Article pubs.acs.org/ac
Development of a Universal Metabolome-Standard Method for Long-Term LC−MS Metabolome Profiling and Its Application for Bladder Cancer Urine-Metabolite-Biomarker Discovery Jun Peng,† Yi-Ting Chen,‡ Chien-Lun Chen,§ and Liang Li*,† †
Department of Chemistry, University of Alberta, Edmonton, Alberta, Canada Department of Biomedical Sciences, College of Medicine, Chang Gung University, Taoyuan, Taiwan § Department of Urology, Chang Gung Memorial Hospital, Taoyuan, Taiwan ‡
S Supporting Information *
ABSTRACT: Large-scale metabolomics study requires a quantitative method to generate metabolome data over an extended period with high technical reproducibility. We report a universal metabolome-standard (UMS) method, in conjunction with chemical isotope labeling liquid chromatography−mass spectrometry (LC−MS), to provide long-term analytical reproducibility and facilitate metabolome comparison among different data sets. In this method, UMS of a specific type of sample labeled by an isotope reagent is prepared a priori. The UMS is spiked into any individual samples labeled by another form of the isotope reagent in a metabolomics study. The resultant mixture is analyzed by LC−MS to provide relative quantification of the individual sample metabolome to UMS. UMS is independent of a study undertaking as well as the time of analysis and useful for profiling the same type of samples in multiple studies. In this work, the UMS method was developed and applied for a urine metabolomics study of bladder cancer. UMS of human urine was prepared by 13C2-dansyl labeling of a pooled sample from 20 healthy individuals. This method was first used to profile the discovery samples to generate a list of putative biomarkers potentially useful for bladder cancer detection and then used to analyze the verification samples about one year later. Within the discovery sample set, three-month technical reproducibility was examined using a quality control sample and found a mean CV of 13.9% and median CV of 9.4% for all the quantified metabolites. Statistical analysis of the urine metabolome data showed a clear separation between the bladder cancer group and the control group from the discovery samples, which was confirmed by the verification samples. Receiver operating characteristic (ROC) test showed that the area under the curve (AUC) was 0.956 in the discovery data set and 0.935 in the verification data set. These results demonstrated the utility of the UMS method for long-term metabolomics and discovering potential metabolite biomarkers for diagnosis of bladder cancer.
M
as an external calibration standard where the sample signal was extrapolated from the external calibrants to compensate for signal drifts. This approach may not be accurate and precise in all cases, even if a QC injection is done after every one sample run. It is desirable to use an internal calibration method to achieve good precision of large-scale studies. In this work, we report a universal metabolome-standard (UMS) method to provide long-term analytical reproducibility for LC−MS-based metabolome profiling work. This method is applicable to the chemical isotope labeling (CIL) LC−MS metabolomics platform, while the external calibration method of injecting a QC sample is useful for nonlabeling LC−MS methods. However, CIL has been shown to provide much improved separation and detection for group-based submetabolome profiling (e.g., dansylation can be used to target the
etabolomics, attempting to profile the whole set of metabolites in a biological sample, holds enormous promises for discovery of metabolite biomarkers for diagnosis of diseases (e.g., cancer).1 Although sample size in a discovery study is usually small, analyzing a large number of samples is essential in the verification and validation phases in order to increase the statistical power and address the important issue related to heterogenesis of phenotypes of a disease.2−4 This longer term metabolomics study that often involves the analysis of different sets of samples of the same type requires an analytical method to provide high technical reproducibility over an extended period. While LC−MS is increasingly used in metabolomics research, achieving high long-term reproducibility is still an analytical challenge,5 due to issues such as ion suppression, MS performance drift, LC column contamination and aging, etc.6,7 To address this challenge, previous studies utilized intermittent injections of a quality control (QC) sample and developed algorithms to correct for signal drifts in the sample runs.7−9 In these methods, the QC sample was used © 2014 American Chemical Society
Received: March 24, 2014 Accepted: May 30, 2014 Published: May 30, 2014 6540
dx.doi.org/10.1021/ac5011684 | Anal. Chem. 2014, 86, 6540−6547
Analytical Chemistry
Article
amine- and phenol-containing submetabolome).10−13 The reported method of CIL LC−MS for metabolomics used a pooled sample equally aliquoted from all samples (e.g., control samples and disease samples) within a specific study.14,15 This pooled-control method requires an aliquot taken from each sample to generate the pooled sample, which may not be feasible if the original sample amount is limited. In addition, it cannot be conveniently applied to a large-scale metabolomics study where multiple batches of samples are analyzed at different times or even different laboratories. In the UMS method reported herein, the UMS of a specific type of sample (e.g., human urine) is prepared a priori by labeling a representative sample (e.g., a pooled urine sample from a number of individuals independent of any study) using one of the isotopic form of a labeling reagent targeting a group-based submetabolome. This standard is then spiked to individual samples labeled with the other form of the labeling reagent in any metabolomics study. LC−MS analyses of the resultant mixtures provide the relative quantification of the submetabolomes of individual samples. Since the same UMS is used for different batches of samples, the submetabolome data sets generated over a long period or from different laboratories and different studies can be readily compared. To illustrate the applicability and performance of the UMS method, we applied this method for a bladder cancer metabolomics study aimed to discover potential metabolite biomarkers for diagnosis of bladder cancer. Bladder cancer has the fourth highest incidence among all cancers in men and sixth among all men and women in the United States (www.cancer. gov). The current standard of diagnosis is cystoscopy, which is invasive, costly, and not sufficiently sensitive for early detection.16 Metabolomics has been reported to search for metabolite biomarkers of bladder cancer, including the studies of using human urine,17−21 serum,22,23 and tissues samples.24,25 Animal model studies have been also carried out.26 Urine biomarkers for detecting bladder cancer have attracted more attention due to noninvasive sample collection and urine’s direct contact with the tissue. Although these reported studies have shown the promise of metabolomics in search for biomarkers, most of these studies used healthy subjects as normal controls, which are very different from cancer patients. In addition, there was no external verification of the discovery results except one recent study using NMR.22 NMR is not sensitive, compared to LC−MS. In this work, we used both hernia and urinary tract infection (UTI) or hematuria (HU) as controls which have been shown to be particularly relevant to bladder cancer in recent proteomics studies.27−29 We demonstrated that the UMS method could differentiate these groups, and the results could be verified using a second set of samples analyzed at a different time using a different MS instrument.
information on sample collection, storage, and processing is given in Supplemental Note N1 of the Supporting Information. The sample information on bladder cancer patients and noncancer controls is shown in Supplemental Table T1 of the Supporting Information. The discovery data set included the samples collected from two noncancer control groups [i.e., 68 hernia and 31 urinary tract infection (UTI) or hematuria (HU)] and the bladder cancer group (i.e., 91 bladder cancer patients). The verification data set included the samples collected from 14 hernia, 37 UTI or HU, and 44 bladder cancer patients. These samples were analyzed using the UMS method one year after the discovery samples were analyzed. Universal Urine-Metabolome Standard. The overall workflow of the UMS method is shown in Figure 1. To
Figure 1. Workflow for the universal metabolome-standard (UMS) method. In this particular case, urine UMS was prepared from the pooled urine of healthy subjects. Individual samples from different batches were analyzed by LC−MS after mixing with an aliquot of the urine UMS.
produce the urine standard, a pooled urine sample with a 6 mL volume was generated by taking the same aliquot from each of 20 healthy individuals. One milliliter of the pooled urine sample was used to perform the dansylation reaction in a 20 mL centrifuge plastic tube. The reaction volume was scaled up proportionally based on the reported labeling protocol (see Supplemental Note N1 of the Supporting Information).10 Instead of using 12C2-dansyl chloride (20 mg/mL), 13C2-dansyl chloride (20 mg/mL) was used to label the pooled urine sample. There were 6 replicate reactions, and after the reactions the solutions were mixed. The final volume was 40 mL by diluting with 50% ACN solution. Only 20 μL of this UMS was spiked into a 12C2-dansyl-labeled individual urine sample. UMS was stored at −80 °C; we have not investigated the stability of the UMS when it is stored at low (e.g., −20 °C) or room temperature. Sample Concentration Normalization. An LC-UV method previously reported30 was used to normalize the total
■
EXPERIMENTAL SECTION Supplemental Note N1 of the Supporting Information provides information on sample collection, dansylation labeling, sample concentration normalization, quality control (QC) samples, blank run and sample storage effect, and additional LC−MS conditions used in this work. Some of the key experimental procedures are described below. Samples. All the cancer and control samples were collected in Taiwan and then shipped to Edmonton, Alberta, Canada, for analysis, while the samples of healthy individuals used for preparing the UMS were collected in Edmonton. The 6541
dx.doi.org/10.1021/ac5011684 | Anal. Chem. 2014, 86, 6540−6547
Analytical Chemistry
■
metabolite concentrations of individual samples before mixing with the UMS. Supplemental Note N1 of the Supporting Information provide more details on the normalization experiment. LC−MS. An Agilent 1100 series binary system (Agilent, Palo Alto, CA) and an Agilent reversed-phase Eclipse plus C18 column (2.1 mm × 100 mm, 1.8 μm particle size, 95 A pore size) were used for LC−MS. LC solvent A was 0.1% (v/v) LC−MS grade formic acid in 5% (v/v) LC−MS grade ACN, and solvent B was 0.1% (v/v) LC−MS grade formic acid in LC−MS grade ACN. The gradient elution profile was as follows: t = 0 min, 20% B; t = 3.0 min, 35% B; t = 16 min, 65% B; t = 18.6 min, 95% B; t = 21 min, 95% B; t = 21.3 min, 98% B; t = 23.0 min, 98% B; and t = 24.0 min, 20% B. Two microliters of sample was injected into LC−MS. The flow rate was 150 μL/min. The flow from LC was split 1:2, and a 50 μL/ min flow was loaded to the electrospray ionization (ESI) source of a mass spectrometer, while the rest of the flow was delivered to waste. A flow rate of 50 μL/min was optimal for the microspray interface used. All MS spectra were obtained in the positive ion mode. For the discovery work, Bruker 9.4 T ApexQe Fourier transform ion-cyclotron resonance (FT-ICR) mass spectrometer (Bruker, Billerica, MA) was used for LC−MS analysis of the 12C2-dansyl-labeled individual sample mixed with the 13C2-dansyl-labeled UMS. For the verification work, Bruker Maxis Impact high-resolution quadrupole time-of-flight (QTOF) mass spectrometer (Bruker, Billerica, MA) was used. Additional information on LC−MS conditions are shown in the Supplemental Note N1 of the Supporting Information. Data Analysis. The XCMS software31 was used for peak picking from the LC−MS data. An in-house written R program was used to find the 12C2-/13C2-dansyl labeled peak pairs based on the mass difference of 2.00671 Da of isotopic pairs and the mass accuracy tolerance of 2 ppm. The relative ion intensity of the pair was calculated. The redundant peaks of each metabolite, such as natural isotopic peaks, sodium/potassium/ ammonium adduct peaks, doubly or triply charged peaks, and dimer/multimer peaks, were automatically removed by the program. An in-house written Perl program was used to align the peak pairs across the different urine samples based on the mass accuracy tolerance of 5 ppm and retention time shift tolerance of 60 s. Multivariate statistical analysis was carried out using SIMCAP+ 11.5 (Umetrics AB, Umea, Sweden). Principal component analysis (PCA), partial least-squares discriminant analysis (PLSDA), and orthogonal projections to latent structures discriminant analysis (OPLS-DA) were used to analyze the data. A web-based software Metaboanalyst 2.032 was also used to perform other statistical analyses. Receiver operating characteristic (ROC) analysis was conducted using a web-based software ROCCET (http://www.roccet.ca/ROCCET/).4 The putative metabolite biomarkers were selected based on a list of variable importance from the random forest model and also VIP ranking from the OPLS-DA model. After subtracting the dansyl group mass, the MycompoundID program33 was used to search the measured metabolite mass against the human metabolome database (HMDB)34 or the Evidencebased Metabolome Library (EML)33 with a mass accuracy tolerance of 5 ppm to identify the metabolites putatively.
Article
RESULTS AND DISCUSSION
Figure 1 shows the overall workflow of the CIL LC−MS platform using a UMS. The UMS of a specific sample type is prepared by 13C2-dansyl labeling a large volume of a standard sample that should be sufficient for analyzing many individual samples. In this particular case, since we focused on urine metabolomics, a urine UMS was prepared from a pooled urine of healthy subjects. We have investigated several analytical issues related to the workflow. Reproducibility. One of the main objectives of developing the UMS method was to improve the technical reproducibility of a large-scale metabolome profiling work. We first assessed the reproducibility of this method over three months using a quality control (QC) urine sample (see Supplemental Note N1 of the Supporting Information on QC sample preparation). In this work, 190 individual samples in the discovery phase were randomly and unevenly divided into 4 batches, depending on the instrument availability on a given period. The first (47 samples), second (66 samples), and third (42 samples) batch was run in week 1, 3, and 5, respectively, and the last batch (35 samples) was run 3 months after the first batch started. During this 3-month period, the LC−MS instrument was running samples continuously for this and other projects. Within a batch, one QC injection was normally done after 9 sample injections except the last few samples. Supplemental Table T2 of the Supporting Information summarizes the results obtained on the precision of the QC data consisting of an average of over 580 peak pairs or putative metabolites per run. The mean CV of peak ratios within a batch ranges from 6.2% to 7.9%, and the median CV ranges from 4.3% to 4.8%. The mean CV from the four batches over the 3-month period is 13.9%, and the median CV is 9.4%. We also evaluated the 3-month reproducibility using a PCA score plot. Supplemental Figure S1 of the Supporting Information shows the plot from all the QC sample injections (for clarity, only the QC data are shown). It is clear that all the QC data sets are clustered together. Thus, the technical variation was small during three months of intermittent data acquisition. In a separate experiment, we assessed the overall experimental variation including sample preparation by using 4 experimental replicates of the QC sample. Supplemental Figure S2 of the Supporting Information shows the distribution of CV values of peak ratios for all the peak pairs detected. The mean CV is 9.4%, and more than 93% of the peak pairs have their CV of
