Drug screening linked to molecular profiling identifies novel dependencies in patient-derived primary cultures of paediatric high grade glioma and DIPG

Paediatric high grade glioma and diffuse midline glioma (including DIPG) are comprised of multiple biological and clinical subgroups, the majority of which urgently require novel therapies. Patient-derived in vitro primary cell cultures represent potentially useful tools for mechanistic and preclinical investigation based upon their retention of key features of tumour subgroups under experimental conditions amenable to high-throughput approaches. We present 17 novel primary cultures derived from patients in London, Dublin and Belfast, and together with cultures established or shared from Barcelona, Brisbane, Rome and Stanford, assembled a panel of 52 models under 2D (laminin matrix) and/or 3D (neurospheres) conditions, fully credentialed by phenotypic and molecular comparison to the original tumour sample (methylation BeadArray, panel/exome sequencing, RNAseq). In screening a subset of these against a panel of ~400 approved chemotherapeutics and small molecules, we identified specific dependencies associated with tumour subgroups and/or specific molecular markers. These included MYCN-amplified cells and ATM/DNA-PK inhibitors, and DIPGs with PPM1D activating truncating mutations and inhibitors of MDM2 or PARP1. Specific mutations in PDGFRA were found to confer sensitivity to a range of RTK inhibitors, though not all such mutations conferred sensitivity to targeted agents. Notably, dual PDGFRA/FGFR and downstream pathway MEK inhibitors showed profound effects against both PDGFRA-sensitising mutant and FGFR1-dependent non-brainstem pHGG and DIPG. In total, 85% cells were found to have at least one drug screening hit in short term assays linked to the underlying biology of the patient’s tumour, providing a rational approach for individualised clinical translation.


INTRODUCTION
Paediatric high grade glioma (pHGG), including pontine and other diffuse midline glioma (DMG), is a remarkably heterogeneous collection of diseases, comprised of multiple biological subgroups arising in distinct anatomical locations within the central nervous system (CNS), and at different ages 1 . In almost all instances, the clinical outcome remains extremely poor, with a median overall survival of 9-18 months 2 . Making progress in this disease has been hampered historically by a lack of primary patient material for study, firstly in terms of defining the underlying biology, and latterly in terms of having appropriate model systems to develop novel therapeutic approaches. As progress has advanced rapidly in recent years to further refine our understanding of the critical genetic and epigenetic drivers of these diverse subgroups [3][4][5][6] , it has become increasingly apparent that reappraisal of pHGG clinical trial structure is needed to evaluate novel agents in the appropriate patient populations 7 , and that subgroup-specific models are required to generate the necessary preclinical evidence to prioritise novel therapies for trial 8 . Given the rarity of the disease, this represents a significant challenge that can only be overcome collaboratively 9 .
We published the first molecular characterisation of available cell line models of pHGG (n=3) more than a decade ago, before the advent of comprehensive genomic data on the tumours themselves 10 . Grown as traditional monolayer cultures in serum-containing media, they represented amenable models given their rapid turnover in vitro, and ability to proliferate indefinitely, and with the later discovery of the histone H3 mutations which define a large proportion of pHGG/DIPG cases 11,12 , even represented the first such model system available for a histone mutant tumour in the form of the H3.3G34V KNS42 cell line 13 . Such models are however extremely limited, and concerns about how predictive they are of the patient disease have long been raised across all human cancer types 14-16 17,18 .
More recently, the establishment of short-term cultures in serum-free conditions directly from tumour cells collected at biopsy, surgery, or autopsy has allowed for the generation of a wealth of novel models across multiple cancer types. Such patient-derived (or 'primary') cell cultures are believed to be more reflective of the biology of the disease, given they preserve the genetic characteristics and heterogeneity of the tumours found in patients and could thus represent more clinically relevant models 17,[19][20][21] .
The first DIPG primary culture was established in 2011, in three-dimensional neurosphere conditions tailored to promote stem cell growth and restrict differentiation 22,23 . Later, twodimensional adherent culture of patient-derived glioma stem cell cultures was also employed 24 . The experience in generating such cultures in DIPG specifically was consolidated in an international collaborative study identifying specific protocols predictive of success, including a source from biopsy versus autopsy, acquisition and transport in DMEM media versus Hibernate A, as laminin-coated 2D versus 3D neurospheres 25 . Attempts to systematise the generation of and access to such models across childhood brain tumours in the form of collaborative biobanks have taken advantage of this increasing effort worldwide 26 . Despite this, many of the pHGG/DIPG subgroups observed in the clinic remain poorly represented, and comprehensive molecular and phenotypic data on the models available is incomplete.
We have been prospectively collecting fresh tumour tissue to generate novel models from pHGG/DIPG patients treated at our hospitals, and have amassed a panel of 52 unique in vitro cultures, spanning multiple biological subgroups of the disease. Linking drug screening to comprehensive molecular analysis has identified numerous rational genetic dependencies that may be worthy of further investigation, and highlight the utility of such panels to inform future therapeutic development in this disease.

Establishment of novel pHGG/DIPG primary cultures
To facilitate the assembly of as large a panel of pHGG / DIPG cultures as possible, we took three approaches to sample collection ( Figure 1A). We initiated prospective collection of fresh both two-(laminin-coated flasks) and three-dimensions (neurospheres in ultra-low attachment plates) under stem cell conditions. Models were considered successfully established once they had reach passage five. Doubling times ranged from 1 to more than 6 days. Orthotopic implantation in the anatomical locations from which the primary tumour arose was attempted for 15 samples, of which 10 formed tumours with a median survival of 175 days, though with a wide variability amongst models (range 53-404 days) (Supplementary Figure S1). A summary of these data is provided in Supplementary Table S1.

Molecular credentialing of established cells
Molecular analysis was carried out on cells once beyond at least passage five where possible.
If available, analysis was also carried out on matched frozen tumour tissue (exome sequencing, RNAseq, methylation BeadArray) and blood (exome sequencing) from the same patient. Excluding one case identified as a hypermutator (ICR-CXJ-016, n=4043 somatic mutations in the tumour, n=4288 in the cells), we found an average of 24 and 19 alterations in the tumours and cells, respectively (Supplementary Table S2). There was no statistical difference in the number of alterations in tumours and cells (median = 24, range 6-229 and median = 19, range 5-163, respectively, p=0.645, t-test), with the vast majority of putative driver alterations, as defined by those most frequently observed in our large retrospective 1 and prospective 8 tumour series, found in both. These included 2 samples with H3F3A G34R, 22 with H3F3A K27M and 9 with HIST1H3B K27M mutations. There was one case of histone wild-type DIPG metastasizing to the spine, found instead to harbour EZHIP overexpression by RNAseq 27 . We additionally had multiple samples with alterations in TP53 (n=31), PDGFRA Using the Heidelberg methylation classifier 28 v11b4, cells (and where possible their matching tumour samples) were assigned scores according to which of 92 brain tumour subgroups they most closely resembled (Supplementary Table S2). In addition, the samples were clustered using a tSNE projection along with a reference pan-glioma set of 1766 tumour samples derived from the Heidelberg reference and validation sets 28 and our own studies 1,8,29 . Although formal classifier scores for the cells was often equivocal, the primary cultures tended to align closely with both their relevant subgroups, as well as the patient-matched tumour specimens from which they were derived ( Figure 1C). As such, our panel represented all major subgroups of pHGG/DIPG, with over half representing DMG_K27 (n=30), including the K27M wild-type DIPG spinal metastasis overexpressing EZHIP, and far fewer GBM_G34 (n=2). The remaining H3 wild-type hemispheric models represent HGG_MID (including the hypermutator case), GBM_MYCN, GBM_RTK_III, GBM_RTK_II as well as lower-grade subgroups such as PXA and the NTRK-fusion infant cases in the continuum between IHG and desmoplastic infantile ganglioglioma/astrocytoma (LGG_DIG_DIA) 29 .

High-throughput drug screening
For 20 of the cultures that we were able to establish as 3D neurospheres, we carried out 384 well plate drug sensitivity screening, using cell viability after five days drug exposure (estimated by CellTiter Glo) as the primary readout; this allowed us to calculate drug sensitivity Z scores, SF50 and AUC values for each drug in each 3D neurosphere culture (

Single outlier responses
Certain cell models were found to be dramatically more sensitive to some classes of drugs compared to the rest of the panel. An exemplar of this was ICR-CXJ-001, which displayed screening hits on the basis of POC Z scores against a range of chemotypes of toolbox DNA repair inhibitors, specifically those targeting ATM (e.g. KU0060019) and DNA-PK (e.g. KU0057788) ( Figure 3A). These cells were derived from a IDH wild-type grade IV glioblastoma, arising in the frontal lobe of a 4-year-old boy. The methylation subclassification was GBM_MYCN, and the DNA copy number profile from the methylation array revealed a high-level MYCN amplification ( Figure 3B). This was confirmed by metaphase FISH analysis with specific MYCN and chromosome 2 centromeric probes, highlighting amplification via double minutes ( Figure 3C). In validation, both compounds were found to have a significantly more potent effect on cell viability in ICR-CXJ-001 cells compared to a mini-panel of five pHGG/DIPG cells chosen as insensitive (and MYCN wild-type) (2.9-fold GI50, p<0.001, t-test) ( Figure 3D), confirming the results of the initial screen. Similar outlier responses were observed in other cell models, such as DUB-D001 (H3.3K27M mutant DIPG), which was uniquely sensitive to inhibition of a senescence-associated axis of sphingosine 1-phosphate receptor (fingolimod) and Bcl-2/Bcl-xL (ABT-263, ABT-737) (Supplementary Table S3).

Identification of genetic dependencies
We next sought to explore specific genetic dependencies across our panel by looking for correlations between drug class hits and pathway alterations in multiple cells. Strikingly, we observed consistent hits in drugs targeting the p53-mediated DNA damage response in cultures harbouring activating truncating mutations in PPM1D ( Figure 4A). Three PPM1D mutant cultures (HSJD-DIPG-007, HSJD-DIPG-008, HSJD-DIPG-014) were significantly more sensitive on the basis of AUC to multiple, structurally distinct, PARP inhibitors (such as olaparib, p=0.0105, t-test), as well as the MDM2 inhibitor nutlin-3 (p=0.0165, t-test), compared to PPM1D wild-type cells ( Figure 4B). This was also clearly seen with POC Z scores for both compounds ( Figure  Similarly, H3.3K27M mutant cells were found to have differential sensitivities to compounds acting on the epigenetic processes dysregulated by the mutation. This included the wellestablished effects of HDAC inhibitors, in this instance most significantly by belinostat, but also a novel hit in the form of the acetaldehyde dehydrogenase inhibitor disulfiram, also thought to act on DNA methyltransferases ( Figure 4D). There was a significantly increased sensitivity in H3.3K27M mutant cells compared to wild-type in terms of AUC (belinostat -p=0.0358, t-test; disulfiram -p=0.0088, t-test) ( Figure 4E), as also seen with Z score POC ( Figure 4F).

Functional mutation annotation
We were also able to assign distinct and consistent responses to a range of similarly targeted agents to different genetic alterations targeting PDGFRA in our panel. Specifically, the two cultures showing a significantly different sensitivity to a wide range of compounds known to inhibit PDGFRA were derived from tumours harbouring A384ins (HSJD-GBM-001) and D846N (HSJD-DIPG-008) mutations, spanning both extracellular and kinase domains, respectively ( Figure 5A). Conversely, cells harbouring the previously reported resistance mutation D842Y (HSJD-GBM-002) were insensitive as expected ( Figure 5B Figure S3D) compared to other mutations, and the mini-panel of wild-type controls (6.0-fold, p=0.0035; 136-fold, p=0.0041, respectively, t-test). Notably, the D846N mutation in HSJD-DIPG-008 has been shown to be subclonal in multi-region autopsy sequencing of the patient's tumour 33 , and we were unable to detect this in the cell passages used for validation.

Pathway-level dependencies
Finally, we were additionally able to identify critical dependencies on multiple nodes of the same signalling pathway through our integrated drug screening and molecular data.

DISCUSSION
The establishment of methods for generating patient-derived cultures of pHGG and DIPG is providing a rapidly expanding resource for the study of the disease, with such models being used for mechanistic evaluation of epigenetic reprograming associated with the histone H3 mutations [36][37][38][39][40] , tumour-tumour cell 33 and tumour-microenvironmental interactions 41,42 , invasion/migration 43 and the evaluation of novel drug targets by preclinical efficacy assays [44][45][46][47][48][49][50] and high-throughput screening 30,51,52 . Here we present our experience with deriving a new, well-characterised, prospectively-collected panel of such models alongside more established cells, in order to inform such initiatives spanning pHGG/DIPG subgroups. In addition, we demonstrate the utility of a subset of these models to identify novel, or refined, candidates for biologically rational therapeutic approaches.
We present around 40 models that have not previously had their fundamental molecular features published, though many of which have already been shared collaboratively across research groups worldwide. These cells present a wide range of behaviours in vitro, in both 2D and 3D serum-free culture, and many are problematic to include in high-throughput approaches. Given the additional practical difficulties associated with protracted latency times when orthotopically implanted in immunocompromised mice, often between 12-18 months in our hands, we do not yet have systematic in vivo characterisation of the full panel, particularly given the criteria of deriving a serially xenografted P2 model in order to consider a PDX model 'established' 26 . Such efforts are, however, ongoing. These data are important to record and make available, so that researchers are clear about which models may be most suitable for their particular experiments, in addition to the key molecular features they wish to explore.
Where models have been successful, in vitro and in vivo, they can be seen to recapitulate key phenotypic features of the human disease in terms of morphology, marker expression 33 , and infiltrative growth.
An important criterion to assess the usefulness of these models is how well they retain the key molecular features of the human disease in general, and specifically the patient-matched tumour samples from which they were derived. Here we show that the major driving alterations are retained, and that epigenetic profiles of cell-based models closely resemble the tumours themselves. These data can be reliably integrated to match these models to emerging and By applying a subset of the most amenable models to high-throughput drug screening, we have been able to identify novel treatment candidates via a variety of means. This includes the systematic assessment of genotype-phenotype correlations in the form of drug sensitivities linked to specific genetic alterations, and can be done at the gene-or pathway-level. Hits identified this way included the plausible synthetic lethalities of drugs targeting p53-mediated DNA repair in DIPG patients with activating truncating mutations in PPM1D. Such tumours represent ~10% DIPG 3-6 and 37.5% adult midline glioma 53 , and have been associated with hypersensitivity to NAMPT inhibitors 54 , or a direct target for drug inhibition in combination with radiotherapy 55 . PPM1D (or Wip1) is described as a negative regulator of the p53-mediated DNA damage response 56 , and somatic mutations are mutually exclusive with those found in and NF-KB signalling 58 . Intriguingly, its copper-containing metabolite CuET (which spontaneously forms in cell culture media) is thought to kill cancer cells through aggregation of the p97/VCP segregase subunit NPL4 59,60 , which removes ubiquitinated proteins from the chromatin fraction and may indicate a more direct effect on histones in K27M cells. It has also been shown to be a non-nucleoside DNMT1 inhibitor that can reduce global 5me C content, and reactivate epigenetically silenced genes in prostate cancer cells 61 . Such agents, like 5azacytidine, have been suggested as treatments for H3K27M cells owing to the observed pervasive H3K27ac deposition across multiple loci including endogenous retroviral elements (ERVs), priming them for activation by DNA methylation, and rendering them differentially sensitive to DNA demethylating agents 39 . Notably, the combination of such drugs with HDAC inhibitors presented a specific therapeutic vulnerability in H3K27M cells, and repurposed agents such as belinostat and disulfiram may contribute to such approaches.
By contrast, it is apparent that not all genetic alterations in any given gene necessarily convey the same level of sensitivity to targeted inhibition. Numerous PDGFRA inhibitors have been trialled in pHGG/DIPG since the earliest molecular profiling studies identifying a relatively high frequency of amplification and mutation 62,63 . The results have been generally / universally disappointing, in part through likely poor blood-brain barrier penetration of certain compounds, and also due to a lack of appropriate patient selection 64 . We show this is compounded by the fact that certain mutations (and amplification of the wild-type) may be non-responsive to RTK inhibition.
It does also represent an opportunity, whereby patient-derived models can be used to determine the relative sensitivity to these agents of the distinct mutations -either through retrospective screening panels like this one, or through co-clinical trials 65 . Given the rarity of the disease, such variant-level sensitivities may pose a problem for running clinical trials, however our data also highlight the utility of identifying specific downstream pathway vulnerabilities, and/or multi-target inhibitors in which common susceptibilities to drugs may be identified across genotypes -the example here whereby patients with both brainstem and non-brainstem tumours with sensitizing alterations in PDGFRA or FGFR1, who may benefit from dual RTK or MEK inhibitors and potentially be included in the same study .
There are now many initiatives worldwide to develop, characterise and utilise such patientderived models of pHGG / DIPG. Aggregating ongoing efforts, even if only virtually, to provide the most appropriate, well-characterised models in a systematic way to adequately resource the whole research community remains a priority.

From the South Thames paediatric neurosurgical centres (Kings College Hospital and St
George's Hospital NHS Trusts), where the oncology care is delivered at the Royal Marsden Hospital Children and Young People's Unit, we collected tumour tissue directly from theatre for any suspected high grade glial tumours from patients under the age of 25 years at first diagnosis. These were taken from any anatomical site, and whenever possible, as excess to routine diagnostics, were collected from biopsies as well as resections. If the pathological diagnosis was not a WHO grade III or IV glioma, the specimen was banked for any future appropriate Ethical Committee-approved project. The median age of the high grade glioma Patient-derived cultures were established either immediately after collection (biopsy, resection or autopsy) or from live cryopreserved tissue, with authenticity verified using short tandem repeat (STR) DNA fingerprinting 22,30 (Supplementary Table S2) and certified mycoplasmafree. Fresh tumour tissue was first finely minced with the use of sterile scalpels followed by gentle enzymatic dissociation with LiberaseTL (Roche, Basel, Switzerland, 5401020001) for 10 min at 37°C. After incubation, an additional 5ml of fresh media were added and the sample was centrifuged at 1300rpm for 5 minutes. The digested tissue was then resuspended in fresh media and triturated gently with a pipette (5-10 times). The cell suspension was then  Centre, Bambino Gesù Children's Hospital, or DKFZ Heidelberg. Data was pre-processed using the minfi package in R (v11b4). DNA copy number was recovered from combined intensities using the conumee package. The Heidelberg brain tumour classifier (molecularneuropathology.org) 28 was used to assign a calibrated score to each case, associating it with one of the 91 tumour entities which feature within the current classifier.
Clustering of beta values from methylation arrays was performed based upon correlation distance using a ward algorithm. DNA copy number was derived from combined log2 intensity data based upon an internal median processed using the R packages minfi and conumee to call copy number in 15,431 bins across the genome.       Concentration of compound is plotted on a log scale (x axis) against Z score plotted as a percentage of control (Z score POC) (y axis).