Molecular Classification of Ependymal Tumors across All CNS Compartments, Histopathological Grades, and Age Groups

DNA Methylation Profiling of Ependymal Tumors Identifies Nine Distinct Molecular Subgroups

There is a growing body of evidence that DNA methylation patterns of tumor cells, specifically of promotor regions, represent a very stable molecular memory of the respective cell of origin throughout the disease course (Hoadley et al., 2014; Hovestadt et al., 2014), thus making this assessment particularly suitable for tumor classification. To attempt such a classification for ependymal tumors across all age groups and anatomical compartments, we generated genome-wide DNA methylation profiles for 500 ependymal tumors using the Illumina 450k methylation array. Clinical parameters of the entire cohort are presented in Figure S1 and Table 1.

Unsupervised hierarchical clustering of the DNA methylation data using 10,000 probes with highest SD across the entire dataset identified nine distinct subgroups of ependymal tumors, three within each major compartment of the CNS: spine (SP), PF, or ST (Figure 1A). Cluster analyses using different numbers of probes (5,000, 20,000, or 50,000) gave the same results (data not shown). Subgroup stability was further confirmed by bootstrap analysis (bootstrap probability 0.999; p = 0.001). Based on associations with anatomical location, histology, and genetic alterations, as outlined below, we annotated these nine subgroups as SP-SE (n = 7 samples), SP-MPE (n = 26), SP-EPN (n = 21), PF-SE (n = 33), PF-EPN-A (n = 240), PF-EPN-B (n = 51), ST-SE (n = 21), ST-EPN-YAP1 (n = 13), and ST-EPN-RELA (n = 88). One of the subgroups within each compartment was enriched with grade I SEs, and we therefore labeled them as SP-SE, PF-SE, and ST-SE. Interestingly, some grade II EPNs, confirmed after pathology re-review, also clustered with these SE subgroups (Figures 1A and S2A). We then turned our attention to the non-SE-like subgroups (a total of six) within each of the three major anatomical compartments. Within the spinal compartment, SP-MPE and SP-EPN showed a relatively good concordance with the histopathological subtypes MPE (grade I) and EPN (grade II) (Figure 1A). However, for the two remaining PF and ST subgroups, we found no concordance with histological grading. Within the PF, the two non-SE-like subgroups represented previously described Group A and Group B EPNs of the PF (Mack et al., 2014; Wani et al., 2012; Witt et al., 2011), which for standardization purposes we termed PF-EPN-A and PF-EPN-B. In the ST compartment, we suspected that the largest subgroup probably represented EPNs with recurrent C11orf95-RELA fusions, as described by Parker et al. (2014). Indeed, when we performed RNA sequencing and/or targeted analysis for the two major types of C11orf95-RELA fusion transcripts (Figure S2B; Table S1), we found these in 22 of 25 (88%) ST-EPN-RELA EPNs tested for RELA fusion types 1 and 2, suggesting that the majority of cases in this subgroup harbor these specific types of fusions. Interestingly, in one case of the ST-EPN-RELA subgroup, negative for RELA fusion types 1 and 2, we detected a PTEN-TAS2R1 fusion leading to a frame shift and subsequent disruption of PTEN (Table S1). Loss of PTEN expression/activity has previously been demonstrated to induce NF-kB activity through activation of Akt/ mTOR (Dan et al., 2008), suggesting that alternative mechanisms may exist that lead to activation of this key pathway in the ST-EPN-RELA subgroup. We did not systematically test for other less frequent fusions involving RELA as described by Parker et al. (2014) (RELA fusion types 3-7), which likely account for the majority of RELA fusion type 1 and 2 negative cases within the ST-EPN-RELA subgroup.

Molecular Subgroups Remain Stable at Relapse

To evaluate further whether molecular subgroup classification remains stable at disease recurrence, we performed hierarchical clustering of DNA methylation profiles of 45 primary ependymal tumors and their corresponding recurrent tumor specimens. In addition to primary tumors representing SP-MPE, PF-SE, PF-EPN-A, PF-EPN-B, or ST-EPN-RELA subgroups, the cohort comprised of 48 tumors from the time of first or second recurrence. Notably, inspection of cluster groups relative to patient identification number (1–45) showed that all recurrent ependymal tumors clustered without exception directly next to their corresponding primaries or very close to them, in all cases within the same subgroup that was attributed to the primary tumor (Figure 1B). These results strongly suggest that DNA methylation profiling remains a reliable tool for subgroup identification not only at initial diagnosis, but also at the time of recurrence.

YAP1 Gene Fusions Define a Distinct Subgroup of ST EPNs

As RELA fusions were identified exclusively in the ST-RELA-EPN subgroup, we next asked which genetic alterations might underlie the other ST EPN subgroup in children, especially as we identified several cases in this subgroup (9 of 13) with copy number aberrations on chromosome 11 around the YAP1 locus (Figure 2A). Parker et al. (2014) had already identified YAP1 fusions in two ST EPN samples that were negative for a RELA fusion. Therefore, it was tempting to speculate that these cases might represent a separate molecular subgroup of ST EPN. To test this hypothesis, we performed RNA sequencing for 7 of the 13 cases that clustered with this subgroup and for which sufficient high-quality RNA was available. Consistent with our findings on the DNA copy-number level, we detected the previously identified YAP1-MAMLD1 fusion (Parker et al., 2014) in six of seven tumors and another gene fusion containing YAP1 and an uncharacterized gene, FAM118B, in the remaining tumor (Figures 2B and 2C). All YAP1 fusions identified by RNA sequencing could be verified by Sanger sequencing (Figure 2B). In the YAP1-MAMLD1 fusions, exons 1–5 or 1–6 (out of 9) of YAP1 (according to reference sequence GenBank: NM_001130145) are fused in frame to exons 2–7 or 3–7 of MAMLD1 (according to reference sequence GenBank: NM_005491). In the other fusion, YAP1 exons 1–7 are fused in frame to the entire coding region encoded by exon 3–9 of FAM118B (according to reference sequence GenBank: NM_024556). In both fusion proteins, most of the N-terminal domains of the YAP1 protein are retained, and for MAMLD1, most of the C-terminal part is retained (Figure 2C). Importantly, no single YAP1 fusion was identified by RNA sequencing in any of 48 additional ependymal tumors of the other molecular subgroups of the PF or ST region. This demonstrates that YAP1 fusion-positive cases characterize another molecular subgroup of ST EPNs, annotated as ST-EPN-YAP1, that are distinct from RELA fusion-positive ST EPNs.

Molecular Subgroups Show Distinct Copy Number Profiles

Genome-wide DNA copy number profiles, generated by using the combined intensity values of the methylation probes (Hovestadt et al., 2013), showed strong differences in numbers and patterns of DNA copy number alterations (CNAs) between the molecular subgroups (Figure 3). Strikingly, gene amplifications, frequently seen in other brain tumors like medulloblastoma and glioblastoma, were almost totally absent in ependymal tumors. The same is true for homozygous gene deletions with the exception of CDKN2A deletions, which were commonly found in the ST-EPN-RELA subgroup (14 of 88; 16%), but not in other subgroups (p < 0.001). This subgroup also shows frequent loss of the entire chromosome 9 or chromosomal arm 9p (46 of 88; 52%), which is not seen in any other subgroup (p < 0.001). Most CNAs across all subgroups involved gains or losses of whole chromosomes indicating aneuploidy. Gains and losses of chromosomal arms were much less frequent. Interestingly, ST-EPN-RELA tumors, but not tumors from the other eight subgroups, frequently displayed dramatic copy number changes, reminiscent of chromothripsis (38 of 88; 43%), i.e., rearrangements of entire chromosomes or chromosomal arms with alternating copy number states (Korbel and Campbell, 2013; Parker et al., 2014; Rausch et al., 2012). Chromosome 11, where the C11orf95 and RELA genes are located, was always involved in these events, but sometimes other chromosomes were affected as well, yet to a much lesser extent (Figures 3 and S3; Table S2). Since Parker et al. (2014) extensively demonstrated based on whole-genome sequencing data that C11orf95-RELA fusions resulted from chromothripsis of chromosome 11q, we inferred that the observed massive fragmentations of chromosome 11q in ST-EPN-RELA tumors are based on this phenomenon. No chromothripsis was detected in tumors classified as ST-EPN-YAP1, even though YAP1 is also located on chromosome 11. Instead, these cases frequently displayed focal CNAs around the YAP1 locus (Figure 2A). Surprisingly, besides the CNAs on chromosome 11, the remainder of the genome is relatively stable in most ST-EPN-YAP1 tumors, which is in contrast to the ST-EPN-RELA EPNs that typically show an abundance of CNAs. Predominantly stable genomes were also seen in SEs of all three anatomical compartments (SP-SE, PF-SE, ST-SE). As expected, this was also true for the PF-EPN-A subgroup. The only CNA that was frequently observed in both SP-SE and PF-SE, but not ST-SE, was a complete or partial loss of chromosome 6, in line with previous findings (Kurian et al., 2008). Loss of chromosome 6 is also frequently seen in PF-EPN-B tumors, where it is the most frequent CNA (42 of 51 tumors, 82%), and to a lesser extent in SP-EPN, PF-EPN-A, and ST-EPN-RELA tumors. For PF-EPN-A tumors, the most frequent CNA observed was gain of 1q (60 of 240; 25%), which was also seen in PF-EPN-B (9 of 51; 18%), and ST-EPN-RELA tumors (21 of 88; 24%). Many more CNAs are seen in both SP-MPE and SP-EPN, with the most frequent event being loss of 22q in SP-EPN tumors (19 of 21; 90% show loss of 22q). This frequent involvement of the 22q locus is not surprising as it includes NF2, which is known to be frequently mutated in spinal EPNs (Ebert et al., 1999; Slavc et al., 1995). However, loss of 22q is not entirely restricted to SP-EPN, as it is also seen at lower frequencies in other subgroups (Figure 3). It remains to be seen whether NF2 also plays a role in these subgroups or whether other genes on 22q are involved. Finally, aside from ST-EPN-RELA tumors displaying chromothripsis, tumors in the PF-EPN-B subgroup show by far the highest degree of genomic instability, with many gains and losses of entire chromosomes or (less frequent) chromosomal arms in each tumor (Figure 3). Altogether, these data show that ependymal tumors of the nine molecular subgroups indeed represent genetically highly distinct entities.

Transcription Profiles Reveal Distinct Druggable Pathways in Molecular Subgroups

To investigate whether (epi)genetically defined molecular subgroups of ependymal tumors also show subgroup-specific transcriptional differences, we generated gene expression profiles for all tumor samples, for which sufficient high-quality RNA was available. This resulted in 209 gene expression profiles generated on the Affymetrix U133 Plus2.0 array. Unfortunately, ependymal tumors of the SP-SE subgroup were not included because no RNAs were available. Unsupervised hierarchical cluster analysis of the gene expression data almost perfectly recapitulated the molecular subgroups derived by DNA methylation profiling (Figure S4). To identify genes that are specifically (over)expressed in only one of the subgroups, we performed supervised analyses by comparing gene expression patterns of each subgroup against all other subgroups. This resulted in the identification of multiple genes that are either exclusively expressed or at least highly overex pressed specifically in one of the subgroups. A heat map showing the expression of these signature genes across all molecular subgroups is displayed in Figure 4A. In addition, gene expression patterns of two representative genes per subgroup are displayed (Figure 4B). The full list of signature genes (Table S3) provides a first hint of the underlying biology, pathway activation, and potential drug targets in the distinct molecular subgroups. For instance, KIT, targetable by several multikinase inhibitors, is only expressed at high levels in PF-SE tumors and not in other subgroups (Figure 4B). Another example is RELA, the transcription factor involved in activating the NF-kB pathway, which is only expressed in ST-EPN-RELA tumors (Parker et al., 2014). In order to get an even better insight in the biological processes and pathways that play a role in each of these subgroups, we performed gene-ranked pathway enrichment analysis with g:Profiler (Reimand et al., 2011) and used Cytoscape Enrichment Map to visualize processes and pathways that distinguished the different molecular subgroups or anatomical compartments (Figure 4C). Gene sets were compiled from the Gene Ontology (GO) biological processes, as well as Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and CORUM databases of pathways and protein complexes. Significant gene sets (false discovery rate [FDR] < 0.05, p < 0.01) (Table S3) were visualized as interaction networks (Figure 4C). Interestingly, while several gene sets involved in brain development clearly play a role in all subgroups, as expected, other gene sets were specific to certain subgroups. For instance, cell cycle genes, cell migration genes or genes involved in MAPK signaling were most active in PF-EPN-A tumors and ST-EPN-RELA tumors, while cAMP/carbohydrate metabolism or dopamine signaling genes were more active in ST-EPN-YAP1 tumors. Ciliogenesis genes were exclusively found in PF-EPN-B tumors. The clear differences in active biological processes and pathways identified between the various molecular subgroups of ependymal tumors, in line with previous data for the PF subgroups (Witt et al., 2011), further support their distinct origins and may suggest possible avenues for future subgroup-specific targeted therapies.

Molecular Subgroups Correlate with Clinicopathological Variables

The nine molecular subgroups of ependymal tumors described herein were closely associated with specific age groups (Figures 1 and S1; Table 1). All three SE-like subgroups were exclusively found in adults (18 years and older, range 22–76), with median ages of 49 years (SP-SE), 59 years (PF-SE), and 40 years (ST-SE), respectively. The other two SP subgroups, SP-MPE and SP-EPN, were also mainly found in adults, but some occurred in children (SP-MPE median age 32 years, range 9–66; SP-EPN median age 41 years, range 11–59). Patients in the PF-EPN-B subgroup were found mostly in the adolescent and young adult populations (median age 30 years, range 10–65). In contrast, tumors in the other PF subgroup, PF-EPN-A, were almost exclusively found in young children (median age 3 years, range 0–51). Only two patients in this subgroup were older than 18. The remaining two ST subgroups, ST-EPN-YAP1 and ST-EPN-RELA, were much more common in children (ST-EPN-YAP1 median age 1.4 years, range 0–51; ST-EPN-RELA median age 8 years, range 0–69), but a significant portion (23%) of the ST-EPN-RELA tumors was also found in adults. Gender distribution in our cohort revealed a preponderance of males over females (1.6:1). The gender distribution of all subtypes is presented in Table 1, with some significant differences between subgroups (p = 0.004). For example, among PF tumors, there was a preponderance of males represented in the PF-EPN-A (male:female = 1.8:1) and PF-SE (male:female = 3:1) subgroups, whereas in the PF-EPN-B subgroup, males were in the minority (male:female = 0.7:1).

Risk Stratification by Molecular Subgrouping Is Superior to Histopathological Grading

Survival analyses for all patients with available outcome data (n = 388) showed remarkable differences in OS and progression-free survival (PFS) between the molecular subgroups (Figure 5). Both the PF-EPN-A and ST-EPN-RELA subgroups show a dismal outcome, with 10-year OS rates of rv50% and PFS rates of rv20%. Notably, patients comprising these subgroups are mostly pediatric. All other subgroups have a much better outcome, with 5-year OS rates around 100% and 10-year OS rates ranging from 88%–100% (Table 1). It is well known that patients with ependymal tumors within the spinal region can usually be cured by complete neurosurgical resection alone. In our study, the number of spinal tumors was not sufficient to derive clinically meaningful conclusions, but the favorable OS rates (OS 100% for all three subgroups) are consistent with published literature (McGuire et al., 2009) (Figure S5). As one example for the clinical utility of molecular subgrouping over conventional histopathology, the SE-like molecular subgroups from the PF and ST compartments all showed favorable outcomes on OS analysis. Since a substantial proportion of molecular SEs was classified as grade II and even grade III EPN (Figures 1A and S2; Table 1), molecular classification in these instances reveals tumor subgroups with favorable outcome that would not be yielded by histopathological analysis. Notably, albeit OS rates are excellent for PF-SE and ST-SE, only SE tumors from the PF show frequent progression (Figure 5). This clinical course is reminiscent of PF-EPN-B tumors, which cluster close to the PF-SE subgroup, suggesting also some biological similarities between these two subgroups. Examining other clinical variables, for either compartment-specific or subgroup-specific prognostic value, showed that extent of resection was associated with OS in PF-EPN-A (p = 0.01) and with PFS in both PF-EPN-A (p = 0.0004) and PF-EPN-B (p = 0.019) (Figure S6A). Other clinical variables that showed prognostic value in univariate analyses across all subgroups included age (outcome of children is worse than adults) and chemotherapy (patients who received chemotherapy did worse than patients who did not receive chemotherapy; Table S4). However, both these factors were strongly biased by the molecular subgrouping as the two subgroups that showed the worst outcome, PF-EPN-A and ST-EPN-RELA, were also the subgroups that were mainly pediatric and in which most patients received chemotherapy as compared with the other subgroups. Within each of these subgroups, however, patients who received chemotherapy showed no significant survival improvement compared with patients who did not receive chemotherapy (data not shown). WHO grading and radiotherapy had no prognostic value in these univariate analyses, neither within anatomical compartments nor within molecular subgroups (Figure S6B; Table S4). In line with previous reports (Godfraind et al., 2012; Korshunov et al., 2010), gain of chromosome 1q also showed strong prognostic value in the present series. However, a highly significant difference in OS between tumors with and without gain of 1q was only seen for the PF-EPN-A tumors (p = 0.00001; Figure S6C), while no significant differences in OS were seen for PF-EPN-B or ST-EPNRELA tumors, even though all three subgroups displayed similar frequencies of 1q gain (PF-EPN-A: 25%; PF-EPN-B: 18%; ST-EPN-RELA: 24%; Figure 3). Surprisingly, however, 1q gain showed a strong prognostic association regarding PFS not only in PF-EPN-A but also in PF-EPN-B tumors, but not in the ST-EPN-RELA subgroup (PF-EPN-A: p = 0.0068; PF-EPN-B: p = 0.0075; Figure S6C).

Finally, we performed a multivariate Cox regression analysis to see whether molecular subgrouping, level of resection, and 1q gain are independent prognostic parameters for ependymal tumors. For comparison, we also included the clinical variables WHO grade and radiotherapy, even though they showed no prognostic value in the univariate analyses (Figure S6; Table S4). Results presented in Table 2 show that molecular subgrouping, level of resection, and 1q gain are all independent prognostic parameters for both OS and PFS. Additional parameters examined, including age and chemotherapy, did not have independent prognostic value in multivariate analysis. In addition, a likelihood-ratio test was performed to compare the full model including all variables with a multivariate Cox model that does not include molecular subgrouping. The resulting p values were p = 0.039 for OS and p = 0.012 for PFS, indicating that adding molecular subgrouping significantly improved the model fit. In contrast, comparing the full model with a model that omits WHO grading led to non-significant p values for OS (p = 0.79) and PFS (p = 0.56), indicating that WHO grading did not improve the model when other variables were already included (Table 2).

Tumor Material and Patient Characteristics

Clinical samples and data were collected after receiving written informed consent according to protocols approved by the institutional review boards, including University Hospital Heidelberg, NN Burdenko Neurosurgical Institute, The Hospital for Sick Children, University of Utah, Schneider Children’s Medical Center of Israel, University of Colorado Denver, University Hospital Brno, University of California San Francisco, NYU Langone Medical Center, University of Cambridge, University of Bonn, and MD Anderson Cancer Center. At least 80% of tumor cell content was estimated in all tumor samples by staining cryosections (~5 μm thick) of the piece from which nucleic acid extraction was performed with H&E. Diagnoses were confirmed by histopathological assessment by at least two independent neuropathologists, including a central pathology review at the departments of neuropathology at the MD Anderson Cancer Center, the University Hospital Heidelberg, or the University Hospital Bonn that utilized the 2007 WHO classification for CNS tumors. The following tumor types were included: SE (WHO grade I), MPE (WHO grade I), EPN (WHO grade II), and anaplastic EPN (WHO grade III). No patient underwent chemotherapy or radiotherapy prior to the surgical removal of the primary tumor. Detailed clinical and patient characteristics are shown in Figure S1 and Table 1.

Biostatistics

Independence of chromosomal aberrations along molecular subgroups was analyzed by chi-square tests, where p values have been computed by 100,000 Monte Carlo simulations. The distribution of gender within molecular subgroups was tested against a null proportion of 0.5 using exact binomial test based on Clopper-Pearson intervals. Distribution of survival times was estimated by using Kaplan-Meier estimates and compared with the log-rank test. Prognostic impact of covariates on PFS and OS was evaluated on the basis of hazard ratios and 95% CIs from Cox’s proportional hazards regression model. Multivariate Cox’s proportional hazards regression models were used to adjust effects for additional covariates. In addition, Firth’s correction was applied to account for the monotone likelihood problem arising due to total separation of group and events (Heinze and Schemper, 2001). To compare nested Cox proportional hazards models Likelihood-ratio tests were applied.