A Compendium of Syngeneic, Transplantable Pediatric High-Grade Glioma Models Reveals Subtype-Specific Therapeutic Vulnerabilities

H3.3^G34R and PDGFRA Alterations Drive Tumorigenesis from the Ganglionic Eminences, with PDGFRA^C235Y-Derived Cells Uniquely Sensitive to the FGFR Inhibitor Infigratinib

To systematically assess the effect of histone mutations in combination with other alterations seen in patients, we optimized a delivery system to introduce these mutations at embryonic day 12.5 (E12.5) into the mouse dorsal or ventral pallium, targeting the cortex or GE, respectively, using IUE of piggyBac transposons and CRISPR vectors (Fig. 1B; Supplementary Table 1) (32). Transient expression of the piggyBac transposase and Cas9 ensures a brief period of clonal induction, allowing small populations of genetically recombined and spatially restricted premalignant progenitors to give rise to gliomas in an immunocompetent environment. The vector carrying the histone mutant expressed GFP downstream from a modified 2A peptide (PQR), and the mutant histone itself was C-terminally tagged with an HA label, allowing immunochemical detection of tumor cells with GFP and the mutant histone with HA (Supplementary Fig. 1A) (33). The vectors carrying the overexpressed partner alterations were similarly tagged with an EBFP/V5 label (Supplementary Fig. 1B). Vector concentrations were optimized for co-transfection efficiency considering the size of each plasmid (NEBioCalculator). Finally, to enable non-invasive bioluminescence imaging (BLI), we engineered and co-delivered a vector expressing Akaluc luciferase, which enables BLI in C57BL/6J mice harboring tumors in deep brain structures such as the ventral forebrain or brainstem without the need for shaving or depilation (Supplementary Fig. 1C) (34). We confirmed our ability to target these anatomic locations, demonstrating that GFP-encoding vectors can be specifically and reproducibly introduced into the GE or cortex at E12.5 (Supplementary Fig. 2A–D).

To dissect the contributions of H3.3^G34R, specific PDGFRA alterations, and the environmental niche of the GE in vivo, we created mouse models carrying H3.3^G34R, ATRX^LOF and p53^LOF (GPA), together with PDGFRA^WT (GPAP), PDGFRA^C235Y (GPAC), or PDGFRA^D842V (GPAD) (Fig. 1B; Supplementary Figs. 3A–B; 4A–B; 5A–B; 6). PDGFRA^WT was used to mimic the focal amplification of the normal gene that is often seen in these tumors, while PDGFRA^C235Y modeled the highly prevalent mutation observed specifically in H3.3^G34R tumors. PDGFRA^D842V was included as a positive control to guarantee tumor development and allow comparison to other published models (35–38).

GPAD tumors showed 100% penetrance with a median survival of 67 days, as expected for the highly oncogenic PDGFRA^D842V mutation (Fig. 1C–E; Supplementary Table 2–3). Tumor penetrance in GPAP and GPAC mice was lower, with 55% of the mice developing neurological symptoms and tumors (Fig. 1C–E). GPAC tumors were more aggressive and had a shorter latency, with a median survival of 296 days compared to 453 days for GPAP tumors (Supplementary Table 2–3). The GPA combination without PDGFRA was unable to drive tumor development (Fig. 1E; Supplementary Fig. 6). We validated the presence of each introduced mutation in all models (Fig. 1D; Supplementary Figs. 3A–B; 4A–B; 5A–B; 6), confirming that co-delivery of separate vectors using piggyBac IUE leads to stable integration and co-expression of each mutation in the majority of tumor cells. Histologically, H3.3^G34R tumors possessed both densely compacted and diffuse components, and all maintained low levels of ATRX expression (Fig. 1D; Supplementary Figs. 3A–B; 4A–B; 5A–B; 6). Recapitulating human tumor biology, H3.3^G34R tumors displayed lower levels of Olig2 than K27M tumors and expressed high levels of the interneuron progenitor markers Dlx1 and Dlx2 (Fig. 1F) (13).

Introducing PDGFRA mutations into the GE induced lethality in a proportion of electroporated embryos. Co-delivery of H3.3^WT with PDGFRA^C235Y or PDGFRA^D842V saw only 34% and 38% of pups surviving to weaning, respectively (Supplementary Fig. 7). This decrease in survival was not observed with PDGFRA^WT (Supplementary Fig. 7). Interestingly, survival was higher when H3.3^G34R was co-delivered with PDGFRA^C235Y into the GE, with 65% of H3.3^G34R/PDGFRA^C235Y pups surviving to weaning (Supplementary Fig. 7). This suggests that H3.3^G34R can alleviate some of the lethality induced by PDGFRA^C235Y, allowing survival of a greater proportion of embryos expressing this combination of mutations until H3.3^G34R/PDGFRA^C235Y cells undergo transformation. This effect is also niche-specific, occurring only in the GE and not in the cortex (Supplementary Fig. 7).

We generated gliomasphere (GS) cell lines from symptomatic GPAP, GPAD and GPAC mice in serum-free media. GS cells expressed HA-tagged H3.3^G34R and maintained ATRX and p53 downregulation with PDGFRA overexpression, as expected (Fig. 1F). We assessed the sensitivity of GPAP, GPAD and GPAC GS cells to eight different small molecule inhibitors, each targeting a different pathway: PDGFRA (avapritinib), PIK3CA (alpelisib), FGFR1 (infigratinib), MEK (trametinib) (39), MDM2 (idasanutlin) (40), NAD biosynthesis (FK866) (41), histone lysine methylation (GSK-J4) (42) and histone acetylation (corin) (43). While GPAC cells were largely insensitive to the PDGFRA inhibitor avapritinib (Supplementary Fig. 8) (44), they displayed a remarkable, sub-micromolar sensitivity to the FGFR inhibitor infigratinib (Fig. 1G). This sensitivity to infigratinib was specific and more pronounced in GPAC cells compared to GPAD, GPAP and wild-type neurospheres, indicating a unique selectivity for the PDGFRA^C235Y mutation. We confirmed human tumor-derived GBM002 cells that carry the G34R mutation are also more sensitive to infigratinib than H3.3^WT glioma cells (Fig. 1G).

Collectively, these data describe the first H3.3^G34R models developed by targeting the GE in embryonic mice, experimentally validating the GE as a niche capable of giving rise to these tumors. Three models were developed, driven by H3.3^G34R and three different PDGFRA alterations. Additionally, the GPAC model describes the first GE-targeted, H3.3^G34R-mutant model driven by the recently discovered PDGFRA^C235Y mutation. Finally, we demonstrate that GPAC tumors are unexpectedly sensitive to the FGFR inhibitor infigratinib, which may offer alternative strategies to sensitize these tumors to treatment, known to be insensitive to traditional PDGFRA inhibitors.

K27M Models Driven by ACVR1, PIK3CA, PPM1D, NF1 and FGFR1 are Highly Penetrant

Most K27M DMGs harbor partner alterations, with the majority possessing 3–10 additional mutations from over 30 recurrently altered genes (2, 45). To assess which alterations can cooperate with K27M to induce tumorigenesis, we first evaluated mutations most frequently found in human DMGs. We focused on modeling four main subtypes driven by combinations of distinct partners (Fig. 2A). Three of these models (KPP, H3.1KACVPIK and KPPMPIK) represent the most common DMG subtypes, harboring partner alterations in PDGFRA, ACVR1, PIK3CA and PPM1D. Additionally, we modeled a rarer midline subtype that occurs outside the hindbrain, in the thalamus (7, 22), harboring NF1^LOF and FGFR1^N457K (KNF, Fig. 2A). FGFR1 mutations and NF1^LOF represent interesting counterpoints to p53^LOF and PDGFRA alterations, allowing evaluation of different signaling pathways that have thus far not been investigated in the context of K27M.

We introduced these mutation combinations into the LRL of E12.5 mice using the piggyBac transposon system, as before. Bicistronic HA-tagged H3.1/3^WT or H3.1/3^K27M-GFP vectors were delivered together with vectors expressing various partner alterations, the Akaluc vector to allow BLI, and transient piggyBac transposase and Cas9, to mediate permanent recombination. We confirmed our ability to target the correct anatomical location, demonstrating specific and reproducible induction of GFP expression in the hindbrain at P0 and at weaning (Supplementary Fig. 9A–C).

High-grade, 100% penetrant tumors developed with all four combinations (Fig. 2B–D), with large GFP⁺ lesions apparent in the hindbrains of symptomatic mice (Supplementary Figs. 10A–B; 11A–B; 12A–B; 13A–B; 14; 15). We verified the presence of each partner alteration in these tumor models (Fig. 2C; Supplementary Figs. 10A–B; 11A–B; 12A–B; 13A–B; 14; 15). Tumors were diffusely infiltrative, and GFP⁺ tumor cells clearly colocalized with H3^K27M and Ki67 and expressed high levels of neural stem/progenitor cell markers Olig2 and Nestin (Fig. 2C; Supplementary Figs. 10A–B; 11A–B; 12A–B; 13A–B; 14; 15). GFP⁺/HA⁺ cells in tumors displayed marked reduction in H3K27me3 levels, as expected in cells expressing the K27M mutation (Supplementary Figs. 10A–B; 11A–B; 12A–B; 13A–B; 14; 15). These data highlight the similarity of these models to human tumors, as these histologic features are used in the clinic to differentiate DMGs from other brain tumor entities. Each mutation combination also demonstrated differences in the latency of tumor progression, as measured by the length of time needed for the development of brain-tumor related neurological symptoms requiring euthanasia (Fig. 2D, Supplementary Table 2–3). KNF tumor-bearing mice demonstrated the shortest median survival at 28 days, followed by KPPMPIK (42 days), H3.1KACVPIK (65 days) and KPP (230 days).

These data show that p53^LOF and PDGFRA^WT, PIK3CA^E545K and PPM1D^ΔC (PPM1D^W420*, orthologous to human PPM1D^W427*, the most frequent truncation in (2)), and NF1^LOF and FGFR1^N457K are sufficient to induce DMGlike tumors in the context of H3.3^K27M. In addition, in the context of H3.1^K27M, ACVR1^G328V and PIK3CA^E545K are able to induce transformation. Taken together, these four mutation combinations induce brainstem tumors in mice that recapitulate the genetic and histologic characteristics of DMG subtypes, as well as their lethality. Importantly, in contrast to previously published H3^K27M-driven models (21, 46–49), the models described here are the first and only brainstem-targeted models that do not rely on constitutively active PDGFRA or PDGF overexpression to drive tumor development, while maintaining high penetrance (31). Exogenously high levels of the ligand or constitutive activation of PDGFRA is rare in the context of H3^K27M in human tumors and can confound efforts to identify selective targets for precision therapy, as these can be oncogenic independently of H3^K27M. The models described here allow evaluation of the unique tumor biology produced by specific partner alterations and investigation of their selective therapeutic targeting, without the confounding presence of PDGF overexpression or mutant PDGFRA.

FGFR1 and NF1 Mutations Promote an Early OPC Tumor Signature while p53 Loss with PDGFRA Amplification Fuels Myeloid Infiltration

To define the transcriptional landscape of H3.3^K27M tumors at single-cell resolution, we profiled the KNF (n=3), KPP (n=2) and KPPMPIK (n=2) models using 10X Chromium single-nuclei RNA sequencing (snRNA-seq). In total, we obtained 137,695 cells passing quality control with an average of 2,255 UMIs per cell (KNF n=93,960; KPP n=31,479; KPPMPIK n=12,256, Supplementary Table 4–6). After quality control, dimension reduction, and clustering, we first distinguished clusters of immune cells from cells of neuroectodermal origin based on canonical markers (Supplementary Fig. 16A–C). Next, we performed an automated annotation at the individual cell level using a consensus of 4 classifiers trained on a single-cell reference atlas of the developing brain for neuroectodermal cells (27), and on a single-cell reference of CD45⁺ mouse brain cells for immune cells (50) (Supplementary Fig. 16D–E). This analysis identified well-defined populations of cells along the oligodendroglial lineage, as well other cell types including astrocytes, vascular, neuronal, and immune cells (Supplementary Fig. 16E). Finally, we corrected for batch effects using Harmony, subsampling KNF replicates to avoid biases resulting from the large differences in cell numbers. This resulted in well integrated datasets with biological replicates showing similar profiles (Fig. 3A, Supplementary Fig. 17A–B).

To assess the relevance of these models to study human disease, we performed single-sample gene set enrichment analysis (ssGSEA) in each individual cell, assessing gene signatures specific for K27M and G34R/V human tumors, or posterior fossa ependymoma (PFA) as controls. For all H3.3^K27M mouse models, OPC-like cells showed a very specific enrichment for H3.3^K27M signatures (Fig. 3B, Supplementary Fig. 17C), significantly higher than for H3.1^K27M, H3.3^G34R, and PFA signatures, confirming their molecular profiles recapitulate human tumor biology.

We next compared cell compositions of the KNF and KPP models, to study the effects of the rarer, thalamus-specific NF1^LOF and FGFR1^N457K mutations in relation to the more frequently brainstem-specific PDGFRA^WT amplification and p53^LOF context (Fig. 3B–D). First, we observed a marked difference in cellular hierarchy between the two phenotypes, with the KNF models predominantly composed of OPC-like cells, while the KPP models showed a different cellular composition with more mature-like cells predominating (Fig. 3E–G). Reconstruction of gene regulatory networks using SCENIC (51) revealed a different dynamic of transcription factor (TF) activation along differentiation. For the oligodendroglial lineage, several TFs active in early OPC-like cells in the KNF models (e.g. Foxp2, Sox17) were activated in later differentiation stages (myelinating oligodendrocyte-like cells) in the KPP model (Fig. 3H, Supplementary Fig. 17D). Conversely, Maf, Gli1, Bcl11, among others, showed later activation in the KNF genotype. Second, KPP models showed much higher levels of myeloid infiltration (Fig. 3F, I–J), with higher proportions of cells expressing pan-immune and myeloid markers (Fig. 3I). Furthermore, microglia clusters in these models showed increased expression of disease associated microglia (DAM) signature genes (Fig. 3J). We also profiled KPPMPIK using snRNA-seq, which displayed more similarity to KPP than KNF tumors, with fewer immature OPCs and more mature-like cells (Supplementary Fig. 17A–D). Altogether, these results underscore the distinct downstream effects of NF1^LOF and FGFR1^N457K on cell intrinsic programs, as well as cell on extrinsic factors such as the tumor immune microenvironment.

p53^LOF, NF1^LOF, FGFR1^N457K and CCND2^WT are Sufficient to Induce Transformation in the Context of K27M

Next, we evaluated H3^K27M-specific partner alterations individually and in combination with other rarer partners to identify those that had transformation potential. H3 mutations arrest progenitors in a proliferative, preneoplastic state but cannot induce transformation alone, thus we interrogated which additional mutations are needed at minimum to induce tumor development. We focused on mutations linked to discrete brain regions and key biological pathways altered in DMGs, including mutations specific to thalamic and brainstem tumors and affecting RTKs, BMP, PI3K/mTOR, and MAPK signaling, the cell cycle, and DNA repair. Therefore, we investigated alterations in FGFR1, ACVR1, PIK3CA, NF1, CCND2, PPM1D and p53, in combinations of 2 or 3 co-expressed mutations, and in the context of 2 different histone mutations (H3.1^K27M and H3.3^K27M) (Fig. 4A). These alterations, together with the 4 subtype models described above, reflect a large proportion of the genetic heterogeneity seen in DMGs while also representing tumor subtypes that do not rely on PDGF/PDGFRA signaling.

When delivered into the LRL at E12.5, p53^LOF (KP), FGFR1^N457K (KF), and NF1^LOF (KN) were sufficient to induce tumorigenesis as the sole second hit together with H3.3^K27M, with median survivals of 275 days, 209.5 days, and 199 days, respectively (Fig. 4B–C; Supplementary Figs. 18A–B; 19A–C; 20A–B). However, PPM1D^ΔC or PIK3CA^E545K were not able to drive tumor development in our system when delivered as the only partners with H3.3^K27M. Only 1 of 4 H3.3^K27M/PPM1D^ΔC mice included in the survival curve in Figure 2D developed a brain tumor, the rest developed hydrocephaly or other non-brain tumor-related conditions requiring euthanasia. In the context of H3.1^K27M, p53^LOF could function as the sole additional transforming hit with a median survival of 153 days (H3.1KP; Fig. 4B–C; Supplementary Fig. 21A–B), but ACVR1^G328V could not without co-introduction of PIK3CA^E545K as a third hit. In the context of H3.3^K27M and p53^LOF; FGFR1^N457K (KPF) and CCND2^WT (KPC) were able to accelerate tumor development as third hits, with median survivals of 47 days and 217 days, respectively (Fig. 4B–C; Supplementary Figs. 22A–C; 23A–B). The KP, H3.1KP, KF and KPF models were 100% penetrant, while the KN and KPC models were 70% and 65% penetrant, respectively (Fig. 4C; Supplementary Table 2–3). GFP⁺ cells in these tumors colocalized with K27M and Ki67 and expressed high levels of Olig2 (Fig. 4B; Supplementary Figs. 18A–B; 19A–C; 20A–B; 21A–B; 22A–C; 23A–B). We confirmed the presence of the histone mutation and partner alterations in each model (Supplementary Figs. 18A–B; 19A–C; 20A–B; 21A–B; 22A–C; 23A–B). The different combinations of partner alterations produced different latencies of tumor development, as shown by their Kaplan-Meier survival curves depicted in Fig. 4C and summarized in Supplementary Table 2–3.

Of note, H3.1^K27M and H3.3^K27M expression alone (Supplementary Figs. 14–15), and H3.3^WT/p53^LOF, H3.3^WT/NF1^LOF, H3.3^WT/FGFR1^N457K, and H3.3^WT/PPM1D^ΔC/PIK3CA^E545K (Supplementary Table 3), were not able to induce tumor development in mice when introduced into the LRL at E12.5. Although our previous work described H3.3^K27M and p53^LOF-driven gliomagenesis in both the cortex and hindbrain (20), the KP model described here is the first 100% penetrant model of this 2-hit combination in the brainstem. A previous H3.3^K27M/p53^LOF transgenic model produced cerebellar rather than brainstem tumors (21). In contrast, our approach restricts tumor development to the brainstem, where these tumors occur in patients. In addition, H3.1KP is also the first model to describe H3.1^K27M’s oncogenic capacity with just one additional alteration: p53^LOF. Finally, mutations impacting cell-cycle components are present in 25% of pHGGs, with CCND2 amplifications being the most common in DMG (2). Thus, the KPC model is unique in allowing evaluation of cell-cycle checkpoint deregulation in the pathogenesis of DMG.

We also compared the H3.3^K27M, p53^LOF and PDGFRA^WT (KPP) model to iterations that included ATRX^LOF and PDGFRA^D842V, to identify unique effects of these partners. We compared the KPP model to models harboring the constitutively active PDGFRA^D842V mutation as this mutation and another similarly constitutively active PDGFRA mutation (PDGFRA^V544ins) have been utilized in recently published models of H3^K27M and H3.3^G34R gliomas (21, 38, 48, 52). We also wanted to evaluate whether ATRX^LOF impacted tumor latency, as this alteration is found in a proportion of brainstem tumors. We therefore modeled: 1) H3.3^K27M, p53^LOF and PDGFRA^D842V (KPD), 2) H3.3^K27M, p53^LOF, ATRX^LOF and PDGFRA^WT (KPAP), and 3) H3.3^K27M, p53^LOF, ATRX^LOF and PDGFRA^D842V (KPAD) (Fig. 4B–C; Supplementary Figs. 24A–B; 25A–B; 26). As in KPP, we found KPD, KPAP and KPAD tumors were 100% penetrant in the brainstem, with median survivals of 34.5 days, 207 days, and 45 days, respectively, in contrast to KPP’s 230-day median survival (Supplementary Table 2–3). It is important to note that H3.3^WT/PDGFRA^D842V expression produced tumors with a similar median survival to H3.3^K27M/PDGFRA^D842V (34.5 days vs 37 days), and the two were not significantly different (Supplementary Table 2–3). This suggests that PDGFRA^D842V is oncogenic regardless of H3^K27M expression in our system and may obscure the role of the histone mutation and other partner alterations.

These data demonstrate that the K27M histone mutation and at least one additional partner are necessary for tumorigenesis, and that neither the histone mutation alone, nor the partner alterations in the context of H3.3^WT, are sufficient. In the context of H3.3^K27M and H3.1^K27M, p53^LOF, NF1^LOF, FGFR1^N457K and CCND2^WT can act as second and third hits capable of inducing and/or accelerating tumorigenesis. These data also demonstrate that the 2-hit combinations of H3.1^K27M/ACVR1^G328V, H3.3^K27M/PIK3CA^E545K and H3.3^K27M/PPM1D^ΔC are not oncogenic using this approach, potentially requiring additional partners for transformation. Altogether, these data describe nine minimal, 2 and 3-hit mutation combinations (KP, H3.1KP, KN, KF, KPF, KPC, KPD, KPAP and KPAD) capable of inducing tumor development in the context of H3^K27M. Six of these models are able to transform without co-expression of PDGF/PDGFRA, and therefore represent valuable preclinical systems reflecting patient alterations.

Specific Partners Associate with Spinal Dissemination, Exophytic Spread and Cranial Nerve Invasion

Further phenotypic characterization of these models revealed intriguing differences in spinal dissemination, exophytic spread or blood vessel co-option, cranial nerve invasion, and parenchymal invasion. Spinal dissemination, and more specifically thoracic spinal dissemination (rather than medullar or cervical), was exclusive to KNF mice, and was very apparent, occurring in every mouse in this condition and in no other DMG model (Fig. 5A). We also observed what appears to be co-option of the ventral blood vessel tracks of the brain, sometimes referred to as exophytic spread, in 40% – 80% of mice harboring FGFR1^N457K, PIK3CA^E545K, PPM1D^ΔC and NF1^LOF (Fig. 5B). Invasion into highly myelinated cranial nerves was also a noteworthy phenotype, occurring most frequently in the KF, KPP, H3.1KACVPIK and KPPMPIK conditions (Fig. 5C). In terms of invasion in brain parenchyma, FGFR1^N457K, NF1^LOF, PPM1D^ΔC and PIK3CA^E545K-expressing conditions demonstrated significantly higher levels of thalamic/supratentorial and cerebellar invasion. Conditions with ATRX^LOF and 2-hit models KP and H3.1KP displayed lower levels of invasion, with the H3.1KP model in particular being more restricted to the brainstem proper in 60% of mice of this condition (Fig. 5D).

We characterized the mitotic index and levels of H3K27me3 across the 16 models representing different pHGG subtypes (Fig. 5E–F). The mitotic index was consistently >10 across all conditions, with H3.1KACVPIK, KPF, KNF and KPC possessing the highest mitotic activity at >40 for each (40% of GFP⁺ tumor cells being Ki67⁺). GPAP, GPAC and GPAD tumors had mitotic indices of 13, 30 and 55, respectively (Fig. 5E). H3K27me3 levels were significantly reduced in all models harboring the H3^K27M mutation, while H3.3^G34R tumors maintained high levels of this epigenetic mark (Fig. 5F). Altogether, these data describe specific phenotypes that are unique to certain partner alterations, such as spinal dissemination, exophytic spread, cranial nerve invasion and extra-pontine spread. These data also demonstrate differences in proliferation rates in models driven by different partner alterations and confirm the primary impact of the H3^K27M mutation in every tumor (reduced H3K27me3 levels).

Ex Vivo Cells Derived from De Novo Models Engraft in Syngeneic, Immunocompetent Mice

Syngeneic, engraftable pHGG models are needed to identify novel treatment strategies targeting microenvironmental and immune interactions. These models have been difficult to generate and represent a major technical obstacle within the pHGG field, limiting identification of novel drug targets. To develop syngeneic models, we derived a minimum of two lines from symptomatic mice for each model and grew them in serum-free media as gliomaspheres (GS). We confirmed expression of HA-tagged H3^K27M as well as downregulation of H3K27me3 in each GS line derived from the 2, 3 and 4-hit models, compared to WT LRL neural stem cells (NSCs) (Fig. 6A–B). We confirmed downregulation of p53 in KP and H3.1KP cells, and NF1 in KN cells (Fig. 6A). KF cells overexpressed FGFR1 as expected (Fig. 6A), and KPP, KPD, KPAP and KPAD cells overexpressed PDGFRA, the latter two together with ATRX downregulation (Fig. 6B). KPPMPIK cells overexpressed the truncated form of PPM1D and PIK3CA, while H3.1KACVPIK cells overexpressed ACVR1 and PIK3CA (Fig. 6B). KNF and KPF cells overexpressed FGFR1 as expected, and KPF cells demonstrated knockdown of p53 levels, while KNF cells had diminished levels of NF1 (Fig. 6B). These data demonstrate that each model retains expression of driver and partner alterations introduced in the embryo and corroborates our in vivo findings regarding their co-expression (Supplementary Fig. 10–15; 18–26).

Next, we evaluated the engraftment potential of these GS lines in syngeneic C57BL/6J mice. H3^K27M GS cells were engrafted into the pons, while H3.3^G34R tumor spheres were engrafted into the striatum. GS lines derived from every condition except for KNF were able to engraft in syngeneic mice, and most H3.3^K27M lines were 100% penetrant with very short latencies (Fig. 6C–D). It is unclear why KNF GS lines were unable to engraft, however this combination of mutations produces the most aggressive tumors, with the shortest latencies in the de novo setting and rapid growth in vitro. It is possible that this condition is less adaptable to changes in growth conditions. KF, H3.1KP, H3.1KACVPIK, GPAP and GPAC were less than 100% penetrant when orthotopically engrafted, with GPAP consistently demonstrating 40%, GPAC 60%, H3.1KP and H3.1KACVPIK 60%, and KF 80% penetrance (Fig. 6D). In contrast, KP, KPP, KPPMPIK, KN, KPF and GPAD were 100% penetrant, with all mice developing symptoms in 20–100 days (Fig. 6D). Engrafted tumors appeared high-grade in H&E-stained sections and were positive for GFP, H3^K27M or H3.3^G34R and possessed high levels of Ki67 labeling (Fig. 6D). These engraftable models, representing numerous different pHGG subtypes, will be useful tools for investigating tumor interactions with immune and other stromal cells. Importantly, introducing these mutation combinations into wild-type GE or LRL NSCs in vitro was not able to produce similarly engraftable GS lines, suggesting that prior transformation in an in vivo, immunocompetent environment is required.

Drug Screening in Patient and Murine Model-Derived Cell Lines Reveals Sensitivity to Combination Therapy Targeting MEK and PI3K Signaling

To utilize these models for identifying subtype-specific vulnerabilities that are pharmacologically tractable, we assessed sensitivity to the same eight different small molecule inhibitors used to assess sensitivity in GPAC cells. Each molecule targeted a different pathway component impacted by one of the partner alterations: PDGFRA (avapritinib), PIK3CA (alpelisib), FGFR1 (infigratinib), MEK (trametinib) (39), MDM2 (idasanutlin) (40), NAD biosynthesis (FK866) (41), histone lysine methylation (GSK-J4) (42) and histone acetylation (corin) (43). We evaluated these inhibitors in 10 different mouse H3^K27M GS lines (KP, H3.1KP, H3.1KACVPIK, KPP, KPD, KPPMPIK, KNF, KPF, KN, and KF).

We identified specific sensitivity to PDGFRA inhibition by avapritinib in KP, H3.1KP and KPP cells, and to PIK3CA inhibition by alpelisib in KP, KF, KNF and GPAC cells. Cells from KF, KNF, KPF, GPAC and KPAD showed specific sensitivity to FGFR1 inhibition by infigratinib, while H3.1KP, H3.1KACVPIK, KPPMPIK, KN and KNF cells were sensitive to MEK inhibition by trametinib (Fig. 7A; Supplementary Fig. 27). We also saw increased sensitivity to corin in KNF and H3.1KACVPIK cells (Fig. 7A; Supplementary Fig. 27). Interestingly, we noticed differences in sensitivity in KPD (PDGFRA^D842V) compared to KPP (PDGFRA^WT) cells, with PDGFRA^D842V cells being more sensitive to infigratinib and corin, and less sensitive to GSK-J4, avapritinib and alpelisib, than PDGFRA^WT cells (Supplementary Fig. 27). This suggests that PDGFRA mutation status may be an important consideration in evaluating responses to treatment in DMG models. Statistical analyses are summarized in Supplementary Table 7. We confirmed the selective sensitivity of murine KPP cells to avapritinib in vivo (Fig. 7B). We orally delivered 30 mg/kg of avapritinib to syngeneic allografts over 2–3 weeks. Avapritinib treatment significantly extended survival in the KPP model (median survival 18 days with vehicle vs 26 days with avapritinib, p=0.0023).

To confirm these partner alteration-specific vulnerabilities in the human context, we first assessed sensitivity to alpelisib and trametinib in patient-derived cells carrying NF1 mutations (BT416), p53 mutations and PDGFRA^WT (SU-DIPG-VI), and PIK3CA^E544K/PPM1D^ΔC (BT869). We confirmed sensitivity to alpelisib in p53 mutant/PDGFRA^WT and PIK3CA^E544K/PPM1D^ΔC cells and to trametinib in PIK3CA^E544K/PPM1D^ΔC and NF1 mutant cells (Fig. 7C). These results demonstrate that our mouse model-derived GS cells can aid in identifying therapeutic vulnerabilities unique to specific partners, which also replicate in human tumor-derived cells with similar mutation profiles.

To identify combination therapy approaches that could increase likelihood of therapeutic success, we combined the lowest effective concentrations of trametinib and alpelisib and evaluated this combination in mouse and human cells carrying p53^LOF/PDGFRA^WT and PIK3CA^E545K/PPM1D^ΔC alterations. In cells from both species, trametinib and alpelisib synergized and reduced viability with stronger effects than the application of either agent alone (Fig. 7D). Therefore, a combination therapy approach targeting MEK and PI3K signaling can sensitize DMG subtypes carrying partner alterations in p53/PDGFRA^WT and PPM1D/PIK3CA.

Finally, to identify agents that can limit the hallmark invasiveness of DMGs, we investigated whether in vivo administration of trametinib and alpelisib could limit tumor dissemination in the short term. We administered the agents directly into the CNS using mini-osmotic pumps, and included a low dose of the efflux transporter inhibitor elacridar to retain high trametinib levels within the CNS. At 1 week following engraftment with KPP GS cells, when tumors demonstrated moderate levels of BLI signal, pumps containing either vehicle/elacridar or alpelisib/trametinib/elacridar were implanted into mice for continuous drug delivery into the fourth ventricle. As shown in Fig. 7E, 4 days of direct CNS delivery dramatically reduced GFP⁺ cell dissemination in the alpelisib/trametinib-treated animals, relative to vehicle. Importantly, tumor volume was not affected at this early stage (Fig. 7E), indicating the specific effects on dissemination and invasion. These results highlight the utility of our in vivo models for evaluating pharmacological strategies to curtail DMG invasiveness, which may make them more amenable to surgical resection. Taken together, these data demonstrate the utility of our syngeneic models for preclinical testing of treatment efficacy, providing a proof of principle for investigations now possible for many different pHGG subtypes.

Vector Construction

The piggyBac donor and helper vector system was used to transduce NSCs in utero, as previously described (32). CAG-PBase and PBCAG-GFP were a kind gift from F. Chen and J. LoTurco. CRISPR/Cas9 pX330 vectors containing negative control (5’-GCGACCAATACGCGAACGTC-3’), Trp53-targeting (5’ACAGCCATCACCTCACTGCA-3’) or Nf1-targeting gRNA (5’-AGTCAGCACCGAGCACAACA-3’) sequences were a kind gift from J. Gronych (59). The piggyBac vectors carrying the histone mutations and partner alterations were cloned into PBCAG-GFP using a combination of classical restriction enzyme-based cloning, gene synthesis and the NEBuilder HiFi DNA Assembly Cloning System (New England BioLabs). All vectors drive target gene expression from the CAG promoter, and shRNAs from either the U6 promoter or an intronic eSIBR cassette (60, 61). The mutant histone, PDGFRA, ACVR1 and Akaluc vectors expressed GFP downstream from a PQR 2A peptide, and the mutant histone was C-terminally tagged with an HA label (33). The piggyBac vectors carrying other partner alterations were tagged with an EBFP/V5 label upstream from the PQR 2A peptide. The piggyBac and CRISPR vectors used in this study were: PBGAG-H3.3^WTHA-PQR-GFP, PBCAG-H3.3^K27MHA-PQR-GFP, PBCAG-H3.1^K27MHA-PQR-GFP, PBCAG-H3.3^G34RHA-PQR-GFP, PBCAG-PDGFRA^WT-PQRGFP, PBCAG-PDGFRA^D842V-PQR-GFP, PBCAG-PDGFRA^C235Y-PQR-GFP, PBCAG-ACVR1^G328V-PQR-GFP, PBCAG-EBFP^V5-PQR-PIK3CA^E545K, PBCAG-EBFP^V5-PQR-PPM1D^W420*, PBCAG-EBFP^V5-PQR-FGFR1^N457K, PBCAG-EBFP^V5-PQR-CCND2^WT, PBU6-ATRXshRNA, PBCAG-Akaluc-PQR-GFP, CAG-PBase, PX330-Nf1gRNA-CBh-Cas9, PX330-p53gRNA-CBh-Cas9, PX330-controlgRNA-CBh-Cas9.

In Utero Electroporation

In utero electroporation was performed as previously described (20), with minor modifications for targeting the hindbrain or ganglionic eminences. Timed-mated, pregnant C57BL/6J (RRID:IMSR_JAX:000664) mice were acquired from (Charles River Laboratories, Kent, United Kingdom) and maintained under pathogen-free conditions, in individually ventilated cages, with food and water provided ad libitum. All procedures were approved by the University of Cambridge Animal Welfare and Ethical Review Body and carried out under a UK Home Office License (PPL PP2303899) in accordance with the Animals (Scientific Procedures) Act 1986. Pregnant females at E12.5 (hindbrain) or E13.5 (forebrain) were anesthetized using 2.5% isoflurane and 1.5L O₂/min, with analgesic support provided pre-operatively via subcutaneous delivery of Buprevet at 0.1 mg/kg. Uterine horns were exposed through a 1 cm incision and individual embryos were digitally manipulated into the correct orientation for intraventricular injection. Pulled borosilicate capillaries were loaded with endotoxin-free DNA and Fast Green dye (0.05%, Sigma) for visualization, and a microinjector (Eppendorf) was used to inject either the lateral or fourth ventricles with the DNA-dye mixture. 3–5 plasmids were injected simultaneously, each up to a final concentration of 2 μg/μl and 1–2 μl of total solution was injected per embryo. DNA was electroporated into ganglionic eminence progenitors using 5 mm tweezertrodes (BTX), or into lower rhombic lip progenitors using 3 mm tweezertrodes, applying 5 square pulses at 35 V and 25V respectively, 50 ms each with 950 ms intervals. The embryos were returned into the abdominal cavity, the muscle and skin were sutured, and the animal was monitored until fully recovered from the procedure.

Bioluminescence Imaging

Tumor-bearing mice received intraperitoneal administration of 100ul of 15mM AkaLumine-HCl (HY-112641A, MedChemExpress). Approximately 5 min after substrate administration, the mice were anesthetized with 2.5% isoflurane and 1.5 L O₂/min, and bioluminescence images were acquired using an IVIS Spectrum (PerkinElmer, Boston, MA, USA; RRID:SCR_018621). The following conditions were used for image acquisition: open for total bioluminescence, exposure time = 60 sec, binning = medium: 8, field of view = 13.5 × 13.5 cm, and f/stop = 1. Bioluminescent images were analyzed and exported using Living Image 4.7 software (PerkinElmer; RRID:SCR_014247).

Perfusion Fixation

Mice were transcardially perfused under terminal anesthesia with ice-cold PBS followed by ice-cold 4% paraformaldehyde. Brains were dissected and postfixed in 4% paraformaldehyde overnight before transferring to a gradient of 15–30% sucrose-PBS solutions overnight to ensure full cryoprotection of the tissue. After complete permeation with sucrose solution, brains were frozen in O.C.T. (Sakura) on dry ice and sectioned on a Leica 3050S cryostat (Leica Biosystems; RRID:SCR_020206) into 30 μm thick free-floating sections.

Single-nuclei RNA-seq Data Preprocessing

The 10x Genomics Cell Ranger pipeline (v3.1.0) was used to trim, demultiplex, and align reads to the reference genome, distinguish cells from background, and count unique transcripts per cell. Reads were aligned to mm10 customized with added vector sequences: piggyBac transposase (with CAG promoter), piggyBac LTR-3’, Akaluc, PQR, GFP, piggyBac LTR-5’, H3.3WTHA, P2A. Sequences for U6-Trp53gRNA-Cas9 and shATRX were also added to the reference used for KPP and KPPMPIK genotypes, while U6-Nf1gRNA-Cas9 were added to the reference for KNF.

Quality control of expression matrices and clustering of individual samples was performed in R (v3.6.1) with methods from the Seurat package (62) (v3.2.1). Genes with expression in fewer than 3 cells and cells with fewer than 200 detected genes were excluded. Low quality cells were filtered based on the following metrics: number of detected genes, number of unique molecular identifiers (UMIs), and percent of mitochondrial transcripts. Filtering thresholds are summarized in Supplementary Table 4.

Libraries were scaled to 10,000 UMIs per cell and log-normalized. Number of UMIs and mitochondrial content were regressed from the normalized gene counts and the residuals were z-scored gene-wise. Dimensionality reduction was performed using principal component analysis (PCA), with 100 PCs computed using the top 2,000 most variant genes. The first 30 PCs were then used as input for visualization in two dimensions using uniform manifold approximation and projection (UMAP) (arXiv 1802.03426v3) and as input for clustering. Cells were clustered using a shared nearest neighbor (SNN) modularity optimization algorithm (63) using the Louvain algorithm on a k-nearest neighbor graph with k = 20 and the resolution parameter set to 1. Cell cycle scores for G2/M and S phases were computed as implemented in Seurat using G2/M and S phase-associated signatures converted to murine homologs (64). Scores were obtained by calculating average expression of phase-associated genes and subtracting the average expression of control gene lists. Control gene lists were derived by binning genes in each input list into 24 bins according to expression levels and randomly selecting 100 control genes from within each expression bin. Post-clustering quality control was performed to remove clusters with low sequencing depth and high mitochondrial content.

Integration of Samples Within and Across Genotypes

To identify cell populations shared by KNF and KPP samples, samples were integrated using Harmony (65) (v1.0). Due to imbalanced cell numbers between genotypes, each KNF replicate was first downsampled to match the KPP replicates (N=15,537), such that proportions of clusters identified in individual sample analysis were maintained. Number of UMIs and mitochondrial content were regressed from the merged normalized gene counts, and PCA of variant genes was performed as described above. The first 30 PCs were aligned by sample using the following parameters: theta = 2, lambda = 1, sigma = 0.1. The “harmonized” PCs were used as input for visualization with UMAP, and for clustering using a shared nearest neighbor (SNN) modularity optimization algorithm using the Louvain algorithm on a k-nearest neighbor graph with k = 20 and resolution = 0.5. Integration of the KPPMPIK replicates was performed with the same workflow.

Consensus Cell Type Identification

To identify cell types in the single-nuclei RNA-sequencing data, we implemented an automated, reference-based annotation workflow based on four prediction methods: SciBet (66), SingleCellNet (67), SingleR (68), and Spearman correlation. A consensus annotation was assigned when at least two methods agreed without a tie. Cell type annotation was performed using two murine brain cell type references. Immune cells were first labeled at the cluster level in individual samples based on expression of canonical markers (Ptprc, Ly86, Hexb, Cd84, Ccr5, Cd247, Skap1, Zfp366, Btla). These clusters were annotated with a single-cell dataset of normal mouse brain CD45+ immune cells as the training reference (50). Non-immune cells were then annotated with a single-cell dataset of embryonic and early postnatal murine brain (27), with the consensus labels defined without age information. The consensus labels were summarized as broad cell classes: radial glial cells (RGC), oligodendrocyte precursor cells (OPC), glial progenitors, oligodendrocytes, astrocytes, ependymal cells, neuronal progenitors, neurons, vascular cells, immune cells, and other. This ontology is detailed in Supplementary Table 5. Only cells with a consensus prediction from agreement of at least two methods were used for downstream analysis.

Gene Set Enrichment of Human Tumor Signatures

To assess expression of human tumor-derived signatures in the murine models, we performed single-sample gene set enrichment analysis (ssGSEA) with code adapted from the GSVA package (v.1.27.0) as described (27). Enrichment in single cells was scored for four signatures derived from previous studies of bulk RNA sequencing and differential expression analysis between several tumor entities (13, 27). This includes signatures from comparing: H3.3 K27M and H3.1 K27M pontine high-grade glioma (N=20) (4, 27, 69, 70), H3.3 K27M high-grade glioma and EZHIP posterior fossa ependymoma (PFA) (N=28) (4, 27, 69–71), G34R/V and H3.3 K27M high-grade glioma (N=26) (13, 27). All differential expression analyses yielding the signatures were performed using DESeq2 (v1.14.1). Signatures were processed by removing mitochondrial genes (defined as beginning with “MT-”), selecting the top 100 genes based on logFC/variance(logFC), and converting to mouse homologs. The top 200 genes based on logFC/variance(logFC) were selected for the H3.1 K27M signature to obtain a number of mouse homologs comparable to the other signatures. The signatures used for ssGSEA scoring are found in Supplementary Table 6.

Transcription Factor Activity Inference

The activity of transcription factors (TFs) and their regulated genes were inferred from normalized expression data with the python implementation of SCENIC (51) (v.0.10.0) through the Singularity (v.3.8) container. Briefly, modules of genes coexpressed with TFs were detected through reconstructing a gene regulatory network. TFs with binding motifs enriched in their corresponding modules were retained, and genes containing the binding motifs as potential direct targets were retained in the modules. The AUCell algorithm (51) was then used to compute an activity score for each TF’s module in each cell. The above workflow was performed with the merged data across replicates for KNF and KPP genotypes separately.

Ex vivo NSC/GS Isolation and Culture

Tumor-bearing C57BL/6J mice were euthanized by CO₂ exposure. Brains were rapidly dissected in ice-cold dissociation medium (DM) containing: 20 mM glucose, 81.8 mM Na₂SO₄, 30 mM K₂SO₄, 5.8 mM MgCl₂, 250 μM CaCl₂, 1 mM HEPES, 160 μM NaOH, 0.8 mM kynurenic acid, 50 μM D-APV, 100 U/mL penicillin, 100 μg/mL streptomycin, 5 μg/ml plasmocin and 100 μg/ml primocin. Coronal sections were cut using a brain matrix and GFP⁺ (tumor) and GFP⁻ regions (stroma) were microdissected under an epifluorescence stereomicroscope (Leica M205, Leica Biosystems). Microdissected tissue was then enzymatically digested into a single-cell suspension using the Papain Dissociation System according to the manufacturer’s instructions (Worthington Biochemicals). The dissociated cell solution was separated on an OptiPrep density gradient to remove debris, following which GFP⁺ cells were sorted using a FACSAria II (BD Biosciences). Sorted cells were plated into NeuroCult NSC proliferation media (STEMCELL Technologies) containing 20 ng/ml EGF (Miltenyi Biotec), 20 ng/ml bFGF (Miltenyi Biotec), 10 ng/ml PDGF-AA (Shenandoah Biotechnology), 10 ng/ml PDGF-BB (Shenandoah Biotechnology) and 2 μg/ml Heparin (STEMCELL Technologies). Cells were grown as spheroids (gliomaspheres) using ultra-low attachment plates (Corning). Cell lines were authenticated using STR profiling and contamination with mycoplasma was tested using the qPCR-based PhoenixDX Mycoplasma Detection Kit. The results were recorded with the cell line authentication service at the CRUK Cambridge Institute. Cell lines at or below passage number 15 were used in experiments.

Patient-Derived Cell Culture

Patient-derived cell lines were provided by N. Jabado. BT416 (RRID:CVCL_WW64), BT869 (RRID:CVCL_C1MH), SU-DIPG-VI (RRID:CVCL_IT40), G477, and GBM002 cells were grown at 37°C, 5% CO₂ in stem cell media consisting of Neurocult NS-A Basal Medium (human, StemCell Technologies) and Neurocult NS-A Proliferation Supplement (human, StemCell Technologies). The media was supplemented with 20 ng/mL EGF (StemCell Technologies), 10 ng/mL bFGF (StemCell Technologies) and 0.2% heparin (Stem Cell Technologies). Cells were cultured in flasks pre-coated with 0.01% poly-L-ornithine and 1 mg/mL laminin diluted in PBS.

Drug Sensitivity Assays

Mouse tumor-derived GS cells and patient-derived cell lines were plated in 96-well plates and treated with 8 compounds at different concentrations, either for 3 days (avapritinib, FK-866, idasanutlin, trametinib) or 6 days (alpelisib, corin, GSK-J4, infigratinib). At the endpoint cell viability was measured using CellTiter-Glo (2.0 Promega). Relative luminescence units (RLU) for each well were normalized to the median RLU from the DMSO control wells as 100% viability. Three technical replicates per drug condition were performed as well as two independent biological replicates. IC50 values (drug concentration causing 50% inhibition of cell proliferation) were calculated using GraphPad Prism (RRID:SCR_002798), and the curves show the mean ± SD of the replicates per condition measured. Statistics are summarized in Supplementary Table 7. All compounds were purchased from MedChemExpress (MCE) and were diluted in DMSO to a final concentration of 10 mmol/L.

RNA Extraction and qPCR

RNA was extracted using the RNeasy Plus kit (Qiagen) according to the manufacturer’s instructions. RNA quality and quantification was carried out using a Nanodrop spectrophotometer (Thermo Fisher) or a Bioanalyzer (Agilent). 500 ng of total RNA was used for reverse transcription using TaqMan Reverse Transcription Reagents (Thermo Fisher) and real-time PCR was performed using the SsoAdvanced Universal SYBR Green Supermix (Bio-Rad) on a CFX96 Touch Real-Time PCR Detection System (Bio-Rad; RRID:SCR_018064). Primers are described in Supplementary Table 8. The 2-^ΔΔCT method was used to calculate relative gene expression levels and gene expression was normalized to mouse B2M levels.

Western Blotting

Total protein extraction was performed using lysis buffer (Tris-HCl 50 mM pH 8, NaCl 150 mM, Triton X-100 1%, NaF 100 mM, EDTA 1 mM, MgCl₂ 1 mM, Glycerol 10%) containing a protease inhibitor cocktail (Sigma-Aldrich). Equal amounts of total protein were subjected to Bis-Tris gel electrophoresis (Nupage) and then transferred to nitrocellulose membranes using iBlot2 Dry Blotting System (Thermo Fisher). The membranes were blocked with Intercept PBS Blocking Buffer (Li-Cor) and incubated with primary antibodies overnight at 4°C (Supplementary Table 9). After washing, membranes were incubated with Li-Cor Fluorescent IRDye secondary antibodies at RT (800 CW and 680 CW dyes) for 1 hour. Detection was performed with Li-Cor Odyssey FC Imaging System. Histone extraction was performed using the Histone Extraction Kit (Active Motif). Cells were lysed according to the One-step Extraction protocol provided by the manufacturer and histone lysates then underwent Bis-Tris gel electrophoresis, transfer and detection as described above.

Histology

For histological analysis of tumors, paraformaldehyde-fixed brains were embedded in paraffin, sectioned at a thickness of 3 μm, and processed for hematoxylin-eosin (HE) staining. Immunohistological detection of various antigens was performed on the Leica automated Bond-III platform in conjunction with the Polymer Refine Detection System (DS9800) using biotin-free secondary antibodies coupled to streptavidin-horseradish peroxidase and diaminobenzidine as a chromogen (Supplementary Table 9). Slides were then mounted using DPX Mountant for Histology (#06522, Sigma Aldrich), digitized on a Leica Aperio AT2, then viewed and exported using the Aperio ImageScope software (Leica Biosystems).

Immunofluorescence In Vivo

For immunofluorescence, free-floating sections were incubated in a blocking solution (10% goat or donkey serum, 3% BSA, 0.3% Triton-X in PBS) for 1 hr at RT and then incubated with primary antibodies at 4°C overnight (Supplementary Table 9). Sections were washed in PBS-Tween (0.05%) before addition of Hoechst 33342 and Alexa Fluor-conjugated secondary antibodies in blocking solution for 1 hr at RT. Following washing, sections were mounted in ProLong Diamond Antifade mountant (Thermo Fisher Scientific) and imaged on a confocal microscope (Leica SP8, Leica Biosystems).

Quantification of H3K27me3 and Ki67 in Tissue Sections

H3K27me3 fluorescence signal in HA⁺ cells was quantified using both ImageJ (RRID:SCR_003070) and Metamorph software (Molecular Devices; RRID:SCR_002368). 2–4 single-plane confocal images with a z-step of 2 μm were combined for fluorescence intensity analysis (total depth less than 10 μm). GFP⁺Ki67⁺ cells in the striatum and pons were quantified using ImageJ and QuPath (RRID:SCR_018257). Briefly, a DAPI mask was generated and used to quantify the total number of GFP⁺ tumor cells, and the total number of GFP⁺Ki67⁺ cells were then counted and used to calculate the mitotic index of the tumor. All cell counting and fluorescent intensity analyses were carried out on images from a minimum of 3 mice per condition.

Orthotopic Allotransplantation

8-week-old C57BL/6J mice (Charles River Laboratories, Kent, United Kingdom) were maintained under pathogen-free conditions, in individually ventilated cages, with food and water provided ad libitum. All procedures were approved by the UK Home Office (PPL PP2303899) and carried out in accordance with the Animals (Scientific Procedures) Act 1986. Analgesia was provided pre-operatively via a subcutaneous injection of Buprevet (0.1 mg/kg). Mice were anesthetized (2.5% Isoflurane and 1.5 L O₂/min) and the head of the mouse was fixed in a stereotactic frame (#51730, Stoelting Europe). A midline incision was made along the scalp to expose the skull and a small burr hole was made using a high-speed drill at defined stereotaxic coordinates; 0.5 mm anterior and 1.8 mm lateral from bregma for striatal targeting, and 0.8 mm posterior and 1.1 mm lateral from lambda for pontine targeting. A total of 1.5 – 3 ×10⁵ GS cells resuspended in 5μl of PBS were then delivered at a depth of 3.2 mm or 5 mm respectively using a 26-gauge (2mm, AS point style) Hamilton Syringe (Bonaduz, GR, Switzerland) at a controlled rate of 2 μl/min before the needle was then removed at a rate of 0.5 mm/min. The scalp was then closed with sutures and mice were placed in a heat chamber until fully recovered before being returned to their home cage.

In Vivo Drug Treatment

Animals were intracranially engrafted in either striatum or pons with GS cells and received either oral administration four days post-engraftment, or were implanted with an ALZET mini-osmotic pump (Durect) seven days post-engraftment. Mice were randomized based on bioluminescence signal prior to treatment initiation. For oral gavage treatment, avapritinib was dissolved in 10% dimethyl sulfoxide (DMSO) and 90% saline (containing 20% SBE-β-CD saline), and mice received either avapritinib (30mg/kg) or vehicle for 15 consecutive days via this route. For osmotic pump treatment, a 14-day ALZET pump (Model 2002) was connected to an ALZET brain infusion cannula (ALZET brain infusion kit 2), filled, and primed with the treatment or vehicle (elacridar in 20% SBE-β-CD saline) as per manufacturer’s instructions prior to implantation. The pump was inserted into a subcutaneous pocket slightly posterior to the scapulae and the cannula installed in the fourth ventricle. The cannula pedestal was secured in place using cyanoacrylate gel (Loctite 454) and pumps were removed following treatment as per manufacturer’s recommendation.