In this study, we used array technology for quantitative expression and methylation profiling in a well characterized cohort of newly diagnosed GBM patients. We describe (i) the relationship between DNA methylation pattern and gene expression in GBM and (ii) the association between DNA methylation and clinical outcome in a subgroup of patients given uniform treatment in accordance with the STUPP protocol.
The methylation analysis identified 616 CpG sites DM between GBM and control brain and revealed considerable heterogeneity between GBMs, particularly for hypermethylated CpG sites. Hypo- and hypermethylated CpG sites were preferentially located outside and within CpG islands, respectively. This clearly confirms that cancer cells are characterized by both a loss of methylation in CpG-depleted regions and gains of methylation at CpG islands . Consistent with the findings of Martinez et al. , the hypermethylated gene set was found to be significantly enriched in PRC2 targets, highlighting the putative role of polycomb group proteins in de novo methylation in GBM. However, our data were not entirely consistent with this hypothesis. Indeed, there is no strong methylation pattern among the PRC2 targeted promoters and the changes in expression of the PRC2 and DNMT genes do not follow the hypermethylation gradient observed between low- and high-Δβ GBM clusters. This suggests that other genes may be linked to polycomb-associated de novo methylation.
The integrated analysis of DNA methylation and gene expression showed that DNA methylation only partly regulated gene expression. Indeed, almost a quarter of the DM genes also displayed concordant differential expression (chi-square test p -value < 0.01) (Additional file 9) and, in GBM samples, only 3% of the genes displayed an inverse correlation between promoter methylation and expression levels. This finding is consistent with published data for GBM . Moreover, many other well known mechanisms are involved in the regulation of gene expression (e.g. copy number alterations [2, 3], transcription factor production and recruitment, histone modifications, micro-RNA expression ). Nevertheless, our analysis led to the identification of 13 genes displaying concordant differential methylation and differential expression in GBM and control brain, and whose methylation and expression patterns were anti-correlated. The expression patterns of these genes may therefore be tightly regulated by epigenetic mechanisms, and their in-depth analysis may help us to understand the contribution of DNA methylation to glioblastomagenesis. Most of these genes have already been implicated in cancer-related processes. For example, ZNF217 (encoding zinc finger protein 217) is an important oncogene in many cancer types and its overexpression has been implicated in cell immortalization and resistance to chemotherapy . A recent study demonstrated that the ZNF217 protein forms nuclear complexes with several histone-modifying proteins (including EZH2) with synergistic effects in transcriptional repression . Another example is provided by FABP7 (brain fatty acid binding protein 7), which is expressed by the radial glia and involved in glia-guided neuronal migration . This protein has been associated with pure GBM histology, invasion and poor prognosis . Yet another example is provided by TSPYL5 (encoding testis-specific Y-like protein), which is a potent tumor suppressor gene and a frequent target of epigenetic silencing in glial tumors and gastric cancers [17, 36]. This gene has been shown to play a role in cell growth and resistance to radiation, through regulation of the p21(WAF1/Cip1) and PTEN/AKT pathway .
Noushmehr et al.  described a rare subgroup of GBMs with a CpG Island Methylator Phenotype. These G-CIMP tumors are a subclass of the GBM proneural subtype defined by Phillips et al. and Verhaak et al. [38, 39]. They were shown to be associated with secondary and recurrent GBMs, IDH1 somatic mutation, younger age at diagnosis and longer survival. Based on the G-CIMP 8-gene signature they describe (ANKRD43, HFE, MAL, LGALS3, FAS-1, FAS-2, RHO-F, and DOCK5), we identified three G-CIMP-positive tumors in the 55 patients of our cohort. This proportion (5.5%) is similar to that reported in the context of the TCGA (7.6%). We also confirm the association of G-CIMP status with IDH1 somatic mutation (Fisher's exact test p -value = 2e-4) and younger age at diagnosis (Wilcoxon rank sum test p -value = 0.01). However, we were unable to test the association between G-CIMP-positive status and OS, due the low frequency of this phenotype (three patients, two with survival data available).
Survival analysis was performed on a cohort of 50 patients uniformly treated by radiotherapy combined with concomitant and adjuvant temozolomide (STUPP protocol) . To our knowledge, this is the largest uniformly treated GBM cohort ever to be studied over such a large number of CpG loci. As expected, MGMT promoter methylation was strongly associated with longer survival, in both the microarray and pyrosequencing approaches. The chosen cutoff point for the β-value (10%) is similar to frequently used values (9%) . For the 27,578 CpG sites tested, MGMT methylation status remained one of the most powerful predictors of response to temozolomide-based treatment in GBM. Nevertheless, we have also identified two different types of prognostic markers. The first type stratifies the patients similarly to MGMT, but with a higher AUC. There is an association between the methylation level of MGMT and SOX10 promoters (chi-square test p -value < 0.01). The SOX10 gene is one such marker, and the hypermethylation of its promoter was associated with shorter survival in our cohort. Interestingly, the SOX10 protein is a marker of oligodendrocytes , and the presence of oligodendroglial differentiation areas in GBM has also been associated with longer survival . The second type of prognostic marker (FNDC3B, TBX3, DGKI, and FSD1) identifies patients with MGMT -methylated tumors not responding to STUPP treatment (Additional file 10). This second group of markers need to be validated on a larger cohort.