Metabolic profiling, metabolomic and metabonomic procedures for NMR spectroscopy of urine, plasma, serum and tissue extracts

Transcription

1 PROTOCOL Metabolic profiling, metabolomic and metabonomic procedures for NMR spectroscopy of urine, plasma, serum and tissue extracts Olaf Beckonert, Hector C Keun, Timothy M D Ebbels, Jacob Bundy, Elaine Holmes, John C Lindon & Jeremy K Nicholson Department of Biomolecular Medicine, Faculty of Medicine, Imperial College London, Sir Alexander Fleming Building, South Kensington, London, UK. Correspondence should be addressed to J.K.N. (j.nicholson@imperial.ac.uk) Nature Publishing Group natureprotocols Published online 25 October 2007; doi: /nprot Metabolic profiling, metabolomic and metabonomic studies mainly involve the multicomponent analysis of biological fluids, tissue and cell extracts using NMR spectroscopy and/or mass spectrometry (MS). We summarize the main NMR spectroscopic applications in modern metabolic research, and provide detailed protocols for biofluid (urine, serum/plasma) and tissue sample collection and preparation, including the extraction of polar and lipophilic metabolites from tissues. 1 H NMR spectroscopic techniques such as standard 1D spectroscopy, relaxation-edited, diffusion-edited and 2D J-resolved pulse sequences are widely used at the analysis stage to monitor different groups of metabolites and are described here. They are often followed by more detailed statistical analysis or additional 2D NMR analysis for biomarker discovery. The standard acquisition time per sample is 4 5 min for a simple 1D spectrum, and both preparation and analysis can be automated to allow application to high-throughput screening for clinical diagnostic and toxicological studies, as well as molecular phenotyping and functional genomics. INTRODUCTION Metabolomics 1 and metabonomics 2 encompass the comprehensive and simultaneous systematic profiling of multiple metabolite concentrations and their cellular and systemic fluctuations in response to drugs, diet, lifestyle, environment, stimuli and genetic modulations, in order to characterize the beneficial and adverse effects of such interactions Multiparametric metabolic profiling technologies are mainly centered on NMR spectroscopy 22 27,29,31,32 and mass spectrometry (MS) (usually with a chromatographic separation step) 1,29,32 34, because both spectroscopic platforms can give extensive structural and conformational information on multiple chemical classes in a single analytical procedure. Multivariate spectroscopic data are typically analyzed using chemometric and pattern recognition techniques to extract latent metabolic information, and enable sample classification and biomarker discovery 2,28, NMR spectroscopy has been used extensively for multivariate metabolic profiling of cells, tissues and biological fluids since the 1970s 22 27,31. Many NMR-based applications of metabonomics have been published, including the extensive study of physiological variation in experimental animals, such as male/female differences, age-related changes, dietary modulation, diurnal effects and phenotyping of mutant and transgenic animals 3 10,27,29,30. This has led to the development of toxicological applications in which metabonomic biofluid profiles were used to identify specific biomarkers of organ toxicity, an important economic factor of attrition in the preclinical pharmaceutical discovery process 10,28,41. A recently completed large-scale project in the COMET group (COnsortium on MEtabonomics in Toxicology), where the effects of 147 different model toxins and treatments were studied, was successful in implementing this technology for safety screening, initially for liver and kidney toxicity, across five pharmaceutical companies (Bristol-Myers- Squibb, Eli Lilly and Co., Hoffmann La Roche, NovoNordisk and Pfizer Inc.) 41,42,43. Metabonomics and metabolomics have also found application not only in the study of many diseases 12 14,44 46 but also of factors such as nutrition and gut microflora 47,48. Since the possibility of predicting postdose drug effects from predose metabolic profiles has been shown, this pharmaco-metabonomics approach has been identified as a potential precursor for personalized medicine 49. These and other large-scale applications, for example, in epidemiological studies 50, show the necessity for standardized protocols to ensure high reproducibility. In this protocol, we describe the methodology that is required to perform NMRbased metabonomic analysis of biofluids and tissue samples for metabolite profiling. Background Analytical approaches in metabonomics and metabolomics. The main analytical techniques that are employed for metabonomic studies are based on NMR spectroscopy and MS, the latter requiring a preseparation of the metabolic components using either gas chromatography after chemical derivatization, or LC 29. Capillary electrophoresis coupled to MS has also shown some promise. Other more specialized techniques such as Fourier transform infrared (FTIR) spectroscopy and arrayed electrochemical detection have been used, too. Both MS and NMR are suitable techniques for metabonomic analysis but have different analytical strengths and weaknesses and give complementary information, and those have been comprehensively reviewed 32. All metabolomic and metabonomic studies result in complex multivariate data sets that require visualization software and chemometric and bioinformatic methods for interpretation. The aim of these procedures is to produce biochemically based fingerprints that are of diagnostic or other classification value. A second stage, crucial in such studies, is to identify the substances causing the diagnosis or classification, as these become the potentially complex set of biomarkers that define 2692 VOL.2 NO NATURE PROTOCOLS

2 PROTOCOL 2007 Nature Publishing Group natureprotocols the biological or clinical context and help explain the mechanisms related to tissue damage or disease. NMR-based metabolite profiling. High-resolution NMR spectroscopy is a quantitative nondestructive, noninvasive, nonequilibrium perturbing technique that provides detailed information on solution-state molecular structures, based on atomcentered nuclear interactions and properties. NMR spectroscopic methods can also be used to probe metabolite molecular dynamics and mobility (such as ligand protein binding) through the interpretation of NMR spin relaxation and molecular diffusion properties 51. NMR is a robust and reliable technique for metabonomic applications in which high reproducibility is paramount 50. It allows the detection of a wide range of structurally diverse metabolites simultaneously, providing a metabolic snapshot at a particular time point. Metabolite concentrations down to the low micromole per liter range are readily detected in B4 5 min acquisition time using current-generation high-field spectrometers. Usually, the whole spectral information is used for further chemometric analysis, but specific NMR pulse sequences can be employed to select subsets of metabolites, if necessary. On the other hand, using no selection or extraction of metabolites up-front to investigate all possible variables is advantageous in metabonomics studies in which no prior information about metabolites is known or in which it still remains to be established whether several metabolites are linked to factors such as toxicity, disease. Given the diversity of applications of this technology across a broad swathe of biomedicine, efforts have been underway to standardize the reporting of metabonomic and metabolomic data, and inter alia 52,53 the Standard Metabolic Reporting Structures (SMRSs) group has recently provided a comprehensive report ( and summary publication of the major issues involved 54. Following SMRS, the Metabolomics Standards Initiative (MSI), which through a set of working groups is building comprehensive descriptions of the reporting needs for the technologies and various applications of the subject ( has been set up. This process is designed not to be proscriptive and does not recommend detailed protocols for data acquisition, for example. In fact, although standardization of reporting experimental data is highly desirable, it would be detrimental to the exploratory nature of the subject to allow only validated or approved procedures to be used in experimental metabolism studies. Biological sample types. Metabonomic studies generally use biofluids or cell or tissue extracts as primary sources of metabolic fingerprint data. Biofluids are usually relatively easy to obtain, particularly urine and serum or plasma, and this is important in animal and human studies. A wide range of fluids have been studied in addition to urine and plasma, including seminal fluids, amniotic fluid, cerebrospinal fluid, synovial fluid, saliva and other digestive fluids, blister and cyst fluids, lung aspirates and dialysis fluids 30.In addition, a number of NMR-based studies have used nondestructive analysis of tissue biopsy samples and their lipid and aqueous extracts. Extensive evaluations of extraction methods for NMRbased metabolic analyses have been performed 55,56. Metabolic profiling can also be used to characterize in vitro cell systems such as yeast 57, tumor cells 58 and tissue spheroids 15, which can be used as model systems for liver or tumor investigations, for example. Sample preparation approaches. Automatic sample preparation is possible for NMR spectroscopy, and standard NMR spectra typically take only a few minutes to acquire using robotic flow-injection methods 59. For large-scale studies, bar-coded vials containing the biofluid can be used and the contents of these can be transferred and prepared for analysis using robotic liquid handling technology into 96-well plates (volume ¼ 1 ml) under Laboratory Information Management System (LIMS) control. Larger density plates are becoming available for high-throughput measurements, too, especially when using capillary microprobes. Currently, using such approaches, hundreds of samples per day can be measured on one spectrometer, each taking a total data acquisition time of typically 5 min with minimal sample handling or pretreatment, although it is possible to acquire data faster than this. Alternatively, for more precious samples or for those of limited volume, conventional 5-mm diameter (or less) NMR tubes are usually used, either individuallyorusingacommercialsampletubechangerandautomatic data acquisition. For more specialist microsample applications, it is possible to use much smaller tubes, for example, 1 3-mm diameter, and even microprobes with volumes as low as 10 ml. NMR data. A typical 1 H NMR spectrum of urine contains thousands of sharp lines from predominantly low molecular weight metabolites (Fig. 1). Blood plasma and serum contain both low and high molecular weight components, and these give a wide range of signal line widths (Fig. 2): Broad bands from protein and lipoprotein signals contribute strongly to the 1 H NMR spectra, with sharp peaks from small molecules superimposed on them. The large interfering NMR signal arising from water in all biofluids is easily eliminated by the use of appropriate standard NMR solvent suppression methods. The reference compound used in aqueous media is usually the sodium salt of 3-trimethylsilylpropionic acid (TSP) with the methylene groups deuterated to avoid giving rise to peaks in the 1 H NMR spectrum. Other reference standards are DSS 9 Aromatic signals Hippurate 8 t-aconitate 7 6 Urea Allantoin Taurine * TMAO Creatinine Hippurate Citrate 2-Oxoglutarate Alanine branched-chain amino acids and organic acids TSP Sugars, polysols, amino acid CH Figure MHz 1 H NMR spectrum of control rat urine, displaying hundreds of resolved peaks. All peaks are referenced to the resonance of TSP (sodium salt of 3-trimethylsilylpropionic acid) at 0 p.p.m. The spectrum displays a wide range of metabolites such as aromatic, aliphatic compounds, sugars, amino acids and other osmolytes. The asterisk denotes the suppressed signal of water. TMAO, trimethylamine N-oxide. NATURE PROTOCOLS VOL.2 NO

3 PROTOCOL 2007 Nature Publishing Group natureprotocols [2,2-dimethyl-2-silapentane-5-sulfonate sodium salt], or for organic solvents, tetramethylsilane (TMS). The peak integrals relate directly to the number of protons giving rise to the peak, and hence to the relative concentrations of the substances in the sample. Absolute concentrations can be obtained if the sample contains an added internal standard of known concentration, or if a standard addition of the analyte of interest is added to the sample, or if the concentration of a substance is known by independent means (e.g., glucose in plasma can be quantified by a conventional biochemical assay). Recently, a synthetic electronic reference signal (ERETIC, Electronic REference To access In vivo Concentrations) has been introduced for quantitation purposes, and this method does not require any internal standards 60. Standard NMR pulse sequences, where the observed peak intensities are edited on the basis of molecular diffusion coefficients or on NMR relaxation times [such as the Carr Purcell Meiboom Gill (CPMG) spin-echo sequence], can be used to largely enhance only the contributions from macromolecules, or to selectively highlight the signals from the small molecule metabolites, respectively 20 (Fig. 3). Identification of biomarkers can involve the application of a range of techniques, but 1 H NMR spectra of urine and other biofluids, even though they are complex, allow many resonances to be assigned directly 16 based on their chemical shifts, signal multiplicities and by adding authentic material, and further information can be obtained by using spectral editing techniques or interrogation of spectral databases of authentic substances. Structure elucidation using 2D NMR spectroscopy. 2D NMR spectroscopy can be useful for increasing signal dispersion and for elucidating the connectivities between signals, thereby helping to identify biochemical substances. These include the 2D J-resolved experiment 61, which yields information on the multiplicity and coupling patterns of resonances, aiding molecule identification. Correlation spectroscopy (COSY) 62 and total correlation spectroscopy (TOCSY) experiments 63 provide spin spin coupling connectivities, giving information on which hydrogens in a molecule are close in chemical bond terms. Use of other types of nuclei of spin I ¼ 1/2, such as naturally abundant 13 C, 15 Nor 31 P, can be important to help assign NMR peaks making use of heteronuclear correlation NMR experiments. The lower sensitivity or less abundant nucleus NMR spectrum (such as 13 C) is detected indirectly using the more sensitive/abundant nucleus ( 1 H) by making use of spin spin interactions such as the one-bond 13 C 1 Hspin spin coupling between the nuclei to effect the connection. Generally, 2D experiments require much longer acquisition times than the standard 1D pulse sequences (in the range of hours while 1D 8 Tyrosine Glucose Lipid =CH * Glucose, amino acid CH Lactate Creatine Citrate Lipid = CH-CH 2 CH= Lactate Alanine Lipid =CH-CH 2 Lipid CH 2 (VLDL and LDL) Lipid CH 3 (VLDL and LDL) Acetate Valine Figure MHz 1 H NMR spectrum of blood serum sample. The spectrum shows signals of low molecular weight metabolites as well as of larger molecules such as lipoproteins. The large molecules give rise to much broader resonances in both the aromatic and the aliphatic area due to their fast relaxation properties in the NMR experiment. The signal intensities of low molecular weight compounds from d are low in this type of acquisition. LDL, low density lipoprotein; VLDL, very low density lipoprotein. spectra can generally be performed within minutes). These experiments, in particular, have benefited from the development of cryogenic probes where the detector coil and preamplifier (but not the samples) are cooled to B20 K. This has provided an improvement in spectral signal-to-noise ratios of up to a factor of 4 by increasing the detection sensitivity of the radio frequency (RF) coil and by reducing the thermal noise in the electronics of the spectrometer. With optimized cryoprobe design, direct 13 C detection at natural abundance also can be made practical 64. Hyphenated technologies. For biomarker identification, it is also possible to separate out the substances of interest on a larger scale from a complex biofluid sample using techniques such as solidphase extraction or HPLC. For metabolite identification, directly Unedited spectum (1D NOESY) T2 edited (CPMG) 0 Figure 3 Example of edited spectra of a plasma sample [all were acquired with solvent (water) suppression]. 1D nuclear Overhauser enhancement spectroscopy (NOESY) spectra generally give the best overview over all types of molecules in biofluids. Diffusion properties can be used to select mainly macromolecular signals (diffusion-edited), while Carr Purcell Meiboom Gill (CPMG) spectra use the fast relaxation of protons in macromolecules (short T 2 ) to filter particularly those signals out and leave peaks from small molecules or signals from molecules with significant segmental motion. Spectral projections from J-resolved spectra show signals from small metabolites, too; but here resolution of peaks is improved, since resonances do not show any multiplicity and appear as individual singlets. Diffusion-edited ( 10 scale) J-resolved (2D) spectral projection ppm ppm 2694 VOL.2 NO NATURE PROTOCOLS

4 PROTOCOL 2007 Nature Publishing Group natureprotocols coupled chromatography NMR spectroscopy methods can also be used. The most general of these hyphenated approaches is HPLC NMR MS in which the eluting HPLC fraction is split, with parallel analysis by directly coupled NMR and MS techniques 65. This can be operated in on-flow, stopped-flow and loopstorage modes, and thus can provide the full array of NMR- and MS-based molecular identification tools. These include 2D NMR spectroscopy as well as MS MS for the identification of fragment ions and FT-MS or time-of-flight MS (TOF-MS) for accurate mass measurement and hence derivation of molecular empirical formulae. The analytical strategies for metabolic profiling have recently been reviewed 32. High-resolution NMR of intact tissues. The development of high-resolution 1 H magic angle spinning (MAS) NMR spectroscopy has made feasible the acquisition of high-resolution NMR data on small pieces of intact tissues with no pretreatment Rapid spinning of the sample (typically at B4 6 khz) at an angle of relative to the applied magnetic field serves to reduce the loss of information caused by line broadening effects seen in nonliquid samples such as tissues. MAS NMR spectroscopy has straightforward, but manual, sample preparation. NMR spectroscopy on a tissue sample in an MAS experiment is the same as solution-state NMR, and all common pulse techniques can be employed in order to study metabolic changes and to perform molecular structure elucidation and molecular dynamic studies. Statistical spectroscopy. A new method for identifying multiple NMR peaks from the same molecule in a complex mixture, based on the concept of Statistical Total Correlation Spectroscopy (STOCSY), has been demonstrated 39,66. This takes advantage of the multicolinearity of the intensity variables in a set of spectra to display the correlation among the intensities of the various peaks across the whole sample. This method is not limited to the usual connectivities that are deducible from more standard 2D NMR spectroscopic methods. Added information is available by examining lower correlation coefficients or even negative correlations, since this leads to the connection between two or more molecular species involved in the same biochemical process. In an extension of the method, the combination of STOCSY with supervised chemometrics methods offers a new framework for the analysis of metabonomic data. In a first step, a supervised multivariate discriminant analysis can be used to extract the parts of NMR spectra related to the discrimination between two sample classes. This information is then combined with the STOCSY results to help identify the molecules responsible for metabolic variation. The STOCSY approach has also been extended to include heterogeneous forms of data. In this form, it is known as statistical heterospectroscopy (SHY) and this has been applied to the coanalysis of both NMR and mass spectra in metabonomic toxicity studies 40. This allowed better assignment of biomarkers of the toxin effect by using the correlated, but complementary, information available from the NMR and mass spectra taken on a whole sample cohort. Recently, a new approach, statistical diffusion-ordered spectroscopy (S-DOSY), has been presented, which combines diffusionordered (DO) NMR spectroscopy with STOCSY for the analysis of complex biofluids to give enhanced information recovery using the diffusion properties of biomolecules. Furthermore, a visualization tool was developed in which the apparent diffusion coefficients from DO spectra are projected onto a 1D NMR spectrum (diffusion-ordered projection spectroscopy, DOPY) 67. Experimental design Before performing the spectroscopic analysis, it is necessary to consider aspects such as sample numbers per group and randomization. If this is a pilot study, smaller sample numbers per group are often sufficient to identify trends between groups. In order to be able to validate results, sufficient sample numbers should be analyzed to achieve statistically significant results (consult and involve a statistician as early as possible in the study planning process). Generally, NMR-based metabonomics studies of biofluids have shown a high reproducibility when using NMR 21,50.Therefore, in most cases it is sufficient to have one sample per time point. Using flow automation, problems with the injection of samples are rare (and often due to problems of sample volume), but can occur. This implies that it is important to keep aliquots of samples at the sample collection stage in order to be able to repeat acquisitions or prepare fresh samples for additional 2D acquisitions. When performing the work under Good Laboratory Practice (GLP) conditions, it may be necessary to analyze multiples of every sample, which should be randomized across the whole run. In work with tissue samples, it is important to keep the time of tissue in the defrosted stage as short as possible to avoid enzymatic activity and degradation processes. Different collection and extraction protocols for tissue samples will lead to observation of different fractions in the metabolite profile. If it is important for a study to get as accurate a possible reflection of in vivo metabolism, then tissue needs to be frozen as rapidly as possible to prevent any enzymatic changes (e.g., labile phosphates and glycolytic intermediates can change on a millisecond timescale) and loss of volatiles. In most cases, it is sufficient to freeze samples directly in liquid nitrogen; it may be decided to freeze larger samples by freeze-clamping. It needs to be considered that these conditions are not always easy to achieve, for example, in clinical surgery where the care of the patient has priority, but it is important to discuss and arrange these practical aspects with personnel to achieve optimum results. Grind tissue samples in the frozen stage in a mortar and pestle or a ball mill cooled with liquid nitrogen, or transfer them directly into organic solvents (preferably ice-cold) for homogenization while still frozen. The ground samples must not be allowed to thaw before coming into contact with perchloric acid/solvents. Do not use TSP as a reference standard in plasma, serum or any other samples with a high protein content, since the compound binds to proteins resulting in a much reduced signal with a very broad line width. An alternative is the use of formate. When preparing samples into well plates, substitute H 2 Ofor urine or serum/plasma in the first and last sample in each plate to produce blank samples. This provides a check for carry-over between samples and prevents biofluid remaining in the NMR probe at the end of the run. It is also advisable to run one to two aliquots of a representative biofluid sample per well plate across the whole run this serves as a quality control (QC) measure. Limits of applicability and practical considerations NMR spectroscopy of biofluids and tissue extracts is most efficient in the millimole per liter down to the micromole per liter range. Strong ph variations between urine samples (which can have NATURE PROTOCOLS VOL.2 NO

5 PROTOCOL considerable buffering capacity) can lead to signal shifts, despite additional buffering. In most studies this is not problematic, but it should be considered in any study in which acidosis or alkalosis is induced. Dilute samples (e.g., urine) can be difficult to analyze, in particular in cryoprobes where a strong water signal gives rise to radiation damping. Generally, water suppression through presaturation has to be employed to reduce the water signal of biofluids, and this helps targeting the dynamic range to the metabolites of interest. Where simple presaturation is not efficient, excitation sculpting has been found to be beneficial (see EQUIPMENT SETUP). High salt concentrations and samples of high ionic strength (e.g., buffered urine samples) can affect the tuning and matching of probes, which may make longer pulse lengths necessary; smaller NMR tube diameters can be chosen to reduce these effects, which are particularly prominent when using cryoprobes. Very recently, autotune devices have become available on the market, which may optimize tuning and matching for each sample by option, therefore assuring optimum conditions. Sample storage: Blood serum and plasma 68 ( 80 1C) and urine (own results, 40 1C) samples can be stored for 9 months without problems, and do not show any significant differences when analyzed Nature Publishing Group natureprotocols MATERIALS REAGENTS. Na2 HPO 4, 99+% ACS, anhydrous (Sigma-Aldrich). NaH2 PO 4, 99%, anhydrous (Sigma-Aldrich). 3-(trimethylsilyl)propionic-2,2,3,3-d4 acid sodium salt, 98 atom %D (Sigma-Aldrich). Sodium azide (NaN3 ; Sigma-Aldrich). Perchloric acid (PCA), ACS reagent, 70% wt/wt (Sigma-Aldrich). K2 CO 3, ACS reagent, Z99.0% (Sigma-Aldrich). D2 O, DM-4LC (Goss Scientific Instruments). Water, HiPerSolv for HPLC, BDH (VWR International Ltd.). Chloroform, AnalaR, Z99%, BDH (VWR International Ltd.). Methanol, HiPerSolv, Z99.8%, BDH (VWR International Ltd.). Acetonitrile, NMR-Chromasolv, Z99.6%, Riedel-de-Ha n (Sigma-Aldrich). CDCl3, 100, Z99.96 atom %D, contains 0.03 vol/vol TMS (Sigma-Aldrich). Methanol-d4, 99.8% (CD3 OD; Goss Scientific) EQUIPMENT. 600 MHz Avance DRX NMR spectrometer (Bruker Biospin) or similar (Varian, Jeol, etc.). 600 MHz is not the only frequency that can be used in metabonomics applications, but it is a good compromise for cost-benefit. Flow-injection probe FI TXI 600 SB 5 mm with Z gradient (Bruker Biospin) or similar. Tube probe TXI 600 MHz S3 5-mm XYZ gradient (Bruker Biospin) or similar. Gilson 215 flow-injection system with Icon-NMR or similar. Autosampler BACS60 for tube NMR (Bruker Biospin) or similar. Gilson 215 sample preparation robot with SampleTrack (Bruker Biospin) or similar. Tissue homogenizer T25 basic (IKA Labortechnik), Potter homogenizer (B. Braun Biotech Co.) or similar. Genevac EZ2 solvent evaporator (Genevac Ltd.) or similar. Eppendorf tubes, 1.5 ml (VWR International Ltd.). 96 deep well-plate, Ritter Riplate, 1 ml, with sealing mats (Ritter) REAGENT SETUP Urine samples For animal urines or human urines possibly containing bacterial contamination: collect into labeled tubes containing NaN 3 (to result in a total concentration of azide of min 0.05% wt/vol) over ice or in a refrigerator ( 2 to+21c), and store frozen at 40 1C until analyzed. For details about sample storage see ref. 69. Prepare urine samples into 96-well plates for NMR spectroscopy using a Bruker SampleTrack system and a Gilson 215 preparation robot or equivalent technology. Alternatively, samples can be made up manually into standard 5 mm NMR tubes. In order to provide some stabilization of the urinary ph, mix urine aliquots with phosphate buffer (ph 7.4), see PROCEDURE. Blood samples Collect B0.80 ml into lithium heparin tubes (e.g., BD vacutainer, Li-heparin) to give plasma, or leave on ice to coagulate (e.g., BD vacutainer, no additive) and, after centrifugation, to result in serum. For these processes, use standard site-specific procedures. Avoid EDTA, citrate and other added stabilizers since they give additional signals in the NMR spectra. Also, collection tubes should be avoided, which use gel to separate blood cells from plasma. The time before separation of blood cells should ideally not exceed 30 min. Centrifuge samples at 1,600g for 15 min at 4 1C. Store supernatant at 40 1C until analyzed. More details about sample storage can be found in ref. 17. Prepare blood samples into 96-well plates for NMR spectroscopy using a Bruker SampleTrack system and a Gilson 215 preparation robot or equivalent technology. Alternatively, they can be made up manually into standard 5-mm NMR tubes. In order to provide diluted samples including deuterated lock solvent for NMR spectroscopy, mix sample aliquots with saline solution, see PROCEDURE. Tissue samples Freeze the tissue sample rapidly in liquid nitrogen after collection to immediately stop any enzymatic or chemical reactions (see EXPERIMENTAL DESIGN). Store samples at 80 1C. The typical sample size is 100 mg (wet mass) although as little as 20 mg can be used. Use a manual grinding method with a mortar and pestle to disrupt the tissue or, alternatively, use an electric homogenizer for homogenization in solvents. These approaches are consistent with previous evaluation studies using wet and dry, ground or homogenized, tissue samples, and in general it has been found that homogenizing provides less inter-sample variability compared with grinding wet tissue 56. Phosphate buffer Prepare phosphate buffer (ph 7.4) by weighing g Na 2 HPO 4,5.25gNaH 2 PO 4, 1 mm TSP and 3 mm NaN 3 into a 1 l volumetric flask. Add 200 ml of D 2 O and fill up to 1 l with water. Shake thoroughly, and leave in a sonicator at 40 1C, interspersed by shaking the flask, until the salts are dissolved. Saline solution Prepare a 0.9% NaCl (wt/vol) solution by weighing 9 g of NaCl into a 1-l volumetric flask. Add 100 ml of D 2 O and fill up to 1 l with water, then shake until salt is dissolved. EQUIPMENT SETUP General NMR setup Experiment setup for the individual samples: Depending on the questions asked about the biochemical properties of the compounds, the user can choose from a range of experiments, which are exemplified here for Bruker Avance spectrometers: normal one-pulse sequence (zg), 1D nuclear Overhauser enhancement spectroscopy (NOESY)-presat (noesypr1d), CPMG-presat (cpmgpr), J-resolved and diffusion-edited experiments. The most common are 1D NOESY-presat for urine, 1D NOESY-presat and CPMG-presat for blood and a normal one-pulse sequence (zg) for tissue extracts. The details of these sequences, which are used unmodified, can be found in the Supplementary Note (NMR Pulse Sequences) and in PROCEDURE. Other spectrometer manufacturers provide the same pulse sequences, but they have different names and syntax, please discuss with the supplier. Generally, all NMR experiments should be acquired at a constant temperature; we perform experiments at 300 K. Ensure that temperature has been calibrated using a standard methanol sample, for example. All NMR experiments require that the correct 901 pulse length is determined on a representative sample this is generally the first biofluid sample in the run. The same sample is used to determine the offset of the water signal for the water suppression. Both parameters are then used for the whole data set. Where an automatic pulse calibration per sample is possible, choose this option to improve quality. Water suppression. In order to observe the dynamic range of metabolite concentrations efficiently, the water signal needs to be suppressed, ideally to or below the highest metabolite peak. Generally, the normal water presaturation is effective and the power level is optimized so that optimum suppression is achieved without suppressing metabolite resonances next to it. Due to the chemical exchange with water, the urea peak is influenced by water presaturation this requires the elimination of this peak in later chemometric 2696 VOL.2 NO NATURE PROTOCOLS

6 PROTOCOL 2007 Nature Publishing Group natureprotocols analysis. In very dilute samples, sufficient water suppression may not be achieved using presaturation, and excitation sculpting has proven useful to eliminate the water peak 70,71. The user may decide to run all experiments either with the same receiver gain (urine, blood) or with an automatically adjusted receiver gain when the sample concentration varies too much (urine or tissue extracts). The following procedure describes recommendations for the number of scans per experiment for the individual biological matrices. The user may decide to increase the number of scans per experiment for the individual biological matrices if the signal-to-noise ratio is too low for the metabolites of interest, but this will have an impact on the overall experiment time per sample. Several aspects of the NMR acquisition can be automated in programs that are linked with the pulse sequence (Bruker: au programs). Among these are, in consecutive order, a waiting delay for the sample to reach a constant temperature (0.5 1 min on flow-injection probes with a heated transfer line, 4 5 min on tube probes), automatic locking onto the reference signal of the deuterated solvent, homogenization of the magnetic field ( shimming ), receiver gain adjustment, acquisition and automatic processing (apodization, Fourier transformation, phasing and baseline correction). Recent developments of NMR vendors now include, besides automated pulse calibration, automatic tuning and matching on every sample in automation. This is especially of advantage in samples where salt concentration varies between samples (e.g., urines). Choose this option, where possible. General maintenance of the flow-injection NMR system The push solvent is 0.1% NaN 3 in H 2 O (wt/vol; deionized). After each run, flush the system using hydrogen peroxide (10%)(wt/vol), followed by HCl (0.1 M) and finally H 2 O containing azide (0.1%). Once a week, flush the probe manually with disinfectant and allow the solution to stand in the probe for 1 h. Periodically, as required by continuous monitoring of performance, thoroughly clean the probe by flushing for several hours after removal from the magnet. PROCEDURE Sample preparation 1 Prepare urine, blood and tissue samples using the guidelines described in options A (Biofluids) and B, C or D (extraction of tissues with acetonitrile, perchloric acid or methanol/chloroform/water, respectively). Again, quantities apply to Bruker 5 mm tube NMR and the above-described flow-nmr conditions, respectively. Adjust the quantities accordingly depending on different vendor requirements and/or when varying tube size, for example. (A) Biofluids (urine and blood plasma/serum) (i) Rat urine, human urine: Mix 400 ml of urine with 200 ml of phosphate buffer (ph 7.4). Adjust these volumes according to the NMR probe used. Prepare samples either into well plates (flow-injection) or Eppendorf tubes (tube NMR). (ii) Mouse urine: Mix 200 ml of urine with 200 ml H 2 O and 200 ml of phosphate buffer (ph 7.4). Adjust these volumes according to the NMR probe used. Prepare samples either into well plates (flow-injection) or Eppendorf tubes (tube NMR). (iii) Plasma or serum: Add aliquots (200 ml) to 400 ml of 0.9% saline. Adjust these volumes according to the NMR probe used. Prepare samples either into well plates (flow-injection) or Eppendorf tubes (tube NMR). (iv) Well plates: Substitute H 2 O for urine or serum/plasma in the first and last sample in each plate to produce blank samples. Prepare 1 2 QC biofluid aliquots into each well plate. Centrifuge the well plate at 1,800g for 5 min to remove insoluble material, before positioning in the NMR flow-injection Gilson robot. Inject 500 ml of sample into the probe (Bruker flow-injection probe, this volume may vary between different makes of probes); depending on the requirements and make of equipment, it may be decided to use air gaps in between sample and push solvent or similar. (v) NMR tubes: Centrifuge Eppendorf tubes at 12,000g for 5 min at 4 1C, and transfer 550 ml of sample into 5 mm NMR tubes. (vi) Proceed to Step 3 (NMR DATA ACQUISITION AND PREPROCESSING). (B) Extraction of polar metabolites from tissues using acetonitrile (i) Prepare ice-cold solvents. (ii) Weigh intact frozen tissue, then homogenize in 50% acetonitrile/50% H 2 O (vol/vol) (5 ml g 1 tissue). Centrifuge at 12,000g for 10 min at 4 1C. (iii) Collect the supernatant and lyophilize. PAUSE POINT Store samples at 80 1C until NMR acquisition 18,19. (iv) Proceed to Step 2. (C) Extraction of polar metabolites from tissues using perchloric acid (i) Weigh the frozen tissue and grind it in a mortar cooled with liquid nitrogen. Then add 5 ml g 1 of 6% ice-cold perchloric acid and grind further to completely mix the acid and sample, and then allow it to thaw in the mortar. Alternatively weigh the frozen tissue and homogenize in the ice-cold solution straight away. (ii) Vortex the sample and then place on ice for 10 min. (iii) Centrifuge for 10 min at 12,000g at 4 1C. (iv) Neutralize the supernatant (e.g., 500 ml) to ph 7.4 with 2 M K 2 CO 3 and leave for 30 min on ice to precipitate the potassium perchlorate salts, taking care to add K 2 CO 3 slowly to avoid loss of sample during effervescence. (v) Check and adjust the ph if necessary, then centrifuge the sample and freeze dry the supernatant. PAUSE POINT Store samples at 80 1C (refs. 72, 73). (vi) Proceed to Step 2. (D) Combined extraction of polar and lipophilic metabolites from tissues using methanol/chloroform/water (i) Prepare ice-cold solvents: methanol, chloroform, water. (ii) Weigh intact frozen tissue, then transfer into a glass vial. (iii) Homogenize after adding 4 ml of methanol per gram of tissue and 0.85 ml g 1 water to the sample. Vortex the sample, then add 2 ml g 1 chloroform to the sample and vortex again. (iv) Add 2 ml g 1 chloroformand2mlg 1 water to the sample and vortex again. NATURE PROTOCOLS VOL.2 NO

7 PROTOCOL (v) Leave sample on ice or in the fridge for 15 min. Centrifuge at 1,000g for 15 min at 4 1C. The solutions should now separate into an upper methanol/water phase (with polar metabolites) and a lower chloroform phase (with lipophilic compounds), separated by protein and cellular debris. Centrifuge again if still no clear separation. (vi) In turn, transfer the upper and lower layers of each sample into separate glass vials. Remove the solvents from the samples using a speed vacuum concentrator or under a stream of nitrogen. PAUSE POINT Store aqueous extract samples at 80 1C until required. Keep the lipophilic phase in deuterated organic solvent (see Step 2(B); to reduce effects of oxidation) in the refrigerator until required, but preferably not longer than 2 3 d before NMR acquisition 56, (vii) Proceed to Step Nature Publishing Group natureprotocols Processing of tissue extracts for NMR 2 Process the tissue extracts for NMR acquisition using option A for aqueous extracts/water soluble metabolites and option B for lipophilic extracts/lipid metabolites. (A) Preparation of aqueous extracts/water-soluble metabolites for NMR spectroscopy (i) Before NMR acquisition, resuspend the polar tissue extracts in either 580 ml NMR buffer (100 mm sodium phosphate buffer, ph 7.4, in D 2 O, containing mm TSP and optionally 0.2% NaN 3 )orind 2 O containing TSP as a chemical shift reference (d ¼ 0 p.p.m.). (ii) Vortex samples and then centrifuge at 12,000g for 5 min. (iii) Transfer 550 ml of the supernatant into an NMR tube and proceed to Step 3. (B) Preparation of lipophilic extracts/lipid metabolites for NMR spectroscopy (i) Resuspend the lipophilic tissue extracts in 580 ml deuterated NMR solvent (2:1 mixture of chloroform-d (CDCl 3 ) containing 0.03 vol/vol TMS, and CD 3 OD) and then vortex. This method works well when running NMR experiments manually. For automated runs, we have found that resuspending the lipophilic tissue extracts in 580 ml deuterated CDCl 3 containing 0.03 vol/vol TMS only is robust and reliable with regards to locking and shimming. (ii) After centrifugation (1,000g, 5 min), transfer 550 ml of the supernatant into an NMR tube and proceed to Step 3. NMR data acquisition and preprocessing 3 Set temperature to 300 K (one may decide to set the temperature higher for some plasma analysis). 4 Load sample into probe and leave sufficient time to equilibrate (see EQUIPMENT SETUP). 5 Using a representative sample (this could be the QC sample): (i) Tune and match the probe. (ii) Set the RF carrier frequency offset value to the H 2 O resonance and determine the water saturation power. For excitation sculpting, adjust the power for the shaped pulse. (iii) Determine the 901 pulse length at a given power level. (iv) Readjust the frequency offset for water signal suppression, if necessary. (v) Transfer these settings to the experiments to be submitted. 6 Select suitable experiments for the samples: 1D NOESY-presat (option A) for urine, 1D NOESY-presat and CPMG-presat (option B) for plasma and serum [optionally also J-resolved (option C) and diffusion-edited (option D)] and a single-pulse sequence (option E) (or 1D NOESY-presat or CPMG-presat) for extracts. Processing parameters: If not mentioned otherwise, the 1D spectra are generally processed by applying a line broadening of Hz and zero-filling by a factor of 2 to give 64k frequency domain data points. (A) 1D NOESY-presat sequence (i) Measure the 1 H NMR spectra of biofluids and aqueous extracts using a specified water suppression pulse sequence such as 1D NOESY-presat, which employs the first increment of a NOESY pulse sequence with water irradiation during the relaxation delay and also during the mixing time (on a Bruker instrument, this is called noesypr1d). This has the form RD-901-t-901-t m -901-ACQ, where RD is the relaxation delay, t is a short delay typically of B3 ms, 901 represents a 901 RF pulse, t m is the mixing time and ACQ is the data acquisition period 16. The mixing time is t m ¼ ms (but the presaturation quality is not very sensitive to this value), and other parameters are spectral width ¼ 20 p.p.m., number of time domain data points ¼ 32,768, relaxation delay, RD ¼ 2.0 s, acquisition time ¼ 1.36 s, number of scans ¼ 64 for urine and n ¼ 128 for plasma, serum and tissue extracts, and the receiver gain is set to fill the digitizer as closely as possible. This results in a total acquisition time of B4 5 min per sample (64 scans). In addition to these routine settings, there has been some suggestion that a similar sequence could be employed, which uses gradients in the sequence and improves solvent suppression quality (e.g., noesygppr1d). Then, only a short mixing time of 10 ms can be used, which is of benefit for suppressing the relaxation effects for quantification. (ii) For processing, apply a line broadening of Hz to the FID and zero-fill to double the number of Fourier domain points to 64k. A target spectral resolution is required, typically a line width at half height on the TSP resonance of not 2698 VOL.2 NO NATURE PROTOCOLS

8 PROTOCOL 2007 Nature Publishing Group natureprotocols 42.5 Hz (with a line broadening of 1 Hz) and sufficient resolution in the manufacturer s hump test (Bruker, compare to similar procedures from other providers) to ensure that the 29 Si satellite peaks are evident. For urine samples containing high levels of protein or salt concentration, this will not be achieved. Failure to reach this resolution need not cause the data acquisition to be halted, but it can be used as a cut-off for subsequent data analysis. A target signal-to-noise ratio on the TSP signal is not necessary, although this might be useful for making comparisons of spectrometer performance at different field strengths and NMR probe volumes. A target signal-to-noise ratio on the peaks from the endogenous metabolites is not possible due to some toxins causing large changes in urinary volumes and hence dilution. (B) CPMG relaxation-editing sequence with presaturation (i) In addition for serum or plasma, acquire relaxation-edited spectra using the CPMG-presat sequence, J-resolved spectra and diffusion-edited spectra. At times, it might also be appropriate to use the CPMG-presat sequence for water-soluble metabolites to attenuate the NMR signals of any remaining proteins. The CPMG-presat pulse sequence 77 has the form RD-901-(t-1801-t) n -ACQ, where the definitions are as above, plus 1801 is a 1801 RF pulse, t is the spin-echo delay and n represents the number of loops. Typically, the number of scans is 8 n n, number of dummy scans ¼ 16, number of scans ¼ 128, number of time domain points ¼ 32k, spectrum width ¼ 20 p.p.m., relaxation delay ¼ 2 s, acquisition time ¼ 1.36 s, number of loops, n ¼ 80 or higher (depending on the intensity of macromolecular peaks), spin-echo delay, t ¼ 400 ms, giving a total echo time ¼ 64 ms. Irradiation is at the water peak during RD. (C) J-resolved (i) Set up the J-resolved pulse sequence 61 in the form RD-901t t 1 -ACQ, where t 1 is an incremented time period, number of scans ¼ 4 n n, number of dummy scans ¼ 16, number of scans per increment ¼ 16, number of time domain points ¼ 32k, number of increments ¼32, spectral widths ¼ 20 p.p.m. (F2) and 40 Hz (F1), acquisition time ¼ 1.36 s, relaxation delay ¼ 2 s. (ii) After Fourier transformation (it might be necessary to use linear prediction to improve results), perform baseline correction, then tilt by 451 to ensure that all the J-coupled multiplets (the F1 axis) are orthogonal to the chemical shift axis (F2), baseline-correct again, then symmetrize the spectra about the J ¼ 0 Hz line. Although the 2D J-resolved spectrum is obtained, pattern recognition is often carried out on the sum or skyline F2 projection. (D) Diffusion-edited (i) Acquire diffusion-edited spectra using a pulse sequence 78 with bipolar gradients and the LED scheme that has the form -RD-901-G G G 2 -T-901-G G G 2 -t-901-acquire FID, where RD is a relaxation delay, 901 is a 901 RF pulse, G 1 is the pulsed-field gradient that is applied to allow editing, 1801 is a 1801 RF pulse, G 2 is a spoil gradient applied to remove unwanted magnetization components. The diffusion delay D is the time during which the molecules are allowed to diffuse this is the period (901-G G G 2 -T-); and t is a delay to allow the longitudinal eddy currents caused within the sample to decay. The settings are relaxation delay ¼ 2 s, number of scans ¼ 64, number of time domain points ¼ 32k, spectral width ¼ 20 p.p.m., acquisition time ¼ 1.36 s, D ¼ 0.1 s, t ¼ 5 ms, the length of G 1 is 1 ms (recovery delay: 100 ms) and of G 2 is 2 ms, the ratios of the gradients are given in Supplementary Notes. (E) Normal one-pulse sequence (i) The 1 H NMR spectra of lipidic extracts do not require presaturation of the water resonance. Here, use a simple 901 pulse-acquire sequence, if measuring fully relaxed spectra. Alternatively, the Ernst angle can be used to acquire a higher number of scans in shorter time and to obtain an optimum signal-to-noise ratio 79,80. Typically, collect 64 transients into 32,768 data points, other parameters are relaxation delay ¼ 2 s, spectral width ¼ 20 p.p.m., acquisition time ¼ 1.36 s.? TROUBLESHOOTING TIMING Sample preparation Sample preparation of biofluids can be automated and takes B1 1.5 h per 96-well plate. Tissue extract preparation is generally performed manually and takes B4 6 h per sample batch; the process takes longer if drying down of organic solvents with nitrogen gas is involved. NMR acquisition For the NMR experiments, a throughput of 120 urine samples a day, 48 serum/plasma samples (1D NOESY-presat & CPMG-presat) or 72 extracts can be achieved.? TROUBLESHOOTING Main problems can be avoided by ensuring that water offset and 901 pulse length are adjusted on representative samples. Also very important: ensure that sample volumes are constant to avoid problems with automatic locking or shimming. Problems with baseline rolling ( wiggles ) can occur when sample concentrations vary too much and the receiver may be overloaded. In this case, select automatic receiver gain adjustment. NATURE PROTOCOLS VOL.2 NO

9 PROTOCOL 2007 Nature Publishing Group natureprotocols Another problem can be the water presaturation when sample concentration is low. This can be specifically problematic when the aqueous samples are run in a cryoprobe where radiation damping effect is much stronger. In these cases, excitation sculpting has proven very useful (a possible sequence in Bruker terminology is called zgesgp). ANTICIPATED RESULTS Reproducibility Over the past decade, NMR-based metabonomics has been applied successfully to investigate many different areas of physiological variation, toxicity, disease and, recently, response prediction. All studies have in common that they require thorough study planning and experimental execution. Studies, which can be treated as pilot studies have smaller sample numbers and are sufficient to spot trends or strong responses in metabolism. Studies with larger sample numbers are necessary to evaluate findings using, for example, training and test set classifications. For all these applications, it is paramount that the analytical technology has a high reproducibility, implying that the metabolic effects of disease or treatment as detected with metabonomic technology, as well as the intersubject variation, are much higher than the analytical variability. Two large-scale studies have been used to validate the reproducibility of the metabonomic platform, as well as for the investigation of metabolic effects. One is a population study (INTERMAP 50 ) analyzing phenotypic differences between 24-h urine samples obtained from volunteers in Japan (n ¼ 259), USA (n ¼ 315) and China (n ¼ 278). The study was used to investigate analytical reproducibility, urine sample storage procedures, inter-instrument variability and split-specimen detection. It was found that the multivariate analytical reproducibility of the NMR screening platform was 498%, and most classification errors were due to urine specimen handling variability. As a further result of this study, novel combinations of biomarkers were identified, which separated the samples from different populations. These findings were attributed to differences in genetic, dietary and gut microbiological factors. Particularly in the analysis of toxicity, a comprehensive exploration was achieved by the COMET consortium formed between five pharmaceutical companies and Imperial College London, UK, with the aim of developing methodologies for the acquisition and evaluation of metabonomic data generated using 1 H NMR spectroscopy of urine and blood serum from rats and mice for preclinical toxicological screening of candidate drugs 41. Numerous methodological publications on the use of metabonomics in developing screening modules arose from the COMET project 42. The COMET group showed that it is possible to construct predictive and informative models of toxicity using NMR-based metabonomic data, delineating the whole time course of toxicity 43. The project goals of the generation of comprehensive metabonomic databases (B35,000 NMR spectra covering a wide range of model toxins and treatments, 147 in total) and successful and robust multivariate statistical models (expert systems) for prediction of toxicity, initially for liver and kidney toxicity in the rat and mouse, were achieved 42,43. As part of this project, the analytical reproducibility of metabonomic protocols, sample preparation and NMR data acquisition were investigated. These were performed at two sites (one using a 500 MHz and the other using a 600 MHz NMR system) analyzing two identical (split) sets of urine samples from an 8-d acute study of hydrazine toxicity in the rat 21. Figure 4 shows the response profile after treatment with a low and a high dose of hydrazine in a principal component analysis (PCA) plot. A position on the map symbolizes one metabonomic fingerprint (an NMR spectrum) of a urine sample. Figure 5 Score (t) scatter plots for PC1 versus PC2 from a principal component analysis model using the data acquired at two different sites using the same protocols for the analysis of split aliquots in a study on hydrazine toxicity (triangles and circles represent the two different datasets, respectively). The ellipse denotes the 95% significance limit of the model. An asterisk indicates a putative outlier. The plot clearly highlights that interaliquot differences are very low compared to the physiological changes seen at different time points: the composition changes from 16 h (right hand corner) to h (left hand corner) and back again to 168 h (right-hand corner), following an L-shape 21. Reprinted with permission from ref. 21. Copyright (2002) American Chemical Society. t[2] t[1] Ctrl Low dose High dose Figure 4 Mean score trajectories of principal component analysis of urinary NMR spectral data for each dose group (control, low-dose, high-dose) showing progression of metabolic effects of hydrazine treatment. Time of sampling is indicated. High-dose data also show standard deviations across replicates (n ¼ 8upto48h;n ¼ 4 post 48 h). NMR data from one site only is shown 21. Reprinted with permission from ref. 21. Copyright (2002) American Chemical Society. t[2] * t[1] 2700 VOL.2 NO NATURE PROTOCOLS

10 PROTOCOL 2007 Nature Publishing Group natureprotocols If this position changes in the course of a study, it means that at least the concentration of one of the metabolites in the urine NMR spectrum changes (often, several metabolites change at the same time). In this particular example, every dot is the average of either four or eight spectra (four or eight individuals) at a certain time point. It can be clearly seen that the magnitude of response to the toxin is dose-dependent, but follows a similar trajectory (time profile). The same metabonomics protocol for sample preparation and NMR acquisition was performed at another site, preparing and analyzing split aliquots of the same sample. Figure 5 illustrates that, despite analysis on different spectrometers, the obtained results were very similar and gave near-identical descriptions of the metabolic response (trajectory) to hydrazine treatment. The main consistent difference between the datasets was related to the efficiency of water resonance suppression in the spectra. In this model of both datasets combined, describing all systematic dose- and time-related variation, differences between the two datasets (split-sample differences) accounted for only 3% of the total modeled variance compared to B15% for normal physiological (predose) variation. Furthermore, Mercury chloride o3% of spectra displayed distinct inter-site differences, and these were clearly identified as outliers in their respective dose group PCA models. It was noted that no samples produced clear outliers in both data sets, suggesting that the outliers observed did not reflect an unusual sample composition, but rather sporadic differences in sample preparation leading to, for example, very dilute samples. Any variation in the comparison of split samples needs to be seen in relation to the intersample variation between control animals, and, eventually, to the metabonomic effect of a treatment or disease. Both, interanimal variation between controls as well as the metabonomic effect, were much higher, and this underlines the robustness and high reproducibility of the NMRbased metabonomic technology. Metabonomics is well suited for the detection of different toxic episodes and other metabolic alterations. Our example in Figure 6 highlights that different toxins, for example, affecting kidney, liver and pancreas, give rise to different metabolic trajectories with different magnitude, direction, onset and recovery. This metabolic information was used to establish an expert system for the predicition of toxicity within COMET 43. The protocols described in this publication for the analysis of urine, blood samples and tissue extracts have been optimized to provide high-quality data with a range of sample types; but experiments can always be tuned to give optimal performance for particular samples. Hence, they are not to be regarded as being unchangeable. However, if large-scale statistical analysis is to be performed on such data sets, then it is imperative to maintain a constancy of detailed experimental protocol; otherwise, systematic analytical differences will impair data set comparison and may give rise to false biomarker reporting. PC2 score PC1 score Hexachlorobutadiene Hydrazine (Study B) Hydrazine (Study E) Streptozotocin Figure 6 Metabonomics identifies different responses from the urine profiles of animals treated with different toxins. A high similarity between trajectories can be seen where toxins which affect the same organ were given (hydrazine: liver; mercury chloride and hexachlorobutadiene: kidney; streptozotocin: pancreas). Each dot on the PCA plot symbolizes the mean of ten urine samples up to 48 h, and five urine samples until 168 h (courtesy of H. Keun, IC). Note: Supplementary information is available via the HTML version of this article. ACKNOWLEDGMENTS We thank our academic and industrial collaborators for helpful discussions in the formulation of this paper, including those participating in the COMET project. Published online at Reprints and permissions information is available online at reprintsandpermissions 1. Fiehn, O. Metabolomics the link between genotypes and phenotypes. Plant Mol. Biol. 48, (2002). 2. Nicholson, J.K., Lindon, J.C. & Holmes, E. Metabonomics : understanding the metabolic responses of living systems to pathophysiological stimuli via multivariate statistical analysis of biological NMR spectroscopic data. Xenobiotica 29, (1999). 3. Bell, J.D., Sadler, P.J., Morris, V.C. & Levander, O.A. Effect of ageing and diet on proton NMR spectra of rat urine. Magn. Reson. Med. 17, (1991). 4. Bollard, M.E. et al. Investigations into biochemical changes due to diurnal variation and estrus cycle in female rats using high resolution 1 HNMR spectroscopy of urine and pattern recognition. Anal. Biochem. 295, (2001). 5. Phipps, A.N., Wright, B., Stewart, J. & Wilson, I.D. Use of proton NMR for determining changes in metabolite excretion profiles induced by dietary changes in the rat. Pharm. Sci. 3, (1997). 6. Gavaghan, C.L., Wilson, I.D. & Nicholson, J.K. Physiological variation in metabolic phenotyping and functional genomic studies: use of orthogonal signal correction and PLSDA. FEBS Lett. 530, (2002). 7. Holmes, E. et al. Chemometric models for toxicity classification based on NMR spectra of biofluids. Chem. Res. Toxicol. 13, (2000). 8. Gartland, K.P., Bonner, F.W. & Nicholson, J.K. Investigations into the biochemical effects of region-specific nephrotoxic lesions. Mol. Pharmacol. 35, (1989). 9. Beckwith-Hall, B.M. et al. Nuclear magnetic resonance spectroscopic and principal components analysis investigations into biochemical effects of three model hepatotoxins. Chem. Res. Toxicol. 11, (1998). 10. Robertson, D.G., Reily, M.D., Lindon, J.C., Holmes, E. & Nicholson, J.K. Metabonomic technology as a tool for rapid throughput in vivo toxicity screening. NATURE PROTOCOLS VOL.2 NO

11 PROTOCOL 2007 Nature Publishing Group natureprotocols in Compr. Toxicol., Molecular and Cell. Toxicol., 14, , eds. Vanden Heuvel, J.P., Perdew, G.H., Mattes, W.B., Greenlee, W.F., Elsevier (2002). 11. Garrod, S. et al. High resolution magic angle spinning 1 H NMR spectroscopic studies on intact rat renal cortex and medulla. Magn. Reson. Med. 41, (1999). 12. Moka, D. et al. Biochemical classification of kidney carcinoma biopsy samples using magic-angle-spinning 1 H nuclear magnetic resonance spectroscopy. J. Pharm. Biomed. Anal. 17, (1998). 13. Cheng, L.L. & Pohl, U. The role of NMR-based metabolomics in cancer. in Handbook of Metabonomics and Metabolomics (eds. Lindon, J.C., Nicholson, J.K. & Holmes, E.) (Elsevier, Amsterdam, 2006). 14. Swanson, M.G. et al. Proton HR-MAS spectroscopy and quantitative pathologic analysis of MRI/3D-MRSI-targeted postsurgical prostate tissues. Magn. Reson. Med. 50, (2003). 15. Bollard, M.E. et al. Metabolic profiling of the effects of D-galactosamine in liver spheroids using 1 H NMR and MAS NMR spectroscopy. Chem. Res. Toxicol. 15, (2002). 16. Nicholson, J.K., Foxall, P.J., Spraul, M., Farrant, R.D. & Lindon, J.C. 750 MHz 1 H and 1 H- 13 C NMR spectroscopy of human blood plasma. Anal. Chem. 67, (1995). 17. Teahan, O. et al. Impact of analytical bias in metabonomic studies of human blood serum and plasma. Anal. Chem. 78, (2006). 18. Waters, N.J., Holmes, E., Waterfield, C.J., Farrant, R.D. & Nicholson, J.K. NMR and pattern recognition studies on liver extracts and intact livers from rats treated with a-napthylisothiocyanate. Biochem. Pharmacol. 64, (2002). 19. Coen, M. et al. An integrated metabonomic investigation of acetaminophen toxicity in the mouse using NMR spectroscopy. Chem. Res. Toxicol. 16, (2003). 20. Wang, Y. et al. Spectral editing and pattern recognition methods applied to highresolution magic-angle spinning 1 H nuclear magnetic resonance spectroscopy of liver tissues. Anal. Biochem. 323, (2003). 21. Keun, H.C. et al. Analytical reproducibility in 1 HNMR-basedmetabonomic urinalysis. Chem. Res. Toxicol. 15, (2002). 22. Brown, F.F., Campbell, I.D., Kuchel, P.W. & Rabenstein, D.C. Human erythrocyte metabolism studies by 1 H spin echo NMR. FEBS Lett. 82, (1977). 23. Brindle, K.M., Brown, F.F., Campbell, I.D., Grathwohl, C. & Kuchel, P.W. Application of spin-echo nuclear magnetic resonance to whole-cell systems. Membrane transport. Biochem. J. 180, (1979). 24. Nicholson, J.K., Buckingham, M.J. & Sadler, P.J. High resolution 1 H n.m.r. studies of vertebrate blood and plasma. Biochem. J. 211, (1983). 25. Nicholson, J.K. et al. Proton nuclear-magnetic-resonance studies of serum, plasma and urine from fasting normal and diabetic subjects. Biochem. J. 217, (1984). 26. Nicholson, J.K., Timbrell, J.A. & Sadler, P.J. Proton NMR spectra of urine as indicators of renal damage. Mercury-induced nephrotoxicity in rats. Mol. Pharmacol. 27, (1985). 27. Nicholson, J.K. & Wilson, I.D. High resolution proton magnetic resonance spectroscopy of biological fluids. Prog. Nucl. Magn. Reson. Spectrosc. 21, (1989). 28. Nicholson, J.K., Connelly, J., Lindon, J.C. & Holmes, E. Metabonomics: a platform for studying drug toxicity and gene function. Nat. Rev. Drug Discov. 1, (2002). 29. Lindon, J.C., Nicholson, J.K. & Holmes, E. (eds.) The Handbook of Metabonomics and Metabolomics (Elsevier, Amsterdam, 2006). 30. Bollard, M.E., Stanley, E.G., Lindon, J.C., Nicholson, J.K. & Holmes, E. NMR-based metabonomics approaches for evaluating physiological influences on biofluid composition. NMR Biomed. 18, (2005). 31. Bales, J.R., Sadler, P.J., Nicholson, J.K. & Timbrell, J.A. Urinary excretion of acetaminophen and its metabolites as studied by proton NMR spectroscopy. Clin. Chem. 30, (1984). 32. Lenz, E.M. & Wilson, I.D. Analytical strategies in metabonomics. J. Proteome Res. 6, (2007). 33.van der Greef,J.et al. Evaluation of field desorption and fast atom-bombardment mass spectrometric profiles by pattern recognition techniques. Anal. Chim. Acta 150, (1983). 34. Wilson, I.D. et al. High resolution ultra performance liquid chromatography coupled to oa-tof mass spectrometry as a tool for differential metabolic pathway profiling in functional genomic studies. J. Proteome Res. 4, (2005). 35. Gartland, K.P. et al. Pattern recognition analysis of high resolution 1 HNMR spectra of urine. A nonlinear mapping approach to the classification of toxicological data. NMR Biomed. 3, (1990). 36. Gartland, K.P., Beddell, C.R., Lindon, J.C. & Nicholson, J.K. The application of pattern recognition methods to the analysis and classification of toxicological data derived from proton NMR spectroscopy of urine. Mol. Pharmacol. 39, (1991). 37. Ebbels, T. et al. Toxicity classification from metabonomic data using a density superposition approach: CLOUDS. Anal. Chim. Acta 490, (2003). 38. Antti, H. et al. Statistical experimental design and partial least squares regression analysis of biofluid metabonomic NMR and clinical chemistry data for screening of adverse drug effects. Chemom. Intell. Lab. Syst. 73, (2004). 39. Cloarec, O. et al. Statistical total correlation spectroscopy: an exploratory approach for latent biomarker identification from metabolic 1 H NMR data sets. Anal. Chem. 77, (2005). 40. Crockford, D.J. et al. Statistical HeterospectroscopY (SHY), a new approach to the integrated analysis of NMR and UPLC-MS datasets: application in metabonomic toxicology studies. Anal. Chem. 78, (2006). 41. Lindon, J.C. et al. Contemporary issues in toxicology the role of metabonomics in toxicology and its evaluation by the COMET project. Toxicol. Appl. Pharmacol. 187, (2003). 42. Lindon, J.C. et al. The Consortium for Metabonomic Toxicology (COMET): aims, activities and achievements. Pharmacogenomics 6, (2005). 43. Ebbels, T.M.D. et al. Prediction and classification of drug toxicity using probabilistic modeling of temporal metabolic data: The Consortium of Metabonomic Toxicology Screening approach. J. Proteome Res. 6, (2007). 44. Lindon, J.C. & Holmes, E. A survey of metabonomics approaches for disease characterization. In Handbook of Metabonomics and Metabolomics (eds. Lindon, J.C., Nicholson, J.K. & Holmes, E.) (Elsevier, Amsterdam, 2006). 45. Moolenaar, S.H., Engelke, U.F. & Wevers, R.A. Proton nuclear magnetic resonance spectroscopy of body fluids in the field of inborn errors of metabolism. Ann. Clin. Biochem. 40, (2003). 46. Brindle, J.T. et al. Rapid and noninvasive diagnosis of the presence and severity of coronary heart disease using 1 H NMR-based metabonomics. Nat. Med. 8, (2002). 47. Nicholls, A.W., Mortishire-Smith, R.J. & Nicholson, J.K. NMR spectroscopic based metabonomic studies of urinary metabolite variation in acclimatizing germ-free rats. Chem. Res. Toxicol. 16, (2003). 48. Nicholson, J.K., Holmes, E. & Wilson, I.D. Gut microorganisms, mammalian metabolism and personalized health care. Nat. Rev. Microbiol. 3, (2005). 49. Clayton, T.A. et al. Pharmaco-metabonomic phenotyping and personalised drug treatment. Nature 440, (2006). 50. Dumas, M.E. et al. Assessment of analytical reproducibilty of 1 H nmr spectroscopy based metabonomics for large-scale epidemiological research: the INTERMAP study. Anal. Chem. 78, (2006). 51. Liu, M., Nicholson, J.K. & Lindon, J.C. High resolution diffusion and relaxation edited one- and two-dimensional 1 H NMR spectroscopy of biological fluids. Anal. Chem. 68, (1996). 52. Jenkins, H. et al. A proposed framework for the description of plant metabolomics experiments and their results. Nat. Biotechnol. 22, (2004). 53. Bino, R.J. et al. Potential of metabolomics as a functional genomics tool. Trends Plant Sci. 9, (2004). 54. Lindon, J.C. et al. Summary recommendations for standardization and reporting of metabolic analyses. Nat. Biotechnol. 23, (2005). 55. Le Belle, J.E., Harris, N.G., Williams, S.R. & Bhakoo, K.K. A comparison of cell and tissue extraction techniques using high-resolution 1 H-NMR spectroscopy. NMR Biomed. 15, (2002). 56. Ching, Y.L., Wu, H., Tjeerdema, R.S. & Viant, M.R. Evaluation of metabolite extraction strategies from tissue samples using NMR metabolomics. Metabolomics 3, (2007). 57. Villas-Boas, S.G., Hojer-Pedersen, J., Akesson, M., Smedsgaard, J. & Nielsen, J. Global metabolite analysis of yeast: evaluation of sample preparation methods. Yeast 22, (2005). 58. Ippolito, J.E. et al. An integrated functional genomics and metabolomics approach for defining poor prognosis in human neuroendocrine cancers. Proc. Natl. Acad. Sci. USA 102, (2005). 59. Spraul, M. et al. Flow injection proton nuclear magnetic resonance spectroscopy combined with pattern recognition methods: implications for rapid structural studies and high throughput biochemical screening. Anal. Comm. 34, (1997). 60. Akoka, S., Barantin, L. & Trierweiler, M. Concentration measurement by proton NMR using the ERETIC method. Anal. Chem. 71, (1999). 61. Aue, W.P., Karhan, J. & Ernst, R.R. Homonuclear broad band decoupling and twodimensional J-resolved NMR spectroscopy. J. Chem. Phys. 64, (1976). 62. Aue, W.P., Bartholdi, E. & Ernst, R.R. Two-dimensional spectroscopy. Application to nuclear magnetic resonance. J. Chem. Phys. 64, (1976). 63. Braunschweiler, L. & Ernst, R.R. Coherence transfer by isotropic mixing: application to proton correlation spectroscopy. J. Magn. Reson. 53, (1983). 64. Keun, H.C. et al. Cryogenic probe 13 C NMR spectroscopy of urine for metabonomic studies. Anal. Chem. 74, (2002). 65. Lindon, J.C., Nicholson, J.K. & Wilson, I.D. Directly-coupled HPLC-NMR and HPLC-NMR-MS in pharmaceutical research and development. J. Chromatogr. B 748, (2000) VOL.2 NO NATURE PROTOCOLS

12 PROTOCOL 2007 Nature Publishing Group natureprotocols 66. Cloarec, O. et al. Virtual chromatographic resolution enhancement in cryoflow LC-NMR experiments via statistical total correlation spectroscopy. Anal. Chem. 79, (2007). 67. Smith, L.M. et al. Statistical correlation and projection methods for improved information recovery from diffusion-edited NMR spectra of biological samples. Anal. Chem. 79, (2007). 68. Deprez, S., Sweatman, B.C., Connor, S.C., Haselden, J.N. & Waterfield, C.J. Optimisation of collection, storage and preparation of rat plasma for 1 HNMR spectroscopic analysis in toxicology studies to determine inherent variation in biochemical profiles. J. Pharm. Biomed. Anal. 30, (2002). 69. Maher, A.D., Zirah, S.F., Holmes, E. & Nicholson, J.K. Experimental and analytical variation in human urine in 1 H NMR spectroscopy-based metabolic phenotyping studies. Anal. Chem. 79, (2007). 70. Hwang, T.L. & Shaka, A.J. Water suppression that works: excitation sculpting using arbitrary wave-forms and pulsed-field gradients. J. Magn. Reson. A 112, (1995). 71. Aranibar, N., Ott, K.H., Roongta, V. & Mueller, L. Metabolomic analysis using optimised NMR and statistical methods. Anal. Biochem. 355, (2006). 72. Sonnewald, U., Isern, E., Gribbestad, I.S. & Unsgard, G. UDP-N-acetylhexosamines and hypotaurine in human glioblastoma, normal brain tissue and cell cultures: 1H/NMR spectroscopy study. Anticancer Res. 14, (1994). 73. Henke, J., Willker, W., Engelmann, J. & Leibfritz, D. Combined extraction techniques of tumour cells and lipid/phospholipid assignment by two dimensional NMR spectroscopy. Anticancer Res. 16, (1996). 74. Bligh, E.G. & Dyer, W.J. A rapid method of total lipid extraction and purification. Can. J. Biochem. Physiol. 37, (1959). 75. Tyagi, R.K., Azrad, A., Degani, H. & Salomon, Y. Simultaneous extraction of cellular lipids and water-soluble metabolites: evaluation by NMR spectroscopy. Magn. Reson. Med. 35, (1996). 76. Beckonert, O., Monnerjahn, J., Bonk, U. & Leibfritz, D. Visualizing metabolic changes in breast-cancer tissue using 1H-NMR spectroscopy and self-organizing maps. NMR Biomed. 16, 1 11 (2003). 77. Meiboom, S. & Gill, D. Modified spin-echo method for measuring nuclear relaxation time. Rev. Sci. Instrum. 20, (1958). 78. Wu, D., Chen, A. & Johnson, C.S. An improved diffusion-ordered spectroscopy experiment incorporating bipolar-gradient pulses. J. Magn. Reson. A 115, (1995). 79. Ernst, R.R. & Anderson, W.A. Application of Fourier transform spectroscopy to magnetic resonance. Rev. Sci. Instr. 37, (1966). 80. Jones, D.E. & Sternlicht, H. Fourier transform nuclear magnetic resonance I. Repetitive pulses. J. Magn. Reson. 6, (1972). NATURE PROTOCOLS VOL.2 NO

13 Anal. Chem. 2006, 78, Scaling and Normalization Effects in NMR Spectroscopic Metabonomic Data Sets Andrew Craig, Olivier Cloarec, Elaine Holmes, Jeremy K. Nicholson, and John C. Lindon* Biological Chemistry, Faculty of Natural Sciences, Imperial College London, Sir Alexander Fleming Building, South Kensington, London SW7 2AZ U.K. Considerable confusion appears to exist in the metabonomics literature as to the real need for, and the role of, preprocessing the acquired spectroscopic data. A number of studies have presented various data manipulation approaches, some suggesting an optimum method. In metabonomics, data are usually presented as a table where each row relates to a given sample or analytical experiment and each column corresponds to a single measurement in that experiment, typically individual spectral peak intensities or metabolite concentrations. Here we suggest definitions for and discuss the operations usually termed normalization (a table row operation) and scaling (a table column operation) and demonstrate their need in 1 H NMR spectroscopic data sets derived from urine. The problems associated with binned data (i.e., values integrated over discrete spectral regions) are also discussed, and the particular biological context problems of analytical data on urine are highlighted. It is shown that care must be exercised in calculation of correlation coefficients for data sets where normalization to a constant sum is used. Analogous considerations will be needed for other biofluids, other analytical approaches (e.g., HPLC- MS), and indeed for other omics techniques (i.e., transcriptomics or proteomics) and for integrated studies with fused data sets. It is concluded that data preprocessing is context dependent and there can be no single method for general use. The application of metabonomics has increased dramatically over the past few years, and many research groups are now attempting to process complex metabolic data sets. Metabonomics has been defined as the quantitative measurement of the multiparametric metabolic response of biological systems to pathology or genetic modification, 1 and the related subject of metabolomics has been described as a comprehensive analysis in which all the metabolites of a biological system are identified and quantified. 2 In these approaches, many samples from a biological origin (biofluids such as urine or plasma, tissue or plant extracts, in vitro culture supernatants, etc.) are analyzed using techniques that produce simultaneous detection and, in some cases, quantitation of many metabolites. The two main technologies that have been * To whom correspondence should be addressed. j.lindon@ imperial.ac.uk. Phone: Fax: (1) Nicholson, J. K.; Lindon, J. C.; Holmes, E. Xenobiotica 1999, 29, (2) Fiehn, O. Plant Mol. Biol. 2002, 48, used so far have been 1 H NMR spectroscopy 3 and mass spectrometry with a prior on-line separation step such as HPLC 4 or GC (in this case, the metabolites generally have to be chemically derivatized). 5 Studies to date have operated to a wide variety of experimental designs that are not always reported comprehensively. In addition, the data sets are usually preprocessed to make them amenable to multivariate statistical analysis. There have been a number of initiatives recently to build consensus on the standardization of metabolic experiments including the output from the SMRS group, 6,7 the ArMet scheme for reporting data and metadata, 8,9 and the output from the recent NIH-sponsored workshop. 10,11 Thus, considerable effort has been expended recently in order to define best practice in the field of metabonomics. These include comprehensive collection of background data to define the study samples (metadata), the techniques and their parameters used for data collection, and the chemometric and statistical procedures for data analysis. The steps involved in the analysis of metabonomics data have been well described 6 and typically involve at a minimum: (a) postinstrument processing of acquired spectroscopic data, such as polynomial baseline correction by removal of offsets, calculation of intensity values either on each data point, on each peak, or summed over segmented regions (binning); (b) production of a data table from the analytical measurements such that there are m rows (observations, samples) and n columns (variables, (3) Lindon, J. C.; Holmes, E.; Nicholson, J. K. Prog. NMR Spectrosc. 2001, 39, (4) Wilson, I. D.; Plumb, R.; Granger, J.; Major, H.; Williams, R.; Lenz, E. J. Chromatogr., B 2005, 817, (5) Szopa, J.; Wilczynski, G.; Fiehn, O. Phytochemistry 2001, 58, (6) Lindon, J. C.; Nicholson, J. K.; Holmes, E.; Keun, H. C.; Craig, A.; Pearce, J. T.; Bruce, S. J.; Hardy, N.; Sansone, S. A.; Antti, H.; Jonsson, P.; Daykin, C.; Navarange, M.; Beger, R. D.; Verheij, E. R.; Amberg, A.; Baunsgaard, D.; Cantor, G. H.; Lehmann-Keeman, L.; Earll, M.; Wold, S.; Johansson, E.; Haselden, J. N.; Kramer, K.; Thomas, C.; Lindberg, J.; Schuppe-Koistinen, I.; Wilson, I. D.; Reily, M. D.; Robertson, D. G.; Senn, H.; Krotzky, A.; Kochhar, S.; Powell, J.; van der Ouderaa, F.; Plumb, R.; Schaefer, H.; Spraul, M. Nat. Biotechnol. 2005, 23, (7) (8) Jenkins, H.; Hardy, N.; Beckmann, M.; Draper, J.; Smith, A. R.; Taylor, J.; Fiehn, O.; Goodacre, R.; Bino, R. J.; Hall, R.; Kopka, J.; Lane, G. A.; Lange, B. M.; Liu, J. R.; Mendes, P.; Nikolau, B. J.; Oliver, S. G.; Paton, N. W.; Rhee, S.; Roessner-Tunali, U.; Saito, K.; Smedsgaard, J.; Sumner, L. W.; Wang, T.; Walsh, S.; Wurtele, E. S.; Kell, K. B. Nat. Biotechnol. 2004, 22, (9) (10) Castle, A. C.; Fiehn, O.; Kaddurah-Daouk, R.; Lindon, J. C. Briefings Bioinformatics. In press. (11) Analytical Chemistry, Vol. 78, No. 7, April 1, /ac CCC: $ American Chemical Society Published on Web 02/18/2006

14 frequencies, integrals); (c) normalization of the data or some related adjustment to the spectral intensities (a row operation); (d) scaling of the data (a column operation); (e) multivariate statistical modeling of the data. However, one major area where considerable interlaboratory variation exists is in the operating procedures used in stages c and d above to preprocess the spectroscopic data for chemometric analyses. One approach that has shown promise has been to adopt a pragmatic methodology for large-scale human studies. 12 An attempt has been made recently to define an optimum approach for NMR spectra of urine, 13 but the data analysis and results were not conclusive, and we show here that data preprocessing has to be context dependent. One particular approach that has been widely used is binning of spectra to produce a reduced set of parameters. This usually involves integration of peak values within defined specific spectral ranges, and the earlier attempt at defining an optimum approach used such binned data. 13 The data preprocessing needs were also addressed by the SMRS group 6 and the NIH workshop, 10 but in the light of conflicting practices, it is thought timely to summarize best practice in this area and show that a single approach will not be optimum in all cases even for a given spectroscopic technique such as 1 H NMR spectroscopy. Thus, data for each study will have to be processed in an appropriate manner according to the study and type of sample. Here we use 1 H NMR spectroscopy as a vehicle for the discussions, but similar considerations will also apply for HPLC-MS metabonomic studies and in research involving other multivariate omics data, such as from transcriptomics and proteomics. The problems will be particularly acute for fused data sets such as combined NMR and MS data 14 or for data from combinations of different MS experiments. 15 Given considerable confusion in the literature, we suggest here a standardized definition for, and explanation of several widely used and generally accepted practices. Binning. Historically, pattern recognition of NMR-based metabolite data was performed using either quantitative or scored integrals of specific spectral peaks. 16,17 This approach does not work well in crowded regions of spectra with substantial peak overlap and is not easily automated for application to large sample sets. Calculating the peak areas within specified segments of a spectrum (binning or bucketing) was introduced originally to allow comparison of NMR data measured at different magnetic field strengths by minimizing but not eliminating the effects of secondorder spectral differences. 18 In NMR-based metabonomic studies, it also reduces the effect of ph-induced changes in chemical shift, (12) Bijlsma, S.; Bobeldijk, I.; Verheij, E. R.; Ramaker, R.; Kochhar, S.; Macdonald, I. A.; van Ommen, B.; Smilde, A. K. Anal. Chem. 2006, 78, (13) Webb-Robertson, B. J.; Lowry, D. F.; Jarman, K. H.; Harbo, S. J.; Meng, Q. R.; Fuciarelli, A. F.; Pounds, J. G.; Lee, K. M. J. Pharm. Biomed. Anal , (14) Crockford, D. J.; Holmes, E.; Lindon, J. C.; Plumb, R. S.; Zirah, S.; Bruce, S. J.; Rainville, P.; Stumpf, C. L.; Nicholson, J. K. Anal. Chem. 2006, 78, (15) Smilde, A. K.; van der Werf, M. J.; Bijlsma, S.; van der Werff-van der Vat, B. J. C.; Jellema, R. H. Anal. Chem. 2005, 77, (16) Gartland, K. P.; Sanins, S. M.; Nicholson, J. K.; Sweatman, B. C.; Beddell, C. R.; Lindon, J. C. NMR Biomed , (17) Gartland, K. P.; Beddell, C. R.; Lindon, J. C.; Nicholson, J. K. Mol. Pharmacol. 1991, 39, (18) Spraul, M.; Neidig, P.; Klauck, U.; Kessler, P.; Holmes, E.; Nicholson, J. K.; Sweatman, B. C.; Salman, S. R.; Farrant, R. D.; Rahr, E.; et al. J. Pharm. Biomed. Anal. 1994, 12, ensuring that the same species is always counted correctly across samples with such variation. Binning offers a rapid and consistent method by which to generate data sets automatically for modeling purposes, and a determination of the effects of changing the bin width and other variables for 1 H NMR spectra of biofluids such as urine and blood plasma or serum had been made previously. 18,19 A typical bin width of 0.04 ppm is frequently used for 1 H NMR spectra of urine as it is a good compromise between resolution and the difficulties related to positional variation in the position of some analyte species NMR peaks, most notably those of citrate, which are a function of ph and ionic strength of the sample. It also encompasses the typical width of an NMR resonance taking into account spin-spin splittings and line widths. By no means, is this bin width the only one that should be used. Binned data should only be used for development of chemometric classification models, and it is necessary to examine and analyze the full resolution spectra for biomarker identification. Since the use of binned data can lead to inaccuracies in peak intensities (e.g., by inclusion of variable amounts of baseline offset) and because of the removal of a number of computational limitations on data matrix sizes, more attention has recently been paid to methods that forego the need for binning of metabonomics data. These have involved methods of analyzing data in full resolution 20 both with and without additional preprocessing methods such as peak alignment Full spectral resolution data analysis is a significant advance allowing direct interpretation of any chemometric-derived model, since at full resolution these have a high similarity in appearance to real spectra, and this aids interpretability. Normalization. This is a row operation that is applied to the data from each sample and comprises methods to make the data from all samples directly comparable with each other. A common use is to remove or minimize the effects of variable dilution of the samples. This is not usually a problem for samples from a biofluid such as plasma, where metabolite concentrations are highly regulated by homeostasis and changes observed in pathological situations are small but significant, or for studies on tissue extracts, where a constant or known weight of tissue can be analyzed. In such cases, it is desirable that the data should reflect directly the concentrations or relative concentrations of metabolites. In fused MS data sets 15 or in NMR spectra where each substance can give several peaks, it is conceivable that the same metabolite will be measured more than once and hence there is no a priori reason these data should be different. A major effect can be observed in NMR spectroscopic studies of urine where many drugs, toxins, and treatments can cause both large increases and decreases in urinary volume and hence in urinary concentration. Hence, for studies on urine, a good working definition requires an awareness of the biological context. Henceforth in this section, we confine ourselves to the discussion of (19) Holmes, E.; Foxall, P. J.; Nicholson, J. K.; Neild, G. H.; Brown, S. M.; Beddell, C. R.; Sweatman, B. C.; Rahr, E.; Lindon, J. C.; Spraul, M.; et al. Anal. Biochem. 1994, 220, (20) Cloarec, O.; Dumas, M. E.; Trygg, J.; Craig, A.; Barton, R. H.; Lindon, J. C.; Nicholson, J. K.; Holmes, E. Anal. Chem. 2005, 77, (21) Stoyanova, R.; Nicholson, J. K.; Lindon, J. C.; Brown, T. R. Anal. Chem. 2004, 76, (22) Stoyanova, R.; Nicholls, A. W.; Nicholson, J. K.; Lindon, J. C.; Brown, T. R. J. Magn. Reson. 2004, 170, (23) Forshed, J.; Torgrip, R. J.; Aberg, K. M.; Karlberg, B.; Lindberg, J.; Jacobsson, S. P. J. Pharm. Biomed. Anal. 2005, 38, Analytical Chemistry, Vol. 78, No. 7, April 1,

15 normalization of data derived from NMR-based metabonomic analysis of urine. The variable nature of urinary volume is often overlooked in metabonomic studies of urine. 13 Renal excretion and urine production serves two purposes, to eliminate waste organic and inorganic species and also to help regulate blood volume. This can be conceived as being achieved through independent control of both solute and solvent excretion rates. The need for normalization in NMR-based metabonomics studies of urine arises primarily from this biological imperative. To correctly obtain true values for the excretion of urinary metabolites, it is necessary to measure the absolute values in molar concentration units per unit time per unit of body weight (e.g., millimoles, per hour, per kilogram of body weight). In animal studies, this of course requires measurement of urinary volumes and animal weights as a study proceeds. For clinical studies, the same criteria also apply. In practice, this method is almost never fully employed. However, one procedure that can help to experimentally normalize such NMR data is to note the volume of urine from which the sample came, freezedry the urine, and then reconstitute it in an appropriate volume of water or buffer to account for any urinary volume differences for a given collection period. 24 Of course, depending on the context of the problem, in many studies, it is perfectly possible to use relative concentrations of metabolites and changes of these as biomarkers of a biological effect. For studies of excretion balance or for flux calculations, often absolute metabolite concentrations and amounts are required, but for investigations of the differential effects of drugs or toxins, for example, it might be that relative effects are sufficient. Normalization to a housekeeping metabolite has also been attempted, typically using a creatinine peak area as a reference. 25 Creatinine clearance has long been considered a constant and used to assess renal function though this may not always be the case as, in some cases, creatinine clearance may be more related to muscle mass particularly in children 25 and the elderly. 26 This results in a metabolite excretion rate relative to that of creatinine. Alternatively, in some cases, the concentration of a specific metabolite can be determined by an independent means (e.g., glucose using conventional clinical chemistry) and this then provides a reference value. 27 It is more common, however, for other preprocessing methods to be used before chemometric modeling, where the biological meaning of the procedure is not directly apparent. One common method of normalization involves setting each observation (spectrum) to have unit total intensity by expressing each data point as a fraction of the total spectral integral. We refer to this method as normalization to a constant sum (CS). It might be useful to consider the volume of urine produced by an individual in a given time (V total ) as being composed of two parts. The first is the volume of water required to dissolve all excreted solutes (V solute ). The second is the volume (V balance ), which could be positive or negative, that serves to balance fluid intake (24) Gartland, K. P.; Bonner, F. W.; Nicholson, J. K. Mol. Pharmacol. 1989, 35, (25) Prevot, A.; Martini, S.; Guignard, J. P. Rev. Med. Suisse Romande 2002, 122, (26) Burkhardt, H.; Bojarsky; Gretz, N.; Gladisch, R. Gerontology 2002, 48, (27) Anthony, M. L.; Sweatman, B. C.; Beddell, C. R.; Lindon, J. C.; Nicholson, J. K. Mol. Pharmacol. 1994, 46, in order to maintain blood volume. In principle, it would be best to compare the metabolite concentrations in V solute between subjects, but this is not possible since only V total is known. The value of V balance will vary between individuals and will be affected by the particular study in progress. For a series of spectra with highly similar internal peak ratios but differing in total intensity because of such dilution or concentration effects, CS normalization of each spectrum can be considered to approximate the relative concentration of species (i.e., as in V solute ). Importantly, this approximation will break down when large perturbations occur to intensities in some spectra (e.g., those from certain toxin-treated animals or the use of diuretic drugs, for example). This is easily seen because if some peak areas are increased and the total is normalized to a constant, others will appear to have decreased, and this effect in closed data sets has been noted previously. 28 This can cause a major difficulty in the interpretation of loadings from chemometric analyses such as principal components analysis (PCA) or PLS when such CS normalization has been used. Scaling. This operation is performed on the columns of data (i.e., on each spectral intensity across all samples). A number of such scaling methods are commonly used. Each column of the table can be given a mean of zero by subtracting the column mean from each value in the column (mean-centering). This is typically done so that all the components found by PCA have as their origin the centroid of the data, resulting in a parsimonious model. Second, each column of the table can be scaled so that it has unit variance, by dividing each value in the column by the standard deviation of the column. If the data are mean centered, the weighting reflects the covariance of the variables, while in unit variance scaling, the weighting reflects their correlation. Other forms of scaling are possible 29 and are used, such as Pareto scaling, where each variable is divided by the square root of the standard deviation of the column values, 30 or logarithmic scaling, when values relating to order of magnitude scores are desired. 27 Furthermore, it should be clear that as normalization and scaling operations serve different purposes it is possible (and is, in fact, the usual practice) to use a combination of normalization and scaling methods. RESULTS To demonstrate the consequences of normalization to a constant sum and column scaling, a set of simulated NMR spectra were constructed and visualized using Matlab routines and analyzed using PCA. In this example, each of 40 samples contains two components, each of which results in a single Lorentzianshaped peak at NMR chemical shifts of 3 and 7 ppm, with a full width at half-height of 0.1 ppm set in a spectral range of 1-10 ppm, digitized in 0.01 ppm steps (Figure 1A). The samples are divided into two classes (such as control and diseased) with peak 1 at 3 ppm with a variance of 10 and a mean value of 100 in class 1 and a mean value of 120 in class 2. Peak 2 at 7 ppm is not class (28) Wold, S.; Johansson, E.; Cocchi, M. In 3D QSAR in Drug Design: Theory, Methods and Applications; Kubinyi, H., Folkers, G., Martin, Y. C., Eds.; Perspectives in Drug Discovery and Design; ESCOM Science: Leiden, (29) Holmes, E.; Nicholls, A. W.; Lindon, J. C.; Ramos, S.; Spraul, M.; Neidig, P.; Connor, S. C.; Connelly, J.; Damment, S. J.; Haselden, J.; Nicholson, J. K. NMR Biomed. 1998, 11, (30) Eriksson, L.; Johansson, E.; Kettaneh-Wold, N.; Wold, S. Multi- and Megavariate Data Analysis. Principles and Applications; Umetrics AB: Umeå, Sweden, Analytical Chemistry, Vol. 78, No. 7, April 1, 2006

16 Figure 1. Simulated spectrum comprising two Lorentzian peaks. (A) The two peaks before the additional of noise, (B) superposition of all 40 spectra, (C) superposition of all 40 spectra, with random dilutions, (D) same as (C) but all spectra normalized to a constant sum, (E) the meancentered true data set, (F) the mean-centered normalized data set, (G) the mean-centered and unit variance scaled true data set, and (H) the mean-centered and unit variance scaled normalized data set. discriminating and has an intensity that is normally distributed across all observations with a mean of 100 and a variance of 50. The true data matrix, X (Figure 1B), was constructed as X ) C P + E, where C is a concentration matrix, P is a matrix of pure compound spectra, and E is a matrix of white noise with values between (1 representing thermal noise from the spectrometer. A diluted version of the data matrix X (Figure 1C) was also produced where X d ) D C P + E, and D is a diagonal matrix containing random dilution factors between 0.5 and 1.0 to represent a variation in urinary volumes that is independent of any class differences. The diluted data set was normalized using the CS method, and the result is shown in Figure 1D. Because Analytical Chemistry, Vol. 78, No. 7, April 1,

17 Figure 2. Results of PCA on the simulated spectral data sets. (A) Scores plot PC1 vs PC2 for the mean-centered true data, (B) scores plot PC1 vs PC2 for the mean-centered normalized data, (C) PC1 vs PC2 loadings plot for the mean-centered true data, (D) PC1 vs PC2 loadings plot for the mean-centered normalized data, (E) PC1 vs PC2 loadings plot for the mean-centered and unit variance scaled true data, (F) PC1 vs PC2 loadings plot for the mean-centered and unit variance scaled normalized data, (G) STOCSY correlation plot for the true data, and (H) STOCSY correlation plot for the normalized data. the instrumental noise has a constant average value, when very dilute spectra are normalized to a constant sum, the apparent signal-to-noise ratio is decreased. The effect of mean-centering the variables is shown on the true data (Figure 1E) and on the normalized data (Figure 1F). The effect of scaling each variable to have unit variance is shown on both data sets (Figure 1G and H, respectively). Both data sets were then modeled using PCA. The expected result is a two-component model showing the two classes clearly separated in PC1 with the corresponding PC1 loadings identifying the pure compound spectrum (i.e., peak 1) of the species responsible for the difference between the two classes. Indeed, for the PC scores for mean-centered true data, as shown in Figure 2A, this is the result obtained. The unit variance scaled true data 2266 Analytical Chemistry, Vol. 78, No. 7, April 1, 2006

18 gave a model with a similar result (not shown). The corresponding loadings are shown for the mean-centered true data (Figure 2C) and for the unit variance scaled true data (Figure 2E). In both cases, the loadings for PC1 identify peak 1 as the feature that describes the variance causing the separation between the two classes and the PC2 loadings describe the variation in peak 2 intensity. Mean-centering results in a model with loadings that have a pseudospectral appearance retaining Lorentzian line shapes. The use of unit variance scaling gives equal weight to each data value, allowing systematic changes with small variance to be more readily detected. However, when the loadings are plotted as a pseudospectrum, the apparent signal-to-noise ratio is degraded and there is a tendency toward square-shaped peaks since intensity values within the same peak will be highly correlated and will thus receive similar weights in the PCA. In the case of NMR data, this can confound the useful information obtainable on peak height ratios and peak multiplicities. The recently reported back-scaling method for some chemometrics procedures (O-PLS) attempts to combine the advantages of both of these approaches. 20 The scores for the normalized data are given in Figure 2B, and again, the first PC largely discriminates between the two classes, but an outlier can be seen in the second PC. This is explained as arising from a very dilute sample with a low signalto-noise ratio, since PC2 is only modeling noise in the data. The corresponding loadings are shown for the mean-centered model (Figure 2D) and for the unit variance scaled model (Figure 2F). After normalization, in both cases, the loadings indicate that it is apparently a combination of an increase in peak1 intensity and a decrease in peak 2 intensity that is the discriminating feature. This is because the CS normalization has mixed the variation in one peak in with that from the other peak and changed the correlation structure of the data. Hence, the real discriminating factor in the data set has been obscured. This change in correlation structure is readily apparent when the STOCSY technique 31 is used, as shown in Figures 2G for the true data set and Figure 2H for the normalized set. STOCSY, as expected, does not show any crosspeak indicating covariance for the true data set, but a cross-peak is observed for the normalized data set. Artifactual cross-peaks that arise as a result of the choice of preprocessing method used (31) Cloarec, O.; Dumas, M. E.; Craig, A.; Barton, R. H.; Trygg, J.; Hudson, J.; Blancher, C.; Gauguier, D.; Lindon, J. C.; Holmes, E.; Nicholson, J. Anal. Chem. 2005, 77, (32) Tseng, G. C.; Oh, M. K.; Rohlin, L.; Liao, J. C.; Wong, W. H. Nucleic Acids Res. 2001, 29, (33) Craig, A. Parallel metabonomic and genomic characterisation of experimental hepatotoxicity in the rat. Ph.D. Thesis, University of London, 2004; p 293. may thus confound the use of STOCSY for structural interpretation, if used inappropriately. CONCLUSIONS We acknowledge that there is little exploration of normalization of intersample volume-related variation within the metabonomic and metabolomic literature. In many cases, exploratory data analysis is carried out by varying normalization and scaling procedures in order to obtain an optimum separation of two or more sample classes using chemometrics methods such as PCA, PLS, or neural networks. While this is useful for deriving classification models for predicting the class of subsequent samples, the interpretation of the biochemical factors responsible for the classification is not straightforward. Thus, we arrive at a more appropriate definition such that normalization of urinary metabolic data should be best considered as a data transformation which minimizes intersample variation due to differences in gross urinary concentration between samples caused by volume differences. The effect of CS normalization as demonstrated here primarily affects the interpretation of chemometric model coefficients. The CS normalization method has been used to good effect to produce models and expert systems with good predictive power. 29 Methods analogous to the invariant set normalization for microarray data 32 as well as scaling by the gradient of a robust regression to the median sample 33 might yield some success as alternative preprocessing methods in metabonomics. It should also be clear that any normalization should be a linear transformation. Finally, with the increased interest in relating data sets collected from different analytical platforms or across different levels of molecular biology, it is important to be aware of the effects of preprocessing on statistical outcome and to be mindful of the consequences of a chosen method of preprocessing and the limitations that this will place on the interpretation of any chemometric model. ACKNOWLEDGMENT We thank the Wellcome Trust for financial support (to O.C. and A.C.) from the project, Biological Atlas of Insulin Resistance ( Received for review October 28, Accepted January 13, AC Analytical Chemistry, Vol. 78, No. 7, April 1,

19 Anal. Chem. 2005, 77, Statistical Total Correlation Spectroscopy: An Exploratory Approach for Latent Biomarker Identification from Metabolic 1 H NMR Data Sets Olivier Cloarec, Marc-Emmanuel Dumas, Andrew Craig, Richard H. Barton, Johan Trygg, Jane Hudson, Christine Blancher, Dominique Gauguier, John C. Lindon, Elaine Holmes, and Jeremy Nicholson*, Biological Chemistry Section, Faculty of Medicine, Biomedical Sciences Division, Imperial College London, South Kensington, London, U.K., Research Group of Chemometrics, Institute of Chemistry, Umeå University, Umeå, Sweden, and The Wellcome Trust Centre for Human Genetics, University of Oxford, Oxford, U.K. Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x We describe here the implementation of the statistical total correlation spectroscopy (STOCSY) analysis method for aiding the identification of potential biomarker molecules in metabonomic studies based on NMR spectroscopic data. STOCSY takes advantage of the multicollinearity of the intensity variables in a set of spectra (in this case 1 H NMR spectra) to generate a pseudo-twodimensional NMR spectrum that displays the correlation among the intensities of the various peaks across the whole sample. This method is not limited to the usual connectivities that are deducible from more standard twodimensional NMR spectroscopic methods, such as TOC- SY. Moreover, two or more molecules involved in the same pathway can also present high intermolecular correlations because of biological covariance or can even be anticorrelated. This combination of STOCSY with supervised pattern recognition and particularly orthogonal projection on latent structure-discriminant analysis (O- PLS-DA) offers a new powerful framework for analysis of metabonomic data. In a first step O-PLS-DA extracts the part of NMR spectra related to discrimination. This information is then cross-combined with the STOCSY results to help identify the molecules responsible for the metabolic variation. To illustrate the applicability of the method, it has been applied to 1 H NMR spectra of urine from a metabonomic study of a model of insulin resistance based on the administration of a carbohydrate diet to three different mice strains (C57BL/6Oxjr, BALB/cOxjr, and 129S6/SvEvOxjr) in which a series of metabolites of biological importance can be conclusively assigned and identified by use of the STOCSY approach. In general, metabonomic and metabolomic studies are based on spectroscopic or spectrometric data of complex biosamples, mainly from 1 H NMR spectroscopy but more recently from liquid or gas chromatography with mass spectrometry. 1-4 Multivariate statistical and pattern recognition methods have been developed * Corresponding author. Tel.: j.nicholson@ imperial.ac.uk. Imperial College London. Umeå University. University of Oxford. to extract sample classification and associated biomarker information from NMR spectroscopic data because of the high complexity of biofluids containing potentially thousands of different metabolites. 5 A well-established way to analyze NMR spectral data has involved first a reduction of these data by integration of spectral sections into bins (frequency windows), which have generally corresponded to a typical spectral width of ppm. This serves to stabilize effects of the peak position variation due to physicochemical environment differences (ph, ionic concentration) and allow a smaller, more manageable, number of variables for statistical processing. 6 Analysis is then carried out using chemometric tools, such as principal components analysis (PCA) and projection to latent structures (PLS, also called partial least squares), to discover the variables describing the metabolic variation involved in the particular study and to allow categorization of the samples from the study. 7 Finally, the parts of the spectra corresponding to the most discriminatory variables are displayed in order to allow identification of the varying metabolites or biomarkers for a particular condition. 8 Recently, it has been shown that it is possible to use full spectral resolution, including all intensity values in a full spectrum, and that inclusion of variable peak position data (such as caused by ph differences between samples) can even be beneficial. 9 The interpretation of autoscaled chemometric models combining back- (1) Brindle, J. T.; Antti, H.; Holmes, E.; Tranter, G.; Nicholson, J. K.; Bethell, H. W.; Clarke, S.; Schofield, P. M.; McKilligin, E.; Mosedale, D. E.; Grainger, D. J. Nat. Med. 2002, 8, (2) Lindon, J. C.; Holmes, E.; Nicholson, J. K. Anal. Chem. 2003, 75, 384A- 391A. (3) Nicholson, J. K.; Lindon, J. C.; Holmes, E. Xenobiotica 1999, 29, (4) Plumb, R. S.; Stumpf, C. L.; Gorenstein, M. V.; Castro-Perez, J. M.; Dear, G. J.; Anthony, M.; Sweatman, B. C.; Connor, S. C.; Haselden, J. N. Rapid Commun. Mass Spectrom. 2002, 16, (5) Nicholson, J. K.; Foxall, P. J. D.; Spraul, M.; Farrant, R. D.; Lindon, J. C. Anal. Chem. 1995, 67, (6) Holmes, E.; Foxall, P. J. D.; Nicholson, J. K.; Neild, G. H.; Brown, S. M.; Beddell, C. R.; Sweatman, B. C.; Rahr, E.; Lindon, J. C.; Spraul, M.; Neidig, P. Anal. Biochem. 1994, 220, (7) Holmes, E.; Nicholls, A. W.; Lindon, J. C.; Connor, S. C.; Connelly, J. C.; Haselden, J. N.; Damment, S. J.; Spraul, M.; Neidig, P.; Nicholson, J. K. Chem. Res. Toxicol. 2000, 13, (8) Holmes, E.; Nicholson, J. K.; Nicholls, A. W.; Lindon, J. C.; Connor, S. C.; Polley, S.; Connelly, J. Chemom. Intell. Lab. Syst. 1998, 44, (9) Cloarec, O.; Dumas, M.-E.; Trygg, J.; Craig, A.; Barton, R. H.; Lindon, J. C.; Nicholson, J. K.; Holmes, E. Submitted for publication in Anal. Chem Analytical Chemistry, Vol. 77, No. 5, March 1, /ac048630x CCC: $ American Chemical Society Published on Web 01/25/2005

20 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x scaled PLS coefficient plots and variable weights demonstrated that this peak position variation can be handled successfully, and can in fact provide additional useful information on the physicochemical variations in metabonomic data sets. This method proved to be useful tool in identifying the 1 H NMR resonances corresponding to the most influential metabolites without the need to reconsult the initial spectra. However, when the number of different resonances is high, as in a biofluid such as urine, the identification of the molecules can be difficult. To address this difficulty in interpretation, we have developed and applied the concept and applications of statistical total correlation spectroscopy (STOCSY) for NMR spectra. This background of this method was introduced by Sasic et al. 10 and is based on a method proposed by Noda, for generalized two-dimensional correlation spectroscopy Successful applications of this correlative approach include infrared, Raman, near-infrared, and fluorescence spectroscopies, 15 where correlations between different spectral features could be identified. However, such correlation methods have not yet been applied to NMR spectroscopy of complex mixtures where the information density and resolution is much higher than that obtained from other spectroscopic techniques. This approach, if successful, would allow identification of highly correlated peak intensities that would lead directly to identification of peaks from the same molecule and hence aid molecule identification. Additionally, in principle, identification of lower or even negative correlations could lead to identification of substances in the same metabolic pathway whose concentrations are interdependent or under some common regulatory mechanism. Both types of information are important for biomarker analysis and identification. This paper describes a new framework for the data analysis of metabonomic data combining NMR spectroscopy and orthogonal projection on latent structure. 16 This combination allows the rapid visualization and identification of the molecules involved in the differentiation between metabolic states arising from strains of animal, toxicity, disease or therapeutic intervention (biomarkers), etc. To demonstrate the potential of this new framework, a metabonomic study taken from the Biological Atlas of Insulin Resistance (BAIR, Wellcome Trust Grant ) project comparing the metabolism of three mouse strains (BALB/c, C57BL/6, and 129S6) normal diet is presented. EXPERIMENTAL SECTION Animal and Samples. A panel of three inbred adult male mouse strains, namely C57BL/6Oxjr, BALB/cOxjr, and 129S6/ SvEvOxjr, was used in this study. All mice had free access to water and standard laboratory chow pellets (ERB, Whitam, U.K.) and were maintained on a 12-h light/dark cycle. Experiments were conducted under a U.K. Home Office License approval and according to the rules of animal use in scientific experiments in the U.K. Urine samples were collected from mice maintained for 12 h (8 p.m.-8 a.m.) in metabolic cages. Samples were collected into vials containing a 1% sodium azide solution to minimize (10) Sasic, S.; Muzynski, A.; Ozaki, Y. J. Phys. Chem. A 2000, 104, (11) Noda, I.; Dowrey, A. E.; Marcott, C.; Story, G. M.; Ozaki, Y. Appl. Spectrosc. 2000, 54, 236A-248A. (12) Noda, I. J. Am. Chem. Soc. 1989, 111, (13) Noda, I. Appl. Spectrosc. 1993, 47, (14) Noda, I. Appl. Spectrosc. 1990, 44, (15) Osaki, Y.; Wang, Y. J. Near Infrared Spectrosc. 1998, 19, (16) Trygg, J.; Wold, S. J. Chemom. 2002, 16, microbiological contamination, centrifuged for solid particle removal, and stored at -80 C until NMR acquisition. NMR Spectroscopy. An aliquot (200 µl) of each urine sample was added to 200 µl of 0.2 M sodium phosphate buffer (ph 7.4) containing 1 mm TSP (sodium trimethylsilyl [2,3,3,3-2 H 4 ]propionate) and 20% D 2 O as a chemical shift reference standard and lock signal, respectively, and 200 µl of water (MilliQ quality). All samples were centrifuged at 3000 rpm for 10 min to remove any solid debris. 1 H NMR spectra were measured at 600 MHz and 300 K using a flow injection system (Bruker Biospin, Karlsruhe, Germany). The water resonance was suppressed by using a 90-3 µs ms-90 pulse sequence with irradiation during a 2-s relaxation delay and also during the 100-ms mixing time. For each sample 64 transients were collected into data points using a spectral width of ppm. The total acquisition time was around 4 min per sample. Prior to Fourier transformation, an exponential line-broadening factor of 1 Hz was applied to each free induction decay. A spin-lock of 100 µs was used for the TOCSY experiment (total correlation spectroscopy). 18 The spectra were phased, baseline-corrected, and referenced to TSP (δ 0.0) automatically using an in-house routine written in MATLAB (Mathworks, Natick, MA). 17 The regions δ ppm and δ ppm were removed to eliminate baseline effects of imperfect water saturation and the nonquantitative contribution of urea, respectively. Statistical Total Correlation Spectroscopy. Statistical total correlation spectroscopy (STOCSY) is based on the properties of the correlation matrix C, computed from a set of sample spectra according to C ) 1 n - 1 X t 1 X 2 where X 1 and X 2 denote the autoscaled experimental matrices of n v 1 and n v 2, respectively; n is the number of spectra (one for each sample) and v 1 and v 2 are the number of variables in the spectra for each matrix. C is therefore a matrix of v 1 v 2, where each value is a correlation coefficient between two variables of the matrices X 1 and X 2. The simplest case is the autocorrelation analysis where X 1 ) X 2. Because the different resonance intensities from a single molecule will always have the same ratio, if the spectrometer conditions are kept identical between samples the relative intensities will be theoretically totally correlated (correlation coefficient r ) 1). In real samples of biofluids, r will be always inferior to 1 because of spectral noise or peak overlaps from other molecules. However, in practice, the correlation matrix from a set of spectra containing different amounts of the same molecule shows very high correlations between the variables corresponding to the resonances of the same molecule. Plotting the correlation matrix provides a graphic representation of the multisample spectroscopic data set comparable to that of a two-dimensional (2D) correlation NMR experiment conducted on one sample containing all the molecules of all the samples. The closest NMR experiment to STOCSY is TOCSY (total correlation spectroscopy), the signals of which arise from protons within a spin system. 18 In principle, concentrations of other molecules can also be correlated to the (17) Ebbels, T. M. D.; Lindon, J. C.; Nicholson, J. K.; Holmes, E. US ; US, (18) Braunschweiler, L.; Ernst, R. R. J. Magn. Reson. 1983, 53, Analytical Chemistry, Vol. 77, No. 5, March 1,

21 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x (19) Beebe, K. B.; Pell, R. J.; Seasholtz, M. B. ChemometricssA practical guide; John Wiley & Sons: New York, (20) Trygg, J. J. Chemom. 2002, 16, (21) Trygg, J.; Wold, S. J. Chemom. 2003, 17, Analytical Chemistry, Vol. 77, No. 5, March 1, 2005 initial molecule of interest, and quantitative relationships between molecules can therefore be highlighted. For example, molecules in the same biochemical pathway may exhibit a similar or even codependent response to a stimulus. In this case the correlation between resonances from different molecules would be high but not usually as strong as for resonances on the same molecule. The method is not restricted to the 1 H- 1 H correlation but, in principle, can be applied to different nuclei. If these involve different NMR-active nuclei ( 13 C- 13 C, 1 H- 13 C, 13 C- 31 P, etc.), X 1 * X 2, then heteronuclear correlation is also possible, yielding novel molecular connectivity information using both types of nuclear spin properties. Finally, it should be noted that STOCSY can be used to derive NMR spectral splittings and J couplings with the same theoretical precision of the one-dimensional (1D) spectral properties from which the 2D data set was derived and it is not limited by low resolution in the F 1 domain of most correlation 2D experiments, which are typically much lower than the standard 1D spectrum. This is, of course, provided that any physicochemical environment variation between samples does not induce variation of the peak positions. Pattern Recognition. Data analysis was carried out in two steps with all the variables mean centered and autoscaled by dividing each variable by its standard deviation. In the first stage a principal components analysis (PCA) was conducted in order to select out the distinct outliers by comparing the spectral residuals from sample to sample using the method of the F calc plot. 19 The selected spectra are then checked and rejected only if they showed inconsistency related to a baseline problem, bad phasing, or a very low signal-to-noise ratio. The second stage is a supervised pattern recognition method called orthogonal projection on latent structure (O-PLS), which was developed by Trygg et al. 16,20,21 An O-PLS model can be seen as a factor analysis model, where the variation in the matrix X (the NMR spectra) and the matrix Y (the descriptive variables) is separated into three parts. The first part contains the variation common in X and Y, the second one contains the specific variation for X, so- called structured noise, and the last one contains the residual variance. The O-PLS method provides a prediction similar to that of PLS (projection on latent structure). However, the interpretation of the models is improved because the structured noise is modeled separately from the variation common in X and Y. Therefore, the O-PLS loading and regression coefficients provide more straightforward and accurate interpretation than PLS, which models the structured noise together with the correlated variation between X and Y. Furthermore, the orthogonal loading matrices provide the opportunity to interpret the structured noise. To test the validity of the model against overfitting, the cross-validation parameter Q 2 was computed. 21 For our applications, each line of the X matrix is an NMR spectrum corresponding to one sample and each column of Y defines a class (or group) whose values are dummy variables as used in discriminant analysis. The method can therefore be defined as O-PLS-DA. To improve the interpretability of the O-PLS model, the method described by Cloarec et al. has been applied. 9 It consists of combining the back-scaled O-PLS-DA coefficients from an autoscaled model with the variable weight of the same model in the same plot. For this purpose, each O-PLS coefficient is first multiplied by the standard deviation of its corresponding variable and then plotted as a function of its related chemical shift but with a color code linked to the weight of the variable in the model, highlighting in this way the resonances of the most important metabolites involved in the discrimination among the different groups (classes). This tool can also be applied to direct structural identification of biomarkers. Furthermore, the result of STOCSY can also been combined with pattern recognition results in one plot. From the O-PLS coefficients, it is possible to select one significant variable and to replot the coefficients as previously, but this time with a color code corresponding to the correlation between the selected variable and other variables, revealing in this way, and according to the level of the correlation, the structural or physiological relationship existing between different resonances. In this way the discriminant resonances between the groups can be highlighted in a first step by the O-PLS-DA, and therefore, due to the intrinsic correlation between resonances from the same molecule, they can be separated to provide easier identification of the discriminant compounds. Computer and Software. NMR processing and pattern recognition were carried out using a Power Mac G5 with dual 64-bit 2-GHz processors and 2 GB of synchronous dynamic random access memory (SDRAM). NMR processing and pattern recognition routines were written in-house in the MATLAB 6.5 environment (Mathworks, Natick, MA). RESULTS 1 H NMR Spectra. To validate the robustness of the new analytical approach and demonstrate its applicability in a functional genomic context to the definition of strain-specific metabolic phenotype characteristics in mice, we acquired and processed a total of H NMR spectra of urine samples corresponding to the different mouse strains (216 from BALB/c, 263 from C57BL/ 6, and 133 from 129S6 strains). The principal components analysis (PCA) of the 1 H NMR spectra data set in combination with the F calc plot allowed us to highlight 13 outliers in the set of H NMR spectra. 19 For all of these spectra, the reason for their isolation from the main body of samples was due to either bad water resonance suppression or a very dilute sample, providing a very low signal-to-noise ratio. These outliers were, therefore, removed from the data set for the rest of the study. In contrast with many previous 1 H NMR based metabonomic studies using reduced data, 2 the method proposed here used the full resolution of the 600-MHz 1 H NMR spectra to extract the biological information related to the differences in the metabolism of different mouse strains. Examples of 1 H NMR spectra of urine corresponding to the three different mouse strains are presented in Figure 1. Many metabolites can already be recognized in these spectra, and a difference in patterns among these three spectra can readily be observed. However, this may be due to variation not related to the strain discrimination (variance between groups) but to variation within the strain (variance within groups), and obviously the comparison of three samples only cannot be conclusive. Statistical Total Correlation Spectroscopy. All 612 rat urine NMR spectra from all the mouse strain samples were used for the computation of the correlation matrix. The result is shown in Figure 2 as a contour plot indicating the highest correlation and

22 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x Figure 1. Example of partial 600-MHz 1 H NMR spectra for three different mouse strains showing some assignments of the most abundant metabolites: (A) BALB/c; (B) C57BL/6; (C) 129S6. Figure 2. STOCSY two-dimensional representation of 1 H NMR spectra for three different mouse strains (BALB/c, C57BL/6, and showing δ/δ 1 H connectivities): (1) water signal suppression imperfection; (2) protein; (3) valeramide; (4) glucose; (5) hippurate; (6) 2-oxoglutarate; (7) 3-hydroxyphenylpropionate. can be interpreted as two-dimensional NMR maps. For instance, in the region between δ 1.5 ppm and δ 4.2 ppm, it is possible to recognize the spin structure of valeramide and glucose (for these compounds and all other direct assignments in this paper, the chemical shift and the multiplicity of molecules have been compared with 1 H NMR spectra of water solution of the pure compound). Just as with 2D correlation spectroscopy plots, each peak in the NMR data set will appear on the diagonal of the correlation matrix. Each data point has an autocorrelation value of 1 and a very high correlation with the other data points from the same peak; for this reason the peaks on the diagonal are visible and other peaks with data points having a significant correlation with the diagonal peak will appear at the appropriate chemical shift, that is, off the diagonal. Simply reading the two chemical shifts of an off-diagonal peak then allows the determination of the chemical shifts of the two correlated peaks. If more than two such peaks are intercorrelated, then it is possible to pick out a network of correlated peak intensities. Several other correlated peaks are also present on this large-scale representation (Figure 2), and an expansion to give the small region shown in Figure 3 reveals the structure of the glycerate ABX system with a high digital resolution such that it is possible to measure the different coupling constants ( 2 J AB ) 11.4 Hz, 3 J AX ) 5.7 Hz, and 3 J BX ) 3.0 Hz). Figure 3. STOCSY two-dimensional representation of 1 H NMR spectra for three different mouse strains (BALB/c, C57BL/6, and 129S6) for the glycerate spin system ABX δh X, δh B, and δh A indicate chemical shifts of three protons and J AB, J AX, and J BX indicate the 1 H- 1 H coupling patterns. Analytical Chemistry, Vol. 77, No. 5, March 1,

23 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x Figure 4. (A) Localized two-dimensional representation of STOCSY of 1 H NMR spectra for three different mouse strains (BALB/c, C57BL/6, and 129S6). (B) Example of TOCSY NMR spectrum obtained from a single urine sample of a 129S6 strain mice for the same chemical shift region. Keys: (1) 2-oxoglutarate; (2) citrate; (3) 3-hydroxyphenylpropionate; (4) methylamine and dimethylamine correlation; (5) dimethylamine and trimethylamine correlation. However, the spin system of a molecule can overlap with other spin systems, and this reduces the correlation that exists for the resonances of both molecules. For this reason, only three out of four peaks of δh A are present in the shown figure. In the region between δ 2.2 and 3.2 ppm, many resonances can be assigned easily with the correlation method (Figure 4). The two correlated triplets at δ 2.45 and 3.01 ppm with the coupling constant 3 J AX ) 7.2 Hz are attributed to 2-oxoglutarate. The AB spin system of citrate can also be recognized, but the correlation is weakened by the peak position variation due to physicochemical differences (ph and metal ion concentration) across samples, which particularly affects citrate. This produces correlated lines instead of spots for the cross-peaks and diagonal peaks with STOCSY (Figure 4). Therefore, in the case of citrate, the cross Analytical Chemistry, Vol. 77, No. 5, March 1, 2005

24 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x Figure 5. One-dimensional STOCSY analysis for the selected variable corresponding to δ ppm. The degree of correlation across the spectrum has been color coded and projected on the spectrum that has the maximum for this variable. (a) Full spectrum; (b) same spectrum between 7.1 and 7.5 ppm; (c) same spectrum between 2.4 and 3 ppm. peaks are not parallel to the diagonal peaks, which indicates that the coupling constant of both AB spin systems varies with the differences of physicochemical environment across the samples. An AX spin system of two triplets can be noticed at δ 2.91 and 2.51 ppm with a coupling constant 3 J AX ) 7.7 Hz. This spin system is strongly correlated to others resonances in the aromatic region of the spectrum. It is difficult to display the correlation between two distant resonances because the resulting peaks are too narrow relative to the large frequency difference between the peaks. However, there is a way to approach this problem as follows. Computing only the correlation between one of the data points representing the maximum of one of the triplets and all the other variables leads to obtaining one vector, which has the size of the number of variables used. Then, by selecting the spectrum with the maximum value of this selected variable, it is possible to plot it with a color code corresponding to the correlation between the selected resonance and all the other points of the spectra. A typical result is shown in Figure 5 and highlights all those correlated metabolite resonances that are not strongly overlapping with resonances from other molecules. Here the potential of the approach can be seen since correlations can be observed between resonances with no NMR-based spin-coupling connectivity. Thus, in the aromatic region, it is possible to recognize the resonances of a meta-substituted benzene ring (one triplet, two doublets, and one singlet). One of the triplet peaks is overlapped with a singlet from another molecule; however, the remaining correlation (plotted in light blue) is enough to reveal its position. Thus this molecule can be identified as a derivative of a meta-substituted phenylpropanoic acid and is probably 3-hydroxyphenylpropionic acid. 22 Furthermore, Figure 4A shows a correlation between methylamine and dimethylamine. The origin of this cross-peak is obviously not from a correlation between two parts of the same molecule but is from the fact that the concentrations of these two molecules are Figure 6. O-PLS cross-validated scores for the discrimination among 1 H NMR urine spectra of three mouse strains. Table 1. O-PLS Model Summary for Discrimination among the 1 H NMR Spectra of Three Mouse Strains a component R 2 X corr R 2 X yo R 2 X R 2 Y Q a R 2 X corr is the part of the modeled variance of X correlated to Y, and R 2 X yo is the part of the modeled variance of X orthogonal to Y. highly correlated, in that they vary in the same way because they are involved in the same pathways. This means that STOCSY is not only able to reveal the whole NMR peak set of single metabolites but is also able to highlight molecules involved in the related pathways. Analytical Chemistry, Vol. 77, No. 5, March 1,

25 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x Figure 7. O-PLS regression coefficient plot corresponding to the strain 129S6. The color bar corresponds to the weight of the corresponding variable in the discrimination between this strain and the other two strains. Finally, although the context of these two representations is different, if we compare the STOCSY plot with a TOCSY spectrum 23 of a single sample (Figure 4A and Figure 4B), the resolution of the STOCSY spectrum is much higher than that of a TOCSY spectrum. This phenomenon derives from the fact that the information provided by STOCSY is related to the whole sample set and not only to one sample. However, STOCSY is very adapted to a large number of samples and the two-dimensional NMR experiments (TOCSY, COSY, JRES,...) remain the best choice in the case of individual samples. With this example, the ability to highlight varying molecular concentrations in a metabonomic sample set has been demonstrated for this method. However, it does not provide any information about the differences between the mouse strains by assigning the metabolites to a specific class of urine sample. Pattern Recognition. The next step of the data analysis process employed a supervised pattern recognition procedure to bring out the specific variation of the urine composition according to the mouse strain. However, because a supervised data analysis method is used, the quality of the model had to be checked before any further interpretation. The O-PLS model enables very good prediction ability (Table 1), and two orthogonal-to-y components were taken on the basis of the cross-validation (maximum Q 2 ). These orthogonal components model the variations of the NMR spectra not correlated to the difference between the groups but interfering with the prediction: the structured noise. 21 The total explained variation of X for this model is relatively low (R 2 X ) 36%) because many regions of the spectra contained only instrumental noise, and the autoscaling of the corresponding variables contributes to increasing the random variance, which (22) Stanley, E. G. Ph.D., University of London, London, (23) Claridge, T. D. W. High-Resolution NMR techniques in Organic Chemistry; Elsevier: Amsterdam, is impossible to model. However, taking into account only the explained variation of X, 33% of the variation of the 1 H NMR spectra (X) is linearly correlated to the discrimination between the mouse strains (Y) and 80% of the variation of Y can be related to the variation of X. Good separation was achieved between the 1 H NMR spectra classes corresponding to three mouse strains and is illustrated by the cross-validated score plot (Figure 6). The discrimination between the strain 129S6 and the other two strains is clearer than the discrimination between the BALB/c and C57BL/6 strains, in which a slight overlap can be noticed. Moreover, the intragroup variance is larger for the BALB/c and C57BL/6 strains than for the 129S6 strain. The summary of the O-PLS model shows that the discrimination between the urine 1 H NMR spectra corresponding to the three mouse strains is clear and makes the interpretation of the O-PLS coefficients possible. The main purpose of this paper is to describe the potential of this methodology and to identify the varying metabolites rather than to focus on all the biological interpretations, which will be addressed in future publications. For this reason only the interpretations for the 129S6 strain O-PLS coefficients are presented. The number of coefficients in the O-PLS model is very large (30 K), but the postprocessing step, which combines back-scaled coefficients with the variable weights, allows the selection of the more important peaks for the discrimination. Figure 7 shows these coefficients plotted as a function of their corresponding chemical shifts, allowing their interpretation in the same way as a conventional NMR spectrum. Among all the different peaks, and according to the color coding, different resonances can be selected according to their weight in the discrimination between the strain 129S6 and the other strains (Table 2). With this list, it is already possible to nominate candidate molecules corresponding to these resonances. For example, the spin system of the glycerate matches 1288 Analytical Chemistry, Vol. 77, No. 5, March 1, 2005

26 Downloaded by DANISH UNIV OF PHARM SCI on October 27, Publication Date (Web): January 25, 2005 doi: /ac048630x Figure 8. Combination of O-PLS coefficients with 1D STOCSY analysis for three exemplar sets of resonances (maximum intensity correlation of peaks is color coded and projected into statistical difference spectra constructed as Figure 7): (A) δ ppm, glycerate; (B) δ ppm, isovalerate; (C) δ ppm, glutarate. Table 2. List of Resonances with O-PLS Weight >0.4 δ (ppm) multiplicity O-PLS weight variation assigned metabolite doublet 0.84 v isovalerate multiplet 0.87 v isovalerate quintuplet 0.69 v glutarate overlapped resonances 0.84 v glutarate and isovalerate singlet 0.56 V dimethylglycine doublet 0.74 v glycerate doublet 0.95 v glycerate doublet of doublets 0.93 v glycerate very well with the resonances at δ 4.095, 3.820, and ppm. The assignment is less obvious for the other resonances, particularly for those involved in the overlap at δ ppm. The information from the strain-nmr correlation (O-PLS) can be crossed with information from the NMR-NMR correlation (STOCSY) to provide highly interpretable models. From O-PLS coefficients plotted with the postprocessing highlighting the more important variables in the discrimination, it is possible to pick up a resonance of interest and replot the same O-PLS coefficients but with a color scheme corresponding to the correlation between the selected resonances and the other resonances of the spectra, as shown previously. This permits rapid visual identification for the experimental spectroscopist. Figure 8 presents the O-PLS coefficients of the 129S6 strain displayed with the correlations corresponding to three of the significant resonances and reveals very clearly the spin systems of the metabolites (glycerate, glutarate, and isovalerate). This shows the ability of combining O-PLS pattern recognition with the result of STOCSY in highlighting the metabolites that most vary among the different groups. CONCLUSION The results presented in this paper have illustrated a new framework for analysis of metabonomic data involving STOCSY and O-PLS based pattern recognition. 1 H NMR statistical total correlation spectroscopy has demonstrated the ability to decipher the structure of many metabolites from biofluid samples. Moreover, in combination with the O-PLS method, it provides a powerful tool for classification, prediction of sample class based on spectral features, and rapid interpretation of metabolic variation and identification of biomarkers, raising the level of interpretability and dissecting the finer features of spectra. ACKNOWLEDGMENT This work was funded by the Wellcome Trust Functional Genomics Initiative BAIR (Biological Atlas of Insulin Resistance) (066786) ( D.G. holds a Wellcome Senior Fellowship in Basic Biomedical Science (057733). We thank Dr. Roger Tatoud and Prof. James Scott for the coordination of the project and the other members of BAIR for scientific discussion. Received for review September 16, Accepted November 11, AC048630X Analytical Chemistry, Vol. 77, No. 5, March 1,

27 Review Received: 18 September 2008 Accepted: 7 May 2009 Published online in Wiley Interscience: 5 June 2009 ( DOI /mrc.2461 NMR metabolomics and drug discovery Robert Powers NMR is an integral component of the drug discovery process with applications in lead discovery, validation, and optimization. NMR is routinely used for fragment-based ligand affinity screens, high-resolution protein structure determination, and rapid protein-ligand co-structure modeling. Because of this inherent versatility, NMR is currently making significant contributions in the burgeoning area of metabolomics, where NMR is successfully being used to identify biomarkers for various diseases, to analyze drug toxicity and to determine a drug s in vivo efficacy and selectivity. This review describes advances in NMR-based metabolomics and discusses some recent applications. Copyright c 2009 John Wiley & Sons, Ltd. Keywords: NMR; metabolomics; drug discovery; disease biomarkers; drug toxicity; principal component analysis; differential NMR metabolomics S2 Introduction NMRspectroscopyis playing an integral and continuallyexpanding role in the pharmaceutical industry, especially since highthroughput screening [1 5] and structure-based drug discovery [6 8] have evolved to be the driving forces behind the discovery process for new therapeutics. [9] This process can be divided into three major steps: lead discovery, drug optimization, and clinical validation, and NMR makes invaluable contributions at all stages. [10 15] NMR is the primary analytical tool used to confirm the chemical structure and composition of both synthetic and natural product chemical leads. [16,17] NMR high-throughput ligand affinity screens high-throughput ligand affinity screens (HTS), especially given the growing popularity of fragment-based libraries, are a well-established component of the discovery process. [18,19] NMR HTS are routinely used to both validate and identify novel chemical leads. [20,21] The universal adoption of the fragmentbased approach means that the rapid screening of small chemical libraries by NMR enables an exponential growth in the exploration of structural space, well beyond traditional HTS methods. [22 24] In addition to the validation and identification of chemical leads, NMR continues to contribute to lead optimization by determining highresolution protein solution structures and rapid protein ligand co-structures. [25,26] The recent expansion into the analysis of the metabolome has also enabled NMR to contribute to the clinical validation step. [13,27,28] By far, this stage is the most challenging and expensive component of the drug discovery process, where a significant number of failures occur. [29,30] From the analysis of biofluids, tissues, and cell extracts, NMR can measure changes in the metabolome resulting from the biological activity of the drug lead. [31 33] The relative concentration and flux of the hundreds to thousands of small-molecular-weight compounds that comprise the metabolome reflect the state of the system. [34 37] As an illustration, a compound designed to inhibit a specific enzyme will result in changes in the concentration of substrates and products associated with the enzyme s activity. Thus, perturbations in the metabolome result from drug efficacy, selectivity, and toxicity. Additionally, the comparative analysis of the metabolome between healthy and diseased individuals identifies metabolites that can be used as biomarkers for the disease. [38 43] A major advantage of NMR-based metabolomic studies is the general ease and simplicity of the methodology. [44] In general, biofluids or cell lysates are simply added to a deuterated aqueous buffer to maintain ph and provide a lock signal before transferring to an NMR sample tube to collect a one dimensional (1D) 1 H NMR spectrum. [45 47] Because of the inherent variability in biological samples, it is necessary to obtain replicates and collect a similar number of NMR spectra so that any observed trends are statistically relevant. This collection of NMR spectra is typically analyzed using an unsupervised statistical technique, such as principal component analysis (PCA). [48,49] PCA reduces the multivariable NMR spectra into the lower dimensional PCA space. Specifically, an NMR spectrum is reduced to a single point in a standard two dimensional (2D) or three dimensional (3D) scores plot. The clustering of NMR spectra in a scores plot determines the relative similarity between the data, where spectra that cluster together indicate a similar metabolome. Accurately interpreting the PCA analysis of NMR spectra requires consistency of sample preparation, data collection, and data processing. [50] It is essential that the observed clustering pattern in the PCA scores plot reflects the anticipated perturbations in the metabolome due to drug activity instead of an artifact from data handling [51 53] Thus, or processing. an additional benefit of NMR-based metabolomics is the minimal sample manipulation, which reduces errors in PCA clustering patterns. In addition to monitoring global perturbations in the metabolome based on the statistical analysis of NMR spectra, the identity and concentration of the major metabolites affected by the drug are also explored by NMR. [54,55] This enables specific [38 43] metabolites to be identified as potential disease biomarkers, to determine if the drug therapy has toxic side effects, [56,57] and to identify metabolic pathways affected by the drug. [58,59] The ability to rapidly and easily monitor the in vivo activity of potential drug candidates at the early stage of drug discovery has significant Correspondence to: Robert Powers, Department of Chemistry University of Nebraska-Lincoln, 722 Hamilton Hall, Lincoln, NE , USA. rpowers3@unl.edu Department of Chemistry, University of Nebraska-Lincoln, 722 Hamilton Hall, Lincoln, NE , USA Magn. Reson. Chem. 2009, 47, S2 S11 Copyright c 2009 John Wiley & Sons, Ltd.

28 NMR metabolomics and drug discovery benefits for the effective treatment of human diseases. [60] Clearly, identifying compounds that exhibit diminished in vivo activity, poor specificity, or toxicity prior to conducting clinical trials is highly desirable and extremely cost effective. [61] Similarly, using NMR to develop accurate and non-invasive protocols for early disease diagnostics by identifying biomarkers is tremendously beneficial to human health. This review will discuss recent developments and applications of NMR metabolomics that are striving to achieve these goals. Methodology Processing of NMR spectra The obvious appeal of NMR-based metabolomics is the relative ease of the methodology, [28] but the success of the approach requires judicious attention to the uniform preparation of samples and consistent data analysis. [50] Specifically, issues such as long-term storage, [62,63] protein removal, [46] selection of extraction solvent, [45] and tissue preparation [47] can all affect the quality and reliability of the analysis. The advantage of PCA is the extreme sensitivity of the method to subtle spectral differences. This sensitivity may be problematic if the variability in clustering patterns within 2D scores plot result from changes in experimental conditions instead of monitoring perturbations in the metabolome. One such example is the unexpected contribution of NMR spectral noise to PCA clustering. [51] The principal component (PC) analysis of ideal metabolomics data consisting of two NMR samples containing either adenosine 5 triphosphate (ATP) or an ATP glucose mixture is shown in Fig. 1. The 10 duplicate spectra were obtained by repeatedly collecting an NMR spectrum utilizing a single sample. Surprisingly, a significant amount of dispersion was observed along the PC2 axis despite the essentially identical data. Even more disturbing was the observation that a single spectrum fell outside the 95% confidence level for the PCA model. Eliminating the noise from the NMR spectra resulted in a 5 tighter clustering pattern and, more importantly, removed the erroneous data point. Given the inherent sensitivity of PCA to spectral noise, changes in NMR chemical shifts and peak widths due to ph, temperature, or instrument fluctuation may result in undesirable changes in clustering patterns in 2D scores plots [64] A common approach to minimize these problems is the use of binning, where NMR spectra are divided into regions or buckets with widths of ppm. [65,66] The total peak intensity within these buckets is summed, which results in reduced resolution, but variations between spectra are smoothed out. The binning process itself may also induce errors that are caused by the definition of the bin edge. Ideally, the bin edge should correspond to a baseline region, but variations between spectra may cause a bin edge to occur at a peak. This occurs if a simple and constant bin definition is used for each spectrum. Recent techniques use intelligent binning to optimize the bin edge definition, which does not require a constant bin width. De Meyer et al. [67] describe an automated binning protocol that uses variable bin widths and a bin quality factor but does not require reference spectra or user-defined parameters (Fig. 2a). The bin quality factor strives to maximize the peak intensity within the bin while minimizing the peak intensity at the bin edges. Noise bins are discarded. Figure 1. PCA scoring plots of the set of 10 ATP ( )andatp glucose( ) NMR spectra (a) with noise and (b) removal of the spectral noise by only binning NMR resonances. The results clearly demonstrate the increased variability and dispersion in the scores plot due to noise (Reprinted with permission from Ref. [51], Copyright 2006 by Elesvier). S3 Magn.Reson.Chem.2009, 47, S2 S11 Copyright c 2009 John Wiley & Sons, Ltd.

29 R. Powers Figure 2. (a) Adaptive Intelligent (AI)-binning clearly isolates the two large R-1 acid glycoprotein peaks (AAG1 and AAG2) in separate bins (solid lines), in large contrast to the bins obtained after standard, equidistant binning (dashed lines), where only equidistant bin 3 is not a mixture of different peaks. (Reprinted with permission from Ref. [67], Copyright 2008 by American Chemical Society). (b) Left: the spectral region of interest in nine spectra from the hydrazine dataset, one doublet from the AB type spectrum of citrate, a triplet from 2-oxoglutarate, and the singlet from succinate. Center: the first three PCs and their corresponding normalized eigenvalues of the spectral region of interest. Right: the spectral region of interest after application of the procedure for the individual peak alignment. (Reprinted with permission from Ref. [68], Copyright 2008 by Elesvier). S4 Peak alignments between spectra are an alternative approach to binning and assist in the identification and quantification of metabolites present in the metabolome sample. [55,69] Anumber of approaches have been described to align 1D and 2D 1 HNMR spectra collected on metabolomic samples. [68,70 72] Stoyanova et al. [68] describe a protocol using PCA to identify regions of a spectrum that experience frequency or phase shift. Specifically, the second PC ( P 2 ) is sensitive to frequency shifts and will display a derivative shape when frequency shifts are a dominate factor. Subsections of the spectra shown to correlate with P 2 are then aligned by shifting the frequencies of the peaks to an average frequency. The procedure is repeated until all subsections of the spectrum that correlate with P 2 are aligned. The analysis of H NMR spectra of rat urine for a hydrazine toxicity study demonstrates the frequency shift observable in a metabolomics study (Fig. 2b). The PC analysis of the NMR spectra without peak alignment indicates the derivative shape P 2 to dominate, and frequency shifts contributes 15% to the total variance. Aligning the NMR spectra and adjusting for the frequency shifts results in a loss of the derivative shape and a drop in contribution to the total variance to 2%. Thus, the PC analysis is reflecting changes in the metabolome composition (desired outcome) instead of subtle chemical shift changes due to minor changes in ph, salt concentration, and temperature, or instrument stability. Assigning NMR metabolome spectra PC analysis of NMR metabolomics data provides a rapid approach to identify global trends and relationships. Alternatively, detailed analysis of the identity and concentration flux of metabolites provides specific comparisons that enable the determination of disease biomarkers and the identification of affected metabolic pathways. This is a relatively challenging endeavor due to the complexity of the metabolome and the lack of reference NMR spectra. First, the metabolome is not completely defined, may contain an infinite number of compounds, and is species dependent, [73] where the number of plant metabolites has been estimated to be [74] The Kyoto encyclopaedia of genes and genomes (KEGG) ( [75] MetaCyc ( [76] and the Human Metabolome [37] databases contain the extent of what is known regarding metabolic pathways. Similarly, the Madison Metabolomics Consortium database ( [36] Human Metabolome database ( [37] and COLMAR Metabolomics Web Server ( [35] are recent efforts to accumulate both 1 Hand 13 CNMRspectra assignments for known metabolites. These resources are enabling reliable assignments of NMR spectra to determine both the identity and concentration for the majority of metabolites in an NMR sample. [54] Nevertheless, assigning 1D 1 H NMR spectra for metabolomic samples is still Copyright c 2009 John Wiley & Sons, Ltd. Magn. Reson. Chem. 2009, 47, S2 S11

30 NMR metabolomics and drug discovery Figure 3. (a) One-dimensional 1 H NMR spectrum of an equimolar mixture of the 26 small-molecule standards. (b) Two-dimensional 1 H 13 CHSQCNMR spectra of the same synthetic mixture (red) overlaid onto a spectrum of aqueous whole-plant extract from A. thaliana (blue). (Reprinted with permission from Ref. [69], Copyright 2008 by American Chemical Society). considerably challenging because of significant peak overlap and the presence of uncharacterized metabolites. [77] Instead, the use of 2D NMR techniques is commonly used to analyze the composition of metabolomic samples. The fast metabolite quantification (FMQ) by NMR method described by Lewis et al. [69] uses a series of 2D 1 H 13 C heteronuclear single quantum coherence (HSQC) spectra collected for mixtures of standard metabolites over a range of concentrations. [69] An experimental biological sample is then used to collect a 2D 1 H 13 C HSQC spectrum, where peak intensity and chemical shifts are compared against the reference set to identify the metabolites and their corresponding concentration (Fig. 3). Fifty metabolites were identified in the biological extracts from Arabidopsis, alfalfa sprouts, and yeast with concentrations ranging from 230 mm to 40 µm. Integrating NMR and MS metabolomic data Mass spectroscopy (MS) has traditionally been used to detect perturbations in the metabolome, [78,79] where NMR and MS provide complimentary approaches to the analysis of metabolomic data. [80] An advantage of MS is its relatively high sensitivity and ability to monitor concentration fluxes for minor components that are typically undetected by NMR. [78,79] Conversely, MS typically requires a hybrid approach because of the low-molecularweight distribution of metabolites (Fig. 4a). Including gas or liquid chromatography to separate compounds with similar molecular weight (MW) may remove or perturb the relative concentration of metabolites. Also, MS is limited to detecting metabolites that are able to ionize well. NMR has similar limitations and is generally restricted to observing metabolites of high concentration. As a result, a number of techniques have been proposed that combine [81 85] NMR and MS data for the analysis of metabolomic samples. A PCA approach that combines 1D 1 H NMR data with desorption electrospray ionization mass spectrometry (DESI-MS) data was described by Chen et al. [81] The approach was applied to urine samples collected from mice to differentiate between healthy mice and mice with lung cancer (Fig. 4b). Simply, 2D scores plot are calculated separately for the NMR and DESI-MS datasets using a reduced compound dataset. The reduced compound dataset is simply subregions from both the NMR and MS spectra that corresponds to peaks associated with six compounds that distinguish the biological samples. Since the PCs from the NMR 2D scores plots are independent of the DESI-MS data, the NMR PC1 values are simply added to the DESI-MS PC values and become the third dimension in a 3D scores plot. The result is a higher separation of the biological samples in the PC space. S5 Magn.Reson.Chem.2009, 47, S2 S11 Copyright c 2009 John Wiley & Sons, Ltd.

31 R. Powers (a) (a) (b) (b) S6 Figure 4. (a) Histogram of molecular weights of typical microbial metabolites (Reprinted with permission from Ref. [73], Copyright 2004 by Elsevier). (b) 3-D score plot combining PCA of NMR and DESI-MS data comparing healthy mice (C1 and C3) with mice with lung cancer (T2 and T4) (Reprinted with permission from Ref. [81], Copyright 2006 by John Wiley & Sons). Applications Disease biomarkers One major promise of NMR metabolomics is the identification of biomarkers from biofluids for early disease diagnosis. [28] The approach is straightforward in concept: compare biofluids from healthy and diseased individuals to identify metabolites uniquely correlated with the disease state. Furthermore, it has the added advantage of being rapid and non-invasive, requiring the simple collection of urine, blood, or saliva samples from patients. Of course, there are inherent challenges and limitations in the use of biomarkers. [86] Fundamental variabilities in an individual s metabolome resulting from age, gender, genetics, environmental exposure, behavior, or diet differences may mask the impact of a disease or incorrectly imply the presence of a disease. Other factors, such as the collection, storage, and handling of the biological samples [62,63] or measurement errors, [51,64] may also compromise the correct identification and utility of biomarkers. NMR-based metabolomics have been used to identify biomarkers associated with a variety of diseases including asthma, [87] Figure 5. (a) Results of supervised principal component discriminant analysis (PC-DA) of plasma samples from run-in visits (visits 1 and 2). A good distinction between diabetic patients and healthy volunteers as well as separation by gender is observed (Reprinted with permission from Ref. [95], Copyright 2006 by Blackwell Publishing Ltd). (b) 3D PCA scores plot based on covariances of the five NMR bin intensities used in the cancer models. Pancreatic control samples are shown as open circles, while pancreatic cancer samples are shown as solid black circles (Reprinted with permission from Ref. [89], Copyright 2006 by Springer Science). arthritis, [88] cancer, [89 91] cardiovascular, [92] diabetes, [93 95] neurodegenerative, [96,97] and pathogen infections. [98] As an illustration, a clinical study described by van Doornet al. [95] demonstrates the use of 1D 1 H NMR analysis of blood serum samples to distinguish between healthy volunteers and type 2 diabetes mellitus (T2DM) patients. Eight healthy male and female volunteers and eight male and female patients diagnosed with T2DM had blood serum drawn twice a week over a 6-week period. A PCA of the NMR spectra (Fig. 5a) shows a large differentiation based on both the disease state and gender of the participants in the study. The discrimination is maintained even if the glucose resonances are removed from the NMR spectra. The T2DM biomarkers permitted a further study to determine the effect of thiazolidinedione therapy to treat T2DM. A similar clinical study described by Beger et al. [89] was conducted to identify biomarkers for pancreatic cancer, a disease with a high mortality rate (1-year survival rate of 20%) because of difficulties related to early diagnosis. Lipid extracts from plasma samples were collected from 90 healthy volunteers and 100 patients with pancreatic cancer. A subset of only four or five bins from the complete NMR spectra was used to create a partial least squares-discriminant function (PLS-DF) model that statistiwww.interscience.wiley.com/journal/mrc Copyright c 2009 John Wiley & Sons, Ltd. Magn. Reson. Chem. 2009, 47, S2 S11

32 NMR metabolomics and drug discovery cally discriminated between healthy individuals and patients with pancreatic cancer (Fig. 5b). Drug toxicity Drug toxicity is a very challenging, costly, and pervasive problem in drug discovery, [29,30,99 101] which is primarily caused by the inherent variability in a patient s response to a specific [ ] therapy. Even in some recent high-profile cases that resulted in the removal of drugs from the market, [105] the vast number of individuals administered the drug did not suffer serious side effects. [106] Generally, only a small percentage of the population suffers serious complications caused by a drug. Ideally, it would be best to identify these individuals prior to starting a drug therapy. [107,108] This would permit general access to the drug and its corresponding benefits to the majority of the population. It is also highly desirable to identify potential drug toxicity events prior to a treatment progressing to serious injury or death. Similar to its application in identifying biomarkers, NMR metabolomics is becoming an essential tool for the identification and evaluation of drug toxicity. [28,38,56,57] The approach is comparable to the identification of biomarkers: biofluids from animals or patients are analyzed before and after treatment with a drug candidate by 1D 1 H NMR and PCA. Any differences in the metabolome that have been associated with serious toxic events, such as liver damage, would be used to identify a toxicity problem with the drug candidate. The approach is demonstrated by a study conducted by Robertson et al., [109] where Wistar rats were treated with two known hepatotoxicants (CCl 4, α-naphthylisothiocyanate) and two known nephrotoxicants (4-aminophenol, 2-bromoethylamine (BEA)). Urine samples were collected daily from the rats and analyzed using 1D 1 H NMR and PCA. The 3D scores plot comparing drug treated rats (a) (b) Figure 6. (a) 1D 1 H NMR spectra of urine from Wistar rats dosed with two known hepatotoxicants and two known nephrotoxicants. Combined PCA analysis of all toxicant treatments (filled symbols) and all untreated samples (open circles). Toxicant data are as follows: squares = CCl 4,hexagons= α-naphthylisothiocyanate (ANIT), diamonds = 4-aminophenol (PAP), and triangles = 2-bromoethylamine (BEA). The results demonstrate the clear onset of toxicity (Reprinted with permission from Ref. [109], Copyright 2000 by the Society of Toxicology). (b) Score plots from 1D 1 HNMRspectraofurine collected from 7-week-old male Han Wistar rats comparing the postdosed samples of the control group (group 1) and the groups dosed with the five compounds (groups 2 6). The samples are colored according to group. Group 3 was determined to have an extreme excretion of choline and the two compounds were excluded as viable drug candidates (Reprinted with permission from Ref. [110], Copyright 2006 by American Chemical Society). S7 Magn.Reson.Chem.2009, 47, S2 S11 Copyright c 2009 John Wiley & Sons, Ltd.

33 R. Powers Figure 7. Illustration of the differential NMR metabolomics method. Different clustering patterns in PCA scores plot determine the activity and selectivity of drug candidates. Hypothetical PCA scores plot depict the following scenarios: (a) inactive compound, (b) active and selective inhibitor, (c) active, non-selective inhibition of target and secondary protein, and (d) active, non-selective preferential inhibition of secondary protein. Labels correspond to wild-type cells (wt) and mutant cells (mut) (Reprinted with permission from Ref. [31], Copyright 2006 by American Chemical Society). S8 Figure 8. (a) Analysis of the in vivo activity of 8-azaxanthine (AZA) in A. nidulans targeting urate oxidase. The PCA scores plot comparing A. nidulans inactive urate oxidase mutant (uaz14) ( ), wild-type with AZA ( ), uaz14 mutant with AZA ( ), and wild-type cells ( ). Results clearly demonstrate the selective activity (see Fig.7b) of AZA (Reprinted with permission from Ref. [31], Copyright 2006 by American Chemical Society). (b) Analysis of the in vivo activity of D-cycloserine (DCS) in mycobacteria targeting alanine racemase. PCA scores plot comparing wild-type (mc 2 155) ( ), inactive D-alanine racemase mutant (TAM23) ( ), DCS resistant mutants (GPM14 ( ), GPM16 ( )), restored D-alanine racemase activity mutant (TAM23 ptamu3) ( ) mc with DCS ( ), and TAM23 with DCS ( ), GPM14 with DCS ( ), GPM16 with DCS ( ), and TAM23 ptamu3 with DCS ( ). The results clearly demonstrate the active, non-selective inhibition of DCS (see Fig. 7c). The secondary target of DCS is predicted to be D-alanine-D-alanine ligase (Reprinted with permission from Ref. [32], Copyright 2006 by American Chemical Society). Copyright c 2009 John Wiley & Sons, Ltd. Magn. Reson. Chem. 2009, 47, S2 S11