Ependymoma relapse goes along with a relatively stable epigenome, but a severely altered tumor morphology

Abstract The molecular biology of ependymomas is not well understood and this is particularly true for ependymoma relapses. We aimed at finding out if and to which extent, relapses differ from their corresponding primary tumors on the morphological, chromosomal and epigenetic level. We investigated 24 matched ependymoma primary and relapsed tumor samples and, as a first step, compared cell density, necrosis, vessel proliferation, Ki67 proliferative index, trimethylation at H3K27 and expression of CXorf67. For the investigation of global methylation profiles, we used public data in order to analyze copy number variation profiles, differential methylation, methylation status and fractions of hypo‐ and hypermethylated CpGs in different epigenomic substructures. Morphologically, we found a significant increase with relapse in cell density and proliferation. H3K27 trimethylation and CXorf67 expression remained stable between primary and relapse tumor samples, and the analysis of DNA methylation profiles neither revealed significant differences in copy number variations nor differentially methylated regions. Significant differences in the methylation status were found for CpG islands, but also in N Shelves or S Shelves, depending on the molecular subgroup. The fraction of probes changing their methylation in the epigenomic substructures appeared subgroup‐specific. Most changes occur in CpG islands, for which relapsed tumors demonstrate higher methylation values than primary tumors. The morphological differences reflect increased aggressiveness upon ependymoma relapse, but, despite slight changes, this observation does not appear to be sufficiently explained by epigenetic changes.


INTRODUCTION
Ependymomas are neuroepithelial tumors of the central nervous system (CNS). They occur in the supratentorial or infratentorial region of the brain, but also in the spinal cord. According to the WHO classification of tumors of the CNS, ependymomas are classified into WHO grades I, II, or III (19). However, WHO grading is often difficult due to ill-defined diagnostic criteria, intratumoral heterogeneity and inter-observer variability. Especially, the distinction between grade II and grade III ependymomas is hardly reliable, which makes it difficult to use histopathological grading as a prognostic marker (7). Recently, nine distinct molecular subgroups were described based on global DNA methylation profiling with three subgroups in each of the three compartments (supratentorial region, posterior fossa and spinal cord). This new molecular classification correlates much better with clinical findings than histopathological classification, making it relevant for future clinical trials (23,25). Later, posterior fossa ependymomas of group A (PF-EPN-A) were further classified into two major subgroups comprising nine underlying subtypes (24). In addition, spinal ependymomas with MYCN amplification were suggested to establish an own molecular subgroup (8). Ependymal tumors can affect every age group, and location as well as prognosis are strongly associated with age. Ependymomas occurring in adults are rather located in the spinal cord, whereas ependymomas occurring in children or adolescents are more likely to be found in the brain. Prognosis is relatively poor for children aged 0-19 years and patients aged 65+ (35). In a retrospective multi-center analysis of 103 pediatric patients with WHO grade II/III intracranial ependymoma, 10-year overall survival (OS) was 50 ± 5%, whereas progression-free survival (PFS) was 29 ± 5%. Even a decade or more after diagnosis, OS and PFS decreased continuously, reflecting the poor long-term outcomes of pediatric ependymoma. Relapses were almost all confirmed to be recurrent ependymomas instead of radiation-induced secondary malignant gliomas and were most likely to occur at the primary site, with occasional spread to the spinal cord or other sites of the brain (20). These findings demonstrate that there is still a huge lack of appropriate therapy, especially when it comes to ependymoma relapses. Although relapse patterns and molecular biology of other brain tumors, such as medulloblastomas or glioblastomas (10,16,31), are already well investigated, there is only very limited insight into the molecular mechanisms of ependymoma progression and relapse. A recent study found methylation profiles of recurrent ependymomas to cluster into the same molecular subgroup as the corresponding primary tumor, indicating that molecular subgroups remain stable over time (25). However, further understanding of the histomorphology and molecular biology of ependymoma relapses is needed in order to establish effective therapy for ependymoma relapses, particularly in consideration of the new molecular classification (23,25).

Samples
Tumor samples were contributed by the Institute of Neuropathology, University Medical Center Hamburg-Eppendorf (n = 19), the Division of Neuropathology, University Hospital Basel (n = 4) and the Institute of Neuropathology, University of Magdeburg (n = 3). Inclusion criterion was the diagnosis of an ependymoma with at least one ependymoma relapse. Diagnosis of two cases turned out to be incorrect with one case being a papillary tumor of the pineal region and the other one being an IDH-mutated glioma after careful reevaluation. Altogether, we investigated 24 primary ependymal tumors and 38 corresponding relapsed ependymal tumors. The use of biopsy-specimens for research upon anonymization was in accordance with local ethical standards and regulations at the University Medical Center Hamburg-Eppendorf.
Age at diagnosis ranged from 0.4 to 62.7 years. Molecular subtype was determined by DNA methylation profiling for cases #1, #10 and #21 and by panel sequencing for case #2. For cases #4, #16 and #17, DNA methylation profiling was performed, but did neither match to a specific molecular subtype of ependymoma nor to one of the previously described CNS tumor DNA methylation classes (4). Still, careful histological reevaluation confirmed the diagnosis of ependymoma. For the remaining posterior fossa ependymoma cases #5-9 and #11-15, molecular subtype was inferred from the H3K27 trimethylation status (26). For the remaining supratentorial ependymoma case (#3), the molecular subtype was deduced from p65 status (27). Additionally, we confirmed by FISH analysis that RELA break apart is present in the primary as well as in one relapsed ependymoma sample for case #2 and #3 (data not shown). Copy number profiling for spinal ependymoma case #19 revealed MYCN amplification. Thus, this case was classified as SP-EPN-MYCN. For the remaining spinal ependymomas, molecular subgroup was not determined. Cases #2, #7, #8, #10, #11, #12, #15, #16, #19, #23 have been tested for the presence of histone H3K27M mutation and turned out to be negative. Altogether, the cohort consists of four supratentorial ependymomas, hereof three ST-EPN-RELA and one tumor with no match to a known molecular subgroup, 13 posterior fossa tumors, hereof six PF-EPN-A, five PF-EPN-B and two tumors with no match to a known molecular subgroup, and seven spinal tumors, including one SP-EPN-MYCN ependymoma and one SP-MPE ependymoma. Further epidemiological and clinical details of patients are summarized in Table 1. Tumor samples were obtained as formalin-fixed, unstained tissue sections from paraffin blocks. H&E staining as well as immunohistochemistry using antibodies against Ki67, H3K27me3 and CXorf67 were performed according to standard protocols.

Histomorphological and immunohistochemical evaluation
Using light microscopy, we scored cell density, necrosis and microvascular proliferation on H&E-stained sections on a scale from 1 to 3 (cell density) and from 0 to 3 (necrosis and vessel proliferation) for primary (n = 12) and corresponding relapsed tumors (n = 21). Furthermore, we compared Ki67 (n = 14), H3K27me3 (n = 19) and CXorf67 immunohistochemistry (n = 15) between ependymoma primary and corresponding relapsed tumors. Ki67 was scored as percentage of positively stained tumor cells. Non-tumor cells in the microenvironment including endothelial cells and immune cells served as internal positive controls for H3K27me3 immunohistochemistry. An unpaired t-test was used for statistical analysis of the collected scores.

Data processing
The DNA methylation dataset (GSE65362) contains Betavalues, the ratio of the methylated probe intensity and the overall intensity. To perform statistical analyses, these  Alive For case #4, case #16 and case #17, DNA methylation was analyzed, but did not match to a known molecular subtype. n.a. = not available.
values were converted into M-Values, that is, the log2 ratio of the intensities of methylated probe vs. unmethylated probe (5).

Copy number variation analysis
For copy number variation (CNV) analysis, the Bioconductor package ChAMP was used (34). It used the HumanMethylation450 data to identify copy number alterations (CNA) by utilizing the intensity values for each probe to count copy number (CN) and determine whether CNAs are present. Here, the CNAs were compared between all primary and relapsed tumor samples as well as for every subgroup separately. For visualization, a heatmap for all patients was generated by computing the log-fold change of intensity values between primary and relapsed tumors of the top 1000 loci. Conumee plots for individual samples were generated using a free online classifier tool (www.molec ularn europ athol ogy.org) (4).

Differential methylation
Differentially methylated probes (DMPs) were computed applying a linear model using the R/limma package. DMPs were determined for relapse tumors compared to primary tumors with paired samples for all patients and within each molecular subgroup.

Methylation status of different epigenomic substructures
Analysis was performed as previously described (30). For each probe, we have computed the average Beta-value across all primary tumors (n = 45) and across all relapse tumors (n = 48). The dataset contains in total 485 577 probes. After mapping the probes onto epigenomic substructures, we got the following number of probes for each substructure: CpG Island n = 150 254, N Shelve n = 24 844, N Shore n = 62 870, S Shelve n = 22 300, S Shore n = 49 197, undefined/unmapped n = 176 112. In the following, we excluded all undefined/unmapped probes from the analysis. Next, we compared such computed mean Beta-values between all primary and relapse samples for each substructure. Then, we repeated this comparison for each subgroup individually. A two-sided t-test of Beta-values was applied. Significance was defined as 0.05.

Fractions of hypo-and hypermethylated CpGs
Analysis was performed as previously described (30). For assessing fractions of hypo-and hypermethylated CpGs, only significantly methylated CpGs with P-values < 0.05 were used, revealing 58 505 significant probes. Here, we used the uncorrected P-value as we did not compare the probes with each other and just searched for small P-values without ranking. Out of 58 505 probes, we identified 21 746 hypomethylated and 36 759 hypermethylated probes, respectively.
Subsequently, the fractions of the substructures were computed, revealing percent values. The same analysis was performed for each subgroup. CpGs with log-fold changes < 0 were defined as hypomethylated and CpGs with log-fold changes > 0 as hypermethylated.

RESULTS
We first aimed at elucidating potential morphological changes occurring during relapse/progression of ependymoma. To answer this question, we compared H&E-stained sections of matched primary (n = 12) and relapsed ependymomas (n = 21) regarding typical features of malignancy ( Figure 1A,C,E,G,I,K,M,O). Cell density was evaluated on a scale from 1 (low) to 3 (high). Similarly, the extent of necrosis and vessel proliferation were evaluated on a scale from 0 (absent) to 3 (plenty). We observed that cell density increased noticeably in relapsed tumors ( Figure 1Q (Table 1) and material was available to score and compare Ki67 labeling for multiple samples from five such cases (Supporting Figure  S1B). In general, proliferation increased over time in all patients (Supporting Figure S1). Together, we conclude from these results that ependymoma cells proliferate faster in a relapse situation in comparison to cells of the corresponding primary tumors.
Loss of trimethylation at lysine 27 of the histone H3 gene and alterations in CXorf67 level, recently renamed into EZH Inhibitory Protein (EZHIP) (13), are important events in PF-EPN-A tumorigenesis (24,26). By immunohistochemistry, we evaluated the association between the expression of these markers and ependymoma relapse by comparing all available H3K27me3-(n = 19) and CXorf67-stained sections (n = 15) for the matched tumor samples. As expected, PF-EPN-A tumors showed a major loss of H3K27 trimethylation ( Figure 2E,F), but stained positive for CXorf67 ( Figure 2G-H

37
relapsed ependymoma samples as shown exemplarily for four cases from different anatomical sites in Figure 2.
As a next step, we aimed at investigating CNA, differential DNA methylation as well as methylation status and fractions of hypo-and hypermethylated CpGs within different epigenomic substructures in primary and relapsed ependymomas.
First, we exemplarily analyzed three cases from our cohort ( Table 1). CNV plots of primary and relapsed ependymoma samples (cases #2, #15) and two relapsed ependymoma samples from the same case (case #10) turned out to be very similar (Supporting Figure S2A). Also, based on their global DNA methylation, the same above-mentioned three

38
cases clustered together in a t-SNE plot, suggesting a similar global DNA methylation profile (Supporting Figure S2B). Finally, Nmyc overexpression was present in both relapsed tumor samples of case #19, a SP-EPN-MYCN ependymoma (Supporting Figure S2C).
In order to analyze these aspects in more depth, we used public DNA methylation data from 45 patients (25) (GSE65362). CNVs at ependymoma relapse compared to the corresponding primary tumor remained relatively stable as shown in a heatmap for each patient individually (n = 45), although some of the patients seemed to demonstrate more changes in CNV by comparing relapsed and primary ependymoma samples than the rest ( Figure 3A). Representative CNV plots for pairs of samples from each of the five represented subgroups in this series are shown in Figure 3B.
Differential DNA methylation within the chromosomes was analyzed for the whole cohort (n = 45) as well as for each of the represented subgroups ST-EPN-RELA (n = 11), PF-SE (n = 2), PF-EPN-A (n = 23), PF-EPN-B (n = 7) and SP-MPE (n = 2). No significant difference with respect to the overall DNA methylation was detected after multiple testing for ependymoma relapse compared to primary tumor samples, neither for the whole cohort ( Figure 4A) nor for any of the above-mentioned molecular subgroups ( Figure 4B-F). Figure 4 shows the log-transformed ratios of DNA methylation differences for each probe in ependymoma relapse compared to primary tumor samples.
Next, we investigated the methylation status of several different epigenomic substructures, including CpG islands, shelves and shores for both primary and relapse ependymoma samples. While the regions located up to 2 kb from CpG islands are termed shores (14), the 2 kb regions flanking the shores up-and downstream are termed shelves (3). Our analysis revealed significantly higher methylation values in relapsed compared to primary tumor samples within CpG islands when looking at the whole cohort ( Figure 5A, P = 6.04e-6) as well as at ST-EPN-RELA ( Figure 5B, P = 1.17e-6), PF-SE ( Figure 5C,  To address the question how these changes in DNA methylation look like in more detail and whether unchanged overall methylation values resulted from hypo-and hypermethylation at different sites, we particularly looked at hypoand hypermethylation in distinct subgroups and genomic regions. As shown in Figure 6A, DNA methylation changes mainly affect the CpG islands. Moreover, in line with the results shown in Figure 5, hypermethylation mainly occurs in CpG islands, whereas hypomethylation similarly happens in all regions ( Figure 6A). Regarding the distinct ependymoma subgroups, CpG changes appear to be similarly distributed over the genomic regions with most changes in the CpG islands and least changes in shelves for most of the subgroups ( Figure 6B). Finally, each subgroup shows a different pattern of fractions of hypo-or hypermethylated probes in the different epigenomic substructures ( Figure 6C-D), indicating that relapse of ependymoma proceeds differently on epigenetic level for every subgroup.

DISCUSSION
In our study, we focused on the investigation of histomorphological, immunohistochemical, chromosomal, and epigenetic differences during ependymoma relapse. We compared ependymoma primary and relapsed tumor samples of 24 patients histomorphologically and immunohistochemically and used public methylation data of 45 patients with primary and relapsed ependymal tumors (25) for chromosomal and epigenetic comparison. By hierarchical clustering of DNA methylation profiles, Pajtler et al. used their data to demonstrate that molecular subgroups remain stable at ependymoma recurrence (25), which we exemplarily confirmed for three of our cases (Supporting Figure S2). However, a comparison of primary and relapsed ependymomas on chromosomal and epigenetic level was not previously conducted, which is what we used the data for in the current work.
In contradiction to Pajtler et al. (25), an earlier study reported subgroup switching between the two posterior fossa groups A and B. More precisely, an analysis of seven group A and seven group B paired samples revealed switching of three group A and one group B posterior fossa ependymoma to the other subgroup at first recurrence, respectively. Here, the transcriptionally distinct subgroups were identified by hierarchical clustering of the top 5% of differentially expressed genes (883 genes in total) (12). However, more reliable classification tools for distinguishing the two posterior fossa ependymoma groups A and B, namely methylation profiling (25) and H3K27me3 immunohistochemistry, which has a sensitivity of 99% and a specificity of 100% in segregating PF-EPN-A from PF-EPN-B tumors (26), were discovered in the meantime. For this reason, we assume that the reported changes in subgroup rather reflected an incorrect classification of some of the cases either at diagnosis or at recurrence than a real subgroup switch. Indeed, the authors failed to show an effect of subgroup switching on clinical outcome, which one would expect given the clearly distinct clinical features of both subgroups (25). Admittedly, the small number of cases made it difficult to prove such a point. Still, consensus nonnegative matrix factorization in the earlier paper only demonstrated a concordance of 92.8% and 76.3% for group A and B posterior fossa ependymomas, respectively (12), strongly supporting our hypothesis.
It was reported previously that histological features such as cellularity, necrosis, vascular proliferation, as well as mitotic count taken together are associated with a dismal outcome in ependymoma. High proliferation indices were encountered in ependymal tumors that behaved in a more aggressive manner (28). We found a significant increase in cell density and proliferation in relapsed tumors compared to the corresponding primary tumors, hypothesizing that relapsed ependymomas behave more aggressively. However, we did not detect a significant change in the degree of vessel proliferation or necrosis, which could be attributed to the relatively small sample size.
It has been shown that H3K27me3 staining is globally and exclusively reduced in PFA ependymomas (26). Furthermore, a recent study found high levels of CXorf67 expression in almost all PFA subtypes except PFA-1f, a subtype that shows high frequencies of H3K27M mutations (24). Although only three out of six PFA ependymomas in this study were tested for the histone H3K27M mutation, these cases turned out to be negative, being in line with our observation that all PFA ependymoma cases stained negative for H3K27me3 and positive for CXorf67 in contrast to all other subgroups (Figure 2).
We did not observe any significant changes in H3K27me3 and CXorf67 during relapse in any of the paired samples. This is in agreement with the assumption that the molecular subgroup, based on DNA methylation profiling, remains stable at disease recurrence (25). Still, we did find some focal intratumoral differences within samples, especially regarding H3K27me3 staining, which probably reflect the high histological variability and intratumoral heterogeneity of ependymomas (7).
It should be noted that the public methylation data (25) we used to look for chromosomal and epigenetic differences do not represent all of the known molecular ependymoma subgroups. So, further research is needed with respect to the four molecular subgroups ST-SE, ST-EPN-YAP1, SP-SE and SP-EPN, that are not represented in this study, as well as for the additional molecular subgroups and subtypes of ependymoma that have subsequently been described (8,24).
CNV analysis did not reveal any significant differences between primary and relapsed ependymoma samples, which is in line with a former study using SNPs to investigate CNA in 11 matched primary and recurrent posterior fossa ependymomas. This study also proposed that transcriptomic and microenvironmental, rather than genomic changes, might be of relevance at recurrence. In particular, the authors referred to differences in the immunophenotype of group A and B posterior fossa ependymomas in general as well as within the respective subgroups at diagnosis and recurrence (12). However, Puget et al. (29) found increased genomic instability in a cohort of 26 relapsed tumors compared to 33 primary ependymomas. Among other imbalances, including gain of 1q and loss of 6q, gain of 9q34 was found to occur more often in relapsed ependymal tumors. Moreover, gains of 9q33 and 9q34 were significantly associated with tumor recurrence, age older than 3 years and posterior fossa location. It was suggested that two potential oncogenes, NOTCH1 and Tenascin-C, located on 9qter and found to be overexpressed might play a role in ependymoma progression. Additionally, a small number of posterior fossa ependymomas with increased genomic instability on chromosome 9, but also on other chromosomes at relapse were reported previously (9,32). Regarding these reported observations for individual patients, we assume, in light of our findings (Figure 3), that they cannot be generalized for a specific subgroup or for ependymomas at relapse.
Although overall DNA methylation did not change (Figure 4), CpG island methylation increased significantly in relapsed ependymomas for all represented subgroups except for PF-EPN-B and SP-MPE. In contrast, relapsed compared to primary ependymomas demonstrated significant hypomethylation of N Shelves in the subgroups ST-EPN-RELA and PF-SE and of S Shelves in ST-EPN-RELA ( Figure 5). While hypermethylation of CpG islands is widely considered to be characteristic for cancers (1,2,6), not much is known about methylation of CpG islands in relapsed tumors. Increased methylation of certain genes has been reported for AML at relapse (17) and aberrant DNA methylation for five genomic loci was shown for a cohort of 25 relapsed ependymomas before (36). Hypermethylation in general seems to occur more often in CpG islands, whereas hypomethylation rather affects other areas, including shores and shelves. This was demonstrated for sessile serrated adenomas with dysplasia, which can progress to colorectal cancer, compared to ordinary sessile serrated adenomas (18), for breast cancer compared to normal breast tissue (15) and for hepatocellular carcinoma, associated with hepatitis C virus or chronic alcohol abuse, compared to normal liver tissue (11,21,33). Furthermore, a recent study compared primary colorectal cancer with nonmatched liver metastases. Hypermethylated DMRs occurred more often in CpG islands than in shelves or shores in both primary colorectal cancers and liver metastases compared to normal colon mucosa, whereas hypomethylated DMRs where distributed equally among CpG islands, shores and shelves. Interestingly, hypomethylation seemed to occur continually in metastatic lesions, whereas hypermethylation remained similar between primary colorectal cancer and liver metastasis. Thus, it was speculated that hypomethylation in liver metastases is linked to drug resistance (22). As for our study, the relevance and impact of CpG island hypermethylation and hypomethylation of shelves in the relapse situation remain unclear and need to be further elucidated.
In conclusion, relapsed ependymomas demonstrate increased features of malignancy, such as increased cell density and high proliferation rate. Although we did find slight changes in the epigenome of relapsed tumors, namely CpG island hypermethylation, these findings cannot fully explain the altered tumor morphology. Thus, further research is needed to identify the molecular mechanisms of ependymoma progression.