Author Correction: Astroblastomas exhibit radial glia stem cell lineages and differential expression of imprinted and X-inactivation escape genes

Astroblastomas (ABs) are rare brain tumors of unknown origin. We performed an integrative genetic and epigenetic analysis of AB-like tumors. Here, we show that tumors traceable to neural stem/progenitor cells (radial glia) that emerge during early to later brain development occur in children and young adults, respectively. Tumors with MN1-BEND2 fusion appear to present exclusively in females and exhibit overexpression of genes expressed prior to 25 post-conception weeks (pcw), including genes enriched in early ventricular zone radial glia and ependymal tumors. Other, histologically classic ABs overexpress or harbor mutations of mitogen-activated protein kinase pathway genes, outer and truncated radial glia genes, and genes expressed after 25 pcw, including neuronal and astrocyte markers. Findings support that AB-like tumors arise in the context of epigenetic and genetic changes in neural progenitors. Selective gene fusion, variable imprinting and/or chromosome X-inactivation escape resulting in biallelic overexpression may contribute to female predominance of AB molecular subtypes. Astroblastoma (AB) is an uncommon brain tumour and its origin remains unknown. Here, the authors perform integrative molecular analysis of 35 AB-like tumours and provide evidence that these arise in the context of epigenetic and genetic changes in neural progenitors occurring during brain development.

A stroblastomas (ABs) are uncommon and historically controversial CNS neoplasms characterized by tumor cells surrounding a central capillary forming a flower petal-like structure called the astroblastic pseudorosette (Fig. 1a) [1][2][3] . ABlike tumors typically occur in the cerebral hemispheres of children and young adults and show considerable female predominance [3][4][5] . Their clinical behavior ranges from frequently surgically-amenable benign lesions to occasionally malignant. A long-standing question is whether ABs are of astrocytic or ependymal lineage 2 .
AB-like tumors are comprised of at least two genetic types based on the presence of the pV600E mutation of the B-Raf serine/threonine kinase gene (BRAF V600E ) 2,4,6 or rearrangement of the meningioma (disrupted in balanced translocation) 1 (MN1) gene 4,6 . Fusions between MN1 at chromosome region 22q12.1 and BEN domain containing 2 (BEND2) at Xp22.13 are reported in MN1-rearranged tumors 6 .
Embryonic NSCs include highly proliferative neuroepithelial cells (NEC), which comprise the primitive neural tube and give rise to early neuronal progenitors (neuroblasts). Around five postconception weeks (pcw) NEC begin to transform into ventricular zone radial glia (vRG), bipolar cells whose apical and basal processes contact the ventricular and pial surfaces of the developing brain, respectively 7,8 (Fig. 1b). NEC and vRG exhibit interkinetic nuclear migration (IKNM), which helps maintain the tightly packed cellular structure of the ventricular lining during cell division. Like NEC, vRG can directly give rise to neuronal progenitors through asymmetric division. Further such vRG divisions produce B1 astrocytes (future adult subventricular zone NSCs), followed by ependymal cells 9 . The latter replace vRG in lining the brain ventricles.
vRG also produce outer radial glia (oRG), whose nuclei reside in the outer subventricular zone (OSVZ), and which appear at about 12 pcw 10 . Unlike vRG, oRG lack a prominent apical process and demonstrate mitotic somal translocation (MST) in which the cell body migrates basally prior to mitosis, facilitating expansion of the OSVZ and cortical plate 11 . They generate intermediate progenitor (IP) cells capable of further division termed transiently amplifying progenitors which generate the majority of cortical neurons. Lastly, oRG also give rise to astrocytes and oligodendrocytes.
Nowakowski et al. 12 observed that the long basal processes of vRG become truncated in the human fetus at about 15 pcw. Unlike vRG and oRG, truncated radial glia (tRG) extend their basal processes only into the OSVZ where their end-feet typically surround blood capillaries similar to classic astroblastoma cells 1,2 (Fig. 1a, b).
DNA methylation in the human brain is dynamic during development. Some genes on the inactive X chromosome in females can escape X inactivation [13][14][15][16] . These X-inactivation escape (XIE) genes, can acquire promoter hypomethylation in somatic cells resulting in biallelic expression from both X chromosomes, potentially doubling their gene dosage 13 .
Here we show that AB-like tumors with MN1-BEND2 fusions appear to exclusively occur in females, and nearly always in children. They exhibit overexpression of IGF2-H19 and DLK1-DIO3 region, and other imprinted genes, XIE genes, and genes highly expressed during brain development prior to 25 pcw, including genes enriched in vRG and ependymal tumors. Other AB-like tumors overexpress mitogen-activated protein kinase (MAPK) pathway, phosphatidylinositol-3-kinase (PI3K) pathway and alternate XIE genes, and tend to occur in young adults of both sexes depending on molecular subtype. These tumors highly express oRG and tRG genes, and genes expressed after 25 pcw that are often involved in neuronal or astrocytic differentiation. Our findings strongly support that ABlike tumors arise in the context of epigenetic and genetic changes in neural progenitors during brain development: early ependymal tumors with MN1-BEND2 from early vRG-derived cells, and MAPK pathway tumors demonstrating more classic astroblastoma histomorphology from oRG-or tRG-derived progenitors, respectively. Data also suggest that increased gene fusion, variable imprinting, and XIE may contribute to the female predominance of certain AB-like tumor molecular subtypes. Lastly, we found that like MN1-BEND2 tumors, nascent immature ependymal cells express IGF2 and may represent an important source of this growth factor in the fetal lateral ventricular zone niche.

Results
Tumors with AB histology segregate into distinct subtypes based on transcriptomic and mutational profiles. We performed RNAseq on 24 supratentorial AB-like cases with available material, including six BRAF V600E -mutant tumors 2 , six MN1rearranaged tumors 4 and twelve tumors without these genetic lesions (Table 1). Three tumors diagnosed as supratentorial ependymoma (STE) and four papillary tumor of the pineal region (PTPR) 17 were also studied for comparison. Affymetrix gene microarray (HTA-2) expression analysis was additionally used to study 15 of the above AB-like tumors, one MN1-rearranaged tumor recurrence (C10r), the four PTPR, six histologically classic pleomorphic xanthoastrocytomas (PXA), four fetal brain samples (28-30 pcw) and 14 normal pediatric and adult brain samples.
Dimensionality reduction of the two expression datasets was then performed individually using tSNE (23,937 mappable genes in the Affymetrix data and 24,917 mappable and encodable RNAseq gene transcripts). Next, we identified 1793 genes differentially expressed among the resultant expression clusters that were common to both datasets using analysis of variance (P < 0.05).
Unsupervised hierarchical clustering of these 1793 genes using the RNAseq data yielded seven tumor clusters (Fig. 1c), similar to tSNE (Fig. 1e). Tumors with MN1 rearrangement, including five containing the MN1-BEND2 gene fusion (Table 1), formed a discrete RNA cluster, as did all BRAF V600E -mutant tumors combined with three tumors without BRAF V600E . The latter showed mutations of other MAPK, and frequently PI3K/AKT/ mTOR pathway genes (Fig. 2, Supplementary Data 1). We designated this RNA cluster MAPK-A. Three additional tumors without BRAF V600E mutation, but otherwise showing similar mutation profiles as MAPK-A tumors, closely clustered with them and were designated MAPK-B (Fig. 1c, e). Hence forth, we refer to these two groups collectively as MAPK-AB lesions. Three tumors lacking both BRAF V600E mutation or obvious PI3K pathway mutations, designated MAPK-C, also formed a subgroup within the MAPK supercluster (MAPK-ABC).
Three histologically atypical AB-like tumors (ATAB) also formed a transcriptomic cluster (Fig. 1c, e). These occurred in older adults and showed high-grade histology (Table 1). Lastly, three STEs (two containing the ZFTA-RELA fusion), and the four PTPR grouped into respective RNA clusters.
MN1-BEND2 tumors and PTPR share ependymal gene expression and low expression of oRG, glial and neuronal genes. SOX14, IRX2, CELSR1, KCNJ5 encoding the inward rectifier potassium channel Kir3.4, and TFF3 encoding an epithelial cell mobility protein, were overexpressed by MN1-rearranged tumors and PTPR almost exclusively (Fig. 1c, Figs. 1-3). Overall, these findings argue for ependymal differentiation of both MN1-BEND2 tumors and PTPR. The latter differ from MN1-BEND2 tumors and many STEs, however, in that they are IGF2 relative under-expressers. This may be related to their 3rd ventricular diencephalic derivation, in contrast to the likely telencephalic origin of MN1-BEND2 tumors and STE. Moreover, the fact that PTPR more highly express the NSC markers PAX6 and OTX2 argues they may be more closely related to a RG stem-like cell 28 , perhaps derived directly from such a cell and lacking an IP possibly required in the genesis of MN1-BEND2 tumors and STE.
In contrast to MAPK supercluster tumors, MN1-rearranged tumors, PTPR and STE showed relatively low expression of the neuronal genes GRIA2 (glutamate ionotropic receptor AMPA type subunit 2), MAOA (monoamine oxidase A, an XIE gene) and MAPT (microtubule associated protein tau). They also demonstrated low expression of the astrocyte markers OLIG2, GFAP, S100β and ALDH1, the oRG markers IDH1 and TNC 10 , and other genes highly expressed in MAPK-ABC tumors, e.g., TCF4,     TCF4 is a WNT transcription factor. LRP1B and CTNND2 are WNT pathway regulatory genes 29 . The latter encodes δ-catenin 2, which affects dendritic spine and synapse formation, and is a marker of oRG and neurons 10 . DISC1 encodes disrupted in schizophrenia 1, which is involved in neurogenesis and astrogenesis, and is a positive regulator of WNT signaling and negative regulator of PI3K/AKT/mTOR signaling 30,31 . The relatively high expression of these genes in MAPK-ABC tumors suggests that the WNT pathway is important in their biology.
MN1-BEND2 and MAPK supercluster tumors show characteristic mutational profiles. In addition to indel, stop gain and stop lost mutations, only missense or splice site mutations found to be "deleterious" by SIFT, "possibly or probably damaging" by PolyPhen and/or deleterious by FathmmMKL 32 (defined as FathmmMKL scores above 0.5), are considered herein unless otherwise specified (Supplementary Data 1). MN1-rearranged tumors demonstrated multiple recurrently mutated genes, including ABCC1, IGF2, ABCF1, and MSH3 (Fig. 2). ABCC1 encodes the ATPase cassette transporter protein multidrug resistance 1, a stem cell marker and MYCN transcriptional target conferring resistance to chemotherapeutics. An ABCC1-SSU72 fusion was present in case C10, and ABCC1-LUCAL3 and ABCC1-ZHHC14 fusions were found in C3 (Fig. 2, Supplementary Data 2). A heterozygous ABCC1 single nucleotide substitution (SNS), p1088P/L was present in the MN1rearanged tumor lacking MN1-BEND2 fusion from a male patient (C16). Predicted deleterious ABCC1 mutations were additionally noted in ZFTA-RELA STE and one ATAB in which ZFTA-RELA was detected by a single read (C4). ABCC1 was also relatively overexpressed in MN1-rearanged and ZFTA-RELA tumors (Fig. 1c, d and Supplementary Figs. 1, 2), suggesting they share a common ependymal stem cell lineage. The mismatch repair gene MSH3 was mutant in four of six MN1-rearranged cases.
PARP8, PTEN, ANK2, ANKFY1, and KIAA1109 non-fusion mutations were unique to non-MN1-BEND2, non-BRAF V600E tumors. KIAA1109 is important in cortical development 38 and mutant in glioblastoma 39 . ANK genes encode ankyrins, which link membrane proteins to the actin cytoskeleton and are involved in dendritic spine growth and synaptic plasticity 40 .
Mutations in CTTN (cortactin) or CTTNBPNL (cortactin-binding protein 2 N-terminal-like) occurred in non-BRAF V600E MAPK-ABC tumors. Cortactin is an actin binding protein involved in cytoskeletal remodeling and implicated in tumor invasion and WNT-dependent synaptic plasticity 41,42 . Three tumors contained VEZT mutations. Its protein product vezatin binds the cadherin-catenin complex linking adherens junctions to the actin cytoskeleton 43 .
The most frequently mutated gene in non-MN1-BEND2, non-BRAF V600E tumors was AHNAK, mutated in six such tumors, including MAPK-C tumors. AHNAK encodes neuroblast differentiation-associated protein, a cadherin-associated actin binding protein implicated in neural cell differentiation, migration, and tumor metastasis 44 .
DNA methylome analysis reveals a distinct methylation class for MAPK-C tumors. We performed genomic DNA methylation analysis of twelve AB-like tumors not previously analyzed. When these data were combined with that from our prior study 4 , tSNE analysis revealed that all MN1-BEND2 tumors clustered with the methylation class "MN1 high-grade neuroepithelial tumor (MN1-HGNET)" (Fig. 3).
MAPK-A AB tumors containing the BRAF V600E mutation again loosely grouped with the "pleomorphic xanthoastrocytoma (PXA)" methylation class 4 (Fig. 3). BRAF-wildtype MAPK-A tumors C14, C20 and C39, and an additional AB-like tumor not studied by RNAseq (C15), appeared on the periphery of the PXA methylation cluster. The three MAPK-C transcriptional cluster tumors, one MAPK-B tumor (C8) and two tumors not studied by RNAseq were among six cases (C8, C18, C19, C22, C27, and C41) forming a distinct genomic DNA methylation class, designated the Group C methylation class.
AB-like tumor histology is characterized by astroblastic pseudorosettes, and oRG/tRG-like classic astroblastoma cytomorphology in Group C tumors. It is not possible to precisely differentiate AB-like tumor genetic subtypes solely by histologic features, as tumor cytology is variable from cuboidal to tapered-columnar to spindleshaped, often within the same tumor 2,4 . In general, however, MN1-BEND2 tumor cellular morphology tends to be more cuboidal to stout-columnar and these tumors more often demonstrate vascular sclerosis, and lack eosinophilic granular material, in comparison to MAPK supercluster tumors 4 ( Fig. 4a [a-c]). MAPK-ABC tumors ( Fig. 4a [d-i]) more frequently demonstrate tapering basal processes and inconspicuous or absent apical processes, especially Group C tumors ( Fig. 4a [g-i]). The latter contained elongate monopolar cells whose endfeet contacted the brain pial surface or basal lamina of blood vessels mimicking oRG and tRG, respectively, similar to astroblastoma cells originally depicted by Bailey and Bucy 1 (Fig. 4b [a, b]). Like MAPK-AB tumors, Group C tumors also demonstrated eosinophilic granular bodies and occasional multinucleated cells 2 , as previously noted in astroblastomas by Bailey and Bucy 1 . Strikingly, all MN1-BEND2 tumors demonstrated diffuse nuclear MN1 positivity ( Fig. 4d [a-c]). Some MAPK-ABC tumors showed probable nonspecific cytoplasmic MN1 immunoreactivity; however, none exhibited diffuse MN1 nuclear immunoreactivity ( Fig. 4d [d-f]). MN1 and BEND2 immunoreactivity were also confirmed by western blot in a MN1-BEND2 case with available frozen tumor (C29) (Supplementary Fig. 4). MN1-BEND2 tumors also demonstrated strong cytoplasmic IGF2 immunoreactivity ( Fig. 4e [a-c]), which was absent or weak in non-MN1-BEND2 AB-like tumors ( Fig. 4e [d-f]).
MN1-BEND2 and MAPK-ABC tumor gene expression overlaps that of NEC and vRG, and tRG and oRG, respectively. We next  pcw when cellular proliferation in the developing brain peaks. Indeed, H19 message is enriched in human vRG as determined by single cell RNAseq 47 . The period prior to 25 pcw corresponded to the highest initial expression of NEC genes within the database, e.g., SOX1, DACH1, PROM1, NOTCH1, NOTCH2, ITGA6 and NRARP; vRG genes, e.g., H19, FOXJ1 and HES1; the tRG gene PALLD, and a few oRG genes, e.g., LIFR and SEZ6L. In contrast, genes more highly expressed or mutated in MAPK-ABC tumors, e.g., GFAP, BRINP3 and TRIM9, and certain oRG genes, e.g., FAM107A and HOPX, are more highly expressed subsequent to 25 pcw when genes related to dendritic spine and synapse development, e.g., CLU, DLG1, KALRN, SLITRK2, become more highly expressed 48 (Fig. 5). The latter period corresponds to the end of widespread cortical neurogenesis and the beginning of cerebral gliogenesis. In contrast to genes highly expressed in MN1-BEND2 tumors that tended to cluster with NEC and vRG marker genes, those selectively expressed in MAPK-ABC tumors tended to cluster with markers of vRG and oRG (Fig. 5). These data suggest that MN1-BEND2 tumors are derived from cells closely related to early vRG that maintain high expression of many NEC genes, and that MAPK-ABC tumors are derived from cells related to oRG that retain expression of some vRG markers.
In order to further explore possible neural stem/progenitor cells which AB-like tumors might share ontologic lineage, we compared published neural stem/progenitor cell gene expression sets 10 The most significant and consistent GSEA findings for NSC genes were enrichment of PTPRs for tRG genes and PXAs for oRG genes. PTPR were significantly enriched for tRG gene expression compared to controls and all other tumor types, except PXA: to fetal brain Both MAPK-ABC tumors and PXA were significantly enriched for only tRG and oRG compared to fetal brain ( Supplementary  Fig. 6Q, Supplementary Data 5). Compared to MN1-BEND2 tumors, MAPK-ABC tumors were significantly enriched only for oRG genes, but showed more frequent mutation of tRG and oRG genes ( Supplementary Fig. 7). PXA was enriched for tRG, and more so for oRG compared to MN1-BEND2 and MAPK-ABC tumors (Supplementary Fig. 6P and T, pFEWER < 0.001, Supplementary Data 5). ATAB showed no clear stem/progenitor cell gene enrichment pattern likely attributable to the heterogeneity of this group.
The MN1-BEND2 tumor markers MN1 and IGF2 are expressed by differentiating ependymal cells in the P0-P6 mouse ventricular zone. MN1-immunoreactivity was detected in the nucleus and cytoplasm of nascent immature ependymal cells of the mouse pup ventricular zone (Fig. 6e, g, l-o). These MN1positive ependymal precursors are in an intermediate stage of differentiation, referred to as the individualized stage, and exhibit centrioles and few to multiple short cilia. This stage occurs between the earliest stage of ependymal differentiation, i.e., the halo stage, involving generation of multiple pro-centrioles, and fully mature ependyma demonstrating mature centrioles and long motile cilia 51 (Fig. 6). Cytoplasmic staining for IGF2 was also observed in these cells (Fig. 6b, h-k), suggesting that MN1-BEND2 or other ependymal tumors may originate from similar immature ependymal cells derived from vRG 9,52 . IP cells basal to the OSVZ also stained for MN1, but not IGF2. Nevertheless, such cells could also theoretically give rise to MN1-BEND2 tumors. Fig. 4 Astroblastoma-like tumor subtypes exhibit characteristic histologies and genetic alterations. a AB tumor histology is characterized by astroblastic pseudorosettes, and oRG-like morphology in Group C tumors. a-i H&E-stained AB-like tumor cases. Top row MN1-BEND2 cases; Middle row BRAF V600Emutant MAPK-A cases; Bottom row Group C methylation class tumor cases. Subpial accumulations of tumor cells in C22 (g) demonstrated long basal processes extending to the pial surface and inconspicuous apical processes, morphologically resembling outer radial glial 10 . Similar cells extend basal processes to blood vessels in C27 and C19 (h-i). b Astroblastoma tumors cells within astroblastic pseudorosettes from Bailey and Bucy's 1 original cohort. a, b Tumor cells with long basal processes resembling those in Group C tumors (Plate III, Figs. 1. and 2). c Tumor cells from a case more similar in appearance to those in MAPK-AB tumors (Plate IV, Fig. 1). Mallory-Davidoff stain (a, b 850× and c, 600×) 1 (Fig. 7a, b and Supplementary  Figs. 8, 9). Examples include genes from the IGF2-H19 imprinted locus on chromosome 11p15, e.g., IGF2, H19, and miR483, and DLK1, DIO3, and miR134, which reside within the DLK1-DIO3 imprinted locus at 14q32 and whose misregulation is linked to schizophrenia and bipolar disorder 56,57 . miR483 was also found to be highly expressed in MN1-BEND2 tumors and RELA STE compared to non-MN1-BEND2 AB-like tumors independently by Nanostring methodology (Supplementary Fig. 10). GNAS, an imprinted NEC gene encoding a G-protein α-subunit activating adenylate cyclase, was also relatively overexpressed in MN1-BEND2 and some other AB-like tumors, but only in female patients ( Supplementary Figs. 7, 8). Most of these genes were also upregulated in MN1-rearranged tumors compared to normal and fetal brain (Fig. 7b and Supplementary Fig. 9).
Fusions involving imprinted and chromosome X genes occur predominantly in AB-like tumors from female patients. Fusions containing the imprinted gene maternally expressed 8 (MEG8), encoding a small nucleolar RNA, were only present in tumors from females, in four MAPK-AB lesions and four of five MN1-BEND2 tumors (Fig. 4f, Supplementary Fig. 11A, Supplementary Data 2 Fig. 11A, Supplementary Data 2).
Fusions involving imprinted genes overall occurred at similar frequency in MN1-BEND2 (2.6/case, n = 5) and MAPK-ABC tumors (2.1/case, n = 15), but were almost twofold more frequent in BRAF V600E cases (4.2/case, n = 6), and approximately tenfold  Fig. 5 MN1-BEND2 and MAPK-ABC tumor-associated genes demonstrate fetal age-specific expression and cluster with vRG and oRG marker genes, respectively. The developmental expression time course of NSC markers and select genes overexpressed or mutated in MN1-BEND2 and MAPK-ABC ABlike tumors was obtained from the Allen Human Developmental Transcriptome database and subjected to unsupervised hierarchical clustering. MN1-BEND2 tumor-associated genes (relatively overexpressed or mutated) are depicted in red and MAPK-ABC tumor associated genes are in blue. Genes associated with NEC, vRG, tRG and oRG are indicated. The x-axis age bar progressive color scheme is arbitrary. Transcript expression is normalized by reads per kilobase of transcript per million mapped reads (RPKM) to compensate for RNAseq generation of more sequencing reads from longer RNA molecules. Data is from up to 16 brain regions from 42 specimens (Allen Human Brain Atlas Developmental Transcriptome, Institute for Brain Science, available from: human.brain-map.org). The figure source data is available in the Source Data File. Considering all genes in all AB-like tumors, fusions were 4.9fold more frequent in female versus male patient tumors (Table 2). Fusions involving chromosome X were 5.3-times more common in tumors from females (P = 0.017) and occurred almost entirely in MN1-BEND2 and BRAF V600E -mutant tumors. Fusions  EIF2S3  USP9X  RPS4X  PLS3  SH3BGRL  COX7B  GPM6B  ATP6AP2  TCEAL9  SYAP1  KDM6A  CTPS2  PHEX  IL3RA  VSIG4  CSF2RA  ZFX  XIST  JPX  FHL1  HCFC1  FTX  RBM3  SMC1A  KDM5C  DDX3X  UBA1  ZRSR2  PRKX  TBL1X  CD99P1  ZBED1  CD99  CDK16  TIMP1  RBBP7  SLC25A6  MXRA5  GYG2  BEND2 Table 2). *Statistically significant p < 0.05. X-inactivation escape genes are more widely expressed by and mutant in MAPK supercluster AB-like tumors. We also compiled a list of 95 XIE genes and 105 genes exhibiting variable XIE status (Supplementary Data 6) 56 .
Expression of AB-like tumor genes correlate with gene promoter methylation. In order to confirm that differential expression of AB-like tumor genes can be, in some cases, attributable to differential promoter methylation, we integrated RNAseq expression and DNA methylation data using the mCSEA R package 59 , which correlates differentially methylated regions (DMRs) with expression of nearby downstream genes. The expression of several MN1-BEND2 overexpressed genes within the IGF2-H19 and DIO3-DLK1 imprinted regions and other MN1-BEND2 and MAPK-ABC overexpressed genes significantly correlated with their promoter methylation status (Supplementary Fig. 14).
BEND2's function is unknown however BEN domaincontaining homologs participate in chromatin remodeling and transcriptional repression 66,67 . As a putative chromatin remodeling protein, BEND2 dysregulation could potentially influence gene expression. We identified MN1-BEND2 fusion transcripts in five tumors from females, all over-expressing MN1 and BEND2 mRNA. BEND2 overexpression has also been documented in an additional MN1-BEND2 tumor 68 . An MN1-rearranged tumor from a male patient in our series, lacking MN1-BEND2 (or other MN1 or BEND2 fusions), did not overexpress MN1, but did overexpress BEND2. It clustered with MN1-BEND2 tumors by RNA expression, but not DNA methylation.
Of eight recently reported AB-like tumors demonstrating EWSR1-BEND2 fusions (Supplementary Data 8), one had documented high BEND2 expression 69 . Most were infratentorial, Fig. 7 MN1-BEND2 AB-like tumors exhibit relative overexpression of imprinted genes while MAPK-ABC tumors show relative overexpression of X-inactivation escape genes. a Heatmap of unsupervised hierarchical clustering of RNAseq expression data for a subset of imprinted genes. b Similar data from the Affymetrix analysis including controls and PXA. c Heatmaps for expression of subsets of XIE genes from the Affymetrix data. IGF2 is absent in the NCBI annotation for the HTA 2.0 array as of April 2021. IGF2-AS (arrow) may serve as a surrogate marker for IGF2 and was upregulated in MN1-rearranged tumors and normal choroid plexus (choroid plxs). The latter develops as an out-pouching of the ependyma and is known to express IGF2 128 . The XIST gene message is indicated by the arrowhead. MN1-rearranged tumors are highlighted in red, MAPK-AB in blue, and MAPK-C in purple. Complete source data for heatmaps are available as Supplementary Fig. 8-9   involved the medulla or spinal cord and some clustered near HGNET-MN1 by tSNE analysis of DNA methylation 70 . All but one in these anatomic regions occurred in males. Two frontal lobe cases were in females. EWSR1 is located at chromosome 22q12.2 near MN1, thus both translocations involve the same 22q12 general breakpoint region. BEND2 overexpression may thus be significant in the tumorigenesis of MN1-BEND2 and similar tumors. Indeed, BEND2 appears to be normally expressed in the brain only at very low levels (Fig. 5). All reported tumors with confirmed MN1-BEND2, and available demographic data, occurred in females aged 4-18 years (mean 10.4 yr, n = 13) (Supplementary Data 8). Tumors occurred in the frontal, parietal or occipital lobe, predominantly the parietal lobe, but not the temporal lobe or infratentorial as seen in non-MN1-BEND2 AB-like tumors (Fig. 8a, Table 1, Supplementary Data 8). In contrast, the mean age of presentation of BRAF V600Emutant AB tumors in our series was 24.7 years (n = 6), and 29.1 years for MAPK-ABC cases (n = 15). It, therefore, appears that AB-like tumors related to earlier developmental stages of NSCs (RG), i.e., MN1-BEND2 tumors, clinically present earlier in life compared to tumors transcriptionally related to later RG, e.g., MAPK-ABC ABs and PTPR 17 .
The transition of embryonic stem cells to NSCs is accompanied by changes in DNA methylation. Additional methylation changes allow expression of glia-specific genes during the mid-gestational shift from primarily neurogenesis to gliogenesis 28,71,72 . Further epigenetic changes occur during increased transcriptional activity and synaptic remodeling in childhood through young adulthood 73,74 . DNA methylation and gene expression patterns in MN1-BEND2 and MAPK-ABC tumors may reflect these developmental dynamics and partially explain their different ages of clinical presentation. Dysregulation or mutation of chromatin remodeling genes, e.g., MN1, BEND2, ATRX, CXXC5, CREBBP, and SMARCAD1; histone modifier genes, e.g., CHD3, KMT2C; and/or possibly X-inactivation effectors could alter DNA methylation and potentially contribute to AB tumorigenesis. The terms early-, mid-, and late-RG have been used in model systems 28,75 . Ziller et al. 28 defined mid-RG as corresponding to peak neurogenesis and early gliogenesis, and late-RG to mostly gliogenesis. Like vRG, early-RG expressed PAX6 and FOXJ1. Late-RG highly expressed OLIG2 and TCF4, similar to oRG and MAPK-ABC ABs.
MN1-BEND2 tumors exhibit high expression of several NEC and vRG markers and low expression of the oRG markers FAM107A, LIFR, and HOPX ( Supplementary Fig. 5A, B, D). Unlike MN1-BEND2 tumors, MAPK-ABC tumors and tRG are enriched for CRYAB 12 (Supplementary Fig. 5C). The latter is more highly expressed after 25 pcw, consistent with tRG expressing a later transcriptional program than vRG (Fig. 5).
Our data suggest that MN1-BEND2 tumors ultimately originate from early vRG-derived ependymal progenitors or a residual population of related cells. Such cells may become neoplastic due to epigenetic and/or genetic events in utero or during the dramatic growth of the brain during childhood. The correlation of MN1-BEND2 tumor transcriptomes with early NSC types (NEC and vRG) and MAPK-ABC tumor, PTPR, and PXA transcriptomes with later NSC types (tRG and oRG) suggests that mutagenic events in early brain development may manifest as tumors in both childhood and adulthood, or that some RG or RG-derived progenitors persist beyond adolescence.
NEC and vRG project an apical primary cilium into the lateral ventricle, facilitating IGF2 binding from the CSF 76 . Unlike adult rodent lateral ventricle NSCs (B1 astrocytes), NSCs in the adult rodent dentate gyrus express IGF2, which is required for their proliferation in an apparent autocrine fashion 73,77 . This regional difference in IGF2 expression was attributed to differential imprinting 73 . We observed IGF2 immunoreactivity in immature ependymal cells of the fetal mouse lateral ventricular zone. Thus, IGF2 produced by differentiating ependymal cells may possibly be a critical pro-proliferative signal in the fetal lateral ventricular zone. This seems plausible considering the requirement of the ependymal layer by B1 astrocytes 78 . Alternatively, this additional IGF2 may be needed by the developing ependyma themselves or other nearby cells.
MAPK supercluster tumors demonstrated high expression, mutation, and fusion of genes reflecting glial and neural differentiation. GSEA supports that these tumors express transcriptomes reflective of tRG and oRG. PXA was more highly enriched for oRG genes compared to tRG genes. The mixed pattern of tRG and oRG gene expression by MAPK-ABC and PXA might be explained by the possibility that gene expression of late RG types and/or their progeny IPs might vary with development over time, and that the published tRG and oRG gene sets represent developmental timeline-dependent snapshots.
MAPK-ABC tumors exhibit overexpression and mutation of genes involved in neurocognitive disorders characteristic of children and young adults. TCF4 is implicated in schizophrenia and Pitt-Hopkins autism-like syndrome 79,80 . ANK3 has been associated with autism, bipolar disorder, and schizophrenia 81 . KALRN, TRIO, and DISC1 misexpression or mutation are implicated in attention deficit hyperactivity disorder, autism, bipolar disorder, schizophrenia, and other conditions [82][83][84][85] . KALRN protein (kalirin) isoforms show increased expression in adolescence and early adulthood when schizophrenia and many MAPK supercluster AB-like tumors occur 86,87 .
Group C and other non-MN1-BEND2/non-BRAF V600E -mutant AB-like tumors exhibit AHNAK and SYNE1 mutations. Although AHNAK has been characterized as a tumor suppressor 88 , AHNAK and SYNE1 have also been implicated as cancer genes 89 . AHNAK is necessary for pseudopod formation, a key step in the epithelial to mesenchymal transition and tumor cell invasion 44 . SYNE1 encodes spectrin repeat-containing nuclear envelope family member 1, which is involved in nucleolemmal cytoskeletal anchoring 23,90 .
An intriguing aspect of MN1-BEND2 and MAPK-AB tumors is their striking female predominance of up to 9:1 or greater 4 . We previously observed chromosome X gains in seven AB-like tumors from female patients and only one in a male patient; X loss was seen in four tumors from females and none from males 4 . Here, fusions involving chromosome X were over five times more common in female patient tumors, and almost entirely in MN1-BEND2 and BRAF V600E -mutant tumors. Chromosome 22 fusions were even more frequent in tumors from females. Meiotic recombination of chromosome 22 is greater in females than males 91 . How this relates to non-homologous mitotic recombination is unclear.
High expression of imprinted genes in MN1-BEND2 tumors and XIE genes in MAPK-ABC tumors suggest a potential role for differential DNA methylation in driving their tumorigenesis. BEND2 itself exhibits variable XIE status 15 . XIE genes can be activated by epigenetic changes in chromatin structure and are frequently mutated in cancer, especially female predominant cancers 92,93 .
The imprinted homeobox gene DLX5 may escape imprinting and be biallelically expressed in mice and patients with the nearly entirely female disease Rett syndrome 94,95 . The latter is characterized by mutations in chromosome X gene MECP2 encoding methyl-CpG binding protein 2, which binds methylated DNA and suppresses transcription. We found frequent overexpression of imprinted genes DIO3, IGF2, H19, and others, correlating with decreased methylation of their promoters in MN1-BEND2 tumors. Our data suggest that epigenetic and genetic alterations associated with AB-like tumors possibly involves loss of imprinting, leading to increased expression and sometimes subsequent mutation of affected genes 96,97 . Loss of imprinting is linked to several cancers and pediatric syndromes, e.g., Beckwith-Wiedemann syndrome involving alteration of the IGF2 imprinting control region and increased risk for Wilms tumor and other neoplasms 98 . High expression of imprinted genes in MN1-BEND2 tumors could alternatively be due to retention of fetal methylation patterns. Interestingly, DNA methylation and expression of imprinted genes in the brain is not only developmentally time dependent, but is also anatomic region dependent [99][100][101] , which may possibly relate to anatomic locations in which AB-like tumors occur (Fig. 8a).
We provide multiple lines of evidence that MN1-BEND2 and MAPK-ABC astroblastomas are of radial glia lineage and likely derived from vRG, and oRG and/or tRG or their close progeny, respectively (Fig. 8b, c). This interpretation is limited however by the mostly correlative and non-mechanistic nature of the data. For instance, the mostly adult glioma, glioblastoma, is thought to originate from subventricular zone B1 astrocytes 102 . We cannot completely rule out that some AB-like tumors may be derived from B1 astrocytes, as the transcriptome of B1 astrocytes is largely undefined.
PTPR has been hypothesized to be derived from the subcomissural organ (SCO) located in the third ventricle. GSEA data strongly suggest that PTPR are closely related to tRG and thus derived from them. tRG resemble an incompletely defined RGlike ventricular zone "ependymoglial" cell known as the tanycyte, which may also project a single cilium into the ventricle and an elongated basal process to a subependymal capillary 103 . Such cells are notably found in various third ventricle locations including the SCO. Tanycytes may thus represent persistent tRG capable of generating PTPR. Notably, PTPR strongly express PAX6, which is expressed by tRG 12 . It has also been proposed that ABs may be derived from tanycytes 104 , potentially consistent with our findings of partial enrichment of tRG genes in MAPK-ABC tumors.
Recently, tumor stem cells expressing the oRG marker PTPRZ1 capable of MST were found in glioblastoma 105 . This supports the possibility that residual oRG or oRG-derived IP cells persist into adulthood and lead to tumor formation. Invasive properties of these oRG-like tumor stem cells 105 may explain why oRG-related MAPK supercluster ABs have a poorer prognosis than MN1-BEND2 tumors 4 . It may also be that B1 astrocytes or other glioblastoma initiating cells can differentiate to attain oRG-like properties. The relevance of either possibility in adult gliomagenesis remains to be determined.
Because MN1-BEND2 containing AB-like tumors are generally associated with excellent patient survival 4 , they should not be termed high-grade neuroepithelial tumors 6 . Because MN1-BEND2 tumors are most likely derived from cells destined to become ependymal cells, they also should not be called astroblastomas, but rather early ependymal tumors with MN1-BEND2 fusion (EET MN1-BEND2). Of practical clinical diagnostic importance, we show that MN1-BEND2 tumors can be differentiated from other AB-like tumors by their strong diffuse MN1 nuclear immunoreactivity and absent diffuse GFAP staining. We recommend that MN1 and GFAP immunohistochemistry, DNA methylation analysis, and confirmation of the MN1-BEND2 fusion (and if negative, MN1-CXXC5 or EWSR1-BEND2 fusions) by FISH, RT-PCR, or sequencing comprise an integrated approach to the diagnosis of EET MN1-BEND2 and other ABlike tumors.
As oRG, and likely tRG, give rise to IPs destined to become neurons or glia, the coinage of "astroblastoma" by Bailey and Cushing 106 , indicating derivation from an astrocyte progenitor, was amazingly insightful. We propose that the term astroblastoma be reserved for MAPK supercluster AB-like lesions, particularly DNA methylation Group C tumors, which may exhibit classic astroblastoma histomorphology.
Lastly, our data identify the IGF2 pathway and the multidrug transporter ABCC1 as candidate therapeutic targets in EET MN1-BEND2 107,108 . While BRAF and downstream MEK and/or PI3K/ AKT/mTOR pathway inhibitors may be useful in treating MAPK-ABC astroblastomas 109,110 .

Methods
Tumors. Construction of our clinical cohort of histologically-defined AB-like cases has been described 2 . The cohort was augmented with three additional AB cases with available material for further study (C39, [43][44] and an additional five tumors with AB-like histological features, for a total of 35 AB-like cases in the current cohort. All control brain tissue was retrospectively obtained from archived material from routine hospital autopsies. Equivalent fetal age control brain tissue was obtained from autopsies of spontaneous abortuses or deceased premature infants at Nationwide Children's Hospital, Columbus, Ohio. Autopsies were requested by patient's families or their physicians for medical reasons and performed with the consent of the patient's next of kin, completely independent of the studies described herein. Informed consent for the use of archival biopsy or autopsy tissue was not obtained, nor required by the institutional review board. No patients or patient families were compensated for these studies. Institutional review board approval from the University of Louisville was obtained for use of control and case material (Approval number 17.0984). MN1-rearrangement was determined by FISH (cases C3, C7, C10, C13, C16, C21, and C29 5 ) and confirmed by CNV analysis using the Illumina EPIC BeadChip 850K microarray data from these and two additional samples (C43 and C44).
The tSNE plot was generated using the 32,000 most variably methylated probes of the study samples and 2801 German Cancer Center (DKFZ) reference samples as previously described 111 . The later consists of 2729 tumors (comprising 76 pathological diagnoses) and 72 normal brain tissues.
RNA isolation. Total RNA was extracted from three 30 μm scrolls cut from FFPE tissue blocks containing 90% tumor or more, using the RecoverAll total nucleic acid isolation kit (Ambion/Thermo Fisher Scientific, Waltham, MA) per the manufacturer's instructions. RNA was then passed through concentration and clean-up columns (Norgen Biotek, Thorold, ON) and quantified using a Nanodrop spectrophotometer (Thermo Fisher Scientific, Waltham, MA).
RNAseq. Capture-based RNAseq was performed using a modified SureSelect strand-specific RNA library prep and target enrichment protocol (Agilent Technologies, Santa Clara, CA). Briefly, 200 ng of FFPE-derived total RNA was heated to 65°C for 5 min and used for first-strand cDNA synthesis. Second-strand cDNA was synthesized, and the ends repaired. The 3ʹ ends of ds-cDNA were adenylated, ligated with adapters and fragments were PCR-amplified. PCR product (700 ng) was hybridized overnight to SureSelect XT Human All Exon V5 + lncRNA baits. The hybridized fragments were captured on magnetic beads and PCR-amplified using uniquely barcoded index primers. Barcoded libraries were pooled and underwent paired-end 150 bp sequencing on an Illumina HiSeq4000 instrument (Illumina, San Diego, CA).
For gene expression, read alignment was performed with hisat2-2.0.5 113 to human genome GRCh38 and aligned files were converted to bam and sorted with samtools v1.3.1 114 . Alignment QC was generated with RSeQC v2.6.4 115 . Counts from the subread package v1.5.1 116 were used to count the number of reads aligning to Ensembl IDs according to GRCh38.86 annotation.
Fusion detection was performed using JAFFA 117 after merging across replicate lanes. Variant calling was performed following GATK's best practices for variant calling in RNAseq using a 2-pass STAR alignment. Alignments were made with STAR v2.5.2a 118 to human reference genome GRCh38. The junctions detected in all samples were merged and used to make a custom STAR genome index for a second pass alignment. Picard v2.4.1 was used to preprocess the alignment files, add read groups, mark duplicates, sort, and index. GATK v3.6 119 was used to split reads containing junctions, realign around indels, and recalibrate the base qualities. Samtools v1.3.1 was used to generate mpilup format. BRAF mutations were identified by a visual inspection of the alignment BAM in the Integrated Genomic Viewer 120 since they were not detected by VarScan analysis. VarScan v2.4.1 121 was used to call and filter variants from mpilup format from the whole genome. Called variants were annotated with Ensembl Variant Effect Predictor (VEP) 122 . Variants were eliminated if listed in dbSNP alone. Variants with MEDIUM or HIGH impact as designated by VEP were retained. After filtering variants detected by the same set of reads were eliminated. PolyPhen, SIFT, and FathmmMKL predictions were determined using respective versions available July 2018.
cDNA microarrays. Affymetrix GeneChip Human Transcriptome Array 2.0 (HTA-2) microarray analysis was performed according to the manufacturer's instructions using the GeneChip WT Pico and Sensation Plus sample preparation kits (Thermo Fisher Scientific). In brief, 50 ng of total RNA was converted to single-stranded cDNA with T7 promoter sequence at the 5′ end. A 3′ Adaptor, as a template for double-stranded cDNA synthesis, was added and the single-stranded cDNA was converted to double-stranded cDNA. Six amplification cycles were applied as recommended by the protocol to enrich the samples. The amplified product was converted to cRNA and purified using purification binding beads. Eluted cRNA yield was measured by NanoDrop. cRNA (20 μg) were reverse transcribed to sense-strand cDNA containing dUTP at a fixed ratio relative to dTTP. The cRNA was removed using RNase H and the 2nd-cycle single-stranded cDNA was purified using purification binding beads. The eluted single-stranded cDNA yield was determined using the NanoDrop and 5.5 μg was fragmented by uracil-DNA glycosylase and apurinic/apyrimidinic endonuclease 1 (APE 1). The fragmented ss-cDNA was labeled by terminal deoxynucleotidyl transferase, and then used to prepare the Hybridization Cocktail, injected and hybridized in the array for 16 h at 45°C in an oven with rotation at 60 rpm. After overnight hybridization, the hybridization cocktail mix was extracted from the array. Each array was washed and stained using Affymetrix GeneChip Fluidics Station 450 according to the user manual (P/N 08-0295) and the arrays were scanned using GeneChip Scanner 7G. Each array image was collected using the GeneChip Command Console, QC, saved as.cel files and saved for further analysis.
RNA expression analysis, dimensionality reduction, and clustering. For the RNAseq data, transcripts with 0 counts > 50% were removed. Data were normalized with Limma-voom R package with quantile normalization 123 . The Ensembl IDs were mapped to unique gene symbols using the Bioconductor org.Hs.eg.db package (ver: 3.13.0) using median summarization.
For the microarray data, probe-level data were generated by analyzing the raw signal intensities from the raw Affymetrix CEL data files for each probe set using the rma algorithm for background intensity correction and normalization in the oligo package (ver: 1.56). Batch effect related to the sample preparation kit was reduced using the combat function of the sva package (ver: 3.4). The probes were mapped to unique gene symbols using the hta20transcriptcluster.db package (ver: 8.8) with median summarization.
Differential gene expression between groups of samples in RNAseq and Affymetrix datasets was performed using analysis of variance thresholding with a two-sided p-value to 0.05. Expression values of the genes differentially expressed in both datasets were scaled and denoted on a heatmap with the respective samples. Heatmaps in Figs. 1, 5, 8, S1-S3, S7, S8, and S11, S12 were generated by unsupervised hierarchical clustering.
Dimensionality reduction of the expression data with t-Distributed Stochastic Neighbor Embedding (tSNE) was performed to facilitate visualization using the following parameters (dimensions = 2, perplexity = 4, iterations = 5000) using the Rtsne package (ver: 0.15). All analyses were performed using R software (ver: 4.1).
Differential gene promoter methylation and expression correlation. Raw methylation files were read and normalized using the sesame package (ver: 1.10.4) with the default settings. Differentially methylated gene promoters were identified using the mCSEA package (ver: 1.12). Correlation between methylation of gene promoters and expression levels was quantified using the mCSEAIntegrate function. This correlates differentially methylated regions (DMRs) with the expression of nearby downstream genes 59 .
Nanostring MicroRNA. MicroRNA (miR) profiling of tumor samples was performed using the Nanostring nCounter system (Nanostring Technologies, Seattle, WA) 126 . One hundred nanograms of total RNA was annealed with multiplexed DNA tags (miR-tags) complementary to specific target miR sequences. Mature miRs were bound to specific miR-tags using a ligase and excess tags were enzymatically removed. Tagged miRs were diluted 1-5, and 5 µl was combined with 5 µl of capture probes and 20 µl of reporter probes in hybridization buffer. Probes were allowed to complex with specific target sequences overnight (16-20 h) at 65°C. Excess probe was removed using a two-step magnetic bead-based purification (nCounter Prep Station) and target/probe complexes were immobilized onto cartridges. Reporter probe counts were obtained for each sample using the Gen1 nCounter Digital Analyzer, which utilizes a CCD camera-equipped microscope to image the immobilized fluorescent reporters via a high-density scan encompassing 600 fields of view. Raw data output was imported into nSolver for normalization (http://www.nanostring.com/products/nSolver). For analysis, data are the geometric mean normalized and then log10. For heatmap generation, data were scaled and truncated as follows. For scaling, the data are mean centered and divided by the SD of the miR. Truncation was done at a value of 4.0 to adjust for rare very large outlier values in the matrix.
Animals. Animal studies were performed in accordance with the guidelines of the European Community and French Ministry of Agriculture and were approved by the "Direction deṕartementale de la protection des populations -Paris (authorization 9343201702211706561). Mice used in this study are Centrin2-GFP (CB6-Tg(CAG-EGFP/CETN2)3-4Jgg/J; The Jackson Laboratory). Animals were grouphoused with free access to food and water in controlled temperature conditions (room temperature controlled at 21-22°C, humidity between 40 and 50%), and exposed to a conventional 12-h light/dark cycle. Experiments were performed on embryos at embryonic (E) days E16, and neonate pups at postnatal day 0-10 (P0-10) of both sexes. All procedures were approved by the Ethical Committee CEEA-005 Charles Darwin (authorization 9343-201702211706561) and conducted in accordance with EU Directive 2010/63/EU. All efforts were made to reduce animal suffering and minimize the number of animals, in compliance with all relevant ethical regulations for animal testing and research.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.