Anal. Chem. 2005, 77, 2455-2463
Use of Selective TOCSY NMR Experiments for Quantifying Minor Components in Complex Mixtures: Application to the Metabonomics of Amino Acids in Honey Peter Sandusky and Daniel Raftery*
Department of Chemistry, Purdue University, 560 Oval Drive, West Lafayette, Indiana 47906
The application of the traditional methods of multivariate statistics, such as the calculation of principle components, to the analysis of NMR spectra taken on sets of biofluid samples is one of the central approaches in the field of metabonomics. While this approach has proven to be a powerful and widely applicable technique, it has an inherent weakness, in that it tends to be dominated by those chemical species present at relatively higher concentrations. Using a set of commercial honey samples, a comparison of this classical metabonomics approach to one based on the use of the selective TOCSY experiment is presented. While the NMR spectrum of honey and its classical metabonomic analysis is completely dominated by a very few chemical species, specifically r-glucose and fructose, the statistical signal carried by minor honey components, such as amino acids, may be accessed using a selective TOCSY-based approach. This approach has the intrinsic virtue that it focuses the statistical analysis on a set of predefined chemical species, which might be chosen for their metabolic significance, and could be composed of either major or minor mixture constituents. Furthermore, the selective TOCSY method allows for more certain chemical identification, acquisition times of ∼1 min, and accurate quantification of the species contributing to the statistical discriminatory signal. The central idea of the field of metabonomics is the application of the classical methods of multivariate statistics, such as principle component analysis (PCA),1,2 to data sets derived from the intensities of 1D proton NMR or LC/MS data.3 This has proven to be a very powerful approach for the rapid analysis of sets of biofluid mixtures, such as urine and body fluids, as well as the related liquid foods, where each member of the sample set is, in gross terms, chemically similar to any other member, but where interesting differences between members of the set may be observed at a more subtle level. * Corresponding author. E-mail:
[email protected]. (1) Krzanowski, W. J. Principles of Multivariate Analysis: A User’s Perspective, revised edition; Oxford University Press: Oxford, U.K., 2000. (2) Johnson, R. A.; Wichern, D. W. Applied Multivariate Statistical Analysis, 4th ed.; Prentice Hall: Upper Saddle River, NJ, 1999. (3) Lindon, J. C.; Holmes, E.; Nicholson, J. K. Anal. Chem. 2003, 75, 384A391A. 10.1021/ac0484979 CCC: $30.25 Published on Web 03/08/2005
© 2005 American Chemical Society
NMR-based metabonomics incorporates two key virtues. First, it is an integrated or whole sample approach, thus avoiding the time, the difficulty, and the possibility of differential fractionation, which are inherent problems with hyphenated techniques. Second, it looks at all the components of the sample at one time and thus might be expected in principle to pull out the complete profile of subtle chemical differences distinguishing different subsets of a given biofluid. Thus, PCA of 1D proton spectra may, for instance, distinguish the urine of chemically stressed rats.4 The key limitation of the metabonomic approach is also that it looks at all the components of a mixture at once. This is problematic because it is almost always the case that a given biofluid, though composed of hundreds or thousands of different organic compounds, will contain a few species present at high concentration levels that will dominate the PCA. Thus, in a correlation PCA study of apple juices, only four species, glucose, fructose, sucrose, and malate, contributed significantly to the principle components.5 In general, this is unfortunate since it is clear that the biochemical and physiological properties of biofluids are often determined by the minor components. Also, because of the dominance of major components, it is difficult to understand the patterns observed in empirical metabonomic studies in terms of metabolism. For instance, while it is very interesting to learn from a metabonomic study that the urine of a chemically stressed rat has decreased levels of the major urine component hippurate,4 it would also be interesting to know how the chemical stress affects the minor components that are metabolically linked to hippurate. It is clear that there is a need for an alternative analytical approach to biofluid mixtures that would focus on a predefined set of compounds and allow for the accurate relative quantification of members of the set regardless of whether they are major or minor components. Selective total correlation spectroscopy (TOCSY) NMR experiments, developed 20 years ago, would seem to be able to isolate from a complex mixture the spectral peaks (4) Holmes, E.; Nicholls, A. W.; Lindon, J. C.; Connor, S. C.; Connelly, J. C.; Haselden, J. N.; Damment, S. J. P.; Spraul, M.; Neidig, P.; Nicholson, J. K. Chem. Res. Toxicol. 2000, 13, 471-478. (5) Belton, P. S.; Colquhoun, I. J.; Kemsley, E. K.; Delgadillo, I.; Roma, P.; Dennis, M. J.; Sharman, M.; Holmes, E.; Nicholson, J. K.; Spraul, M. Food Chem. 1998, 61, 207-213.
Analytical Chemistry, Vol. 77, No. 8, April 15, 2005 2455
from individual chemical species.6 However the intensities of the TOCSY peaks are very strongly affected by the mixing time, relaxation times, and coupling constants of the target spin system.7 After reviewing the present literature, it is unclear as to whether the resulting TOCSY peaks will give data quantitative enough to be of use in metabolic profiling. This paper presents a study in which the classical spectral intensity-based metabonomics approach is compared to one based on specific species quantification using the selective TOCSY experiment. The specific analytical problem chosen to test this new approach on is a particularly sticky one: the determination of amino acids in honey. Our results clearly indicate that the approach shows considerable promise. The TOCSY peaks do in fact give good quantification of the target species, even when the target species are found at concentrations 1000 times below those of the major components. Further, the selective TOCSY approach allows the subsequent PCA calculations to access the discriminatory potential of the minor component variances and is thus a more sensitive method to distinguish the origin of honey than the classical metabonomics approach. Finally, the selective TOCSY experiment is rapid and provides a more certain identification of the chemical species contributing to the sample discrimination than principle component loading plots.
Table 1. Amino Acids Observed in Honey by 1D TOCSY amino acid
excitation (ppm)
Pro
β1
(2.33)
Pro
γ
(1.98)
Ala Thr
β CH3 γ
(1.46) (1.31)
Tyr Tyr ethanol Ile
CH CH CH3 CH2 γ1
(6.90) (7.19) (1.17) (1.46)
Ile
CH2 γ2
(1.25)
Phea,b
nac
TOCSY peaks (ppm) R δ1 δ2 β2 γ R δ1 δ2 β1 R R β CH CH CH2 β CH2 γ2 CH3 γ CH3 δ β CH2γ1 CH3γ CH3δ nac
(4.12) (3.40) (3.32) (2.06) (1.98) (4.12)a (3.40) (3.32) (2.33) (3.80)a (3.60)a (4.25) (7.19)a (6.90) (3.65)a (1.98) (1.25) (0.99) (0.91) (1.98) (1.46) (0.99) (0.91)
a Used in PCA analysis. b Phe quantitation based on peak height at 7.42 ppm. c Not applicable.
EXPERIMENTAL SECTION Honey Samples. Seven sources of commercial honey were purchased locally from food markets. Of these, six were homogenized and one was labeled “organic” and “unhomogenized”. Three of these seven, constituting the “subgroup II” NMR samples described below, were different botanical varieties of homogenized honey sold under different labels by one manufacturer. An eighth source of honey, constituting the “subgroup I” NMR samples described below, was from the local “Bee’s Knees” apiary and was the generous donation of Dr. Peter Kissenger of Bioanalytical Systems, Inc. (West Lafayette, IN). To reduce the viscosity of the honey to make it fluid enough for liquid-state NMR analysis, 1.0 ( 0.1 g of honey was weighed and diluted with 6 ( 0.2 g of 99.9% D2O (Cambridge Isotopes Laboratory, Inc.). Once a homogeneous solution was achieved using a vortex rotor, 600 µL of dilute honey solution was transferred to a standard 5-mm NMR tube for analysis. To account for subtle variations in concentration, each of the eight honey sources was prepared in triplicate, resulting in a total of 24 7-fold dilute honey NMR samples. NMR Spectroscopy. All NMR spectra were taken on a Bruker Avance DRX 500-MHz spectrometer (Bruker-Biospin, Fremont, CA), using a 5-mm inverse HCN triple resonance probe equipped with triple axis XYZ gradient coils. All spectra were acquired at 25 °C. Simple proton spectra were acquired using a 1D NOESY pulse sequence incorporating presaturation for water suppression during the relaxation delay and mixing time.4,5 The relaxation delay and mixing time were set to 2 s and 300 ms, respectively, and the presaturation power used was the minimum needed to effect complete suppression of the water peak. To achieve high S/N for amino acid minor components, 120 FID transients were averaged, resulting in a total acquisition time of 15 min. Selective TOCSY experiments used the standard pulse sequence found in
the Bruker XWINNMR pulse program library. This sequence employs a z gradientsselective 180° pulsesz gradient train to achieve selective excitation of the target peak, followed by a MLEV-17 TOCSY spin lock.6,8,9 Gaussian-shaped pulsed z-field gradients were 1 ms in duration and 14 mT/m at maximum strength. Secant-shaped selective 180° pulses with pulse durations of 40 ms were found to be most effective for selective excitations. TOCSY mixing times were 70 ms. Thirty-two 16K point FID transients were averaged in each selective TOCSY experiment, resulting in an acquisition time of 1 min. Classical Metabonomics PCA Calculations. Proton NMR spectra were acquired using the 1D presaturation NOESY sequence for each of the 24 dilute honey samples and were transformed and phased using XWINNMR. Each real portion of the transformed spectrum was converted to a XY plot format JCAMP file. The JCAMP file was text edited to remove header text and X data and read in as a column into EXCEL (Microsoft Corp., Redmond, WA). In EXCEL the 16K points of each spectrum were 8-fold binned to yield a 2K data column, and in this way a 2K by 24 matrix, containing the spectra of the 24 samples, was constructed. Different portions of this master data matrix, alternatively comprising the complete spectra, the sugar dominated portion of the spectra, or the high-field portion of the spectra were used as input data for correlation PCA calculations in MINITAB 13 (MINITAB Inc., State College, PA). Correlation PCA calculations were performed rather than covariation calculations, since the former have been demonstrated to be somewhat more sensitive to minor constituents.5 In all cases, fewer than 20 principle components were found to be adequate to account for 100% of the variance.
(6) Kessler, H.; Oschkinat, H.; Griesinger, C. J. Magn. Reson. 1986, 70, 106133. (7) Marx, R.; Glaser, S. J. J. Magn. Reson. 2003, 164, 338-342.
(8) Stott, K.; Stonehouse, J.; Keeler, J.; Hwang, T.-L.; Shaka, A. J. J. Am. Chem. Soc. 1995, 117, 4199-4200. (9) Bax, A.; Davis, D. G. J. Magn. Reson. 1983, 65, 355-360.
2456 Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
Figure 1. 1D proton spectrum of 7-fold dilute honey acquired using 1D NOESY sequence with presaturation for water suppression during the relaxation delay and mixing times. The three regions presented are the sections of spectrum used in the three spectral intensity-based PCA calculations presented in Figures 2 and 3: (A) full spectrum from 0 to 8 ppm, (B) sugar-dominated region from 2.8 to 5.8 ppm, and (C) high-field region from -0.6 to 2.8 ppm.
PCA and ANOVA Using Selective TOCSY Data. One set of selective TOCSY experiments was performed for each of the target amino acids listed in Table 1, where each member of the set was one of the 24 honey samples. The resulting set of 24 FIDs was transformed as a set using uniform processing and phasing parameters (exponential apodization with a 1-Hz line broadening). For each target amino acid, one intense, well-isolated TOCSY peak was chosen for integration. The integration of this peak over the set of 24 TOCSY spectra was carried out using a XWINNMR software macro. This resulted, for each target amino acid, in a sequence of numbers representing the relative concentrations of the target as they vary over the 24 honey samples. Five of these sequences, representing the relative concentrations of five target amino acids, were then read into the columns of an EXCEL spread sheet. 1D proton NMR peak intensity data for phenylalanine was
then added as a sixth column (see below), and the resulting 24 by 6 matrix was read into MINITAB 13. Correlation PCA calculations were performed for six principle components. Subsequently, selected sections of the 24 by 6 data matrix, containing data from the subgroup II and subgroup III samples, were used in MINITAB 13 univariate ANOVA calculations.
RESULTS Classical Spectral Intensity-Based Metabonomics Study. At a gross level of consideration honey is a three-component system, composed of R-glucose, fructose, and water. Together these three species account for, on average, 86.7% of the sample by weight. Other simple sugars comprise, on average, 13.2% of the sample weight, while amino acids and other natural products Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
2457
are present as minor components. The amino acid composition of honey has received considerable attention and has been shown to be a good indicator for the place of origin. However, determination of amino acids in honey by HPLC is difficult because of the high sugar concentration and the presence of other chemical species, which will poison HPLC columns.10-13 The results from classical metabonomics calculations on the set of 24 honey samples are presented in Figures 1-3. The 500MHz proton NMR spectrum of 7-fold dilute honey is dominated by the R-glucose and fructose peaks occurring between 5.4 and 3.0 ppm (Figure 1A and B). At a 200-fold increase of the display amplitude, amino acids and other minor components become evident at higher field, 0.8-3.0 ppm (Figure 1C), and in the aromatic region. Although the minor component regions of the spectrum are continuous forests of peaks, peaks from a large number of species, including proline, alanine, threonine, leucine, isoleucine, valine, phenylalanine, tyrosine, and ethanol, are readily identified. Because spectral features of the major and minor constituents are isolated into different chemical shift regions in the honey spectrum, it is possible to perform spectral intensity-based PCA calculations independently on the major and minor constituents.5 This allows a direct assessment of the sensitivity of classical spectral intensity-based metabonomics studies to the presence of highly variable constituents present as minor constituents in biofluid mixtures. Thus, the PCA results presented in Figure 2A were calculated using data inputs from the entire spectrum, 0-8 ppm, while the results shown in Figure 2B were calculated using data inputs only from the major sugar component region of the spectrum, 3.0-5.4 ppm. In both cases, the correlation PCA calculation is completely dominated by the major sugar components, R-glucose and fructose. This point is readily apparent in the loading plots shown in Figure 3A and B. This is in fact true to the extent that exclusion of the high-field minor component region of the spectrum has no effect on the PC2 versus PC1 score plots (compare Figure 2A to Figure 2B). It is obvious that this is not the case because of any greater intrinsic variance in the major sugar constituents. In fact, the proton spectra of all eight brands of honey examined are very similar in the sugar region, if not almost identical. In contrast, very high levels of variability are apparent in the amino acid regions of the spectrum, which essentially constitute a fingerprint for each individual brand of honey, as was pointed out in HPLC studies by Gilbert et al.12 The score plot generated by the PCA analysis of the high-field amino acid region of the spectrum is quite different from that generated in the sugar-dominated calculations (compare Figure 2C to Figure 2A and B). However, what is more subtle and interesting in the context of amplifying the discriminatory potential of the variance of minor constituent species is that all three PCA calculations (full spectrum, sugar only, and high field) are essentially equivalent in terms of their ability to discriminate between the honey samples from different commercial brands. It will be shown below that PCA calculations performed using a data matrix derived from (10) Davies, A. M. C. J. Apic. Res. 1975, 14, 29-39. (11) Davies, A. M. C. J. Food Technol. 1976, 11, 515-523. (12) Gilbert, J.; Shepherd, M. J.; Wallwork, M. A.; Harris, R. G. J. Apic. Res. 1981, 20, 125-135. (13) Pirini, A.; Conte, L. S.; Francioso, O.; Lercker, G. J. High Resolut. Chromatogr. 1992, 15, 165-170.
2458 Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
Figure 2. PC2 vs PC1 score plots from correlation PCA calculations on the set of 24 dilute honey samples, performed using spectral intensity data sets from different sections of the 1D proton spectra: (A) score plot for calculation performed using full spectrum from 0 to 8 ppm (note three subgroup I samples in upper right), (B) score plot for calculation performed using sugar-dominated region of the spectrum from 2.8 to 5.8 ppm (note three subgroup I samples are in upper right), and (C) score plot for calculation performed using highfield region of the spectrum from -0.6 to 2.8 ppm (note three subgroup I samples at higher PC2 values).
selective TOCSY-based quantifications of a selected set of amino acids can divide the initial group of eight brands of honey into three subgroups, which are labeled as subgroup I, subgroup II, and subgroup III. In this context, all three spectral intensity-based PCA calculations described here have generated PC1 versus PC2 score plots (Figure 2A-C) that allow for the discrimination of subgroup I samples from those of subgroups II and III but do not allow for the discrimination of subgroup II samples from subgroup III samples. Selective TOCSY-Based Quantification and Metabonomics Study. The high-field region of 7-fold dilute honey shows an essentially continuous forest of overlapping peaks (Figure 4A). Nonetheless, a 1-min acquisition selective TOCSY experiment, performed with the selective pulse frequency centered at the chemical shift of the proline γ position, isolates the proline peaks cleanly from the obscuring features of the other honey components (Figure 4B). These results should be compared to those obtained on a mixture of 10 mM proline and 10 mM arginine (Figure 4C). The S/N for the proline R position in the honey
Figure 3. PC1 and PC2 loading plots from correlation PCA calculations on the set of 24 dilute honey samples, performed using spectral intensity data sets from different sections of the 1D proton spectra: (A) loading plots from calculation performed using full spectrum from 0 to 8 ppm, (B) loading plots from calculation performed using sugar-dominated region of the spectrum from 2.8 to 5.8 ppm, and (C) loading plots from calculation performed using high-field region of the spectrum from -0.6 to 2.8 ppm.
sample is ∼6.5, while the same peak in the proline-arginine model system, obtained and processed using the same parameters, has a S/N ratio of ∼100. Correcting for the differences in proline concentration, 650 µM for the honey sample and 10 mM for the model system, it would appear that the intrinsic sensitivity of the selective TOCSY experiment in the two samples is the same. In other words, no attenuation of the proline R signal has occurred in the 7-fold dilute honey sample. However, at higher honey concentrations, significant reduction in sensitivity of the proline R position peak is observed. This can be attributed to the higher viscosity of the honey samples and consequent increase in the
relaxation rate during the TOCSY spin lock. (Bax and Davis have observed that the relaxation rate of a spin system magnetization during a MLEV-17 spin lock is the average of 1/T1 and 1/T1F,9 and both 1/T1 and 1/T1F should increase with increasing viscosity as the honey concentration is raised.) Similar results were obtained with isoleucine. A quantitative investigation of the effect of relaxation on the TOCSY intensity of the alanine R-carbon peak showed that at a dilution level of 1:3 (where the sugar concentration is still well above 1 M) the T1 was reduced by ∼25% compared to a dilution level of 1:7. The TOCSY intensity was reduced by a similar amount. However, at higher dilutions, up to 1:27, there Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
2459
Figure 4. (A) 1D proton spectrum of 7-fold dilute honey acquired using 1D NOESY sequence with presaturation for water suppression during relaxation delay and mixing time. (B) Selective TOCSY spectrum of 7-fold dilute honey with selective excitation on the proline γ peak (1.98 ppm). (C) Selective TOCSY spectrum for mixture of 10 mM L-proline and 10 mM L-arginine with selective excitation on the proline γ peak (1.98 ppm).
was essentially no effect on the TOCSY intensity. And importantly, different honey samples behaved essentially identically at the same dilution level. For any given TOCSY mixing time, the intensities of TOCSY peaks observed are generally related to a number of complex factors, including ones deriving from the stoichiometry and coupling constants of the target spin system, and ones deriving from the relaxation properties of the system.7 Therefore, it should be expected that the peaks observed in the honey selective TOCSY experiments will display considerable distortions relative to the 1D proton spectrum, and in general, the intensities of the TOCSY peaks will not be stoichiometric representations of the protons in the spin system. The intensity of a particular TOCSY peak should nonetheless be expected to be proportional to the concentration of the chemical species containing the target spin system, as long as variation in the composition of the samples in the sample set does not vary to such a degree that significant changes are effected in the relaxation properties of the spin system. This is in fact the case with amino acids in honey. Titration of isoleucine into a honey sample shows that the intensity of the TOCSY peaks produced by selective irradiation of either of the γ 2460 Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
methylene protons is proportional to the isoleucine concentration over a range of 1.5 mM (Figure 5). Similar results were obtained for a proline titration. These are significant results, which indicate that, employed judiciously, selective TOCSY measurements can be used to quantify species in a dilute mixture. To evaluate the effectiveness of selective TOCSY quantification as the basis for metabonomics studies, a selective TOCSY data matrix was determined for a chosen set of constituent amino acids, as they occur at varying concentrations over the 24 dilute honey samples. A preliminary survey of the proton NMR spectra of the samples indicated that there was a good deal of variation between honey brands with respect to the concentrations of alanine, proline, threonine, phenylalanine, tyrosine, and ethanol. All of these species were readily quantifiable using selective TOCSY, with the exception of phenylalanine, where the peaks of the spin system were too close together (Table 1). Integrated intensities of selected TOCSY peaks for each of the other species were used, along with the peak heights for phenylalanine, as the input data matrix for a correlation PCA calculation. The score plot of the first two principle components calculated in this analysis divides the set of 24 samples into three subgroups (Figure 6A). PC1 is most significant in distinguishing the three subgroups, and examination of the PC1 loadings presented in Table 2A indicates that it contains roughly equal weighting from each of the six chemical species. In these PCA results, a high negative value of PC1 for a particular sample will indicate high concentrations of the target amino acids. The first subgroup (subgroup I) contains three samples from only one brand of honey (Bee’s Knees) and is characterized by high negative values of PC1. The second subgroup (subgroup II) contains nine samples from three different brands of honey, all sold by a single company, and is characterized by near-zero values of PC1. The last subgroup (subgroup III) contains 12 samples, from 4 brands of honey sold by 4 different companies, and is characterized by positive values of PC1. Univariate ANOVA calculations indicate that the PCA-generated discrimination between subgroup II and subgroup III observed in Figure 6 derives from statistically significant differences in three chemical species; threonine, tyrosine, and phenylalanine (Table 2B). DISCUSSION The results presented here indicate that selective TOCSY spectra of biofluid mixtures produce data that are quantitative measurements of selected constituent components. This should be a very general result. Because metabolic profiling in its nature is interested in natural variations of metabolic components present in some biofluid mixture at low absolute concentrations, the variations in the relaxation properties occurring over a set of samples in any particular study should not be significant. The viscosity and relaxation properties of a 7-fold dilute honey sample should be very similar to that of any other 7-fold dilute honey sample, as our experiments on a limited set of samples show. Barring dramatic increases in concentration, the viscosity and relaxation properties of a suitable prepared rat urine sample should be very similar to those of any other neat rat urine sample, etc. Because of the very large difference in concentrations between the major components and minor components, and because of the unfortunate effects of the sugars on the viscosity and NMR relaxation properties, honey is a very challenging subject for
Figure 5. Selective TOCSY L-isoleucine titration into 7-fold dilute honey. (A) High-field 1D proton spectrum of 7-fold dilute honey with 120 µM added L-isoleucine. Spectrum was acquired using 1D NOESY pulse sequence with presaturation for water suppression during relaxation delay and mixing time. (B) Selective TOCSY spectrum of 7-fold dilute honey with 120 µM added L-isoleucine. Selective excitation at L-isoleucine 1.25 ppm CH2 γ peak. (C) Plot of TOCSY peak integrated intensity vs added L-isoleucine concentration. 2 Integrated intensity of γ and δ CH3 peaks at 1.0 and 0.9 ppm (integrated together). 9 Integrated intensity of γ CH2 peak at 1.46 ppm. [ Integrated intensity of β CH peak at 1.90 ppm.
selective TOCSY experiments. We expect that most other biofluids or liquid foods of interest would in fact give better results. Preliminary results from our laboratory on selective TOCSY experiments in human urine show that indeed the selective TOCSY approach outlined here applies straighforwardly, and these results will be published soon. It is also important to note that while metabolic profiling studies often employ quantification of selected species based on signal intensities or integrals,14,15 the selective TOCSY experiment isolates the features of a single spin system from a crowded forest of peaks, thus allowing for more certain identification of a species, and cleaner and more accurate quantification. In a recent chemo(14) Brescia, M. A.; Kosˇir, I. J.; Caldarola, V.; Kidricˇ, J.; Sacco, A. J. Agric. Food Chem. 2003, 51, 21-26. (15) Nord, L. I.; Vaag, P.; Duus, J. Ø. Anal. Chem. 2004, 76, 4790-4798.
metric study on beer, considerable effort was made by the authors to determine the accuracy of NMR spectral intensity-derived quantifications.15 Comparing NMR intensity-derived quantifications of various major components found in beer to determinations of these components made by HPLC and CE, the authors found a wide range of inaccuracies, which were in a number of cases as high as 20-50% in error at concentrations of 2 mM. In comparison, the errors in the selective TOCSY quantifications presented here were less than 10% at 100 µM (Figure 5). While the samples are not directly comparable, our results do indicate that the judicious use of selective TOCSY may be an excellent and rapid means to obtain relative concentrations for a selected set of metabolites over a set of biofluid samples. The technique can also be used to make accurate absolute concentration determinations by calibrating the TOCSY peak intensities using a standard addition cocktail. Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
2461
Figure 6. Results from correlation PCA calculation on the honey samples using selective TOCSY peak intensity data for the six target amino acids (see Table 1). (A) Score plot for first two principle components showing separation of 24 honey samples into three subgroups. Black symbols are subgroup I samples, gray symbols are subgroup II samples, and open symbols are subgroup III samples. Shapes indicate different honey samples that were measured in triplicate. (B) Plot of eigenvalue vs principle component number, indicating that the first several eigenvalues explain the variation seen in the data. Table 2 (A) TOCSY-Based PCA Analysis of Honey Samples (from MINITAB 13 Calculation)
eigenvalues proportion cumulative variable PHE TYR PRO ALA THR ETH
PHE TYR PRO
PC1
PC2
PC3
PC4
PC5
PC6
4.727 0.788 0.788
0.800 0.133 0.921
0.345 0.058 0.979
0.077 0.013 0.992
0.048 0.008 1.000
0.002 0.000 1.000
PC1
PC2
Loadings PC3
PC4
PC5
PC6
-0.402 -0.449 -0.254 -0.439 -0.429 -0.443
0.238 0.038 -0.920 0.003 0.307 -0.027
-0.715 -0.135 -0.207 0.486 -0.016 0.437
-0.237 -0.359 0.210 -0.129 0.839 -0.226
0.463 -0.804 0.040 0.278 -0.122 0.215
0.016 0.066 0.011 0.691 -0.057 -0.718
(B) ANOVA F-Ratios between Subgroups II and III 100.7 ALA 31.0 THR 0.6 ETH
0.1 35.0 1.1
The selective TOCSY-based approach also appears to provide more information regarding the origin of the honey samples than the classical metabonomics approach. Although the score plots generated by the spectral intensity-based PCA calculations do repeatedly distinguish subgroup I from the other honey samples, 2462 Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
only the selective TOCSY experiments distinguish subgroup II from subgroup III (compare Figure 2 to Figure 6). Examination of the loadings plots for PC1 and PC2 generated in the high-field spectral intensity-based PCA calculation indicates that the observed PCA separation between subgroup I and subgroups II and III in this calculation derives from statistical differences in the concentrations of alanine, proline, and ethanol (Figure 3C). These three species have in effect dominated the high-field spectral intensity-based calculation in a manner similar to the way R-glucose and fructose dominated those calculations in which the sugar-dominated chemical shift region was included (Figures 2A. 2B, 3A, and 3B). Limiting the spectral intensity-based PCA calculation to the sugar-free high-field region in effect creates a new tier of dominant “major components”: alanine, proline, and ethanol. However, the ANOVA calculations presented in Table 2B, which excluded data from subgroup I samples, indicate that of the six species used in the selective TOCSY-based PCA calculations (Figure 6), only threonine, tyrosine. and phenylalanine are significant in distinguishing subgroup II from subgroup III samples. This again illustrates the problem, inherent in spectral intensity-based metabonomics studies, of major components, in this case the “second tier” major components, alanine, proline, and ethanol, dominating the PCA calculation and effectively suppressing the signal carried by variances of minor components, in this case, threonine. Looking at the amino acids that discriminate the data, two of them (phenylanaline and tyrosine) have resonances in the lowfield region of the spectrum, such that a PCA analysis of the 1D NMR data using just that spectral region might suffice for separation of the honey samples. This analysis was carried out using the low-field region 6.5-8.5 ppm and is provided as Supplementary Information (Figure 1S). As the figures show, however, while there is some separation of the honey samples along PC2, the sample clusters are not separated nearly as cleanly as they are by the selective TOCSY experiments. In particular, PC1 is dominated by variance from the noise in the spectra. In a classical spectral intensity-based metabonomic study of apple juices, it was observed that, although apple juices must contain hundreds or thousands of different organic species, the correlation PCA calculations were dominated by just four species; glucose, fructose, sucrose, and malate.5 The classical metabonomic results on dilute honey presented here in Figures 1-3 again illustrate that the major components, when they are included in a spectral intensity-based PCA calculation, will dominate and that the minor components will be suppressed. In the selective TOCSY approach, where input data are TOCSY peak integral intensities introduced into the PCA calculation as discrete numbers, the data for each different species may be acquired to some predefined level of S/N. Thus in the selective TOCSY approach there need be no difference, at the level of the PCA calculation, between major and minor components of a biofluid mixture. CONCLUSIONS Selective TOCSY experiments show promise in their ability to zero in on relevant spectral markers for differentiating and classifying different biofluid samples. While the experimental procedure is somewhat more complex than simply acquiring 1D proton spectra, the benefits include an improved ability for spectral indentification, the exclusion of large, uncorrelated, and potentially
overlapping components, and quantification capabilities. We believe that a further exploration of the potential for selective TOCSYtype experiments, or other more advanced NMR approaches, may yield improved discriminatory power in metabonomics studies. ACKNOWLEDGMENT The authors thank Dr. Pete Kissinger for kindly providing the Bee’s Knees honey sample. This work was supported by the National Institutes of Health.
SUPPORTING INFORMATION AVAILABLE PCA scores and loading plots for the 1D NMR honey spectra using the low-field spectral region. This material is available free of charge via the Internet at http://pubs.acs.org.
Received for review October 12, 2004. Accepted January 24, 2005. AC0484979
Analytical Chemistry, Vol. 77, No. 8, April 15, 2005
2463
