Metabolomic Patterns in Glioblastoma and Changes during

Mar 21, 2010 - (BAT). Reference samples were collected subcutaneously from the patients' .... within the contrast enhancing tumor and one in the BAT r...
0 downloads 0 Views 1MB Size
Metabolomic Patterns in Glioblastoma and Changes during Radiotherapy: A Clinical Microdialysis Study Carl Wibom,†,‡,§ Izabella Surowiec,‡ Lina Mörén,† Per Bergstro ¨ m,† Mikael Johansson,† ,‡ Henrik Antti,* and A. Tommy Bergenheim§ Institution for Radiation Sciences, Department of Oncology, Umeå University Hospital, Umeå, Sweden, Department of Chemistry, Computational Life Science Cluster, Umeå University, Umeå, Sweden, and Department of Neurosurgery, Umeå University Hospital, Umeå, Sweden Received November 26, 2009

We employed stereotactic microdialysis to sample extracellular fluid intracranially from glioblastoma patients, before and during the first five days of conventional radiotherapy treatment. Microdialysis catheters were implanted in the contrast enhancing tumor as well as in the brain adjacent to tumor (BAT). Reference samples were collected subcutaneously from the patients’ abdomen. The samples were analyzed by gas chromatography-time-of-flight mass spectrometry (GC-TOF MS), and the acquired data was processed by hierarchical multivariate curve resolution (H-MCR) and analyzed with orthogonal partial least-squares (OPLS). To enable detection of treatment-induced alterations, the data was processed by individual treatment over time (ITOT) normalization. One-hundred fifty-one metabolites were reliably detected, of which 67 were identified. We found distinct metabolic differences between the intracranially collected samples from tumor and the BAT region. There was also a marked difference between the intracranially and the subcutaneously collected samples. Furthermore, we observed systematic metabolic changes induced by radiotherapy treatment among both tumor and BAT samples. The metabolite patterns affected by treatment were different between tumor and BAT, both containing highly discriminating information, ROC values of 0.896 and 0.821, respectively. Our findings contribute to increased molecular knowledge of basic glioblastoma pathophysiology and point to the possibility of detecting metabolic marker patterns associated to early treatment response. Keywords: chemometrics • gas chromatography-mass spectrometry • glioblastoma • metabolomics • predictive metabolomics • radiotherapy • treatment response

Introduction In the search for new treatments for malignant glioma, there is an imminent need for an improved understanding of the basic tumor pathophysiology as well as for identification of predictive biomarkers for assessment of therapeutic response. Both the transcriptome and the proteome hold the promise to harbor candidate biomarkers associated to treatment response in malignant glioma.1,2 Lately, also the metabolome has gained interest in this regard. The metabolome is generally defined by all of the low molecular weight metabolites in a system and is considered to be downstream of both gene and protein expression, and as such reflect processes on both transcriptional and translational levels. There are several options to study the metabolism in various types of tumors. Magnetic resonance spectroscopy (MRS) is one clinically used technique to study selected metabolites in vivo. It holds * To whom correspondence should be addressed. Henrik Antti, Email: [email protected]. Address: Department of Chemistry, Computational Life Science Cluster, Umeå University, SE-901 87, Umeå, Sweden. Phone: +46 90 7865359. Fax: +46 90 7867655. † Department of Oncology, Umeå University Hospital. ‡ Umeå University. § Department of Neurosurgery, Umeå University Hospital. 10.1021/pr901088r

 2010 American Chemical Society

the advantage of being noninvasive and may therefore be used to repeatedly assess a patient throughout the course of the disease and its treatment. Although metabolomic investigations by MRS in brain tumor have demonstrated promising results related to diagnosis, prognostic markers and potential markers for therapeutic response,3 no metabolomic marker nor pattern has so far had a major impact on clinical decision making. MRS may be a powerful tool to study single metabolites in vivo but the limited number of metabolites possible to identify does not allow MRS to be used for global studies of the metabolome. Using MRS in vitro on cells or tissue extracts may provide an improved resolution compared to in vivo MRS, and thereby also an increased number of detectable metabolites. This approach has for instance been used to demonstrate metabolomic patterns discriminating between different types of brain tumors.4 A fundamentally different approach to study treatment effects at a metabolite level is to perform an unbiased global screening of the metabolome. Gas chromatography hyphenated with mass spectrometry (GC-MS) is an established method for this purpose,5,6 and has successfully been used to detect metabolite abnormalities in cancer in general5,7 as well as to discriminate brain tumor tissue from normal brain tissue.8 Journal of Proteome Research 2010, 9, 2909–2919 2909 Published on Web 03/21/2010

research articles

Wibom et al. 12

Figure 1. Coronar reconstruction of dose planning CT showing implanted microdialysis catheters. Two catheters were placed within the contrast enhancing tumor and one in the BAT region (see inset for preoperative contrast enhanced coronar CT reconstruction).

Brain tissue analysis is however not a feasible approach for longitudinal studies of glioma patients, since tissue collection normally only can be performed once in every patient. In this study, we employed stereotactic microdialysis for sample collection, which allows for continuous study of metabolic events in the extracellular space of tumors such as high-grade gliomas.9 Microdialysis samples were collected from the contrast enhancing tumor as well as from the brainadjacent to-tumor (BAT) region in freely mobilized patients, before and during five days of radiotherapy. The main objective of the study was to investigate the metabolome of malignant glioma and asses the metabolic response to radiotherapy by a predictive metabolomics approach.10

Methods Surgery and Microdialysis. Eleven patients with radiological suspicion of high-grade glioma considered not suitable for surgical resection were included in the study. The patients underwent a stereotactic biopsy to obtain a tissue diagnosis before nonsurgical treatment. The biopsy procedure was carried out under general anesthesia using the Leksell stereotactic frame (Elekta, Stockholm, Sweden). Biopsies, as well as the implantation of microdialysis catheters, were planned via a stereotactic CT investigation.11 The diagnoses were confirmed by a frozen section before implantation of two microdialysis catheters along the biopsy trajectory, one into the contrast enhancing tumor tissue and one into the brain adjacent to tumor (BAT) region (Figure 1). One catheter was also placed in the abdominal subcutaneous tissue as a reference. The catheters had a 10 mm semipermeable membrane with 100 kDa cutoff (CMA 71; CMA Microdialysis, Stockholm, Sweden). The position of the catheters was postoperatively confirmed on the CT performed for dose-planning. The catheters were connected to a 2.5 mL syringe placed in a micro infusion pump with a flow rate of 0.3 µiL/min (CMA 106 or CMA 107; CMA Microdialysis). All catheters were perfused with a Ringer solution (Perfusion fluid T1; CMA Microdialysis) mixed with Dextran (30 g Dextran 60 1000 mL-1) to prevent microfiltra2910

Journal of Proteome Research • Vol. 9, No. 6, 2010

tion. The samples were collected in microvials every second hour, thereafter frozen and kept at -80 °C until analyzed. Patient Care and Radiotherapy. The patients were allowed to recover from surgery for approximately 24 h at the neurointensive care unit before a CT for dose planning was performed. All patients received routinely perioperative bethametasone and the blood glucose was kept under 8 mmol L-1. Eight patients were planned for a standard radiation schedule of 2 Gy × 30. Three elderly patients in poor general condition were planned for a faster fractionation using 3 Gy × 13 (2 cases) or 3.4 Gy × 10 (1 case). Thus, the dose per fraction during the microdialysis sampling was 2 Gy (8 cases), 3 Gy (2 cases), or 3.4 Gy (1 case). The gross tumor volume was defined as the contrast enhancing part of the tumor and a margin of 2 cm was added for the planning target volume. The radiotherapy was started within two to five days after the biopsy and catheter implantation. Sampling of the microdialysis fluid continued during five days of irradiation including the morning after the fifth fraction. In a few cases, the microdialysis had to be terminated earlier due to malfunctioning catheters (Table 1). Sample Selection. Samples for analysis on GC-MS were selected relative to each patient’s individual treatment schedule. Depending on when the catheters were implanted in relation to treatment initiation, as well as on how long a given catheter stayed operational, we were able to follow the patients longitudinally for different durations of time (Table 1). Where applicable, we chose to analyze 3 samples collected before the first radiotherapy session (time points 1-3) and 3 samples collected thereafter; specifically the afternoon after the patient received his or hers first (time point 4), third (time point 5) and fifth (time point 6) radiotherapy fraction. The samples analyzed were collected between 4 and 8 pm, and were stored in two separate microdialysis vials. In a separate step, the two vials from each time point of interest were thawed on ice and pooled together to yield sufficient sample volumes for analysis. The pooled samples were then once again stored at -80 °C prior to extraction. Chemicals. The chemicals used for sample preparation were all of analytical grade, except where otherwise stated. The stable isotope-labeled internal standard compounds (IS) [13C5]-proline, [2H4]-succinic acid, [13C5,15N]-glutamic acid, [1,2,3-13C3]myristic acid, [2H7]-cholesterol and [13C4]-disodium R-ketoglutarate were purchased from Cambridge Isotope Laboratories (Andover, MA); [13C12]-sucrose, [13C4]-palmitic acid and [2H4]butanediamine · 2HCl were from Campro (Veenendaal, The Netherlands); [13C6]-glucose was from Aldrich (Steinheim, Germany) and [2H6]-salicylic acid was from Icon (Summit, NJ). Stock solutions of the IS were prepared either in purified and deionized water (Milli-Q, Millipore, Billerica, MA) or in methanol (J.T. Baker, Deventer, Holland) at the same concentration, 0.5 µg µL-1. Methyl stearate, was purchased from Sigma (St. Louis, MO). N-Methyl-N-trimethylsilyltrifluoroacetamide (MSTFA) with 1% trimethylchlorosilane (TMCS) and pyridine (silylation grade) were purchased from Pierce Chemical Co., heptane was purchased from Fischer Scientific (Loughborough, U.K.). Sample Preparation. Just prior to extraction, the samples were allowed to thaw at 37° for 15 min. 450 µL of the extraction solution (methanol/water (8:1) with 11 IS, each of the concentration 7 ng µL-1) was then added to 50 µL of the microdialysate. The mixtures were vortexted for approximately 10 s thereafter vigorously extracted at a frequency of 30 Hz for 1 min, using a MM301 vibration Mill (Retsch GmbH & Co. KG,

research articles

Metabolomic Patterns in Glioblastoma a

Table 1. Overview of Samples

day -5

Pat 1 T BAT SC Pat 2 T BAT SC Pat 3 T BAT SC Pat 4 T BAT SC Pat 5 T BAT SC Pat 6 T BAT SC Pat 7 T BAT SC Pat 8 T BAT SC Pat 9 T BAT SC Pat 10 T BAT SC Pat 11 T BAT SC

-4

-33

-22

-11

I x x x

x x x

x x x

x x x

x x x

x x x

I

I x x x

x x x

x x x

x x x

x x x

x x x

x x x

x x x

x x x

I

I

I x x x

x x I x x x

x x x x x x

I x x x

x x x I x x x I x x x

x x x

x x x

x x x

x x x x x x

I x x x

0

1

2

3

4

rt1

rt2

rt3

rt4

rt5

x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x rt1 x x x

rt2

rt2

rt2

rt2

rt2

rt2

rt2

x x rt3 x x x rt3 x x x rt3 x x x rt3 x

rt4

8

rt5

rt4

rt4

rt5 x x x rt5 x rt5 x x

rt4

rt5 x x

rt4

rt5 x x x rt3 x x x

rt2

rt4

rt3 x x x rt3 x x x

7

x rt4

rt2

rt2

6

x x

x rt3

rt3 x x x rt3 x x x

5

rt4

rt5 x x x rt4

rt5 x x x

rt5 x x x

a Summary: 172 samples in total (T ) 57, BAT ) 56, SC ) 59). I ) Catheter Implantation (typically around 12 noon). x ) Sample collected for analysis (between 4 and 8 pm)

Haan, Germany). After 120 min on ice, the samples were centrifuged at 19600 g for 10 min at 4 °C. A 200 µL aliquot of supernatant was transferred to a GC vial and evaporated to dryness. Methoxymation with 30 µL of methoxyamine solution in pyridine (15 µg µL-1) was carried out at room temperature for 16 h. Finally, the samples were trimethylsilylated with 30 µL of MSTFA at room temperature for 1 h, after which 30 µL of heptane (containing 0.5 µg of methyl stearate as injection IS) were added. GC-MS. Prior to analysis by GC-MS, the samples were divided into two separate batches, where samples from patients 1-5 constituted batch 1 (82 samples), and samples from patients 6-11 constituted batch 2 (116 samples). The samples in batch 1 were analyzed in the first GC-MS run and the samples in batch 2 in the second GC-MS run, which took place the following day. The run order within each batch was randomized. A 1 µL aliquot of derivatized sample was injected splitless by an Agilent 7683 Series autosampler (Agilent, Atlanta, GA) into an Agilent 6980 GC equipped with a 10 m × 0.18 mm i.d. fused-silica capillary column chemically bonded with 0.18 µm DB5-MS stationary phase (J&W Scientific, Folsom, CA). The

injector temperature was set at 270 °C. Helium was used as carrier gas at a constant flow rate of 1 mL min-1 through the column. For every analysis, the purge time was set to 60 s at a purge flow rate of 20 mL min-1 and an equilibration time of 1 min. The column temperature was initially kept at 70 °C for 2 min and then increased from 70 to 320 at 30 °C min-1, where it was kept for 2 min. The column effluent was introduced into the ion source of a Pegasus III TOFMS (Leco Corp., St Joseph, MI). The transfer line temperature was set at 250 °C and the ion source temperature at 200 °C. Ions were generated by a 70 eV electron beam at a current of 2.0 mA. Masses were acquired from m/z 50 to 800 at a rate of 30 spectra s-1, and the acceleration voltage was turned on after a solvent delay of 165 s. Retention indexes were calculated from the retention times obtained from the injection of a homologous series of n-alkanes (C12-C32) for each batch. All samples were run in randomized order. Hierarchical Multivariate Curve Resolution. Files of acquired data were exported to MATLAB 7.3 (R2006b) (Mathworks, Natick, MA) in NetCDF format for further data processing and analysis. All data pretreatment procedures, such as baseline correction, chromatogram alignment, time-window Journal of Proteome Research • Vol. 9, No. 6, 2010 2911

research articles setting, and hierarchical multivariate curve resolution (HMCR)6 were performed in MATLAB, using in-house scripts. The data acquired from the samples in batch 1 were subjected to H-MCR. Alignment and smoothing using a moving average was performed prior to dividing the chromatograms into 66 time windows from which a total of 183 chromatographic profiles (peaks, i.e., putative derivatized metabolites) with corresponding mass spectra were resolved. The samples from the second batch were then “predictively resolved” according to the H-MCR parameters obtained from the first batch, meaning that the same metabolites were quantified in the same way using the resolved mass spectra and retention indices from batch 1. This is a fast and efficient procedure for processing large series of metabolomic GC-MS data, with maintained high data quality, that has proven useful in global screening and for building diagnostic systems based on metabolic patterns.10,13,14 Prior to further multivariate analysis, all peak areas were normalized using the peak areas of the 11 IS that eluted over the whole chromatographic time range. Ultimately, the complete H-MCR process results in a data table (X), where samples are represented by rows and metabolites by columns. Each cell in the table corresponds to the calculated area under the deconvoluted chromatographic peak, of a specific metabolite in a specific sample. Furthermore, each resolved compound is associated with a corresponding mass spectral profile and a calculated retention index, which can be used in combination for identification of metabolites. For identification mass spectra of all detected compounds were mainly compared with spectra in our in-house mass spectra library database established at Umeå Plant Science Centre containing spectra and retention indexes from authentic standards run on the same instrument under the same experimental conditions as reported here. In addition we consulted the NIST library 2.0 (as of January 31, 2001), and the mass spectra library maintained by the Max Planck Institute in Golm (http://csbdb.mpimp-golm.mpg.de/csbdb/gmd/gmd. html). Positive identification was, in all cases, obtained by combining spectral match values with retention time index, calculated based on the alkane series (C12-C32). Orthogonal Partial Least-Squares. Orthogonal partial leastsquares (OPLS) is a supervised multivariate data projection method used to relate a set of predictor variables (X) to one or more responses (Y).15 By employing a response matrix (Y) that represents predefined sample classes, this method can be used for discriminant analysis (OPLS-DA) to predict class identity and to extract specific features distinguishing between the predefined sample classes. OPLS operates by dividing the systematic variation in X into two parts: one part that is linearly related to Y, and thus can be used to predict Y, and one part that is uncorrelated (i.e., orthogonal) to Y. In this process, each variable in X is associated with a weight, w*, which represents the variable’s covariation with Y. All calculated OPLS components were validated by 7-fold full cross-validation.16 In addition, the cross-validation process was used to assess the models ability to predict the response variation, expressed by the term Q2. Data Pretreatment and Analysis Strategies. The data matrix (X) produced by H-MCR was scaled to unit variance and screened for inconsistencies, using both principal component analysis (PCA)17 and orthogonal partial least-squares discriminant analysis (OPLS-DA) with leave-one-out cross-validation. Whenever evidently deviating samples or metabolites were 2912

Journal of Proteome Research • Vol. 9, No. 6, 2010

Wibom et al. encountered, the original data was properly scrutinized and appropriate action was taken. Data from samples collected prior to treatment initiation (time point 1-3) was studied for difference between sample types using PCA and OPLS-DA. Thereafter, to enable for analyses focused on treatment induced changes, the data in X was subjected to an additional pretreatment step. Samples from the same patient and sample type were categorized as either treated or untreated. Within each group of untreated samples, the obtained value from the first collected sample (time point 1 or 2) was subtracted from all samples in the group (including itself). Thereafter, the original value from time point 3 was copied and included as the first value in the group of treated samples, before the treated samples were subjected to the same transformation. The described procedure was carried out individually for each metabolite, and will herein be referred to as individual treatment over time (ITOT) normalization. Thereby, all metabolites were set to begin at zero for the first sample in both the treated and the untreated group, for each patient and sample type. The result was stored in a separate matrix (XITOT). XITOT was subsequently modeled by OPLS-DA in a series of steps to extract metabolites affected by treatment. The analyses were performed on tumor samples and BAT samples individually, and consisted of the following steps: (i) Exclusion of variables unaffected by treatment, that is, excluding variables with low model weight values (|w*| < 0.05) in an OPLS-DA model against treatment. (ii) Exclusion of variables displaying similar correlation to treatment also in the subcutaneously collected reference samples. This was achieved by means of two different OPLS-DA models with treatment as response, one based on SC samples and one on tumor/BAT samples. In a specific versus unique structures (SUS) plot,18 the w* vectors from the two models were compared and variables affected by treatment in both models were discarded. The cutoff values used were |w*| > 0.05 for correlation in tumor/BAT samples and |w*| > 0.1 for correlation in SC samples. (iii) Finally metabolites whose variation in concentration was similar before and after treatment was excluded. The remaining variables were analyzed in two separate OPLS-DA models, where the response variables were treatment and sampling sequence, respectively. By means of the SUS-plot, variables correlated to both treatment and sampling sequence were excluded (using the same cutoff limits as in step ii). Following this series of exclusion steps, the remaining metabolites were considered interesting as descriptors of treatment effect. To visualize the metabolic pattern associated with treatment, a final OPLS-DA model for each of the sample types (tumor and BAT) based on the selected metabolites was calculated and the cross-validated score plots from these models are presented herein for visualization and interpretation. Furthermore, to estimate the magnitude by which each of the selected metabolites was affected by treatment, we calculated normalized effect of treatment (NET) values. The NET value represents the quota between the observed treatment induced difference and the average baseline value, as calculated on ITOT normalized data. The treatment induced difference was calculated as the mean of all untreated samples subtracted by the mean of all treated samples. The achieved difference was subsequently divided by the absolute value transformed mean of all untreated samples. This combination of predictive data processing using H-MCR and multivariate predictions, here by means of cross-validated OPLS-DA scores, is known as predictive metabolomics and has

Metabolomic Patterns in Glioblastoma been developed for, and recently successfully applied to, studies of the human metabolome.13,14,19 In other words, predictive metabolmics in this sense refers to the application of chemometric strategies to all steps of a study,and not to the prediction of clinical end points. Validation and Evaluation. The metabolites that were highlighted as interesting in the analyses described above were further evaluated by two different methods for assessing significance. The first was based on a 95% confidence interval (CI) for each variable’s model loading (w*). The CI was estimated by jack knifing, and variables whose CI did not contain the value zero were considered as significant. The second significance test was a standard paired Student’s t test, where p-values