Gemcitabine-Induced TIMP1 Attenuates Therapy Response and Promotes Tumor Growth and Liver Metastasis in Pancreatic Cancer

Pancreatic ductal adenocarcinoma (PDAC) is one of the most aggressive cancers with dismal prognosis for overall survival, due to late presentation and limited responses to therapy (1, 2). About 80% of cases are inoperable at diagnosis leaving chemotherapy and/or chemoradiotherapy as the main treatment options. Gemcitabine constitutes one of the chemotherapy backbones but patients commonly show either poor or even complete lack of response to this agent (1). Gemcitabine is a DNA nucleoside analogue that irreversibly inhibits ribonucleotide reductase activity, preventing the synthesis of dioxyribonucleotides required for DNA replication and repair (3). Treatment induces DNA damage and apoptosis of growing cells, thus is frequently employed in combinatorial chemotherapy regimens to treat ovarian, breast, and bladder cancers in addition to pancreatic cancer. While patients have been shown to benefit from gemcitabine, poor responses in pancreas imply that intrinsic resistance and poor patient response to this chemotherapeutic limits efficacy (4). In studies addressing gemcitabine sensitivity, it was recently shown that acute cytokine release following DNA damage may contribute to resistance acquired during the course of treatment as well as to treatment failure (5).

A number of cytokines have been associated with PDAC and have been investigated for potential as early diagnostic or predictive biomarkers, and are beginning to demonstrate some power in diagnosis (1, 4). TIMP1 was originally highlighted as having potential to distinguish patients with PDAC from controls; however, subsequent evidence demonstrated that although patients display elevated levels, TIMP1 does not identify early disease (6).

Tissue inhibitor of matrix metalloproteases (TIMP1) is known to inhibit metalloproteases (MMP) disintegrin and metalloproteinase domain-containing protein 10 and 17 (ADAM-10 and ADAM-17; ref. 7). TIMP1 has also been shown to activate prosurvival signaling, independent of its MMP inhibitory activity, through binding with its receptor CD63 and subsequent engagement of integrin β1–mediated activation of PI3K signaling (8–11).

TIMP1 can influence a number of tumorigenic biological processes, such as proliferation, apoptosis, and metastasis (7, 12, 13). Although TIMP1 expression is suggested as a prognostic marker in several malignancies (14–16), and its expression is associated with hyperproliferation of K-Ras (G12D)-mutated pancreatic cells (17), the functional role of TIMP1 within the PDAC microenvironment remains largely unexplored. Intriguingly, tumor-derived TIMP1 has been proposed to support neutrophil infiltration to the normal liver and the creation of a premetastatic niche to which circulating tumor cells adhere and seed metastasis in colorectal cancer (13). Here, we found that TIMP1 is directly elevated in response to gemcitabine treatment that was predominantly associated with a protumorigenic phenotype.

KPC cells derived from pancreatic tumors of KRas^G12D; Trp53^R172H; Pdx-1 Cre (KPC) mice and cultured in DMEM supplemented with 10% FCS and authenticated as previously reported (18, 19). The human PDAC cell line PANC1 was purchased from European Collection of Authenticated Cell Cultures and cultured in DMEM supplemented with 10% FCS. PANC1 cells were authenticated by STR profiling. To generate stable knockdowns (KD) of mouse TIMP1 and human TIMP1, KPC and PANC1 cells were transduced with lentivirus-mediated shRNA or the control vector (MISSION shRNA, Sigma). To generate stable overexpression, PANC1 cells were transduced with retroviral particles. The p6610 MSCV-IP N-HA only TIMP1 (retroviral plasmid) was purchased from Addgene (catalog number 35009) and used to generate live retroviral particles in the packaging cell line HEK 293T. The retroviral particles were then used to transduce PANC1 wild-type (WT) and knockdown (KD) cells, alongside an empty retroviral control vector. Immortalized pancreatic CAF-stellate cells were obtained from Neuromics, and cultured in VitroPlus III, low serum, complete (Neuromics). Pancreatic CAFS were phenotypically tested by Vitro Biopharma for a range of markers including CD105, CD90, CD44, CD326, CD133, FAP, GFAP, FSP1, α-SMA, and vimentin. KPC and PANC1 cells were selected with 5 μg/mL and 2.5 μg/mL of puromycin, respectively. All cell lines were obtained in 2015, and fresh aliquots were thawed and used up for a limited amount of time to the sixth passage, except gemcitabine-resistant cells that were maintained for two months in cell culture. All cell lines in our laboratory were tested for Mycoplasma every 4 weeks using the MycoAlert kit (Lonza).

All animal experiments were carried out according to the UK Animal (Scientific Procedures) act of 1986. SCID and C57BL/6 mice were purchased from Charles River Laboratories. KPC transgenic mice (20) were bred at the premises of the University of Oxford (Oxford, United Kingdom). Genotyping was carried out at Transnetyx Inc. When KPC mice reached 12 weeks of age, palpation, sonography, and/or MRI was used to identify and measure the size of intra-abdominal pancreatic tumors in KPC mice.

For the subcutaneous models, human PANC1 and murine KPC cells (1 × 10⁶ cells in 100 μL of PBS) were injected subcutaneously into SCID (immunosuppressive) and C57BL/6 (immunocompetent) mice, respectively. Mice were randomized when tumors reached 100 mm³. Where specified, mice were treated with gemcitabine (100 mg/kg, i.p.) or 8 Gy. Tumor volume was measured by calipers as described previously (21). Non-tumor–bearing C57BL/6 mice were also treated with gemcitabine.

For the liver metastasis model, SCID or C57BL/6 mice were injected intrasplenically with 5 × 10⁵ PANC1 or KPC cells, respectively, in PBS as described previously (22). C57BL/6 mice were sacrificed at day 14 after inoculation, whereas SCID mice were imaged by MRI and sacrificed when liver metastasis was observed 21–26 days after inoculation. For all animal experiments, 5–7 mice were randomized in each experimental group.

The MRI method (23), including its technical characteristics, is described in detail in Supplementary Methods.

Proteome profiler antibody array and phosphokinase array (R&D Systems) were performed according to manufacturer's protocol. Human and mouse TIMP1 Quantikine and DuoSet ELISA kits (R&D Systems) were used according to manufacturer's protocol.

Tissue sections were stained for hematoxylin/eosin (H&E) and CD31 staining, and scored as described previously (21, 24). Immunofluorescence, Western blot analysis, and flow cytometry are described in Supplementary Methods.

Clonogenic survival was carried out as reported previously (24). Wound scratches across a monolayer of KPC and PANC1 cells were conducted out using a sterile 200 μL pipette tip. Bright-field images were taken at 16 and 22 hours until wound closure. Boyden invasion chamber inserts (Costar, 12-μm pore size) were coated with 100 μg/cm² of Matrigel. Cells were seeded onto the chamber insert and placed into a 24-well plate containing a cell line monolayer acting as an attractant in its respective media, and invasion was measured after 16 hours.

These assays are described in Supplementary Methods. In the microarray experiment, standard MIAME guidelines were followed. The full microarray data (GSE94891) are available on GEO.

GraphPad Prism 5 was used for all statistical analyses besides Kaplan–Meier survival curves for which IBM SPSS software Version 21 (SPSS Inc.) was used. By default, values were expressed as mean ± SD and statistical significance was measured using one-way ANOVA with Bonferroni posttest, or using unpaired t test. The log-rank test was used to compare Kaplan–Meier curves, while unpaired nonparametric comparisons were made using the Mann–Whitney test where appropriate. In all tests, P values <0.05 were considered significant.

To gain a better understanding of cytokine release following DNA damage, we treated transgenic KPC mice bearing tumor volumes between 70 and 120 mm³ with saline or gemcitabine (100 mg/kg, i.p.) at day 1, 4, and 8. To identify both early and sustained response genes following gemcitabine treatment, mice were divided into three groups; control mice were treated with saline and culled at their terminal endpoint. Gemcitabine “short-term” mice were sacrificed at day 9 from treatment initiation, that is, 1 day upon completion of gemcitabine, and gemcitabine “long-term” mice were allowed to progress and were sacrificed at their terminal endpoint (Fig. 1A). There was no clear separation, in the mice treated with gemcitabine relative to control treated mice, suggesting no survival advantage with gemcitabine (Fig. 1B; Supplementary Fig. S1A). We did not cull any mice at day 9 in the saline group but believe this is irrelevant to the current work as some of the transgenic KPC mice became terminally ill at this time point. However, the mean tumor volume was only marginally decreased in the gemcitabine short-term group and remained the same at the terminal endpoint compared with controls (Supplementary Fig. S1B and S1C, respectively). To identify potential secreted biomarkers that could reveal signaling mechanisms downstream of gemcitabine treatment, cytokine profiling was carried out on the serum from the three groups (Fig. 1C). Several cytokines including TIMP1, IL6, G-CSF, granulocyte–macrophage CSF (GM-CSF); and chemokine (C-X-C motif) ligand 13 (CXCL13), were upregulated early upon treatment with gemcitabine. The levels of TIMP1 were particularly high and its increased levels were sustained until mice reached their terminal endpoint. TIMP1 levels in serum of KPC mice after gemcitabine were confirmed by ELISA in n = 6 mice (Fig. 1D). IHC staining for TIMP1 also confirmed elevated expression of TIMP1 in PDAC tumors following gemcitabine treatment, irrespective of stromal content of the tumors (Fig. 1E), as quantified in (Supplementary Fig. S1D). In addition, TIMP1 mRNA levels were found to be higher in tumors following gemcitabine treatment (Supplementary Fig. S1E). Alternatively, there was no increase in TIMP1 levels in the serum of non-tumor–bearing mice following gemcitabine treatment (Supplementary Fig. S1F), indicating the tumor to be the source of TIMP1. Furthermore, to determine whether TIMP1 played a role in gemcitabine resistance, gemcitabine-resistant KPC cells were generated by culturing cells in progressively increasing concentrations of the drug over a period of 2 months, confirmed by clonogenic survival assay (Supplementary Fig. S1G). Compared with the parental sensitive cell line (KPC^S), gemcitabine-resistant cells (KPC^R) showed higher expression of TIMP1 in both conditioned media (Fig. 1F) and lysates (Fig. 1G). Moreover, TIMP1 levels further increased in conditioned media (Fig. 1F) following treatment with gemcitabine in both the sensitive and the resistant cell lines. In addition, online datamining confirmed the increased expression of TIMP1 in gemcitabine-resistant lung cancer cells, relative to gemcitabine-sensitive lung cancer cell lines (Fig. 1H; ref. 25). These data indicate that TIMP1 is induced by gemcitabine and that TIMP1 levels are intrinsically higher in gemcitabine-resistant cells compared with its sensitive counterparts, and may play a role in tumor cell resistance to gemcitabine.

Interestingly, analysis of microarray data from 4 different PDAC cohorts revealed a significant upregulation of TIMP1 in human PDAC relative to normal matched pancreatic tissue. In the Iacobuzio-Donahue dataset (26) comprising 12 resected infiltrating PDACs compared with 5 samples of normal pancreas, TIMP1 is among the top 6% of genes upregulated in PDAC versus normal pancreas (P = 0.003; Fig. 2A). In the Segara dataset comprising multiple samples of pancreatic tissue from 11 surgical resections and six matched normal appearing pancreas tissue (27), TIMP1 is in the top 5% of genes upregulated in PDAC relative to normal tissue (P = 4.95E–4; Fig. 2B). In the Badea dataset (28) comparing samples of whole PDAC tissue and matched samples of normal pancreas (n = 39), TIMP1 was overexpressed in the top 1% of upregulated genes (P = 5.03E–15; Fig. 2C). Buchholz and colleagues (29) microdissected surgically resected tissue to remove normal tissue effects and in this dataset we observe TIMP1 is upregulated 3-fold and belongs to the top 2% of upregulated genes in both PanIN-3 lesions and PDAC compared with normal duct and acinar cells, PanIN-1b, and PanIN-2 lesions (Fig. 2D). This indicates that TIMP1 is upregulated at the later stages of PDAC progression, in agreement with recent reports recent that have clearly dismissed the TIMP1 as a marker of early disease (6).

To elucidate the role of TIMP1, we stably knocked down TIMP1 expression using five different lentiviral shRNA constructs each in the human PDAC cell line PANC1 and the murine PDAC cell line KPC (Supplementary Fig. S2A and S2B). The cell lines with the highest knockdowns were confirmed by qPCR (Supplementary Fig. S2C and S2D) as well as by Western blotting (Supplementary Fig. S2E) besides ELISA (Fig. 3A and B). TIMP1 knockdown resulted in a decrease in colony formation in both cell lines (Fig. 3C and D); quantified in (Supplementary Fig. S2F and S2G). In addition, the surviving fraction in TIMP1 knockdown sensitized both cell lines to gemcitabine and radiation compared with untreated/DMSO controls, (Fig. 3E–H). We overexpressed TIMP1 retrovirally in our WT and KD cell line alongside the appropriate empty retroviral vector (Fig. 3I). Overexpression of TIMP1 resulted in a rescue of cell growth and desensitized cells to radiation and gemcitabine (Fig. 3J–L). Representative images are shown in (Supplementary Fig. S2H–S2J). We injected PANC1 and KPC TIMP1 WT and KD subcutaneously into SCID and C57bl/6 mice, respectively, and noted that the time it took for the tumors to reach a volume of 300 mm³ was significantly longer in the TIMP1 KD group compared with TIMP1 WT cells (Fig. 3M). We found no differences in the amount of necrosis or stroma between the two groups in PANC1 and KPC tumors (Supplementary Fig. S2K–S2N). With the KPC cells, the time taken for tumors to reach maximum volume (700 mm³) was significantly longer in the TIMP1 KD group (Fig. 3N). To confirm whether our in vitro chemo/radiosensitization was replicated in vivo, we treated mice with 8 Gy or 100 mg/kg gemcitabine. In keeping with our in vitro data, tumor growth was significantly delayed (Fig. 3O and P). Taken together, these results show that TIMP1 facilitates primary tumor growth, associated with lower response to gemcitabine and radiation both in vitro and in our in vivo subcutaneous models.

To identify potential tumorigenic pathways affected by TIMP1 signaling, we carried out a microarray and analyzed gene expression in TIMP1 WT versus KD PANC1 cells. In total, 136 differentially regulated genes were identified (Fig. 4). Gene annotation studies indicated enriched expression of inflammatory transcripts, cell adhesion transcripts, and chemotaxis in TIMP1 KD cells (Table 1). Representative genes were validated by real-time qPCR (Supplementary Fig. S3A–S3D).

To verify whether the role of TIMP1 in chemotaxis translated to an increased migratory capacity, we compared TIMP1 WT and KD cells in a wound scratch assay. TIMP1 expression resulted in a stronger migratory phenotype in both PANC1 (Fig. 5A) and KPC cells (Fig. 5B) in the scratch wound assay quantified in (Supplementary Fig. S4A). Also, both PANC1 and KPC TIMP1 KD cells showed decreased invasive capacity towards CAF compared with their WT counterparts (Supplementary Fig. S4B). TIMP1 did not affect the expression of phosphokinases in PANC1 cells in vitro (Supplementary Fig. S4C). TIMP1 has been previously reported to enhance angiogenesis (30, 31). In our primary subcutaneous animal models, tumors derived from both PANC1 and KPC TIMP1 KD cells showed decreased blood vessel density staining compared with WT groups (Fig. 5C and D). As our gene enrichment analysis identified transcripts involved in cell adhesion and chemotaxis, we examined whether TIMP1 expression in PDAC cells could affect their metastatic potential in vivo. As the liver is the most common site of metastatic spread in PDAC, we injected PANC1 TIMP1 WT and KD cells intrasplenically into SCID mice as reported (32). MRI imaging was used to monitor liver metastasis (Supplementary Fig. S4D and S4E). TIMP1 KD resulted in significantly reduced hepatic metastatic burden (Fig. 5E and F) with lower vascular density within the metastatic regions (Fig. 5G). To verify our results in a different mouse model, we injected KPC TIMP1 WT and KD cells intrasplenically into C57Bl/6 mice. The hepatic metastatic burden was significantly lower following TIMP1 suppression (Fig. 5H). H&E staining confirmed the presence of metastasis in the liver (Fig. 5I) and blood vessel density was also significantly reduced (Fig. 5J). To address pathways that may be involved in the metastatic process, we performed a protein array chIP, but found that pathway activation was indistinguishable from controls (Supplementary Fig. S4C)

TIMP1 level in blood has been considered as a prognostic factor in patients with liver metastasis (33). We harvested blood from C57Bl/6 mice at days 4, 7, and 14 postintrasplenic injection of KPC cells and found significantly increased circulating TIMP1 with increased metastatic burden (Fig. 6A). We conclude that TIMP1 expression correlates with enhanced vascular density and is essential for tumor cell migration and metastatic growth.

TIMP1 can promote immunosuppression (13). Our microarray data identified inflammatory pathways altered by TIMP1. Hence, we queried whether there were any differences in immunosuppressive myeloid populations in the tumor microenvironment that are known to promote malignant growth. While we failed to identify significant changes in primary subcutaneous PANC1 and KPC mouse models following TIMP1 suppression (Supplementary Fig. S5A and S5B), we noted distinct changes in the immune profile of metastatic livers. When we examined the immune profile in livers harboring metastases at day 4, 7, and 14 following intrasplenic injection of KPC cells in C57Bl/6 mice, no macroscopic metastasis was observed at day 4 or day 7, but metastases were evident by day 14 and CD45⁺CXCR2⁺ neutrophils increased with progressive metastatic development (Fig. 6B). Taken together, these results suggest that increasing TIMP1 levels during metastasis formation are associated with increased liver infiltration by myeloid cells.

When we investigated the immune profile in liver metastases following intrasplenic injection of TIMP1 WT and KD PANC1 cells in SCID mice (Fig. 5E) we noted a significant decrease in CD11b⁺Gr1⁺ myeloid-derived suppressor cells (MDSC) and neutrophil infiltration following TIMP1 suppression (Fig. 6C). Similarly, MDSCs and CD4⁺CD25⁺FoxP3⁺ regulatory T cells were decreased in liver metastases of C57Bl/6 mice intrasplenically injected with TIMP1 KD KPC cells compared with WT controls (Fig. 6D). Taken together, our results suggest that the presence of TIMP1 expressing cells is associated with hepatic recruitment of immunosuppressive populations. To query whether the differences in immune cell infiltration was due to differences in metastatic burden, we compared immune cell infiltration in metastasis of equal size and found most of the MDSCs to indeed localize to the tumor and to be of equal percentage. This is indicative of a linear relationship between TIMP1 expression, liver metastatic burden, and immune cell infiltration (Supplementary Fig. S5C and S5D). This would imply that the presence of tumor cells themselves may be required for immune cell infiltration.

Gemcitabine constitutes one of the backbones in the treatment of advanced PDAC but patients often respond poorly to this agent. We identified TIMP1 as a gemcitabine response gene in serum taken from transgenic KPC mice. IHC staining revealed increased TIMP1 protein expression independently of the stromal content in tumor after gemcitabine treatment. This was also confirmed by transcriptional analysis showing an increase in TIMP1 mRNA levels in tumors from gemcitabine-treated mice. Although no direct evidence in PDAC, previous analysis of 264 breast cancer patients demonstrated longer overall survival after gemcitabine-incorporating chemotherapy in patients with TIMP1-negative tumors compared with TIMP1-positive cases (34). Additional studies have correlated high TIMP1 expression with rapid tumor progression after treatment with chemotherapy agents such as cyclophosphamide methotrexate, 5-fluorouracil, epirubicin, and doxorubicin in breast and colon cancers (35, 36). Furthermore, we found TIMP1 expression to be among the top 1% of upregulated genes in human PDAC cohorts and it was significantly upregulated in human PDAC compared with normal pancreatic tissue. Of note, treatment of nontumor–bearing mice with gemcitabine did not alter TIMP1 levels, suggesting that the presence of tumor cells is necessary for TIMP1 upregulation upon chemotherapy with gemcitabine. These data indicate that TIMP1 expression is correlated with tumor progression and decreased response to gemcitabine in PDAC.

We found that TIMP1 expression is important for tumor cell survival as evidenced from decreased colony formation following stable knockdown of TIMP1 in both human and murine PDAC cell lines. Also, overexpression of TIMP1 enhanced clonogenic survival in untreated cells, and also after gemcitabine and radiotherapy that further supports its protumorigenic role. TIMP1 can activate prosurvival and invasive signaling through engaging CD63, activating PI3K/Akt signaling (8–11). Besides this, TIMP1 can regulate cell proliferation by translocating to the nucleus and interacting with the zinc finger protein PLZF (37). Of note, TIMP1 expression occurs downstream of the K-Ras signaling in the transgenic KPC model (17). Inhibition of TIMP1 resulted in improved response of PDAC cells to ionizing radiation and gemcitabine. Accordingly, Oncomine database analysis revealed significant TIMP1 upregulation in gemcitabine-resistant lung cancer cell lines compared with gemcitabine-sensitive counterparts. This is in line with our findings and others demonstrating a correlation between TIMP1 and decreased sensitivity of cell lines to chemotherapy (38).

Previous studies have reported a link between TIMP1 and angiogenesis. TIMP1-overexpressing tumors displayed increased VEGF-A expression (39), augmented growth rates with increased vessel density (40), and metastatic colonization in murine models of lung cancer (41). Inhibition of secreted TIMP1 decreased myofibroblast-induced angiogenesis (42). Also, TIMP1-positive cells have been observed adjacent to CD34⁺ blood vessels in colorectal liver metastasis (43). In agreement, we now show that TIMP1 suppression results in decreased vessel density, both in human and murine PDAC xenograft and liver metastasis models. Furthermore, we identified alterations in chemotaxis pathways upon knockdown of TIMP1, thus supporting a protumorigenic and prometastatic role for TIMP1 via alteration of chemotaxis and the possible promotion of angiogenesis.

We also found that expression of genes playing a role in inflammatory pathways appeared most dependent on TIMP1 levels. Inflammation is a critical component of tumor progression (44). The protumorigenic role of TIMP1 has been previously attributed to its cytokine-like functions that form an inflammatory milieu (45). In addition to TIMP1 upregulation, we noted an increase in the inflammatory cytokine IL6 expression following gemcitabine treatment in KPC mice. Of note, IL6 promotes progression of PanINs into PDAC (46, 47), whereas TIMP1 expression is downstream of IL6 in rat hepatocytes (46). Also, the TIMP1 promoter is a direct target of the inflammatory mediator NFκB (46). Thus, administration of gemcitabine appears to induce an inflammatory response, including TIMP1 upregulation in PDAC.

TIMP1 can promote an immunosuppressive environment via SDF-1–dependent recruitment of neutrophils to the liver and creating a premetastatic niche (13). Suppression of TIMP1 reduced PDAC cell migration, and also invasion of PDAC cells toward CAFs (immortalized PSCs), indicating a prometastatic role for TIMP1. Furthermore, by using an intrasplenic injection model, we observe that circulating PDAC tumor cells require TIMP1 expression to support hepatic colonization associated with recruitment of CXCR2-positive neutrophils, CD11b⁺Gr1⁺ MDSCs and CD4⁺CD25⁺FOXP3⁺ regulatory T cells to the liver. Inhibition of CXCR2 abrogates metastatic seeding (19), whereas Tregs are often encountered in fibrotic tissue and can promote liver metastases (48, 49). Notably, examination of small liver metastases with comparable size between TIMP1 WT and KD groups failed to reveal a significant difference of MDSCs or Tregs between the two groups, indicative of a linear relationship between TIMP1 expression, liver metastatic burden, and immune cell infiltration. This is suggestive that the presence of tumor cells is essential for immune cell infiltration (1, 49, 50).

To summarize, treatment of PDAC with gemcitabine led to upregulation of the cytokine TIMP1. TIMP1 expression correlates with tumor progression in a number of patient cohorts analyzed. Suppression of TIMP1 enhanced tumor response to gemcitabine and radiotherapy, and significantly decreased clonogenic survival, migration, and invasion both in vitro and in vivo. Importantly, TIMP1 suppression was associated with decreased vascular density in the tumor xenografts and liver metastasis, and reduced liver infiltration by immunosuppressive cell populations, resulting in an overall reduction in liver metastasis formation. Our study therefore provides a rationale for the design and addition of TIMP1-specific inhibitors to chemo/radiotherapy.

No potential conflicts of interest were disclosed.

Conception and design: Z.D'Costa, W.G. McKenna, E. O'Neill, E. Fokas

Development of methodology: Z.D'Costa, R.V. Stiphout, S.Y. Lim, P. Kinchesh, S.C. Smart, W.G. McKenna, E. O'Neill, E. Fokas

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Z.D'Costa, K. Jones, A. Azad, S.Y. Lim, A.L. Gomes, S.C. Smart, E. O'Neill

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Z.D'Costa, K. Jones, A. Azad, R.V. Stiphout, S.Y. Lim, A.L. Gomes, S.C. Smart, W.G. McKenna, F.M. Buffa, O.J. Sansom, E. O'Neill, E. Fokas

Writing, review, and/or revision of the manuscript: Z.D'Costa, R.V. Stiphout, S.Y. Lim, A.L. Gomes, P. Kinchesh, S.C. Smart, W.G. McKenna, O.J. Sansom, E. O'Neill, E. Fokas

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Z.D'Costa, E. Fokas

Study supervision: W.G. McKenna, R.J. Muschel, E. O'Neill, E. Fokas

We thank Mick Woodcock, Angela Diana, and Graham Brown for the technical support.

This work was funded by Cancer Research UK (CRUK C5255/A15935; CRUK A19277). Z.C. D'Costa, A.K. Azad, W.G. McKenna, and E. Fokas were also funded by the Kidani Memorial Trust.

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.