E-Cadherin/ROS1 Inhibitor Synthetic Lethality in Breast Cancer

E-cadherin defects are frequently found in breast cancer (>13%) and gastric cancer (>14%) and are particularly prevalent in lobular breast cancers, which account for 15% of all mammary carcinomas (1). CDH1 encodes a calcium-dependent plasma membrane–bound cell–cell adhesion glycoprotein (2). In epithelial cells, E-cadherin forms homotypic adhesive complexes, known as adherens junctions (AJ) that control cell–cell contact, the contractility of cells, and ultimately the integrity of epithelial cell layers (3). Although the extracellular domain of E-cadherin interacts with E-cadherin molecules on adjacent cells, the intracellular domain interacts with, and controls, a number of proteins including p120 catenin (p120), α-catenin, γ-catenin, β-catenin, receptor tyrosine kinases, and a series of plasma membrane–associated receptors and cytoskeletal proteins (2). Loss of E-cadherin function causes a wide variety of phenotypes ranging from defects in cell migration and the orientation of the mitotic spindle, as well as dysregulation of cell–cell adhesion and anoikis resistance (reviewed in ref. 3).

In lobular breast cancer, loss of E-cadherin expression occurs early on in the tumorigenic process and is seen in up to 90% of cases, often co-occurring with mutations in the PI3K coding gene PIK3CA (4). Lobular breast cancers tend to be estrogen receptor (ER)– and progesterone receptor (PR)–positive and ERBB2 amplification–negative and to have a low Ki67 index and a luminal A–intrinsic subtype (1, 5–7). Although these biomarkers might predict a favorable response to adjuvant endocrine therapy, retrospective analyses of two recent clinical trials (BIG 1-98 and ABCSG-8) suggest that a subset of patients with invasive lobular breast cancer (ILC) have poorer responses to endocrine therapy when compared with those with invasive ductal carcinomas (IDC) that display similar biomarkers (8, 9). Furthermore, pathologic complete response rates after neoadjuvant chemotherapy are low in ILC (10, 11), suggesting that additional approaches are required to target this disease. In other breast cancer subtypes, E-cadherin expression might also influence patient outcome. For example, in triple-negative breast cancer, the prognosis of patients with E-cadherin–negative tumors is significantly worse than those with E-cadherin–positive disease (12, 13).

At present, it is not clear whether actionable or pharmacologically tractable E-cadherin synthetic lethal effects can be identified that are likely to work clinically. Such clinically actionable synthetic lethal effects might be expected to be relatively resilient to additional molecular changes and to operate in the face of the high degree of molecular diversity that exists in cancer (i.e., hard synthetic lethal effects; ref. 14). In the data presented below, we illustrate that the combined use of multiple, distinct, in vitro, ex vivo, and in vivo model systems and the exploitation of different functional profiling modalities (genetic and chemical screens) can be used to identify robust and actionable E-cadherin synthetic lethal interactions. The most notable synthetic lethality we identified in this way was between E-cadherin and the ROS1 receptor tyrosine kinase, an effect that is clinically actionable using ROS1 inhibitors such as crizotinib or foretinib.

To identify candidate therapeutic targets for breast cancers with loss of E-cadherin, we used CRISPR/Cas9 mutagenesis in MCF7 breast tumor cells (ERα-positive, luminal A, PIK3CA mutant; described hereafter as MCF7^Parental cells) to generate daughter clones (MCF7^A02, MCF7^B04, and MCF7^B05) with frameshift mutations in CDH1 and loss of E-cadherin expression (Fig. 1A and B; Supplementary Fig. S1). Compared with MCF7^Parental cells, E-cadherin–defective cells displayed a rounded morphology also seen in breast tumor cells harboring naturally occurring E-cadherin mutations (Fig. 1C). We used MCF7^A02 and MCF7^Parental cells in two parallel functional screens to identify E-cadherin synthetic lethal effects: (i) a drug-sensitivity screen where we assessed the relative sensitivity of cells to an in-house curated library of 80 small-molecule inhibitors that are either in clinical use for the treatment of cancer or in late-stage clinical development (Fig. 1D; Supplementary Tables S1 and S2); and (ii) a parallel siRNA sensitivity screen, using siRNA SMARTpools (four different siRNAs targeting a single gene in each well) targeting >1000 cell-cycle control genes, kinase-coding genes or DNA repair–related genes (see Methods; Fig. 1E; Supplementary Table S3). The drug-sensitivity screens identified a series of candidate E-cadherin synthetic lethal drugs, including PF-03758309 (a PAK inhibitor), PF-03814735 (an Aurora kinase inhibitor), PI3K/mTOR inhibitors (BEZ-235, PF-04691502, and everolimus), and the ROS1/MET/ALK inhibitors (15) crizotinib (PF02341066, Pfizer) and foretinib (GSK1363089, GSK; Fig. 1D; Supplementary Tables S1 and S2). In order to identify E-cadherin synthetic lethal effects from our MCF7 isogenic cell line siRNA screen, we calculated the difference in siRNA Z scores between E-cadherin–defective and E-cadherin–proficient cells and identified 104 E-cadherin synthetic lethal effects (P < 0.05, Fig. 1E; Supplementary Table S3). Gene ontology analysis of this 104-gene list using EnrichR (16) highlighted gene sets associated with myosin light chain kinase activity (Supplementary Table S4, adjusted P value = 8.52 × 10⁻⁹, PLK3, AAK1, HUNK, CSNK1A1, NEK4, PAK4, CPNE3, PIM1, CLK3, RPS6KA2, SBK1, STK38L, TGFBR1, and MYLK2), gene sets related to GTP-dependent protein kinase activity (Supplementary Table S4, adjusted P value = 1.35 × 10⁻⁸, PLK3, AAK1, HUNK, CSNK1A1, NEK4, PAK4, CPNE3 PIM1, CLK3, RPS6KA2, SBK1, STK38L, and TGFBR1), and a set of candidate synthetic lethal genes associated with RHO-dependent protein serine/threonine kinase activity (Supplementary Table S4, adjusted P value = 1.35 × 10⁻⁸, PLK3, AAK1, HUNK, CSNK1A1, NEK4, PAK4, CPNE3, PIM1, CLK3, RPS6KA2, SBK1, STK38L, and TGFBR1). Other gene sets identified in this analysis included genes associated with RNA polymerase II carboxy-terminal domain kinase activity (Supplementary Table S4, adjusted P value = 8.52E−09), protein serine/threonine kinase activity (Supplementary Table S4, adjusted P value = 8.52 × 10⁻⁹), and calmodulin-dependent protein kinase activity (Supplementary Table S4, adjusted P value = 8.85 × 10⁻⁹).

Amongst the most significant hits in the siRNA screen was ROS1 (ROS proto-oncogene 1, receptor tyrosine kinase; Fig. 1E; Supplementary Table S3). We selected ROS1 for further study as (i) the synthetic lethality observed with ROS1 siRNA was statistically significant and (ii) the parallel small-molecule inhibitor screen identified ROS1 inhibitors (foretinib and crizotinib) as candidate synthetic lethal drugs, suggesting that the ROS1 effect might be clinically tractable. The siRNA screen did not identify other targets of crizotinib or foretinib (e.g., MET, ALK, KDR, and AXL) as being candidate E-cadherin synthetic lethal effects (Supplementary Fig. S2). We also noted that in both MCF7 and MCF10A isogenic systems, loss of E-cadherin expression caused upregulation of ROS1 protein but did not elicit changes in either MET or ALK expression (Fig. 1F and G; Supplementary Fig. S3), suggesting that enhanced ROS1 expression could represent a homeostatic response to E-cadherin loss; reexpression of E-cadherin in E-cadherin–defective MCF7^A02 cells reversed this ROS1 induction (Fig. 1H), suggesting causality.

In validation experiments, we found that each of the four individual ROS1 siRNAs from the ROS1 SMARTpool caused silencing of ROS1 and preferentially inhibited the E-cadherin–deficient MCF7^A02 clone and also an E-cadherin–defective MCF10A CDH1^−/− nontumor cell line (Fig. 2A–C). Two further E-cadherin–defective clones derived by MCF7 CRISPR/Cas9 targeting, MCF7^B04 and MCF7^B05, were also significantly more sensitive to either foretinib or crizotinib than the parental cells (Fig. 2D; Supplementary Fig. S1), suggesting these effects were not private to MCF7^A02. Restoring E-cadherin expression in MCF7^A02 using a FLAG epitope–tagged E-cadherin cDNA expression construct reduced foretinib sensitivity as well as sensitivity to an additional ROS1/MET/ALK inhibitor, TAE673 (Fig. 2E and F), confirming that E-cadherin influences the response to these agents. By assessing the effects of additional ROS1/MET/ALK inhibitors on E-cadherin isogenic cells, we found that foretinib and crizotinib gave the greatest difference in drug sensitivity between E-cadherin wild-type and E-cadherin–defective cells (Fig. 2G), warranting their further investigation. The sensitivity of E-cadherin–defective cells to foretinib was similar to that observed in HCC78 cells (Fig. 2H and I), which harbor an SLC34A2–ROS1 gene fusion rearrangement, rendering the cell line highly addicted to ROS1 kinase activity (15). We also found E-cadherin–defective cells to be sensitive to a recently described ROS1 kinase inhibitor, PF-06463922 (ref. 17; Fig. 2J). Finally, transfection of E-cadherin–defective MCF7^A02 cells with increasing amounts of ROS1 siRNA enhanced the cell-inhibitory effects of foretinib (Fig. 2K). Conversely, expression of a crizotinib-refractory p.G2032R mutant ROS1 fusion cDNA (18) caused crizotinib resistance in E-cadherin–defective cells (Fig. 2L), suggesting that ROS1 could be a critical foretinib/crizotinib target in these cells.

Many synthetic lethal effects are private to a small number of model systems and do not operate universally in the face of the molecular diversity found in breast and other cancers (14). To assess whether the synthetic lethal effect of ROS1 inhibition could apply more widely in breast cancer with E-cadherin loss, we interrogated recently described in-house siRNA “Achilles' heel” screen data describing the kinase dependencies in 34 breast tumor cell lines (19). We first determined the E-cadherin protein expression in each of these 34 models by Western blotting, classifying cell lines as either “E-cadherin–defective” (n = 12) or “E-cadherin wild-type” (n = 22) and used these classifications to identify siRNAs that selectively targeted the E-cadherin–defective cohort of breast tumor cell lines (Fig. 3A and B). We found that the “E-cadherin–defective” or “E-cadherin wild-type” status defined by Western blotting was consistent with E-cadherin protein expression determined by mass spectrometry (MS; ref. 20), as well as CDH1 mRNA expression and CDH1 mutation data (Fig. 3C and D; Supplementary Table S5). For example, each of the tumor cell line models with CDH1 truncating mutations or homozygous deletions lacked E-cadherin protein expression; similarly, breast tumor cell lines with CDH1 promoter hypermethylation (1, 21) also lacked full-length E-cadherin protein (Fig. 3A). Using the siRNA Achilles’ heel screen data (ref. 19; Fig. 3B), we identified 31 siRNAs that selectively targeted the E-cadherin–deficient cohort (P < 0.05, median permutation t test), including ROS1 (Fig. 3E; Supplementary Fig. S4A and B median permutation t test P = 0.04; Supplementary Table S6). We also assessed whether MET or ALK siRNA selectively targeted E-cadherin–defective breast tumor cells but an analysis of 34 breast tumor cell lines, classified according to E-cadherin status (Supplementary Table S5), did not identify a significant association (Supplementary Fig. S5).

We also interrogated previously published siRNA Achilles' heel screen data describing the kinase dependencies in an additional 69 non–breast cancer derived tumor cell lines (19). After annotating these according to E-cadherin status, we also noted a ROS1/E-cadherin synthetic lethal effect, suggesting that this effect was not necessarily specific to breast tumor models (Supplementary Fig. S6A–S6C; Supplementary Table S7). Validation experiments confirmed the ROS1/E-cadherin synthetic lethality in the breast tumor cell line models (Supplementary Fig. S7A–S7E), suggesting that this effect was not isolated to isogenic models, but also operated in molecularly diverse models of breast cancer.

We also used an orthogonal analytic approach, REVEALER (repeated evaluation of variables conditional entropy and redundancy; ref. 22), to identify the molecular features found in breast tumor cell lines that were most associated with sensitivity to ROS1 siRNA. REVEALER uses a set of molecular features (e.g., the presence or absence of mutations or other defects in key cancer driver genes or proteins) and target inhibition data (in this case sensitivity to ROS1 siRNA) from tumor cell line profiling experiments to identify multiple, often complementary, molecular features that correlate with target inhibition, quantifying these effects as nonlinear information correlation coefficients (IC), between 1 (perfect correlation/features associated with resistance to target inhibition) and −1 (perfect negative correlation/features associated with sensitivity to target inhibition); REVEALER ICs ≤0.1 or >0.1 are regarded as profound correlations (22). We carried out a REVEALER analysis using the ROS1 siRNA Z score data in the breast tumor cell lines as the measure of target inhibition and the mutational status of the 23 recurrently mutated cancer driver genes in the breast tumor cell line panel plus the E-cadherin protein classification described above as molecular features. Using this approach, we found E-cadherin protein deficiency to have a far greater correlation with ROS1 siRNA Z score (IC −0.5) than any of the other molecular features analyzed (Fig. 3F), supporting the hypothesis that E-cadherin status is an important determinant of sensitivity to ROS1 inhibition in breast cancer. We also assessed the sensitivity of the breast tumor cell line panel to foretinib or crizotinib and found that the E-cadherin–defective models were more sensitive to both inhibitors (Fig. 3G and H; Supplementary Table S8). The median sensitivity to either foretinib or crizotinib in the E-cadherin–defective cohort was also similar to that in the SLC34A2–ROS1-translocated HCC78 tumor cell line model (Fig. 3G and H). The responses to ROS1 inhibitors observed in the tumor cell line panel appeared to be independent of epithelial or mesenchymal status. For example, MCF10A cells do not undergo EMT upon losing E-cadherin expression (23) but did exhibit both foretinib and crizotinib sensitivity (Fig. 2H–J). Similarly, SUM149, a mesenchymal breast tumor cell line that has wild-type E-cadherin expression (24), was not sensitive to ROS1 inhibition, whereas E-cadherin–defective, epithelial cells (e.g., SUM44 and SKBR3) were (Supplementary Fig. S8). Furthermore, we did not find that expression of MET, ALK, pAKT, AKT, pERK, or ERK correlated with E-cadherin status nor drug sensitivity (Supplementary Fig. S9), suggesting that these were unlikely to be determinants of the synthetic lethal effects we identified. As ROS1 fusion genes are well-established biomarkers of ROS1 inhibitor sensitivity, we also analyzed paired-end RNA-sequencing data using Chimerascan in order to identify ROS1 fusion genes in our tumor cell line panel (25, 26). Although Chimerascan identified the SLC34A2–ROS1 fusion event in HCC78 lung tumor cells (the positive control), ROS1 fusions were not identified in any of the breast tumor cell lines (Supplementary Table S9). In addition, ROS1 copy number or ROS1 variants, although rare in our tumor cell line panel, did not appear to correlate with crizotinib/foretinib sensitivity (Supplementary Table S9). We did note an increase in ROS1 expression and pROS1 levels in E-cadherin–defective cells, similar to those seen in E-cadherin isogenic systems (Fig. 1F–H; Supplementary Fig. S10), suggesting that enhanced ROS1 expression could represent a homeostatic response to E-cadherin loss. Finally, we interrogated recently published data describing drug-sensitivity effects in ex vivo–cultured breast cancer explants (27) to identify drug-sensitivity effects associated with loss of E-cadherin. When querying these data for drug-sensitivity effects associated with CDH1 gene copy-number loss, we found sensitivity to crizotinib (PF-02341066; ref. 27) to be the most significant effect (Fig. 3I, P value 0.00127). The mean AUC for crizotinib was 0.0713 in explants with CDH1 copy-number loss (n = 5) and 0.138 in explants without CDH1 copy-number loss (n = 12; FDR 0.214).

Given the central role of endocrine therapy in the treatment of ER⁺ breast cancers and the frequency at which endocrine resistance develops, especially in the advanced disease setting (28), we assessed whether CDH1 siRNA caused foretinib or crizotinib sensitivity in previously validated, endocrine therapy–resistant, MCF7 LTED cells (29). We found that CDH1 siRNA caused both foretinib and crizotinib sensitivity in MCF7 LTED cells (P < 0.0001, ANOVA, Fig. 4A–E) as it did in endocrine-sensitive MCF7 cells (Fig. 2G), suggesting that the E-cadherin foretinib/crizotinib synthetic lethal effects were somewhat independent of endocrine therapy sensitivity. We also found that ER⁺, de novo endocrine-resistant and E-cadherin–defective MDAMB134VI tumor cells, which were derived from a case of invasive lobular carcinoma (30), were sensitive to ROS1 inhibitors in both in vitro (Supplementary Fig. S11A) and in vivo experiments (Supplementary Fig. S11B–S11D), suggesting that ROS1 inhibition could also target E-cadherin–defective breast tumor cells in this particular setting.

We next assessed whether crizotinib or foretinib sensitivity extended to models of E-cadherin–defective diffuse gastric cancer, given the elevated frequency of E-cadherin defects in this cancer subtype. We found that either crizotinib or foretinib selectively targeted E-cadherin–defective gastric tumor cell lines (Supplementary Fig. S12A and S12B). The effects of crizotinib in additional gastric tumor cell lines have also been recently assessed (27, 31). We reanalyzed these data and found that those tumor cell lines with reduced E-cadherin expression (defined by CDH1 mRNA levels) were more sensitive to crizotinib than those with higher E-cadherin expression (Supplementary Fig. S12C).

In investigating the cellular phenotypes associated with these E-cadherin–selective effects, we found that exposure of E-cadherin–defective cells to foretinib caused an increase in the proportion of cells with >4N DNA content (Fig. 5A and B); an increase in the frequency of cells with abnormal mitoses, particularly cells with multiple nuclei (Fig. 5C–F); an increase in expression of the G₂–M DNA-damage biomarker p21 (Fig. 5G); and an increase in cellular apoptosis as assessed by PARP cleavage and caspase 3/7 activity (Fig. 5H–L; Supplementary Fig. S13). These phenotypes were reminiscent of those seen in rhabdomyosarcoma tumor cells exposed to crizotinib (32) or chronic myeloid leukemia (CML) tumor cell lines, where foretinib elicits phenotypes associated with mitotic catastrophe (MC), including multinucleated giant cells, increased DNA content, and apoptotic cell death (33).

To assess what might cause such multinuclear phenotypes, we visualized cells exposed to crizotinib or foretinib using live cell microscopy. We found that in both E-cadherin wild-type and E-cadherin–defective cells, either foretinib or crizotinib extended the time cells spent in mitosis (Supplementary Fig. S14A and S14B). This extended mitosis had dichotomous effects, depending upon the status of E-cadherin. In E-cadherin wild-type cells, exposure to a ROS1 inhibitor resulted in the formation of two mononuclear daughter cells (Fig. 6A). In contrast, when exposed to a ROS1 inhibitor, E-cadherin–defective cells initiated cytokinesis but failed to complete invagination of the cell membrane at the cleavage furrow, resulting in the formation of multinuclear cells (Fig. 6A). We noted that cytokinesis in E-cadherin–defective cells exposed to foretinib or crizotinib was characterized by prolonged membrane oscillation, starting at the onset of anaphase, cleavage furrow regression resulting in multinucleated cells, and also the formation of cells with lagging chromosomes (Fig. 6A; Supplementary Fig. S15; Supplementary Videos S1–S5).

Cytokinesis failure and a multinuclear phenotype have previously been associated with defective p120 catenin activity (p120; refs. 34, 35). Actomyosin contractility at the ingression and cleavage furrow during anaphase and telophase is controlled by p120, which mediates these effects via binding to the centralspindlin component MKLP1 and the GTPase RHOA (35). p120 catenin also normally interacts with E-cadherin at the cell membrane where it is tyrosine phosphorylated (36–38), raising the possibility that loss of E-cadherin could impair p120 function. We found that E-cadherin–defective tumor cells exhibited a reduction in p120 catenin levels (Fig. 6B), consistent with recent findings from a large-scale analysis of breast tumors with/without E-cadherin defects (39). Furthermore, as well as ROS1 immunoprecipitating with p120 (Fig. 6C), we found that ROS1 inhibition (i) exacerbated the p120 reduction seen in E-cadherin–defective cells; (ii) reduced tyrosine phosphorylation of p120; and (iii) reduced p120 levels at the cleavage furrow (Fig. 6B and C; Supplementary Fig. S16A and S16B). We also found that siRNA-mediated gene silencing of p120 (or ROS1) caused a multinuclear phenotype and elicited synthetic lethality in E-cadherin–defective but not E-cadherin wild-type cells (Fig. 6D–G). In normal epithelial cells, two opposing forces facilitate cytokinesis: (i) force provided by E-cadherin–dependent AJ at the apical membrane and (ii) force provided by the contraction of the actomyosin ring at the cleavage furrow (40). One model to explain our observations might be that loss of one of these opposing forces in E-cadherin–defective cells causes a greater reliance on processes and proteins that control cleavage furrow formation, such as p120 catenin. Our data suggest that ROS1 inhibitors exacerbate an existing p120 defect in E-cadherin–defective cells and impair cytokinesis to such an extent that abnormal mitoses form; this elicits a DNA-damage response (e.g., induction of p21 and γH2AX) and ultimately apoptosis. ROS1 inhibition has lesser effects when buffered by the activity of wild-type E-cadherin.

Consistent with the concept that E-cadherin defects might cause a dependency upon processes controlled by actomyosin networks, such as cleavage furrow maturation, we also noted that the RHO GTPase effector kinase CDC42BPA (CDC42-binding kinase, MRCKα), which controls actomyosin function (41), was also identified as a robust synthetic lethal effect in our earlier siRNA screens (Supplementary Fig. S4A and S4B). We validated this synthetic lethal effect in subsequent experiments, including those using a toolbox small-molecule MRCKα inhibitor (ref. 42; Supplementary Fig. S17A–S17D), suggesting that additional targets associated with these processes might exist in E-cadherin–defective cancers.

To assess the in vivo therapeutic potential of foretinib and crizotinib, we tested the ability of these drugs to affect the growth of E-cadherin–defective invasive mammary carcinomas derived from the K14cre;Cdh1^F/F;Trp53^F/F (KEP) mouse ILC model; these mammary carcinomas show a strong resemblance to human ILC (43, 44). E-cadherin–defective mammary tumors from KEP female mice were orthotopically transplanted into recipient mice and, once established, animals were treated with foretinib, crizotinib, or drug vehicle. In mice that received the drug vehicle alone, tumor growth continued unabated; in contrast, either foretinib or crizotinib treatment had a strong antitumor effect, reduced tumor volume, and extended the survival of tumor-bearing mice (Fig. 7A–H, ANOVA P < 0.0001 in each case). Foretinib or crizotinib treatment also elicited a reduction in the proliferative index of tumors (as estimated by Ki67 immunohistochemistry) and an apoptotic response (Fig. 7D). The reduction in tumor volume in the early stages of crizotinib treatment is highlighted in Fig. 7F. Similar studies in mice bearing MCF7^A02 or MCF7^Parental derived xenografts also established that the antitumor effect of foretinib was significantly enhanced by the absence of E-cadherin (Supplementary Fig. S18A–S18E). We also assessed the antitumor effect of foretinib treatment in an E-cadherin–defective patient-derived breast tumor xenograft (PDX) model, BCM2665, which was derived from a female with ER-negative, ERBB2-negative, basal-like breast cancer (Fig. 7I; ref. 45). As before, foretinib significantly inhibited the growth of established tumors (P < 0.0001, ANOVA) and extended the survival of mice (Fig. 7J and K). E-cadherin–defective tumors from mice treated with foretinib also consistently exhibited a profound reduction in levels of Ki67, suggesting a severe impairment in proliferative rate; showed increased levels of cleaved caspase-3, a marker of tumor cell apoptosis; and had reduced hematoxylin and eosin (H&E) staining, suggesting tumor necrosis (Fig. 7L). Given the availability of an immunohistochemical assay for measuring total and phosphorylated ROS1 in human tumors, we also established that foretinib treatment resulted in a decrease in pROS1 and total ROS1 (Supplementary Fig. S19).

E-cadherin defects are a common characteristic of human tumors. Here, we show that ROS1 inhibition constitutes an E-cadherin synthetic lethal interaction that can be elicited with clinical inhibitors such as crizotinib or foretinib. These effects appear to be robust in the face of considerable molecular heterogeneity that exists among different isogenic models, distinct tumor cell lines, ex vivo breast tumor explants, and different mouse models of E-cadherin–defective breast cancer. The profound antitumor effects seen in mice with E-cadherin–defective tumors suggest that the observed synthetic lethal effect might also have translational utility. Several distinct approaches could be used to assess the potential of ROS1 inhibitors in E-cadherin–defective cancers. For example, clinical trials in advanced forms of ILC, where E-cadherin defects are common, might seem appropriate. In such trials, absence of E-cadherin using immunohistochemical analysis could be used to preselect appropriate patients for treatment. As crizotinib is already used in the treatment of non–small cell lung cancer and has been the subject of prior clinical safety assessment, this drug seems a reasonable candidate for testing in such trials.

Mechanistically, our data suggest that loss of E-cadherin imparts upon cells a dependency upon ROS1 that is likely related to the ability to undergo cytokinesis. Our work suggests that p120 defects could play a role in these phenotypes, but it also seems possible that other proteins contribute to the synthetic lethal phenotype. For example, we also found that other proteins associated with actomyosin control such as MRCKα are also synthetic lethal with E-cadherin defects; these could conceivably play a part in the ROS1/E-cadherin synthetic lethal effect. Although drug-like inhibitors of RHO effector kinases such as MRCKα are still in development (46), licensed drugs such as crizotinib might provide an actionable approach to targeting E-cadherin–defective cancers. Similarly, although we were able to elicit E-cadherin synthetic lethal effects with multiple different ROS1 siRNA reagents and to partially reverse crizotinib sensitivity with a p.G2032R-mutant ROS1 fusion cDNA (Fig. 2), the antitumor therapeutic effect of drugs such as crizotinib or foretinib could also be influenced by inhibition of kinases other than ROS1. Although we were unable to elicit synthetic lethality with siRNAs for MET, ALK, AXL, and KDR in isogenic or nonisogenic systems (Supplementary Figs. S2 and S5), we cannot formally exclude the possibility that other kinase targets of foretinib or crizotinib also contribute to the E-cadherin synthetic lethality. It seems possible that clinical studies where mechanisms of resistance to crizotinib in E-cadherin–defective tumors are assessed will inform this area.

Small-molecule inhibitors were obtained from SelleckChem. siRNAs were obtained from GE Dharmacon. Cell lines were obtained from the ATCC and European Collection of Cell Cultures in 2010 and 2011 and maintained according to the supplier's instructions, as described previously (19, 47). The MCF10A CDH1^−/− and MCF10A CDH1^+/+ isogenic cell lines were obtained in 2014 from Sigma Aldrich and maintained according to the supplier's instructions. At 6-month intervals and prior to storage, the identity of each cell line was confirmed by short tandem repeat profiling of 10 loci using the GenePrint 10 system (Promega). At monthly intervals, Mycoplasma testing of cell cultures was carried out using the MycoAlert Mycoplasma Detection Kit (Lonza).

MCF7 cells were targeted using the Edit-R-CRISPR-CAS9 gene engineering kit (GE Dharmacon) according to the supplier's instructions. The following crRNA sequence was used: 5′-GCUGAGGAUGGUGUAAGCGAGUUUUAGAGCUAUGCUGUUUUG-3′. MCF7 cells were transfected in 24-well plates (100,000 cells/well) with tracerRNA, crRNA, and Cas9 plasmid. Seventy-two hours after transfection, cells were plated in 15-cm dishes and continuously cultured until colonies formed. Colonies were recovered and profiled using PCR and Sanger sequencing to determine the presence of CDH1 mutations.

siRNA screens and small-molecule screens were performed as described previously (19, 48). A list of small-molecule inhibitors and their molecular targets used in the screens is shown in Supplementary Table S10. See also Supplementary Materials and Methods for details.

Cell survival analysis was conducted as previously described (49). See also Supplementary Materials and Methods for details.

Whole-cell protein extracts were prepared from cells lysed in NP250 buffer (20 mmol/L Tris pH 7.6, 1 mmol/L EDTA, 0.5% NP40, 250 mmol/L NaCl), supplemented with protease inhibitor cocktail tablets (Roche). Protein concentrations were measured using Bio-Rad Protein Assay Reagent (Bio-Rad). For Western blot analysis, 50 μg of whole cell lysates was electrophoresed on Novex 4% to 12% gradient bistris precast gels (Invitrogen) and immunoblotted overnight at 4°C with antibodies listed in Supplementary Table S11.

Cells were plated on coverslips and exposed to drugs the following day. After drug exposure, cells were fixed in 4% (v/v) paraformaldehyde for 10 minutes, washed, permeabilized in 0.2% (v/v) Triton X-100 in PBS for 20 minutes, washed, and blocked in immunofluorescence buffer IFF [1% (w/v) bovine serum albumin, 2% (v/v) FBS in PBS] then immunostained with primary antibodies targeting F-Actin conjugated to Alexa Fluor 488 and Tubulin (Santa Cruz Biotech) and detected with a Texas red conjugated secondary antibody (Supplementary Table S11). DAPI staining was used to detect nuclei. Mitotic and nuclear phenotypes of at least 200 cells per condition were scored in each replicate experiment. Confocal experiments were imaged using Zeiss CLM700.

Time-lapse microscopy was performed in 6-well plates using a Diaphot inverted microscope (Nikon), in a humidified CO₂ chamber at 37°C, using a motorized stage (Prior Scientific), controlled by Simple PCI software (Compix). Cells were first transfected with a mCherry-H2B plasmid, FACS sorted for mCherry-H2B to facilitate DNA visualization, and then exposed to foretinib or crizotinib for a 24-hour period, over which they were microscopically imaged. Sorting of cells is described in Supplementary Materials and Methods.

Measurements took place on a BD LSR II SORP flow cytometer (BD Biosciences) equipped with a 488-nm blue laser, a 561-nm yellow laser, a 633-nm red laser, and a 404-nm violet laser. Propidium iodide was measured with 600 LP 610/20 BP. Cell population was gated in a FSC/SSC dot plot and doublets were gated out based on a DNA dye area/width dot plot. This cell population was further analyzed regarding its cell distribution (G₁, S, and G₂–M phase) using FlowJo V software.

In vivo efficacy studies were conducted using KEP or BCM2665 PDX tumor-bearing mice as previously described (45, 50). Experiments on KEP tumor-bearing mice were carried out at the Netherlands Cancer Institute according to local and international regulations and ethical guidelines and were approved by the local animal experimental committee at the Netherlands Cancer Institute (DEC-NKI AvD:30100 2015 407 Appendix 1, WP 5791 and WP 5900). BALB/c nude mice were implanted with KEP tumor sections in their fourth mammary gland on the right side as described previously (50). Once tumors reached 200 mm³ in volume, mice were randomized into cohorts who received either crizotinib/foretinib 25 or 50 mg/kg a day via oral gavage for 28 days or the drug vehicle. Operators were blinded to which cohort received foretinib or crizotinib and which received vehicle. A surrogate of animal survival was used and defined by the amount of time taken for tumors to reach a pre-agreed volume of 1,500 mm³, at which point mice were sacrificed. Experiments using the BCM 2665 PDX model were conducted at the Institute of Cancer Research (ICR) according to local (UK Home Office) and international regulations and ethical guidelines and were approved by the local animal experimental committee at the ICR. Viably frozen fragments of the BCM2665 model (45) were obtained from Dr. Michael Lewis, Baylor College of Medicine, Houston, Texas, under material transfer agreement and propagated in SCID-Beige and NSG hosts. SCID-beige host mice were obtained from Charles River Laboratories at 21 to 28 days of age and were implanted with BCM2665 tumor fragments (∼2 mm in length) subcutaneously into the inguinal mammary fat pad following standard procedures. Once tumors measured 2 mm in diameter (assessed by palpation and caliper measurement), mice were randomized into cohorts who received either foretinib (25 mg/kg) or the drug vehicle (1% hydroxypropylmethylcellulose, 0.2% SDS in H₂O) every other day (4 days per week) by oral gavage. Operators were blinded to which cohort received foretinib and which received vehicle. Tumor growth was monitored over time (assessed by palpation and caliper measurement) and a surrogate of animal survival defined by the amount of time taken for tumors to reach a pre-agreed 12-mm diameter, at which point mice were sacrificed and tumors excised. Tumor volume was calculated using the formula V = (π × length × width²/6), where the length is the largest tumor diameter and width is the perpendicular diameter. Statistical analysis was performed using Prism. Tumors were formalin-fixed and paraffin-embedded, and slides were stained with H&E, or immunohistochemistry with antibodies against Ki67 and cleaved caspase-3 was undertaken using standard procedures.

M. Dowsett reports receiving commercial research grants from Pfizer and Radius, has received honoraria from the speakers bureau of Roche, is a consultant/advisory board member for Radius, and has received remuneration from ICR Rewards for Inventors. No potential conflicts of interest were disclosed by the other authors.

Conception and design: I. Bajrami, M.D. Gurden, P.W.B. Derksen, A.N.J. Tutt, A. Ashworth, C.J. Lord

Development of methodology: I. Bajrami, F. Song, R. Kumar, M.D. Gurden, C.J. Ryan, C.J. Lord

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): I. Bajrami, R. Marlow, M. van de Ven, R. Brough, H.N. Pemberton, J. Frankum, F. Song, R. Rafiq, A. Konde, D.B. Krastev, M. Menon, R. Kumar, S.J. Pettitt, M.D. Gurden, M.L. Cardenosa, I. Chong, P. Gazinska, F. Wallberg, L.-A. Martin, M. Dowsett, J. Jonkers, A.N.J. Tutt, C.J. Lord

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): I. Bajrami, R. Marlow, J. Campbell, A. Gulati, R. Kumar, M.D. Gurden, R. Natrajan, C.J. Ryan, C.J. Lord

Writing, review, and/or revision of the manuscript: I. Bajrami, R. Marlow, R. Kumar, S.J. Pettitt, E.J. Sawyer, L.-A. Martin, M. Dowsett, C.J. Ryan, J. Jonkers, A.N.J. Tutt, A. Ashworth, C.J. Lord

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): I. Bajrami, R. Marlow, F. Song, R. Kumar, S.J. Pettitt, P.W.B. Derksen, C.J. Lord

Study supervision: A.N.J. Tutt, A. Ashworth, C.J. Lord

Other (contribution toward the study of the mitotic phenotype and MOA): M.D. Gurden

Other (generating data): S. Linardopoulos

This study was funded by Programme Grant funding to C.J. Lord from Breast Cancer Now (CTR-Q4-Y2) as part of funding to the Breast Cancer Now Toby Robins Research Centre, and program grant funding to C.J. Lord from CRUK (C30061/A24439). P.W.B. Derksen is supported by a grant from the Royal Dutch Cancer Society (KWF UU 2011-5203). M.L. Cardenosa is funded by a predoctoral grant from the Carlos III Health Institute (PFIS16/000246) and CIBERONC, an initiative of the Carlos III Health Institute (CB16/12-00473). C.J. Ryan is a Sir Henry Wellcome Fellow jointly funded by Science Foundation Ireland, the Health Research Board, and the Wellcome Trust (103049/Z/13/Z) under the SFI–HRB–Wellcome Trust Biomedical Research Partnership. We thank the late Chris Marshall FRS and members of his laboratory at The Institute of Cancer Research, London, for helpful discussions. We thank members of the Breast Cancer Now Toby Robins Breast Cancer Research Centre Core Pathology Laboratory for pathology support. We also thank members of the Preclinical Intervention Unit of the Mouse Clinic for Cancer and Ageing (MCCA) at the Netherlands Cancer Institute and members of the Biological Services Unit at the ICR for in vivo study support. We acknowledge NHS funding to the NIHR Royal Marsden Hospital Biomedical Research Centre.