Research ArticleCancer

Lowered H3K27me3 and DNA hypomethylation define poorly prognostic pediatric posterior fossa ependymomas

See allHide authors and affiliations

Science Translational Medicine  23 Nov 2016:
Vol. 8, Issue 366, pp. 366ra161
DOI: 10.1126/scitranslmed.aah6904

Epigenetics helps find the bad tumors

Ependymomas are brain tumors that can occur in people of all ages and in different parts of the central nervous system. The prognosis of these tumors does not necessarily correlate with clinical characteristics or even tumor grade, and there are no recurrent genetic mutations that can be used to classify these tumors. To address this problem, Bayliss et al. examined the epigenetics of ependymoma. By doing this, the authors identified some characteristic methylation patterns that correlate with prognosis, including one specific pattern that is also seen in childhood gliomas and associated with more invasive tumors.


Childhood posterior fossa (PF) ependymomas cause substantial morbidity and mortality. These tumors lack recurrent genetic mutations, but a subset of these ependymomas exhibits CpG island (CpGi) hypermethylation [PF group A (PFA)], implicating epigenetic alterations in their pathogenesis. Further, histological grade does not reliably predict prognosis, highlighting the importance of developing more robust prognostic markers. We discovered global H3K27me3 reduction in a subset of these tumors (PF-ve ependymomas) analogous to H3K27M mutant gliomas. PF-ve tumors exhibited many clinical and biological similarities with PFA ependymomas. Genomic H3K27me3 distribution showed an inverse relationship with CpGi methylation, suggesting that CpGi hypermethylation drives low H3K27me3 in PF-ve ependymomas. Despite CpGi hypermethylation and global H3K27me3 reduction, these tumors showed DNA hypomethylation in the rest of the genome and exhibited increased H3K27me3 genomic enrichment at limited genomic loci similar to H3K27M mutant gliomas. Combined integrative analysis of PF-ve ependymomas with H3K27M gliomas uncovered common epigenetic deregulation of select factors that control radial glial biology, and PF radial glia in early human development exhibited reduced H3K27me3. Finally, H3K27me3 immunostaining served as a biomarker of poor prognosis and delineated radiologically invasive tumors, suggesting that reduced H3K27me3 may be a prognostic indicator in PF ependymomas.


Pediatric brain tumors are the most common solid malignancy of childhood and consist of diverse tumor types. Among these, pathogenesis of ependymomas, compared to that of medulloblastomas and pediatric glioblastomas (GBMs), remains poorly understood. The World Health Organization (WHO) classifies ependymomas as grade II or III (anaplastic) based on histopathologic criteria (1). However, histopathologic grading correlates poorly with patient outcome and treatment response (2). Further, ependymomas show variable outcomes to therapy, with some patients responding well and others refractory to treatment (3, 4). Therefore, there is an urgent and unmet need to better understand the biology of ependymomas to develop more robust disease biomarkers for prognosis and to create more effective treatments.

Ependymomas develop throughout the neuraxis but predominantly occur in three distinct regions: posterior fossa (PF), supratentorial (ST; cerebral hemispheres), and spinal cord (SC) (1). PF ependymomas are frequently found in children, whereas ST and SC ependymomas are more common in adults (1, 2). Histologically, ependymomas within each of these locations have similar morphology but behave differently based on molecular heterogeneity. For example, ST ependymomas with poor prognosis have C11orf95-RELA fusions (1, 2, 5). Similarly, PF ependymomas have been subgrouped into group A (PFA) and group B (PFB) tumors based on gene expression profiles and CpG island (CpGi) methylation phenotype (CIMP) status (1, 2, 68). PFA ependymomas mainly occur in children, have a distinct gene expression profile, exhibit CpGi hypermethylation (CIMP+ve), are more invasive, and bear a worse prognosis. In contrast, PFB ependymomas occur in older individuals and younger adults, are CIMP-ve and less invasive, and carry a better prognosis (1, 2, 68).

Two independent groups using whole-exome and whole-genome sequencing have demonstrated that PF ependymomas do not harbor recurrent genetic alterations (5, 7). CpGi hypermethylation in the absence of recurrent genetic changes in PFA ependymomas suggests that epigenetic mechanisms are central to PF ependymoma pathogenesis. Because DNA methylation and histone methylation are closely linked, we hypothesized that further elucidation of epigenomics in PF ependymomas, specifically evaluating histone modifications, would be key to understanding the biology of these tumors.


H3K27me3 is reduced in ~80% of pediatric PF ependymomas

To characterize epigenetic alterations, we first extracted histones from PF and ST ependymomas and performed mass spectrometry and Western blotting analyses. Mass spectrometry (Fig. 1A; P = 0.016) and Western blotting analysis showed that H3K27me3 was substantially reduced (Fig. 1, B and C; P = 0.002) in PF ependymomas compared to ST, SC, and spinal myxopapillary ependymomas. A corresponding increase in H3K27ac (P = 0.005) in tumor samples with reduced H3K27me3 was also noted (Fig. 1, B and C). Other histone modifications, including H3K9me3, H3K4me3, H3K36me2, H3K79me2, H3S28p, H4K20me3, H4K5ac, and H4K16ac, were not significantly different between PF ependymomas and controls (Fig. 1A and fig. S1, A to E).

Fig. 1. H3K27me3 is globally reduced in a subset of pediatric PF ependymomas.

(A) Mass spectrometry of H3K27 and H3K4 modifications in pediatric PF (n = 5) and ST (n = 3; except for H3K27ac, n = 2) ependymomas. *P = 0.016. a.u., arbitrary units. (B) Quantification of H3K27me3 and H3K27ac Western blots normalized to total histone H3 content in PF (n = 6), ST (n = 6), and SC (n = 3) ependymomas. **P = 0.002 for H3K27me3 and **P = 0.005 for H3K27ac. (C) Western blots for H3K27me3, H3K27ac, and total histone H3 quantified in (B). (D to I) Representative H3K27me3 IHC staining in pediatric PF (D to F) ependymomas, H3K27M mutant GBM (G), pediatric ST (H), and adult PF (I) ependymomas. Arrows indicate H3K27me3-positive endothelial and immune cells as internal control. Scale bars, 200 μm. (J) Quantification of H3K27me3 IHC in pediatric and adult ependymomas and of H3K27M and H3WT GBM (for n, see table S7). ***P < 0.0001. (K and L) Representative composite immunofluorescence images for endothelial marker (CD31, red) (K) or leukocyte common antigen (CD45, red) (L), H3K27me3 (green), and 4′,6-diamidino-2-phenylindole (DAPI; blue) in a PF-ve ependymoma. Arrowheads indicate H3K27me3-positive nontumor cells. Scale bars, 20 μm. Data are means ± SEM. Statistical significance in (A), (B), and (J) was determined by two-sided unpaired two-tailed Student’s t test. MP, myxopapillary spinal ependymoma; SE, subependymoma; WT, H3 wild-type glioma.

Because our Western blotting results suggested some degree of heterogeneity in H3K27me3 (reduction in five of six samples), we further characterized H3K27me3 by immunohistochemistry (IHC) in three (two pediatric and one adult) independent ependymoma tumor cohorts (Fig. 1, D to L, and Table 1). We observed a significant reduction (P < 0.0001) in H3K27me3 (PF-ve) in tumor cells in 86% of PF samples in the first pediatric cohort and in 76% of PF samples in the second pediatric cohort (total 82% of pediatric PF ependymomas) (Fig. 1, D to L, and Table 1). H3K27me3 was not reduced in the rest of the pediatric PF ependymomas (PF+ve; 18% of total) or any other subtypes of pediatric and adult (third cohort) ependymomas (Fig. 1, F to J, and Table 1). Reduction in H3K27me3 in PF-ve ependymomas was similar to that seen in histone H3K27M (K27M) mutant gliomas (Fig. 1, D, E, G, and J) (9, 10). Likewise, cells that retained H3K27me3 in PF-ve tumors were mainly endothelial cells and infiltrating immune cells (Fig. 1, D, E, G, K, and L, and fig. S1, F and G, arrows), serving as an internal control for IHC (Fig. 1, D and E, arrows). No significant differences in H3K9me3, H3K36me3, or H3K36me2 immunostaining were observed in these samples (fig. S1, C to E).

Table 1. Demographics of patients providing ependymoma tumor samples.

PF-ve, posterior fossa H3K27me3-ve; PF+ve, posterior fossa H3K27me3+ve; GTR, gross total resection; STR, subtotal resection; NTR, near total resection; NA, not available; CHLA, Children’s Hospital Los Angeles; CHOP, Children’s Hospital of Philadelphia; UCSF, University of California, San Francisco.

View this table:

PF-ve/PF+ve are clinically and biologically similar to PFA/PFB ependymomas

We compared age, gender, and WHO grades between PF-ve and PF+ve ependymomas and included data from previously well-characterized PFA and PFB tumors in parallel (6). In pediatric tumors, both PF-ve (median, 3.7 years) and PFA ependymomas (median, 2.5 years) occurred in younger children in contrast to older age groups seen in PF+ve (median, 8.7 years) and PFB ependymomas (median, 11 years) (Fig. 2A). The gender distributions of PF-ve/PFA/PF+ve/PFB ependymomas were similar (Fig. 2B). As previously reported (6), tumor grade did not show a clear distinction between PF-ve and PF+ve ependymomas (fig. S2A).

Fig. 2. PF-ve tumors share clinical and biological features with PFA ependymomas.

(A and B) Age (A) and gender (B) distribution in ependymomas (6) immunostained with H3K27me3. *P = 0.026 (for n, see table S7; number of cases is indicated when n < 3). ns, not significant. (C) Comparison of gene expression between PF-ve/PF+ve (n = 11/4; y axis) and PFA/PFB (n = 18/19; x axis) (6) ependymomas. FD, fold difference. (D) Principal components (PC) analysis of CpGi methylation in PF-ve (n = 8), PF+ve (n = 4), and ST (n = 4) ependymomas. (E) CpGi methylation per chromosome (bar plot and ideogram) indicating hypermethylated (pink) and hypomethylated (green) regions from PF-ve (n = 8) and PF+ve (n = 4) ependymomas. (F) H3K27me3 IHC from previously defined PFA (n = 13) and PFB (n = 3) pediatric ependymomas. ***P < 0.0001. Arrowheads indicate H3K27me3-positive endothelial and immune cells used as internal control. Scale bars, 200 μm. Data are means ± SEM. Statistical significance was determined in (A) and (F) by two-sided unpaired two-tailed Student’s t test, in (B) by two-tailed χ2 test, and in (C) by two-tailed Pearson’s correlation.

We next assessed gene expression and CpGi methylation in these tumors. Differences in gene expression between PF-ve and PF+ve [RNA sequencing (RNA-seq)] significantly correlated (r = 0.5671, P < 0.0001) with differences in gene expression between PFA and PFB tumors (Fig. 2C and fig. S2B) (6). PF-ve ependymomas showed significantly higher CpGi methylation (P = 3.045 × 10−07) compared to PF+ve tumors (Fig. 2, D and E, and fig. S2, C to F), as seen with PFA versus PFB ependymomas (7). Moreover, the expression of genes that showed similar CpGi promoter differential methylation (7) in PF-ve versus PF+ve and PFA versus PFB ependymomas showed a significant correlation (r = 0.6157, P < 0.0001) (fig. S2D). Finally, H3K27me3 immunostaining was significantly lower (P < 0.0001) in previously defined (8) PFA ependymomas compared to PFB tumors (Fig. 2F). These data together suggest that PF-ve/PF+ve ependymomas share many clinical and biological similarities with PFA/PFB ependymomas (table S1).

PF-ve ependymomas exhibit global DNA hypomethylation correlating to changes in gene expression

Because our DNA methylation profiling was performed using enhanced reduced representation bisulfite sequencing, which provides a comprehensive genomic coverage, we extended our DNA methylation analysis beyond CpGi regions. PF-ve tumors showed significant genome-wide DNA hypomethylation in comparison to PF+ve (P = 1.56 × 10−93) (Fig. 3, A to E, and fig. S3, A to G). To determine how DNA methylation affected gene expression, we identified differentially methylated regions (DMRs) at genomic areas corresponding to gene promoters [5 kb upstream of transcriptional start sites and 5′ untranslated region (5′UTR)], including 213 genes that were differentially expressed in PF-ve tumors (Fig. 3, E and F). Promoter differential methylation showed a significant relationship with gene expression (P < 0.0001, fig. S3H). Hypomethylated DMRs in promoters were noted in 82% of up-regulated genes (121 of 148; Fig. 3F), including genes implicated in PF ependymomas such as AQP1 (11) and ALDH1L1 (6). Therefore, PF-ve ependymomas exhibit CpGi hypermethylation and global DNA hypomethylation correlating with increased gene expression compared to PF+ve ependymomas.

Fig. 3. PF-ve ependymomas exhibit global DNA hypomethylation that drives gene expression.

(A and B) Chromosomal methylation bar plot and ideogram indicating hypermethylated (pink) and hypomethylated (green) regions. (C) Cumulative distribution frequency plots of global DNA methylation in PF-ve, PF+ve, and ST ependymomas (arrows indicate direction of hypomethylation or hypermethylation). ****P = 1.56 × 10−93 for PF-ve versus PF+ve and negligible (reported as 0.00) for PF-ve versus ST. (D) Percent methylation (overall, CpG islands, CpG shores, exons, introns, and promoters) in ependymomas (P values are provided in table S7). Data are means ± SD. (E) DMRs in CpG shores, CpG shelves, 5’UTR, 3’UTR, introns, coding DNA sequences (CDSs), gene bodies, and promoters. (F) Integration of promoter DNA methylation (x axis) and gene expression (y axis) for differentially expressed genes (n = 213), where DMRs at promoters were identified. For all graphs, PF-ve (n = 8), PF+ve (n = 4), and ST (n = 4). For (C) and (D), statistical significance was determined by Mann-Whitney U or Wilcoxon test. Other P values are provided in table S7.

H3K27me3 genomic distribution correlates with DNA methylation

Reduced H3K27me3 in PF-ve ependymomas is not genetically driven, because targeted sequencing on 71 ependymoma samples (cohort 1) did not show recurrent mutations in factors regulating H3K27me3, similar to results from two other independent groups (table S2) (5, 7). Similarly, IHC, protein, and gene expression for factors that regulate H3K27me3 could not explain reduced H3K27me3 in PF-ve tumors (fig. S4, A to E), in agreement with EZH2 immunostaining data from others (12).

Recent studies have demonstrated that unmethylated CpGis recruit the PRC2 complex to increase H3K27me3 (13, 14). We therefore sought to assess the relationship between H3K27me3 and DNA methylation. First, we assessed H3K27me3 genomic distribution using chromatin immunoprecipitation followed by sequencing (ChIP-seq) and compared this with gene expression (RNA-seq). H3K27me3 marked many genetic loci in PF-ve ependymomas despite global reduction in H3K27me3 (Fig. 4, A and B). Consistently, gene expression showed an inverse relationship with the amount of H3K27me3 and was significantly (P < 0.0001) lower in genes that were enriched for H3K27me3 at their promoters (Fig. 4, A to C, and fig. S4, F and G). These genes included SIX1 and BAMBI, which are implicated in neuroglial differentiation (Fig. 4, B and C, and fig. S4F). Moreover, low SIX1 and BAMBI expression was associated with worse prognosis in PF ependymomas (Fig. 4D).

Fig. 4. H3K27me3 distribution inversely correlates with CpGi methylation.

(A) Gene expression (log2 FD; PF-ve/ST, n = 5/3, ***P < 0.0001 and PF-ve/PF+ve, n = 11/4, **P = 0.0022) in H3K27me3 high versus H3K27me3 low regions. (B) H3K27me3 distribution in the SIX1 locus. (C) Comparison of BAMBI and SIX1 gene expression (log2 FD) between PF-ve/ST (n = 5/3), PF-ve/PF+ve (n = 11/4), and PFA/PFB [n = 18/19 (6)] ependymomas. (D) Kaplan-Meier survival curves in PF ependymomas (6) with low (red, n = 16) or high (blue, n = 21) expression of SIX1/BAMBI mRNA. (E) FD (PF-ve/ST) of H3K27me3 enrichment for hypermethylated or hypomethylated CpGi in PF-ve tumors. ***P < 0.0001. (F) DMRT2A locus on chromosome 1 with CpGi hypomethylation and enriched H3K27me3. (G) LHX2 locus on chromosome 9 with CpGi hypermethylation and low H3K27me3. (H) Correlational analyses between H3K27me3 distribution (FD, PF-ve/ST, y axis) and DNA methylation (mean methylation difference, x axis). (I) Functional groupings of all genes that are epigenetically deregulated. Red indicates pathways related directly to neuronal development and cell differentiation (altered promoter DNA methylation and/or marked by promoter H3K27me3) and shows altered gene expression. Data in (A) and (E) are means ± SEM. Statistical significance was determined by two-sided unpaired two-tailed Student’s t test; the log-rank test was used in (D); and two-tailed Pearson’s correlation was used in (H).

We then compared genomic H3K27me3 distribution with DNA methylation. H3K27me3 was significantly more enriched in hypomethylated CpGi (P < 0.0001) and showed an inverse relationship with CpGi methylation (r = −0.4827, P < 0.0001) (Fig. 4, E to H, and fig. S4H). For example, DMRT2A, a gene that regulates neurogenesis (15), showed CpGi hypomethylation (mean methylation difference, −33.3) and was enriched for H3K27me3 (7.4-fold increase) in PF-ve ependymomas (Fig. 4F). In contrast, LHX2, another neurogenesis regulator (16), showed low H3K27me3 (1.43 difference) but CpGi hypermethylation (mean methylation difference, 48.4) (Fig. 4G). Because PF-ve/PFA ependymomas exhibit higher CpGi methylation compared to PF+ve/PFB ependymomas, the inverse relationship between H3K27me3 distribution and CpGi methylation suggests that aberrant DNA methylation may drive low H3K27me3 levels. Genes that showed differential DNA methylation or enrichment for H3K27me3 were involved in various neurodevelopmental functions when subjected to Database for Annotation, Visualization, and Integrated Discovery (DAVID; analysis (Fig. 4I and table S3).

PF-ve/PFA ependymomas share epigenetic similarities with H3K27M gliomas

Similar to ependymomas, pediatric GBMs occur in PF (including diffuse pontine and cerebellar gliomas) and ST (or hemispheric) regions. PF gliomas carrying H3K27M mutations (60 to 80%) are more common in children and, similar to PF-ve ependymomas, show global loss in H3K27me3. In contrast, H3K27me3 is retained in ST gliomas [which bear H3G34R/V (G34) mutations] and in ST ependymomas (with C11orf95-RELA fusions) (5, 17, 18). Moreover, H3K27M gliomas, in addition to exhibiting reduced H3K27me3, are associated with genome-wide DNA hypomethylation (19). Because PF-ve/PFA ependymomas share many clinical (age and PF location) and molecular features with H3K27M gliomas, we compared genome-wide DNA methylation and H3K27me3 distribution in PF-ve/PFA ependymomas and H3K27M gliomas to gain further insights into the disease mechanism. Unsupervised cluster analyses of DNA methylation data derived from previously published data sets (7, 20) demonstrated that PFA ependymomas grouped mainly with H3K27M gliomas (Fig. 5, A and B). Genomic regions with similar methylation patterns (n = 228) in both PFA and H3K27M but not in G34R/V or PFB tumors included many promoter regions (n = 46 of 228) (fig. S5A), including those of two genes important in radial glia biology: PCDH9 (cadherin expressed in radial glia) and S100B (glial calcium-binding protein, marker for radial glia). These two genes exhibited promoter DNA hypomethylation and increased expression in PFA and H3K27M but not PFB or G34 tumors (Fig. 5C and fig. S5B). Further, increased S100B and PCDH9 expression together correlated with poor prognosis in PF ependymomas (Fig. 5D).

Fig. 5. PFA/PF-ve ependymomas show similarities to H3K27M gliomas in DNA methylation and H3K27me3 genome-wide distribution.

(A) Unsupervised clustering analyses of Illumina 450K array DNA methylation data from PFA (n = 14) and PFB (n = 11) ependymomas (7) and H3K27M (n = 18), G34 (n = 18), and IDH1 mutant (n = 19) gliomas (20). All methylation sites were included in the analysis. (B) Heat map of representative hypermethylated and hypomethylated CpGi from (A). (C) Comparison of S100B and PCDH9 gene expression [log2 FD between PFA/PFB (n = 18/19, left y axis) (6), H3K27M/H3WT (n = 12/21, right y axis) (20), and G34/H3WT (n = 10/21, right y axis) (20)]. (D) Kaplan-Meier curves in PF ependymomas with (6) high (n = 25) or low (n = 12) expression of S100B/PCDH9 mRNA. (E) Heat map displaying H3K27me3 peak intensities in PF-ve (n = 3), ST (n = 3), and NPCs expressing H3.3K27M (K27M) or H3.3 WT (22). Unguided hierarchical clustering was used for clustering samples based on H3K27me3 enrichment. **P = 0.0075. (F) H3K27me3 distribution in the EMX2 locus in PF-ve and ST ependymomas and H3K27M-expressing and H3WT NPCs. (G) Numbers of specific H3K27me3 peaks common to PFA (n = 6) or PFB (n = 4) ependymomas (7) and H3K27M or H3WT NPCs (22). ***P < 0.0001. (H) Comparison of EMX2 and HOXA7 gene expressions (log2 FD) between PFA/PFB [brown, left y axis, n = 18/19 (6)], H3K27M/H3WT [orange (20), right y axis, n = 12/21], and G34/H3WT [green, right y axis (20), n = 10/21]. (I) Kaplan-Meier survival curves in PF ependymomas (6) with high (n = 27) or low (n = 10) expression of EMX2/HOXA7 mRNA. (J) Overlap of differentially DNA- or H3K27-methylated genes in PFA and H3K27M tumors with those in human radial glial gene expression data sets (23). Statistical significance in (D) and (I) was determined by the log-rank test. In (E) and (G), the significance of differences in peak numbers was evaluated by using two-tailed χ2 test.

Despite genome-wide H3K27me3 decreases in H3K27M gliomas, certain genetic loci, including genes implicated in neuroglia lineage specification, display increased H3K27me3 (19, 21). We therefore compared our H3K27me3 ChIP-seq from PF-ve and ST ependymomas with two independently published H3K27me3 ChIP-seq data sets generated from H3.3 WT (H3WT) or H3K27M-expressing neural progenitor cells (NPCs) (22) or H3K27M mutant and H3WT glioma tumor samples (19). Unsupervised clustering of H3K27me3 ChIP-seq peaks showed that PF-ve tumors clustered with H3K27M NPCs but not with ST ependymomas or H3WT NPCs (P = 0.0075) (Fig. 5, E and F). Similarly, compared to ST ependymomas, PF-ve tumors had significantly more H3K27me3 peaks (P = 0.0041) that were also found in H3K27M gliomas but not in H3WT gliomas (fig. S5E). We confirmed these results in previously published H3K27me3 ChIP-seq data from PFA and PFB ependymomas (7), where common peaks between PFA and H3K27M were significantly higher than those between PFB and H3K27M or H3WT NPCs (P < 0.0001) (Fig. 5G). H3K27me3 peaks (n = 255), which are common to PFA and H3K27M NPCs, corresponded largely to promoters (248 of 255, 97%) (fig. S5C). Among them, genes that were down-regulated in PFA and H3K27M have been implicated in radial glial development (EMX2) and neuronal patterning (HOXA7), which together correlated with poor prognosis (Fig. 5, F, H, and I, and fig. S5, D and E). To assess the functional relevance of genes that are epigenetically deregulated in both PF-ve and H3K27M tumors in an unbiased manner, we combined genes whose promoters exhibit differential DNA methylation (n = 46; fig. S5, A and B) or H3K27me3 enrichment (n = 248; fig. S5, C and D) and found that they showed ~88% (n = 258 of 294) overlap with published human radial glial signatures (Fig. 5J and table S4) (23). Functional analyses revealed that the top enriched categories were involved in neurodevelopment (fig. S5G and table S5). Together, PFA/PF-ve ependymomas shared substantial similarities in DNA methylation and H3K27me3 genomic distribution with H3K27M glioma, corresponding to genes involved in radial glial development and function.

Radial glia in the developing cerebellum show reduced H3K27me3

Our findings that factors involved in radial glial biology were epigenetically deregulated in PF-ve ependymomas prompted us to assess H3K27me3 in PF radial glia during development. Ependymomas are thought to arise from radial glia and express glial fibrillary acidic protein (GFAP) and S100B (7, 11, 24, 25). Radial glia have dual functions, serving as central nervous system progenitor stem cells and mediating neuronal migration (26). PF ependymomas mainly occur in the fourth ventricular region (fig. S6A), where the cerebellum forms the roof and the brainstem forms the floor (fig. S6B). We therefore examined H3K27me3 in GFAP-positive radial glia in this region in nonpathologic human developing PF brain tissues (Fig. 6, fig. S6, and table S6).

Fig. 6. Radial glia in early cerebellar development show reduced H3K27me3.

(A and B) Representative confocal immunofluorescence images of human cerebellar tissues stained with antibodies against GFAP (green), H3K27me3 (red), and DAPI (blue) when the EGL is present before birth (A) or when EGL migration is complete (B). Arrows indicate H3K27me3-negative (A) or H3K27me3-positive (B) nuclei in GFAP-positive radial glia. Arrowheads indicate radial glial processes. Dotted circles indicate granular cells. Solid lines indicate the pial surface, and the dotted lines indicate the extent of the EGL. Scale bars, 20 μm. (C and D) Higher-power images of radial glia. Large arrowheads indicate H3K27me3-negative (C) or H3K27me3-positive (D) nuclei in GFAP-positive radial glia. Small arrowheads indicate radial glial processes. Dotted circles indicate granular cells. Scale bars, 8 μm. (E and F) Representative confocal immunofluorescence images from the human fourth periventricular region. Scale bars, 20 μm. Arrows indicate H3K27me3-positive nuclei in GFAP-positive radial glia. (G) Percentage of GFAP-positive/H3K27me3-negative or GFAP/H3K27me3-positive radial glia before and after development (n = 3 each). ***P < 0.0001. (H) Amount of H3K27me3 in each cell (H3K27me3 pixel units) quantified using ImageJ (National Institutes of Health) from various cell types in the PF before and after development. ***P < 0.0001; for n, see table S7. (I) Expression (log2 FD) of factors that differentiate PFA from PFB ependymomas in PFA/PFB [n = 18/19 (6)] and radial glia from P6/P30 mice [n = 5 each (29)]. ND, not detected. Data in (G) and (H) are means ± SEM; statistical significance was determined in (G) using two-tailed χ2 test and two-sided unpaired two-tailed Student’s t test in (H). RG, radial glia; ML, molecular layer; PC, Purkinje cell; IGL, internal granular layer; GL, granular layer; 4V, fourth ventricle.

In the developing PF, cerebellar radial glia mediate external granular layer (EGL) neuronal precursor migration to their destination in the granular layer, completed by 2 to 3 months after birth (27). GFAP-positive cerebellar radial glia in the fourth ventricle roof exhibited reduced H3K27me3 when the EGL was prominent prenatally but gained H3K27me3 when EGL migration was complete postnatally (P < 0.0001) (Fig. 6, A to D, G, and H, and fig. S6, C to F). In contrast, ependymal cells lining the fourth ventricle floor showed comparatively higher H3K27me3 before and after development (Fig. 6, E to H, and fig. S6, G and H). Human and mouse radial glia substantially differ in their gene expression patterns (28). Despite this, gene signatures that distinguish PFA from PFB ependymomas (6) showed similar trends in data derived from mouse developing (P6/P30) PF radial glia (Pearson’s r = 0.4610, P = 0.0354) but showed the opposite trend in developing mouse EGL neurons (granular layer precursors/mature granular neurons; Pearson’s r = −0.5462, P = 0.0433) (Fig. 6I and fig. S6, I and J) (29). Reduced H3K27me3 in PF-ve ependymomas and PF radial glia in early development is consistent with data that suggest radial glia as ependymoma candidate cells of origin (11, 25).

H3K27me3 global reduction delineates poorly prognostic PF ependymomas

We compared radiologic findings and overall survival in PF-ve and PF+ve ependymomas. PF-ve tumors were radiologically and histologically more invasive than PF+ve tumors (P < 0.0001) (Fig. 7, A to D). Furthermore, PF-ve tumors were associated with overall higher rates of incomplete resection (subtotal and near total resection; Fig. 7E), consistent with their invasive nature. PF-ve tumors also exhibited significantly worse prognosis compared to PF+ve ependymomas in the pediatric cohort (P = 0.026, Fig. 7F) and in a larger sample size combining all three cohorts including adult ependymomas (P = 0.0068, Fig. 7G). We next determined whether H3K27me3 immunostaining could be a clinically important prognostic biomarker in PF ependymomas. Cox proportional hazards analysis confirmed H3K27me3 reduction, but not age, sex, or tumor grade, as a significant prognostic indicator (P = 0.027, Fig. 7H; extent of tumor resection was not included in this analysis because this information was available only in the first pediatric tumor cohort, Table 1; recurrence was not associated with poor prognosis possibly because of the relatively small sample size). In contrast, H3K9me3, H3K36me2, and H3K36me3 immunostaining did not show any relationship with overall survival (fig. S7, A to C). Finally, the extent of H3K27me3 was reduced in matched primary and recurrent PF-ve when compared to that observed in matched primary and recurrent ST ependymomas (Fig. 7I). These data together suggest that reduced H3K27me3 by IHC staining is an independent prognostic indicator in pediatric PF ependymomas.

Fig. 7. Decreased H3K27me3 is a poor prognostic biomarker in pediatric PF ependymomas.

(A and B) Representative axial and coronal T2-weighted magnetic resonance imaging (MRI) images from PF-ve (A) and PF+ve (B) ependymomas (white arrows). (C) Percentage of cases with or without invasion in PF-ve (n = 20) versus PF+ve (n = 4). ***P < 0.0001. (D) Tumor invasion and infiltration into surrounding brain tissue were compared between PF+ve (n = 4) and PF-ve (n = 28) by detecting neurofilaments in axons within tumor tissues (arrowheads) using IHC (representative images from PF-ve and PF+ve tumors and quantification). Scale bars, 60 μm. ***P = 0.0017. (E) Extent of resection (GTR, gross total resection/no detectable tumor on postsurgical MRI; STR, subtotal resection/portion of tumor is detectable on postsurgical MRI; NTR, near total resection/detection of minimal tumor on postsurgical MRI) in pediatric PF-ve (n = 30) and PF+ve (n = 4) ependymomas. **P = 0.0047. (F and G) Kaplan-Meier survival curves from pediatric (F) or combined pediatric and adult (G) cohorts of PF-ve [red; n = 50 in both (F) and (G)] and PF+ve [blue; n = 11 in (F) and n = 15 in (G)]. (H) Cox proportional regression analyses for overall survival in PF-ve (n = 50) and PF+ve (n = 11) ependymomas. Exp(B) is the measure of the hazard ratio. CI, confidence interval. (I) H3K27me3 in matched primary and recurrent PF-ve and ST ependymomas (n = 3 for each). Representative images and quantification are shown. Scale bars, 200 μm. Data are means ± SEM in (D) and (I). Statistical significance in (C) and (E) was determined by using two-tailed χ2 test, in (D) and (I) by unpaired two-sided two-tailed t test, and in (F) and (G) by the log-rank test.


We examined histone modifications to better understand the epigenome of childhood PF ependymomas. Our mass spectrometry, Western blotting, and IHC analyses demonstrate that ~80% of PF ependymomas exhibit reduced H3K27me3. Although the small sample size is a limitation of our study and our findings will need to be confirmed in a larger tumor cohort to assess the relationship between H3K27me3 and PFA/PFB ependymomas, PF-ve ependymomas occurred in younger children, were more radiologically invasive, and defined poorly prognostic tumors sharing many similarities with previously described PFA ependymomas based on gene expression, CIMP status, and overall prognosis (6, 7). Our data, along with previous studies, suggest that the molecular and epigenetic childhood PF ependymoma stratification may be prognostic (68).

The 2016 WHO Classification of Tumors of the Central Nervous System update includes molecular characteristics of ST ependymomas (RELA fusion), underscoring the importance of integrating molecular and histopathologic criteria (1). In this context, H3K27me3 IHC in PF ependymomas is easy to obtain and cost-effective across many clinical settings and surgical neuropathology practices, where methylation and gene expression arrays may not be readily available. IHC staining for H3K27me3 can therefore be used as a readily translatable prognostic molecular surrogate on routine formalin-fixed sections in pediatric PF ependymomas.

H3K27M mutations in PF gliomas dominantly inhibit EZH2 function, resulting in H3K27me3 global reduction (10). Two independent groups have shown that pediatric PF ependymomas do not bear recurrent genetic mutations unlike H3K27M gliomas (5, 7). Consistent with these data, we found no recurrent mutations in factors that influence H3K27me3 in PF-ve ependymomas. Recent studies have documented a role for unmethylated CpGis in PRC2 recruitment. Genome-wide analyses have revealed a strong positive correlation between methylation-free CpGi and PRC2 binding, and in vitro studies demonstrate that the recruitment and PRC2 activity depend on DNA methylation (13, 14, 30). In our analyses, PF ependymomas showed an inverse relationship with CpGi methylation and H3K27me3 enrichment. Thus, the lowered H3K27me3 in these tumors may result from the inability of PRC2 to access chromatin because of increased CpGi methylation. However, the mechanisms that regulate aberrant DNA methylation in PF ependymomas are not known and require further elucidation. We also found that human PF radial glia displayed low H3K27me3 during early development. Therefore, reduced H3K27me3 in PF-ve ependymomas may also reflect the developmental window during which tumor formation occurs.

We found unexpected similarities in DNA methylation patterns and H3K27me3 genome-wide distribution between PF-ve/PFA ependymomas and H3K27M gliomas, which correlated with a characteristic expression of genes critical for radial glial function and development. Because PF-ve/PFA ependymomas and H3K27M gliomas share many clinical features, including patient age and location, our observations suggest that a chromatin state defined by DNA and H3K27 hypomethylation may play a critical role in transformation and/or growth of tumor cells originating within the developing PF. Elucidating regulators of this shared chromatin state may be critical in understanding the biology of PF tumors and taking advantage of this potential therapeutic target. Our data suggest that H3K27me3 reduction may be of substantial importance in identifying ependymoma patients who would benefit from epigenetic therapies.


Study design

The objective of this study was to assess epigenetic modifications in childhood PF ependymomas. We observed a reduction in H3K27me3 in PF ependymomas, which led us to hypothesize that reduced H3K27me3 is central to PF ependymoma pathogenesis. Because ependymomas are rare tumors, no a priori power analysis was conducted to determine sample size. Sample size was determined on the basis of frozen tissue availability and ranged from n = 3 to 11 (biological replicates) for methylation analyses, Western blotting, and ChIP-seq studies. No samples were excluded from data analyses. Tumor samples were obtained from CHLA, CHOP, UCSF, and The Hospital for Sick Children, University of Toronto, after institutional review board approval. All studies were conducted in compliance with REMARK (Reporting Recommendations for Tumor Marker Prognostic Studies) guidelines for prognostic studies. All identifiers from cases were removed before analysis. Three independent neuropathologists reviewed hematoxylin and eosin sections in a nonblinded manner and assessed H3K27me3 immunostaining in a blinded manner. Image capturing and analyses were performed in a blinded manner.


Statistical analyses were performed in consultation with the Department of Computational Medicine and Bioinformatics (University of Michigan). Graphs were plotted, and statistical analyses were performed using Prism software (version 6, GraphPad). Unpaired two-tailed two-sided Student’s t test, two-sided χ2 test, two-sided analysis of variance (ANOVA) followed by post hoc Bonferroni multiple comparison analysis, and two-sample Mann-Whitney U/Wilcoxon test were used to analyze data as indicated. Z scores were used to segregate high and low expression cutoffs for each marker based on methodology from The Cancer Genome Atlas. Overall survival data were plotted as Kaplan-Meier curves and analyzed using Mantel-Cox log-rank test. Multivariate survival analysis was performed using Cox proportional hazards model in SPSS (version 22, IBM). The hazard ratio represented by the SPSS software as Exp(B) is defined as the ratio of death probabilities in PF-ve versus PF+ve ependymomas. Data were considered significant if P values were below 0.05 (95% confidence intervals); n and P values are reported in figure legends and table S7.


Materials and Methods

Fig. S1. Other histone modifications in pediatric PF ependymomas are not different.

Fig. S2. PF-ve tumors share similarities with PFA ependymomas.

Fig. S3. PF-ve tumors show lower DNA methylation compared to PF+ve ependymomas.

Fig. S4. CpGi methylation inversely correlates with H3K27me3 distribution.

Fig. S5. PFA/PF-ve ependymomas share epigenetic similarities with H3K27M gliomas.

Fig. S6. Radial glia in early cerebellar development show reduced H3K27me3.

Fig. S7. H3K9me3, H3K36me3, and H3K36me2 do not correlate with overall prognosis.

Table S1. Comparison of PF-ve/PFA and PF+ve/PFB ependymomas.

Table S2. Focused sequencing panel.

Table S3. DAVID analysis of differentially methylated and regulated genes in PF-ve ependymomas.

Table S4. Overlap of genes that exhibited common epigenetic alterations between H3K27M gliomas and PFA ependymomas with human radial glial signatures.

Table S5. DAVID analysis of genes that exhibited common epigenetic alterations between H3K27M gliomas and PFA ependymomas.

Table S6. Characteristics of human postmortem brain samples.

Table S7. Data values in tabular format.


  1. Acknowledgments: We thank D. Yarilin, J. Zhang, A. Viale, N. Socci [Memorial Sloan Kettering Cancer Center (MSKCC)], C. Edwards, C. Paran (University of Michigan), A. Alonso (Weill Medical College of Cornell University), and K. Miller (CHLA) for experimental and data analysis help. We thank A. Heguy and I. Dolgalev (New York University) and the New York Genome Center for help with sequencing studies. We thank T. Lindsten and D. Pozega for critical reading of the manuscript. Funding: This work was supported by grants from the National Cancer Institute (K08 CA181475 to S.V.), Matthew Larson Foundation (to S.V.), Sidney Kimmel Foundation (to S.V.), Doris Duke Foundation (to S.V.), and RO1GM110174 and P01CA196539 (to B.A.G.). R.C.M. is supported by the University of Michigan Bioinformatics Core. C.L. is the Kandarian Family Fellow supported by the Damon Runyon Cancer Research Foundation (DRG-2195-14). C.B.T. is supported by the Cancer Center Support Grant (MSKCC) (NIH P30 CA008748). Author contributions: J.B., P.M., C.L., S.U.J., C.C., G.L.R., D.M., A.P., and S.V. performed experiments and analyzed data; A.S.M., M.P., and M.S. provided clinical data; C.C. and J.B. captured images in a blinded manner; R.C.M., P.M., B.S., P.P., and M.C. performed bioinformatics; B.A.G. performed histone mass spectrometry; B.T. analyzed MRI scans in a blinded manner; S.V., A.R.J., M.S., C.H., and M.P. are neuropathologists who provided and assessed tumor samples; P.W.L., A.M., C.D.A., C.B.T., A.M.C., A.R.J., and S.V. provided critical insights; J.B., C.L., A.R.J., and S.V. wrote and edited the paper; all authors provided edits and comments. Competing interests: C.B.T. is cofounder of Agios Pharmaceuticals and has financial interest in Agios. C.B.T. is also on the board of directors of Merck and CRL. The other authors declare that they have no competing interests. Data and materials availability: The data for this study have been deposited in the National Center for Biotechnology Information database as GSE87779 and GSE89452.
View Abstract

Navigate This Article