Glioblastoma, the most devastating of the primary brain tumors, is characterized by deregulation of multiple pathways, such as EGFR/PTEN/Akt/mTOR, TP53/MDM2/p14ARF and p16INK4a/RB. Several signaling molecules in such cascades are already the targets of therapies for glioblastomas but improvements remain modest from a clinical standpoint (Stupp et al., 2007). Identification of further tumor biomarkers is thus needed to provide new molecular targeted therapies. Important insights into tumor suppressor genes and oncogenes will probably be provided by identifying genomic aberrations inducing direct changes in gene expression, that is, genes with expression levels either significantly concordant or correlated with changes in DNA copy number.
DNA copy number alterations (CNAs) are generally more numerous in malignant tumors than in benign ones, and could be classified as both causal and random genetic events. Some CNAs have a direct effect on gene expression and are likely to be more critical than others in the biology of cancer. Such CNAs can result either in the loss of tumor suppressor gene function or in the over-expression and activation of oncogenes. Both mechanisms constitute putative early oncogenic steps. A better understanding of the initiating molecular determinants of malignant tumors will require the identification of the CNAs that are functionally significant.
This task is particularly challenging for glioblastomas because of their highly rearranged genome (Bredel et al., 2005a; Maher et al., 2006), and by the large number of genes that have been implicated at the transcriptome level (Nutt et al., 2003; Tso et al., 2006). Recently, some genome-scale studies of glioblastoma described the relationship between DNA dosage and gene expression (Nigro et al., 2005; Liu et al., 2006a,b; Phillips et al., 2006), but with some weaknesses: (1) intratumor variability was not taken into account; (2) direct gene level correlations were not really possible due to the use of low resolution arrays; and most importantly (3) genome and transcriptome data sets were unpaired, prone to provide false positives, and so to miss targeted genes.
In this study, we performed paired genome-wide analyses of glioblastoma, focusing on genes that showed concordant CNAs and expression patterns. High-resolution maps of chromosomal alterations were obtained by performing array-based comparative genomic hybridizations (arrays CGH) on 19 glioblastomas. Gene expression profiling was carried out on the same tumor samples, and compared to those obtained on non-neoplastic brain samples. We validated and investigated our result in an independent publicly available microarray data set of 81 glioblastomas and 23 normal brains. The associated annotation and network analyses provide insights into the genetic mechanisms driving glioblastoma.
MATERIALS AND METHODS
A total of 76 glioblastomas and eight normal brain samples were used in this study. Fresh-frozen glioblastoma samples from patients admitted to the Neurosurgery Departments of Brittany University Hospitals (Rennes and Brest) were collected with informed consent and subjected to WHO classification. Histology was confirmed by hematoxylin-eosin staining of paraffin-embedded blocks. The non-neoplastic brain tissues were obtained from normal white matter area taken from patients undergoing surgery for chronic epilepsy. For microarray and array CGH hybridization, a set of 19 glioblastoma and four normal brain samples were used. Blood was available for all these patients. Each snap-frozen tumor block was cryodissected in 10 μm sections. The first section as well as sections obtained every 100 μm were stained to select tissue with at least 70 percent of tumor cells and to exclude necrotic areas and widespread blood vessels. To allow a paired and accurate comparison between closely related biological materials, the adjacent sections were alternatively pooled in different tubes for RNA and DNA extraction. For subsequent real-time reverse transcription-PCR validation of selected genes, we used an independent set of 57 glioblastoma samples.
Nucleic Acid Preparation
Tumor DNA was extracted according to manufacturer's instructions (NucleoSpin Tissue Kit, Macherey Nagel, Düren, DE). Blood DNA (reference) was isolated from peripheral blood leukocytes using a classical saline extraction. Total RNA was isolated using Macherey-Nagel NucleoSpin RNAII Kit. RNA integrity was confirmed using the Agilent 2100 Bioanalyzer (Agilent, Palo Alto, CA).
Arrays CGH Profiling
Genomic analyses were performed on Human 44K Agilent arrays CGH (Agilent Technologies, Santa Clara, CA) according to the manufacturer's protocol. Briefly, 5 μg of DNA were double-digested (AluI and RsaI, Promega, Madison, WI). Tumor DNA and reference DNA were both labeled by random priming with Cy3-dCTP and Cy5-dCTP for dye-swap experimental design. Tumor and reference DNA were pooled and hybridized (65°C, 48 hr). Arrays were washed and scanned on an Agilent G2565BA microarray scanner. Data were extracted and flagged with the Feature Extraction software (FE v9.4.1, CGH_44k_1005 protocol). Data preprocessing was carried out using limma (R package) from Bioconductor (Smyth, 2005). Values were median normalized and fluorescence log2 ratios were calculated as the average of two-paired arrays (dye-swap) except for one pair of arrays from which only one array met the quality criteria. Missing values were imputed with the k-nearest neighbors method implemented in impute (R package).
Array CGH Data Analysis
Log2 ratios (tumor versus reference) of probe intensities were plotted according to their genomic position, chromosome by chromosome. DNA copy number alterations were identified using the Gain and Loss Analysis of DNA algorithm (Hupé et al., 2004) implemented in GLAD (R package). This method uses adaptive weights smoothing (AWS) procedure to detect breakpoints from array CGH profiles, and assigns a copy number status to each altered or normal chromosomal region. A “segmented” data set was generated by determining uniform copy number segment boundaries and by replacing normalized log2 ratio for each probe by the calculated smoothing values. As done in the multiple myeloma study of Aguirre et al., (2004), distribution of the smoothing values was used to define thresholds for the analysis of hemizygous deletions, homozygous deletions, gains, and amplifications. Thresholds for low-copy number gain and hemizygous deletion were set at −0.15 and +0.15, respectively (±6 SD of the middle 75% of the data). Thresholds for high amplitude events were chosen at +0.9 for amplification and at −0.40 for homozygous deletion. Minimal common regions (MCRs) were defined as loci with CNAs in at least two samples; one of them showing an extreme CNA event defined by thresholds +0.29 and −0.29, <99% and <1% quantiles. Unsupervised hierarchical clustering was done on the segmented data matrix using the Support Tree option in MeV (http://www.tigr.org/software/tm4/midas.html). Consensus clusters (average linkage and Pearson correlation metric) were built by bootstrap.
Gene Expression Profiling
Gene expression profiling was performed using the Agilent Whole Human oligo-Microarray Kit 4 × 44K multiplex format (Agilent Technologies, Santa Clara, California), with manufacturer's recommended procedures for microarray-based one-color (http://www.chem.agilent.com/temp/rad37FF4/00064034.pdf). Briefly, 350 ng of total RNA with control RNA Spike In were amplified and labeled with Cy3-CTP (Agilent Technologies, Santa Clara, California). 1.65 μg of Cy3-labeled RNA was hybridized (65°C, 17 hr) per array. The processed Multiplicative Detrend FE data (FE v9.4.1, GE1-v5_91_0806 protocol) of scanned images were median normalized and missing values were imputed (limma and impute, R packages).
Expression Data Analysis
Analysis of gene expression was conducted to highlight the genes differentially expressed between glioblastomas and normal brain tissues. The significance in differential gene expressions was determined using standard Student's t test of two groups. To account for multiple hypothesis testing, we computed adjusted P values by controlling the false discovery rate (FDR) with the Benjamini & Hochberg (BH) procedure implemented in multtest (R package). Differentially expressed genes were defined as follows: BH adjusted P value < 0.01 and absolute mean log2 ratio (glioblastoma versus mean normal brain) greater than two.
The complete dataset has been submitted to the gene expression omnibus data (GEO) public database at NCBI, and the accession number is GSE10878.
Integrated Copy Number and Expression Analyses
Combination of genome and transcriptome datasets was done gene-by-gene for all of the annotated genes that were present on both arrays. We used two approaches to identify all of the genes whose transcription levels were potentially affected by DNA alterations. A schematic of our approach is provided in Figure 1. In the targeted study (Fig. 1A), we identified the genes differentially expressed between glioblastoma and normal brain (BH P value < 0.05), located in MCR, and concordant to the corresponding CNA, for example, we selected those that were over-expressed or under-expressed and located in a region of gain or loss, respectively. In the Correlation study (Fig. 1B), we evaluated the direct influence of copy number alterations on gene expression in MCR. We identified the genes with highly correlated DNA segmented values and expression patterns (Pearson's correlation coefficient up to 0.7).
Gene Ontology, Canonical Pathway, and Functional Network Analyses
Functional annotation analyses were performed with the NIH-DAVID software (version 2.1b, http://david.abcc.ncifcrf.gov/) and the method developed by Aubry et al., (2006). We used the first method with the parameters: GOTERM_BP_ALL, KEGG_PATHWAY and SP_PIR_Keywords; the significance threshold was set on a P value < 0.05 with Benjamini multiple testing correction. The second one was used to provide deeper informative annotations by combining evidence and literature with gene ontology annotation database and PubGene biomedical literature index. Functional network analyses were executed using the web-delivered application from Ingenuity Pathways Analysis4 that enables the visualization and exploration of molecular interaction networks in gene expression data.
Real-time quantitative PCR (Q-PCR) for PCDH9 and STARD13
Q-PCR reactions were done with the 7900HT Fast Real-Time PCR System using the SYBR™Green PCR Master Mix (Applied Biosystems®). B2M, β-2 micoglobulin, RNA was chosen as internal control. Calibration was done with FirstChoice® Human Brain Reference Total RNA (Applied Biosystems®). The relative amounts of the gene transcripts were determined using the ΔΔCt method, as described by the manufacturer. The following forward (F) and reverse (R) primers were designed using Primer3 (v.0.4.0): B2M, F: 5′-TCCAACATCAACATCTTGGT-3′ and R: 5′-TCCCCCAAATTCTAAGCAGA-3′; PCDH9, F: 5′-GCATATTGTCACTTAGGTCAAACCA-3′ and R: 5′-GTCATGCCTTAACAAAAACCTCCT-3′; STARD13, F: 5′-TGCTAATGGATCGAATGTGCTT-3′ and R: 5′-TTCTCCAACACCAGTTGCTAAATC-3′.
Comparison with Published Expression Data
We evaluated the robustness of gene selections on a publicly available microarray data set. The expression data from Sun et al., (2006) (GDS1962) were used to generate the comparison data set of gene expression changes between glioblastoma (n = 81 samples) and normal brain (n = 23 samples). The GDS1962 was downloaded from the Gene Expression Omnibus database and managed with GEOquery (R package). A data matrix was generated by R programming, with the mean centered and scaled values corresponding to the processed data per array. We downloaded the annotation of the Affymetrix U133 Plus 2.0 Array platform from the Ensembl website. We retrieved the genes, and we were interested in by identifying probe sets with gene symbols. For heatmap representation, values were expressed as log2 ratio of glioblastoma versus mean of all normal brain samples.
Recurrent and Novel Genomic Changes in Glioblastoma
Glioblastomas (n = 19) were analyzed by array CGH in a dye-swap experiment design to support the normalization step and to provide a robust evaluation of the genomic profiles. Only somatic changes were defined because each tumor DNA was hybridized with the corresponding patient blood DNA. Segmentation analysis identified large aberrations at the genome level as well as focal higher-amplitude recurrent CNAs (MCRs). A summary of the CNAs is shown in Figure 2A. The most frequent imbalances were: gain of chromosomes 7 (73%, gain of both 7p and 7q in 47%, 7p alone 10%, and 7q alone in 16%) and 20 (16%); and loss of 9p (58%), chromosome 10 (58%), parts of chromosome 13 (31%), and 22q (21%).
The MCRs (2285 DNA segments and 4816 genes) corresponding to either amplification (log2 ratio >0.9) or homozygous deletion (log2 ratio < −0.40) are presented in Supp info. Table 1 (A and B). These high-amplitude events span a median size of 8.4 Mb with an average of 11 known genes. The following previously well-characterized amplicons were identified: coamplification of PDGFRA, KIT, and KDR (three tumors), coamplification of EGFR, SEC61G (six cases), coamplification of CDK4 and MDM2 (one tumor) and amplicon at 1q32.1 (two cases) including a gene encoding a catalytic subunit of Pi3k (PIK3C2B) (Knobbe and Reifenberger, 2003). The well-known deletion of CDKN2A and CDKN2B with codeletion of the putative tumor suppressor gene MTAP (Schmid et al., 2000) was found in four cases. The recently identified glioblastoma deletion of tumor-suppressor gene CDKN2C was also present in one tumor (Solomon et al., 2008). In addition, we identified the following new amplicons: at 20q11-13, including POFUT1 known to be essential for Notch function (Kroes et al., 2007), at 9p22.1, including DNAJA1 that encodes a cochaperone of heat shock protein 70 (WaNg et al., 2006) and at 11p13, including PAX6 (Daugaard et al., 2007). New homozygous deletions included particularly DKK1 (10q11.2) that is a Wnt/β-catenin pathway inhibitor shown to be proapoptotic in brain tumor cells (Shou et al., 2002) and other genes (CNTNAP3, 9p13.1; GLUD2, in Xq24, and, BAGE and BAGE4, in 21p11.1).
MCRs Cluster Glioblastoma Around EGFR and STARD13 Status
To probe the organization of MCRs across the tumor set, we performed unsupervised hierarchical clustering of glioblastomas in the space of MCRs (without taking into account the sex chromosomes). Cluster analysis (Fig. 2B) emphasized common recurrent aberrations, including the gain of whole chromosome 7 and losses of chromosome 10 and 9p. Notably, cluster analysis highlighted two sample groups depending apparently on whether EGFR (7p11.2) was amplified or not (subtype-1 and subtype-2, respectively). In addition, it showed that three glioblastomas (one with EGFR amplification, the others without gain of chromosome 7) had particular genomic profiles (see Supp info. Table 2). For these three patients, neither reanalysis of the histological nor clinical data were able to phenotypically appreciate these particularities. To more directly assess the relation between MCRs and the two groups (subtype-1 and subtype-2), we applied a supervised analysis using a standard Student's t test. Supervised analysis showed that the two subtypes exhibited distinct patterns of MCRs (Supp info. Table 3). In particular, it confirmed that all of the subtype-1-glioblastomas but none of the subtype-2-glioblastomas carried EGFR amplification at 7p11.2 locus (P = 0.006). It also showed that: (i) subtype-1 presented more frequent hemizygous deletions at 13q31-13q34 (P = 0.01) and 13q12-13q21 (P = 0.03); (ii) gains at 7p21.3-p21.2, 7q21.11, 7q21.2 and 7q31.1-q34 were more frequent (P < 0.01) in subtype-2 (77–88%) when compared with subtype-1 (28–42%). We also investigated these differences at the transcriptome level (Student's t test with BH correction, comparing expression values between subtype-1 and subtype-2 for genes located in the aberrant regions on chromosome 7 and 13q). Only two differential gene expressions were highlighted: EGFR over-expression (P = 0.011) and STARD13 under-expression (P = 0.036).
Glioblastoma Expression Profiling Identifies Huge Amount of Alterations
Evaluation of genes that are differentially expressed in glioblastoma versus normal brain was undertaken using a standard Student t test with BH correction (P < 0.01) and absolute average log2 ratio greater than two. Expression profiling identified 664 over-expressed genes in glioblastoma, including 56 genes over-expressed greater than 30-fold, and 1224 under-expressed genes, including 157 genes under-expressed greater than 30-fold. A list of all of the identified genes is provided in Supporting information File 1. From the top ten of the over-expressed genes, we identified the antiapoptotic gene BIRC5/survivin (Blum et al., 2006) and the transcription factor E2F2 (Okamoto et al., 2007), the activities of which have been already linked to glioblastoma. We also found two highly expressed mitotic kinases, PBK and BUB1, and a potential cell cycle regulator, DLG7/HURP, that have not been previously reported in glioblastoma. Proliferation-related genes were also included among the highly expressed genes. Some of them were previously well described as participating in glioblastoma progression, such as UBE2C (Bredel et al., 2005b), MKI67 (Raghavan et al., 1990), TOP2A (van den Boom et al., 2003), TNC (Sarkar et al., 2006), CHI3L1/YKL-40 (Tanwar et al., 2002), MELK (Liu et al., 2006a,b), and CD44 (Ylagan and Quinn, 1997). The others, NCAPG, KIF20A, CENPA, and RRM2, are novel glioblastoma-associated genes with reported functional roles in cytokinesis and/or cell proliferation (Geiman et al., 2004; Wonsey and Follettie, 2005). Regarding the under-expressed genes, we identified some with known tumor suppressor functions in glioma (CHD5 and LGI1). CHD5 was identified last year as a tumor-suppressor that controls proliferation, apoptosis, and senescence via the p19(Arf)/p53 pathway in glioma (Bagchi et al., 2007). It has been suggested that the leucine-rich, glioma-inactivated gene 1 (LGI1) gene is a candidate tumor suppressor gene involved in progression of glial tumors (Chernova et al., 1998). In addition, four components of the Wnt/β-catenin signaling pathway were highly under-expressed: two Wnt antagonists, WIF1 and SFRP1, and PPP2R2C (PP2A) and WNT10B (Kirikoshi and Katoh, 2002). Such disruptions may result in an improper function of the Wnt/β-signaling pathway leading to aberrant cell proliferation and therefore explaining part of the glioblastoma progression.
Functional annotation of the highly differentially expressed genes (greater or less than 30-fold) underlined distinct biological processes according to groups of over- or under-expression. Highly over-expressed genes were significantly associated with the regulation of the mitotic cell cycle, and more precisely with the following GOTERM_BP_ALL: M phase (P = 5.7e-7), microtubule-based process (P = 3.9e-3), and sister chromatid segregation (P = 0.01). Highly under-expressed genes were significantly associated with cell communication (P = 1.8e-3), cell-cell signaling (P = 3.1e-9), and neurophysiological process (P = 4.9e-5).
Identification of the Glioblastoma DNA Targeted Genes
To identify targeted genes, we used the paired array CGH and transcriptome data sets, measured on the same glioblastomas and restricted to MCRs. We combined these two paired data sets to focus only on the genes for which the expression was affected by aberrant copy number variations. Two possibilities were evaluated: genes with expression levels significantly concordant with changes in DNA copy number (Targeted study) and genes with expression levels directly correlated with such changes but not necessarily with a significant differential expression in all glioblastomas (Correlated study) (see Materials and Methods, Fig. 1, for schematic representation).
The Targeted study showed that DNA copy number influenced gene expression across a 13.8% range of MCRs, corresponding to 261 significant targeted genes (STGs) that were differentially expressed and concordant with MCR patterns. Representation of DNA aberrations and corresponding gene expression on two mirrored heatmaps clearly shows the unbalanced distribution of concordant over- and under-expressed genes on chromosomes (Fig. 3A). The 95 STGs found to be highly expressed in gain of regions were located in chromosome 1 (1p32.1), 4 (4q11-4q12), 7, 12 (12p11), and X (Xq24) and the 166 STGs under-expressed in deleted regions were located in 9p, chromosome 10, and 13q. Functional annotation analyses of the STGs were performed by taking into account under- and over-expressed groups separately. Significant enrichments were found with DAVID: the over-expressed STGs were particularly related to developmental processes (GOTERM_BP_ALL: development, morphogenesis, and SP_PIR_KEYWORDS: developmental protein) and the under-expressed STGs to mRNA splicing (SP_PIR_KEYWORDS: alternative splicing). The method developed by Aubry et al., (2006) provided more details on tumorigenesis-related processes linked to STGs including cell cycle (mitosis and apoptosis), cell adhesion, DNA repair, and angiogenesis (Fig. 3B). To understand how STGs are related, they were replaced in pathways and molecular interactions established on the ingenuity knowledge base. Networks and associated functions and disease were scored and checked for significance. The most relevant one was built around EGFR and displayed high-level functions in cancer and neurological disease. It contained 27 STGs (Fig. 3C), including the GBM-related genes CAV1, DMBT1, EGFR, KDR, and MGMT. The over-expressed STGs were notably linked to cell proliferation and movement.
The Correlated study focused on the DNA alterations affecting gene expression on a distinct, individual, and isolated manner. To achieve this, we performed direct correlation analyses of copy number and expression data on a gene-by-gene basis throughout the genome. Transcription levels were highly correlated to genomic patterns for 159 correlated genes (CGs), representing only 4.1% of the genes located in MCRs. The CGs were located on chromosomes 1, 4, 7, 9, 10, and 13, as for the STGs, but also on chromosomes 17, 20, and 22. The strong influence of DNA copy number on the CGs expression is evident by examination of the heatmap representations of CGs data for DNA aberrations and gene expression (Fig. 4A). The overall patterns of gene amplification and increased gene expression are concordant, that is, a significant fraction of the highly amplified genes appear to be correspondingly highly expressed. Such concordance is also found for part of the genes located in losses. Functional annotation of CGs highlighted gene ontology terms (biological processes) mostly associated with cell cycle, DNA repair, RNA processing, and brain development (Fig. 4B). The top-scoring networks (Fig. 4C), built around EGFR and PDGF, contained glioma-related genes (CDKN2A, CDKN2B, EGFR, and MLL3), genes that play a role in hematological disorder (CDKN2A, CDKN2B, EGFR, GIT1, MPDZ, NT5C3, OPRS1, PDGFA, PLCG1, RALA, SLC25A13, SRPK2, and TIMM23) and genes involved in cell division process (CDKN2A, CDKN2B, DMTF1, PLCG1, and ZC3HC1).
A Targeted Genes Signature for Glioblastoma
Following the results of both targeted and correlated analyses, we determined a set of 406 genes constituting the glioblastoma targeted genes signature. Fourteen genes were identified in both analyses. To evaluate the relevance of this glioblastoma genomic signature, we used a publicly available microarray data set (GDS1962) of 81 glioblastomas and 23 normal brain samples, from the study by Sun et al., (2006). We were able to map 369 of our 406 selected genes on the Affymetrix U133 Plus 2.0 Array platform, and generated the corresponding data matrix. We performed a Student t test on this data table to identify the genes that were differentially expressed between glioblastoma and normal brain. With a risk level (BH) of 5%, we confirmed that 86% of the 369 genes were differentially expressed in glioblastoma (92% for the Targeted study and 73% for the Correlated study). More details are available in Supporting information File 2. A heatmap representation of all of the glioblastomas (81 from the Sun et al., study [external] and 19 from the present study [local]) is provided in Figure 5A. The strong similarity of the red:green profiles between the two data sets (external and local) illustrate the strong robustness of the targeted genes signature. Network analysis provided two top scoring networks: the first one (Fig. 5B) was built around Rb and NfκB and the second (Fig. 5C) around Akt, EGFR, and PDGFR. Overlaid functions (“proliferation of cells,” “hematological disorder,” and “developmental process of tumor cell lines”) were mostly associated to over-expressed genes. Both networks included a substantial number of genes that have been implicated in glioblastomagenesis.
To go further in the validation of the glioblastoma targeted genes signature, we probed separately the organization of the targeted genes across the tumor set from the correlated genes (Fig. 6). We performed unsupervised hierarchical clustering of all of the tumors in the space of each set of genes. The targeted cluster analysis (Fig. 6A) defined two groups of tumors. However, the genes signature did not provide a clear distinction between the two groups, illustrating the homogeneous expression of the targeted genes in glioblastoma. The correlated cluster analysis (Fig. 6B) emphasized the heterogeneous part of the signature, also grouping glioblastomas into two subtypes, mainly depending on whether genes mapping to 7p11.2 were highly over-expressed or not and whether CDKN2A and CDKN2B were deleted or not.
PCDH9 and STARD13 were considered interesting due to their biological function and because their under-expression was strongly correlated to the corresponding genomic state. PCDH9 and STARD13 under-expression was validated by RT Q-PCR on an independent set of 57 glioblastoma samples and on the same panel of glioblastomas (Fig. 7). It was also confirmed by the GDS1962 analysis.
Glioblastomas show numerous DNA alterations of varying size and location as well as large transcriptome modifications. The functional annotation analyses suggest that the molecular alterations occurring in glioblastoma promote cell growth, proliferation, survival, and apoptotic resistance. However, identifying high-priority biomarkers by this way is still challenging. Consequently, we sought to identify targeted genes by integrating data on the genome and transcriptome levels. To achieve this task, we implemented a dual strategy that delineates genes with either significantly concordant or correlated changes in expression and in copy number.
A major advantage of choosing a dual strategy is that it bypasses limitations due to a small number of cases and due to inter-tumor heterogeneity. Indeed, a great part of the genes with copy number driving changes in expression are likely to be functionally significant even if occurring in only few cases. In our study, such genes were surveyed by the correlation study. This is well illustrated in our analysis of CDKN2A and CDKN2B. These two genes are altered at the DNA level with an impact on transcription. The correlation study but not the targeted one identified these two genes, which were also validated by external data analysis. Moreover, the correlation study and the validation step help to balance the lack of statistical power due to our small sample size cohort.
The dual strategy delineated two sets of genes that appear to have distinct biological relevance. The genes identified by the targeted study had homogeneous transcriptome modifications in glioblastomas. Even although gene expression was affected in almost all tumors, it was not related to CNAs in all glioblastomas. The homogeneous transcriptome signature must involve other regulatory mechanisms such as regulations by transcription factor, miRNA, and DNA methylation. These widespread homogeneous changes are related to mitotic defects, chromosome segregation errors, disruption of cell adhesion processes, activation of DNA repair, escape from apoptosis and angiogenesis. Moreover, our network analysis revealed a top-scoring pathway potentially affecting glioblastoma cell invasion. It is therefore tempting to assume that the targeted genes correspond to changes highly essential for glioblastoma development and that they are the most likely candidate oncogenes and tumor-suppressor genes linked to tumorigenesis.
In contrast, the genes identified by the correlation study have an expression directly and highly linked to the DNA copy number status, consequently representing more sporadic events. The two clusters of glioblastoma suggest distinct pathways leading to glioblastoma development, raising the question whether these different sets of genes combinations result from compensatory functions within genetic pathways or whether they represent different clinical subclasses of glioblastoma. Our functional annotation and network analyses provided some information, suggesting that part of the heterogeneity was due to genes associated general cancer biology, as well as to genes that control normal brain.
We also identified PCDH9 and STARD13 as potential tumor suppressor genes in glioblastoma. PCDH9 is a protocadherin predominantly expressed in the nervous system and is highly similar to cadherin-related tumor suppressor precursor. Three other protocadherins, PCDH8, PCDH20, PCDH21, were shown to be under-expressed in the targeted study; PCHD20 is a candidate tumor suppressor in non-small-cell lung cancers (Imoto et al., 2006). As these genes are likely to be involved in cell–cell interactions in the glial cell compartment, it will be of interest to investigate the links between their low expression levels, adhesion of glial cells, and the promoted invasive growth of glioblastoma. StAR-related lipid transfer (START) domain containing 13 (STARD13) belongs, with DLC1, to the family of RhoGap proteins with START domains (Soccio and Breslow, 2003). The corresponding genes are frequently lost in hepatocellular and breast carcinomas (Ching et al., 2003; Nagaraja and Kandpal, 2004). Both STARD13 and DLC1 are suggested to be candidate tumor suppressor genes (Yuan et al., 1998; Ng et al., 2006). Moreover, Hatch et al., have recently suggested that impairments of STARD13 gene can participate in etoposide resistance via its role in ceramide signaling to the RhoA pathway (Hatch et al., 2007). A temozolomide/etoposide chemotherapy regime is currently being evaluated in trials for glioblastoma (Korones et al., 2003), and it will be of great interest to study response to this treatment with regard to presence or absence of STARD13 deletions.
We would like to thank Cyrille Surbled for technical assistance and members of the CNRS-UMR6061 for discussions and comments on the manuscript. We acknowledge the Departments of Neurosurgery (CHU-Rennes, CHU-Brest) for contributing tumor samples. We also thank Phillip Jordan for his help in writing this manuscript.