Characteristics of Samples and Central Pathologic Review

A neuropathologist independently blindly reviewed all samples to ensure consistent classification based on contemporary guidelines from the WHO.²¹ Only grade 4 astrocytomas (GBM) were used. Clinical findings of all patients with GBM and control brains are shown in Tables 1 and 2 and the Appendix (online only).

Cell Line, Protein Extraction, Sodium Dodecyl Sulfate–Polyacrylamide Gel Electrophoresis, Immunoblot, and Immunohistochemical Analysis

U87 was grown as described.²⁵ Extracts were prepared from cell pellets or from human brain tissue and processed as previously described.¹⁰ Immunohistochemical analyses for pErk, pAkt, glial fibrillary acidic protein (GFAP), p53, and Y-Box-Protein-1 (YB1) were performed as described in the Appendix (online-only) and in Pollack et al²¹ and Sutherland et al.²⁶

Laser Capture Microdissection RNA Extraction, and Linear Amplification

Frozen sections were processed as described previously.²⁷ The neuropathologist identified tumor cells for capture (Fig A1A, online only). The rest of the block was used for protein analysis. RNA was extracted from all samples and subjected to two rounds of T7-RNA-polymerase amplification. Fidelity and reproducibility of RNA amplification has been shown²⁸ and was further validated as described in the Appendix and Figures A1B and A1C (online-only).

Microarray Analyses

RNA from 14 pGBM and seven aGBM fresh-frozen (FF) samples was compared with the same reference pooled RNA extracted from three healthy brains from children ages 1, 7, and 15 years. Cy3- or Cy5-labeled cDNA probes from samples and pooled controls were hybridized to Human 19K cDNA spotted arrays (19,008 genes/ESTs, University Health Network, http://www.microarrays.ca). Slides were scanned and intensity quantified using QuantArray (GSI Luminomics, Billerica, MA). Inversion of fluors in distinct cDNA probes were performed (dye swap) to account for nonspecific dye-associated effects on hybridization and signal detection. The LOWESS scatter-smoothing algorithm in the GeneSpring 7.0 software package (Agilent, Santa Clara, CA) normalized the raw fluorescence data. We analyzed 60 hybridizations consisting of dye-swap hybridizations of 30 biologic replicates for all 21 GBM samples (analysis of several samples was duplicated to assess for reproducibility from two different RNA extractions and amplifications; Fig A1B-A1C; Appendix). “Filter on Confidence” and analysis of variance (ANOVA) statistical tools in GeneSpring identified genes with reproducible changes in transcript abundance. GeneSpring was also used to perform hierarchical clustering and principal components analysis (PCA). Quantitative real-time polymerase chain reaction (RT-PCR) was performed to validate relative gene expression (Appendix).

Activation of Ras and Akt Pathways Occurs in Only a Subset of pGBM and Is Associated With Poor Outcome

We investigated activation of the Ras and Akt pathways in 18 frozen pGBM tumors by Western blot analysis using phosphorylation-specific antibodies against known effectors of these pathways. U87 aGBM cell-line, and a pooled protein lysate from three pediatric control brains (1, 7, and 15 years) were used as positive and negative controls respectively. Slides from the same blocks were stained with hematoxylin and eosin, confirming that only lysates from pGBM tumor populations were being analyzed. Phosphorylation of Raf, Mek1/2, and Erk1/2 was seen in 12 of 18 pGBM samples and in the U87 cell line, indicating activation of the Ras pathway (Fig 1; data not shown). Phosphorylation of these downstream effectors of Ras was not observed in six of 18 pGBM samples. Phosphorylation of Akt and GSK3 was observed in 10 samples (Fig 1; data not shown), all of which were also active for Ras and had lower levels of phosphate and tensin homolog (PTEN), the dual-activity phosphatase known to participate in inactivating the Akt pathway.²⁹

We validated results obtained by Western blot using an immunohistochemical approach on 18 formalin-fixed, paraffin-embedded (FFPE) pGBM samples that included four samples previously tested by Western blot on their frozen counterparts, and another new set of 14 pGBM for which no frozen counterparts were available (Appendix). Sections from samples were tested for pErk (associated with Ras activation), pAkt (associated with Akt activation), and glial fibrillary acidic protein (GFAP; astrocytic marker) immunoreactivity. Five samples showed no staining for pErk, while showing normal staining for GFAP of adjacent sections, excluding problems related to tissue preservation (Fig A1D). The remaining samples showed strong staining for pErk (Fig 2). pErk-positive regions contained spindle-like tumor cells showing atypical elongated nuclei and a fibrillary staining typical of GBM and active Ras.³⁰ pAkt was present in 13 of 18 samples that were also positive for pErk (Fig 2). Where material for both technical approaches was available, the same results were obtained.

When investigating putative prognostic factors, striking results were obtained for Ras activation. Children with active Ras (21 of 32) had poor survival with only one survivor, whereas five of 11 with no Ras activation are alive and disease free, with a follow-up of at least 4 years (log-rank P < .009). This result reflects Akt activation because most patients with active Ras also had active Akt (19 of 21). Sex, younger age, treatment, and p53 expression were not associated with better survival with the limitation of sample size.¹⁸ Availability of frozen tissue from pGBM is limited. To determine whether the data obtained on Ras and Akt activation in pGBM has potential prognostic value, we performed the same immunohistochemical analysis on an independent data set of 21 FFPE pGBM samples (Table 2; Appendix). pErk and pAkt were highly immunoreactive in 13 of 21 of these samples, whereas both stainings were negative in eight of 21 samples (Fig 2). Survival data further validated our data on the association of Ras and Akt activation with poor survival in pGBM (Table 2). Results from all patients suggest that there are at least two forms of pGBM: one form with poor survival, with only one patient alive, associated with an active Ras/Akt pathways (35 [66%] of 53 samples for Ras and 31[58.5%] of 53 for Akt) and a second form with better prognosis (nine of 18 children alive and disease free), without Ras or Akt activation (18 [34%] of 53; log-rank P < .0001; Fig 3; Tables 1 and 2).

Transcriptional Profiling Distinguishes Two Subsets of pGBM on the Basis of Their Association With Ras/Akt Activity and Reveals a Molecular Signature for pGBM That Is Distinct From aGBM

To study gene expression changes, we selected an approach that allows us to compare changes between tumor and nontumor brain as well as changes between pGBM samples. RNA from 14 pGBM frozen samples was isolated and hybridized to Human 19K cDNA spotted arrays together with the same reference pooled RNA extracted from three control pediatric brains. Laser capture microdissection (LCM) was used to selectively capture malignant astrocytes. RNA was subjected to two rounds of T7 linear amplification (aRNA) to circumvent the limited amount of frozen material.^31-33 GeneSpring's Filter on Confidence tool identified 2,593 transcripts with statistically significant changes in abundance in the 14 pGBM samples compared with the pooled control (Welsch t test P < .0001; multiple testing correction, Benjamini and Hochsberg; false-discovery rate, 3.4%). Two-dimensional hierarchical clustering organized and visualized the profiles of these transcripts (y-axis) from each of the 14 samples (x-axis; Fig 4) and indicated a high degree of homogeneity in pGBM. Other statistical algorithms (the Wilcoxon-Mann-Whitney test of significance analysis of microarrays) were tested with the same results. We analyzed the data set using a module-level view obtained from a cancer compendium³⁴; and also organized the gene sets using GoMiner, a computer resource that incorporates the hierarchical structure of the Gene Ontology Consortium³⁵ to automate a functional categorization of gene lists based on biologic processes. Both methods aim to distill a higher order for the analysis of complex data sets. They yielded similar results, showing, as expected, that pGBMs are actively proliferating tumors (Fig 5; Appendix Tables A1 to A4, online only).

To visualize similarities within samples, we used PCA, a method of data reduction in which the high dimensionality of the data is reduced to two to three viewable dimensions representing linear combinations of genes that account for most of the variance of the data set. PCA separated samples into two groups, indicating the presence of at least two distinct populations of pGBM (Fig 6A). These populations were associated with Ras/Akt activity in a sample (Ras active color coded in yellow, inactive in red) and not with other known markers for pGBM (necrosis, proliferation index, p53 expression, age).^18,36 ANOVA testing identified 1,437 transcripts that could distinguish tumors associated with differing Ras activity (Welch t test P cutoff of .0001; multiple testing correction, Benjamini and Hochsberg). Most transcripts were distinct from those that distinguished pGBM from control brain (Fig 6A; Fig A2A, online only). Gene ontology terms analysis shows that transcripts associated with Ras/Akt active tumors exhibit enhanced rates of nucleic acid, protein synthesis, and metabolism (Tables A1 and A2, online-only). With the limitation of sample size, ANOVA testing of expression profiles did not reveal a difference based on treatment, necrosis, age, or p53 overexpression.^15,21

To determine whether pGBM is molecularly distinguishable from aGBM, we performed microarray analysis on seven aGBM samples using the same LCM/RNA amplification approach and the same reference control aRNA. Using scatter plots, several of the changes in transcript abundance that distinguish tumors from control brain are maintained between aGBM and pGBM (Fig A2, online-only). Using PCA, components 2 and 3 showed that aGBM (coded in blue) cluster separately from both types of pGBM (Fig 6A). Furthermore, ANOVA testing using the same stringency as before identified a sharp signature of 1,569 transcripts that distinguished pGBM from aGBM (Figs 6A and A2; Tables A1-A4; Appendix). These transcripts combined with the 1,437 genes that distinguished pGBM on the basis of their association with Ras activity were separated by conditional hierarchical clustering (Fig 7; Table A5, online only). A first tree-branching separated Ras-nonactive pGBM samples from Ras-active samples. A second branching in samples associated with Ras activation separated pGBM from aGBM. One long-term survivor aGBM clustered with Ras-nonactive samples. Sample showed weak pErk staining (Fig A3, online only).

Data from transcriptional profiling mirror data obtained on the protein level indicate that there are at least two subgroups of pGBM and that both differ molecularly from aGBM.

Validation of the Data Set and Identification of Targets Involved in pGBM Gliomagenesis

There are no reports of gene expression profiling that specifically target pGBM. We compared our results to recent series on adult high-grade gliomas (aHGG). Using the common gene name as a correspondence marker for an independent data set of 31 aHGGs,⁹ we found an overlap between both data sets (Fig A3, online only). Phillips et al¹³ identified prognostic subgroups of aHGG: tumors segregated into subclasses on the basis of their preferential expression of genes characteristic of neural tissue (PN; favorable prognosis), proliferating cells (Prolif; poor-risk disease), or mesenchymal tissues (Mes; poor-risk disease). Authors likened Mes, Prolif, and PN signatures to those of neural stem cells, transit-amplifying cells, and immature neurons, respectively. We applied their classification scheme to our data set (Figs 8 and 9). PCA using their 108 probes also separated pGBM samples from aGBM samples and both subsets of pGBM. aGBM, as expected, had the Mes or Prolif signatures, as did Ras-active pGBM (Fig 8). However, this subset of pGBM could still be distinguished from aGBM. Conditional clustering with Pearson correlation using the Mes and Prolif probe sets separated aGBM from both subsets of pGBM (Fig 9). PGBM associated with no Ras/Akt activation lacked the PN signature associated with better prognosis in aGBM. These data validate our data set and show that, aside from the lack of Akt activation, the pGBM subset associated with a better survival has a unique molecular profile that is distinct from known markers of long-term survivors in aGBM.

PGBM Associated With No Ras/Akt Activation Express No Neural Stem-Cell Markers and Upregulate Apoptosis-Associated Transcripts

Recent evidence suggests that HGG may arise from stem cell–like cancer cells at multiple stages of differentiation.³⁷ Tumors associated with Ras/Akt activation and poor prognosis overexpressed markers associated with neural stem cells, including CD133, nestin, maternal embryonic leucine zipper kinase (MELK), vimentin, and Dlx2, whereas the subset of pGBM associated with no activation of Ras/Akt and better outcome only showed increased vimentin expression (Fig 10). ^13,37,38 Expression levels were validated by quantitative RT-PCR (qRT-PCR) for all of these transcripts (data not shown). Using GoMiner, differentially modulated transcripts associated with Ras-nonactive pGBM showed overexpression of gene sets involved in apoptosis. Conversely, aGBM and pGBM associated with a Ras-active pathway showed overexpression of gene sets involved in protein synthesis, translation, transcription, DNA repair, and synthesis (Tables A1 to A4, online-only).

YB1 Is Overexpressed in pGBM and May Increase EGFR Expression in Akt-Active Samples

YB1 showed increased expression (3.8-fold compared with healthy brain) in 12 of 14 pGBM samples, but not in aGBM (Fig 6B). YB1 was further investigated because it is involved in brain embryogenesis and contributes to oncogenesis in a range of epithelial cancers, potentially through Akt mediated phosphorylation.^26,39-41 We validated expression of YB1 by qRT-PCR and by immunohistochemistry (Figs 11 and 12; Tables 1 and 2). YB1 was overexpressed in 26 of 32 samples with mostly nuclear localization in samples with active Akt, and was cytoplasmic in samples with no active Akt. These data were corroborated by results obtained on the independent data set of 21 pGBM samples: 12 of 21 positive with nuclear staining for YB1 in Akt-active samples (Table 2). Nuclear YB1 increases expression of several genes including EGFR.⁴² We found EGFR overexpression by qRT-PCR and immunohistochemistry mainly in Akt-active samples (Tables 1 and 2).

Characteristics of GBM Samples and Controls

Thirty-nine samples had a positive central review of GBM and included 25 FF and 14 FFPE specimens. In addition, FFPE sections were available from the same surgical procedure for four of the FF specimens. Samples and complete pathology reports were obtained from the Pediatric Cooperative Human Tissue Network, the London/Ontario Tumor Bank, and from collaborators in other centers. Patients' ages ranged from 6 months to 16 years for pediatric samples (n = 32) and from 30 to 72 years for adult samples (n = 7). FF and FFPE healthy brain specimens were obtained from children aged between 1 and 15 years from the same sources after surgery for epilepsy. All samples were obtained with informed consent after approval of the institutional review board of the respective hospitals in which the patients were treated. Only the three adult patients with secondary GBM had previously received radiotherapy.

Sections (5 μmol/L) for immunohistochemical analysis from an independent data set of 21 FFPE pGBMs were obtained from European collaborators. Clinical characteristics of these patients are provided in Table 2.

Pediatric control brains were obtained from the London Ontario Tumor Bank and from the Montreal Children's Hospital. All control samples were also reviewed by the neuropathologist (S.A.) to ascertain for quality of tissues, astrocytic content, and absence of tumor material. Six control brains from children aged between 1 and 15 years were included in this study because they had both frozen and FFPE blocks available, an astrocytic content above 80% as ascertained by GFAP staining, and absence of reactive gliosis. Three of these samples were pooled and used as controls for Western blot (Fig 1) and microarray analysis (Figs 2 to 4) and individually stained with the various antibodies in the immunohistochemical analysis. The three other samples were used as additional controls in the microarrays analysis.

Cell Lines and Antibodies

Human GBM cell line U87 (a gift from Del Maestro, MD, PhD, McGill University) was grown as previously described (Shi Q, Bao S, Maxwell JA, et al. J Biol Chem 279:52200-522009, 2004). Unless otherwise stated, all antibodies used were obtained from Cell Signaling (New England Biolabs, Pickering, Canada). Anti-GFAP monoclonal antibody was obtained from Advanced ImmunoChemical Inc (Long Beach, CA).

Total Protein Extraction, Sodium Dodecyl Sulfate–Polyacrylamide Gel Electrophoresis, and Immunoblot Analysis

Extracts were prepared from cell pellets or from human brain tissue using ice-cold RIPA lysis buffer supplemented with phosphatase and protease inhibitors as previously described (Rajasekhar VK, Viale A, Socci ND, et al. Mol Cell 12:889-901, 2003). One hundred micrograms of protein lysates were separated in a 10% sodium dodecyl sulfate (SDS) -polyacrylamide gel, proteins transferred onto polyvinylidene fluoride (PVDF) filter type immobilon-P transfer membranes (Millipore, Billerica, MA), and immunoblotted with the recommended dilution of antibodies overnight at 4°C as described elsewhere (Yim SH, Ward JM, Dragan Y, et al. Toxicol Pathol 31:295-303, 2003). Antirabbit or mouse horse-radish peroxidase conjugated secondary antibodies (Boehringer Manheim, Manheim, Germany) were used and the cross reactivity was visualized by enhanced chemoluminescence (Amersham, Piscataway, NJ) on Biomax MR films (Kodak; Amersham).

Immunohistochemical Analysis

Immunohistochemical analyses for pErk, GFAP (astrocytic marker), and p53 were performed as described previously (Pollack IF, Finkelstein SD, Woods J, et al. N Engl J Med 346:420-427, 2002; and Choe G, Horvath S, Cloughesy TF, et al. Cancer Res 63:2742-2746, 2003). Slides were counterstained with hematoxylin and mounted. Positive (U87) and negative controls (immunoglobulin G) were included with each batch of sections to confirm the consistency of the analysis. GFAP, a histologically verifiable internal positive control antigen, was used to identify cases in which a lack of immunoreactivity for pErk and pAkt might indicate problems linked to the labeling of the tissue and, thus, tissue preservation, rather than a lack of protein phosphorylation. Importantly, all assays were carried out at the same time with the same reagents. The neuropathologist, blinded to outcome and histology, evaluated the degree of staining. For YB1 expression, staining and scoring of slides were performed as previously described (Sutherland BW, Kucab J, Wu J, et al. Oncogene 24:4281-4292, 2005).

RNA Extraction and Amplification From Samples

Approximately 30 ng of RNA were extracted from one slide (PicoPure RNA Isolation Kit; Arcturus, Sunnyvale, CA) and subjected to two rounds of T7 RNA polymerase amplification (RiboAmp RNA Amplification Kit; Arcturus) following the manufacturer's protocol. DNase treatment was included. Isolated RNA was quantified by the Ribogreen RNA Quantitation Kit (Molecular Probes, Carlsbad, CA) on an LS50 luminescence spectrometer (PerkinElmer, Waltham, MA). After the first-strand synthesis of the first round, the integrity of the starting material was assessed by PCR reaction with β-actin primers. aRNA yield was also measured with Ribogreen RNA Quantitation Kit and RNA integrity by the BioAgilent RNA assay for each sample. An average amplification yield of 80 μg of aRNA was obtained, and the size of the amplification products checked by loading 2 μL on a formaldehyde agarose gel. We were able to have quality RNA for microarray analysis from 14 of the 18 FF pGBM samples. Fidelity of linear aRNA has been previously shown (Hu X, Pandolfi PP, Li Y, et al. Neoplasia 7:356-368, 2005). We also validated our material by comparing microarray analysis performed on total RNA and amplified RNA extracted from the same sample (Fig A1B) and for three samples on amplified RNA extracted from two separate LCM/amplification procedures from the same sample, as previously described (Fig A1C; Watanabe H, Tanaka F, Doyu M, et al. Hum Genet 107:452-457, 2000).

qRT-PCR

Fifty nanograms of aRNA from each tumor sample were submitted to one-step RT-PCR. Reverse transcription and amplification of aRNA were performed (Quantitect SYBR Green RT-PCR kit, Qiagen, Mississauga, Ontario, Canada). Experimental conditions for each set of gene-specific primers were optimized with the one-step RT-PCR Kit (Qiagen). Primers sequences were designed in the 3′ end of each gene with the help of Primer3 Software (http://frodo.wi.mit.edu/primer3/input.htm).

Each reaction was done in triplicate on a Mx4000 Multiplex Quantitative PCR System (Stratagene, Cedar Creek, TX). After 40 cycles of amplification, an additional step allowing dissociation curve analysis was performed. Specificity of the amplification process was evaluated by dissociation curve analysis and agarose gel electrophoresis. Fold changes were calculated using the standard curve method. Total RNA from three aGBM cell lines (U87, SF-126, and SF-539) were pooled. A 10-fold serial dilution of this mixture was used to construct a standard curve for both reference and target genes. Amplification efficiencies (E) were calculated as E = 10 ^(−1/S), where S represents the slope of the standard curve. Only efficiencies within 90% and 110% were accepted. Quantities of target were calculated by plotting the Ct values to the corresponding standard curve. Tumoral quantities were then normalized to an endogenous control (β-actin). Fold change is given by dividing the normalized target quantity by the value of the calibrator (nontumoral tissue control). The normalized target amount of the calibrator is set to the value 1.

⇓

⇓

⇓

⇓

⇓

⇓

⇓

⇓