A new subtype of diffuse midline glioma, H3 K27 and BRAF/FGFR1 co-altered: a clinico-radiological and histomolecular characterisation

The first part of the cohort is composed of 60 patients diagnosed with H3-K27M or EZHIP-overexpressing tumours harbouring a coding mutation in BRAF or FGFR1 genes, from the Necker Enfants-Malades/GHU-Sainte Anne Hospital/Gustave Roussy center and Biomede 1 Trial (NCT02233049) (n = 29), or from other published cohorts (n = 31) [3, 18, 23, 25, 30, 31, 37]. The control cohort includes patients with wild-type FGFR1/BRAF, in the above-mentioned cohorts of H3K27M or EZHIP-overexpressing DMG (flowchart in Supplementary Fig. 1, online resource). Tumour tissue and clinical data were collected under informed consent obtained from the parents or guardian according to the IRB approved protocol (CNIL 1176643).

All available radiology outcomes at diagnosis (CT and MRI) of patients with H3.3-K27M tumours from our cohort were reviewed centrally by three experts (VDR, NB, and JG). Parameters specifically recorded were: radiological presentation (diffuse, circumscribed or nodular and diffuse), presence of contrast enhancement (yes/no) and calcifications (yes/no). For circumscribed tumours, the pattern at evolution was also analysed.

Formalin-fixed paraffin-embedded (FFPE) tissue samples for each patient were retrieved and Haematoxylin–Phloxine–Saffron (HPS)-stained slides were analysed by two experienced neuropathologists (PV and ATE) to confirm morphological diagnoses. Micro-calcifications were noted as present or absent as well as granular bodies, ganglion neurons, necrosis, and microvascular proliferation. Mitotic activity (per 2 mm²) and tumour growth architecture were analysed within the inherent limits of a stereotaxic biopsy exploration. The latter was labelled as diffuse, compact tumoral areas or both. Morphological aspects were evaluated as ganglioglioma-like (GG-like), HGG with piloid astrocytic component or DMG-like. The infiltration pattern was also assessed by NF70 immunostaining (i.e. residual NF70 network or not). Immunostaining was performed from 3 μm-thick representative FFPE sections using a Dako OMNIS automate. The following primary antibodies were used: H3K27me3 (1:2500, polyclonal, Diagenode), H3-K27M (1:5000, clone EPR18340, Abcam), Neurofilament Protein NF70 (1:100, clone 2F11), CD34 (1:40, clone QBEnd-10, Dako), BRAFV600E (1:100, clone VE1, Abcam), ATRX (1:200, polyclonal, Diagomics), Ki-67 (1:200, clone MIB-1, Dako). Antigen retrieval was performed at 95 °C, pH9 (GV80011-2, Dako) or pH6 (GV805, Dako). External positive and negative controls were used for antibody validation.

DNA and RNA were extracted from frozen tumours using Allprep DNA/RNA kit (Qiagen) and were quantified using, respectively, the Qubit Broad Range double-stranded DNA assay (Life Technologies) or the Qubit RNA high sensibility (Life Technologies). When no frozen material was available for DNA methylation profiling, tumour DNA was extracted from formol-fixed paraffin-embedded (FFPE) block sections using dedicated protocols at Diagenode or Integragen. Targeted DNA sequencing (≥ 6000× coverage) or whole exome sequencing (≥ 130× coverage) was performed as previously described [10, 41]. WES was aligned with BWA according to GATK best practice guidelines, then Mutect2 was used for the DNA calling. RNAseq on DMG primary tumours was performed at Integragen (Evry, France). PolyA mRNA molecules were initially purified from at least 100 ng total RNA (NEBNext® Poly(A) mRNA Magnetic Isolation Module, NEB) and libraries were then prepared using the NEBNext Ultra II Directional RNA Library Prep Kit (NEB). Paired-end reads of 100 bp were generated on an Illumina NovaSeq reaching an average sequencing depth of 60 million reads.

Genome-wide DNA methylation analysis was performed using either the Illumina HumanMethylation450 BeadChip (450 k) or EPIC arrays as previously published [2, 3]. Data were obtained from different platforms (DKFZ Heidelberg; Integragen; Diagenode; published data) and were analysed with R (v4.0.4). For t-Distributed Stochastic Neighbour Embedding (t-SNE) analysis, the minfi package was used to load idat files and preprocessed with the function preprocess.illumina for dye bias and background correction. Probes located on sex chromosomes or not uniquely mapped to the human reference genome were removed. Probes containing single-nucleotide polymorphisms or that were not present in both EPIC and 450 k methylation array were also eliminated. A batch effect correction was done with removebatchEffect function from limma package, to remove difference between formalin-fixed paraffin-embedded and frozen samples. The probes were sorted by standard deviation. The 10,000 most variable probes were used for subsequent clustering analysis and to compute the 1-variance weighted Pearson correlation between samples. The distance matrix was used as input in t-SNE from Rtsne package. For the second analysis, DNA methylation-based classification of CNS tumours from DKFZ-Heidelberg was used in order to predict the CNS tumour class based on the V12.7 of the classifier (www.​molecularneuropa​thology.​org).

Reads were pre-processed using the nf-core RNAseq pipeline (v3.0), mapped to the reference genome GRC38/hg38 with the STAR tool (v2.6.1d), annotated with GENCODE v36 and counting was performed with the Salmon quantification tool (v1.4.0). Differential gene expression analysis was performed with the DESeq2 package (v1.30.0, minReplicatesForReplace = 7, betaPrior = TRUE) with a threshold of 0.01 for Benjamini–Hochberg adjusted p value (adj-p). For gene set enrichment analysis (GSEA), hypergeometric tests were used to identify overrepresented gene sets from the MSigDB v7.4 database, amongst genes ranked by significance and fold-change in differential expression analysis, with Benjamini–Hochberg multiple testing correction using the package Clusterprofiler. Catalogues considered included Hallmark and C2. Differences were considered as significant when false discovery rate adj-q value was < 0.02. Gene expression comparison was evaluated with Wald test using DESeq2.

Overall survival (OS) was estimated with the Kaplan–Meier method and median overall survival was computed using a log-rank test. OS was obtained from the post-diagnosis until death patient or last known information. The analysis was realised in Prism9 software. Multivariable Cox proportional hazards regression model on OS was performed including histone H3, BRAF, FGFR1, TP53 status, age at diagnosis and tumour location with R software using the function coxph() of the survival package (Version 3.2–13).

Distribution of age at diagnosis according to different parameters was accessed by Mann–Whitney test. Presence of macro-calcification, contrast enhancement, radiologic profile and sex ratio were evaluated by Fisher’s exact test. Chi-square test for trend was used to evaluate tumour type and location. All statistic tests were performed using Prism 9 software (GraphPad).

We analysed a study cohort of H3K27-altered gliomas harbouring mutations in BRAF (n = 22), FGFR (n = 37), or both (n = 1; Supplementary Fig. 1 and Table 1, online resource). Of these, 43 were children (< 18 years) and 16 adults. In parallel, we analysed control cases of patients with BRAF^WT/FGFR^WT DMG H3-K27 from the corresponding cohorts. All DMG H3-K27M with BRAF alterations harboured a somatic V600E substitution, but no fusion (Fig. 1). Concerning FGFR, all cases displayed FGFR1 hotspot substitutions N546K/D (73%, 27/37) and/or K656E/M (23%, 11/37) preferentially occurring in CNS tumours, but neither FGFR1 fusion/duplication nor FGFR2/3 alterations (Fig. 1) [20]. All these mutations in BRAF and FGFR1 are widely known to induce an aberrant MAPK activation [12, 20, 27, 38, 43]. We estimated the overall prevalence of DMG BRAF^V600E and DMG FGFR1^MUT within the H3.3-K27M tumours in our composite paediatric–adult cohort at 7.6% and 12.3%, respectively. Thus, DMG BRAF^MUT/FGFR1^MUT together could represent close to 20% of all DMG H3.3-K27M.

We further analysed the genomic landscape of tumour for which material or data were available and observed that BRAF^MUT or FGFR1^MUT was mostly associated with H3.3-K27M mutation but not H3.1-K27M mutation. One tumour harboured a FGFR1^MUT in the context of a DMG H3-K27 wild-type with EZHIP overexpression presenting an ACVR1 mutation (case #60). Second, BRAF and FGFR1 mutations were mutually exclusive, as only one tumour presented both hits (case #23; Fig. 1). However, no clonality information was available for this case to confirm subclonality of the two mutations. Finally, we observed that in 90% (9/10) of FGFR1^MUT cases for which we had the information, FGFR1 mutation was clonal to H3-3A mutation with a similar variant allele frequency suggesting a role in the early steps of oncogenesis of these tumours. It was less frequently the case for patients with DMG H3-K27M BRAF^V600E with only 33% (2/6) tumours where BRAF mutation appeared clonal to H3-K27M mutation. In addition, H3-K27 FGFR1^MUT tumours presented often other hits in the MAPK pathway with NF1 (13/31; 42%) or PTPN11 in (3/22; 13.6%) as the topmost mutated genes (Fig. 1). Additional MAPK-activating mutations seemed to be less frequent for BRAF^MUT tumours with 1/12 (8.3%) case harbouring an NF1 mutation. TP53 mutations were found in 5% (1/20) and 9.3% (3/32) of the BRAF^MUT and FGFR1^MUT tumours, respectively but not PPM1D (Fig. 1). For one H3.3-K27M FGFR1^MUT patient, TP53 mutation was clonal to H3.3-K27M and it was sub-clonal for the other case for which information was available. In addition, ATRX was mutated, or its expression was lost on IHC in 59.3% (16/27) of FGFR1^MUT DMG, albeit never in BRAF^MUT tumours, and this was not associated to TP53^MUT in all but one case (#24). A mutation in a member of the PI3K/AKT/mTOR signalling pathway was present in 19% (4/21) of H3-K27M-FGFR1^MUT tumours (Fig. 1).

We next analysed the histo-pathological profile of cases of our own cohort (Supplementary Table 1). These 29 tumours were initially diagnosed as DMG H3K27 (n = 13), ganglioglioma grade 1 (n = 5), anaplastic ganglioglioma grade 3 (n = 4) or other gliomas (GBM n = 3; pilocytic astrocytoma n = 1; LGG n = 1 or oligo-astrocytoma grade 3 n = 2). All BRAF^MUT/FGFR1^MUT DMG presented a H3K27 trimethylation loss as expected and were positive for H3K27M staining with the exception of the case #60, harbouring EZHIP overexpression. Tumour growth pattern evaluated morphologically and based on NF70 immunostaining revealed that 12 tumours were only diffuse (n = 3 in FGFR1^MUT group and n = 9 in BRAF^MUT), 4 were only composed of compact tumoral tissue and 13 were mixed diffuse with nodular compact areas. We observed three major histological patterns: GG-like in 41.3% (12/29), DMG-like in 34.4% (10/29), and HGG with a piloid astrocytic component (piloid-HGG) in 24.3% (7/29). BRAF^MUT and FGFR1^MUT tumours presented a classic DMG histological profile only in 25% (4/16) and 46% (6/13) of cases, respectively (Figs. 1, 2, Supplementary Fig. 2a, online resource). A GG-like profile, characterised by mixed neuronal and glial tumour cells, was present in 50% (8/16) and 30% (4/13) of BRAF^MUT or FGFR1^MUT cases, respectively (Fig. 2b–i). Finally, 25% (4/16) of BRAF^MUT and 23% (3/13) of FGFR1^MUT DMG presented a piloid-HGG profile with piloid cell morphology (Fig. 2j, k). The remaining FGFR1^MUT DMG case (3/13) harboured a mixed DMG-like and GG-like or piloid-HGG, and the last GG-like and piloid-HGG profile (Fig. 2l). Micro-calcifications, albeit rare in classic DMG, were observed in 56% (9/16) of BRAF^MUT and 38% (5/13) FGFR1^MUT DMG (Fig. 2j). Ganglion cells and eosinophilic granular bodies were seen in 37% (11/29) and 31% (9/29) mainly in the BRAF^MUT group (7/16) (Fig. 2b). GFAP whorls defined as a small gliofilament tangle, represented a rare but particular aspect (n = 4) (Fig. 2m–o). Perivascular lymphoid infiltrates were frequent (48%, 14/29) and especially enriched in the BRAF^MUT group (62%, 10/16) (Fig. 2c). CD34 extravascular staining was observed in the majority of tested cases: 50% (4/8) in FGFR1^MUT and 60% (9/15) in BRAF^MUT. ATRX loss of expression was frequently seen in FGFR1^MUT cases tested (8/11) but absent in BRAF^MUT (0/16) (Figs. 1, 2l). The histopathological profile of BRAF^MUT or FGFR1^MUT DMG thus appeared very heterogeneous and distinct compared to classical DMG H3-K27M.

Radiologically, we observed that DMG H3-K27 BRAF^MUT or FGFR1^MUT were mostly diffuse but presented a significant enrichment of a mixed nodular-diffuse aspect compared to control DMG H3-K27 (Fig. 3a–b; (9/11) 82% and (5/9) 56% versus (4/47) 8%; p value =  < 0.0001, chi-square test for trend). Few tumours were completely circumscribed (Fig. 3a, b—(2/11) 18% and (2/9) 22% versus (0/47) 0%; p value =  < 0.0001, Chi-square test for trend), and radiological analysis at progression identified a diffuse evolution pattern in half (2/4). Second, these tumours were more often contrast enhancing (Supplementary Fig. 2b, online resource; (11/11) 100% and (9/9) 100% versus 47% (22/47); p value = 0.0013 and 0.0029, respectively; Fisher’s exact test). Third, they developed more calcifications than DMG H3-K27 (Fig. 3a, c; (6/11) 55% and (5/9) 56% versus (1/38) 3%; p value =  < 0.0001, Fisher’s exact test). Presence of micro-calcifications detected in histological analyses correlated with macro-calcifications seen by CT scan in 4/8 (50%) DMG BRAF^MUT and 2/5 (33.3%) FGFR1^MUT (Supplementary Fig. 3c, d, online resource). In total in histologic and radiologic analyses, DMG H3K27M BRAF^MUT or FGFR1^MUT was calcified in 8/11 (72.7%) and 6/9 (66.7%), respectively. To evaluate how radiological features could detect DMG with MAPK alterations, we classified tumours according to a ‘classic’ (diffuse) or ‘atypical’ (nodular-diffuse or circumscribed aspect and/or presence of macro-calcifications) radiological profile independently of their genotype. This showed that 91% (10/11) of DMG with BRAF^V600E and 78% (7/9) DMG with FGFR1^MUT H3K27M DMG are classified as atypical versus only 8.5% (4/47) classical DMG H3K27 (Supplementary Fig. 3e, online resource; p value =  < 0.0001, Fisher’s exact test).

Given the disparities between DMG_K27-BRAF/FGFR1 and classical DMG H3K27-altered, we hypothesised that DMG H3 K27-altered with MAPK-activating mutations might correspond to either (i) a new subtype of DMG or (ii) atypical aggressive MAPK-driven low-grade gliomas/glioneuronal tumours. To test these hypotheses, we analysed the DNA methylation profile of the whole cohort. Based on the Heidelberg DNA methylation brain classifier V12.8, 54% (7/13) of BRAF^MUT and 67% (10/15) of FGFR^MUT tumours classified as DMG_K27 and the remaining corresponded to other classes or were undefined (score < 0.9) (Supplementary Fig. 3a, online resource), which was consistent with the hypothesis that DMG H3 K27-altered with MAPK-activating mutation constitute a unique subtype of DMG. In contrast, all DMG H3.3-K27M BRAF^WT/FGFR^WT clustered as DMG_K27. We then performed an unsupervised clustering based on DNA methylation profiles of these samples together with reference gliomas from the literature [2]. All DMG H3-K27 BRAF^MUT or FGFR1^MUT but one (case #42) separated from DMG H3-K27 on the tSNE, independently of their H3 mutational status or brain classifier score (Fig. 4, and supplementary Fig. 3b, online resource). They also separated from midline BRAF^V600E LGG and GG grade 1 (Supplementary Fig. 3c, online resource). Furthermore, they were grouped in subclasses according to the nature of the secondary MAPK mutation in BRAF or FGFR1 (Fig. 4, and supplementary Fig. 3b, online resource).

TP53 mutations and BRAF^V600E/FGFR1 mutations are mostly mutually exclusive in DMG. In order to avoid the confounding effect of TP53 mutations on the outcome of H3.3-K27M DMG without MAPK-activating alterations, we stratified patients on Histone H3 and TP53 genotypes of both subgroups in Kaplan–Meier overall survival (OS) analyses. This analysis showed a significant better OS for paediatric and adult DMG H3-K27 patients with activating BRAF (median OS 37 mo.) or FGFR1 mutations (median OS 36 mo.) compared to DMG H3.3-K27M TP53^WT (median OS 12 mo.) and other DMG subtypes (Fig. 5a; p value < 0.0001, global log-rank test). Further, we showed that there was no impact of histopathological features such as microvascular proliferation, necrosis or mitotic index, on the OS of patients with BRAF or FGFR1-mutated DMG (supplemental Fig. 2f-h, online resource). To pursue, we performed a multivariable analysis to evaluate the association of Histone H3, BRAF, FGFR1 and TP53 mutational status, age at diagnosis and tumoral location with survival. As expected from previous publications, histone H3 and TP53 status were significantly associated with OS (Table 1) [4, 41]. Age was significantly associated with prognosis, but its overall impact was only marginal compared to other variables. We also showed that BRAF^V600E (HR: 0.2132, 95% CI 0.1098–0.4140, p value = 5.03e-06) and FGFR1^MUT (0.3414, 95%CI 0.1963–0.5939, p value = 0.0001) are strong and independent prognostic markers in H3K27-altered DMG (Table 1).

We next compared other clinical parameters. The sex ratio was balanced in DMG with MAPK-activating mutations (Supplementary Fig. 4a, online resource). However, we identified a significant difference in age at diagnosis between H3.3-K27M DMG, BRAF^MUT and FGFR1^MUT DMG (Fig. 5b). More precisely, H3.3-K27M BRAF^V600E DMG developed only in children (< 18 years) with a median age at onset of 7.2 years lower than DMG H3.3K27M with 9.85 years (Fig. 5b; p value = 0.0123, Mann–Whitney test) but similar to DMG H3K27-altered paediatric cases with 7.6 years in a restricted paediatric cohort (supplementary Fig. 4b, online resource). In contrast, patients with FGFR1^MUT have a significant higher age at diagnosis compared to those FGFR1^WT with median of 14.8 and 9.8 years respectively (Fig. 5b; p value = 0.0001, Mann–Whitney test). The age at diagnosis of H3-K27M DMG can vary according to initial tumour locations, with a higher age of onset for thalamic versus pontine tumours [18, 34]. We found this difference of age according to tumour location in DMG H3.3-K27M and in DMG BRAF^V600E but not in DMG FGFR1^MUT (Supplementary Fig. 4c, online resource). Finally, DMG with BRAF and FGFR1 mutations were significantly more frequent in the thalamus compared to DMG H3.3-K27M (Fig. 5c; 70% (16/23) and 58% (21/36), respectively versus 28% (55/199); p value = 0.0004, and p value < 0.0001, Chi-square test for trend). No variation in term of OS, age at diagnosis and tumour location was noted according to FGFR1-mutated variant: FGFR1^N546K/D versus FGFR1^N656E (Supplementary Fig. 5, online resource).

We next wondered if the sequence of appearance of mutations was identical in these tumours, and more largely if they correspond to (i) DMG (H3-K27M as a first hit) or (ii) atypical aggressive low-grade gliomas (BRAF^V600E as first hit). To address this question, we analysed BRAF copy number variation (CNV) by digital droplet Polymerase Chain Reaction (ddPCR) on genomic DNA from one BRAF^V600E H3.3-K27M tumour (case #2) and from a H3.3-K27M BRAF^WT clone derived from this primary tumour in vitro (Supplementary Fig. 6, online resource). The BRAF^WT clone had two BRAF alleles and thus resulted from the re-amplification of an ancestral H3.3-K27M-only clone, but not from a genetic loss of the BRAF^V600E allele, demonstrating that, in one DMG H3.3-K27M BRAF^V600E, H3.3-K27M was the first hit. Interestingly, DNA methylation profiles from this in vitro amplified BRAF^WT H3.3-K27M ancestral clone from case #2 and also from a tumour relapse enriched in H3.3-K27M BRAF^WT clone from patient #7 (initially diagnosed with DMG H3.3-K27M BRAF^V600E), clustered with DMG H3-K27 with MAPK alterations instead of classical DMG H3-K27 (Supplementary Fig. 3b, online resource).

In order to investigate the possible specificities of the DMG H3.3-K27M with BRAF^MUT or FGFR1^MUT, we compared their transcriptome to regular DMG H3.3-K27M TP53^WT separately. We found 676 significantly up-regulated and 633 down-regulated genes in the contrast H3.3K27M BRAF^WT versus H3.3-K27M BRAF^V600E DMG and 228 up-regulated and 274 down-regulated genes in the contrast H3.3-K27M FGFR1^WT versus H3.3-K27M FGFR1^MUT DMG (adj. p value ≤ 0.001) (Supplementary Fig. 7a, b, online resource). Ninety-four up-regulated and 111 down-regulated genes were common to the two comparisons. Using gene set enrichment analyses (GSEA), we observed an enrichment for MAPK signalling and PI3K/AKT/MTOR signalling signatures in both BRAF^MUT and FGFR1^MUT DMG (Fig. 5d; Supplementary Fig. 7e, f, online resource) as well as angiogenesis and hypoxia signatures (Supplementary Fig. 7 g-j, online resource). In addition, transcriptomic signatures highlighted activation of senescence and P53 signalling pathway in both comparisons (Fig. 5d). P53 protein is a tumour suppressor implicated in the permanent cell cycle arrest by inducing senescence or apoptosis in response to stress like oncogene activation [6, 14]. TP53 pathway activation was validated at the protein level by immunohistochemistry (IHC) in 71% (10/14) of DMG H3.3-K27M TP53^WT with BRAF/FGFR1 alterations, showing heterogeneous, weak to strong TP53 staining. Moreover CDKN1A, encoding the senescence marker P21, was overexpressed at the RNA level in both BRAF^MUT and FGFR1^MUT gliomas compared to those only H3.3-K27M mutated (adj. p value < 0.0001) (Supplementary Fig. 7 k online resource). CDKN2A, which encodes the tumour suppressor P16, was overexpressed only in BRAF-mutant gliomas (Supplementary Fig. 7 l online resource; adj. p value = 0.0048). As a whole, based on transcriptome and IHC, DMG H3-K27 with BRAF^MUT/FGFR1^MUT were characterised by a senescence programme, likely induced trough a P16/P21-P53 axis.