Introduction

Glioblastoma (GBM) is the most common primary brain tumor in adults and ranks among the deadliest human cancers. Despite advances in surgical technique and improved chemoradiotherapy [11], median survival is less than 15 months [50]. Intriguingly, a small proportion (16 %) of GBM patients receiving intensified therapy survives for more than 36 months [49]. These so-called long-term survivors (LTS) are usually younger and have a higher Karnofsky Performance Score (KPS) at the time of diagnosis [31, 58]. Recently, extending prior attempts to subclassify glioblastoma [44], The Cancer Genome Atlas (TCGA) have suggested four subtypes (proneural, neural, classical and mesenchymal) based on gene expression patterns [55]. In addition, classification based on DNA methylation has been proposed [51, 52]. However, there are only marginal differences in survival between these subtypes and imperfect overlap with key genetic aberrations. The long-term clinical impact of these classifications is thus unclear. Further studies by the TCGA have revealed that a number of proneural GBMs show genome-wide hypermethylation [42]. Tumors with this so-called glioma CpG island methylator phenotype (G-CIMP) almost uniformly carry mutations in isocitrate dehydrogenase isoenzymes (IDH1 and IDH2). Indeed, the vast majority of mutations (>95 %) occur in IDH1 and among these, R132H substitutions are by far the most common [22]. It is now recognized that remodeling of the active center of IDH enzymes leads to production of 2-hydroxyglutarate (2-HG) [57]. Elevated levels of this metabolite in turn interfere with a number of enzymes involved in epigenetic regulation, resulting in global hypermethylation [54]. Thus, a causal link between IDH mutations and G-CIMP has been established. As IDH mutations are observed in the vast majority of lower grade gliomas (LGGs) and secondary GBMs (sGBMs) [59], it has been speculated that GBMs harboring this specific lesion are in fact derived from clinically silent precursor lesions and do not constitute true primary glioblastomas (pGBMs). This is further supported by results from the TCGA consortium and others who have shown that IDH mut and G-CIMP tumors have a much better outcome and lack prototypical genetic alterations of pGBM such as EGFR amplification, CDKN2A deletions, and PTEN loss [2, 42, 52, 55]. Unsurprisingly, a large number of studies have found a higher proportion of IDH mut [21] and G-CIMP tumors [47] among LTS. These data imply that available studies on long-term survival may have been confounded by IDH mutations, and further research is warranted to elucidate the mechanisms contributing to improved survival in primary, IDH wt glioblastoma.

Materials and methods

Patient data

All patients received surgery at the Department of Neurosurgery (University Hospital Heidelberg, Germany) and were treated with post-operative chemoradiotherapy as well as six cycles of adjuvant temozolomide-based chemotherapy. To ensure completion of the therapy regimen, only patients who survived for more than 7 months after surgery were included. Diagnosis of glioblastoma was confirmed through histopathological review by independent, board-certified neuropathologists (AvD, DC, CH). In cases where IDH1 data was not available as part of routine diagnostic workup, mutations were ruled out by sequencing and immunohistochemistry as described elsewhere [5, 22]. Clinical and outcome data for patients treated at the Department of Neurosurgery (University Hospital Heidelberg, Germany) were obtained through review of patient’s charts. Detailed patient information is provided in Table S1.

Patient material, quality control, and RNA extraction

Tumor material was obtained following surgical resection. Tissue was snap-frozen and stored at −80 °C until further processing. RNA was extracted with TRIzol® RNA Isolation Reagents (Life Technologies) from samples found eligible in terms of tumor cell content (>60 %) and necrosis (<20 %). Analyte concentration and quality were determined using the Nanodrop 2000 spectrophotometer (Thermo Scientific) and Bioanalyzer 2100 (Agilent).

Microarray experiments

All specimens were assayed on the Agilent SurePrint G3 Human Exon 2 × 400 K platform. Labeling and hybridization reactions were performed according to the manufacturer’s protocols. Probe sequences were aligned to the Ensembl transcript database, filtered based on BLAST quality metrics and combined into transcript expression estimates. Median transcript expression was used as an expression estimate (n = 21,389). The resulting data were normalized using VSN [26], transformed to log2 scale and median centered. Differential expression between LTS and STS was assessed using Student’s t test with correction for unequal variances. Exploratory data analysis (Principal Component Analysis) was conducted within the R statistical software environment [53]. GeneGo’s Metacore software tool from Thomson Reuters was used for all pathway analyses.

Statistical analyses

GraphPad Prism was used for statistical analyses for patient data of the discovery cohort. All other analyses were conducted within the R software environment. Numerical data were compared using Student’s t test with correction for unequal variances. Ordinal data were analyzed using non-parametric statistics (Mann–Whitney U test). Overlap of categorical variables was assessed using the χ 2 test or Fisher’s exact test where applicable. Survival analysis was conducted using the “survival” R package. Log-rank tests and Cox’ proportional hazards model were used for univariate and multivariate comparisons, respectively.

Assessment of subtypes in discovery cohort

Centroids established by Verhaak et al. [55] for subtyping of GBM expression data were downloaded from the TCGA website (accompanying data freeze released with aforementioned publication). Pearson’s r for each of the four centroids was calculated for all samples and all available genes (n = 776). Each sample was then assigned the subtype of the centroid it was most highly correlated with (Fig. S1).

Multicolor immunostaining, image analysis and evaluation

Acetone-fixed cryostat sections (5–7 μm) were used for triple immunofluorescence staining. Microglia were detected using mouse anti-CD68 (Dako). M2 Polarization of microglia cells was evaluated by mouse anti-CD163 (AbD Serotec) and rabbit anti-CD204 (Sigma-Aldrich) stainings [45]. Primary antibodies were diluted in Antibody Diluent (Dako) and incubated for 1 h at room temperature. For the detection of the primary antibodies CD163 and CD204, anti-mouse AlexaFluor647 (Invitrogen) and anti-rabbit AlexaFluor555 (Invitrogen) were used and diluted in PBS (Gibco). DAPI (Invitrogen) was used to counterstain nuclei. CD68 was detected using Zenon technology according to the manufacturers’ protocols (Invitrogen).

Specificity of primary and secondary antibodies was ensured by isotype-matched controls and negative controls (without primary antibody), respectively. After application of both primary and secondary antibodies, three washing steps were performed in PBS containing 0.05 % Tween. Slides were mounted with Elvanol (Calbiochem). Evaluation was performed with the TissueQuest system (TissueGnostics) which allows for automated quantification of multiple markers in whole tissue sections. Complete tissue sections were recorded with the Olympus IX51 microscope (at 20 × resolution) equipped with a XM10 camera and the Cell Sense software (all Olympus). Tissue sections were analyzed with the TissueQuest 4.0 software (TissueGnostics) through evaluation of median staining intensity of cells in FACS-like scattergrams [34]. All analyses have been conducted by an independent, blinded researcher.

Macrophage expression signature

A gene expression signature reported by Martinez and coworkers [37] was downloaded from the supplemental files provided with the publication. Signature genes (n = 97) were filtered for those also available in our dataset (n = 83). Fold changes were provided as positive and negative values for higher expression in M1 and M2 polarized macrophages, respectively. To indicate M2 phenotype by higher correlation values, fold changes were multiplied by −1. Pearson’s r for the correlation of these fold changes and expression data was calculated for all samples in the study cohort.

TCGA data

Level 3 gene expression data, summarized mutation data, and patient-centric copy number data released by Brennan et al. [2] were downloaded from the TCGA website (https://tcga-data.nci.nih.gov/docs/publications/gbm_2013/). Clinical data were acquired from the supplementary files provided with the publication. Gene expression microarray data were filtered for (1) non-tumor samples, (2) duplicate measurements, and (3) samples lacking clinical annotation, which were then removed (n = 519 remaining). Log2 transformed gene expression data were then normalized to z-scores and all samples classified as G-CIMP were removed (n = 468 remaining).

All genes with differential expression (p < 0.01, Student’s t test; n = 130) between LTS and STS in the discovery cohort were used as an input for the analysis (for gene names, refer to Table S2). Pearson’s r was calculated for all possible combinations of TCGA samples and tumors in the discovery cohort, resulting in a 468 × 16 correlation matrix.

Subsequently, each TGCA sample was classified as “LTS-like” if it was most highly correlated with a LTS tumor. Conversely, TCGA samples most highly correlated with a STS tumor were classified as “STS-like”. Z-score normalized expression values were used to define a “high” and “low” expression group for each gene using the median expression as a cutoff. p values for differences in survival between the groups were calculated with the Log-rank test. Genes were classified as “rOS genes” (reduced overall survival) or “iOS genes” (increased overall survival) if median OS (overall survival) was higher in the “low” and “high” expression groups, respectively.

To correct for multiple testing, the false discovery rate for overlap of differentially expressed genes between “LTS-like” and “STS-like” samples on the one hand and survival-associated genes on the other hand was estimated through a permutation-based approach. In total, 1000 permutations of group labels were performed. Differentially expressed genes for these random groupings were identified for each permutation and compared to survival-associated transcripts. Here, the χ 2 statistic served as a parameter for the degree of enrichment. The resulting false discovery rate was calculated as n sig/n total. Here, n sig corresponds to the number of permutations with a χ 2 statistic greater than observed in the original analysis while n total denotes the total number of permutations.

To facilitate further analysis, the TCGA patient-centric CNA data were simplified as follows: all broader gains and focal amplifications (indicated by +1 and +2 in the table) were denoted by +1; conversely, all broader losses and focal deletions were denoted by −1. All genes with multiple entries were removed. Gains and losses of whole chromosomes were assumed to be present if >50 % of genes on a given chromosome were affected.

Genome-wide expression for co-gain tumors and the rest of the data set were estimated by first sorting for chromosomal position based on coordinates retrieved from the ENSEMBL data base and approximating mean values for each group by a loess regression curve (smoothing parameter α = 0.05). Due to preprocessing (z-score normalization), expression patterns for each group represent deviations as compared to the average across the whole data set. Thus, higher expression denoted by an upwards deviation of the curve might reflect either higher proportions of gains or lower proportions of deletions, depending on the distribution of these copy number aberrations among the different groups. Accordingly, lower expression may be due to a higher proportion of deletions or a lower proportion of gains in these regions.

Validation of chromosome 19/20 co-gain

An independent cohort of GBM samples with extensive molecular annotation including IDH1 mutation status and outcome data was collected from the Sahlgrenska University Hospital and the Karolinska Institute in Sweden. Diagnosis was performed according to the latest WHO guidelines by an experienced neuropathologist (VPC). These samples were subjected to analysis by aCGH for copy number assessment (unpublished data, platforms and processing as described [27, 38]). A lift-over to the latest release of the human genome (hg19) was performed for genomic coordinates and data for all samples were plotted along the genome. Whole chromosome gains of chromosomes 19 and 20 were assessed using visual inspection. The data set was filtered for IDH wt samples with a confirmed diagnosis of primary glioblastoma. All samples which passed this filtering step (n = 124) were used for subsequent survival analyses.

Fluorescence in situ hybridization

Two-color interphase FISH was performed on deparaffinized sections using CEP probes for chromosome 20 (spectrum orange; Catalogue No. 06J36-030, Abbott) and chromosome 17 (spectrum aqua; Catalogue No. 06J38-027, Abbott) as a reference. Pretreatment of slides, hybridization, post-hybridization processing, and signal detection were performed as previously described [33]. Samples showing sufficient FISH efficiency (90 % nuclei with signals) were evaluated. Signals were scored in 200 non-overlapping, intact nuclei of each sample. A tumor was considered to carry a gain of chromosome 20 when the mean ratio of CEP20/CEP17 was >1.25 ± SD per sample. Moreover, the percentage of nuclei with detectable gain of chromosome 20 was calculated for each sample.

Expression data for aCGH validation cohort

Gene expression array data (Affymetrix U133A) was available for a subset (n = 32) of the aCGH validation cohort. These data had been deposited in the GEO database (Accesion No. GSE1993) as part of a larger study on astrocytic tumors [43]. All analyses were conducted within the R programming environment. Series matrix files were acquired through the “GEOquery” package, available in bioconductor [10]. Expression estimates were transformed to log2 scale and median centered. Repeat measurements of genes were combined into a single value (mean expression). Genes with differential expression (p < 0.01) between LTS and STS in the discovery cohort were used to classify cases as “LTS-like” and “STS-like” as outlined above (refer to “TCGA data” within the “Materials and methods”).

Results

Clinical characteristics of GBM long- and short-term survivors

Current standard in the treatment of primary glioblastoma consists of surgery and temozolomide (TMZ) based post-operative chemoradiotherapy, followed by six cycles of adjuvant TMZ [49]. To rule out any treatment-related bias and categorically exclude secondary GBMs, we focused on IDH wt GBM patients who completed the full treatment regimen. Patients with a survival of more than 36 months after diagnosis (LTS, n = 10) were compared to patients who survived for less than 10 months (STS, n = 6). Clinicopathological characteristics of the study cohort are summarized in Table 1 (for more detailed information, refer to Table S1). There were no significant differences in age (p = 0.19, Mann–Whitney U test), MGMT methylation status (p = 0.3, Fisher’s exact test) or sex (p = 0.63, Fisher’s exact test). Surprisingly, there was a trend towards better pre-operative performance (KPS) in STS (p = 0.08, Mann–Whitney U test). Furthermore, neither location of the tumor nor extent of resection differed between the groups (p = 0.51, χ 2 test and p = 0.61, χ 2 test, respectively). To conclude, there were no statistically significant differences in clinical parameters between LTS and STS GBM patients in our study cohort.

Table 1 Clinicopathological characteristics of LTS and STS discovery cohort
Full size table

Transcriptional profiling of GBM long- and short-term survivors

Assessment of gene expression subtypes according to the TCGA [55] splits our study cohort into mesenchymal (n = 7), classical (n = 7), and proneural (n = 2) groups (Table 1; Fig. S1). We did not observe enrichment of any subtype among LTS or STS (p = 0.8, χ 2 test). Transcriptional profiles of LTS and STS were investigated with a number of unsupervised and supervised methods. Principal component analysis (Fig. S2) did not result in a clear separation of the groups. However, out of 21,389 genes we found 1504 to be differentially expressed (p < 0.05, Student’s t test) between LTS and STS (Table S3). Of these, 755 showed higher expression in STS and 749 in LTS, respectively. From this, we concluded that while there are conserved differences in gene expression between STS and LTS, there is no global, uniform LTS phenotype detectable by microarray profiling of bulk tissue.

Next, pathway analysis was conducted for a better understanding of the biological processes underlying these expression differences (for detailed results, refer to Table S4 and S5). We noted that six out of seven significantly enriched pathways (FDR corrected p < 0.01) for genes upregulated in STS were involved in immune processes. Consequently, we found a number of genes well known for their role in tumor-promoting inflammatory processes and macrophage/microglia education. Among other genes identified through pathway analysis, PLA2G2A, CHI3L2, CCL2, CCL18 and CCL20 were overexpressed in STS (for more detailed information, refer to Tables S3 and S5).

Differential macrophage/microglia activation in LTS and STS

Recently, evidence has surfaced for divergent and even antagonistic effects of macrophages with different polarization. While the so-called M1 cells exhibit anti-tumor activity, the M2 phenotype supports tumor cell growth and invasion, a finding which also applies to brain tumors (reviewed in [18]).

We hypothesized that qualitative rather than quantitative differences regarding microglia might govern the LTS and STS phenotypes. Therefore, a previously published gene expression signature [37] was employed to investigate the activation status of microglia in our study cohort. Martinez et al. investigated transcriptional differences between macrophages with M1 and M2 polarization (reviewed in [40]). This signature consists of genes which were found to be differentially regulated between these phenotypes. Since it also includes information regarding fold changes, i.e., the quantitative amount of change, we were able to calculate correlation (Pearson’s r) between signature genes and expression data for all patients in our cohort. Here, positive and negative correlations imply M2 and M1 polarization, respectively. We found correlation values among STS to be significantly higher than those for LTS (p < 0.05, Student’s t test), indicating differential polarization and a shift towards the M2 phenotype among STS (Fig. 1a).

Fig. 1
figure 1

Microglial infiltration and activation status in LTS and STS. A previously published gene expression signature based on transcriptional differences of M1 and M2 activated macrophages was used to assess microglial polarization in the study cohort. a Shows correlation with this signature for LTS and STS samples. STS tumors exhibited significantly greater correlation with the M2 signature (p < 0.05, Student’s t test). These findings were followed up with multicolor immunofluorescence staining (see text). Staining for CD68, CD163 and CD204 allowed for the identification and quantification of macrophages (CD68+) and their polarization. b Congruent with transcriptomic results, M2 microglia were more abundant in STS tumors (p < 0.02, Mann–Whitney U test). The number of microglial cells, however, did not differ between the groups (refer to Fig. S3). c, d Show representative stainings for a LTS and STS case, respectively. CD cluster of differentiation, LTS long-term survivor, STS short-term survivor

Full size image

To examine whether these findings from bulk tissue gene expression data also translated to the protein level, we additionally performed multicolor immunofluorescence stainings. The TissueFAXS method [34] was applied to simultaneously detect and quantify CD68 (macrophage/microglia marker) as well as CD163 and CD204, two established M2 markers [30, 45]. We were thus able to assess both the amount and polarization of microglia for all cases in the study cohort. Staining for CD68 revealed no significant differences in the number of microglial cells (p = 0.43, Mann–Whitney U test, Fig. S3). The proportion of pro-tumorigenic M2 cells among tumor microglia, however, was significantly higher in STS as compared to LTS (p = 0.02, Mann–Whitney U test, Fig. 1b, d). Conversely, LTS tumors contained larger proportions of M1 cells (p = 0.02, Mann–Whitney U test, Fig. 1b, c). In summary, protein level data were consistent with the concept of differential polarization and, furthermore, confirmed prior transcriptomic analyses.

LTS and STS expression profiles are prognostic in non-G-CIMP GBMs

A number of recent publications on the transcriptome of GBM LTS have assessed mesenchymal, proneural, classical and neural subtypes among their patient cohorts [16, 46]. However, none of these studies validated differences in gene expression between LTS and STS in a larger number of tumors. As our study cohort was highly selected, yet small, we aimed to project the transcriptional phenotypes of LTS and STS onto a larger data set. Gene expression data were downloaded from the TCGA. As global hypermethylation is a sensitive parameter for IDH mutations, all G-CIMP cases were removed for further analyses (n = 468 remaining). Using genes which were differentially expressed between LTS and STS (n = 130, p < 0.01, Student’s t test; for gene names, see Table S2), we calculated pair-wise correlations of TCGA tumors and samples in our study cohort. The resulting correlation matrix is shown in Fig. 2a. Subsequently, each TCGA tumor was classified as either “LTS-like” or “STS-like”, according to the sample in the discovery cohort it was most highly correlated with.

Fig. 2
figure 2

Classification of TCGA samples based on LTS and STS expression profiles. a The heatmap shows all pair-wise correlations between LTS/STS samples of the discovery cohort (columns) and non-G-CIMP GBMs in the TCGA data set (rows) for genes with differential expression between LTS and STS (p < 0.01, Student’s t test, n = 130). Red and blue indicate high and low correlation, respectively (range −0.53 to 0.63). Each TCGA tumor was classified as “LTS-like” or “STS-like”, depending on the sample from the discovery cohort it was most highly correlated with. b Kaplan–Meier plot for non-G-CIMP TCGA samples according to the classification based on LTS/STS expression profiles. The blue and orange lines denote tumors classified as “LTS-like” and “STS-like”, respectively. The p value was calculated using the Log-rank test. For illustration purposes, a cutoff of 5 years was used for the x-axis. G-CIMP glioma CpG island methylator phenotype, LTS long-term survivor, STS short-term survivor, TCGA The Cancer Genome Atlas

Full size image

We noted a significant difference in survival between “LTS-like” and “STS-like” glioblastomas in the TCGA data set (p = 0.007, Log-rank test, Fig. 2b). We found no enrichment for classical, neural, mesenchymal or proneural tumors in either group (p = 0.67, χ 2 test), implying our classification was independent of these subtypes. There was also no significant association with established prognostic factors such as age (p = 0.38, χ 2 test) or extensive therapy (radiotherapy or chemoradiotherapy followed by adjuvant temozolomide treatment, p = 0.78, χ 2 test). We found a trend towards more frequent MGMT hypermethylation in the “LTS-like” group. However, this did not reach statistical significance (p = 0.1, χ 2 test). We further investigated how our classification performed for LTS in the TCGA data set. Out of 23 non-G-CIMP patients with survival of more than 36 months in the TCGA data set, 17 (73.9 %) were correctly classified as “LTS-like” by our approach (p = 0.03, χ 2 test). Pathway analysis (Table S6 and S7) again revealed inflammatory response as a major topic among genes with higher expression in “STS-like” tumors. We also found a difference in correlation with the aforementioned microglial signature in TCGA samples, similar to our discovery set. “STS-like” tumors exhibited significantly greater correlation (Pearson’s r) with the M2 signature than “LTS-like” tumors (p < 0.0001, Student’s t test; Fig. S4).

Large datasets such as those from the TCGA are often used to assess the prognostic value of certain target genes. A common approach is the median split, whereby a continuous variable is turned into a categorical one. More specifically, the median gene expression is used to dichotomize into a “low” and “high” expression group. Then, a statistical measure (in most cases, the Log-rank test) is used to test whether a significant difference in survival exists between these groups. Presumably, multiple mechanisms with an impact on patient outcome act in GBM tumors. Thus, various expression signatures associated with outcome might yield prognostic, yet different groupings. We therefore reasoned that a classification scheme might be considered more reliable if it has high correlation with the unbiased identification of survival-associated genes.

Thus, we extended the approach outlined above and assessed the overlap of survival-associated genes and those with differential expression between “LTS-like” and “STS-like” TCGA samples. First, the median split was used to identify genes with significant survival association in non-G-CIMP GBM (n = 1093; p < 0.05, Log-rank test). We nominally classified them as “iOS genes” (improved overall survival, n = 404) or “rOS genes” (reduced overall survival, n = 689). We hypothesized that genes with overexpression in “LTS-like” samples should be enriched for “iOS genes”. Conversely, we expected transcripts with higher expression in “STS-like” to be enriched for “rOS genes” (see Fig. 3a for a brief summary of the approach). Indeed, enrichment analysis revealed a striking overlap in both cases (p < 10−12 , χ 2 test, Fig. 3b–d). Furthermore, after accounting for multiple testing by estimation of the false discovery rate (FDR), the highly significant (q = 0.008) interrelation of our classification as “LTS/STS-like” and single-gene prognosticators was confirmed.

Fig. 3
figure 3

Overlap of survival-associated and differentially expressed genes in the TCGA data set. a Gives an overview of the approach. Survival association was investigated for all genes using a median split and Log-rank testing (see text). Genes were classified as “iOS genes” (improved OS) and “rOS genes”(reduced OS) if their expression correlated significantly with better and worse prognosis, respectively. We then investigated whether genes with significant survival association were more likely to be overexpressed in “STS-like” or “LTS-like” samples. b Provides a graphical representation of the results. The x- and y-axis show p values on a negative log10 scale for survival association (Log-rank test) and differential expression (Student‘s t test), respectively. For illustration purposes, p values for survival-associated “rOS genes” and transcripts overexpressed in “STS-like” samples were multiplied by −1. The dashed lines represent cutoffs for p = 0.05. The majority (81 %) of survival-associated “iOS genes” was also overexpressed in “LTS-like” tumors (blue circles in the upper right part). Conversely, “rOS genes” were also commonly (92 %) overexpressed in “STS-like” tumors (orange circles in lower left part). c The left contingency table provides a numerical representation for the highly significant (p < 10−12 , χ 2 test) enrichment of “iOS genes” among genes with overexpression in “LTS-like” tumors. d The right contingency table summarizes results for “rOS genes” and overexpression in “STS-like” tumors. Again, there was highly significant overlap (p < 10−12, χ 2 test). Colors in contingency tables correspond to those in (b). iOS improved overall survival, LTS long-term survivor, rOS reduced overall survival, STS short-term survivor

Full size image

Co-gain of chromosomes 19 and 20 is a prognostic factor in non-G-CIMP glioblastoma

We also explored the association of genetic lesions with LTS- or STS-like status in the TCGA data set. We identified a number of mutations (n = 20), which were significantly enriched (p < 0.05, Fisher’s exact test) in one or the other group (Table S8). However, each of these mutations only occurred in a small number of samples and did not show any prognostic effect by itself.

We further assessed whether chromosomal aberrations were more prevalent in one of the groups (for detailed results, refer to Table S9). We observed that prototypical GBM lesions such as amplifications on chromosome 7 and deletions of CDKN2A/CDKN2B as well as chromosome 10 were present in both groups at similar frequencies. However, LTS-like tumors were enriched for patients with concurrent gain of chromosomes 19 and 20, defined as duplications affecting more than half of each chromosome. Here, 55 of 86 samples (64 %) with co-gain were classified as “LTS-like”, whereas 31 (36 %) belonged to the “STS-like” group (p = 0.004, χ 2 test). While isolated gains exist, aberrations of chromosomes 19 and 20 co-occur frequently (p < 0.0001, χ 2 test) and are a common finding in non-G-CIMP GBM (18.5 %).

Notably, co-gain of chromosomes 19 and 20 is a highly significant prognostic factor (p < 0.0001, Log-rank test; Fig. 4a). However, it appears that only simultaneous gains are associated with a better prognosis. Cases with isolated gains of chromosomes 19 or 20 did not show improved outcome as compared to GBMs harboring neither aberration in the TCGA data set (p = 0.2 and p = 0.82, respectively, Log-rank test; Fig. 4b). Intriguingly, taken together, “LTS-like” status and chromosome 19/20 co-gain showed an even stronger association with survival (p = 2 × 10−6, Log-rank test) as compared to all other samples (Fig. S5).

Fig. 4
figure 4

Prognostic significance of chromosome 19/20 co-gain. a Shows a Kaplan–Meier curve for the TCGA data set (n = 465) comparing patients with concurrent gains of chromosome 19 and 20 to the rest of the cohort. b Provides a more comprehensive view comparing isolated gains of chromosomes 19 (gray line) or chromosome 20 (black line), co-gain of both chromosomes (blue line) and those samples with no gains on chromosomes 19 or 20 (orange line). Isolated gains of chromosomes 19 or 20 did not show improved outcome as compared to samples without these gains (p = 0.2 and p = 0.82, respectively, Log-rank test). Samples with co-gain show improved outcome compared to all other groups. c Shows aCGH data for a representative sample from a second, independent cohort of primary GBMs harboring co-gain of chromosomes 19 and 20 (with probes colored blue and orange, respectively). This tumor also exhibits additional copy number alterations typical for primary GBM, including high-level amplification of EGFR (chromosome 7) and homozygous deletion of CDKN2A on chromosome 9. d In congruence with our findings in the TCGA data set, co-gain of chromosomes 19 and 20 showed prognostic benefit in this second, independent validation cohort (p < 0.05, Log-rank test; n = 124). aCGH array comparative genomic hybridization, CDKN2A cyclin-dependent kinase inhibitor 2A, EGFR epidermal growth factor receptor, GBM glioblastoma multiforme, TCGA The Cancer Genome Atlas

Full size image

Co-gain of both chromosomes also remained a highly significant prognostic factor in multivariate analysis, correcting for age and extensive therapy (Table 2). However, in concordance with findings by the TCGA [2], we did not detect significant association with survival for MGMT hypermethylation after correction for other known prognostic factors. In line with these findings, LTS in the TCGA data set were also more likely to carry the co-gain. Here, 8 of 22 (36 %) patients with available copy number data exhibited the combined lesion. This corresponds to a significantly higher proportion as compared to the rest of the data set (18 %; p < 0.05, Fisher’s exact test).

Table 2 Multivariate analysis of OS for co-gain of chromosomes 19 and 20
Full size table

Independent validation of prognostic significance of chromosome 19/20 co-gain

To further interrogate and corroborate the prognostic importance of the chromosome 19/20 co-gain, we investigated a second, independent study cohort. This validation set consisted of previously unpublished aCGH data, including molecular annotation as well as clinical outcome information (for patient data, refer to Table S10). Again, we exclusively investigated IDH wt cases with confirmed diagnosis of primary glioblastoma (n = 124). Figure 4c shows a representative copy number profile for a sample with typical GBM-associated alterations and co-gain of chromosomes 19 and 20. We confirmed the highly significant co-occurrence of gains on chromosomes 19 and 20 (p < 0.0001, χ 2 test). Also, the frequency of co-gains (15.3 %) was similar to the TCGA data set (18.5 %). Albeit the considerably smaller number of samples, co-gain of chromosomes 19 and 20 was again significantly associated with patient survival (p < 0.05, Log-rank test, Fig. 4d).

For a subset of these samples, mRNA microarray data had been released as part of a prior study (GSE1993, [43]). This allowed us to also infer “LTS-like” and “STS-like” status for a limited number of samples (n = 32) in the aCGH validation cohort. We found that LTS-like tumors (n = 13) included all samples with co-gain (n = 3), whereas none belonged to the STS-like group (n = 19). Despite the very small sample numbers, this trend almost reached significance (p = 0.057, Fisher’s exact test).

Extent and transcriptional effect of chromosome 19/20 co-gain

We further assessed the extent of gains on chromosomes 19 and 20 in the TCGA dataset. Here, we noted a nearly bimodal distribution in both cases, suggesting that the majority of samples show either no aberration or gain of the whole chromosome (Fig. 5a–b).

Fig. 5
figure 5

Extent, clonality and transcriptional impact of gains on chromosomes 19 and 20. a, b Give an overview of the extent of gains on chromosomes 19 and 20, respectively. Histograms show the number of samples (y-axis) in which gains of different extent (x-axis) occur. Zero and one correspond to normal gene dose and whole chromosome gain, respectively. These data indicate a nearly bimodal distribution in both cases. c Depicts the genomic landscape in a typical co-gain sample. Comparable signals for gains on chromosome 19 and 20 suggest similar dynamics regarding their distribution among tumor cells. FISH was used to further differentiate between two models which might account for this observation (high amplitude gains in a subset of malignant cells vs. low amplitude gains in the majority of the tumor cell population). d Shows representative cases for FISH in a cohort of co-gain cases (n = 4, upper panel) and controls without gains on chromosome 19 or 20 (n = 2, lower panel). The majority of tumor cells in co-gain samples exhibited chromosome 20 gains, while controls were uniformly negative. e Shows normalized expression in the TCGA data set, approximated by a loess regression curve (smoothing parameter α = 0.05), plotted along the genome for samples harboring co-gain of chromosomes 19 and 20 (blue line) and the rest of the cohort (orange line). The dashed gray line represents average expression across the whole dataset. While minor changes were observed in different regions, only those for chromosome 19 and 20 reached statistical significance after correction for multiple testing (horizontal, dotted lines correspond to a FDR with q < 0.05). FDR false discovery rate, FISH fluorescence in situ hybridization, TCGA The Cancer Genome Atlas

Full size image

Recent studies have highlighted the extensive genetic heterogeneity within tumors of the same patient [17]. The current model of tumor evolution states that early, initiating drivers are present in all cancer cells whereas later events only occur in subpopulations of the tumor (reviewed in [19]). We thus interrogated the extent of the co-gain and the proportion of affected cells by FISH in four IDH wt GBMs harboring this alteration and two controls.

These samples were selected based on copy number data generated from DNA methylation microarrays [25, 52]. A copy number profile for a representative co-gain sample of this cohort is shown in Fig. 5c (for other cases, refer to Fig. S6). Of note, all samples were IDH wt and exhibited genomic landscapes typical for primary GBM. As these tumors show almost identical signal intensities for chromosomes 19 and 20, similar dynamics regarding their distribution among tumor cells is likely. For this reason and technical limitations, we used chromosome 20 as a surrogate parameter for the co-gain. Our results confirmed gains in the majority of tumor cells in all selected samples, while controls were uniformly negative (data for two representative cases are shown in Fig. 5d). For most affected cells, 2–3 additional copies of chromosome 20 were observed per nucleus.

Upon analysis of gene expression across the genome, in addition to low amplitude variations in several regions of the genome, we found a strong upregulation of genes located on chromosomes 19 and 20 for tumors harboring the co-gain, indicating a gene dosage effect (Fig. 5e). Gene expression also showed smaller changes across other chromosomes. Indeed, certain copy number aberrations such as deletions on chromosomes 13, 14 and 15 seem to be almost mutually exclusive with chromosome 19/20 co-gain. However, these copy number changes were themselves not associated with patient survival (data not shown). Yet, upon FDR correction, only expression changes on chromosomes 19 and 20 remained significant (q < 0.05).

We conclude that co-gains of chromosomes 19 and 20 (1) frequently involve the whole chromosome, (2) likely affect the majority of tumor cells and (3) lead to upregulation of genes located in these regions.

FISH analysis in the discovery cohort

Co-gain of chromosome 19 and 20 was identified in the TCGA data set based on a transcriptional signature learned from our initial STS and LTS cohort. We thus hypothesized the co-gain might also be present in LTS cases of the discovery cohort. Therefore, we again performed FISH analysis for chromosome 20. Intriguingly, the experiment revealed gains among eight out of ten (80 %) LTS samples (Table S11). Due to the lack of additional data, we cannot preclude that some of these cases might harbor isolated gains of chromosome 20. However, even when considering the cumulative frequency (33 %) for co-gains and isolated chromosome 20 gains, this finding represents highly significant enrichment (p = 0.003, Binomial test).

Discussion

Due to the grim prognosis of GBM, LTS have intrigued researchers for decades and reports about these patients have been published as early as 1950. However, most results are heavily biased by IDH mutations. In the present study, we accounted for this confounding variable and identified a subgroup of GBM, characterized by co-gain of chromosomes 19 and 20 and markedly better survival. Moreover, we provided considerable evidence for differential polarization of microglia with higher abundance of the M2 phenotype in STS. Most importantly, these findings have been confirmed in an independent patient cohort and by different experimental approaches.

As of now, cytogenetic studies and genome-wide screens in GBM have focused mostly on common copy number alterations (CNAs) such as amplifications of EGFR, losses on chromosome 10 (including the tumor suppressor PTEN) and homozygous deletions of CDKN2A (reviewed in [51]). Furthermore, these studies have produced conflicting results. For example, EGFR amplification has been associated with both improved [24] and poor [48] outcome, while some reports indicated no prognostic effect at all [23]. Consequently, none of these CNAs are assessed in routine diagnostics due to their inconclusive nature. Isolated gains of chromosomes 19 or 20 are also common CNAs in GBM [2]. Despite their frequency, there is only anecdotal evidence of their co-occurrence [32]. Here, we not only consolidate this finding, but also provide compelling evidence for the strong prognostic benefit for tumors with concurrent gain of chromosomes 19 and 20. Also, this co-gain could easily be assessed as part of a routine neuropathological workup in the near future as high-throughput methods are finding their way into diagnostics. For instance, array-based methylation profiling allows for simultaneous detection of G-CIMP tumors and MGMT hypermethylation while generating copy number profiles at the same time, including information about chromosome 19/20 co-gain [52]. Thus, identification of primary GBMs and stratification according to prognostic factors such as the co-gain could soon become a useful complement to histopathological evaluation.

While many prognostic classifications have been proposed for GBM [7, 15, 29, 36, 44], none have adequately corrected for IDH mutations. This is also supported by results from Reifenberger et al., which indicate that identification of IDH mut tumors is a hallmark feature of previously published signatures [46]. By accounting for both IDH status and treatment, we are among the first to provide a classification scheme capable of predicting prognosis in purely IDH wt tumors.

Combined gains of chromosome 19 and 20 had been identified through a transcriptional signature learned from LTS and STS. As the same samples from which the signature was initially derived were later shown to also harbor gains, these findings are most likely interrelated. This is also supported by the cumulative prognostic effect of “LTS-like” status and presence of the co-gain. Possible causative effects, however, should be addressed in future studies.

Chromosome 19 has the highest gene density among the genome and contains several large gene families, including zinc finger transcription factors and cytochrome P450 enzymes [20]. Both chromosomes also contain genes implicated in diseases of the nervous system [12, 20] such as PRNP which is linked to Creutzfeldt–Jakob disease. Furthermore, different alterations of these chromosomes have been implicated in cancer. For instance, loss of 1p and 19q is a hallmark CNA of oligodendroglioma [35], while gains have been described for leukemia [1]. As the observed co-gains involve all of chromosomes 19 and 20, the prognostic effect is unlikely caused by single transcripts. As a consequence, our study clearly delineates the urgent need for further research addressing the functional consequences of whole chromosome gains rather than single genes.

Pathway analysis of differentially regulated genes in our study cohort implicated tumor-promoting, microglia-driven inflammatory processes in STS. Active anti-tumor response by the adaptive immune system constitutes a favorable prognostic factor in GBM [34] and has also been identified as a feature of LTS tumors [13]. Yet, the ambivalent role of the innate immune system in glioma is well documented (reviewed in [18]). Recruited macrophages/microglia and myeloid progenitors can exert pro-tumorigenic effects such as angiogenesis [41], invasion [56], and proliferation [9, 28].

These cancer-enabling inflammatory processes have previously been described as a salient feature of the mesenchymal subgroup. Differentiation towards this phenotype is driven by a conserved transcriptional network [6] and was shown to correlate with poorer survival [44] in early high-throughput studies. Subsequent efforts have also provided evidence for higher numbers of infiltrating macrophages and greater extent of necrosis in this subtype [8, 14]. In addition to these findings, abundance of M2 macrophages has been found to correlate with glioma grade [45].

However, after exclusion of IDH mut/G-CIMP samples, the largest study in GBM to date [2] has not confirmed inferior prognosis for mesenchymal tumors. Indeed, there even was a trend towards better outcome as compared to IDH wt proneural samples, highlighting the confounding effect of IDH mutations. It is important to realize the imminent impact of these results, as a considerable number of studies have made use of the very same data for validation purposes.

Recently, a pro-tumorigenic inflammation signature has been described for IDH wt STS [16]. In addition to confirming this finding, we have resolved these expression differences further. Transcriptomic analyses and multicolor immunofluorescence stainings have provided considerable evidence for divergent polarization with higher numbers of M2 microglia in STS tumors. Thus, our multilevel approach provides important insights into the association of innate immune response and survival in IDH wt GBM.

The substantial impact of unrecognized IDH mutations on research in GBM is also one of the most important shortcomings of previous studies on LTS. As only recently became evident, these tumors have an utterly different biology despite histological similarities with GBM. Those differences include, but are not limited to, (1) younger age at diagnosis, (2) a different pattern of CNAs, (3) an almost uniform proneural expression subtype and (4) G-CIMP. Their confounding effect on gene expression signatures has been outlined above. At the same time, studies on CNAs usually detected hallmark lesions of primary GBM among STS and common aberrations of lower grade lesions among LTS [3, 4]. This bias extends across all levels of cell biology including epigenetics, which is why methylation screening found G-CIMP tumors to be overrepresented among LTS [47]. Additional proof comes from the reportedly high histopathological misclassification rates [39] in LTS and the high prevalence of IDH mutations among them [21]. Therefore, most studies on LTS to date have identified prototypical features of lower grade gliomas for which better survival is well documented, whereas we have taken great care to avoid this bias.

In summary, we provide strong evidence for M2 polarization of microglia as a feature of particularly aggressive tumors. Furthermore, we have shown co-gain of chromosomes 19 and 20 to constitute a bona fide marker for a subgroup of primary GBM with better outcome. It is detectable both on the single-cell and tissue level with different technical platforms and robustly associated with patient survival. In addition, we utilized extensive validation cohorts as part of our study. As clinical studies necessitate appropriate stratification according to known prognostic biomarkers, our findings will have important clinical implications and can, at the same time, be easily integrated into modern molecular diagnostics.