85 meningiomas were successfully analyzed using the Affymetrix Human Mapping 100k SNP arrays and 68 meningiomas were successfully hybridized independently onto U133 Plus 2.0 expression arrays. 65 meningiomas had data generated from both assays.
Chromosomal abnormalities are more common in higher grade meningiomas
The vast majority of chromosomal aberrations consisted of whole chromosome arm losses. Of the 85 tumors, 32% (27/85) had no observed chromosomal arm losses, and 68% (58/85) of these meningiomas revealed significant genomic abnormalities of one or more whole chromosomal arm losses. Consistent with previous descriptions, there was a strong association of increasing frequency of chromosomal losses with increasing histological grade () (22
). Of grade I meningiomas, 44% had no evidence of any chromosomal arm loss while 56% (32/57) had at least one whole arm loss. Among the grade I tumors, 33% (19/57) had a single chromosome arm loss, 9% (5/57) had two losses, and 14% (8/57) had three losses or more. In contrast, only 10% of grade II meningiomas had no detectable losses (2/20). Further, if a grade II tumor had a chromosomal loss (18/20), all had more than one chromosomal arm loss: 25% (5/20) had two chromosomal arm losses while 65% (13/20) of the grade II tumors had three chromosomal arm losses or more. All of the grade III meningiomas had three or more whole chromosomal arm losses. In aggregate, grade I meningiomas had a median of 1.0 chromosomal arm loss per tumor, grade II tumors had 3.0 chromosomal arm losses, and grade III tumors had a median of 9.5 chromosomal arm losses per sample demonstrating a strong correlation of genomic losses with increasing grade. Pair-wise Kruskal Wallis rank sum tests confirmed significant differences in chromosomal arm losses between grade I vs. II (p-value = 1.36 × 10−5
), grade II vs. III (p-value = 4.07 × 10−3
), and grade I vs. III (p-value = 1.39 × 10−5
Chromosomal arm losses are more common in higher grade meningiomas
From these data a pattern of molecular hierarchy is suggested in which increased aggressiveness of the meningiomas is caused by yet undetermined genes within the large chromosomal losses. The pattern of loss while not uniform reveals general patterns and reveals the relative importance of specific chromosomal loss events. Of the grade I tumors, the only recurring chromosome loss was the loss of chromosome 22q which was detected in 49% (). Grade II meningiomas displayed a higher frequency of chromosome 22q monosomy (85%), but additionally accumulated frequently recurring losses on 14q (60%) and 1p (55%). Grade III meningiomas acquired a much more diverse chromosome arm loss pattern but retained the core features of grade II tumors with frequent losses on 22q (75%), 1p (75%) and 14q (38%). Moreover, frequent losses of 18q (75%), 6q (63%), 10q (63%), 11p (50%), 7p (38%), and 4p (38%) were observed which were either not observed among the lower grades or were substantially more frequently appearing in the grade III tumors. The non-random nature of the chromosomal losses highlights these chromosomes for further genetic analysis in meningiomas, and implies that multiple genetic events are necessary in the generation of grade III meningiomas. Further, a common mutational mechanism is whole chromosomal arm loss implying the sequential loss of function of specific genes. Unfortunately, insufficient smaller chromosomal arm losses were observed to resolve the location of these genes. Moreover, there were no statistically significant associations detected between genetics and location (frontal, parietal, occipital, et al. etc) or side (left, right, bottom). In aggregate, these observations indicate a mutational hierarchy with chromosome 22q loss occurring as the primary event followed by chromosome 1p and/or 14q loss, and then mutations of genes on 18q, 6q, 10q, 11p, 7p, and 4p. The cumulative genetic mutations that lead to grade III histologic features and higher propensity for recurrence are thus multiple.
Molecular correlates of recurrent meningiomas
Given the retrospective nature of the current study and insufficient length of follow up of individual patients (post gross or sub-total resections), we are not able to comment meaningfully on true predictors of recurrence. However, we were able to compare tumors observed to be recurrent to those tumors at first presentation which identified a higher number of total chromosomal arm losses among recurrent tumors relative to newly diagnosed tumors (Kruskal wallis p-value = 7.7×10−3). Only eight primary samples analyzed in this dataset were sampled from patients who had a later recurrence. None of the recurrent biopsies were available for analysis. After comparing these 8 primary samples with the remaining set of primary tumors that did not yet relapse, none of the specific chromosomal losses that we found among our recurrent tumors were determined to be significantly enriched among these eight primary tumors over the 17 primary tumors that did not recur (Fisher’s exact test p-value > 0.05). These results suggest that chromosomal losses observed to be more common in the recurrent tumors may be acquired during disease progression rather than being present at the time of initial diagnosis.
Within our dataset and as expected, Ki-67 labeling was highly related to the WHO grade. Ki-67 expression is defined as percentage of positive cells. The mean and median expression of Ki-67 in WHO grade I was 0.27 and 0.17, respectively. The mean and median expression of Ki-67 in WHO grade II was 1.19 and 0.53, respectively. The mean and median expression of Ki-67 in WHO grade III was 5.21 and 3.25, respectively. There was a statistically significant difference between the Ki-67 expression across the three meningioma grades. Also as expected, there was a significant correlation between Ki-67 positivity and recurrence (n = 85 tumors, Kruskal-Wallis (chi-square = 11.9256, df = 1, p-value = 5.5×10−4). High or medium Ki-67 expression was significantly more common in the recurrent tumors relative to the newly diagnosed meningiomas (n = 85 tumors, p-value = 4.0×10−3). However, when Ki-67 levels were used to distinguish between grade II and grade III newly diagnosed meningiomas versus the recurrent meningiomas, Ki-67 expression failed to classify which of the grade II or III samples were recurrent meningiomas (p-value = 0.44). We use this as a proxy of the ability to classify recurrence status to search for evidence of underlying molecular mutations with recurrence. Based on chromosomal loss patterns, recurrent grade II and III meningiomas had an average of 8.3 chromosomal arm losses versus 3.9 chromosomal losses observed among grade II and III newly resected meningiomas (Kruskal-Wallis p-value = 0.0077). Seven chromosomal arm losses were significantly more common in the recurrent meningiomas: 18q (Fisher's Exact two-tailed p-value = 0.0002), 6q (p-value = 0.0021), 10q (p-value = 0.0099), 16q (p-value = 0.017), 2p (p-value = 0.017), 14q (p-value = 0.017), and 18p (p-value = 0.03). In order to assess if the great increase in chromosomal anomalies in the recurrent samples were potentially due to artifacts of brain irradiation induced chromosomal loss, we compared recurrent tumors that had not received radiation with non-recurrent tumors not treated with radiation and identify that 6q loss was enriched in recurrent samples (p-value = 8.0 × 10−4) and 14q loss was enriched in recurrent samples (p-value = 2.8 × 10−3). This indicates that some of the genetic abnormalities in the recurrent tumors are not artifacts introduced by radiation. These data suggest that chromosomal arm loss of a series of specific chromosomes is a more reliable indicator of recurrence than Ki-67 staining.
To further explore the unique aspects of the recurrent meningiomas, we filtered genes that were differentially expressed between the recurrent tumors and those of comparable grade at first diagnosis. The recurrent samples significantly over-expressed four cell cycle genes (MLF1IP, CKS2, CDC2, and PRC1) (68 tumors, fold change ≥ 2.0, p-value ≤ 0.01, FDR 7.1%) while the newly diagnosed tumors significantly over-expressed 116 other genes (). However, network analysis did not produce significantly enriched map elements nor networks consistent for any significant functional theme (significance threshold 0.01, p-value = 0.05; Supplementary Information S2
). Since Ki-67 is an excellent marker of proliferation, we also used the genome-wide expression data to search for genes correlated with Ki-67 levels in meningiomas. Ten genes were significantly over-expressed in the mid and high Ki-67 positive meningiomas which were significantly over represented with cell cycle related genes RRM2, NCAPG, MLF1IP, CCNB1, CKS2, CDC2, BUB1B, NUSAP1, and PRC1 (data not shown). Conversely, there were only three genes (FZD7, AASS, and C11orf41) that were significantly over-expressed in the low Ki-67 positive tumor samples.
Identification of gene expression correlated with recurrent meningioma
Tumor chromosome 18q loss is more common in females with meningiomas than males
With the clinically observed higher frequency of meningiomas in females relative to males, we explored if there were any significant chromosomal loss differences in the tumors from females versus males. Chromosome 18q loss was detected in 8 of 16 females and 2 of 18 males in samples with a chromosomal deletion not solely restricted to 22 monosomy (Fisher’s Exact test p-value = 0.02). These data may indicate alternate pathways of tumorigenesis in meningiomas dependent on the sex of the affected individual for yet undetermined reasons.
Gene expression analysis identifies five main types of meningiomas that correlate broadly with histological grade and chromosomal loss
Meningiomas are histologically quite varied in appearance even within histologic grades. We sought to identify underlying molecular themes in an unsupervised manner to reveal the complexity of meningiomas even within this modest sample set. For these analyses, we included all 68 samples for which gene expression data were generated. 1,316 genes were produced from the high CV (≥ 1.0) filtering criteria for expression values greater than the lowest quartile in 20 percent or more of the samples. Unsupervised clustering of the 68 meningiomas revealed three branches of samples organized by their gene co-expression in a hierarchical clustering dendrogram. The first two sample branches revealed highly cohesive expression clusters, while the third sample branch was more heterogeneous and the samples only loosely related to each other. Thus, we reanalyzed the samples in the third branch by hierarchical clustering of probesets that had high CV (≥ 0.8) with expression values greater than 20 in at least 20 percent or more of these samples. Three distinct hierarchical subgroups were identified within this third branch based on gene expression. The first two sample branches are labeled “group 1” and “group 2”, while the remaining three sample branches are labeled “group 3”, “group 4”, and “group 5”. An aggregate set of 355 probesets (302 genes) were identified (Supplementary Information S3
) and combined to form a five group expression signature panel (). In order to determine how the 302 genes from our 5 group expression panel compared to previously published studies, we found that 26/55 genes up-regulated in grade III over grade I meningiomas from Carvalho et al were found among our 5 group expression classifier (hypergeometric p-value = 0.01) (Supplementary Information S3
). In addition, 24/62 genes down-regulated in grade III over grade I meningiomas were also found among our 5 group expression classifier (hypergeometric p-value = 0.0096). Carvalho et al did not identify specific grade II meningioma expression signatures over grade I or grade III tumors due to what they explained as heterogeneity amongst grade 2 meningioma genomic expression. We concur that grade II meningiomas display broad expression heterogeneity as the majority of our grade 2 meningiomas (9/14 = 64%) comprise the bulk of expression groups 3, 4, and 5 which prompted our ad hoc secondary analysis. To further confirm the presence of five gene expression based groups among meningiomas, we collected publicly available meningioma microarray samples to see if our 302 differentially expressed genes could identify the 5 meningioma groups in other studies. 87 additional meningioma microarrays of varying grades (WHO I = 32, WHO II = 20, WHO III = 3, WHO unknown = 32, Supplementary Information S1
) were collected from two separate institutional studies (GEO accession GSE4780 and GSE9438) and clustered with our samples based on the five group 302 gene panel. Hierarchical agglomerative clustering was able to confirm the presence of five distinct expression groups (Supplementary Information S6
Five group gene expression based categorization of meningiomas
The 5 gene expression based meningioma groups correspond with histological grade (Kruskal Wallis p-value = 3.6×10−3
), however, samples did not track precisely with ascending grade and ascending gene expression group assignment (). Expression groups 1 and 2 were predominantly populated by grade I samples (30/42), while group 3 was much more heterogeneous and highly enriched in grade II and grade III tumors (17/23). In particular, group 1 consisted solely of grade I tumors (WHO I: 100%), while group 2 contained grade I tumors by a large majority (WHO I: 84%) in addition to minor amounts of higher grade samples (WHO II: 11%, WHO III: 5%). Group 3 was more heterogeneous with respect to histological grade as it contained fewer grade I tumors than in groups 1 and 2 (70% vs. 100% and 84%, respectively) and more grade II (20%), and grade III (10%) meningiomas. Group 4 predominantly contained grade II meningiomas (58%) and minor populations of grade I (33%) and grade III (8%) tumors. Finally, group 5 had the largest proportion of grade III meningiomas (18%) and a substantial number of grade I (47%), and grade II (35%) tumors. Thus, while a trend was evident, there was no perfect correspondence of higher grade tumors with the gene expression categories (Supplementary Information S1
). However, while there existed no significant correlation between ascending gene expression categories with age (Kendall’s rank correlation p-value = 0.86), a significant correlation was found with ascending gene expression categories and gender. Lower gene expression categories appeared to demonstrate bias towards female tumor enrichment while male tumors appeared more frequently in higher gene expression categories after controlling for initial gender selection bias (group 1 (female:male ratio = 2:1), group 2 (female:male ratio = 3:1), group 3 (female:male ratio = 1:1.6), Kendall’s rank correlation p-value = 0. 01).
Metacore network pathway analysis identified 30 significant networks among the 168 genes distinctly expressed in groups 1 and 2 with the 129 genes distinctly expressed in group 3 (Hypergeometric p-value range: 2.04 × 10−39
to 9.69 × 10−4
, Supplementary Information S3
). Almost one quarter of these networks (23%) were found to involve a kinase signaling pathway where group 1 and 2 meningiomas were found to express a different set of kinase pathway members (MYLK, PRKD1, NTRK2, ROR1, TNIK, and PRKG1) than those expressed by group 3 meningiomas (EPHA3, DCLK1, PDK1, MET, EPHA7, INSR, and ABP1). Moreover, the remaining 23 networks demonstrated GO processes spanning cancer related themes: regulation of cell death (11/50, hypergeometric p-value = 8.7 × 10−8
), regulation of cell proliferation (20/50, p-value = 3.5 × 10−15
), developmental regulation: anatomical structure morphogenesis (11/50, p-value = 1.5 × 10−11
), nervous system development (22/50, p-value = 1.2 × 10−6
), organ development (8/50, p-value = 8.8 × 10−19
) and several metabolic processes: protein amino acid O-linked glycosylation via threonine (7/50, p-value = 1.8 × 10−3
), regulation of adenylate cyclase activity (14/50, p-value = 1.7 × 10−10
), and fatty acid biosynthetic process (11/50, p-value = 2.6 × 10−8
The expression based groups also clustered all of the recurrent meningiomas among group 3 (Fisher’s exact p-value = 3.0×10−4) however, this is likely due to the preponderance of higher grade tumors clustering among this expression group as 8/9 recurrent tumors were grade II or grade III from the outset. Between the five gene expression based groups, the chromosomal loss patterns observed across the groups were strikingly different (). In aggregate, the median number of chromosome arms lost per group was significantly smaller in groups 1 and significantly higher in group 3 than the others (group 1 = 0, group 2 = 1.0, group 3 = 2.5, group 4 = 2.5, and group 5 = 6.0; Kruskal-Wallis rank sum test p-value = 1.2×10−3).
Median number of chromosome losses varies by gene expression-based group
Combined use of gene expression data and chromosomal loss data to identify potential tumor suppressors
We attempted to use the combined SNP/gene expression data to suggest candidate tumor suppressor genes within the more commonly deleted chromosomes other than chromosome 22. First, we determined the minimal regions of common loss on the commonly deleted chromosomal regions among the grade III samples, the grade III and II samples, and grade II and grade I samples using the “minimal loss rank ordering” and “drill down” features systematically with Nexus copy number analysis software. Shared minimal loss regions were easily identified through Nexus’ output files (Supplementary Information S4
). We then hypothesized that the causative genes that are mutational targets are more likely to be down-regulated in expression either by mutations on the non-deleted chromosome (that would lead to a premature stop and lower the mRNA abundance through nonsense mediated decay) or by smaller scale deletions/mutations invisible by the SNP arrays that lead to complete lack of the mRNA. Using this combination approach, we highlight a series of genes within the chromosomal regions more frequently deleted in the grade III tumors (Supplementary information S4
). As an example, the minimal region of loss on 11q (11q23.1–q23.2) contains 42 genes. However, only one of these genes, TIMM8B was lower in expression in the 3 deleted samples than in the other non-deleted grade III tumors. No mutations of this gene in meningioma have yet been described. Similarly, 20p was deleted in 3 of the grade III tumors and contained 21 genes, but only one, C20orf7, was down-regulated in the grade III samples relative to the lower grade samples. Using this approach to search for mutated genes, we looked at genes that may be highlighted by comparing chromosome losses between expression group profiles which highlighted CDC14A at 1p21 in a deleted interval between 1p36.13–1p36.11. Decreased expression of CDC14A was identified from comparing expression differences among the 44 genes at this locus between group 4 and 5 samples carrying the deletion (12/17) and group 4 and 5 samples (5/17) lacking the deletion. This is an attractive candidate tumor suppressor in meningiomas based on previously published functional data (18
Identification of meningioma specific genes
The determination of five separate expression-based subclasses and the tremendous variability of chromosomal losses between meningiomas analyzed here highlight the diversity of molecular types of meningiomas and prompted us to determine if there were features shared in common across all meningiomas relative to other brain cancer types and normal brain and non-brain tissues. We thus compared the 65 meningioma expression samples to available data from 533 other expression samples which consisted of various glioma samples from prior published work of ours and other colleagues (7
), which are available and co-normalized within the Celsius database. Comparing meningiomas to this panel of 389 gliomas and 144 normal tissues revealed a large set of genes, which can robustly differentiate all of the meningiomas from any of the other neoplastic and non-neoplastic tissues. There are a surprisingly large number of genes (n = 4,912) that were identified as higher expressed in meningiomas than in normal tissues and gliomas (fold change = 2.0, p-value = 0.05) with an overall FDR of 0.1%. In order to identify genes strongly and specifically correlated with meningiomas, regardless of grade, genes were filtered based on a fold increased expression in meningiomas of at least 10 and p-value less than 0.001. These criteria identified 130 highly meningioma specific genes (FDR 0%) relative to all of the other tissues (Supplemental Information S5
). These genes were consistently expressed across all 65 meningiomas indicating many common features of the histologically and genetically diverse panel of meningiomas (). To better characterize this observation, these genes were analyzed to identify enriched biological themes using the network analysis mining tool available through MetaCore. 130 MetaCore objects were recognized and thirteen annotated map themes contained statistically significant ontological terms among these genes. Prominent themes consisted of Development: WNT signaling pathway (5/53 map objects; hypergeometric p-value = 1.69 ×10−2
), cytoskeleton remodeling: role of PDGFs in cell migration (3/21 map objects; p-value = 2.12×10−2
), and GTP metabolism (4/40 map objects; p-value = 2.66 ×10−2
). Finally, transcripts that encoded known secreted proteins were identified as potential targets for blood based diagnostics and are listed in Supplementary Information S5
Figure 5 Directed comparison of all meningioma samples relative to combined glioma and normal tissues identifies meningioma specific genes. Normal tissues (n = 144) and gliomas (n = 389) were compared against meningiomas (n = 65) with a 10.0 fold and significance (more ...)