Combining NMR and LC/MS Using Backward Variable Elimination: Metabolomics Analysis of Colorectal Cancer, Polyps, and Healthy Controls

Abstract

Both nuclear magnetic resonance (NMR) spectroscopy and mass spectrometry (MS) play important roles in metabolomics. The complementary features of NMR and MS make their combination very attractive; however, currently the vast majority of metabolomics studies use either NMR or MS separately, and variable selection that combines NMR and MS for biomarker identification and statistical modeling is still not well developed. In this study focused on methodology, we developed a backward variable elimination partial least-squares discriminant analysis algorithm embedded with Monte Carlo cross validation (MCCV-BVE-PLSDA), to combine NMR and targeted liquid chromatography (LC)/MS data. Using the metabolomics analysis of serum for the detection of colorectal cancer (CRC) and polyps as an example, we demonstrate that variable selection is vitally important in combining NMR and MS data. The combined approach was better than using NMR or LC/MS data alone in providing significantly improved predictive accuracy in all the pairwise comparisons among CRC, polyps, and healthy controls. Using this approach, we selected a subset of metabolites responsible for the improved separation for each pairwise comparison, and we achieved a comprehensive profile of altered metabolite levels, including those in glycolysis, the TCA cycle, amino acid metabolism, and other pathways that were related to CRC and polyps. MCCV-BVE-PLSDA is straightforward, easy to implement, and highly useful for studying the contribution of each individual variable to multivariate statistical models. On the basis of these results, we recommend using an appropriate variable selection step, such as MCCV-BVE-PLSDA, when analyzing data from multiple analytical platforms to obtain improved statistical performance and a more accurate biological interpretation, especially for biomarker discovery. Importantly, the approach described here is relatively universal and can be easily expanded for combination with other analytical technologies.

Graphical abstract

Metabolomics provides an important approach in systems biology to investigate biological states as well as the effects of internal and external perturbations through the study of changes in metabolite concentrations and fluxes.^1–9 Complex metabolic processes in living systems respond to many stimuli, including diseases and drugs, resulting in alterations in metabolic profiles; metabolomics aims to detect these changes at the molecular level using advanced analytical chemistry techniques and multivariate statistical analysis. Metabolomics studies have resulted in a number of important findings, including a deeper understanding of cancer metabolism^10,11 and drug toxicity,^12,13 the potential for improved early disease detection^14–16 or therapy monitoring,^4,17 and successful applications in environmental science,¹⁸ nutrition,¹⁹ etc.

The two most commonly used analytical technologies in metabolomics are nuclear magnetic resonance (NMR) spectroscopy and mass spectrometry (MS).^20,21 NMR is well-known as a premier method for structural identification and for the analysis of multicomponent mixtures as it is rapid and nondestructive, requires little or no sample preparation, and provides highly reproducible (coefficients of variation, CVs, of a few percent) and quantitative results.^22–25 MS is another essential method for identifying and quantifying metabolites, especially those of low abundance in complex biosamples, because of its intrinsically high sensitivity and high selectivity.^26,27 Notably, metabolomics data from NMR and MS experiments are complex because they usually contain signals from many metabolites; therefore, multivariate statistical analysis plays an important role in metabolomics for reducing data dimensions, differentiating similar spectra, building predictive models, etc.^28,29

NMR and MS generate different metabolic profiles from the same sample; thus, their combination can be very valuable in metabolomics. One thereby can obtain a more comprehensive profile of detectable metabolites, and as shown below, one can potentially improve the reliability and predictive accuracy of statistical models. However, currently the vast majority of metabolomics studies use either NMR or MS separately, although a growing number of studies do combine NMR and MS analysis to advantage.^30–44 For example, we developed principal component (PC)-directed partial least-squares (PLS) analysis to combine one-dimensional (1D) ¹H NMR and direct analysis in real time (DART) MS data to improve breast cancer detection.³¹ Powers and co-workers and Karaman et al. have proposed multiblock PLS approaches to integrate data from MS and NMR.^43,44 Thus far, however, the potential benefits of combining NMR and MS for biomarker discovery and statistical modeling are still not well recognized. In particular, optimizing variable selection is one of the major challenges for multiblock data because of the extra number of variables. Variable selection is often performed using univariate analysis or as a byproduct of full-scale multivariate statistical analysis;^32,43 the contribution of each individual variable to multivariate modeling is rarely studied. Therefore, it is highly desirable to further investigate and develop approaches in metabolomics to make better use of both NMR and MS data.

In this study focused on method development, we examined the performance of combining ¹H NMR and targeted LC/MS/MS metabolite profiles from patients with colorectal cancer (CRC), patients with polyps, and healthy controls. It is important to improve CRC and polyp detection, as CRC is one of the most prevalent and deadly cancers in the United States and worldwide.⁴⁵ To date, numerous metabolic alterations have been found in CRC tissue,^37,46–48 serum,^49–55 urine,⁵⁶ and fecal water⁵⁷ from the metabolite profiles measured by NMR and/or MS, and therefore, one might surmise that combining NMR and MS data could result in improved metabolite panels. Here, we analyzed a total of 127 serum samples from three groups of subjects, and potential biomarkers were selected using a backward variable elimination⁵⁸ approach that was incorporated into multiblock PLS discriminant analysis (BVE-PLSDA). Monte Carlo cross validation (MCCV)^59,60 was performed and demonstrated the robust diagnostic power of this NMR-and MS-based metabolomics approach in differentiating healthy controls from patients with CRC and polyps.

EXPERIMENTAL PROCEDURES

Chemicals

Deuterium oxide (D₂O, 99.9% D), L-tyrosine-¹³C₂, and sodium L-lactate-¹³C₃ were purchased from Cambridge Isotope Laboratories, Inc. (Andover, MA). Trimethylsilylpropionic-2,2,3,3-d₄ acid sodium salt (TSP) was obtained from Sigma-Aldrich (Milwaukee, WI). Acetonitrile (LC/MS grade), ammonium acetate (LC/MS grade), methanol (LC/MS grade), and acetic acid (LC/MS grade) were all purchased from Fisher Scientific (Pittsburgh, PA). Deionized (DI) water was provided in house by a Synergy Ultrapure Water System from EMD Millipore (Billerica, MA).

Serum Samples

All work was conducted in accordance with protocols approved by the Indiana University School of Medicine and Purdue University Institutional Review Boards. All subjects in the study provided informed consent according to institutional guidelines. Patients undergoing colonoscopy for CRC screening or CRC surgery were evaluated, and blood from the patients was obtained after overnight fasting and identical bowel preparation but prior to their procedure (colonoscopy or surgery). Healthy or polyp status was determined after screening. In total, blood samples from 28 CRC patients, 44 individuals with polyps, and 55 healthy controls were collected and analyzed by NMR and MS. The detailed demographic and clinical parameters for the patients and healthy controls included in this study are listed in Table 1. Each blood sample was allowed to clot for 45 min and then centrifuged at 2000 rpm for 10 min. Sera were collected, aliquoted into separate vials, transported under dry ice, and then stored at −80 °C until they were used.

Table 1.

Summary of Demographic and Clinical Information of Subjects Recruited and Analyzed for This Study

	healthy controls	polyps	CRC
no. of samples (no. of patients)	55 (55)	44 (44)	28 (28)
age, mean (range)	52.8 (21–74)	56.1 (39–68)	55.3 (27–86)
gender
male	25	23	16
female	30	21	12
no. of polyps
0–5	–	26	–
6–9	–	4	–
10–15	–	3	–
≥16	–	11	–
cancer stage
stage I/II	–	–	3
stage III	–	–	8
stage IV	–	–	17
diagnosis
colon cancer		–	19
rectal cancer		–	9
ethnicity
Caucasian	50	23	15
African American	4	1	2
Asian	0	1	0
NAa	1	19	11

^aInformation not available.

¹H NMR Spectroscopy

Five hundred thirty microliters of each serum sample was transferred into a 5 mm NMR tube. To minimize the interference of macromolecules with the TSP signal, a 60 μL TSP solution (20.9 nmol in D₂O) sealed in a coaxial glass capillary that served as a chemical shift reference (δ 0.00) was placed in the NMR tube during the experiment. All ¹H NMR experiments were performed at 25 °C on a Bruker DRX 500 MHz NMR spectrometer equipped with a TXI probe. ¹H NMR data for each sample were acquired using the one-dimensional CPMG (Carr–Purcell–Meiboom–Gill) pulse sequence with water presaturation. For each spectrum, 128 transients were collected and 16K data points were acquired using a spectral width of 6000 Hz. The data were Fourier transformed after an exponential weighting function corresponding to 0.5 Hz line broadening had been applied. The spectra were phased and baseline corrected using Bruker TopSpin software (version 3.5). The spectral data were deposited into The Metabolomics Consortium Data Repository and Coordinating Center (DRCC, Study ST000285).⁶¹ We performed manual metabolite identification and peak integration (based on characteristic chemical shift and multiplicity). Table S1 shows the list of metabolites detected by NMR and their chemical shift regions used for peak integration.^23,24

LC/MS/MS

A robust targeted LC/MS/MS method has recently been developed and used in a number of studies in the Northwest Metabolomics Research Center (NW-MRC).^{11,15,53,62–67} In this study, the MS data from the same samples as those measured by NMR, collected in our previous study,¹⁵ were used to demonstrate the performance of the NMR–MS combination. Briefly, the LC/MS/MS experiments were performed on an Agilent 1260 LC (Agilent Technologies, Santa Clara, CA) AB Sciex QTrap 5500 MS (AB Sciex, Toronto, ON) system. Table S2 shows the LC gradient conditions. We monitored 99 and 59 MRM transitions in negative and positive mode, respectively (158 transitions in total). The optimized MS conditions are listed in Table S3. The extracted MRM peaks were integrated using MultiQuant version 2.1 (AB Sciex). The LC/MS/MS data were also deposited into the DRCC (Study ST000284).⁶¹

Data Analysis

A TSP solution of known concentration contained in a separate glass capillary was used to provide a chemical shift reference and to normalize the NMR data. In this study, we used the same MS data that have been investigated in a previous study;¹⁵ however, here we analyzed the samples that were measured by both NMR and MS. In the published study,¹⁵ we observed that normalization using the QC data on a metabolite-by-metabolite basis reduced technical variation in the data. For example, we obtained a median QC CV of 8%, ranging from 5 to 31% with ~80% metabolites having a CV of <15% after QC normalization. Both NMR and MS data were autoscaled prior to multivariate statistical analysis.

Figure 1 shows the flowchart of MCCV-BVE-PLSDA, describing the process of identifying a mixed panel of NMR and MS markers to improve CRC and polyp detection. We compared the performance of NMR and MS data, both individually and in combination (NMR–MS). On the basis of the concept of BVE,⁵⁸ we developed a MCCV-BVE-PLSDA algorithm to obtain the variable subset that generated improved prediction. The analysis started by including all metabolite signals (variables) and stopped once every variable had been examined. In each iteration, one variable (metabolite) was removed, and the remaining variables were used for PLS-DA. The variables with the highest prediction accuracy for the test samples in MCCV^59,60 were kept for the next iteration. During MCCV (which was repeated 500 times for each variable), all samples were randomly divided into two sets, 70% as the training set and 30% as the test set. PLS-DA was performed on the training set, and then the resulting model was used to predict the classification of test set samples. The sample membership could be either correctly assigned, termed true class, or randomly assigned (permutation). Several metrics were used to estimate and compare the performance of the models. First, classification accuracy was used to measure the performance; for example, if 1 was assigned as the dummy index for the CRC group and 0 was used for the control group, 0.5 would be used as the threshold to determine the classification for both the training and test samples. Classification accuracy was calculated as the ratio between the number of correctly predicted samples and the total number of samples, which takes both false positives and false negatives into consideration. In addition, the area under the receiver operator characteristics curve (AUROC) was also used as a performance metric.

Figure 1.

Flowchart indicating the steps for MCCV-BVE-PLSDA.

All analyses were conducted in Matlab (R2008a, Mathworks, Natick, MA) installed with the PLS Toolbox (version 4.11, Eigenvector Research, Inc., Wenatchee, WA) using scripts written in house.

RESULTS AND DISCUSSION

¹H NMR and MS Spectra

Figure S1a shows a typical 500 MHz 1D ¹H NMR spectrum of a serum sample from a colon cancer patient. The spectrum contained NMR signals from small molecules, including formate, phenylalanine, tyrosine, histidine, glucose, lactate, creatinine, glutamine, alanine, valine, isoleucine, leucine, etc. In total, 70 metabolite variables (Table S1) were obtained from the ¹H NMR spectra by manual peak integration.

The targeted LC/MS/MS system allowed the detection of 158 (including two ¹³C-labeled internal standards) MRM transitions (Table S3), for the metabolites from more than 20 different metabolite classes (e.g., amino acids, carboxylic acids, pyridines, etc.) and from >25 important metabolic pathways (e.g., TCA cycle, amino acid metabolism, glycolysis, etc.). Panels b and c of Figure S1 show the overlapped extracted ion chromatograms (EICs) of the metabolites that were detected in the serum sample from a colon cancer patient with positive and negative ionization, respectively. In total, we detected 113 metabolites that were present in the serum samples, and they are highlighted in the ID column of Table S3.

It can be clearly seen that NMR and MS metabolic profiles have some overlap (Tables S1 and S3), such as lactate and glucose that were detected on both instruments. Meanwhile, many metabolites were detected only on one platform or the other; for example, citrate could be detected by only NMR, and reduced glutathione was measured by LC/MS/MS alone. In Figure 2a, we show the correlation between all the NMR and MS variables, while Figure 2b presents the correlation between the subset of metabolites that can be detected by both NMR and MS. Many metabolites had low correlation values between NMR and MS in Figure 2a; however, most of the overlapped metabolites had large correlation coefficients (Figure 2b). Nevertheless, a few metabolites in Figure 2b had weak correlations, which are probably in part due to the presence of matrix effects in the MS data or peak overlap in the NMR spectra.

Figure 2.

(a) Correlation between all NMR and MS variables. (b) Correlation between the subset of metabolites (labeled in the figure) that can be detected by both NMR and MS. The X axis provides an index of all NMR variables in the data matrix, and the Y axis provides an index of all MS variables in the data matrix. The metabolites are listed in Tables S1 (NMR) and S3 (MS, highlighted).

CRC versus Healthy Controls

Panels a and b of Figure 3 show the results of MCCV-BVE-PLSDA in selecting a subset of metabolite markers for differentiating CRC patients from healthy controls, using NMR, MS, and NMR–MS data. As shown in Figure 3a, the highest classification accuracy of the NMR–MS data was clearly better than that using the models derived from either NMR or MS alone. As could be anticipated, an excessive number of variables led to the deterioration of statistical models, and there was a number and/or range of variables that could produce the best statistical performance. This was consistently observed for all the NMR-, MS-, and NMR–MS-based models.

Figure 3.

(a) Results of MCCV-BVE-PLSDA for selecting a subset of variables to differentiate CRC from healthy controls, using NMR (solid red line), MS (dashed green line), and NMR–MS (dotted blue line) data. (b) Classification accuracy of true class (red dots) and random permutation (blue squares) models in MCCV, when the selected set of variables was used for MCCV-BVE-PLSDA modeling in panel a. (c) Results of MCCV-BVE-PLSDA for selecting a subset of variables to differentiate healthy controls from polyp patients, using NMR (solid red line), MS (dashed green line), and NMR–MS (dotted blue line) data. (d) Classification accuracy of true class (red dots) and random permutation (blue squares) models in MCCV, when the selected set of variables was used for MCCV-BVE-PLSDA modeling in panel c.

Table 2 summarizes the selected sets of metabolites resulting from MCCV-BVE-PLSDA and their statistical performance in the pairwise comparisons among CRC, polyps, and healthy controls. In the case of CRC versus healthy controls, a set of seven NMR variables provided the best classification accuracy of 0.84 ± 0.07, compared to 0.71 ± 0.08 for all 70 variables. The MS data generated a classification accuracy of 0.93 ± 0.05 using 19 variables compared to a value of 0.80 ± 0.07 for all 113 variables. Interestingly, it was observed that simply putting the NMR and MS data together does not guarantee better statistical performance (0.79 ± 0.08 for all 183 variables), because too many poorly performing variables will reduce prediction accuracy. However, after MCCV-BVE-PLSDA using the combined set of NMR–MS data, the highest classification accuracy of 0.95 ± 0.05 was achieved using 31 variables. The complementary information provided by NMR and MS was beneficial in improving the statistical analysis. Therefore, we recommend incorporating appropriate variable selection in multivariate statistical analysis to minimize data redundancy, a step that is currently not often performed.

Table 2.

Summary of the Subsets of Metabolites and Their Statistical Performance in the Pairwise Comparisons among CRC, Polyps, and Healthy Controls classification accuracy^a

samples	data set	total no. of variables	no. of MCCV-BVE-PLSDA-selected variables	all variables	classification accuracya
					subset selected by MCCV-BVE-PLSDA
CRC vs controls	NMR	70	7	0.71(0.08)	0.84(0.07)
	MS	113	19	0.80(0.07)	0.93(0.05)
	NMR–MS	183	31	0.79(0.08)	0.95(0.05)
CRC vs polyps	NMR	70	11	0.72(0.08)	0.83(0.07)
	MS	113	21	0.82(0.07)	0.95(0.04)
	NMR–MS	183	30	0.82(0.08)	0.98(0.02)
polyps vs controls	NMR	70	3	0.55(0.07)	0.67(0.08)
	MS	113	6	0.56(0.07)	0.71(0.07)
	NMR–MS	183	13	0.57(0.08)	0.74(0.07)

^aThe numbers in parentheses are the standard deviation values of the classification accuracy values from iteration calculations.

Figure 3b compares the classification accuracy of true class models and random permutations in MCCV, when the selected set of variables was used in MCCV-BVE-PLSDA. As expected, the average classification accuracy of random permutations was very close to 0.5, regardless of whether NMR, MS, or NMR–MS data were used. The classification accuracy values of true class models were clearly higher than those of random permutations, which further confirmed that the NMR and/or MS variables did contain variations related to CRC.

Table S4 shows the alterations of selected metabolite markers by MCCV-BVE-PLSDA from the NMR and/or MS data that were involved in the comparison between CRC patients and healthy controls. The metabolite markers from the combined NMR–MS data had a significant overlap with those derived from the models based on NMR or MS data alone. It can be seen that both NMR (seven variables) and MS (24 variables) contributed to the mixed panel of biomarker candidates (NMR–MS). The important NMR and MS metabolites showed no overlap, with the exception of histidine (which decreased in CRC serum for both NMR and MS), providing the evidence that NMR and MS can make unique contributions to statistical modeling in metabolomics. The unpaired Student’s t test was also performed on each metabolite to assess its statistical significance between the two groups. In the NMR–MS data, five NMR variables and nine MS variables had P values of <0.05. We also list the adjusted P values in Table S4, with the false discovery rate (FDR) controlled at 0.05.

CRC versus Polyps

Similarly, we examined this combined NMR–MS metabolomics approach to differentiate CRC from polyp patients. As shown in Figure S2a and Table 2, 11 NMR variables were required to obtain the highest classification accuracy of 0.83 ± 0.07 in MCCV-BVE-PLSDA, and 21 selected MS variables provided a classification accuracy of 0.95 ± 0.04. The combination of 30 NMR and MS variables from the NMR–MS data produced a significantly better classification accuracy of 0.98 ± 0.02. The NMR–MS combination was more efficient in improving CRC and polyp separation compared to NMR or MS alone. It was also seen in Figure S2a that more variables did not necessarily provide better statistical performance, for NMR, MS, or NMR–MS data. Again, the use of variable selection in the statistical analysis of metabolomics data is highly effective in improving the modeling.

The results of MCCV for the selected variables from each data set showed that the average classification accuracy of random permutations (~0.5) was clearly lower than that of the true class models (Figure S2b). This result again indicated that the NMR and/or MS variables extracted a high degree of the biological variation related to CRC or polyps. From Table S5, one can see that the 30 metabolite markers from the NMR–MS data included 11 NMR variables and 19 MS variables. Most of the NMR (MS) variables selected by MCCV-BVE-PLSDA from the NMR–MS data overlapped with those from the NMR (MS) data alone, while a number of important variables were unique to the different data sets.

Polyps versus Healthy Controls

Although of great importance for preventing the development of CRC, the metabolic profiles of polyp patients and healthy controls have not been compared as often as those for CRC.⁶⁸ Differentiating polyp patients from healthy controls is a challenging problem, and diagnostic tests other than colonoscopy generally show poor performance. Using the approach described above, classification accuracies of 0.67 ± 0.08, 0.71 ± 0.07, and 0.74 ± 0.07 were achieved using the selected NMR (three variables), MS (six variables), and NMR–MS (13 variables) data sets, respectively (see Figure 3c and Table 2). Again, the statistical performance of NMR–MS data was significantly better than that of NMR or MS data alone, and variable selection was successful in producing a subset of metabolites that provided the best classification accuracy. The MCCV results shown in Figure 3d indicate that the classification accuracy of true class models was clearly higher than that of random permutations, although the performance is not as good as it is for distinguishing CRC. Table S6 shows the 13 important variables (four from NMR and nine from MS) identified in the NMR–MS data for separating polyps and healthy controls. For example, the level of lipids [NMR (1.209–1.302 ppm)] was increased in polyp patient serum while the level of orotate (MS) was decreased.

Interestingly, a number of important metabolites overlapped in the pairwise comparisons among polyps, CRC, and controls (Tables S4–S6). For example, proline (NMR–MS data) was important in the comparisons of CRC versus controls (Table S4) and CRC versus polyps (Table S5). However, we did not observe a metabolite that was important in the NMR–MS data for all the three pairwise comparisons, although lactate was important in all the MS analyses alone, as indicated in Tables S4–S6. CRC had the lowest adenosine level (NMR–MS data), and polyp patient serum had the highest level of adenosine. The level of orotate was increased in CRC compared to that in polyps, and it was decreased in polyps compared to controls, such that polyp patients had the lowest levels of orotate. The level of this metabolite (and a few others) did not continuously increase or decrease from controls to polyps, and then to CRC, which indicates that CRC disease progression, as reflected in metabolism, is likely a very complex process.

Figure 4 shows the results of MCCV-BVE-PLSDA, but based on AUROC to estimate the classification performance. These results confirmed that variable selection is highly useful for improving multivariate statistical analysis, and the combination of NMR and MS has a better diagnosis performance than NMR or MS alone. While AUROC and classification accuracy are highly correlated, they do not measure performance identically.

Figure 4.

Results of MCCV-BVE-PLSDA (AUROC) in selecting a subset of variables using the NMR (solid red line), MS (dashed green line), and NMR–MS (dotted blue line) data to compare (a) healthy controls vs CRC, (b) polyps vs CRC, and (c) healthy controls vs polyps. MCCV-BVE-PLSDA (AUROC) is the same as MCCV-BVE-PLSDA but uses AUROC values to estimate the classification performance.

Metabolic Pathways

Although detailed biological analysis is beyond the scope of this paper, a number of metabolite changes in important metabolic pathways were observed in this study that are of potential significance to CRC and polyp and are consistent with those reported in previous studies.^15,49–52 These pathways include glycolysis, the TCA cycle, fatty acid metabolism, amino acid metabolism, glutaminolysis, etc. In Figure 5, biomarker candidates discovered from the NMR–MS data are highlighted for the pairwise comparisons among CRC, polyps, and controls. Both NMR (red stars) and MS (blue circles) significantly contribute to the altered metabolism shown in Figure 5, which should be helpful for improving our understanding of metabolite perturbations and the mechanisms related to CRC and polyp development.

Figure 5.

Metabolism pathway diagram showing important metabolites in the NMR–MS data for the pairwise comparisons among CRC, polyps, and healthy controls. The NMR metabolites are highlighted with red stars; the MS metabolites are highlighted with blue circles.

In particular, glucose (detected by NMR) was found to be upregulated in CRC compared to controls (Table S4). This could be due to the need for cancer cells to take up glucose to maintain a high rate of glycolysis, which produces lactate even under aerobic conditions to fulfill the cells’ large demand for carbon substrates.^10,69 In fact, some other glycolysis intermediates, such as PEP (MS) and pyruvate (MS), are also highlighted in Figure 5. Cancer cells also use glutamine as an important energy source (glutamine addiction or glutaminolysis),^10,70 which explains the perturbed glutamine levels (MS) indicated in Figure 5. Amino acid metabolism was significantly impacted by CRC and polyps, as well. For example, changes in alanine (MS), histidine (NMR and MS), aspartate (MS), etc., were emphasized in our statistical modeling. Alterations of amino acid levels can indicate the altered cancer cell activities, e.g., synthesis of proteins or catabolism to provide energy and/or other metabolite substrates. Fumarate (MS), citrate (NMR), and oxaloacetate (MS) are important metabolites in the TCA cycle, and their altered levels (Figure 5) fit well with the hypothesis that the TCA cycle is altered by CRC and polyp formation. Purine metabolism and fatty acid/lipid metabolism changes are also linked to CRC and polyps, based on the significant changes in the levels of adenosine (MS), lipids (NMR), and linolenic acid (MS). It is clear from Figure 5 that both NMR and MS are valuable methods for identifying metabolic changes that occur in patients with CRC and polyps.

Overall, both NMR and MS have advantages and disadvantages as predominant analytical methods in metabolomics, and their combination can make use of their strengths that include NMR’s reproducibility and quantitative nature, along with MS’s high sensitivity and broad coverage. In this study, considering the complementary analytical features of NMR and MS, we believe that leveraging both methods will provide new insights for biomarker discovery and disease diagnosis. Given the large number of detectable metabolites, we recommend using an appropriate variable selection step, such as MCCV-BVE-PLSDA, to extract a useful set of metabolite markers from both the NMR and MS data, instead of simply concatenating them together. This new approach can improve statistical performance and provide more comprehensive biological interpretation. While performing both NMR and MS experiments requires more effort and expense, on the basis of the examples provided in this study, we believe that the benefits outweigh the costs, especially at the biomarker discovery stage.

The aim of this study was not to determine the best variable selection method, but to demonstrate the importance of variable selection, especially in the case of combining NMR and MS data, which is infrequently investigated in metabolomics. MCCV-BVE-PLSDA in this study is an expansion of the BVE-PLSDA approach based on leave-one-out cross validation.⁵⁸ Notably, MCCV-BVE-PLSDA is different from the methods based on a predefined variable ranking list,^71,72 which may lead to filtering out a variable that performs poorly alone but becomes highly useful when combined with other variables. In each iteration of MCCV-BVE-PLSDA, each variable is combined with n – 2 other variables (n is the total number of variables in this iteration) in PLSDA modeling n – 1 times (each time a different variable is excluded from our analysis). We remove the variable (one for each iteration) without which the remaining variables produce the highest prediction accuracy. In addition, we performed regression analysis with a number of previous variable selection methods.^73–77 For example, Figure S3 shows the results of MCCV-BVE-hierarchical PLSDA, and Figure S4 shows the results of the variable importance in the projection (VIP)-based stepwise selection method [MCCV-BVE-PLSDA (VIP) comparing healthy controls vs polyps]. All these results indicated that variable selection is very important in multivariate statistical analysis in picking a subset of variables that provide the best prediction accuracy. These results also show that the combination of NMR and MS exhibits a statistical performance better than that of NMR or MS alone. MCCV-BVE-PLSDA is thus a valuable and complementary approach to previous variation selection methods,^73–77 especially for combining NMR and MS data, and provides a set of significant variables worth further investigation. MCCV-BVE-PLSDA is also straightforward, is easy to implement, and can identify the contribution of each individual variable to multivariate statistical models. Because of the limited number of samples, MCCV was used for internal cross validation in this study; however, it can be easily adapted to external cross validation when a larger number of samples is available. In addition, our data analysis approach is relatively universal and can be expanded to combine other analytical technologies.

CONCLUSIONS

In this study, we developed and applied the MCCV-BVE-PLSDA approach to examine the performance of combining NMR and MS for discovery metabolomics. We profiled serum metabolites from CRC patients, patients with polyps, and healthy controls, which were measured by NMR and LC/MS/MS. MCCV-BVE-PLSDA identified the subsets of metabolites with good diagnostic performance that could be initially validated using MCCV. Further validation will require more samples and would benefit from additional efforts to fully quantify the metabolite biomarkers and verify their robustness, which we are pursuing. Importantly, it was found that the combination of NMR and MS showed a statistical performance better than that of NMR or MS alone. Both NMR and MS contributed significantly to the achievement of a comprehensive biological interpretation for understanding CRC and polyp development mechanisms. Therefore, when possible, we recommend the combined use of NMR and MS along with appropriate variable selection methods in metabolomics, especially for the purpose of discovering biomarker candidates.