Patients and tumour samples
Patients were selected based on classical clinical and radiological criteria, i.e. short clinical history (<3 months of symptom duration) and the presence of a pontine tumour infiltrating at least 50 % of the pons (suppl. Table S1; suppl. Fig S1) . All patients underwent systematic stereotactic or surgical biopsy at Necker Hospital (Paris, France). Diagnosis of glial infiltrative non-pilocytic neoplasm was histologically confirmed in all patients and we obtained snap-frozen tumour material from 91 children. A smear of each biopsy was performed before freezing and the presence of tumour cells was assessed before their use for the genomic analyses. Informed consent for the translational research programme was obtained from the parents or guardian according to the IRB approved protocol (number DC-2009-955 for tumour banking). Site of the biopsy (routinely the junction between the pons and the cerebellar peduncle where there was an hypersignal on FLAIR sequences) was checked on the post-biopsy imaging .
Biopsies from patients with pHGG in a non-brainstem location (n = 93) were obtained during the same period at Necker hospital. Their histone H3.1/H3.3 mutation status was determined by Sanger sequencing.
Histone H3 genes were analysed by direct sequencing of PCR-amplified products from tumour DNA using primers listed in suppl. Table 2.
Morphological sequences of the MRI were reviewed independently by three clinicians (neurosurgeon, neuroradiologist and oncologist). The following criteria were scored on the diagnostic MRI: location, contrast enhancement, ring contrast enhancement, large area of necrosis, cysts, presence of stripes in the infiltrated brain on T2/FLAIR sequences and tumour size. In case of discrepancy, definitive scoring was obtained by consensus. Response to radiotherapy was judged by the clinical improvement only in patients with stable or decreasing doses of steroids to avoid misinterpretation of the radiology due to pseudoprogression . Disease extension was registered throughout the follow-up (i.e. local only, loco-regional or metastatic). Patients without an MRI in the last 2 months were excluded, considering that data were incomplete.
For the diffusion maps acquired at diagnosis, regions of interest (ROI) were drawn over the T2 hyperintensity corresponding to the tumour in each slice, creating a volume of interest (VOI) carefully avoiding necrotic areas, by an experienced neuroradiologist blind to the clinical data. These VOIs were transferred to the co-registered diffusion maps and every voxel value was individually registered. Histograms were created from the apparent diffusion coefficient (ADC) and distributed diffusion coefficient (DDC) voxel data .
The performance in terms of prognostication of the recently published “DIPG survival model” was evaluated in comparison to other biological stratification. This score is based on the assessment of the following parameters: age at diagnosis, interval between first symptoms and diagnosis, presence of a ring enhancement and use of adjuvant chemotherapy in addition to radiotherapy .
Histology, immunohistochemistry (IHC) and FISH analyses
Tumours were histologically classified according to WHO 2007 criteria whenever possible. Emphasis was put on the presence of an oligodendroglial component in the tumour cells (morphology, negativity of the tumour cells for vimentin, positivity for OLIG2), presence of interstitial oedema (semi-quantitative) and the presence or absence of necrosis.
A systematic panel of IHC markers was routinely performed: OLIG2, vimentin, GFAP, p53 (DO-7), PTEN, EGFR and MIB-1 as previously described . Additional stainings were developed to detect the loss of nuclear expression of the trimethylation mark at position K27 of the histone 3 tail (1:1000, polyclonal rabbit antibody, Diagenode, Belgium), the nuclear expression of the K27M form of histone H3 (1:1000, polyclonal rabbit antibody, Millipore, CA) and loss of ATRX nuclear expression (1:200, polyclonal rabbit antibody, Sigma-Aldrich, MO).
PDGFRA gene copy number was assessed by fluorescent in situ hybridization (FISH) using PDGFRA/CEN4 Dual Color Probe (Abnova, Tapei, Taiwan) on interphase nuclei. In brief, four-micron sections of tumour were mounted on SuperFrost Plus slides (Erie Scientific CA., Portsmouth, NH) and the probed area determined in accordance with haematoxylin and eosin-stained section. The sections were deparaffinised in xylene, rehydrated through an ethanol series, air-dried and incubated in pre-treatment solution (1 M NaSCN-Tris) at 80 °C for 25 min. Slides were then treated with a 0.01 % pepsin solution (Sigma-Aldrich, Saint Louis, USA) at 37 °C for 8 min. After dehydration, 10 µl of probe mixture was applied to each sample, slides were coverslipped and co-denatured at 75 °C for 5 min and hybridized at 37 °C for 48 h using thermobrite system (Leica Biosystems, Richmond, IL). A post-hybridization wash was performed in 2 × SSC at 73 °C for 2 min. Preparations were dehydrated and counterstained with 4,6-diamidino-phenyl-indole (DAPI). Signals were scored in at least 100 non-overlapping interphase nuclei. PDGFRA gene amplification was considered as positive in (A) specimens that have ≥40 % of cells displaying ≥4 copies of the PDGFRA signal, (B) specimens that display PDGFRA gene amplification, according to one of the following criteria: (a) a PDGFRA to CEN4 ratio ≥2 over all scored nuclei and calculated using the sum of PDGFRA divided by the sum of CEN4 when mean CEN4 per cell is ≥2 copies; (b) the presence of gene cluster (≥4 spots) in ≥10 % of tumour cells; (c) at least 15 copies of the PDGFRA signals in ≥10 % of tumour cells. Results were recorded using a DM600 imaging fluorescence microscope (Leica Biosystems, Richmond, IL) and digital imaging software (Cytovision, v7.4).
Genomic and statistical analysis
Gene expression (GE) profiling and comparative genomic hybridization on array (aCGH) were conducted for patients with enough material available of required quality on an Agilent platform. For all statistical analysis, the level of significance was 5 %.
Gene expression microarrays analysis
The data analysed is the result of the gathering of DIPG samples belonging to three different cohorts of young patients with high-grade glioma. In such experimental design, a well-known undesired bias is the batch effect, which is purely technical. To correct for this effect, we replicated some samples in at least two of three batches and performed the ComBat analysis  as recommended in Chen et al. . A normal brainstem sample from commercial source, hybridized on chips in the three batches, was used as a reference to normalize the data for the intensity bias as recommended in Do et al. . The normalization of fluorescence intensities is performed in two steps, inspired by Bolstad et al. . The first one is a loess normalization using the normal brainstem samples as a common baseline array for all arrays in the cohort. The second step is a quantile normalization.
Differential analysis was then performed using a moderated t test, comparing mean of log2(intensities) in both H3.1- and H3.3-mutated DIPG samples, implemented in the limma R package. The null hypothesis H0 for each gene is that the mean of log2(intensities) is the same in both H3.1- and H3.3-mutated DIPG samples. The alternative hypothesis H1 for each gene is that they are different. The obtained p values were adjusted for multiple testing using the Benjamini–Hochberg procedure. The a priori defined level of significance was 5 % after correction for multiple testing.
Comparative genomic hybridization array analysis
Log2(ratio) between raw signals of the reference DNA and DIPG DNA was first normalized for the dye and intensity effects and also for the local GC % content, using loess procedure, implemented in the limma R package . Data were then centred according to normalized log2(ratio) distribution values by an in-house script using the EM (expectation maximization) approach and aberrations status calling was automatically performed. Normalized centralized values were then segmented using the Circular Binary Segmentation (CBS) algorithm , implemented in the DNAcopy R package. The normal copy number interval [log2(ratio) = 0 which means 2 DNA copies as in the DNA reference] was calculated for each sample by multiplying with a define factor the median absolute deviation (MAD) of the normalized data for each single sample.
Survival curves comparisons
Survival functions were estimated with the Kaplan–Meier method and all survival function estimate comparisons were performed using a log-rank test. The null hypothesis H0 was that the two considered survival function estimates were the same. The alternative hypothesis H1 was that they were different.
Multivariate survival analysis
The multivariate survival analysis was conducted on H3.1- or H3.3-mutated patients only. The other patients (wild type or H3.2 mutated) were excluded from this analysis. The pool of initial covariates to include in the Cox model was: (1) H3-variant mutation, (2) presence of metastasis, (3) MRI contrast enhancement, (4) treatment type and (5) DIPG Risk score. We first checked for multicollinearity. H3-variant mutation and presence of metastasis are correlated (cor = 0.32) as well as MRI contrast enhancement and the DIPG Risk score (cor = 0.56). We started the analysis with 4 different pools of covariates to separate the correlated pairs of variables and then performed a backward stepwise variable selection using the pec R package. At the end of the procedure, two models retained no variables while the other retained two (‘H3-variant mutation’ and ‘DIPG Risk score’) and three variables (‘H3-variant mutation’, ‘treatment type’ and ‘MRI contrast enhancement’), respectively. For both models, the coefficient for ‘H3-variant mutation’ was 1.4109 and 1.6389, respectively, while the other coefficients were much lower (DIPG Risk score coef = 0.0873, radiotherapy coef = 0.1101, Tarceva coef = −0.4963, Temodal coef = 0.5126, MRI contrast enhancement coef = 0.6573). The hazard ratios were 3.98 and 5.00, respectively, and the confidence intervals were (0.25; 0.92) and (0.67; 1.93), respectively. The log-likelihood p values were 3.39e−5 and 7.69e−5, respectively. The H3-variant mutation is then the most important variable in the multivariate Cox model.
Distribution analyses (diffusion data and age at diagnosis)
To test if both ADC and DDC values were drawn from the same distribution for H3.1- and H3.3-mutated patients, a Mann–Whitney test was performed in each case. The null hypothesis H0 is that ADC (or DDC, respectively) values are drawn from the same distribution law for both H3.1- and H3.3-mutated patients. The alternative hypothesis H1 is that they follow two different distribution laws.
The same methodology was adopted to compare distributions of age at diagnosis between H3.1- and H3.3-mutated patients.
To compare sex ratios in H3.1- and H3.3-mutated patients, we performed a Chi-squared proportion test as none of the theoretical headcounts was below 5. The null hypothesis H0 was that the male–female proportions are the same in the two considered groups. The alternative hypothesis H1 was that they are different.
For the response to radiotherapy comparison, there was one theoretical headcount below 5, so we performed a Fisher exact test instead of the Chi-squared proportion test. The null hypothesis H0 was that the proportions of good and bad response to radiotherapy were the same in the two considered groups.