A quantitative multimodal metabolomic assay for colorectal cancer

This study was approved by the Conjoint Health Research Ethics Board at the University of Calgary (Ethics ID number E21805) and conforms to the Helsinki Declaration (October 2008). Colorectal cancer serum samples from patients in stage I, stage II, and stage III (locoregional CRC), and stage IVa (liver-limited metastasis) along with the relevant clinical information were collected in a prospective cohort of colorectal cancer patients, diagnosed at the Foothills Medical Center or referred to this center for resection and management of their primary adenocarcinoma or metastatic CRC and provided written consent, between 2006 and 2012. Patients with extrahepatic metastases, any acute inflammatory state, sepsis, and familial colorectal adenomatosis or cancer cases were excluded from this study. Colorectal adenoma samples and control samples were collected prospectively by the Forzani & MacPhail Colon Cancer Screening Centre (CCSC) at the University of Calgary. Disease-free controls consisted of individuals who underwent screening colonoscopy and were found not to have CRC or adenoma. CRC, adenoma, and control cases were all fasting for a minimum of 8 h before surgery or endoscopic procedure. Surgical pathology for CRC patients and endoscopic pathology for adenoma and control cases were obtained and thoroughly reviewed, and all diagnoses were confirmed. Control samples are matched for age and gender with adenoma and cancer patients. Samples were collected in plastic gold top vacutainer tubes (BD Biosciences, Mississauga, Ontario, Canada), which contain a clot activator and a proprietary gel for serum separation. Samples were processed in 6-h time from collection and were stored in −20°c freezers until the day of analysis [8, 9].

In these studies, we used quantitative Biocrates AbsoluteIDQ p150 Kit (Biocrates Life Sciences AG, Austria) which can measure the concentration of 163 endogenous metabolites from 4 biochemical classes, using a panel of 27 validated internal controls. These metabolites are selected from four classes of acylcarnitines, biogenic amines, sphingolipids, and glycerophospholipids (lyso-phosphatidylcholines and phosphatidylcholines). Detailed preparation steps have been previously described [10]. In brief, 10 μl of each serum sample was prepared and submitted to mass spectrometric analysis on an API4000 Qtrap® tandem mass spectrometry instrument (Applied Biosystems/MDS Analytical Technologies, Foster City, CA) equipped with a solvent delivery system, in both positive and negative ion modes. Selective multiple reaction monitoring (MRM) detections of chemical homologs in conjunction with stable isotope-labeled internal controls, provided in the kit plate filter, was utilized for metabolite quantification. This method is shown to be in compliance with the FDA guidance for industry- Bioanalytical method validation [11], indicating proof of reproducibility in the given range. External quality control samples consisted of four commercially available human serum samples (Sigma-Aldrich, Germany) as well as four pooled serum samples from our controls for each kit. The final metabolite concentrations were automatically calculated by MetIDQ, the proprietary software of the kit. Metabolites were checked for the coefficient of variance >0.25 in quality controls, and all passed the check. The concentrations of 146 metabolites were above the limit of detection, and these metabolites were selected for further analysis. Missing values were imputed with the minimum value in the dataset. The pre-processed data were then transferred for further statistical analyses.

Throughout this study, wherever a two group statistical comparison were desired, a two-sided Student’s t-test was used. We considered a priori p-value smaller than 0.05 as statistically significant. Where required, the significance thresholds adjusted by Holm-Bonferroni correction method were used. For analysis of stage-dependent variations in more than two groups, Bonferroni-corrected Kruskal-Wallis Test [8] (non-parametric approach) was computed by Multi-Experiment Viewer (MeV), version 4.9 (The TM4 Software Development Team) [9]. To generate heatmaps, we used the Spearman’s Rank Correlation distance metrics and complete linkage method.

The pre-processed data from MetIDQ were log-transformed and autoscaled (unit variate scaled and centered) before importing the SIMCA multivariate analytical software (Version 14.0.0, Umetrics AB, Sweden). Owing to the quantitative nature of the assay, we require median fold change normalization to account for inter-sample analytical biases. To evaluate the risk of overfitting bias on a dataset with small sample size, a pre-analysis permutation test of 10,000 iterations was applied on sample classes in the training set, as described before by Westerhuis et al. [12]. As for the combinatorial analysis, the relevant datasets were joined and then block-transformed [13‐16]. For each comparison, an exploratory Principal Component Analysis (PCA) with up to three components was used for discovering intrinsic clusters and revealing potential outliers. After exclusion of outliers, subsets of potentially significant metabolites for each comparison were selected by performing Welch’s t-test (assuming unequal variances). This filtering procedure was performed by setting a pre-test maximum p-value threshold of 0.30 in the Welch test, which removes clearly uniformative metabolites from further analysis [4, 17, 18]. Selected subsets were used for orthogonal partial least squares discriminant analysis (OPLS-DA) or O2-PLS-DA. Further refinement was applied through excluding metabolites with variable importance on projection (VIP) of less than a threshold. This VIP threshold was set separately for each analysis, so as the maximum for R²Y and Q²Y are obtained and their difference is at minimum. This approach has shown to be sufficiently reliable for the purpose of multivariate statistical analyses, including OPLS-DA [4, 17].

To assess the performance of supervised multivariate models, including OPLS-DA and O2PLS-DA, R²Y and Q²Y scores were used for measurement of the dataset variance covered by the model, and the predictability of the model in 7-fold cross-validation [4]. Models with the difference of more than 0.2 between R²Y and Q²Y were reevaluated. Potential confounders in each model were evaluated for their unwanted effects, as described in the Results section. Also, to examine the PLS-DA and OPLS-DA models for validity and potential overfit, a permutation test of 999 iterations was applied to each model, and the results were reported as Q2-intercept for that model [19]. Q2-intercept is the intercept of a line fitted to Q²Y scores versus the correlation of the permutated Y-vector and original Y-vector for each iteration. The model is valid and non-random if Q2 intercept is at or below zero [20].

Predictive performance of the generated models in external validation were evaluated by the area under the receiver operating characteristic curves (AUROC), which were calculated by GraphPad Prism (version 6.01 for Windows, GraphPad Software, La Jolla California USA, www.​graphpad.​com).

The characteristics of the study cohort are summarized in Table 1. Samples were randomly assigned to the training set and validation set, in a stratified design for locoregional CRC (stages I, II, and III), and liver-limited metastatic CRC (stage IVa). In patients with stage IVA disease, 17 (45%) received chemotherapy within 3 months before sampling; 8 patients (33%) with non-metastatic disease received chemotherapy before sampling.

To quantitatively evaluate serum composition of amino acids, acylcarnitines, and lipid compounds, including glycerophospholipids and sphingolipids, we submitted samples to semi-quantitative FIA-MS/MS. Out of 163 measured metabolites, 146 metabolites could be reliably found in all groups of patients and controls (Additional file 1: Table S1). Three samples were identified as outliers on PCA, and they were excluded from further analysis. We also evaluated for a potential confounding effect of pre-sampling chemotherapy on the metabolomic profile of CRC patients (Additional file 2: Fig. S1A to E). We could not identify any cluster linked to the chemotherapy status. The training set consisted of 53 samples (controls (N = 21), stage I (N = 5), stage II (N = 5), stage III (N = 5) and stage IVa (N = 17) cases). Following filtering by p-value and VIP (>1), 48 metabolites were employed in the metabolomic model, which resulted in an encouraging model: R²Y was 0.83 and Q²Y was 0.75; CV-ANOVA was <0.00001 (Fig. 1a and b, and Table 2). The model was then tested on an independent validation set (controls (N = 20), stage I (N = 3), stage II (N = 3), stage III (N = 3) and stage IVa (N = 19) cases). Sensitivity in this external validation set was 93%, and specificity was 95%; accuracy was 94%, and precision was 96%; AUROC was 98%.

To evaluate whether the identified CRC profile is significantly derived from the stage IVa subgroup of patients, we generated a discriminant model for separation of 5 classes: 4 CRC stages and disease free controls,. This analysis included all samples contained in the training and validation sets (N = 100) (Additional file 3: Fig. S2a). The model could strongly distinguish all CRC stages from the control group on the first predictive component (R²Y = 0.46 and Q²Y = 0.30, p-value <0.00001). On the second component, stage IVa patients were distinguished from all locoregional CRC patients. The overlap observed in locoregional CRC stages could be either due to the similarity of their metabolomic profile (relative to their difference with control or stage IVa groups; Additional file 1: Table S2), or it may be secondary to a smaller sample size in each of the three locoregional classes. In all, the data demonstrate that there are stage-associated signatures, but there is also a signature that distinguishes CRC patients as a whole from disease-free controls.

Figure 1b describes the correlation coefficients of the most important metabolites contributing to the CRC metabolomic profile. Glutamine was the most important metabolite in the projection derived from FIA-MS/MS (decreased in CRC), followed by phosphatidylcholine acyl –alkyls C38:1 and C40:3 (increased and decreased in CRC, respectively). In univariate t-test, glutamine was the most differentially abundant metabolite, as well (p-value = 2.69e-10). Ornithine and serine had the highest positive correlations with CRC, while decenoylcarnitine (C10:1) had the highest negative correlation with CRC (Fig. 1b). Of 14 amino acids explored, methionine, valine and tryptophan were significantly reduced in CRC, while ornithine and serine had increased levels. The remaining amino acids did not demonstrate any considerable significance in our model. There was no clear association with stage (Fig. 1c and d). Figure 1e and f illustrate the distribution of changes in the PCs in CRC, based on the length of carbon chains and the number of unsaturated bonds. Five lysophosphatidylcholines (Lyso-PC) were particularly elevated in CRC; none was decreased. This has not been previously reported. Phosphatidylcholines are important constituents of cell membranes, as well as mediators in lipid metabolism. Alterations of the phosphatidyl acyl-alkyl compounds with 38 to 44 carbons in CRC were also observed. These perturbations are represented as fold changes per carbon content in Fig. 2a and are discussed later.

To further enhance the capability of this signature biomarker, we evaluated a combinatorial biomarker composed of the 41 compound GC-MS driven signature we previously defined [5] and this 48 metabolite FIA-MS/MS signature. The inclusive model had an R²Y score of 0.91 and a Q²Y score of 0.84 (CV-ANOVA p-value <0.00001). When we regenerated the model with the most informative metabolites (46 compounds, Table 3), the model had an R²Y score of 0.89 and a Q²Y score of 0.80 with CV-ANOVA p-value <0.00001 (Fig. 2b). With permutations consisting of random assignments of samples to diagnostic classes [12], permutated models performed poorly. That is, with 10,000 permutations, the Q²Y = −0.11 (compared to 0.84 in our model); and root mean square error of prediction (RMSEP) = 0.56 (compared to 0.11 in our model). In all, the permutated models and our model were significantly different (P < 0.00001), and our model had a superior performance than models based on random class assignment. Moreover, in a permutation test consisting of random metabolite combinations in actual diagnostic classes, the Q2-intercept was −0.59, which is a reflection of a model with a high degree of robustness and reliability.

Sporadic colorectal adenocarcinoma is preceded by the formation of adenomatous lesions (polyps) as its precursor. Sera from 31 average-risk adenoma patients and 31 age- and gender- matched healthy controls in the age range of 50 to 70 years old, to evaluate whether metabolomic profiling can also detect this very early stage of the disease. In all cases, one adenoma <1 cm in diameter was detected by colonoscopy, except for one case, that had two adenomas. Clinical factors for these groups are summarized in Table 1.

Adenomas and disease-free controls were analyzed on the FIA-MS/MS platform. An unsupervised multivariate analysis, using 146 quantitative concentrations revealed no latent pattern or clustering (8 components, R²X = 0.77) (Fig. 3a). A supervised analysis (OPLS-DA) with a focused analysis of 9 metabolites identified by data filtration (p < 0.30, Table 4 and Additional file 1: Table S3) provided the basis of a model discriminating adenoma and disease-free controls (R²Y = 0.30, Q²Y = 0.20, CV-ANOVA p-value = 0.01, one orthogonal and one predictive component) (Fig. 3b). The AUROC on internal cross-validation was 82% (95% CI, 0.72–0.93) (Fig. 3c). Therefore, the quantitative analysis using this set of 9 metabolites is comparable to the metabolomic biomarker derived previously by GC-MS, comprised of 17 metabolites [5].

Finally, we sought to determine whether the adenoma profile identified by the quantitative approach could be combined with the adenoma profile defined by GC-MS [5] to enhance the performance of the biomarker. The combined signature composed of 14 GC-MS metabolites and 9 quantitative metabolites could generate a model which was slightly better than the quantitative signature alone (R²Y = 0.46, Q²Y = 0.26, CV-ANOVA p-value = 0.003, Q2-intercept = −0.31). Selecting for the most important metabolites, we generated a biomarker pattern consisting of 18 compounds (R²Y = 0.33, Q²Y = 0.21, CV-ANOVA p-value = 0.01, Q2-intercept = −0.26) (Fig. 3d and Table 5). Overall, the combinatorial signature had a better performance in the training study.