Multiomic and clinical analysis of multiply recurrent meningiomas reveals risk factors, underlying biology, and insights into evolution

Abstract

An important subset of meningiomas behaves aggressively and is characterized by multiple recurrences. We identify clinical, genetic, and epigenetic predictors of multiply recurrent meningiomas (MRMs) and evaluate the evolution of these meningiomas in patient-matched samples. On multivariable binomial logistic regression, MRMs were significantly associated with male sex (P = 0.012), subtotal resection (P = 0.001), higher number of meningiomas on presentation (P = 0.017), and histopathological sheeting (P = 0.002). Multiomic analysis of primary meningiomas revealed that MRMs have greater copy number losses (P = 0.0313) and increased DNA methylation (P = 0.0155). In meningioma cells with knockdown of EDNRB, a locus with greater promoter methylation and decreased gene expression in MRMs had increased proliferation (P < 0.0001). MRM recurrences were found to be similar to primaries but have a greater burden of copy number gains (P < 0.0001) and increased methylation (P = 0.0045). This clinical and multiomic investigation of MRMs harbors implications for the future development of biomarkers and therapeutic agents for these challenging tumors.

Clinical and mutiomic analysis reveals key biomarkers and evolution of multiply recurrent meningiomas.

INTRODUCTION

Although meningiomas are typically considered benign lesions, a subset exhibits an aggressive clinical course. Treatment of symptomatic primary meningiomas includes surgical resection with or without adjuvant radiation depending on lesion size, location, extent of resection, the World Health Organization (WHO) grade, and treating team preference. Patient characteristics previously associated with meningioma recurrence following treatment include older age, male gender, Simpson grade, and subtotal resection or no surgical procedure (1, 2). On imaging, meningioma recurrence has been negatively associated with focal or diffuse calcification and positively associated with peritumoral edema, larger tumor volume, and proximity to a major sinus (3, 4). From a histopathological perspective, brain invasion, intra-tumoral spontaneous necrosis, and high mitotic index have been associated with increased recurrence risk and shorter disease-free survival (5–7). However, these studies do not distinguish tumors characterized by a single recurrence that respond well to treatment and are thus clinically manageable from truly aggressive multiply recurrent tumors resistant to multiple treatments, which are especially challenging to manage.

The WHO grading system for meningiomas includes grades 1 to 3, with WHO grade 3 lesions having the worst prognosis. The 2021 update to the WHO classification of meningiomas added TERT promoter mutation or CDKN2A/B homozygous loss as independently classifying a meningioma to be WHO grade 3 (8). Several groups have proposed alternative classification systems of meningiomas according to molecular criteria, including transcriptomic, genomic, and epigenomic information (9–12). Following an integration of these classifiers, molecular classification yielded three major groups—MenG A, MenG B, and MenG C—with MenG A exhibiting the best prognosis and MenG C the worst, as defined by time to first recurrence (13). However, these molecular classification schemas do not account for MRMs or identify molecular features predictive of multiple recurrences. Furthermore, molecular analysis of meningiomas has focused on primary tumors, and, thus, there is a paucity of longitudinal molecular data on patient-matched primary and recurrent meningiomas.

Thus, although the meningioma literature has identified many clinical and molecular markers of early recurrence in meningiomas, the clinical and molecular predictors of aggressive MRMs remain to be identified. In addition, whether MRMs are molecularly distinct from less aggressive tumors at initial diagnosis and whether the molecular features of MRMs evolve with subsequent recurrences are fundamental questions that have not yet been addressed. In this dual-institution study, we aim to identify (i) clinical and molecular features associated with MRMs and (ii) molecular profiles for patient-matched primary and recurrent MRMs.

RESULTS

Clinical predictors of MRMs

A total of 1186 patients with primary meningiomas were retrospectively reviewed in this study. Thirty-one of these primary tumors went on to be multiply recurrent, defined by ≥2 radiologic recurrences following treatment, and 1155 remained nonrecurrent (Fig. 1A). Patients with MRM were significantly associated with longer follow-up time (P < 0.001) and higher WHO grade (P < 0.001) compared to patients with nonrecurrent meningioma (NRM). There was a relatively equal distribution of patients with MRM and NRM from each institution before matching (P > 0.999; Table 1). To compare these groups of meningiomas, we thus performed propensity score matching (PSM) to match patients with MRM and NRM based on follow-up time, WHO grade, and institution.

Fig. 1. PSM of patients from two institutions reveals clinical predictors of MRMs.

(A) Schematic representing patient institutional distribution before and after PSM. (B) Histogram depicting WHO grade distribution for nonrecurrent meningioma (NRM) and multiply recurrent meningioma (MRM) groups before and after propensity score matching (PSM). (C) Density plot depicting the follow-up time (months) distribution for NRM and MRM groups before and after PSM. (D) Representative non-contrast head computed tomography and associated T1-weighted post-contrast magnetic resonance imaging highlighting dystrophic calcification and bone invasion. (E) Clinical characteristics of 115 patients with meningioma included in clinical cohort after PSM separated by multiple recurrent (left, dark green) and nonrecurrent (right, light green) meningiomas. (F) Univariate analysis revealed a significant association between MRMs and male sex (*P = 0.042), subtotal resection (***P < 0.001), sheeting (***P < 0.001), macronucleoli (*P = 0.035), necrosis (**P = 0.009), hypercellularity (*P = 0.015), and bone invasion on imaging (*P = 0.015). NRMs were significantly associated with psammoma bodies (**P = 0.005) and dystrophic calcification on imaging (**P = 0.006).

Table 1. Propensity score matching variables.

Values are presented as median (interquartile range) or n (%). Statistical tests performed: Mann-Whitney U test, Fisher’s exact test, and chi-square test of independence. SMD, standardized mean differences. *P ≤ 0.05.

	Unmatched (n = 1186)			Matched (n = 115)
	Multiply recurrent (n = 31)	Nonrecurrent (n = 1155)	P value	Multiply recurrent (n = 31)	Nonrecurrent (n = 84)	P value
WHO grade			<0.001*			0.586
1	11 (35.5)	825 (71.4)		11 (35.5)	38 (45.2)
2	14 (45.2)	295 (25.5)		14 (45.2)	34 (40.5)
3	6 (19.4)	35 (3.0)		6 (19.4)	12 (14.3)
Follow-up time (months)	82.9 (57.6, 115.2)	25.0 (6.13, 60.0)	<0.001*	82.9 (57.6, 115.2)	83.4 (31.51, 126.7)	0.451
Institution			>0.999			>0.999
A	11 (35.5)	395 (34.2)		11 (35.5)	28 (33.3)
B	20 (64.5)	760 (65.8)		20 (64.5)	56 (66.7)

Following 1:3 nearest neighbor PSM, 31 patients with MRM were matched to 84 patients with NRM for a total of 115 patients. Follow-up time, WHO grade, and institution were well-balanced after PSM (Fig. 1, B and C). Following PSM, there were no longer significant differences between MRM and NRM groups for follow-up time (P = 0.451), WHO grade (P = 0.586), and institution (P > 0.999; Table 1). All standardized mean differences (SMDs) were less than 0.1 except for WHO grade (SMD of 0.211). In the total PSM cohort, 66.1% (76 of 115) of patients were from institution B, median age at diagnosis was 51 years (interquartile range, 48.5 to 65.0), 67.8% (65 of 115) were female, 82.6% (95 of 115) were white, 90.4% (104 of 115) were non-Hispanic, and 39.1% (45 of 115) had convexity meningiomas. Of MRMs, median number of recurrences was 2, mean was 2.77, and the range was 2 to 7 (Table 2 and data S1).

Table 2. Characteristics of propensity score–matched dual-institution cohort.

Values are presented as median (interquartile range) or n (%). Statistical tests performed: Mann-Whitney U test, Fisher’s exact test, and chi-square test of independence. HPFs, high-power field. *P ≤ 0.05.

Characteristic	Total (n = 115)	Multiply recurrent (n = 31)	Nonrecurrent (n = 84)	P value
Institution				>0.999
A	39 (33.9)	11 (35.5)	28 (33.3)
B	76 (66.1)	20 (64.5)	56 (66.7)
Age at diagnosis	51 (48.5, 65.0)	56 (44.0, 65.0)	49.5 (37.8, 66.0)	0.162
Follow-up time (months)	82.9 (43.2, 123.3)	83.4 (31.51, 126.7)	82.9 (57.6, 115.2)	0.451
Sex				0.042*
Male	37 (32.2)	15 (48.4)	22 (26.2)
Female	65 (67.8)	16 (51.6)	62 (73.8)
Race				0.245
White	95 (82.6)	24 (77.4)	71 (84.5)
Black/African American	14 (12.2)	5 (16.1)	9 (10.7)
Asian/Pacific Islander	3 (2.6)	2 (6.5)	1 (1.2)
Other	3 (2.6)	0 (0)	3 (3.6)
Ethnicity				>0.999
Hispanic	11 (9.6)	3 (9.7)	8 (9.5)
Non-Hispanic	104 (90.4)	28 (90.3)	76 (90.5)
Number of meningiomas				<0.001*
1	22 (89.6)	22 (89.6)	81 (96.4)
2	2 (3.5)	2 (3.5)	2 (2.4)
3	4 (12.9)	4 (12.9)	0 (0)
4	1 (3.2)	1 (3.2)	0 (0)
5	1 (3.2)	1 (3.2)	0 (0)
6	0 (0)	0 (0)	1 (1.2)
7	0 (0)	0 (0)	0 (0)
8	1 (0.9)	1 (0.9)	0 (0)
Adjuvant radiation	22 (19.1)	9 (29.0)	13.0 (15.4)	0.101
Symptoms on presentation
Altered mental	18 (15.7)	5 (16.1)	13 (15.5)	>0.999
Status/cognitive decline
Cranial nerve palsy	4 (0)	2 (6.5)	2 (2.4)	0.294
Dizziness	12 (10.4)	2 (6.5)	10 (11.9)	0.509
Gait disturbance	16 (13.9)	6 (19.4)	10 (11.9)	0.364
Headache	41 (35.7)	8 (25.8)	33 (39.3)	0.263
Muscle weakness	16 (13.9)	6 (19.4)	10 (11.9)	0.364
Neck pain	3 (2.6)	0 (0)	3 (3.6)	0.562
Seizure	20 (17.4)	5 (16.1)	15 (17.9)	>0.999
Sensory disturbances	6 (5.2)	2 (6.5)	4 (4.8)	0.660
Speech disturbance/aphasia	6 (5.2)	3 (9.7)	3 (3.6)	0.341
Tremor	3 (2.6)	2 (6.5)	1 (1.2)	0.176
Vision disturbance	28 (24.4)	13 (41.9)	15 (17.9)	0.015*
Vomiting	9 (7.8)	1 (3.2)	8 (9.5)	0.441
Location
Convexity	45 (39.1)	11 (35.5)	34 (40.5)	0.786
Parasagittal/falcine	16 (13.9)	5 (16.1)	11 (13.1)	0.763
Tentorial	3 (2.6)	0 (0)	3 (3.6)	0.562
Intraventricular	3 (2.6)	1 (3.2)	2 (2.4)	>0.999
Anterior fossa	21 (18.3)	7 (22.6)	14 (16.7)	0.648
Middle fossa	16 (13.9)	5 (15.1)	11 (13.1)	0.763
Posterior fossa	11 (9.6)	2 (6.5)	9 (10.7)	0.725
Imaging features
Tumor volume (cm3)	24.2 (7.6, 50.3)	44.8 (31.3, 67.7)	13.4 (4.1, 42.1)	<0.001*
Dystrophic calcification	17 (14.9)	0 (0)	17 (20.5)	0.006*
Necrosis	5 (4.4)	1 (3.2)	4 (4.8)	>0.999
Brain invasion	4 (4.4)	3 (9.7)	2 (2.4)	0.120
Bone invasion	7 (6.1)	5 (16.1)	2 (2.4)	0.015*
Extent of resection				<0.001*
Gross total resection	80 (70.2)	13 (41.9)	68 (81.0)
Subtotal resection	34 (29.6)	18 (58.1)	16 (19.1)
WHO grade				0.719
1	40 (43.5)	12 (38.7)	38 (45.2)
2	47 (45.9)	13 (41.9)	34 (40.5)
3	18 (15.7)	6 (19.4)	12 (14.3)
Pathologic features
Sheeting	23 (20.0)	14 (45.2)	9 (10.7)	<0.001*
Macronucleoli	16 (13.9)	8 (25.8)	8 (9.5)	0.035*
Nuclear pleomorphia	3 (2.6)	1 (3.2)	2 (2.4)	>0.999
Necrosis	33 (28.7)	15 (48.4)	18 (21.4)	0.009*
Chronic inflammation	3 (28.7)	0 (0)	3 (3.6)	0.562
Hypercellularity	31 (27.0)	14 (45.2)	17 (20.2)	0.015*
Small cells	28 (24.4)	12 (38.7)	16 (19.1)	0.053
Psammoma bodies	36 (31.3)	3 (9.7)	33 (38.3)	0.005*
Brain invasion	16 (13.9)	6 (19.4)	10 (11.9)	0.364
Bone infiltration	10 (8.7)	3 (9.7)	7 (8.3)	>0.999
Dura infiltration	7 (6.1)	1 (3.2)	6 (7.1)	0.672
Intranuclear inclusions	20 (17.4)	6 (19.4)	14 (16.7)	0.952
Prominent nucleoli	24 (20.9)	7 (22.6)	17 (20.2)	0.987
Atypia	5 (4.4)	1 (3.2)	4 (4.8)	>0.999
Calcification	8 (6.7)	2 (6.5)	6 (7.1)	>0.999
Mitosis per 10 HPFs	2 (0, 5.0)	5 (3.5, 9.5)	1.0 (0, 4.8)	0.001*
Ki67 index	5.2 (0.11, 11.1)	9.1 (3.3, 19.0)	2.7 (0.04, 8.7)	0.003*

Demographic, clinical, histopathological, and imaging characteristics were first analyzed to identify predictors of MRM on univariable analysis (Fig. 1, D to F, and Table 2). Male sex was associated with MRM (MRM: 48.4% versus NRM: 26.2%; P = 0.042). However, there was no association between MRM and race (P = 0.245), ethnicity (P > 0.999), or age at diagnosis (P = 0.162). Patients with a higher number of meningiomas on presentation were more likely to have an MRM (P < 0.001). There was no significant difference between patients with MRM and NRM receiving adjuvant radiation. The only presenting symptom significantly associated with an MRM was visual disturbance (P = 0.015). Visual disturbances may be caused by direct cortical disturbance, compression of the optic apparatus (including the optic nerve, chiasm, or tract and the cranial nerves controlling the extraocular muscles), or elevated intracranial pressure due to tumor volume or cerebral edema. Of patients with visual disturbances, 60.7% (17 of 28) had tumors located either in the anterior/middle fossa, likely affecting the optic apparatus directly, or over the occipital lobe, likely affecting the visual cortex directly. Other neurologic symptoms such as cognitive decline, gait abnormalities, headache, weakness, pain, tremor, or vomiting were not significant predictors of MRM (data S1).

On imaging, greater tumor volume (MRM: 44.8 cm³ versus NRM: 13.4 cm³; P < 0.001) was a predictor of MRM. No MRMs had imaging evidence of dystrophic calcification on presentation (MRM: 0% versus NRM: 20.5%; P = 0.006). However, other imaging features such as location of meningioma, necrosis on imaging, and brain invasion on imaging were not correlated with MRM (Fig. 1E and Table 2). Subtotal resection was significantly associated with MRM (MRM: 58.1% versus NRM: 19.1%; P < 0.001). Analysis of histopathological features revealed that the presence of sheeting (MRM: 45.2% versus NRM: 10.7%; P < 0.001), macronucleoli (MRM: 25.8% versus NRM: 9.5%; P = 0.035), necrosis (MRM: 48.4% versus NRM: 21.4%; P = 0.009), and hypercellularity (MRM: 45.2% versus NRM: 20.5%; P = 0.015) were all independently associated with MRM. In addition, average mitosis per 10 high-power field (MRM: 5 versus NRM: 1; P = 0.001) and Ki67 index (MRM: 9.1% versus NRM: 2.7%; P = 0.003) were both also independently associated with MRM. The presence of psammoma bodies was inversely correlated with MRM (MRM: 9.7% versus NRM: 38.3%; P = 0.005; Fig. 1, E and F, and Table 2). Several other histologic features were not associated with MRMs such as nuclear pleomorphia (P > 0.999), chronic inflammation (P = 0.562), brain invasion (0.364), bone infiltration (P > 0.999), dura infiltration (P = 0.672), intranuclear inclusions (P = 0.952), prominent nucleoli (P = 0.987), atypia (P > 0.999), and calcifications (P > 0.999; Fig. 1E and Table 2).

On multivariable binomial logistic regression, MRM was significantly associated with male sex [odds ratio (OR) = 4.136, P = 0.012], subtotal resection (OR = 6.151, P = 0.001), higher number of meningiomas at presentation (OR = 2.019, P = 0.017), and histopathological sheeting (OR = 6.725, P = 0.002; Table 3).

Table 3. Multivariable logistic regression for multiply recurrent meningiomas.

OR, odds ratio; CI, confidence interval. *P ≤ 0.05.

Variable	Beta estimate	SE	OR (95% CI)	P value
Constant	−3.621	10.648	0.027	<0.001*
Sex
Female				Reference
Male	1.420	0.566	4.136 (1.365, 12.532)	0.012*
Extent of resection
Gross total resection				Reference
Subtotal resection	1.817	0.0.553	6.151 (2.081, 18.180)	0.001*
Number of meningiomas	0.703	0.295	2.019 (1.132, 3.600)	0.017*
Histopathological sheeting	1.906	0.604	6.725 (2.059, 21.964)	0.002*

WES yields no specific genomic variants predictive of MRMs

Following clinical analysis of predictors of MRMs, we investigated molecular predictors of MRMs. We obtained sequencing data from two institutions including whole-exome sequencing (WES), RNA sequencing (RNA-seq), and DNA methylation array data. Ninety-seven meningiomas (29 MRMs and 68 NRMs) had multiomic data available for analysis. Among these, 12 tumors had all three omic datasets available for analysis (Fig. 2, A and B).

Fig. 2. Sample representation across omic methods.

(A) Oncoprint representing sample distribution across omic methods. (B) Venn diagram showing omic sample distribution across methods.

First, we investigated the utility of omic analysis to predict multiple recurrences in meningiomas. To perform a controlled comparison, only primary molecular MenG C meningiomas, the molecular group associated with recurrence based on integrative molecular classification (13, 14), were used to compare MRMs to NRMs (Fig. 3A). First, WES analysis was performed on this sample set made up of 12 MRMs and 16 NRMs. There were 10 WHO grade 3, 10 WHO grade 2, and 8 WHO grade 1 meningiomas (Fig. 3B). Non-synonymous, meningioma-associated genomic variants for each lesion were identified. Because of lack of matched normal tissue to tumor samples for filtration of germline variants, this study does not probe for novel single-nucleotide variants or structural variants and instead focuses on variants previously associated with meningioma. There were no significant differences in any specific genomic variant between MenG C MRMs and NRMs on Fisher’s exact tests, including TERT promoter mutations and CDKN2A/B loss (Fig. 3C). Genomic instability in the form of chromosomal alterations, especially 1p, 14q, and 22q chromosomal arm loss, has been widely established as a marker of recurrence in meningiomas (14). Therefore, we compared genomic instability of MenG C MRMs to MenG C NRMs by determining the burden of chromosomal arm losses, gains, or either losses or gains (Fig. 3D). Consistent with the literature, we observed that a greater burden of chromosomal arm loss was associated with MRMs compared to NRMs (two-sample t test, P = 0.03128). Genomic instability was also evaluated at a more granular level by considering the percentage of the genome altered as a continuous variable and inspecting segment-level copy number alterations (CNAs; fig. S1, A and B). Unlike at the arm level, no statistically significant changes were observed, but the percentage of the genome lost in MRMs was, on average, higher than in NRMs (23% in MRMs versus 18% in NRMs). To evaluate the role of chromosomal arm alterations in MenG C MRMs, chromosomal alterations for all arms were subjected to unsupervised clustering to determine whether MRMs and NRMs cluster separately (Fig. 3E). Notably, almost all MenG C meningiomas have 1p and 22q loss, so these losses were noncontributory to the clustering analysis. Rather than separating from each other, MRMs and NRMs were intermixed on clustering analysis, suggesting that the pattern of chromosomal arm alteration alone cannot predict whether a MenG C meningioma will be multiply recurrent (Fig. 3E). Sub-analysis of only gross totally resected meningiomas (n = 9 MRMs and n = 15 NRMs) yielded similar results although statistical significance was not achieved when comparing arm losses, potentially due to decreased statistical power (two-sample t test, P = 0.07163; fig. S2). Thus, within MenG C meningiomas, although no single structural abnormality was associated with MRMs versus NRMs, MRMs exhibited greater genomic instability at the chromosomal arm level in the form of chromosomal arm losses.

Fig. 3. WES reveals a greater burden of chromosomal losses in MRMs.

(A) Schematic representing the MRM and NRM MenG C meningiomas that underwent WES and methylation array analysis from formalin-fixed paraffin-embedded (FFPE) tissue. t-SNE, t-Distributed Stochastic Neighbor Embedding. (B) Patient information for the MenG C meningioma samples including molecular class, institution, recurrence status, sex, and WHO grade. (C) Status of non-synonymous genomic variants commonly reported in the meningioma literature from WES in MRM (n = 12) and NRM (n = 17). (D) Arm-level copy number differences between MRMs (n = 12) and NRMs (n = 16) (two-sample t test; losses, P = 0.0313; gains, P = 0.7894; and alterations, P = 0.1377). (E) Unsupervised clustering based on chromosomal arm alterations calculated from WES of all MenG C tumors is unable to separately cluster MRMs and NRMs. n.s., not significant.

Methylation profiling reveals distinct molecular features of MRMs

DNA methylation array profiling has become an important method to classify and prognosticate central nervous system (CNS) tumors (15). Therefore, we subjected the MenG C cohort of tumors (MRM, n = 12; and NRM, n = 16) that underwent WES to methylation array profiling (Fig. 3A). Supervised clustering based on the top 10,000 most variable probes (MVPs) for this cohort of MenG C meningiomas was performed (Fig. 4A). ß Values (0.0 to 1.0) represent the degree of methylation at a given locus with higher ß indicating greater methylation. To understand whether there was an overall difference in global methylation level of MRMs versus NRMs, we calculated the average ß value of each sample and found that MRMs had a significantly higher degree of methylation of the 10,000 MVPs compared to NRMs (two-sample t test, P = 0.0155; Fig. 4B). When excluding subtotally resected tumors, the 10,000 MVPs of MRMs were still significantly more methylated than NRMs (P = 0.0016; fig. S2C). Next, we wanted to understand whether any specific CpG sites had methylation level differences between MRMs and NRMs, specifically those in promoter regions, which may affect downstream gene transcription. Comparing differentially methylated probes (DMPs) between groups yielded 1034 significant DMPs, of which 395 were located in promoter regions of genes. Overall, we observed 257 genes with increased promoter methylation in MRMs and 74 genes with decreased promoter methylation in MRMs (Fig. 4C). These data suggest that robust global and local methylation differences exist between molecular MenG C MRMs and NRMs.

Fig. 4. Methylation array profiling elucidates differences between MRMs and NRMs.

(A) Supervised clustering heatmap comparing top 10,000 most variable probes (MVPs) for MRMs (n = 12) versus NRMs (n = 17) in MenG C meningiomas. (B) Comparison of average beta values of top 10,000 MVPs between MRMs (n = 12) and NRMs (n = 17). Two-sample t test performed (P = 0.0155). (C) Quantification of differentially methylated probes (DMPs) between MRMs and NRMs focused on promoter site probes associated with genes. (D) Overlap of genes from methylation analysis in (C) and RNA-seq analysis (MRM, n = 3; and NRM, n = 63) comparing MRMs and NRMs. Up-regulated RNA expression was defined as differentially expressed genes (DEGs) with P < 0.05 and log₂ fold change (log₂FC) > 1, and down-regulated DEGs had P < 0.05 and log₂FC < −1. (E) Reverse transcription–quantitative polymerase chain reaction (qPCR) validation of knockdown efficiency for three distinct EDNRB small interfering RNA (siRNA) constructs in the CH157 meningioma cell line (n = 3 independent experiments). GAPDH, glyceraldehyde-3-phosphate dehydrogenase. (F) Cell proliferation of each siRNA construct in CH157 cells (n = 4 independent experiments). All comparisons ****P < 0.0001 on one-way analysis of variance (ANOVA) with multiple comparisons.

Next, we wanted to determine whether these methylation differences at specific sites had an impact on gene expression. We performed bulk RNA-seq of a cohort of MenG C meningiomas to identify differentially expressed genes (DEGs) between MRMs and NRMs. Typically, we would expect genes to have an inverse relationship between promoter site methylation and RNA expression (16, 17); therefore, we identified two groups of genes: (i) genes with hypermethylated promoter sites and down-regulated RNA expression and (ii) genes with hypomethylated promoter sites and up-regulated RNA expression (Fig. 4D). Genes found to have significantly increased promoter methylation and decreased RNA expression in MRMs included BNIP2, CD34, CLVS1, CXADR, EDNRB, EVX1, GNG2, IRX1, MAPRE2, PCDH17, PLVAP, TF, and ZBTB12. Conversely, there were three genes found to have significantly decreased promoter methylation and increased RNA expression in MRMs: LIG1, SLC9A8, and ZNF512B (Fig. 4D). Given the differences between MRM and NRM in burden of CNAs, we also performed a parallel analysis of the methylation data using the SeSAMe preprocessing pipeline, which essentially masks regions with CNA from subsequent analysis (18). Comparing DMP analysis from SeSAMe preprocessed data to our original DMP analysis revealed conserved promoter site hypermethylation and decreased RNA transcription for 8 of the 13 genes of the original analysis (EDNRB, GNG2, MAPRE2, ZBTB12, BNIP2, CXADR, PCDH17, and TF) and conserved promoter site hypomethylation and increased RNA transcription for all three genes of the original analysis (fig. S3). The genes with hypermethylation and decreased RNA expression may represent important markers to predict multiple recurrences in MenG C meningiomas with potential therapeutic implications. Together, our molecular analysis suggests that, although specific genomic variants and CNAs did not reliably predict multiple recurrences in meningiomas, methylation profiling and RNA-seq analysis may provide more useful predictive biomarkers and insights into the biology of MRMs.

Next, we aimed to functionally validate our in silico findings using an in vitro model to understand whether manipulation of identified MRM gene targets could confer a more aggressive biological phenotype. We were able to successfully knockdown the expression of EDNRB, a candidate gene with increased promoter methylation and decreased RNA expression in MRMs (Fig. 4E). Notably, EDNRB promoter hypermethylation was present even when excluding patients with subtotal resection of their primary tumor and when processing our data with the SeSAMe preprocessing pipeline. Prior work in triple-negative breast cancer and hepatocellular carcinoma highlighted a correlation between down-regulation of EDNRB and worse cancer prognosis, but the function of EDNRB has not been studied in meningiomas (19, 20). Therefore, we performed small interfering RNA (siRNA)–mediated knockdown of EDNRB using the NF2 mutant meningioma cell line CH157 and evaluated cell proliferation using the CellTiter-Glo assay. siRNA control cells had the same level of EDNRB RNA expression as non-transfected cells. EDNRB knockdown had a mean knockdown efficiency of 78% (EDNRBi.1), 95% (EDNRBi.2), and 62% (EDNRBi.3) for each of the three siRNA constructs used (Fig. 4E). Compared to control, EDNRB knockdown cells had significantly higher proliferation 36 hours after transfection for each of the three siRNA constructs (P < 0.0001; Fig. 4F), suggesting that the down-regulation of EDNRB may drive an aggressive meningioma phenotype.

MRMs exhibit an increase in copy number gains from primary tumor to recurrences

We next focused on identifying longitudinal molecular changes in MRMs from primary tumors to subsequent patient-matched recurrences. Eleven subjects with matched primary-recurrent samples were identified, 10 with WES data and 10 with methylation profiling from formalin-fixed paraffin-embedded (FFPE) tissue (Fig. 5, A and B). All patients except one had at least one round of radiation (fractionated or stereotactic) following the first resection before resection of recurrences (Fig. 5B). Of the 10 patient-matched primary samples in each cohort, 70% (7 of 10) were gross total resections, and 30% (3 of 10) were subtotal resections. Most meningiomas were molecular class MenG C (only one patient had a MenG A tumor, and one had a MenG B tumor).

Fig. 5. Recurrent MRMs exhibit no change in genomic variants or subclones but bear an increased burden of copy number gains.

(A) WES and methylation array profiling was performed on longitudinal matched primary and recurrent FFPE samples from patients with MRM. (B) Patient information for the 11 primary, 11 first recurrence, and 2 second recurrence lesions. (C) Non-synonymous genomic variants commonly reported in the meningioma literature from WES of 10 primary and matched recurrent pairs. (D) Quantification of NF2 variants from (C) showing no significant difference between primary and first recurrence for any specific variant type. (E) Arm-level copy number differences between primary (n = 10) and matched first recurrence (n = 10) tumors (paired t test; losses, P = 9302; gains, P < 0.0001; and alterations, P = 0.0836). (F) Chromosomal arm gains and losses for primary and matched recurrent tumors of 10 patients. (G) Histogram representing total chromosomal arm gains and losses between primary and first recurrence of MRMs. (H) Subclonal analysis of whole-exome data using variant allele frequencies (VAFs) from a representative patient comparing primary versus recurrence 1, recurrence 1 versus recurrence 2, and primary versus recurrence 2.

First, commonly reported meningioma-specific genomic variants were identified from the WES cohort (Fig. 5C). Because NF2 was the most common variant observed in the dataset, sub-analysis was performed to assess whether there was an association with NF2 variant types in primary versus recurrence tissue. There was no significant difference in NF2 variant type between recurrences and primary tumors using Fisher’s exact test (P > 0.999; Fig. 5D). Further analysis from WES data included evaluation of CNA changes from primary to recurrent tumors in MRMs (Fig. 5E). We hypothesized that recurrences may harbor greater genomic instability compared to the primary tumor, especially in the setting of radiation treatment. Therefore, we performed pairwise comparisons of copy number losses, gains, and overall alterations (Fig. 5, E and F, and fig. S4) for primary versus matched first recurrence MRMs (molecular data from second recurrence subjects were available for only two samples, so these were excluded from statistical analysis but are included in descriptive analyses). Here, we observed a statistically significant increase in copy number gains, both at the arm level (paired t test, P < 0.0001) and as a percentage of the genome altered (paired t test, P = 0.0031). The same observations were made even when excluding the three patients with subtotal resection from this analysis (fig. S5, A to D). Evolution of genomic features from primary to recurrent tumors can also be evaluated through subclonal analyses from whole-exome data. To determine whether there was a subclonal shift from primary meningioma to recurrent tumors in patients with MRM, subclonal analysis was performed using variant allele frequencies (VAFs) from tissue from a subject who underwent gross total resection of their primary tumor with molecular data from the primary and two subsequent tumor recurrences available. This yielded four clusters of subclones that remained similar from the primary tumor to the first and second recurrences (Fig. 5H). Thus, while whole-exome analysis revealed no new genomic variants or subclonal shift, it did reveal an increase in the burden of copy number gains from primary to recurrent MRMs.

Methylation profiling is similar in MRMs from primary to recurrences

Next, we investigated MRMs using epigenetic methylation profiling of the matched primary and recurrent tumors. Unsupervised clustering using the 10,000 MVPs grouped each patient’s primary and first and second recurrent lesions alongside one another (Fig. 6A), suggesting primary and recurrent tumors in patients with MRM harbor similar epigenetic (DNA methylation) profiles even with intervening radiation treatment. Notably, in the three tumors with an upgrade in WHO grade from primary tumor to recurrence (PT2, PT4, and PT6), the methylation profile of the recurrence remained similar to the primary tumor, highlighting the stability of epigenetic profiles of these tumors (Fig. 6A). In an orthogonal analysis identifying the percent of the 10,000 MVPs that had conserved methylation status between samples, we redemonstrate findings that show patient-matched samples have similar methylation profiles from primary to recurrences (Fig. 6B), further underscoring the longitudinal stability in epigenetic profiles of MRMs. However, comparison of the average ß values of the 10,000 MVPs showed a significant increase in methylation in first recurrence tumors compared to matched primary tumors (paired t test, P = 0.0045; Fig. 6C), although it is unclear whether this hypermethylation is due to radiation exposure or tumor recurrence alone as most recurrent tumors were radiated before resection.

Fig. 6. Matched primary and recurrent meningiomas have similar methylation profiles.

(A) Unsupervised clustering heatmap comparing 10,000 MVPs for primary and matched recurrence from patients with MRM (n = 10). (B) Heatmap of percent of 10,000 MVPs with conserved methylation status across samples (n = 10 patients). Unsupervised clustering groups primary and recurrent samples together. (C) Pair-wise comparison of average methylation ß values of the 10,000 MVPs between matched primary (n = 10) and recurrence 1 (n = 10) of each MRM tumor. Paired analysis was performed using paired t test (P = 0.0045).

While clustering and ß value comparison showed similar methylation profiles of matched primary and recurrent tumors, pair-wise comparison of primary and recurrent tumors revealed relative hypermethylation in recurrent tumors. Therefore, we investigated whether there were any specific CpG loci that harbored different methylation states between primary and recurrent lesions. When performing DMP analysis both with a paired approach, comparing each patient’s first recurrence to their own primary tumor as well as an unpaired approach, comparing all first recurrences to all primary tumors, we found no significant DMPs between matched primary and recurrences. Although limited by sample size, for the two patients with second recurrence data also available, there were no significant DMPs between second recurrence and primary tumors. These methylation findings held true even after exclusion of patients with subtotal resection (STR) of their primary lesion (fig. S5, E to G). Together, these analyses suggest that, although there is hypermethylation of the 10,000 MVPs in first recurrences compared to primary MRMs, there is no specific CpG locus that is differentially methylated.

DISCUSSION

We present this study to evaluate predictors of aggressive, multiply recurrent meningiomas (MRMs). In multivariable binomial logistic regression analysis, MRMs were significantly associated with male sex (P = 0.012), subtotal resection (P = 0.001), higher number of meningiomas on presentation (P = 0.017), and histopathological sheeting (P = 0.002). MRMs were associated with greater global chromosomal loss (P = 0.0313) and increased methylation at promoter sites (P = 0.0155) but were not associated with specific chromosomal arm-level alterations. A subset of nine genes were identified to have significantly greater promoter methylation and decreased RNA expression in MRMs (compared to NRMs), and three genes had a significantly lower promoter methylation and increased RNA expression in MRMs. In vitro validation confirmed that down-regulation of EDNRB, which had significantly greater promoter methylation and decreased gene expression in multiomic analysis, in meningioma cells generated a more aggressive tumor phenotype. Compared to primary tumors, MRM recurrences were found to have increased copy number gains (P < 0.0001) and percent genome gained (P = 0.0031) as well as significant hypermethylation of the top 10,000 MVPs (P = 0.0045).

The predictive characteristics of MRMs identified in this study complement those in previous work on meningioma recurrence. Notably, there was no association between adjuvant radiation and MRMs, suggesting that, now, treatment teams are not able to adequately predict which meningiomas will be more aggressive during initial presentation. Although meningiomas have been reported to have a higher incidence in females, male patients often have more aggressive lesions with more complex clinical courses (21, 22). Further, the majority of patients with MenG C meningiomas were reported to be male [11]. Our study strengthens these findings with the majority (66.1%) of the cohort being female but with male sex being a predictor of MRMs. In addition, our finding of increased number of meningiomas on presentation being associated with MRMs complements prior national database work, highlighting that the increasing number of meningiomas had a negative impact on overall survival (22). The extent of resection has been extensively studied in many cranial tumors, and, in meningiomas, the consensus is that subtotal resection predicts worse overall survival and progression-free survival (23). However, even when subtotally resected meningiomas were excluded from our analysis, our molecular analysis conclusions remained unchanged. In our dual-institution PSM cohort, we also find that subtotal resection is an independent predictor of more aggressive MRMs.

The WHO classification, which relies heavily on histological features, is now the most standardized approach to predict meningioma prognosis. In the 2016 WHO grading update, evidence of brain invasion independently qualified a meningioma to be a WHO grade 2 lesion (24). Although brain invasion was not an independent predictor of MRMs in our study, this is likely, in part, due to our cohort being propensity score matched by WHO grade. Because there were similar proportions of WHO grade 2 lesions in both the MRM and NRM cohort, this likely correlates to a similar proportion of lesions with brain invasion in both groups. Another criterion for a WHO grade 2 meningioma is exhibiting three of the five atypical histological characteristics: necrosis, sheeting, prominent nucleoli, hypercellularity, and small cells (8). Although the cohort was matched for WHO grade, sheeting was significantly associated with more aggressive MRM on multivariable analysis.

On the basis of modern molecular classification of meningiomas, tumors with the worst prognosis are typically characterized by 1p and 22q chromosomal loss (MenG C). In this study, we evaluated whether we can further identify which MenG C meningiomas will become multiply recurrent. We did not find any specific chromosomal arm loss associated with MRMs within the MenG C subtype. However, we did identify a significantly greater global loss of chromosomal arms in MRMs compared to that in NRMs. (25, 26) Prior work has shown that global chromosomal instability is associated with a worse prognosis in other cancer types and CNS tumors (27, 28).

With evolving sequencing technologies, methylation array profiling is playing an increasingly critical role in the molecular classification of meningiomas and other CNS tumors (13, 15). We showed that MRMs are globally hypermethylated compared to NRMs. In addition, we identified nine genes with significantly hypermethylated promoter sites and low RNA expression changes and a set of three genes with hypomethylated promoter sites and higher RNA expression levels when comparing MRMs to NRMs. These findings suggest that these specific genes may be driving meningioma aggressiveness and lead to a multiply recurrent and treatment-resistant phenotype. In our in vitro model system, decreased expression of the gene EDNRB, which was found to have increased promoter methylation and decreased expression, was associated with higher cell proliferation, as one might expect in more aggressive tumors. Although the role of EDNRB in multiple recurrences of meningioma has not been previously evaluated, prior work has shown that lower expression of EDNRB predisposes patients with hepatocellular carcinoma to have a poorer prognosis (20). Similarly, patients with triple-negative breast cancer with low EDNRB expression have been shown to have shorter disease-free survival (19).

We found that, while recurrent MRMs have no specific pattern of chromosomal alterations and generally remain epigenetically similar to the primary tumor, they do have an increased burden of chromosomal arm gains and increased methylation at the MVPs. Notably, almost all patients in this study had at least one type of radiation treatment in between primary and recurrences, so it is not possible to distinguish whether these changes are due to radiation treatment effect or other evolutionary pressures in multiply recurrent tumors. If these changes are in response to radiation, then this finding would be consistent with prior investigation into other CNS tumors and their response to radiation. For instance, in the context of glioblastoma, tumor cells transcriptionally convert to a more mesenchymal phenotype following radiation treatment (29, 30). Alternatively, if these findings, particularly of nonspecific hypermethylation in recurrent tumors, reflect evolutionary pressure in multiply recurrent tumors, then this might be consistent with previously reported findings of locally disordered methylation in cancers such as CLL (31, 32).

Although this study details a robust analysis of MRMs from a clinical and multiomic perspective, it is not without limitations. While data were collected from patients treated at two institutions, the cohort did not have large enough sample size to increase the number of features included in the PSM algorithm or to include an external validation arm. In vitro validation to confirm computational findings is critical; however, future work can build upon our findings and perturb relevant target genes to assess meningioma tumor cell phenotype and resistance to therapeutics. In addition, expansion of multiomic methods to include other more granular omic methods such as single-cell RNA and single-nuclear assay for transposase-accessible chromatin (ATAC) sequencing may yield additional insights into the biology of MRMs.

MATERIALS AND METHODS

Study design

Following Institutional Review Board (IRB) approval and informed consent at each institution (institution A, IRB no. H43183; and institution B, IRB no. 201409046), a total of 1315 patients with meningioma from two large academic institutions from 2005 to 2022 were queried for MRMs and NRMs. MRMs were defined as meningiomas with two or more recurrences. Inclusion criteria included any patient with MRM or NRM with primary lesion treated at home institution with surgical resection or biopsy to ensure the availability of WHO grading information. Exclusion criteria included age less than 18 years old at time of initial meningioma treatment as well as patients with any genetic syndromes such as neurofibromatosis. A total of 31 patients with MRM (institution A: n = 11 versus institution B: n = 20) and 1155 patients with NRM (institution A: n = 395 versus institution B: n = 760) were identified for a total of 1186 patients (Fig. 1).

PSM was used to match patients with MRM to patients with NRM (33, 34). Variables selected for PSM multivariable logistic regression model were WHO grade, follow-up time, and institution. These variables were chosen to normalize comparison groups to identify predictive variables of multiple recurrences in patients with meningioma. One-to-three nearest neighbor matching without replacement was performed using the “MatchIt” package in R (34). Optimal matching was performed to minimize the total within-pair difference of propensity scores (33). In line with prior recommendations, an optimal caliper width of 0.2 times the SD of the logit of the propensity score was used (35). Following matching, the R packages “tableone” and “cobalt” were used to assess the balance of individual covariates using SMD, where an SMD between −0.1 and 0.1 following PSM would reflect a perfectly balanced cohort (36). Following PSM, 31 patients with MRM were matched to 84 patients with NRM, for a total cohort of 115 patients (Fig. 1).

A subset of the 115 patients with available FFPE tissue was identified for molecular analysis. For the primary MenG C meningioma analysis, 12 MRMs and 23 NRMs underwent both whole-exome and methylation analysis. Bulk RNA-seq was performed on 3 MRMs and 63 NRMs.

Clinical analysis

Retrospective chart review was conducted on all 115 patients for demographic, clinical, histopathological, and imaging information to explore as possible predictors of multiple recurrences. Recurrence was defined as radiological progression of the meningioma following any treatment type, including subtotal or gross total resection, radiation treatment, stereotactic radiosurgery, or chemotherapy treatment. All patients’ primary meningioma WHO grades were regraded according to the updated 2021 WHO grading criteria using molecular data when available. Demographic variables such as age, sex, race, and ethnicity were collected. Clinical variables included the number of meningiomas present on initial diagnosis, symptoms at diagnosis, extent of resection, postoperative complications, and follow-up time. Histopathological variables like WHO grade, mitotic index, Ki67, and pathological features such as sheeting, macronucleoli, spontaneous necrosis, chronic inflammation, hypercellularity, small cells, brain infiltration, bone infiltration, intranuclear inclusions, prominent nucleoli, and calcifications were queried for from the original pathology report. Imaging evaluation included location of meningioma, tumor volume, brain invasion, bone invasion, necrosis, and intratumoral dystrophic calcifications.

WES and analysis

Automated dual-indexed libraries were constructed with ~100 to 300 ng of genomic DNA from FFPE using the KAPA HTP Kit (Roche Diagnostics, catalog no. 07962363001) on the SciCloneG3 NGS instrument (PerkinElmer) targeting 250–base pair inserts. Libraries were pooled at equimolar ratio yielding 4 to 5 μg of library pools. Each library pool was hybridized overnight with the xGen Exome Hyb Panel v2 (Integrated DNA Technologies, catalog no. 10005153) that spans a 34-Mb target region (19,433 genes) of the human genome. For TERT promoter targeted panel sequencing, libraries were pooled as close to equimolar ratio as possible to yield a 4.7-μg library pool. Each library pool was hybridized overnight with the TERT promoter probes only.

The concentration of each capture library pool was determined through quantitative polymerase chain reaction (qPCR) using the KAPA library Quantification Kit according to the manufacturer’s protocol (KAPA Biosystems/Roche) to produce cluster counts appropriate for the Illumina NovaSeq6000 instrument. Normalized pools were sequenced on a NovaSeq6000 S4 Flow Cell using the XP workflow and a 151 × 10 × 10 × 151 sequencing recipe according to the manufacturer’s protocol. Target sequencing depth was determined before sequencing, and pools were normalized in ratios on the basis of the targeted depth of 100×. Notably, patients from institution A had matched blood and patients from institution B did not.

The samples were analyzed on a DRAGEN Bio IT processor using DRAGEN software version 4.0.3 with a GRCh38 reference genome. We processed the sample separately in tumor-only mode with germline tagging enabled. Alignments were generated in CRAM format with duplicates marked. Structural and copy number variant calls were also generated. CNA calling was based on a panel of normal samples. Small variants that passed all QC filters in the exome target region were annotated using ANNOVAR. Synonymous variants were not included.

Copy number calling from the WES data was performed using CNVkit (37). Because of the absence of paired normal samples, a flat reference was created from the hg38 human reference genome for copy number calling. CRAM files were converted to the BAM format using SAMtools (38). Log₂ ratio estimates were calculated for each sample BAM file using the batch command. Segment calls were filtered to exclude those with fewer than 10 probes on the segment, and segment copy number calls were then made as follows: If seg.mean < −1, then copy number = 0; if −1 ≤ seg.mean < −0.3, then copy number = 1; if −0.3 ≤ seg.mean ≤ 0.3, then copy number = 2 (copy neutral); if 0.3 < seg.mean ≤ 1, then copy number = 3; and if seg.mean > 1, then copy number = 4+. The percent gain or percent loss of the genome as a continuous variable was computed for each sample by taking the sum of the length of the gained or lost segments and dividing it by the length of the genome covered by WES for each sample. We then calculated the percent gain or percent loss for each chromosomal arm for each sample by dividing the length of segments gained or lost on each arm by the length of the arm covered by WES for each sample. Copy number gain or loss of an arm was defined as 50% gain or loss in that chromosomal arm.

The R package SciClone was used to infer tumor subclonal architecture using VAFs to evaluate subclonal evolution from primary to recurrence 1 and recurrence 2 using default parameters (39). The analysis was performed on one representative sample with primary, recurrence 1, and recurrence 2 tumor data available.

Methylation array processing and analysis

Illumina 850K EPIC BeadChip was used to conduct DNA methylation analysis on extracted tumor DNA from FFPE tissue, as per the manufacturer’s instructions (Montreal, Canada). For each sample, 500 ng of DNA in a final volume of 45 μl of water was input for each sample in the Infinium methylEPIC platform. The EZ DNAMethylation Kit (Zymo Research) was used to bisulfate convert the normalized DNA samples. The standard Illumina Infinium HD Restore protocol was used to repair the degraded FFPE samples. The samples were immediately processed through the normal Illumina methylation protocol. Methylation chips were processed using a Tecan instrument and read using the iScan instrumentation. All data were processed using Illumina’s Genomestudio software. Samples from institution A were processed using EPIC BeadChip v1, and EPIC BeadChip v2 was used for samples from institution B.

Using the Chip Analysis Methylation Pipeline (ChAMP) R package, raw .idat files were imported and filtered only for common probes between EPIC BeadChip v1 and v2 (40). Briefly, data were then processed and normalized using the standard ChAMP pipeline; probes were filtered out if they failed to hybridize (detection, P > 0.01), had <3 beads in 5% of samples, were not at CpG sites, were defined as multihit probes, were located on sex chromosomes, or were associated with single-nucleotide polymorphisms (41). Sample quality was assessed, and one sample was excluded because of a lack of bimodal distribution of ß value density plot. ß values were calculated and normalized using the beta mixture quantile dilation method (41). Singular-value decomposition analysis was completed to evaluate the need for batch effect correction, and champ.runCombat was used to correct for batch effect due to institution/platform type (42). Normalization was performed using the champ.norm function, yielding a normalized b matrix to use for visualization. The ß values of the top 10,000 MVPs were used for heatmap construction and clustering analysis. The average ß value of the top 10,000 MVPs was calculated for each sample. Orthogonal analysis was performed by binarizing the methylation status of the top 10,000 MVPs (hypermethylated probe = b > 0.5 and hypomethylated probe = b ≤ 0.5) and determining the percent of probes with conserved methylation statuses across samples.

To identify DMPs between MRM and NRM primary MenG C meningiomas, the champ.DMP function was used. DMPs on promoter sites were identified as 5′ untranslated, first exon, transcriptional start site 200 (TSS200) or TSS1500 regions. Unsupervised clustering of MRM and NRM primary MenG C meningiomas using the top 10,000 MVPs was performed using “Ward.D2” as the row clustering and default column clustering in the ComplexHeatmap R package (43, 44). Champ.PairedDMP was also used to analyze DMPs matched primary and recurrent meningiomas.

The SeSAMe methylation array analysis pipeline was used to preprocess the methylation data in parallel to ChAMP processing as it is able to account for artifacts caused by germline or somatic chromosomal deletions. We loaded our raw data into the openSesame function with default settings, which yield a b matrix. The mliftover function was used to decrease differences between the samples run on EPIC V1 versus EPIC V2 (18). All subsequent analyses using SeSAMe preprocessed data were completed in the same way as completed for ChAMP preprocessed data.

RNA sequencing

RNA was isolated under the aegis of a Baylor College of Medicine IRB–approved protocol. We reviewed RNA-seq data from 66 patients with primary MenG C meningiomas (MRMs, n = 3; and NRMs, n = 63). Of these, RNA-seq data from 25 patients had been previously reported and made publicly available (11, 13). In 24 samples, RNA-seq had been performed as previously described using the Illumina platform (11). We obtained RNA-seq data for the remaining 42 tumors from Tempus (Chicago, IL), which entailed sending tumor samples along with saliva for processing according to their protocol (45).

Raw reads from RNA-seq data were processed with an in-house pipeline using STAR for read alignment, FastQC and RSeQC for read and alignment quality assessment, and FeatureCount for expression count (46–48). The reads were aligned to the GRCh38 human reference genome and then mapped to the human transcriptome based on University of California Santa Cruz gene annotations. After alignment and mapping, the RNA-seq read counts were normalized and applied a variance stabilizing transformation using the DESeq2 package (49). Differential gene expression analysis was performed using the DESeq2 package by applying a false discovery rate cutoff of 0.05.

In vitro model validation

CH157 meningioma cells (RRID:CVCL_5723) were used for in vitro validation. We used the Dicer-substrate RNAi kit for EDNRB knockdown (TriFECTa RNAi, Integrated DNA Technologies, NJ) using 10 μM for each experimental well: siRNA 1, CAUGUCAGUAU CAUGUUCUCUAATT; siRNA 2, AGUAUUGACAGAUAUCGAGCUGUTG; and siRNA 3, AAGAUUGGUGGCUAUUCAGUUUCTA (50). Negative control siRNA was provided (Integrated DNA Technologies, NJ). Reverse transfection was performed using Lipofectamine RNAiMAX (Invitrogen, CA) with 50,000 CH157 cells in a 96-well plate. After 36 hours, reverse transcription–qPCR was performed to evaluate of knockdown efficiency. The expression was compared to untransfected CH157 cells, n = 3 biological replicates. CellTiter-Glo Assay (Promega, WI) was used to evaluate phenotypic proliferation differences between control siRNA cells and knockdown cells. For each experiment, 12 wells per condition were used, leaving empty the outermost rows and columns (rows A and H and columns 1 and 12). The median absorbance of these 12 wells was used as true value of the experiment. As control siRNA and non-transfected CH157 cells had identical expression of EDNRB, we used control siRNA as control in the cell proliferation assay, n = 4 biological replicates.

Statistical analysis

Differences between multiply recurrent and nonrecurrent groups were evaluated using Mann-Whitney U tests or two-sample t tests for continuous variables and chi-squared test for independence or Fisher’s exact test for categorical variables. Paired comparisons were done using the Wilcoxon test. Covariates, which were independently predictive of MRM, were evaluated and adjusted for the multivariable binary logistic regression model. All analyses were performed using R (Vienna, Austria), IBM SPSS Statistics version 28.0 (Armonk, NY), or GraphPad Prism version 9.5.0 (San Diego, CA) (51–53).

Supplementary Materials

The PDF file includes:

Figs. S1 to S5

Legends for data S1 to S3

Other Supplementary Material for this manuscript includes the following:

Data S1 to S3