Sample cohort

To investigate biological associations underlying malignancy of IDH-mutant astrocytomas, four large multi-domain and multi-centre omics datasets were leveraged: CATNON, TCGA, GLASS-NL primary (GLASS-NL-P) and GLASS-NL recurrent (GLASS-NL-R) [1, 7, 41]. We obtained DNAm (CATNON: n = 430, TCGA: n = 256, GLASS-NL-P: n = 98, GLASS-NL-R: n = 137) and DNA-sequencing (CATNON: n = 424, TCGA: n = 253, GLASS-NL-P: n = 97, GLASS-NL-R: n = 133) data for each of the studies. RNA-sequencing of primary IDH-mutant astrocytomas included in the CATNON trial was successfully performed for 138 samples and extended with transcriptomic data from the TCGA (n = 247), GLASS-NL-P (n = 65) and GLASS-NL-R (n = 102) datasets.

The median age at diagnosis was significantly lower in the GLASS-NL (32 [18–70]) dataset in comparison to both CATNON (37 [18–82], p = 5.92e−07) and TCGA (37 [14–74], p = 9.95e−06). The percentage of primary samples with WHO CNS5 grade 4 was higher in CATNON (13%), compared to GLASS-NL (9%) and TCGA (8%).

Continuous grading coefficient as a measure for grading/malignancy

For all datasets, methylation data were uploaded into the DNAm-based CNS-tumour classifier for classification and copy number variation estimations. Three samples from the TCGA dataset exhibited 1p/19q codeletion (TCGA-CS-5394, TCGA-FG-7637, TCGA-VM-A8CA) and were therefore removed from all analyses (Supplementary Fig. 1a).

The vast majority of the samples were classified as IDH-mutant astrocytoma (CATNON: 412/430, TCGA: 244/253, GLASS-NL-P: 93/98, GLASS-NL-R: 116/137). The highest fraction of A_IDH_HG samples were observed in the CATNON and GLASS-NL-R datasets (CATNON: 98/412, TCGA: 25/244, GLASS-NL-P: 4/93, GLASS-NL-R: 45/116) (Supplementary Fig. 1a). This aligns with the inclusion criteria of CATNON and the higher proportion of progressed recurrent IDH-mutant astrocytoma in GLASS-NL [44].

Samples not classified as A_IDH_LG/A_IDH_HG were most often classified as oligodendroglioma (O_IDH, n = 11) and oligosarcoma (OLIGOSARC_IDH, n = 15) (Supplementary Fig. 1a/b). Oligosarcoma is not recognised as a distinct tumour type by WHO CNS5, but represents oligodendroglioma with mixed oligodendroglial and sarcomatous morphology [31, 36]. Oligosarcoma forms a unique distinct methylation class that has been incorporated in the latest version (v12.8) of the CNS-tumour classifier [36]. Samples classified as oligosarcoma did not harbour 1p/19q codeletion and had a lower tumour-purity estimate compared to other IDH-mutant methylation subclasses (Supplementary Fig. 1c, Supplementary Methods). TERT promoter mutations were absent in TCGA and CATNON samples classified as oligosarcoma [36]. Moreover, four GLASS-NL cases, classified as astrocytoma (A_IDH_LG) in the first resection, were later classified as oligosarcoma in their matched recurrent samples. All samples classified as oligosarcoma harboured astrocytoma-like features, including ATRX and/or TP53 mutations.

We utilised the calibrated classification probabilities derived from the CNS-tumour classifier to generate a DNAm-based continuous grading coefficient (CGC). Formally, the CGC is calculated as the natural logarithm of the calibrated classification probabilities between A_IDH_LG and A_IDH_HG. We then defined three revised astrocytoma subtype classes (low: CGC < − 4.5, medium: CGC [− 4.5 to 4.5], high: CGC > 4.5) based on the relation between the CGC and the calibrated probability scores from the CNS-tumour classifier in CATNON (Fig. 1a). Clinical (age, sex, treatment), histological (necrosis and/or microvascular proliferation) and molecular (CNS-tumour classifier class, CNV load and CDKN2A/B HD) characteristics of patients within each CGC class are presented in Supplementary Table 1.

As may be expected, CGC class was positively associated with the number of samples being classified as WHO CNS5 grade 4 (CATNON: p < 0.0001, TCGA: p < 0.0001, GLASS-NL: p < 0.0001, Fisher Exact Test, Fig. 1b, Supplementary Table 1). However, a substantial proportion of WHO CNS5 grade 4 tumours were present in the CGC low and medium subgroups. Similarly, not all CGC-high tumours were WHO CNS5 grade 4. Our CGC subgroups were strongly associated with OS: median OS in CGC-low not reached (95% CI [8.2–not reached]), with CGC-medium 6.9 years (95% CI [5.7–not reached]) and CGC-high 3.4 years (95% CI [3.0–not reached]). All three CGC subgroups showed significantly different OS (CGC-medium vs CGC-low: HR 1.90, 95% CI [1.31–2.76]; p < 0.001, CGC-medium vs CGC-high HR: 2.12 95% CI [1.33–3.36]; p = 0.001). Importantly, CGC cut-off values determined on CATNON also showed prognostic significance in TCGA (p < 0.0001) and GLASS-NL-R (p < 0.005) (Fig. 1c). In these independent datasets, the CGC subgroups were significantly associated with CDKN2A/B HD (p < 0.0001, Fisher Exact Test, Supplementary Table 1).

Multivariable Cox PH-regression analysis on CATNON showed that the CGC subgroups were an independent prognostic factor (CGC-medium vs CGC-low: HR: 1.68, 95% CI [1.15–2.46] and CGC-high vs CGC-medium: HR: 3.42, 95% CI [2.12–5.51]) when adjusted for age, sex, CDKN2A/B HD, necrosis and/or microvascular proliferation and treatment with adjuvant/concurrent temozolomide (Fig. 1d). The CGC was also significantly associated with survival in multivariate analyses, outperforming WHO CNS5 in both the TCGA and GLASS-NL datasets (Supplementary Fig. 2, TCGA: p = 0.003, GLASS-NL: p = 0.007, Wald test).

Fig. 1

Evaluation of the three DNAm-based CGC-based subgroups (low, medium, high) across CATNON, TCGA and GLASS-NL. a Cut-off values for the three CGC groups (low: CGC < − 4.5, medium: CGC [− 4.5, 4.5], high: CGC > 4.5) which were determined based on the association between the CNS-tumour classifier probability score (A_IDH_HG) and the CGC. b Bar plot depicting the fraction of samples within each CGC subgroup according to WHO CNS5. c Kaplan–Meier overall survival curves stratified by CGC subgroup. p values were determined by log-rank test. d Survival forest plots showing the results of Cox proportional hazard models for CGC subgroups corrected for age, sex, WHO CNS5 criteria (CDKN2A/B HD and histological features) and treatment arms according to the CATNON trial. e Unsupervised principal component 1 and 2 of DNAm data demonstrates spatial segregation of IDH-mutant astrocytomas. CNS-tumour classifier subtype assignment (A_IDH_LG/A_IDH_HG) and CGC subgroups are indicated

High CNV load is associated with worse outcome in IDH-mutant astrocytoma [34, 37]. We found that CGC groups were significantly associated with CNV load in all datasets (p < 0.0001, Fisher Exact Test, Supplementary Table 1). CNV load was also associated with OS in a univariate analysis (CATNON: HR: 1.42 95% CI [1.02–1.97]; p = 0.038, TCGA: HR: 2.42 95% CI [1.38–4.23]; p = 0.0021, GLASS-NL-R: HR: 2.39 95% CI [1.45–3.96]; p < 0.001). Earlier work reported worse survival for gliomas with heterozygous deletion of CDKN2A/B [17]. To further investigate this, we manually assessed CNV profiles in the CATNON dataset, blinded to all other clinical and molecular parameters. This manual inspection was necessitated by the difficulty in distinguishing heterozygous deletion using conservative cut-off values. We identified 7 cases with a heterozygous deletion of CDKN2A/B. Patients without a deletion had a median OS of 9.45 years (95% CI [7.52–not reached]), whilst those with a heterozygous deletion showed a median OS of 4.46 years (95% CI [2.97–not reached]). Patients with CDKN2A/B HD conferred worse survival, with a median OS of 3.11 years (95% CI [2.83–5.60]). However, limited by the insufficient sample size, there was no significant difference in survival between patients with a heterozygous deletion and those with wild-type CDKN2A/B (p = 0.12, log-rank test).

We wondered to what extent our CGC captured the overall variability of DNAm profiles in our samples. To test this, we summarised the global DNAm profile of samples classified as IDH-mutant astrocytoma by PCA on the 10,000 most variable probes. Summarising the DNAm profile by our three CGC subgroups captured the global methylation profile better compared to A_IDH_LG and A_IDH_HG (Fig. 1e).

The CGC for samples in the CATNON, TCGA, GLASS-NL-P, GLASS-NL-R datasets and common copy number events are illustrated in Fig. 2a. We observed that the spatial distribution of methylation profiles along PC1 and PC2 was captured more effectively by the CGC for all datasets, emphasising the efficacy of a continuous approach (Fig. 2b). High CGC values were not restricted to CDKN2A/B HD tumours alone (Fig. 2a/c) and we next explored associations between the CGC and copy number events (Supplementary Fig. 3). In addition to CDKN2A/B HD (CATNON: p = 6.60e−07, TCGA: p = 6.78e−05, GLASS-NL: p = 1.88e−08), we also found associations with SMARCA2 HD (CATNON: p = 7.70e−05, TCGA: p = 3.62e−04, GLASS-NL: p = 6.68e−04), RB1 HD (CATNON: p = 6.40e−04, TCGA: p = 0.033, GLASS-NL: p = 0.019), PTEN HD (CATNON: p = 2.73e−06, TCGA: p = 0.12, GLASS-NL: p = 1.69e−03), CDK4 amplification (CATNON: p = 5.15e−05, TCGA: p = 0.016, GLASS-NL: p = 0.046) and PDGFRA amplification (CATNON: p = 5.15e−05, TCGA: p = 0.019, GLASS-NL: p = 0.024).

Fig. 2

Continuous Grading Coefficient (CGC) as a tool to study malignancy in IDH-mutant astrocytoma. a Samples of individual cohorts (CATNON, TCGA, GLASS-NL) ranked according to their CGC. As can be seen, specific copy number of events (RB1 HD, PDGFRA amplification, CDK4 amplification and CDKN2A/B HD) and overall survival were correlated with the CGC. Samples are coloured based on astrocytoma subtype assignment (blue: A_IDH_LG, red: A_IDH_HG). b Unsupervised PC1 and PC2 on the DNAm data showing spatial segregation of samples classified as IDH-mutant astrocytoma. c Distribution of the CGC based on CDKN2A/B HD status across the different datasets. Samples with CDKN2A/B HD showed a significantly higher CGC. p values determined by Wilcoxon signed-rank test

Global decrease in DNA-methylation and hypermethylation of CpG islands associates with continuous grading coefficient

To elucidate which CpG sites associate with malignant transformation as defined by the CNS-tumour classifier, we applied DMP analysis on the CGC using linear regression modelling for each dataset separately. In total, 8% of all tested probes were differentially methylated in the CATNON dataset and 8% in the TCGA dataset. In both datasets, the vast majority of differentially methylated probes were all hypomethylated in higher grade malignancies (CATNON: 99%, TCGA: 98%). Interestingly, although the vast majority of CpG sites had decreased methylation levels, those present on CpG islands were more often hypermethylated than hypomethylated (CATNON: p < 2.2e−16, TCGA: p < 2.2e−16, Fig. 3a, Supplementary Table 2). This observation is further supported by the finding that a larger fraction of hypermethylated probes belonged to the TSS200 region (Supplementary Fig. 4). Hypomethylated probes were typically found in the open sea, located more than 4 kb away from CpG islands. Focussing on DMP present in both the 450k (TCGA) and 850k (CATNON) array, we found that ~ 60% of the hypermethylated probes were shared.

Fig. 3

Supervised DNAm and RNA analysis on CATNON and TCGA with validation on the GLASS-NL dataset identifies gene clusters associated with malignancy. a Distribution of probes belonging to CpG islands, shelfs (N_Shelf, S_Shelf), shores (N_Shore, N_Shelf) and the open sea across all hypomethylated and hypermethylated probes. Distributions are displayed separately for the DMPs resulting from independent analyses conducted on CATNON (850k chip) and TCGA (450k chip). Genome-wide distribution of all probes is shown on the left as a reference. p values were determined by Fisher’s Exact Test. b Correlation between the DNAm-based signature scores of the hypermethylated and hypomethylated probes with the CGC in GLASS-NL (850k chip). c Spearman correlation between CGC and unsupervised PC2 of the transcriptomic data in all datasets. d Volcano plots showing the per-gene RNA log2FoldChange on the CGC and the corresponding FDR adjusted p value for both the CATNON and TCGA datasets. C0, C1, C2, and C3 genes are indicated. e Recursive-based correlation plot on VST expression of the differentially expressed genes in CATNON. Three upregulated (C1: green, C2: yellow, C3: red) and one downregulated (C0: brown) cluster were distinguished. Gene Ontology enrichment analysis resulted in significant hits for three clusters (C1: cell cycling, C2: embryonic development, C3: ECM). f Correlation between the DGE tests on the CATNON and TCGA datasets. For each gene the log2FoldChange divided by its standard error (Wald statistics) is indicated. g Spearman correlation between the RNA-based signature scores for C0, C1, C2, and C3 and the CGC in GLASS-NL.

We subsequently validated the overlapping DMPs identified in the TCGA and CATNON datasets on GLASS-NL. The median M-value of the overlapping hypermethylated (n = 149) and hypomethylated (n = 12,708) probes both showed a correlation with the CGC and resection number in the GLASS-NL cohort (hypermethylated: ρ = 0.57, hypomethylated: ρ = − 0.89, Fig. 3b). These correlations may be explained by an increased malignancy over time [44].

Distinct transcriptional features are associated with continuous grading coefficient

We included RNA-sequencing data of IDH-mutant 1p/19q non-codeleted CATNON and TCGA samples and findings were further validated on GLASS-NL-P and GLASS-NL-R. First, we conducted unsupervised PCA on the 1000 most variably expressed genes for each dataset independently. In all datasets, the primary source of variation (PC1) was likely associated with tumour purity as demonstrated by the expression of neuronal genes, with neuron marker genes like SLC12A5, TMEM130 and SV2B ranking among the top 50 contributors [25]. To confirm that the neuronal marker signature is associated with tumour purity, we used the top 50 most contributing genes of PC1 from each dataset and calculated the enrichment score for these genes. We projected this enrichment score onto our snRNA-seq data. These data indeed revealed that expression was primarily derived from neurons (Supplementary Fig. 5a/b). PC2 was however strongly associated with the DNAm-based CGC (CATNON: ρ = 0.67, TCGA: ρ = 0.50, GLASS-NL-P: ρ = 0.38, GLASS-NL-R: ρ = 0.75, Fig. 3c).

Gene-level differential regression analysis on samples from the CATNON and TCGA was then performed to find genes associated to the CGC. To ensure we are mainly investigating tumour-intrinsic signals, we estimated tumour purity using two methods and evaluated the outcomes by comparing their correlation with RNA expression levels of neuron markers (Supplement Methods, Supplementary Fig. 6a). We did not observe a significant difference in tumour purity based on CDKN2A/B HD status (CATNON: p = 0.46, TCGA: p = 0.39, Supplementary Fig. 6b). Also, there was no clear dependency of tumour purity on the outcome of the DGR analysis (CATNON: ρ = − 0.23, TCGA: ρ = 0.28, Supplementary Fig. 6c). Therefore no additional purity corrections were performed.

DGR on the CGC revealed a total of 650 differentially expressed genes in the CATNON dataset (Supplementary Table 2). The differential gene expression was skewed, with a higher proportion (77%) of genes showing increased expression in samples with a higher CGC (Fig. 3d). We then performed recursive correlation-based clustering [14] and identified four distinct gene clusters (Fig. 3e, C0/C1/C2/C3). The clusters that showed increased expression in more malignant samples (high CGC) were associated with specific biological functions as determined by gene ontology: C1 (cell cycling, n = 175), C2 (embryonic development, n = 176) and C3 (ECM, n = 97) (Fig. 3e). C0 (n = 149), which contained genes with a decreased expression in more malignant samples, could not be attributed to a biological function. DGR analysis on the TCGA dataset resulted in 667 differentially expressed genes (Fig. 3d), with a strong overlap with those identified with the CATNON analysis (58%). Also, the majority of the differentially expressed genes were upregulated (72%). Despite the differences in study designs, we observed a strong correlation across both datasets in outcome of the tests performed (ρ = 0.71, Fig. 3f).

Subsequently, we further examined the genes present within each of the transcriptional clusters. C1 contained genes associated with histones (e.g. H3C2, H2BC9, H2BC11), transcription factors (e.g. E2F1, E2F8, FOXM1), kinesins (e.g. KIF14, KIF15, KIFC1), DNA replication (e.g. RAD51, EXO1, CENPK) and cell cycle control (e.g. CHEK1, CCNB1/2, CDK6). Of note, established markers for cell cycling, such as MKI67 and TOP2A were found. C2 contained developmental transcription factor genes, such as members of the HOX (n = 18), PAX (n = 3) and TBX (n = 2) gene family. Within C3, genes associated with ECM formation were found, such as COL1A1 and COL1A2. On the contrary, C0 comprised many long non-coding RNA genes (n = 81). We compared expression levels of C0 genes amongst purified human CNS cell types (Supplementary Methods, Supplementary Fig. 7), and observed high expression levels in mature astrocytes as compared to foetal astrocytes (p < 0.001). Conversely, upregulated genes (C1-C3) were higher expressed in foetal astrocytes (p < 0.001). These combined findings suggest tumour dedifferentiation during malignant transformation.

We further examined genes present in our transcriptional clusters on protein level by IHC using DAB staining in CGC-low/medium (n = 5) and CGC-high (n = 5) samples. Quantification of Ki-67-positive cells revealed a significantly higher fraction of positive cells in the CGC-high group (Supplementary Fig. 8a, p = 0.03). HOXD10 was also selected for IHC since it emerged as one of the top hits in the DGR analysis. Whilst HOXD10-positive cells were detected, the staining exhibited high background signal and was predominantly cytoplasmic, which complicated accurate quantification. Nonetheless, HOXD10 was more prominent in CGC-high regions compared to CGC-low regions (Supplementary Fig. 8a).

PCA was conducted individually for each cluster (C0–C3) to derive a per tumour-sample representative value. This value summarised the expression pattern of genes within the respective cluster, as represented by PC1. Of note, the differentially expressed genes used to define this RNA signature were identified solely based on results from CATNON. When these values were calculated on the GLASS-NL dataset, we found correlations between the CGC and all RNA signature scores (C0: ρ = − 0.73, C1: ρ = 0.55, C2: ρ = 0.65 and C3: ρ = 0.52, Fig. 3g).

We utilised our snRNA-seq dataset (vanHijfte_2023), which included six matched primary (A_IDH_LG) and recurrent (A_IDH_HG) IDH-mutant astrocytoma samples from three patients (Fig. 4a). Uniform Manifold Approximation and Projection (UMAP) showed distinct cell types and the previously reported tumour cell states (astro-like, oligo-like and stem-like, Fig. 4a, Supplementary Fig. 5a). Interestingly, at tumour recurrence/high-grade, the oligo-like and astro-like cell states were less pronounced and a substantial tumour cell cluster could not reliably be defined using our marker gene set. We then overlaid our transcriptional signatures (C0-C3) with this dataset and found that C0, C1 and C2 were predominantly expressed in tumour cells (Fig. 4b/c). More specifically, C0 and C1 showed high expression in astro-like and proliferating tumour cells respectively [16, 45]. Genes from C3 were equally expressed by endothelial cells/pericytes and undetermined tumour cells (Fig. 4b/c). In line with our RNA-sequencing results, expression of C0 genes decreased and C1/C2/C3 increased during malignant transformation. In summary, our data suggest that malignant transformation of IDH-mutant astrocytoma is associated with an upregulation of cell cycling, embryonic development and ECM genes. This was accompanied by decreased expression of astro-like state genes.

Fig. 4

Single-nucleus RNA-sequencing (A_IDH_LG: n = 3, A_IDH_HG: n = 3) validates bulk DGE results and shows enrichment of gene clusters in select cell subpopulations. a UMAP projection illustrating cell-type annotations for all tumours combined. b Dotplot displaying the enrichment score of each of the bulk RNA-sequencing clusters (C0, C1, C2, and C3) identified in our DGR analysis. c UMAP projection showing enrichment scores of the downregulated (C0), cell cycling (C1), embryonic development (C2), and ECM (C3) clusters for A_IDH_LG and A_IDH_HG separately

Hypermethylation and upregulation of embryonic development genes in more malignant IDH-mutant astrocytoma

We next integrated the methylome differences with transcriptional changes of the corresponding genes (Fig. 5a). The differentially hypermethylated probes (n = 149) were associated with 63 genes from our RNA expression data (Supplementary Table 2). Pathway enrichment analysis on these hypermethylated genes revealed significant hits for high-CpG-density promoter (HCP) genes bearing the histone H3 trimethylation mark at K27 (H3K27Me3) in brain (FDR-adjusted p value = 4.8e−14).

Fig. 5

Supervised analysis reveals co-existence of upregulated and hypermethylated embryonic development genes. a Correlation between the DMP (x-axis) and DGR (y-axis) analyses across CATNON (850k chip) and TCGA (450k chip). b Log2FoldChange of the CGC on RNA VST expression across the HOXD (chr2), HOXA (chr7) and HOXC (chr12) loci. Chromosomal positions are indicated in megabases (Mb). c Log2FoldChange of the CGC on the DNAm M-values for the HOX loci (HOXA, HOXC, and HOXD) and the surrounding regions in CATNON (850k chip) and TCGA (450k chip). d Correlation between the embryonic development RNA signature scores (C2) and the median M-value of the hypermethylated probes. CDKN2A/B HD status is indicated in red

Comparison of the hypermethylated genes with transcriptional changes of the corresponding genes revealed a strong overlap with the upregulated embryonic development cluster (C2, n = 27/63 genes). Thus, although the vast majority of the genes were hypomethylated with increased malignancy/CGC, we found hypermethylated CpGs to be associated with genes that were, paradoxically to what may be expected, transcriptionally upregulated. These hypermethylated and upregulated C2 genes encode for key developmental transcription factor genes, such as members of the HOX (n = 11), PAX (n = 2) and TBX (n = 1) family of genes. When we further examined the expression of the HOX genes of C2, we observed gene-ordered differential expression along the different gene clusters, where genes positioned at the 5’ end of the loci showed a higher expression with higher CGC scores and gradually decreased downstream in both CATNON and TCGA (Fig. 5b). This suggests coordinated derepression of enhancer sequences up/downstream of the gene cluster. Indeed, regions surrounding the HOX loci were hypomethylated, supporting the hypothesis of derepression of enhancers (Fig. 5c). Importantly, in both the bulk and single-cell data increased HOX expression of those present in C2 was not restricted to tumours with CDKN2A/B HD (Supplementary Fig. 5c).

The per-sample median M-value of the hypermethylated probes correlated strongly with the C2 signature score on both the CATNON (ρ = 0.72) and TCGA (ρ = 0.73) datasets (Fig. 5d). We also found a modest correlation in GLASS-NL-P (ρ = 0.19), which may be expected due to the low number of high-grade primary samples. Conversely, a strong correlation was evident GLASS-NL-R (ρ = 0.60) (Fig. 5d).

Histopathological features associated with continuous grading coefficient

We investigated the correlation between thirteen histological features and the CGC through a multivariate linear model to gain insight into their association with tumour malignancy. For the CATNON dataset, thirteen histological features were scored by a panel of seven international neuropathologists: cell density, calcifications, increased number of blood vessels, microvascular proliferation, neoplastic-appearing astrocytes and oligodendrocytes, giant cells, (miniature) gemistocytes, nuclear pleomorphism, microcysts and mucoid degeneration, necrosis, and mitotic count [18]. Of these, the mitotic index (p = 0.0093) and giant cells (p = 0.042) were significantly associated with the CGC. This is in line with earlier work, which showed prognostic significance of the mitotic index [18]. We next correlated these variables with our RNA signatures and found that samples with a high mitotic index (> 2 mitoses per 10 40× consecutive high-power fields) and presence of giant cells had a higher C1 cell cycling (mitotic index: p = 0.015, giant cells: p = 0.029) and C2 embryonic development (mitotic index: p = 0.003, giant cells: p = 0.0019) signature value. In summary, our data revealed that increased cell cycling at the gene level is correlated with an elevated mitotic index.

Hypermethylation phenotype and high expression of embryonic development genes are associated with poor survival

We categorised samples into low- and high-risk groups for the C2 and hypermethylation signatures individually, based on the sign (negative or positive) of the C2 signature and PC1 associated with the M-value of hypermethylated probes (n = 149), respectively. Both high RNA expression of C2 genes (p < 0.0001, HR: 3.90 95% CI [2.13–7.13]) and the hypermethylation phenotype (p < 0.0001, HR: 2.42 95% CI [1.73–3.38]) were significantly associated with shorter survival in a univariate analysis on the CATNON dataset (Fig. 6a/b). These associations remained present after adjusting for age, sex, CDKN2A/B HD, microvascular proliferation/necrosis and adjuvant/concurrent temozolomide using a multivariable Cox PH model (C2 signature: p = 0.014, HR: 2.42 95% CI [1.20–4.87], hypermethylation phenotype: p = 0.0048, HR: 1.73 95% CI [1.18–2.52], Fig. 6c/d). These molecular signatures hold more prognostic value compared to the mitotic index (p = 0.053, HR: 1.66 95% CI [0.99–2.77]) in a univariate analysis.

Fig. 6

Survival analysis of CATNON based on RNA and methylation signatures. a/c Kaplan–Meier overall survival curves stratified by RNA C2 signature (a) and hypermethylation phenotype (c) risk groups. b/d Survival forest plots of predictive Cox proportional hazard models on the C2 signature (b) and hypermethylation phenotype (d) corrected for age, sex, CDKN2A/B HD, treatment and histology

The C2 RNA signature was associated with shorter survival in both TCGA (p < 0.0001, HR: 3.11 95% CI [1.7–5.6] and GLASS-NL-R (p = 0.0047, HR: 2.58 95% CI [1.30–5.11], Supplementary Fig. 9). The hypermethylation phenotype also showed prognostic significance in both TCGA (p < 0.0001, HR: 3.43 95% CI [1.93–6.11], and GLASS-NL-R (p = 0.012, HR: 1.92 95% CI [1.15–3.23], Supplementary Fig. 9).