Synthetic Lethal Targeting of ARID1A-Mutant Ovarian Clear Cell Tumors with Dasatinib

New targeted approaches to ovarian clear cell carcinomas (OCCC) are needed, given the limited treatment options in this disease and the poor response to standard chemotherapy. Using a series of high-throughput cell-based drug screens in OCCC tumor cell models, we have identified a synthetic lethal (SL) interaction between the kinase inhibitor dasatinib and a key driver in OCCC, ARID1A mutation. Imposing ARID1A deficiency upon a variety of human or mouse cells induced dasatinib sensitivity, both in vitro and in vivo, suggesting that this is a robust synthetic lethal interaction. The sensitivity of ARID1A-deficient cells to dasatinib was associated with G1–S cell-cycle arrest and was dependent upon both p21 and Rb. Using focused siRNA screens and kinase profiling, we showed that ARID1A-mutant OCCC tumor cells are addicted to the dasatinib target YES1. This suggests that dasatinib merits investigation for the treatment of patients with ARID1A-mutant OCCC. Mol Cancer Ther; 15(7); 1472–84. ©2016 AACR.


Introduction
Ovarian clear cell carcinomas (OCCC) comprise between 5 and 25% of all epithelial ovarian cancers and are often associated with endometriosis (1). Patients with advanced OCCC generally respond poorly to standard platinum-based chemotherapy and have a median 5-year survival rate of less than 32% (2,3). Recently, comprehensive DNA sequencing of OCCCs has led to the identification of likely driver mutations in this disease (4,5). The most commonly recurrent genetic event in OCCC is somatic mutation in the tumor suppressor gene ARID1A (AT-Rich Inter-active Domain-containing protein 1A) which is present in up to 57% of patients (4,5). A significant proportion of the tumorassociated somatic ARID1A mutations in OCCC are frame-shift insertion/deletion mutations or nonsense mutations that are predicted to result in premature truncation of the protein (4,5). In addition to being recurrently mutated in OCCC, frame-shift insertion/deletion mutations and nonsense mutations in ARID1A have also been identified in multiple other cancer types such as gastric cancer, renal clear cell cancer, and pancreatic tumors (6). Loss of ARID1A expression has also been associated with a shorter progression-free survival and resistance to platinum-based chemotherapy in OCCC patients (7).
The best-characterized role of the ARID1A protein is as a component of the BAF SWI/SNF chromatin-remodeling complex. The BAF complex plays a key role in modifying the position of nucleosomes on DNA (8)(9)(10) and in doing so likely regulates the access of additional proteins to DNA (8). The composition of the BAF complex includes: (i) an ATP-dependent DNA helicase, encoded by the SMARCA4/BRG1 or SMARCA2/BRM genes; (ii) ARID1A or ARID1B, which each encompass an ARID DNAbinding domain; and (iii) a series of additional accessory subunits such as SMARCB1, SMARC1 and 2, SMARCD1-3, SMARCE1, ACTL6A and DFP1-3 (8,11). ARID1A dysfunction has been associated with a relatively diverse set of phenotypes including defects in cell differentiation, alterations in the control of cell proliferation, as well as defects in the repair of DNA (9,10,12,13). However, a precise molecular understanding of how defects in ARID1A lead to each of these phenotypes is not yet clear.
Given the high prevalence of ARID1A mutations in OCCC, identifying drugs that selectively target ARID1A-mutant tumor cells could potentially inform the development of new therapeutic approaches to this disease. Recently the methyltransferase, EZH2 and PARP inhibitors have been proposed as therapeutic targets in ARID1A-mutant tumors (14,15) and synthetic lethality between ARID1A and ARID1B has also been described (16). Here, we describe a high-throughput functional genomics approach to screen a series of clinically used drugs with the intention of identifying novel ARID1A synthetic lethal effects (17) that operate in OCCC. By profiling drugs that are either already used in the clinical management of cancer or compounds with targets that are currently in late-stage development, we also aimed to identify ARID1A synthetic lethal approaches that could potentially be rapidly translated in the clinic.

Reagents and cell lines
Dasatinib was purchased from Selleck chemicals. Additional drugs and small-molecule inhibitors used in this study are listed in Supplementary Table S1. ES2 and TOV21G were obtained from the ATCC. RMG-1, SMOV2, KOC7C, HCH1, OVAS, OVISE, OVMANA, OVTOKO, and KK were provided by Dr. Hiroaki Itamochi (Tottori University School of Medicine, Yonago, Japan). Ovarian clear cell lines were grown in RPMI with 10% FCS. The identity of cell lines was confirmed by STR typing using the StemElite Kit (Promega) in March 2013 and where appropriate, profiles were confirmed using the Children's Oncology Group (COG) Cell Culture and Xenograft Repository bank (http://strdb. cogcell.org) or JCRB cell bank (http://cellbank.nibiohn.go.jp/ english/). ARID1A HCT116 were grown in McCoy media with 10% (v/v) FCS. Arid1a-null and wild-type mouse embryonic stem cells were provided by Dr. Zhong Wang (Harvard Medical School, Boston, MA; ref. 9) and grown on gelatin-coated plates in DMEM with 10% (v/v) FCS supplemented with 0.1 mmol/L nonessential amino acids, 1 mmol/L sodium pyruvate, 0.1 mmol/L betametcaptothanol, and 2,000 U LIF/mL. All cell lines were routinely confirmed as being mycoplasma negative using the MycoAlert Kit (Lonza) throughout experimentation.

High-throughput drug screen
Cell lines were profiled using a customized drug library (Supplementary Materials and Methods) containing 68 compounds (Supplementary Tables S2 and S3). Cells were plated at a density of 250 or 500 cells per well. Twenty-four hours later, media containing drug library were added to adherent cells. On day seven, cell viability in each well was estimated using CellTiter-Glo (Promega). Luminescence data was log2 transformed and centered on a per plate basis according to the plate median value. Z prime >0.3 and r 2 >0.75 were used to define acceptable screen data. Where appropriate, surviving fractions were calculated relative to DMSO-treated wells and these data were used to generate AUC and SF 50 data.

Cell-based assays
Short-term drug exposure assays were performed in 96-well plates. Cells were plated at a density of 250-500 cells/well and drug added at the indicated concentration 24 hours later. Cell viability was estimated after seven days using Cell-Titre Glo (Promega). Surviving fractions (SF) were calculated relative to DMSO-treated wells, and drug sensitivity curves plotted.
Clonogenic assays were performed in triplicate in 6-well plates. OCCC cell lines were plated at a density of 500 cells/ well and mouse embryonic stem (ES) cells were plated at a density of 2,000 cells per well on gelatin-coated plates. Media containing dasatinib at the indicated concentrations were replaced every three days. After 14 days, cells were fixed in 10% trichloroacetic acid and stained with sulforhodamine-B prior to counting.

siRNA experiments
A 384-well plate arrayed siRNA library targeting 784 genes (Dharmacon) was used (gene list described in Supplementary  Table S4) or a bespoke Dharmacon siRNA library of dasatinib targets. Each well either contained a SMART pool (four distinct siRNA species targeting different sequences of the target transcript combined) or a single siRNA species. Additional positive (siPLK1) and negative [siCON1, siCON2 and Allstar (Dharmacon and Qiagen, respectively)] controls were also added to each plate. Cells were reverse transfected using Dharmafect4 (Dharmacon). Forty-eight hours after transfection, media containing dasatinib or the drug vehicle, DMSO, were added to the plates. After seven days, cell viability in each well was estimated using a CellTiter-Glo assay. Data were processed as described in Supplementary Materials and Methods and as in ref. 18.

Dasatinib bead proteomics
Inhibitor bead proteomics was performed as described previously (19,20). Dasatinib was covalently linked to ECH sepharose 4B using EDC and then 1 mg of precleared cell lysate was rotated with 50 mL of dasatinib bead slurry for three hours. Beads were washed twice with cold binding buffer and then an additional three times with binding buffer with reduced NaCl minus detergent. Proteins were eluted with 6 mol/L urea and trypsinized, desalted, and analyzed by LC-MS/MS on a Thermo LTQ Orbitrap Velos. Spectra were searched using Protein Prospector and labelfree quantitation performed with Skyline. Significance of effects was determined using MS Stats (21).

Protein analysis
Cells were lysed, electrophoresed, and immunoblotted as described previously (22). We used the following antibodies: ARID1A [Bethyl A301-040A (human) and Santa Cruz Biotechnology sc-32761 (mouse)] a-Actinin (A5044, Sigma), Actin (sc-1616, Santa Cruz Biotechnology), b-Tubulin, (2146, Cell Signaling Technology), YES1 (3201S, Cell Signaling Technology), and CDKN1A (2947, Cell Signaling Technology). All secondary antibodies were horseradish peroxidase conjugated. Primary antibodies were used at a concentration of 1:1,000 and secondary antibodies at a concentration of 1:10,000. Protein bands were visualized using enhanced chemiluminescence (ECL, GE Healthcare) and Kodak BioMAX XAR film (Kodak). Alternatively, lysates were probed with the primary antibody and a fluorescent dyelabeled secondary antibody and images were taken using the Odyssey Infrared Imaging system from LI-COR.

Exome sequencing
BWA (Burrows-Wheeler Aligner, bio-bwa.sourceforge.net) was used to align short reads to a reference sequence (GRCh37). Duplicate sequence reads (PCR-derived duplicates) were removed from further analysis at this point. Base quality recalibration, realignment around indels, and variant calling were performed using the Genome Analysis Tool Kit (GATK) using the Broad best practice variant detection workflow (www.broadinstitute.org/ gatk/guide/best-practices). Small insertions and deletions detected in the tumor cell lines that were absent in the reference genomes were considered to be candidate somatic mutations. The set of candidate somatic mutations was refined using the following list of heuristic rules: (i) variants called in regions not covered by the exome capture probes were excluded; (ii) variants marked as low quality (QUAL below 20) were excluded; (iii) variants with fewer than 10 reads covering the locus in all samples were excluded. Common SNPs (those reported to have a global minor allele frequency of greater than 5% in any of the 1000 genomes project data sets) were also removed from the main analysis. The remaining variants were annotated using the Ensembl variant effect predictor script (Ensembl v61).

Cell-cycle analysis
Cells were plated at a density of 2 Â 10 6 cells per well of a 6well plate and incubated for 24 hours after which dasatinib or 0.1% (v/v) DMSO was added to the media and cells cultured for a further 24 hours. After incubation, both adherent and freefloating cells were harvested and then fixed overnight with cold 70% (v/v) ethanol. Cells were then treated with RNase A for 30 minutes prior to nucleic acid staining with propidium iodide (PI, Sigma). Samples were analyzed on a BD LSR II flow cytometer using BD FACSDiva software (BD Biosciences).

Apoptosis assay
After 48-hour dasatinib exposure, cells were analyzed using the ApoTox-Glo Triplex Assay (Promega) as per the manufacturer's instructions.

In vivo efficacy studies
For the biomarker study, TOV21G cells were used to generate subcutaneous xenografts in nu/nu athymic female mice. Once tumors were established (!500 mm diameter), treatment with dasatinib was initiated for 72 hours. Dasatinib was either dissolved in 80 mmol/L sodium citrate at pH 3 for administration via oral gavage or dissolved in DMSO for intraperitoneal injections. Three animals received oral dasatinib at a dose of 15, 30, or 45 mg/kg/day and three animals received dasatinib at a dose of 5, 10 or 15 mg/kg/day i.p., the remaining animals served as a control and were not treated. Two hours after the final drug administration, animals were culled; tumor and normal tissue (spleen and liver) were then collected. Tissue was then thinly sliced and mixed with RIPA buffer in a cryotube overnight at 4 C. Samples were then homogenized and spun in a centrifuge at 15,000 rpm for 15 minutes. The supernatant was then removed and protein quantified using Bio-Rad Protein Assay Reagent (Bio-Rad).
TOV21G cells were transfected with a luciferin expression lentivirus (LVP433 Amsbio) and maintained in blasticidin selection (25 mg/mL) for 10 days prior to injection into female athymic nu/nu mice. A total of 1.0 Â 10 6 TOV21G cells in 100 mL of PBS were injected directly into the peritoneal cavity of 40 nu/nu athymic female mice. Treatment with dasatinib (n ¼ 20) or vehicle (n ¼ 20) was initiated 24 hours after tumor injection. Mice received either dasatinib 15 mg/kg in sodium citrate (80 mmol/L) via oral gavage or vehicle treatment (sodium citrate) daily. Prior to IVIS imaging, luciferin 150 mg/kg (PerkinElmer) was injected into the peritoneum. Mice were culled when they exhibited signs of distress or >20% weight loss or gain (secondary to ascites).

Results
Identification of ARID1A-selective drugs using a focused drug screen We aimed to identify existing drugs that could selectively target ARID1A-mutant OCCC. To do this, we designed a high-through-put drug sensitivity screen (Fig. 1A), where we profiled the in vitro sensitivity of a range of tumor cell models of OCCC to 68 drugs many of which are currently used in the treatment of cancer or are in late-stage development. To facilitate this, we first characterized a panel of commonly used OCCC tumor cell models according to their ARID1A gene mutation and protein expression status. Using whole exome sequencing and immunoblot analysis, we confirmed the previously documented (23,24) presence of truncating ARID1A mutations and loss of full-length ARID1A protein expression in SMOV2, OVISE, TOV21G, OVTOKO, OVMANA, KOC7C, OVAS, and HCH1 OCCC cell line models (Supplementary Figs. S1A and S1B and S2; Supplementary Table S5). We also confirmed the expression of full-length ARID1A protein and the absence of ARID1A gene mutations in ES2, KK, and RMG-1 models of OCCC (Supplementary Figs. S1A and S1B and S2; Supplementary Table  S5). This characterization allowed us to define two cohorts of OCCC tumor cell models for further analysis: ARID1A mutant (deficient; SMOV2, OVISE, TOV21G, OVTOKO, OVMANA, KOC7C, OVAS and HCH1) and ARID1A wild -type (ES2, KK, and RMG-1).
Using these OCCC tumor cell lines, we carried out a series of parallel high-throughput drug sensitivity screens (HTS). As a screening library, we used an in-house curated collection of 68 drugs that were selected on the basis of being either already used in the treatment of cancer or being in late-stage development (Supplementary Table S2). Each tumor cell line was plated in a 384well plate format and then 24 hours later, exposed to each drug for a subsequent five days. At the end of this five-day period, we estimated cell viability by the use of a CellTiterGlo (CTG) assay (Fig. 1A). Each of the 11 OCCC tumor cell lines was screened in triplicate, with replica drug sensitivity data from each cell line being highly reproducible as shown by Pearson correlation coefficients between screen replicas of > 0.75 ( Fig. 1B and Supplementary Table S6). The dynamic range of each screen was also estimated by calculating Z prime (Z') values for data from each 384-well plate used in the screen. To calculate Z' values, we compared CTG luminescent readings from DMSO (cell inhibition negative control) exposed cells to CTG luminescence readings from wells where cells were exposed to 5 mmol/L puromycin (cell inhibition-positive control). Each 384-well plate in the screen delivered a Z 0 > 0.5 ( Fig. 1C; Supplementary Table S6), confirming a suitable dynamic range for each plate used in the screen. To maximize the potential for identifying ARID1A synthetic lethal effects, we screened each cell line using four different drug concentrations (1 nmol/L, 10 nmol/L, 100 nmol/L, and 1 mmol/L) and used these data to generate dose/response survival data for each drug in each cell line model (Supplementary Table S7). These data were then used to calculate AUC for each drug in each tumor cell line as well as SF 50 (surviving fraction 50-the concentration of drug required to elicit a 50% inhibition of the cell population) estimates of drug sensitivity (Supplementary Fig. S3A and S3B) of each OCCC tumor cell line.
By comparing the median AUC values in ARID1A wild-type and mutant cohorts, we identified drugs predicted to deliver an ARID1A-mutant-selective effect. The most profound effect identified in this way was elicited by dasatinib (BMS-354825; Fig. 1D and E). When defining the differential drug sensitivities in ARID1A-mutant versus wild-type cohorts by comparing median AUC values, dasatinib showed a distinct effect, with an AUC of 4 Â 10 À6 in the ARID1A-mutant cohort versus 1 Â 10 À5 in the ARID1A wild-type cohort ( Fig. 1D and E), an effect also observed by the  Identification of ARID1A-selective drugs using a focused drug screen. A, focused drug high-throughput screen (HTS) schematic. Eight ARID1A-mutant and three ARID1A wild-type OCCC cell lines were plated in triplicate 384-well plates and, 24 hours later, media containing compound library were added to the 384-well plates. Cells were continuously cultured for a subsequent five days after which cell viability was estimated using a luminescence assay (CellTiter-Glo, Promega). After processing data (see Materials and Methods), the screen quality control was assessed by examining the correlation between data from replica screens (example shown in B) and Z prime statistics for each plate and each screen replica (example shown in C). B, example, scatter plot of drug sensitivity data from replica ES2 cell line screens. C, example, Z prime values for replica ES2 screens (R1 ¼ replica 1). The distribution curve on the left represents the data from positive control (siPLK1) and the curve on the right is from negative controls (siCON). D and E, waterfall and box-whisker plots of dasatinib AUC data from the high-throughput screen. Median ARID1A-mutant AUC ¼ 4 Â 10 6 and wild-type ¼ AUC 1 Â 10 5 . Waterfall (F) and box-whisker (G) plots of dasatinib SF 50 data from the high-throughput screen. Median ARID1A mutant SF 50 ¼ 271 nmol/L as opposed to wild-type SF 50 ¼ 707 nmol/L. comparison of dasatinib SF 50 data ( Fig. 1F and G). Finally, by using an ANOVA calculation to assess the overall difference in surviving fraction across the 1 nmol/L-1 mmol/L dasatinib concentration range used in the screen, we found the ARID1A-mutant cohort to have a significantly lower survival that the ARID1A wildtype cohort (P < 0.0001, ANOVA). In addition to dasatinib, we also noted that drugs such as everolimus caused moderate ARID1A selective effects in the drug screen (Supplementary Table  S7; Supplementary Fig. S3A). Elements of the PI3K/mTOR signaling cascade have recently been shown to be constitutively active in ARID1A-mutant endometrial and OCCC cancers (25)(26)(27), suggesting that ARID1A-mutant tumors might be addicted to PI3K signaling (28). On the basis of this earlier work and our screen data, we assessed these moderate ARID1A-selective effects in subsequent validation experiments ( Supplementary Fig. S4). These confirmed that ARID1A-mutant tumor cell line models were modestly more sensitive to drugs such as the everolimus, but these effects were not as profound as those caused by dasatinib ( Supplementary Fig. S4).
Dasatinib is a synthetic lethal drug in ARID1A-mutant OCCC tumor cell line models To validate the ARID1A selectivity of dasatinib identified in the high-throughput screens, we carried out both short-term as well as long-term drug sensitivity experiments in our panel of OCCC tumor cell lines. We found that a relatively short six-day exposure to dasatinib was selective for ARID1A-mutant OCCC models ( Fig.  2A and B and Supplementary Fig. S5A and S5B, two-way ANOVA, ARID1A-mutant vs. wild-type cohorts; P < 0.05) as was a longerterm, 15-day drug exposure (Fig. 2C and D and Supplementary  Fig. S5C and S5D, two-way ANOVA, ARID1A-mutant vs. wildtype cohorts in Fig. 2C; P < 0.0001).
Although we found dasatinib to preferentially target the ARID1A-mutant cohort of OCCC models, the possibility existed that other genetic variants in the tumor cell line panel might explain the sensitivity to dasatinib. To independently establish whether ARID1A was indeed a determinant of dasatinib sensitivity, we assessed drug sensitivity in three different isogenic model systems in which we engineered either ARID1A depletion or mutation. In the first instance, we exploited the previously validated mouse ES cell model of Arid1a deficiency ( Fig. 2E and Supplementary Fig. S2) generated by Gao and colleagues (9), where both alleles of Arid1a were rendered dysfunctional by gene targeting. We found dasatinib to be selective for the Arid1a-null ES cell model, compared with the wild-type clone, consistent with the hypothesis that ARID1A is a determinant of dasatinib sensitivity (P < 0.0001 two-way ANOVA; Fig. 2F). We also used gene silencing to suppress ARID1A expression in the two ARID1A wildtype, dasatinib-resistant, human OCCC models and ARID1A wild-type breast cancer and colorectal cancer cell lines models ( Fig. 2G and H and Supplementary Figs. S2 and S5E-S5G). Not only did the siRNA SMARTpool designed to target ARID1A elicit dasatinib sensitivity in the ARID1A wild-type OCCC cell lines but this was also caused by multiple independent ARID1A siRNA species (Fig. 2G and H), suggesting that this was unlikely to be an off-target effect of the RNA interference reagents used.
In addition to these two systems of ARID1A perturbation, we also generated an isogenic human colorectal tumor cell model (HCT116) in which both copies of the ARID1A gene were inactivated by a p.Q456 Ã truncating mutation. This model was generated by AAV-mediated somatic gene targeting (29). As expected, mutation of both copies of ARID1A (Supplementary Fig. S6A) caused loss of full-length ARID1A expression ( Supplementary  Figs. S2 and S6B). Using this model (ARID1A Q456 Ã /Q456 Ã ) and the parental ARID1A wild-type clone (ARID1A WT/WT ), we carried out a subsequent high-throughput drug screen using a second compound library which included the original screening library with the addition of a number of supplementary drugs (Supplementary Tables S3 and S8). Three different concentrations of dasatinib (1,000 nmol/L, 500 nmol/L, and 100 nmol/L) were ranked within the ten most profound ARID1A Q456 Ã /Q456 Ã selective effects (Supplementary Fig. S6C). This finding was confirmed in a subsequent validation assay ( Supplementary Fig. S6D, two-way ANOVA, P < 0.001; ARID1A Q456 Ã /Q456 Ã cells compared with the parental ARID1A WT/WT model).
Having confirmed the validity of the ARID1A/dasatinib synthetic lethality, we assessed whether this effect could be explained by hyperactivity and/or addiction to dasatinib targets. In the first instance, we used an unbiased activity-based proteomics approach to estimate the relative activity of all dasatinib-binding kinases in a subset of our OCCC panel. This "inhibitor bead" approach is based upon the use of an ATP-competitive kinase inhibitor (in this case dasatinib) covalently linked to sepharose. As the ATP-binding avidity of protein kinases is increased by the allosteric changes that follow kinase activation, the relative amount of particular kinases bound to sepharose-dasatinib beads (measured by mass spectrometry) can be used to estimate kinase activity ( Fig. 3A; refs. 19,30,31). Using this approach, we profiled OVISE, KK, RMG1, and TOV21G cells and found that five dasatinib targets (EPHA2, MAP4K5, ABL2, YES1, and ABL1) were significantly (P < 0.01) enriched in the ARID1A-mutant cell lines compared with the wild-type cells ( Fig. 3B; Supplementary Tables S9 and S10).
Although these data suggested that ARID1A-mutant tumor cells exhibited enhanced activity of a number of dasatinib targets, it did not formally confirm addiction to EPHA2, MAP4K5, ABL2, YES1, or ABL1. Therefore, in parallel with this proteomic assessment, we also used a genetic approach and assessed the addiction of 10 OCCC tumor cell lines (three ARID1A wild-type and seven ARID1A-mutant models) to dasatinib targets (Fig. 3C). We selected 14 dasatinib targets that have dasatinib dissociation constants (K d ) of <1 nmol/L (32) and are inhibited at clinically relevant concentrations (ref. 33; Supplementary Table S11) and profiled each tumor cell model with this library. Using this data, we were able to calculate median NPI values for both the ARID1A-mutant OCCC cell line cohort (SMOV2, OVISE, TOV21G, OVMANA, KOC7C, OVAS, and HCH1) as well as the ARID1A wild-type cohort (KK, ES2, and RMG-1; Fig. 3C). By comparing the cellinhibitory effects caused by each siRNA in ARID1A wild-type and mutant cohorts, we identified likely ARID1A synthetic lethal effects (Supplementary Table S11; Fig. 3D and Supplementary  Fig. S7A). We noted that the most consistent ARID1A-selective effect (i.e., where the ARID1A-selective effect was observed with multiple different siRNAs) was caused by siRNA designed to target YES1 (Fig. 3D). Three of the YES1 individual siRNA species and the YES1 siRNA SMARTpool ranked within the ten most profound ARID1A-selective effects, as defined by the median difference in NPI (Fig. 3E and F Table S11). We also found that the YES1 siRNA SMARTpool and individual YES1 siRNAs selectively targeted the ARID1A Q456 Ã /Q456 Ã HCT116 clone, as opposed to the ARI-D1A WT/WT parental clone ( Supplementary Fig. S7C), suggesting that ARID1A mutation might indeed cause dependency upon   Dasatinib is a synthetic lethal drug in ARID1A mutant OCCC tumor cell line models. A, median dasatinib SF dose-response curves in ARID1A-mutant and wild-type cohorts. Cells were plated in 96-well plates and exposed to dasatinib for five days after which SF were calculated. Data in A represents median of six replicates per cell line. Error bars, SEM. Dasatinib response in ARID1A-mutant versus wild-type cohort; two-way ANOVA, P < 0.0001. B, waterfall plot showing the individual dasatinib SF 50 for each OCCC cell line (nmol/L). C, clonogenic (15-day exposure) dasatinib response in ARID1A-mutant vs. wild-type cohort; two-way ANOVA, P < 0.0001. D, representative clonogenic assay images from two ARID1A wild-type and two ARID1A-mutant cell lines exposed to dasatinib E, Western blot analysis of Arid1a expression from whole-cell lysates in Arid1a isogenic mouse ES cells. Tubulin is included as a loading control. F, dasatinib dose-response survival data from clonogenic experiments in isogenic ES cells. Cells were plated on gelatin-coated plates and exposed to dasatinib for 14 days. Error bars, SEM from six replicas per cell line. SF for the Arid1a-null versus Arid1a wild-type mouse ES cells; two-way ANOVA, P < 0.001. G, dasatinib dose-response survival data from the ARID1A wild-type cells (ES2) transfected with siRNA targeting ARID1A. Forty-eight hours after transfection, cells were plated into 96-well plates and exposed to dasatinib for five days. Error bars, SEM from six replica experiments. H, Western blot analysis of ARID1A expression from whole-cell lysates (ES2) transfected as described in G. Tubulin is included as a loading control.     Fig. S2 and S2E, S2F).
Dasatinib sensitivity in ARID1A-mutant OCCC is p21 and RB dependent and is characterized by an apoptotic response One of the known phenotypic effects of dasatinib exposure in tumor cell lines is the induction of G 1 cell-cycle arrest followed by cell apoptosis (34)(35)(36)(37)(38)(39). To understand whether a similar effect might explain the ARID1A/dasatinib synthetic lethality, we assessed cell-cycle progression and the extent of apoptosis in two ARID1A-mutant and two wild-type OCCC models. This was carried out using propidium iodide and Annexin V staining, followed by FACS. We observed a modest but significant increase in dasatinib-induced G 1 arrest in the ARID1A-mutant models, HCH1 and OVISE compared with wild-type models, ES2 and KK (Fig. 4A, P < 0.05 in each wild-type vs. mutant comparison, Student t test). There was, however, a more profound apoptotic response to dasatinib in the ARID1A-mutant models suggesting that although a cell-cycle response could form part of the  Dasatinib sensitivity in ARID1A-mutant OCCC is characterized by G 1 arrest and an apoptotic response. A, column chart illustrating the proportion of cells in different phases of the cell cycle when exposed to dasatinib. Exposure to dasatinib led to a significant increase in the proportion of cells in the G 1 phase of the cell cycle in the ARID1A-mutant cell lines HCH1 and OVISE but not the ARID1A wild-type cell lines ES2 and KK. Cells were exposed to 50 nmol/L dasatinib for 24 hours and fixed in 70% ethanol prior to propidium iodide staining and FACS analysis. Error bars, AE SD from n ¼ 3 independent experiments. S-phase fraction in ARID1A-mutant tumor cells is significantly reduced by dasatinib exposure (Student t test: P < 0.05, for both OVISE and HCH1). B, bar chart illustrating caspase-3/7 activity in response to dasatinib. Cells were treated as in A and caspase-3/7 activity assessed using the ApoTox Glo triplex assay (Promega). P values for ARID1Amutant cells (OVISE and TOV21G) cells versus ARID1A wild-type (ES2 and KK) ( ÃÃÃ , P < 0.001; ÃÃ , P < 0.01; ns, not significant; Student t test). dasatinib effect, apoptosis might also play a role in the cell growth inhibition observed [P values for ARID1A-mutant cells (OVISE and TOV21G) cells versus ARID1A wild-type (ES2 and KK) P < 0.05, Student t-test; Fig. 4B and Supplementary Fig. S8).
To further investigate what the key determinants of the ARID1A/dasatinib synthetic lethality might be, we used a synthetic rescue genetic screen to identify genes which when inactivated drove dasatinib resistance in ARID1A-null OCCC models. To do this, we transfected the ARID1A-mutant (p.542fs), dasatinib-sensitive (SF 50 ¼ 23 nmol/L) OVISE OCCC model with a siRNA library targeting 784 genes, predominantly protein kinase-coding genes, and assessed whether we could induce dasatinib resistance. OVISE cells were reverse transfected in 384-well plates containing the siRNA library. In total, nine replica transfections were performed. Forty-eight hours after siRNA transfection, cells were exposed to either 25 nmol/L dasatinib (three replicates), 120 nmol/L dasatinib (three replicates), or the drug vehicle DMSO (three replicates). Cells were then continuously exposed to drug for five days after which cell viability was estimated by use of a CTG assay. In cells transfected with a control nontargeting siRNA (siCON), exposure to 25 nmol/L dasatinib resulted in a SF 50 , whereas exposure to 120 nmol/L dasatinib resulted in a SF on 16% (SF 16 ). The effect of each siRNA on dasatinib resistance was then estimated by calculating drug effect (DE) Z scores (ref. 18; see Materials and Methods) with positive DE Z scores representing resistance -causing effects. For each gene, the DE Z scores at 25 and 125 nmol/L dasatinib were calculated and rank ordered ( Fig. 5A; Supplementary Table S12). By calculating the average DE Z scores from both SF 50 and SF 16 screens, we identified those siRNAs that elicited the most robust dasatinib resistance-causing effects.
Using this relatively unbiased approach, we found the most profound dasatinib resistance-causing effects in both SF 50 and SF 16 screens to be caused by siRNAs targeting the key G 1 -S cellcycle regulators CDKN1A (p21) and RB1 (DE Z scores of 4.41 and 4.96 respectively, Fig. 5A). Dasatinib has been shown to increase the expression of p21, the protein encoded by CDKN1A, as well as enhancing the protein expression of another cyclin-dependent kinase inhibitor, p27 (34,36,(38)(39)(40)(41). Considering this, as well as the results from the synthetic rescue screen described above, we assessed whether the dasatinib/ARID1A synthetic lethality could be reversed by inactivation of CDKN1A or the downstream mediator, RB1. We found that an siRNA pool designed to target CDKN1A reversed the dasatinib/ARID1A synthetic lethality in the ARID1A-mutant OCCC cell line OVISE but had negligible effects in an ARID1A wild-type model ES2 (Fig. 5B and Supplementary  Fig. S9A). To confirm that the dasatinib resistance observed was due to CDKN1A silencing and not an off-target effect, three individual siRNA species and the SMARTpool were transfected into two ARID1A-mutant OCCC cell lines and the ARID1A isogenic clones. Transfection of each individual siRNA and the SMARTpool led to dasatinib resistance in both of the cell lines examined suggesting an on-target effect (Fig. 5C) and Western blot analysis confirmed CDKN1A silencing (Fig. 5D and Supplementary Fig. S2). We also found that silencing of RB1, a key modulator of the G 1 -S checkpoint also caused dasatinib resistance in multiple ARID1A-mutant OCC cell lines and the isogenic HCT116 ARID1A Q456 Ã /Q456 Ã clone (Supplementary Figs. S2 and S9B-S9E), supporting the hypothesis that in ARID1A-null tumor cells, the inhibitory effect of dasatinib is mediated via the activity of CDKN1A and RB1.

Dasatinib is selective for ARID1A-mutant OCC tumors in vivo
We generated an in vivo model of ARID1A-mutant OCCC by subcutaneously xenografting the TOV21G ARID1A-mutant OCCC cell line (ARID1A p.548fs/p.756fs) into immunocompromised recipient mice. This led to the growth of very aggressive, highly proliferative, subcutaneous tumors. This approach allowed us to assess the effect of dasatinib on these tumors. Published reports on the use of dasatinib in mice vary greatly in terms of the route of administration (oral gavage or intraperitoneal injections) as well as in the dasatinib dose used (42)(43)(44). Therefore, to identify a dose and route of administration that would minimize the impact of deleterious side effects while still delivering an effective dose, we first conducted a biomarker study. Subcutaneous TOV21G tumors were established in nu/nu athymic mice and animals were treated with dasatinib using either oral gavage or intraperitoneal (i.p.) routes of administration at a range of concentrations (oral administration of 15, 30, or 45 mg/kg per day or intraperitoneal administration of 5, 10, or 15 mg/kg per day). By estimating the extent of SRC phosphorylation as a biomarker of dasatinib activity, we were able to predict whether dasatinib elicited a mechanistic effect in both tumor and normal tissue ( Supplementary Fig. S10). At each dasatinib dose tested and using either oral or intraperitoneal administration, SRC phosphorylation was completely abolished in both tumor and normal tissue (Supplementary Fig. S10) and was well tolerated in each case. We therefore selected the lowest dose of dasatinib (15 mg/kg) for subsequent studies and administered this via an oral route.
On the basis of these data, we assessed the in vivo antitumor efficacy of dasatinib. To do this, we first labeled ARID1A-mutant TOV21G cells with a luciferase-expressing construct so that we could later visualize the extent of tumor burden in live animals by the use of an In Vivo Imaging System (IVIS, PerkinElmer). These labeled cells were then xenografted into recipient mice where they generated widespread miliary (disseminated) peritoneal disease and ascites formation, reminiscent of the clinical scenario in OCCC (Fig. 6A). Treatment of these mice with dasatinib led to a significant reduction in TOV21G-related luminescence ( Fig. 6A and B, P < 0.01 two-way ANOVA), suggesting a therapeutic response. This effect was also reflected in an increase in overall survival in the dasatinib-treated cohort of mice (P ¼ 0.004; Mantel-Cox test).

Discussion
Ovarian cancer is the most lethal gynecologic malignancy. The OCCC subtype presents a particular clinical management challenge as it is invariably characterized by resistance to standard chemotherapy as well as a poor prognosis (1). In the work presented here, we have attempted to understand drug sensitivities associated with ARID1A mutation in OCCC by using a functional genomics approach. This has led us to the identification of dasatinib as a candidate synthetic lethal drug in ARID1Amutant OCCC. Dasatinib is not only selective for ARID1A-mutant OCCC models, but experimental induction of ARID1A deficiency drives dasatinib sensitivity in ARID1A wild-type OCCC models as well as in isogenic cell systems (both mouse and human) where ARID1A has been rendered dysfunctional by gene targeting. The selectivity of dasatinib in this context appears to be characterized by a cell-cycle arrest/apoptotic phenotype, which can be reversed with silencing of either CDKN1A or RB1.  Dasatinib sensitivity in ARID1A-mutant OCCC is dependent upon G 1 -S checkpoint effectors. A, drug effect (DE) Z score distribution plots from two independent dasatinib resistance siRNA screens in OVISE tumor cells. Data from SF 50 (left) and SF 16 (right) dasatinib-resistant screens are shown. CDKN1A and RB1 siRNA effects are highlighted, driving dasatinib resistance in each screen. B, dasatinib dose-response curves for the ARID1A-mutant OCCC cell line OVISE and the ARID1A wild-type cell line ES2 following transfection with the CDKN1A siRNA SMARTpool. Silencing of CDKN1A expression results in dasatinib resistance. Each point represents the mean and SEM of six replicates. CDKN1A siRNA dasatinib survival curve versus control siRNA two-way ANOVA; P < 0.0001 for OVISE and P ¼ 0.02 for ES2. C, dasatinib dose-response curves for the ARID1A-mutant OCCC cell line OVISE and the ARID1A Q456 Ã /Q456 Ã isogenic clone after transfection with four individual CDKN1A siRNA oligos and the CDKN1A SMARTpool. Each of the individual siRNAs results in dasatinib resistance suggesting that the effect observed is not due to an off-target effect. For each CDKN1A siRNA, dasatinib survival curve versus control siRNA (two-way ANOVA, P < 0.001). D, Western blot analysis demonstrating the on-target nature of the CDKN1A siRNA. OVISE cell line was transfected with CDKN1A siRNA and whole-cell lysates collected 48 hours later. Lysates were probed with the p21 antibody and a fluorescent dye-labeled secondary antibody used.
ARID1A has previously been demonstrated to act as a negative regulator of the cell cycle (10,24,45). In part, this effect has been ascribed to an impairment of p21-mediated cell-cycle control in ARID1A-null cells (24). Taking this into account, along with the data presented here, and the known role of dasatinib in p21 induction (34,36,46), one possibility is that dasatinib targets ARID1A-mutant tumor cells because it reverses p21 dysfunction caused by ARID1A deficiency. The work presented here suggests that the dasatinib sensitivity of ARID1A-null models is somewhat dependent upon p21 and RB1, observations that are consistent with this hypothesis. It is possible that the sensitivity to dasatinib might be related to a dependency upon the dasatinib target YES1 that we have observed, although we cannot discount the possibility that other dasatinib targets might also be involved or that combinations of dasatinib targets might play a role in the phenotype we observe. We do note that an analysis of microarraybased mRNA profiles from the ARID1A wild-type and mutant tumor cell lines described in this report and publicly available transcriptomic profiles of ARID1A-mutant OCCC did not identify a canonical signaling process that was distinct between wild-type and mutant cohorts (R.E. Miller, C.J. Lord, A. Ashworth, unpublished observations). Although the correlation between ARID1A genotype and dasatinib sensitivity are striking, being present in not only in the panel of OCCC models but also in isogenic systems, we noted that one of the ARID1A-mutant OCCC cell lines, KOC7C, was relatively resistant to dasatinib but addicted to YES1. It is possible that while the ARID1A defect in KOC7C is sufficient to drive YES1 addiction, other genetic or epigenetic factors in this cell line modulate the ARID1A/dasatinib synthetic lethality.
Our assessment of dasatinib efficacy suggests that there are promising signs that dasatinib can inhibit ARID1A-mutant tumors in vivo, although it is clear that further work is required to optimize how dasatinib might be used to elicit a profound, long-lasting antitumor response in women with OCCCs. For example, using additional OCCC cell lines (both ARID1A wildtype and mutant) as established orthotopic tumors (as opposed to nonestablished tumors shown in Fig. 6) would extend the observations made using the TOV21G model. It also seems reasonable to suggest that combination therapy approaches involving dasatinib might be used to maximize the therapeutic window as well as minimizing the impact of signaling feedback loops that might impair the therapeutic effect of dasatinib. In this regard, one clear objective for future in vivo assessment will be to assess the possibility that the therapeutic window caused by dasatinib could be enhanced by combining this kinase inhibitor with other proposed ARID1A synthetic lethal drugs such as EZH2 and PARP inhibitors (14,15). It might also be pertinent to assess the impact of tumor heterogeneity on the dasatinib therapeutic response. In this case, subsequent studies assessing the effect of dasatinib in mice bearing ARID1A-mutant human patient derived xenografts (PDX) might be appropriate, given that PDX material has the capacity to reflect the molecular heterogeneity of human tumors.
Dasatinib is already licensed for use for the treatment of chronic myelogenous leukemia (CML) and Philadelphia chromosomepositive acute lymphoblastic leukemia (AML; refs. 47,48). Several clinical trials have also been conducted, or are at least underway, in patients with solid tumors. In these solid tumor clinical trials, dasatinib is being assessed either as a single-agent therapy or is being assessed when used in combination with cytotoxic or other targeted therapies (49). However, to date, there has been limited success in using dasatinib in the treatment of ovarian cancer patients. When dasatinib was assessed as a single-agent therapy in patients with relapsed epithelial ovarian cancer, minimal antitumor activity was observed, although we note there were only two patients with OCCC in this study, neither of which had known ARID1A status (50,51). In a recent phase I clinical trial using dasatinib in combination with standard cytotoxic chemotherapy (carboplatin and paclitaxel), the drug combination regime could be delivered safely and some evidence of clinical efficacy was achieved (50,51). Of course, neither of these trials were designed to test the hypothesis that ARID1A-mutant OCCC might respond to dasatinib therapy and based upon the work we describe in this manuscript, we believe that testing this hypothesis is warranted.

Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed. In vivo assessment of ARID1A-mutant xenograft models. A, representative IVIS luminescent images of mice with TOV21G peritoneal xenografts after 14 days dasatinib treatment. TOV21G cells were infected with a luciferase expressing lentivector and injected into the peritoneum of BALB athymic mice (n ¼ 40). Twenty-four hours later, dasatinib (15 mg/kg/day, n ¼ 20) or vehicle (n ¼ 20) was initiated 10 minutes prior to IVIS imaging, and mice were administered luciferin (10 mL/mg) intraperitoneally. Colored bar, photon count/cm 2 /sec. B, bar chart of mean luminescence values and SEM of photons/cm 2 /sec. Dasatinib-treated versus vehicle-treated (two-way ANOVA, P < 0.01).