Abstract
Treatment of the tumor and dural margin with surgery and sometimes radiation are cornerstones of therapy for meningioma. Molecular classifications have provided insights into the biology of disease; however, response to treatment remains heterogeneous. In this study, we used retrospective data on 2,824 meningiomas, including molecular data on 1,686 tumors and 100 prospective meningiomas, from the RTOG-0539 phase 2 trial to define molecular biomarkers of treatment response. Using propensity score matching, we found that gross tumor resection was associated with longer progression-free survival (PFS) across all molecular groups and longer overall survival in proliferative meningiomas. Dural margin treatment (Simpson grade 1/2) prolonged PFS compared to no treatment (Simpson grade 3). Molecular group classification predicted response to radiotherapy, including in the RTOG-0539 cohort. We subsequently developed a molecular model to predict response to radiotherapy that discriminates outcome better than standard-of-care classification. This study highlights the potential for molecular profiling to refine surgical and radiotherapy decision-making.
Similar content being viewed by others
Main
Meningiomas are the most common primary intracranial tumor in adults1,2. Maximal resection of both the meningioma and its dural attachments has historically been the primary goal of surgery, with Simpson grade used as a metric to define extent of resection (EOR)3,4,5,6,7,8. However, contemporary studies have challenged the universal benefit of aggressive surgical strategies, including dural resection in every case, particularly if it places critical neurovascular structures at risk7,8,9. Aside from surgery, radiotherapy (RT) remains the sole alternative treatment for these tumors, generally reserved as adjuvant therapy for aggressive, recurrent or incompletely resected meningiomas10,11,12,13,14,15,16,17,18,19,20. However, rates of tumor control after adjuvant RT are highly variable, creating a need to determine better predictors of response17,21,22.
Although several new molecular classifications and prognostic systems have been described for meningiomas, response to surgery and RT continues to vary considerably among patients, and the role of treatment in the context of these molecular biomarkers has not been fully explored23,24,25,26,27,28,29,30. Furthermore, although ongoing randomized trials are examining the efficacy of RT in a subset of patients with meningioma, molecular features were not used for treatment stratification, as most were discovered after the inception of these studies31,32.
To address these critical gaps, a large cohort of clinically annotated, molecularly profiled meningiomas, including invaluable data from a prospective clinical trial (NRG Oncology RTOG-0539; NCT00895622), were used in propensity score matching (PSM) to (1) determine the benefit of EOR and additive dural treatment (Simpson grade) in the context of molecular biomarkers; (2) determine the capability of molecular classification to predict response to RT; and (3) create and validate molecular predictive models of response to RT using prospective clinical trial samples that are made publicly available to help inform RT treatment decisions17,22,33.
Results
Clinical and molecular cohort
A total of 2,824 retrospective meningiomas were assembled, of which 1,686 tumors from 10 institutions had molecular and clinical outcomes data (Extended Data Fig. 1, Extended Data Table 1 and Supplementary Table 1). Subsets of this cohort were used for specific analyses for EOR, Simpson grade and RT, respectively (Fig. 1a,b), including a separate, prospectively collected cohort of 100 meningiomas from the NRG Oncology RTOG-0539 phase 2 clinical trial used for both validation of findings from the retrospective RT analysis and for independent biomarker discovery and predictive modeling (Table 1 and Extended Data Table 2). Given the key differences in baseline covariates between the different treatment arms compared throughout the study (for example, in comparing gross total resection (GTR) versus subtotal resection (STR), there were more recurrent meningiomas, meningiomas in skull base locations and tumors that received adjuvant RT in the latter group compared to the former), PSM was performed to balance these covariates in a systematic and standardized manner for our analyses of GTR versus STR, Simpson grades 1/2 versus 3, Simpson grade 1 versus grade 2 and receipt of adjuvant RT versus observation.
a,b, CONSORT diagram demonstrating the cohort of meningiomas used in this study, the sequence of analyses and results presented and the respective institutions from where these samples were obtained for our EOR, Simpson grade, and RT analyses (a) and for RT model building (b). c,d, Sankey plot (c) and bar plots (d) showing the relationship of WHO grading with molecular classifications (Molecular Group (MG) for primary analysis and DKFZ methylation subclass to confirm the molecular diagnosis of meningioma). e,f, Kaplan–Meier survival curves showing PFS of meningiomas based on WHO grade (e) and MG (f). g–h, OS of meningiomas used in our study based on classification by WHO grade (g) and MG (h). Inset shows the P values from the pairwise log-rank test between groups with Benjamini–Hochberg correction for multiple comparisons.
All meningiomas in our study with DNA methylation data were first classified using the DKFZ Central Nervous System tumor classifier v.12.5 (https://www.molecularneuropathology.org/) to confirm the molecular diagnosis of meningioma before being stratified into other meningioma molecular classification and prognostic systems (Fig. 1c,d and Extended Data Fig. 2)24,25,26,27,28,29,34. Primary analyses used the Toronto Molecular Groups (henceafter referred to as just ‘Molecular Groups’ in the manuscript) as it performed optimally for outcome prediction (Extended Data Fig. 2b and Extended Data Table 1), whereas sensitivity analyses were performed and detailed for other molecular classification and prognostic systems. Overall, Hypermetabolic and Proliferative meningiomas had a higher burden of copy number variants (CNVs), including losses of chromosomes 1p, 10, 14, 18 and 22q, compared to Immunogenic and NF2-wild-type (NF2-wt) meningiomas irrespective of institution (Extended Data Fig. 1). In keeping with previous work, progression-free survival (PFS) outcomes declined in a stepwise manner from Immunogenic to NF2-wt to Hypermetabolic to Proliferative meningiomas, respectively, with overall survival (OS) following a similar trend (Fig. 1f,h)24,27.
Impact of surgical resection across Molecular Groups
Although EOR is known to be important for meningiomas, its benefit in the context of molecular classification is unclear. A significant interaction was found between EOR and Molecular Group (P = 0.018 and P = 0.035 for PFS and OS, respectively). GTR appeared to confer a significant PFS benefit for meningiomas in all Molecular Groups (Extended Data Fig. 3a–d). When controlling for World Health Organization (WHO) grade, primary/recurrent tumor status and receipt of adjuvant RT, local tumor control after GTR was significantly less durable for Hypermetabolic (STR:GTR hazard ratio (HR) = 1.77, 95% confidence interval (CI): 1.28–2.43, P = 4.7 × 10−4) and Proliferative meningiomas (HR = 1.58, 95% CI: 1.14–2.22, P = 0.013) than for Immunogenic and NF2-wt tumors (Extended Data Fig. 3e). When OS outcomes were similarly examined across Molecular Groups, STR was primarily associated with significantly worse OS in Proliferative meningiomas (HR = 1.90, 95% CI: 1.28–2.82, P = 0.002), but not in meningiomas from other Molecular Groups (Extended Data Fig. 3f–j). However, when all cases were considered in multivariable analyses, regardless of the molecular classification or prognostic system used to risk stratify these tumors, including when prognostic CNVs were considered, GTR was consistently associated with improved PFS and OS compared to STR (Extended Data Fig. 4 and Supplementary Figs. 2 and 3). Results were concordant when only recurrent meningiomas were considered (Supplementary Fig. 4). This suggests that, regardless of tumor biology, complete surgical resection remains the critical determinant of local control, even in the recurrent setting, but its benefits may be attenuated in the most biologically aggressive meningiomas.
To examine the relationship between EOR and Molecular Group in more granular detail while further controlling for baseline covariate differences, PSM was performed for age, sex, WHO grade, primary/recurrent tumor status, receipt of adjuvant RT, tumor location and Molecular Group between the GTR and STR groups (Fig. 2a–f and Supplementary Tables 2 and 3). After PSM, STR remained significantly associated with worse PFS (HR = 2.02, 95% CI: 1.49–2.73, P = 5.2 × 10−6; Fig. 2e,f). This finding remained consistent when analyses were repeated using other molecular classification and prognostic systems (Extended Data Fig. 5 and Supplementary Tables 4–9), including when sensitivity analyses were performed using multiple imputation and PSM to address missing data (Supplementary Tables 10–23).
a, Schematic of a meningioma and its dural attachments/origin preoperatively, with legend indicating different methods of dural margin treatment (Simpson grades 1–3). b, Schematic of treatment arms being compared in PSM analyses: GTR versus STR (above), Simpson grades 1/2 versus grade 3 (middle) and Simpson grade 1 versus grade 2 (below). c–f, Kaplan–Meier survival plot of PFS before PSM (c) and covariate balance before PSM (unadjusted) and after PSM (adjusted) (d); Kaplan–Meier survival plot of PFS after PSM (e); and results of multivariable Cox regression model of PFS for GTR versus STR comparison (f). g–j, Kaplan–Meier survival plot of PFS before PSM (g); covariate balance before PSM (unadjusted) and after PSM (adjusted) (h); Kaplan–Meier survival plot of PFS after PSM (i); and results of multivariable Cox regression model of PFS for Simpson grade 1/2 versus grade 3 comparison (j). k–n, Kaplan–Meier survival plot of PFS before PSM (k); covariate balance before PSM (unadjusted) and after PSM (adjusted) (l); Kaplan–Meier survival plot of PFS after PSM (m); and results of multivariable Cox regression model of PFS for Simpson grade 1 versus grade 2 comparison (n). Horizontal error bars in the forest plots of f, j and n represent the 95% CI of the HRs for each of the covariates included in the multivariable Cox regressions. P values for each covariate in the multivariable Cox regression were derived from Waldʼs test (two-tailed) without adjustments for multiple comparisons. Panels a and b were created with BioRender.
Impact of treating dural margins
A benchmark for meningioma surgery historically has been aggressive resection of the tumor’s dural margins in addition to the tumor itself, with the view that this will further reduce recurrence risk. In instances where aggressive excision of these dural elements may not be safe or feasible (Simpson grade 1), they can be either thermocoagulated (Simpson grade 2) or left in situ (Simpson grade 3). To first determine the impact of any type of dural margin treatment, meningiomas that received either a Simpson grade 1 or 2 resection were grouped together. These meningiomas had significantly improved PFS compared to tumors that were completely macroscopically removed without any adjunctive dural excision or coagulation (Simpson grade 3; Fig. 2b,g). Conversely, among meningiomas with dural treatment (Simpson grades 1 or 2), there was no significant PFS difference between tumors that received complete dural margin excision (Simpson grade 1) versus those with dural margins treated with thermocoagulation alone (Simpson grade 2; Fig. 2b,k).
After PSM, meningiomas that received a Simpson grade 3 resection still had significantly shorter time to recurrence compared to those that received a Simpson grade 1/2 resection (HR = 1.64, 95% CI: 1.03–2.62, P = 0.038; Fig. 2h–j and Supplementary Table 24). Additionally, achieving a Simpson grade 1 resection still did not confer a significant PFS benefit compared to a Simpson grade 2 resection after PSM (HR = 0.87, 95% CI: 0.52–1.44, P = 0.58; Fig. 2l–n). These findings were consistent across other molecular classification and prognostic systems (Extended Data Fig. 5 and Supplementary Tables 4–9) and again even with PSM after multiple imputation (Supplementary Tables 10–23).
On multivariable analysis, like EOR, there was a significant interaction between Simpson grade and Molecular Group (P = 0.049). After controlling for age, sex, WHO grade, tumor location and receipt of adjuvant RT, obtaining a GTR without any dural treatment (Simpson grade 3) was associated with worse PFS in NF2-wt (HR = 3.32, 95% CI: 1.42–7.74, P = 0.005) and Immunogenic meningiomas (HR = 5.88, 95% CI: 0.96–36.16, P = 0.056; Extended Data Fig. 6a,c) compared to Hypermetabolic and Proliferative tumors; however, there was no significant PFS benefit for specifically obtaining a Simpson grade 1 resection versus a Simpson grade 2 resection for meningiomas in any Molecular Group (Extended Data Fig. 6b,d).
Overall, these findings suggest that while there is clear additive benefit to treating the dural margins in addition to complete tumor resection, pursuing aggressive dural excision in every case may not appreciably improve oncologic outcomes compared to just thermocoagulation of these attachments alone.
Defining a group of RT-resistant meningiomas
Adjuvant RT is currently the only adjunctive standard-of-care treatment aside from surgery for meningiomas and is generally selected for ‘high-risk’ patients based on WHO grade and/or the presence of residual tumor after surgery. However, response to RT in meningiomas is highly variable, and contemporary clinical practice reflects this uncertainty by including RT as a treatment option for meningiomas across all WHO grades.
To look at this more specifically, all meningiomas in the retrospective cohort were first examined. Some of the heterogeneity in overall outcomes were found to be resolved by Molecular Group classification, even for cases within each WHO grade (Fig. 3a–h). When the retrospective cohort was then stratified based on EOR (GTR versus STR) and receipt of adjuvant RT (RT versus observation), adjuvant RT significantly prolonged PFS after STR for Immunogenic meningiomas (P = 0.036) and moderately for NF2-wt meningiomas (P = 0.14) (Fig. 3i,j). For Hypermetabolic meningiomas, only marginal improvement of PFS was observed when RT was added to either GTR (P = 0.950) or STR (P = 0.136) (Fig. 3k) cases on Kaplan–Meier analysis. Proliferative meningiomas, however, had almost universally short PFS regardless of treatment paradigm, although tumors that received both GTR and adjuvant RT appeared to have incrementally better outcomes (Fig. 3l).
a, Heatmap of CNVs organized by chromosome number (vertical) and Molecular Group (MG) (horizontal) called by DNA methylation in the complete cohort of retrospective meningioma cases with available DNA methylation data. b, Principal component analysis (PCA) of all meningiomas based on DNA methylation, colored and circled by their respective MG. c–e, Molecular reclassification of meningiomas into each of the four MGs, including the proportion of cases within each respective WHO grade belonging to each MG. f–h, PFS outcomes stratified by MG of meningiomas within each WHO grade: WHO grade 1 (f), grade 2 (g) and grade 3 (h). i–l, PFS outcomes stratified by EOR in molecularly defined Immunogenic (i), NF2-wt (j), Hypermetabolic (k) and Proliferative (l) meningioma cases, with PFS groups stratified based on treatment modality (EOR and receipt of adjuvant RT). Inset shows the P values from the pairwise log-rank test between groups with Benjamini–Hochberg correction for multiple comparisons. PC, principal component.
In the retrospective cohort, the meningiomas treated with adjuvant RT after surgery had similar PFS outcomes across Molecular Groups as in the complete cohort that included cases without RT (Fig. 4a). However, there were also significant baseline differences between the RT-treated meningiomas compared to those that were observed after surgery without RT, reflecting the real-world biases of preferentially irradiating higher WHO grade, recurrent and STR cases (Supplementary Table 25). PSM was again performed to control for these specific covariates (which were the same as those used for risk stratification in the RTOG-0539 clinical trial) between the RT and observation (Obs) arms in the retrospective cohort (Fig. 4b,c). Following this and stratification of matched cases into individual Molecular Groups, adjuvant RT was now associated with significantly prolonged PFS in molecularly defined Immunogenic, NF2-wt and Hypermetabolic meningiomas but not in Proliferative meningiomas (Fig. 4c). When this analysis was repeated in only incompletely resected (STR) meningiomas—a challenging cohort to manage clinically (n = 512; Supplementary Fig. 5 and Supplementary Table 26)—the significant benefit of adjuvant RT was retained in STR Immunogenic and NF2-wt meningiomas (Supplementary Fig. 5c,d) with a signal toward longer PFS in Hypermetabolic tumors although not reaching statistical significance in this group (P = 0.150; Supplementary Fig. 5e). STR Proliferative meningiomas clearly still did not benefit from adjuvant RT (P = 0.20; Supplementary Fig. 5f).
a, PFS of cases in the retrospective molecular meningioma cohort that received adjuvant RT after surgery stratified by Molecular Group. b, Covariate balance before PSM (unadjusted) and after PSM (adjusted) for meningiomas treated with adjuvant RT (+RT) versus observation (Obs) after surgery. c, Above, PFS of the PSM retrospective cohort comparing meningiomas that received adjuvant RT (+RT) after surgery to those that were observed (Obs) in each respective Molecular Group; below, the proportion of these matched cases that received GTR and STR in each Molecular Group. d,e, Molecular cohort of RTOG-0539 clinical trial cases used as a validation cohort here, including their originally defined risk groups and their reclassification into Molecular Groups after molecular profiling. f, Genome-wide chromosomal arm-level copy number alterations called based on DNA methylation of molecularly profiled RTOG-0539 cases organized by Molecular Group. g, PFS of meningiomas from the RTOG-0539 trial that received adjuvant RT after surgery (defined as the original ‘intermediate-risk’ and ‘high-risk’ treatment arms of the trial) stratified based on Molecular Group. h, Schematic of separate PSM analysis mimicking a randomized controlled trial with the RT arm composed entirely of RT-treated meningiomas from the prospective RTOG-0539 cohort and the control/observation arm composed of retrospective meningioma cases that did not receive adjuvant RT after surgery. i, Covariate balance before and after the PSM schema outlined in k. j, PFS analysis of PSM meningioma cohort stratified based on receipt of adjuvant RT (+RT; comprising only cases from the RTOG-0539 trial) versus observation (Obs; non-RTOG-0539 cases only) in each Molecular Group.
To validate these findings using the NRG Oncology RTOG-0539 cohort, DNA methylation (n = 95) and RNA sequencing (RNA-seq) (n = 100) data were generated on these meningioma samples, and they were stratified into the four Molecular Groups in the same manner as the retrospective cohort (Fig. 4d–f and Extended Data Table 2). Outcomes of meningiomas in the prospective RTOG-0539 cohort that received surgery and adjuvant RT resembled those of the retrospective meningioma cohort, notably with Hypermetabolic and Proliferative meningiomas accounting for nearly all progressive tumors in the trial (Fig. 4g). To further validate our findings above, the prospective RTOG-0539 cases that received adjuvant RT were treated as an ‘RT’ arm, and PSM was performed for WHO grade, EOR, primary/recurrent tumor status and Molecular Group, with cases from our retrospective cohort of meningiomas that did not receive RT as the ‘observation’ arm to simulate ‘randomization’ (Fig. 4h,i). After this PSM, adjuvant RT was still significantly associated with prolonged PFS in Immunogenic, NF2-wt and Hypermetabolic meningiomas compared to observation alone but not in Proliferative meningiomas (Fig. 4j), validating that Proliferative meningiomas may be largely RT resistant. In the context of these findings, if the risk stratification of the original RTOG-0539 risk groups is re-examined, 23 of 58 (40%) cases that received adjuvant RT (‘intermediate-risk’ and ‘high-risk’) could be classified as RT-resistant Proliferative meningiomas, whereas 35 of 36 (97%) of the ‘low-risk’ (observation) group were non-Proliferative tumors that could be deemed RT responsive (Fig. 4e). This is particularly important as RT-resistant Proliferative meningiomas can be found across all WHO grades (Fig. 3c–e), providing further rationale for future molecularly informed clinical trials that may be able to triage these RT-resistant meningiomas to systemic/experimental therapies.
RT biomarker discovery in the RTOG-0539 meningiomas
In addition to validating RT results from the retrospective cohort above, the prospective RTOG-0539 trial cases were also utilized to discover new DNA methylation and gene expression biomarkers of RT response, given their uniform RT treatment plans and standardized outcome measures.
First, a DNA methylation-based model of RT response was trained using probes that were differentially methylated between the RTOG-0539 meningiomas that did and did not recur after RT within a 3-year window. This model was tested in two independent validation cohorts comprised of RT-treated meningiomas from the retrospective cohort (n = 154 and n = 122, respectively; Fig. 5a–e). This DNA methylation model was able to predict 3-year PFS after RT significantly more accurately than a model trained using standard-of-care WHO grade alone (P = 0.043, paired two-tailed DeLongʼs test; Fig. 5b,d) and WHO grade combined with other clinical covariates (Extended Data Fig. 7a,e–g). Additionally, this model was able to stratify RT-treated meningiomas into an ‘RT-resistant’ and an ‘RT-responsive’ group even for tumors within each WHO grade (Fig. 5c,e and Extended Data Fig. 7c,d). Given its design, this predictive DNA methylation model is intended to be specifically utilized for RT-treated meningiomas (to predict PFS following surgery and fractionated RT) and is made publicly available for use and testing at https://www.meningiomaconsortium.com/models/.
a, Principal component analyis (PCA) based on DNA methylation of meningiomas colored for model training (RTOG-0539 cohort) and model testing/validation (retrospective cohort). b,d, Area under the receiver operating characteristic curve (AUC ROC) and its 95% CI (shaded) for predicting 3-year PFS after RT for the RT DNA methylation model in the first (b) and second (d) retrospective validation cohorts compared to a WHO grade model. c,e, PFS after RT for meningiomas in the first (c) and second (e) retrospective validation cohort stratified based on RT DNA methylation risk score (high-risk and low-risk). f, Differential gene expression analysis on RT-treated cases in the RTOG-0539 cohort that recurred early (≤3 years; n = 16) versus cases that remained stable (n = 46). Labeled 26 genes met the cutoff of FDR-adjusted P < 0.05 (Waldʼs test with Benjamini–Hochberg adjustment) and |logFC| > 0.57 and were used for gene expression model building. g,h, The top 20 upregulated (g) and downregulated (h) gene expression pathways in the progressive RTOG-0539 meningiomas after GSEA. Unadjusted P values were obtained from permutation testing. i, AUC ROC and its 95% CI (shaded) for the prediction of 3-year PFS after RT for the RT gene expression model compared to the RT DNA methylation model and a WHO grade model tested in the combined retrospective validation cohort with matched DNA methylation and RNA-seq data. j, AUC ROC and its 95% CI (shaded) for predicting 3-year PFS after RT for the combined RT DNA methylation and gene expression model compared to a WHO grade model tested in the same cohort. k, PFS of meningiomas in the combined validation cohort dichotomized into an RT-resistant and an RT-responsive group based on combined DNA methylation and gene expression risk scores and stratified across WHO grade. l, Distribution of predicted risk scores across Molecular Group (Immunogenic n = 18; NF2-wt n = 23; Hypermetabolic n = 45; Proliferative n = 34) generated by the combined DNA methylation and gene expression model. For all box plots, the center line shows the median risk score (annotated on the plot); the box limits are the interquartile range (IQR); and whiskers show the range of data (Q1 −1.5× IQR, Q3 +1.5× IQR), with additional points as outliers. P values here were obtained from two-tailed Wilcoxon rank-sum test for pairwise comparison of means between groups. NES, normalized enrichment score; PC, principal component; yr, years.
Next, a gene expression-based model of RT response was trained using analogous RT-treated meningiomas from the RTOG-0539 clinical trial. First, differential RNA-seq analysis uncovered 26 differentially expressed genes among meningioma cases that progressed after RT versus those that remained stable (false discovery rate (FDR)-adjusted P < 0.05, log fold change (FC) > 0.57). These 26 genes were then used to build this gene expression model of RT response (Fig. 5f and Extended Data Table 3). The RT gene expression model only marginally outperformed WHO grade (P = 0.371, paired two-tailed DeLongʼs test) and appeared less discriminative of outcome (area under the curve (AUC) = 0.725, 95% CI: 0.633–0.817; Fig. 5i) than the RT DNA methylation model (AUC = 0.777, 95% CI: 0.682–0.873; Fig. 5i) in the same retrospective validation cohort of meningiomas with matched DNA methylation and RNA-seq data (n = 120) on the same tumors. Gene expression patterns in these RT-resistant meningiomas showed upregulation of cell cycle and DNA repair pathways with concomitant downregulation of cell death/apoptosis pathways (Fig. 5g,h), supportive of mechanisms underlying RT resistance.
When the predictive DNA methylation and RNA-seq signatures derived from RTOG-0539 meningiomas were combined however, model performance further improved beyond that of either model alone (0.809, 95% CI: 0.722–0.897; Fig. 5j) and significantly outperformed WHO grade (P = 7.1 × 10−3, paired two-tailed DeLongʼs test). This combined molecular predictor could similarly identify RT-resistant meningiomas within each WHO grade (Fig. 5k). Median risk scores generated using this model were also significantly higher in Proliferative meningiomas than meningiomas from all other Molecular Groups, orthogonally supporting the above finding that RT-resistant meningiomas have a largely proliferative biology (Fig. 5l).
Discussion
For this study, we leveraged the largest meningioma cohort to date with multiplatform molecular, treatment and 20-year outcome data in addition to samples from a prospective phase 2 clinical trial to identify molecular predictors of treatment response in meningiomas. We used PSM to mimic a randomized trial design to characterize the benefits of differential degrees of tumor resection and dural margin treatment across different molecular classifications, in addition to identifying a group of molecularly defined RT-resistant meningiomas. These findings have the potential to meaningfully impact clinical decision-making and treatment selection for patients in today’s molecular era of meningioma management.
Our results show that GTR imparts favorable PFS across all Molecular Groups compared to STR and although the benefits of GTR are less durable with respect to local control in more biologically aggressive meningiomas, complete resection does confer an OS benefit for these molecularly defined Proliferative meningiomas. Moreover, we show that local disease control is further improved when the meningioma’s dural margin is treated either by surgical resection or thermocoagulation in addition to complete tumor resection (either Simpson grade 1 or grade 2) compared to no marginal treatment at all (Simpson grade 3), particularly for more benign Immunogenic and NF2-wt meningiomas. Notably, however, aggressive resection of dural margins (Simpson grade 1) does not appear to provide a substantial, durable benefit compared to thermocoagulation alone (Simpson grade 2). Although delineation of the ideal extent of dural resection or coagulation to eliminate all neoplastic or pre-neoplastic disease is beyond the scope of our study, our results do suggest that the dural origin and attachments of a meningioma need to be treated in some manner to optimally delay recurrence, be it resection or cauterization. Although most treating clinicians subscribe to the maximal resection of both the meningioma and its dural attachments when possible, this had not been thoroughly investigated in a sufficiently large clinical cohort that included molecular data and the use of PSM until now8,35,36. The findings of our study revisit this historical dogma and provide rationale for a more nuanced consideration of addressing dural margins, particularly when it may be associated with higher surgical risk.
The effectiveness of RT for gross totally resected WHO grade 2 meningiomas is currently being investigated by the ongoing NRG-BN003 and ROAM/EORTC-1308 clinical trials. The results of these trials are expected to mature over the next several years and will be practice changing. Molecular biomarkers were not used for treatment stratification in these trials, as most relevant molecular classifications were discovered only after the inception of these studies37,38. To address this gap, we leveraged the largest cohort of molecularly profiled meningiomas treated with surgery and adjuvant RT, including an invaluable cohort from the prospective RTOG-0539 phase 2 clinical trial, to show that molecular stratification of meningiomas provides predictive information beyond WHO grade when considering response to RT. Our analyses suggest that adjuvant RT provides robust PFS benefits for Immunogenic and NF2-wt meningiomas, including after incomplete resection, moderate benefit for Hypermetabolic meningiomas but little-to-no benefit for aggressive, Proliferative cases. These findings support the rationale for investigating RT results for meningiomas in the context of molecular classification and considering future molecular-pathology informed clinical trials to investigate systemic treatments for RT-resistant meningiomas.
To further define patients who might benefit from RT and to provide an individualized, probabilistic risk of tumor progression/recurrence after RT, we developed molecular models following biomarker discovery in the prospective RT-treated RTOG-0539 cases using both DNA methylation and gene expression and validated the performance of these models in independent cohorts of meningioma patients who received adjuvant RT after resection. We found that, although models using either DNA methylation or gene expression alone could outperform WHO grade in predicting response to RT, DNA methylation appeared to have better performance as a standalone platform, but a model combining both molecular modalities was optimal. These models also orthogonally supported our findings that RT-resistant meningiomas have a largely proliferative biology, enriched for cell cycle, and DNA-damage repair pathways with downregulation of apoptotic processes.
The results of this study should be considered in the context of some limitations, such as the exclusion of cases without molecular data that may be biased toward different institutions; the retrospective nature of most of the cases analyzed, including for PSM analysis, which may be ideally performed with prospective cases when available; and lack of granular data on postoperative complications, which may potentially influence outcomes and RT treatment decisions.
Methods
Clinical cohort
Clinical data were collected for patients with meningiomas from 10 different participating institutions of the International Consortium on Meningiomas (ICOM) by expert clinicians, including neurosurgeons, neurooncologists and radiation oncologists, specializing in the treatment of meningiomas at their respective institution. Meningiomas were operated on between 2000 and 2020 and enriched for clinical aggressiveness or higher WHO grade tumors. Specific data elements collected were in accordance with pre-established common data elements designed for studies on meningiomas and included age, sex, tumor status (recurrent or primary at the time of surgery), WHO grade, Simpson grade, receipt of adjuvant RT, histological subtype and location39. Meningiomas were graded by experienced neuropathologists from each institution in accordance with either the 2016 or the 2021 WHO classification, and central pathological review was performed at the University Health Network in Toronto, Canada, when slides or tissue were available. Data from additional publicly available repositories were also collected. All meningiomas with molecular data were centrally re-graded in accordance with 2021 WHO classification based on the presence of CDKN2A/B homozygous deletion or TERTp mutation (Supplementary Table 27)40. Our primary outcome was PFS, defined as the time from the index surgery to the time that progression or recurrence was seen on imaging noted by either the reporting neuroradiologist at each respective institution or another treating clinician, death from any cause if the patient died without documentation of progression or the last date of neurosurgical or radiographic follow-up if the patient was subsequently lost to follow-up. Secondary outcome was OS, defined as the date from the index surgery to either the date of confirmed death or date of last follow-up, at which point the patient was censored if death was not confirmed. Meningiomas without data on WHO grade, clinical follow-up to determine PFS or OS or documented EOR were excluded (Fig. 1a). The NRG Oncology RTOG-0539 (NCT00895622) is a prospective, phase 2, non-randomized clinical trial with fully mature clinical follow-up data that stratified patients with meningiomas to adjuvant RT or observation treatment arms. In this trial, patients were divided into low-risk, intermediate-risk and high-risk groups based on WHO grade, primary/recurrent tumor status and EOR. Patients in the low-risk group were managed with observation, whereas those in the intermediate-risk and high-risk groups were treated with uniform RT protocols after surgery17,22,33. Tissue for molecular profiling and clinical data from the RTOG-0539 trial were obtained through the Molecular Profiling to Predict Response to Treatment (MP2PRT) program as part of the National Cancer Institute’s Cancer Moonshot Initiative. Ethics approval for this study was provided by the Research Ethics Board via the Coordinated Approval Process for Clinical Research at the University Health Network (Toronto, Canada) locally under CAPCR 18-5820 and by relevant institutional review boards (IRBs) at all included institutions (Indiana University, Case Western Reserve University, University of Tubingen, Vanderbilt University, Vancouver General Hospital, Fred Hutchinson Cancer Center and Northwestern University). As part of institutional policy and routine surgical consent, patients whose meningiomas were included in this study provided consent for their tumor sample, tumor data and de-identified clinical data to be used for clinical or translational research projects.
EOR
EOR was reported by the treating, experienced neurosurgeon at each respective institution as either GTR, indicating complete macroscopic removal of the tumor with or without dural coagulation or excision, or STR, indicating residual macroscopic tumor postoperatively. Postoperative magnetic resonance imaging (MRI) scans were reviewed centrally by experienced neuroradiologists for all Toronto cases with EOR data (n = 358). When available, postoperative radiographic imaging was reviewed within each institution to verify EOR (n = 680). In the remainder of cases with documented EOR, including for publicly available datasets (n = 769), it was indeterminant whether postoperative MRI was reviewed to confirm the surgeons’ impression. More granular reporting of Simpson grade was also performed based on the reporting surgeon on whether the overlying dura was excised completely (Simpson grade 1), thermocoagulated (Simpson grade 2) or left in situ without either of these measures performed (Simpson grade 3). For equivocal cases, the Simpson grade was omitted and these cases were not included in Simpson grade analyses. All meningiomas that underwent STR were designated as Simpson grade 4, and all tumors where only a biopsy was performed for tissue diagnosis (Simpson grade 5) were excluded.
DNA and RNA extraction
DNA and RNA were extracted from a combination of fresh-frozen and paraffin-embedded tissue for meningiomas using a DNeasy Blood and Tissue kit (QIAGEN, 69506), a QIAamp DNA FFPE Tissue kit (QIAGEN, 56404), a RNeasy Mini kit (QIAGEN, 74104) and an AllPrep DNA/RNA FFPE kit (QIAGEN, 80234) and quantified using a NanoDrop 1000 instrument (Thermo Fisher Scientific) and an Agilent Bioanalyzer (Agilent Technologies).
DNA methylation
The Illumina Infinium MethylationEPIC BeadChip array (Illumina) was used to obtain genome-wide DNA methylation profiles on 250–500 ng of DNA after bisulfite conversion with an EZ DNA Methylation kit (Zymo Research, D5002). General quality control (QC) measures were performed as previously described and in accordance with the manufacturerʼs instructions. Processing of methylation files (IDAT) was performed as previously described using the University of California Santa Cruz Genome Browser (GRCh38/hg38 assembly) and the minfi package (Bioconductor v.1.46.0). Differentially methylated probes/CpGs were identified using the limma package in Bioconductor41.
RNA-seq
cDNA libraries were generated as previously described on RNA samples that passed standard QC measures on the Agilent Bioanalyzer and sequenced on an Illumina NovaSeq 6000 S4 flow cell (paired end 2 × 150 base pairs (bp)) to obtain 70 million reads per sample. The quality of raw reads was initially assessed using FastQC (v.0.11.5)42,43. After this, adaptors were trimmed using Trim Galore (v.0.5.0)44. Next, STAR aligner (v.2.4.2a) was used to align reads to the human reference genome (hg38), and, subsequently, HtSeq (v.0.11.0) was employed to count reads over gene exons45,46. Gene annotation was then carried out using GENCODE (v.33)47.
Copy number alterations
Copy number alterations for each tumor were inferred from their DNA methylation data using the conumee package in Bioconductor, as previously described48,49. Chromosomal arm-level CNVs were called based on the average log2 ratio of methylated and unmethylated probe intensities across the entire arm with <−0.2 as a ‘loss’, >0.2 as a ‘gain’ and otherwise as ‘neutral’. All losses and gains were manually confirmed by independent visual inspection of the genome-wide CNV plot. Homozygous loss of CDKN2A/B was determined by manual inspection of the genome-wide CNV plots for each sample as previously described, and cases were upgraded to ‘WHO grade 3’ where homozygous loss was present in accordance with the 2021 WHO classification criteria for meningiomas40,50,51.
TERTp sequencing
Amplification of the TERT promoter region containing the hotspots C228T and C250T was performed on 973 meningioma tumor samples. TERTp mutation status was available for an additional 109 cases from the previously published study by Bayley et al.25. Platinum SuperFi II PCR Master Mix (Thermo Fisher Scientific, 12368010) with primers 5′-AGTGGATTCGCGGGCACAGA-3′ and 5′- CAGCGCTGCCTGAAACTC-3′ was used to produce a 235-bp amplicon52. Polymerase chain reaction (PCR) products were separated by gel electrophoresis to verify a product at the correct size. After purification with a ZR-96 DNA Clean-Up kit (Zymo Research, D4018), PCR products were sent for Sanger sequencing at the Centre for Applied Genomics (Toronto, Canada). Chromatograms were analyzed in Geneious Prime, and all C228T and C250T mutations were confirmed by sequencing of a second PCR reaction53.
Molecular Group stratification
The DKFZ methylation class (benign, intermediate and malignant) and subclasses (benign-1, benign-2, benign-3, intermediate-A, intermediate-B, malignant and SMARCE1-altered) were determined for each tumor through manually inputting IDAT files generated through DNA methylation into the publicly accessible brain tumor classifier v.12.5 (https://www.molecularneuropathology.org/mnp/)27. For cases that had a classification calibrated score below 0.80, a t-distributed stochastic neighbor embedding (t-SNE; Rtsne R package (v.0.16)) and a uniform manifold approximation and projection (UMAP) were plotted using reference cases from the brain tumor classifier v11.4 and institutional meningioma samples to confirm sample clustering with other known meningioma reference cases. Cases that did not classify or cluster with meningiomas in this manner or that classified or clustered with other pathologies (for example, solitary fibrous tumor/hemangiopericytoma and inflammatory cortex) were excluded from downstream molecular classification. For primary analysis, meningiomas were assigned to the four Molecular Groups originally published by Nassiri et al.24—Immunogenic (MG1), NF2-wt (MG2), Hypermetabolic (MG3) and Proliferative (MG4)—using a combination of DNA methylation, CNV and gene expression data (where available)24. The former two groups consist of predominantly benign meningiomas, whereas the latter two groups are enriched for clinically aggressive meningiomas. For sensitivity analyses, meningiomas were also assigned to other methylation/molecular classifications previously published by Sahm et al., Choudhury et al. and Bayley et al. based on the original methodologies described in each respective publication and confirmed using multiple, independent orthogonal approaches (Extended Data Figs. 1 and 2)24,25,26,28,29,34,54. In brief, DNA methylation was used to derive the UCSF MethG and Baylor MenG. Additional confirmation of the fidelity of each tumor’s molecular classification was performed using three additional, common methods. First, a multinomial elastic net (GLMnet) classifier was trained on the cohort from which the original molecular classification was derived from using DNA methylation and RNA-seq data where available and then applied to all other datasets from the other cohorts to predict the probability of each sample belonging to a specific molecular class or group55. Next, non-negative matrix factorization (NMF R package (v.0.27)) was used to extract metagenes from the DNA methylation and RNA-seq data of each of the original molecular classification’s training cohort and applied these metagene signatures to cluster all other samples into molecular/methylation groups56,57. In rare instances where there were discrepancies between the predicted molecular/methylation groups across any of these methodologies, two independent blinded reviewers (V.P. and A.P.L.) manually examined the CNV plots and determined a consensus classification based on the degree of changes pathognomic to each molecular/methylation group (for example, chr 1p loss, 14 loss, 10 loss, 5 gain, 12 gain, 22q loss and CDKN2A/B deletion). Meningiomas were also stratified to the following three integrated molecular prognostic systems—Integrated Grade, Morphomolecular Risk Group and Gene Expression Risk—based on the methodologies described in each group’s original publication28,29,34. For the former two prognostic systems that integrate chromosomal arm-level CNVs into classification, losses and gains of respective chromosomal arms were defined based on the cutoffs specified above (see ‘Copy number alterations’ subsection) and confirmed by manual inspection of the genome-wide CNV plots. Chromosomal arm-level changes were defined as ‘gain’, ‘loss’ or ‘neutral’, and points were allocated for specific losses in keeping with what was previously published for Integrated Grade and Morphomolecular Risk Group. Even partial chromosomal arm losses or gains were considered, in keeping with previous reports54. For Gene Expression Risk, only cases with available RNA-seq (including all cases with matched DNA methylation) were able to be used. The Gene Expression model was trained in the publicly available UCSF cohort (n = 185), using their methodologies as originally published, by computing the product of the LASSO regularized Cox regression coefficient (from the glmnet and cv.glmnet functions of the glmnet package (v.4.1-8) in R) of the predefined 34-gene signature and the normalized counts of the 34 prognostic genes55. This value was then rescaled from 0 to 1 to define a probabilistic risk metric. Bootstrap aggregation was used to train 500 ridge-regression submodels to determine a bootstrap aggregated risk score for each training sample, and these risk scores were then applied to the validation cohorts with available RNA-seq data34.
PSM
PSM was performed using a 1:1 or 2:1 (when Simpson grades 1 and 2 were grouped together to compare to Simpson grade 3) nearest neighbor matching algorithm without replacement with a predefined caliper width ranging from 0.10 to 0.20. PSM was performed based on the same following variables unless otherwise defined in the text: age, sex, WHO grade, primary/recurrent tumor status, receipt of adjuvant RT, tumor location and one of each of the respective molecular classification schemes (Molecular Group, UCSF MethG and Baylor MenG) and integrated molecular prognostic systems (Integrated Grade, Morphomolecular Risk and Gene Expression Risk)24,25,26,28,29,34. Covariate balance was assessed by computing the standardized mean difference after PSM. To address missing clinical data (non-outcome and non-molecular data), multiple imputation was performed using the ‘mice’ package in R to impute five separate complete datasets; PSM was performed within each imputed dataset followed by averaging the treatment effects of a multivariable Cox regression within each dataset; and then the Rubin formula was used to combine estimates from these multiply imputed full datasets into a single set of results (results of the pooled Cox multivariable model) using the MatchThem package in R58,59.
RT DNA methylation model building
Meningiomas from the intermediate-risk and high-risk groups of the prospective RTOG-0539 trial were used to develop molecular models capable of accurately predicting response to RT on an individual patient level17,22. These tumors received consistent and uniform adjuvant RT treatment plans, allowing us to develop a specific predictor of RT response. DNA methylation and RNA-seq data were generated on all RTOG-0539 meningiomas to stratify tumors into their respective Molecular Groups as detailed above. Both DNA methylation and gene expression data were used to build predictive models that could individualize risk of early recurrence after surgery and adjuvant RT. For feature selection in our DNA methylation model, univariable Cox regression analysis for PFS in the RTOG-0539 cohort of meningiomas that received surgery and adjuvant RT (intermediate-risk and high-risk groups from the original trial; n = 58 meningioma cases) were computed for all probes covered by the EPIC 850K and v.2.0 array to create a rank list of prognostic, discriminative probes. A total of 1,791 probes that were significantly correlated with PFS after surgery and adjuvant RT at a significance level of P < 0.001 were selected to move forward to the subsequent feature selection step. Next, differential methylation analysis was performed between meningiomas from this same RTOG-0539 cohort that progressed/recurred after surgery and adjuvant RT within 3 years of treatment versus those that remained stable within the same period. This was done to further refine the specificity of DNA methylation probes associated with progressive/recurrent disease cases versus stable disease after surgery and RT. These 15,137 significant (FDR-adjusted P < 0.05), highly differentially methylated probes (delta beta ≥0.10 or ≤−0.10) were then overlapped with probes selected from the univariable analysis (1,791 abovementioned probes), and a set of common probes specific for response to RT were selected as the final features for model building (685 probes). To build the molecular RT model, extreme gradient boosting (XGBoost; xgboost R package (v.1.7.5.1)) modeling was performed using these final 685 probes in the RTOG-0539 training cohort to create and tune a molecular DNA methylation-based predictor of early (3-year) recurrence/progression after surgery and adjuvant RT. A 3-year cutoff was used as this was also the primary outcome of interest in the RTOG-0539 clinical trial. This model was then evaluated on two independent validation cohorts consisting exclusively of meningiomas from the University Health Network (Toronto) that received both surgery and RT (n = 154) and a cohort with cases combined from multiple other institutions (Indiana University, Fred Hutchinson Cancer Center, Case Western, Tubingen, Unity Health and Vancouver General Hospital; n = 122). Model performance for predicting PFS at 3 years (primary outcome of the RTOG-0539 trial) was assessed by generating receiver operating characteristic (ROC) curves and computing the average AUC and its 95% CI, using bootstrap resampling with 10,000 resamples (pROC R package (v.1.18.4)). The performance of this model was compared with an analogously trained XGBoost model using the 2021 WHO grade alone. There was no overlap between the samples used to train these models and the samples used in the testing/validation cohorts. The 95% CIs for the AUCs of both models were generated using bootstrap resampling with 10,000 resamples for the validation cohorts only. The difference in performance between the two models was compared by computing an average ΔAUC (AUCMolecular RT Model − AUCClinical Standard of Care Model) and its 95% CI in the same independent validation cohorts. Determination of an optimal cutpoint to dichotomize the validation cohorts into an RT-resistant and an RT-sensitive group was determined using the cutpointr package for a risk score cutoff that maximized the Youden index for predicting PFS after surgery and RT (RT-specific PFS)60.
RT gene expression model building
To develop a prognostic set of genes that could predict response to RT, differential gene expression analysis was performed between meningioma cases in the RTOG-0539 cohort with gene expression data that recurred early (≤3 years) after surgery and adjuvant RT (n = 14 cases) and cases that remained stable after this timepoint (n = 45). Differential gene expression was assessed using DESeq2 (v.1.26.0), using negative binomial generalized linear models for each gene61. Subsequently, Waldʼs test was applied to obtain P values, followed by Benjamini–Hochberg adjustment for controlling multiple testing. Genes were ranked based on a combined metric incorporating both FC direction and computed P values, computed as sign(logFC) × −log10(P value). Here, the sign of logFC indicates the direction of change (positive for upregulation and negative for downregulation), whereas −log10(P value) indicates the significance level. Subsequently, Gene Set Enrichment Analysis (GSEA, v.3.0) was conducted as previously detailed, using the ranked scores to discern potential involvement of differentially expressed genes in shared biological pathways. For feature selection of specific genes that may determine response to RT, an FDR cutoff of 0.05 and a |logFC| cutoff of 0.57 (representing a 1.5× change in expression) were set, above which genes would be considered as significantly prognostic. This produced a predictive set of 26 genes (Extended Data Table 3) that was subsequently passed into a LASSO Cox regression model in our RTOG-0539 training cohort to obtain a set of coefficients for each gene. Risk scores were generated by multiplying the coefficients for each gene by their normalized counts (counts per million (CPM)) and scaled to a value between 0 and 1. In keeping with previous publications, bootstrap aggregation was then used to train 500 ridge-regression submodels using normalized and log-transformed gene counts as the input variable and risk scores generated in the training cohort as the output to obtain an aggregated risk score for each sample (rms R package (v.6.7-1) and hdnom R package (v.6.0.2))34. These risk scores were then applied to all cases in the retrospective validation cohort (n = 120) comprising cases with matched DNA methylation and RNA-seq data, none of which was used to train the model. Model performance was assessed in the same manner as the DNA methylation model and benchmarked against a model built using standard-of-care WHO grade alone.
Combined DNA methylation and RNA-seq model
To develop a model that used both DNA methylation and gene expression data, the risk scores generated by the two respective models above and the same cases in the RTOG-0539 cohort with matched DNA methylation and RNA-seq data (n = 58) were used. The probability scores derived from both the XGBoost methylation model and the RNA expression risk score derived from the 26-gene expression model were normalized to a range of 0–1 and served as the exclusive training features for a glmnet model. Logistic regression with the ‘binomial’ family was employed for model training using the glmnet method. This model was subsequently applied to a matched retrospective validation cohort with corresponding DNA methylation and RNA-seq data for meningioma analysis. Assessment of model performance and discrimination of an optimal cutpoint to dichotomize the validation cohort were completed using the same methodologies as above for each of the individual molecular models.
Clinical and molecular nomogram development
Two distinct elastic net Cox proportional hazards models were developed to forecast PFS after surgery and RT, leveraging methylation data and clinical variables, correspondingly. The choice of the elastic net regularization method aimed to address potential multicollinearity and identify the most informative predictors. Model training was conducted using a training dataset derived from the RTOG cohort. Cross-validation employing three folds facilitated the optimization of model hyperparameters, with the lambda.min rule used to ascertain the optimal penalization parameter. Nomograms were subsequently devised based on the trained models to visually represent the anticipated probability of 3-year PFS after RT for individual patients. These nomograms encompassed the chosen predictors, including methylation probabilities, grade, EOR and RT dose, alongside their corresponding coefficients derived from the Cox models. Predicted probabilities were specifically generated at a time point of 3 years after RT. The predictive efficacy of the methylation-based model underwent evaluation using an external validation dataset. Time-dependent AUC statistics were computed to gauge the discriminatory capacity of the model across varying time intervals. Additionally, the concordance index (C-index) and calibration plots were scrutinized to assess model performance. AUC ROC values were calculated for these models using the same methodology as described above for the DNA methylation and gene expression models, along with their respective 95% CIs, achieved through bootstrap resampling comprising 10,000 resamples.
Statistical analysis
All statistical analyses were carried out using R v.4.2.2 (ref. 12). PFS was compared between groups using a Cox proportional hazards model. Statistical significance was taken to be P < 0.05 unless otherwise specified. Survival analyses were performed using Kaplan–Meier methods and fitting Cox proportional hazards models. The proportional hazards assumption was tested by plotting the scaled Schoenfeld residuals of covariates in the model against time and using both individual and global Schoenfeld tests (Supplementary Fig. 1).
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
DNA methylation data and gene expression data (RNA sequencing) generated for this study are deposited in the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) under the superseries accession number GSE270375. Unprocessed DNA methylation data from the retrospective meningioma tumors are available under the accession number GSE270371 and gene expression counts data are available under the accession number GSE270638. Unprocessed RNA sequencing data (FASTQ) are deposited in the NCBI Sequence Read Archive (https://www.ncbi.nlm.nih.gov/sra) under the project number PRJNA1127224 for the retrospective meningiomas. Unprocessed DNA methylation data and gene expression data (RNA sequencing) for the prospective NRG RTOG-0539 clinical trial cases are deposited in the database of Genotypes and Phenotypes (dbGaP) under the project no. phs003707.v1.p1 entitled MP2PRT-MNG: identifying novel molecular markers of response to radiotherapy in meningiomas using samples from the RTOG-0539 (NCT00895622). Previously published, publicly available data were downloaded from the GEO database (https://ncbi.nlm.nih.gov/geo) under the following accession numbers: GSE189521 (Bayley et al. DNA methylation) and GSE183656 (Choudhury et al. DNA methylation and RNA sequencing)25,26.
Code availability
The open-source software, tools and packages used for data analysis in this study, as well as the version of each program, are as follows: ImageJ (v.2.1.0), R (v.4.3.1), FastQC (v.0.11.5), STAR (v.2.4.2a), caret R package (v.6.0-94), Rtsne R package (v.0.16), survival R package (v.3.5-5), xgboost R package (v.1.7.5.1), glmnet R package (v.4.1-8), rms R package (v.6.7-1), hdnom R package (v.6.0.2), pROC R package (v.1.18.4), minfi (Bioconductor v.1.46.0), DESeq2 (Bioconductor v.1.26.0), HtSeq (v.0.11.0), MatchIt R package (v.4.5.5) and MatchThem R package (v.1.1.0).
References
Ostrom, Q. T. et al. CBTRUS statistical report: primary brain and other central nervous system tumors diagnosed in the United States in 2015–2019. Neuro Oncol. 24, v1–v95 (2022).
Vernooij, M. W. et al. Incidental findings on brain MRI in the general population. N. Engl. J. Med. 357, 1821–1828 (2007).
Simpson, D. The recurrence of intracranial meningiomas after surgical treatment. J. Neurol. Neurosurg. Psychiatry 20, 22 (1957).
Gallagher, M. J., Jenkinson, M. D., Brodbelt, A. R., Mills, S. J. & Chavredakis, E. WHO grade 1 meningioma recurrence: are location and Simpson grade still relevant? Clin. Neurol. Neurosurg. 141, 117–121 (2016).
Gousias, K., Schramm, J. & Simon, M. The Simpson grading revisited: aggressive surgery and its place in modern meningioma management. J. Neurosurg. 125, 551–560 (2016).
Heald, J. B., Carroll, T. A. & Mair, R. J. Simpson grade: an opportunity to reassess the need for complete resection of meningiomas. Acta Neurochir. (Wien) 156, 383–388 (2014).
Nanda, A. et al. Relevance of Simpson grading system and recurrence-free survival after surgery for World Health Organization grade I meningioma. J. Neurosurg. 126, 201–211 (2017).
Sughrue, M. E. et al. The relevance of Simpson grade I and II resection in modern neurosurgical treatment of World Health Organization grade I meningiomas. J. Neurosurg. 113, 1029–1035 (2010).
Ehresman, J. S. et al. The relevance of Simpson grade resections in modern neurosurgical treatment of World Health Organization grade I, II, and III meningiomas. World Neurosurg. 109, e588–e593 (2018).
Lee, G. et al. Timing of adjuvant radiotherapy in atypical meningiomas. Int. J. Radiat. Oncol. Biol. Phys. 108, S189 (2020).
Sarhan, N. et al. Short course hypofractionated radiotherapy for frail or elderly patients with meningioma. Cureus 12, e8604 (2020).
Chen, W. C. et al. Radiotherapy for meningiomas. J. Neurooncol. 160, 505–515 (2022).
Wang, J. Z. et al. Outcomes and predictors of response to fractionated radiotherapy as primary treatment for intracranial meningiomas. Clin. Transl. Radiat. Oncol. 41, 100631 (2023).
Press, R. H. et al. Outcomes and patterns of failure for grade 2 meningioma treated with reduced-margin intensity modulated radiation therapy. Int. J. Radiat. Oncol. Biol. Phys. 88, 1004–1010 (2014).
Fleming, C. W., Parsai, S. & Suh, J. H. A management dilemma: adjuvant radiotherapy after gross total resection of atypical meningioma. Transl. Cancer Res. 8, 1–3 (2019).
Soldà, F. et al. Long-term efficacy of fractionated radiotherapy for benign meningiomas. Radiother. Oncol. 109, 330–334 (2013).
Rogers, C. L. et al. High-risk meningioma: initial outcomes from NRG Oncology/RTOG 0539. Int. J. Radiat. Oncol. Biol. Phys. 106, 790–799 (2020).
Zhu, H. et al. Efficacy of adjuvant radiotherapy for atypical and anaplastic meningioma. Cancer Med. 8, 13–20 (2019).
Maclean, J., Fersht, N. & Short, S. Controversies in radiotherapy for meningioma. Clin. Oncol. (R. Coll. Radiol.) 26, 51–64 (2014).
Lee, G. et al. Adjuvant radiation therapy versus surveillance after surgical resection of atypical meningiomas. Int. J. Radiat. Oncol. Biol. Phys. 109, 252–266 (2021).
Goldbrunner, R. et al. EANO guideline on the diagnosis and management of meningiomas. Neuro Oncol. 23, 1821–1834 (2021).
Rogers, L. et al. Intermediate-risk meningioma: initial outcomes from NRG Oncology RTOG 0539. J. Neurosurg. 129, 35–47 (2018).
Nassiri, F. et al. DNA methylation profiling to predict recurrence risk in meningioma: development and validation of a nomogram to optimize clinical management. Neuro Oncol. 21, 901–910 (2019).
Nassiri, F. et al. A clinically applicable integrative molecular classification of meningiomas. Nature 597, 119–125 (2021).
Bayley, J. C. T. et al. Multiple approaches converge on three biological subtypes of meningioma and extract new insights from published studies. Sci. Adv. 8, eabm6247 (2022).
Choudhury, A. et al. Meningioma DNA methylation groups identify biological drivers and therapeutic vulnerabilities. Nat. Genet. 54, 649–659 (2022).
Sahm, F. et al. DNA methylation-based classification and grading system for meningioma: a multicentre, retrospective analysis. Lancet Oncol. 18, 682–694 (2017).
Maas, S. L. N. et al. Integrated molecular-morphologic meningioma classification: a multicenter retrospective analysis, retrospectively and prospectively validated. J. Clin. Oncol. 39, 3839–3852 (2021).
Driver, J. et al. A molecularly integrated grade for meningioma. Neuro Oncol. 24, 796–808 (2021).
Choudhury, A. et al. Hypermitotic meningiomas harbor DNA methylation subgroups with distinct biological and clinical features. Neuro Oncol. 25, 520–530 (2022).
Schwartz, T. H. & McDermott, M. W. The Simpson grade: abandon the scale but preserve the message. J. Neurosurg. 135, 488–495 (2020).
Lemée, J. M. et al. Extent of resection in meningioma: predictive factors and clinical implications. Sci. Rep. 9, 5944 (2019).
Rogers, C. L. et al. Low-risk meningioma: initial outcomes from NRG Oncology/RTOG 0539. Neuro Oncol. 25, 137–145 (2023).
Chen, W. C. et al. Targeted gene expression profiling predicts meningioma outcomes and radiotherapy responses. Nat. Med. 29, 3067–3076 (2023).
Ammendola, S. et al. The histopathological diagnosis of atypical meningioma: glass slide versus whole slide imaging for grading assessment. Virchows Arch. 478, 747–756 (2021).
Nagashima, G. et al. Dural invasion of meningioma: a histological and immunohistochemical study. Brain Tumor Pathol. 23, 13–17 (2006).
Jenkinson, M. D. et al. The ROAM/EORTC-1308 trial: radiation versus observation following surgical resection of atypical meningioma: study protocol for a randomised controlled trial. Trials 16, 519 (2015).
Jenkinson, M. D. et al. Radiotherapy versus observation following surgical resection of atypical meningioma (the ROAM trial). Neuro Oncol. 16, 1560–1561 (2014).
Nassiri, F. et al. Consensus core clinical data elements for meningiomas (v2021.1). Neuro Oncol. 24, 683–693 (2022).
Louis, D. N. et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol. 23, 1231–1251 (2021).
Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015).
Ward, C. M., To, T. H. & Pederson, S. M. ngsReports: a Bioconductor package for managing FastQC reports and other NGS related log files. Bioinformatics 36, 2587–2588 (2020).
Andrews, S. FastQC: a quality control tool for high throughput sequence data. http://www.bioinformatics.babraham.ac.uk/projects/fastqc/ (2010).
Krueger, F. Trim Galore! http://www.bioinformatics.babraham.ac.uk/projects/trim_galore/ (2019).
Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).
Anders, S., Pyl, P. T. & Huber, W. HTSeq—a Python framework to work with high-throughput sequencing data. Bioinformatics 31, 166–169 (2015).
Harrow, J. et al. GENCODE: the reference human genome annotation for The ENCODE Project. Genome Res. 22, 1760–1774 (2012).
Mah, C. K., Mesirov, J. P. & Chavez, L. An accessible GenePattern notebook for the copy number variation analysis of Illumina Infinium DNA methylation arrays. F1000Res. 7, ISCB Comm J-1897 (2018).
Capper, D. et al. Practical implementation of DNA methylation and copy-number-based CNS tumor diagnostics: the Heidelberg experience. Acta Neuropathol. 136, 181–210 (2018).
Sievers, P. et al. CDKN2A/B homozygous deletion is associated with early recurrence in meningiomas. Acta Neuropathol. 140, 409–413 (2020).
Wang, J. Z. et al. Increased mRNA expression of CDKN2A is a transcriptomic marker of clinically aggressive meningiomas. Acta Neuropathol. 146, 145–162 (2023).
Spiegl-Kreinecker, S. et al. TERT promoter mutations are associated with poor prognosis and cell immortalization in meningioma. Neuro Oncol. 20, 1584–1593 (2018).
Kearse, M. et al. Geneious Basic: an integrated and extendable desktop software platform for the organization and analysis of sequence data. Bioinformatics 28, 1647–1649 (2012).
Hielscher, T. et al. Clinical implementation of integrated molecular-morphologic risk prediction for meningioma. Brain Pathol. 33, e13132 (2022).
Engebretsen, S. & Bohlin, J. Statistical predictions with glmnet. Clin. Epigenetics 11, 123 (2019).
Frigyesi, A. & Höglund, M. Non-negative matrix factorization for the analysis of complex gene expression data: identification of clinically relevant tumor subtypes. Cancer Inform. 6, 275–292 (2008).
Tamayo, P. et al. Metagene projection for cross-platform, cross-species characterization of global transcriptional states. Proc. Natl Acad. Sci. USA 104, 5959–5964 (2007).
Zhang, Z. Propensity score method: a non-parametric technique to reduce model dependence. Ann. Transl. Med. 5, 7 (2017).
Zhang, Z. Multiple imputation with multivariate imputation by chained equation (MICE) package. Ann. Transl. Med. 4, 30 (2016).
Thiele, C. & Hirschfeld, G. cutpointr: improved estimation and validation of optimal cutpoints in R. J. Stat. Softw. 98, 1–27 (2021).
Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).
Acknowledgements
We would like to thank our colleagues, including investigators and site personnel, at each respective institution who contributed to this study. At the University of Toronto, we would like to thank members of our laboratory—Q. Wei, S. Mansouri, O. Singh and T. Tung—for their help in sample procurement. We thank Julissa Tsao, Troy Ketela, and all the staff of the Princess Margaret Genomics Centre (https://pmgenomics.ca), as well as Karen Ho and the staff at The Centre for Applied Genomics (https://tcag.ca) for their expertise in generating some of the sequencing data used in this study. At the Indiana University School of Medicine, we would like to thank medical students P. Virtanen, N. Burket, M. Blackwell, J. Koenig, A. Alfonso and J. Davis for their help with clinical data collection. At Vanderbilt University, we would like to acknowledge B. C. Mobley and R. Thompson for their efforts in providing clinical and pathologic samples. For the NRG Oncology RTOG-0539 cases, total RNA sequencing was performed by the High Throughput Sequencing Facility at the University of North Carolina School of Medicine, under the supervision of C. Jones and K. Hoadley. Additionally, we would like to thank key members of NRG Oncology, the Cancer Therapy Evaluation Program (CTEP) and Leidos Biomedical Research for their collaborative efforts as part of the Molecular Profiling to Predict Response to Treatment (MP2PRT) program, including C. Szot, Z. Abdullaev, S. DeVries, L. Longoria, S. Marcus and K. Johnson. This project was funded in part with federal funds from the National Cancer Institute, National Institutes of Health, under the Cancer Moonshot Initiative (contract number HHSN261201500003I; task order number HHSN26100039). The content of this publication does not necessarily reflect the views or policies of the US Department of Health and Human Services nor does mention of trade names, commercial products or organizations imply endorsement by the US Government. Additional funding for this study was provided by the Canadian Institutes of Health Research (CIHR) Project Fund (RN482811-481519; G.Z., F.N. and J.Z.W.), Brain Tumor Charity United Kingdom (GN-000430; GN-000693; G.Z. and F.N.), the UHN Foundation, Mary Hunter Meningioma Research Fund, the V Foundation for Cancer Research (G.Z.), the CIHR Vanier Scholarship (F.N., J.Z.W. and C.G.), the American Association of Neurological Surgeons Neurosurgery Education & Research Foundation Research Fellowship (F.N. and J.Z.W.), the Congress of Neurological Surgeons Tumor Section (F.N. and J.Z.W.) and the Princess Margaret Hospital Foundation Hold ‘em For Life Oncology Fellowship (F.N., J.Z.W., A.P.L. and C.G.).
Author information
Authors and Affiliations
Consortia
Contributions
Manuscript writing, figure preparation, general statistical analysis and general bioinformatics were performed by J.Z.W. Molecular classification, model building and bioinformatics support was provided by V.P., A.P.L. and J.L. TERT promoter sequencing was performed by C.G. Copy number variant analysis and calling were performed by A.P.L. Molecular analyte extraction, quantification and sequencing were performed by J.Z.W., A.P.L., C.G., A.A. and Z.P. DKFZ methylation class and subclass determination was performed by R.Y. and R.K. using the publicly available online tool created by F.S. RNA sequencing analysis was performed by V.P. and J.L. Biostatistical review was performed by O.S. Clinical data from the NRG Oncology RTOG-0539 meningioma cases were provided by S.L.P. and M.W. Clinical meningioma samples and matching clinical data were provided by C.W., A.C.G., M.A.Z., G.T., M.T., F.B., D.A.A.B., E.C.H., J.B.S., A.E.S., C.H., S.C., L.B.C., A.R., S.M. and S.Y. Pathologic review was performed by A.G. and K.A. Radiotherapy data of institutional retrospective meningioma cases were provided by D.S.T. Study design and conception were done by J.Z.W, F.N. and G.Z. All authors participated in manuscript review and revisions in some capacity.
Corresponding authors
Ethics declarations
Competing interests
S.Y. is a member of advisory boards and has received honoraria from AstraZeneca, Amgen, Bayer, Janssen, Pfizer, Roche and Servier. The remaining authors declare no competing interests.
Peer review
Peer review information
Nature Medicine thanks Annette Molinaro, E. Antonio Chiocca and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Ulrike Harjes, in collaboration with the Nature Medicine team.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended Data Fig. 1 Copy number variation (CNV) plot in each molecular cohort.
CNVs were derived based on inference from DNA methylation data organized based on molecular group (MG) assignment and by institution, including publicly available datasets utilized as well as the prospective RTOG-0539 clinical trial cohort.
Extended Data Fig. 2 Reclassification of meningiomas into different previously published molecular classifications and integrated prognostic systems.
a. Heatmap of cases with matched DNA methylation and RNA sequencing on the same meningioma utilized for outcome prediction and generation of Brier error curves b. Brier error curves over time for prediction of PFS at 5-years in the complete molecular cohort with matched DNA methylation and RNA-sequencing data demonstrating that Molecular Group (MG) had the lowest Brier error across multiple different molecular classifications and prognostic systems with integrated molecular features. Integrated Brier score or cumulative prediction error/cumulative rank probability score for each molecular classification and prognostic system are listed in the parentheses. (c-e) Sankey diagrams showing the distribution of cases from each WHO Grade and MG into the DKFZ methylation subclasses (c), the UCSF Methylation Groups (MethG) (d), and the Baylor Meningioma Groups (MenG) (e). (f-h) Kaplan-Meier (KM) survival curves of PFS based on cases classified into the DKFZ methylation subclasses (f), MethG (g) and MenG (h). (i-k) KM survival curve of OS based on cases classified into the DKFZ methylation subclasses (i), MethG (j) and MenG (k). (l-n) Sankey diagram showing the distribution of cases from each WHO Grade and MG stratified into Integrated Grade (l), Morphomolecular Risk (m), and Gene Expression Risk groups (n). (o-q) KM survival curves of PFS based on cases stratified by Integrated Grade (o), Morphomolecular Risk (p), and Gene Expression Risk Group (q). (r-t) KM survival curve of OS stratified by Integrated Grade (r), Morphomolecular Risk (s), and Gene Expression Risk Group (t). Insets shows the P values obtained from the pairwise log-rank test between groups with Benjamini–Hochberg correction for multiple comparisons.
Extended Data Fig. 3 Benefits of extent of surgical resection (EOR) across Molecular Groups (MG).
(a-b) Kaplan-Meier (KM) survival curve showing PFS of meningiomas that received a GTR vs STR belonging to the following MG: Immunogenic (a), NF2-wildtype (b), Hypermetabolic (c), Proliferative (d). e. Results of the multivariable Cox regression analysis assessing the PFS benefits of EOR while controlling for age, sex, WHO grade, primary/recurrent tumor status, and receipt of adjuvant RT with an interaction term between EOR and MG. (f-i) KM survival curve showing OS of meningiomas based on EOR (GTR vs STR) in each MG: Immunogenic (f), NF2-wildtype (g), Hypermetabolic (h), 4: Proliferative (i). j. Results of multivariable Cox regression analysis on the effect of EOR on OS while controlling for age, sex, WHO grade, primary/recurrent tumor status, and receipt of adjuvant RT with an interaction term between EOR and MG. Box size for the forest plots in (e) and (j) are relative to the weight of the effect size with its center representing the HR and the horizontal error bars representing the 95% CI.
Extended Data Fig. 4 Univariable and multivariable analyses of PFS and OS across molecular classifications and prognostic systems.
(a, i) Univariable Cox proportional hazards model of PFS (a) and OS (i) in the complete retrospective cohort of meningiomas. (b-h) Multivariable Cox proportional hazards models of PFS before PSM with meningiomas classified in the following different molecular strata: MG (b), DKFZ Methylation Subclass (c), UCSF MethG (d), Baylor MenG (e), Integrated Grade (f), Morphomolecular Risk (g), and Gene Expression Risk (h). (j-p) Multivariable Cox proportional hazards models of OS before PSM with meningiomas classified in the following different molecular strata: MG (j), DKFZ Methylation Subclass (k), UCSF MethG (l), Baylor MenG (m), Integrated Grade (n), Morphomolecular Risk (o), and Gene Expression Risk (p). Horizontal error bars in the forest plots of represent the 95% confidence interval of the hazard ratios for each of the covariates included in the Cox regression models presented. P-values for each covariate in the multivariable Cox regression were derived from the Wald test (two-tailed) without adjustments for multiple comparisons.
Extended Data Fig. 5 PSM analyses using other molecular classifications and integrated molecular prognostic systems.
(a-x) From left to right: love plot, KM survival curve, and results of multivariable Cox regression analyses after PSM for GTR vs STR, Simpson grades 1/2 vs 3, Simpson grade 1 vs 2, and RT vs observation using the DKFZ methylation subclasses (a-d), UCSF MethG (i-l), Baylor MenG (q-t), Integrated Grade (e-h), Morphomolecular Risk (m-p), and Gene Expression Risk (u-x). Horizontal error bars in the forest plots of represent the 95% confidence interval of the hazard ratios for each of the covariates included in the multivariable Cox regression models presented. P-values for each covariate in the multivariable Cox regression were derived from the Wald test (two-tailed) without adjustments for multiple comparisons.
Extended Data Fig. 6 Prognostic role of Simpson Grade resection across Molecular Groups.
a. PFS in each Molecular Group (MG) based on grouping meningiomas that received a Simpson grade 1 or 2 resection together vs those that received a Simpson grade 3 resection. b. PFS in each Molecular Group for meningiomas that received a Simpson grade 1 resection vs a Simpson grade 2 resection. (c-d) Results from the multivariable Cox proportional hazards model showing the effect of Simpson grade on PFS when controlling for age, sex, WHO grade, receipt of adjuvant RT, and tumor location with an interaction term between Simpson Grade and Molecular Group. Box size for the forest plots in (c) is relative to the weight of the effect size with its center representing the HR and horizontal error bars representing the 95% confidence interval of the hazard ratios for each of the covariates included in the multivariable Cox regression models presented.
Extended Data Fig. 7 RTOG-0539 trained DNA methylation model of RT response and molecular nomogram.
a. Representative screenshot of DNA methylation .idat files uploaded for 2 separate samples and their respective probabilistic risk of recurrence within 3 years of surgery and adjuvant RT on the publicly available RTOG DNA methylation predictor site. b. representative magnetic resonance images of the same sample cases uploaded to the predictor in panel (a) with one RT-resistant case demonstrating recurrence within 3 years of clinical follow-up after gross total resection and adjuvant radiotherapy (above), and a different case in another patient (below) demonstrating interval stability after 3-years following similarly gross total resection and adjuvant RT. (c-d). RT-specific PFS outcomes of meningiomas within each WHO grade of the first (c) and second (d) retrospective validation cohorts stratified into an RT-resistant and RT-responsive group based on the predicted DNA-methylation based RT risk score. P-values generated from Log-rank test. e. Molecular nomogram built using the DNA methylation based RT predictor, WHO grade, EOR, and RT dose to predict PFS post-surgery and adjuvant RT using the RTOG-0539 cohort as the training cohort. Each variable in the nomogram is scaled accordingly and values/scores for each variable are assigned points which cumulatively add up to estimate a probabilistic risk of recurrence within 3-years of treatment. f. clinical nomogram built using the same variables and training cohort as in (e) except without DNA methylation risk score. g. AUC and 95% CI demonstrating the predictive accuracy of the molecular and clinical nomogram respectively validated in the same retrospective cohort of RT-treated meningiomas (N=276).
Supplementary information
Supplementary Information
Supplementary Figs. 1–5 and Supplementary Tables 1–27
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Wang, J.Z., Patil, V., Landry, A.P. et al. Molecular classification to refine surgical and radiotherapeutic decision-making in meningioma. Nat Med 30, 3173–3183 (2024). https://doi.org/10.1038/s41591-024-03167-4
Received:
Accepted:
Published:
Version of record:
Issue date:
DOI: https://doi.org/10.1038/s41591-024-03167-4
This article is cited by
-
Two molecular subtypes of skull base meningiomas with extracranial extension
Acta Neuropathologica Communications (2025)
-
Collision tumor, a metastatic melanoma within a meningioma: a case report
Acta Neuropathologica Communications (2025)
-
Integrated molecular and detailed anatomical profiling identifies a prognostically adverse subtype of posterior fossa meningiomas: high-risk copy number alterations are associated with midline predilection and predict poor prognosis
Acta Neuropathologica Communications (2025)
-
The spectrum of radiation therapy options for craniopharyngioma: a systematic review
Journal of Neuro-Oncology (2025)
-
Overall survival and progression-free survival in pediatric meningiomas: a systematic review and individual patient-level meta-analysis
Journal of Neuro-Oncology (2025)