Radiation-induced glioblastomas (RIGs) represent a significant proportion of glioblastomas (GBMs) seen in children and young adults and manifest poor prognosis. Little is known about their underlying biology, although limited studies have suggested no unique histologic or cytogenetic characteristics to distinguish them from de novo GBMs. In this study, we confirmed that a series of 5 RIGs showed no unique histologic or cytogenetic features compared with de novo pediatric GBMs, prompting us to further investigate RIGs using gene expression microarray profiling and Western blot analysis. Despite the inability of histologic and molecular genetic studies to identify distinguishing features between RIGs and pediatric GBMs, gene microarrays suggested significant differences between these 2 tumor types, at least those occurring in pediatric patients. Pediatric RIGs show greater homogeneity of gene expression than do de novo pediatric GBMs. Greater overlap was detected in gene expression patterns between RIGs and pilocytic astrocytomas than between RIGs and GBMs, medulloblastomas, ependymomas, atypical teratoid rhabdoid tumors, or rhabdomyosarcomas, suggesting a common precursor cell for RIG and pilocytic astrocytoma. Western blot analyses confirmed that ErbB3, Sox10, and platelet-derived growth factor receptor-α proteins were consistently expressed in RIGs but rarely in pediatric GBMs.
In pediatric oncology, CNS radiation therapy is often essential but carries risks of neurotoxicity. As more pediatric cancer survivors reach adulthood, there is growing concern that they are experiencing a number of chronic health conditions, the most devastating of which are secondary radiation-induced malignancies (1). Little is known about the etiology underlying secondary malignancy tumorigenesis. Pediatric patients are more susceptible to radiation-induced complications than adult patients and, generally, the younger the patient, the higher the incidence of and the shorter the interval to complications (2). A single-institution study of 1,612 patients with acute lymphoblastic leukemia treated at St. Jude Children's Hospital showed that there was a statistically significant increased risk of development of a secondary brain tumor with increased radiation dose (3). This institution experienced a particularly high incidence of secondary brain tumors when CNS radiotherapy was combined with antimetabolite agents (4).
Radiation-induced tumors of the nervous system include glioblastoma (GBM), anaplastic astrocytoma, sarcoma, meningioma, and schwannoma (5). Criteria for defining a radiation-induced tumor, presented by Cahan et al (6) and modified slightly since then, include the following: 1) the secondary tumor must originate in the field that was previously irradiated; 2) there is a histologically proven difference between the primary and secondary neoplasm; 3) there is a sufficiently long time period between irradiation and the onset of a secondary tumor presentation (usually >5 years); 4) the second differing tumor type is histologically confirmed; and 5) the patient must not have a mutator phenotype, such as that seen in Li-Fraumeni syndrome or retinoblastoma, for example, which favors the development of such neoplasms.
Most radiation-induced gliomas are high grade and have a particularly poor prognosis; sarcomatous components may be identified. Radiation-induced GBMs (RIGs) are particularly challenging for pediatric neuro-oncologists to treat and, in our experience, manifest a more aggressive course than de novo pediatric GBMs. Current chemotherapy regimens have little effect, and radiation therapy may not be an option in patients with RIGs because their maximum life-time dosages may have been exceeded with their prior irradiation (7). New therapies are clearly needed to improve survival and quality of life.
Defining the molecular characteristics of RIGs may provide novel targets for their treatment or prevention. The best study to date on RIG was conducted by Brat et al (8) who performed molecular analysis of 9 RIGs, in patients aged 9 to 60 years, to ascertain whether these lesions arose either from unique radiation-induced genetic alterations or matched those identified in de novo GBM. They found no significant difference between de novo GBM and RIG, apart from the possible absence of PTEN loss in RIG. Thus, they were unable to identify a diagnostic fingerprint for RIG. Comparative classification of gene expression profiles has been used previously to elucidate CNS embryonal tumor biology (9). Building on these previous studies, we have now used genome-wide gene expression microarray technology to characterize pediatric RIGs by comparison with other pediatric tumor types.
Materials and Methods
The following studies were performed using CNS tumor samples from patients who presented for treatment at The Children's Hospital (Denver, CO) between 1995 and 2006. All studies were conducted in compliance with internal review board regulations (COMIRB #95-500). Tumors were classified according to the World Health Organization international histologic tumor classification.
The 5 RIGs used in this study were obtained from patients who had originally presented with one of the following: Burkitt's lymphoma, medulloblastoma (MED), pilocytic astrocytoma (PA), ependymoma (EPN), and acute lymphoblastic leukemia (Table 1). Patient age ranged from 12 to 23 years at presentation of a RIG. RIGs developed 3 to 15 years after chemotherapy and radiotherapy as treatment for the original, primary tumor. All 5 patients died 1 to 10 months after presentation of their RIGs.
An additional 48 pediatric nervous system tumor specimens were used for comparison purposes in this study, including 5 de novo pediatric GBMs, 6 PAs, 15 EPNs, 9 MEDs, and 5 atypical teratoid/rhabdoid tumors (AT/RT) (all confirmed by loss of INI-1 protein by immunohistochemistry). We also compared RIGs with 8 rhabdomyosarcomas (RMS) from systemic locations to include a sarcoma comparison group.
Histologic and Cytogenetic Analysis of Tumor Specimens
All tissue sections were fixed in 10% formalin, embedded in paraffin wax, and stained with routine Harris hematoxylin and eosin, with immunohistochemistry for TP53 (Dako, Carpinteria, CA) and MIB-1 (Dako). MIB-1 cell cycle labeling indices were performed by 1,000 cell manual count. Fluorescence in situ hybridization analysis for epidermal growth factor receptor (EGFR) and loss of heterozygosity for 10q23 (PTEN locus) was conducted as described previously (10). Karyotype analysis was performed using classic cytogenetic culture-based techniques at the Colorado Genetics Laboratory (Denver, CO).
Gene Expression Microarray Analysis of Tumor Specimens
Tumor resections were snap-frozen in the operating room at the time of surgery and stored in liquid nitrogen until further use. RNA was extracted from these tumor specimens, labeled, and hybridized to Affymetrix HG-U133 Plus 2 microarray chips (Affymetrix, Santa Clara, CA) according to manufacturer's instructions, as described previously (11).
Data Transformation and Analysis
Data were exported to GeneSpring GX 7.3 bioinformatics software (Silicon Genetics, Redwood City, CA), which normalizes data from different experiments using multifilter comparisons. Data were normalized by 1) transformation of set measurements <0.01 to 0.01 and 2) "per-chip" normalization in which each chip is normalized to the 50th percentile of the measurements from that chip.
Three types of comparative analysis were then performed using the normalized microarray data: 1) analysis of homogeneity of RIG; 2) analysis of overexpressed genes in RIG, pediatric GBM, PA, and other pediatric CNS tumors; and 3) analysis of gene expression of molecules overexpressed in RIG that have previously been implicated in CNS development and/or tumorigenesis.
Comparative Analysis of Homogeneity in RIG Versus Other Tumor Types
The first analysis was performed to measure the homogeneity of RIG specimens compared to the homogeneity or heterogeneity of pediatric GBM, PA, EPN, MED, and AT/RT. This was achieved by comparison of the number of "signature" genes (i.e. those genes that are consistently over- and underexpressed) in the above tumor types. If the homogeneity in a given tumor type is high, then that tumor will have gene expression profiles that are relatively well conserved and will consequently have a high number of signature genes that are consistently over- or underexpressed. Conversely, in a heterogeneous tumor type, the number of signature over- or underexpressed genes will be low. This same type of analysis has been used successfully to assess intratumor sampling heterogeneity in gene expression profiles from patients with cervical, gastric, and lung cancers (12-14).
Over- and underexpressed genes were identified by pooling 5 gene expression profiles for the test tumor type (i.e. 5 RIGs) and comparing this to the control group, which consisted of the pooled gene expression profiles for all other tumor types (i.e. 5 GBMs, 5 PAs, 5 EPNs, 5 MEDs, and 5 AT/RTs). The number of gene expression profile replicates for each tumor type was set at 5 to avoid disproportionate contributions from either group in this analysis. A list of those genes that were greater than 2-fold over- and underexpressed in the test group versus the control group was generated. Analysis of variance was then applied to limit the list of genes to those with a statistically significant difference (p < 0.05). The same analysis was then performed for each of the tumor types consecutively. The numbers of signature genes for each tumor type were then compared as a measure of relative homogeneity within the group.
Venn Diagram Analysis of Overexpressed Genes in Radiation-Induced Glioblastoma Versus Other Pediatric CNS Tumors
The second analysis performed using microarray data was a comparative meta-analysis of the 100 highest overexpressed genes in RIGs (n = 5), pediatric GBMs (n = 5), PAs (n = 6), EPNs (n = 15), MEDs (n = 9), AT/RTs (n = 5) and RMSs (n = 8) to further measure the similarity and differences between RIGs and other pediatric CNS tumor types. The first step of this analysis was a rapid screen designed to identify those genes that were uniquely and highly overexpressed in the selected tumor versus other pediatric CNS tumors in a nonbiased manner. To achieve this comparison, we divided our gene expression profiles into 2 groups. The first group contained pooled gene expression profiles for the test group (i.e. RIG) and the second control group contained pooled gene-expression profiles of all other tumors used in the analysis (i.e. GBM, PA, EPN, MED, AT/RT, and RMS). By pooling tumors other than the selected tumor as a control, a baseline was created against which significantly overexpressed genes in that selected tumor could be identified. A larger number of tumor gene expression profiles were used in this analysis than in the homogeneity analysis to create a more robust control group. The magnitude of overexpression was ranked according to fold increase in expression of a particular gene in the test tumor that was statistically significant versus control by analysis of variance (p < 0.05). The 100 highest overexpressed genes according to this fold-change ranking were selected for the second step in the meta-analysis. The process was then repeated to identify the 100 highest overexpressed genes for each of the other tumor types. This approach has previously been used to identify differentially expressed genes in microarray data sets (10,15,16).
In the second step, GeneSpring software Venn diagram analysis was used to compare the overlap in expression in the 100 highest overexpressed genes for RIG versus each of the other tumor types as a measure of similarity of gene expression. Venn diagram analysis was also used to obtain a list of genes that are uniquely overexpressed in RIG by subtracting from the top 100 overexpressed genes in RIG any genes that were also in the top 100 overexpressed genes in all of the other tumor types. A list of genes that are shared between RIG and PA and between RIG and pediatric GBM was also created using the Venn diagram analysis.
Analysis of Selected Genes Overexpressed in Radiation-Induced Glioblastoma
The third analysis was to compare expression of genes identified by Venn analysis among all tumor types to validate and expand on the results of the above analysis. Genes were selected because they had been previously identified as having involvement in CNS development and/or tumorigenesis. Specifically, expression values for ERBB3, SOX10, platelet-derived growth factor receptor-α (PDGFRα), oligodendrocyte transcription factor 1 (OLIG1), oligodendrocyte transcription factor 2 (OLIG2), NK2 transcription factor-related locus 2 (NKX2.2), and BMP/retinoic acid-inducible neural-specific protein 3 (BRINP3) were obtained for individual specimens of RIGs (n = 5), GBMs (n = 5), PAs (n = 5), EPNs (n = 15), MEDs (n = 9), AT/RTs (n = 5) and RMSs (n = 8) tumors. Gene expression values, measured as normalized hybridization intensity, were averaged for each tumor type and are displayed as bar graphs.
Western Blot Analysis
Tumor protein lysates were obtained from snap-frozen tumor specimens of 5 RIGs, 5 pediatric GBMs, and 6 PAs. Approximately 50 mg of frozen tumor specimen was homogenized using a Rotor-Stator homogenizer (Kinematica, Lucerne, Switzerland) on ice in 500 μL of lysis buffer (50 mM HEPES, pH 7.5, 150 mM NaCl, 10 mM EDTA, 10% glycerol, 1% Triton X-100, 1 mM sodium vanadate, and 1 mM sodium molybdate with a complete Mini Protease Inhibitor Cocktail [Roche, Indianapolis, IN]). Twenty-five micrograms of the resulting protein lysates were then incubated with a loading dye according to the method of Laemmli (17), separated on precast 10% Criterion sodium dodecyl sulfate Tris-glycine gel (BioRad, San Diego, CA) and transferred to a polyvinylidene difluoride membrane (Millipore, Billerica, MA). Membranes were incubated with rabbit polyclonal anti-ErbB3 (sc-285; Santa Cruz Biotechnology, Santa Cruz, CA), rabbit polyclonal anti-Sox10 (CeMines, Golden, CO), rabbit polyclonal anti-PDGFRα (sc-338; Santa Cruz Biotechnology), or goat polyclonal anti-β-actin (sc-1616; Santa Cruz Biotechnology) sera at dilutions of 1:3000, 1:2000, 1:3000, and 1:1500, respectively. After incubation with secondary horseradish peroxidase-conjugated donkey anti-rabbit antibody (1:3000 dilution) (Jackson ImmunoResearch Laboratories, West Grove, PA) or donkey anti-goat antibody (1:4000 dilution) (Santa Cruz Biotechnology), immunoreactive signals of the protein bands were detected using enhanced chemiluminescence (PerkinElmer, Boston, MA) detection solution and exposed to X-Omat Blue XB-1 imaging film (Eastman Kodak, Rochester, NY).
Data for Affymetrix gene expression comparisons are presented as mean values ± SEM. p < 0.05 was considered statistically significant. Statistical analysis was performed using GraphPad Prism 4 statistics software (GraphPad Software, Inc., San Diego, CA).
Histologic and Cytogenetic Features of Radiation-Induced Glioblastoma
All 5 RIG specimens were classified as GBMs, showing fibrillary to small cell phenotypes (Fig. 1), with mitotic activity and either microvascular proliferation, necrosis, or both. One showed prominent pseudopalisading necrosis, and all manifested high levels of TP53 immunostaining affecting >25% of tumor cells. MIB-1 cell cycle labeling indices (1,000 cell manual count) ranged from 12.2% to 67.7% (12.2%, 27.6%, 42.1%, 56.4%, and 67.7%), with a median of 42.1%.
Fluorescence in situ hybridization analysis for EGFR and loss of heterozygosity for 10q23 (PTEN locus) was conducted. No evidence of EGFR amplification was found in 4 of 4 informative cases, and loss of 10q, including the PTEN locus, was found in 3 of 3 informative cases.
The striking difference between the primary tumor and the secondary RIG is typified by karyotype analyses on tumor samples from Patient 145 (Table 2). Cytogenetic studies of 3-day suspension cultures from Patient 145 from their original medulloblastoma in 1999 showed relatively simple cytogenetics. In contrast, their RIG manifested an exceedingly complex karyotype. Although cytogenetic data were not available for the primary tumors of Patients 12 and 478, the highly complex karyotype of the RIGs for Patients 12 and 478 showed no overlap with standard cytogenetic results for Burkitt's lymphoma or EPN, the primary tumors for Patients 12 and 478, respectively.
Gene Expression Profile Homogeneity Analysis of Pediatric CNS Tumors
Gene expression microarray profiles for RIGs, de novo pediatric GBMs, PAs, EPNs, MEDs, and AT/RTs were evaluated for signature over- and underexpressed genes to determine the relative level of homogeneity within different tumor types. The number of conserved signature genes in RIGs was found to be approximately 2-fold higher than that in pediatric GBMs (Fig. 2). The number of conserved genes seen in RIGs was very similar to that seen in PAs and EPNs. This finding demonstrates that RIGs represent a more conserved entity and have a more tightly defined gene expression profile than de novo pediatric GBMs, which have more variable gene expression profiles. Thus, it can be concluded that RIGs show significantly more homogeneity than pediatric GBMs.
Meta-Analysis of Highest Overexpressed Genes in Radiation-Induced Glioblastoma Versus Other Pediatric CNS Tumors
Meta-analysis of the 100 highest overexpressed genes in RIGs versus those in other pediatric CNS tumors was performed using Venn diagram analysis (Table 3). Surprisingly, significant overlap was found in the gene expression signatures for RIG and PA. Thirty-nine of the 100 overexpressed genes were found to be shared between these 2 tumor types. By contrast, pediatric GBM only shared 2 genes with RIG and 8 genes with PA. MED showed a small amount of overlap with RIG (4 genes), and EPN, AT/RT, and RMS showed no overlap. To evaluate the contribution of false discovery to this type of analysis, the same analysis was performed but with randomized samples. The highest overlap observed between 2 randomized groups was 10 of 100 genes. The strong overlap observed between the 100 highest overexpressed genes in RIG and PA is therefore greater than those genes that can be accounted for by false discovery. Furthermore, the overlap observed between RIG and pediatric GBM, between RIG and MED, and between pediatric GBM and PA cannot be considered significant.
Three gene lists were generated from the Venn diagram analysis of: 1) genes uniquely overexpressed in RIG (Table 4); 2) genes with shared overexpression in RIG and PA (Table 5); and 3) genes with shared expression in RIG and pediatric GBM (Table 6).
Analysis of ErbB3 and Sox10 Expression in Radiation-Induced Glioblastoma
Gene lists generated from the Venn diagram analysis were examined for genes that had previously been implicated in astrocyte development. It was found that both ERBB3 and SOX10 were highly overexpressed in RIG and to a lesser extent in PA compared with all other brain tumors, both astrocytic and nonastrocytic (Fig. 3A, B). This was a significant finding, as paired ErbB3 and Sox10 overexpression had previously been identified in PA and is thought to be involved with the tumorigenesis of this astrocyte lineage (11).
Expression of ErbB3 and Sox10 was further evaluated in RIGs at the protein level using Western blot analysis (Fig. 4). This analysis demonstrated that ErbB3 protein was expressed in all (5 of 5) RIGs and was highly expressed in 3 of 5 specimens. In contrast, ErbB3 protein expression in pediatric GBMs was marginal in 1 of 5 specimens and absent from the remainder. ErbB3 protein was highly expressed in 5 and absent from 1 of 6 PA samples. Sox10 protein expression showed a similar pattern of expression to ErbB3, being moderately expressed in all 5 of 5 RIGs. Moderate Sox10 protein expression was observed in 1 of 5 pediatric GBMs, but only marginally expressed in the remainder. In PA, Sox10 protein was moderately expressed in 4 of 6 samples and marginally expressed in the remainder.
Analysis of Platelet-Derived Growth Factor Receptor-α Expression in Radiation-Induced Glioblastoma
PDGFRα, which is targeted by the small molecule inhibitor imatinib (Gleevec) (18), was present in the list of genes that were overexpressed only in RIGs (Table 4). PDGFRα showed the highest overexpression in RIGs and was overexpressed to a lesser extent in PA and AT/RT (Fig. 3C).
Western blot analysis of PDGFRα protein in 5 RIG, 5 pediatric GBM, and 6 PA specimens demonstrated that PDGFRα was very highly expressed in 2, moderately expressed in 2, and marginally expressed in 1 of 5 RIGs (Fig. 4). In contrast, PDGFRα protein was marginally expressed in 1 of 5 pediatric GBMs and absent from the remainder. Moderate PDGFRα expression was observed in 2, marginally expressed in 2, and absent from 2 of 6 PAs.
Gene Expression of Molecules Involved in CNS Development and/or Tumorigenesis
Examination of those genes overexpressed in RIGs revealed a number of genes that are potentially involved in CNS development and/or tumorigenesis. It was found that oligodendroglial lineage markers OLIG1, OLIG2, and NKX2.2 and neuronal marker BRINP3 were highly overexpressed in RIGs compared with pediatric GBMs (Fig. 3D-G). Furthermore, PAs also exhibited overexpression of these markers, although at a lower level than for RIGs.
Development of effective molecular targeted therapy for RIG mandates better characterization of the molecular biology of this tumor. The few biologic studies of RIG to date have identified no molecular markers with which to characterize RIG (8,19). In the largest molecular analysis of RIGs to date, Brat et al (8) found that alterations in p53, KRAS, EGFR, and p16 in RIG were similar to alterations in de novo GBM, apart from an absence of PTEN mutations in RIG. This led them to conclude that "the spectrum of molecular genetic alterations [in RIG] appear to be similar to that described in spontaneous high grade astrocytomas, especially those of the de novo type." The findings of the present study correlated with those of Brat et al, in that no histopathologic or cytogenetic features of RIG were found to distinguish them from de novo GBM. Similar to that study, in the present study we also noted the absence of EGFR amplification. Brat et al performed mutational analysis of the PTEN gene in 9 RIGs and found no mutations, contrasting with the relative frequency of such mutations in de novo GBMs, but noted their small sample size. We also suffered from small sample size in our study, but did identify loss of heterozygosity of the PTEN site in 3 of 3 RIGs examined. Because Brat et al did not investigate the allelic status of chromosome 10, there is no discrepancy between our findings and theirs. We did not test for KRAS point mutations or p16 deletions, as did Brat et al, because these parameters had not been shown to be distinctive in RIGs. We turned to global gene expression analysis as a possible means by which distinguishing biologic characteristics of RIG might be identified, and our microarray analytical approach demonstrated marked dissimilarities between RIG and de novo pediatric GBM.
Histologic heterogeneity is a hallmark of de novo GBMs (20), and comparison of the entire transcriptome of de novo pediatric GBMs did indeed show gene expression pattern heterogeneity in this category of tumors. In contrast, despite the varied histologic appearance also observed in the 5 RIGs in our series, analysis of gene expression profiles demonstrated that these 5 cases showed a significantly higher level of homology with each other than that observed among pediatric GBMs. In other words, RIG as an entity displays considerably less genetic heterogeneity than de novo pediatric GBMs. In fact, we found that the degree of homogeneity of RIGs paralleled the genetic homogeneity observed for PAs and EPNs. The finding that RIGs have a more tightly conserved molecular pathology than de novo pediatric GBMs may suggest a common, shared tumorigenic origin and pathway for RIGs. This may be related to the fact that RIGs have a common initiating event, i.e. radiation, versus the varied etiologic mechanisms that are likely to cause de novo pediatric GBMs.
This study also underscores the ability of gene expression microarray analyses to suggest possible cellular lineages that are being recapitulated in RIG tumorigenesis. Surprisingly, analysis of genes overexpressed in RIGs demonstrated a significant overlap with genes overexpressed by PAs (39%) and minimal overlap with genes overexpressed by de novo pediatric GBMs (2%). Almost no overlap in gene expression was detected between RIGs and either MEDs, EPNs, AT/RTs, or RMS. ErbB3 and Sox10, which have been shown to regulate development of the neural crest (21), were both found to be overexpressed in RIG and PA, but not in de novo pediatric GBM. This finding is significant, as paired ErbB3 and Sox10 overexpression was recently identified as a molecular characteristic of PA, suggesting possible recapitulation of developmental pathways in the tumorigenesis of PA (11). Oligodendroglial lineage markers OLIG1, OLIG2, PDGFRα, and NKX2.2 (22,23) and neuronal lineage marker BRINP3, a developmentally regulated neural-specific gene (24,25) are also overexpressed in RIG and PA. Collectively, these data suggest that RIG and PA may share a common precursor cell, although obviously the grade differences (World Health Organization grade I vs IV) and very different clinical outcomes for these 2 tumors indicate significant differences in their tendencies to accrue additional genetic alterations, once tumorigenesis is initiated.
Finally, the conserved molecular pathology of RIG that we identified in this study lends hope that a common therapeutic target might be possible for these otherwise poor-prognosis tumors. One of the most interesting features of our study was our identification of elevated protein expression in PDGFRα and ErbB3 in RIGs compared to that in de novo pediatric GBMs. Although speculative, the possibility of using novel therapeutic agents such as imatinib, a small molecule inhibitor of PDGFR (18), to treat RIGs is raised. Additionally, ErbB3, which requires heterodimerization with either ErbB1 or ErbB2 for activity, could be indirectly targeted by using the dual tyrosine kinase inhibitor lapatinib (26).