Stability of the CpG island methylator phenotype during glioma progression and identification of methylated loci in secondary glioblastomas

Forty DNA samples from 20 astrocytoma/glioma patients were used in this study. These patient samples consisted of; 10 WHO grade II astrocytomas, 15 WHO grade III astrocytomas and 15 WHO grade IV glioblastomas. The 40 DNA samples represent 20 cases of paired early and late lesions from the same patient. The DNA was extracted from tissue samples consisting of a minimum of 80% tumor. The DNA from four non-disease brain samples was used to provide the normal, expected levels of methylation. Ethical guidelines were followed for patient sample collection and all samples have been anonymised. Research was conducted according to the principles expressed in the Declaration of Helsinki. Patients gave written informed consent for analysis of tumor samples. The study was approved by the Institutional Ethics Committees of University of Technology Dresden and University of Birmingham.

The Illumina Infinium HumanMethylation450 array (Illumina, San Diego, CA, USA) was performed on 0.5 μg bisulfite modified patient DNA according to manufacturers’ instructions. Bisulfite modification of DNA and array hybridization was carried out by Cambridge Genomics Services. Raw data was obtained using Genome Studio software from Illumina. The raw data were processed using the lumi R [17] package to correct for the color bias present due to the use of different dye on the array. To correct this bias, Infinium type I and type II are separated, then both channel are also separated and the color bias is corrected using a within array smooth quantile normalization. After correction the two channels and probe types are combined and a between array quantile normalization is performed. The beta score are then calculated. The raw files have been deposited in NCBI’s Gene Expression Omnibus [18] and are accessible through GEO Series accession number GSE58298.

Probes demonstrating detection p-values greater than 0.01 in any sample were removed along with probes located on the X and Y chromosomes. To ensure tumor specific hypermethylation, probes showing a beta value ≥0.25 in any of the four normal samples were also removed. Hypermethylation was subsequently determined as a beta value ≥0.5. This was considered relevant if present in >30% tumor samples. Additional filtering was achieved limiting selection to genes for which the hypermethylation criteria were met in ≥3 probes associated to that gene.

The top 2000 most variable loci for each clustering event were determined by selecting the 2000 probes with the greatest standard deviation across all the given samples. Clustering was performed using the Cluster3 program (http://bonsai.hgc.jp/~mdehoon/software/cluster/software.htm#ctv) and visualized using the Java TreeView program (http://jtreeview.sourceforge.net/). Unsupervised hierarchical clustering was performed using the Euclidean based algorithm.

Clone sequencing was used for array validation. 0.5 μg of DNA for each sample was bisulfite modified using the Qiagen EpiTect kit (Qiagen, Heidelberg, Germany) according to manufacturers’ instructions. PCR reactions were performed using FastStart Taq DNA polymerase (Roche, West Sussex, UK) on a semi-nested basis for all genes using the primers listed in Additional file 1. A touchdown PCR program for primary and secondary reactions using gene specific annealing temperatures was performed. Selected PCR products were cloned into the pGEM-T easy vector (Promega, Madison, WI, USA) according to manufacturers’ instructions and cultured overnight at 37°C. Up to 12 colonies were selected for single colony PCR using primer sequences F: 5′- TAATACGACTCACTATAGGG -3′ and R: 5′- ACACTATAGAATACTCAAGC -3′. PCR products were cleaned for sequencing using thermosensitive alkaline phosphatase (Fermentas UK, York, UK) and Exonuclease I (NEB, Ipswich, MA, USA) and then sequenced using cycle sequencing on an ABI 3730 (Applied Biosystems, Carlsbad, CA, USA). Methylation indexes were calculated as a percentage of the number of methylated CpGs out of the total number of CpGs sequenced.

Previously described primers were used to amplify 129 bp and 150 bp fragments of the IDH1 and IDH2 genes [19]. The IDH1 forward primer 5′-CTCCTGATGAGAAGAGGGTTG-3′ and IDH1 reverse primer 5′-TGGAAATTTCTGGGCCATG-3′ were used to sequence codon 132 and the IDH2 forward primer 5′-TGGAACTATCCGGAACATCC-3′ and IDH2 reverse primer 5′-AGTCTGTGGCCTTGTACTGC-3 were used to sequence codon 172 of IDH2. Twenty nanograms of genomic DNA were used as starting material for a 25 μl total volume PCR reaction using Go Taq polymerase. An annealing temperature of 58°C was used for 35 cycles. PCR products were bi-directionally sequenced using cycle sequencing on an ABI 3730x (Applied Biosystems, Carlsbad, CA, USA).

Illumina Infinium HumanMethylation450 BeadChip array data was used for the following 19 TCGA primary glioblastomas: TCGA-06-5416, TCGA-06-0171, TCGA-26-5136, TCGA-06-0190, TCGA-06-5418, TCGA-06-0210, TCGA-26-5135, TCGA-26-5134, TCGA-26-5132, TCGA-12-5295, TCGA-06-5414, TCGA-06-0211, TCGA-26-5133, TCGA-06-5417, TCGA-06-0221, TCGA-26-1442, TCGA-06-6389, TCGA-06-6701, TCGA-15-1444. All array data was downloaded from the TCGA Data Portal (https://tcga-data.nci.nih.gov/tcga/tcgaHome2.jsp). IDH1 and IDH2 mutation status for these tumors was identified using the cBioPortal for Cancer Genomics (http://www.cbioportal.org/public-portal/).

Unsupervised clustering of the 2000 most variable loci in all 40 samples plus normal controls produces two major clusters: major cluster 1 (n = 20 samples; mean beta value = 0.21) and major cluster 2 (n = 24 samples; mean beta value = 0.60) (p < 0.001; ANOVA) (Figure 1a, b). Each major cluster can be further sub-divided into 2 sub-clusters: sub-clusters 1a and 1b (n = 13 and n = 7 samples respectively; mean beta values 0.14 and 0.34 respectively) and sub-clusters 2a and 2b (n = 12 samples in each cluster; mean beta values 0.50 and 0.69 respectively) (p < 0.001; ANOVA). Mean beta values for samples within each sub-cluster differ significantly in all comparisons (p < 0.05; ANOVA) (Figure 1b). Samples within major cluster 2 demonstrate a high level of methylation throughout the most variable 2000 loci indicating the CpG island methylator phenotype (CIMP) and these samples were designated CIMP^+ve with all but one sample (P19E) demonstrating an IDH1 mutation (Figure 1). Within our most variable 2000 loci were probes for genes previously associated with a CIMP phenotype in GBM [10]. Samples in major cluster 1 appear to be negative for the CIMP phenotype and were designated CIMP^–ve with sub-cluster 1a appearing to be notably normal-like, including all the control normal samples, whilst sub-cluster 1b has low level methylation. Interestingly, major cluster 1 included several IDH1 mutation positive samples as well as all the IDH1 mutation negative samples except for P19E (Figure 1a). In general, IDH1 mutation negative samples (P2, P3, P4 and P9) demonstrated very similar methylation patterns between early and late grades (Figure 1a). For all but one (P16) of the IDH1 positive samples, the lower grade sample demonstrated distinct CIMP and this is suggestive that it is a very early event in secondary gliomagenesis. In addition, no sample gained CIMP during progression suggesting that it occurs early on or not at all. In progression to the later grades the IDH1 positive sample split into two categories; those samples that retain a very similar methylation profile after progression (P5, P8, P10, P11, P15, P17) and those that demonstrate a partially remaining CIMP^+ve between early and late lesions or greatly reduced (becoming CIMP^-ve) degree of methylation after progression (P1, P7, P12, P13, P14, P18, P20) (Figure 1a). Thus, in total, progression through to higher grades had little effect on the genome-wide methylation for 10 of the 20 pairs (50%) and no effect on CIMP status for 16 of the 20 pairs (80%). The P19 sample acted as though an IDH mutation was present and that it fell into the second category of IDH1 mutation positive samples. Although the sample was negative for IDH1 or IDH2 mutation it could possibly have another mutation capable of causing a similar effect, such as a TET2 mutation, that was not assessed for. The IDH1 positive P16 sample acted more like an IDH1 negative sample for unknown reasons and was retained for further analysis.

To initially discern a list of differentially methylated loci between normal and tumor samples we first split the samples into grade II, III and grade IV groups and identified hypermethylated loci within each group. Following removal of all probes showing a β-value ≥0.25 in any of the four normal samples, the remaining probes were considered hypermethylated if >30% of tumor samples showed a β-value of ≥0.5. When using these criteria: 6024 CpG loci were identified as being hypermethylated in grade II astrocytomas of which 4374 were associated with a gene; 5295 CpG loci were identified as being hypermethylated in grade III astrocytomas of which 3772 were associated with a gene; 3329 CpG loci were identified as being hypermethylated in grade IV glioblastomas of which 2397 were associated with a gene. This trend of decreasing methylation levels is in agreement with our clustering data above and with previous studies [11, 20]. Further analysis was carried out only with probes that were associated with genes. The location of differentially methylated loci with respect to gene features was very similar for each grade, the majority being within the gene body (34.9%, 34.9% and 36.1% for grades II, III and IV respectively) and within 1500 bp of the transcription start site (22.6%, 21.7% and 20.2% for grades II, III and IV respectively), this largely followed the distribution of analyzed CpG probes as determined by array design (Additional file 3: Figure S1). However, we saw a very different distribution of hypermethylated CpG loci compared to design array when assessing the genomic location. In this case, the majority of hypermethylated probes fell within CpG islands (67.9%, 72.7% and 73.8% for grade II, III and IV respectively) while only 35.6% of total analyzed probes fell within these regions. In contrast, we saw very few hypermethylation events within open sea locations (7.1%, 7.1% and 5.5% for grades II, III and IV, respectively) compared to the total number of CpG loci analyzed within these locations (31.4%) (Additional file 3: Figure S1). Due to the large number of probes per gene in the Infinium HumanMethylation450 BeadChip array we were able to further refine our gene lists by removing genes that had limited CpG hypermethyaltion events. Removal of genes that were not represented by ≥3 probes resulted in 2189 relevant hypermethylated probes representing 496 genes in astrocytoma grade II samples, 1837 relevant hypermethylated probes representing 427 gene in astrocytoma grade III samples and 1208 relevant hypermethylated probes representing 279 genes in grade IV glioblastomas. Of the 2189 loci that are hypermethylated in grade II astrocytomas, approximately 24.9% (n = 544) are specifically hypermethylated within this group, in contrast, grade III and IV samples showed a lower level of specific grade methylation (10.8%, n = 198 and 8.4%, n = 102) respectively (Figure 2).

To try and identify genes important throughout secondary gliomagenesis it was assumed that genes hypermethylated in all glioma grades would be the most relevant. This analysis identified 939 hypermethylated CpG loci across all grades for further analysis. This list represents 232 genes and was, as before, reduced to 218 genes (represented by 914 CpG loci) by selecting genes that were represented in the list by ≥3 CpG loci probes. The gene list and beta values for these probes are provided in Additional file 4: Tables S2 and S3 respectively. Three genes (ALS2CL, GNMT and WNK2) were chosen from the list of 218 genes to confirm array values with regard to methylation. We chose two genes that had not previously been shown to be methylated in GBM (ALS2CL and GNMT) and one gene that has (WNK2) [21] for this technical validation of array results. Results from clone sequencing confirmed β-values >0.5 are representative of methylation and that very low β-values correspond to no methylation (Additional file 5: Figure S2; Additional file 1). Use of the Ingenuity Pathway Analysis software identified 47.7% (104/218) of genes as falling within five molecular and cellular function groups; cell morphology, cellular movement, cellular development, cellular growth and proliferation, and cellular assembly and organization. Of these 104 genes, 39 have been previously associated with cancer (Additional file 6: Table S4-S5).

Since there is evidence for primary and secondary gliomas having different genetic attributes and this is one of the first examples of the Illumina Infinium HumanMethylation450 BeadChip arrays on secondary gliomas we used a subset of the publically available Infinium HumanMethylation450 BeadChip primary GBM TCGA datasets to compare methylation in primary and secondary grade IV glioblastomas to determine any global methylation differences. To avoid any bias due to our sGBM data being adjusted for Illumina Type I/II probes (see Methods), we chose to use our sGBM data prior to adjustment for this particular analysis. Using 15 of TCGA grade IV pGBM and our 15 grade IV sGBM data, we clustered the most variable 2000 loci and observed that the pGBM samples largely clustered together, while the sGBM samples clustered into two separate groups dependent upon their CIMP phenotype (data not shown). This suggests there is also epigenetic heterogeneity between pGBM and sGBM, at least in terms of DNA methylation, but interestingly there were three of the pGBM samples that clustered within the sGBM CIMP^+ve group, one of which had the IDH1 p.R132H change. To further expand this analysis we then downloaded all pGBM IDH1 p.R132H mutated samples that had available Infinium HumanMethylation450 BeadChip methylation data (4 additional samples, there were no IDH2 mutated samples). Clustering of the total 19 pGBM samples with our 15 grade IV sGBM samples showed the CIMP phenotype within all five pGBM IDH1 mutated samples, which clustered together with our sGBM CIMP^+ve IDH1 mutation positive samples, indicating the CIMP^+ve phenotype induced by IDH1 mutation in pGBM is similar in sGBM. Two additional pGBM samples also clustered within this group, completing the smaller of the two major cluster groups (Figure 3). Of the larger, CIMP^-ve cluster, samples split into four sub-clusters, dependent predominantly on pGBM/sGBM status. Sub-clusters 1c and 1d contained all but two pGBM samples and only one sGBM sample whilst the remaining two sub-clusters contain all but one sGBM CIMP^-ve samples (Figure 3), indicating methylated targets of CIMP^-ve primary and secondary grade IV glioblastomas differ significantly. Three of the four sGBM CIMP^-ve IDH1 mutation positive samples are the samples that exhibited CIMP in the earlier lesion but not the later lesion, whilst the fourth sGBM and its paired earlier lesion were CIMP^-ve. Comparison of methylated gene lists for sGBM and pGBM samples (irrelevant of CIMP status) identified 180 genes that were only methylated in sGBM samples according to our criteria (Additional file 7: Table S6). We also identified 338 genes that were only methylated in pGBM samples (Additional file 7: Table S6) and 123 genes methylated in both. Reassuringly, only two genes were present in the pGBM specific list and our earlier list of 218 genes that were methylated across grade II, III and IV secondary gliomas. Of the 180 genes that were sGBM specific from this analysis, 115 were present in our list of 218 genes across grade II, III and IV secondary gliomas (Additional file 8: Figure S3). This discrepancy is most probably due to a combination of looking only at grade IV samples and using data unadjusted for Type I/II Illumina probes. Ingenuity analysis identified a substantial number of genes associated with cancer in both lists, with substantially more in the pGBM only list; 20% and 54% of genes within the sGBM and pGBM lists respectively. Considerable differences were observed between the molecular and cellular functions of genes within each list (Table 1). The pGBM gene list is enriched for genes that alter or control gene expression which in turn may affect cellular development and growth and proliferation. In contrast, the sGBM only list is enriched for genes that affect cell death, survival and maintenance pathways that would need to be altered or abrogated for tumorigenesis. Thus, the differing patterns of methylation between these two subtypes of glioma may provide differing advantages to these tumor cells.