DOI:10.1158/0008-5472.

CAN-04-3016

Proteomic-Based Prognosis of Brain Tumor Patients Using Direct-Tissue Matrix-Assisted Laser Desorption Ionization Mass Spectrometry
Sarah A. Schwartz, Robert J. Weil, Reid C. Thompson, et al. Cancer Res 2005;65:7674-7681. Published online September 1, 2005.

Updated Version

Access the most recent version of this article at: doi:10.1158/0008-5472.CAN-04-3016

Cited Articles Citing Articles

This article cites 35 articles, 15 of which you can access for free at: http://cancerres.aacrjournals.org/content/65/17/7674.full.html#ref-list-1 This article has been cited by 10 HighWire-hosted articles. Access the articles at: http://cancerres.aacrjournals.org/content/65/17/7674.full.html#related-urls

E-mail alerts Reprints and Subscriptions Permissions

Sign up to receive free email-alerts related to this article or journal. To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department at pubs@aacr.org. To request permission to re-use all or part of this article, contact the AACR Publications Department at permissions@aacr.org.

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016

Research Article

Proteomic-Based Prognosis of Brain Tumor Patients Using Direct-Tissue Matrix-Assisted Laser Desorption Ionization Mass Spectrometry
Sarah A. Schwartz, Robert J. Weil, Reid C. Thompson, Yu Shyr, Jason H. Moore, 8 5 1,6 Steven A. Toms, Mahlon D. Johnson, and Richard M. Caprioli
Departments of 1Biochemistry, 2Neurosurgery, 3Biostatistics, 4Molecular Physiology and Biophysics, and 5Pathology and 6Mass Spectrometry Research Center, Vanderbilt University School of Medicine, Nashville, Tennessee; 7Surgical Neurology Branch, National Institutes of Neurological Disorders and Stroke, NIH, Bethesda, Maryland; and 8Brain Tumor Institute, Cleveland Clinic Foundation, Cleveland, Ohio
1,6 2,7,8 2 3 4

Abstract
Clinical diagnosis and treatment decisions for a subset of primary human brain tumors, gliomas, are based almost exclusively on tissue histology. Approaches for glioma diagnosis can be highly subjective due to the heterogeneity and infiltrative nature of these tumors and depend on the skill of the neuropathologist. There is therefore a critical need to develop more precise, nonsubjective, and systematic methods to classify human gliomas. To this end, mass spectrometric analysis has been applied to these tumors to determine glioma-specific protein patterns. Protein profiles have been obtained from human gliomas of various grades through direct analysis of tissue samples using matrix-assisted laser desorption ionization mass spectrometry (MS). Statistical algorithms applied to the MS profiles from tissue sections identified protein patterns that correlated with tumor histology and patient survival. Using a data set of 108 glioma patients, two patient populations, a short-term and a long-term survival group, were identified based on the tissue protein profiles. In addition, a subset of 57 patients diagnosed with high-grade, grade IV, malignant gliomas were analyzed and a novel classification scheme that segregated short-term and long-term survival patients based on the proteomic profiles was developed. The protein patterns described served as an independent indicator of patient survival. These results show that this new molecular approach to monitoring gliomas can provide clinically relevant information on tumor malignancy and is suitable for high-throughput clinical screening. (Cancer Res 2005; 65(17): 7674-81)

Introduction
Primary central nervous system tumors present significant challenges for diagnosis and treatment. Gliomas are the most common primary brain tumor with f25,000 new cases per year. Half of these tumors exhibit aggressive, infiltrative behavior, respond poorly to treatment, and are characterized as one of the more fatal human malignancies. Clinical diagnosis and treatment decisions for these tumors are based almost exclusively on tissue histology. Accurate glioma classification is therefore critical. Histologic diagnosis predominantly uses the WHO classification scheme (1),

which classifies gliomas based on the principal cell type and assigns them a grade of I to IV by increasing degrees of malignancy. Current methods for diagnosing gliomas, however, are highly subjective and depend on the skill of the neuropathologist. Considerable diagnostic variability also occurs because of the heterogeneous and highly infiltrative nature of these tumors. Moreover, no reliable molecular markers are known to exist that accurately categorize tumors in a clinically relevant fashion. There is therefore a critical need to develop more precise, nonsubjective, and systematic methods to classify human gliomas based on molecular markers. Recent advances have used cDNA microarrays to identify tumor gradespecific gene patterns (28) and potential markers of survival (911). In addition, two-dimensional gel electrophoresis technology has been used to analyze known tumor biomarkers for grade-specific trends (12) and monitor protein changes relative to tumor grade and patient survival (13). Most of these approaches rely on histopathologic analysis for accuracy verification and are low throughput, and some require a foreknowledge of the markers of interest. Few efforts have focused on developing a grade-independent, high-throughput, patient prognostic tool applicable in a clinical setting. We have developed a direct-tissue protein profiling approach to tumor analysis using matrix-assisted laser desorption ionization mass spectrometry (MALDI MS) to correlate protein patterns obtained directly from tumor biopsies with patient survival trends. This technology is characterized by high mass measurement accuracy (100-200 ppm; ref. 14), high sensitivity (attomoles to femtomoles), and a potential for high throughput (prepared samples can be analyzed in <5 minutes). The mass spectra acquired from tissue biopsies reflect a portion of the protein content within the tissue (i.e., f300-500 proteins). This technology has been used previously for imaging protein localization within a tumor biopsy (15), monitoring protein changes in mouse prostate (16) and rat pituitary and pancreas (17), and identifying tumor and prognostic specific biomarkers for patients with lung carcinomas (18). We report the identification of MALDI MS prognostic-specific protein patterns based on the analysis of 162 tissue biopsies from 127 patients. Protein patterns were identified that accurately classify glioma subtypes and distinguish patients into two prognostic groups, a short-term survival (STS) group and a long-term survival (LTS) group. In addition, we studied a well-characterized subset of patients with grade IV gliomas and identified protein patterns that predict differential survival.

Requests for reprints: Richard M. Caprioli, Mass Spectrometry Research Center, Vanderbilt University, 465 21st Avenue South, Room 9160, Medical Research Building 3, Nashville, TN 37232-8575. Phone: 615-322-4336; Fax: 615-343-8372; E-mail: r.caprioli@vanderbilt.edu. I2005 American Association for Cancer Research. doi:10.1158/0008-5472.CAN-04-3016

Materials and Methods

Materials. The MALDI matrix compound 3,5-dimethoxy-4-hydroxycinnamic acid (sinapinic acid), H&E, phenylmethylsulfonyl fluoride (PMSF),

Cancer Res 2005; 65: (17). September 1, 2005

7674

www.aacrjournals.org

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Brain Tumor Patient Prognosis Based on Protein Patterns NaCl, and ammonium bicarbonate were purchased from Sigma Chemical Co. (St. Louis, MO). DMEM was from Life Technologies, Inc. (Rockville, MD) and fetal bovine serum (FBS) was from Gemini Bio-Products (Woodland, CA). T-PER extraction buffer was purchased from Pierce Biotechnology, Inc. (Rockford, IL). Sucrose, ammonium acetate, and ultrapure Tris were obtained from J.T. Baker (Phillipsburg, NJ). Sequencing-grade trifluoroacetic acid (TFA) was from Burdick and Jackson (Muskegon, MI). Highperformance liquid chromatography (HPLC)grade acetonitrile was purchased from EM Science (Merck, Darmstadt, Germany). Sequencing-grade trypsin was from Promega (Madison, WI) and anti-PEA-15 was from Santa Cruz Biotechnology, Inc. (Santa Cruz, CA). Collecting and processing clinical material and patient information. Tissues were obtained, with informed consent and institutional review board approval, from patients undergoing tumor resection or other surgical procedures at Vanderbilt University Medical Center, Cleveland Clinic Foundation, and the NIH. A total of 162 tissue samples from 127 patients, including 19 patients undergoing resective surgery for nonneoplastic disease and 29 grade II, 22 grade III, and 57 grade IV glioma patients, were analyzed. Patient information was collected, including gender, age, treatment received before and after surgery, extent of surgery, current status (alive, alive with progressive disease, deceased, and cause of death), and survival from the time of original pathologic diagnosis. Samples were collected at the time of surgery, immediately snap frozen in liquid nitrogen, and stored at 80jC until analysis. Histopathologic diagnoses were made by a neuropathologist, blinded to the original clinical diagnosis, from subsequent H&E-stained sections according to the 2000 WHO classification (1) as described previously (19). Samples were prepared for MALDI analysis as described previously (19, 20). Briefly, frozen tissues were sectioned and transferred to MALDI target plates. Matrix droplets (0.1 AL saturated sinapinic acid in 50:50 acetonitrile/0.1% TFA in water, v/v) were blindly deposited on the surface of the sample, and the sections were dried. Optical section images were taken to align MS analysis regions with cellular morphology determined by histology. Samples were analyzed in a blinded fashion without knowledge of histologic diagnosis or clinical data. Mass spectrometry analysis and data processing. Each matrix droplet was analyzed on a MALDI time-of-flight (TOF) Voyager DE-STR mass spectrometer (Applied Biosystems, Foster City, CA) as described previously (19). Spectra were internally mass calibrated using the singly and doubly charged ions for a-hemoglobin (m/z 7,564.2 and 15,127.4, respectively), ubiquitin (m/z 8,565.8), and thymosin h4 (m/z 4,964.5, previously identified in human glioblastoma xenographs; ref. 15). Mass spectra were baseline corrected, smoothed, and normalized. The peak lists from each individual biopsy or patient, depending on the analysis approach, were averaged to generate one general protein profile. Statistical data analysis. Two independent supervised methods, symbolic discriminant analysis (SDA; ref. 21) and weighted flexible compound covariate method (WFCCM; ref. 22), were used to analyze the protein profiles. SDA applies genetic programming to determine discriminatory signals and builds functions using these signals that distinguish sample populations based on their classification. WFCCM applies multiple statistical tests to determine discriminatory markers. A linear combination of these markers is then generated that differentiates the sample groups. Further information is included in the supplementary data. Protein marker identification. Two samples were used for protein identification, a glioma cell line and a human glioma tissue sample. The glioblastoma cell line, U118 MG (American Type Culture Collection, Manassas, VA), was cultured in DMEM supplemented with 10% FBS and harvested using an extraction buffer [0.25 mol/L sucrose, 0.01 mol/L TrisHCl, 0.1 mmol/L PMSF (pH 7.6) at 4jC]. A cell aliquot was mixed 1:1 (v/v) with the extraction buffer, homogenized in an ice-chilled Duall homogenizer, and centrifuged at 10,000 g for 10 minutes at 4jC. The supernatant was collected for protein identification. A glioblastoma (grade IV glioma) tissue was collected at the time of surgery, frozen, and stored at 80jC. The tissue was homogenized in T-PER extraction buffer (50 mg tissue/1 mL T-PER) in an ice-chilled Duall homogenizer and centrifuged at 16,000 g for 30 minutes at 4jC. The supernatant was collected for protein identification. Both samples were separated in two dimensions, first by ion exchange chromatography followed by reverse-phase HPLC. The cell line supernatant was separated by anion exchange chromatography using a HiTrap Q HP anion exchange column (Amersham Biosciences, Uppsala, Sweden) and a NaCl gradient (0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.45, 0.55, and 1 mol/L NaCl) based on the extraction solution. HPLC separation for selected fractions was achieved over a Vydac (Hesperia, CA) 214MS52 reverse-phase C4 column (5 Am particles, 2.1 mm 25 cm) at 40jC using a linear gradient of 5% B to 20% B over 11 minutes, 20% B to 30% B over 15 minutes, 30% B to 55% B over 90 minutes, and 55% B to 95% B over 10 minutes. For fraction separation, solvent A was 0.1% TFA and solvent B was 0.1% TFA in acetonitrile. The tissue sample supernatant was separated by cation exchange chromatography using a HiTrap SP HP cation exchange column (Amersham Biosciences) with a linear gradient of 0% B to 100% B over 15 minutes, where A was 10 mmol/L ammonium acetate and B was 1 mol/L NaCl in 10 mmol/L ammonium acetate (pH 3.8 at room temperature). Selected fractions were separated over a Vydac 214MS5115 reverse-phase C4 column (5 Am particles, 1 mm 15 cm) at 40jC using a linear gradient of 5% B to 25% B over 5 minutes, 25% B to 60% B over 50 minutes, and 60% B to 95% B over 20 minutes. Fractions were analyzed by MALDI MS for the markers of interest after each separation. HPLC fractions of interest were reconstituted in 0.1 mol/L ammonium bicarbonate and digested with trypsin (1:50, trypsin/protein, w/w; 37jC; 16-20 hours). Digested fragments were analyzed using either an Applied Biosystems 4700 MALDI TOF/TOF mass spectrometer or a ThermoLTQ ion trap mass spectrometer equipped with a Thermo Surveyor LC pump and a microelectrospray source (Thermo Electron, San Jose, CA). MS and MS-MS spectra from MALDI TOF/TOF analysis were collected and the proteins were identified as described previously (23). The data were searched against the human National Center for Biotechnology Information database using the Mascot 9 database search algorithm. A significance cutoff score of 65 was used. Analysis on the ThermoLTQ mass spectrometer was done using one full MS scan followed by three MS-MS scans of the three most intense ions. MS-MS spectra were searched against the human database using Sequest (Thermo Electron) and the Sequest search outputs were filtered using a custom-designed software tool called Complete Hierarchical Integration of Protein Searches using the following filtering criteria: cross-correlation (X corr) value of >1.0 for singly charged ions, >1.8 for doubly charged ions, and >2.5 for triply charged ions. In addition, a RSp (ranking of preliminary score) value of <5 and a Sp value (preliminary score) >350 were required for positive peptide identifications. A minimum of two peptide matches and a positive correlation between the m/z ratio detected and the molecular weight of the intact protein (including post-translational modifications) were also required for protein identification. Immunohistochemistry. For immunofluorescence histochemistry, 18 Am thick sections were cut on a cryostat and incubated for 24 hours with PEA-15 antibodies (1:1,000). The sections were washed and incubated with Cy3-conjugated anti-mouse secondary antibodies (1:1,600; Jackson ImmunoResearch, West Grove, PA), washed, mounted, and coverslipped.

Results
Mass spectrometric profiling of human brain tissues. A total of 162 tissue samples from 127 patients (108 glioma and 19 nontumor patients) were collected and analyzed by MS. The general protocol is presented (Fig. 1). Tissue sections were coated with matrix droplets (typically 5-10 droplets were deposited on each section) and each droplet was directly analyzed by MALDI MS; serial sections were collected and stained with H&E for histopathology. Over 1,000 individual mass spectra, representing nontumor cell populations from nontumor patients or tumor cell populations from glioma patients, were used for comparative analysis. Between 300

http://www.matrixscience.com

www.aacrjournals.org

7675

Cancer Res 2005; 65: (17). September 1, 2005

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Cancer Research

Figure 1. Schematic of method. Protein profiles were collected directly from tissue sections and analyzed using classification programs for prognostic specific patterns. Tissue sections from tumor samples are collected and thawed onto MALDI target plates. Subsequent sections are collected for histologic staining and analysis. Multiple matrix droplets are deposited on each section and each droplet region is analyzed by MALDI MS. Collected spectra are processed and the resulting peak lists are used in a classification analysis.

and 500 individual protein signals in the m/z range of 2,000 to 70,000 were detected. The spectra were processed and multiple spectra were averaged to generate one peak list per patient or tissue sample depending on the statistical analysis done. To determine the variability of our tissue profiling approach, the average intraclass agreement rate (protein pattern variation within a patient sample) was measured to be 91.3 F 2.6% (95% confidence interval, 87.295.4%). This measurement suggests a relative spectrum-to-spectrum consistency within a given tissue sample. Correlating protein pattern changes to glioma classification. Initial data analysis focused on verifying that direct-tissue MALDI MS analysis could be used for tissue classification. Supervised classification analysis was done to identify tumor classification-distinctive biomarker patterns verified by histology. Data were processed, averaged by biopsy, and grouped into one of four categories: nontumor tissue (26 samples), grade II tumor (35 samples), grade III tumor (28 samples), and grade IV tumor (73 samples). Nontumor tissue refers to samples collected from patients undergoing surgical resection for nonneoplastic processes. For statistical analysis, biopsies were separated into training and testing data sets, consisting of two-thirds and one-third of the samples per classification, respectively. Pairwise comparisons were done on the training set to identify a subset of differentially expressed protein signals that best separated each classification: nontumor versus grades II, III, and IV biopsies; nontumor versus each individual tumor grade; grade II versus III; grade II versus IV; grade III versus IV; and grade II and III versus IV. Two independent methods were used for data analysis: SDA (21) and WFCCM (22). SDA uses genetic programming to build functions based on determined discriminatory signals, which distinguish sample classifications. WFCCM generates a model based on a linear combination of statistically determined discriminatory markers, which distinguishes sample groups. For each analysis approach, a model was defined that best classified samples in the training data set. Based on the model, each patient was assigned a score using the expression, or signal intensity, of the determined biomarker signals; the accuracy of this classification scheme was verified on a blind data set (testing data set). The results from these analyses are summarized (Table 1). Classification and prediction accuracies are defined as the number

of samples in the training and testing data sets, respectively, correctly classified. Biopsy protein patterns reflect a strong separation between tumor and nontumor tissues that extends to individual tumor grades. In all cases, nontumor tissues could be distinguished from gliomas with >92% classification accuracy. When comparing gliomas of different grades, the best separation was seen when comparing grade II and IV tumors (>93% classification accuracy), with slightly lower values for the grade III versus IV and grade II and III versus IV. The most difficult comparison was between grades II and III, which recapitulates the clinical situation. Accuracy limitations in protein profiling are due in part to the infiltrative nature of these tumors, the heterogeneous nature of the cells that comprise gliomas, and some methodologic limitations. Nonetheless, the results compare favorably with studies of interclass observer variability in pathology and neuropathology (24, 25). Interestingly, both analyses exhibited similar abilities in segregating the individual classifications. Models from SDA and WFCCM did well in distinguishing nontumor from tumor and generally separating individual tumor grades but did poorly in segregating grade II from grade III. Although the marker patterns determined by SDA and WFCCM were distinct, 35% of the classification-specific markers determined using SDA were also selected by WFCCM. These results suggest that classification based on protein profiling may be independent of the statistical analysis technique used. An independent agglomerative hierarchical clustering algorithm verified the statistically significant discriminator protein patterns, determined by WFCCM, in the training cohort for each of the classifications done. The results of three of these, nontumor versus tumor, nontumor versus grade IV tumor, and grade II versus grade IV tumor are shown (Fig. 2A-C , respectively). Clustering patterns reflect the strong correlation between the MS protein profile and the tissue classifications. Correlating protein pattern changes to glioma patient survival. Statistical analysis was then applied to the entire tumor data set of 108 glioma patients, with the spectra averaged by patient, to identify biomarker patterns that correlate to patient survival trends. Using WFCCM, a summary survival score was determined for each patient based on the statistically determined significant protein signals. Patients were then separated into short-term and

Cancer Res 2005; 65: (17). September 1, 2005

7676

www.aacrjournals.org

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Brain Tumor Patient Prognosis Based on Protein Patterns

long-term prognostic groups using a sensitivity analysis according to their correlated protein and survival patterns (Fig. 3A ). A proteomic pattern of 24 distinct MS signals distinguished patients based on survival trends from the time of pathologic diagnosis into two groups, STS (mean survival, <15 months) and LTS (mean survival, >90 months) groups. Seventeen of these markers were not determined as tumor- or tumor gradespecific discriminatory markers in the previous analyses. Several of these survival signal differences are displayed (Fig. 3C-E ), including an overexpression of m/z 9,747 and 10,092 in the STS group and an overexpression of m/z 10,262 in the LTS group. Within the total tumor patient population, analysis identified 52 patients in the STS prognostic group and 56 patients in the LTS prognostic group, with P < 0.0001. Univariate analysis showed a positive correlation (P < 0.01) between advanced patient age, increasing tumor grade, and tumors with an astrocytic lineage and shorter survival trends. Taking these factors in account, a multivariate Cox proportional hazards model showed a strong correlation between MS protein pattern and patient survival after adjustment for patient age, gender, tumor subtype, tumor grade, extent of tumor resection, and use of radiation and chemotherapy treatments before and after surgery. Therefore, the protein pattern served as an independent indicator of patient survival. The key patient survival variables are shown in a modified model (Table 2). Survival analysis of the glioma population from the time of surgery, in which the analyzed sample was resected, was also done with similar results (data not shown). Glioblastoma multiforme, the most common and malignant form of glioma, is also one of the most rapidly fatal of all human malignancies; median survival after diagnosis for these tumors is measured in months. For patients with a glioblastoma multiforme, age, clinical performance status, and extent of surgical resection are the principal, well-validated prognostic variables. WFCCM analysis was used to determine whether protein patterns could further differentiate patients based on survival from the time of glioblastoma multiforme presentation. A proteomic pattern of two distinct MS signals was identified that segregated the patients into a STS group (average survival, 10.9 months) and a LTS group (average survival, 16.8 months; Fig. 3B ). Neither of these signals was identified as a significant discriminatory marker in the previous analyses. A

total of 28 of the 57 patients were classified into the STS group and 29 patients were classified into the LTS group (P < 0.0001). Although an independent correlation existed between increasing patient age and shorter survival trends, the protein pattern was a powerful, independent predictor of patient survival. A multivariate Cox proportional hazards model showed a strong correlation between MS protein pattern and patient survival after adjustment for patient age, gender, extent of tumor resection, histologic subtype, and use of radiation and chemotherapy treatments. A modified form of this model is presented (Table 2). Identifying glioma biomarkers. Discriminatory protein identification was done using two protein sources, the malignant human glioma cell line, U118 MG, and a primary human grade IV glioblastoma sample. Both the cells and the tissue sample were homogenized and proteins were separated using a two-dimensional liquid chromatography approach, consisting of an ion exchange separation followed by reverse-phase HPLC separation. Fractions were monitored by MALDI MS during separation for the m/z signals of interest. Selected fractions were digested and analyzed by either an Applied Biosystems 4700 MALDI TOF/TOF mass spectrometer or a ThermoLTQ ion trap mass spectrometer. Six proteins were identified, including calcyclin (m/z 10,092), dynein light chain 2 (m/z 10,262), calpactin I light chain (m/z 11,073), astrocytic phosphoprotein PEA-15 (m/z 15,035), fatty acidbinding protein 5 (m/z 15,076), and tubulin-specific chaperone A (m/z 17,268; Fig. 4A ). Calcyclin, calpactin I light chain, and tubulin-specific chaperone A were identified as overexpressed in grade IV gliomas. On the other hand, astrocytic phosphoprotein PEA-15 was overexpressed in grade II and III tumors as opposed to grade IV gliomas and fatty acidbinding protein 5 was overexpressed in grade III tumors as opposed to grade IV. Calcyclin and dynein light chain 2 also discriminated between glioma survival subgroups with calcyclin predominant in STS patients and dynein light chain 2 overexpressed in LTS patients. The presence and relative expression levels for several of these proteins were verified by immunohistochemistry on intact tumor sections. For example, PEA-15 is shown to be in higher abundance in grade II astrocytomas compared with grade IV glioblastomas as recognized by the antibody staining pattern (Fig. 4B and C ). This

Table 1. Results of pairwise biopsy comparisons based on histopathology

Analysis Data set (no. biopsies training; testing set) SDA WFCCM

No. biomarkers Classification Prediction No. biomarkers Classification Prediction determined accuracy (%) accuracy (%) determined accuracy (%) accuracy (%) NT/T NT/TII NT/TIII NT/TIV TII/TIII TII/TIV TIII/TIV TII, TIII/TIV 18/91;8/45 18/24;8/11 18/17;8/11 18/50;8/23 24/17;11/11 24/50;11/23 17/50;11/23 41/50;22/73 2 2 4 2 2 3 4 1 92 92 94 96 76 93 85 79 89 84 95 84 77 82 80 80 28 26 38 41 61 17 32 62 96 100 97 99 88 97 96 93 92 84 89 87 50 82 76 78

NOTE: Biopsies were separated with two-thirds of the samples in the training data set and one-third in the testing data set and the displayed pairwise classifications were done. NT, nontumor tissue; TII, grade II glioma; TIII, grade III glioma; TIV, grade IV glioma. A subset of biomarkers that best distinguished the two groups was determined and the classification accuracy for both the training data set (classification accuracy) and the testing data set (prediction accuracy) was measured.

www.aacrjournals.org

7677

Cancer Res 2005; 65: (17). September 1, 2005

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Cancer Research

increase in protein expression correlates well with the presence of a MS signal at m/z 15,035 collected from a consecutive grade II tissue section as opposed to the loss of this signal in the grade IV section (Fig. 4D ).

Discussion
Direct-tissue profiling by MALDI MS has been used to analyze the protein patterns within human gliomas and correlate these patterns to tumor classification and patient survival. These data can be used to distinguish tumor tissue from nontumor brain tissue, define protein profiles specific to tumor grade, and identify differential patient survival patterns based on protein expression patterns. Because the accepted standard for glioma classification is histopathologic grading, we initially sought to validate the MS approach by identifying grade-specific biomarkers that correlated to histopathologically determined classifications. Tumors were subclassified by two neuropathologists, blinded to the original diagnosis, and analyzed by MALDI MS without knowledge of the original classification. Only samples with coincident clinical diagnoses were included in the analysis. Based on two independent classification approaches, WFCCM and SDA, MALDI MS provided specific proteomic patterns that classified glial tumors and nontumor brain tissue with high accuracy and precision. Proteomic profiles were used to discriminate between normal brain tissue and gliomas >92% of the time, with individual classification

accuracies between normal tissues and individual grades ranging from 92% to 100%. In addition, MALDI MS patterns were used to distinguished glioma grades with high accuracy ranging from 76% to 97%. The most difficult distinction was between WHO grade II and III tumors, which mimics the clinical situation. Statistical analysis identified >100 potential, tumor-specific biomarkers. Validation of MS-based tumor classification using two different statistical techniques highlights the power of protein profiling for tissue characterization independent of the analysis approach. WFCCM was also applied to identify MS-derived protein patterns that correlate to patient survival trends for all glioma patients and for a subgroup of patients with histologically confirmed glioblastoma multiforme. For all patients, standard treatment regimens were followed, including surgical tumor resection plus adjuvant radiotherapy and chemotherapy, as clinically indicated and tolerated. We show that a relatively small number of proteins can be used to distinguish between STS and LTS patients within the glioma patient population as a whole (P < 0.0001). Although these results are in line with previous clinical and pathologic studies, showing that the WHO grading system possesses discriminating predictive power, the protein pattern was an independent indicator of patient survival. In addition, MALDI MS protein profiling was used to analyze a large group of patients with the most malignant form of glioma, glioblastoma multiforme, and found that the MS pattern from two

Figure 2. Hierarchical cluster analysis of normal brain and tumor tissues. NT versus T (A ), NT versus TIV (B ), and TII versus TIV (C ) tissues in the training cohort were clustered according to the protein expression patterns determined by WFCCM analysis. Clustering was based on 28, 41, and 17 differentially expressed protein signals, respectively. Each row represents an individual protein signal characterized by a unique protein ID number (right ), and each column represents an individual patient tissue sample. The dendrogram (top ) clusters tissue samples based on similarity in protein expression profiles. The intensity of each MS signal is indicated by a gray scale, with black denoting the lowest intensity. N, nontumor; TII, grade II tumor; TIV, grade IV tumor.

Cancer Res 2005; 65: (17). September 1, 2005

7678

www.aacrjournals.org

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Brain Tumor Patient Prognosis Based on Protein Patterns

Figure 3. Kaplan-Meier survival curves and corresponding discriminatory mass signals for patient groups with a short-term or long-term prognosis according to MS proteomic patterns. Analyses were done to determine discriminatory protein patterns that stratified, based on patient survival trends, all glioma patients from the time of initial pathologic diagnosis using 24 protein signals (A ) and patients with grade IV glioblastoma from the time of glioblastoma multiforme presentation using 2 protein signals (B ). P < 0.0001 for each analysis. Several discriminatory signals that distinguish STS and LTS patients from the entire glioma population are shown, including m/z 9,747 (C ), 10,092 (D ), and 10,262 (E ). A total of six averaged spectra from three STS patients (solid line ) and three LTS patients (dashed line ) are presented.

m/z signals could further stratify patients into a STS group and a LTS group (P < 0.0001). For both analyses, the MALDI MS profile was the strongest determinant of survival in both univariate and multivariate analyses, stronger than most previously identified predictive variables, such as age, extent of resection, tumor grade, and use of adjuvant therapy. As expected, for the full glioma population, some overlap exists between grade-specific biomarkers and survival markers. Of the 24 discriminatory patient survival biomarkers for the entire glioma population, 17 were unique to the survival stratification. On the other hand, analysis of the glioblastoma multiforme population determined two unique markers that segregated the STS and LTS patients. These results suggest a novel approach to tissue classification based not on histopathologic features requiring visual analysis but on a molecular analysis of the protein patterns specific to the tissue sample. Based on statistical analysis, several discriminatory proteins were identified, including calcyclin, dynein light chain 2, calpactin I light chain, astrocytic phosphoprotein PEA-15, fatty acid binding protein 5 and tubulin-specific chaperone A. The MS signals from these proteins serve to discriminate gliomas from normal brain tissue and tumors of differing grade from one another; calcyclin and dynein light chain 2 also discriminated between glioma survival subgroups. These proteins are thought to be involved in several aspects of tumorigenesis. Calcyclin (S100A6), which plays a potential role in cell cycle progression and cell differentiation (26), is overexpressed in many tumors, especially at the margins of invasive cancers (2730). Dynein light chain 2, a subunit of the microtubule-associated dynein motor complex, binds and sequesters Bim, a proapoptotic protein, to negatively regulate its apoptotic function (31). Calpactin I light chain (p11, S100A10) is expressed in many cancer cell lines (32, 33) and is thought to bind and stimulate plasminogen conversion to plasmin, a cell surface proteinase involved in tumor cell

invasion and metastasis (34). PEA-15, an apoptosis inhibitor involved in several cell growth pathways (35, 36), is overexpressed in several tumor cell lines, including breast, larynx, cervix, and skin (36, 37), whereas studies have suggested overexpression of the fatty acidbinding protein 5 gene in prostate cancer tissue and cell lines (38, 39). Tubulin-specific chaperone A is a cofactor required for proper h-tubulin folding (40).

Table 2. Multivariate Cox proportional hazards model for the grade II, III, and IV tumor population and the grade IV tumor population alone
Variable Hazard ratio (95% confidence interval) P

Grade II, III, and IV gliomas MS protein pattern Age Gender Chemotherapy Radiation Tumor grade Grade IV gliomas MS protein pattern Age Gender Chemotherapy Radiation

1.002 1.034 0.926 0.296 1.402 4.581 1.014 1.063 1.551 0.898 0.799

(1.001-1.003) (1.006-1.062) (0.445-1.928) (0.107-0.814) (0.561-3.500) (1.440-14.575) (1.007-1.021) (1.026-1.021) (0.620-3.879) (0.327-2.467) (0.281-2.268)

<0.0001 0.0177 0.8366 0.0184 0.4697 0.0100 0.0001 0.0006 0.3485 0.8349 0.6728

NOTE: Analysis shows a strong correlation between MS protein pattern and patient survival after adjustment for several patient variables, including patient age, gender, tumor grade, and use of radiation and chemotherapy treatments before sample resection.

www.aacrjournals.org

7679

Cancer Res 2005; 65: (17). September 1, 2005

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Cancer Research

Figure 4. Summary of proteins identified and PEA-15 immunohistochemical staining. A, MS and MS/MS analysis of separated and digested fractions of interest identified six discriminatory proteins summarized in the table. The label m/z designates the m/z ratio monitored during direct-tissue profiling, classification group identifies which tissue group overexpressed the marker (whether a grade II, III, or IV glioma or the STS or LTS group determined from survival analysis of the entire glioma patient population), source material specifies whether the protein was identified from the glioma cell line (cell) or tumor biopsy (tissue), and the Swiss-Prot accession no. is as stated. B-D, for identification validation, two human glioma samples, a grade II astrocytoma and a grade IV glioblastoma multiforme, were sectioned and analyzed to correlate MALDI MS protein profiling with PEA-15 immunohistochemical staining. Fluorescent immunohistochemical images collected from a grade II astrocytoma (B ) and grade IV glioblastoma (C ) after staining for PEA-15 are presented. D, MS profiles collected by direct-tissue MALDI MS analysis of serial glioma tissue sections are also displayed; the protein profiles from the grade II glioma (gray trace ) and the grade IV glioma (black trace ) are presented. The intensity of m/z signal 15,035, identified in glioma cell lines as PEA-15, is increased in the MS profiles obtained from grade II compared with grade IV gliomas. Images from serial tissue sections, stained with anti-PEA-15, corroborate this by demonstrating a stronger PEA-15 staining pattern in the grade II tumor specimen compared with the grade IV sample.

Identification of these proteins was done in both a human glioblastoma cell line and a human glioblastoma tissue sample. These studies showed that, although cell lines are not ideal sources for protein identification, due to potential post-translational modifications and genomic mutations specific to the cell line, a positive correlation between the proteins identified from a cell line versus a tissue sample can exist. The identification of proteins from cell lines followed by further characterization of these proteins using traditional immunohistochemistry methods in intact tissues should serve as a valuable tool for protein identification and biomarker validation when resources are limited. Our analysis has several potential limitations. A rank cutoff was used in WFCCM to determine the number of protein signals used in each classification. Therefore, the number of peaks reported is based not on the smallest or largest number of signals that could discriminate the classes but rather on an intermediate number based on statistical evidence. It may be possible to achieve a similar classification rate using a different subset of peaks. Although a variety of variables could lead to the misclassified samples, potential limitations include the diffuse cellular nature of the tumors as well as histopathologic inaccuracy. Furthermore, the tumors were not analyzed for genetic alterations suspected of playing a role in gliomagenesis, which may have prognostic significance, nor did we control for histologic homogeneity or require a specific treatment regimen. Although it may have been useful to focus this study on a homogeneous study population, the mixed nature of the tumors more faithfully

corresponds to the clinical situation. Thus, this study represents an initial but important attempt to describe an unbiased molecular diagnostic tool that also possesses the power to predict overall outcome. In summary, MALDI MS protein profiling has been used to determine protein expression patterns that distinguish primary gliomas from normal brain tissue and one grade of gliomas from another, with high sensitivity and specificity. In addition, we have shown that a small subset of protein signals can be used to predict survival of glioma patients as well as to identify differential survival patterns within a more homogenous population of glioblastoma multiforme patients. Additional studies are under way to enhance and expand the MALDI MS analysis of a larger spectrum of human gliomas as well as to identify and characterize additional known and suspected peptide/protein biomarkers. Once identified, characterization of the function and biological role of the specific proteins involved in the origin as well as the progression of gliomas may permit their use as diagnostic and prognostic markers. Because MALDI MS technology is capable of analyzing numerous samples, with analysis times of f5 minutes per sample, this technology is amenable to high-throughput tissue screening in a clinical setting.

Acknowledgments
Received 8/20/2004; revised 6/30/2005; accepted 6/30/2005. Grant support: NIH/National Institute of General Medical Sciences grant GM 58008; NIH/National Cancer Institute/National Institute on Drug Abuse grant CA 86243; and Vanderbilt Physician Scientist Development Award, Vanderbilt-Ingram

Cancer Res 2005; 65: (17). September 1, 2005

7680

www.aacrjournals.org

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research

DOI:10.1158/0008-5472.CAN-04-3016
Brain Tumor Patient Prognosis Based on Protein Patterns
Cancer Center, and National Institute of Neurological Disorders and Stroke, NIH intramural research program (R.J. Weil). The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact. We thank Aurea Pimenta for growing the glioma cell lines, Salisha Hill and Lisa Zimmerman for assistance with the protein identification, Malin Andersson and Ariel Deutch for the immunohistochemistry, and Bashar Shakhtour, Jeremy Roberts, Bill White, and Huiming Li for their help in the bioinformatics analysis of these data.

References
1. Klieihues P, Cavenee WK. World Health Organization classification of tumours of the nervous system. Lyon: WHO/IARC; 2000. 2. Sallinen SL, Sallinen PK, Haapasalo HK, et al. Identification of differentially expressed genes in human gliomas by DNA microarray and tissue chip techniques. Cancer Res 2000;60:661722. 3. Shai R, Shi T, Kremen TJ, et al. Gene expression profiling identifies molecular subtypes of gliomas. Oncogene 2003;22:491823. 4. Rickman DS, Bobek MP, Misek DE, et al. Distinctive molecular profiles of high-grade and low-grade gliomas based on oligonucleotide microarray analysis. Cancer Res 2001;61:688591. 5. Markert JM, Fuller CM, Gillespie GY, et al. Differential gene expression profiling in human brain tumors. Physiol Genomics 2001;5:2133. 6. Kim S, Dougherty ER, Shmulevich I, et al. Identification of combination gene sets for glioma classification. Mol Cancer Ther 2002;1:122936. 7. Louis DN, Holland EC, Cairncross JG. Glioma classificationa molecular reappraisal. Am J Pathol 2001; 159:77986. 8. van den Boom J, Wolter M, Kuick R, et al. Characterization of gene expression profiles associated with glioma progression using oligonucleotide-based microarray analysis and real-time reverse transcription-polymerase chain reaction. Am J Pathol 2003;163:103343. 9. Nutt CL, Mani DR, Betensky RA, et al. Gene expression-based classification of malignant gliomas correlates better with survival than histological classification. Cancer Res 2003;63:16027. 10. Simmons ML, Lamborn KR, Takahashi M, et al. Analysis of complex relationships between age, p53, epidermal growth factor receptor, and survival in glioblastoma patients. Cancer Res 2001;61:11228. 11. Smith JS, Tachibana I, Passe SM, et al. PTEN mutation, EGFR amplification, and outcome in patients with anaplastic astrocytoma and glioblastoma multiforme. J Natl Cancer Inst 2001;93:124656. 12. Chakravarti A, Delaney MA, Noll E, et al. Prognostic and pathologic significance of quantitative protein expression profiling in human gliomas. Clin Cancer Res 2001;7:238795. 13. Iwadate Y, Yamaura A, Sato Y, Sakiyama S, Tagawa M. Induction of immunity in peripheral tissues combined with intracerebral transplantation of interleukin-2producing cells eliminates established brain tumors. Cancer Res 2001;61:876974. 14. Guilhaus M, Mlynski V, Selby D. Perfect timing: timeof-flight mass spectrometry. Rapid Commun Mass Spectrom 1997;11:95162.

15. Stoeckli M, Chaurand P, Hallahan DE, Caprioli RM. Imaging mass spectrometry: a new technology for the analysis of protein expression in mammalian tissues. Nat Med 2001;7:4936. 16. Todd PJ, Schaaff TG, Chaurand P, Caprioli RM. Organic ion imaging of biological tissue with secondary ion mass spectrometry and matrix-assisted laser desorption/ionization. J Mass Spectrom 2001;36:35569. 17. Caprioli RM, Farmer TB, Gile J. Molecular imaging of biological samples: localization of peptides and proteins using MALDI-TOF MS. Anal Chem 1997;69:475160. 18. Yanagisawa K, Shyr Y, Xu BGJ, et al. Proteomic patterns of tumour subsets in non-small-cell lung cancer. Lancet 2003;362:4339. 19. Schwartz SA, Weil RJ, Johnson MD, Toms SA, Caprioli RM. Protein profiling in brain tumors using mass spectrometry: feasibility of a new technique for the analysis of protein expression. Clin Cancer Res 2004; 10:9817. 20. Schwartz S, Reyzer M, Caprioli R. Direct tissue analysis using matrix-assisted laser desorption/ionization mass spectrometry: practical aspects of sample preparation. J Mass Spectrom 2003;38:699708. 21. Moore JH, Parker JS, Olsen NJ, Aune TM. Symbolic discriminant analysis of microarray data in autoimmune disease. Genet Epidemiol 2002;23:5769. 22. Shyr Y, KyungMann K. Weighted flexible compound covariate method for classifying microarray data. In: Berrar DP, editor. A practical approach to microarray data analysis. Norwell: Kluwer Academic Publishers; 2003. 23. Friedman DB, Hill S, Keller JW, et al. Proteome analysis of human colon cancer by two-dimensional difference gel electrophoresis and mass spectrometry. Proteomics 2004;4:793811. 24. Aldape K, Simmons ML, Davis RL, et al. Discrepancies in diagnoses of neuroepithelial neoplasms: the San Francisco Bay Area Adult Glioma Study. Cancer 2000;88:23429. 25. Castillo MS, Davis FG, Surawicz T, et al. Consistency of primary brain tumor diagnoses and codes in cancer surveillance systems. Neuroepidemiology 2004; 23:8593. 26. Tonini GP, Fabretti G, Kuznicki J, et al. Gene expression and protein localisation of calcyclin, a calcium-binding protein of the S-100 family in fresh neuroblastomas. Eur J Cancer 1995;31A:499504. 27. Camby I, Nagy N, Lopes MB, et al. Supratentorial pilocytic astrocytomas, astrocytomas, anaplastic astrocytomas and glioblastomas are characterized by a differential expression of S100 proteins. Brain Pathol 1999;9:119. 28. Komatsu K, Murata K, Kameyama M, et al. Expression of S100A6 and S100A4 in matched sam-

ples of human colorectal mucosa, primary colorectal adenocarcinomas and liver metastases. Oncology 2002; 63:192200. 29. Bronckart Y, Decaestecker C, Nagy N, et al. Development and progression of malignancy in human colon tissues are correlated with expression of specific Ca(2+)-binding S100 proteins. Histol Histopathol 2001;16:70712. 30. Stulik J, Osterreicher J, Koupilova K, et al. Differential expression of the Ca2+ binding S100A6 protein in normal, preneoplastic and neoplastic colon mucosa. Eur J Cancer 2000;36:10509. 31. Puthalakath H, Huang DC, OReilly LA, King SM, Strasser A. The proapoptotic activity of the Bcl-2 family member Bim is regulated by interaction with the dynein motor complex. Mol Cell 1999;3:28796. 32. Zhi H, Zhang J, Hu G, et al. The deregulation of arachidonic acid metabolism-related genes in human esophageal squamous cell carcinoma. Int J Cancer 2003;106:32733. 33. El-Rifai W, Moskaluk CA, Abdrabbo MK, et al. Gastric cancers overexpress S100A calcium-binding proteins. Cancer Res 2002;62:68236. 34. Zhang L, Fogg DK, Waisman DM. RNA Interference-mediated silencing of the S100A10 gene attenuates plasmin generation and invasiveness of Colo 222 colorectal cancer cells. J Biol Chem 2004;279: 205362. 35. Condorelli G, Vigliotta G, Cafieri A, et al. PED/ PEA-15: an anti-apoptotic molecule that regulates FAS/TNFR1-induced apoptosis. Oncogene 1999;18: 440915. 36. Ramos JW, Hughes PE, Renshaw MW, et al. Death effector domain protein PEA-15 potentiates Ras activation of extracellular signal receptor-activated kinase by an adhesion-independent mechanism. Mol Biol Cell 2000;11:286372. 37. Condorelli G, Vigliotta G, Iavarone C, et al. PED/ PEA-15 gene controls glucose transport and is overexpressed in type 2 diabetes mellitus. EMBO J 1998; 17:385866. 38. Vaarala MH, Porvari K, Kyllonen A, Vihko P. Differentially expressed genes in two LNCaP prostate cancer cell lines reflecting changes during prostate cancer progression. Lab Invest 2000;80:125968. 39. Vanaja DK, Young CYF. Protein and nucleic acid sequences of differentially expressed genes of human prostate cancer and methods and compositions for diagnosis, staging and prognosis of prostate cancer. Patent WO 2005007830. 2004 July 14. 40. Melki R, Rommelaere H, Leguy R, Vandekerckhove J, Ampe C. Cofactor A is a molecular chaperone required for h-tubulin folding: functional and structural characterization. Biochemistry 1996;35:1042235.

www.aacrjournals.org

7681

Cancer Res 2005; 65: (17). September 1, 2005

Downloaded from cancerres.aacrjournals.org on December 19, 2010 Copyright 2005 American Association for Cancer Research