E-cadherin-deficient cells have synthetic lethal vulnerabilities in plasma membrane organisation, dynamics and function

We previously carried out a genome-wide siRNA screen in isogenic breast MCF10A cells with and without E-cadherin expression and observed that GPCRs and cytoskeletal proteins frequently displayed an SL relationship with E-cadherin [7]. To elucidate the mechanisms underpinning synthetic lethality in E-cadherin-deficient cells, we have now extended our analysis of the siRNA screen data to include genes with the reverse effect to synthetic lethality, that is, genes whose siRNAs reduced the viability of MCF10A cells more than MCF10A-CDH1^−/− cells. We have termed these genes ‘reverse synthetic lethal’ (RSL). We hypothesise that RSL proteins have direct or indirect interactions with E-cadherin, but CDH1^−/− cells have compensated for the loss of this interaction, making them less sensitive to knockdown of RSL proteins (relative to wild-type cells). This adaption is most likely to have occurred through the activation of one or more proteins with overlapping function (i.e., functional redundancy). In contrast to these RSL proteins, we hypothesise that MCF10A-CDH1^−/− cells have been unable to fully compensate for the loss of the interaction between E-cadherin and any given SL protein, exposing a vulnerability (summarised in Fig. 1a). Although RSL proteins may not represent useful drug targets on their own, compounds which also inhibit the functional homologue may prove to be SL. Regardless, the identity of RSL proteins provides a fuller understanding of the impact of E-cadherin loss on epithelial cells. 1648 genes met our RSL threshold of a MCF10A-CDH1^−/−/MCF10A viability ratio of ≥ 1.3. Using the DAVID Functional Annotation Clustering tool v6.8 [15], the most enriched functional cluster identified amongst RSL genes was a group of terms associated with phosphatases (enrichment score = 14.88; Table 1). Of 239 phosphatases in the Human Dephosphorylation Database, 224 were included in our siRNA screen. A density distribution of the MCF10A-CDH1^−/−/MCF10A viability ratios for these 224 phosphatases produced a single peak with the highest density occurring at a viability ratio of 1.34 (Fig. 1b). This distribution was shifted significantly (p < 2.2 × 10⁻¹⁶) in the RSL direction compared to the distribution of all 18,120 genes in the siRNA screen which displayed a single peak centered around a MCF10A-CDH1^−/−/MCF10A ratio of 1.00. In contrast, the overall density distribution of 496 protein kinases in the siRNA screen showed a small but highly significant shift (p < 2.2 × 10⁻¹⁶) towards synthetic lethality with the peak of the density distribution at 0.95 (Fig. 1c). The contrasting cell viability distribution of kinase and phosphatase siRNAs is likely to be related to the greater functional promiscuity of phosphatases [16]. Ion channels were also highly enriched in the gene set analysis of RSL genes (enrichment score = 10.93; Table 1). Consistent with this observation, the MCF10A-CDH1^−/−/MCF10A density distribution of 161 voltage-gated ion channels in the siRNA screen was significantly right-shifted from the overall distribution of all 18,120 genes with a peak at 1.14 (p < 2.2 × 10⁻¹⁶; Fig. 1d). In contrast, the distribution of 370 solute carrier genes was not significantly different to the distribution of all 18,120 genes (p = 0.45; Fig. 1e). The RSL phenotype of the ion channels was observed in most voltage-gated ion channel subgroups with the exception of small families of gap junction (n = 20), bestrophin (n = 4) and leucine-rich repeat channel (n = 5) proteins which showed a contrasting trend towards a SL phenotype (Fig. 1f). Terms associated with ribosomes, splicing and proteasomes were also enriched in the RSL gene set analysis (Table 1). Cytoplasmic ribosomal proteins (n = 78) and spliceosome proteins (n = 132) showed a corresponding significant RSL shift in the MCF10A-CDH1^−/−/MCF10A density distributions (Fig. 1g, h), as did ubiquitin-specific peptidases (n = 55), E2 ubiquitin conjugating enzymes (n = 36), HECT E3 ubiquitin ligases (n = 25) and proteasome complex proteins (n = 43) (Fig. 1i, j, k, m). In contrast, the RING E3 ubiquitin ligases (n = 299) were not RSL, but instead featured a large SL shoulder in the MCF10A-CDH1^−/−/MCF10A ratios centered at approximately 0.8 (p < 2.2 × 10⁻¹⁶) (Fig. 1l). Surprisingly, GPCR signalling terms were highly enriched amongst the RSL gene functions (enrichment score = 14.33; Table 1) just as they were in the SL gene functions [7]. Of the 1409 GPCRs identified by the Human Gene Nomenclature Committee, 721 were represented in the siRNA screen, including 245 non-sensory, 86 orphan, 363 olfactory GPCRs and 27 taste GPCRs. Consistent with the gene ontology analyses, the MCF10A-CDH1^−/−/MCF10A density distributions for the 245 non-sensory GPCRs had a bimodal distribution, with peaks at viability ratios of 0.82 and 1.37 (p = 1.0 × 10⁻¹²; Fig. 2a). The left peak contained the SL GPCRs identified previously and the right peak comprised the set of RSL GPCRs. A similar bimodal distribution was also observed for the 86 orphan GPCRs (peaks at 0.82 and 1.39; Fig. 2b). In contrast, the viability ratio distribution of the taste and olfactory GPCRs (which have negligible expression in breast cells) predictably presented as single peaks centered at viability ratios of 0.97 (Fig. 2c) and 1.07, respectively. The density distributions for a set of 35 guanine nucleotide binding protein (G-protein) genes showed a single major peak that was shifted in the SL direction (max = 0.92, p = 0.05; Fig. 2d). Together, the above results demonstrate that whole classes of membrane receptors, channels and protein-modifying enzymes are influenced by E-cadherin loss.

To determine whether we could also detect SL effects in major signalling pathways, we first examined the MCF10A-CDH1^−/−/MCF10A viability ratio density distributions for three cell survival pathways that are associated with E-cadherin-mediated cell-to-cell adhesion: PI3K/AKT, WNT and HIPPO [17‐22]. Using KEGG pathway genes, the PI3K/AKT pathway (n = 224) showed a significant SL shift (peak = 0.90, p = 1.3 × 10^− 4; Fig. 2e) while no significant shift in distribution was observed for either the WNT or HIPPO pathways (p = 0.48 and 0.91, respectively; Fig. 2f, g). Separation of the PI3K/AKT pathway genes into functional subgroups showed that the SL effect occurred across most functional groups in the pathway including plasma membrane receptors, integrins and G proteins (Fig. 2h). The only exception to these SL effects was phosphatases which were strongly RSL. The NFKB and MAPK pathways, which are indirectly associated with adherens junction signalling, showed distributions that were not significantly different to the overall 18,120 gene distribution (p = 0.18 and 0.86, respectively; results not shown). Of the other major KEGG cancer pathways, MTOR, RAS, TNF, JAK-STAT and focal adhesion signalling pathways also showed significant SL distributions, but this synthetic lethality could be largely attributed to genes which overlapped with the canonical PI3K/AKT pathway (Suppl. Fig. 1). The observed synthetic lethality was further validated by testing the PI3K antagonist PI103 and the pan-AKT antagonist NS3728 on both the MCF10A isogenic cell line pair and a recently derived CDH1 isogenic gastric NCI-N87 pair (A. Chen, manuscript in preparation). Both drugs caused a significant SL effect in the two cell line pairs (Fig. 2i, j), consistent with the siRNA data.

The highly significant SL and RSL effects of diverse GPCR’s, voltage-gated ion channels, and ubiquitinylation enzymes with roles in vesicle trafficking and receptor recycling, prompted us to investigate the possibility that disruption of plasma membrane organisation or dynamics in E-cadherin-null cells might unify these distinct effects. To do this, we first examined the density distributions of groups of proteins involved in membrane trafficking, using the lists provided in the KEGG Membrane Trafficking hierarchies [23]. 36 functional groups (with a minimum of 20 proteins/group) from the exocytosis, SNARE, endocytosis, endocytosis-lysozyme transport, protein recycling and endosome-golgi transport hierarchies were examined. The three most SL distributions, (i.e., the distributions with the lowest MCF10A-CDH1^−/−/MCF10A density distribution peaks) comprised the family of ‘endocytosis/ARF GTPases and associated proteins’ (n = 23; peak = 0.85; p = 0.007; Fig. 3a) and two membrane curvature protein families: F-BAR (n = 20; peak = 0.84; p = 0.09; Fig. 3b) and ESCRT (n = 24, peak = 0.87; p = 0.02; Fig. 3c). These associations prompted us to test directly for the presence of an SL interaction between E-cadherin and endocytosis, by measuring the uptake of fluorescently labelled cholera toxin B in the MCF10A isogenic cell line pair. The B subunit of cholera toxin, which is non-toxic, is frequently used to measure internalisation by various endocytic mechanisms [24]. After 30 min, uptake of Alexa Fluor 488-labelled cholera toxin B was significantly lower (p = 0.04) in the CDH1^−/− cells compared to the wild-type MCF10A cells (Fig. 3d). Together, these data demonstrate that membrane trafficking is perturbed in MCF10A-CDH1^−/− cells, perhaps due to an underlying disruption of plasma membrane organisation and dynamics. Comparable disruption of exocytosis may explain the accumulation of mucins in signet ring cells, a common feature of DGC and a hallmark of early stage HDGC (Fig. 3e).

To determine whether the disrupted cytoskeletal and membrane function of CDH1-deficient cells was reflected in altered drug sensitivity, we treated the isogenic MCF10A and gastric NCI-N87 cell lines with selected compounds that impact on cytoskeletal function, plasma membrane dynamics and lipid raft composition. The inhibitors of actin polymerisation, cytochalasin D and latrunculin B caused a significant SL differential (i.e., greater inhibition of the CDH1-deficient line) in the isogenic MCF10A cells but an RSL differential (i.e., greater inhibition of the CDH1-expressing line) was observed in the isogenic NCI-N87 cells (Fig. 4a, b). The endocytosis inhibitor bafilomycin A1, an antagonist of vacuolar ATPase [25], caused an SL differential in both the isogenic MCF10A and NCI-N87 lines (Fig. 4c). Lipid rafts are dynamic, relatively ordered membrane nanodomains that are enriched in cholesterol and sphingolipids and have a probable role in signal transduction [26]. Disruption of these domains with the membrane cholesterol-depleting drug methyl-β-cyclodextrin (MβCD) and the inhibitor of cholesterol synthesis, atorvastatin, caused an SL differential in both the MCF10A and NCI-N87 isogenic lines (Fig. 4d, e). Amphotericin B, an antifungal agent that binds to cholesterol and leads to perforation of the plasma membrane [27] showed contrasting effects, with an RSL effect in the MCF10A pair and an SL effect on the NCI-N87 isogenic lines (Fig. 4f). Regardless of the direction of the effect (SL or RSL), the differential activity of membrane-disrupting compounds strongly suggest that E-cadherin loss is leading to changes in plasma membrane dynamics. A fundamental shift in plasma membrane organisation or dynamics is consistent with our observation that RNAi inhibition has either an SL or RSL effect on the majority of GPCR and ion channel proteins. To explore the possible utility of pharmacological inhibition of these classes of membrane proteins, we tested the sensitivity of our isogenic cell line pairs to the chloride channel inhibitor NS3728 and the CB1 cannabinoid receptor antagonist Otenabant. NS3728 was SL in the MCF10A lines but RSL in the NCI-N87 lines, whereas Otenabant showed a significant SL effect on both isogenic pairs (Fig. 4g, h). These results point to the potential utility of selected GPCR or ion channel inhibitors for the treatment of CDH1-deficient cancers but highlight the importance of cellular context.

The isogenic gastric cancer and non-malignant breast cell line data provide valuable insights into possible targets for the chemoprevention and treatment of CDH1-deficient tumours. To enable us to prioritise functions and pathways that are most likely to be robust clinical targets, we statistically interrogated tumour gene expression data for patterns that were consistent with E-cadherin-associated synthetic lethality. In this approach, we hypothesised that if a genuine SL interaction exists between E-cadherin and any given pathway or protein, there would be fewer tumours than expected by chance with low expression (relative to all tumours) of both CDH1 and the pathway or protein. RNA-seq data from 415 gastric cancers was obtained from TCGA [28] and expression metagenes were generated for each of the 2069 pathways in Reactome [29] (version 58) by taking the first eigenvector of the singular value decomposition [30]. Pathway metagene values were separated into tertiles, and a chi-squared test statistic was used to measure association with CDH1 expression tertiles. Empirical p values were calculated for each pathway via resampling (500,000 iterations per pathway), and statistical significance was assessed after FDR correction (Fig. 5a). Using an adjusted p value threshold of 0.2, 20 pathways were identified as being associated with CDH1 expression (Table 2), although several of these pathways overlapped in terms of gene composition (Fig. 5b). The majority of these candidate SL pathways were involved in cell–cell adhesion, membrane trafficking, membrane lipid composition, apoptosis of cell adhesion proteins and GPCR signalling. Consistent with the metagene selection method, hierarchical clustering of gene expression data from each of these pathways showed variably sized clusters of genes with either strikingly high or low expression in CDH1-deficient tumours (Fig. 5c). The expression changes were most pronounced in the 10% of tumours with the lowest CDH1 expression by rank. Notably, the 10% rank level corresponded to a distinct inflexion point in the absolute CDH1 expression of all samples (Fig. 5d). The 10% rank position corresponds to 36% of the overall median expression of CDH1 and may mark a functional threshold for E-cadherin expression [31]. The low expression of CDH1 in this bottom 10% of tumours cannot simply be attributed to high stromal cell contamination of DGCs, since tumours included in the TCGA analysis were restricted to those with a minimum tumour nuclei content of 60% [28]. The genes in the hierarchical clustering regions with the highest relative expression in CDH1-deficient tumours contained 92 known receptors or enzymes (excluding orphan GPCRs). The MCF10A/MCF10A CDH1^−/− RNAi cell viability ratios for these 92 genes showed a bimodal distribution, with the major SL peak occurring at a viability ratio of 0.87 (p = 0.002; Fig. 5e). This distribution demonstrates that our two approaches to identify candidate SL proteins produce an overlapping set of SL candidates which can now be prioritised for further validation (Suppl. Table 1).