Methylation profiling identifies 2 groups of gliomas according to their tumorigenesis

Abstract

Extensive genomic and gene expression studies have been performed in gliomas, but the epigenetic alterations that characterize different subtypes of gliomas remain largely unknown. Here, we analyzed the methylation patterns of 807 genes (1536 CpGs) in a series of 33 low-grade gliomas (LGGs), 36 glioblastomas (GBMs), 8 paired initial and recurrent gliomas, and 9 controls. This analysis was performed with Illumina's Golden Gate Bead methylation arrays and was correlated with clinical, histological, genomic, gene expression, and genotyping data, including IDH1 mutations. Unsupervised hierarchical clustering resulted in 2 groups of gliomas: a group corresponding to de novo GBMs and a group consisting of LGGs, recurrent anaplastic gliomas, and secondary GBMs. When compared with de novo GBMs and controls, this latter group was characterized by a very high frequency of IDH1 mutations and by a hypermethylated profile similar to the recently described glioma CpG island methylator phenotype. MGMT methylation was more frequent in this group. Among the LGG cluster, 1p19q codeleted LGG displayed a distinct methylation profile. A study of paired initial and recurrent gliomas demonstrated that methylation profiles were remarkably stable across glioma evolution, even during anaplastic transformation, suggesting that epigenetic alterations occur early during gliomagenesis. Using the Cancer Genome Atlas data set, we demonstrated that GBM samples that had an LGG-like hypermethylated profile had a high rate of IDH1 mutations and a better outcome. Finally, we identified several hypermethylated and downregulated genes that may be associated with LGG and GBM oncogenesis, LGG oncogenesis, 1p19q codeleted LGG oncogenesis, and GBM oncogenesis.

Fig. 1.

Comparison with Methylight. (A–C) Methylation levels (β-value) of MGMT, RASSF1, and RBP1 were compared with the methylation status obtained using Methylight (U for unmethylated and M for methylated). For each gene, the β-values in methylated samples were significantly higher when compared with unmethylated samples (Wilcoxon's test, α = 5%). When using a 0.2 β-value cut-off (gray dashed line) to define methylated (β-value >0.2) and unmethylated samples (β-value <0.2), the consistency between both methods was 70.1%, 81.2%, and 92.1% for MGMT, RASSF1, and RBP1 methylation status assessment. (D–F) Comparison of MGMT, RASSF1, and RBP1 methylation frequencies in LGG and GBM using Ilumina's bead arrays and Methylight and comparison with previously reported methylation frequencies for MGMT and RASSF1.^,,

Fig. 2.

Distribution of β-values in LGG, GBM, and control samples. The distribution of the β-values is given as a density plot for CpG within (A) and outside (B) CpG Islands according to sample types. The β-value, between 0 and 1, indicates the proportion of methylation of CpG. It shows that CpGi were hypermethylated (>0.8) in both LGGs and GBMs compared with control samples, whereas CpGo were hypermethylated in LGGs when compared with both control samples and GBMs.

Fig. 3.

Unsupervised hierarchical clustering of the 33 LGGs, 36 GBMs, 8 paired initial and recurrent gliomas, and 9 controls according to their methylation profile. Genes (lines) and samples (columns) are ordered according to a hierarchical clustering. β-values increase from green (unmethylated) to yellow to red (methylated). The following annotations are displayed: (i) recurrent: indicates paired initial and recurrent gliomas, each pair has a distinct color; (ii) grade: normal samples (red), grade 2 (blue), grade 3 (green), and grade 4 (orange); (iii) phenotype: normal samples (blue), astrocytoma (red), oligodendroglioma (orange), and mixed glioma (green); (iv) Verhaak classification: Verhaak et al. classes in tumors for which the gene expression profile was available, neural (green), proneural (orange), mesenchymal (blue), and classical (red); (v) 1p19q codeletion: yes (black) and no (white); (vi) IDH1 mutation: yes (black) and no (white); (vii) EGFR amplification assessed using CGH array: yes (black) and no (white); (viii) MGMT methylight: MGMT methylation according to Methylight, methylated (white) and unmethylated (black); (ix) MGMT _P281_F: MGMT methylation according to Illumina's bead arrays, methylated (white) and unmethylated (black). This clustering suggests that LGG as well as IDH1-mutated GBM and anaplastic gliomas are characterized by a hypermethylated profile in comparison with normal brain and de novo GBM. It also shows that the methylation profile of gliomas remains remarkably stable across their evolution even at the time of anaplastic transformation.

Fig. 4.

PCA of LGG samples. PCA distinguishes LGG samples with and without 1p19q codeletion.

Fig. 5.

Unsupervised hierarchical analysis with the 220 TCGA GBM samples. Genes (lines) and samples (columns) are ordered according to a hierarchical clustering. β-values increase from green (unmethylated) to yellow to red (methylated). The following annotations are displayed: (i) Data set: our data set (white), TCGA data set (black); (ii) recurrent: paired initial and recurrent gliomas in our data set. Each pair has a distinct color, and secondary GBM in the TCGA data set (pink); (iii) grade: normal samples (red), grade 2 (blue), grade 3 (green), and grade 4 (orange); (iv) Verhaak: Verhaak et al. classes in tumors for which the gene expression profile was available, neural (green), proneural (orange), mesenchymal (blue), and classical (red); (v) IDH1 mutation: IDH1 mutation in tumors for which the information was available, yes (black) and no (white). This clustering shows that a subset (n = 19 of 220) of TCGA GBM samples characterized by a hypermethylated profile and by a high rate of IDH1 mutation clustered with our LGG samples.

Fig. 6.

Overall survival of GBM samples with an LGG-like methylation profile in TCGA data set. Kaplan–Meier survival curve of TCGA GBM samples demonstrating that TCGA GBM samples with a hypermethylated LGG-like profile (gray dashed curve) had a longer overall survival than the GBM samples that clustered into the de novo GBM cluster (black curve).