SOX9 Defines Distinct Populations of Cells in SHH Medulloblastoma but Is Not Required for Math1-Driven Tumor Formation

This article is featured in Highlights of This Issue, p. 1793

Hedgehog signaling plays a key role in regulating embryonic development and constitutive activation of the pathway has been linked to the onset of a wide variety of tumors (1). These include basal cell carcinoma (BCC), the most common of form of human skin cancer, and medulloblastoma, the most common malignant pediatric brain tumor. Advances in molecular and genetic profiling have revealed four distinct medulloblastoma molecular subgroups (WNT, SHH, Group 3 and Group 4), which can be further separated into 12 molecular subtypes (2), each associated with their own unique set of oncogenic drivers. The SHH-activated subgroup (SHH-MB) account for 25% to 30% of all medulloblastomas. A common feature to all BCCs and SHH-MB is prolonged activation of the Hedgehog signaling pathway, referred to as ligand-independent constitutive Hedgehog pathway activation (3), making these cancers excellent candidates for Hedgehog pathway inhibitors. Despite Hedgehog pathway inhibitors having shown promising results in the treatment of metastatic and locally advanced BCC (4), they have failed to produce a durable benefit in the treatment of medulloblastoma (5, 6). Hence, the search to discover new therapeutics for the treatment of Hedgehog-driven medulloblastoma.

Although the cellular of origin of SHH-MB was the subject of debate for many years, it is now well recognized that granule cell precursors (GCP) function as SHH-MB–initiating cells. GCPs originate from the rhombic lip and migrate along the outside of the cerebellum forming a transient structure referred to as the external granule layer (EGL; ref. 7). Actively proliferating GCPs reside in the outermost layer of the EGL (outer EGL, oEGL), a cell population defined by MATH1 expression. GCPs subsequently exit the cell cycle (occupying the middle EGL, mEGL) (8) and start to differentiate as they migrate toward the inner EGL (iEGL; ref. 9). Differentiating GCPs subsequently undergo radial migration through the molecular layer (ML) toward their final destination in the inner granular layer (IGL), where they differentiate into mature granule neurons (10). The majority of other cells that make up the cerebellum arise from neural stem cells (NSC) residing in the ventricular zone (11). Although Hedgehog pathway activation in multipotent NSCs also results in medulloblastoma, it has been shown tumor formation only occurs once the targeted NSCs have committed to the GCP lineage (12).

Several lines of evidence exist to suggest that the stem cell transcription factor, SOX9, functions as an oncogene in the genesis of Hedgehog-derived tumors. SOX9 has been shown to function as a downstream target of the Hedgehog signaling pathway (13–15) and elevated SOX9 expression has been identified in the majority of human BCC samples (16), and medulloblastoma (17), with more recent analysis revealing SOX9 to be subgroup specific, with high levels observed specifically in SHH-MB (18). SOX9 has also been shown to regulate SHH-MB self-renewal, promote specification of an SHH-MB subtype (18), and linked to increased brain stem and spinal cord metastasis and thus poor patient outcome (19). Taken together, these studies suggest that SOX9 functions as a critical transcription factor in human SHH-driven medulloblastoma, suggesting that development of specific SOX9 inhibitors or compounds that attenuate SOX9 expression, may prove useful as a potential therapeutic strategy for the treatment of SHH-MB (19). However, the functional requirement for SOX9 in the genesis of primary medulloblastoma remains to be determined.

We set out to functionally test whether SOX9 functions as an oncogene in SHH-MB, similar to that defined for BCC (20). Using multiple approaches, including scRNA analyses of human medulloblastoma, we define the presence of three distinct SOX9⁺ cell populations both in the normal developing cerebellum and in SHH-MB. In addition, we find that loss of Sox9 in MATH1⁺ GCPs, which function as the medulloblastoma-initiating cells of SHH-MBs (12, 21), has no adverse effect on cerebellar development or SHH-MB tumor formation, thereby suggesting that if SOX9 is required for medulloblastoma initiation or growth then it must be acting via a cerebellar cell type other than MATH1⁺ GCPs.

Math1Cre (22), GFAPCre (23), Ptch1^lox (24), Ptch1^lacZ (25), Sox9^lox (26) and Math1-GFP (27) mice have been previously described and were all maintained on a C57BL/6 background. Animals were treated in accordance with the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes. The University of Queensland Animal Ethics Committee approved all experiments. In all our time analyzing the phenotypes described within this manuscript we have always observed Ptch1 heterozygous to be nonphenotypic and indistinguishable from wild-type animals. “Control”, as referred to in the manuscript, therefore refer to either wild-type or Ptch1 heterozygous embryos or adult mice.

Patient-derived xenograft (PDX) mouse models were originally generated in the Olson laboratory (Fred Hutchinson Cancer Research Center) using pediatric patient tumor tissue obtained from Seattle Children's Hospital with approval from the Institutional Review Board. PDX lines were generated by implanting tumor cells directly into the cerebellum of immunocompromised mice (NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/Szj (NSG), or athymic nude) within hours of surgical removal from the patient, and then propagating them from mouse to mouse exclusively without in vitro passaging. PDX lines screened include: Med-1712FH (SHH) and Med-411FH (Group3), maintained in our laboratory; and Med-210FH (Group 3) and Med-2312FH (Group 4), maintained by the Olson laboratory.

Experimental MB tumor animals were anaesthetized, transcardial perfusion performed and the brain samples post-fixed in 4% PFA and embedded in paraffin. Math1:GFP brain samples were perfused and fixed before embedding in OCT. Antibody markers were analyzed via standard immunofluorescence techniques using the following antibodies: KI67 (ab15580: abcam), Tag1 (4D7: Hybridoma Bank), DCX (AB2253: Merck), NEUN (MAB377: Merck), PAX6 (AB2237: Merck), SOX9 1:100 (BAF3075: R&D systems), SOX9 1:100 (AB5535: Chemicon), SOX9 1:100 (ab76997: abcam), GFP 1:1000 (ab13970: abcam), GFAP (Z0334: Dakocytomation), GFAP (MAB360: Chemicon), SOX2 (MAB4343: Merck), SOX2 (ab97959: abcam). Primary antibodies were detected with species-specific secondary antibodies conjugated to Alexa Fluor 488, 594 or 647 1:250 (Molecular Probes).

Imaging analysis of all microscopy data was performed using ImageJ software (version 1.53; National Institutes of Health, MD, RRID:SCR_003070). A line was drawn across multiple cells starting from a high intensity SOX9 position within the image. An intensity profile plot returned the fluorescence intensity of each pixel along the line. The relative intensity of each fluorescence channel at each pixel was assessed to determine the relative expression of each fluorescently tagged protein.

GSE85217 dataset from the GEO database including 763 medulloblastoma samples was used to assess SOX9 mRNA expression on R2: Genomics Analysis and Visualization Platform (http://r2.amc.nl). Two gene expression datasets were downloaded to illustrate tumor type and subtype context. Tumors of the central nervous system and normal brain (28), GEO accession GSE50161 and a medulloblastoma subtype study (2), GSE85217. The reference data were downloaded as raw CEL files and preprocessed and RMA normalized using the R package oligo. Cross-platform normalization and principal component analysis were performed equivalent to a previous study by (29) using the Metagene code for cross-platform, cross-species characterization of global transcriptional states (30).

Math1-GFP pups were collected at P7 and GCPs were isolated and stained with TAG1 primary antibody and sorted using a FACSAria cell sorter (BD Biosciences). RNA was extracted using the RNeasy micro kit (QIAGEN, 7404).

RNA sequencing (RNA-Seq) libraries were prepared using the Illumina TruSeq stranded total RNA LT (with Ribo-Zero Gold) Set A (Illumina, RS-122–2301) and Set B (Illumina, RS-122–2302). Sequencing was performed using the Illumina NextSeq500 (NextSeq control software V1.4/Real Time Analysis V2.1). All RNA sequence reads were mapped against the mouse genome assembly mm9 using STAR v2.5.3a (RRID:SCR_004463). The alignment was performed with a two-pass approach where splice junctions discovered during a first alignment guides the forming of a final second alignment (31). Unmapped reads were realigned using Bowtie2 v2.3.4.3 (RRID:SCR_016368) before duplicate reads were removed with Piccard v2.10.3. The expression profile was quantified with Subread v1.5.2 (RRID:SCR_009803) and TPM normalized using R v3.6.0. Differential expression analysis of RNA-seq data obtained from the oEGL and the mEGL GNP population was performed using DESeq (RRID:SCR_000154) and edgeR (RRID:SCR_012802).

Data were downloaded from the Gene Expression Omnibus (RRID:SCR_005012) (accession GSE119926) and processed and normalized using Seurat. A cell was considered positive for a marker if the normalized expression value (normalized log-counts) of the gene in that cell was greater than 0.

MR images were acquired on a Bruker 7T Clinscan interfaced with a Siemens spectrometer running Numaris 4 VB17 using a 23 mm mouse head volume coil. T₂-weighted (T2) imaging sequences were performed at resolution 0.078 × 0.078 × 0.700 mm³; TR/TE 2750/45 ms/ms; flip angle 180°. Tumor development was monitored from P40 through to tumor endpoint. T2 maps were imported into Horus and binary masks of the region of interest (ROI) were manually delineated around the entire cerebellar structure to calculate tumor size.

To examine SOX9 expression in human MB we interrogated the microarray expression data of a cohort of 763 human patient MB samples encompassing all 12 MB subtypes. These data reveal higher levels of SOX9 transcript expression across all WNT and SHH subtypes than group 3 or group 4 subtypes (Fig. 1A). In support of this observation, immunofluorescence analysis of human PDX tumors revealed prominent SOX9 protein expression in the bulk of SHH-MB (Fig. 1B) and little SOX9 immunoreactivity in group 3 or group 4 PDX tumors (Fig. 1C and D). In order to examine SOX9 expression in murine medulloblastoma, we screened a variety of SHH-MB genetic mouse models including GCP-targeted early onset (Math1Cre:Ptch1^lox/lox) and late onset (Math1Cre:Ptch1^lox/+) tumors, NSC-targeted early onset (GFAPCre:Ptch1^lox/lox) and late onset (GFAPCre:Ptch1^lox/+) tumors and the conventional adult-onset medulloblastoma model (Ptch1^lacZ). We consistently observed widespread distribution of SOX9 expression throughout the bulk of all GCP-targeted tumors (Fig. 1E–J; Supplementary Table S1), NSC-targeted tumors (Supplementary Table S1), and S60% of Ptch1^lacZ-derived tumors (Supplementary Table S1). These data clearly indicate that, similar to human SHH-MB, SOX9 expression is associated with murine SHH-MB. We subsequently interrogated the 763 human patient medulloblastoma cohort dataset in search for prognostic data linked to differences in SOX9 expression, however, no association was found in survival, metastasis status, or any other prognostic indicator annotated within this dataset (data not shown).

Upon close examination of all murine SHH-MB tumors we observed varying degrees of SOX9 fluorescence intensity (Figs. 1E–J and 2E). Similar fluorescence intensity and distribution patterns were observed using three different SOX9 antibodies. Cells expressing relatively high levels of SOX9 fluorescence (SOX9^high) were observed dispersed throughout the tumor (Fig. 1E and F; Supplementary Fig. S3A and B, E and F). SOX9^high cells co-expressed the glial differentiation marker GFAP (Fig. 1G) and mature astrocyte marker S100β (Fig. 1H; Supplementary Fig. S3E–S3H), indicating a glial/astrocyte cell fate. In contrast, relatively low levels of SOX9 fluorescence (SOX9^low) were observed in palisading tumor nuclei across the bulk of the tumor (denoted by dotted lines in Fig. 1E and F: increased exposure of identical frames depicted in Fig. 1E and F, respectively). This morphologic pattern, referred to as rhythmic palisades, is commonly encountered in a wide range of CNS neoplasms (32). Palisading SOX9^low cells did not exhibit evidence of glial cell fate (GFAP/S100β; Fig. 1G and H), early neuronal specification (PAX6; Fig. 1I), postmitotic/migrating GCP cell fate (DCX; Fig. 1J) or mature granule neuron fate (NEUN; data not shown). Taken together, these data reveal that SHH-MB tumors consist of three distinct populations, SOX9^high, SOX9^low and SOX9⁻ cells. We subsequently calculated the fluorescence intensity of SOX9 expressed by individual cells in SHH-MB (Supplementary Fig. S3A, S3B, S3E, and S3F). This quantifiable data confirmed the existence of three distinct SOX9 populations and allowed us to define the fluorescence intensity that define a cell as SOX9^high, SOX9^low or SOX9⁻.

The single-cell transcriptomes of 25 patients with medulloblastoma (33) were analyzed for their level of SOX9, GFAP, and ATOH1 (MATH1) gene expression. Cells expressing any combinations of the three marker genes were identified and revealed two SOX9 populations unique to SHH-MB (SOX9⁺MATH1⁺ and SOX9⁺MATH1⁺GFAP⁺) and another SOX9 population present in SHH-MB yet largely absent in remaining medulloblastoma subgroups (SOX9⁺GFAP⁺; Fig. 1K). Specifically, approximately 25% of SHH-MB tumor mass cells were SOX9⁺MATH1⁺, representing tumor cells reminiscent of proliferating oEGL GCPs; approximately 16% of cells coexpress SOX9⁺GFAP⁺, representing tumor astrocytes; and approximately 11% of the tumor mass express SOX9⁺MATH1⁺GFAP⁺, reminiscent of the recently described tumor-derived transdifferentiated astrocytes (34). Further analyses of the relative expression level of SOX9 in each of cell population revealed SOX9⁺GFAP⁺ and SOX9⁺MATH1⁺GFAP⁺ express higher levels of SOX9 than SOX9⁺MATH1⁺ cells (Fig. 1l). We next sought to identify whether SOX9⁺GFAP⁺, SOX9⁺MATH1⁺GFAP⁺, or SOX9⁺MATH1⁺ populations correlate to putative stem cell compartments by screening for PROM1 (CD133), FUT4 (CD15), NGFR (CD271) or SOX2 gene expression. However, we observed no evidence of stem-cell marker enrichment in any of the three SOX9 compartments described (Supplementary Fig. S4).

To characterize the developmental origins of SOX9^high and SOX9^low tumor cells, we set out to define the temporal and spatial expression profiles of both SOX9 cell populations during cerebellar development. We noted that the SOX9⁺ cell populations previously described by Schedl and colleagues (14) comprising of NSCs residing in the ventricular zone of wild-type E13.5 (Supplementary Fig. S1A), E16.5 (Supplementary Fig. S1B), and E18.5 cerebellum (Fig. 2A, red box); scattered glial cells of E18.5 cerebellum (Fig. 2A, yellow box); and Bergmann glia of postnatal (Fig. 2B) and mature cerebellum (data not shown), all correspond to SOX9^high immunoreactivity (defined by their quantified fluorescence intensity, data not shown). In contrast, cells calculated to express SOX9^low immunoreactivity featured a restricted temporal-spatial expression pattern. SOX9^low cells were exclusively expressed in postnatal GCPs located in the EGL (Fig. 2C). More specifically, SOX9^low cells appear to spatially reside in the oEGL, excluded from TAG1-expressing cells of the mEGL (data not shown) and DCX expressing cells of the iEGL (Fig. 2D). To further confirm that SOX9^low is restricted to the oEGL, we screened for expression of the proliferation marker KI67 and observed SOX9^low expression was indeed confined to proliferating GCPs (Fig. 2F–H). Of note, we observed regions within Math1Cre:Ptch1^lox/lox tumors (Fig. 2E) that revealed a SOX9 expression pattern characteristic of wild-type P7 cerebellum (Fig. 2D).

Although MATH1 is the standard marker of GCPs, a lack of MATH1-immunostaining antibodies prohibits its use in immuno-screening procedures. We therefore made use of Math1-GFP mice, in which GFP expression is driven by the endogenous Math1 promoter, to isolate oEGL and mEGL cells from P7 cerebellum via FACS. Proliferating GCPs in the oEGL were defined as expressing high levels of MATH1 (MATH1-GFP^high) and cells of the mEGL defined as MATH1-GFP^low/TAG1-positive (Fig. 2J). We subsequently performed RNA-seq and identified 6333 differentially expressed genes between the two GCP populations. More specifically, we identified that Sox9 was expressed at significantly higher levels in the oEGL (1142 ± 347.7 sequence hits) than the mEGL (215.5 ± 22.28 sequence hits; Fig. 2K). To spatially verify that SOX9^low expression corresponded with a MATH1-expressing GCP cell fate, we screened the cerebellum of P7 Math1-GFP mice and observed SOX9^low expression completely overlapped with Math1-driven GFP expression (Fig. 2I). To functionally address whether advanced SHH-MB tumors exhibit a GCP cell fate, we crossed Math1-GFP mice to Math1Cre:Ptch1^lox/lox mice and observed GFP expression in the majority of tumor cells (Supplementary Fig. S2) and confirmed that only SOX9^low cells coexpress Math1-GFP (Supplementary Fig. S3A–S3D) and KI67 (data not shown). Overall, these data indicate that SOX9^low acts as a surrogate marker for MATH1 protein expression in GCPs, suggesting it can prove useful in IHC staining procedures to mark cells of a proliferative GCP fate.

In order to address the functional role of SOX9^low expression in the developing cerebellum we specifically ablated Sox9 in GCPs using Math1Cre. We observed no defects in the morphological structure (Fig. 3A) or granule neuron lineage specification (PAX6; Fig. 3B) in E18.5 Math1Cre:Sox9^lox/lox cerebellum. Despite efficient deletion of SOX9 from the GCP population of P7 Math1Cre:Sox9^lox/lox cerebellum (Fig. 3C and C’), GCP proliferation (KI67: Fig. 3D), early GCP differentiation (TAG1: Fig. 3E) and late stage GCP maturation (NEUN: Fig. 3F) were all indistinguishable between control and Math1Cre:Sox9^lox/lox cerebellum. These results indicate that SOX9^low expression in developing GCPs does not play a role in regulating the GCP lineage.

We next investigated whether SOX2, previously reported to be expressed in early developing GCPs (35), was dysregulated in Math1Cre:Sox9^lox/lox cerebellum, thereby potentially concealing the effects of SOX9 loss due to functional redundancy between the two Sox family members. However, we only observed SOX2 staining in the NSCs of the VZ, Bergmann glia and cerebellar white matter, with no SOX2 expression detected in the GCP population of E13.5, E16.5, and P7 cerebellum (Supplementary Fig. S1A–S1D). Furthermore, expression of SOX2 in the EGL of Math1Cre:Sox9^lox/lox mice was not detected (data not shown).

We next asked whether SOX9 is required as a downstream effector of Hedgehog signaling in SHH-MB tumor cells by specifically ablating Sox9 in MATH1⁺GCPs, which function as the cellular origin of SHH-MB. Immunofluorescence staining (Fig. 4A and B) and quantification of SOX9 fluorescence intensity (data not shown) confirmed efficient deletion of SOX9^low within the bulk of Math1Cre:Ptch1^lox/lox;Sox9^lox/lox tumor tissue compared to Math1Cre:Ptch1^lox/lox. The SOX9⁺ cells remaining in Math1Cre:Ptch1^lox/lox;Sox9^lox/lox tumor tissue were confirmed to be SOX9^high/S100β+ astrocytes (Fig. 4B’). By 75 days, 100% of Math1Cre:Ptch1^lox/lox (SOX9 wildtype) mice had succumbed to MB and were euthanized due to clinical manifestations of the disease, including head swelling, weight loss, poor grooming, inactivity and circling (Fig. 4C, black line). Similarly, 100% of Math1Cre:Ptch1^lox/lox;Sox9^lox/lox (SOX9 ablated) mice succumbed to MB by 65 days (Fig. 4C, pink line). Both Math1Cre:Ptch1^lox/lox and Math1Cre:Ptch1^lox/lox;Sox9^lox/lox tumors presented with comparable mitotic indices (P = 0.30), evident by KI67 expression (Fig. 4D–F) and a similar spatial distribution of DCX, a marker of migrating GCPs, TAG1, a marker of differentiating GCPs and NEUN, a marker of mature granule neurons (data not shown). MRI analysis of Math1Cre:Ptch1^lox/lox and Math1Cre:Ptch1^lox/lox;Sox9^lox/lox tumors from 6 to 8 weeks of age revealed no statistical difference in tumor growth rate (Fig. 4G) and overall tumor size (Fig. 4H). RNA sequencing (RNA-seq) data was compared against a data set encompassing a variety of different CNS tumors and PCA analysis revealed both Math1Cre:Ptch1^lox/lox;Sox9^lox/lox tumors cluster with medulloblastomas, and that they both continue to exhibit a SHH subtype profile (Fig. 4I). Moreover, we observed no differential gene expression across 70 previously identified GCP and MB Gli1 target genes (Supplementary Table S3; (36)), indicating that Sox9 ablation does not perturb the Hedgehog signaling activity of SHH-MB tumor cells. Taken together, these data demonstrate that ablation of SOX9 activity within SHH-MB tumor cells does not affect MB formation.

Given that genetic loss of Sox9 in tumor cells did not impede MB formation, we set out to address whether expression of SOX9 within the tumor microenvironment (SOX9^high astrocytes) is required for MB formation. We have previously utilized the GFAPCre mouse model system, whereby gene deletion will be targeted to NSCs and all glial/astrocyte lineages that constitute the tumor and as previously reported, all GFAPCre:Ptch1^lox/lox tumor mice succumb to MB within 28 days (12). Accordingly, since SOX9 expression in the cerebellum and medulloblastoma is also present in cells of the GFAP⁺ lineage we set out address the role of SOX9 in GFAP⁺ cells. However, we observed that GFAPCre-driven ablation of Sox9 was not compatible with life, with no live GFAPCre:Sox9^lox/lox pups recorded (Supplementary Table S2).

Previous studies concluded that SOX9 expression in murine medulloblastoma is dependent on cellular origin, with SOX9 only expressed in NSC-targeted transplantation medulloblastoma models, and not expressed in GCP-targeted transplantation MB models (37). However, our data clearly demonstrate the presence of scattered SOX9^high glial cells and rhythmic palisading structures of SOX9^low GCPs in GCP-targeted, NSC-targeted and Ptch1^lacZ SHH-MB mouse models. Furthermore, we demonstrate that SOX9^low can be used as a surrogate marker for MATH1 in the developing cerebellum and that advanced SHH-MB tumors consist largely of morphological rhythmic palisade structures, characteristic of superfluous folding of persistent EGL. Our scRNA data analysis also revealed the presence of three distinct SOX9⁺ cell populations present in SHH-MB that were noticeably absent in remaining medulloblastoma subgroups. Taken together, these data show the association of relatively high overall SOX9 mRNA transcript expression in SHH tumors is due to the fact that SHH-MB tumors are largely comprised of SOX9^low GCPs and SOX9^high astrocytes.

To our knowledge, this is the first report to describe the presence of two distinct SOX9-expressing cell populations in the developing cerebellum and in SHH-MB. The dual SOX9 expression pattern of SOX9^high and SOX9^low observed in the cerebellum is reminiscent of SOX2 expression in the forebrain where SOX2 is differentially expressed between radial glial and intermediate progenitor populations (38). Our data and other studies (39) have demonstrated that complete knockout of Sox9 in the cerebellum, driven by En1-Cre, Pax2-Cre or GFAPCre, results in lethality shortly after birth alongside defects in the neurogenesis-to-gliogenesis switch. In contrast, Math1Cre-directed ablation of Sox9 in the developing cerebellum results in viable, fertile mice with no discernible granule neuron lineage defects. A common theme among SOX proteins is functional redundancy, with SOX9 shown to exhibit redundant activity with other members of the SOXE subgroup, SOX8 and SOX10 (40–42). However, interrogation of our RNA-seq data did not reveal any evidence of endogenous Sox8 or Sox10 transcripts in P7 GCPs (data not shown), arguing against the absence of phenotype being attributable to possible SOXE family functional redundancy. Previous studies have reported SOX2 expression in the early rhombic lip and EGL of murine cerebellum at E13.5 and E14.5 (35), suggesting the lack of phenotype observed following SOX9 ablation might be attributable to SOX2 functional redundancy. However, our immunofluorescence studies using two separate SOX2 antibodies, could not confirm such expression data. Thus, despite expression of SOX9 in actively proliferating postnatal GCPs, our loss-of-function data reveal that SOX9 does not play an essential role in the specification, proliferation or differentiation program of GCPs.

No disclosures were reported.

C. Adolphe: Conceptualization, formal analysis, investigation, writing–original draft, writing–review and editing. A. Millar: Investigation. M. Kojic: Investigation, writing–review and editing. D.S. Barkauskas: Formal analysis, investigation, methodology. A. Sundstrom: Formal analysis, investigation, writing–review and editing. F.J. Swartling: Supervision, investigation, writing–review and editing. S. Hediyeh-zadeh: Formal analysis, investigation, writing–review and editing. C. Tan: Formal analysis, methodology. M.J. Davis: Formal analysis, supervision, investigation, writing–review and editing. L.A. Genovesi: Formal analysis, investigation, writing–review and editing. B.J. Wainwright: Conceptualization, resources, formal analysis, supervision, funding acquisition, writing–original draft.

The authors would like to thank Pengxiang Ji for providing Med411 and Med1712 PDX tissue, and Madison Nakamoto and Prof. Jim Olsen for providing Med210 and Med2312 PDX tissue. Sox9^lox mice were kindly provided by Prof. Gerd Scherer. Confocal microscopy was performed at the Australian Cancer Research Foundation Dynamic Imaging Centre for Cancer Biology. This work was supported by a Strategic Research grant from the University of Queensland.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.