Non-invasive fecal metabonomic detection of colorectal cancer

Abstract

Colorectal cancer (CRC) is a major cause of mortality in many developed countries. Effective screening strategies were called for to facilitate timely detection and to promote a better clinical outcome. In this study, the role of fecal metabonomics in the non-invasive detection of CRC was investigated. Gas chromatography/time-of-flight mass spectrometry (GC/TOFMS) was utilized for the metabolic profiling of feces obtained from 11 CRC patients and 10 healthy subjects. Concurrently, matched tumor and normal mucosae surgically excised from CRC patients were profiled. CRC patients were differentiated clearly from healthy subjects based on their fecal metabonomic profiles (orthogonal partial least squares discriminant analysis [OPLS-DA], 1 predictive and 3 Y-orthogonal components, R²X = 0.373, R²Y = 0.995, Q² [cumulative] = 0.215). The robustness of the OPLS-DA model was demonstrated by an area of 1 under the receiver operator characteristic curve. OPLS-DA revealed fecal marker metabolites (e.g., fructose, linoleic acid, and nicotinic acid) that provided novel insights into the tumorigenesis of CRC. Interestingly, a disparate set of CRC-related metabolic aberrations occurred at the tissue level, implying the contribution of processes beyond the direct shedding of tumor cells to the fecal metabotype. In summary, this work established proof-of-principle for GC/TOFMS-based fecal metabonomic detection of CRC and offered new perspectives on the underlying mechanisms.

Keywords: :

• colorectal cancer
• feces
• metabonomics
• metabolomics
• metabolic profiling
• gas chromatography/time-of-flight mass spectrometry

Introduction

Colorectal cancer (CRC) is one of the leading causes of cancer-related death worldwide, accounting for 8% of all cancer mortality.¹ The prognosis of CRC worsens significantly with disease progression, with a drastic decline in 5-y survival rate from 90% in stage I to 12% in stage IV.² Timely detection of CRC by population screening hence becomes a crucial intervention that could reduce CRC mortality.³ However, established screening methods such as colonoscopy or immunochemical-based fecal occult blood test (FOBT) are either invasive or lack sensitivity and specificity.⁴ In recent years, fecal DNA⁵ and mRNA testing⁶ have emerged as potential non-invasive screening tools, but further optimization and validation are required with respect to their clinical diagnostic value. These limitations have fueled ongoing research into complementary fecal-based approaches that would facilitate non-invasive screening for CRC.

While previous studies have explored DNA and mRNA aberrations as potential fecal markers of CRC, a knowledge gap exists regarding the role of post-transcriptional metabolic perturbations. As a result of the dysregulated cellular metabolism and transport that supports their untethered cell growth, cancer cells often carry unique metabolic fingerprints.^7-10 Clinical metabonomic studies comparing matched tumor and healthy tissue from CRC patients have consistently revealed alterations in the levels of amino acids, nucleosides and carbohydrates.^11-17 Besides host-related factors, dysbiosis of gut microbiota is another key feature of CRC. For instance, a lower abundance of Eubacterium rectale and an increased population of Enterococcus faecalis have been observed in the feces of CRC patients.^18,19 Collectively, the repertoire of metabolic derangements derived from both cancer cells and gut microbes may culminate in a distinct metabolic phenotype (metabotype) that characterizes the pathology of CRC. During its direct transit through the gut lumen, it is likely that the feces assimilate this unique metabotype. We hence hypothesized that fecal metabonomics could distinguish CRC patients from healthy subjects and may be applied to the non-invasive characterization of the disease.

Previous studies have demonstrated the potential utility of nuclear magnetic resonance (NMR) spectroscopy in the metabolic profiling of fecal water in CRC patients.^20,21 Notably, low fecal levels of acetate and butyrate were found in these patients. While promising, the inherently low sensitivity of NMR precludes the detection of low-abundance metabolites. To broaden the metabolic coverage, complementary analytical platforms such as gas chromatography/mass spectrometry (GC/MS) could be employed. Importantly, GC/MS offers several advantages in terms of reproducibility, sensitivity and reliable structural identification of metabolites.²² Recently, gas chromatography/time-of-flight mass spectrometry (GC/TOFMS)-based analyses of plasma, serum, urine, and exhaled breath have shown promise in the non-invasive characterization of diabetes,^23,24 hepatocellular carcinoma,²⁵ bladder cancer,²⁶ and asthma.²⁷ These studies showcased the feasibility of GC/TOFMS-based metabonomics in discovering diagnostic biomarkers of various diseases. At the time that our work was performed, no clinical studies have explored the use of GC/TOFMS for the non-invasive fecal metabolic profiling of CRC.

To fill this knowledge gap, we investigated the metabolic composition of feces from CRC patients and healthy subjects and determine if GC/TOFMS-based metabonomics could be a viable approach for the non-invasive detection of the disease. Additionally, tumor specimens and their matched normal mucosae from the same cohort of CRC patients were analyzed and compared with examine the concordance and variation of metabolic alterations at the fecal and tissue levels.

Results

Immunochemical FOBT

Fecal specimens from 3 CRC patients (CRC 4, CRC 6, and CRC 7, Table S1) contained levels of blood equivalent to more than 1 mg Hb/g stool. The remaining patients had occult levels (less than 1 mg Hb/g stool) of blood in their stool except for CRC 3 whose FOBT yielded undetectable level of blood in stool (Table S1). All 10 healthy subjects had undetectable or occult levels (less than 1 mg Hb/g stool) of blood in their stool.

Fecal metabonomic profiling of CRC

Figure 1A shows representative GC/TOFMS chromatograms of fecal specimens from a CRC patient and a healthy subject. After the exclusion of artifactual and irreproducible peaks, 398 well-resolved chromatographic peaks were detected.

Figure 1. Fecal metabonomic profile of CRC patients and healthy subjects. (A) Typical GC/TOFMS chromatograms of fecal specimens from a CRC patient and a healthy subject (Hl). (B) PCA model 2 differentiating CRC patients from healthy subjects (Hl). (C) OPLS-DA model 2 differentiating CRC patients from healthy subjects (Hl). (D) Validation plot of OPLS-DA model 2.

Separate models were constructed with and without the inclusion of patients with blood in stool beyond 1 mg Hb/g stool (models 1 and 2 respectively). The performance statistics of the models were displayed in Table 1. For PCA models 1 and 2, QC samples were clustered closely confirming the analytical validity of the method. This observation also confirmed that the clustering trends observed among the clinical observations (CRC patients vs. healthy subjects) were due to biological rather than analytical variations. While partial discrimination between CRC patients and healthy subjects was observed in the PCA scores plots (see Fig. 1B for an illustration of PCA model 2), both OPLS-DA models clearly distinguished the two groups. As shown in Table 1, OPLS-DA model 1 showed the performance statistics of Q² (cumulative) = 0.215, R²X = 0.373, and R²Y = 0.995 (1 predictive and 3 Y-orthogonal components) and the CRC patients and healthy subjects were classified accurately. On the other hand, OPLS-DA model 2 showed improved performance statistics of Q² (cumulative) = 0.35, R²X = 0.407 and R²Y = 0.996 (1 predictive and 3 Y-orthogonal components) (Table 1; Fig. 1C). The robustness of both OPLS-DA models 1 and 2 was demonstrated by an area of 1 under the ROC curves. Permutation testing further indicated that both models were valid, as exhibited by the negative y-intercept of the Q² regression lines (see Fig. 1D for validation plot of OPLS-DA model 2). Nevertheless, it is important to note that OPLS-DA model 2 showed superior performance in terms of the number of permuted classifications with Q² exceeding that on the actual classification of samples as compared with OPLS-DA model 1 (2 vs. 17, Table 1). A total of 9 differential analytes characterizing feces of CRC patients were uncovered based on OPLS-DA model 2, among which, 3 marker metabolites (fructose, linoleic acid, and nicotinic acid) were putatively identified. All 3 identified metabolites were found at lower levels in the feces of CRC patients compared with healthy subjects (Table 2).

Table 1. Performance statistics of models 1 (all samples included) and 2 (fecal samples containing blood beyond 1 mg Hb/g stool excluded)

Model	No of samples included	Unsupervised PCA			Supervised OPLS-DA
		R^2X (cum)	Q^2 (cum)	Clustering of QC samples	R^2X (cum)	R^2Y (cum)	Q^2 (cum)	AUC of ROC^a	Permutation testing^b
									N (Q^2 permuted > Q^2 original)^c	y intercept^d
1	CRC (n = 11), Hl (n = 10)	0.212	0.139	Yes	0.373	0.995	0.215	1	17	Negative
2	CRC (n = 8), Hl (n = 10)	0.222	0.138	Yes	0.407	0.996	0.35	1	2	Negative

^a ROC analysis was performed using the cross-validated predicted Y-values. Corresponding AUCs were determined. ^bPermutation testing was performed on the parent PLS-DA models that contained the same number of components. ^cNumber of permuted classifications with Q² values exceeding that on the actual classification of samples. ^dy intercept of the regression line fitted among the Q² values.

Table 2. Fecal and tissue marker metabolites characterizing CRC

Fecal marker metabolites differentiating CRC patients and healthy subjects
Metabolite identity	Fold change^a	S.I.^b	R.I.^c
Nicotinic acid	0.25	91	1269.3
Fructose ^d	0.26/0.27	84/82	1928.4/1937.8
Linoleic acid	0.31	92	2200.9
Tissue marker metabolites differentiating CRC tumor and matched normal mucosae
Metabolite identity	Fold change^e	S.I.^b	R.I.^c
Glucose	0.10	88	1940.0
Galactose	0.11	78	1894.9
3-phosphoglycerate	0.32	82	1824.2
Citric acid	0.35	85	1843.7
Uracil	4.06	90	1328.6
Uridine	3.21	60	2445.2
Inosine	0.46	82	2583.4
Proline	2.12	88	1155.9
Creatinine	0.49	87	1558.3

^a Fold change less than 1 indicates lower level in the feces of CRC patients compared with healthy subjects; ^bS.I.: Similarity Index (the percentage by which the mass spectrum of the analyte matches that in the library); ^cR.I.: Retention Index of the derivatized metabolites; ^dDerivatized metabolites were present as multiple peaks in the chromatogram; ^eFold change more than 1 indicates higher level in the tumor tissue compared with matched normal mucosae; fold change less than 1 indicates lower level in the tumor tissue compared with matched normal mucosae.

Tissue metabonomic profiling of CRC

Tumor tissue was clearly distinct from matched normal mucosae in the PCA scores plot (Fig. 2A). Based on a validated OPLS-DA model (see Fig. 2B for OPLS-DA plot and Fig. 2C for ROC curve), 33 analytes that differentiated the two groups were elucidated. The putative identities of 9 marker metabolites are presented in Table 2. Glucose, galactose, 3-phosphoglycerate, citric acid, inosine, and creatinine were reduced while uracil, uridine, and proline were elevated in the tumor tissue compared with matched normal mucosae (Table 2).

Figure 2. Tissue metabonomic profile of tumor tissue and matched normal mucosae. (A) PCA scores plot discriminating CRC tumor (T) from matched normal mucosae (M). (B) OPLS-DA scores plot discriminating CRC tumor (T) from matched normal mucosae (M). (C) Validation plot of OPLS-DA model.

Discussion

Fecal metabonomic profiling of CRC

Developing an accurate and non-invasive screening tool has been a longstanding goal in the management of CRC. In this respect, fecal metabonomics represents a conceptually promising approach, given the intricate changes in the molecular environment of the gut lumen which could be induced by neoplasia and the underlying dysbiosis and reflected as metabolic alterations in the feces. Herein, we evaluated the use of a GC/TOFMS-based fecal metabonomic approach in detecting CRC.

Preliminary unsupervised PCA revealed a partial separation between the CRC patients and healthy subjects. The lack of complete separation between the two groups was not unexpected as the large inter-individual variability in terms of diet and gut microbial activity might possibly dilute the disease-related metabotype. To circumvent the systematic variation unrelated to pathological status and enhance class separation, a supervised OPLS-DA was employed.^28,29 The resultant OPLS-DA models were robust in discriminating CRC patients from healthy subjects, as evident from the good Q² and area under the ROC curves. Of note, CRC 3, a CRC patient who had undetectable blood in stool and might thus constitute a false-negative in conventional FOBT-based screening of CRC, was identified correctly in the OPLS-DA model (Table S1). Our current findings illuminated the potential utility of GC/MS-based fecal metabonomics in accurately identifying CRC patients. Furthermore, it reiterated the role of metabonomics as a powerful systems biology tool in disease classification even with a small sample size of approximately 20 subjects.^30-32 Future studies may involve an external validation on a larger series of patients to further assess the predictive power of the model. Subjects with other cancer stages may also be included to ascertain the generalizability of the data.

Gut bleeding is a clinically prevalent phenomenon associated with CRC and other gut diseases. One of the strengths of our study was the additional consideration of the influence of blood in stool on the fecal metabonomic analysis. This crucial concept was first elucidated in a study demonstrating that gross bleeding (100 mg Hb/g feces) could potentially confound GC/MS-based fecal metabonomic studies while occult blood (1 mg Hb/g feces) exerted a negligible effect on the fecal metabotype.³³ Efforts were hence undertaken in the current study to construct and compare models with and without the inclusion of fecal samples containing blood beyond 1 mg Hb/g stool. Our data clearly verified the ability of fecal metabonomic analysis to reliably differentiate CRC subjects from healthy subjects regardless of the presence of samples containing blood beyond 1 mg Hb/g stool. Nevertheless, further comparison of Q² and the outcome of permutation testing between models 1 and 2 indicated that class separation was slightly obscured in the presence of fecal samples containing blood beyond 1 mg Hb/g stool. The latter finding reinforced the variation conferred by gross blood that could influence the fecal metabotype characterizing CRC. Importantly, to accurately glean mechanistic insights into CRC pathogenesis, marker metabolites were extracted from model 2 that was constructed independently of the presence of blood-related metabolites.³³

Comparison of metabolic alterations at fecal and tissue levels

Fecal markers can be broadly classified as exfoliated markers, secreted markers, and leaked markers.⁵ Exfoliated markers are derived from the shedding of colonocytes at the gut luminal surface; secreted markers are emanated from the epithelial cells lining the colonic lumen; and leaked markers are products of disturbed blood or lymph vessels following tumor growth. The use of model 2 precluded the possibility of leaked markers being elucidated. To shed light on whether the unique fecal metabotype of CRC patients originated, specifically, from the exfoliated tumor cells, the metabolic aberrations at both fecal and tissue levels were examined. Corroborating previous studies,^11-16 our tissue metabonomic analysis showed elevated amino acid (e.g., proline) and nucleic acid constituents (e.g., uracil and uridine), and decreased carbohydrates (e.g., glucose and galactose) in the tumor reflective of their higher proliferative state. Interestingly, this set of marker metabolites was not manifested at the downstream fecal level. The discordant set of tissue and fecal markers implied that the unique metabotype differentiating exfoliated tumor cells from normal colonocytes was possibly diluted in the feces by other contributors to the fecal profile, for instance the abundant shedding of normal colonocytes. In other words, the fecal metabotype of CRC patients were possibly defined more extensively by processes beyond the direct shedding of tumor cells. These complex processes remained to be determined, although possible postulations were cancer-related dysbiosis of gut microbiota or the occurrence of distinct metabolic fluxes between tumor cells and the colonic lumen. Following the recently expounded theory of “tumor field effects”,³⁴ it may also be possible that morphologically normal colonocytes in CRC patients harbored metabolic changes induced by adjacent tumor activity and the exfoliation of these cells might contribute to the discrete fecal metabotype of CRC patients. Importantly, our intriguing finding opened new perspectives for future investigations into the origin of fecal metabolic markers and the multifactorial manifestation of CRC.

Marker metabolites from fecal metabonomic profiling

Concurrent with our publication, Weir et al. also described the GC/MS-based fecal metabonomic profiling of CRC but the influence of blood in stool was not considered in the elucidation of marker metabolites.³⁵ In our study, non-blood-related fecal marker metabolites were extracted from model 2 constructed using fecal samples that do not contain gross levels of blood. Based on these marker metabolites, further insights into the pathogenesis of CRC were yielded. While it might be premature to draw any conclusive association between the marker metabolites and their dysregulated pathways, the relevance of these markers was underscored by several prior studies.

Extensive expression of GLUT-5, a fructose transporter, has been implicated in different types of cancers, including CRC,³⁶ presumably in response to their heightened energy consumption. As both CRC and healthy subjects were not restricted in terms of dietary carbohydrate intake, the reduced fecal level of fructose observed in CRC patients might correspond with the augmented uptake via GLUT-5 in the colon. Linoleic acid has been shown recently to inhibit the growth of CRC cells by mediating oxidative stress and mitochondrial dysfunction.³⁷ The finding of a decreased level of linoleic acid in the feces of CRC patients, consistent with the observation by Weir et al.,³⁵ was potentially significant given its reported anti-oncogenic effect. Nevertheless, the mechanistic cause of the alteration remained unclear and may be of considerable interest in future studies. Finally, the diminished fecal level of nicotinic acid in CRC patients was in agreement with the previously established link between nicotinic acid deficiency and genomic instability.³⁸ The former raised the possibility of a suppression in the activity of nicotinic acid-producing bacteria such as bifidobacteria³⁹ in relation to CRC. While further studies are required to substantiate the functional significance of these metabolic aberrations, the current work provided novel insights into the role of both tumor cells and gut microbes in CRC tumorigenesis.

Few previous reports on NMR-based fecal water analyses have identified metabolites such as acetate and butyrate as predominant markers differentiating CRC patients from healthy subjects.^20,21 Notably, a different set of discriminating metabolites was uncovered in our study. The use of GC/MS hence contributed toward a broadened metabolic space for the discovery of fecal CRC markers.

Conclusion

This proof-of-concept study highlighted the ability of a GC/MS-based fecal metabonomic approach to accurately identify patients with CRC. Despite the small sample size, the discovery of a unique metabotype characterizing CRC opened new avenues in the non-invasive detection of the disease. In addition, the elucidation of marker metabolites provided a mechanistic handle to delineate the biochemical underpinnings of the disease.

Materials and Methods

Clinical population and samples

Eleven CRC patients (Duke stages B–C) and 10 healthy subjects were recruited at the outpatient specialist clinic from September 2011 to June 2012. All subjects were Han Chinese. The inclusion criteria for CRC patients were a diagnosis of sporadic CRC and an intention to undergo surgical tumor resection. Exclusion criterion was treatment with neoadjuvant chemotherapy or radiotherapy. Randomly selected healthy subjects were eligible for this study if a colonoscopy was performed within the past 5 mo to confirm the absence of CRC or pre-cancerous polyps. Exclusion criteria consisted of a history of CRC, a history of polyps within the past 3 y, the presence of inherited CRC syndromes (e.g., Lynch syndrome and familial adenomatous polyposis), and a family history of hereditary CRC. In both groups, subjects did not have coexistent inflammatory bowel disease. The clinical characteristics of the enrolled CRC patients and healthy subjects were summarized in Table 3. This study was approved by the Institutional Review Board at the Singapore General Hospital (2009/723/B). All subjects provided informed consent prior to participation in the study.

Table 3. Clinical characteristics of CRC patients and healthy subjects

Characteristics	CRC patients (n = 11)	Healthy subjects (n = 10)	P value
Age, years
Mean	64.5	57.4	0.090^a
Range	56–80	48–79
Sex, no. (%)
Male	7 (63.6)	4 (40.0)	0.395^b
Female	4 (36.4)	6 (60.0)
Tumor site, no. (%)
Sigmoid	2 (18.2)
Rectosigmoid	2 (18.2)
Rectum	7 (63.6)
Dukes’ stage no. (%)
B	6 (54.5)
C	5 (45.5)
Grade
Well differentiated	1 (9.1)
Moderately differentiated	10 (90.9)

^aIndependent samples t test; ^bFisher exact test.

Fecal specimens were collected domestically by CRC patients prior to surgical resection and by healthy subjects within a month of enrolment into the study. Subjects were given detailed instructions for sample collection. Briefly, a representative aliquot of feces was collected in a sealed plastic container. Care was taken to avoid contact with toilet water and urine. The specimens were transported to the hospital on the same day, or otherwise, kept at 4 °C and brought to the hospital within 2 d. All specimens were stored at −80 °C immediately upon receipt at the hospital.

Surgical tissues resected from the same cohort of CRC patients were snap-frozen in liquid nitrogen, microdissected, and stored at −80 °C until analysis. Careful microdissection of the excised tissue ensured that at least 90% of the tumor specimen comprised cancer cells. Matched normal mucosa samples were obtained from excised tissue at least 5 cm away from the edges of the tumor. The tumor specimens were also subjected to routine histological examination by a blinded pathologist to identify the stage and grade of the cancer. Due to limited sample availability, surgical tissues from 8 out of the 11 CRC patients were subjected to metabonomic profiling.

Materials

RIDASCREEN^® Hemoglobin was purchased from R-Biopharm. Spectroscopy grade methanol, chloroform, and toluene were obtained from Tedia. N-methyl-N-(trimethylsilyl) trifluoroacetamide (MSTFA) with 1% trimethylchlorosilane (TMCS) and methoxyamine hydrochloride in pyridine (MOX) were products of Thermo Fisher Scientific. Standard alkane series (C10–C40) and sodium sulfate were purchased from Sigma-Aldrich Inc. Water was purified by a Milli-Q system (Millipore).

Immunochemical FOBT

Fecal hemoglobin (Hb) levels were measured using RIDASCREEN^® Hemoglobin, an enzyme-linked immunosorbent assay (ELISA)-based immunochemical FOBT, according to the manufacturer’s instructions. Absorbance (extinction at 450 nm, reference wavelength of 620 nm) was measured on an Infinite M200 microplate reader (Tecan). RIDA®SOFT Win Software was used to convert absorbance values into Hb concentrations (mg Hb/g stool). Fecal Hb levels were stratified into the following categories: (1) not detected, (2) below 0.05 mg Hb/g stool, (3) 0.05–1 mg Hb/g stool, and (4) higher than 1 mg Hb/g stool.

GC/TOFMS metabonomic profiling of feces

A previously developed sample preparation protocol was adopted with slight modifications.³³ Fecal samples were freeze-dried and 80 mg of each lyophilized sample was ultrasonicated with 1 mL of ice-cold methanol:water (8:2) mixture in a bath ultrasonicator (Elma Transsonic 460) at 4 °C for 30 min, and vortex-mixed for 2 min. The samples were centrifuged subsequently at 18 000 g at 4 °C for 20 min. A 0.5 mL of the supernatant was obtained carefully. Additionally, quality control (QC) samples were prepared by combining equal volumes of supernatant from all CRC and healthy subjects. Each sample and QC supernatant was concentrated to complete dryness at 50 °C under a gentle stream of nitrogen gas (Turbovap LV, Caliper Life Sciences). Toluene (100 μL, kept anhydrous with sodium sulfate) was added to the dry residue, and dried fully again at 50 °C under nitrogen gas. Methoximation of the dried metabolic extract was performed with 50 μL of MOX (20 mg/mL) at 60 °C for 2 h. Following methoximation, 100 μL of MSTFA with 1% TMCS was introduced and the mixture was incubated at 60 °C for 45 min to produce trimethylsilyl derivatives. Finally, 100 μL of the cooled products was subjected to GC/TOFMS analysis.

GC/TOFMS metabonomic profiling of tissue

Tissue samples were processed using the method described by Mal et al.,¹³ with slight modifications. Fresh frozen tissue (approximately 20 mg) was ultra-sonicated with 1 mL of ice-cold chloroform:methanol:water (2:5:2) mixture in a bath ultrasonicator at 4 °C for 150 min. After vortex-mixing for 2 min, the samples were centrifuged at 18 000 g at 4 °C for 6 min. A 0.65 mL of supernatant was extracted. Separately, QC samples were prepared by pooling equal volumes of supernatant from all tumor tissues and normal mucosae. Drying of the supernatants was accomplished as described above. Derivatization was performed using 100 μL of MSTFA with 1% TMCS followed by incubation at 70 °C for 30 min to produce trimethylsilyl derivatives. Finally, 50 μL of the cooled products was subjected to GC/TOFMS analysis.

GC/TOFMS analysis

GC/TOFMS analysis was performed on a Pegasus GC/TOFMS (LecoCorp.) connected to an Agilent 7890 gas chromatograph. A DB-1 (30 min × 250 µm i.d.) fused silica capillary column (Agilent J&W Scientific), with 0.25 µm film thickness, was operated with open split interface. Helium was employed as the carrier gas at a constant flow rate of 1.5 mL/min. The injector split ratio was set at 1:20 and 1:6 for fecal and tissue metabonomic analysis, respectively. A volume of 1 μL was injected. The injector, transfer line, and ion source temperatures were held at 220, 280, and 250 °C, respectively. Oven temperature was programmed at 70 °C for 0.2 min, escalated at 9 °C/min to 270 °C where it remained for 5 min, and further increased at 40 °C/min to 310 °C where it was maintained for 11 min. The MS functioned in the electron impact ionization mode at 70eV. Mass data was acquired in a full scan mode from m/z 40 to 600 with an acquisition rate of 20 spectra per second. Detector voltages were 1500 V and 1650 V for fecal and tissue metabonomic analysis, respectively. During the analysis, samples from CRC and healthy subjects were randomized, along with QC samples, to avoid analytical bias. To detect retention time shifts and facilitate kovats retention index (RI) calculation, standard alkane series mixture (C10–C40) was injected at regular intervals. RIs are relative retention times normalized to n-alkanes eluted adjacently.⁴⁰

Data pre-processing

All gas chromatograms were subjected to baseline correction, de-convolution, noise reduction, smoothing, library matching, and area calculation using ChromaTOF software (version 4.41, LecoCorp.). Only peaks with signal-to-noise ratio larger than 150 were analyzed. The following peaks were further discarded: artifactual peaks that were similarly present in blank samples, poorly resolved peaks, and peaks with poor shapes.

Analytes were identified putatively by comparing both the MS spectra and RI with those in the National Institute of Standards and Technology (NIST) library. If RIs were unavailable in the NIST library, RIs from the Fiehn library, Golm Metabolome Database (http://csbdb.mpimp-golm.mpg.de/gmd.html), Human Metabolome Database (http://www.hmdb.ca/) as well as reported literature⁴¹ were utilized for comparison.

Statistical analysis

Normalization to a constant total chromatographic peak area was made prior to chemometric analyses. Irreproducible peaks defined by a coefficient of variation greater than 30% among QC samples were excluded from further analysis. Chemometric analysis was performed using SIMCA-P version 12.0 software (Umetrics). The normalized data was mean-centered, unit variance scaled,⁴² and subjected to unsupervised principal component analysis (PCA) to observe for outliers and inherent grouping trends. Following PCA, supervised orthogonal partial least squares discriminant analysis (OPLS-DA) was applied to build discriminating models and elucidate differential analytes that contributed to the separation. For fecal metabonomic profiling, two separate models were constructed. The first model (model 1) was built using the fecal metabolic profiles of all subjects regardless of the levels of blood in their stool measured using the FOBT. In the second model (model 2), fecal samples containing blood beyond 1 mg Hb /g stool were excluded. The performance of the models was indicated by the parameters Q² and R². The goodness-of-prediction parameter Q² described the predictive ability of the model under 7-fold cross validation. A quantitative measure of the fraction of the variation in the data set that could be explained by the model, also known as goodness-of-fit, was given by the parameter R².

To examine the robustness of the OPLS-DA models in classifying CRC patients and healthy subjects, and distinguishing tumor tissue from normal mucosae, receiver operator characteristic (ROC) analysis was performed using the cross-validated predicted y values. Corresponding areas under the curves (AUC) were determined. To check for overfitting of data in the OPLS-DA models, response permutation testing was performed with 100 repetitions. The validation was performed on the parent partial list squares discriminant analysis (PLS-DA) models that contained the same number of components. During this process, the sample classes (also known as y variables) were shuffled randomly while the x data were left intact. The validity of the models was (1) inversely related to the number of permuted classifications with Q² values exceeding that on the actual classification of samples, and (2) indicated by a negative y intercept of the regression line fitted among the Q² values.

Based on OPLS-DA, a list of differential analytes was identified using the following criteria: (1) variable importance for projection score of more than 1, (2) P value of less than 0.05 in the relevant t tests, and (3) absolute fold change of more than 2. Specifically, the independent Welch t test was used to compare analytes from fecal samples of CRC patients and healthy subjects, while paired samples t test was used to compare analytes from tumor tissue and matched normal mucosae (Microsoft Excel 2007).

Abbreviations:
CRC	=	colorectal cancer
FOBT	=	fecal occult blood test
NMR	=	nuclear magnetic resonance
GC/MS	=	gas chromatography/mass spectrometry
GC/TOFMS	=	gas chromatography/time-of-flight mass spectrometry
MSTFA	=	N-methyl-N-(trimethylsilyl) trifluoroacetamide
TMCS	=	trimethylchlorosilane
MOX	=	methoxyamine
Hb	=	hemoglobin
QC	=	quality control
RI	=	retention index
NIST	=	National Institute of Standards and Technology
PCA	=	principal component analysis
OPLS-DA	=	orthogonal partial least squares discriminant analysis
ROC	=	receiver operator characteristic
AUC	=	area under the curve
PLS-DA	=	partial list squares discriminant analysis

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Acknowledgments

The authors thank Ms Jiamin Koh, Dr Ming Hian Kam, Ms Pauline Ngiik Hung Wong, and the Department of Colorectal Surgery at Singapore General Hospital for their support in subject recruitment, as well as Ms Elya, Ms Michelle Shu Mei Lo, Ms Wei Lin Goh, and Ms Grace Yu Hui Wong for their technical assistance and retrieval of the clinicopathological data. This project is funded by the Singapore Ministry of Education’s (MOE) Academic Research Grant (R-148-000-135-112 to ECY Chan and R -148-000-133-112 to HK Ho). LC Phua is supported by the NUS President Graduate Fellowship.

10.4161/cbt.27625