Author Correction: Atrx inactivation drives disease-defining phenotypes in glioma cells of origin through global epigenomic remodeling

Mutational inactivation of the SWI/SNF chromatin regulator ATRX occurs frequently in gliomas, the most common primary brain tumors. Whether and how ATRX deficiency promotes oncogenesis by epigenomic dysregulation remains unclear, despite its recent implication in both genomic instability and telomere dysfunction. Here we report that Atrx loss recapitulates characteristic disease phenotypes and molecular features in putative glioma cells of origin, inducing cellular motility although also shifting differentiation state and potential toward an astrocytic rather than neuronal histiogenic profile. Moreover, Atrx deficiency drives widespread shifts in chromatin accessibility, histone composition, and transcription in a distribution almost entirely restricted to genomic sites normally bound by the protein. Finally, direct gene targets of Atrx that mediate specific Atrx-deficient phenotypes in vitro exhibit similarly selective misexpression in ATRX-mutant human gliomas. These findings demonstrate that ATRX deficiency and its epigenomic sequelae are sufficient to induce disease-defining oncogenic phenotypes in appropriate cellular and molecular contexts. ATRX inactivation frequently occurs in glioma. Here, the authors explore the role of ATRX inactivation in oncogenesis, highlighting ATRX deficiency driven epigenomic changes that influence the expression of genes crucial to the oncogenic phenotype.

S omatic mutations in genes encoding epigenetic regulators have now been widely reported across a number of cancer types 1 . Diffusely infiltrating gliomas, the most common primary brain tumors, feature loss-of function mutations in the SWI/SNF chromatin remodeler gene ATRX (α-thalassemia mental retardation X-linked) as defining molecular alterations delineating major adult and pediatric disease subtypes [2][3][4][5] . These incurable glioma variants tend to exhibit morphological and immunohistochemical features of astrocytes and, accordingly, are classified as "astrocytomas". Within this disease context, ATRX deficiency invariably co-occurs with mutations in TP53 and in genes encoding either isocitrate dehydrogenase enzymes (IDH1 and IDH2) in adults or H3.3 histone monomers (H3F3A and HIST1H3B) in children. However, inactivating ATRX mutations have also been identified in neuroblastoma, pancreatic neuroendocrine tumor, and multiple sarcomas, where they arise in distinct molecular and cellular contexts, indicating that oncogenic mechanisms of broad physiological relevance are likely mobilized by ATRX deficiency 6 .
What these mechanisms precisely entail remains unclear. Germline mutations in ATRX cause a rare, congenital neurodevelopmental condition associated with intellectual disability (ATR-X syndrome) 7 , which likely results from extensive p53dependent apoptosis within the neuronal precursor Atrx Vinculin p53 Tp53 -/-Atrx-Tp53-/-Atrx+ LGG R1 LGG R3 Proneural PG Mesenchymal Neural Classical Neuroblastic LGG R2 LGG Glioma subtypes exhibiting high rates of ATRX deficiency shown in red. EPL early progenitor-like, PG pre-glioblastoma, Oligo oligodendrocyte, Astro astrocyte, OPC oligodendroglial precursor cell, Astro (cult) cultured astrocyte compartment 8 . Such cell death may derive from genomic instability, given that ATRX deficiency has been repeatedly linked to replication stress, DNA damage, and aneuploidy 6 . Recent work suggests that the deleterious effects of ATRX deficiency on genomic integrity induce the alternative lengthening of telomeres (ALT) phenotype exhibited by tumors harboring ATRX mutations 9,10 . ALT is a telomerase-independent mechanism of telomere maintenance based on homologous recombination 11 , and as such, may serve to potentiate cellular immortality. However, while the potential for unlimited self-renewal obviously represents a requirement for oncogenesis, its sufficiency in and of itself is questionable. By this reasoning, ATRX deficiency must induce additional oncogenic phenotypes in affected tumors, likely involving shifts in transcriptional profiles, given the established role of ATRX in chromatin biology. Recent studies have extensively implicated ATRX in the regulation of chromatin state and composition [12][13][14][15] . In particular, ATRX is known to form a complex with DAXX (death-associated protein 6) to modulate H3.3 histone composition at sites across the genome 16,17 . ATRX also appears to be required for the normal expression of specific genes. For instance, impaired ATRXdependent transcription at the α-globin locus is thought to induce α-thalassemia in ATR-X syndrome 18 . Whether and how the epigenomic consequences of ATRX deficiency promote gliomagenesis and disease-relevant phenotypes has not been explored.
To model the cellular and molecular context of ATRX-mutant gliomagenesis, we inactivated Atrx in Tp53-intact and deficient murine neuroepithelial progenitors (mNPCs). We report that Atrx deficiency, particularly when combined with Tp53 loss, dramatically alters mNPC phenotypes, promoting cellular migration and shifting differentiation markers toward an astrocytic lineage profile. In this way, Atrx loss in mNPCs recapitulates the two biological features most classically associated with diffusely infiltrating astrocytomas. Investigating gene expression changes underlying these phenotypes, we establish compelling correlations with widespread shifts in chromatin accessibility and H3.3 composition, whose positioning strikingly corresponds to the normal distribution pattern of Atrx binding. Finally, we confirm that Atrx-dependent target genes found to mediate migratory and differentiation phenotypes in mNPCs are similarly misexpressed and functionally implicated in ATRX-mutant human glioma tumors and cell lines. Taken together, our results demonstrate that ATRX deficiency drives glioma-relevant biology by directly modulating chromatin architecture and composition, thereby influencing the expression of phenotypically crucial gene sets.
To better ascertain how the transcriptional programs mobilized by Atrx deficiency reflect glioma biology, we employed single sample gene set enrichment analysis (ssGSEA) to correlate mNPC gene expression patterns with established tumor signatures from TCGA (the Cancer Genome Atlas) and other sources 2,25,26 . We found that Atrx deficiency, in the setting of inactivated Tp53, shifted mNPC transcriptional profiles such that they were strongly correlated with the gene expression signatures of several glioma subclasses (Fig. 1d). Neither Tp53 nor Atrx loss alone elicited this response. Particularly robust associations were found with early progenitor-like (EPL), LGG R1, and LGG R3 signatures, all of which are derived from glioma subclasses featuring high rates of ATRX mutation 2,4 . Moreover, using published neuroepithelial lineage profiles derived from mice 27 , we demonstrated that inactivating Atrx in Tp53 −/− mNPCs shifted transcriptional patterns toward mature astrocytic and oligodendroglial signatures, at the expense of those corresponding to neurons and oligodendrocyte precursor cells (OPCs; Fig. 1e). These findings indicate that transcriptional alterations induced by ATRX deficiency in cultured mNPCs at least partially recapitulate those occurring in ATRX-mutant gliomas, while also implicating cellular developmental programing in the mediation of this biology.
Atrx deficiency modulates differentiation state and motility. To further explore Atrx-dependent transcriptional alterations and their functional consequences, we focused all subsequent analyses on Tp53 −/− mNPCs and those genes differentially expressed (fold change>2, Q<0.01; ANOVA FDR test) between Atrx+ and Atrx− states (N = 2012, Supplementary Fig. 2, Supplementary Data 1). Consistent with our ssGSEA findings, we found that the pool of overexpressed transcripts (N = 1105) exhibited significant overlaps with gene ontology (GO) functional sets associated with glial and neuronal development (Fig. 2a). These strong correlations prompted us to investigate the impact of Atrx deficiency on cellular differentiation state and capacity. We examined the expression of established lineage markers in Atrx+ and Atrx-mNPCs in the context of induced differentiation. We found that, over the course of five days in differentiating conditions, Atrx deficiency abrogated the expression of both the neuronal marker tubulin β 3 (Tubb3) and the oligodendrocytic marker Olig2, and instead increased expression of the astrocytic marker glial fibrillary acidic protein (Gfap; Fig. 2b, c). Evaluating other lineage markers yielded similar results ( Supplementary Fig. 3). Moreover, Atrx deficiency reduced the number of morphologically mature neurons over the five-day differentiation period in culture (Supplementary Fig. 4), an observation confirmed by Tubb3 immunofluorescence (Fig. 2d, e). We also used our gene expression data to investigate the precise lineage states whose expression patterns best correlated with Atrx-mNPCs. Recent work has identified marker-designated subpopulations in the embryonic and neonatal mouse brain corresponding to astrocytic and oligodendrocytic precursors at different developmental stages 28 . Using gene In all cases, data are scaled relative to the highest expression level for each marker across the sample set. Error bars reflect standard deviation. c western blot showing levels of histiogenic markers in Atrx+ and Atrx− mNPCs in proliferation media (P) and in differentiation media for 1, 3, and 5 days (d1, d3, d5; Vinculin loading control). d, e Tubb3 (green) immunofluorescence images (d) and quantification (e) showing reduced neuronal processes and a lower number of Tubb3+ cells in Atrx-mNPCs after 5 days in differentiating conditions. In upper plot, data represent the average percentage of Tubb3 + cells counted in 15 fields at ×20 magnification for each genotype; in lower plot, data represent the average length of Tubb3 + protrusions, 30 cells/genotype analyzed (DAPI counterstain; error bars reflect standard deviation; P values determined by unpaired two-tailed t-test). f heat map showing how differentially transcribed genes in Atrx-mNPCs (upregulated and downregulated) are expressed in distinct astrocytic and oligodendrocytic precursor compartments. Embryonic days 14.5, E16.5, and E18.5 (E14.5, E16.5, and E18.5) and postnatal days 4 and 7 (P4 and P7) are displayed expression data derived from these distinct cellular compartments, we demonstrated that Atrx-dependent transcriptional alterations were strongly correlated and anti-correlated with those of developing astrocytes and oligodendrocytes, respectively (Fig. 2f). Taken together, these findings indicate that Atrx deficiency alters the differentiation profile of mNPCs, favoring astrocytic over neuronal and oligodendrocytic histiogenesis. We also found significant correlations between transcripts overexpressed in the setting of Atrx deficiency and gene sets implicated in cellular motility (Fig. 2a). To validate these associations experimentally, we demonstrated that Atrx-mNPCs migrated faster than Atrx+ isogenics in both transwell and wound healing (scratch) assays (Figs 3a-c and Supplementary Fig. 5). Moreover, IP-western blot showed increased levels of activated, GTP-bound RhoA in Atrx-mNPCs ( Fig. 3d-e). Rho GTPase signaling has been repeatedly implicated in normal as well as cancer cell migration 29 . Finally, treatment with the RhoA inhibitor C3 exoenzyme, as expected, significantly reduced mNPCs motility in both Atrx-and Atrx+ contexts, notably with a dose-dependent response only evident in Atrx-mNPCs (Fig. 3f). These findings show that Atrx deficiency promotes mNPC migration, at least in part through Rho GTPase-dependent mechanisms. Intriguingly, the increased cellular motility and astrocytic histiogenic profile induced by Atrx deficiency in this in vitro model recapitulate two of the hallmark features associated with diffuse astrocytoma, the predominant ATRX-mutant human glioma subtype 30 .
Atrx exhibits a gene regulatory binding pattern in mNPCs. Having documented widespread transcriptional alterations and disease-relevant phenotypes emerging with Atrx deficiency, we sought to characterize whether and how global epigenomic remodeling might be involved in mediating these effects. Earlier work had shown that ATRX binds to tandem repeat sequences in telomeric regions as well as euchromatin, where its loss can impact gene expression 18 . To establish the binding distribution of Atrx in mNPCs, we optimized chromatin immunoprecipitation (ChIP) in the murine context, using published sequences where Atrx is known to coordinate. Preliminary studies revealed robust enrichment for Atrx at these predicted binding sites (Fig. 4a). We then proceeded with ChIP-high throughput sequencing (ChIPseq) in our mNPC lines, using Atrx-isogenics as negative controls. Our findings ( Fig. 4b-e, Supplementary Data 2, 3) contrasted sharply with what had previously been reported in mESCs 18 , despite an almost identical degree of overlap (66% versus 65%) between Atrx enrichment peaks and tandem repeat sites in the murine genome. In mNPCs, Atrx exhibited a much larger number of binding sites (73,723 and 76,164 in Tp53 −/− and Tp53 +/+ lines, respectively) than in mESCs (19,103). Notable qualitative differences were also apparent, with only 8.0% overlap between mESC and Tp53 −/− mNPC peaks. By contrast, enrichment peaks for Tp53 +/+ and Tp53 −/− mNPCs overlapped extensively (83.2%), indicating that p53 loss has, at best, only a modest effect on the genomic positioning of Atrx.
The distribution of Atrx binding with respect to open reading frames (ORFs) was also quite different in NPCs relative to mESCs. Whereas Atrx binding in mESCs was largely restricted to intergenic sites, the majority of Atrx peaks in mNPCs (67%) involved either gene bodies or promoter regions (Fig. 4b).
Given the extensive and gene-associated distribution of Atrx in mNPCs, we hypothesized that its deficiency would promote widespread, transcriptionally relevant shifts in chromatin state.
To explore this possibility, we performed assay for transposaseaccessible chromatin with high-throughput sequencing (ATACseq) 31 in both Atrx+ and Atrx− mNPCs, identifying genomic loci whose accessibility were significantly altered in the setting of Atrx deficiency. This analysis revealed thousands of sites across the genome where chromatin exhibited either significantly increased (N = 11607) or decreased (N = 7887) accessibility upon Atrx inactivation (Fig. 5a, Supplementary Data 4, 5). We subsequently referred to these sites as "open" and "closed", respectively.
Remarkably, 82.3% of ATAC-seq-altered sites (80.4% for open and 84.8% for closed) exhibited overlap with the genomic coordinates of Atrx enrichment peaks (Fig. 5b). These findings indicate that altered chromatin accessibility arising in the context of Atrx deficiency is largely restricted to the immediate vicinity of Atrx binding sites. Not surprisingly, we also documented genomic associations between ATAC-seq-altered sites and both TSSs and enhancer regions, reflecting their intimate relationship with Atrx binding loci (Fig. 5c). We then examined whether these shifts in chromatin accessibility were associated with differential gene expression, focusing on ORFs harboring ATAC-seq open and/or closed sites within 10 kb of their TSSs (Supplementary Data 1). This analysis revealed highly significant correlations between open sites and upregulated genes, and between closed sites and downregulated genes (Fig. 5d). The converse relationships were not observed. Reassuringly, key genes associated with astrocytic lineage (Gfap, Id3, and Stat3) [32][33][34] and upregulated in the context of the differentiation phenotype described above were also associated with ATAC-seq open sites (Fig. 5e, Supplementary Data 1). These findings reveal intriguing links between Atrx deficiency, chromatin structure, and gene expression. Moreover, they suggest that the majority of transcriptionally relevant epigenetic events occurring with Atrx deficiency derive from locally dysregulated chromatin composition at vacant Atrx binding loci.
Gna13 promotes Atrx-deficient glioma cell motility. Given the association of ATAC-seq open sites with upregulated astrocytic lineage markers (see above), we hypothesized that regions of increased chromatin accessibility would also arise in spatial proximity to genes mediating the observed motility phenotype of Atrx-mNPCs. Accordingly, we cross-referenced genes harboring ATAC-seq open sites and Atrx enrichment peaks within 10 kb of their TSSs with upregulated transcripts also bearing functional associations with motility, migration, and/or invasion by either ingenuity pathway analysis (IPA) or GO (Fig. 6a). This process identified 38 genes (Supplementary Table 1), which we then subjected to systematic siRNA knockdown in the context of the transwell migration assay. We found several genes for which significant repression reproducibly correlated with abrogation of cell migration (Fig. 6b). Among these targets, knockdown of Gna13, which encodes the α subunit of the heterotrimeric Gprotein Gα13, had the greatest effect. Several studies have linked the G12 G-protein subfamily, which includes Gα12 and Gα13, to cancer cell proliferation and motility 35 . Moreover, G12 Gproteins signal extensively through downstream RhoA GTPases 35 , which our earlier findings had already implicated in Atrxdeficient mNPC motility (Fig. 3d-f).
To assess whether GNA13 might facilitate similar processes in ATRX-deficient human disease, we examined TCGA gene expression data for lower-grade glioma (LGG), focusing our analysis on those tumors harboring an IDH1 or IDH2 mutation. As indicated above, ATRX mutation in adult glioma arises almost exclusively in the context of concurrent IDH mutation. Moreover, recent work has shown that IDH-mutant gliomas are composed of two predominant subtypes, those with and those without 1p/ 19q chromosomal codeletion (IDHmut-codel and IDHmutnoncodel), with the IDHmut-noncodel subclass not only harboring the vast majority of ATRX-mutant tumors, but also exhibiting uniformly low-level ATRX expression 2 . Our analysis confirmed this latter association, while also demonstrating significantly higher levels of GNA13 transcript in IDHmut-noncodel gliomas (Fig. 6c, d). Notably, two additional motility regulators, AGT and EMILIN1, implicated in our siRNA screen were similarly overexpressed as were the astrocytic markers/regulators GFAP, ID3, and STAT3 ( Fig. 6b and Supplementary Fig. 6). In parallel, we performed immunohistochemical studies on ten IDH-mutant gliomas with known ATRX expression status. We found that GNA13 levels were consistently higher in ATRX-tumors than in ATRX+ counterparts (Fig. 6e).
We then evaluated the expression of GNA13 protein in primary glioma stem cell (GSC) lines derived from patient tumors. GS 5-22, an IDH1-mutant (IDH1 R132H) GSC also featuring ATRX mutation, exhibited higher GNA13 expression than TS 603, an IDH1-mutant, ATRX-wild type line, as well as two IDH-wild type GSCs (GS 600 and TS 543; Fig. 7a, b). Moreover, GS 5-22 exhibited increased transwell migration relative to TS 603 (Fig. 7c). These relationships were recapitulated in a second IDH1-mutant, ATRX-mutant GSC line (GS 8-18) and elevated levels of GNA13 transcript were also seen in third IDH1mutant ATRX-mutant GSC line (JHH-273) only capable of forming subcutaneous xenografts 36 (Supplementary Fig. 7). Finally, we used electroporation to transiently restore human ATRX expression in both mNPCs and GS 5-22 cells. While our protocol resulted in only~20-30% transfection efficiency, we were nevertheless able to demonstrate significant reversion of the migratory phenotype accompanied by a modest, but significant decrease in Gna13/GNA13 expression in both cell lines (Fig. 7d-j).
Taken together, these data firmly implicate GNA13 and downstream RhoA signaling, along with potentially other ATRXresponsive genes, in the process of ATRX-dependent glioma cell migration. In a larger sense, they also suggest that functionally relevant epigenetic and transcriptional alterations mobilized by Atrx deficiency in murine mNPCs are also operative in ATRXmutant human gliomas.
Atrx-dependent epigenomic shifts correlate with H3. 3. ATRX has been shown to regulate incorporation of H3.3 histone monomers into nucleosomes 16,17 . To determine whether ATRXdependent shifts in chromatin accessibility involve altered H3.3 composition, we performed H3.3 ChIP in Atrx+ and Atrx-mNPCs, focusing on genomic loci previously demonstrated to be both Atrx-bound and associated with differentially expressed genes in the context of Atrx deficiency. We found significant changes in H3.3 enrichment within these regions, which included binding peaks approximating several phenotypically relevant transcripts, such as Gna13 and Gfap (Fig. 8a, b and Supplementary Fig. 8a-8d). By contrast, sampled genes not exhibiting differential expression with Atrx deficiency that were either Atrx-bound (Tchp) or unbound (Cd163) failed to exhibit significant H3.3 enrichment over IgG control (Supplementary Fig. 8e, f). We also performed H3.3 ChIP-seq in both Atrx+ and Atrx-mNPCs to ascertain the effects of Atrx deficiency on H3.3 composition genome-wide. While this global analysis lacked the sensitivity of our focused, RT-qPCR-based approach, we were still able to identify 560 sites of differential H3.3 incorporation and demonstrate 77.2% overlap with Atrx binding peaks (Fig. 8c, Supplementary Data 6). Finally, focused ChIP analysis in ATRX+ (TS 603) and ATRX-(GS 5-22) GSCs revealed differential H3.3 incorporation in association with GNA13 (Fig. 8d), extending our findings into bona fide human disease models. These results indicate that Atrx deficiency abnormally alters H3.3 composition at vacant Atrx binding sites, which in turn mediates, at least in part, shifts in chromatin accessibility and gene expression driving glioma-relevant phenotypes.

Discussion
ATRX inactivation defines large subsets of adult and pediatric glioma 2-5 , where its sheer frequency argues for a major role in oncogenesis. Considering the established role of ATRX as an epigenetic regulator, we hypothesized that its inactivation induces glioma-relevant phenotypes by way of chromatin remodeling and downstream shifts in gene expression. To investigate this possibility, we re-engineered the basic molecular features of ATRXdeficient glioma in a disease-relevant cellular context. Primary GSCs harboring ATRX mutation have thus far proven difficult to culture and genetically manipulate, and the one Atrx-deficient murine model generated to date layers Atrx loss upon an already high-grade, NRAS-driven glioma 37 , affording limited opportunity for the study of more subtle phenotypes. As indicated above, a number of investigations have implicated mNPCs as potential glioma cells of origin 19 . Although no in vitro system can perfectly recapitulate the process of oncogenesis, these primary cells, when not excessively passaged, offered an opportunity to study the gliomagenic effects of Atrx deficiency in a molecular background not extensively altered by other cancer-promoting physiological dysruptions. The sufficiency of ATRX inactivation, either alone or in combination with TP53 loss, to induce proliferative gliomagenesis has yet to be demonstrated. Our work, along with that described in another recent report 38 , supports the notion that the oncogenic effects of ATRX deficiency are more measured and may be dependent on as yet uncharacterized molecular factors. Nevertheless, Atrx inactivation induced widespread gene expression changes in mNPCs, especially when paired with Tp53 loss, recapitulating its core mutational context in human gliomas. Prior functional analysis of ATRX deficiency in the central nervous system has occurred primarily in the TP53-intact setting 8,21,22 , which tends to result in cell death and, as such, may not represent an ideal model for glioma biology. Indeed, we also observed increased cell death in Atrx-intact mNPC isogenics. Reassuringly, combined Atrx and Tp53 loss altered mNPC transcriptional patterns such that they strongly correlated with several glioma signatures, particularly those derived from ATRXdeficient tumors.
Atrx-dependent gene expression changes in mNPCs were accompanied by altered cellular morphology, increased motility, and a shift in differentiation markers toward an astrocytic histiogenic profile. These latter two phenotypes recapitulate two cardinal features of ATRX-mutant adult and pediatric gliomas, namely widespread tumor cell invasion into surrounding brain and astrocytic histopathology 30 . Until quite recently, primary brain tumors were classified solely on the basis of presumed

GNA13
Relative enrichment * * Fig. 8 Atrx-dependent epigenetic and transcriptional alterations correlate with H3.3 composition. a, b H3.3 ChIP qPCR from Atrx-bound genomic loci approximating Gna13 and Gfap reveals differential enrichment between Atrx+ and Atrx-mNPCs at many points. Positioning of loci subjected to qPCR relative to Atrx binding peaks and open reading frames (Gna13 and Gfap) is also shown (left panels). c sample trace showing relationship of Atrx ChIP-seq peaks (black) to sites of significantly differential H3.3 composition (ΔH3.3) arising with Atrx deficiency (red). d H3.3 ChIP qPCR from Atrx-bound genomic loci approximating GNA13 reveals differential enrichment between ATRX+ (TS 603) and ATRX-(GS 5-22) human GSCs (right panels; *P < 0.05; **P < 0.001; ***P < 0.0001. In all cases, error bars reflect standard deviation and P values determined by unpaired two-tailed t-test histiogenesis 39 . Within this schema, the morphological and immunohistochemical features of diffuse astrocytoma were thought reflect fundamental derivation from an astrocytic precursor, although direct evidence supporting this hypothesis was limited 40 . By demonstrating that ATRX deficiency directly mobilizes the expression of key astrocytic markers and regulators, we provide compelling evidence that this disease-defining molecular alteration, in and of itself, directs precursor cells of unspecified lineage toward an astrocytic phenotype at the expense of normal neuronal and/or oligodendrocytic differentiation. These findings support the notion that apparent lineage derivation in cancer is a fluid concept, dependent on molecular as well as cellular context, while also suggesting that mobilization of latent developmentally relevant pathways may underlie a significant subset of ATRX-deficient biology. The migratory behavior of diffusely infiltrating gliomas, including ATRX-mutant astrocytomas, represents a major determinant of their malignant potential, effectively rendering them incurable by surgical resection. Accordingly, an improved understanding of the molecular events driving this phenotype would have tangible therapeutic implications. Our analysis implicated multiple Atrx target genes in glioma cell motility, including Gna13, whose direct transcriptional upregulation and downstream signaling through RhoA GTPases promoted mNPC migration. Moreover, our analysis of human tissue samples and multiple primary GSCs revealed that this molecular mechanism might also be operative in ATRX-deficient gliomas. In this way, we established compelling mechanistic links between ATRX deficiency and targetable molecular networks classically implicated in cancer cell motility. Prior work using LN 229 glioblastoma cells as a model system had found reduced cell migration arising with ATRX deficiency 41 . However, a more recent report has validated our findings in neural stem cells also harboring molecular alterations (IDH1 mutation and TP53 inactivation) characterizing ATRX-deficient astrocytoma 38 .
To better understand how Atrx, or lack thereof, modulates disease-relevant transcriptional profiles in murine mNPCs, we assessed its genomic distribution by ChIP-seq. We found an unexpectedly extensive binding pattern that contrasted sharply, both quantitatively and qualitatively, with what had been previously reported for mESCs 18 . Although the different number of Atrx binding peaks identified in our mNPC profiles may partially reflect technical variability between the two studies, the notable correlation of Atrx distribution with gene bodies and promoter regions, not seen in mESCs, speaks to a hitherto unappreciated, cell line-specific functionality for Atrx. In particular, a limited set of histiogenic lineages appear to feature epigenomic states unusually dependent on Atrx for normal gene regulation, and by extension, exquisitely sensitive to Atrx inactivation. Intriguingly, they also reflect the relatively narrow spectrum of tumor types exhibiting high rates of ATRX mutation, which are notably enriched for neuroepithelial derivation (e.g., adult and pediatric glioma, neuroblastoma, and pancreatic neuroendocrine tumor) 6 . In this way, the oncogenic effects of ATRX deficiency likely require specific cellular contexts to achieve their full expression.
The molecular context permitting ATRX-deficient gliomagenesis may be similarly restricted. Indeed, for reasons described above, ATRX-mutant gliomas almost invariably harbor TP53 mutations. Moreover, the vast majority of adult and pediatric gliomas with ATRX deficiency also contain mutant IDH and H3.3 genes, respectively, both of which have been shown to significantly and distinctively impact global epigenomic landscapes [42][43][44] . In the present study, we chose to focus our analysis on the core signature of ATRX and TP53 inactivation, both to identify the precise epigenetic and transcriptional consequences of ATRX deficiency and also obtain findings broadly generalizable across adult and pediatric glioma. We acknowledge, however, that including either mutant IDH or H3.3 proteins in our mNPC system could promote additional epigenomic modifications, impacting transcriptional and phenotypic profiles. That being said, our validation studies in ATRX-deficient human gliomas suggest that the cell migratory and differentiation phenotypes identified in our mNPC models are driven primarily by ATRX loss.
Atrx inactivation in mNPCs had profound effects on the global epigenome. We observed widespread changes in chromatin accessibility whose positioning strikingly reflected both the normal binding distribution of Atrx as well as Atrx-dependent transcriptional alterations. These data support a model in which ATRX loss modulates chromatin structure primarily in the immediate vicinity of its vacant binding sites, dysregulating local gene expression and inducing glioma-relevant phenotypes (Fig. 9). These localized shifts in chromatin architecture imply underlying alterations in nucleosome composition. Supporting this conjecture, and consistent with prior reports 16,17 , we found altered levels of H3.3 enrichment in Atrx+ and Atrx-isogenic mNPCs by both focused and global ChIP analysis at geneassociated Atrx-bound loci whose differential expression mediated key Atrx-deficient phenotypes. Findings related to GNA13 were further validated in human GSCs, albeit in a non-isogenic context. Prior studies have shown that H3.3 compositional changes occur during neuroepithelial differentiation, and that ATRX mediates H3.3 deposition at telomeres as well as heterochromatin sites associated with the silenced alleles of imprinted genes 15,45 . Our findings expand on these conclusions, demonstrating a novel, cell type-specific role for ATRX in nervous system development grounded in broad H3.3 regulation at ORFs genome-wide. Moreover, our documentation of genomic sites exhibiting either increased or decreased H3.3 content in the setting of Atrx deficiency suggests that Atrx may play a more general role in precisely titrating H3.3 composition at target site nucleosomes, rather than simply promoting H3.3 incorporation. Recent work has demonstrated that ATRX deficiency also modulates the composition of macroH2A histones in chromatin as well as the levels of H3K27 and H3K9 trimethylation marks [12][13][14][15]46 . Although our data do not exclude the involvement of these regulatory mechanisms, we note that none of these prior studies localized the effects of ATRX deficiency so precisely to vacant ATRX binding sites. Accordingly, our findings support a distinct mechanism of ATRX-based transcriptional control at least partially dependent on differential H3.3 incorporation into geneassociated nucleosome complexes. Moreover, they establish firm links between ATRX-mediated H3.3 regulation and functionally relevant gene expression programs on a global scale.
Prior reports have revealed compelling associations between ATRX deficiency and ALT 9,10,37,47 , a process that likely enables cellular immortality in ATRX-mutant tumors. That virtually all IDH-mutant adult gliomas feature either ATRX deficiency or activating mutations in TERT 2 , which encodes the catalytic core of telomerase, speaks to the absolute necessity of pathological telomere maintenance across the tumor group. However, only a small fraction of IDH-mutant, TERT-mutant tumors do not also feature 1p/19q codeletion 2 . Accordingly, this same genetic evidence implies that ATRX deficiency must induce other essential phenotypes in affected tumors. Moreover, while ATRX mutations are frequently found in H3.3 mutant pediatric gliomas, TERT mutations are rare 48 , further supporting the notion that the oncogenic sequelae of ATRX deficiency are not limited to ALT. Our data support this conjecture by demonstrating specific, ALTindependent molecular mechanisms mediating disease-relevant biological behavior. That Atrx inactivation was insufficient to induce a full ALT phenotype in our mNPCs is consistent with earlier studies in both murine and human experimental systems 23,24,37 . These findings illustrate the multifactorial influences of ATRX deficiency on the oncogenic process.
In summary, we describe, for the first time, the epigenomic consequences of ATRX deficiency in putative glioma cells of origin, along with their effects on gene expression programs and cellular differentiation and motility phenotypes. In doing so, our work characterizes targetable molecular pathways mediating wellestablished malignant behavior in diffusely infiltrating astrocytic gliomas, while also revealing discrete examples of how mutational disruption of chromatin regulation impacts the expression of tumor-specific biology. mNPC isolation and culture. Forebrains were removed from p1 mouse pups and mechanically homogenized by repetitive pipetting in NeuroCult Basal Medium containing NeuroCult Proliferation Supplement, 20 ng/ml EGF, 10 ng/ml basic FGF, 2 μg/ml heparin (Stemcell Technologies), 50 units/ml penicillin, and 50 μg/ml streptomycin (Thermo Fisher Scientific). Homogenized samples were then cultured in a humidified incubator at 5% CO 2 and 37°C. To inactivate Atrx, cells were infected with a Cre-IRES-Puro lentiviral plasmid (a gift from Darrell Kotton; Addgene plasmid #30205) 50 and selected with 1 μg/mL of puromycin (Sigma-Aldrich). To induce differentiation, cells were plated on dishes coated with 10 μg/ ml laminin (Sigma-Aldrich) and cultured in NeuroCult Basal Medium containing NeuroCult Differentiation Supplement (Stemcell Technologies). For experiments, cells were used at passages 5-10.
Proliferation assays. Cell proliferation was assessed by multiple methods. (1) 2D proliferation assay: 5 × 10 3 mouse mNPCs were grown in triplicate on laminincoated (10 μg/ml) 96-well plates and stopped at different time points. Proliferation was evaluated using the CellTiter-Glo Luminescence Cell Viability Assay (Promega), following manufacturer instructions. (2) 3D-proliferation assay: mNPCs in triplicate wells were embedded in 0.5% 2-hydroxyethyl agarose (Agarose VII, Sigma-Aldrich) and cultured for 1 week. Colonies were then stained with crystal violet and counted. (3) Cell Cycle analysis: cell cycle profiles were analyzed by propidium iodide staining (50 μg/ml propidium iodide and 100 μg/ml RNase A) followed by fluorescence-activated cell sorting (FACS) analysis. (4) In vivo growth of mNPCs was assessed by intracranial injection of 1 × 10 5 luciferase expressing cells in 8-week old Nu/Nu mice (Charles River). Growth was monitored by Xenogen IVIS imaging following luciferin injection and data were analyzed by IVIS Lumina LivingImage software.
Cell motility assays. To assess transwell migration, cells were labeled with 5 μg/ml of the green fluorescent tracer DiO (Life Technologies). A total of 2 × 10 5 cells were added to the upper compartment of a 24-well FluoroBlok insert (Corning) coated with laminin 20 μg/ml, in NeuroCult Basal Medium plus 0.1% BSA (Sigma-Aldrich). 20 ng/ml EGF was used as a chemoattractant in the lower chamber. After 24 h, invading cells were detected with a Cytation 3 Imaging System microplate reader (BioTek). Displayed data reflects 5 biological replicates. For wound-healing assays, exponentially growing cells were seeded (1 × 10 5 ) in a 96-well laminincoated ImageLock plate (Essen BioScience) in 3 biological replicates to create a dense monolayer and then scratched with the Wound Maker (Essen BioScience), following manufacturer instructions. NeuroCult Basal Medium plus 0.1% BSA and 20 ng/ml EGF was added after washing with PBS, and wound closure was monitored by an IncuCyte Zoom System (Essen BioSciences) for 24 h. Images were analyzed with Incucyte Scratch Wound Cell Migration Software (Essen BioSciences) to determine cellular density within the wound relative to surrounding cellular density. When applicable, RhoA inhibitor I (C3 exoenzyme; Cytoskeleton, Inc.) was used at the indicated concentrations (see above). For siRNA-based screening, mNPCs were transfected (DharmaFECT 1) with 25 nM siRNA pools (ON-TARGETplus SMARTpool; Dharmacon RNA Technologies) corresponding to genes of interest (Supplementary Table 1). After 36 h, cells were detached and assayed for transwell migration as described above.
Immunohistochemistry. Immunohistochemistry for GNA13 was performed using a polymeric biotin-free horseradish peroxidase method on a Leica Microsystems Bond RXM automated stainer. The Refine polymer detection kit (Leica Biosystems) was used for detection. For antibody information, please see Supplementary  Table 2.
Immunofluorescence. Cells were grown on polylysine (Sigma) treated and laminin coated Labtek chamber slides (Nunc), and fixed with 4% paraformaldehyde in PBS for 15 min. They were then permeabilized (with PBS, 1% BSA, 0.1% Triton X-100, and 2% serum) for 5 min and saturated with the blocking buffer (PBS, 1% BSA, and 2% serum) for 30 min. Primary (Supplementary Table 2) and secondary antibodies were successively incubated at room temperature for 1 h each. Images were acquired with a Nikon A1R MP confocal microscope. Quantifications were performed using NIH ImageJ software (http://rsb.info.nih.gov/ij/). For telomere FISH, 5 μm frozen sections were fixed with 4% paraformaldehyde for 10 min, dehydrated, and denatured at 85°C for 5 min in hybridization mixture consisting of 10 mM Tris-HCL pH 7.2, 70% formamide, 0.5% blocking solution reagent (Roche), and a complementary Cy3-labeled peptide nucleic acid probe specific to the mammalian telomere repeat sequence TTAGGGTTAGGGTTAGGG 3′ (Applied Biosystems). Hybridization was performed at room temperature for two hours in the hybridization mixture. Slides were then washed twice with 70% formamide, 10 mM Tris-HCl pH 7.0-7.5, 0.1% BSA followed by multiple washes in PBS. Finally, slides were mounted in Prolong Gold with DAPI (Invitrogen, Grand Island, NY, USA) and examined using a Leica CTR5000 inverted fluorescent microscope.
RhoA activity assay. Exponentially growing cells were harvested and lysed. 500 μg of protein were subjected to rhotekin-RBD bead pulldown (Rho Activation Assay Biochem Kit, Cytoskeleton Inc.) for 1 h at 4°C. After thorough washing, the samples were boiled for 5 min in Laemmli sample buffer to detach active GTPbound Rho and then loaded onto 12% SDS-PAGE gels for immunoblotting with an anti-RhoA antibody (Supplementary Table 2).
RNA-Seq and data analysis. Whole transcriptome sequencing was performed at the MSKCC Integrated Genomics Operation. Total RNA was extracted from three biological mNPC replicates at passage 5 following cre-mediated Atrx inactivation. Cells were cultured in proliferation conditions and total RNA isolated using the RNeasy Mini Kit (Qiagen) according to manufacturers instructions. Following ribogreen quantification and quality control using an Agilent BioAnalyzer, 500 ng of total RNA was subjected to polyA selection and Truseq library preparation (Illumina) with six cycles of PCR. Samples were barcoded and sequenced on a HiSeq 2500 in a 50 bp/50 bp paired end run using the TruSeq SBS Kit v3 (Illumina). An average of 42 million paired reads was generated per sample. At most, ribosomal reads represented 0.1% of total and mRNA accounted for 74.9% on average. Raw sequencing data was converted to gene expression values using CLC genomics workbench (Qiagen). Data were then quantile normalized (Partek), clustered, and used to derive differentially expressed gene sets (2-fold expression change, Q < 0.01). Enrichment and pathway analyses were performed using Panther for GO correlations, along with IPA, both with standard parameters. ssGSEA 52 was performed using publically available glioma 2,25,26 and neuroepithelial ontology 27 signatures. Briefly, gene expression values for each sample were rank-ordered and an enrichment score determined using the Empirical Cumulative Distribution Function (ECDF) of the signature genes and remaining genes. Enrichment scores were obtained by integrating the differences between signature and non-signature ECDF results. Prior to undertaking ssGSEA, principal component analysis (PCA) of transcriptional data was performed using custom R scripts. This revealed one outlier replicate from the entire sample set ( Supplementary Fig. 9), which was excluded prior to further processing to optimally assess stratification between the different cellular genotypes. For neuroepithelial ontology signatures, the 500 most highly expressed genes were used. Glioma subtype-specific genes included the 250 most highly expressed transcripts as defined by a 1-versus rest 2-class SAM 53 . Analysis of Atrx-dependent transcriptional signatures in the external expression dataset of astrocytic and oligodendrocytic precursors at different stages of development 28 was carried out as follows. For visual display as heat maps (JavaTree-View), expression profiles were centered within each time point on the average of astrocytic and oligodendrocytic groups, and biological replicates within each experimental group and time point were then averaged together.
Quantitive PCR (qPCR). For RT-qPCR, RNA was prepared with RNeasy Mini Kits (Qiagen) and quality was checked by microfluidics separation using a Bioanalyser 2100. RNA was used for first-strand cDNA synthesis using random primers and SuperScript II Reverse Transcriptase (Invitrogen). qPCR was performed using Power SYBR Green PCR Master Mix (Applied Biosystems). Results were analyzed by the ΔΔCt method 54 using Gapdh as a housekeeping gene. Primers for RT-qPCR are listed in Supplementary Data 7, and for ChIP-PCR in Supplementary Data 8.
TCGA gene expression data. Normalized gene expression data were obtained from the Cancer Genome Atlas lower-grade glioma publication page (https://tcgadata.nci.nih.gov/docs/publications/lgg_2015/), and sample annotations were taken from supplementary appendix 2 of the lower-grade glioma publication 2 .
ChIP-Seq and analysis. For ATRX ChIP 18 , cells in three biological replicates cultured under proliferation conditions were fixed with 2 mM EGS (Pierce) for 45 min at room temperature in PBS. Formaldehyde was then added to 1% for 20 min and quenched with 125 mM glycine. Chromatin was sonicated with a Bioruptor Plus (Diagenode) to generate DNA fragments under 500 bp and lysates were immunoprecipitated with 40 μg ATRX H300 antibody or rabbit IgG control (Santa Cruz Biotechnology). Histone H3.3 chromatin immunoprecipitation was performed following the manufacturer instructions of Magna ChIP A Chromatin Immunoprecipitation Kit using 10 μg of Anti-Histone H3.3 antibody (Millipore). Sequencing libraries were prepared using the Kapa Hyper library Prep kit (Roche) with library amplification primer mix and sequenced as 50 base pair reads by single-read format on an Illumina HiSeq 2500 (v4 chemistry) with a maximum of 7 samples per lane (Genome Technology Center Facility, New York University). Fastq files were aligned to the mm9 genome using Bowtie2. The resulting BAM files for input and IP were subjected to peak calling using MACS2 with default parameters. For Atrx ChIP-seq, significant peaks derived from Atrx-cells were subtracted out. Promoter/gene body pie chart analysis was performed using in-house R scripts, defining promoters as regions within 1 kb of TSSs 18 . TSSs were downloaded from the University of California at Santa Cruz (UCSC) genome browser. In all cases the most upstream TSS was used for each gene. For display, SGR files were generated from BAM files and viewed in the Integrated Genome Browser. Previously published data sets for mESCs 18 were downloaded as fastq files and subjected to identical processing pipelines as described above for mNPCs.
ATAC-Seq and analysis. ATAC-Seq 31 was performed on mNPCs (three biological replicates per genotype) as follows. Nuclei were prepared by two 500 g centrifugation steps, each 5 min, separated by a 5 min wash in cold PBS. Nuclei were then lysed in cold 10 mM Tris-HCL, pH 7.4, 10 nM NaCl, 3 mM MgCl 2 , and 0.1% IGEPAL CA-630, and spun again at 500 × g for 10 min. Nuclear pellets were then resuspended in transposase reaction mix (25 µl 2 × TD buffer, 2.5 µl transposase (Illumina) and 22.5 µl nuclease-free water) for 30 min at 37°C. Samples were then purified using Qiagen MinElute kits. Libraries were amplified from purified samples using customized primer sets 31 and sequenced as paired end 50 reads on a HiSeq 2500 with version 4 chemistry and 6 samples per lane (~50 million paired end reads per sample). Fastq files were aligned to mouse genome build mm9 using BWA 55 . Polymerase chain reaction duplicated reads were filtered out using the MarkDiplicates.jar Picard software tool (http://picard.sourceforge.net). Significant peaks were then identified using the Zero Inflated Negative Binomial Algorithm (ZINBA) 56 . Resulting peak intensity information was combined with peak density data using DiffBind 57 to identify differential open chromatin sites between data derived from Atrx-and Atrx+ cell lines. Novel genes with open chromatin regions were then annotated using ANNOVAR 58 .
TSS and enhancer heat maps. TSS/promoter and enhancer regions for mouse E14.5 brain were obtained from enhancer-promoter units identified by the ENCODE project. Triplicates from ATAC-seq data were merged to produce unified aligned files and which were then transformed to heat maps using deepTools2 (version 2.2.4) with a 10 kb window 59 . For enhancers, a scale factor of 5 was used.
Statistics. Methods utilized for transcriptional and epigenomic analysis are described elsewhere (see above). Enrichment significance for ATAC-seq and/or Atrx ChIP-seq associated genes relative to differentially expressed transcripts was determined using the hypergeometric distribution. Analyses involving multiple datasets with replicates, including those for functional assays, were subjected to either paired or unpaired, two-tailed t-test.