1. Introduction
Meningioma is the most common primary tumor in the central nervous system with mesodermal-arachnoid origin. Around 80% of meningiomas are benign and curable by surgical resection alone. However, 20% of meningioma will recur after initial surgical operations and a further comprehensive treatment regimen is required [
1]. Currently, the histoimmunochemistry-based WHO pathology grade system is the main predictor for patient outcomes with meningioma. WHO grade I meningiomas are usually defined as “benign” tumors while grade II (atypical) and grade III (anaplastic) meningiomas are more aggressive and to recur [
2]. Meningiomas of high grade (grade II and III) are considered as having higher risk of recurrence than that in grade I meningiomas. Although the 2016 WHO classification of central nervous system tumors histologically subdivides meningioma into 15 subtypes, it does not indicate the prognosis of patients precisely [
3]. To precisely diagnose meningioma and complement the grade system, a few investigations on the genetics and epigenetics of meningioma have been performed recently.
Precise cancer diagnoses are critical for the suitable treatment strategy for cancer patients, therefore, a few studies were under investigation on the biological biomarkers for meningioma. Neurofibromin 2 (NF2) mutation is the first characterized alteration in meningioma and is observed in approximately 80% of high grade meningiomas, suggesting that it is potentially a prognostic biomarker. Structural variants including loss of chromosomes 6q, 9p, 10q, 14, and 18q, as well as gains in 17q and 20q are also associated with recurrence of meningioma [
4,
5]. Gene expression signature shows its association of patient survivals and another panel of two gene (PTTG1 and LEPR) expressions demonstrates association with the recurrent meningioma [
6,
7,
8]. More recently, a DNA methylation-based classification defining six distinct subtypes with clinical relevance reveals a stronger association of clinical outcomes than conventional WHO classification. Unsupervised clustering of DNA methylation profiles in meningiomas identifies two distinct subgroups associated with distinct recurrence-free survival [
9]. Similarly, Sahm et al. also identified two major and six minor subgroups of meningioma from DNA methylation profile with significantly different clinical behaviors and non-NF2 mutated meningiomas are clustered in one “benign” subgroup [
10]. In addition, a recent study displays a 64-CpG loci-based predictor that indicates the risk of meningioma recurrence [
11]. Overall, these studies highlighted the significance of biological biomarkers for the purpose of precise diagnosis and individualized management of meningioma.
The consensus clustering of cancer transcriptomes yields a robust and precise diagnosis for heterogeneous tumors, especially in brain tumors. Glioblastoma multiforme, the most malignant glioma, is reclassified into four subtypes with distinct transcriptome and clinical outcomes [
12]. Similarly, children with wingless (WNT) subtype of medulloblastoma have the most favorable outcomes than children with others subtypes of medulloblastoma [
13]. In addition, patients with posterior fossa group A or supratentorial REL-associated protein (RELA)-positive ependymoma show dismal prognosis of all subtypes of ependymoma and the risk stratification by comprehensive molecular subgrouping is superior to histological grading [
14]. These subtyping of heterogeneous tumors shed a light on precision diagnosis and treatment in cancer patients. However, whether meningioma could be robustly clustered into subtypes by transcriptomes with clinical disperse behaviors remains unclear.
Immunotherapy is recently emerging as a new hope for cancer treatment and management. As vasculature is enriched in meningioma, peripheral and central system immune cells are present in meningioma, which enable the application of immunotherapy. Exclusion of cytotoxic CD8+ T cells and CD4+ helper T cells are observed in high-grade meningioma, suggesting the enhanced T cells therapy is potentially treatment for recurrent meningioma [
15]. The abundance of PD-L1 expression in meningioma also suggests that meningioma patients likely benefit from the immune checkpoint inhibition [
16]. Therefore, a few clinical trials are under investigation to explore the efficacy of checkpoint inhibition in high-grade and/or recurrent meningioma [
17]. Despite that, we are still not clear whether the immune infiltration has shared the same pattern across all types of meningioma.
In this study, we consolidated the total RNA sequencing results from 179 WHO I, II or III meningiomas and identified four biologically distinct subtypes of meningioma using consensus clustering. These subtypes of meningioma also have mutually different frequency of DNA methylation pattern, gene fusion and infiltrating immune cell profiles. Further, our results of random-forest-based machine learning identified Alkaline Phosphatase (ALPL) is the feature gene in the most aggressive subtype of meningioma. We established and validated Meningioma Progression score (MPscore) to characterize the risk of progression in meningioma. Our subclassification provides a further insight on the understanding of biological behaviors in meningioma.
2. Materials and Methods
2.1. Data Preprocessing and Tissue Sample Validation
RNA sequencing profiles of 179 meningioma were included in this study [
18,
19]. The raw reads were preprocessed prior to the transcriptome analysis. Trim-galore was utilized to remove adapters and low-quality reads with the follow parameters (-q 25 --length 50 -e 0.1 --stringency 5). The trimmed reads were subjected to alignment against hg38 by STAR 2.5.3a with default settings. FeatureCounts of Subread 1.6.4 was utilized for quantifying the counts of the RNA transcripts. Batch effects were corrected by removeBatchEffect. The Transcripts per kilobase million (TPM) rate was calculated for normalization and subjected to downstream analyses.
For DNA methylation array datasets, a cohort of 39 meningiomas from a previous study was utilized for the investigation of DNA methylation alteration between subtypes [
20]. Raw signals were quantile-normalized before the removal of the probes located in sexual chromosomes and the single nucleotide polymorphisms (SNP). The crossing-reactive probes were also removed before the downstream analyses. The M values were used for statistical analyses while the beta values were used for the data visualization and biological interpretation. We utilized the manifest “IlluminaHumanMethylation450kanno.ilmn12.hg19” for the CpG probes annotation.
A small cohort of six meningioma samples along with two independent datasets were utilized for the meningioma progression scoring validation. The detailed methods were described in the supplementary methods.
2.2. Consensus Clustering
Principal component analysis (PCA) was employed for dimension reduction exploration. The distances of samples were determined by the root-mean-square deviation (Euclidean distance) of the top 2000 genes. Hierarchical clustering with agglomerative average linkage was performed in this study, as our basis for consensus clustering, to detect the robust clusters. The distance metric 1-(Pearson’s correlation coefficient) was used for variances detection between samples. SigClust was performed to establish the significance of the clusters in a pairwise fashion. All subtype identification was performed by the package “ConsensusClusterPlus”.
2.3. Differentially Expressed Gene (DEG) Analysis
Differentially expressed gene (DEG) analysis was performed for identification of featured genes in each subtype. DEG was performed on linear modelling of indicated (co-) variates on expression values by limma (Ritchie et al. 2015). p-values generated from limma modelling were corrected for multiple hypothesis testing by Benjamini and Hochberg false discovery rate (FDR) adjustments. Each subtype was tested against all other groups to generate this subtype featured genes. The FDR-adjusted p-values < 0.05 and |log Fold Change (FC)| > 2 were considered statistically significant. The DEGs between subtypes were visualized by heatmaps.
2.4. Stemness Index (SI) Prediction
A stemness index (SI) model utilizing an OCLR algorithm on pluripotent stem cells was generated by Malta et al. to predict the proportion of stem cells per given cancer sample [
21]. We applied this stemness index model to the 179 meningiomas using Spearman correlation for RNA expressions. The whole workflow is available from:
https://bioinformaticsfmrp.github.io/PanCanStem_Web/ (accessed on 1 August 2020).
2.5. Copy Number Alteration (CNA) and DNA Methylation Analysis
The copy number alteration (CNA) was calculated by the package “conumee” using DNA methylation dataset. All chromosomes alterations of each included samples were calculated and plotted by “CNV.genomeplot” with default setting. For structural variation, the chr1p or chr22q loss (mean of chromosomal arm less than 0.1) was selected by a reduction of copy number in chr1q or chr22q [
19]. The stemness index was predicted for each meningioma. The CpG island methylation phenotype (CIMP) was defined as that most variable CpG loci (a standard deviation larger than 0.2 in a certain subtype) were hypermethylated. Subtype-specific CpG loci signatures were determined by the package “limma” using M value in a pairwise fashion and the |logFC| > 2 and adjusted
p value < 0.05 was considered as the statistical significance.
2.6. Immune Cell Infiltration Prediction
By applying the Microenvironment Cell Populations-counter (MCPcounter) method, the abundances of eight immune cells infiltrating meningioma were predicted to explore the feasibility of immunotherapy in meningioma [
22]. Eight immune cell proportions were compared in each subtype of meningiomas. PD-L1 expression in each subtype of meningiomas was also compared.
2.7. Fusion Genes Identification
We employed the STAR-Fusion 1.8.1 against hg38 to detect the fusion genes in meningiomas with default settings. The most common fusions in all and subtypes of meningiomas were compared to uncover the subtype featured fusion. Fusions spanning two mutually different chromosomes were considered as the interchromosomal fusions.
2.8. Random Forest Model
The most importance of transcripts in subtype of meningiomas were identified by random forest (RF) using the “randomForest” package with default settings. The cross-validated prediction performance of this model was iterated by the function “rfcv” with 10-fold cross-validation and a removal of 1.5 variables in each step. The discrimination of the most important gene (variable) for meningioma in the given subtype was predicted “pROC” against all other subtypes. The area under the curve (AUC) was utilized here as the accuracy for subtype prediction.
2.9. Meningioma Progression Score (MPscore)
The R package “ssGSEA” was utilized to construct the meningioma progression score (MPscore) per given sample using the “GSVA” package [
23]. Prior to MPscore construction, we selected DEGs of subtype 3 meningioma as the progressive gene signature and reference. For accurately surrogating the progressive phenotype, we sub-divided the gene list into up- or down-regulated gene lists in subtype 3, calculated the scores through ssGSEA, respectively. The MPscore was the sum of the difference of ssGSEA-predicted scores from up- or downregulated gene list.
2.10. Statistical Analysis
R software version 3.5.1 (R Core Team, Vienna, Austria) was used for all statistical analyses. Student’s t test was used for the statistical comparison of two groups. ANOVA was performed to test the statistical significance between more than three groups and Tukey’s honestly significant difference (HSD) test was conducted as a post hoc test when the results of ANOVA indicated significance. A p value less than 0.05 was considered statistically significant.
4. Discussion
Accumulating evidence has indicated meningioma is a heterogeneous tumor but the feature of clinical recurrent meningioma is still unclear. Through the consensus clustering of meningioma transcriptomes, we identified four mutually distinct subtypes of meningioma and all grade III and most grade II are clustered in subtype 3 of meningioma. This result suggests that the characterization of the differentially expressed genes in subtype 3 will uncover the biological features of meningioma progression (malignancy and recurrence). Meningioma in subtype 3 has distinct gene expression pattern to other subtypes, which confirms the heterogeneity of meningioma. Our analyses of noncoding RNAs and DNA methylation also further confirms the biological complexity of the tumorigenesis and progression in meningioma. As machine learning is widely used for biomarker discovery and potential therapeutic targets screening in cancer research [
28,
29] and random forest (RF)-based classification and featured variable identification has demonstrated the advantages of non-overfitting and robustness over the conventional differentially expressed gene analysis [
30], we employed RF for integration of feature gene in subtype 3.
ALPL is identified as the top featured gene of all subtype 3 DEGs contributing to the features in subtype 3 of meningioma. The product of ALPL is a membrane bound glycosylated enzyme that broadly participating in phosphatase activity and alkaline phosphatase activity, therefore, the alteration of ALPL is found to be associated with hypophosphatasia and prostate cancer bone metastasis [
31,
32]. Our results suggest ALPL is likely a subtyping and recurrence diagnostic biomarker with a significant accuracy, which is consistent with the previous studies [
33,
34]. Although RF predicts TIMP3, INMT and SLC16A1 followed by ALPL as the important feature genes in subtype 3, the expressions of these three genes are not consistently lower or higher than all other subtypes of meningioma, which restricts their potent of being biomarkers for meningioma. The molecular pathway alterations caused by the reduced ALPL in meningioma is still needed to be investigated further though a few studies uncover the association of ALPL and NOTCH1 regulation in human epithelial cells and ALPL is one of the key hub genes in glioblastoma [
35,
36].
In order to identify the progressive subtype in meningioma, we constructed MPscore based on the featured genes in subtype 3. Our MPscore demonstrates a remarkably discriminative capacity for progressive meningioma in two independent cohorts. A few attempts have been made to explore the recurrence or progression related genes in meningioma. However, there is an inconsistence of meningioma progressive gene lists between studies [
6,
37,
38,
39]. Direct comparison between grade I and anaplastic or atypical meningioma unlikely provides the generic progressive genes because high-grade transformation has already occurred in grade I meningioma in early stage [
40,
41]. We construct a reference-based MPscore for the prediction of meningioma progression. We also validated our MPscore in three independent cohorts, highlighting the clinical utility of MPscore across different profiling sources (RNA seq or microarray). A total of 53 genes are included as the progressively phenotypic references in our study. In our MPscore gene reference, LEPR is another well-characterized prognostic biomarker and independent predictive biomarker for meningioma. Loss of function of LEPR is associated with the elevated leptin levels and obesity, showing its participating in adipose biogenesis [
42]. Although the biological regulation of LEPR in meningioma is still unclear, a few observative studies uncover patients with meningioma are prone to be obese [
43,
44]. Our clustering is consistent with the main findings of previous studies, which also suggests the robustness of our analyses. Patel et al. find a subgroup of meningioma has a shorter recurrence-free survival by clustering of WHO grade I and II meningioma while the other two subgroups have relatively low risks of recurrence. Our results of clustering echo their findings where a few grade I meningiomas (subtype 3) have similar transcriptome pattern to the recurrent meningioma while most grade I meningiomas are clustered into two distinct subtypes (subtype 1 and 2). Notably, our results further reveal these grade I meningiomas that likely recur share similar transcriptome patterns with grade III meningioma, confirming that progressive transformation happens in grade I meningioma. Another clustering of WHO grade I meningioma indicates there are five subgroups in meningioma but four of five subgroups are enriched in the WNT pathway. In our analyses, WNT-pathway-related genes such as DKK2 are involved in subtype 1 signature gene. Together with previous studies, the clustering by transcriptomes highlights the heterogeneity and the genetic variation in meningioma and subtyping of meningioma enable us to identify the meningioma with high risk of recurrence.
One previous study found some novel gene fusions involved in NF2, the most common mutation in meningioma. However, the NF2 gene fusion-based biomarker for the prediction of meningioma progression or recurrence might not be reliable in clinical practice as radiation therapy could induce new NF2 mutations or structural variants in meningioma [
10,
20]. Patients may not benefit from these biomarkers targeting postradiation NF2 gene fusions as these NF2 genes do not predict the risk of recurrence in the initial diagnosis of meningioma. Although our study fails to identify the subtype or recurrence specific biomarkers, we discovered that gene fusion with EYA1 is the most common fusion across all subtypes of meningioma, indicating EYA1 fusion is likely an early event in tumorigenesis in meningioma. EYA1 is a protein phosphatase and a transcriptional coactivator for SIX1 that regulates gene expression and cellular proliferation. In meningioma, EYA1 is also a key molecule by regulating the cell viability and cell cycle [
45]. Compared with other types of brain tumor, meningioma has significantly higher expression of EYA1. Our results support the critical role of EYA1 in meningioma though further investigation is required to validate altered activated pathways caused by EYA1 fusion. These results shed a light on the tumor formation and potential therapeutic targets for meningioma.
The four subtypes of meningioma we identified also have different immune cell infiltration and PD-L1 expression. Previous small-scale studies display T cells and B cells infiltrated in meningiomas are antigen-experienced and monocytes are also present in meningioma, indicating the feasibility of immunotherapy application in meningioma [
46,
47]. Our results show that a broad spectrum of immune cells infiltration in meningioma, and moreover, we show monocytic lineage is the most predominant of all immune cell types. Of all subtypes, subtype 1 has the least immune cells (including monocytes, T cells and B cells) infiltration. That is likely because all subtype 1 meningiomas are benign and composed from WHO grade I meningioma. As noted, subtype 3 has the highest cytotoxic T cells infiltration of all subtypes, suggesting the enhanced T cells could be a potential therapeutic for patients in subtype 3 while NK-cell-based therapeutics may benefit patients in subtype 4 [
48]. Our analysis of PD-L1 in meningioma is partially consistent with previous study where higher grade meningioma has higher levels of PD-L1 [
45]. Our results show the levels of PD-L1 in subtype 3 and 4 are higher than that in subtype 2. However, we also found subtype 1 composed from WHO grade I meningioma has relatively higher level of PD-L1. We presume there probably is a post-transcriptional modification or regulation of PD-L1 in meningioma so it is rarely detected by antibody-based immunohistochemistry [
27,
45,
49,
50].
We noticed there are a few limitations in this study. Firstly, due to the accessibility to the materials and datasets, we do not crossvalidate our subtype clustering with the clustering by DNA methylation and mutation information. A comprehensive landscape of meningioma integrating mutation, structural variant, DNA methylation, RNA (mRNA and ncRNA) transcripts and proteomics will help us understand the biological behaviors of meningioma recurrence. Histopathological analyses of immune cells infiltration and cell cycle markers will also complement the microenvironment landscape of meningioma. Secondly, although we demonstrated that grade II meningiomas is likely presented in Subtype 3 and Subtype 3 meningiomas usually have a high expression level of mitotic genes, the association of biological subtypes with histopathological types remains unclear [
51]. More studies on the distinct biological features within different histopathological types are under investigation. Thirdly, the prediction capacity of ALPL for meningioma recurrence needs to be validated externally. In this study, the high-grade meningioma surrogates the recurrent meningioma, which might be a selection bias. Therefore, prior to the clinical application, the assay designed for ALPL also should be elaborately tested for clinical utility in a large perspective cohort where meningioma patients developing to recurrence are recruited.