To identify novel glioma-associated pathomechanisms and molecular markers, we performed an array-based comparative genomic hybridization analysis of 131 diffuse astrocytic gliomas, including 87 primary glioblastomas (pGBIV), 13 secondary glioblastomas (sGBIV), 19 anaplastic astrocytomas (AAIII) and 12 diffuse astrocytomas (AII). All tumors were additionally screened for IDH1 and IDH2 mutations. Expression profiling was performed for 74 tumors (42 pGBIV, 11 sGBIV, 13 AAIII, 8 AII). Unsupervised and supervised bioinformatic analyses revealed distinct genomic and expression profiles separating pGBIV from the other entities. Classifier expression signatures were strongly associated with the IDH1 gene mutation status. Within pGBIV, the rare subtype of IDH1 mutant tumors shared expression profiles with IDH1 mutant sGBIV and was associated with longer overall survival compared with IDH1 wild-type tumors. In patients with IDH1 wild-type pGBIV, PDGFRA gain or amplification as well as 19q gain were associated with patient outcome. Array-CGH analysis additionally revealed homozygous deletions of the FGFR2 gene at 10q26.13 in 2 pGBIV, with reduced FGFR2 mRNA levels being frequent in pGBIV and linked to poor outcome. In conclusion, we report that diffuse astrocytic gliomas can be separated into 2 major molecular groups with distinct genomic and mRNA profiles as well as IDH1 gene mutation status. In addition, our results suggest FGFR2 as a novel glioma-associated candidate tumor suppressor gene on the long arm of chromosome 10.
Diffusely infiltrating astrocytic gliomas are the most common primary brain tumors in adults and cover a spectrum of malignancy grades, ranging from diffuse astrocytoma of World Health Organization (WHO) grade II (AII), over anaplastic astrocytoma WHO grade III (AAIII) to glioblastoma WHO grade IV (GBIV). Clinically, 2 types of glioblastomas are distinguished, i.e., primary glioblastomas (pGBIV) that arise de novo with a short clinical history and secondary glioblastomas (sGBIV) that develop by progression from a preexisting lower grade glioma.1 Both types share a similarly poor prognosis when patients are matched for age and clinical status at the time of glioblastoma diagnosis.2 However, pGBIV and sGBIV demonstrate distinct profiles of genetic aberrations. For example, pGBIV show frequent EGFR amplification, PTEN mutation and typically loss of an entire chromosome-10 copy,3, 4 while sGBIV are characterized by more frequent TP53 and IDH1 mutation as well as losses on 10q but not always 10p.1, 5, 6 Several recent studies employed large-scale genomic and mRNA expression profiling for the molecular characterization of astrocytic gliomas, in particular high-grade tumors.4, 7–10 Distinct gene expression signatures were found that may stratify high-grade glioma patients into prognostically distinct subgroups7, 10–12; however, the relationship between such molecular subgroups and histological tumor grade, tumor evolution (primary versus secondary glioblastoma), as well as mutation status of the IDH1 gene is less well defined.
Here, we report on a comprehensive profiling of astrocytic gliomas of WHO grades II, III and IV using array-based comparative genomic hybridization (array-CGH), microarray-based expression profiling, and focused molecular analyses of selected candidate genes.
AII: astrocytoma WHO grade II; AAIII: anaplastic astrocytoma WHO grade III; array-CGH: array-based comparative genomic hybridization; FGFR2: fibroblast growth factor receptor 2; IDH: isocitrate-dehydrogenase; MGMT: O-6-methylguanine-DNA methyltransferase; NBR: non-neoplastic brain; OS: overall survival; PAM: prediction analysis of microarrays; PDGFRA: platelet-derived growth factor receptor alpha; pGBIV: primary glioblastoma WHO grade IV; SAM: significance analysis of microarrays; sGBIV: secondary glioblastoma WHO grade IV; TP53: tumor protein p53
Material and Methods
Frozen samples of 131 astrocytic gliomas with a histologically determined tumor cell content >80% were retrieved from the tumor collections at the Departments of Neuropathology, Heinrich-Heine-University Düsseldorf, Charité Universitätsmedizin Berlin, Germany, and the International Agency for Research on Cancer (IARC), Lyon, France. All samples were histologically classified according to current WHO criteria. Follow-up data were retrospectively determined and linked to molecular data in an anonymized manner, as approved by the institutional review board of the Medical Faculty, Heinrich-Heine-University, Düsseldorf. The clinical data are summarized in Supporting Information Tables S1 and S2. For a more detailed description see the Supporting Information Methods.
Array-CGH and gene expression profiling were performed as previously reported.13, 14 Data were analyzed based on EnsEMBL (version 51) using packages of the Bioconductor project15 implemented in our in-house developed ChipYard framework for microarray data analysis (http://www.dkfz.de/genetics/ChipYard/) and deposited in the NCBI Gene Expression Omnibus (GEO) database, accession no. GSE15698510.
The MGMT and FGFR2 promoter methylation status was assessed as reported.16 DNA samples were sodium bisulfite-treated, PCR-amplified and sequenced. Results were scored according to the ratio of the cytosine (methylated) to thymidine (unmethylated) peak at each CpG site. Transcript levels of FGFR2 variants FGFR2-IIIb (NM_022970.3) and FGFR2-III (NM_000141.4) were determined by qRT-PCR with SYBR Green I fluorescent dye. The presence of mutations affecting IDH1 codon 132 or IDH2 codon 172 was assessed by direct sequencing. The TP53 gene (exons 4–10) was screened for mutations by SSCP analysis, followed by sequencing of aberrant SSCP bands.17
Statistics and bioinformatics
Unsupervised hierarchical clustering was performed with the pvclust algorithm18 using Pearson correlation coefficients and Ward's minimum variance method. For multidimensional scaling, the metaMDS function of the r-package vegan19 was used. Classifier gene signatures were selected using the nearest shrunken centroid classifier algorithm20 with 10-fold cross-validation. Survival analyses were performed using the survival package21 and log-rank tests. For pathway analysis, Ingenuity® software (www.ingenuity.com) was used.
Array-CGH profiling identifies 2 major astrocytic glioma groups
DNA-copy number profiles of 131 astrocytic gliomas were generated using a bacterial artificial chromosome (BAC) microarray. AII, AAIII and sGBIV demonstrated widely distributed but similar chromosome aberrations which increased in frequency from low-grade to higher-grade tumors (Fig. 1a). In contrast, pGBIV showed clear-cut genomic profiles that were distinctly different from those of the other astrocytoma groups. Between pGBIV and sGBIV, particularly striking differences were observed at 10p and 19q. 10p was deleted in 75/86 (87%) of pGBIV as opposed to 3/13 (23%) of sGBIV; 5/13 (38%) of sGBIV even carried gains on 10p. At 19q, pGBIV commonly show copy number gains, while copy number losses are frequent in sGBIV. Additional differences mapped to 7p, 8q, 12p and 20q (Fig. 1a). In individual tumors, several novel multicopy aberrations were identified, for which the respective target genes have not been described so far (Tables S3 and S4).
Unsupervised classification of astrocytic gliomas based on gene expression
Gene expression profiling was performed for 74 astrocytic gliomas and 4 non-neoplastic brain tissue samples. Unsupervised cluster analysis identified 2 major subgroups, cluster-1, containing the vast majority of pGBIV and cluster-2, almost entirely consisting of AII, AAIII and sGBIV (Fig. 1b). Non-neoplastic brain samples (NBR) formed a separate cluster closely related to cluster-2. All clusters were supported with bootstrap probabilities of 85% or higher. Principle component analysis of the data yielded virtually identical results (Supporting Information Fig. S1A). Prediction analysis of microarrays (PAM) revealed a 304-gene classifier signature (heatmap shown below dendrogram) suggesting differential activity of genes related to growth, proliferation and cellular movement (Table S5). Within cluster-2, 2 subclusters were distinguished, 1 consisting of WHO grade III and IV tumors (cluster-2high), the other of WHO grade II and III tumors (cluster-2low). Gene expression differences between these subclusters primarily affect cell-cycle related genes (Table S6).
IDH1 and IDH2 mutation status associates with gene expression signatures
Point mutations of IDH1 codon 132 were found in the majority of AII (10/12), AAIII (16/19) and sGBIV (11/13), as well as a few pGBIV (7/87). One of the 2 IDH1 wild-type sGBIV carried a codon 172-mutated IDH2 gene. Analysis of MGMT methylation and TP53 mutation revealed significant positive correlations of IDH1 mutation with both MGMT methylation (p < 0.001) and TP53 mutation in the entire tumor group (p = 0.018, Fisher's exact test). The frequency of MGMT promoter methylation was higher in expression cluster-2 tumors (28/33) when compared with expression cluster-1 tumors (19/40; p = 0.001). Similarly, TP53 mutation was more common among the investigated cluster-2 tumors (10/28 vs. 6/32, p = 0.054). However, the most striking association was observed between the IDH1 and IDH2 mutation status and the overall gene expression patterns (Fig. 1b). In cluster-1, 39 of 40 tumors had wild-type IDH1 and IDH2 genes, whereas 30 of 34 tumors in cluster-2 carried IDH1 or IDH2 mutations (p < 0.001) (Fig. 1b and Fig. S1). Notable exceptions to this pattern were groups of 3 and 2 pGBIV positioned near the left and right margins of cluster-2high, respectively. The former 3 tumors lacked IDH1 or IDH2 gene mutations and at least 2 of them showed cluster-1-like expression patterns. PAM revealed that these tumor samples probably joined cluster-2 because they share with the other tumors of this cluster the expression of an immune gene signature that is not included in the 304-gene classifier. The other 2 pGBIV in cluster-2 were from young patients (ages 26 and 49), carried IDH1 mutations and showed a cluster-2 gene expression pattern (Fig. 1b). Five additional pGBIV, for which only array-CGH data were available, also carried IDH1 mutations; these tumors also derived from young patients.
Taken together, these data suggest that 5 major classes of diffusely infiltrating astrocytic gliomas may be distinguished based on clinicopathological features and the IDH1 and IDH2 mutation status (Table 1), i.e., IDH1 or IDH2 mutant AII, AAIII and sGBIV, as well as pGBIV with and without IDH1 or IDH2 gene mutations. IDH1-mutant pGBIV patients (pGBIV-IDHmut) represent a novel class characterized by young age and markedly longer OS than IDH1 wild-type pGBIV patients (pGBIV-IDHwt). Only 6 tumors (5%) did not fit into 1 of our 5 categories, including 3 AAIII without IDH1 or IDH2 mutations. Histological review of these cases confirmed the initial diagnosis; however, it cannot be excluded that they represented incompletely sampled and thus underdiagnosed pGBIV.
Table 1. Clinicopathological features
IDH1 mutant pGBIV share characteristics of both sGBIV and IDH1 wild-type pGBIV
For further characterization of the major types of diffuse astrocytic gliomas, we focused on the 124 tumor samples that exactly matched our classification criteria (Table 1). Analysis of DNA-copy number frequencies revealed significant differences (p < 0.01; Fisher's exact test) between sGBIV and pGBIV-IDH1wt at 10 loci (Fig. 2a). For pGBIV-IDH1mut, it currently is unclear whether they are more similar to pGBIV, as suggested by their clinical history, or to sGBIV, as suggested by their IDH1 mutation status and gene expression profiles, or whether they represent an independent group. At 7 of the 10 examined loci, copy number aberrations occurred with significantly different frequencies (p < 0.05) in pGBIV-IDH1wt and pGBIV-IDH1mut, but not pGBIV-IDH1mut and sGBIV. This suggests that pGBIV-IDH1mut tumors might represent sGBIV cases, in which a preexisting lower-grade lesion had not been detected. However, analysis of multicopy number aberrations (Fig. 2b) showed that at least 1 pGBIV-IDH1mut carried EGFR amplification and PTEN homozygous deletion, features that are typical for pGBIV but not sGBIV. Furthermore, pGBIV-IDH1mut lacked amplification at 12p including the CCND2 locus, an aberration more often found in sGBIV (Fig. 2b). However, the fact that most clearly distinguished pGBIV-IDH1mut from other glioblastomas was their long median overall survival of 1,386 days, which was much longer compared with pGBIV-IDH1wt (265 days) and sGBIV (277 days) patients (Table 1).
Associations of a reported prognostic expression signature with tumor group and IDH1 gene mutation status
Our cluster-1 versus cluster-2 classifier (Table S5) considerably overlaps (19/35 genes) a published prognostic classifier for grade III and IV astrocytomas.7, 10 To effectively compare the 2 classifiers, we mapped the genes of the signature reported by Phillips et al.10 to our data set and performed unsupervised cluster analysis based only on the 35 classifier genes. The analysis yielded 2 major tumor clusters that were highly correlated with the IDH gene mutation status (Fig. S2) and similarly composed as those of our own analysis (Fig. 1b). This suggests that, similar to our signature, the signature of Phillips et al.10 classifies tumors largely according to clinicopathological features and IDH1 mutation status.
Molecular markers associated with overall survival in IDH1 and IDH2 wild-type pGBIV
Survival correlations were restricted to pGBIV-IDH1wt tumors, because this was the only group that included a sufficient number of patients with available follow-up data (n = 65 patients). Long-term survivors (>2 years) were deliberately excluded, since pGBIV pathogenesis in these patients might be based on distinct molecular mechanisms.22 Univariate analysis of OS relative to chromosome aberrations detected by array-CGH identified copy number gain (p = 0.031) or amplification (p = 0.035) at the PDGFRA locus on 4q12 to be associated with adverse prognosis, while gains at 19q13 loci were associated with more favorable outcome (p = 0.020) (Fig. 3). In multivariate analysis, congruous results were obtained (gain PDGFRA: p = 0.034, HR = 2.096; ampPDGFRA: p = 0.037, HR = 2.548; gain 19q13: p = 0.008, HR = 0.382), while the parameters patient age, extent of resection and Karnofsky index did not reach significance.
To identify associations of gene expression and OS, we performed a long- versus short-survival PAM based on the data from 29 pGBIV-IDHwt patients. The resulting classifier was enriched in genes with roles in cell migration and the regulation of actin cytoskeleton stability (Table S7). Survival correlations also were tested for each signature gene by univariate analysis.
Homozygous deletion and reduced expression of FGFR2 in glioblastomas
Array-CGH profiling identified 2 pGBIV-IDHwt with homozygous deletions involving the FGFR2 locus at 10q26.13 (Fig. 4a). In addition, we found frequent down-regulation of FGFR2 transcripts in glioblastomas relative to normal brain in the microarray data set. Lower relative expression of FGFR2 was significantly associated with adverse prognosis in univariate analysis of pGBIV-IDH1wt patients (n = 29; p = 0.021) (Fig. 4b). Gene expression analysis by qRT-PCR specific for the epithelial (FGFR2-IIIb) and mesenchymal (FGFR2-IIIc) isoforms of FGFR2 confirmed significant down-regulation in pGBIV relative to non-neoplastic brain tissue for both transcript variants (Figs. 4c and 4d). Sodium bisulfite sequencing of the FGFR2 promoter region showed methylation in more than 50% of 33 investigated CpG sites in 1/10 AII, 1/11 AAIII and 3/11 sGBIV but none of 43 pGBIV (data not shown).
Using array-CGH and gene expression profiling, we provided strong support for the hypothesis that astrocytic gliomas can be subdivided into 2 major molecular subgroups (pGBIV versus AII/AAIII/sGBIV) and that pGBIV and sGBIV are molecularly distinct tumor entities.1 Consistent with tumor progression, unsupervised cluster analysis subdivided the AII/AAIII/sGBIV group into a low-proliferative subgroup consisting of AII and a subset of AIII (cluster-2low) and a high-proliferative subgroup consisting of sGBIV and the remaining AAIII (cluster-2high). The subdivision of cluster-2 may indicate that the histologically defined entity of AAIII consists of 2 molecularly and possibly prognostically distinct subgroups.
Several recent studies reported on frequent IDH1 and IDH2 mutations in diffusely infiltrating gliomas.5, 6, 23–25 The IDH1 and IDH2 proteins catalyze the conversion of isocitrate to alpha-ketoglutarate in the cytoplasm and mitochondria, respectively. Reported mutation frequencies ranged from 74 to 90% in AII, 55–78% in AAIII, 82–88% in sGBIV and 5–7% in pGBIV. In line with these data, we found IDH1 mutations in more than 90% of AII, AAIII and sGBIV but only 8% of pGBIV, while IDH2 mutation was restricted to a single case of sGBIV. Interestingly, we found that the IDH1 mutation status was closely linked to the gene expression signatures discriminating the 2 major subgroups of astrocytomas. This suggests that IDH1 mutations occur early in the development of most AII, AAIII and sGBIV and play an important role in their pathogenesis. One study suggested that IDH1 mutation up-regulates HIF1alpha signaling.26 In contrast to these data, we did not find any evidence that HIF1alpha target genes are up-regulated in IDH1 mutant gliomas when compared with IDH1 wild-type gliomas. Alternatively, loss of IDH1 function might also reduce the capacity to quench radical oxygen species due to the loss of production of redox-equivalents.27 A more recent study showed that the IDH1-R132H protein constitutes a gain-of-function mutant catalyzing the synthesis of 2-hydroxyglutarate (2HG) from alpha-ketoglutarate. Subsequently to IDH1 mutation, 2HG accumulates in the tissue potentially contributing to tumor formation and malignant progression.28
Interestingly, our 304-gene classifier separating the 2 main clusters of cerebral astrocytomas showed considerable overlap with a published prognostic classifier distinguishing so-called “proneural” versus “proliferative” and “mesenchymal” subgroups of malignant gliomas.10 Unsupervised clustering of our data based on this signature of Phillips et al.10 grouped the tumor samples into clusters closely resembling those of our own analysis. Specifically, our cluster-1 largely corresponded to the “proliferative” and “mesenchymal” groups that cannot be clearly separated in our data set, while cluster-2 corresponded to the “proneural” group. These results have important implications. First, most classifier genes of Phillips et al.10 and this study carry comparable predictive power and hence can be interchanged without altering analysis outcome. Second, the “proliferative”/“mesenchymal” and “proneural” signatures originally reported as prognostic signatures in high-grade gliomas may rather be diagnostic signatures distinguishing IDH1 wild-type and IDH1 mutant gliomas, which in fact largely corresponds to pGBIV versus AII, AAIII and sGBIV, respectively. During the preparation of our manuscript, 2 additional studies were published that classified glioblastomas12 and gliomas (including a mixture of astrocytoma, oligoastrocytoma, oligodendroglioma and pediatric pilocytic astrocytoma)11 based on gene expression profiles. While in each of these studies subgroups were defined in which IDH1 mutant tumors were overrepresented, the associations of the IDH1 mutation status and gene expression were less strong compared with our data. The reasons for this difference are not quite obvious but may be related to the different numbers of tumor samples used in the various studies as well as the variable selection of distinct entities, such as glioblastomas,12 respectively a broad range of various histological tumor types from adults and children.11
A special case is represented by the group of IDH1 mutant pGBIV (pGBIV-IDH1mut). Although based on only 7 such tumors, our data suggest pGBIV-IDH1mut as a separate group of glioblastomas characterized by a unique pattern of genomic imbalances, a cluster-2 (˜proneural) expression signature, and distinct clinical features, in particular younger age of onset and more favorable prognosis than pGBIV-IDH1wt. Evidence supporting this hypothesis comes from recent studies demonstrating that younger pGBIV patients typically display a “proneural” expression pattern and longer overall survival29 and that pGBIV-IDH1mut patients are characterized by younger age and significantly longer survival.6, 30, 31 It remains possible that a considerable fraction of pGBIV- IDHmut might actually be sGBIV that had remained clinically inconspicuous before the glioblastoma diagnosis.32 In any case, pGBIV-IDH1mut should be distinguished from pGBIV-IDH1wt tumors when analyzing results of clinical trials or performing translational investigations.
Univariate and multivariate survival analysis in 65 pGBIV-IDHwt patients showed gain/amplification at the PDGFRA oncogene locus and 19q gain to be associated with shorter and longer OS, respectively, We detected PDGFRA amplification in 8/65 (12%) of tumors, comparable to reported numbers.19–21, 24 These studies did not report on an association between PDGFRA amplification and survival of glioblastoma patients. More recently, PDGFRA amplification has been associated with IDH1 mutation and a proneural expression profile in glioblastomas, however, was also detected in subsets of other molecularly defined glioblastoma subtypes.12 Gain of 19q had been associated with shorter survival in glioblastoma patients.33, 34 In contrast to our series, these studies included high proportions of young patients and long-term survivors, which we deliberately excluded from our analysis due to their association with the IDH1mut status in our cohort. The molecular mechanisms by which 19q copy number status impacts overall survival currently are unclear. Univariate analyses of survival relative to gene expression in our pGBIV-IDH1wt patients identified LIMK2 and several other genes associated with actin cytoskeleton stability and cellular migration as being associated with reduced overall survival. Based on these data, one may speculate that survival is in part dependent on NOGO-pathway mediated migration.35, 36 The observed amplification (n = 1) and overexpression (n = 8) of PARD3, which directly interacts with LIMK2, in pGBIV may support this notion.
Another interesting finding obtained by array-CGH in our series of astrocytic gliomas concerns the demonstration of homozygous deletion of FGFR2 in 2 primary glioblastomas. The distal portion of chromosome arm 10q, i.e., the region located distal to the tumor suppressor gene PTEN has been implicated in several studies to carry additional glioblastoma-associated tumor suppressor genes.37 Several candidate genes located within 10q24-qter have been reported, including DMBT1,38MXI1,39LGI1,40WDR1141 and others. Our data implicate FGFR2 as a putative glioblastoma-associated tumor suppressor gene, as already reported in prostate and bladder cancer.42 How FGFR2 exerts anti-tumorigenic activity is not clear, however, evidence points to an involvement of nuclear factor (erythroid-derived 2)-like 2 (NFE2L2), a direct target gene of FGFR2 signaling, which activates cytoprotective enzymes by binding to antioxidant-response elements in their gene promoter regions.43 In line with previous data on gliomas,44 we found reduced levels of FGFR2 transcripts in glioblastomas. However, the molecular mechanisms leading to the frequently reduced FGFR2 expression in glioblastomas remain to be elucidated, as homozygous deletion or aberrant promoter methylation are restricted to a minor fraction of gliomas.
In summary, we reported on the definition, based on unsupervised bioinformatic analyses, of a novel classifier gene-expression signature for astrocytomas that is strongly associated with IDH1 gene mutation status. We showed that PDGFRA gain or amplification as well as 19q gain were associated with patient outcome within the group of patients with IDH1 wild-type pGBIV and identified FGFR2 as a novel glioma-associated candidate tumor suppressor gene in the distal region of the long arm of chromosome 10.
The authors thank Mrs. Britta Friedensdorf, Mrs. Heidi Schlingemann, Mrs. Sandra Meinhardt, Mrs. Heike Seul and Mr. Jörg Schlingemann for excellent technical assistance. This work was supported by Grants from the German Federal Ministry of Education and Research (BMBF) within the National Genome Research Network NGFN2 and NGFNplus.