ALK2 inhibitors display beneficial effects in preclinical models of ACVR1 mutant diffuse intrinsic pontine glioma

Diffuse intrinsic pontine glioma (DIPG) is a lethal childhood brainstem tumour, with a quarter of patients harbouring somatic mutations in ACVR1, encoding the serine/threonine kinase ALK2. Despite being an amenable drug target, little has been done to-date to systematically evaluate the role of ACVR1 in DIPG, nor to screen currently available inhibitors in patient-derived tumour models. Here we show the dependence of DIPG cells on the mutant receptor, and the preclinical efficacy of two distinct chemotypes of ALK2 inhibitor in vitro and in vivo. We demonstrate the pyrazolo[1,5-a]pyrimidine LDN-193189 and the pyridine LDN-214117 to be orally bioavailable and well-tolerated, with good brain penetration. Treatment of immunodeprived mice bearing orthotopic xenografts of H3.3K27M, ACVR1R206H mutant HSJD-DIPG-007 cells with 25 mg/kg LDN-193189 or LDN-214117 for 28 days extended survival compared with vehicle controls. Development of ALK2 inhibitors with improved potency, selectivity and advantageous pharmacokinetic properties may play an important role in therapy for DIPG patients.

D iffuse intrinsic pontine glioma (DIPG) is an incurable infiltrating glioma of the brainstem in children, with a median overall survival of 9-12 months 1-3 . The mainstays of chemotherapy in histologically similar tumours of other anatomical brain regions are ineffective, in part due to the absence of factors which predict efficacy, such as MGMT promoter methylation and response to temozolomide 4 . The uniqueness of the underlying biology of DIPG is most readily demonstrated by the high prevalence (>80%) of lysine-tomethionine substitutions at position 27 (K27M) in genes encoding histone H3.1 (HIST1H3B, HIST1H3C) and H3.3 variants (H3F3A) 5,6 . To date, radiotherapy provides the only therapeutic response, although >90% children suffer a relapse and die from their disease within 2 years 1,3,7 .
Recent collaborative molecular sequencing initiatives have defined the genomic landscape of DIPG. We and others identified recurrent somatic activating mutations in the gene ACVR1 in~25% of DIPG patients [8][9][10][11] . Whilst H3.3 K27M mutations are also present in other midline regions such as the thalamus, spine and cerebellum (diffuse midline glioma with H3K27M mutation in the 2016 WHO classification schema 12 ), ACVR1 mutations are associated with H3.1K27M substitutions, and appear restricted to DIPG 13 . In keeping with the clinicopathological differences between H3.3 and H3.1 K27M mutant subgroups 14 , ACVR1 mutations have been reported at a younger age of diagnosis and with longer overall survival in children with DIPG [8][9][10][11] . Although apparently not found in any other human cancer, these variants are found in the germline of patients with the congenital malformation syndrome fibrodysplasia ossificans progressiva (FOP), in which soft tissue is remodelled to bone in response to (often trauma-related) inflammation 15 . In the brain, activin A-ACVR1 signalling is involved in the process of myelination [16][17][18] , an intriguing association given the putative oligodendroglial precursor origins of DIPG [18][19][20] .
ACVR1 encodes the receptor serine/threonine kinase ALK2, and in models of FOP, it has recently been reported that the characteristic mutations confer an aberrant sensitivity to the ligand activin A, produced as part of the inflammatory response, rather than the canonical BMPs 21 . This results in increased pathway activation and cell signalling via a canonical phosphorylated SMAD1/5/8-SMAD4 pathway to drive expression of target genes including ID1/2/3, SNAIL and HEY1 22 . Inhibition of this response provides the major impetus of preclinical development and clinical trials in FOP.
The first small molecule inhibitor of ALK2 was dorsomorphin, identified through a classical BMP ventralization assay in zebrafish embryos 23 . The pyrazolo[1,5-a]-pyrimidine scaffold of dorsomorphin provided the basis of further compounds demonstrated to target ALK2 and the related BMP receptors ALK3 and ALK6, without affecting the type I TGF-beta receptor ALK5 24 . Continued improvements in selectivity have been reported in a newer series based instead on a pyridine scaffold, the prototype of which is K02288 25 .
In the present study, we demonstrate ACVR1 mutations confer abnormal ligand responsiveness to activin A in DIPG cells, with spatiotemporal expression of activin A in neurodevelopment correlating with tumour origins. Screening mutant and wild-type DIPG cultures with a range of pyrazolo [1,5-a]pyrimidine-and pyridine-based ALK2 inhibitors demonstrated differential effects on cell viability, recapitulating genetic knockdown with shRNA, and prolongation of survival in orthotopic patient-derived xenograft models of DIPG.
Signalling responsiveness and dependencies in ACVR1 mutant DIPG cells. Mapping gene expression of ACVR1 and INHBA (activin A) to publically available databases of neurodevelopment reveals a marked peak of activin A expression in the mid-foetal period at approximately 19 weeks post-conception, particularly in the neocortex and striatum, coinciding with that of a K27M gene expression signature thought to reflect the cellular origins of DIPG 26,27 . This is in contrast to BMP4/6, which are substantially reduced at the same time after peaking considerably earlier in development ( Fig. 2a) (Supplementary Fig. 2a). Notably, a similar pattern of differential expression for Acvr1, Inhba and Inhbb was noted in the mouse pons/medulla compared to Bmp4 and Bmp6 ( Supplementary Fig. 2b). Although employing an artificial exogenous source, using in vitro patient-derived DIPG models 10 , we observe a profound increase in phospho-SMAD1/5/8 in ACVR1 mutant cells relative to wild-type only after activin A, rather than BMP4, stimulation in starved conditions, although the differential effects downstream on ID1 are similar with both ligands (Fig. 2b) ( Supplementary Fig. 3). Knockdown of ACVR1 by shRNA in H3.3 K27M, ACVR1 R206H HSJD-DIPG-007 cells caused a dramatic reduction in cell viability after 4, and especially, 11 days, demonstrating the dependence of these cells on signalling through the receptor (Fig. 2c). Knockdown of ACVR1 also resulted in significant cell death, with increased apoptosis seen by FACS (29.3% shACVR1_5 vs 13.1% shCTRL cells at 7 days) (Fig. 2d).
Such effects were not transient, with wash-out experiments for both compounds showing a persistent inhibition of downstream signalling 24 h post-treatment ( Supplementary Fig. 6a, b). In the presence of ligands, there was no difference in GI50 values for LDN-193189 ( Supplementary Fig. 6c), however, addition of Activin-A conferred enhanced sensitivity of R206H HSJD-DIPG-007 cells to LDN-214117 ( Supplementary Fig. 6d). In order to assess whether treatment of DIPG with these compounds would afford a therapeutic window, we also assessed their effects on normal human astrocytes derived from the brainstem (NHA-BS). For both LDN-193189 ( Supplementary Fig. 6e) and LDN-214177 Fig. 6f), there was a marked lack of sensitivity in the NHA-BS cells, with greater than 10-fold difference in effect on cell viability compared to the DIPG cultures.
Pharmacokinetics and pharmacodynamics of ALK2 inhibitors. We next investigated which of the tested compounds possessed the desired biopharmaceutical profiles in terms of CNS penetration and in vivo effects on downstream signalling. Plasma drug exposure (area under the curve, AUC) was assessed for six compounds after oral and intravenous treatment, and bioavailability (F) was calculated. Those with the highest F were LDN-193189 (0.94), and to a lesser extent, LDN-214117 (0.75) ( Supplementary Fig. 7). Both LDN-193189 and LDN-214117 were found to be well-tolerated in mice at 35 and 25 mg/kg, respectively, despite an initial weight loss with LDN-193189 ( Fig. 5a, e). The concentrations of both compounds in the blood persisted over 8 h post oral administration (Fig. 5b, f), whilst importantly, brain levels were either even higher (LDN-193189 brain:plasma ratio = 1.34, 2 h post-dose, day 14) (Fig. 5c) or only slightly reduced compared with plasma (LDN-214117 brain: plasma ratio = 0.80, 2 h post-dose, day 14) (Fig. 5g). In both cases, the concentrations in mouse brain (3.37 and 10.94 µM) were considerably higher than the in vitro GI 50 values observed (0.70 and 1.57 µM). Treatment at these concentrations also lead to a substantial reduction in pharmacodynamic biomarkers in HSJD-DIPG-007 xenograft tumours, showing a near-ablation of  In vivo efficacy of targeting ALK2. Having identified two candidate compounds of distinct chemotypes with in vitro efficacy linked to effects on DIPG signalling, and brain penetration at doses sufficient to elicit a similar response in vivo, we next investigated their efficacy in orthotopic patient-derived xenografts. We treated established ACVR1 mutant HSJD-DIPG-007 and H3/ACVR1 wild-type ( Supplementary Fig. 8a- (Fig. 6a,  b, e, f). After 28 days, there was a decrease in cellularity of the treated ACVR1 mutant tumours compared with controls, as assessed by counting of positive cells by immunohistochemical staining using anti-human nuclear antigen (p = 0.0079, LDN-214117, t-test) (Fig. 6c, d, g, h), though differences in proliferation (Ki67) (Supplementary Fig. 8d-g), and vascularisation (CD31) ( Supplementary Fig. 8h-k) were not significant. In summary, ACVR1 mutations elicit a dependency of downstream signalling in DIPG cells than can be inhibited by single agents of multiple chemotypes of inhibitor compounds  which penetrate orthotopic tumours at concentrations that produce cell death and survival benefit in orthotopic mouse models as single agents at well-tolerated dosages.

Discussion
The discovery of somatic ACVR1 mutations in DIPG identical to those found in the germline of FOP patients has not only highlighted unexpected links between neurodevelopment and chondrogenesis, but also allowed for the fast-tracked evaluation of specific ALK2 inhibitors developed for the control of heterotopic ossification in the context of childhood brain tumours 24 . Inhibition of ALK2, via pharmacologic or genetic means, led to inhibition of proliferation and induction of apoptosis, as well as a selective reduction in cell viability in patient-derived in vitro tumour models, and enhanced survival of mice bearing orthotopic ACVR1-mutant patient-derived DIPG xenografts in vivo. A major challenge in the context of DIPG (unlike FOP) is the necessity for agents to penetrate the tumour behind what may be a particularly restrictive blood-brain barrier within the brainstem. Short-cutting drug development by using the reported off-target effects on ALK2 of drugs licensed for alternate targets and different indications holds promise given the urgency of the unmet clinical need, but appeared unwarranted for the compounds tested here (perhexiline maleate 28 , saracatinib 29 and momelotinib 30 ).
In our mouse models, derivatives of the earliest pyrazolo[1,5-a] pyrimidine-and pyridine-based compounds (LDN-193189 31 and LDN-214117 32 , respectively) were found to be present within the brain at concentrations sufficient to ablate signalling via phospho-SMAD1/5/8 and ID1, biomarkers indicative of pathway inhibition and efficacy. With their good potency and selectivity, LDN-193189 and LDN-214117 are promising lead compounds. They have attractive physicochemical properties consistent with their CNS penetration and good pharmacokinetics due to low clearance and moderate volume of distribution. Notably, the distinct inhibitors were found to have differences in specificity for ACVR1 mutant vs wild-type cultures, and determining such specificity profiles will be an important part of future drug development. Furthermore, the distinct ACVR1 mutations found in DIPG may themselves respond differently, with R206H more sensitive across a range of compounds, whilst the R258G mutant cells were largely insensitive to the pyrazolo[1,5-a]pyrimidines, though responded well to the pyridine-based compounds. There was also a clear therapeutic window demonstrated in respect of the lack of sensitivity in vitro of normal human astrocytes derived from the brainstem. Future development of inhibitors capable of producing effects across the whole spectrum of variants observed in DIPG patients will be critical to maximise patient benefit.
In FOP, ACVR1 mutations confer aberrant responsiveness to the ligand activin A, an effect we demonstrate in DIPG, albeit using the addition of exogenous ligand in vitro. Of more clinical relevance is the observation that activin A expression also appears to peak at the predicted temperospatial origins of DIPG during neurodevelopment, suggestive of a direct mechanistic link to tumour development. Activin-A and Acvr1 were also substantially elevated in the mouse pons/medulla compared to Bmp4 and Bmp6, implying a feasible mechanism of receptor activation in our in vivo orthotopic xenograft experiments. Enhanced signalling by BMPs in the context of adult glioma is associated with a pro-differentiation programme 33 . ACVR1 / H3.1 K27M mutant tumours appear to have a more astrocytic rather than oligodendroglial gene expression signature and morphological appearance 34 , although in terms of putative developmental origin it should be noted that activin A-ACVR1 signalling is associated with oligodendrocyte differentiation and myelination 16,17 , a developmental process actively occurring at the age and location at which DIPG arises. Notably, genes associated with oligodendrocyte differentiation were elevated in G328E/V/W mutant tumours compared to R206H/G356D. Although Activin A confers enhanced signalling via the same phospho-SMAD1/5/ 8-SMAD4 axis as BMPs, it remains possible that various noncanonical pathways may also be activated within this specific context. Gene expression signatures point towards a delicate balance associated with hypoxia 34 and stem cell niche maintenance that remain unexplored.
Notably, not all ACVR1 mutations appear to confer a better prognosis. The R206H mutation, which in the germline accounts for >95% of classical FOP patients 35 , does not have the striking predilection for H3.1 K27M mutations seen with the other variants, and being found in roughly equivalent numbers of H3.3 K27M tumours likely accounts for the extremely poor survival of these patients. That LDN-193189 and LDN-214117 were able to produce a survival benefit in vivo in a particularly aggressive H3.3 H27M/ACVR1 R206H model of DIPG provides hope that such interventions may be more widely applicable to other genotypes.
Although statistically significant, the effect seen with these compounds was relatively modest, producing no cures in our mouse models, similar to other single agent intervention studies in DIPG [36][37][38][39] . Consequently, it is unclear how much benefit would be translated to the clinical setting of ALK2 inhibitors as single agents. Precisely how dependent DIPG cells are on the cell signalling conferred by ACVR1 mutations at the time of presentation is not known, nor are the possible interactions of these inhibitors with the radiotherapy that all children with DIPG are currently given. Identification of co-segregation of ACVR1 mutation and dysregulation in the PI3K pathway potentially points the way towards combinatorial approaches which may be tested in this subgroup of tumours. The Akt inhibitor perifosine was one of the earliest compounds evaluated in DIPG models 40 and translated to the clinic 41 ; sensitivity to the mTOR inhibitor TAK228 (MLN0128) 42 and dual PI3K-mTOR inhibitor BKM120 36 have also recently been reported and an ongoing stratified medicine trial in DIPG patients includes everolimus as one of three investigational arms (NCT02233049). The genetic dependence and preclinical efficacy demonstrated here with inhibitors of ALK2 raises the prospect of another class of compounds entering the clinical trials toolkit for children with this devastating disease.

Methods
Bioinformatic analysis. Curated gene-level copy number, expression and mutation data from DIPG patients were obtained from a recent meta-analysis of published and unpublished data 14 and is available within the paediatric-specific implementation of the cBioPortal genomic data visualisation portal (pedcbioportal. org). Gene expression data from Agilent WG2.5, Affymetrix U133Plus2.0 or RNA sequencing were platform-centred, and log-transformed expression measures were combined and further normalised using pairwise loess normalisation. Gene Set Enrichment Analysis (software.broadinstitute.org/gsea) was performed using the GSEA java application based upon pairwise comparisons of the major subgroups in the merged dataset. Differential expression analysis was based on a Mann-Whitney U test of centred expression values between cases. Spatiotemporal gene expression data from developing and adult brain samples were obtained from the Human Brain Transcriptome project (hbatlas.org) 43 . ACVR1 shRNA knockdown, qRT-PCR and apoptosis analysis. Lentiviral particles for shRNA knockdown of ACVR1 were produced by co-transfecting the packaging vectors (pMD2.G, pMDLg/RRE and pRSV-Rev), and one individual shRNA expression vector into 293T cells. ACVR1 shRNA vectors (Sigma) used were as follows: TRCN0000382239 (TRC2, shACVR1-1), TRCN0000315051 (TRC2, shACVR1-2), TRCN0000315051 (TRC2, shACVR1-3), TRCN0000000443 (TRC1, shACVR1-4) and TRCN0000000445 (TRC1, shACVR1-5) or the scrambled control shRNA vector (SHC002, shCTRL). Viral particles were concentrated using Lenti-x concentrator (Clontech, Mountain View, CA, USA) and resuspended in stem cell media. HSJD-DIPG-007 cells were plated onto laminin-coated plates and 24 h later transduced with either ACVR1 or control shRNA lentiviral particles at a multiplicity of infection of 1 in the presence of polybrene (4 µg/ml). Viral medium was removed 24 h later and cells were selected with puromycin (5 µg/ml). On day 4 post-selection, cells were collected and RNA extracted for quantitation of ACVR1 expression by qRT-PCR. RNA was reverse-transcribed using Superscript II Reverse Transcriptase (Thermo Fisher), followed by quantitative PCR using the Maxima SYBR Green/ROX qPCR (Thermo Fisher) on an Applied Biosystems 7900HT instrument. Reactions contained 0.5 μg RNA and individual measurements of human ACVR1 and human GAPDH endogenous control were measured in triplicate. On days 4 and 11 cells were counted using a Vi-CELL XR (Beckman Coulter, High Wycombe, UK). Cellular apoptosis was measured using AnnexinV-APC Apoptosis Kit (BD Bioscience, Oxford, UK) on day 7 post-selection. Propidium Iodide (PI) was used as a vital dye. Cells were subjected to Annexin V/PI FACS analysis to assess the proportion of cells undergoing apoptosis. For FACS analyses, the stained cells were analyzed using BD machine (BD Biosciences).
Compound efficacy assays. Cells were plated at a density of 2000-4000 cells/well on laminin-coated 96-well plates in a minimum of triplicates. After three days of incubation, compound was added to each well in concentrations from 0.07-20 µM and incubated at 37°C, in 5% CO 2 , 95% humidity for eight days (192 h). For Activin A / BMP4 studies, cells were supplemented with media containing BMP4 (10 ng/ml, Peprotech, London UK) or Activin A (10 ng/ml, Thermo Fisher) 24 h prior to adding the drug. Drug was added in media supplemented with Activin A and BMP4 in concentrations from 0.07 to 20 µM and incubated at 37°C, 5% CO 2 , 95% humidity for eight days (192 h). Dorsomorphin, DMH1, perhexiline maleate, LDN-193189, LDN-212854 and LDN-214117 were purchased from Sigma-Aldrich; saracatinib and momelotinib were purchased from Selleckchem (Houston, TX, USA); K02288 was purchased from BioFocus (Saffron Walden, UK); LDN-213844 (K03841b), LDN-212838 (K03449c) and K05907 were synthesised at the Structural Genomics Consortium, Oxford (see Supplementary Methods for chemical structures and synthesis). LDN-213819 was a kind gift from Paul Yu and Greg Cuny (Harvard University). Cell viability was assessed by the CellTiter-Glo luminescent cell viability assay (Promega, Madison, WI, USA) and GI 50 values were calculated using GraphPad Prism version 6 as the concentration of compound required to reduce cell viability by 50%.
Statement of compliance with ethical regulations guiding proper use of laboratory animals. All experiments utilizing mice were performed in accordance with institutional and European guidelines (EU Directive 2010/63/EU) and were approved by the local animal care and use committee (Comite Etico de Experimentacion Animal at Universidad de Barcelona, protocol 135/11).
Pharmacokinetics. Nine-week-old female BALB/c mice were treated with LDN-193189, LDN-213844, LDN-212838, LDN-214117, K05907 and saracatinib at 5 mg/kg drug intravenously, 5 mg/kg drug by oral gavage, or vehicle. 20 µl blood was taken from the tail vein, spiked onto Whatman FTA DMPK-B cards and left to dry for at least 6 h. Female NOD.SCID mice were similarly treated with either 35 mg/kg LDN-193189, 25 mg/kg LDN-214117 or vehicle orally, with 20 µl tail vein blood taken at 1, 2, 4, 6 and 8 h after the first dose and prepared as before. Plasma and brain tissue samples were taken in triplicate at 2 h post-dose on day 14. Analysis was carried out by liquid chromatography-tandem mass spectrometry (LC-MS/MS) using a Waters Xevo TQS coupled with an Acquity UPLC H-class system (Waters, Herts, UK). Chromatography was carried out using a Phenomenex (Macclesfield, UK) C18 X-B column (2.6 µm, 50 mm × 2.1 mm). Data acquisition was performed using Targetlynx, version 4.1 and pharmacokinetic modelling was carried out using Phoenix WinNonlin version 6.3 (Certara, NJ, USA).
In vivo efficacy studies. A single cell suspension of each culture was made the day before implantation and cultured overnight. On the implantation day, small tumourspheres in exponential growth were harvested by mild centrifugation. 3week-old female NOD.SCID mice were anaesthetised with 100 mg/kg ketamine and 10 mg/kg xylazine and immobilized in a stereotaxic apparatus (Stoelting, Wood Dale, IL) at coordinates x+0.5 and y−5.4 from the bregma suture. 5 × 10 5 HSJD-DIPG-007 or HSJD-GBM-001 tumourspheres (5 × 10 5 cells), suspended in 5 μl matrigel (BD Biosciences) were injected at 3.1 mm depth (targeting the 4th ventricle) with a dull 22G needle attached to a 50 μl syringe (Hamilton, Bonaduz, Switzerland), using a stereotaxic arm. Mice were stratified (n = 7-10 per group) and treated with 25 mg/kg of either LDN-193189, LDN-214117 or vehicle for 28 days, starting 21-28 days post-inoculation. LDN-193189 was prepared each day in water whilst LDN-214117 was prepared in 10% DMSO diluted in saline. Mice were monitored by daily weighing and were sacrificed by cervical dislocation upon deterioration of condition or 20% weight loss from the maximum weight achieved, with tissue taken for further analysis. Mouse brains collected at the end of the efficacy study were processed for immunohistochemistry. Plasma and brain samples from treated and control mice were taken at 2 h post-dose at the end of the 28day treatment for pharmacokinetic and pharmacodynamic analyses.
Immunohistochemistry. PFA-fixed mouse brains were paraffin embedded and sectioned (4 µm) for immunohistochemical analysis and stained with haematoxylin and eosin (H&E). For immunohistochemistry, sodium citrate (pH 6.0) heatmediated antigen retrieval was performed and staining was carried out using an antibody directed against human nuclear antigen (HNA) (Millipore, #4383, 1:100). Pressure antigen retrieval was performed and staining was carried out using antibodies directed against Ki67 (DAKO, #7240, 1:100), and CD31 (Abcam, #28364, 1:50). All primary antibodies were diluted into 1% Tris buffer solution with 0.05% Tween-20, except Ki67 which was diluted into Dako antibody diluent. Antibodies were incubated for 1 h at room temperature. Novocastra Novolink Polymer Detection Systems Kit (Leica Biosystem RE-7150) was used for the staining. Slides were then mounted using Leica CV Ultra mounting medium and slides were imaged using the high throughput-scanning microscope AxioScan Z1. Five fields of view were selected across sagittal and coronal cuts of the brain at 70% zoom and images quantified using Image J and application Cell Counter 44 .
Statistics and reproducibility. Statistical analysis was carried out using R 3.3.0 (www.r-project.org) or GraphPad Prism version 6. Categorical comparisons of counts were carried out using Fisher's exact test, comparisons between groups of continuous variables employed Student's t-test or ANOVA. Interventional in vitro experiments were carried out with at least n = 3 biological replicates. Differences in survival were analysed by the Kaplan-Meier method and significance determined by the log-rank test. All tests were two-sided and a p-value of less than 0.05 was considered significant. Multiple testing was accounted for using false discovery rate q-values or the Bonferroni adjustment.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The bioinformatic datasets generated analysed during the current study are available in the paediatric-specific implementation of the cBioPortal genomic data visualisation portal (http://www.pedcbioportal.org). Data showing gene expression differences in