Inactivation of Ezh2 upregulates Gfi1 and drives aggressive Myc-driven Group 3 medulloblastoma

Deletion of Ezh2 accelerates de novo MB development and progression of established MBs

Exons 15–20 of Ezh2 encode the SET domain required for catalytic activity (Figure 1A). We used all-in-one CRISPR-Cas9 plasmids encoding green fluorescent protein (GFP) and Cas9, with or without single-guide RNAs (sgRNAs), to simultaneously target both Ezh2 exons 10 and 18 in primary Trp53^−/−;Cdkn2c^−/− GNPs grown as neurospheres. DNA sequencing of GFP-positive, CRISPR-targeted, subclones revealed that virtually all of them contained Ezh2 mutations, which included indels in exons 10 and 18 as well as deletion of intervening sequences (see below).

Pooled GFP-positive neurospheres isolated by fluorescence activated cell sorting (FACS) were infected with retroviral vectors expressing red fluorescent protein (RFP) with or without Myc. Immunoblotting of proteins extracted from CRISPR-Cas9-edited populations demonstrated absence of detectable Ezh2 with decreased global H3K27me3 and increased H3K27 acetylation (ac) marks but unaltered H3K4me3 levels (Figure 1B). Expression of Eed was unchanged, whereas the level of Suz12, another core component of the PRC2 complex, was decreased in Ezh2-deficient cells (Figure 1B). Myc⁺;Trp53^−/−;Cdkn2c^−/− GNPs, whether expressing Ezh2 or not, formed large cell anaplastic MBs (Figure S1A) when injected into the cortices of nude mice, but Ezh2-deficient tumors arose more rapidly (Figure 1C), indicating that Ezh2 restrains MB formation. Importantly, Myc was required for MB formation regardless of Ezh2 status.

We similarly used CRISPR-Cas9 to delete Ezh2 from four individually-derived Myc- induced primary G3 tumors (propagated as MB1-MB4 “tumorsphere” lines). Each MB cell line contained multiple unique Myc retroviral integration sites, arguing against a requirement for specific insertional mutagenesis on tumor development and implying that the primary tumors were polyclonal (Figure S1B). Using established Myc⁺ G3 MB lines in lieu of primary Trp53^−/−; Cdkn2c^−/− GNPs, we again simultaneously targeted exons 10 and 18 with CRISPR-Cas9; 94% of single cell-derived subclones demonstrated Ezh2 indels and/or deletions of intervening DNA between exons 10 and 18. Pooled tumorspheres significantly downregulated full-length Ezh2 protein expression (Figure 1D) with concomitant decreases of global H3K27me3 (Figures 1D, S1C and S1D) and increases in H3K27ac marks compared to those in sham-targeted Ezh2^+/+ MBs that received CRISPR-Cas9 but no sgRNAs (Figure 1D). Equal numbers of cultured primary MB cells, expressing Ezh2 or not, were implanted into the cortices of secondary 6–8 week old recipient mice, where G3 MBs arise equally well outside the cerebellum and maintain their key genotypic and phenotypic characteristics (Kawauchi et al., 2012). Ezh2-deficient MB cells induced tumors significantly faster than those expressing control levels of Ezh2 despite variations in tumor latencies between different MB lines (Figure 1E).

DNA deep sequencing of sham-targeted primary MB1 and MB3 tumors confirmed that more than 99.6% of amplified Ezh2 exon 10 and 18 sequences were wild-type. In contrast, CRISPR-targeted primary tumorsphere populations contained numerous clonal variants each represented at a different frequency, confirming the efficiency of gene editing and tumor polyclonality (Table S1). The observed genetic variations likely account for the heritable but variable tumor latencies of the four independently derived MB lines. After orthotopic transplantation, secondary tumors were harvested from the cortices of recipient mice, triturated, and purified tumor cells were analyzed without intervening culture (Kawauchi et al., 2012). Three randomly selected secondary tumors derived from primary MB1 and MB3 cells also contained numerous, only partially overlapping clonal variants, many of which had not been detected in the primary tumor from which they arose and consistent with continued clonal evolution (Table S1).

Secondary tumors lacked detectable Ezh2 (Figure 1D) and exhibited global decreases in H3K27me3 but no significant changes in H3K4me3 at regions within 3 kb upstream and downstream of the promoter transcriptional start sites (TSS) (Figures S1E and S1F). RNA-Seq and GSEA of pooled Ezh2⁺ and Ezh2⁻ tumors (3 each from MB1, MB2, and MB3) highlighted significant concordance between genes upregulated in Ezh2-deficient tumors and those having high CpG density and promoter H3K27me3 marks in embryonic day-14.5 mouse whole brains or neural precursor cells (Figures S1G and S1H) (Meissner et al., 2008). Among the most commonly affected genes resulting from Ezh2 inactivation in MBs were those affecting tissue development and homeobox (Hox) genes (Figure S1I, and see below).

As an independent approach, TALENs targeting a sequence within Ezh2 exon 14 were used to induce SET domain frame shift mutations; selected single cell-derived MB1 subclones were evaluated one by one for tumor development. Six of 58 clones completely lacked Ezh2 protein expression. Of these, TALEN-targeted single cell clones 5, 23, and 47, each with frame shift mutations of 5, 2, and 11 base pairs, respectively, in one allele and a large deletion of the other, as well as control clones derived from sham-targeted tumorspheres (no sgRNA), were selected for further evaluation of protein expression (Figure 1F) and tumorigenicity (Figure 1G). Each of the three Ezh2-null clones generated secondary MBs faster than control clones expressing Ezh2, although single Ezh2⁺ clones were less aggressive than the polyclonal MB1 tumor from which they were derived (compare with Figure 1E, top left panel).

Injection of MB1 cells engineered to overexpress an Ezh2 mutant lacking the catalytic SET domain (Ezh2ASET) (Varambally et al., 2002) mimicked the effects of Ezh2 deletion on global histone marks (Figure 2A) and accelerated secondary tumor formation (Figure 2B). These findings suggested that Ezh2ASET acts as a dominant-negative mutant that quenches the catalytic activity of wild type Ezh2. In contrast, two Ezh2 gain-of-function mutants (Y637F and Y637S) that increased H3K27me3 levels (Figure 2C) and contribute to oncogenicity in other tumor types (Sneeringer et al., 2010) slowed secondary MB1 tumor formation (Figure 2D). Furthermore, disruption of Suz12, which encodes a critical component required for assembly of the PRC2 complex (Pasini et al., 2004), accelerated MB tumorigenicity as well (Figures 2E, 2F, and S2A–S2C). Finally, when we re-introduced Ezh2-IRES-GFP versus GFP alone into Ezh2⁺ or Ezh2⁻ MB1 and MB3 cells (Figure 2G), accelerated tumor formation accompanying Ezh2 inactivation was significantly reversed (Figures 2H and 2I). Retardation of tumor development after Ezh2 restoration argues that any elicited off-target mutations were not responsible for the consistently observed phenotypic changes following Ezh2 loss. Taken together, inactivation of Ezh2/EZH2 and disruption of PRC2 histone methylase function accelerates G3 MB formation and tumor progression.

Ezh2 inactivation in G3 MB implicates Gfi1 as an oncogenic target gene

Because H3K27me3 is a repressive promoter mark and depletion of Ezh2 accelerated tumor initiation and progression, we reasoned that genes driving proliferation would be up-regulated in G3 MBs lacking Ezh2. Three Ezh2-deficient and three control secondary tumors derived from each of three MB lines (18 total samples) were subjected to comparative RNA-Seq analysis. Heat maps of the top 1,000 differentially expressed genes (not shown) confirmed that all tumors remained more closely related to the G3 MB subgroup than to mouse MBs of the SHH or WNT subgroups or to primary GNPs (Kawauchi et al., 2012). The top nine most significantly upregulated genes in Ezh2-deficient secondary tumors compared to Ezh2-positive ones included at least six Hox genes (also see Figure S1I) known to be regulated by PRC2 (Bracken et al., 2006; Levine et al., 2004; Lund and van Lohuizen, 2004), as well as Igf2bp2, Gfi1 and Erbb2 (Fig. 3A and S3A). HOX protein levels were increased in MB1-MB3 Ezh2-deficient secondary tumor cells (Figure 3B). Gfi1 and Erbb2 were previously revealed to have oncogenic activity in MB (Gilbertson, 2005; Northcott et al., 2014), whereas Igf2bp2 has been implicated as an oncogene in other systems (Baxter, 2014; Bell et al., 2013). With the exception of Erbb2, these genes were downregulated when Ezh2 expression was again reinforced (Figures S3B and S3C). Myc (with Cdh1 used as an additional control) was unaffected. Enforced retroviral vector-mediated overexpression of GFP-tagged Gfi1 and Erbb2, but not Igf2bp2 or Hoxb4, in MB1 and MB2 tumorspheres accelerated MB tumor formation after re-transplantation of sorted GFP-positive cells into the brains of mice (Figures 3C–3E).

To investigate whether these same genes sustained Ezh2-dependent modifications of histone H3, we performed chromatin immunoprecipitation followed by nucleotide sequencing (ChIP-Seq) with antibodies to H3K27me3, H3K4me3 and H3K27ac in secondary MB3 tumors expressing Ezh2 or not (Figure 3F). Although inactivation of Ezh2 led to global changes in H3K27me3 and H3K27ac in Myc⁺;Trp53^−/−;Cdkn2c^−/− GNPs (Figure 1B), in primary MB tumors (Figure 1D), and secondary tumors derived from them (Figures S1E–S1F), the patterns of such modifications across specifically targeted genes were variable. Reciprocal Ezh2-dependent changes in H3K27me3 and H3K27ac were evident across five Hox genes (Figure 3F). Although no global changes in H3K4me3 were observed in these tumors (e.g. Figure S1F), H3K4me3 occupancy increased over all Hox genes studied consistent with their robust activation in response to Ezh2 inactivation (Figure 3F). In contrast, although Erbb2 was upregulated when Ezh2 was deleted, patterns of histone H3 modifications across this gene were unaffected (Figure 3F) and Erbb2 expression was not downregulated when Ezh2 was restored (Figures S3B and S3C). ChIP-Seq analysis of Gfi1 revealed a distinct pattern in which reduction of H3K27me3 in response to Ezh2 loss predominated (Figure 3F). Together, these data suggested that several Hox genes and Gfi1 were direct transcriptionally repressed targets of Ezh2, whereas Erbb2 might be indirectly regulated in response to Ezh2 downregulation.

Functional interactions between Ezh2 and Gfi1

Given that Gfi1, a documented G3 MB oncogenic driver, was up-regulated in response to Ezh2 inactivation (Figure 2) and was a presumptive target of Ezh2 regulation (Figures 3A and 3F), we explored the effects of Gfi1 inactivation on MB tumor progression. To this end, we co-infected three all-in-one CRISPR-Cas9 vectors containing guide RNAs directed to Gfi1 exons 4 and 5 and co-expressing GFP to inactivate the zinc finger DNA binding domain of Gfi1 in highly aggressive Ezh2-negative MB3 cells (Figure 4A). Following retrieval of infected GFP-positive cells by FACS and expansion of the edited cultured tumorspheres, deep sequencing analysis was performed to determine the frequencies of Gfi1 mutations in the edited populations, which manifested frequent deletions within exon 4 and indels affecting exon 5 (Figures S4A–S4C). Simultaneous targeting of these two exons generated a mixture of cells of which some retained Gfi1 expression; immunoblotting confirmed increased Gfi1 levels in Ezh2⁻ MB3 cells versus Ezh2⁺ controls and their partial reduction after CRISPR-Cas9-mediated editing (Figure 4B). Introduction of these Ezh2⁻ and partially Gfi1-deficient MB3 cells into the brains of nude CD1 mice demonstrated that Gfi1 inactivation slowed tumor progression (Figure 4C). When orthotopically transplanted into naive mice, edited individual Gfi1⁻ MB1 subclones 1 and 14 (Figure S4D) similarly exhibited retarded MB tumorigenicity as compared to Gfi1⁺ control clones (Figure 4D).

By contrast, when vectors expressing Gfil-GFP or GFP alone were used to infect Myc⁺;Trp53^−/−;Cdkn2c^−/− MB1 and MB2 tumorspheres, Ezh2 expression was unaffected by Gfi1 while Myc levels were modestly increased (Figure 4E). Enforced Gfil expression in MB1 tumorspheres increased their proliferation rate (Figure S4E). RNA-Seq analysis of six control (GFP only) and seven secondary tumors overexpressing Gfi1-GFP whose onset was accelerated (Figure 4D) showed that more than 2,000 transcripts were differentially regulated more than 2-fold (FDR <0.05) (Figure 4F, 4G). ChIP-Seq analysis revealed that the endogenous Gfil protein binds directly to the Ezh2, Suz12, and Eed promoters in primary non-tumorigenic neurospheres generated from GNPs purified from the cerebellum of 7 day-old Trp53^−/−;Cdkn2c^−/− pups (Figure S4F). Whereas enforced Gfi1 overexpression had little effect on expression of these same three PRC2 genes in MB tumor cells, it negatively regulated expression of PRC2 target genes consistent with its known role as a transcriptional repressor (Duan et al., 2005; Grimes et al., 1996; McGhee et al., 2003; Saleque et al., 2007; Vassen et al., 2006; Zweidler-Mckay et al., 1996). In particular, Hox genes as well as Erbb2 and Igf2bp2 (Figures 4F and S4G) were downregulated in direct contrast to their upregulation in response to Ezh2 inactivation (Figures 3 and S3). Gfi1 can directly repress the homeobox gene Meis1 to regulate expression of a transcription complex consisting of Meis1, Hoxa9 and Pbx1 (Chowdhury et al., 2013; Horman et al., 2009), and, in agreement, all three of these genes were significantly downregulated in response to Gfi1 transduction (Figure S4G). Through GSEA analysis, the global gene expression pattern observed after Gfi1 transduction was consistent with downregulation of Suz12 target genes (Figure 4G) (Pasini et al., 2004). Hence, not only does inactivation of Ezh2 decrease Gfi1 H3K27me3 modification (Figure 3F) and increase Gfi1 expression (Figures 2, 4C, and S3C), but Ezh2 and Gfi1 repress a common set of PRC2 target genes (Figures 4F, 4G, S4G, and S4H). Yet, these two genes play contrasting roles in respectively inhibiting and accelerating Myc-induced G3 MB, implying that Gfi1 has additional relevant targets.

Strikingly, although induction of G3 MBs from primary mouse GNPs depends upon both Myc overexpression and Trp53 loss of function whether Ezh2 is defective or not, overexpression of both Myc and Gfi1 was sufficient to induce large cell anaplastic MBs in primary GNPs that retained functional Trp53 (Figures 4H and 4I). Similar findings have been made using different neural stem and progenitor cells isolated from the neonatal cerebellum (Northcott et al., 2014). Indeed, a subset of human G3 MBs overexpress both MYC and GFI1 but retain wild type TP53, implying that GFI1 may act to subvert p53-dependent tumor suppression. RNA-Seq analysis revealed that expression of several hundred genes was differentially regulated in Myc⁺ MBs generated from 18 primary Trp53^−/− GNPs versus MBs arising from four Trp53^+/+ GNPs overexpressing Gfi1 (Figure 4J). However, there were no significant differences in expression of canonical p53-responsive genes (including Cdkn1a (p21), Cdkn2a (Ink4a/Arf), Mdm2, or Bcl2 family members) between these cohorts (see data accession numbers in Experimental Procedures), providing no mechanistic data directly linking Gfi1 with the p53 pathway. In summary, Ezh2 and Gfi1 play both common and distinctly different cross-regulatory roles in MB tumor development with Ezh2 loss and Gfi1 overexpression contributing to enhanced tumorigenesis.

Animal husbandry

CD-1 nu/nu mice (Charles River Laboratories) were used as recipients for cortical transplants. All animal experiments were approved by, and conducted in accordance with, St. Jude Children’s Research Hospital Animal Care and Use Committee guidelines, as required by the United States Animal Welfare Act and the National Institutes of Health’s policy to ensure proper care and use of laboratory animals for research.

Vectors, PCR analyses, retroviral integration site determination, and immunostaining

See Supplemental Experimental Procedures.

Cell culture, transfection, and protein analysis

GNPs purified from cerebella of 7 day-old Trp53^−/−;Cdkn2c^−/− mice and mouse G3 MB tumorspheres (MB1, MB2, MB3, and MB4 previously designated #19251, #19554, #19568, and #9730 (Vo et al., 2016) were cultured on ultra-low attachment dishes. Human recombinant basic FGF and EGF (PeproTech) were added to the culture medium every 2 days. For GNP and G3 MB transfection, 1 × 10⁶ cells were transfected with 2.5 μg of each TALEN pair or CRISPR-Cas9 plasmids using Lipofectamine 2000 (Invitrogen) or Xfect Transfection Reagent (Clontech). Positive RFP⁺/GFP⁺ cells were expanded in bulk or sorted as single cells into 96-well plates and recovered as subclones. For total protein and histone extractions, cell pellets were lysed in RIPA and hypotonic lysis buffers, respectively, (Morfouace et al., 2014), and proteins separated on denaturing gels were immunoblotted (Ayrault et al., 2010) with primary antibodies listed in Table S2.

Retrovirus infection and orthotopic transplantation

G3 MB tumorsphere cells were infected with retroviral vectors for 2 days, and FACS was used to isolate the RFP⁺/GFP⁺ fraction. Cells were expanded in culture, harvested and suspended in Matrigel (BD Bioscience) before 5 × 10⁵ cells were transplanted into the cerebral cortices of CD-1 nu/nu recipient mice as described (Ayrault et al., 2010; Kawauchi et al., 2012). Animals were examined daily for symptoms of sickness (doming of the head, ataxia, and reduced activity). Moribund mice were humanely sacrificed, tumor cells were purified, and extracted RNAs and proteins were analyzed (Kawauchi et al., 2012). Tumors, brains and spines of mice were reviewed for histopathology to evaluate whether the tumors were G3 large cell/anaplastic and metastatic.

DNA Deep-Sequencing and Ezh2 mutation profiles

Endogenous Ezh2 amplicons of exons 10 and 18 from CRISPR-Cas9-targeted populations were cloned by primers listed in Supplemental Table S3. Amplicons from different samples were mixed and barcoded by another round of PCR using a Nextera Index Kit. The library prepared from the final PCR products of indexed samples was subjected to 300 bp paired-end sequencing with Illumina MiSeq for each of the two exon regions. After quality check (score of 25) and size (50 bp) trimming, reads of each sample were aligned to the sequences of designated regions (the consensus sequences of Ezh2 exons 10 and 18, generated from the control samples within the same sequencing run) with more than 10,000× depth of coverage, and followed with variant calling using CLC Genomic Workbench. Finally, an in-house tool was used to filter and summarize the allele frequencies of variantsand to confirm the allele independency of each mutation type. Frequencies of Gfi1 mutations were determined by analogous methods.

RNA-Sequencing analysis

RNA was extracted from mouse G3 MB and Ezh2-deleted tumorsphere cells and tumor chunks using Trizol, as previously described (Kawauchi et al., 2012). FASTQ sequences were mapped to the mouse mm9 genome by a pipeline that employs STAR (Dobin et al., 2013) and BWA (Li and Durbin, 2009). Mapped reads were counted with HTSEQ (Anders et al., 2015). Raw counts were voom normalized and statistically contrasted using the pipeline limma in R 3.2.3 (Law et al., 2014). An empirical Bayes fit was applied to contrast cells lacking Ezh2 to control cells and to generate p-values and false discovery rates for each gene. For GSEA and heat map visualizations, gene level FPKM values were computed, and the data was log2 transformed. Quality control and visualizations were performed using Partek Genomic Suite 6.6 (St Louis MO, USA) and STATA MP/11.2 (College Station TX, USA). A two factor ANOVA of the cells lacking Ezh2 was performed, and false discovery rates calculated for a model of genotype, tumor of origin and the interaction of the two factors. A two class comparison of Ezh2-deficient versus Ezh2^+/+ control tumors was examined by GSEA, and the 100 most differentially expressed genes were analyzed for GO category enrichment using DAVID. Heat maps were produced using Z transforms of log2 FPKM data in Spotfire Decision Site 9.1.2 from TIBCO (Boston MA, USA).

ChIP-Sequencing analysis

ChIP was performed as previously described (Lee et al., 2006). Briefly, 5–7 × 10⁷ crosslinked cells were resuspended in 10 ml buffer 1 (50 mM HEPES-KOH, pH 7.5, 140 mM NaCl, 1 mM EDTA, 10% glycerol, 0.5% NP-40, 0.25% Triton X-100, 1× Sigma protease inhibitor) and incubated on ice for 10 minutes. Samples were pelleted at 1,350×g for 5 minutes, resuspended in 10 ml buffer 2 (10 mM Tris-HCl, pH 8.0, 200 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 1× Sigma protease inhibitor) and incubated on ice for 10 minutes. Pelleted samples were resuspended in 2 ml Covaris sonication buffer (10 mM Tris-HCL, pH 8.0, 1 mM EDTA, 0.1% SDS, 1× Sigma protease inhibitor) and sonicated using a Covaris E220 focused ultrasonicator. Samples were diluted with 5 volumes of buffer 3 (10 mM Tris-HCl, pH 8.0, 100 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% Na-Deoxycholate, 1% Triton X-100, 1× Sigma protease inhibitor). Chromatin was subjected to ChIP using rabbit polyclonal antibodies directed against H3K27me3, H3K27ac, H3K4me3, and Gfi1 (Supplemental Table S2). DNA was purified using phenol:chloroform as described and sequenced (Lee et al., 2006).

We used BWA (version 0.5.9-r26-dev, default parameter) to align the reads to the mouse (UCSC mm9) or human (UCSC hg19) genome and then marked duplicated reads with Picard (version 1.65), with only nonduplicated reads kept by samtools (parameter “-q 1 -F 1024” version 0.1.18). To control the quality of the data and estimate the fragment size, the nonduplicated version of SPP (version 1.11) was used to draw cross-correlation with support of R (version 2.14.0). Upon manually inspecting the cross-correlation plot generated by SPP, the best fragment size estimates (the smallest fragment size estimated by SPP in all of our cases) were used to extend each read and to generate bigwig files to view on IGV (version 2.3.47). We scaled the bigwig files to 15 M reads so that the heights of peaks between samples are roughly comparable. The heat maps for different histone marks were generated using ngsplot (Shen et al., 2014).

Statistical analysis

The Kaplan-Meier method was used to calculate the significance of mouse survival. Statistical analyses were performed in the GraphPad Prism software version 6.0. P-values were calculated by an unpaired two-tailed t-test from 3 independent experiments.