The parents/guardians of the patient gave their informed consent before the donation of the tumor tissue for research purposes and for retrospective research access to relevant medical records and previously obtained pathology samples. Written informed consent was obtained from the minor’s legal guardian for the publication of any potentially identifiable images or data included in this article. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of the Child and Adolescent Health Service, Western Australia (HREC: 1769/EP (PRN 0000002372) A Perth Children’s Hospital Oncology Protocol for Collecting and Banking Paediatric Research Specimens; approved 21/08/2003).

Animal experiments were approved by the Animal Ethics Committee of the Telethon Kids Institute and performed in accordance with Australia’s Code for the Care and Use of Animals for Scientific Purposes (AEC#263 approved 1/9/2013, AEC#300 approved 18/4/2016, AEC#362 approved 24/4/2020). Immunodeficient BALB/c nude mice were purchased from the Animal Resources Centre (Murdoch, Western Australia, Australia) and J:NU mice were obtained from The Jackson Laboratory (Bar Harbor, Maine, USA). Implantation of tumor cells was performed as previously described (20). Specifically, approximately 1 hour following the patient’s final surgical procedure, tumor tissue (fourth surgical sample) from the patient (ID 801806) was mechanically dissociated, filtered through a 100 μm cell strainer, and suspended in Matrigel (BD Biosciences, San Jose, California, USA). Animals received general anesthesia and pre-operative analgesia (ketamine 100mg/kg intraperitoneally, medetomidine 1mg/kg intraperitoneally) and post-operative analgesia (0.4mg/ml ibuprofen in drinking water for five days). Cells (approximately 10⁶ per mouse in 2μl) were implanted into the cortex (approximately -0.45mm from bregma at a depth of 3mm; n=3) or cerebellum (approximately -6.3mm from bregma at a depth of 2mm; n=2), of five 8-week-old mice using a Hamilton syringe. The implantation process took approximately 5 minutes per mouse. Upon tumor-related morbidity, the brain was bisected at the implantation site and one half of the brain containing the tumor was kept for histology. The remaining tumor was removed, dissociated and reimplanted into the cortex of successive recipients as described above (referred to as PDOX TK-EPN862). At autopsy, no evidence of leptomeningeal spillage of tumor cells from the implantation procedure was observed, nor was there evidence of leptomeningeal metastasis of the tumor.

Tissue samples were fixed in 4% paraformaldehyde in phosphate buffered saline or neutral buffered formalin for 24 hours and embedded in paraffin. Patient and mouse PDOX tissue sections (5 µm) underwent microwave antigen retrieval in a citrate buffer before immunohistochemistry (IHC) using the following antibodies and dilutions: Olig2 (Millipore, Burlington, MA, USA, MABN50; 1:200), Tri-methyl-histone H3 (K27) (Cell Signaling, Beverly, MA, USA, 9733; 1:200), GFAP (Sigma Aldrich, St Louis, MO, USA, G3893-2ML; 1:200), Ki67 (Cell Signaling, 9027; 1:400), p53 (Cell Signaling, 2527; 1:160), synaptophysin (Cell Signaling, 36406; 1:200), EMA (Dako, Santa Clara, CA, USA, M0613; 1:100) and EZHIP (Sigma Aldrich, HPA004003-25UL; 1:200). Sections were incubated with species-specific biotinylated goat anti-IgG secondary antibodies, followed by detection with an Elite ABC kit and NovaRED peroxidase substrate, then counterstained with Gill’s hematoxylin according to manufacturer’s instructions (Vector Laboratories, Burlingame, California, USA). Hematoxylin and eosin (H&E) staining was performed as per standard protocols using a Leica Autostainer XL.

Genomic germline DNA was prepared from peripheral blood mononuclear cells using a QIAamp DNA Mini Kit (Qiagen, Hilden, North Rhine-Westphalia, Germany, 51304) as per the manufacturer’s instructions for DNA extraction from lymphocytes. Genomic tumor DNA and RNA were prepared from fresh frozen patient and PDOX tumor samples using an AllPrep DNA/RNA Mini Kit (Qiagen, 80204) according to the manufacturer’s instructions. DNA quality was determined by gel electrophoresis and spectrophotometry (Nanodrop, Thermo Fisher Scientific, Waltham, MA, USA), and quantified using fluorometry (Qubit, Life Technologies, Waltham, MA, USA, Q32851). RNA quality and quantity were determined using the LabChip GX nucleic acid analyzer (Perkin Elmer, Waltham, MA, USA) (performed by the Australian Genome Research Facility, Perth, Western Australia, Australia).

Genomic DNA (500–1000 ng) was treated with sodium bisulphite using the EZ DNA methylation kit (Zymo Research, Irvine, CA, USA) and bisulphite conversion was confirmed by methylation specific PCR as described previously (21, 22). Quantification of DNA methylation was performed at the Australian Genome Research Facility (Melbourne, Victoria, Australia) using the Human Methylation EPIC BeadChip (Illumina, San Diego, CA, USA) run on an Illumina iScan System (Illumina) using the manufacturer’s standard protocol. Raw idat files were uploaded to an online DNA methylation-based classification of CNS tumors platform (www.molecularneuropathology.org, version 11b4 and version 12.5) (23) and basic copy number variant profiles from methylation array data analyzed using the output generated from this classifier (24). Fisher’s exact test was performed using GraphPad Prism software (version 9.4.0) to determine statistical significance between 1q gain and PDOX generation success.

Whole genome sequencing (WGS) data obtained from the patient germline and tumor DNA samples were analyzed as reported in (25). For the PDOX, an additional step to remove mouse reads using BBSplit version June 11 2018 (26) was done prior to the previously described method. Default parameters were used except for ambiguous2 that was set to `toss` in order to conservatively exclude ambiguously mapped reads to either the mouse or human reference genomes. WGS analysis included the identification of somatic single nucleotide variants, short indels, cytogenetic-scale and gene-level copy number and structural variants, as reported in (25). WGS dataset generated by this study are available from the European Genome-Phenome Archive under accession number EGAS00001006843.

RNA sequencing (RNAseq) analysis and expression profiling was performed as reported in (25). Differential expression analysis was conducted using the R package edgeR. Genes were removed if the counts per million was less than 1 in 2 or more samples. Genes were considered significantly differentially expressed with an absolute fold change (|FC|) ≥ 2 and a false discovery rate (FDR) < 0.05. For the differential analysis between cranial metastasis and the PDOX model a further filtration was performed that removed all genes with a counts per million of 0 due to mouse infiltrating reads. Correlation analysis was performed between the different tumors and PDOX using the R package corrplot on the filtered gene set. KEGG pathway enrichment analysis was performed using DAVID with the significant differentially expressed genes from the PDOX comparing the cranial metastasis as input. Transcripts per million (TPM) expression values were used for plotting and for comparing the patient tumor and PDOX model samples against the ZERO cohort, a reference dataset containing high-risk pediatric brain tumors (25). RNAseq dataset generated by this study are available from the European Genome-Phenome Archive under accession number EGAS00001006844.

A previously well, three-year-old male presented with a three-week history of headache, early morning vomiting, seizures and lethargy. Magnetic resonance imaging (MRI) of the brain revealed the presence of a large posterior fossa mass (71mm by 50mm) within the fourth ventricle and extending into the foramen of Luschka ( Figures 1A, B ). An extraventricular drain was placed to release the pressure followed by subtotal excision of the mass (first surgical sample). Histopathological assessment of the resected mass showed a highly cellular tumor with evidence of widespread perivascular pseudorosettes, moderate nuclear pleomorphism, and multifocal necrosis. Of note, no true ependymal rosettes were identified. The proliferative index, as assessed by Ki67-positivity, was approximately 20% with 9-12 mitotic figures per 10 high power fields of view identified. Tumor cells were predominantly negative for OLIG2 and synaptophysin, positive for GFAP and demonstrated intracytoplasmic dot-like EMA staining. These characteristics are consistent with the diagnosis of WHO grade III EPN with anaplastic features. Immunostaining for p53 was negative. Tumor cells were also negative for histone H3 lysine 27 tri-methylation (H3K27me3) and expressed high levels of EZHIP, consistent with the features of PFA EPN ( Figure 1C ).

Following surgery, the patient suffered from severe posterior fossa syndrome and required intense rehabilitation before being clinically fit to receive radiotherapy. Postoperative imaging confirmed a residual mass (12mm by 8mm) at the right lateral lower pons with extension over the petrous ridge into the middle cranial fossa. The patient was treated as per the ACNS0831 Children’s Oncology Group study protocol (27). He received two cycles of induction chemotherapy (vincristine, carboplatin, cyclophosphamide and etoposide). Imaging assessment post-induction cycles indicated further progression of the residual tumor (increase to 27mm by 10mm). A further surgical attempt achieved a partial resection. Histopathological assessment revealed the residual mass retained the features of the original tumor, however no tissue sample was available from this resection for research. The patient received 59.4 Gy of focal radiotherapy followed by four cycles of maintenance chemotherapy (vincristine, cisplatin, cyclophosphamide and etoposide).

After being in remission for 12 months following the completion of treatment, surveillance imaging revealed the presence of a solitary large drop metastasis in the terminal thecal sac between L4-S2 ( Figure 2A ) with stable residual intracranial disease. Complete resection of the spinal lesion was performed (second surgical sample), followed by 36 Gy craniospinal irradiation with 14.4 Gy focal boost. The histological features of the metastasis were consistent with the primary lesion ( Figure 2B ).

Ongoing surveillance scans 10 months after the completion of craniospinal irradiation detected another metastatic spinal lesion at T12 ( Figure 3A ). The patient then commenced an early phase trial protocol for recurrent malignancies [ACCT007: Rap-CV (28)] involving treatment with rapamycin, cyclophosphamide, and vinorelbine. No response was observed following two cycles of treatment, and the spinal lesion progressed to 50 mm in size. A further metastatic lesion (12 mm) in the mesial occipital region was also discovered at this time ( Figure 4A ). To prevent spinal cord compression, complete excision of the spinal metastasis was performed (third surgical sample), followed by 15 Gy focal radiation. The patient was further treated with fluorouracil according to another early phase clinical trial. The mesial occipital lesion continued to progress ( Figure 4B ) requiring complete resection (fourth surgical sample) followed by a 20 Gy focal boost to the tumor bed. The spinal and cranial lesions were both histologically consistent with previous tumor samples ( Figures 3B , 4C ). The patient had stable disease for four months, following which leptomeningeal metastases were detected throughout the brain and spinal cord. He was treated with one dose of gemcitabine according to an early phase trial treatment without success. The patient was provided with end-of-life care and died a short time later, four years following the primary diagnosis. A summary of the treatment procedures performed and surgical samples collected is illustrated in Supplementary Figure 1 .

Tumor cells from the fourth surgical sample were implanted into the brains of five immunodeficient mice. Three of these five mice developed brain tumors (two from cortical implants and one from cerebellar implants) generating a PDOX model termed TK-EPN862. Upon the development of tumor-related morbidity in these animals, tumor tissues were harvested and serially transplanted into the cortex of a further 11 immunodeficient mice. Of these secondary implant recipients, two mice developed a brain tumor. Upon serial implantation of these tumor cells into the cortex of five further mice (tertiary implant recipients), no further tumors grew, resulting in the loss of the PDOX model ( Figure 5A ). Attempts to resurrect the TK-EPN862 model by orthotopically implanting cryopreserved cells were unsuccessful.

The median time to morbidity across all tumor-related deaths was 262 days. Asymptomatic mice were either euthanized for reasons unrelated to tumor growth (such as rectal prolapse) or at the predetermined experimental endpoint in accordance with animal ethics requirements (defined as 365 days following implant).

Histological assessment of tumor tissue from the TK-EPN862 PDOX ( Figure 5B ) demonstrated that the tumors growing in mice recapitulated many features of the patient tumor from which they were derived. Similar to the matched patient surgical sample ( Figure 4C ), TK-EPN862 tumors were highly cellular with evidence of perivascular pseudorosettes. Immunostaining of TK-EPN862 tumor cells was consistent with the features of EPN, including negative staining for OLIG2 and synaptophysin, positive staining for GFAP, and intracytoplasmic dot-like positivity for EMA. TK-EPN862 tumor cells were also negative for p53 and had a proliferative index of approximately 20%. Consistent with the features of PFA EPN, tumor cells from TK-EPN862 were negative for H3K27me3 and expressed high levels of EZHIP ( Figure 5B ). These histopathological features were maintained across in vivo passages of TK-EPN862.

Methylation profiling of the first surgical sample from patient 801806 indicated it classified clearly as a PFA EPN (calibrated score >0.99 using the Molecular Neuropathology 2.0 classifier versions 11b4 and 12.5) ( Supplementary Table 1 ). The metastatic surgical samples examined (two spinal lesions and one distal cranial lesion) also robustly classified as PFA EPN (calibrated score > 0.99 using the Molecular Neuropathology 2.0 classifier versions 11b4 and 12.5), irrespective of where the tumor recurred, consistent with the findings of others (4, 29). Additionally, the TK-EPN862 PDOX classified as PFA EPN (calibrated score > 0.98 using the Molecular Neuropathology 2.0 classifier versions 11b4 and 12.5), demonstrating faithful recapitulation of the original patient tumor in the mouse ( Supplementary Table 1 ).

PFA EPN can be further divided molecularly into nine subtypes (PFA-1a-f and PFA-2a-c), each with distinct survival outcomes (10). Additional analysis of this patient’s disease using the most recent version of the Molecular Neuropathology 2.0 classifier (v12.5), which includes these subclasses, further classified all surgical and PDOX samples as PFA-2 EPN (calibrated score > 0.99). While the primary (first) surgical sample and the third surgical sample (spinal) were unable to be further subclassified, possibly due to normal tissue contamination, the second surgical sample and the PDOX robustly classified more specifically as PFA-2b (calibrated score >0.9). The fourth surgical sample from which the PDOX was derived also best classified as PFA-2b (calibrated score = 0.89) ( Supplementary Table 1 ). Of note, PFA-2 tumors are associated with a higher rate of distant relapse compared to PFA-1 tumors (10), consistent with the features of this case.

Copy number profiling using methylation data revealed gains of chromosomes 7, 8 and 19 across all surgical samples and in the TK-EPN862 PDOX tumor, supporting the notion that the secondary and subsequent tumors were metastases of the primary tumor, rather than de novo occurrences. Whole chromosome gains, including chromosomes 8 and 19 as observed in this case, are more common in PFA-2 EPN compared to PFA-1 EPN (10), and are consistent with the molecular classifications of the patient and PDOX tumors ( Supplementary Table 1 ). In the fourth surgical sample and the matched PDOX, an additional loss of chromosomes 2q and 6q were observed, suggesting progressive genomic instability of the tumor ( Figure 6 ). Gain of 1q, which is associated with more aggressive disease and poorer outcome in PFA EPN (4, 30), was not observed, consistent with the low frequency of this alteration in PFA-2b EPN (10).

WGS of the fourth surgical sample and matched PDOX confirmed the chromosomal abnormalities observed in the copy number plots (gain of chromosomes 7, 8 and 19 and loss of chromosomes 2q and 6q depicted in the third circle of the CIRCOS plots in Figure 7 ). In addition, loss of chromosome 16 and gain of 17q were observed in the PDOX by WGS ( Figure 7B ), however, no cancer-relevant genes in these locations were found to be significantly over- or under-expressed compared to the matched surgical sample by RNAseq. No single nucleotide variants of clinical significance were identified in either sample, consistent with the low mutation rates observed in PFA ependymomas (11). In particular, the absence of histone H3 K27M mutations [which are observed solely in PFA-1 EPN and absent from PFA-2 EPN (10)] are consistent with the molecular classification of these tumors as PFA-2 EPN.

In an effort to investigate if there were gene expression differences in the metastatic samples compared to the primary tumor that may provide new knowledge about relapsed EPN, we performed RNAseq on the primary tumor (first surgical sample), one subsequent spinal metastasis (third surgical sample), and the cranial recurrence (fourth surgical sample). There was insufficient high-quality RNA available from the second surgical sample to perform RNAseq on this tumor. Gene expression analysis showed little variance between the primary tumor, spinal metastasis and cranial recurrence, with correlation coefficient values above 0.95 between all sample comparisons ( Figure 8A ), despite the marked chromosomal losses observed in the fourth surgical sample compared to earlier samples. These data suggest that this PFA EPN predominantly retained its pattern of gene expression across metastases in different compartments of the CNS.

Comparing the primary tumor with one of the spinal lesions, eight genes were differentially expressed ( Figure 8B and Supplementary Table 2 ), with IFITM1 and ZFHX4 being notable due to their described roles in cancer metastasis (31, 32). Expression levels of IFITM1 increased 16.6-fold in the spinal metastasis compared to the primary cranial lesion ( Supplementary Table 2 ). IFITM1 is associated with glioma cell proliferation, migration and invasion (31), and so may have played a role in the metastatic process in this lesion. Additionally, the expression of ZFHX4 increased over 120-fold in the spinal metastasis ( Supplementary Table 2 ). Higher expression of this gene may have contributed to the progression of this tumor as ZFHX4 has been associated with poor survival and metastasis in ovarian cancer (32) and is reported to play a role in the maintenance of tumor-initiating cells in glioblastoma (33).

When comparing the primary tumor and the cranial recurrence, 21 genes were significantly differentially expressed (|FC|≥2, FDR<0.05; Figure 8C and Supplementary Table 3 ), with expression of the proto-oncogene FOSB increased over 100-fold in the cranial recurrence compared to the primary tumor. FOSB has been reported to be highly expressed in glioma tissue compared to normal brain and is associated with glioma cell proliferation, migration, and invasion (34). The high expression levels of ZFHX4 observed in the spinal metastasis were also observed in the cranial recurrence (43.8-fold increase compared to the primary tumor), reinforcing the possibility this gene may have played a role in the metastatic progression of this disease ( Supplementary Table 3 ).

We next aimed to compare the transcriptome of TK-EPN862 with the matched lesion from which it was derived (the fourth surgical sample). Transcriptome analysis revealed 113 differentially expressed genes between the PDOX and the cranial metastasis ( Figure 8D and Supplementary Table 4 ). We then performed KEGG pathway analysis using this gene list in order to elucidate specific biological pathways that may be altered in the PDOX. This revealed that most of the significantly altered genes were associated with pathways expressed in normal brain tissue ( Supplementary Table 5 ). Additionally, three genes (GRIK2, KCNJ3 and RAPGEF4) located on chromosomes 2q or 6q were highly overexpressed in the PDOX model, which was unexpected given the loss of 2q and 6q demonstrated by copy number estimates in both samples. Taken together, this suggests that the differential gene expression patterns observed are most likely due to normal mouse brain contamination, rather than alterations arising in the tumor cells post-engraftment.

Following this, we evaluated expression levels of EZHIP as a marker of PFA EPN (10), using the ZERO cohort of high-risk pediatric brain tumors as a reference dataset (25). As expected, PFA EPN within the reference cohort expressed high levels of EZHIP ( Figure 9 ; green dots). High expression of EZHIP was observed in the first, third and fourth surgical samples, as well as in TK-EPN862 ( Figure 9 ; red and yellow dots, respectively), which correlates with the high levels of EZHIP protein expression observed by IHC ( Figures 1 , 3 - 5 ). The lower RNA expression level of the PDOX compared to the matched patient tumor (fourth surgical sample) is most likely due to the contaminating normal mouse brain tissue as discussed above.

EPNs are challenging tumors to propagate in the mouse, with other laboratories publishing low success rates with this tumor type compared to other malignant CNS tumors (16, 18). Our combined data on attempts at establishing EPN PDOXs from the Telethon Kids Institute (Perth, Australia), Fred Hutchinson Cancer Centre (Seattle, USA) and Children’s Cancer Institute (Sydney, Australia) show that collectively only five out of 36 attempts (5/36; 13.8%) were successful beyond two passages in vivo (excluding pending attempts that have not yet had the opportunity to be propagated beyond two passages). Indeed, 25 of all attempts (25/39; 69.2%) failed to establish at all from the primary implant ( Supplementary Table 6 ). Furthermore, of the five successful models, three of these have begun to demonstrate loss of tumorigenicity at later in vivo passages, further highlighting the challenges of creating EPN PDOX models.

As the development of these models requires significant time and resource input for a relatively low chance of engraftment success, we sought to identify any biomarkers that may be indicative of increased likelihood of PFA EPN PDOX generation success. Despite the aggressive nature of the tumor in the patient presented in this report, the tumor and the matched PDOX did not demonstrate 1q gain, which is associated with poorer outcomes and more aggressive disease in PFA EPN (4, 30). Of the published PFA EPN PDOX models with molecular data, 1q gain was reported in all but one of these models (16, 18, 35), raising the possibility that 1q gain may be associated with an increased likelihood of a PFA EPN PDOX successfully establishing. To investigate this, we performed DNA methylation array to determine 1q status on the tumors from all historic attempts to establish a PFA EPN PDOX model in the laboratories of Telethon Kids Institute (Perth, Australia) and Children’s Cancer Institute (Sydney, Australia). A lack of primary patient material precluded analysis on unpublished samples from the Fred Hutchinson Cancer Center cohort. Molecular classification as PFA EPN were confirmed for all tumors using the Molecular Neuropathology 2.0 classifier (v11b4 and v12.5). A successful PDOX model was defined as having been successfully propagated beyond two passages in vivo. Including data from published PDOX models, we found that 10/11 successful PFA EPN PDOX models had 1q gain, compared to only 1/7 attempted PFA EPN PDOX models that failed to establish ( Table 1 and Supplementary Figure 2 ), and that this difference was statistically significant (p = 0.0025). Although the sample size is small owing to the rarity of this specific subtype, these data suggest that 1q gain may be an important predictor of PFA EPN PDOX establishment success.