Liquid Chromatography−Mass Spectrometric Analysis of Urinary

Multivariate pattern recognition (PR) analysis combined with LC/MS was utilized to evaluate the feasibility of predicting chemical-induced toxicity in...
2 downloads 0 Views 381KB Size
Chem. Res. Toxicol. 2005, 18, 1887-1896

1887

Liquid Chromatography-Mass Spectrometric Analysis of Urinary Metabolites and Their Pattern Recognition for the Prediction of Drug-Induced Hepatotoxicity Sookie La, Hye Hyun Yoo, and Dong-Hyun Kim* Bioanalysis and Biotransformation Research Center, Korea Institute of Science and Technology, P.O. Box 131, Chungryang, Seoul 136-791, Korea Received July 12, 2005

Multivariate pattern recognition (PR) analysis combined with LC/MS was utilized to evaluate the feasibility of predicting chemical-induced toxicity in rats. Urine samples were collected from rats treated with vehicles or four hepatotoxins, R-naphthyl isothiocyanate (ANIT), carbontetrachloride (CCl4), acetaminophen, and diclofenac, and analyzed by HPLC coupled with electrospray mass spectrometry. Chromatographic data were normalized using modified Z score transformation with those of the control group, to remove the vehicle effects for a further enhanced multivariate analysis. The LC/MS-based profiles of the urine samples showed different levels of endogenous metabolites, which were characteristic of each hepatotoxin. In the principal component (PC) map of the urinary spectra from rats treated with ANIT, the metabolic trajectory moved away from the predose position, reaching a maximum separation at the 32-48 h time period. The metabolic profiles partially recovered to the basal conditions on 7 days postdose. A principal component analysis was performed on the urinary spectra of rats treated with the vehicles or four hepatotoxins. Each group formed a distinct and isolated cluster in the PC map, indicating drug-induced perturbation in the urine profiles. To construct mathematical models for predicting drug-induced hepatotoxicity, supervised analyses, such as linear discriminant analysis and soft independent modeling of class analogy with residual distance (SIMCA-RD), were performed. The SIMCA-RD showed high predictability, over 95%, in the results of cross-validation using the leaving-one-out method. The developed LC/MSPR approach might be a useful tool for the prediction of drug-induced hepatotoxicity and for the understanding of hepatotoxic mechanisms.

Introduction Toxicants disrupt a complex network of metabolic interrelationships, leading to significant changes in the concentrations of a large number of constituents in body fluids, which are not easily attainable by manual data treatment. Therefore, systematic approaches are required to study the in vivo metabolic profile changes induced by toxicants. Metabonomics, where metabolite profiling is combined with chemometric analysis, is a rapidly growing area of scientific research (1, 2). Metabonomic approaches are now being investigated by large pharmaceutical companies to screen compounds for toxicity and to select lead compounds. In metabonomics, the effect of a pharmaceutical candidate on a whole animal or individual organ is investigated by measuring the changes in endogenous metabolites over a period of time following the administration of a compound (3-5). To investigate the complex metabolic consequences of toxic reactions in biological systems, information-rich analytical approaches are required. Most work in this field has been achieved using NMR as the analytical method (1-8). NMR has the advantages of being nondestructive, applicable to intact biomaterials, and intrinsically more information-rich, with respect to the determination of molecular structures, especially in complex* To whom correspondence should be addressed. Tel: +82-2-9585055. Fax: +82-2-958-5059. E-mail: [email protected].

mixture analyses. While very effective in this context, NMR has several disadvantages as compared to MS, including poor sensitivity, the nondetection of chemical classes, such as sulfates, and difficulty in the removal of xenobiotic-related resonances from the NMR spectrum. Recently, LC/MS has emerged as a powerful means of generating multivariate metabolic data. While mass spectrometry will give both high sensitivity quantitation and structural information, the chromatography step will alleviate the issue of overlapping signals by time-resolving them. Plumbs et al. showed the potential of LC/MS to detect differences in the composition of urine from control and dosed animals in a toxicity study (9) and differences within a strain and in diurnal and gender variations (10). Most metabolic profiles are very complex and thus require proper computer-aided pattern recognition (PR)1 methods for the discrimination of abnormal from normal states (11, 12). Simple unsupervised principal component analysis (PCA) of the analytical data enables visualization of biological data sets based on the inherent similarity/dissimilarity of samples with respect to their bio1 Abbreviations: ANIT, R-naphthylisothiocyanate; CDA, canonical discriminant analysis; LC/MS, liquid chromatography/mass spectrometry; LDA, linear discriminant analysis; LOO, leave-one-out; PC, principal component; PCA, principal component analysis; PR, pattern recognition; SIMCA, soft independent modeling of class analogy; SIMCA-RD, soft independent modeling of class analogy with residual distance; ZST, Z score transformation.

10.1021/tx050187d CCC: $30.25 © 2005 American Chemical Society Published on Web 11/03/2005

1888

Chem. Res. Toxicol., Vol. 18, No. 12, 2005

chemical composition. Once a relationship has been identified between metabolic profiles and toxicity from the initial analysis, supervised approaches, such as partial least-squares discriminant analysis (13), canonical discriminant analysis (CDA) (14), linear discriminant analysis (LDA) (14), soft independent of class analogy (SIMCA) (15), and neural networks (16), may be implemented with a view to maximizing the separation between classes, and ultimately construct representative models for predicting the most likely toxicological effects of new compounds (17). In addition, data preprocessing is necessary for optimizing a metabonomic analysis. Unit variance, Pareto scaling, variable stability scaling, and Z score transformation (ZST) have been used for all variables in metabolic profiles to influence the model with equal potential or to reduce the effects of the least stable variables prior to a PR analysis (18). In the current work, data sets were normalized, using modified ZST, to decrease the variations caused in the experimental procedures, such as vehicle effects (19). Unlike usual ZST, which uses the mean and SD of an entire data set, modified Z scores are calculated from the means and SD of a control group (vehicle only-treated groups). In this study, PR analysis combined with LC/MS was applied to test the feasibility of predicting the chemicalinduced hepatotoxicity in rats caused by four prototype hepatotoxins, representing different toxic mechanisms: R-naphthylisothiocyanate (ANIT), carbontetrachloride (CCl4), acetaminophen, and diclofenac. ANIT induces intrahepatic cholestasis, while CCl4 causes a fatty liver, via free radical formation (20). Acetaminophen can result in centrilobular hepatic necrosis, with the occasional observation of nephrotoxicity (21), and hepatic biotransformation of diclofenac leads to immune-mediated chemical hepatotoxicity (20). In the present investigation, the effects of each hepatotoxin on the urinary metabolic profiles in the rat, analyzed by LC-MS, have been shown. In addition, a PR method is suggested for predicting drug-induced hepatotoxicity and the differentiation of hepatotoxic compounds based on their mechanism of action.

Experimental Procedures Animals and Treatments. Seventy-five male SpragueDawley rats (DaeHan Laboratory Animal Research Center Co., Taejeon, Korea; 250-280 g) were individually housed in metabolism cages, in a well-ventilated room at 23 ( 2 °C, with a relative humidity of 55 ( 10% and 12 h light/12 h dark cycle and allowed free access to standard laboratory chow (DaeHan Laboratory Animal, Taejon, Korea) and water. Each rat received a single oral dose of ANIT dissolved in a corn oil (100 mg/kg, n ) 9; 50 mg/kg, n ) 6; and 10 mg/kg, n ) 6), acetaminophen dissolved in a 50% PEG (700 mg/kg, n ) 6), or diclofenac dissolved in 0.9% saline (20 mg/kg, n ) 6). CCl4 was administered as a single intraperitoneal dose at 800 mg/kg (0.503 mL/ kg, dissolved in corn oil, n ) 6). In addition, three control groups were dosed with the corn oil (n ) 9), 50% PEG (n ) 6), or saline (n ) 6) vehicles. All vehicle volumes were adjusted as 10 mL/ kg. All animals were acclimatized in the metabolism cages for 1 day prior to treatment. Urine samples were collected from the ANIT-treated rats for study of the time-related variations at the following time periods: predose (24 h predose), 0-8, 8-24, 24-32, 32-48, 48-72, 72-96, 96-120, 120-144, and 144-168 h. Urine samples were also collected for study of the hepatotoxin-related variations 2 days (32-48 h) after dosing. The rat urine samples were frozen and stored at -20 °C prior to analysis. The urinary concentration of creatinine was measured

La et al. using the Jaffe reaction (22). Ten microliters of diluted urine, with a creatinine concentration of 1 mM, was injected into the LC/MS system to remove concentration differences between urine samples. Clinical Chemistry. An independent experiment was done for the clinical chemistry. Rats (eight rats in each group) were treated with the same doses of hepatotoxins described above and control group (corn oil). Blood samples were collected from four rats per group at each time by cardiac puncture under light ether anesthesia on days 1 and 2. The bloods were allowed to clot at room temperature and then centrifuged at 300 rpm for 10 min to separate the serum. Serum alanine aminotransferase (ALT) and aspartate amino transferase (AST) activities were determined on all samples. All measurements were performed using standard automated clinical chemistry instrumentation. Unpaired Student’s t-tests were used to compare the clinical chemical data between groups. Reversed Phased LC/MS. The HPLC was performed using a LC-10ADvp binary pump system, SIL-10ADvp autosampler, and CTO-10ASvp oven (Shimadzu, Kyoto, Japan). The analytical column was a Capcell Pak C18 (250 mm × 2.0 mm i.d., 5 µm, Shiseido, Japan). The HPLC mobile phases consisted of 5 mM ammonium formate (pH 4.0) (A) and 90% acetonitrile in 5 mM ammonium formate (pH 4.0) (B). A gradient program was used for the HPLC separation, with a flow rate of 0.2 mL/min. The initial composition of buffer B was 0% and increased to 35% in 8.0 min and then to 95% in a further 9.0 min, followed by reequilibration to the initial conditions for 3.0 min. Each run time was 20.0 min. The HPLC was coupled to an API2000 triple-quadrupole mass spectrometer (Applied Biosystems SCIEX, Concorde, Canada) equipped with a Turbo Ion Spray source. Electrospray ionization was performed in the negative ion mode, with nitrogen as the nebulizing, turbo spray and curtain gas, with optimum values set at 40, 35, and 75 (arbitrary values), respectively. The heated nebulizer temperature was set at 420 °C. The mass spectrometer operated with unit resolution for Q1. The instrument was operated in full scan mode, scanning form m/z 100 to 700, with a scan time of 0.5 s. The resolving poser was set to provide unit mass resolution across the entire m/z range. Data were collected from 0 to 20 min using Analyst 1.3.1 software. Data Processing. Each individual LC/MS chromatogram was converted into the netcdf format form and then into txt format using GC and GCMS View version 5.5 software (ChromSW Inc., CA). Data reduction steps were achieved using an in-house program by the statistical software package SAS 8.02 (SAS Institute Inc., Cary, NC). Data reduction was performed as follows. Each individual text file was divided into 22 fractions of 100 scans each. The noise contribution to the data was reduced by removing peaks of intensities below 30000. Mass intensities under the same m/z value in each fraction were combined together. The mass profiles of these 22 fractions were used as variables of individual object data. Peaks originating from the administered compounds and related metabolites were removed. The resulting data set was log transformed to reduce the influence from large components and the heteroscedasticity of noise structure. The log transformed data set was used for the calculation of modified z scores. Modified z scores were calculated according to the following formula:

Z)

X-X h control σcontrol

where X h control and σcontrol denoted the mean and SD, respectively, of a variable, X, for each control group. The log transformed data set and transformed data set by modified ZST, were used as the original and normalized data sets, respectively, in multivariate analysis performed afterward. Data Analysis. The multivariate analysis was performed using SAS and SIMCA-P software (version 10.0, Umetrics AB, Umea˚, Sweden). For the unsupervised methods, PCA was

Prediction of Drug-Induced Hepatotoxicity by LC/MS performed on all subject samples for data reduction and exploratory analysis of high-dimensional data sets. Subsequently, the principal component (PC) score plots of the first three PCs were constructed as the results of PCA. Having established the existence of intrinsic differences between groups, the supervised methods were performed on the data set of the control and all of the hepatotoxin groups at 32-48 h time point with inclusion of class information (control, acetaminophentreated, diclofenac-treated, CCl4-treated, and ANIT-treated). CDA was used for more of an obvious classification among each group. LDA and SIMCA with residual distance to the model (SIMCA-RD) were used to construct the prediction model for the toxin specific hepatotoxicity. These statistical tools give suitable classification rules for the groups and enable the unknown samples to be assigned to the models with the shortest residual distance among the classes. Twenty-one samples of control, nine samples of ANIT, and six samples of each acetaminophen, diclofenac, and CCl4 were used as a data set for LDA and SIMCA-RD analyses. Because each sample number was small, cross-validation using the leaving-one-out (LOO) method was performed for testing the modeling power of each supervised method as described elsewhere (23-25).

Results LC/MS Chromatograms of Urine Samples from ANIT-Treated Rats. The LC/MS spectra shown in Figure 1 illustrate the changes in the levels of urinary endogenous metabolites over a 1 week period after a single oral administration of 100 mg/kg of ANIT to rats. Significant changes in the urinary metabolites profile were observed from 32 to 48 h, with predominant changes including the appearance of bile acid derivatives, such as taurocholic acid isomers (m/z 514), glycocholic acid (m/z 464), cholic acid (m/z 407), tetrahydroxycholanoyl-taurine isomers (m/z 512), and glycomonohydroxycholic acid isomers (m/z 530). The assignment of each bile acid was performed by comparison of the MS/MS spectra and retention times, with reference standard bile acids (data not shown). The altered endogenous metabolic profile of the urine sample collected between 120 and 144 h did not return to the predose state but was similar to that of the vehicle control (corn oil) collected over the same period. The dose-related effect of the ANIT on the urinary metabolites was examined between 32 and 48 h after dosing as the highest perturbations were observed during this period. The intermediate dose of ANIT (50 mg/kg) produced similar alterations in the endogenous metabolites as compared to those of the high dose (100 mg/kg). However, few changes were observed in the urine collected from the rats administered with the low dose (10 mg/kg) (data not shown). PCA of Group Mean Time- and Dose-Dependent Data from ANIT-Treated Rats. The scores plots based on the HPLC/MS chromatograms of the urine samples from control and ANIT-treated rats over the entire time course are shown in Figure 2. First, PCA was performed on the original data set, and the first three PCs were displayed as 3D maps (Figure 2A). There was a clear difference between the PCA metabolic trajectories of the control and ANIT-treated groups, mainly in PC1. The ANIT-treated group showed a shift to the left from the control group on the PC1 axis and reached a maximum separation at 32-48 h time point. Subsequently, the ANIT-treated group did not return to the predose origin but gradually to the control region of the plot, indicating a metabolic recovery between 72 and 168 h. Both the

Chem. Res. Toxicol., Vol. 18, No. 12, 2005 1889

control and the dosed groups were furthest separated from the origin at 0-8 h time point in PC2. The positions of the two groups at 0-8 h can be seen to be close to each other, suggesting that the alteration at 0-8 h did not result from the drug treatment but from the vehicle. To remove the vehicle effect and many other effects such as individual variations, which can also affect the metabolic profile of the control group, normalization of the original data set was performed using modified ZST. When PCA was performed on the normalized data set, each time point of control group, concentered to one point, and the ANIT-treated group exhibited a more apparent shift in PC3. The trajectory of the ANIT group, likewise, reached a maximum separation at 32-48 h time point, with the later time points mapped toward the predose origin. Unlike using the original data set, the metabolic trajectory using the normalized data set showed metabolic variations, which are truly driven by ANIT treatment because the vehicle-originated changes were subtracted (Figure 2B). PCA was performed on the original data set of the control and the high, mid, and low dose groups (10, 50, and 100 mg/kg, respectively) during the 32-48 h time period. The score plot (Figure 3) revealed a clear separation between the control and the treated animals in PC1. Outlying samples in the low dose group illustrated the presence of interanimal variation. The low dose group was separated from the control group in PC2 and PC3. Nevertheless, the mid and high dose groups showed apparent separation from the low dose group, as well as from the control group, in PC1, demonstrating a more remarkable perturbation in the urinary metabolic profile. However, the mid and high dosed groups were not separated in the PCA plot. Plasma Biochemistry. Changes in the serum markers of hepatic damage after treatment of ANIT, CCl4, acetaminophen, and diclofenac are shown in Figure 4. ANIT (100 mg/kg) and CCl4 (800 mg/kg) gave rise to increases in the ALT and AST activities at 24 h, and the elevation was more pronounced at 48 h. The levels of AST at 24 h were similar to those at 48 h in the case of CCl4, but the levels of AST and ALT at 48 h were dramatically increased in the case of ANIT, suggesting a delayed onset of ANIT-induced hepatotoxicity. In rats administered 700 mg/kg acetaminophen, the levels of AST and ALT were marginally increased at 24 and 48 h. Diclofenac did not affect any of the serum markers of hepatotoxicity, although it did cause considerable interanimal variation in AST levels at 48 h, suggesting that diclofenac-like hepatotoxicant cannot be identified from a plasma biochemical analysis. LC/MS Chromatograms of Urine Samples from Rats Treated with Acetaminophen, Diclofenac, ANIT, and CCl4. The urinary metabolic profile of each hepatotoxin-treated group was analyzed by LC/MS using the samples at 32-48 h time point. The peaks of endogeneous metabolites were identified and confirmed by comparison with retention time and MS/MS chromatogram of the corresponding authentic standards. The LC/MS spectra of urine collected from ANIT-treated group showed marked and consistent changes in the levels of endogenous metabolites. A decrease more than 80% was observed in the compounds with protonated molecular ions at m/z 162, 194 (hydroxyhippuric acid), 218, 236, and 268. On the contrary, a significant increase in the urinary levels was seen in the bile acid derivatives

1890

Chem. Res. Toxicol., Vol. 18, No. 12, 2005

La et al.

Figure 1. Series of LC/MS spectra of urine samples obtained predose and at various time points, up to 168 h (as indicated), following the administration of ANIT (100 mg/kg) to rats.

(m/z 407, 464, 512, 514, and 530) and the compounds with protonated molecular ions at m/z 192 (phenylacetylglycine) and 269 (Table 1). CCl4 gave rise to the increased urinary excretion of the compounds having the ions at m/z 339, increase of 178 (hippuric acid) and 273, but dramatically decreased the levels of compounds having the ions at 160 and 207, as compared with the vehicle control. In acetaminophen-treated rats, large amounts of acetaminophen metabolites (acetaminophen

glucuronide and acetaminophen sulfate) were identified in the LC/MS spectra. These spectral parts were eliminated during the data construction process for the subsequent multivariate analysis. Among the endogenous compounds, a profound increase of the compound with a molecular ion at m/z 207 was identified in the acetaminophen-treated group as compared to vehicle control (50% PEG-treated). The changes of the endogenous metabolites in the urine samples collected from diclofenac-treated

Prediction of Drug-Induced Hepatotoxicity by LC/MS

Chem. Res. Toxicol., Vol. 18, No. 12, 2005 1891

Figure 3. PC scores plot of the urine spectra from rats, collected between 32 and 48 h postadministration of ANIT (10, 50, and 100 mg/kg) and control vehicle.

Figure 2. PC maps of the urine sample spectra from rats treated with ANIT (100 mg/kg). The mean values of each time point were plotted. (A) Original data set and (B) normalized data set.

rats were less severe than those observed in other hepatotoxicants-treated rats. 3-Hydroxy indole was increased by 40% as compared with that in the vehicle control (saline) (Figure 5). PCA of Group Mean Data from Rats Treated with Four Prototype Hepatotoxicants. Unsupervised PCA was performed to the original data set obtained from the above LC/MS chromatogram. The scores plots of the first three PCs showed that each of the four hepatotoxicant groups formed distinctive and isolated clusters, suggesting characteristic perturbations in the urine profiles due to each toxicant. The control (vehicle-treated) groups were also separated into three distinct clusters according to the vehicle types (50% PEG, 0.9% saline, and corn oil, respectively) in PC3, indicating that the vehicle type can affect the metabolic profiles in metabonomic studies. The diclofenac group was not separated from its vehicle control (saline) group. Metabolic variations induced by diclofenac were not revealed by the first three PC values in the original data set (Figure 6A). When the PC score plots were generated from the normalized data set using the modified ZST, the class separation in the PC score plots improved (Figure 6B). The control groups administrated the different vehicles formed one cluster. The PC

Figure 4. Effects of dosing ANIT (100 mg/kg), acetaminophen (700 mg/kg), diclofenac (20 mg/kg), and CCl4 (800 mg/kg) on the serum levels of (A) AST and (B) ALT on 1 day and 2 day postdose. Each data represent means ( SD (n ) 4, *p < 0.05).

score plots showed that each group was more clearly separated, and furthermore, the diclofenac-treated group was partially separated from the control group in PC2. CDA, a supervised analysis, was performed to obtain maximum separation between predefined classes. This approach is more focused on the class discriminating

1892

Chem. Res. Toxicol., Vol. 18, No. 12, 2005

La et al.

Table 1. Relative Changes in Major Endogenous Metabolite Concentrations between the Hepatotoxin-Treated Groups and Each Control Groupa % of increase or decrease m/z

RT (min)

160 162 165 178 (hippuric acid) 189 (benzenediol sulfate) 192 (phenylacetylglycine) 194 201 207 212 (3-hydroxyindole) 218 222 223 231 236 242 245 (3-hydoxyphenylpropionic acid sulfate) 268 269 273 (ferulic acid sulfate) 295 339 343 407 (cholic acid) 464 (glycocholic acid) 512 (tetrahydroxycholanoyltaurine) 514 (taurocholic acid) 530 (glycomonohydroxycholic acid)

10.4 11.1 11.5 9.0 9.8 9.7 7.0 13.6 6.7 10.9 7.2 10.3 6.5 14.9 12.1 10.1 13.5

acetaminophen

13.9 14.9 9.5 9.3 8.7 13.3 16.1 14.0 13.1

diclofenac

CCl4 70 V 40 V

40 V

300 v 60 V 61 V

500 v

70 V 40% v 40 V 60 V 60 V

30 V

70 v 30 V 2000 v

ANIT 70 V 85 V 80 V 70 V 300 v 91 V 73 V 80 V 40 V 92 V 70 V 80 V 75 V 90 V 70 V 50 V 89 V 400 v 60 V 60 V 80 V S** S** S**

13.8 13.4

S** S**

a Relative percentages were calculated for metabolites that showed statistically significant differences from those in the relevant control group (p < 0.05). *The largest peak of isomers. **Specific in the ANIT-treated group.

variation in the data as compared to the unsupervised approaches. The first 23 PCs (explained over 90% of the total variance of data set) were used as input variables. Five ellipsoidal circles were drawn around each group cluster at three SD boundaries, with 99.7% confidence levels, from each group center. When CDA was applied to PCs from the original data set, the urine samples were separated into five clusters, in more condensed forms as compared to the unsupervised PCA data, although the 99.7% confidence areas of the acetaminophen and control groups somewhat overlapped (Figure 7A). However, when CDA was applied to PCs from the normalized data set, all groups containing acetaminophen groups were completely and more definitely separated in canonical plots (Figure 7B). Consequently, classification between groups became more obvious with the CDA than with the unsupervised PCA. LDA and SIMCA-RD were applied to the original and normalized data sets to construct prediction models for drug-induced hepatotoxicity and evaluated their prediction power by calculating the correct group classification ratio of each class using LOO validation. The LDA and SIMCA-RD methods showed predictive abilities of 82.2 and 73.3%, respectively, in the original data set and 86.7 and 96.9%, respectively, in the normalized data set (Table 2). Especially, when the SIMCA-RD method was used, three of six acetaminophen samples and five of six diclofenac samples were misclassified as the control group in the original data set but this was improved in the normalized data set, in that only one of the six CCl4 samples was misclassified as the ANIT group. Consequently, the normalization process resulted in apparent

improvements in the predictive performance of the LDA and SIMCA-RD methods.

Discussions In the present investigation, the LC/MS, coupled with PR, was adapted to construct mathematical models describing the physiological variation of the urinary metabolic profiles due to the four model hepatotoxins. At first, the time- and dose-related effects of ANIT on the urinary metabolites were investigated. ANIT is a representative model hepatotoxin, which is known to induce cholestasis. Many metabonomic studies on ANIT using NMR have been reported (4-6), but the study using LC/ MS has not been reported yet. Over the 8 days of the study, many fluctuations in the concentrations of the endogenous metabolites, such as hippuric acid or bile acids, were observed in the urinary profiles of the ANITtreated rats by LC/MS analysis (Figure 1). It is considered that cholestasis-induced hepatocyte necrosis and the subsequent membrane breakdown caused by ANIT resulted in the elevation of bile acids. A bile duct obstruction would lead to a reflux of bile acids into the plasma compartment (6, 26), and as a consequence, abnormally high levels of bile acids would then be excreted into the urine. Besides an increase of urinary bile acids, wellknown by the previous NMR approach (4-6, 13), an increase of urinary phenylacetylglycine was observed additionally. Phenylacetylglycine has been suggested to be a potential biomarker for phospholipidosis (27), but further work is needed to conclusively link the marker to the effect of ANIT.

Prediction of Drug-Induced Hepatotoxicity by LC/MS

Chem. Res. Toxicol., Vol. 18, No. 12, 2005 1893

Figure 5. LC/MS spectra of the urine samples collected between 32 and 48 h postadministration of hepatotoxin and each vehicle: (A) acetaminophen, (B) diclofenac, and (C) ANIT and CCl4.

In addition to the ANIT-treated group, a clear difference was observed between the urinary metabolic profiles of the predose and 7 day postdose samples in the vehicletreated group (data not shown). One of the prominent metabolites showing large variation in the vehicle-treated group was hippuric acid. Variations in urinary hippuric acid excretion have previously been noted and are often

associated with changes in feeding regimes and the metabolism of gut microflora (28, 29). Likewise, the PCA shifts shown in the metabolic trajectory of the control urine samples (Figure 2A) are supposed to be caused by experimental procedures, such as acclimatization in the metabolism cages or vehicle effects. These metabolic variations can interfere with the detection of toxin-

1894

Chem. Res. Toxicol., Vol. 18, No. 12, 2005

Figure 6. PC scores plots of the urine sample spectra from rats collected between 32 and 48 h postadministration of acetaminophen (700 mg/kg), diclofenac (20 mg/kg), CCl4 (800 mg/kg), ANIT (100 mg/kg), and each control vehicle. (A) Original data set and (B) normalized data set.

related alterations. The normalization of data using the modified ZST method could effectively remove the interferences and vehicle effects over time. By down weighting the variables that showed similar concentrations between two groups via the modified ZST, the distinction between the classes in the subsequent multivariate models was improved. As shown in Figure 2, the ANIT-induced changes in the urinary profile over times were more distinct after normalization of the data. The alterations observed during the first 48 h period became gradually less pronounced between 48 and 144 h postdosing, which was consistent with the time-dependent changes in the urinary metabolic profiles in the ANIT-treated rats, as analyzed by NMR (4-6). The onset of the recovery process was observed in the metabolic trajectory as the course of the trajectory turned back toward the predose position at 48 h postdosing. PCA analysis of the urinary LC/MS spectra demonstrated the development of toxicity due to dose increase. A PCA shift was observed in the low dose group (10 mg/ kg), even though the chromatographic profiles were not significantly distinguished from those obtained in the

La et al.

Figure 7. Canonical plots of the PCs from rats collected between 32 and 48 h postadministration of acetaminophen (700 mg/kg), diclofenac (20 mg/kg), CCl4 (800 mg/kg), ANIT (100 mg/ kg), and each control vehicle. (A) Original data set and (B) normalized data set. Table 2. Cross-Validation (LOO) Results for Hepatotoxin Specific Classification in LDA and SIMCA-RD Modelsa predictability of constructed models original data set acetaminophen diclofenac CCl4 ANIT control total (%)

normalized data set

LDA

SIMCA-RD

LDA

SIMCA-RD

5/6 6/6 3/6 7/9 21/21 82.2

3/6 1/6 6/6 9/9 21/21 73.3

4/6 6/6 4/6 9/9 21/21 86.7

6/6 6/6 5/6 9/9 21/21 96.7

a Predictability of constructed models ) number of correct classifications/total number.

vehicle group. The inability to distinguish between the 50 and 100 mg/kg treated groups suggests that the metabolic perturbations caused by ANIT were not further pronounced at doses higher than 50 mg/kg. For comparison of the metabolic changes induced by each of the four models of hepatotoxin, the urine samples collected between 32 and 48 h were used. Although each toxin has a different onset time and duration of its toxicity, analysis at the same time point is preferable for

Prediction of Drug-Induced Hepatotoxicity by LC/MS

ultimate construction of toxicity prediction model. The 32-48 h time point was selected since the urinary metabolite profile of the ANIT-treated group showed the maximum change at that time point, and furthermore, the highest elevations of serum ALT and AST were also observed on the second day after dosing in other hepatotoxin-treated groups. The PCA results obtained from the urine samples of the selected time point showed successful classification by each hepatotoxin group. Even though the time showing the highest toxicity may vary among hepatotoxicants, the sample collection should be done during specific time periods in order to predict the type of hepatotoxicity of unknown compounds. The nonsteroidal antiinflammatory drug, diclofenac, is known to cause mitochondrial damage, with a delayed onset of toxicity (30). In this experiment, the levels of serum ALT and AST were only marginally elevated with acute diclofenac treatment, and the PCA scores plot revealed no distinctive separation between the vehicle and the diclofenac-treated animals. This phenomenon might be partially due to vehicle effects, which can mask marginal changes in the urinary compounds induced by diclofenac. The PC scores plot of the normalized data set, with the vehicle effects removed, revealed more apparent separation between the control and the model hepatotoxintreated rats. In addition, separation between the control and the diclofenac-treated rats was observed in the PC scores plot after normalization of the data set. Although PCA is a highly informative method for the metabonomic approaches, it is an unsupervised method of data analysis and consequently does not provide a suitable system for the prediction of the toxicity induced by unknown compounds. Meanwhile, a supervised approach allows for the construction of automated expert systems, which is capable of analyzing LC/MS data set and determining whether a particular urine sample is normal or abnormal. Furthermore, supervised analysis enables the construction of multivariate models for discriminating hepatotoxic mechanisms. As with PCA analysis, normalization also improved the predictabilities of supervised analysis. Especially, the four different hepatotoxin classes were distinguished with 96.7% predictability using the SIMCA-RD method. These results suggest the possible prediction of toxic compounds based on the toxic mechanism as well as the target organ. In this study using LC/MS, xenobiotic-related peaks could be easily removed from the PCA analysis based on MS/MS and sulfates such as benzenediol sulfate, 3-hydroxyphenylpropionic acid sulfate, and ferulic acid sulfate were observed mainly as markers for hepatotoxicity differently to NMR approaches. Especially, 3-hydroxyindole, which is part of the tryptophan metabolism cycle, was newly found as a potential biomarker, and increases of m/z 339 by CCl4 and m/z 207 by acetaminophen are also remarkable as putative biomarkers. MS methods alone are limited in terms of identifying compounds of completely unknown structure. Further study with FTICR mass or LC NMR-MS is required to identify the unknown metabolites shown as possible markers of hepatotoxicity, although only a few metabolites were identified and confirmed in this study due to the limitation of authentic standards available. Once the database of urinary metabolic profiles for LC/MS analysis has been compiled, the toxicological significance of these metabolites and their value as marker compounds can be more fully evaluated. Further experiments are undergoing for

Chem. Res. Toxicol., Vol. 18, No. 12, 2005 1895

the structural elucidation of urinary endogenous metabolites. In the present study, a multivariate PR method coupled with LC/MS was applied to urine samples of rats administrated each of four model hepatotoxins with different hepatotoxic mechanisms. The normalization using the modified ZST enhanced the recognition of urinary metabolic changes and the classification of hepatotoxin-related metabolic variations in the PCA pattern and also improved the predictive power of a supervised analysis. These results demonstrated that LC/MS coupled with SIMCA-RD can be a useful tool for the prediction of drug-induced hepatotoxicity and the elucidation of toxic mechanisms in the early period of drug development and safety evaluation.

Acknowledgment. The work presented was supported, in part, by National Research Laboratory grants funded by the Korean Ministry of Science and Technology and, in part, by an intramural grant from the Korea Institute of Science and Technology.

References (1) Nicholson, J. K., Connelly J., Lindon, J. C., and Homes, E. (2002) Metabonomics: A platform for studying drug toxicity and gene function. Nat. Rev. 1, 153-161. (2) Kenney, B., and Shockcor, J. P. (2003) Complementary NMR and LC-MS technologies for metabonomic studies. Pharmagenomics Nov (1), 56-63. (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 of and recovery from nephrotoxic lesions in the rat induced by mercury(II) chloride and 2-bromoethanamine. Mol. Pharmacol. 42, 922-930. (4) Beckwith-Hall, B. M., Nicholson, J. K., Nicholls, A. W., Foxall, P. J., Lindon, J. C., Connor, S. C., Abdi, M., Connelly, J., and Holmes, E. (1998) Nuclear magnetic resonance spectroscopic and principal component analysis investigations into biochemical effects of three hepatotoxins. Chem. Res. Toxicol. 11, 260-272. (5) Lindon, J. C., Homes, E., and Nicholson, J. K. (2003) So what is the deal with metabonomics? Anal. Chem. 75 (17), 385A-391A. (6) Waters, N. J., Homes, E., Williams, A., Watrfield, C. J., Farrant, D., and Nicholson, J. K. (2001) NMR and pattern recognition studies on the time-related metabolic effects of R-naphthylisothiocyanate on.liver, urine, and plasma in the rat: An integrative metabonomic approach. Chem. Res. Toxicol. 14, 1401-1412. (7) Mortishire-Smith, R. J., Skiles, G. L., Lawrence, J. W., Spence, S., Nicholls, A. W., Johnson, B. A., and Nicholson, J. K. (2004) Use of metabonomics to identify impaired fatty acid metabolism as the mechanism of a drug-induced toxicity. Chem. Res. Toxicol. 17, 165-173. (8) Williams, R. E., Williams, R. E., Jacobsen, M., and Lock, E. A. (2003) 1H NMR pattern recognition and 31P NMR studies with D-serine in rat urine and kidney, time- and dose-related metabolic effects. Chem. Res. Toxicol. 16, 1207-1216. (9) Plumb, R. S., Stumpf, C. L., Gorenstein, M. V., Castro-Perez, J. M., Dear, G. J., Anthony, M., Sweatman, B. C., Connor, S. C., and Haselden, J. N. (2003) Metabonomics: The use of electrospray mass spectrometry coupled to reverse-phase liquid chromatography shows potential for the screening of rat urine in drug development. Rapid Commun. Mass Spectrom. 17, 2541-2549. (10) Plumb, R. S., Granger, J., Stumpf, C., Wilson, I. D., Evans, J. A., and Lenz, E. M. (2003) Metabonomic analysis of mouse urine by liquid-chromatography-time-of-flight mass spectrometry (LC/ TOFMS): Detection of strain, diurnal and gender differences. Analyst 128, 819-823. (11) Rhodes, G., Miller, M., McConnell, M. L., and Novotny, M. (1981) Metabolic abnormalities associated with diabetes mellitus, as investigated by gas chromatography and pattern-recognition analysis of profiles of volatile metabolites. Clin. Chem. 27, 580585. (12) Jellum, E., Harboe, M., Bjune, G., and Wold, S. (1991) Interpreting complicated chromatographic patterns. J. Pharm. Biomed. Anal. 9, 663-669.

1896

Chem. Res. Toxicol., Vol. 18, No. 12, 2005

(13) Azmi, J., Griffin, J. L., Antti, H., Shore, R. F., Johnson, E., Nicholson, J. K., and Homes, E. (2002) Metabolic trajectory characterization of xenobiotics-induced hepatotoxin lesions using statistical batch processing of NMR data. Analyst 127, 271-276. (14) Tate, A. R., Damment, S. J. P., and Lindon, J. C. (2001) Investigation of the metabolite variation in control rat urine using 1H NMR spectroscopy, Anal. Biochem. 291, 17-26. (15) Brindle, J. T., Nicholson, J. K., Schofield, P. M., Grainger, D. J., and Homes, E. (2003) Application of chemometrics to 1H NMR spectroscopic data to investigation a relationship between human serum metabolic profiles and hypertension. Analyst 128, 32-36. (16) Homes, 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. (17) Homes, E., and Antti, H. (2002) Chemometric contributions to the evolution of metabonomics: Mathematical solutions to characterizing and interpreting complex biological NMR spectra. Analyst 127, 1549-1557. (18) Keun, H. C., Ebbels, T. M. D., Antti, H., Bollard, M. E., Beckonert, O., Holmes, E., Lindon, J. C., and Nicholson, J. K. (2003) Improved analysis of multivariate data by variable stability scaling; application to NMR-based metabolic profiling. Anal. Chim. Acta 490, 265-276. (19) Beckwith-Hall, B. M., Holmes, E., Lindon, J. C., Gounarides, J., Vickers, A., Shapiro, M., and Nicholson, J. K. (2002) NMR-based metabonomic studies on the biochemical effects of commonly used drug carrier vehicles in the rat. Chem. Res. Toxicol. 15, 11361141. (20) Moslen, M. T. (1996) Toxic response of liver. In Casarett and Doull’s Toxicology: The Basic Science of Poisons, 5th ed. (Kassen, C. D., Amdur, M. O., and Doull, J. D., Eds.) pp 403-415, McGrawHill, New York. (21) Coen, M., Ruepp, S. U., Lindon, J. C., Nicholson, J. K., Pognan, F., Lenz, E. M., and Wilson, I. D. (2004) Integrated application of trasncriptomics and metabonomics yields new insight into the toxicity due to paracetamol in the mouse. J. Pharm. Biomed. 35, 93-105. (22) Guder, W. G., Hoffmann, G. E., Hubbuch, A., Poppe, W. A., Siedel, J., and Price, C. P. (1986) Multicentre evaluation of an enzymatic

La et al.

(23)

(24)

(25)

(26)

(27)

(28)

(29)

(30)

method for creatinine determination using a sensitive colour reagent. J. Clin. Chem. Clin. Biochem. 24, 889-902. Vapnik, V. N. (1998) Statistical learning theory. Wiley Series on Adaptive and Learning Systems for Signal Processing, Communications and Control, Wiley, New York. Webb-Robertson, B. J., Lowry, D. F., Jarman, K. H., Harbo, S. J., Meng, Q. R., Fuciarelli, A. F., Pounds, J. G., and Lee, K. M. (2005) A Study of spectral integration and normalization in NMRbased metabonomic analyses. J. Pharm. Biomed. Anal. 39, 830836. Beckonert, O., Bollard, M. E., Ebbels, T. M. D., Keun, H. C., Antti, H., Holmes, E., Lindon, J. C., and Nicholson, J. K. (2003) NMRbased metabonomic toxicity classification: Hierarchical cluster analysis and k-nearest-neighbour approaches. Anal. Chim. Acta 490, 3-15. Clayton, T. A., Lindon, J. C., Everett, J. R., Charuel, C., Hanton, G., Le Net, J.-L., Provost, J.-P., and Nicholson, J. M. (2003) An hypothesis for ad mechanism underlying hepatotoxin-induced hypercreatinuria. Arch. Toxicol. 77, 208-207. Nicholls, A. W., Nicholson, J. K., Haselden, J. N., and Waterfield, C. J. (2000) A metabonomic approach to the investigation of druginduced phospholipidosis: An NMR spectroscopy and pattern recognition study. Biomakers 5, 410-423. Phipps, A. N., Wright, B., Stewart, J., and Wilson, I. D. (1997) Use of proton NMR for determining changes in metabolite excretion profiles induced by dietary changes in the rat. Pharm. Sci. 3 (3), 143-146. Gregersen, N., Mortensen, P. B., and Koelvraa, S. (1983) On the biologic origin of C6-C10-dicarboxylic and C6-C10-ω-1-hydroxy monocarboxylic acids in human and rat with acyl-CoA dehydrogenation deficiencies: In vitro studies on the ω- and ω-1-oxidation of medium-chain (C6-C12) fatty acids in human and rat liver. Pediatr. Res. 17 (10), 828-834. Boelsterli, U. A. (2003) Diclofenac-induced liver injury: a paradigm of idiosyncratic drug toxicity. Toxicol. Appl. Pharmacol. 192 (3), 307-322.

TX050187D