biofluids can be increased greatly by the application of multivariate statistical analysis that both allows a reduction in the NMR data complexity and can be used to generate models for classification of samples. The value of applying data reduction and multivariate statistics to NMR spectral analysis has been well-documented (2, 9, 10), and we have termed this approach “metabonomics” (4). In particular, for biofluids, principal components analysis (PCA) and nonlinear mapping (NLM) have been used to aid in the interpretation of data from 1D 1H NMR spectra
10.1021/tx990210t CCC: $19.00 © 2000 American Chemical Society Published on Web 05/06/2000
472
Chem. Res. Toxicol., Vol. 13, No. 6, 2000
obtained from experimental animals (4, 11) and to identify biomarkers of toxicity (12, 13). In addition, similar metabonomic approaches have been used to evaluate the time-related urinary biochemical response of rats to model toxins (2, 3). However, unsupervised chemometric methods such as PCA and NLM have limited capabilities of classification, particularly where there are a large number of classes or subsets within a data set. To refine the definition of the class of toxicity and to optimize sample class discrimination, supervized methods such as soft independent modeling of class analogy (SIMCA) analysis have been used (11). Recent advances in NMR spectroscopic technology, such as the development of flow injection probes coupled with the use of robotic-sample handlers, have led to a significant improvement in the throughput of measurement of 1H biofluid spectra (14). The actual acquisition of NMR spectra has been increased via expanded automation to accommodate the acquisition of several hundred spectra in a 24 h period. However, processing and interpretation of the data have so far remained under direct manual control and are generally inefficient. To further improve throughput and data analysis, we have attempted to develop automated “expert” systems for sample classification in in vivo metabonomic toxicology studies. We describe here the utility of automated NMR systems for predicting the site and/or mechanism of toxicity of novel pharmacologically active compounds. Before the biochemical effects of chemical-induced toxicity using NMR-PR methods were assessed, the limitations of the approach were explored by assessing the extent of biochemical variation in a population of control animals. The study presented here has been conducted in three stages (i) to define normal physiological variation in control urine samples obtained from HW and SD rats, (ii) to determine strain-related differences in urine composition for commonly used rat strains, and (iii) to evaluate the derived prototype expert systems using selected model toxins.
Materials and Methods Animals and Sample Treatment. Control urines were collected from SD (n ) 58) and HW (n ) 58) rats as part of a series of ongoing toxicity studies. Animals were placed in metabolism cages for an 8 day period, and 24 h urine samples were obtained from each animal over eight consecutive 24 h time intervals. Animals were allowed free access to food (SQC Rat and Mouse Maintenance Diet 1, Special Diet Services Ltd.) and water throughout the study. Urine samples were also collected from animals treated with model toxins. HW rats (n ) 12) were administered a single dose of hydrazine (either 75 or 90 mg/kg in 0.9% saline po), and urine samples were collected over an 8 day period at 24 h intervals. However, only samples collected between 0 and 48 h pd were used for statistical analysis as this was the time period, determined by histopathology and clinical chemical assays, during which hydrazine exhibits a consistent hepatotoxic effect at these dose levels. SD rats (n ) 5) were treated with a single dose of HgCl2 (0.75 mg/kg in 0.9% saline ip), and urine samples were obtained for the same time period as the control samples. On the basis of the clinical chemical and histopathological evidence for toxic effect, urine samples obtained between 24 and 96 h following HgCl2 treatment were used in this study. Urine samples were collected during continuous 24 h time periods, with the exception of samples obtained between 24-32 h and 32-48 h pd. Following HgCl2 treatment, many differences in urine profile are known to occur between 24 and 48 h pd (3), and therefore, the number of samples collected from animals during this time period was increased
Holmes et al. to better define the time profile of HgCl2 toxicity. All urine samples were stored at -40 °C and analyzed using 1H NMR spectroscopy within 2 months of sample collection. Samples were made up from 400 µL of urine mixed with 200 µL of buffer solution [0.2 M Na2HPO4 and 0.2 M NaH2PO4 (pH 7.4)], and the resulting solution was left to stand for 10 min. The buffered urine was then centrifuged at 13 000 rpm for 10 min to remove precipitates. An aliquot of the resulting supernatant (500 µL) was then placed into a 5 mm NMR tube to which 50 µL of a sodium 3-(trimethylsilyl)[2,2,3,3-2H4]-1-propionate (TSP) solution in D2O was added. The TSP acted as a chemical shift reference (δ 0.0), and the D2O provided a lock signal for the NMR spectrometer. NMR Spectroscopy of Urine Samples. Conventional 1H NMR spectra of urine were measured at 600.13 MHz on a Bruker DRX-600 spectrometer using the following standard pulse sequence: RD-90°-t1-90°-tm-90°-acquire free induction decay (FID) (15). Here, RD represents a relaxation delay of 2 s during which the water resonance is selectively irradiated and t1 corresponds to a fixed interval of 3 µs. The water resonance is irradiated for a second time during the mixing time tm (150 ms). For each sample, 64 FIDs were collected into 64K data points using a spectral width of 7002.8 Hz, an acquisition time of 4.68 s, and a total pulse recycle delay of 7.68 s. The FIDs were multiplied by an exponential weighting function corresponding to a line broadening of 0.3 Hz prior to Fourier transformation. Data Reduction and Principal Components Analysis. Each NMR spectrum was corrected for phase and baseline distortions using XWINNMR (version 2.1, Bruker GmbH, Rheinstetten, Germany) and reduced to 256 integrated regions of equal width (0.04 ppm) corresponding to the δ region from -0.2 to 10.0 using AMIX (version 2.5, Bruker GmbH, Karlsruhe, Germany). These data were collected into a single data table, and multivariate analysis was performed using Pirouette software (version 2.03, Infometrix, Inc., Woodinville, WA). Results were corroborated using Unscrambler, an alternative chemometric software package (version 6.11, Camo ASA, Oslo, Norway). The δ region between 4.5 and 6.0 was removed prior to any statistical analyses to remove any spurious effects of variability in the suppression of the water resonance and any cross-relaxation effects (mediated by chemical exchange of protons) on the urea signal (16). Following removal of this region, each spectral data set was normalized to the total sum of the integrals in AMIX. Control data for the two individual rat strains were analyzed using PCA to determine those samples that were outliers. PCA was conducted both on mean-centered data and on autoscaled, mean-centered data. Mean centering data involves subtracting the calculated average of a variable from these data so that the mean for each variable is 0, whereas autoscaling refers to the division of each variable by the standard deviation for that variable. Although mean centering without scaling is the most common procedure when all data are the result of one spectroscopic technique, autoscaling can extract important changes in metabolites present in low levels in biofluids as an equal weight is conferred on all variables. Spectra from samples which mapped separately from the main cluster in the scores plots of PC1 versus PC2, or which showed a large distance to model in the outlier diagnostics plot, were examined for abnormalities and, where spectra were found to be genuinely atypical to the normal population, removed. The outlier diagnostics plot allows the identification of samples that are abnormal and are not representative of the sample population and is based on two parameters. The first is called the “sample residual” and is the difference in the value of the multivariate data in the original data set and the corresponding value in the reduced PC data set, i.e., that proportion of the data not explained by the model (17). The second discriminator, the Mahalanobis distance, is the distance between a given sample point in the reduced PC data representation and the multivariate centroid of the data points (18). The outlier diagnostics plot uses a confidence limit of 95% to define normal
Chemometric Analysis of Biofluid NMR Data
Figure 1. Typical 600 MHz 1H NMR spectra (δ 0.5-9.0) of whole rat urine from (A) Han-Wistar and (B) Sprague-Dawley animals. Abbreviations: DMG, dimethylglycine; HOD, residual water; m-HPPA, m-(hydroxyphenyl)propionic acid; NAGs, Nacetylglycoprotein fragments; 2-OG, 2-oxoglutarate; TMAO, trimethylamine N-oxide. distribution; therefore, it is expected that 5% of a population of samples will fall outside this confidence limit and that such samples will not be truly “abnormal”. A further test set of samples containing HgCl2-treated rat urines (n ) 20) and hydrazine-treated rat urines (n ) 20) was also analyzed by PCA. SIMCA Analysis. Five “training” sets of control urine data containing 100, 200, 400, 600, or 800 urine spectra were tabulated with each set comprising equal numbers from each strain. These five training sets of different size were used to assess the number of control urine samples required to adequately define the boundaries for a population of control urine samples. SIMCA models of normal rat urine were constructed for each training set (i) to define biochemical boundaries for the control urine sample population regardless of strain and (ii) to characterize control urine composition for the SD and HW urine populations individually. The SIMCA models were constructed for each set of data using between 1 and 20 principal components (PCs). To validate the models, a further data set composed of 50 HW and 50 SD urine samples was created as a test set to evaluate the predictive accuracy from the SIMCA models. Urine samples used in the test set were independent of those used in the training set; i.e., separate sets of animals were used for training and testing the model. The models defining the populations of control rat urine were then tested further using an additional data set that contained data from 1H NMR spectra of urine from SD rats that had received a single dose of HgCl2 or HW rats that had received a single dose of hydrazine. Separate SIMCA models were built for urine samples from rats treated with HgCl2 or hydrazine using only eight samples per group. These models were tested using a further data set of urine samples (n ) 12 per group), which were obtained from a separate group of rats which had been treated with either hydrazine or HgCl2.
Results 1H
NMR Spectroscopy of Control Urine from SD and HW Rats. The 600 1H NMR spectra of the whole urine from both strains were compared, and few systematic differences could be detected. Figure 1 shows the 600 MHz 1H NMR spectra of whole rat urine from HanWistar (A) and Sprague-Dawley (B) rats. Many of the endogenous urinary components were assigned on the basis of previous literature data (1, 7, 15). Visual comparison of the control urine spectra obtained from both strains of rat indicated that, overall, the biochemical
Chem. Res. Toxicol., Vol. 13, No. 6, 2000 473
Figure 2. Outlier diagnostic plot (Mahalanobis distance vs sample residual). Samples, which have a large distance to the PCA model, and are not normally distributed are mapped in the northeast corner of the plot.
profiles were broadly similar. Nonetheless, a small number of strain-related variations in the biochemical profile of the 1H NMR urine spectra were noted. For example, HW rat urine contained higher concentrations of acetate, lactate, and taurine in relation to the total urine profile, with lower concentrations of hippurate in comparison with the SD rat (Figure 1). However, the urinary concentrations of citrate, taurine, hippurate, and trimethylamine N-oxide were found to be variable in urine samples obtained from the same strain of rat, making simple visual comparisons of the spectral profiles difficult. PCA of Control 1H NMR Data. To assess the distribution of samples obtained from the two strains of rat, and to obtain a more detailed analysis of strainrelated variations in urinary composition, PCA was performed on combined data from both strains. Spectra from 450 control SD urine samples and 450 control HW urine samples were analyzed. An outlier diagnostics plot, as described in Materials and Methods, was used to identify urine samples that did not fall within the multivariate boundary of normality for the population of control animals (Figure 2). Those samples that were classified as outliers mapped in the NE region of the plot. The 1H NMR spectra of these samples were investigated and found to be atypical. For example, several of the urine samples identified as abnormal contained multiple resonances from chlorogenic acid metabolites, which are thought to arise due to modulations in the populations of gut microflora (19). An example of the 1H NMR spectrum for such a sample which contained depleted levels of hippurate and elevated levels of chlorogenic acid metabolites, such as m-(hydroxyphenyl)propionic acid, is shown in Figure 3. All samples classified as “aberrant” were removed prior to further analysis. The scores plot of PC1 versus PC2 for 1H NMR spectra of the control urine samples (Figure 4) shows that, although the two populations of samples heavily overlapped, some strain-related subclustering existed. PCA afforded a visible degree of separation between the strains based on ratios of the concentrations of metabolite resonances which were characteristic for each strain.
474
Chem. Res. Toxicol., Vol. 13, No. 6, 2000
Holmes et al.
Figure 3. 1H NMR spectra (aromatic proton region, δ 6.28.9) selected from the outlier diagnostic plot (Figure 2) of (A) a typical urine spectrum and (B) an atypical urine spectrum, obtained from control SD rats.
Figure 5. The 600 MHz 1H NMR spectra (δ 0.5-4.5) of typical whole rat urine from (A) control HW rat, (B) hydrazine-treated HW rat, (C) control SD rat, and (D) HgCl2-treated SD rat.
Figure 4. Principal components scores plot of PC1 vs PC2 from mean-centered data-reduced 1H NMR spectra of 900 control urine samples from HW (b) and SD (0) rats. Table 1. PCA-Detected 1H NMR Spectral Regions That Differ between SD and HW Rat Urine
region (δ)
major metabolite
strain in which metabolite is present in a relatively higher concentration
1.32-1.36 1.92-1.96 2.40-2.44 2.44-2.48 2.56-2.58 2.68-2.72 2.92-2.96 3.00-3.04 3.04-3.08 3.24-3.28 3.40-3.44 3.56-3.80 3.96-4.0 7.52-7.64 7.80-7.84
lactate acetate succinate 2-OG citrate citrate DMG 2-OG creatinine taurine taurine glucose and sugars hippurate hippurate hippurate
HW HW HW SD HW HW HW SD SD HW HW SD SD SD SD
Those regions which were noted to give rise to strain separation are shown in Table 1. Metabolite excretion levels that appeared to differentiate strains included acetate, lactate, succinate, citrate, dimethylglycine, and taurine, all of which were larger as a ratio with the total spectral intensity in HW rats. Conversely, hippurate, creatinine, and 2-oxoglutarate were found to be present in relatively greater ratios in the urine of SD rats.
However, despite the evidence of partial clustering according to strain, the large number of samples that overlapped in the region between the two groups indicated the existence of a high degree of commonality. No evidence of clustering related to time of sampling over the 8 day collection period was noted in either strain. 1H NMR Spectral Analysis of Urine Samples Obtained from Rats Treated with Model Toxins. To assess the effect of normal physiological variance and strain differences in toxicology studies, two sets of urine samples obtained from rats treated with organ specific toxins were included in the data set. Hydrazine is a model hepatotoxin used to induce steatosis in experimental animals. Spectra obtained from rats treated with hydrazine (Figure 5B) showed evidence of elevated urinary creatine, taurine, β-alanine, and 2-aminoadipate concentrations, together with lower urinary levels of citrate, succinate, 2-oxoglutarate, and creatinine when compared with spectra from control animals (e.g., Figure 5A). Increased urinary levels of taurine and creatine are known to be indicative of hepatotoxicity (20, 21). In contrast, the spectra obtained from rats treated with the renal cortical nephrotoxin, HgCl2 (Figure 5D), showed elevations in urinary levels of amino acids, organic acids, and glucose together with depletion of citrate, succinate, hippurate, and 2-oxoglutarate levels as compared to the levels in the appropriate control (Figure 5C). This pattern of metabolic perturbation is consistent with renal cortical damage (3, 7). The differences in the metabolic profiles of urine spectra obtained from rats treated with toxins in comparison with those obtained from control spectra were much more marked than the differences in biochemical composition arising from strain, and many of the perturbations could be identified by visual inspection of the spectra. PCA of 1H NMR Data Obtained from Rats Treated with Model Toxins. PCA was performed on the com-
Chemometric Analysis of Biofluid NMR Data
Chem. Res. Toxicol., Vol. 13, No. 6, 2000 475
Figure 6. PCA plot derived from 1H NMR urine spectra of (A) control HW urine samples (b) mapped with samples from hydrazinetreated HW rats (4) and (B) control SD urine samples (0) mapped with urine samples from HgCl2-treated SD rats (×). Table 2. PCA-Detected 1H NMR Chemical Shift Regions Separating Urine Spectra from Spectra of Control and Hydrazine-Treated HW Ratsa region (δ)
major metabolite
difference between urine samples from control and toxin-treated rats
1.32-1.36 1.60-1.68 1.92-1.96 2.04-2.08 2.20-2.28 2.40-2.44 2.44-2.48 2.56-2.58 2.68-2.72 3.00-3.04 3.04-3.08 3.16-3.20 3.24-3.28 3.40-3.44 3.72-3.80 3.92-3.96 3.96-4.0 7.52-7.64 7.80-7.84
lactate 2-amino adipate 2-amino adipate acetoacetate 2-amino adipate succinate 2-OG citrate citrate 2-OG creatine unidentified taurine taurine 2-amino adipate creatine hippurate hippurate hippurate
+ + + + + + + + + + + -
a + indicates an increase and - a decrease in the relative metabolite concentration in hydrazine-treated rat urine.
plete data set, and clear separation between control and toxin-treated urine samples was achieved (Figure 6A,B). Examination of the loadings used to construct the eigenvectors enabled a table of biochemical markers to be formulated for hydrazine- and HgCl2-induced toxicity (Tables 2 and 3). SIMCA Analysis of 1H NMR Data. Since some degree of inherrent clustering relating to strain of rat was observed in the control data using PCA, these data were analyzed further using SIMCA analysis. Models were calculated for both SD and HW rat strains using the training sets of 100, 200, 400, 600, and 800 urine samples. These models were then evaluated using a test set comprised of 50 SD urine samples and 50 HW urine samples. SIMCA models were constructed from the first 20 PCs with a view of establishing the class membership for each
Table 3. PCA-Detected 1H NMR Chemical Shift Regions Separating Urine Spectra from Spectra of Control and HgCl2-Treated SD Ratsa
region (δ)
major metabolite
difference between urine samples from control and toxin-treated rats
0.96-1.04 1.20-1.24 1.32-1.36 1.44-1.52 1.72-1.76 2.40-2.44 2.44-2.48 2.56-2.58 2.68-2.72 3.00-3.04 3.04-3.08 3.56-3.80 3.96-4.0 4.04-4.08 6.88-6.92 7.18-7.22 7.52-7.64 7.80-7.84
valine and isoleucine 3-hydroxybutyrate lactate alanine lysine succinate 2-OG citrate citrate 2-OG creatinine glucose hippurate creatinine tyrosine tyrosine hippurate hippurate
+ + + + + + + + -
a + indicates an increase and - a decrease in the relative metabolite concentration in HgCl2-treated rat urine.
Table 4. Cumulative Variance for Each Training Set (representing progressively larger population sizes) cumulative variance (no. of samples in training sets) no. of PCs
100
200
400
600
800
1 2 3 4 5 6 10 20
54.56 69.43 76.07 81.60 86.35 89.72 94.89 98.63
54.72 68.54 75.31 80.97 85.10 87.83 88.84 98.87
49.71 59.37 67.59 72.81 78.66 82.09 90.56 97.24
50.29 59.87 67.76 73.64 78.32 81.31 90.18 97.12
48.69 59.22 67.15 72.82 77.48 81.35 89.86 96.91
of the samples in the test set. The cumulative variance for each of the PCs used (i.e., the proportion of the original data explained by successive PCs) is described in Table 4. Figure 7A shows the Cooman’s plot of membership of CS1 (HW) and CS2 (SD) calculated from the first three PCs for the 800-sample model. The plot is
476
Chem. Res. Toxicol., Vol. 13, No. 6, 2000
Holmes et al.
Figure 7. (A) Cooman’s plot of 100 control urine spectra describing the class boundaries for the control HW urine sample population (CS1, b) and the control SD sample population (CS2, 4). (B) Cooman’s plot of 100 control urine spectra together with urine spectra obtained from hydrazine-treated rats (n ) 20, 0) and HgCl2-treated rats (n ) 20, ×). Control data are shown clustering in the southwestern region of the plot, while NMR spectra of urine from animals following treatment of toxins occur in the northeastern region, indicating that they do not fit either population of control urine samples. Table 5. Percentage of Correct Prediction of Sample Class/Strain from the SIMCA Models % of correct prediction of control samples in test set (regardless of strain)
% of correct prediction of strain (based on class of best fit)
no. of samples in training seta
no. of samples in training seta
no. of PCs
100
200
400
600
800
100
200
400
600
800
1 2 3 4 5 6 10 20
97 97 94 96 94 94 87 85
100 98 98 100 100 100 96 90
100 100 100 100 100 100 99 96
100 100 98 100 100 100 97 97
100 98 97 100 100 100 98 98
72 73 73 84 71 77 76 73
70 69 66 72 74 72 88 82
71 76 73 72 77 73 88 94
68 75 80 75 78 75 89 99
70 77 79 75 80 77 81 99
a
Samples in training sets are comprised of 50% SD urine samples and 50% HW urine samples.
divided into four regions by the intersection of 95% confidence limit lines for each class of urine samples. These regions identify the class membership with the NW region defining HW only, the SE region defining SD only, the SW region containing both HW and SD, and the NE region containing neither control HW nor SD. The distribution of samples within this plot indicated that using SIMCA models based on 3 PCs (explaining 67.15% of the total variance in the data set), 3 of the 100 test samples were classified as not lying within the boundaries of the control model. Moreover, the incorrectly classified test samples were re-examined and found to differ slightly from other control urine samples. For example, glucose and lactate levels appeared to be elevated in the urine spectra from two of the misclassified samples. Furthermore, although a high degree of commonality between the two strains existed, i.e., the SIMCA analysis assigned most control samples to both strain-related classes, this model correctly classified 79% of the test samples according to the class or strain of best fit (Table 5). By increasing the number of PCs in the SIMCA model to 20, we could classify 99% of the test data according to
strain. However, cross validation of the model suggested that a lower number of PCs was more appropriate as there was less danger of overfitting the data. Although the variance explained by the PCA model (r2) will increase with subsequent PCs, the predictive power of the model (q2) does not automatically increase with incremental PCs, and thereafter, the inclusion of subsequent PCs into the model will only add noise (22). The SIMCA models were recalculated using smaller training sets of 100, 200, 400, and 600 urine samples (1:1 HW:SD) with a view of establishing the optimum number of samples required to describe the sample population. The models were evaluated with the same test set, and the results are summarized in Table 5. Little difference was observed in the predictive accuracy of the models calculated from the five training sets, with the exception of the smallest sample set, indicating that a class of 200 urine samples was sufficient to characterize a control population of animals. In each case, the model was able to give a better prediction for urine samples obtained from SD rats. The test set was extended by the addition of 40 urine samples from animals that had received a dose of either
Chemometric Analysis of Biofluid NMR Data Table 6. Prediction of Toxin-Treated Test Samples no. of PCs in the SIMCA model
no. of misclassified samples obtained from HgCl2-treated rats
no. of misclassified samples obtained from hydrazine-treated rats
1 2 3 4
1 of 12 1 of 12 1 of 12 1 of 12
0 of 12 4 of 12 4 of 12 6 of 12
hydrazine or HgCl2 (Figure 5). Figure 7B shows the Cooman’s plot of membership of HW (CS1) and SD (CS2) classes using the 800 sample model and the extended test set. Using a two-class SIMCA analysis trained on control data from the two strains of rat, 99.2% of the samples from the extended test set were correctly classified; i.e., most of the control samples fell within the boundary of one or both of the control classes, while samples obtained from toxin-treated animals did not fit either model. Separate SIMCA models were built for both sets of toxin-treated rat (n ) 8 per class). The remaining toxintreated samples (n ) 12 per class) were combined with the control test data and used to assess the four class SIMCA model where class 1 is the HW control, class 2 is the SD control, class 3 is made up of HW rats treated with hydrazine, and class 4 is made up of SD rats treated with HgCl2. All of the samples obtained following hydrazine treatment were correctly classified, and only one of the samples in the HgCl2 test set was misclassified as a control SD sample (Table 6). This sample was collected between 72 and 96 h post-treatment and was found to have a spectral profile similar to those of control urine samples, suggesting that the urine had been collected from an animal that was exhibiting regeneration at an early stage. Comparison of the Cooman’s plots in panels A and B of Figure 7 indicated that although the method was powerful enough to distinguish between the two classes of control data, the variation associated with samples that were obtained following toxic insult by hydrazine or HgCl2 was so substantial as to minimize the strain-related differences in urine composition.
Discussion This study has used 1H NMR spectroscopy and PR methods to construct mathematical models describing the “normal” physiological variation of metabolic profiles in urine for two rat strains. We have shown that by using outlier diagnostic plots (Figure 2), aberrant control samples can be detected. Many factors can be responsible for generating biochemical alterations in the baseline urinary profile, including change in diet, poor health, genetic drift, and bacterial contamination of urine, and some of these factors may have implications on the toxicologic response of an animal. In many routine toxicology studies, the number of experimental animals used is typically in the range of 5-10 animals per class (2, 23). The distorting effect of including an abnormal animal in studies using small numbers of animals can be significant, and makes statistical interpretation of such studies difficult. The NMR-based metabonomic methods developed in this study would provide a means of prescreening animals prior to their inclusion in toxicological studies. In this way, a more homogeneous group of animals can be selected, allowing the possibility of excluding abnormal animals from toxicology studies. In addition, NMR-PR-based analysis of control urine can
Chem. Res. Toxicol., Vol. 13, No. 6, 2000 477
allow the monitoring of biochemical changes due to genetic drift or dietary modification over time. SIMCA models based on the 1H NMR spectral profiles of urine samples from as few as seven rats per strain allowed the characterization and prediction of strain in an independent set of test samples. However, since successful prediction is reliant upon training sets containing well-represented sample populations (describing the full-range naturally occurring variation within a class), the models built from the larger training sets will be more robust. The sensitivity of these chemometric methods in differentiating between the 1H NMR urine profiles obtained from SD and HW rats highlights the future potential of this technique as a tool for investigating genetic manipulation in experimental animals. Although strain-related differences in the biochemical composition of urine could be detected using metabonomic methodology, perturbations in the urinary profile caused by administration of the two toxins studied (Figure 7A,B) were significantly larger than any strainrelated urinary variation. The model toxins that were chosen (hydrazine and HgCl2) were used here to exemplify the difference in the magnitude of biochemical perturbation induced by toxicological episodes in comparison with factors such as physiological variation and strain differences. Previously, we have shown that 1H NMR spectroscopy and PCA-based pattern recognition allow for the classification of urine samples following a toxic insult (11, 13). However, although PCA is a highly informative method for the determination of toxicity and toxicity types, it does not provide a suitable system for the prediction of a class of toxicity for novel drugs. SIMCA analysis enables the construction of separate PC models describing each strain of control animal and each class of toxicity. Furthermore, it enhances automation capability and provides an improved statistical basis by allowing the inclusion of confidence limits within models. Thus, the SIMCA approach allows the construction of automated expert systems capable of analyzing novel toxicological data and determining whether a particular urine sample is normal or abnormal. The current study reports the extent of variation in the metabolic composition of control urine samples obtained from standard laboratory strains of rat. Moreover, this system offers the possibility of characterizing and classifying different types of toxic lesion on the basis of the 1H NMR-detected perturbations in biofluid composition. Having established a database of 1H NMR spectral urine profiles for control rats, we are carrying out ongoing studies by applying similar chemometric methods to establish strain differences in response to various toxicological insults. In the present study, the urine samples obtained from toxin-treated animals were used to exemplify the potential of this methodology in toxicological screening and classification; therefore, the time period over which samples were collected was relatively small. However, since the metabolic response of organisms to toxicological challenge is dynamic, care must be taken to incorporate a representative time course that accommodates the development of a lesion in toxicological studies. The chemometric models derived in this study enabled the classification of control urine from two different strains, allowed the identification of aberrant control samples, and aided the classification of samples following a toxic insult. Together with the recent developments in
478
Chem. Res. Toxicol., Vol. 13, No. 6, 2000
NMR spectrometer technology, automated chemometric analysis of NMR biofluid data should enable the construction of efficient systems for toxicological screening in vivo.
Acknowledgment. We gratefully thank GlaxoWellcome (A.W.N.) and SmithKline Beecham Pharmaceuticals (E.H.) for funding. We are also grateful to Dr. Scott Ramos (Infometrix, Inc.) for statistical advice.
References (1) Nicholson, J. K., and Wilson, I. D. (1989) High-Resolution NMR spectroscopy of biofluids. Prog. NMR Spectrosc. 21, 444-501. (2) Beckwith-Hall, B., Nicholson, J. K., Nicholls, A. W., Foxall, P. J. D., Lindon, J. C., Connor, S. C., Abdi, M., Connelly, J., and Holmes, E. (1998) Nuclear magnetic resonance spectroscopic and pattern recognition investigations into the biochemical effects of hepatotoxins. Chem. Res. Toxicol. 11, 260-272. (3) Holmes, E., Bonner, F. W., Sweatman, B. C., Lindon, J. C., Beddell, C. R., Rahr, E., and Nicholson, J. K. (1992) Nuclear magnetic resonance spectroscopy and pattern recognition analysis of the biochemical processes associated with the progression and recovery from nephrotoxic lesions in the rat induced by mercury II chloride and 2-bromoethanamine. Mol. Pharmacol. 42, 922930. (4) 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. (5) Kruse, S., Kvalheim, O. M., Gadeholt, G., Halsteinslid, L., and Sletten, E. (1991) Multivariate analysis of proton NMR spectra of serum from rabits. Monitoring progressive growth of implanted VX-2 carcinoma. Chemom. Intell. Lab. Syst. 11, 191-196. (6) Maxwell, R. J., Martinez-Perez, I., Cerdan, S., Cabanas, M. E., Arus, C., Moreno, A., Capdevila, A., Ferrer, E., Bartomeus, F., Aparicio, A., Conesa, G., Roda, J. M., Carcellar, F., Pascual, J. M., Howells, S. L., Mazucco, R., and Griffiths, J. R. (1998) Pattern recognition analysis of 1H NMR spectra from perchloric acid extracts of human brain tumour biopsies. Magn. Reson. Med. 39, 869-877. (7) Gartland, K. P. R., Bonner, F. W., and Nicholson, J. K. (1989) Investigations into the biochemical effects of region-specific nephrotoxins. Mol. Pharmacol. 35, 242-250. (8) Wilson, I. D., Nicholson, J. K., Ghauri, F. Y. K., and Blackledge, C. A. (1991) Use of high-field nuclear magnetic resonance spectroscopy for the analysis of biological fluids. Anal. Proc. 28, 217-223. (9) El-Deredy, W. (1997) Pattern recognition approaches in biomedical and clinical magnetic resonance spectroscopy: a review. NMR Biomed. 10, 99-124. (10) Somorjai, R. L., Nikulin, A. E., Pizzi, N., Jackson, D., Scarth, G., Dolenko, B., Gordon, H., Russell, P., Lean, C. L., Delbridge, L., Mountford, C. E., and Smith, I. C. P. (1995) Computerized
Holmes et al.
(11)
(12)
(13)
(14)
(15)
(16)
(17) (18) (19)
(20)
(21)
(22)
(23)
consensus diagnosis: A classification strategy for the robust analysis of MR spectra. I. Application to 1H spectra of thyroid neoplasms. Magn. Reson. Med. 33, 257-263. Anthony, M. L., Sweatman, B. S., Beddell, C. R., Lindon, J. C., and Nicholson, J. K. (1994) Pattern recognition of the site of nephrotoxicity based on metabolic data derived from nuclear magnetic resonance spectra of urine. Mol. Pharmacol. 46, 199211. Holmes, E., Connor, S. C., Nicholls, A. W., Polley, S., Nicholson, J. K., Lindon, J. C., 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, 251-261. Holmes, E., Nicholls, A. W., Lindon, J. C., Ramos, S., Spraul, M., Neidig, P., Connor, S. C., Connelly, 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. Spraul, M., Hofman, M., Ackermann, M., Nicholls, A. W., Damment, S. J. P., Haselden, J. N., Shockor, J. P., Nicholson, J. K., and Lindon, J. C. (1997) Flow injection proton nuclear magnetic resonance spectroscopy combined with pattern recognition methods: Implications for rapid structural studies and high throughput biochemical screening. Anal. Commun. 34, 339-341. Nicholson, J. K., Foxall, P. J. D., Spraul, M., Farrant, R. D., and Lindon, J. C. (1995) 750 MHz 1H and 1H-13C NMR spectroscopy of human blood plasma. Anal. Chem. 67 (5), 793-811. Connor, S. C., Nicholson, J. K., and Everett, J. R. (1987) Chemicalexchange and paramagnetic T2 relaxation agents for water suppression in spin-echo proton nuclear magnetic resonance spectroscopy for biological fluids. Anal. Chem. 59, 2885-2891. Beebe, K. R., Pell, R. J., and Seahsholt, M. B. (1998) Chemometrics: A practical guide, Wiley, New York. Pirouette Manual Version 2.6 (1999) Infometrix, Inc., Woodinville, WA. Phipps, A. N., Stewart, J., Wright, B., and Wilson, I. D. (1998) Effect of diet on the urinary excretion of hippuric acid and other dietary-derived aromatics in rat. A complex interaction between diet, gut microflora and substrate specificity. Xenobiotica 28, 527537. Sanins, S. M., Nicholson, J. K., Elcombe, C., and Timbrell, J. A. (1990) Hepatotoxin-induced hypertaurinuria: a proton NMR study. Arch. Toxicol. 64, 407-411. Waterfield, C. J., Turton, J. A., Scales, M. D. C., and Timbrell, J. A. (1993) Investigations into the effects of various hepatoxic compounds on urinary and liver taurine levels in rats. Arch. Toxicol. 67, 244-254. Eriksson, L., Johansson, E., Kettanah-Wold, N., and Wold, S. (1999) Introduction to multi and megavariate data analysis using projection methods (PCA and PLS) UMETRICS. Connor, S. C., Hughes, M. G., Moore, G., Lister, C. A., and Smith, S. A. (1997) Antidiabetic efficacy of BRL49653 a potent orally active insulin sensitising agent, assessed in the C57BL/KsJ db/ db diabetic mouse by non invasive proton NMR studies in urine. J. Pharm. Pharmacol. 49, 336-344.
TX990210T
