Solution nuclear magnetic resonance (NMR) spectroscopy is widely used to analyze complex mixtures of organic compounds such as biological fluids and tissue extracts. Targeted profiling approaches with reliable compound quantitifcation are hampered, however, by signal overlap and other interferences. Here, we present a tool named MetaboQuant for automated compound quantification from pre-processed 1D and 2D heteronuclear single quantum coherence (HSQC) NMR spectral data and concomitant validation of results. Performance of MetaboQuant was tested on a urinary spike-in data set and compared with other quantification strategies. The use of individual calibration factors in combination with the validation algorithms of MetaboQuant raises the reliability of the quantification results. MetaboQuant can be downloaded at http://genomics.uni-regensburg.de/site/institute/software/metaboquant/ as stand-alone software for Windows or run on other operating systems from within Matlab. Separate software for peak fitting and integration is necessary in order to use MetaboQuant.
Metabolomic analyses of biological fluids and tissue extracts have become a major application of nuclear magnetic resonance (NMR) (1). Targeted profiling approaches are used to quantify predefined compounds (2), yet signal overlap due to the large number of endogenous and exogenous compounds present in biological specimens makes this process difficult (3). One way to address this problem is by fitting expected line shapes to an observed signal (4), although this approach is less useful in heavily overlapped regions. One may exclude strongly overlapped signals and integrate only signals that appear free of overlap by eye, but an automatic routine would present a better solution since it would not rely on the skill and experience of the researcher inspecting the spectra. In addition, automation saves time when quantifying large spectra sets containing hundreds of signals as is often the case in metabolomic studies.
Another problem encountered when performing quantitative NMR analyses is that signal intensities of equally concentrated compounds–even different signals from the same compound– might vary considerably. Different strategies have been proposed to circumvent this issue in HSQC spectra, for example, using experimentally determined (3, 5) or theoretically calculated (6) calibration factors, or reconstruction of time-zero HSQC spectra (7). With the exception of time-zero HSQC spectra, these approaches rely on the multiplication of peak integrals by individual calibration factors.
Here, we describe a software tool named MetaboQuant that calculates accurate compound concentration values from 1D and 2D NMR peak intensities using individual calibration factors and different outlier detection algorithms. MetaboQuant performs the quantification procedure automatically once the required parameters have been set.
MetaboQuant does not perform peak picking, fitting, or integration on NMR spectra, steps for which other software such as the proprietary Amix (Bruker BioSpin, Rheinstetten, Germany) or the free ACD/NMR Processor Academic Edition (www.acdlabs.com/resources/freeware/nmr_proc) is available.
MetaboQuant offers a novel approach for the accurate quantification of metabolites from NMR data by using individual calibration factors and by automatically excluding unreliable and overlapped signals. Materials and methods Test data
For performance tests, a latin square spike-in experiment was prepared where the sum of the molar amounts of spike-ins added to aliquots of a biological specimen is the same (Table 1). Eight metabolites commonly found in human urine, namely alanine, creatinine, histidine, betaine, acetate, phosphocreatine, β-hydroxybutyrate (BHBA), and lactate (all purchased from Sigma- Aldrich, Steinheim, Germany) were added in varying amounts to aliquots of a urine specimen collected from a healthy volunteer. Briefly, each added metabolite was dissolved in pure water and these solutions were diluted to yield eight different concentrations of each metabolite. The urine specimen was split into eight 200 µL aliquots and 25 µL each of the previously prepared metabolite solutions (all with differentconcentrations) were added, yielding a volume of 400 µL.
le> Sample preparation and NMR measurements
Two hundred µL of 0.1 M phosphate buffer (pH 7.4) and 50 µL deuterium oxide containing 0.75% (w/v) sodium 3-trimethylsilyl-2,2,3,3-tetradeuteropropionate (TSP) as a reference substance (Sigma-Aldrich) were added to 400 µL of spiked urine. 1D 1H NOESY spectra and 1H-13C-HSQC spectra were recorded on a 600 MHz Avance III spectrometer (Bruker BioSpin) employing a triple-resonance cryoprobe and z-gradients. For HSQC spectra, 128 increments were recorded. The number of increments was doubled using complex linear prediction and then zero-filled to a total of 1024. All other measurement and processing details were performed as described previously (3). As the addition of acidic substances may cause changes in pH and, consequently, chemical shifts, sample pH was assessed using the citrate signal at 2.62 ppm in the 1D spectra (8). Mean and standard deviation of the calculated pH were 7.13 ± 0.01, indicating minimal changes in chemical shift. For details on pH and TSP calculations, see the Supplementary Material.
Peaks were picked and integrated using Amix 3.9.13. Spectral positions of the peaks are given in Supplemental Table S1. Raw and processed NMR data are available at http://genomics.uni-regensburg.de/site/institute/software/NMR/Latin_Square.zip. Statistical analyses were performed using Excel 2007 (Microsoft, Redmond, WA). Software development
MetaboQuant was developed using Matlab 2007b (MathWorks, Natick, MA). Executable files for Windows (Microsoft), a Matlab script for use on other operating systems and a detailed manual are available at http://genomics.uni-regensburg.de/site/institute/software/metaboquant. MetaboQuant is accessed via an intuitive graphical user interface (Figure 1) providing access to all parameters and files from one program window.
Results and discussion
Functions and data flows within MetaboQuant are visualized in Figure 2A following a structured analysis design (9). Each circle in Figure 2A corresponds to one panel in the main program window (Figure 1). In the following paragraphs, the work-flow of MetaboQuant is described briefly.
A file containing annotated peak integrals is used as data input. The file may be a tab delimited text file, an Excel file, or an Amix output file. Adjusting integrals of individual signals
First, peak integrals are divided by the number of nuclei contributing to each peak. The number of nuclei is taken from a peak information file provided by the user. A peak integral may be scaled to a reference substance and given as the fraction of the integral (FI) of the reference substance. The FIs may then be multiplied by a correction factor, for instance for differences in reference substance concentration across samples.
Next, each FI may be scaled by an individual calibration factor taken from the peak information file. A set of calibration factors for commonly observed compounds is included in the software package. However, these values are only applicable to measurements performed using exactly the same measurement and sample parameters. The parameters used can be found at http://genomics.uni-regensburg.de/site/institute/software/NMR/Latin_Square.zip in the files acqus and acqu2s (2D data) attached to each spectrum and may be accessed most conveniently by opening the respective spectra within TopSpin (Bruker BioSpin). In general, it is recommended that users create their own peak calibration factors. A detailed description on how to create such factors is given in the MetaboQuant manual. Alternatively, in the absence of individual calibration factors, each FI may be multiplied by the reference substance concentration to obtain semiquantitative values. Reliability checks
As most compounds give rise to more than one NMR signal, several values are available for a given compound. To obtain the concentration, one may calculate an average over all values for the compound. However, individual signals may be compromised by overlap with other signals, artifacts, or various other sources of error. Therefore, it is important to check the reliabilityof each signal belonging to a compound. The next bundle of algorithms (the third circle in Figure 2A) contains several checks for this purpose. Details are shown in Figure 2B as an UML2 activity diagram (10).
For a given sample matrix, some of the observed signals are prone to overlap. Therefore, the user can specify the signals to be used for quantification of each compound. Imagine a metabolite signal overlapped by another signal of low intensity. If the concentration of the compound in question is high, this may not pose a problem, as the low-intensity overlap will not change the peak integral too much. Conversely, when the compound's concentration is too low to yield a visible NMR signal, the overlapping signal might still be visible, leading to an incorrect interpretation of the results. Consequently, a presumably non-overlapped signal of this compound may be marked as obligatory. If this signal is not observed in the spectrum because it is below the noise level, the compound is marked as not detected. Obligatory peaks have previously been used for analysis of NMR spectra but only for compound identification, not quantification (11). The obligatory peak feature, however, requires experience and specific knowledge about single compounds and, therefore, is reinforced by more generally applicable reliability checks.
In the case of low-abundance compounds, some of the expected NMR signals may not be visible in the spectrum. To ensure a high likelihood of correct quantification, the user may define a threshold for the ratio of observed to expected signals (set by default to 0.66), below which compounds are not quantified. For a compound with three signals, at least two signals have to be visible. However, in the presence of signals stemming from multiple nuclei, such as those generated by methyl groups, quantification results for low-abundance compounds may still be accurate even if fewer than the defined number of signals is observed. In cases where a compound has been excluded for not meeting the threshold, an additional reliability check should be performed. This check is based on the assumption that for low concentrations signals with many contributing nuclei will be visible, while peaks with fewer nuclei will remain below the noise level. For all visible signals of a compound, the number of contributing nuclei is taken from the peak information file and the maximum number observed (MNO) is stored. The same procedure is applied to non-visible signals and the maximum number of atoms contributing to a non-observed signal (MNN) is stored. In case MNO exceeds MNN, the observed signals of this compound are considered further. Otherwise, the compound concentration is set to zero. This reliability check is, to the best of our knowledge, a novel feature in NMR quantification. In some circumstances, the reliability check might cause quantification errors, such as for substances like glycerophosphocholine where strong peaks are prone to overlap. In such cases, the overlapped peaks should be excluded from quantification or a weaker peak should be marked as “obligatory” as described above.
Peak integrals may be compromised by overlap or errors in peak picking. When comparing several peak integrals for a compound, compromised peaks will differ strongly from the other peaks. Peaks are therefore excluded as outliers from quantification if their integrals differ by more than a chosen threshold from the median of all peaks obtained for a compound. The default threshold is 40%. Both the threshold for missing peaks and the outlier threshold were manually optimized on a separate biological data set in a trade-off between excluding as many unreliable values and keeping as many reliable values as possible.
Following removal of outliers and questionable peaks, the mean of the remaining peaks is calculated to yield the concentration. If only a single peak is left after outlier removal, the median of all observed peaks is chosen as the concentration value. For compounds with only two available signals showing deviations above the threshold, the smaller peak is chosen since overlap usually results in increased integrals.
Optionally, the resulting concentration values may be compared with lower limits of quantification taken from a file provided by the user and values below these limits are discarded. Scaling and normalization of calculated concentrations
Dilution factors may be used to correct concentration values in those instances where limited amounts of specimen require dilution to achieve the volume needed for NMR measurements. This correction is achieved either by multiplying all values with the same dilution factor or by using individual dilution factors provided in the spectrum title.
The obtained quantitative values may be normalized to a chosen compound (i.e., divided by the concentration of this compound). For replicate samples, means and technical errors may be calculated. Replicate samples have to be marked in the spectrum title. Report of results
The calculated concentrations are saved to an output file and the values of all excluded peaks are stored for manual inspection. All parameters, file names, and the program version are stored in the results file. These data allow the exact reproduction of the quantification procedure at a later time point and enable reproducible results. Evaluation
To test the performance of MetaboQuant, a latin square spike-in experiment was performed. First, 1H-13C HSQC spectra were generated as these generally yield excellent quantification results upon correction of each peak integral with an individual calibration factor (3). If the quantification works well, a slope of 1 should be observed for each compound when performing a linear regression of measured concentrations against spiked-in concentrations. In addition, the standard deviation of the 8 slopes should be close to 0.
For comparison, quantification was performed using both Amix 3.9.13 and MetaboQuant 1.2 (using peak integrals calculated by Amix). Five different quantification approaches were used, as listed in Table 2, and the results are presented and discussed in the following paragraphs. For a list of the retrieved concentration values, see Supplemental Table S2.
le> Quantitative results using Amix
To retrieve absolute concentrations, the mean of the peak integrals for a compound calculated by Amix has to be manually multiplied by the reference concentration. When using the exact reference concentration to calculate quantitative results, the mean of the observed slopes deviates significantly from the expected slope of 1 (Table 2). In addition, the slopes vary strongly as indicated by the high standard deviation. For most compounds, this indicates a poor correlation between spiked-in and measured concentrations. This deviation is explained by the fact that in HSQC spectra signal intensities are influenced by compound-dependent factors such as different INEPT transfer efficacies. The considerable deviations between measured and expected values are also evidenced by the Bland-Altman plot (12) shown in Figure 3A.
As noted above, an individual calibration factor may be calculated for each C-H group. The calibration factors of the eight spiked-in metabolites were averaged to obtain a mean calibration factor. As expected, using a mean calibration factor brings the slopes of the regression lines (Table 2) closer to 1 with a relative standard deviation of 14.2%. The Bland-Altman plot (Figure 3B) shows improved quantification results as compared with the Amix exact reference approach. However, a trumpet-shaped distribution is evident, indicating high variances for large concentrations. The correlation of variance to metabolite concentration value is a well-known phenomenon of NMR data (13). Of 64 possible concentration values (8 compounds in 8 samples), 48 values were determined. Amix automatically excluded compounds for which one or more peaks were not found in the spectrum, affecting mostly histidine. Due to a bad line shape for the histidine peak at 8.01/137.9 ppm (1H/13C), Amix did not integrate this peak in 5 out of 8 spectra. In the remaining spectra, the peak was integrated but with drastically decreased intensities. Quantitative results using MetaboQuant
The basic mode of MetaboQuant offers several outlier detection algorithms, but requires no compound-specific values. When using the basic mode in combination with a mean calibration factor, 59 values were calculated, yielding a mean slope close to 1 with a relative standard deviation of 16.5% (Table 2). The resulting Bland-Altman plot (Figure 3C) is similar to that obtained with Amix using a mean calibration factor, although more concentration values were obtained. The four lowest BHBA values and the lowest acetate value were not determined. For the lowest concentration of BHBA and acetate, no peaks were observed above noise level. The three other missing BHBA values were excluded as the number of observed peaks was below the threshold. The histidine peak at 8.01/137.9 ppm was automatically excluded in all samples by outlier detection. Thus, quantitative results rested on the remaining histidine signals. It is not surprising that the standard deviation is higher in this case than in the Amix results mentioned above, as more low-abundance compounds were quantified.
Next, the advanced mode of MetaboQuant without individual lower limits of quantification was applied. This mode includes scaling peak integrals by individual calibration factors and reliability checks. The number of calculated concentration values rose to 62 and the relative standard deviation of the calculated slopes (Table 2) was 6.87%. The Bland-Altman plot in Figure 3D shows considerably smaller deviations between measured and expected values when compared with the previously discussed approaches. Additionally, the trumpet-shape (indicating stronger deviations) starts at higher concentration levels and is less pronounced than for the basic mode and Amix, showing improved quantification results. Figure 3D shows that all strongly deviating values were over-quantified. This is probably the result of peak overlap still present in the data even after employing all reliability checks. The lowest concentration values of acetate and BHBA were missing since no peaks were observed above the noise level. The standard deviation is lower compared with both Amix and basic mode MetaboQuant results, indicating high reliability of the quantification results.
Finally, concentration values falling below predefined individual lower limits of quantification were excluded automatically. This yielded 61 concentration values, excluding the lowest acetate value and the 2 lowest BHBA values. The other performance values (Table 2) and the Bland-Altman plot (Figure 3E) were similar to advanced mode without checking quantification limits.
The fact that the reliability checks removed badly integrated signals of histidine indicates that the routines are robust in excluding outliers due to both signal overlap (causing increased integrals) and technical reasons such as imperfect peak picking or integration (causing decreased integrals).
Quantification was repeated on 1D 1H spectra of the same samples, using Amix peak shape analysis. Results are similar to the results for HSQC spectra (Supplemental Tables S3 and S4; Supplemental Figure 1). The slightly worse performance is probably due to the fact that 1D 1H spectra are more prone to peak overlap than HSQC spectra. Using line shape analysis methods reduces this issue to some extent. Therefore, we recommend line shape analysis to create peak integrals for use in MetaboQuant.
As a note of caution, a thorough investigation of all peaks used for quantification should be performed on test samples to identify peaks that are prone to creating non-reliable values. Such peaks should be excluded from quantitative analysis. When setting up a quantification protocol, comparison to an orthogonalquantification platform such as GC-MS or LC-MS is strongly recommended for result validation.
In conclusion, MetaboQuant enables reliable quantification from 1D 1H and 2D 1H-13C HSQC spectra. Basic versions of MetaboQuant have been successfully applied to metabolite quantification in urine, blood, plasma, milk, and tissue extracts (14-18).
This work was funded by the MeGA-M and the NGFN Leukemia grants of the German Federal Ministry of Education and Research, the Bavarian Genome Network BayGene and by the German Research Foundation DFG in context of the clinical research unit KFO 262.
The authors thank Helena Zacharias, Ann-Kathrin Immervoll, and Caridad Louis for helpful remarks on using the software and Margit Gratzl for helpful discussions on user-friendliness and stability.
The authors declare no competing interests.
Address correspondence to Wolfram Gronwald, Institute of Functional Genomics, Regensburg, Germany. Email: [email protected]
1.) Nicholson, J.K., and J.C. Lindon. 2008. Systems biology: metabonomics. Nature 455:1054-1056. 2.) Weljie, A.M., J. Newton, P. Mercier, E. Carlson, and C.M. Slupsky. 2006. Targeted profiling: quantitative analysis of 1H NMR metabolomics data. Anal. Chem. 78:4430-4442. 3.) Gronwald, W., M.S. Klein, H. Kaspar, S.R. Fagerer, N. Nürnberger, K. Dettmer, T. Bertsch, and P.J. Oefner. 2008. Urinary metabolite quantification employing 2D NMR spectroscopy. Anal. Chem. 80:9288-9297. 4.) Mierisová, S., and M. Ala-Korpela. 2001. MR spectroscopy quantitation: a review of frequency domain methods. NMR Biomed. 14:247-259. 5.) Lewis, I.A., S.C. Schommer, B. Hodis, K.A. Robb, M. Tonelli, W.M. Westler, M.R. Sussman, and J.L. Markley. 2007. Method for determining molar concentrations of metabolites in complex solutions from two-dimensional 1H-13C NMR spectra. Anal. Chem. 79:9385-9390. 6.) Rai, R.K., P. Tripathi, and N. Sinha. 2009. Quantification of metabolites from two-dimensional nuclear magnetic resonance spectroscopy: application to human urine samples. Anal. Chem. 81:10232-10238. 7.) Hu, K., W.M. Westler, and J.L. Markley. 2011. Simultaneous quantification and identification of individual chemicals in metabolite mixtures by two-dimensional extrapolated time-zero (1)H-(13)C HSQC (HSQC(0)). J. Am. Chem. Soc. 133:1662-1665. 8.) Miyataka, H., T. Ozaki, and S. Himeno. 2007. Effect of pH on 1H-NMR spectroscopy of mouse urine. Biol. Pharm. Bull. 30:667-670. 9.) DeMarco, T. 1979. Structured Analysis and System Specification. Prentice Hall, Upper Saddle River, NJ. 10.) Rumbaugh, J., I. Jacobson, and G. Booch. 2004. Unified Modeling Language Reference Manual, 2nd Ed. Pearson Higher Education. 11.) Xia, J., T. Bjorndahl, P. Tang, and D. Wishart. 2008. MetaboMiner - semi-automated identification of metabolites from 2D NMR spectra of complex biofluids. BMC Bioinformatics 9:507. 12.) Bland, J.M., and D.G. Altman. 1986. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1:307-310. 13.) Kohl, S.M., M.S. Klein, J. Hochrein, P.J. Oefner, R. Spang, and W. Gronwald. 2012. State-of-the art data normalization methods improve NMR-based metabolomic analysis. Metabolomics 8:146-160. 14.) Klein, M.S., M.F. Almstetter, G. Schlamberger, N. Nürnberger, K. Dettmer, P.J. Oefner, H.H.D. Meyer, S. Wiedemann, and W. Gronwald. 2010. Nuclear magnetic resonance and mass spectrometry-based milk metabolomics in dairy cows during early and late lactation. J. Dairy Sci. 93:1539-1550. 15.) Klein, M.S., C. Dorn, M. Saugspier, C. Hellerbrand, P.J. Oefner, and W. Gronwald. 2011. Discrimination of steatosis and NASH in mice using nuclear magnetic resonance spectroscopy. Metabolomics 7:237-246. 16.) Gronwald, W., M.S. Klein, R. Zeltner, B.-D. Schulze, S.W. Reinhold, M. Deutschmann, A.-K. Immervoll, C.A. Böger. 2011. Detection of autosomal dominant polycystic kidney disease by NMR spectroscopic fingerprinting of urine. Kidney Int. 79:1244-1253. 17.) Klein, M.S., N. Buttchereit, S.P. Miemczyk, A.-K. Immervoll, C. Louis, S. Wiedemann, W. Junge, G. Thaller. 2012. NMR metabolomic analysis of dairy cows reveals milk glycerophosphocholine to phosphocholine ratio as prognostic biomarker for risk of ketosis. J. Proteome Res. 11:1373-1381. 18.) Zacharias, H., G. Schley, J. Hochrein, M.S. Klein, C. Köberle, K.-U. Eckardt, C. Willam, P.J. Oefner, and W. Gronwald. 2012. Analysis of human urine reveals metabolic changes related to the development of acute kidney injury following cardiac surgery. Metabolomics Online First Article.