260
Chem. Res. Toxicol. 1998, 11, 260-272
Nuclear Magnetic Resonance Spectroscopic and Principal Components Analysis Investigations into Biochemical Effects of Three Model Hepatotoxins B. M. Beckwith-Hall,† J. K. Nicholson,† A. W. Nicholls,† P. J. D. Foxall,† J. C. Lindon,† S. C. Connor,‡ M. Abdi,§ J. Connelly,§ and E. Holmes*,† Department of Chemistry, Birkbeck College, University of London, Gordon House, 29 Gordon Square, London WC1H 0PP, U.K., Department of Analytical Science, SmithKline Beecham Pharmaceuticals, New Frontiers Science Park, 3rd Avenue, Harlow, Essex CM19 5AW, U.K., and Department of Safety Assessment, SmithKline Beecham Pharmaceuticals, The Frythe, Welwyn, Herts AL6 9AR, U.K. Received May 1, 1997 1H
NMR spectroscopy of urine combined with pattern recognition (PR) methods of data analysis has been used to investigate the time-related biochemical changes induced in SpragueDawley rats by three model hepatotoxins: R-naphthyl isothiocyanate (ANIT), D-(+)-galactosamine (GalN), and butylated hydroxytoluene (BHT). The development of hepatic lesions was monitored by conventional plasma analysis and liver histopathology. Urine was collected continuously postdosing up to 144 h and analyzed by 600-MHz 1H NMR spectroscopy. NMR spectra of the urine samples showed a number of time-dependent perturbations of endogenous metabolite levels that were characteristic for each hepatotoxin. Biochemical changes common to all three hepatotoxins included a reduction in the urinary excretion of citrate and 2-oxoglutarate and an increased excretion of taurine and creatine. Increased urinary excretion of betaine, urocanic acid, tyrosine, threonine, and glutamate was characteristic of GalN toxicity. Both GalN and ANIT caused increased urinary excretion of bile acids, while glycosuria was evident in BHT- and ANIT-treated rats. Data reduction of the NMR spectra into 256 integrated regions was used to further analyze the data. Mean values of each integrated region were analyzed by principal components analysis (PCA). Each toxin gave a unique time-related metabolic trajectory that could be visualized in two-dimensional PCA maps and in which the maximum distance from the control point corresponded to the time of greatest cellular injury (confirmed by conventional toxicological tests). Thereafter, the metabolic trajectories changed direction and moved back toward the control region of the PR map during the postdose recovery phase. The combination of urinary metabolites which were significantly altered at various time points allowed for differentiation between biliary and parenchymal injury. This NMRPR approach to the noninvasive detection of liver lesions will be of value in furthering the understanding of hepatotoxic mechanisms and assisting in the discovery of novel biomarkers of hepatotoxicity.
Introduction The main objective of the present study was to apply automatic data reduction techniques for the first time to the 1H NMR1 spectroscopic analysis of toxicological data. By using automatic data reduction techniques to convert 1H NMR biofluid spectra into a set of spectral integrals, the necessity of making any a priori assump* Address correspondence to Dr. E. Holmes. † University of London. ‡ SmithKline Beecham Pharmaceuticals, New Frontiers Science Park. § SmithKline Beecham Pharmaceuticals, The Frythe. 1 Abbreviations: R-naphthyl isothiocyanate (ANIT), alkaline phosphatase (ALP), alanine aminotransferase (ALT), butylated hydroxytoluene (BHT), 1H-1H correlation NMR spectroscopy (COSY), free induction decay (FID), D-(+)-galactosamine (GalN), reduced glutathione (GSH), methylamine (MA), nuclear magnetic resonance (NMR), nuclear Overhauser enhancement spectroscopy (NOESY), pattern recognition (PR), postdose (pd), principal component (PC), principal components analysis (PCA), solid-phase extraction chromatography (SPEC), uridine 5′-monophosphate (UMP), uridine 5′-diphosphate (UDP), uridine 5′triphosphate (UTP).
tions as to the relative importance of urinary metabolites in characterizing toxicity is removed. Routine toxicity screening generally relies on performing a batch of clinical chemical assays for selected analytes. High-frequency 1H NMR spectroscopy of biofluids enables the simultaneous measurement of changes in the levels of a wide range of endogenous metabolites. However, although information on hundreds of endogenous metabolites can be generated using 1H NMR spectroscopy, the subsequent analysis of spectra is often subjective with certain prominent resonances being selected for quantitation. In the current study we have generated a data set of 1H NMR spectra of urine following the administration of either control vehicle or model liver toxins. Automatic data reduction was then used to reduce each spectrum to a series of integrated regions of equal spectral width. The integrated regions were used as input descriptors for principal components analysis (PCA) with a view to classifying the toxins on the basis of their metabolic effects.
S0893-228x(97)00067-2 CCC: $15.00 © 1998 American Chemical Society Published on Web 04/03/1998
NMR Studies in Experimental Hepatotoxicity
Chem. Res. Toxicol., Vol. 11, No. 4, 1998 261
Table 1. Histopathological and Biochemical Alterations Recorded in Rats Given r-Naphthyl Isothiocyanate (ANIT), Butylated Hydroxytoluene (BHT), and Galactosamine (GalN) biochemical changes compound/regimen ANIT 150 mg/kg po
histopathology hepatocanalicular changes [up to 24 h] (13); biliary epithelial cell necrosis and multifocal bile duct obstruction [48-72 h] (13); decreased bile flow [by 24 h] (13); recovery by 7 days
increase plasma ALP, ALT, γ-GT, and 5′N; hepatic taurine (14); plasma bilirubin and bile acids (15)
decrease urinary taurine (15)
BHT
1000 mg/kg po × 1 centrilobular lesions (16) 500 mg/kg po × 7 periportal necrosis (17) GalN 200-1000 mg/kg ip acidophilic cytoplasmic degeneration [4 h]; hepatocyte necrosis [6 h] (18); scattered foci of hepatocyte necrosis, influx of inflammatory filtrates, and increased bile ductules [24 h] (19); recovery within 7-12 days
High-frequency 1H NMR spectroscopy of biofluids has been shown to be useful for exploring time-related biochemical changes following toxic insult. However, high-frequency 1H NMR spectra of biofluids (acquired at 500 MHz and above) often contain several thousand 1H NMR resonances providing structural and quantitative information on hundreds of endogenous metabolites. 1H NMR spectroscopy of biofluids by NMR has led to assignment of signals for biomarkers of organ-specific toxicity (1). In addition, NMR spectroscopy has been used to study the biochemical sequelae and mechanisms of toxicity of a range of compounds through noninvasive measurement of body fluids such as urine (2-9). Previous studies have shown the value of applying statistical data reduction and multivariate analysis techniques, such as PCA and nonlinear mapping (10, 11), to the analysis of 1D NMR spectroscopic data. Application of PR methods to 1H NMR data enables the reduction of the complex spectral profiles to two-dimensional maps for the purposes of classifying toxins (10). Using selected quantified NMR resonances as PR input, it has been shown that toxins can be classified according to their specific site and mechanism of action in experimental animals (12). However, like conventional clinical chemical assays, these previous studies involved selecting a set of metabolites and, therefore, involved assumptions as to which metabolites were important in defining the toxicity of the compounds. Methods of automatically reducing 1H NMR spectra to a series of integrated regions have been developed and applied to the investigation of normal physiological variance in humans (11). In this study we have applied automatic data reduction and PCA to the analysis of 600-MHz 1H NMR spectra obtained over a 7-day time course following chemically induced liver damage. Since reversible toxic lesions develop and resolve, urinary perturbations following a toxic insult are timerelated and vary with the progression of the lesion. Furthermore, we have shown that certain types of toxicological lesions are best described by time-related changes in the concentration or ratios of a combination of metabolic biomarkers (1). Using NMR-PR techniques, a time-related “trajectory” of the biochemical events associated with the development of a lesion can be constructed with a view to establishing biochemical markers for the onset and recovery phases of the lesion. This has been shown for the model nephrotoxins mercury(II)
AST, ALT, and LDH (16) GalN 1-phosphate, UDP-galactosamine, UDP-glucosamine, and N-acetylhexosamine monophosphates (20-22); sorbitol dehydrogenase, ALT and plasma bilirubin [peak at 24-48 h] (19, 23)
UTP, UDP, UMP, UDP-glucose, UDP-galactose and UDP-glucuronic acid (24); coagulation factors [24-48 h] (23)
Figure 1. Chemical structures of the hepatotoxins investigated in the present study.
chloride and 2-bromoethanamine in studies on Fischer 344 rats (12), but this approach has not been applied to other organs or tissue types. The model hepatotoxins selected for the purposes of this study were R-naphthyl isothiocyanate (ANIT), D-(+)galactosamine hydrochloride (GalN), and butylated hydroxytoluene (BHT) (Figure 1), which induce liver toxicity by different mechanisms. Changes in biochemistry and in hepatic histopathology previously reported to be produced in rats given ANIT, GalN, or BHT are summarized in Table 1. Oral administration of a single dose of ANIT (150 mg/kg) results in findings characteristic of intrahepatic cholestasis within 16-24 h (25, 26). BHT given orally produces either centrilobular or periportal damage depending on dose and period of administration (16, 17). Metabolism of single doses of GalN by enzymes of the galactose pathway in the liver leads to sequestration of uridine and reversible acute hepatitis; other regimens result in cirrhosis and liver tumors.
Materials and Methods Animals and Treatments. Twenty-eight male SD rats (Charles River UK; 250-300 g) were housed individually in metabolism cages in a well-ventilated room, at a temperature of 21 ( 2 °C and a relative humidity of 50 ( 10%, with a 12-h light/12-h dark cycle. Food and tap water were provided ad libitum. Each rat received a single dose of ANIT in 10 mL/kg of body weight of corn oil (po, 200 mg/kg, n ) 6), BHT in 10 mL/kg of body weight of corn oil (po, 1000 mg/kg, n ) 6), or GalN in 5 mL/kg of body weight of saline (ip, 600 mg/kg, n ) 6). In addition, two control groups were dosed with corn oil (po, n ) 5) or saline (ip, 0.9% w/v, n ) 5). All animals were
262 Chem. Res. Toxicol., Vol. 11, No. 4, 1998 acclimatized for 9 days in grid-based plastic cages prior to group allocation and for 1 day in metabolism cages before treatment. Urine was collected on dry ice during the following periods: predose (24-h period), 0-8, 8-24, 24-32, 32-48, 48-72, 7296, 96-120, and 120-144 h. Urinary volumes and pH values were recorded, and samples were centrifuged at 1700g for 10 min to remove food particles and other debris. For assessment of hepatic and renal histopathology, an extra 4 rats/group were housed in grid-based cages and killed 48 h after dosing by exposure to rising concentrations of CO2. Plasma Biochemistry. Blood samples (0.5-0.8 mL) were collected into heparinized tubes by tail vein puncture 24 and 48 h after treatment (all animals) and 96 and 144 h after treatment (6 animals). The plasma was immediately separated from red cells by centrifugation (1700g, 10 min, at room temperature). For each rat, plasma ALP, ALT, total bilirubin, urea, and creatinine were measured using a BM/Hitachi 717 multianalyzer, set at either 30 or 37 °C, with reagent kits supplied by Boehringer Mannheim. Assays were carried out using standard methods (27-30). Unpaired Mann-Whitney tests were used to compare clinical chemical data between groups. Histopathology. Livers and kidneys were removed from all rats immediately after death (4 rats at 48 h and 6 rats at 144 h), and samples of tissue were fixed in 10% buffered formol saline. After processing, wax-embedded sections (5 µm) were cut, stained with hematoxylin and eosin, and examined by light microscopy. 1H NMR Spectroscopy of Urine. NMR measurements were made on a Bruker AMX600 spectrometer operating at 600.13-MHz 1H frequency at ambient probe temperature (298 ( 1 K). To reduce the pH range of the samples, analysis was performed following the addition of 200 µL of 300 mM sodium phosphate buffer (made up in D2O) to 400 µL of each urine sample. The pH values of all samples then fell within the range 6.80-7.80. Sodium 3-(trimethylsilyl)[2,2,3,3-2H4]propionate (an internal standard, δ 0.0) was added to give a final concentration of 1 mM. For each sample, 64 free induction decays (FIDs) were collected into 65 536 data points using a spectral width of 12 195 Hz and an acquisition time of 2.69 s. An additional delay of 3.0 s was added between pulses to allow T1 relaxation. NMR data were acquired using a standard pulse sequence, noesypr1d (31). This pulse sequence utilizes the first increment of the NOESY pulse sequence to effect suppression of the water signal with a secondary irradiation field applied during the relaxation delay of 3 s and during the mixing time of 100 ms. This NMR experiment provides a convenient and effective solvent suppression method for biofluids (32). To assist metabolite identification, the time domain spectra were zero-filled to 131 072 data points before Fourier transformation (FT). The effects of various preprocessing methods have been extensively studied and are reported in the literature (33). An exponential apodization function was applied to the FID, corresponding to a line broadening of 0.3 Hz, and spectra were adjusted for phase and baseline offset. Resonance assignments of metabolites were confirmed by a combination of chemical shift, spin-spin coupling patterns, coupling constants, literature data (2, 4), and, in certain cases, standard additions. To aid structural identification of metabolites, two-dimensional homonuclear 1H-shift correlated (COSY) NMR experiments (34) were performed on selected urine samples. For each COSY experiment, 2048 data points were collected for each of 512 increments with 8 scans/ increment over a spectral width of 9090 Hz. The data were zerofilled to 1024 data points in the F1 domain, and a sine-bell apodization function was applied in both dimensions before double FT. To simplify the identification of urinary bile acids, solid-phase extraction chromatography with NMR detection (SPEC-NMR) was applied to selected samples (35). Urine was loaded onto C18 Bond Elut columns and eluted with a 0-100% stepwise methanol-water gradient. The eluents were collected and analyzed by 1H NMR spectroscopy. Automated Data Reduction of 1H NMR Spectra. NMR spectral data reduction was achieved using the program AMIX
Beckwith-Hall et al. (31). The spectral region δ 0.2-10.0 was segmented into regions of 0.04 ppm width giving 256 regions/spectrum. The area under the spectrum was calculated for each segmented region and expressed as an integral value (11). The region of the spectrum which included water and urea (δ 4.2-6.0) was removed from the analysis for all groups (both treated and control) to eliminate both the variation in water suppression and the variation in the integral of the urea signal due to partial cross-saturation via the solvent exchanging protons. To avoid including xenobiotic-related metabolites in the analysis, certain regions of the NMR spectra were omitted. For ANIT the entire aromatic region (δ 5.0-10.0) was omitted, whereas for BHT the spectral regions δ 8.84-8.08, 7.20-7.28, 7.04-7.08, and 1.28-1.48 were removed. GalN is rapidly excreted and contributes to the urinary NMR spectrum in the first 8 h pd. Therefore, for the GalN data set, the urine samples collected at 0-8 h were not included in the PR analysis. The resulting data sets were tabulated using the SAS software suite (version 6.10) on a Silicon Graphics Power Indigo R8000 computer. A mean “noise value” was calculated from the first 15 segments (δ 10.00-9.60) which were known to contain no detectable metabolite signals in any sample. The noise contribution to the spectral integral files was removed by subtraction of the calculated mean noise value from each segment. To reduce interference from artifacts, any integral values which were less than 5 times the calculated mean noise value for all of the tabulated spectra were discarded. All remaining segments were scaled to the total integrated area of the spectrum (having first removed the regions containing resonances from water, urea, and drug metabolites). Unsupervised PCA of 1H NMR Spectroscopic Data. To establish the presence of any intrinsic class-related patterns or clusters in the basic NMR data, PCA was performed on these data without any inclusion of information concerning the classification of samples on the basis of either treatment group or the time of urine collection. Each integral value was used as a descriptor for PCA, which was performed on both the covariance and the correlation (the standardized covariance) matrices. PCA methods involve the calculation of linear combinations of the original descriptors, the PCs, such that each PC is orthogonal to all others with the first PC (PC1) containing the largest amount of variance (subsequent PCs contain progressively less variance) (36). Thus, a plot of PC1 versus PC2 provides the most efficient 2D representation of the information contained in the data set. Two-dimensional PC plots of PC1 versus PC2 or PC3 versus PC2 were constructed in order to establish the presence of any intrinsic class-related patterns or clusters in these data. The samples were separated into their respective groups according to treatment, and the mean integral values for each treatment group at each time point were then calculated. PCA was repeated on a data set of mean values. The data sets relating to separate toxins were then merged (excluding any integral areas which were not common to all three toxins) in order to compare the effects of the toxins on urinary composition. Supervised PCA of 1H NMR Spectroscopic Data. Having established that there were intrinsic differences between groups, the samples were assigned to one of five groups (saline vehicle only, corn oil vehicle only, ANIT-treated, BHT-treated, or GalN-treated). Student’s t-tests and analyses of the eigenvectors were used to locate the spectral regions which contained metabolite resonances that varied significantly from the control and were therefore responsible for classification. The spectral regions selected as being significantly different were then examined with a view to identifying the combination of “marker” metabolites indicating the different biochemical effects of the hepatotoxins. To improve classification, the matrix was reduced to contain only those integrated regions which contributed to the separation of any two classes at a level of p < 0.005. PCA was then repeated, and the samples were mapped in plots of PC1 versus PC2 and PC3 versus PC2.
NMR Studies in Experimental Hepatotoxicity
Chem. Res. Toxicol., Vol. 11, No. 4, 1998 263
Figure 2. Effect of dosing ANIT (200 mg/kg), BHT (1000 mg/kg), and GalN (600 mg/kg) on plasma levels of (a) ALP, (b) ALT, and (c) total bilirubin at 24, 48, 96, and 144 h. Statistics: (a) p
