CRISPR/Cas9-Mediated Trp53 and Brca2 Knockout to Generate Improved Murine Models of Ovarian High-Grade Serous Carcinoma

Ovarian high-grade serous carcinoma (HGSC) is the commonest subtype of human ovarian cancer (1), and long-term survival remains poor (2). Mutation in TP53 is essentially universal (3, 4), whereas germline mutation in BRCA1 or BRCA2 is observed in approximately 15% cases (5, 6). These mutations impair the ability to repair DNA double-strand breaks (DSB) via homologous recombination (HR). The Cancer Genome Atlas consortium suggested that HR defects may be present in approximately 50% HGSC through a variety of additional mechanisms, including somatic BRCA1/2 mutation and epigenetic loss of BRCA1 expression (4). Another key development has been the finding that HGSC may arise in the fimbriae of the fallopian tube, rather than the ovarian surface epithelium (OSE; ref. 7). Early serous tubal intra-epithelial carcinoma (STIC) lesions in the tubal fimbriae are observed in up to 10% of women with germline BRCA1/2 mutations undergoing prophylactic salpingo-oophorectomy (8). Although a tubal origin of HGSC is widely accepted (9), STIC lesions are only seen in around half of HGSC cases (10). Recent data from mouse models (11, 12) and clinical HGSC samples (13) suggest that the ovary itself may still be crucial in the pathogenesis of HGSC, and it is possible that there could be more than one origin of this tumor.

The absence of reliable murine models has been a major impediment in HGSC research (14). This is particularly important for investigation of the immune microenvironment. The presence of tumor-infiltrating CD8 T lymphocytes (TIL) and tertiary intraepithelial lymphoid aggregates are both associated with improved prognosis in HGSC (15, 16), whereas intratumoral immunosuppressive myeloid and lymphoid cells (17, 18) are associated with poor prognosis. However, it is unclear whether or how specific genomic events in HGSC influence the immune microenvironment.

The ID8 model, first described in 2000 (19), remains the only transplantable murine model of ovarian cancer routinely available. Whole ovaries from C57Bl/6 mice were trypsin-digested, and the dissociated cells passaged in vitro, initially in the presence of EGF. After approximately 20 passages, cells lost contact inhibition, and ten separate clones were derived, of which ID8 is the most widely used. Following intraperitoneal injection of ID8 in syngeneic mice, diffuse peritoneal carcinomatosis, with blood-stained ascites, develops in approximately 110 days (19). Over 100 publications have utilized the ID8 model, but none has characterized it in light of current understanding of human ovarian cancer biology.

Here, we show that parental ID8 lacks mutations in Trp53, Brca1, and Brca2, and demonstrates HR competence in functional assays. We have used CRISPR/Cas9 gene editing technology to generate single (Trp53) and double (Trp53;Brca2) knockout derivatives of ID8 and evaluated their utility as a model of human HGSC. In particular, we show that loss of individual genes results in significant alterations in immune cell infiltration into the tumor microenvironment.

ID8 cells, obtained from Dr. Katherine Roby (University of Kansas Medical Center, Kansas City, KS), were cultured in DMEM supplemented with 4% fetal calf serum, 100 μg/mL penicillin, 100 μg/mL streptomycin, and ITS (5 μg/mL insulin, 5 μg/mL transferrin, and 5 ng/mL sodium selenite). As ID8 was obtained directly from their original source, separate short tandem repeat validation was not performed. For cytotoxicity assays, cells were plated onto 24 well plates (3 × 10³ cells/well) in triplicate. Survival was assessed by MTT assay (Nutlin-3) or sulphorhodamine B assay (rucaparib) after 72 hours.

Whole-exome sequencing and analysis were performed by Beckman Coulter Genomics. Full details are given in Supplementary Methods. Summary results are presented in Supplementary Tables S1 to S3. Primary sequencing data (BAM and VCF files) are available in the ArrayExpress database (http://www.ebi.ac.uk/arrayexpress) under accession number E-MTAB-4663.

Confirmatory Sanger sequencing of exons 2 to 9 of Trp53 was also performed on genomic DNA extracted from 10⁷ parental ID8 cells in log-growth phase as well as from five separate ID8-microdissected tumors extracted from female C57Bl/6 mice after 110 days of intraperitoneal growth.

Two open-access software programs, CHOPCHOP (https://chopchop.rc.fas.harvard.edu/) and CRISPR design (http://crispr.mit.edu/), were used to design guide RNAs (gRNA) targeted to Trp53 exon 5 and Brca2 exon 3. Three guides were designed per gene. Annealed oligonucleotides were ligated into BbsI-linearized pSpCas9(BB)-2A-Puro [PX459 (20), gifts from Feng Zhang via Addgene]. Version 1 (Addgene # 48139) was employed for Trp53 deletion and version 2.0 (Addgene # 62988) for Brca2. All plasmids were sequenced to confirm successful ligation.

ID8 cells (4 × 10⁵) were plated overnight in antibiotic-free medium, and transfected with 4 μg PX459 using Lipofectamine 2000, selected under puromycin (2.5 μg/mL) for 48 hours and plated onto 96-well plates (10 cells/mL). Single-cell colonies were expanded for DNA extraction, protein extraction, and cryopreservation.

PCR primers spanning potential sites of deletion were designed (Trp53: F 5′-cttccctcacattcctttcttg-3′; R 5′-gctgttaaagtagaccctgggc-3′, Brca2: F 5′-catggagggagtcacctttg-3′, R 5′-gctctggctgtctcgaactt-3′). Clones with large PCR deletions were selected for subsequent analysis. Remaining clones were screened using the Surveyor Nuclease Assay (Integrated DNA Technology). Mutations were confirmed by Sanger sequencing. All sequence alignment was performed using MAFFT version 7 (http://mafft.cbrc.jp/).

Note that 15 μg of total protein was electrophoresed at 140 V for 1 hour, transferred onto nitrocellulose, and blocked in 5% nonfat milk. Antibodies used were as follows: p53 (CM5, Novacastra and Ab26, Abcam) and β-actin (Sigma; A1978). Membranes were exposed on a Chemi-doc MP (Biorad) with ECL (β actin) and ECL prime (p53). For array experiments, 5 mL supernatant was collected from 10⁶ cells plated for 16 hours on a 10-cm plate, and centrifuged (2,000 rpm for 5 minutes). Mouse Cytokine Antibody Array C1 (C-series) was blotted according to the manufacturer's instructions (RayBiotech, Inc.).

RNA was extracted from 3 × 10⁵ cells in log-growth phase, and 2 μg was reverse transcribed (Applied Biosystems High Capacity Reverse Transcription Kit). cDNA (50 ng) was amplified using 10x iTaq qRT-PCR master mix (Biorad), 20X primer probe mix in a total of 20 μL under the following cycle parameters: 2 minutes: 50°C, 10 minutes: 95°C, 40X (15 seconds: 95°C; 1 minute: 60°C). All qRT-PCR primers were purchased from Applied Biosystems and Trp53 custom made from Sigma Aldrich: F 5′-catcacctcactgcatggac-3′, R 5′-cttcacttgggccttcaaaa-3′, probe 5′-ccccaggatgttgaggagt-3′. Values were normalized to Rpl-34.

Cells were seeded on coverslips and treated with rucaparib (10 μmol/L for 24 hours) or irradiated (10 Gy), permeabilized with 0.2% Triton (Sigma) in PBS for 1 minute, then fixed in 3% paraformaldehyde, and 2% sucrose for 10 minutes. Cells were stained with anti-γH2AX antibody (Millipore) and costained with anti-Rad51 (Santa Cruz Biotechnology) antibody for 45 minutes at 37°C. Cells were costained with DAPI. Coverslips were mounted on slides and images captured using a Zeiss 710 confocal microscope and foci counted using ImageJ software. Cells were deemed HR competent if the number of Rad51 foci more than doubled in the presence of DNA DSB damage (≥2-fold increase in number of γH2AX foci) as previously described (21).

All experiments were complied with the UK welfare guidelines (22) and were conducted under specific personal and project license authority. Cells (5 × 10⁶) were inoculated intraperitoneally in 6- to 8-week-old female C57Bl/6 mice (Charles River Laboratories). Mice were monitored regularly and killed upon reaching UK Home Office limits. All decisions about animal welfare and experiment endpoints were made by one of us (D. Athineos) independently of main study investigators to prevent bias. Ascites was collected, and all visible tumor deposits dissected out. Half of tumor material was snap frozen (dry ice) and half fixed in neutral-buffered 4% paraformaldehyde. Sections (5 μm) from formalin-fixed paraffin-embedded tumors were stained (Dako Autostainer, Dako) and quantified as detailed in Supplementary Methods.

Ascites was centrifuged (2,200 rpm, 10 minutes) and supernatant stored at –80°C. The cell pellet was treated with red blood cell lysis buffer (Sigma Aldrich—5 minutes, room temperature), resuspended in 10 mL PBS, recentrifuged, and stored at –80°C in FBS/10% DMSO. Solid tumor deposits in ice-cold PBS/protease inhibitor solution were dissected into pieces less than 1 mm diameter using a scalpel and digested at 37°C for 30 minutes (0.012% w/v collagenase type XI, 0.012% w/v dispase, 0.25% Trypsin in 0.1% BSA in RPMI.) Note that 10 mL of 0.1% BSA/RPMI was added, and the tubes shaken vigorously followed by 100-μm filtration. Cells were centrifuged for 5 minutes at 1,500 rpm, resuspended in PBS containing 5% FBS, and then counted. Flow cytometry and gating strategies are presented in Supplementary Methods.

To assess the genomic landscape of ID8, we undertook whole-exome sequencing, covering approximately 194,000 exons at a mean depth of 400× with 90% exons sequenced >50×. Approximately 6,000 variants were identified, the vast majority of which were nonfunctional. Functional alterations (nonsynonymous coding, stop-gain, and frameshift) were identified in approximately 100 genes, the large majority of which were nonsynonymous coding (See Supplementary Tables S1–S3 for summary and list of all functional alterations). However, we were unable to identify any functional mutations in genes characteristic of HGSC (Trp53, Brca1, Brca2, Nf1, Rb1), and the absence of Trp53 mutations was confirmed by Sanger sequencing (data not shown). In addition, mutations typical of clear cell (Arid1A, Pik3ca), low-grade serous (Braf), endometrioid (Ctnnb1), and mucinous (Kras) carcinomas were also notably absent. We did identify a mutation in Adamts3 (c.1089C>T; pV199I)—a recent analysis of The Cancer Genome Atlas (TCGA) data identified that mutations in ADAMTS genes were associated with platinum sensitivity as well as improved progression-free and overall survival in HGSC (23). In addition, mutation in Gabra6 (c.347T>G; pE22D) was identified. GABRA6 was one of the genes mutated at statistically significant frequency in TCGA analysis—however, transcription was absent in all TCGA tumors, suggesting that GABRA6 mutation is of minimal clinical relevance (4).

Given the centrality of TP53 mutations in HGSC, we also assessed p53 function in ID8. There was a robust increase in p53 protein following treatment with cisplatin and the MDM2 inhibitor Nutlin-3 (Fig. 1A), with marked increases in Cdkn1A (p21) transcription within 4 hours of cisplatin treatment (Fig. 1B), indicating that p53 remains transcriptionally active. Sanger sequencing of intraperitoneal ID8 tumors showed no Trp53 mutation in any tumor, including common hotspot mutations sites (R172, Y217, R245, R270—Fig. 1C), whereas immunohistochemistry (IHC) confirmed a wild-type pattern of p53 expression (Fig. 1D). IHC examination of typical HGSC markers indicated that tumors were strongly and diffusely positive for WT1, but negative for Pax8 (Fig. 1D).

The whole-exome sequencing data identified no functional abnormalities in Brca1, Brca2, or other HR genes. We confirmed that parental ID8 cells were able to form Rad51 foci in response to DNA DSB damage induced by irradiation as well as the PARP inhibitor rucaparib (Fig. 1E), and fulfilled the criteria of competent HR as previously described (21). Together, these data suggest that parental ID8 is poorly representative of human HGSC.

Three separate gRNA constructs targeted to exon 5 of Trp53 (Fig. 2A) were cloned into the PX459 plasmid. Following transfection and screening (Fig. 2B and C), clones were derived from all three guides (F3—guide G; A2—guide K; C7 and M20—guide R), all of which contained bi-allelic deletions in Trp53 exon 5 (Supplementary Fig. S1), ranging from 4 bp (clone M20) to 280 bp (clone A2). All four null clones showed absent basal p53 expression by immunoblot (Fig. 2D), with significantly reduced basal transcription of Trp53 (Fig. 2E), Cdkn1a, and Bax (Fig. 2F). There was also no increase in p53 expression following treatment with cisplatin or Nutlin-3 (Fig. 2G), nor an increase in Cdkn1a transcription following cisplatin (Fig. 2H). Finally, there was a significant reduction in cell death induced by Nutlin-3 in all four clones compared with parental ID8 (Fig. 2I). These results collectively indicate that all four Trp53^−/− clones are functionally p53 deficient. We also isolated control clones that had been exposed to CRISPR plasmids (both empty PX459 and PX459 encoding Trp53 gRNA) but did not contain any Trp53 mutation on sequencing. These cells retained p53 transcriptional activity that was indistinguishable from parental ID8 (Supplementary Fig. S2).

We then assessed intraperitoneal growth of the Trp53^−/− clones. Following intraperitoneal injection into female C57Bl/6 mice, there was a highly significant reduction in time to reach predefined humane endpoints with all four clones. Median survival time ranged from 42 to 57 days (Fig. 3A), compared with 101 days for mice bearing either parental ID8 or CRISPR control cells (P < 0.0001 for all clones compared with both parental ID8 and CRISPR controls). There was no difference in volume of ascites between parental ID8, CRISPR control, or Trp53^−/− tumors (Fig. 3B). Macroscopically, the patterns of growth and spread within the peritoneal cavity were similar, although there was some evidence of increased numbers of miliary deposits on the peritoneum and diaphragm in Trp53^−/− tumors (Fig. 3C). By IHC, we confirmed the absence of p53 expression in Trp53^−/− tumors (Fig. 3D). Trp53^−/− tumors retained strong positivity for WT1, but remained negative for Pax8 (Fig. 3E). We also observed significant increases in Ki67 expression in Trp53^−/− tumors (Fig. 3F), consistent with their more rapid intraperitoneal growth.

We next targeted exon 3 of Brca2, which encodes the PALB2 binding domain (Fig. 4A). Using the F3 Trp53^−/− clone, we generated three separate double Trp53^−/−;Brca2^−/− clones (1.4, 2.14, and 3.15), each derived from a different guide, and each with distinct deletions (Supplementary Fig. S3). All three Trp53^−/−;Brca2^−/− clones fulfilled the criteria for defective HR (Fig. 4B; ref. 21) and lost the ability to form Rad51 foci in response to irradiation (Fig. 4C; Supplementary Fig. S4). In addition, all clones were significantly more sensitive to PARP inhibitor–mediated cytotoxicity (Fig. 4D). By contrast, control cells (Trp53^−/− cells exposed to PX459 encoding Brca2 gRNA but with no Brca2 deletion) had the same sensitivity to rucaparib as parental ID8 and F3 Trp53^−/−, and were HR competent by Rad51 assay (data not shown).

We also assessed intraperitoneal growth of Trp53^−/−;Brca2^−/− clones. When analyzed individually, there was no difference in mouse survival between Trp53^−/− F3 tumors and any of the double null tumors (Fig. 4E). However, when analyzed collectively, there was a small, but significant, increase in survival: mice bearing Trp53^−/−;Brca2^−/− tumors survived 10 days longer than Trp53^−/− F3 tumors (57 vs. 47 days; P < 0.01; Supplementary Fig. S5). In addition, mice had significantly lower ascites volumes than either parental or Trp53^−/− tumors (Fig. 4F). There were large diaphragmatic and peritoneal deposits, and ascites was consistently less hemorrhagic (Supplementary Fig. S6). Ki67 histoscores were significantly higher than ID8 parental tumors (not shown) but not significantly different to those seen in Trp53^−/− (Fig. 4G).

To investigate the utility of our new ID8 derivatives to study the relationship between specific mutations within malignant cells and the tumor microenvironment, we first stained tumors for the presence of T lymphocytes (both CD3 and CD8) and macrophages (F4/80). Intraepithelial CD3⁺ and CD8⁺ cells were both generally sparse, with no significant differences between any of the tumor genotypes (Fig. 5A and B). However, we observed the presence of lymphoid aggregates within Trp53^−/−;Brca2^−/− tumors that were absent from both parental ID8 and Trp53^−/-tumors (Fig. 5C; Supplementary Figs. S7 and S8). These aggregates were composed predominantly of CD3⁺ cells, although CD8⁺ populations were visible at the periphery (Supplementary Fig. S9). We also observed striking and significant increases in macrophage infiltration in Trp53^−/− tumors (Fig. 5D; Supplementary Fig. S10). Macrophage infiltration was more variable in Trp53^−/−;Brca2^−/− tumors—median histoscore was greater than in ID8 parental tumors but lower than in Trp53^−/− tumors, although neither difference was statistically significant (Fig. 5D).

We next investigated the specific role of p53 loss upon myeloid populations. Immunoblot array showed marked increases in the monocyte chemoattractant CCL2, and, to a lesser extent, CCL5 and sTNFR1 in conditioned medium from Trp53^−/− cells (Fig. 6A; Supplementary Fig. S11). Immunophenotyping of disaggregated solid tumor deposits (Fig. 6B; Supplementary Fig. S12) and ascites (Fig. 6C; Supplementary Fig. S12) showed significant increases in CD11b⁺ populations in Trp53^−/− tumors compared with Trp53 wild-type. Monocytic myeloid-derived suppressor cells (defined as CD11b⁺Ly6C⁺Ly6G^–) were the predominant myeloid population in both tumor and ascites compared with polymorphonuclear (PMN) MDSC populations (CD11b⁺Ly6C⁺Ly6G⁺). Consistent with IHC, we found significantly more F4/80-positive cells within Trp53^−/− ascites, although the proportions of iNOS⁺ and CD206⁺ cells did not alter significantly between the two genotypes (Fig. 6D; Supplementary Figs. S13 and S14).

Here, we show that ID8, a widely used murine model of ovarian carcinoma, is poorly representative of HGSC, with a conspicuous absence of mutations in genes associated with the human disease, and evidence of functional p53 activity. The overall mutational burden in ID8 is low (functional variants in only 100 genes in 49 MB of sequenced DNA), which concurs with very recent data from ID8-G7, a subline of ID8 that has been passaged in vivo (24). We did not observe activating mutations in common oncogenes (e.g., Kras, Nras, Myc, Egfr, Pik3ca) that may drive carcinogenesis in parental ID8 cells, but we have not undertaken copy-number analyses, and thus cannot exclude the presence of an oncogenic amplification.

Using CRISPR/Cas9 gene editing, we generated sublines of ID8 bearing loss-of-function deletions in Trp53 and Brca2, and demonstrate that these alter tumor growth in the peritoneal cavity. In preliminary experiments, we also show that single-gene mutations can alter the tumor microenvironment, with a significant increase in immunosuppressive myeloid populations within tumor and ascites upon loss of p53 expression, as well as the appearance of intraepithelial lymphoid aggregates in tumors lacking both p53 and Brca2.

Genetically engineered mouse models of HGSC have been difficult to generate (25). Recently, two HGSC models were described: the Drapkin lab utilized the Pax8 promoter to drive Cre-mediated recombination of Trp53 and Pten with Brca1 or Brca2 in mouse fallopian tube secretory epithelium. This resulted in development of STIC lesions, and subsequent invasive tumors in the ovary and peritoneal cavity within 20 weeks (26). By morphology and IHC, the tumors resembled human HGSC, but no ascites was observed, and all Pten^−/− mice developed endometrial lesions (hyperplasia, dysplasia, or carcinoma). In addition, no transplantable cell lines have been described from these mice. A second fallopian tube model has been described, with SV40 large T antigen under the control of the Ovgp-1 promoter (27, 28). Again, STIC-like lesions with p53 signatures were described, as well as invasive tumors within the ovary. However, no peritoneal dissemination or ascites was seen, and, again, no lines that can be readily transplanted into nontransgenic mice have been described. Both of these models are undoubtedly of great importance. However, we believe that a transplantable model, based on a single genetic background (C57Bl/6), which recapitulates disseminated peritoneal disease with ascites and in which multiple genotypes can potentially be rapidly investigated in parallel, is an important adjunct to transgenic models.

One potential criticism of our models is that they arise from a single cell line. Although other murine ovarian cancer cell lines have been described (29), they derive from chimeric SV40 T-Ag animals, and thus cannot be transplanted into nontransgenic animals, which significantly reduces their utility. ID8 remains the only murine line derived from a single nontransgenic strain from which to base these models. A second potential criticism is the cell of origin of ID8 itself. In the original description of ID8 (19), whole mouse ovaries were trypsinized with the aim of removing the OSE layer, but the exact cell of origin of ID8 is unclear. Although the fallopian tube is undoubtedly the originating site of many HGSC, a potential ovarian origin is possible for a subset of these tumors (13). In addition, data from the Drapkin lab model indicated that oophorectomy reduced peritoneal dissemination, suggesting that the ovary plays a role in promoting metastatic spread of fallopian tube lesions (26). Thus, an OSE origin of ID8 does not preclude the applicability of our new cells as a model of HGSC, and our tumors do express some typical HGSC markers (e.g., WT1). However, the absence of Pax8 staining strongly suggests that ID8 is not of fallopian tube secretory origin. PAX8 staining is used in the diagnosis of HGSC and is usually strongly positive, although this is not universal (30). Therefore, the absence of Pax8 in our ID8 models is important to note, but is not a barrier to their use in HGSC research.

Although CRISPR/Cas9 gene editing is an extremely powerful tool, off-target effects can occur (31, 32), which are difficult to quantify precisely in any given model without whole-genome sequencing. Here, we used multiple different guides and selected at least one clone per guide. In addition, we utilized appropriate negative controls (cells exposed to CRISPR plasmid transfection but that do not contain mutations in target gene). The uniform phenotypes observed in each of our models suggest strongly that any off-target effects are not significant.

Finally, we have undertaken preliminary experiments to investigate the link between tumor-specific mutations and immune composition in the microenvironment. Here, we show that loss of p53 increases CCL2 expression and induces marked increases in immunosuppressive myeloid populations both within solid tumor deposits and ascites. CCL2 is a critical chemokine for attraction of monocyte populations; wild-type p53 can suppress CCL2 expression via direct binding to the CCL2 5′UTR (33) and can reduce CCL2-induced xenograft growth (34). In Trp53^−/− tumors, there appeared to be a bias toward monocytic MDSC populations, which have been characterized as more immunosuppressive than PMN MDSC in tumor-bearing mice (35). Intriguingly, we also observed aggregates of lymphoid cells in tumors lacking both p53 and Brca2. The structures contained hundreds of cells, which were predominantly CD3⁺ but CD8^–. Four different types of lymphoid aggregates in HGSC were recently described, the largest of which resembled activated lymph nodes (16), and which contain a variety of T- and B-cell lineages, including plasma cells. The presence of plasma cells was associated with expression of two genes, IGJ and TNFRSF17, but was independent of mutation in BRCA1 or BRCA2. Our preliminary data suggest that loss of Brca2 and p53 can lead to formation of tertiary lymphoid aggregates, but further work, including gene expression analysis and further flow cytometry, will be required to elucidate this relationship. It is not clear whether these microenvironmental changes explain why loss of Brca2 and p53 function results in slower intraperitoneal growth compared with p53 loss alone. Ki67 staining suggests that there is no significant reduction in proliferation in the Trp53^−/−;Brca2^−/− tumors, suggesting that non–cell-autonomous factors may be contributing. Certainly, in patients with HGSC, BRCA2 mutation and the presence of intratumoral lymphocytes (15) are both associated with longer overall survival (36), although the precise mechanisms for this remain unclear.

In summary, we have used gene editing technology to generate transplantable murine ovarian cancer cell lines that recapitulate critical mutations in human HGSC. Our results suggest that our ID8 models may act as an important tool, alongside genetically engineered mouse models and primary patient material, in understanding the complexities of HGSC biology. Further mutants are under construction, and all our ID8 derivatives are freely available to other researchers.

I.A. McNeish is a consultant/advisory board member for AstraZeneca and Clovis Oncology. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J. Walton, J. Blagih, D. Strathdee, K. Vousden, I.A. McNeish

Development of methodology: J. Walton, E. Leung, D. Stevenson, I.A. McNeish

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Walton, D. Ennis, E. Leung, S. Dowson, M. Farquharson, L.A. Tookman, D. Athineos, S. Mason, K. Blyth, K. Vousden, I.A. McNeish

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Walton, J. Blagih, E. Leung, M. Farquharson, I.A. McNeish

Writing, review, and/or revision of the manuscript: J. Walton, J. Blagih, D. Ennis, E. Leung, S. Dowson, L.A. Tookman, D. Stevenson, K. Blyth, F.R. Balkwill, K. Vousden, M. Lockley, I.A. McNeish

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D. Ennis, S. Dowson, C. Orange

Study supervision: F.R. Balkwill, M. Lockley, I.A. McNeish

We would like to thank Emma Johnson for expert technical assistance with mouse experiments and Colin Nixon and Marion Stevenson for IHC.

This work was funded by the University of Glasgow and Cancer Research UK [grants C16420/A12995 (I.A. McNeish), C596/A18076 (K. Vousden), C608/A15973 (I.A. McNeish), C596/A20921 (K. Vousden), and C16420/A16354 (F.R. Balkwill)]. All animal work was performed in the Biological Services Unit facilities at the Cancer Research UK Beatson Institute (Cancer Research UK grant C596/A17196).

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.