Immunologic Profiling of Mutational and Transcriptional Subgroups in Pediatric and Adult High-Grade Gliomas

Published OnlineFirst July 2, 2019; DOI: 10.1158/2326-6066.CIR-18-0939

Research Article Cancer Immunology Research Immunologic Proﬁling of Mutational and Transcriptional Subgroups in Pediatric and Adult High-Grade Gliomas Michael Bockmayr1,2,3, Frederick Klauschen2, Cecile L. Maire4, Stefan Rutkowski1, Manfred Westphal4, Katrin Lamszus4, Ulrich Schuller€ 1,3,5, and Malte Mohme4

Abstract

Immunologic treatment strategies are under investigation presenting cell (APC)/natural killer (NK) cell/T-cell– for high-grade gliomas. Determining relevant immunologic dominated immune clusters. Immune cell expression profiles pathways is required for invigorating a tumor-specific immune correlated with transcriptional and mutational subgroups response. We therefore investigated the immunologic pheno- but were independent of age and histologic diagnosis. By types within different subgroups of high-grade gliomas, with a includingfunctional pathways andcorrelatingtheexpression of focus on rare genetic subgroups of pediatric and adolescent immunostimulatory and -inhibitory receptor–ligand interac- patients to identify potentially targetable mechanisms. We tions, we were able to define the immunologic microenviron- gathered published gene-expression data from 1,135 high- ment and identify possible immunologic subtypes associated grade glioma patients and applied a machine-learning tech- with poor prognosis. In addition, comparison of overall sur- nique to determine their transcriptional (mesenchymal, clas- vival with the immunologic landscape and with checkpoint sic, neural, and proneural) and mutational [K27, G34, IDH, molecules revealed correlations within the transcriptional and wild type (WT)] subtypes. Gene signatures of infiltrating and mutational subgroups, highlighting the potential applica- immune cells and functional immune pathways were evalu- tion of PD-1/PD-L1 checkpoint inhibition in K27-mutated ated in correlation to histologic diagnosis, age, and transcrip- tumors. Our study shows that transcriptional and mutational tional and mutational subgroups. Our analysis identified subgroups are characterized by distinct immunologic tumor four distinct microenvironmental signatures of immune microenvironments, demonstrating the immunologic hetero- cell infiltration (immune 1–4), which can be stratified into geneity within high-grade gliomas and suggesting an immune- vascular, monocytic/stromal, monocytic/T-cell–, and antigen- specific stratification for upcoming immunotherapy trials.

Introduction apy, tumor recurrence is a certainty (1). Although treatment modalities such as immunotherapeutic approaches with check- High-grade gliomas are defined by aggressive and infiltrative point inhibitors have failed to show efficacy in a phase III study in growth, reducing the life expectancy in pediatric and adult recurrent glioblastomas (2, 3), ongoing studies are trying to patients (1). Despite multimodal treatment, including neurosur- harness the ability of the immune system to counteract tumor gical resection and combination of radiotherapy and chemother- growth and induce long-term remission (4). To identify which patients might benefit from checkpoint inhibition or other immu- 1Department of Pediatric Hematology and Oncology, University Medical Center notherapeutic approaches, such as tumor vaccines, chimeric anti- Hamburg-Eppendorf, Hamburg, Germany. 2Charite – Universitatsmedizin€ Berlin, gen receptor (CAR) T-cell therapy, autologous T-cell transfer, or corporate member of Freie Universitat€ Berlin, Humboldt-Universitat€ zu cytolytic virus injections, we must better understand the glioma- Berlin, and Berlin Institute of Health, Institute of Pathology, Berlin, Germany. specific immune microenvironment (5). Large-scale immunolog- 3Research Institute Children's Cancer Center Hamburg, Hamburg, Germany. ic profiling of gene-expression data in other cancer entities has 4 Department of Neurosurgery, University Medical Center Hamburg- yielded information to improve precision immunotherapy (6–8). Eppendorf, Hamburg, Germany. 5Institute of Neuropathology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany. Unfortunately, the intricate cellular, genetic, and transcriptional heterogeneity of malignant gliomas adds complexity to efforts to Note: Supplementary data for this article are available at Cancer Immunology define targetable immunologic profiles (9, 10). Research Online (http://cancerimmunolres.aacrjournals.org/). The 2016 World Health Organization (WHO) classification € U. Schuller and M. Mohme contributed equally to this article. identified subcategories among high-grade gliomas (11). Molec- Corresponding Authors: Malte Mohme, Department of Neurosurgery, ular profiling, especially by isocitrate dehydrogenase (IDH) gene University Medical Center Hamburg-Eppendorf, Martinistr. 52, 20246 mutation analysis (12), improves understanding of subgroup- Hamburg, Germany. Phone: 49-40-7410-0; Fax: 49-40-7410-58121; E-mail: specific aggressiveness and prediction of overall survival (OS) in [email protected]; and Ulrich Schuller,€ Research Institute Children's Cancer adult patients. Transcriptomic studies have identified additional Center Hamburg, Martinistr. 52, N63 (HPI), 20251 Hamburg, Germany. Phone: 49- 40-4260-51240; Fax: 49-40-7410-40350; E-mail: [email protected] subgroups. Phillips and colleagues described different molecular subgroups of astrocytoma and their prognostic relevance, and Cancer Immunol Res 2019;7:1401–11 then Verhaak and colleagues refined this classification by intro- doi: 10.1158/2326-6066.CIR-18-0939 ducing a gene-expression–based molecular classification of glio- 2019 American Association for Cancer Research. blastoma that stratifies tumors into mesenchymal, classic, neural,

www.aacrjournals.org 1401

Bockmayr et al.

and proneural subgroups (13, 14). These groups share features of (version 19; ref. 29). All data originating from the same center/ biological aggressiveness and response to therapy (13). The study were transformed to median 0, as there were still remaining proneural subgroup was further subdivided into cGIMP and batch effects (Supplementary Fig. S1A). Because the data were non-cGIMP methylated tumors (15). The neural subtype has not homogeneous with respect to mutational profile, localiza- been excluded in many subsequent studies due to suspected tion, and grade, no adjustment of the variability of each gene was contamination of tumor samples with normal tissue (14, 16). performed between series measured on the same microarray. In addition to the groups defined by transcriptional subtypes or However, the mean absolute deviation, which is less influenced mutations in the IDH gene in adult tumors, genomic sequencing by outliers than the standard deviation, was aligned between the studies also identified lineage-defining aberrations, such as the two microarray platforms for each gene separately. After batch- somatic mutations of the histone 3 variants in the positions G34 effect removal, clustering of the data revealed concordance of and K27 (17–19). Furthermore, pediatric glioblastoma can be similar tumors (with respect to diagnosis, mutational profile, or stratified by additional molecular subgroups according to their transcriptional subgroup) from different data sets. RTK or MYCN pathway aberrations (19). The K27 and G34 Duplicates defined by a Pearson correlation > 0.999 were mutations in histone 3 were predominantly found in pediatric removed, yielding a final data set of 79 anaplastic astrocytomas and adolescent glioblastoma, highlighting their biological dis- (AA), 28 diffuse intrinsic pontine gliomas (DIPG), and 1,028 tinctiveness and the need to stratify high-grade gliomas according glioblastomas. Analyses involving both arrays were restricted to to their mutational profile. the 12,061 common genes. Accumulating data suggest that molecularly distinct glioma Regarding the information on the IDH subgroup, the study by subgroups differ in their tumor microenvironment and their Gravendeel and colleagues performed sequencing of the IDH1 immunostromal profiles (20, 21). Investigations of The Cancer locus only (30), whereas TCGA data sets included sequencing Genome Atlas (TCGA) data set, predominantly including IDH information of the complete IDH gene (13, 31). wild-type (WT) glioblastomas, revealed heterogeneity in the composition of the immune environment between mutational Computation of immune and stromal signatures and transcriptional subgroups (22). Wang and colleagues dem- To obtain specific signatures of tumor-infiltrating immune and onstrated a subgroup-dependent macrophage/microglia infiltra- stromal cells, we first defined candidate genes from the signatures tion according to the gene signature, with increased infiltration in published by Becht and colleagues (32) and Danaher and col- NF1-mutated tumors (23). This was also confirmed by gene- leagues (33) for 10 microenvironmental cell populations (MCP): þ enrichment analysis for transcriptional subtypes by Doucette and T cells, CD8 T cells, cytotoxic lymphocytes, B lineage cells, colleagues (24) and by our group by gene-expression profiling for natural killer (NK) cells, monocytic lineage cells, myeloid den- different molecular subgroups in medulloblastoma (25). dritic cells, neutrophils, fibroblast-like cells, and endothelial cells Our analysis evaluates the subtype-specific immune microen- (Supplementary Tables S2 and S3). Second, all candidate genes vironment across a cohort of pediatric and adult high-grade were manually reviewed, and those likely to overlap with glioma gliomas, comparing mutational and transcriptional profiles in signatures were removed (Supplementary Table S2). In a final order to identify immune infiltration patterns and functional step, a correlation network was built by linking all genes pathways mediating local immunosuppression. with Pearson correlation over 0.4 for each signature separately. Only those genes belonging to the largest correlation module were retained as final candidate genes, because uncorrelated genes Materials and Methods are unlikely to be specifically expressed by a unique cell All data analyses were performed using the statistical pro- population (33). gramming language R (26), including the packages survival, caret, Scores of tumor-infiltrating immune and stromal cell signa- randomForest, igraph, gplots, and Rtsne (https://CRAN.R-project. tures were defined as the average values of the corresponding org/package¼survival, https://CRAN.R-project.org/package¼caret, marker genes. Signatures for different immunologic processes http://igraph.org, https://CRAN.R-project.org/package¼gplots, were obtained from Doucette and colleagues (24) and Thorsson http://www.jmlr.org/papers/v9/vandermaaten08a.html, and and colleagues (34) and computed analogously (Supplementary https://cran.r-project.org/package¼randomForest). Table S3). This analysis is based on gene-expression data, and no histologic or flow-cytometric analysis was performed. Conclu- Data sets, preprocessing, and batch-effect removal sions of the presence of different immune cells were drawn based Raw gene-expression data (.CEL files) from high-grade gliomas on expression of genes or gene signatures. including eight series on the Affymetrix Human Genome U133 Plus 2.0 Array and one series on the Affymetrix HT_HG-U133A Inference of mutational status and transcriptional subtype array were downloaded from Gene Expression Omnibus Inference of the mutational status (IDH, K27, G34, and WT) (GEO; Fig. 1A; Supplementary Table S1; ref. 27). and the transcriptional group (classic, mesenchymal, proneural, To minimize batch effects, the fRMA normalization, and neural) were performed with a random forest classifier which includes a batch-effect correction, was used as imple- coupled with a confidence estimation using the R-package ran- mented in the R-package fRMA (28) using the input vectors domForest. The classifier was built on the 1,000 most variable from the R-packages hthgu133ahsentrezgfrmavecs and probes in the training sets with 1,000 trees. Default settings were hgu133plus2-hsentrezgfrmavecs (http://bioconductor.org/ used for the remaining parameters. Selection of the most variable packages/release/data/annotation/html/hthgu-133afrmavecs. genes was based on the standard deviation computed for each html, and http://bioconductor.org/packages/release/data/ training fold separately. annotation/html/hgu133-plus2frmavecs.html) and the custom As the frequency of IDH, K27, and G34 mutations was age CDF files hthgu133ahsentrezgcdf and hgu133plus2hsentrezgcdf dependent in our data set, separate classifiers were built for young

1402 Cancer Immunol Res; 7(9) September 2019 Cancer Immunology Research

Immune Proﬁling of Pediatric and Adult Gliomas

A Study Origin n Histology Chip PMID GEO ID Ceccarelli 2016 TCGA 524 GBM HT-u133A 26824661 GSE83130 Rembrandt Los Angeles 241 GBM u133p2 19208739 GSE108474 Griesinger 2013 Denver 34 Pediatric Gliomas u133p2 24078694 GSE50161 Lee 2008 Los Angeles 26 (267) GBM Mixed 18940004 GSE13041 Gravendeel 2009 Rotterdam 175 All gliomas u133p2 19920198 GSE16011 Sturm 2012 Heidelberg 46 GBM u133p2 23079654 GSE36245 Sturm 2016 Heidelberg 16 (182) PNET, mixed u133p2 26919435 GSE73038 Paugh 2010 Memphis 48 Pediatric HGG u133p2 20479398 GSE19578 Figure 1. Paugh 2011 Memphis 25 DIPG + Brainstem u133p2 21931021 GSE26576 Data collection and in silico prediction of transcriptional and mutational n = 1135 subtype. A, We gathered data from B D Mutational prediction nine different studies using microarray Instances Random forest gene-expression proﬁling. Numbers WT delineate high-grade glioma samples ... *1,000 included in this study, whereas rue class

numbers in parentheses display the IDH K27 T overall number of available data sets Tree 1 Tree 2 Tree n

within the described study. B and C, Class A Class B Class X G34 Next, we performed a random forest WT K27 IDH G34 % of Predict- approach (schematically shown, B), Majority voting Predicted class able cases Final class followed by internal cross-validation Subclass prediction (schematically shown, C) to predict the

mutational and transcriptional C PN subtype. D, Cross tables illustrate Training fold Test fold prediction accuracy compared with known data sets. E, Horizontal bar Cross-validation MN NE diagrams display the origin and cohort Determination of class True model accuracy

composition of mutational and CL

transcriptional subtype before (left) PN NE MN CL % of Predict- and after (right) the application of the Predicted class able cases prediction algorithm. AA, anaplastic astrocytoma; CL, classic transcriptional 0% 50% 100% subgroup; DIPG, diffuse intrinsic Transcriptional Mutation Subtype E Age (years) Diagnosis Mutation pontine glioma; GBM, glioblastoma; subtype predicted predicted G34, G34-mutated glioma; HGG, high- TCGA (n = 524) grade gliomas; IDH, IDH-mutated glioma; K27, K27-mutated glioma; MN, Rembrandt (n = 241) mesenchymal transcriptional subgroup; N/A, not available; NE, Denver (n = 34) neural transcriptional subgroup; PN, proneural transcriptional subgroup; Los Angeles (n = 26) PNET, primitive neuroectodermal tumors; WT, IDH wild-type glioma. The n asterisk () indicates that the random Rotterdam ( = 175) forest algorithm was repeated 1,000 times and marks a fold. Heidelberg (n = 62)

Memphis (n = 73)

0 50 100 0 50 100 0 50 100 0 50 100 0 50 100 0 50 100% n n 0−19 (n = 158) AA (n = 79) G34 (n = 15) CL (n = 154) G34 ( = 17) CL ( = 295) K27 (n = 42) MN (n = 374) 20−39 (n = 151) DIPG (n = 28) K27 (n = 37) MN (n = 173) n n IDH (n = 127) NE (n = 155) 40−59 (n = 425) GBM (n = 1028) IDH ( = 84) NE ( = 61) WT (n = 821) PN (n = 231) 60−89 (n = 401) WT (n = 527) PN (n = 107) n n N/A (n = 472) N/A (n = 640) N/A ( = 128) N/A ( = 80) Prediction accuracy: Prediction accuracy: >91% >91%

patients [reference set: all WT (n ¼ 42), K27 (n ¼ 36), and G34 group size was reduced to at most twice the number of samples (n ¼ 13) patients under 20 years, as well as 20 IDH-mutated of the smallest class for the WT/K27/G34/IDH classifier by patients under 30 years] and older patients [reference set: all downsampling. Group size was reduced to the size of the IDH-mutated (n ¼ 79) and all WT (n ¼ 485) patients over smallest class for the two other classifiers. 20 years]. Patients with a known transcriptional subgroup from Estimates of the classification accuracy were obtained from the TCGA cohort were chosen as reference for the transcriptional 5-fold cross-validation. As no parameter optimization was per- subgroup classifier. To consider imbalanced class sizes, each formed, nested cross-validation was not required. Because raw

www.aacrjournals.org Cancer Immunol Res; 7(9) September 2019 1403

Bockmayr et al.

classification results yielded accuracies under 90% for each task, unknown mutational and transcriptional subtype information. we used tail probabilities to increase accuracy, and samples with a To do so, we applied a random forest approach to separately low proportion of votes for the predicted class were excluded and predict the mutational and transcriptional subgroups of the considered unclassifiable. To this end, the lowest threshold value missing samples (Fig. 1B). Machine learning was able to predict was determined for each classification task such that samples from the mutational subgroup, defined by either histone H3 mutations the validation fold with prediction score over the threshold were at positions G34 or K27, as well as IDH mutations, or tumors correctly assigned with an accuracy of over 91%. This allowed harboring a wild-type for IDH (WT) and no histone H3 increasing accuracies over 91% for each of the classification tasks. mutations (17–19). Next, machine learning predicted the tran- The random forest classifier was compared with a radial basis scriptional subgroup, defined by classic, mesenchymal, neural, function kernel support vector, which did not yield superior and proneural signatures, by extrapolating from the known classification accuracy in this setting. expression profiles. In an additional in silico validation step, the model accuracy was determined by internal cross-validation Heat mapping, differential expression, k-means clustering, (Fig. 1C). Using tail probabilities and applying a threshold of t-distributed stochastic neighbor embedding, and survival >91% prediction accuracy (Fig. 1D), we were able to predict the analysis mutational subgroup in an additional 344 cases and the tran- Data were trimmed at z-scores 3 prior to heat map analysis. scriptional subgroup in 560 additional cases based on the known Heat maps were generated using the average-linkage method and gene-expression profiles (Fig. 1E). The age distribution of the Pearson correlations as similarity measure. Significance of differ- individual groups is shown in Supplementary Fig. S1B. We kept ential expression between subgroups was assessed using the the neural phenotype in our prediction model despite the con- Kruskal–Wallis test. Multiple testing correction was performed troversial literature in this regard (16). We did this, although data with the Benjamini–Hochberg (BH) method. suggest that a neural profile might be contaminated with large Immunostromal clusters were defined using k-means clustering amounts of nonneoplastic brain cells. Including the neural group of the previously defined marker genes. The number of clusters therefore enabled us to investigate the immune microenviron- (n ¼ 6) was selected using the elbow method. For visualization, ment of high-grade gliomas that are particularly diffuse or infil- t-distributed stochastic neighbor embedding (t-SNE) was used trated by normal brain cells. with Euclidian distance, perplexity 50, 1,000 iterations, and a learning rate h ¼ 100. Default values were chosen for the remain- Optimization of immune cell population signature on the ing parameters. Survival analysis was performed using the pro- glioma data set portional hazards model as implemented by the coxph function To analyze the abundance of tumor-infiltrating immune cell from the R-package survival. Expression values for the different populations by inferring from gene-expression data, several gene signatures were transformed to quartiles and included as numer- signatures have been published using different tools (32, 33, 41). ical values. Multiple testing correction for the cell populations/ These approaches are based on gene-expression signatures of the signatures was performed using the BH method where appropri- cell types under consideration and are mostly derived from ate if not otherwise indicated. pancancer analyses. However, some of the known signatures also include genes that are expressed during neuronal development or tumorigenesis of malignant brain tumors. We therefore manually Results reviewed these signatures and tested the correlation of the pub- Prediction of mutational and transcriptional subgroups of lished immune cell population signatures in our data set for high-grade gliomas internal correlation to extract a glioma-specific immune subset Our aim in the present study was to combine gene-expression signature, which is independent of tissue-specific interferences profiling data from the Affymetrix U133A and U133p2 gene array (Fig. 2; Supplementary Table S2). chips and analyze immune cell and immune pathway expression Inspection of diverse marker signatures revealed several genes profiles in pediatric and adult high-grade gliomas according to age (n ¼ 22/144) that were not immune cell specific. Some of these as well as mutational and transcriptional subgroups. We identi- genes are involved in neural development or glioma biology [e.g., fied nine studies with appropriate gene-expression data sets KLRC3 (42), EOMES (43), GREM1 (44); Supplementary Table available on GEO (Fig. 1A; refs. 21, 30, 31, 35–40). After exclusion S2]. Furthermore, many signatures included genes with low or of doublets and other tumor entities not representing gliomas, negative correlation in our data set, which is in contradiction with overall 1,135 cases were included in this study (Supplementary the hypothesis that these genes are specifically expressed by a Table S1). To compare gene-expression profiles from different unique type of cell in high-grade glioma (Fig. 2, gray dots; ref. 33). centers as well as platforms, we corrected for batch effects using As genes with different coexpression patterns are unlikely to be fRMA preprocessing and additional normalization (Supplemen- specifically expressed by a single-cell population, only those genes tary Fig. S1A; Materials and Methods). belonging to the largest connected component were retained as Next, the combined data set was characterized in more detail to final markers for the following analyses (Fig. 2, orange dots). allow correlative analyses between different glioma subgroups. However, transcriptional subtype information was available only Immune cell gene-expression profiles define immunologic in 495 samples [not available (NA): n ¼ 640], and the mutational clusters status was available only in 663 samples (NA: n ¼ 472). We After optimization of the gene-expression signatures of infil- therefore used existing gene-expression data sets with known trating immune cell populations, we analyzed the composition of mutational and transcriptional subtype to train a machine- immune and stromal cell gene-expression clusters in correlation learning algorithm, which was then able to predict the mutational to age, transcriptional subgroup, and mutational subgroup and transcriptional subtype of gene-expression data sets with (Fig. 3A). Unsupervised hierarchical clustering identified four

1404 Cancer Immunol Res; 7(9) September 2019 Cancer Immunology Research

Immune Proﬁling of Pediatric and Adult Gliomas

T cells CD8 T cells Cytotoxic lymphocytes NK cells B lineage IL21R MAL CD160 CD8B GZMH KIR2DL1 MS4A1 FCRL2 CTSW CR2 CD5 CD3E KLRB1 CTLA4 PRF1 TNFRSF17 CD19 KIR2DL3 KIR3DL2 PNOC KLRD1 NCR1 CD6 FLT3LG CD79A CD3G GZMA KIR3DL1 SPIB SIRPG NKG7 TCL1A IGKC CD3D TNFRSF25 GZMB KIR2DL4 ICOS BLK SH2D1A TRAT1 CD8A XCL1 CD22 CD28 GNLY PTGDR BANK1

Monocytic lineage Myeloid dendritic cells Neutrophils Endothelial cells Stroma/fibroblast-like MEGF9 CSF1R CD84 CD1B COL6A1 SIGLEC5 FCGR3B PLXNA2 FPR3 CD1A CXorf36 PAMR1 ADAP2 SLC25A37 CD163 CXCR2 ACVRL1 MS4A4A CD1E HAL ROBO4 TFEC VWF COL3A1 CEACAM3 MYCT1 MMRN2 PGF COL1A1 HDC CD68 CXCR1 MMRN1 FCAR VNN3 TIE1 COL6A2 CLEC10A FAM124B EMCN PLA2G7 STEAP4 MS4A2 SDPR CLIC2 TAGLN CPA3 TNFRSF10C CA4 PALMD DCN TPSAB1 TECPR2 EDN1

Included in signature Excluded

Figure 2. Optimization of glioma-speciﬁc immune cell subset signature. Due to the interference from immune signatures published by Becht and colleagues and Danaher and colleagues (32, 33) with certain genes expressed by tumor cells themselves, we performed an internal correlation analysis of the presumable immune cell gene sets within our compiled glioma cohort. The correlation threshold was set at R ¼ 0.4. Genes not correlating with each other (gray dots) were excluded from the analysis. In addition, 15 genes from the published gene signatures with known glial/neuronal expression were excluded in advance and are not shown. Twenty-ﬁve additional genes published for immune cell analysis were not included due to their unavailability on the microarray chip (Supplementary Table S2).

major clusters, which segregated according to their cellular sig- most prevalent in WT glioma and in tumors with mesenchymal nature into four clusters: vascular (immune 1), monocytic and signatures and were significantly less expressed in histone H3– stromal (immune 2), monocytic and T-cell dominant (immune mutated (K27 and G34) tumors (P ¼ 0.0022; Fig. 3B, first and third 3), and an immune cluster with a gene-expression profile indic- rows). A similar pattern was observed for cytotoxic T-lymphocyte ative of infiltration by antigen-presenting cells (APC), NK cells, (CTL) genes.B lineagecell gene expression was the strongestin G34- and T cells (immune 4; Fig. 3A). Hierarchical clustering revealed mutated tumors but was not influenced by the transcriptional that certain immune clusters correlated with the transcriptional subtype. Monocytic lineage clusters, which include microglia and subtype. Immune 1 primarily correlated with the classic transcrip- tumor-infiltrating macrophages, were underrepresented in G34- tional subtype, whereas immune 2 and 3 were dominated by a mutated tumors (P ¼ 2.5e 7) but overrepresented in gliomas with a mesenchymal signature. Immune 4 demonstrated a mixed picture mesenchymal signature (P ¼ 7.9e 92). Signatures for myeloid with no association to a transcriptional subgroup. Furthermore, we dendritic cells (mDC), neutrophils, and NK cells, on the other observed an increase of immune cluster 4 in G34-mutated tumors, hand, were expressed in G34-mutated and IDH-mutated tumors, as well as the expected increase of immune cluster 2 in mesenchy- although no correlations to a transcriptional subgroup were found. mal tumors (Supplementary Fig. S2A and S2B). There was no The endothelial component was increased in IDH-mutated and WT obvious correlation of the immune clusters with age (Fig. 3A). tumors (P ¼ 0.0018) with a classic signature (P ¼ 2.7e 5), indi- Next, we analyzed the prevalence of each immune/stromal cell cating a higher vascularization in these tumors compared with population, as defined by their abovementioned gene signature histone H3–mutated gliomas. Some of the strongest differences (Supplementary Table S3), individually according to mutational were observed with the stroma/fibroblast-like gene signature, subgroup, age, transcriptional subgroup, and histologic diagnosis which was reduced in IDH-mutated tumors but expressed in WT (Fig. 3B). Age did not have a significant influence on the gene tumors (P ¼ 5.2e 20) and tumors with a mesenchymal signature (P expression of classic immune cell subsets (Fig. 3B, second row). Age ¼ 1.1e 66; Fig. 3B, right column). Taken together, these data show did, however, affect the stromal component, as tumors of pediatric an impact of the mutational and transcriptional subgroups on the and young patients demonstrated significantly lower gene- immune cell composition, demonstrating that the tumor biology, expression profiles of fibroblast-like/stroma (P ¼ 1.9e 5)andin driven by the transcriptional and mutational phenotypes rather part endothelial signatures (P ¼ 0.0016). Low expression of the than the age of the immune system or the age at tumor occurrence, fibroblast-like signature was most frequent in IDH-mutated tumors shapes the immune landscape. (Fig. 3B). Association of stromal expression with age was therefore most likely due to the age distribution of IDH-mutated tumors, t-SNE analysis reveals immune microenvironment profiles which were most prevalent in young adults (Supplementary Fig. Next, we analyzed the immune microenvironment in S1) and not to intrinsically age-related mechanisms, as the stromal more detail by investigating both gene signatures for immune expression signature was not observed in WT gliomas. The stromal cell populations and functional pathways involved in immune component was associated with monocytic cells (Supplementary activation or inhibition, which have been defined by glioma Fig. S2C). This association was primarily seen in the immune 2 or pancancer studies by Thorsson and colleagues and Doucette cluster, which was increased in tumors with a mesenchymal tran- and colleagues (24, 34). First, we performed a t-SNE analysis scriptional subtype (Supplementary Fig. S2). T-cell signatures were using individual gene-expression signatures for immune cell

www.aacrjournals.org Cancer Immunol Res; 7(9) September 2019 1405

Bockmayr et al.

Monocytic lineage Stroma/fibroblast-like Endothelial cells Myeloid dendritic cells B lineage NK cells Neutrophils CD8 T cells T cells Cytotoxic lymphocytes

G34 IDH K27 WT N/A 0−19 20−39 40−59 60−89 CL MN NE PN N/A AA DIPG GBM

Immune 1 Immune 2 Immune 3 Immune 4 Immune cluster (Vascular) (Monocytic/stromal) (Monocytic/T cell) (APC/NK-/T cell) -3 0 3 z-scores

B T cells CTLs B lineage Monocytic Myeloid dendritic Neutrophils NK cells Endothelial Stroma/ lineage cells (mDCs) cells fibroblast-like Mutation 3 0.2

0.2 K27 G34 1 IDH 0.0 WT −1 Mutation −0.5 0.5 −0.4 0.0 0.4 −0.3 0.0 −0.2 −3 −0.3 0.0 0.2 0.4 −2 0 1 2 −0.4 0.0 0.4 −0.3 0.0 0.2 P = 0.0022 P = 0.0023 P = 0.0023 P = 2.5e−07 P = 0.00074 P = 0.00074 P = 0.0018 P = 0.0018 P = 5.2e−20

Age (years) 2 0.4

0.2 0−19 1 20−39 0

0.0 40−59 Age 60−89 −0.2 0.2 −0.5 0.5 1.0 −0.4 −0.3 0.0 0.2 −3 −1 1 3 −0.2 0.0 0.2 −2 −0.4 0.0 0.4 0.8 −0.3 0.0 P = 0.16 P = 0.068 P = 0.25 P = 0.18 P = 0.092 P = 0.79 P = 0.54 P = 0.0016 P = 1.9e−05

Transcriptional 2 1.0 subtype 0.5

0.2 CL 0.5 MN NE subtype 0.0 PN −1 0 1 −2 0 2 4 Transcriptional −0.2 −0.5 −0.2 0.2 −0.2 0.0 0.2 −0.2 0.0 0.2 −0.4 0.0 0.4 P = 7.9e−19 P = 3.4e−47 P = 0.0036 P = 7.9e−92 P = 0.00084 P = 0.0036 P = 0.46 P = 2.7e−05 P = 1.1e−66

Diagnosis

0.2 AA 13 DIPG GBM 0.0 0.4 −1 Diagnosis −0.5 0.5 −0.2 0.0 0.2 −3 −0.2 0.0 0.2 −2 0 1 2 −0.3 0.0 0.2 −0.4 −0.4 0.0 0.4 −0.3 0.0 P = 0.0063 P = 0.062 P = 0.043 P = 0.023 P = 0.01 P = 0.15 P = 0.99 P = 0.00021 P = 1.1e−14

Figure 3. Immune cell signatures stratiﬁed by age, histologic diagnosis, and the transcriptional as well as the mutational subtype. A, Heat map representation of an unsupervised hierarchical clustering of 10 immune cell subsets based on their gene-expression signature (see above). Four distinct immune clusters can be identiﬁed. B, Gene expression of individual immune cell populations is illustrated according to mutational subtype, age, and transcriptional subtype. P values are calculated by the Kruskal–Wallis test. CL, classic transcriptional subgroup; GBM, glioblastoma; G34, G34-mutated glioma; IDH, IDH-mutated glioma; K27, K27- mutated glioma; MN, mesenchymal transcriptional subgroup; N/A, not available; NE, neural transcriptional subgroup; PN, proneural transcriptional subgroup; WT, IDH wild-type glioma.

populations that have been optimized for glioma samples tumors (Fig. 4A and B). The progression-free survival (PFS) for (Fig. 2). We included published signatures for pathways involved clusters 1 to 6 (P ¼ 0.0043) and tumor samples that have been in immune activation/inhibition (Supplementary Table S3; collected after previous treatment, i.e., chemotherapy or radio- Fig. 4A). K-means clustering for all samples (n ¼ 1,135) was therapy, are shown in Supplementary Fig. S3A to S3D. No undertaken to identify microenvironmental subtypes. A Kaplan– signiﬁcant impact of previous treatment on the immune gene- Meier plot of the OS within these clustered subtypes is shown expression landscape, except for a decrease of endothelial cell in Fig. 4B. Here, the different clusters show signiﬁcant differences signatures on pretreated samples (P ¼ 2.3e 5), was observed in (P ¼ 2.1e 10), which may, in part, be explained by the fact that this analysis. clusters 1 (green) and 6 (gray) contained mostly IDH-mutated In analogy to the data presented in Fig. 4A, t-SNE plots were tumors, which have a better prognosis than K27, G34, or WT drawn for the transcriptional and mutational subgroups, as well

1406 Cancer Immunol Res; 7(9) September 2019 Cancer Immunology Research

Immune Proﬁling of Pediatric and Adult Gliomas

ABK-means clustering Overall survival C Transcriptional subtype D Mutation E Age (years) F Diagnosis

Cl. 1 Cl. 4 Cl. 2 Cl. 5 Cl. 3 Cl. 6

.75 P = 2.1*e-10

.25

Cl. 1 Cl. 2 Cl. 3 Cl. 4 Cl. 5 Cl. 6 0 12 24 36 48 60 CL MN NE PN K27 G34 IDH WT 0−19 20−39 40−59 60−89 DIPG AA GBM Time (months)

GHT cells CTL CD8+ T cells NK cells Coinhibitory Costimulatory IImmunosuppressors Immuneactivators JGlioma antigen

Proinflammatory Immunosuppressive ARG1 + IDO T regs Monocytic lineage Myeloid markers Stroma/fibroblast-like Stimulatory ligand Stimulatory receptor cytokines signaling

B cell lineage mDCs Neutrophils Endothelial cells Inhibitory ligand Inhibitory receptor TAM + skewing Antigen presentation Others

−4 −2 0 2 4 Z−scores

Immune cell populations Thorsson et al. Doucette et al.

Functional immune signatures

Figure 4. t-SNE analysis of immune cell signatures and functional immune pathways. A, K-means clustering of all gene-expression proﬁles results in six distinct immunologic subtypes. B, OS demonstrates the impact of IDH-mutated tumors on clusters 1 and 6. C–F, t-SNE analysis according to transcriptional subgroup (C), mutational subgroup (D), age (E), and histologic diagnosis (F). G–J, Individual plots demonstrate distribution of immune cell subsets (G), functional pathways published by Thorsson and colleagues (H) and Doucette and colleagues (I), and other pathways (J). The red dotted circle highlights the prevalent expression of inhibitory receptors in cluster 6. Individual gene signatures for the different immune cell subsets and pathways are listed in Supplementary Table S3, respectively. Certain signatures published by Doucette and colleagues that are not associated with an immunomodulatory function (see Supplementary Fig. S2) are shown in G and J. TAM, tumor-associated macrophages.

as for age and histologic diagnosis (Fig. 4C–F). Gene-expression has a stronger impact on the immune microenvironment than the profiles for the t-SNE plotted samples of the individual signatures mutational IDH status. of immune cell populations and functional pathways from the Functional immune inhibition or activation, described by pancancer (34) and glioma analysis (24) are shown in Fig. 4G–I. pathways of stimulatory and inhibitory receptor/ligand inter- The gene expression of the individual signatures is illustrated in action, was primarily associated with clusters 4 and 2. Although Supplementary Fig. S4A and S4B. Immune gene t-SNE clusters the stimulatory receptor/ligand pathways were both expressed correlated best with the transcriptional subgroup. Here, the mes- inthesametumorsamples,theinhibitoryreceptor/ligandaxis enchymal and classic subtypes demonstrated the strongest expres- was discordantly expressed (Fig. 4H, middle and bottom). sion of immune genes, contributing to polarization of the samples. Inhibitory ligands, including TGFB1, VEGFA/B, and IL10, Cluster 4 was mainly defined by genes indicating an infiltration of among others, were predominantly found in tumors with a regulatory T cells (Treg), fibroblast-like cells, and cells of monocytic mesenchymal transcriptional subgroup and extended beyond and myeloid lineage (Fig. 4G, second row). A strong signature of the monocytic and regulatory T-cell gene environment. Sam- þ CTL or the corresponding CD8 T-cell markers was detected only ples with strong inhibitory ligand gene expression were espe- in a minority of these samples, indicating that a stroma-rich cially low in CTL signatures. Although the expression of inhib- environment correlates with strong Treg and monocytic infiltration itory receptors, such as PDCD1, CTLA4, LAG3, and KIR2DL1/3, þ and precludes infiltration of potentially tumor-specific cytotoxic can usually be attributed to CD3 T cells, their expression did þ CD8 T cells. No clear polarization to any cluster was found for not correlate with tumors rich in T-cell gene-expression clusters other gene signatures, such as mDCs, neutrophils, B cells, or and was mainly associated with tumors with low immune cell endothelial cells, within the combined analysis of the microenvi- signatures and low stimulatory receptor/ligand pathways ronment. Although almost all IDH-mutated gliomas are known to (Fig. 4H). The inhibitory receptor expression represented one express a proneural transcriptional subtype and therefore have a of the most distinct clusters (Fig. 4H; red circle, cluster 6), rather good prognosis, the proneural type is associated with poor distant from immunosuppressor and immunosuppressive sig- prognosis in WT glioblastomas (45). When stratifying by their naling signatures defined by Doucette and colleagues (Fig. 4I) immunostromal gene signatures, these two biologically different or other glioma antigen and ARG1 þ IDO signatures (Fig. 4J; types group together, indicating that the transcriptional phenotype ref. 24). Solely the coinhibitory pathways, which include

www.aacrjournals.org Cancer Immunol Res; 7(9) September 2019 1407

Bockmayr et al.

PDCD1LG2 and SLAMF7, and glioma antigen–rich samples samples with available OS information (n ¼ 1,057) by diagnosis, colocalized to the inhibitory receptor area. These samples, age, transcriptional subgroup, and mutational subgroup and which express high amounts of glioma antigen and harbor depicted the hazard ratio of individual gene signatures as heat cells with strong expression of PDCD1 and CTLA4,might map representations (Fig. 5A; Supplementary Table S4A). Across identify tumors amenable to checkpoint inhibition treatment. all samples, functional immune pathways, including immuno- Moreover, within this region, most IDH-mutated tumors can be suppressors, immunoactivators, immunosuppressive cytokines, found. Although they do not express large amounts of antigen, and checkpoints or immune effectors, among others, are all the tumors seem to depend on the inhibitory receptor pathway. associated with worse OS. These effects are pronounced in tumors with glioblastoma and AA histology and in patients >40 years of Tumor-speciﬁc immune response affects survival associated age. The negative impact of functional immune pathways on the with age and subgroup OS is strongest in tumors with proneural transcriptional pheno- Finally, we wanted to determine the impact of individual type, which also include IDH-mutated glioma, whereas the OS of immune cell population signatures and functional immune path- mesenchymal tumors is independent of their immune microen- ways on the OS of high-grade gliomas. First, we stratiﬁed all tumor vironment. This also holds true for WT tumors, in which no

A Subgroup-specific hazard ratios Diagnosis Age Transcriptional subtype Mutation *** *** *** * ** * *** * * Immunosuppressors ** Immuneactivators *** *** * *** * Immunosuppressive cytokines and checkpoints ** * * * Immune effector *** ** *** ** Immunosuppressive signaling *** ** *** ** TAM + skewing *** * ** * *** Fibroblasts Stimulatory ligand *** * *** ** *** * * Inhibitory ligand ** ** * * ** * Glioma antigen * * Antigen presentation T-cell effector Innate immune * * * Monocytic lineage ** ** * ** Myeloid markers Tregs Endothelial cells * Coinhibitory Proinf cytokines Costimulatory Stimulatory receptor ** * CD8 T cells * Inhibitory metabolism ** * ** * * Myeloid dendritic cells ** ** * * B lineage Neutrophils Cytotoxic lymphocytes ** * *** * T cells ** ** NK cells ** ** * *** * Inhibitory receptor Survival = 113) = 136) = 413) = 145) = 37) = 13) n = 218) = 281) n = 113) = 769) n = 1,057) Better Worse n AA (n = 77) ral ( K27 ( n All ( n G34 ( IDH (n WT (n DIPG ( = GBM26) (n = 954) 1/2 1 2 0−19 y (n hymal (n = 347) Neural (n 20−39 y (n40 −59 y (60−89 y (n = 395) Classic (n Proneu Hazard ratio Mesenc BCOS by immune cluster OS by immune checkpoint molecules Immune 1 (vascular) P = 0.0063 Endothelial cells BTLA

Immune 2 (monocytic/stromal) * HAVCR2 Monocytic lineage Stroma/fibroblast-like * * TGFB1 Endothelial cells LAG3 T cells/CTL * ** * ** * *

Immune 3 (monocytic/T cell) * ** ** CD274 Monocytic lineage Overall survival TIGIT T cells Immune 4 (APC/NK-/T cell) * * * * CTLA4 Myeloid DCs B cells * * * * PDCD1

0.0 0.2 0.4 0.6 0.8 1.0 NK cells ral AA IDH WT Hazard ratio 0 102030405060T cells/CTL GBM ymal K27 G34 DIPG 0−19 y h 8 20−3940−59 y 60−89 y y ClassicNeu

Time (months) All tumors Proneural Count 04 1/2 1 2 Mesenc Value

Figure 5. Subgroup-specific correlation of OS with gene-expression profiles of immunologic subsets and pathways. A, Subgroup-specific hazard ratios according to histologic diagnosis, age, transcriptional subtype, and mutational subtype. Individual gene signatures for the different immune cell subsets and pathways are listed in Supplementary Table S4. B, OS stratified by the four immune clusters identified in Fig. 3. P value calculated by log-rank test. C, Correlation of OS with the expression of immune-checkpoint genes according to the known subgroups (Supplementary Table S4B). P values: , P < 0.05; , significant after BH correction for the tested signatures; , significant after Bonferroni correction for all hypotheses. GBM, glioblastoma; G34, G34-mutated glioma; IDH, IDH-mutated glioma; K27, K27-mutated glioma; Proinf, proinflammatory; TAM, tumor-associated macrophages; WT, IDH wild-type glioma; y, years.

1408 Cancer Immunol Res; 7(9) September 2019 Cancer Immunology Research

Immune Proﬁling of Pediatric and Adult Gliomas

association with worse survival and negative immune regulators TCGA gene-expression data set of 544 patients, previous work was found. Improved survival is observed primarily in association by Doucette and colleagues highlighted that the mesenchymal with gene-expression profiles of cytotoxic (T and NK cells) and subgroup displays a gene-expression profile reflective of a APCs (mDCs and B cells) or inhibitory receptors (PDCD1, CTLA4, proinflammatory antitumor response (24). Our cohort of and LAG3; Fig. 5A). These data show that the immune microen- 1,135 gene-expression profiles further enabled stratification of vironment in tumors with mesenchymal transcriptional subgroup the mesenchymal subgroup into two immunologically distinct or patients at young age (<19 years) does not appear to affect OS, infiltration clusters, as the prevalence of monocytic markers togeth- potentially indicating that these tumors are characterized by an er with a stromal and endothelial composition (immune 2) could immunosuppressed environment. be distinguished from a monocytic lineage profile with reduced To identify the influence of immune cell signatures on OS stromal component but increased cytotoxic T-cell signature (Fig. 5B) as well as PFS (Supplementary Fig. S3D) and to deter- (immune 3), respectively. This observation suggests that the mes- mine a marker profile that could be used in clinical diagnostics to enchymal subgroup may harbor a distinct immunologic subtype classify the different immune microenvironments, we defined with increased gene expression indicative for higher CTL infiltra- four immune clusters (Fig. 3A). When we plotted the OS of each tion, which might be amenable to immune-checkpoint inhibition. immune cluster, we were able to confirm that immune cluster 2 Work published by Luoto and colleagues analyzed the (monocytic/stromal) was associated with a significantly worse immune cell composition of 156 glioblastoma RNA sequenc- prognosis (Fig. 5B). Other immune clusters with an immune ing (RNA-seq) samples by TCGA to investigate the impact microenvironment rich in APC and cytolytic cells (NK and T cells) of genomic alterations on the immune microenvironment in present the best OS. IDH WT and IDH-mutated glioblastoma (22). That study Because of continuing study of immune-checkpoint inhibitors identifies three immune response groups, a "negative" macro- in clinical trials, we analyzed the association of major immune phage, granulocyte, and macrophage-dominated group in checkpoints with the different high-grade glioma groups (Fig. 5C; IDH1-mutated tumors, a "humoral," lymphocyte-enriched sig- Supplementary Table S4B). Here, CD274 (PD-L1), CTLA4, nature in EGFR-gained samples, and a "cellular-like," macro- PDCD1, and LAG3 demonstrated significant correlations. As phage-infiltrated subgroup in EGFR-amplified tumors (22). We expected, PD-L1, a marker that is associated with immune escape, confirm the observation of a lower monocyte/macrophage was associated with a worsened survival (hazard ratio >1), where- signature component in IDH-mutated tumors. However, we as the other markers (LAG3 and CTLA4) correlated with better did not see an IDH-associated immune phenotype, as in our survival. PD-L1 was informative in tumors with AA histology and cohort, K27- and G34-mutated gliomas correlated more with with K27 mutation. Some checkpoints appeared to have a con- an immune cluster 4 phenotype (immune 4: APC/NK cell/T trary impact on OS, as TGFB1 expression resulted in better cell). We showed that IDH WT and mutated tumors displayed prognosis in tumors with classic transcriptional phenotype and the most distinct differences in the stromal/fibroblast-like cell a worse survival in patients under the age of 20 years. Taken signature. Yet, only IDH-mutated tumors seemed to leverage together, these data demonstrate that immunologic phenotype certain functional pathways, as they expressed "inhibitory refines response of high-grade gliomas to immunologic treatment receptors" such as PD-1, CTLA4, LAG3, or KIRs, which concurs approaches. with the finding of a "negative regulation of lymphocyte function" in IDH tumors by Luoto and colleagues (22). In a global pancancer analysis, glioblastoma was classified Discussion as a lymphocyte-depleted, macrophage-enriched tumor entity Molecular stratification of high-grade gliomas is paving the way with a suppressed TH1 axis (34). Although the monocytic macro- to a better understanding of tumor heterogeneity. Integrated molec- phage lineage signature was also evident in our analysis, we ular analyses allow stratification into different subtypes and pre- describe a subset of glioblastoma with low expression of monocytic diction of clinical response to therapy (13). Analyses in adult lineage gene cluster, as the vascular (immune 1: vascular) and APC/ glioblastoma patients suggest that genomically diverse subtypes NK cell/T-cell (immune 4) clusters were depleted of monocyte/ display distinct microenvironmental profiles (20, 22, 24). It is macrophage signatures. The K27 and G34 tumors do not seem to therefore hypothesized that genetic driver mutations promote follow the glioblastoma pattern described by Thorsson and collea- unique immunologic landscapes (20, 23). Given the diverse geno- gues (34). We show that K27 and G34 tumors displayed lower gene- mic alterations between pediatric and adult high-grade glioma, i.e., expression signatures of monocytic lineage cells and harbored K27 and G34 mutations, as well as the potential clinical implica- lower cytotoxic and overall T-cell gene expression compared with tions for the translational use of therapeutic regimens, including IDH-mutated and WT gliomas. The depletion of lymphocytic clus- chemotherapy and immune-checkpoint inhibition, we compared ters in the K27 and G34 tumors is comparable with the "immuno- the immunologic microenvironment of high-grade gliomas across logically quiet" immune subtype described for low-grade gliomas in age groups and stratified by clinically relevant genetic subgroups. the pancancer analysis (34). However, the increased B cell gene- In our study, we demonstrate that different mutational and expression prevalence in G34-mutated tumors in our study points to þ transcriptional subgroups shape the composition of the immune a primary signature indicative for CD4 TH2cellinfiltration, which subset signatures, independent of age or major histologic diag- therefore deviates from the observed dominance of a TH1 response noses. We identified four distinct immunologic patterns of in low-grade glioma by Thorsson and colleagues (34). immune cell profiles, which are defined by either vascular The abovementioned studies demonstrate that the molecular (immune 1), monocytic/stromal (immune 2), monocytic/T-cell and transcriptional heterogeneity in malignant gliomas cor- (immune 3), and APC/NK cell/T-cell (immune 4)–dominant relates with high variability of the immunologic landscape immune clusters. These clusters correlated with the different (20, 22, 24, 34). These studies emphasize the need for further transcriptional subgroups of high-grade gliomas. Based on a stratification to individualize upcoming clinical studies. Our data

www.aacrjournals.org Cancer Immunol Res; 7(9) September 2019 1409

Bockmayr et al.

demonstrate that age and in part histologic diagnosis are not proneural subgroup, on the other hand, which includes IDH-mu- relevant for the immune subset composition. However, tated tumors, displays expression on "inhibitory immune recep- immune cell signatures and functional immune pathways tors" that correlates with OS. K27 and G34 gliomas exhibit diverse are associated with the transcriptional subgroups (13, 16). immune pathway expression, with K27 tumors especially Especially the mesenchymal subgroup demonstrated immune deprived of lymphocyte, NK cell, and antigen-presenting mDC activity and association with monocytic lineage profiles and signatures and presumably other mechanisms of immune escape, immune clusters 2 and 3. The strongest correlation of immune- such as the PD-L1 pathway. Taken together, our study demon- relevant cell signatures and functional pathways was, however, strates that future immunotherapeutic strategies should take into observed in the proneural subgroup. Here, immunosuppressive account the distinct immunologic landscapes that may modulate and tumor-associated macrophage–related pathways, as well as therapeutic efficacy, especially for rare mutational subgroups of stromal/fibroblast-like components correlated with a decreased high-grade gliomas. OS. Although the mesenchymal subgroup displayed an immunologic component, almost no correlation with OS was present. In Disclosure of Potential Conflicts of Interest addition, immune gene expression did not correlate with either M. Westphal is a consultant/advisory board member for AbbVie. No poten- survival in IDH WT gliomas or with patient age below 20 years, tial conflicts of interest were disclosed by the other authors. implying an immunologic "inertness" in these groups. We initially hypothesized that high-grade gliomas in children Authors' Contributions and young adults might demonstrate distinct immunologic pro- € fi Conception and design: M. Bockmayr, C.L. Maire, U. Schuller, M. Mohme les due to a still developing or young immune system. Except for Development of methodology: M. Bockmayr, F. Klauschen, M. Mohme the above-described lack of correlation of immunologic gene Acquisition of data (provided animals, acquired and managed patients, expressions with OS, we did not find any association with provided facilities, etc.): M. Bockmayr, C.L. Maire, S. Rutkowski, a specific immune cell pattern. However, when dissecting high- K. Lamszus, M. Mohme grade gliomas of the young and adolescent patients into their Analysis and interpretation of data (e.g., statistical analysis, biostatistics, molecular subgroups, i.e., K27- and G34-mutated tumors, oppos- computational analysis): M. Bockmayr, M. Westphal, K. Lamszus, M. Mohme Writing, review, and/or revision of the manuscript: M. Bockmayr, ing effects of immunologic genes on OS became apparent. In the F. Klauschen, C.L. Maire, S. Rutkowski, K. Lamszus, U. Schuller,€ M. Mohme K27-mutated tumors, CD274 (PD-L1) and CTLA4 correlated Administrative, technical, or material support (i.e., reporting or organizing with a worse prognosis, whereas the opposite was true for data, constructing databases): M. Bockmayr, M. Mohme G34-mutated tumors. Furthermore, G34 tumors appear to rely Study supervision: F. Klauschen, K. Lamszus, M. Mohme primarily on TGFB1 and HAVCR2 as pathways for immune escape. This finding implicates that checkpoint inhibition of the Acknowledgments PD-1/PD-L1 pathway might be beneficial for patients with diffuse This study was supported by the Werner-Otto-Stiftung (to U. Schuller)€ and midline gliomas harboring histone H3 K27M mutations. the Fordergemeinschaft€ Kinderkrebszentrum Hamburg (to M. Bockmayr and U.- Our study shows that immunologic heterogeneity in high- Schuller),€ the Anni-Hofmann Stiftung (to K. Lamszus), and the Else Kroner-€ grade gliomas is associated with distinct transcriptional and Fresenius Stiftung (to M. Mohme). mutational subgroups, independent of patient age. The immune The costs of publication of this article were defrayed in part by the payment of cell signatures correlate with the transcriptional subgroup and can page charges. This article must therefore be hereby marked advertisement in be clustered into four immunologic groups. The mesenchymal accordance with 18 U.S.C. Section 1734 solely to indicate this fact. subgroup shows an immunologic activation pattern, which does not correlate with OS, reflecting the inert immunobiology of this Received January 5, 2019; revised March 21, 2019; accepted June 27, 2019; transcriptional subtype as well as IDH WT glioblastoma. The published first July 2, 2019.

References 1. Westphal M, Lamszus K. The neurobiology of gliomas: from cell biology to a gene-expression-based retrospective study. PLoS Med 2016;13: the development of therapeutic approaches. Nat Rev Neurosci 2011;12: e1002194. 495–508. 8. Gentles AJ, Newman AM, Liu CL, Bratman S V, Feng W, Kim D, et al. The 2. Reardon DA, Omuro A, Brandes AA, Rieger J, Wick A, Sepulveda J, et al. prognostic landscape of genes and infiltrating immune cells across human OS10.3 Randomized phase 3 study evaluating the efficacy and safety of cancers. Nat Med 2015;21:938–45. nivolumab vs bevacizumab in patients with recurrent glioblastoma: Check- 9. Buerki RA, Chheda ZS, Okada H. Immunotherapy of primary brain tumors: Mate 143. Neuro Oncol 2017;19(Suppl 3):iii21. facts and hopes. Clin Cancer Res 2018;24:5198–205. 3. Filley AC, Henriquez M, Dey M. Recurrent glioma clinical trial, CheckMate- 10. Mohme M, Neidert MC, Regli L, Weller M, Martin R. Immunological 143: the game is not over yet. Oncotarget 2017;8:91779–94. challenges for peptide-based immunotherapy in glioblastoma. Cancer 4. Lim M, Xia Y, Bettegowda C, Weller M. Current state of immunotherapy for Treat Rev 2014;248–58. glioblastoma. Nat Rev Clin Oncol 2018;15:422–42. 11. Louis DN, Perry A, Reifenberger G, von Deimling A, Figarella-Branger D, 5. Rabinovich GA, Gabrilovich D, Sotomayor EM. Immunosuppressive strat- Cavenee WK, et al. The 2016 World Health Organization classification of egies that are mediated by tumor cells. Annu Rev Immunol 2007;25: tumors of the central nervous system: a summary. Acta Neuropathol 2016; 267–96. 131:803–20. 6. Becht E, de Reynies A, Giraldo NA, Pilati C, Buttard B, Lacroix L, et al. 12. Yan H, Parsons DW, Jin G, McLendon R, Rasheed BA, Yuan W, et al. IDH1 Immune and stromal classification of colorectal cancer is associated with and IDH2 mutations in gliomas. N Engl J Med 2009;360:765–73. molecular subtypes and relevant for precision immunotherapy. Clin 13. Verhaak RGW, Hoadley KA, Purdom E, Wang V, Qi Y, Wilkerson MD, et al. Cancer Res 2016;22:4057–66. Integrated genomic analysis identifies clinically relevant subtypes of glio- 7.AliHR,ChlonL,PharoahPDP,MarkowetzF,CaldasC.Patternsof blastoma characterized by abnormalities in PDGFRA, IDH1, EGFR, and immune infiltration in breast cancer and their clinical implications: NF1. Cancer Cell 2010;17:98–110.

1410 Cancer Immunol Res; 7(9) September 2019 Cancer Immunology Research

Immune Proﬁling of Pediatric and Adult Gliomas

14. Phillips HS, Kharbanda S, Chen R, Forrest WF, Soriano RH, Wu TD, et al. 30. Gravendeel LAM, Kouwenhoven MCM, Gevaert O, de Rooi JJ, Stubbs AP, Molecular subclasses of high-grade glioma predict prognosis, delineate a Duijm JE, et al. Intrinsic gene expression profiles of gliomas are a better pattern of disease progression, and resemble stages in neurogenesis. predictor of survival than histology. Cancer Res 2009;69:9065–72. Cancer Cell 2006;9:157–73. 31. Ceccarelli M, Barthel FP, Malta TM, Sabedot TS, Salama SR, Murray BA, et al. 15. Noushmehr H, Weisenberger DJ, Diefes K, Phillips HS, Pujara K, Berman Molecular profiling reveals biologically discrete subsets and pathways of BP, et al. Identification of a CpG island methylator phenotype that defines a progression in diffuse glioma. Cell 2016;164:550–63. distinct subgroup of glioma. Cancer Cell 2010;17:510–22. 32. Becht E, Giraldo NA, Lacroix L, Buttard B, Elarouci N, Petitprez F, et al. 16. Wang Q, Hu B, Hu X, Kim H, Squatrito M, Scarpace L, et al. Tumor Estimating the population abundance of tissue-infiltrating immune and evolution of glioma-intrinsic gene expression subtypes associates with stromal cell populations using gene expression. Genome Biol 2016;17: immunological changes in the microenvironment. Cancer Cell 2017;32: 218. 42–56. 33. Danaher P, Warren S, Dennis L, D'Amico L, White A, Disis ML, et al. Gene 17. Schwartzentruber J, Korshunov A, Liu X-Y, Jones DTW, Pfaff E, Jacob K, et al. expression markers of tumor infiltrating leukocytes. J Immunother Cancer Driver mutations in histone H3.3 and chromatin remodelling genes in 2017;5:18. paediatric glioblastoma. Nature 2012;482:226–31. 34. Thorsson V, Gibbs DL, Brown SD, Wolf D, Bortone DS, Ou Yang T-H, et al. 18. Korshunov A, Capper D, Reuss D, Schrimpf D, Ryzhova M, Hovestadt V, The immune landscape of cancer. Immunity 2018;48:812–30. et al. Histologically distinct neuroepithelial tumors with histone 3 G34 35. Griesinger AM, Birks DK, Donson AM, Amani V, Hoffman LM, Waziri A, mutation are molecularly similar and comprise a single nosologic entity. et al. Characterization of distinct immunophenotypes across pediatric Acta Neuropathol 2016;131:137–46. brain tumor types. J Immunol 2013;191:4880–8. 19. Korshunov A, Schrimpf D, Ryzhova M, Sturm D, Chavez L, Hovestadt V, 36. Madhavan S, Zenklusen J-C, Kotliarov Y, Sahni H, Fine HA, Buetow K. et al. H3-/IDH-wild type pediatric glioblastoma is comprised of molecu- Rembrandt: helping personalized medicine become a reality through larly and prognostically distinct subtypes with associated oncogenic dri- integrative translational research. Mol Cancer Res 2009;7:157–67. vers. Acta Neuropathol 2017;134:507–16. 37. Lee Y, Scheck AC, Cloughesy TF, Lai A, Dong J, Farooqi HK, et al. Gene 20. Chen Z, Hambardzumyan D. Immune microenvironment in glioblastoma expression analysis of glioblastomas identifies the major molecular basis subtypes. Front Immunol 2018;9:1004. for the prognostic benefit of younger age. BMC Med Genomics 2008;1:52. 21. Sturm D, Witt H, Hovestadt V, Khuong-Quang D-A, Jones DTW, Koner- 38. Sturm D, Orr BA, Toprak UH, Hovestadt V, Jones DTW, Capper D, et al. mann C, et al. Hotspot mutations in H3F3A and IDH1 define distinct New brain tumor entities emerge from molecular classification of CNS- epigenetic and biological subgroups of glioblastoma. Cancer Cell 2012;22: PNETs. Cell 2016;164:1060–72. 425–37. 39. Paugh BS, Broniscer A, Qu C, Miller CP, Zhang J, Tatevossian RG, et al. 22. Luoto S, Hermelo I, Vuorinen EM, Hannus P, Kesseli J, Nykter M, et al. Genome-wide analyses identify recurrent amplifications of receptor tyro- Computational characterization of suppressive immune microenviron- sine kinases and cell-cycle regulatory genes in diffuse intrinsic pontine ments in glioblastoma. Cancer Res 2018;78:5574–85. glioma. J Clin Oncol 2011;29:3999–4006. 23. Herting CJ, Chen Z, Pitter KL, SzulzewskyF,KaffesI,KaluzovaM,etal. 40. Paugh BS, Qu C, Jones C, Liu Z, Adamowicz-Brice M, Zhang J, et al. Genetic driver mutations define the expression signature and microen- Integrated molecular genetic profiling of pediatric high-grade gliomas vironmental composition of high-grade gliomas. Glia 2017;65: reveals key differences with the adult disease. J Clin Oncol 2010;28:3061–8. 1914–26. 41. Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust 24. Doucette T, Rao G, Rao A, Shen L, Aldape K, Wei J, et al. Immune enumeration of cell subsets from tissue expression profiles. Nat Methods heterogeneity of glioblastoma subtypes: extrapolation from the cancer 2015;12:453–7. genome atlas. Cancer Immunol Res 2013;1:112–22. 42. Cheray M, Bessette B, Lacroix A, Melin C, Jawhari S, Pinet S, et al. KLRC3, a 25. Bockmayr M, Mohme M, Klauschen F, Winkler B, Budczies J, Rutkowski S, natural killer receptor gene, is a key factor involved in glioblastoma et al. Subgroup-specific immune and stromal microenvironment in medul- tumourigenesis and aggressiveness. J Cell Mol Med 2017;21:244–53. loblastoma. Oncoimmunology 2018;7:e1462430. 43. Arnold SJ, Huang G-J, Cheung AFP, Era T, Nishikawa S-I, Bikoff EK, et al. 26. The R Core Team. R: a language and environment for statistical computing. The T-box transcription factor Eomes/Tbr2 regulates neurogenesis in the Vienna (Austria): R Foundation for Statistical Computing; 2014. Available cortical subventricular zone. Genes Dev 2008;22:2479–84. from: www.R-project.org. 44. Yan K, Wu Q, Yan DH, Lee CH, Rahim N, Tritschler I, et al. Glioma cancer 27. Edgar R, Domrachev M, Lash AE. Gene Expression Omnibus: NCBI gene stem cells secrete Gremlin1 to promote their maintenance within the expression and hybridization array data repository. Nucleic Acids Res tumor hierarchy. Genes Dev 2014;28:1085–100. 2002;30:207–10. 45. Sandmann T, Bourgon R, Garcia J, Li C, Cloughesy T, Chinot OL, et al. 28. McCall MN, Bolstad BM, Irizarry RA. Frozen robust multiarray analysis Patients with proneural glioblastoma may derive overall survival benefit (fRMA). Biostatistics 2010;11:242–53. from the addition of bevacizumab to first-line radiotherapy and temozo- 29. Dai M. Evolving gene/transcript definitions significantly alter the interpre- lomide: retrospective analysis of the AVAglio trial. J Clin Oncol 2015;33: tation of GeneChip data. Nucleic Acids Res 2005;33:e175. 2735–44.

www.aacrjournals.org Cancer Immunol Res; 7(9) September 2019 1411

Immunologic Profiling of Mutational and Transcriptional Subgroups in Pediatric and Adult High-Grade Gliomas

Michael Bockmayr, Frederick Klauschen, Cecile L. Maire, et al.

Cancer Immunol Res 2019;7:1401-1411. Published OnlineFirst July 2, 2019.

Updated version Access the most recent version of this article at: doi:10.1158/2326-6066.CIR-18-0939

Supplementary Access the most recent supplemental material at: Material http://cancerimmunolres.aacrjournals.org/content/suppl/2019/07/02/2326-6066.CIR-18-0939.DC1

Cited articles This article cites 43 articles, 12 of which you can access for free at: http://cancerimmunolres.aacrjournals.org/content/7/9/1401.full#ref-list-1

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department Subscriptions at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://cancerimmunolres.aacrjournals.org/content/7/9/1401. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.