Cross-Species Genomics Reveals Oncogenic Dependencies in ZFTA/C11orf95 Fusion–Positive Supratentorial Ependymomas

Ependymomas (EPN) are neuroepithelial malignancies of the central nervous system (CNS), accounting for 5% of all CNS tumors in children (1). The utility of histologic grading of EPN for risk stratification has been discussed controversially, with no consistent associations of tumor grade and patient outcome (2). However, recent genomic studies have allowed for subdivision of supratentorial (ST), posterior fossa (PF), and spinal (SP) EPN into molecularly distinct groups with clearly distinct clinical features and outcome (3–9). Despite these advances, translation into novel treatment approaches is lagging behind. The mainstay of treatment for almost all EPN remains surgery and radiotherapy, whereas chemotherapy has mostly been found to be ineffective (10, 11).

Within the ST CNS compartment, underlying molecular signatures including DNA methylation and transcriptome analysis define three major subgroups, designated ST-SE, ST-EPN-YAP1, and ST-EPN-RELA (6). ST-SE are fusion-negative molecularly classified subependymoma that are mostly observed in adults. ST-EPN-YAP1 tumors are enriched for gene fusions involving the Hippo effector YAP1 and primarily affect infants. The vast majority of ST-EPN is classified best as ST-EPN-RELA, and the tumors predominantly contain oncogenic fusions between RELA, the principal effector of canonical NFκB signaling, and C11orf95, a less well-characterized neighboring gene on chromosome (chr.) 11 (6, 7, 12).

Apart from chromothriptic events on chr. 11 surrounding the fusion, the genome of human ST-EPN-RELA is generally stable, and additional recurrent alterations other than focal CDKN2A/B deletions have not yet been identified (6). The hypothesis of a single-hit oncogenic event is supported by the fact that the C11orf95–RELA fusion is sufficient to drive tumor formation in vivo using the RCAS/tv-a system (13). In ST-EPN-YAP1, MAMLD1 (the most frequent fusion partner to YAP1) was found to mediate fusion-driven oncogenic transformation of cortical neural progenitors through nuclear translocation and interaction with nuclear factor 1 proteins (14). The role of the RELA fusion partner C11orf95 in ST-EPN-RELA, however, is not yet fully understood. Results from previous studies show that overexpression of wild-type RELA or induction of activating RELA mutations is not sufficient for oncogenesis, despite leading to elevated levels of NFκB target genes. This indicates that both the Rel-homology domain of RELA and the C11orf95 partner gene are critical for ependymal tumorigenesis (7, 13).

Although ST-EPN are mostly unambiguously assigned to a molecular group by the Heidelberg Brain Tumor Methylation Classifier (www.molecularneuropathology.org; ref. 15), one of the following three diagnostically challenging constellations may occur in about 20% of all ST-EPN cases (12, 16–19): (i) prediction as ST-EPN-RELA by DNA methylation–based tumor classification but without evidence for a canonical C11orf95–RELA fusion, (ii) the vice-versa combination with a typical fusion event in the absence of a reliable ST-EPN-RELA score, or (iii) ST tumors histologically diagnosed as EPN that cannot be readily assigned to any of the existing molecular classes. In this study, we aimed to molecularly characterize RELA- and YAP1-fusion–negative ST-EPN tumors, which exist in addition to the known fusion-negative ST-SE group. We demonstrate the existence of both ST-EPN harboring C11orf95 fusions to gene partners other than RELA, and ST tumors that are histologically distinct from EPN but harbor a canonical C11orf95–RELA fusion. In addition, we show that newly identified C11orf95-associated fusions possess transforming capacity in vivo. This study paves the way for a refined molecular classification of ST-EPN in the future, provides representative mouse models, and presents a rationale for preclinical studies aiming at blocking central molecular dependencies and target genes (e.g., GLI2) that are shared by tumors driven by C11orf95-containing fusion genes independent of histologic appearance. On the basis of our findings together with two cosubmitted studies in this issue by Kupp and colleagues and Arabzade and colleagues (20, 21), C11orf95 has now been officially designated zinc finger translocation associated (ZFTA) by the HUGO Gene Nomenclature Committee.

To identify molecular group assignment of diagnostically challenging ST-EPN, we included 23 samples fulfilling any of the conditions (i–iii) in an unbiased clustering approach with a comprehensive data set of DNA methylation profiles covering the entire spectrum of existing molecular CNS tumor classes (>70,000 DNA methylation profiles; Supplementary Figs. S1 and S2A). Exploratory samples were mainly clustering in smaller satellite clusters next to (18/23) or within the ST-EPN-RELA cluster (2/23); 3 of 23 were not assigned to any cluster and excluded from further analysis (Supplementary Fig. S2A). Next, we restricted the clustering approach to cases from established molecular EPN groups and satellites only, revealing four distinct additional clusters (Fig. 1A). Cluster stability was confirmed through a hierarchical density–based clustering scan (HDBSCAN; Fig. 1B; Supplementary Fig. S2B and S2C; Supplementary Table S1). Tumors from the ST-EPN-RELA cluster (445/492; 90.4%) and cluster 1 (9/9; 100%) predominantly reached a calibrated score for reliable ST-EPN-RELA methylation group assignment by the Heidelberg Brain Tumor Methylation Classifier, score ≥0.9 (= red dots; classifier version 11b4; Fig. 1B; ref. 15). In cluster 2, only 22 of 43 (51.1%) reached a calibrated score ≥0.9 for ST-EPN-RELA. All samples of the two remaining clusters 3 and 4 (n = 17 and n = 27) were found unclassifiable with calibrated scores <0.9 for any methylation class (= black dots; Fig. 1B; Supplementary Table S1). Control for potential confounding effects through tumor purity, array or tissue type, and probe detection quality did not reveal significant impact by any of these factors (Supplementary Fig. S3A–S3F). Copy-number alterations (CNA) previously described for ST-EPN-RELA, such as loss of CDKN2A, or chromothriptic events on chr. 11 were found to different extents. Loss of chr. 22 was associated with all clusters (Fig. 1C; Supplementary Table S2; refs. 6, 7). Notably, CNAs were calculated from Illumina 450k or Illumina 850k/EPIC arrays, and chromothriptic events may have been called more frequently applying whole-genome sequencing. Taken together, these data demonstrate that diagnostically challenging ST-EPN (exploratory samples), that is, unclassifiable tumors based on DNA methylation or without evidence for a typical fusion, mainly fall into discrete DNA methylation clusters distinct from ST-EPN-RELA and only partly share structural variations with these.

Next, we further investigated molecular characteristics of satellite clusters 1 to 4 as compared with ST-EPN-RELA. To this end, we performed either RNA sequencing (RNA-seq) or DNA panel sequencing and also incorporated previously generated data (RNA-seq and RT-PCR), resulting in comprehensive fusion gene information from clusters 1–4 (n = 48) and ST-EPN-RELA samples (n = 70; refs. 6, 22; Supplementary Table S1). Sequencing data revealed a previously unrecognized ZFTA–RELA fusion type in cluster 1 (n = 2/2), designated fusion type 8 (ref. 7; Fig. 2A and B). All samples from cluster 3 subjected to DNA panel sequencing (n = 5/13), RNA-seq (n = 6/13), or both methods (n = 2/13) harbored a ZFTA–RELA (type 1, 2, or 3) fusion (Fig. 2A; Supplementary Table S1; ref. 7). Notably, samples from cluster 3 were invariably unclassifiable (ST-EPN-RELA score < 0.9; Fig. 1B), indicating that canonical ZFTA–RELA fusions are present in a significant subset of ST-EPN that are molecularly distinct from classic ST-EPN-RELA based on DNA methylation profiling. Tumors from clusters 2 and 4 contained fusions to ZFTA but without involvement of RELA. Alternative ZFTA fusion partners included MAML2 (n = 15), MAML3 (n = 2), NCOA1 (n = 2), NCOA2 (n = 9), and CTNNA2 (n = 1; Fig. 2A–C; Supplementary Table S1). Within remaining samples of the satellite clusters 2 and 4, no fusion (2/33) or fusions without involvement of ZFTA (2/33) were detected, respectively (Fig. 2A; Supplementary Table S1).

ZFTA–RELA fusions were identified in the majority of samples from the established ST-EPN-RELA group (63/70; 90.0%; Supplementary Table S1). Within ST-EPN-RELA we observed a tendency of samples to cluster according to their respective fusion type (Supplementary Fig. S4A). Occasionally, analyses of samples from ST-EPN-RELA revealed more complex rearrangements including alternative fusions, such as ZFTA–SS18 (n = 1/70), SYVN1–MAJIN (n = 1/70), and RELA–MACROD1–ZFTA (n = 1/70; Supplementary Fig. S4B). In 18 cases belonging to ST-EPN-RELA (n = 14), cluster 1 (n = 3), and cluster 3 (n = 1), more than one genetic fusion harboring either RELA or ZFTA was detected. However, all of these showed low confidence scores and may represent nonfunctional by-products of structural rearrangements on chr. 11(6). For all samples from the satellite clusters (6/48) and for ZFTA–RELA-negative cases from the ST-EPN-RELA cluster (2/4) with sufficient material, alternative rearrangements were validated by RT-PCR followed by Sanger sequencing. In-frame fusion transcripts were confirmed in all cases (Supplementary Fig. S4C; Supplementary Table S1).

To investigate whether (epi)genetically defined clusters of diagnostically ambiguous ST-EPN also show transcriptional differences, we further analyzed available expression profiles (n = 66). Unsupervised hierarchical clustering analysis recapitulated clusters 1–4 derived from methylation data (Supplementary Fig. S5A). Exploration of differentially expressed genes (DEG) and gene ontology (GO) analysis comparing ST-EPN-RELA and clusters 2, 3, and 4 [formalin-fixed, paraffin-embedded (FFPE)–derived RNA-seq data were available only for n = 1 in cluster 1] revealed differences in expression patterns and activated signaling pathways (Supplementary Fig. S5B and S5C). RELA was significantly upregulated in ST-EPN-RELA only. Previously described gene networks for ST-EPN-RELA, such as MAPK signaling and synapse organization, were found predominantly active in this group (6). Cluster 2 showed upregulation of metabolic pathways, especially with regard to amino acid metabolism, while clusters 3 and 4 revealed activation of neuroendocrine signaling (Supplementary Fig. S5C). Integration with DEGs from a previously published data set on signature genes in EPN and with (super)enhancer-regulated genes in ST-EPN-RELA identified only very few shared DEGs with the new clusters but strong overlap with the ST-EPN-RELA group from this study (refs. 6, 22; Supplementary Fig. S5D–S5G). A substantial number of (super)enhancers specific to ST-EPN-RELA were also associated with clusters 2 and 3, implying a shared set of genes regulated by ZFTA-associated fusion proteins (Supplementary Fig. S5D–S5G; Supplementary Table S3).

Tumors harboring alternative rearrangements, or in the case of cluster 3 also canonical ZFTA–RELA fusions, exhibited a broad spectrum of institutionally diagnosed high-grade and undifferentiated histologies including characteristics reminiscent of sarcoma, diffuse high-grade glioma, CNS embryonal tumors, and other primitive tumors (Fig. 3A–E; Supplementary Figs. S6, S7A–S7J, S8A–S8F; Supplementary Table S4). In addition, a ZFTA–RELA fusion was detected in a case primarily diagnosed as a centrally located malignant peripheral nerve sheath tumor as well as in a tumor that histologically appeared as “astroblastoma.” L1CAM, a characteristic histopathologic marker for ST-EPN-RELA (23, 24), and nuclear p65, the protein encoded by RELA, could be evaluated by IHC in a limited number of samples from clusters 2, 3, and 4 as well as in ZFTA–RELA-negative and ZFTA–RELA type 8 cases. Apart from one sample in cluster 2, all cases expressed L1CAM to different extents, demonstrating that L1CAM expression is not restricted to tumors harboring canonical ZFTA–RELA fusions (Fig. 3F–I; Supplementary Table S4). IHC stainings for p65 were negative in cluster 2 (0/3) and in one sample from cluster 3 harboring a ZFTA–RELA fusion (1/4), and positive in cases across cluster 1 (n = 1), cluster 3 (n = 3), and cluster 4 (n = 1; Fig. 3J–M; Supplementary Table S4). Although ZFTA–RELA fusions were detected in five of six p65-positive cases, the p65-positive case from cluster 4 harbored a ZFTA–CTNNA2 fusion, demonstrating that p65 positivity is not confined to ZFTA–RELA fusions. Upon central histopathologic review of available tumors (n = 25), most cases were found to be at least compatible with variants of highly dedifferentiated EPN (Supplementary Table S4).

Collectively, these results suggest that ZFTA is an integrally promiscuous partner within potentially oncogenic fusion genes that drive transcriptionally distinct ST-EPN including cases with atypical histologic characteristics.

The ZFTA–RELA fusion gene has been shown to drive tumor formation when delivered to neonatal mouse forebrain cells positive for either NESTIN, GFAP, or BLBP using the RCAS/tv-a system (13), suggesting that ST-EPN-RELA formation may result from single-hit oncogenesis in cells at an early stage during development. This prompted us to test whether the respective fusions detected in clusters 1–4 are sufficient to cause tumor formation as well. To investigate this, recurrently identified fusion genes encoding ZFTA fused to RELA, MAML2, MAML3, and NCOA2 were inserted into the pT2K-Luciferase-based expression vector flanked by Tol2 cis elements. Genomic integration of the fusion genes into cells of the cortical ventricular zone was achieved byin utero electroporation-based transfection with coexpression of the Tol2 transposase (T2TP) at embryonic day 13.5 (E13.5; Fig. 4A). According to our previous study (14), electroporation of ZFTA–RELA and YAP1–MAMLD1 alone induced tumor formation within the cerebral cortex with a median survival of 44 and 29.5 days (n = 11/11 for ZFTA–RELA and n = 30/30 for YAP1–MAMLD1), whereas no tumors were developed by overexpression of wild-type ZFTA (n = 0/13; Fig. 4B). Overexpression of ZFTA–MAML2 (n = 11/11), ZFTA–MAML3 (n = 5/11), and ZFTA–NCOA2 (n = 5/5) induced tumors with a median survival of 29, 103, and 36 days after birth, respectively (Fig. 4B). Histopathologic analysis of mouse tumors displayed several common histologic features among ZFTA fusion–driven tumors. All tumors presented with high density of monomorphous round to oval cells, similar to human EPN, and were similarly sharply demarcated from the surrounding brain (Fig. 4C–F). Thus, newly identified ZFTA-related fusion genes alone are sufficient to drive tumorigenesis in vivo.

Given that (i) IHC using an antibody against hemagglutinin (HA)-tagged fusion proteins revealed nuclear localization of the ZFTA-associated fusion proteins in all fusion-engineered tumors (Fig. 4G–J) and (ii) the most N-terminal zinc finger DNA-binding domain (ZF1) of ZFTA is shared by all fusion proteins (Fig. 2B), we hypothesized that this ZFTA DNA-binding domain is required for the oncogenic capacity of the fusions. In fact, in utero electroporation of ZFTA fusion genes lacking the ZF1 coding region (ΔZF1) failed to develop tumors (Fig. 4K and L; Supplementary Fig. S9A and S9B). Although ZFTA(ΔZF1)–RELA were not detected in the nucleus, nuclear localization capacity of ZFTA(ΔZF1)–MAML2 and ZFTA(ΔZF1)–NCOA2 proteins was still retained (Supplementary Fig. S9C–S9E), strongly suggesting other roles than nuclear shuttling of the shared ZF1 domain for tumorigenesis. Indeed, a cosubmitted manuscript by Kupp and colleagues in this issue demonstrates that chromatin binding and recruitment of chromatin remodeling complexes is related to the single ZF1 domain in ZFTA–RELA (21).

Transactivation domains (TAD) represented another shared element among oncogenic fusion genes (Fig. 2B). To further investigate the role of TADs for tumor formation, ZFTA was fused to potent TADs, VP64, or EP300 (Supplementary Fig. S9F). None of the animals electroporated with ZFTA–VP64 or ZFTA–EP300 developed tumors during surveillance over 12 months (Supplementary Fig. S9G, n = 0/6). These findings suggest that additional oncogenic mechanisms are associated with the respective fusion partners. Importantly, this does not preclude an oncogenic role for the TAD within ZFTA–RELA and alternative fusion types, as Kupp and colleagues demonstrated that the TAD of RELA contributes to the fusion-associated transcriptional program through recruitment of transcriptional coregulators (21).

For direct comparison of tumor models and human tumors, we applied principal component analysis to transcriptional profiles across species considering human and mouse orthologs. This approach revealed extensive molecular differences between ZFTA fusion– and YAP1–MAMLD1–driven tumors (Fig. 5A; ref. 14). To control for species-specific effects and to provide variance measurements between murine models and human tumors, a hierarchical clustering was performed. This approach revealed high similarity between molecular groups of human ST-EPN and respective murine counterparts at the level of transcription (Fig. 5B). The strong effect on the transcriptome could also be demonstrated for another ZFTA–RELA-driven mouse model generated by Arabzade and colleagues (ref. 20; Supplementary Fig. S10A). Although others found global (Arabzade and colleagues) or focal (Kupp and colleagues) changes of histone marks (H3K27ac and H3K27me3) in murine tumor cells, we observed abundant global H3K27 trimethylation and acetylation in our ZFTA fusion–driven mouse models corresponding well to respective levels in human tumors (Supplementary Fig. S10B–S10R; refs. 20, 21).

Because overexpression of L1CAM and activation of the NFκB signaling pathway are molecular features of ST-EPN-RELA (7), we next examined these characteristics in ZFTA fusion–driven murine tumors. Ccnd1 but not L1cam was highly expressed across all types of the fusion-driven tumors (Fig. 5C and D). However, a global activation of the NFκB pathway was not observed in any model of the alternative fusion types, indicating that aberrant activity of this pathway is not contributing to tumorigenesis (Supplementary Fig. S11A and S11B). In line with these findings, Kupp and colleagues observed that altering the Rel-homology domain in ZFTA–RELA fusions, which represents the DNA-binding domain shared by the NFκB family proteins for their signal transduction, did not result in loss of oncogenicity (21).

Because the DNA-binding domain of ZFTA is required for oncogenicity, we further explored common downstream effectors induced by transactivation of the ZFTA-associated fusion genes. To this end, we chose a cross-species approach to concisely match signaling pathways between human tumors and mouse models. To exclude transcriptional information governing EPN cell identity and programs across molecular groups that we had observed previously (22), we selected DEGs for human primary ST-EPN-RELAs significantly upregulated compared with all other molecular groups of EPN (n = 3,825 genes; Fig. 6A). A similar approach was used to compare gene-expression data from ZFTA-driven mouse tumors against data from murine YAP1–MAMLD1 tumors representing the only available alternative faithful model system (14). We found that 2,637 genes shared by ZFTA fusion–driven murine tumors are significantly more highly expressed in comparison with YAP1–MAMLD1 tumors (Fig. 6A). Filtering for orthologs in both mouse and human data resulted in 535 genes commonly upregulated in ZFTA fusion–related tumors across species (Fig. 6A). We next hypothesized that the list of these 535 genes includes the effector genes of characterized and uncharacterized oncogenic signaling commonly upregulated by ZFTA fusion genes. A GO analysis revealed enrichment for cancer-related signaling pathways and partly convergence into known ST-EPN-RELA group–associated pathways, for example MAPK signaling (Supplementary Table S5; ref. 6). We also found well-known oncogenes, such as the sonic hedgehog (Shh) mediator gene GLI2, the WNT mediator gene LEF1, and the EPN oncogene EPHB2 being shared by ZFTA fusion–driven tumors (Fig. 6B; refs. 3, 25, 26). All three genes were specifically upregulated in human ST-EPN-RELA as compared with other molecular groups of EPN (Supplementary Fig. S12A–S12C). In addition, a comprehensive cross-species analysis by Kupp and colleagues comprising all mouse models deployed by the three independent cosubmitted studies and the two largest published data sets of human ST-EPN-RELA identified a common fusion-associated signature of 93 genes that also included GLI2 and EPHB2 (6, 7, 20, 21). To further explore potential direct interactions of ZFTA fusions with these genetic loci, we performed chromatin immunoprecipitation sequencing (ChIP-seq) with antibodies against HA and H3K27ac as well as assay for transposase-accessible chromatin using sequencing (ATAC-seq) analyses on ZFTA–RELA-driven murine tumor cells (Supplementary Methods). Indeed, the ZFTA–RELA fusion was found to directly bind to H3K27ac-marked open chromatin regions of Gli2, Lef1, and Ephb2 (Fig. 6C–E). Consistent with our observation, the study by Kupp and colleagues found that the ZFTA portion is capable of binding these loci (21). Reanalysis of ChIP-seq data on human tumors further supported these findings (Supplementary Fig. S12D–S12G).

To examine potential functional implications of the revealed genes for ZFTA-driven tumorigenesis, we subsequently electroporated ZFTA–RELA together with genes encoding a dominant-negative form of GLI2, LEF1, and EPHB2, respectively (Fig. 7A). Although the genes encoding the C-terminal portion of LEF1 (27) and the ectodomain of EPHB2 (28) did not attenuate tumor growth (Fig. 7B), the N-terminal portion of GLI2 (dnGLI2) that inhibits GLI2-mediated transactivation (29) prevented tumor formation (Fig. 7B and C), indicating the requirement of GLI2 function for ZFTA fusion–associated tumorigenesis. In line with this finding, GLI2 protein expression was elevated in human primary tumors harboring different types of ZFTA fusion genes as well as in corresponding murine tumor models (Supplementary Fig. S13A–S13H). Moreover, we found that GLI2 transcription factor binding sites were highly enriched in histone H3K27ac-marked enhancers and superenhancers of human ST-EPN-RELAs reported in our previous study (ref. 22; Fig. 7D), further highlighting a decisive role of this oncogene.

To investigate the contribution of GLI2 to ST-EPN-RELA tumor progression, we next infected EP1NS, a ZFTA–RELA-expressing ST-EPN cell line, with doxycycline-inducible short hairpin RNAs (shRNA) against GLI2 (shGLI2_1 and shGLI2_2)- and nontargeting control shRNA (shControl)–encoding lentiviruses. Approximately 40% reduction of GLI2 transcripts was observed 48 hours after administration of doxycycline (2 μg/mL; Fig. 7E). Within 96 hours after shRNA induction, EdU pulse-labeling revealed significant reduction of proliferation in GLI2 shRNA–expressing cells when compared with shControl (Fig. 7F). Annexin V staining also confirmed induction of enhanced apoptotic cell death in GLI2 shRNA–expressing cells (Fig. 7F). To further evaluate GLI2 inhibition in vivo, tumor-bearing animals were treated with arsenic trioxide (ATO), a blood–brain barrier–penetrating drug that includes GLI2 in its target spectrum (30–32). ZFTA–RELA-electroporated mice were treated with either 2.5 mg/kg ATO or vehicle (i.p. injection, 5 times per week) after the luciferase signal reached approximately 5 × 10⁶ photons/second. ATO-treated animals showed extended survival when compared with vehicle-treated controls (Fig. 7G; Supplementary Fig. S13I). Comparable expression levels of Gli2 were detected between control and ATO-exposed tumors (Supplementary Fig. S13J), thus excluding that ATO treatment incidentally downregulated Gli2 in vivo. Together, these data suggest GLI2 as a potential therapeutic vulnerability in ZFTA fusion–positive tumors.

In this study, we aimed to further investigate the biological heterogeneity of ST-EPN as a basis for future improved diagnostic accuracy and target identification. To this end, we performed a comprehensive molecular analysis of ST-EPN that confirmed previously described stable molecular groups of EPN but also identified additional satellite clusters related to ST-EPN-RELA. The RELA fusion partner ZFTA was found to be a recurrent partner in alternative translocations within tumors that constitute these satellite clusters. The clinical significance of these satellite clusters needs to be confirmed in future studies with increased sample size and clinical information. The clusters will be included in the upcoming version 12 of the Heidelberg Brain Tumor Methylation Classifier as part of a novel molecular family of ZFTA fusion–related ST tumors.

Diagnostic assessment of cases within satellite clusters appears to be particularly challenging, as tumors often not only harbor alternative ZFTA fusions but also can present histologic characteristics atypical for EPN. Although tumors in clusters 1 and 2 were predominantly diagnosed as EPN, other institutional histologic diagnoses were reported almost exclusively for clusters 3 and 4. Notably, IHC stainings for both L1CAM⁻ and p65⁻ could not reliably distinguish between ZFTA–RELA and other ZFTA-related fusions. Two previous case reports also described diagnostically ambiguous situations with a ZFTA–RELA fusion shared between a primary ST-EPN and its relapse, which histologically was diagnosed as sarcoma, and appearance of the fusion in an atypical teratoid/rhabdoid tumor (33, 34). These findings further illustrate the diagnostic challenges imposed by these exceedingly rare tumors and indicate the potentially arbitrary role of histomorphology that does not necessarily reflect underlying molecular programs, as described for CNS-PNETs (35). These data further underpin the classification as planned for the upcoming fifth edition of the WHO Classification of CNS Tumors that allows for molecularly defined tumor types rather than adhering to strictly morphology-defined entities. We suggest that these oncogenic alterations may affect cells that had remained in an early stemness state permissive for one-hit tumorigenesis and rendering possible the development of morphologically nonneuronal, nonglial elements. This is supported by data in a cosubmitted article by Arabzade and colleagues identifying transcriptional programs within fusion-driven EPN that are active during embryonic brain development (20).

In this study, we revealed the expression of various ZFTA fusion proteins in ST tumors. Each of these fusion proteins by itself caused tumor formation in the cerebral cortex, implying that they share oncogenic mechanisms. In line with Kupp and colleagues (21), we indeed identified a zinc finger DNA-binding domain of the fusion partner ZFTA as an essential element for tumorigenesis, which also resulted in the new official designation of the gene formerly known as C11orf95. In addition, structural comparison of all ZFTA fusion partners identified the common presence of a transactivation domain, raising the possibility that ZFTA fusion oncoproteins activate oncogenes through recruitment of an activating domain to the ZFTA-bound targets. Notably, each of the newly identified ZFTA fusion genes induced tumors with different penetrance and latency. This may be attributed to variable effects of the fusion partners on the transcriptional machinery in neural progenitors. For instance, MAML2 and MAML3 have been known to be a cofactor of NOTCH, which is responsible for clonal expansion of cortical progenitors in the ventricular zone. However, MAML2 shows much stronger transcriptional activation of Hes genes than MAML3 (36). Therefore, ZFTA–MAML2-mediated enhancement of NOTCH signaling is likely to increase the number of fusion-bearing progenitors more efficiently. Consistent with this idea, we indeed found reduced survival in mice electroporated with ZFTA–MAML2 compared with ZFTA–MAML3 (Fig. 4B). Considering that NFκB signaling is involved in neural stem cell (NSC) proliferation in the cerebral cortex (37, 38), ZFTA–RELA is also likely to expand the progenitor pool of the transfected cells, thus shortening the latency of tumor formation. Because ST tumors associated with different fusions are characterized by variable methylation profiles, it could also be hypothesized that each fusion oncoprotein may exert transformation activity in distinct NSC subtypes already committed to specific progenitors, as was reported for medulloblastoma (39, 40). Notably, applying single-cell RNA-seq to a cohort of ST-EPN-RELA and posterior fossa group A EPN (PF-EPN-A), we and others previously recognized a larger intertumoral heterogeneity for ZFTA–RELA–driven tumors compared with PF-EPN-A (41). Future single-cell studies coupled with technologies for profiling the chromatin landscape may enable the inference of developmental lineages.

A previous animal study revealed the NFκB- and non–NFκB-related impact of ZFTA–RELA fusions on tumor formation by mutagenesis (13). A mutation of the Rel-homology domain failed to drive tumorigenesis, whereas alterations of the transactivation domain still resulted in tumor formation. In our study, we did not observe NFκB pathway activation in tumors without RELA as fusion partner. In keeping with this notion, Arabzade and colleagues demonstrated that a major component of the fusion binding is tumor-specific and not observed in canonical NFκB-related gene expression (20). In addition, Kupp and colleagues found that the Rel-homology domain is not required for fusion-driven gene expression (21). It remains to be further elucidated if at least transactivation domains that represent a shared pattern between fusions that cluster together and lack the Rel-homology domain, such as ZFTA–NCOA1, ZFTA–NCOA2, and ZFTA–MAML2, may contribute to tumorigenesis through binding of transcriptional cofactors. Indeed, integrated cross-species analyses identified downstream targets shared by ST tumors with ZFTA fusions, suggesting similar transcriptional activation processes. Our results stress that GLI2 functions as a relevant downstream oncogene in ZFTA fusion–driven ST tumors, and pharmacologic inhibition could significantly reduce tumor growth.

In summary, we demonstrate the transforming capacity of ZFTA-containing fusions, provide representative mouse models, and present a rationale for further preclinical studies blocking central molecular dependencies of these fusions. Tumors containing a canonical or alternative ZFTA fusion will be classified as ST-EPN, ZFTA fusion–positive in the upcoming fifth edition of the WHO Classification of CNS Tumours.

CD-1 mice used for in utero electroporation were obtained from Charles River and housed in a vivarium with a 12-hour light/dark cycle with access to food and water ad libitum. The day of the plug and the birthdate are designated as embryonic day (E) 0.5 and postnatal day (P) 0, respectively. All animal experiments for this study were conducted according to the animal welfare regulations approved by the Animal Care and Use Committee of the National Institute of Neuroscience, NCNP in Japan (approval number: 2019028R1) and the responsible authorities in Germany (Regierungspräsidium Karlsruhe, approval numbers: G-255/19 and G-260/19).

All experiments in this study involving human tissue or data were conducted in accordance with the Declaration of Helsinki. Tumor material (FFPE tissue, preisolated RNA and/or DNA) or information on molecular tumor characteristics was collected and analyzed after receiving written informed consent from the respective patients or their legal representatives and according to the guidelines of the ethical institutional review boards of the participating institutions, such as Heidelberg University Hospital and the NN Burdenko Neurosurgical Institute. For cases from diagnostic or clinical studies, material was obtained in accordance with the respective study protocol and informed consents. For all cases, a genotype check was performed to exclude the possibility that material from the same patient was received from more than one center. To this end, the Pearson correlation across beta methylation values of 59 rs loci present on both the Illumina Infinium HumanMethylation450 and the Illumina Infinium HumanMethylation EPIC array were calculated. Samples with a correlation ≥ 0.95 were considered as genotype matches.

HEK-293T (CRL-3216) cells were purchased from ATCC. HEK-293T cells were cultivated with DMEM (Thermo Fisher) supplemented with heat-inactivated 10% FBS (Thermo Fisher), 2 mmol/L l-glutamine, 100 U/mL penicillin, and 100 μg/mL streptomycin. The cells were maintained in a humidified 5% CO₂ atmosphere at 37°C and subcultured when cell confluency reached approximately 80%. Mycoplasma contamination was assessed periodically by GATC/Eurofins.

The full or partial coding regions of human ZFTA, MAML2, MAML3, and NCOA2 cDNAs with a C-terminal HA tag were amplified by PCR and cloned into pT2K-IRES-Luc plasmid vectors using In-Fusion HD Cloning kit (Takara Bio). Dominant-negative Gli2 was amplified by RT-PCR on total RNA of mouse granular neural progenitor cells. pT2K plasmids were cotransfected with Tol2 transposase encoded in the pCAGGS plasmid. For the generation of ZFTAΔZF1-RELA/MAML2/NCOA2 cDNA, a sequence of zinc finger domain was chosen based on UniProt prediction. All primers used for PCR are listed in Supplementary Table S6.

The human EPN cell line EP1NS was transduced with lentiviral pLKO-tet-on vector system (plasmid #21915, Addgene) containing a puromycin-resistance gene, and a tet-responsive element for doxycycline-inducible expression of shRNA against GLI2 (shGLI2_1 and shGLI2_2) or a nontargeting control shRNA (shControl). All primers used for cloning are listed in Supplementary Table S6. The dox-inducible vectors were generated according to a publicly available protocol (42, 43). Lentiviral particles were generated in HEK293T cells. Virus-containing supernatant was collected to infect EP1NS cell line. Infected cells were selected with 1 μg/mL puromycin. The shRNA expression for GLI2 knockdown in EP1NS was achieved by adding 1 μg/mL doxycycline every 48 hours to the medium. For proliferation assay, 96 hours after doxycycline administration, the cells were treated with EdU (final concentration: 10 μmol/L) for 12 hours and subsequently harvested with Accutase solution. EdU-incorporated cells were labeled using a Click-iT EdU Alexa Fluor 647 Flow Cytometry Assay Kit (Life Technologies) according to the manufacturer's protocol. The cells were passed through a 35-μm cell strainer yielding a single-cell suspension and analyzed by flow cytometry using a FACS Fortessa flow cytometer (BD Biosciences). For the apoptosis assay, the infected cells were harvested 96 hours after doxycycline treatment, and were subsequently washed twice with Cell Staining Buffer (BioLegend). Cells were then stained with Annexin V–APC and DAPI diluted in Annexin V Binding Buffer using Apoptosis Detection Kits (BioLegend) according to the manufacturer's protocol. Samples were analyzed by flow cytometry using a FACS Fortessa flow cytometer (BD Biosciences).

In utero electroporation was performed as reported previously (44). Specifically, endotoxin-free DNA plasmid mixture (1 μg/μL for each plasmid) was injected into the lateral ventricle of E13.5 embryos, and square electric pulses (32 V, 50 ms-on, 450 ms-off, five pulses) were delivered using 5-mm diameter platinum forceps-like electrodes (BTX). For in vivo tumor formation analysis, electroporated animals were selected at neonatal stages by intraperitoneal (i.p.) injection of d-Luciferin (150 mg/kg) and subsequent bioluminescence imaging. Growth of transfected cells was monitored every week by measurement of intensity of bioluminescence with IVIS Lumina LT Series III Caliper (PerkinElmer). The animals were sacrificed once they exhibited neurologic signs, such as head tilting, abnormal gait, and a hunched posture, or at 1 year of age if showing no symptoms.

A stock solution of 20 mg/mL ATO in 1 mol/L NaOH was prepared. It was further diluted to 0.5 mg/mL ATO with PBS, and the solution was sterile-filtrated. The vehicle solution was prepared the same way but without ATO. When the bioluminescence signal of the electroporated animals reached approximately 5 × 10⁶ photons/second, the animals were allocated randomly to vehicle- and ATO-treatment group and treated five days per week either with 2.5 mg ATO/kg/day (i.p) or the equivalent volume of vehicle solution. Prior to the treatment, 20% mannitol in 0.9% saline was intraperitoneally injected into mice (5 mL/kg) to disrupt the blood–brain barrier. The mice were monitored daily for tumor-specific symptoms and euthanized when they exhibited neurologic symptoms.

Brains with tumor from electroporated mice were dissected and fixed with formalin at 4°C for 48 hours. Five-micron-thick paraffin-embedded murine tumor sections were immunostained according to the procedures in our previous study (45). After deparaffinization, the sections were pretreated with citrate buffer at 100°C for 30 minutes. Then the sections were incubated with the primary antibodies (Supplementary Table S7) diluted with Dako REAL Antibody Diluent (Agilent #S2022) at room temperature overnight. DAB staining was performed the next day using SuperVision 2 HRP-Polymer Kit (DCS PD000POL) following the protocol provided by the manufacturer. Slides were mounted in ProLong Gold Antifade Mountant (Invitrogen #P36930). Nuclei were stained with DAPI (300 nmol/L). Images were acquired with confocal microscopes (ZEISS Cell Observer).

IHC for human samples was performed on a Ventana BenchMark ULTRA Immunostainer (Ventana Medical Systems). Antibodies used in this study are listed in Supplementary Table S7.

HEK293T cells were cultured on glass coverslips one day before transfection. Plasmid constructs were transfected using Fugene (Promega) following the instructions provided by the manufacturer. Forty-eight hours after transfection, cells were fixed with 4% paraformaldehyde for 20 minutes followed by 10-minute permeabilization with Triton buffer (0.1% Triton in PBS). After washing with PBS two times, the primary antibody (Supplementary Table S7) was applied directly on the cells for 1 hour at room temperature. The antibody solution was removed by absorption with Whatman filter paper before washing the coverslips two times for 5 minutes with PBS. The corresponding secondary antibody was applied subsequently, incubated for 30 minutes, and three times washed for 5 minutes in PBS. Finally, cells were washed briefly in ddH₂O to remove salts and pure ethanol before they were mounted on microscopy glass slides with Fluoromount-G containing 1 μg/mL DAPI (Southern Biotech).

The protein expression of the plasmids used in this study was validated by Western blotting according to the following procedures: HEK293T cells were transfected with the plasmids and harvested 48 hours after transfection. The cell pellets were lysed with RIPA buffer, and 20 μg of the protein lysates was used for protein detection. Briefly, proteins were denatured for 5 minutes at 95°C, loaded on NuPAGE Bis-Tris (#NP0301BOX, Invitrogen), and separated at 120 V for 2 hours. Proteins were transferred to methanol-activated polyvinylidene difluoride membrane by tank electrotransfer in Towbin buffer for 1 hour at 110 V. Membrane was blocked with 5% skimmed milk in 0.5% Tween/TBS (TBST) for 1 hour at room temperature prior to overnight incubation with primary antibodies (Supplementary Table S7). After washing with TBST, membrane was incubated with secondary antibody for 1 hour at room temperature. The membrane was developed with either ECL (RPN2106, GE Lifesciences) or ECL Prime (RPN2232, GE Lifesciences) as recommended by the manufacturer followed by exposure to autoradiography films in a dark room.

As for H3K27me3 and H3K27ac analysis, mouse brains were lysed in lysis buffer [150 mmol/L NaCl, 20 mmol/L Tris-HCl (pH 7.4), 2 mmol/L EDTA, 1% NP-40] and sonicated with a Bioruptor. The lysates were collected after centrifugation (13,000 × g for 10 minutes) and then denatured in SDS sample buffer at 95°C for 3 minutes. 0.1 μg and 10 μg of the lysates were used for Histone H3 and for H3K27ac and H3K27me3, respectively. Blotted membranes were blocked with 5% nonfat milk in TBST for 30 minutes and immunoblotted with anti-Histone H3 (Abcam, ab1791, 1:1,000), anti-Tri-methyl-Histone H3 (K27; Abcam, ab6002, 1:300) and anti-Acetyl-Histone H3 (K27; CST, D5E4, 1:300) antibodies. After washing, the membranes were incubated with horseradish peroxidase–conjugated secondary antibodies (GE Healthcare, 1:1,000) for 1 hour at room temperature. Then they were washed at least four times and detected via enzyme-linked chemiluminescence (Immobilon Forte; Millipore) in a cooled CCD camera (LAS-4000 mini, Fujifilm). For quantitative analysis, the signal intensities from murine tumor lysates were measured with the ImageJ software and normalized by a global level of histone H3.

Total RNA was extracted from cryopreserved mouse tissues using an RNeasy Plus Mini Kit together with QIAshredder (QIAGEN) according to the manufacturer's instructions and stored in −80°C until use. cDNAs for downstream application were prepared using the SuperScript VILO cDNA Synthesis Kit (Invitrogen).

qPCR mix was prepared following the manufacturing protocol of Power SYBR Green PCR Master Mix (Applied Biosystems). qPCR was performed using the QuantStudio 5 RT-PCR system (Applied Biosystems). The cycling conditions used were 95°C for 10 minutes and 40 cycles of 95°C for 15 seconds and 60°C for 1 minute following dissociation analysis. All qPCR reactions were done in triplicate and normalized to TBP mRNA levels.

Genome-wide DNA methylation profiling was performed using the Illumina Infinium HumanMethylation450 and the Illumina Infinium HumanMethylation EPIC Kits as described previously and according to the manufacturer's instructions (6).

All computational analyses were performed in R version 3.4.4 (R Development Core Team, 2019). Raw signal intensities were obtained from IDAT files using the minfi Bioconductor package version 1.24.0 (46, 47). Illumina EPIC and 450k samples were merged to a combined data set by selecting the intersection of probes present on both arrays (combineArrays function, minfi). Each sample was individually normalized as described previously (14). Subsequently, a correction for the type of material tissue (FFPE/frozen) and array (450k/EPIC) was performed by fitting univariate, linear models to the log₂-transformed intensity values (removeBatchEffect function, limma package version 3.34.5). The methylated and unmethylated signals were corrected individually before beta-values were calculated. CpG probes selection was performed as described previously (14). In total, 428,230 probes were kept for downstream analysis.

To perform unsupervised nonlinear dimension reduction, the remaining probes were used to calculate the 1-variance weighted Pearson correlation between samples by applying the function wtd.cors function of the R-package weights version 1.0.1. The resulting distance matrix was used as input for t-distributed stochastic neighbor embedding analysis (t-SNE; Rtsne package version 0.13). The following nondefault parameters were applied: theta = 0, pca = F, max_iter = 2,500, perplexity = 20.

To identify fitting samples for this study, an exploratory set of 20 cases was chosen on the basis of the following three conditions: (i) prediction for ST-EPN-RELA according to DNA methylation–based classification but without evidence for a canonical ZFTA–RELA fusion, (ii) the vice versa combination with a typical fusion event in the absence of a reliable ST-EPN-RELA score, or (iii) ST tumors histologically diagnosed as EPN that cannot readily be assigned to any of the existing molecular classes. DNA methylation profiles of these cases were clustered with a cohort of 61,821 samples from different tumor entities and experimental data and compared with a reference set (15). This analysis was subsequently repeated with an increased set of 71,270 samples and with a reference cohort of 507 EPN cases covering all 10 major subgroups (5, 6).

The 613 samples in the cohort were assigned to either ST-EPN-RELA, satellite clusters 1–4, or outlier cases based on a hierarchical density-based scan (HDBSCAN; R-package dbscan version 1.1-5; ref. 48) using the two-dimensional projection resulting from the t-SNE as input and applying a minPts parameter of 5. Cluster stability was assessed by a resampling approach. For each of 500 resampling iterations, t-SNE dimension reduction followed by HDBSCAN cluster assignment was applied to 80% of the samples sampled without replacement. In accordance to Consensus Clustering (49), a consensus matrix was calculated storing pairwise relative frequencies of how often two samples were assigned to the same cluster. A heat map of the consensus matrix was generated applying the pheatmap R-package using the default settings for clustering rows and columns (Supplementary Fig. S2C). The heat map was annotated with the HDBSCAN results for the complete data set as well with the frequency of how often a sample was assigned to the outlier cluster over the resampling iterations. The distribution of the number of clusters detected over the 500 iterations is shown in Supplementary Fig. S2B, indicating 6 as the most frequently identified number of clusters (ST-EPN-RELA, clusters 1–4, and outliers).

Copy-number variation analysis from 450k and EPIC methylation array data was performed using the conumee Bioconductor package version 1.12.0 (Hovestadt V, Zapatka M, 2017). Summary copy-number profiles were created by summarizing these data in the respective sets of cases.

DNA panel sequencing was performed on 29 samples obtained from either fresh-frozen or FFPE material using a customized enrichment/hybrid-capture-based next-generation sequencing gene panel of 130 genes recurrently altered in brain tumors according to the manufacturer's instructions and as described previously (50).

High-throughput sequencing of 66 samples obtained from FFPE material and 28 samples obtained from fresh-frozen material with sufficient quality and quantity of RNA was performed according to the manufacturer's instruction and as described previously (5, 51). General FFPE RNA-seq data processing (reads alignment, quality control, and gene-expression counts computation) was performed as described previously (51).

Unsupervised analysis of tumor samples was performed with principal component analysis, t-SNE, and hierarchical clustering based on the selection of the top 1,000 most variable genes with log₂ RPKM-normalized gene-expression counts. Selection of the NFκB target genes was derived from the corresponding source (https://bioinfo.lifl.fr/NF-KB/).

Fusion genes discovery from RNA-seq data was performed using two independent tools: InFusion (52) and Arriba (https://github.com/suhrig/arriba/). Transcription of fusion identified by RNA-seq was confirmed by RT-PCR. RNA was extracted from frozen tumor samples, then reverse transcription and PCR were carried out by using the OneStep RT-PCR Kit (QIAGEN) using specific primers (Supplementary Table S6). Fusions were confirmed by Sanger sequencing (Eurofins Genomics).

The Affymetrix data cohorts were used for cross-species analysis. Human Affymterix data from corresponding study (6) were integrated from the R2 system. The list of common mouse–human gene orthologs from AGDEX Affymetrix reference (14,635 genes in total) was integrated for gene-probe selection in further comparison between human tumor and mouse model data sets. Initially, differentially expressed orthologous genes between the ST-EPN-YAP1 and ST-EPN-RELA tumors starting from the top 5,000 most evident (min P_adj < 0.0006) were applied as the target reference to confirm the model's correspondence based on unsupervised hierarchical clustering and principal component analysis as it was described previously (14). Furthermore, to increase the specificity for ZFTA-driven effects, evident DEGs of ST-EPN-RELA tumors versus all other EPN subgroups were integrated for target candidate selection (n = 3,825, min. P_adj < 0.05). DEGs between models were detected using limma R-package (53) with P_adj < 0.05.

For the GO and pathway analysis, the common orthologs between mouse models and human tumors were selected from DEGs specific for ST-EPN-RELA against all other EPN subgroups and for each ZFTA-driven model against MAMLD1-YAP1 control. GO analysis was performed using ClueGO tool (54) by focusing the top 300 top evident genes.

The Kaplan–Meier method was applied for survival analysis comparing the different fusion constructs and visualized using R version 3.6.1 (R Core Team, 2020) and the survival- and survminer-R packages(https://github.com/therneau/survival, https://github.com/kassambara/survminer). The paired t test was used for EdU and Annexin V analysis in the shGLI2 experiment and visualized using GraphPad Prism.

Data from methylation profiling, RNA-seq, and DNA panel sequencing will be deposited at the European Genome–phenome archive (https://www.ebi.ac.uk/ega/home).

D.R. Ghasemi reports other support from German Cancer Aid and German Academic Scholarship Foundation during the conduct of the study. P. Benites Goncalves da Silva reports personal fees from CNPq—Conselho Nacional de Desenvolvimento Científico e Tecnológico outside the submitted work. G. Fleischhack reports grants from German Children Cancer Foundation during the conduct of the study; grants from German Children Cancer Foundation and other support from Novartis outside the submitted work. S. Rutkowski reports grants from German Children's Cancer Foundation during the conduct of the study; personal fees from BMS and Roche outside the submitted work. V. Ramaswamy reports personal fees from AstraZeneca outside the submitted work. D. Capper reports a patent for DNA methylation–based method for classifying tumor species pending. T. Milde reports grants from BioMed Valley Discoveries outside the submitted work. R.J. Gilbertson reports grants from Cancer Research UK, Brain Tumour Charity and NCI during the conduct of the study. S.M. Pfister reports grants from Eli Lilly, Roche, Pfizer, Bayer, epo, Charles River, and PharmaMar outside the submitted work. F. Sahm reports grants and other support from Illumina, Agilent, AbbVie, and Medac outside the submitted work. No disclosures were reported by the other authors.

T. Zheng: Conceptualization, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. D.R. Ghasemi: Conceptualization, formal analysis, validation, investigation, visualization, writing–original draft, writing–review and editing. K. Okonechnikov: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, writing–original draft, writing–review and editing. A. Korshunov: Investigation, writing–review and editing. M. Sill: Investigation, writing–review and editing. K.K. Maass: Conceptualization, investigation, writing–review and editing. P. Benites Goncalves da Silva: Investigation, writing–review and editing. M. Ryzhova: Investigation, writing–review and editing. J. Gojo: Investigation, writing–review and editing. D. Stichel: Investigation, writing–review and editing. A. Arabzade: Investigation, writing–review and editing. R. Kupp: Investigation, writing–review and editing. J. Benzel: Resources, investigation, writing–review and editing. S. Taya: Investigation, writing–review and editing. T. Adachi: Investigation, writing–review and editing. R. Shiraishi: Investigation, writing–review and editing. N.U. Gerber: Investigation, writing–review and editing. D. Sturm: Investigation, writing–review and editing. J. Ecker: Investigation, writing–review and editing. P. Sievers: Investigation, writing–review and editing. F. Selt: Investigation, writing–review and editing. R. Chapman: Resources, writing–review and editing. C. Haberler: Resources, writing–review and editing. D. Figarella-Branger: Resources, writing–review and editing. G. Reifenberger: Resources, writing–review and editing. G. Fleischhack: Resources, writing–review and editing. S. Rutkowski: Resources, writing–review and editing. A.M. Donson: Resources, writing–review and editing. V. Ramaswamy: Resources, writing–review and editing. D. Capper: Resources, writing–review and editing. D.W. Ellison: Resources, writing–review and editing. C.C. Herold-Mende: Resources, writing–review and editing. U. Schüller: Resources, writing–review and editing. S. Brandner: Resources, writing–review and editing. P. Hernaiz Driever: Resources, writing–review and editing. J.M. Kros: Resources, writing–review and editing. M. Snuderl: Resources, writing–review and editing. T. Milde: Resources, writing–review and editing. R.G. Grundy: Resources, writing–review and editing. M. Hoshino: Resources, writing–review and editing. S.C. Mack: Supervision, funding acquisition, writing–review and editing. R.J. Gilbertson: Supervision, funding acquisition, writing–review and editing. D.T. Jones: Conceptualization, writing–review and editing. M. Kool: Conceptualization, writing–review and editing. A. von Deimling: Resources, writing–review and editing. S.M. Pfister: Conceptualization, supervision, funding acquisition, writing–review and editing. F. Sahm: Conceptualization, resources, supervision, writing–original draft, project administration, writing–review and editing. D. Kawauchi: Conceptualization, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing. K.W. Pajtler: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

We thank the High-Throughput Sequencing Unit of the Genomics and Proteomics Core Facility, German Cancer Research Center (DKFZ) for providing excellent services regarding all sequencing experiments. We are grateful to Norman Mack, Laura Sieber, Britta Statz, Monika Mauermann, Lukas Schmitt, and Tatjana Wedig, Department of Pediatric Neurooncology, German Cancer Research Center, and Laura Doerner, Lisa Kreinbihl, Jochen Meyer, Lea Hofmann and Moritz Schalles, Department of Neuropathology, Heidelberg University Hospital, for excellent technical assistance. This study was generously supported by the Ein Kiwi gegen Krebs-foundation. We thank the German Childhood Cancer Foundation for funding (“Molecular Neuropathology 2.0–Increasing diagnostic accuracy in paediatric neurooncology”; DKS 2015.01). Furthermore, this work was supported by fellowships of German Academic Exchange Service (DAAD; to T. Zheng), the Mildred-Scheel doctoral program of the German Cancer Aid (to D.R. Ghasemi), the German Academic Scholarship Foundation (to D.R. Ghasemi), the Hertie Network of Excellence in Clinical Neuroscience (to P. Sievers), the Else Kröner Excellence Program of the Else Kröner-Fresenius Stiftung (EKFS; to F. Sahm), a grant from the Japan Agency for Medical Research and Development, AMED (JP20ck0106534h0001), Fund for the Promotion of Joint International Research (19K24687, JSPS; to D. Kawauchi), and the Collaborative Ependymoma Research Network (CERN) fellowship (to K.W. Pajtler). U. Schüller was supported by the Gert und Susanna Mayer Stiftung and the Fördergemeinschaft Kinderkrebszentrum Hamburg. Part of the study was funded by the National Institute for Health Research to UCLH Biomedical research center (BRC399/NS/RB/101410). S. Brandner was also supported by the Department of Health's NIHR Biomedical Research Centre's funding scheme. A. Korshunov is supported by the Helmholtz Association Research Grant (Germany). M. Ryzhova is supported by the RSF Research Grant No. 18-45-06012. The methylation profiling at NYU is in part supported by the Friedberg Charitable Foundation and the Making Headway Foundation grants (to M. Snuderl). We thank Maximilian Harkotte, Department of Psychology, Eberhard Karls University Tübingen, for advice regarding data analysis and Niklas Freund, MRC Laboratory of Molecular Biology and Cambridge University, for fruitful discussions. We are grateful to all patients and their families for taking part in this study.

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