Mean global DNA methylation serves as independent prognostic marker in IDH-wildtype glioblastoma

Abstract

Background

The IDH-wildtype glioblastoma (GBM) patients have a devastating prognosis. Here, we analyzed the potential prognostic value of global DNA methylation of the tumors.

Methods

DNA methylation of 492 primary samples and 31 relapsed samples, each treated with combination therapy, and of 148 primary samples treated with radiation alone were compared with patient survival. We determined the mean methylation values and estimated the immune cell infiltration from the methylation data. Moreover, the mean global DNA methylation of 23 GBM cell lines was profiled and correlated to their cellular radiosensitivity as measured by colony formation assay.

Results

High mean DNA methylation levels correlated with improved survival, which was independent from known risk factors (MGMT promoter methylation, age, extent of resection; P = 0.009) and methylation subgroups. Notably, this correlation was also independent of immune cell infiltration, as higher number of immune cells indeed was associated with significantly better OS but lower mean methylation. Radiosensitive GBM cell lines had a significantly higher mean methylation than resistant lines (P = 0.007), and improved OS of patients treated with radiotherapy alone was also associated with higher DNA methylation (P = 0.002). Furthermore, specimens of relapsed GBM revealed a significantly lower mean DNA methylation compared to the matching primary tumor samples (P = 0.041).

Conclusions

Our results indicate that mean global DNA methylation is independently associated with outcome in glioblastoma. The data also suggest that a higher DNA methylation is associated with better radiotherapy response and less aggressive phenotype, both of which presumably contribute to the observed correlation with OS.

Key Points.

1. GBM patients treated with combination therapy showed a significantly better outcome if tumors showed a higher DNA methylation.

2. Higher DNA methylation was associated with higher radiotherapy response in GBM patients and higher radiosensitivity of cell lines.

Importance of the Study.

Patients with GBM are typically treated with a combined radiochemotherapy, but median overall survival is still below two years after diagnosis. Additional prognostic markers are urgently needed to better understand disease outcome and predict response to therapy. Robust analytical methods, such as global DNA methylation, are suitable for this purpose in order to improve the therapy of GBM patients in the future. Here, we identified mean global DNA methylation as a prognostic marker, which was associated with higher radiosensitivity.

Glioblastoma multiforme, IDH-wildtype (GBM) is the most common malignant brain tumor in adults and is classified as CNS WHO grade 4.¹ The median age of GBM patients is 64 years and the female-to-male ratio is 1:1.6.^2,3 Patients, who do not receive additional therapy after resection have a median overall survival (OS) of less than 6 months after diagnosis.² Standard therapy, consisting of surgery and adjuvant radiochemotherapy,⁴ increases the OS to 12 to 15 months.² However, these values vary individually, with 15% of the patients surviving only 6 months or less after diagnosis and 20% surviving longer than 2 years.³ Known prognostic markers are age and extent of resection.^5,6 Furthermore, MGMT (O6-methylguanine–DNA methyltransferase) promoter methylation status is beneficial for survival by increasing the response to temozolomide-based chemotherapy.^5,7,8 Nevertheless, further clinical and molecular markers are urgently needed to stratify patients and to predict response to therapy.

In general, DNA methylation data can be used through cluster analysis of methylation sites or through the identification of prognostic markers based on copy number profiles inferred from methylation data. In the last years, prognostic markers were examined in order to distinguish tumors according to their aggressiveness and to better predict their response to therapy. These efforts have also included genome-wide DNA methylation by using Illumina arrays,^9–11 and the analysis of such genome-wide DNA methylation signatures has led to the separation of more than 80 tumor entities of the central nervous system. For GBM, several methylation subgroups have been described.¹² The three most frequent subgroups of GBM are mesenchymal (MES), receptor tyrosine kinase I (RTK I), and receptor tyrosine kinase II (RTK II),^12–14 which account for about 90% of the cases and are associated with an equally poor survival.^13,15 RTK II [enriched for epidermal growth factor receptor (EGFR) gene amplifications] occurs most frequently (40%–50%), followed by MES (20%–45%) [enriched for neurofibromin-1 (NF1) aberrations].^14,16 RTK I, which is enriched for platelet-derived growth factor receptor-alpha (PDGFRA) gene amplifications, is the least common of the three subgroups (about 20%).¹⁴ Despite the success of the epigenetic analyses in molecular neuropathology, a prognostic value, which is evident for multiple entities, such as ependymoma and medulloblastoma, has not been identified for GBM so far.^17,18

Using a large cohort of 492 GBM patients homogeneously treated according to standard therapy, we here propose how to make the methylome data usable for robust survival prognosis.

Material and Methods

Patient Cohort

Three different cohorts were analyzed, including a total of 645 primary and 31 relapsed GBM samples. The first patient cohort consists of 492 GBM patients from three German hospitals (Charité Berlin, University Hospital of Frankfurt, University Medical Center Hamburg-Eppendorf) and previously published data (The Cancer Genome Atlas (TCGA), Gene Expression Omnibus database (GEO) under the accession numbers GSE60274, and GSE195640).¹⁴ All samples were classified as GBM that had the highest score as GBM on the Heidelberg brain tumor classifier (v11.6). Age, survival, extent of resection, DNA methylation profiles, and therapy details were available for most patients. The extent of resection (EOR) was divided into gross total resection (GTR), near GTR, and partial resection or stereotactic biopsy. The complete removal or removal of more than 90% of contrast-enhancing parts was defined as a GTR or near GTR, whereas a resection of less than 90% was defined as partial resection or biopsy. The EOR of contrast-enhancing parts was evaluated by MRI performed up to 48 h after surgery.^5,6 Only patients treated with a combination of radiation and temozolomide-based chemotherapy (Stupp regimen) were included in this first cohort. The second cohort, which was derived from previously published data (GEO under the accession numbers GSE60274, GSE195684, and GSE195640)¹⁴ includes only patients, who did not receive chemotherapy (n = 148). The third cohort includes 31 patients who had a second resection because of a recurrence, and the methylome was analyzed for both of their tumors. For this study, 30 patients were treated by the Stupp regimen in the first-line therapy, and one patient was treated with temozolomide-based chemotherapy alone. The use of biopsy-specimens for research was approved and in accordance with local ethical standards and regulations at the University Medical Center Hamburg-Eppendorf (PV4904).

DNA Methylation Profiling

Extracted tumor DNA was analyzed for genome-wide DNA methylation patterns using the Illumina 450k array or Illumina EPIC (850k) array. All data are available at https://www.ncbi.nlm.nih.gov/geo/ under the accession number GEO240704. The DNA methylation data obtained were further processed using previously described approaches.¹² The analysis and visualization of the DNA methylation data were performed with R (version 3.6.0). The IDAT files of the two arrays used (Infinium HumanMethylation450 BeadChip and Infinium MethylationEPIC BeadChip) were processed and combined using minfi Bioconductor package v3.9. Probes located on two nucleotides, containing a single-nucleotide polymorphism, or located on the X or Y chromosome were excluded. The preprocess noob function of the minfi package was employed, and beta values were calculated. For the batch effect correction regarding the array type, the package limma (v3.40.0) was used. To verify the analyses based on the minfi, the sesame package (v1.18.4) was also used to preprocess the IDAT files. The delta beta value described the difference in mean methylation of both groups. Usually, a delta beta value of at least 0.2 is used. Mean DNA methylation was determined from the mean of all beta values per sample.

The package DIMEimmune was used to estimate the number of CD4⁺ and CD8⁺ T cells and the number of tumor-infiltrating lymphocytes (TILs).¹⁹ Analysis of immune cell infiltrates was performed using the DIME (Differential Methylation Analysis for Immune Cell Estimation) score described in Safaei et al.¹⁹ For the analysis of the other cell populations, EpiDISH with the reference file of Grabovska et al. was used (v2.16.0).²⁰ The proportion of myeloid cells was calculated by summing all proportions of eosinophils, neutrophils, and monocytes.

To calculate and visualize the copy number aberrations per sample, the conumee package (v1.34.0) was used. The modified total aberration index (tai) was calculated with the CINmetrics package (v0.1.0). For the tai, the length of a segment was multiplied by the segmentation mean if there was a gain or loss in that segment (threshold: ±0.1). Then, these products were added for all segments and divided by the sum of the segment lengths, describing the tai per sample. If tai was equal to 0, the genome was balanced.

Statistical Analysis

The calculation of the best cut-off was performed with the function maxstat.test from the package maxstat (v0.7-25), the log-rank test by using the ranked statistics and the method Lau94 were used for the calculation of the best cut-off. The package ggplot (v3.4.0) was used to represent the data by boxplots. For survival analysis, Kaplan–Meier plots were generated with the packages survminer (v0.4.4) and survival (v3.4-0). The groups were compared using the log-rank test. The multivariate analysis was performed with the finalfit package (version 1.0.5).

Cell Culture

The GBM cell lines LN229, U87MG, LN827, LN71, U343, and Cas-1 were cultured in DMEM (Sigma–Aldrich) supplemented with 10% FCS (Biochrome), 2 mM L-glutamine, and 1 mM sodium pyruvate (Sigma–Aldrich); DKMGvIII- cells were cultured in RPMI (10% heat inactivated FCS, 2 mM L-glutamine, 1 mM sodium pyruvate) while BS153vIII- cells were cultured in DMEM (10% heat inactivated FCS, 2 mM L-glutamine and 1 mM sodium pyruvate). All cells were cultured at 37°C, 5% CO₂, and 100% humidification and were authenticated using short tandem repeat (STR) profiling by Eurofins.

Cell Survival and X-Irradiation

Cell survival after exposure to X-ray was analyzed by colony formation assay. Two hundred and fifty to 350 cells were placed in six-well plates and irradiated 24 h later (Gulmay RS225; Gulmay Medical Ltd.; 200 kv, 15 mA, 0.8 mm Be + 0.5 mm Cu filtering; dose rate of 1.2 Gy/min). Twenty-four hours after the treatment, the medium was replaced. Only for BS153, AmnioMax C-100 Basal Medium (Life Technologies) containing 10% FCS and C-supplement (Life Technologies) was used instead of the usual medium to improve colony formation. Colonies were allowed to grow for 1.5 to 3 weeks depending on the cell line and irradiation dose. Afterward, the colonies were fixed with 70% ethanol and stained with crystal violet. Colonies with at least 50 cells are counted. Cell survival analysis was performed in PRISM (v6.07). We classified the cell lines according to their cell survival at an irradiation dose of 4 Gy. For this purpose, we chose the best cut-off value for each cohort using their mean global DNA methylation.

Results

Patient Characteristics Are Representative in Demographics and Overall Survival

The aim of this study was to analyze the prognostic value of methylome data for GBM patients. Therefore, we first collected and analyzed DNA methylation and OS data from 492 GBM patients, who were homogeneously treated with standard combination therapy (Table S1). Cohort characteristics including the known prognostic factors age, MGMT promoter methylation, and extent of resection are displayed in Figure 1A. The mean age was 59.5 years with a female-to-male ratio of 1:1.6, while MGMT promoter methylation was present in 47.4% of the cases. Information on the extent of resection was available for 315 patients, with most tumors being resected (84.8%), with the extent of resection ranging from gross total resection (GTR, 37.8%), to near GTR (26.7%) to partial resection (20.3%). The median OS was 15.8 months, with 168 cases being censored (Figure 1B). Twenty percent of the patients survived 7.5 months or less, while another 20% survived at least 35 months. In accordance with the literature, age, MGMT promoter methylation as well as the extent of resection were independent prognostic factors (Supplementary Figure S1).

Figure 1.

Overview of the GBM patient cohort. Analysis based on (A) clinical data and (B) Kaplan–Meier plot with the median overall survival marked with an asterisk.

Mean DNA Methylation Is Significantly Different in Long and Short Surviving Patients

Next, we stratified the cohort into long and short surviving patients based on the median OS of 15.8 months. If patients were censored before 15.8 months of survival, they were excluded (n = 106), as it was unknown, to which group they would be definitively assigned. Within the remaining 386 samples, we did not detect significant differences in the methylation of single CpG sites and, therefore, failed to identify a methylation signature that is associated with better or worse survival of GBM patients (Supplementary Tables S2 and S3).²¹

To further investigate epigenetic prognostic markers of survival, we calculated the mean methylation of each sample and the median methylation of the entire cohort, which was 0.49 (Figure 2A). Dividing the cohort according to this median methylation, a significant difference in OS was detected (Figure 2B). A higher level of mean DNA methylation was associated with a significantly better OS (methylation > 0.49, median OS: 18.0 months), while a lower methylation was associated with a significantly worse OS (median OS: 13.0 months; P = 0.007). Further analysis using the best cut-off with the ranked statistics for the mean methylation (beta value = 0.458) identified a small group of patients, whose tumors had very low mean DNA methylation (n = 73) and showed a highly significant lower OS (P < 0.001) with a median OS of 9.2 months compared to 17.0 months (Figure 2C, D). Subsequent multivariate analysis showed that mean methylation is an additional prognostic factor, independent of age, MGMT promoter methylation, and extent of resection (HR: 1.50; 95% CI: 1.11–2.04; P = 0.009, Figure 2E).

Figure 2.

Analysis of mean methylation of 492 GBM patients. (A) Distribution of mean methylation marking the median (black) and best cutoff (orange). (B, C) Kaplan–Meier curves after division of the cohort at the (B) median mean methylation level and (C) best cutoff using the (D) calculated ranked statistics. (E) Multivariate analysis of mean methylation in combination with known prognostic factors. (F) Combinations of all prognostic markers [Age (threshold: >65 years), EOR (extent of resection), MGMT promoter methylation, MM (mean DNA methylation)] colored in green (negative prognostic marker) and orange (positive) and (G) Kaplan–Meier plots of three risk groups (1: orange, 2: purple, 3: green).

Next, we analyzed a potential bias due to the extent of resection. However, both groups (GTR/near GTR and partial resection/biopsy) showed comparable levels of mean methylation (Supplementary Figure S2A). Moreover, the level of mean methylation was a significant prognostic marker in both groups (Supplementary Figure S2B–C). Furthermore, we investigated the influence of copy number aberrations on the mean methylation by calculating the modified total aberration index (tai). However, there was no significant difference in the survival after splitting the cohort at the best cut-off of the tai (Supplementary Figure S2D). Also, tai scores were not significantly different with respect to mean methylation using either minfi or sesame for preprocessing (Supplementary Figure S2E, F). Most of the samples had a negative tai score implying a negative balance regarding copy number aberrations. This observation was confirmed in cumulative plots of the high or low methylated samples (Supplementary Figure S2G, H). Therefore, copy number aberrations did not show any correlation to the mean methylation in our cohort.

We next classified all cases using the Heidelberg brain tumor classifier. As expected from the literature,^13,15 there was no significant difference in patient survival between mesenchymal (MES), RTK I, and RTK II GBM subgroup (Supplementary Figure S3A). When analyzing the median methylation of the three GBM subgroups, we observed that RTK I tumors had a significantly lower mean methylation level than MES or RTK II tumors (Supplementary Figure S3B; P < 0.001). Still, a high mean methylation was correlated with better survival in patients with RTK I and RTK II GBM (best cut-off; Supplementary Figure S3C–E). When including the GBM subgroup into the multivariate analysis, the mean methylation still remained an independent prognostic factor (Supplementary Figure S3F).

In order to analyze whether mean DNA methylation is generally correlated with tumor aggressiveness, we employed a reference set including 91 brain tumor and control methylation classes¹² and compared the mean methylation of the different (sub-)entities (Supplementary Figure S4). However, comparatively high methylation was also observed in very aggressive tumors, such as atypical teratoid/rhabdoid tumors (ATRT),²² and less deadly brain tumors, such as pituitary adenomas may also have lower mean DNA methylation levels. These findings demonstrate that a high-mean methylation is not generally related to a better outcome.

Since we have shown that mean global DNA methylation is an independent prognostic marker for standard-treated GBM, we next combined all independent prognostic markers to calculate three risk groups considering only patients with available information for all prognostic markers (n = 315, Figure 2F). For this, we defined all combinations where the Kaplan–Meier plots showed no significant difference between them, as a specific risk group. Risk group 1, which had the best survival, included all patients, who had no negative prognostic marker (median OS: 39.0 months; Figure 2G). The second risk group described all patients with either one prognostic factor or one of the combinations high age + methylated MGMT promoter or high age + low mean methylation (median OS: 19.0 months). The worst survival was observed in the risk group 3 (median OS: 8.3 months). It included all patients with at least three prognostic factors, any combination with a low extent of resection or the combination of unmethylated MGMT promoter + low mean methylation. All three risk groups differed significantly in OS (P < 0.001).

Immune Cell Infiltration Does Not Explain the Correlation From OS and Mean Methylation

Since differences in OS can arise from differences in immune cell infiltration with more infiltrating immune cells reportedly being associated with better OS,²³ we next analyzed whether differences in mean methylation are associated with the number of infiltrating immune cells. To this end, we used a deconvolution algorithm (DIME Immune Score¹⁹) to estimate the number of CD4⁺, CD8⁺, and tumor infiltrating lymphocyte (TIL) cells in our series of GBM. In line with the literature, we observed a longer OS of GBM patients with higher DIME scores (best cut-off of for each DIME score: CD4⁺: −0.026; CD8⁺: 0.448; TIL: 0.956; Figure 3A–C). Importantly, the DIME scores for all three immune cell infiltrates were significantly higher in the low methylated group of GBM using the best cut-off as in Figure 2C than in the higher methylated tumor samples (Figure 3D–F; CD4⁺: P < 0.001; CD8⁺: P = 0.003; TIL: P = 0.024). To further investigate the influence of distinct cell populations on mean methylation, we additionally used EpiDISH to analyze the populations of myeloid cells, in particular monocytes, neurons, and glial cells with respect to mean methylation (Supplementary Figure S5). Indeed, samples with high methylation showed significantly more myeloid cells and monocytes (Supplementary Figure S5A, B), whereas numbers of neurons and glia cells were comparable in GBMs with low- or high-mean methylation. Looking at the GBM subgroups, only MES GBM had a high proportion of myeloid cells, whereas RTK-II-GBM did not (Supplementary Figure S5E). However, since mean methylation was not significantly different in MES and RTK II tumors, the proportion of myeloid cells did not appear to be the only reason for high mean methylation. Therefore, our data suggest that infiltrating immune cells were not causative for the better survival of patients with increased methylated GBM.

Figure 3.

Correlation of CD4⁺, CD8⁺, and TIL immune cell counts to mean methylation. (A–C) Kaplan–Meier curves dividing the cohort at the best cutoff of the respective DIME scores and DIME score of (D) CD4⁺, (E) CD8⁺, and (F) TIL for low- and high-methylated tumor samples.

Higher Mean DNA Methylation Is Associated With Higher Radiosensitivity

All patients analyzed so far had received standard therapy including radiotherapy and temozolomide (TMZ). To investigate a correlation of mean methylation to therapy, we performed the analyses with irradiation as monotherapy. To this end, we determined the cellular radiosensitivity of a panel of commercially available GBM cell lines by colony formation assay and stratified them by the survival fraction at 4 Gy (SF4) in radiosensitive and radioresistant cell lines (Figure 4A). DNA methylation was analyzed for these, and an additional panel of 15 stem-like cell lines with published SF4 data.²⁴ The latter were also divided in radiosensitive and radioresistant cell lines as shown in Figure 4B. The analysis of cumulative copy number variations within these cell lines did not show any obvious differences that could cause differences in radiosensitivity (Supplementary Figure SF6). However, the combined analysis of SF4 and mean methylation for both cohorts revealed a significantly higher mean methylation for radiosensitive cell lines compared to radioresistant cell lines (Figure 4C; P = 0.007). This suggested that increased cellular radiosensitivity might be a cause for improved survival of patients with highly methylated GBM. To support this hypothesis, we correlated radiotherapy response with mean DNA methylation in 148 patients, who received radiotherapy but no TMZ (Table S4). Patients divided at the best cut-off with higher mean methylation survived significantly longer (median OS: 12.6 months) than patients with lower mean methylation (median OS: 9.4 months; P = 0.029; Figure 4D). By defining two risk groups based on age, MGMT promoter methylation, and mean DNA methylation the difference in OS even ranged from 8.4 to 12.5 months (Figure 4E, F). With low-mean methylation of the tumors, patients survived significantly shorter and are classified in risk group 2 (combination 4 and 6-8). Patients with at least two negative prognostic markers due to their advanced age and low-mean methylation showed the shortest median survival (median OS (combination 6): 7.6 months and median OS (combination 8): 7.2 months). The risk group 1 showed a significantly longer OS (median OS: 12.5 months) than the risk group 2 (median OS: 8.4 months; P = 0.002; Figure 4F).

Figure 4.

Radiosensitivity (red: radiosensitive, blue: radio resistant) of GBM (A, B) cell lines. Dose response curves of eight established GBM cell lines (surviving fraction at 4 Gy: SF4_est). (B) SF4 of 15 GBM stem-like cells.²⁴ (C) Combination of both cohorts and correlation of the mean methylation with SF4 (radiosensitive: red, radioresistant: blue). (D–F) Patient data. (D) Survival of patients treated with irradiation only (148) analyzed by Kaplan–Meier analysis. The curves are divided at the best cutoff (higher methylation: red, lower methylation: blue). (E) Combinations of all prognostic markers [Age, EOR (extent of resection), MGMT promoter methylation, MM (mean DNA methylation)] colored in green (negative prognostic marker) and orange (positive) and (F) Kaplan–Meier analysis of two risk groups (1: orange, 2: green).

Changes in Mean DNA Methylation During Tumor Recurrence

Finally, we tested if mean methylation might also be a general marker for a lower relapse risk. Since recurrent GBM tumors have a more aggressive phenotype than primary tumors,²⁵ we analyzed changes in the mean DNA methylation during first and second resection in a cohort of 31 tumor pairs (Table S5). The tumor pairs indeed showed a difference in mean methylation between the first and second resection (Figure 5A) with a significantly lower mean methylation in the re-resected samples (P = 0.041). Only eight tumors had higher mean methylation at the second compared to the first resection. Because RTK I tumors have a significantly lower mean methylation level than the other two subgroups (Supplementary Figure S3B), and it is known that a change of the tumor subgroup can occur during therapy,¹⁶ we determined possible subgroup switches (Figure 5B). We identified four RTK I tumors which changed the subgroup between resections, while there was only one non-RTK I tumor that underwent a subgroup change from non-RTK I to RTK I subtype. This indicates that lower mean methylation is not caused by a subgroup switch but is associated with a more aggressive tumor phenotype. Furthermore, using EpiDISH as method for deconvolution, we analyzed and compared the tumor cell content and proportion of myeloid cells in the samples from the two resections. There was no significant difference between the first and second resection in terms of the cell populations examined in the samples (Figure 5C, D). Therefore, lower mean methylation cannot be explained by fewer myeloid cells in this cohort either.

Figure 5.

Changes in mean DNA methylation during disease progression. (A) Analysis of recurrence samples (31) versus matching primary tumor (31) regarding mean DNA methylation and (B) subgroup switching between RTK I and non-RTK I. Analysis of the cell populations in the recurrence samples versus primary tumor of (C) cancer cells and (D) myeloid cells.

Discussion

Fifty percent of glioblastoma patients survive only below 2 years after diagnosis despite intensive therapy. Previously, prediction of response to temozolomide-based chemotherapy was possible by analysis of MGMT promoter methylation. In addition, prediction of patient survival is also possible to a limited degree based on patient age and extent of resection. With this study, we demonstrate the prognostic value of genome wide DNA methylation for GBM. In contrast to the GBM subgroup classification achieved by distinct methylation patterns—which have no clear prognostic value—the prognostic impact lies in the mean methylation of a given tumor. The mean methylation is a simple and robust prognostic marker and was demonstrated to be independent from other prognostic markers, such as age, MGMT promoter methylation, and extent of resection. The combination of all four markers even allows a more precise prediction of OS after standard therapy and a clear stratification of patient subgroups with worst (median OS: 8.3 months), moderate (median OS: 19.0 months), and relatively good prognosis (median OS: 39.0 months).

We tried to determine individual CpG sites, which are informative of patient survival but the methylation of single CpG sites is probably not sufficient to predict OS of GBM patients. It has been described by others that specific differentially methylated CpG sites can be associated with OS.¹⁰ However, this is likely due to a more heterogeneous cohort and patients, who have not received adjuvant radiochemotherapy. The mean age of the 492 patients analyzed in our project reflected the known literature as well as gender distribution, and proportion of MGMT promoter methylation.^3,26,27 Therefore, our cohort was not only particularly large and well annotated but also representative of the demographic distribution and survival of glioblastoma patients.

The three methylation subgroups have no significant difference in survival, but in their mean methylation, which has also been described before.¹⁵ Nevertheless, it does not cause the RTK I subgroup with significantly lower methylation to have shorter survival compared to the other two subgroups. In contrast, a combination of all 4 prognostic factors (age, extent of resection, MGMT promoter methylation, and mean global DNA methylation) is more critical to predict survival whereby a higher number of risk factors was associated with worse survival.²⁸ The combinations are also relevant; hence, the OS of elderly patients is better with low mean methylation than poor extent of resection. In addition, the extent of resection in combination with another prognostic marker had a strong influence on OS; therefore, these patients were classified in risk group 3.

To investigate possible causes of mean global DNA methylation as a prognostic marker, we correlated it with the number of immune cells. Analysis of immune cell infiltrates revealed a correlation between mean methylation and DIME scores (Figure 3D–F). However, the scores are not sufficient for a comparable division of the cohort into longer and shorter surviving patients (Figure 3A–C). Mostafa et al. shows, using 51 glioblastoma patients,²³ that an increased number of CD8⁺ lymphocytes is associated with better OS. We also see that a high number of CD8⁺ immune cells correlated with a better OS of GBM (Figure 3B). Contrary to those results, Zuo et al. analyzed the TCGA database and showed that a high immune cell count is associated with worse OS in GBM.²⁹ They examined 46 immune cell types. This clearly indicates that more analyses need to be performed for exploring the correlation between immune cell infiltrates and OS to fully understand the interplay. Our analyses show a significant correlation of OS to immune cell infiltration, with higher immune cell infiltration associated with better survival. The correlations are significant but not as strong as the correlation of mean methylation with survival. Causality of mean methylation due to more immune cells cannot be discerned from our results. In addition, the analysis of cell populations in terms of mean methylation and survival may also be promising, as a higher proportion of myeloids correlates with higher methylation. Nevertheless, a direct causal relationship cannot yet be observed. Moreover, in addition to deconvolution, further analyses on single cell or histological level should be included in the future, as it is known that deconvolution of bulk data have strong limitations in accuracy. The analyses of cell populations in the samples are crucial for further understanding of the biological causality regarding mean methylation. Biological causality also needs further investigation to explain why mean methylation is not relevant in the mesenchymal subset but is in RTK I and RTK II.

Another possible explanation for the correlation of survival and mean methylation is a better response to radiotherapy, meaning that tumor cells with higher mean DNA methylation are more radiosensitive. This is supported by the better survival of patients treated with radiotherapy alone with a higher mean methylation (Figure 4E–F). Although significant, the difference in survival was not particularly large. This could be explained by a poor performance status of at least some patients, since clinical parameters such as age or health status might be crucial reasons why patients were not treated with chemotherapy or standard of care. Since irradiation varied between the two published cohorts (30 × 2 Gy or 10 × 3.4 Gy), it will be important in the future to study a large cohort of patients, who received the same fractionation and dose with respect to mean methylation.

Although higher radiosensitivity might be explained by immunological mechanisms, such as a high number of M2 macrophages,³⁰ it is more likely that differences between more resistant and more sensitive tumors are due to differences in the cellular radiosensitivity. This hypothesis is supported by the cell line experiments, which show a higher methylation in more radiosensitive cell lines. Cellular radiosensitivity is a highly complex phenotype, regulated not only by gene alternations but also by factors such as protein expression and availability, signal transduction, oxygenation, or chromatin structure.^31,32 The investigation of the relationship between cellular radiosensitivity and mean methylation is, therefore, outside the scope of this project and will be analyzed in a separate future project. Here, isogenetic cell lines with either high- and low-mean methylation will be helpful and are currently under development.

In conclusion, our study shows that survival of GBM patients was correlated with mean global DNA methylation. Patients with higher mean methylation survived longer, so further research is needed to understand the biological background and to provide new therapies to improve survival.

Funding

This work was supported by the Landesforschungsförderung Hamburg, LFF GK10. A.E. is thankful for the support within the interdisciplinary graduate school “Innovative Technologies in Cancer Diagnostics and Therapy” funded by the City of Hamburg. U.S. was supported by the Fördergemeinschaft Kinderkrebszentrum Hamburg. KJW received funding from the Mildred Scheel Young Investigator Program by Deutsche Krebshilfe. FR received funding from Illumina.

Conflict of interest statement

None declared.

Data availability

The raw methylation data we used are available at The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus database (GEO) under the accession numbers GSE60274, GSE195640, GSE60274, GSE195684, and GSE195640.