Genomic DNA methylation
850 K methylation analysis was performed on 23 of the 27 histologically-defined primary AB cases and 1 recurrent tumor (24 total samples). All but one sample (C21) yielded DNA methylation profiles of sufficient quality for subsequent analysis. Unsupervised hierarchical clustering and t-SNE analysis was performed by comparing the top differentially methylated probes first to a comprehensive reference series consisting of 2801 tumors representing all described DNA methylation classes (Additional file 1: Figure S1) [5]. This analysis was then reduced to a subset of 195 tumors consisting of: HGNET-MN1 (n = 21); PXA (n = 44); supratentorial ependymoma with C11orf95-RELA fusion (EPN-RELA; n = 70); supratentorial pilocytic astrocytoma and ganglioglioma (LGG-PA/GG-ST; n = 24); control reactive cortex (CONTR-REACT; n = 23); and control cerebral hemisphere (CONTR-HEMI; n = 13) (Fig. 1; Additional file 2: Figure S2). By unsupervised analysis, the ABs failed to cluster into a single group, and instead mostly distributed into previously defined DNA methylation classes. The results from hierarchical clustering and t-SNE analysis were concordant in 20 of 23 samples (Figs. 1 and 2). Eight tumors grouped with HGNET-MN1, seven with PXA, two with EPN-RELA, and one with LGG-PA/GG-ST. Two tumors clustered with either reactive cerebral cortex (C31) or control cerebral hemisphere (C17) (Fig. 1a). The latter was likely due to contamination from normal brain within the sample. An additional three tumors exhibited discordant grouping between the hierarchical clustering and t-SNE analysis. One of these (C19) was discordant between three methylation classes (CONTR-HEMI, CONTR-REACT, and LGG-PA/GG-ST). The other two tumors (C14 and C20) clustered together by hierarchical clustering, but not within any defined reference methylation class. They did, however, group with PXA by t-SNE analysis (Fig. 1b).
As an additional comparison, we also performed supervised analysis using the www.molecularneuropathology.org website, which employs a random forest methylation class prediction algorithm, using the comprehensive reference set used in the initial unsupervised clustering/t-SNE analysis [5]. For tumors with scores above the threshold values, the supervised analysis was concordant with the unsupervised methods (Fig. 2). Five tumors yielded a probability score below the reporting threshold of 0.90; however, in two of these (C22, and C30), the highest probability was consistent with the unsupervised hierarchical clustering analyses (LGG-PA/GG-ST 0.55 and EPEND RELA 0.89, respectively). Three additional cases (C14, C19, and C20) yielded unreliably low probability scores below 0.15 (Additional file 3: Table S1).
Orthogonal validation of DNA methylation groups
Next, we performed orthogonal molecular analysis to determine if the ABs in specific DNA methylation classes contained the hallmark mutations or gene rearrangements of lesions within those classes. Limited molecular evaluation was also performed on tumors for which insufficient DNA was available for methylation analysis. All eight samples clustering with the HGNET-MN1 methylation class demonstrated evidence of MN1 rearrangement by FISH (Fig. 2; Additional file 4: Figure S3). Support for an MN1 rearrangement was also found by FISH in two additional cases for which insufficient material was available for DNA methylation analysis or for which methylation analysis failed (C16 and C21, respectively).
The BRAFV600E mutation was identified in seven of nine tumors clustering in the PXA methylation group by t-SNE (Table 1). BRAFV600E mutations were not detected in the two tumors in which t-SNE and hierarchical clustering were discordant (C14 and C20) (Figs. 1 and 2). The two tumors that clustered within the EPN-RELA methylation class were also found to have evidence of RELA rearrangement by interphase FISH.
Chromosomal copy number
To further evaluate the relationship between ABs and their respective molecular groups, we evaluated the copy number profiles of ABs compared to the respective reference tumors. ABs clustering in the HGNET-MN1 group variably demonstrated loss of chromosomes 22q, 14, and broad regions of X (three of eight cases), similar to previous findings reported in a small cohort of MN1-rearranged ABs [10] (Fig. 3a, Additional file 5: Figure S4A). These findings were consistent with those of the reference cohort of HGNET-MN1 tumors, with the exception of a slightly increased proportion of chromosome 14 loss in ABs.
BRAFV600E-positive ABs that grouped with PXA showed more extensive chromosomal instability compared to other ABs (Fig. 3; Additional file 5: Figure S4A), with frequent gains of chromosome 5, chromosome 7, and chromosome 19; and loss of chromosomes 10, 18, and 6q. Unlike commonly observed in glioblastoma and in the approximately 20% of PXAs (especially anaplastic PXAs) with chromosome 7 and 10 aberrations [21], gain of chromosome 7 and loss of chromosome 10 were mutually exclusive in ABs (Additional file 5: Figure S4A). While interpretation of data for ABs clustering with EPN-RELA was limited by case number, observed copy number variations were compatible with findings in the reference cohort of RELA ependymomas (Fig. 3c).
Tumors showed few recurrent focal copy number abnormalities within or between DNA methylation classes. Loss of the CDKN2A/B locus was observed in one tumor each grouping with HGNET-MN1 and EPN-RELA (Additional file 5: Figure S4B). Three of seven tumors that both grouped with PXA and contained BRAFV600E mutations showed focal copy number loss at CDKN2A/B (Additional file 5: Figure S4B). Neither BRAF wildtype tumors grouping with PXA in the t-SNE analysis (C14 and C20) showed CDKN2A/B loss, nor did any of the other tumors lacking known driver mutations. No other recurrent focal copy number abnormalities were observed across any of the groups.
Histopathology
MN1-rearranged, BRAFV600E-mutant, RELA-rearranged, and tumors without identified driver mutations all occasionally demonstrated nuclear pseudoinclusions [15] (Fig. 4f and h). MN1-rearranged tumors more often showed vascular and/or generalized sclerosis. Three of the 10 MN1-rearranged tumors demonstrated marked sclerosis and contained hyalinized areas consisting of nearly entirely sclerotic vessels as depicted to the right in Fig. 4b. Although the most marked sclerosis was seen in this group, mild vascular sclerosis was also occasionally seen in tumors in the other molecular groups (Fig. 4e, j, and k). BRAFV600E-mutant ABs tended to have stouter cells (Fig. 4e-g); however, such cells were also occasionally seen in MN1-rearranged and other tumors (Fig. 4a).
When examining AB histological features which we previously cataloged [15], RELA-rearranged tumors often showed clear or signet ring-like cells (Fig. 4k and l); however, these were also observed in select tumors from the other types. Eosinophilic granular body–like structures and spheroid hyaline bodies were absent in RELA- and MN1-rearranged tumors, but were common in BRAFV600E-mutant tumors and tumors without known driver mutations (Fig. 4i and j). Rhabdoid-like cells were not identified in RELA-rearranged tumors, but were found in all other types. All but two MN1-rearranged tumors and both RELA-rearranged tumors lacked lymphocytic infiltrates, which were common in other types. Multinucleated cells were present in almost all tumors except two MN1-rearranged tumors, the low-grade RELA-rearranged lesion (C1) and one other lesion (C2). Overall, RELA-rearranged tumors tended to lack some histologic features common in most other groups; e.g., lymphocytic infiltrates, eosinophilic granular material, and rhabdoid-like cells.
Clinicopathologic correlates
Nine of the 10 consensus tumors with MN1 rearrangement (90%) presented in female patients aged 3–33 years and 1 was from an 8-year-old male (mean age, 12.8 years; median, 12 years) (Table 1). Eight of 10 MN1-rearranged tumors were olig2 immunopositive, and none showed hypermethylation of the O6-methylguanine methyl transferase (MGMT) gene promoter via the EPIC BeadChip (Table 1). Similarly, six of seven BRAFV600E-mutant ABs (87.5%) occurred in females aged 12–38 years and one presented in a 33-year-old male patient (mean, 25.9 years; median, 25 years) (Table 1). All were olig2 immunopositive. Two cases (C5 and C9), from patients 33 and 38 years-of-age, respectively, exhibited MGMT promoter hypermethylation, as previously described [15].
RELA rearrangements were detected in tumors from one female and one male (C1 and C30), aged 10 and 19 years, respectively. Olig2 immunohistochemical staining was equivocal in C1 and negative in C30. Neither showed MGMT promoter hypermethylation. Six other tumors negative for BRAFV600E mutations and MN1 or RELA rearrangements consisted of lesions from three females and three male patients ranging in age from 4 to 71 years. None showed MGMT promoter hypermethylation. All were olig2 immunopositive, except C14 [15].
Survival analysis
Survival analysis was statistically limited by the relatively small number of tumors in each group, but overall showed a significant difference between molecular groups (Mantel-Cox, P = 0.045; Fig. 5). In pairwise analysis, there was no appreciable difference in overall survival between ABs with BRAFV600E mutations and tumors without specific driver mutations (P = 0.398). There did appear to be a difference in survival between MN1-rearranged tumors and tumors without identified driver mutations (other tumors); however, this did not reach statistical significance (P = 0.056). There was, however, a clear and significant survival advantage for MN1-rearranged tumors compared to BRAFV600E-mutant tumors (P = 0.013; Fig. 5). In fact, all MN1-rearranged tumor patients in the cohort are currently alive, despite multiple tumor recurrences in some cases (Additional file 6: Table S2). Four deaths each occurred in the seven BRAFV600E mutation patients and in the eight patients without known driver mutations (Table 1). The overall survival of the MN1-rearranged tumor patients’ ranged from 68 to 221 months (mean, 138 months; n = 7) compared to 2 to 141 months (mean, 61 months; n = 7) for patients whose tumors had BRAFV600E mutation, and 18 to 279 months (mean, 127 months; n = 8) for patients with neither genetic alteration.