Pathogenic Epigenetic Consequences of Genetic Alterations in IDH-Wild-Type Diffuse Astrocytic Gliomas

Gliomas represent approximately 80% of malignant central nervous system tumors. The World Health Organization (WHO) classification system subdivides diffuse astrocytic glioma (DAG) and oligodendroglioma into grade II to IV based on histopathologic findings (1, 2). Recent comprehensive studies revealed that WHO grade II and III DAG and oligodendroglioma, called lower grade gliomas (LGG), are classified by combining histopathologic and molecular features including isocitrate dehydrogenase (IDH) status. Mutations in the IDH gene have been found in approximately 80% of LGGs, which show a characteristic feature of relatively favorable prognosis, while IDH-wild-type LGGs display an aggressive phenotype (3–5). Among LGGs, all cases of oligodendroglioma exhibit IDH gene mutation, while DAGs exhibit either wild- or mutant-types IDH gene status. Therefore, LGGs are considered to be divided into IDH-mutant DAGs or oligodendrogliomas (IDH-mutant LGG), or IDH-wild-type DAGs.

Mutant IDH induces characteristic epigenetic alterations, such as extensive DNA hypermethylation (glioma CpG island methylator phenotype, G-CIMP), which may contribute to the initial step of glioma tumorigenesis (6, 7). In contrast, the molecular mechanisms underlying formation of IDH-wild-type DAGs appear to be heterogeneous. Recently, among IDH-wild-type DAGs, tumors with EGFR amplification, or combined whole chromosome 7 gain and whole chromosome 10 loss, or TERT promoter mutation, which are common characteristic molecular features of IDH-wild-type grade IV glioblastomas (GBM), exhibit a poor prognosis comparable with GBMs (8, 9). This type of DAG is termed diffuse astrocytic glioma, IDH-wild-type, with molecular features of glioblastoma, WHO grade IV (DAG-MF). However, the underlying molecular mechanisms related to the relevant molecular drivers for IDH-wild-type DAG without these features (i.e., IDH-wild-type DAG-nonMF) are poorly understood. It is a significant challenge to develop agents that specifically target this type of glioma.

The mouse model mosaic analysis with double markers (MADM) system is a useful technology to investigate the process of tumorigenesis from early to advanced stages (10). Starting from a heterozygous, colorless mouse, the MADM system generates a small number of homozygous mutant cells via inter-chromosomal mitotic recombination and labels them with green fluorescent protein (GFP) and wild-type cells with red fluorescent protein (RFP; ref. 10). In addition to mimicking the physiologic circumstances of tumor initiation, the unequivocal labeling of mutant cells enables the investigation of the entire tumorigenic process.

In this study, we used a glioma mouse model with a MADM system employing a concurrent knock-out of the p53 and NF1 genes, both of which are the most frequently mutated in human IDH-wild-type DAG-nonMF (40% and 30% of these gliomas, respectively). Notably, the pathologic features, and genetic and gene expression profiles of this MADM tumor match well with human IDH-wild-type DAG-nonMF. The resulting tumors showed an upregulation of enhancer of zeste homolog (EZH2), a catalytic component of polycomb repressive complex 2 (PRC2), together with increased H3K27 trimethylation (H3K27me3) on target genes including stem cell reprogramming-related genes, such as Wnt pathway genes. Furthermore, inhibition of EZH2 activity effectively repressed tumor growth both in vitro and in vivo. Our study clarifies a pathogenic molecular pathway, which depends on EZH2/PRC2 activity, and is partly involved in the generation of IDH-wild-type DAG-nonMF.

MADM mice were generated by crossing MADM-TG, p53 KO, NF1 flox mice, HIPP11^GT, MADM-11^GT mice and Tg (GFAP-cre) 25Mes mice obtained from The Jackson Laboratory (stock numbers: 017530, 013749 and 004600, respectively). Ezh2^flox/flox mice were obtained from RIKEN BRC (RBRC0555 B6-Ezh2 floxed). Oral administration of 160 mg/kg EPZ6438 (27626-3, BPS Bioscience and HY-13803, MedChemExpress) was performed twice daily for 20 days (11). EPZ6438 was formulated as a suspension in 1% carboxymethyl cellulose with 0.2% Tween 80 for oral administration (12). We analyzed MADM tumor mice treated with EPZ6438 or vehicle (1% carboxymethyl cellulose with 0.2% Tween 80) as a control (21 tumors in 8 mice and 14 tumors in 7 mice, respectively). All mice were anesthetized with isoflurane (1%–3% in air) and then placed in a 4.7T AVANCE III Pharma Scan (Bruker). During scanning, the breathing and depth of anesthesia were continuously monitored, with the breathing rate maintained at 80–100 breaths per minute. MRI images were visualized with open-source Osirix software that allowed the volumetric computation of tumors.

GFP-positive cells with the homozygous deletion of p53 and NF1 genes were collected at P8, P20, and P60 as precancerous cells and at P90 and P120 as tumor cells of the early and late phase, respectively, by immunopanning methods with anti-PDGFRα antibody-coated plates for P8 cells as described previously (13) and cell sorting using a FACS ARIA II (BD Biosciences).

Cells were derived from various individual MADM tumors. These cells were plated onto flasks, coated with poly-d-lysine (PDL, P6407, Sigma Aldrich) and cultured in Neurobasal medium (21103, Thermo Fisher Scientific) with B27 (17504-044, Thermo Fisher Scientific), Glutamax (35050061, Thermo Fisher Scientific), 10 ng/mL human recombinant PDGF-AA (100-13-A, PeproTech), and 1% penicillin–streptomycin (15140-122, Thermo Fisher Scientific) at 37°C in a humidified incubator with 5% CO₂. U87 (IDH-wild-type grade IV) and SW1783 (IDH-wild-type LGG) cell lines were obtained from ATCC. The TM31 (IDH-wild-type LGG) and T98 (IDH-wild-type GBM) cell lines were obtained from RIKEN BRC (RIKEN Cell Bank). Onda10, Onda11, KINGS1 (IDH-wild-type LGG), and U251 (IDH-wild-type GBM) cell lines were obtained from the Japanese Collection of Research Bioresources (JCRB) Cell Bank. GL261 cell line was obtained from the NCI tumor-cell-line repository (Frederick, MD). Normal human astrocytes (NHA) were kindly gifted by Dr. R. Pieper (University of California, San Francisco, San Francisco, CA; ref. 14). All of these cell lines were analyzed within 6 months after resuscitation. These cell lines were authenticated in original cell banks and were tested to be Mycoplasma free when resuscitation. U87, T98, GL261, and NHA cells were cultured with DMEM (044-29765, Wako Pure Chemical Industries, Ltd.) containing 10% FBS and 1% penicillin–streptomycin (15140-122, Thermo Fisher Scientific). TM31, Onda10, and Onda11 cells were cultured with Eagle MEM (M4655, Sigma Aldrich) medium containing 10% FBS. U251 and KINGS1 cells were cultured with RPMI1640 (189-02025, Wako Pure Chemical Industries, Ltd.) containing 5% FBS. All cell lines were maintained at 37°C in a humidified incubator with 5% CO₂.

Chromatin immunoprecipitation (ChIP) assays were performed according to previously published methods with minor modifications (15–17). Cells (up to 1 × 10⁶) were treated with 1% formaldehyde for 8 minutes to crosslink histones to DNA. The cell pellets were resuspended in lysis buffer (1% SDS, 10 mmol/L EDTA, 50 mmol/L Tris-HCl pH 8.1, and protease inhibitor) and sonicated using a Covaris S220 system (Covaris). After diluting the cell lysate 1:10 with dilution buffer (1% Triton, 2 mmol/L EDTA, 150 mmol/L NaCl, 20 mmol/L Tris-HCl pH 8.1), diluted cell lysates were incubated for 16 hours at 4°C with Dynabeads protein G (100-03D, Thermo Fisher Scientific), which were precoated with 2 μL of anti-histone H3 antibodies (ab1791, Abcam) or Dynabeads M280 sheep anti-Mouse IgG (112-01D Thermo Fisher Scientific), which were coated with 2 μL of anti-K4 trimethylated and anti-K27 trimethylated histone H3 antibodies (CMA304 16H10 and CMA323 1E7, respectively; ref. 18). ChIP products were analyzed by SYBR Green ChIP-quantitative PCR (qPCR) using the primer sets shown in Supplementary Table S1. Negative control (NC) region for ChIP-qPCR was selected according to the data of reference (19). We analyzed the ChIP products as probes on a SurePrint G3 mouse promoter array 2 × 400K according to the manufacturer's protocol (ChIP-chip; G4858A; Agilent Technologies; ref. 16). Results were analyzed by a Whitehead Per-Array Neighbourhood Model (modified) to identify bound genes in an Agilent Genomics Workbench (version 7.0, Agilent Technologies). A probe was considered to be bound, if the mean P value of 3 neighboring probes was less than 0.05 and if one of the following requirements was met: (i) the P value of the central probe was less than 0.05 and the P value of at least one neighboring probe was less than 0.1; or (ii) the P value of at least 2 of the neighboring probes were less than 0.01 (20). These data were visualized in GeneSpring GX, version 12.6.0 (Silicon Genetics). Neural stem cells (NSC), astrocytes (AST), and P8 cells were collected from 4–6 mice (a single ChIP-chip experiment). MADM tumor cells were collected from a MADM mouse. Two independent ChIP-chip experiments were performed. We determined genes modified by each chemical modification commonly across two individual experiments as genes with these modifications in examined cells. We also carried out ChIP-seq using fragmented ChIP products (0.5 ng to 10 ng) after preparing DNA libraries for paired-end sequencing using the NEBNext Ultra II DNA library Prep kit for the Illumina HiSeq 2500 next generation sequencing (NGS) system (New England Biolabs) to validate ChIP-chip data. ChIP-seq clean reads were obtained from raw reads by the removal of adaptors and low-quality reads. The clean reads were aligned to the mouse genome (mm9) using bowtie (version 0.12.8) with parameters -a –best –strata -m 3 -v 3 (21). The levels of histone modification at gene promoters were estimated by counting the numbers of reads mapped within a 2-kb region around the transcriptional start site using HTSeq (version 0.10.0; ref. 22). ChIP-seq experiments using anti-H3K4me3 and anti-H3K27me3 antibodies were performed in duplicate in NSCs, ASTs, P8 precancerous cells, and P120 MADM tumor cells.

o analyze various gene expression patterns in clinical glioma samples, we used the normalized RNA sequencing (RNA-seq) expression data (Illumina Hiseq) from the UCSC cancer genomics browser (https://genome-cancer.ucsc.edu) and the UCSC Xenabrowser (https://xenabrowser.net) in relation to the Cancer Genome Atlas (TCGA) brain lower grade glioma, TCGA glioblastoma multiforme and TCGA pan-cancer (23, 24) and the level-3 preprocessed RNA-seq expression data of grade II and III gliomas in the TCGA portal (https://tcga-data.nci.nih.gov/). Gene set enrichment analysis (GSEA) was performed using the GSEA desktop application (Broad Institute). We analyzed these datasets with R Statistical Software (version 2.14.0; R Foundation for Statistical Computing). The Z score for the expression level of each gene was calculated using the mean expression value and SD among all grade II, III, IV tumors and normal brains or among all genes in each individual.

mRNA RSEM (RNA-Seq by Expectation Maximization) expression profile and genetic mutation profile were obtained with TCGA GBMLGG data (2015_08_21_stddata Run and 2016_01_28 stddata Run). After stratifying samples according with DAGs-MF and DAGs-nonMF definition, relative changes of expression levels of 544 genes, which correspond to upregulated genes specifically in MADM tumor cells (top 5% upregulated transcripts in MADM tumor cells compared with normal brains) in each tumor sample against mean expression levels among normal samples with log₂[(RSEM _(g,i) of tumor+0.5)/(mean RSEM _(g) of normal samples+0.5)] where “g” and “i” indicate g-th gene and i-th sample. 0.5 was added to all values to avoid dividing by zero counts. These calculated values were directly mapped as heatmap colors. Microarray-based gene expression values of the triplicate MADM tumor samples and the duplicate normal samples were quantile normalized with a background correction using R package limma (25). We detected the most differentially expressed transcripts in MADM tumors compared with those of normal samples. Corresponding mouse gene symbols were converted to human symbols by R package biomaRt (26, 27) and plugged into gene expression data compiled from TCGA data (2015_08_21_stddata Run and 2016_01_28 stddata Run).

Consensus clustering was performed utilizing hierarchical clustering based on Ward and Euclidean correlation algorithms with 1,000 iterations using Bioconductor package ConsensusClusterPlus. We selected 80% item resampling, 90% gene resampling, and a maximum evaluated K of 5. The number of clusters was determined by the relative change in area under a cumulative distribution function curve by consensus clustering.

The statistical significance of differences between two groups was analyzed by paired Student t test and Fisher exact test. Microarray data, RSEM, and fragments per kilobase of exon per million reads mapped (FPKM) analyses were performed by paired Student t test, Wilcoxon signed rank test and Wilcoxon rank sum test. Correlation coefficient was calculated with Pearson correlation test. All reported P values were two-sided, with P < 0.05 considered statistically significant.

The Gene Expression Omnibus accession code for the ChIP-chip microarray data of NSCs and ASTs along with MADM tumors is GSE110926. Microarray-based gene expression data of MADM tumor tissues and normal brain tissues can be found at GSE91050, while that related to Ezh2^Δ/wt MADM tumor tissues is at GSE110927. GSE110926 and GSE110927 are also integrated as GSE110937. ChIP-seq and RNA-seq data are available in the DNA Data Bank of Japan Sequence Read Archive (DDBJ DRA; accession number is DRA008249).

All protocols for animal experiments were approved by the Institutional Animal Care and Use Committee of Nagoya University (Nagoya, Japan; 29476). Experiments with an oral administration procedure were also approved by the Committee of Animal Experiments at the National Center for Geriatrics and Gerontology.

Using the public database of TCGA, we investigated the molecular profiles of IDH-wild-type DAGs (n = 64 and n = 13, DAG-MF and DAG-nonMF, respectively). Although significantly more genes were mutated in IDH-wild-type DAGs-MF than DAGs-nonMF, EGFR, PTEN, NF1, TTN, and p53 are commonly mutated in more than 10% of both IDH-wild-type DAGs-MF and -nonMF (Fig. 1A and B). Of note, EGFR was the most commonly mutated gene with the highest mutation frequency and variant allele fraction (VAF) in IDH-wild-type DAGs-MF (40% and 55%), while p53 and NF1 were the genes with high mutation frequency and VAF in IDH-wild-type DAGs-nonMF (40% and 46%, 30% and 61%, respectively; Fig. 1C). GSEA indicated that highly expressed RAS–MAPK signaling pathway genes were enriched in IDH-wild-type DAG-nonMFs compared with IDH-wild-type DAG-MFs (P < 0.01, Fig. 1D). Concurrent genetic alterations in both p53 and RAS–MAPK signaling pathway-related genes including NF1, a negative regulator of RAS–MAPK signaling, are more frequent in IDH-wild-type DAGs-nonMF (23.1%) than IDH-wild-type DAGs-MF (3.1%; P < 0.05). These data indicated that genetic alterations of p53 and NF1 genes play important roles in IDH-wild-type DAGs, particularly nonMF, formation.

To clarify the molecular significances of IDH-wild-type DAGs, a MADM glioma mouse model system was examined. The MADM system generates cells with the homozygous deletion of p53 and NF1 genes (GFP-positive cells), and MADM mice spontaneously develop IDH-wild-type anaplastic gliomas (MADM tumors) whose expression profiles were concordant with the proneural cell type between postnatal days 90 and 120 (P90–P120; Fig. 1E; Supplementary Fig. S1; ref. 13). Pathologic findings showed that MADM tumors are compatible with anaplastic astrocytic glioma without prominent necrosis and microvascular proliferation, which is consistent with a previous study (Fig. 1F; ref. 13). We identified 401 transcripts (top 1% of highly expressed transcripts) that had concordant and markedly higher expression levels in MADM tumor tissues (n = 3) compared with normal brain tissues (n = 2; Supplementary Table S2). TCGA data (n = 581; 194 grade II, 221 grade III, and 166 grade IV) revealed that the expression levels of the 353 human genes that corresponded to the identified 401 mouse transcripts were significantly higher in IDH-wild-type DAGs and GBMs (P < 0.01, Supplementary Fig. S2A) compared with IDH-mutant LGGs. K-means clustering analysis and consensus cumulative distribution function (CDF) analysis revealed that the expression profile of these genes could be used to divide the tissues into two groups dependent on their IDH-mutant or IDH-wild-type status (P < 0.01, Supplementary Fig. S2B). Furthermore, using the expression profiling of the top 10% of highly expressed transcripts (corresponding to 2,411 genes), we obtained concordant findings (Supplementary Fig. S2C and S2D).

Given that IDH-wild-type DAGs contain different subgroups with characteristic molecular features (i.e., IDH-wild-type DAG-MF and -nonMF, Fig. 1G), we investigated the genetic and gene expression profiling of MADM tumors further to identify which IDH-wild-type DAG subgroups are highly comparable with the MADM tumors. Comprehensive genetic analysis revealed that MADM tumors exhibited no EGFR amplification and no combined whole chromosome 7 gain and whole chromosome 10 loss, except for the partial loss of chromosome 10q (Fig. 1H). Sanger sequencing revealed no TERT promoter mutation within the transcription start site ± 1 kb. These data indicated that MADM tumors were not genetically compatible with IDH-wild-type DAG-MF. Next, we selected 544 genes whose expression levels were upregulated specifically in MADM tumor cells but not in a mouse GBM cell line, GL261, which is comparable with human IDH-wild-type GBM (28). Heatmap analysis showed that the expression status of corresponding human 544 genes could be used to clearly divide the IDH-wild-type DAG into two groups (Supplementary Table S3). Expression levels of 544 genes were significantly higher in human IDH-wild-type DAGs-nonMF (n = 13) compared with IDH-wild-type DAGs-MF (n = 64; P < 0.01, Fig. 1I). These data indicate that gliomas in MADM mice are comparable with human IDH-wild-type DAG-nonMF by means of histopathologic and molecular features and that analyzing this mouse model may provide clues to the underlying tumorigenic processes of IDH-wild-type DAG-nonMF.

Epigenetic dysregulation occurs at early stages of tumorigenesis and causes altered proliferation and differentiation throughout tumor evolution, including cases of glioma (29). Considering that IDH-wild-type DAGs-nonMF exhibit less genetic alterations in comparison with IDH-wild-type DAGs-MF (Fig. 1B), we focused on epigenetic alterations subsequent to p53 and NF1 gene alterations during MADM tumor formations. We examined the expression level of epigenome-associated genes in MADM tumors. Among 162 epigenome-associated genes, which were previously functionally classified in human gliomas (30), we identified 116 that were substantially expressed and altered at a genetic level in either human LGGs (n = 415) or grade IV GBMs (n = 292; Supplementary Table S4; refs. 5, 30). Among the 116 epigenome-associated genes, Ezh2 was the most upregulated gene in MADM tumors (Supplementary Fig. S3A)

EZH2 is the components of PRC2 for its catalytic activity of H3K27me3 modification along with SUZ12 and EED. EZH2 is highly upregulated in MADM tumor cells compared with those in normal NSCs or astrocytes (AST; Fig. 2A). We evaluated the sequential changes of Ezh2, Suz12, and Eed expressions in GFP-positive precancerous cells and glioma cells collected by FACS sorting at different time points (P20 and P60: precancerous cells, P90 to P120: MADM tumor cells). Expression levels of Ezh2, Suz12, and Eed were consistently highly upregulated from the early precancerous stage (P8, P20) to glioma (P90–P120), resulting in gradually increased level of H3K27me3 in GFP-positive cells (Fig. 2B and C). Immunohistochemistry (IHC) assay of MADM mouse brain tissues revealed the high expression of EZH2 in GFP-positive precancerous cells (i.e., homozygous deletion of p53 and NF1 genes) even in P20 mutant mice brain and tumor tissues in P120 mice, compared with GFP-negative cells (Fig. 2D). Next, we examined the expression status of 116 epigenome-associated genes in human IDH-wild-type DAGs-MF (n = 64) and -nonMF (n = 13) compared with those in normal brains (n = 5; Supplementary Table S4). EZH2 was also the most upregulated gene among them compared with normal brains (Supplementary Fig. S3B and S3C). In addition, all human glioma cell lines examined, such as TM31 (IDH-wild-type DAG-nonMF cell line), Onda10, Onda11, KINGS1 (IDH-wild-type DAG-MF cell lines), T98, U251, and U87 (IDH-wild-type GBM cell lines) showed high protein levels of EZH2 (Supplementary Fig. S3D).

Recent analysis revealed the nonsimultaneous overexpression of EZH2, EED, and SUZ12 in human gliomas: the frequent overexpression of EZH2 and relatively lower frequencies of EED and SUZ12 overexpression were observed (31), indicating that specific EZH2 functions are independent of the PRC2-repressive function (32) in a certain population of gliomas. Of note, the expression levels of EZH2, SUZ12, and EED were highly concordant with each other in IDH-wild-type DAG-nonMF compared with IDH-wild-type DAG-MF in the TCGA dataset (Fig. 2E). Consistently, lower expression levels of known PRC2 target genes (904 genes; ref. 33) were observed in IDH-wild-type DAG-nonMF compared with IDH-wild-type DAG-MF (P < 0.01, Fig. 2F). These data suggest the PRC2 dysregulation subsequent to genetic alterations of p53 and NF1 genes play pivotal roles in MADM tumor and IDH-wild-type DAGs-nonMF formation as well.

Epigenomic analysis of the MADM mouse model suggested an important role for EZH2/PRC2 during the tumorigenesis of IDH-wild-type DAG-nonMF tumors. Next, we generated Ezh2^Δ/wt MADM mice, in which the Ezh2 catalytic SET domain was heterozygously deleted by the Cre/loxP system (34). In MADM mice, the GFAP promoter becomes active in NSC at E13.5 days and regulates Cre expression. In Ezh2^Δ/wt MADM tumors (P120), the mRNA expression level of the Ezh2 SET domain region was significantly decreased compared with that in Ezh2 wild-type MADM tumors (P120; Fig. 3A). Consistently, H3K27me3 levels were significantly reduced in Ezh2^Δ/wt MADM tumors compared with those of Ezh2 wild-type MADM tumors (P < 0.01; Fig. 3B). Tumor growth in Ezh2^Δ/wt MADM mice (n = 14) was significantly suppressed compared with that observed in Ezh2 wild-type MADM mice (n = 10), resulting in significantly prolonged survival (P < 0.05; Fig. 3C).

Gene expression analysis revealed that 175 genes and 261 genes were significantly upregulated and downregulated, respectively, in Ezh2^Δ/wt MADM mice compared with Ezh2 wild-type MADM mice (Fig. 3D; Supplementary Table S5). Expression levels of the 175 upregulated genes were significantly lower in Ezh2 wild-type MADM tumors than in normal brain tissues. Heterozygous Ezh2 catalytic inactivation restored the expression of these genes to almost normal levels (P < 0.01; Fig. 3E). We identified genes modified with H3K4me3, H3K27me3, or H3K4me3 and H3K27me3 bivalent modifications in Ezh2 wild-type MADM tumor cells with ChIP-chip (H3K4me3 alone: 4,895 genes, H3K27me3 alone: 690 genes and bivalent: 637 genes; representative genes were listed in Supplementary Table S6). The histone modification status of these genes was reflective of gene expression levels in Ezh2 wild-type MADM tumors (Supplementary Fig. S4A). In addition, repressed genes with H3K27me3 and bivalent H3K4me3 and H3K27me3 modifications in Ezh2 wild-type MADM mice were significantly upregulated in Ezh2^Δ/wt MADM mice (P < 0.01, Fig. 3F). We further performed ChIP-seq analysis in NSCs, ASTs, P8 precancerous cells, and P120 MADM tumor cells. The experimental concordances are shown in Supplementary Fig. S4B. The relationship between each modification and corresponding gene expression revealed a significant positive and negative correlation of H3K4me3 and H3K27me3, with mRNA expression, respectively, in all cell types analyzed in duplicate (Supplementary Fig. S4B). GO analysis revealed that the 175 upregulated genes included many neural differentiation-associated genes (Fig. 3G; Supplementary Table S7), while no significant GO terms were found in 261 downregulated genes (P > 0.05).

Among the 175 significantly upregulated genes caused by Ezh2 gene depletion, Fzd8 showed the lowest expression among genes that were modified by bivalent modification in MADM tumors but were active in P8 cells (Fig. 4A). FZD8, a member of the frizzled protein family of G protein–coupled receptors involved in Wnt signaling, inactivates β-catenin function (35). IHC analysis revealed that FZD8 protein expression levels in Ezh2 wild-type MADM tumors were significantly decreased compared with that in Ezh2^Δ/wt MADM tumors (Fig. 4B). In MADM mice, the Fzd8 gene promoter was consistently modified with a H3K4me3 mark in P8 cells, while it acquired H3K27me3 and was modified with bivalent modification in MADM tumor cells (Fig. 4C). The mRNA and protein expression levels of FZD8 were downregulated during tumor formation (Fig. 4D). Ccnd1 is one of the most studied downstream genes of β-catenin transactivation, which promotes cell proliferation (36). During MADM tumor formation, RNA expression level of Ccnd1 gene was upregulated (Fig. 4E).

Depletion of Fzd8 by siRNA increased the proliferation of P20 GFP–positive precancerous cells, while minimal effects were seen in MADM tumor cells, in which Fzd8 had already been repressed (Fig. 4F). In contrast, the overexpression of Fzd8 decreased the proliferation of MADM tumor cells, while no effects were seen in P20 GFP-positive precancerous cells (Fig. 4G). Notably, the expression levels of FZD8 mRNA in human gliomas were lower than those in normal brains (Fig. 4H). Taken together, our data indicate that the dysregulation of EZH2/PRC2 induces aberrant histone modification and may abolish the normal differentiation program via disruption of important mediator pathways, such as FZD8-Wnt–mediated signaling.

Selective EZH2 catalytic inhibitors are effective against hematopoietic malignancy. Therefore, we examined the effects of pharmacologic EZH2 inhibition on MADM tumors by the post-oral (PO) administration of EPZ6438, a selective small-molecule inhibitor of EZH2 (11). EPZ6438 significantly inhibited the cell growth of MADM tumor cells in vitro. EPZ6438 transcriptionally downregulated Ccnd1 expression together with upregulation of FZD8 expression via reduction of H3K27me3 level in MADM tumor cells (Fig. 5A). Of note, EPZ6438 significantly inhibited the cell growth of TM31, a representative human IDH-wild-type DAG-nonMF cell line, which has no EGFR amplification, no combined whole chromosome 7 gain and whole chromosome 10 loss, and no TERT promoter mutation (Fig. 5B; Supplementary Fig. S5). In TM31 cells, EPZ6438 downregulated β-catenin expression together with upregulation of FZD8 expression via reduction of H3K27me3 level (Fig. 5B). In contrast, EPZ6438 had a minimal effect on the growth of human IDH-wild-type DAG-MF cell lines (Onda10, Onda11, KINGS1, and SW1783; all cells have a TERT gene promoter mutation) and human IDH-wild-type GBM cell lines (T98, U251, and U87), although EPZ6438 reduced the H3K27me3 level in these cells and TM31 at the comparable level (Fig. 5C). Furthermore, EPZ6438 had a minimal effect on the growth of a mouse GBM cell line, GL261 (Fig. 5D).

In MADM mice, substantial tumors can be detected by MRI imaging at P90. We therefore started the treatment of MADM mice from P90 onwards. Treatment arms were the PO administration of EPZ6438 at a dose of 160 mg/kg (n = 8) or PBS as a control (n = 7), twice a day for 20 days (11). Consistent with the genetic inhibition of Ezh2, EPZ6438 treatment significantly suppressed tumor growth compared with controls, along with the downregulation of H3K27me3 and upregulation of FZD8 expression in tumor tissues (Fig. 5E and F).

Precision medicine based on the characteristics of each cancer type is required to deliver the most appropriate treatment. Recent studies, including our own, have extensively demonstrated multiple genetic and epigenetic alterations in glioma formation (3–5, 30, 37). While G-CIMP is a characteristic epigenetic feature of IDH-mutant LGG (6, 7), the molecular basis underlying the tumorigenesis of the often-devastating IDH-wild-type DAG is still largely unclear. In this study, we identified the dysregulation of EZH2/PRC2 in both human IDH-wild-type DAG-nonMF and a MADM glioma mouse model in which the genetic and gene expression profile and histopathologic findings resemble this type of glioma.

In our mouse model, the loss of p53 and NF1 genes may affect the PRC2/EZH2-dependent epigenetic machinery at the early stage of malignancy. Interestingly, Ezh2 expression was not affected in mice with a p53 homozygous deletion alone (Supplementary Fig. S6). Given previous evidence that the genetic loss of NF1 activated RAS/MEK/ERK signaling (38) and that the activation of RAS/MEK/ERK signaling upregulated EZH2 expression (39, 40), we argue that the deletion of NF1 may contribute to the upregulation of EZH2 in our MADM mouse model.

Aberrant activation of the Wnt pathway is involved in gliomagenesis via multiple mechanisms. Various negative regulators of the Wnt pathway, such as FAT Atypical Cadherin 1 (FAT1), and Frizzled-related proteins (FRP) such as Dickkopf (DKK), are downregulated in gliomas (41, 42). WNT5A, a classical component of noncanonical Wnt signaling, is commonly upregulated in gliomas and sometimes promotes its invasiveness (43). In this study, we newly report the aberrant activation of the Wnt pathway via the PRC2/EZH2-mediated epigenetic repression of FZD8. Human FZD8 is expressed in the brain, pancreas, kidney, heart, and muscle (44). It is also involved in hematopoietic stem cell maintenance via activation of the noncanonical Wnt pathway (35). A previous study reported that abundant FZD8 expression in normal pancreas tissues was frequently lost in pancreatic adenocarcinomas (45). Mouse model analysis revealed that oncogenic K-Ras causes a tumorigenic phenotype through the downregulation of noncanonical Wnt/Ca²⁺ signaling via the repression of FZD8. Restoring FZD8 in K-Ras–mutant pancreatic cells suppressed malignancy, indicating that the suppression of FZD8-mediated Wnt signaling was essential to K-Ras malignancy (45). Because the RAS/MAPK signaling suppressor, NF1, is deleted in the MADM mouse model, suppression of Fzd8 via aberrant bivalent modification may also play a key role in gliomagenesis via activation of Wnt signaling. Indeed, the suppression or overexpression of FZD8 in GFP-positive MADM cells significantly altered cell growth.

The balance between H3K4me3 and H3K27me3 modification levels is crucial for proper regulation of differentiation. It is known that pluripotent embryonic stem (ES) cells and multipotent NSCs exhibit characteristic bivalent modifications of concurrent active H3K4me3 and repressive H3K27me3, which are characteristic modifications of suppressive and poised transcription on a set of important differentiation-associated genes. Dysregulation of these modifications, such as inappropriate activation by the removal of H3K27me3 or constitutive repression by the removal of H3K4me3, may contribute to tumor initiation and/or progression (46). Fzd8 gene was modified with bivalent modification in ES cells and was activated by the removal of H3K27me3 in NSCs (47). During glioma tumorigenesis, repression of Fzd8 by aberrant bivalent modification may play a pivotal role in gliomagenesis via dysregulation of neural differentiation together with promoting of cell proliferation.

A previous study showed the oligodendrocyte precursor cell (OPC) was the cell of origin for glioma in the MADM mouse model (13). Furthermore, the deletion of NF1 led to mutant OPC expansion through increased proliferation and decreased differentiation, while the deletion of p53 impaired OPC senescence, indicating that specific aberrations caused by individual gene mutations and altered signaling events occur at specific times during gliomagenesis (48). In this study, we provide further key evidence that aberrant epigenetic phenomena induced by certain tumor-related genes affect the activation of reprogramming-associated genes during gliomagenesis. Recently, a conceptual framework was proposed whereby epigenetic functional disorders in cancers can be driven by epigenetic modulators, epigenetic modifiers, and epigenetic mediators (49). An epigenetic modulator regulates the function of epigenetic modifiers and leads to the regulation of epigenetic mediators whose gene products regulate proliferation, development, and cellular plasticity. Mutations in modulators and modifiers are often selected for during cancer development (49). Our data shows that aberrant histone modifications were reproducibly induced by concomitant p53 and NF1 genetic loss during the tumorigenic process from precancerous cells to MADM tumors within individual MADM mice.

Although EZH2 is upregulated in malignant gliomas (50), the therapeutic benefit of targeting EZH2 as a novel therapy for gliomas is still under investigation. Prolonged EZH2 depletion has been reported to lead to more aggressive phenotypes of gliomas (51). In contrast, EZH2 inhibitors appear to be effective in certain types of gliomas such as diffuse intrinsic pontine glioma, which is the most lethal form of pediatric glioma (52, 53). The noncanonical roles of EZH2 independent of PRC2 activity and other PRC2 components, should be addressed before targeting EZH2 as a treatment (32, 54). In IDH-wild-type DAGs-nonMF, three core components of PRC2 are more closely associated with each other than in IDH-wild-type DAGs-MF, suggesting the importance of PRC2 activity in this type of glioma.

Comprehensive mRNA analysis also revealed that gliomas can be divided into four subtypes based on gene expression profiling, namely classical, neural, mesenchymal, and proneural subtypes (55). A recent study demonstrated that the high expression of EZH2 was preferentially found in proneural type glioma stem cells and that the inhibition of EZH2 preferentially reduced the growth of this type of cells (56). In MADM tumors, gene expression profiles were concordant with the proneural cell type, as reported previously (13). Given that the genetic and pharmacologic inhibition of EZH2 efficiently suppresses gliomas in MADM mice and a human IDH-wild-type DAG-nonMF cell line, EZH2 chemical inhibitors may be applicable for the treatment of IDH wild-type DAG-nonMF. In contrast, the chemical inhibitor of catalytic activity of EZH2 had the minimal effects on cell proliferation of IDH-wild-type DAGs-MF cell lines (Onda10, Onda11, KINGS1, and SW1783) and IDH-wild-type grade IV GBM cell lines (T98, U251, and U87) although it reduced H3K27me3 levels in these cells at comparable level with that in TM31. This may be explained by the accumulation of other genetic and epigenetic alterations in IDH-wild-type DAGs-MF and GBMs (Supplementary Fig. S7). Specific EZH2 functions, which are independent of the PRC2 repressive function, may contribute to gliomagenesis of IDH-wild-type DAGs-MF and GBMs. Further clinical trials are required to examine this possibility.

To achieve the essential requirements of precision medicine for patient-specific therapy, our combined analyses of human IDH-wild-type DAGs and the MADM mouse system will help improve the stratification of cancers by utilizing biological biomarkers and the development of effective treatment strategies. Our data clarify a pathogenic molecular pathway of IDH-wild-type DAG-nonMF, which provides a strong rationale for targeting EZH2 to treat this subtype of glioma.

F. Ohka has received speakers bureau honoraria from Daiichi Sankyo. No potential conflicts of interest were disclosed by the other authors.

Conception and design: F. Ohka, K. Katsushima, A. Suzumura, H. Zong, A. Natsume, Y. Kondo

Development of methodology: F. Ohka, K. Shinjo, Y. Okuno, K. Katsushima, K. Aoki, N.A. Rayan, H. Kimura, H. Zong

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): F. Ohka, K. Shinjo, S. Deguchi, Y. Okuno, A. Kato, N. Ogiso, K. Aoki, R. Maruyama, S. Takahashi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): F. Ohka, K. Shinjo, Y. Matsui, Y. Okuno, M. Suzuki, N. Ogiso, K. Aoki, H. Suzuki, N.A. Rayan, S. Prabhakar, J. Göke, T. Shimamura

Writing, review, and/or revision of the manuscript: F. Ohka, Y. Okuno, K. Aoki, S. Sato, A. Natsume, Y. Kondo

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): F. Ohka, Y. Okuno, A. Yamamichi, S. Sato, H. Kimura, H. Zong

Study supervision: A. Suzumura, T. Wakabayashi, H. Zong, A. Natsume, Y. Kondo

This study was performed as a research program of P-CREATE, Japan Agency for Medical Research and Development (18cm0106108h0003 to Y. Kondo), and of the Grant-in-Aid for Scientific Research, the Japan Society for the Promotion of Science (17H03582 to Y. Kondo; 17928985 to A. Natsume).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.