Functional Viability Profiles of Breast Cancer

Abstract

The design of targeted therapeutic strategies for cancer has been driven by the identification of tumor specific genetic changes. However, the large number of genetic alterations present in tumor cells means that it is difficult to discriminate between genes that are critical for maintenance of the disease state from those that are merely coincidental. Even when critical genes can be identified, directly targeting these is often challenging, meaning that alternative strategies such as exploiting synthetic lethality may be beneficial. To address these issues, we have carried out a functional genetic screen in over 30 commonly used models of breast cancer to identify genes that are critical for the growth of specific breast cancer subtypes. In particular, we describe potential new therapeutic targets for PTEN mutated cancers and for ER+ve breast cancers. We also show that large-scale functional profiling allows the classification of breast cancers into subgroups distinct from established subtypes.

INTRODUCTION

Central to the design of novel therapeutic strategies for cancer is the identification of genes that are critical to the survival of tumor cells but which are largely redundant in normal cells. Correlating molecular changes with tumorigenesis has provided one route to the identification of potential drug targets and provides the rationale behind efforts to characterise genetic variation and gene expression in tumors. However, the correlative nature of these data means that it is frequently not possible to determine whether the observations are causative or merely an effect of the disease state. For example, tumor cells generally exhibit anywhere between 10⁴ and 10⁵ genetic changes compared to germline DNA but theoretical estimates suggest that only a few of these mutations (probably less than 10) are critical drivers of tumor formation and survival (reviewed in (1)). In breast cancer, defining the critical genes involved in tumor cell survival can in some cases lead to the development of novel therapeutic approaches to the disease, the most notable recent example being the development of agents such as trastuzumab and lapatinib, that target the reliance of some breast tumors upon the oncogene ERBB2 (2). However, despite the wealth of molecular, genetic and histological characterization of breast tumors and cell line models, our understanding of the genetic dependencies in this disease is relatively poor.

RNA interference (RNAi) screen technology has already enabled the identification of genetic dependencies in cancer cells (1, 3) but has not as yet been applied to a comprehensive study of breast cancer models. We describe here the first attempt to comprehensively define the genetic dependencies for a set of potentially “druggable” genes in a wide range of breast tumor cell line models using a library of short interfering (si)RNA targeting the kinome. In doing so, we not only reaffirm the impact of PI3-kinase and ERBB2 signalling in the disease but importantly we show that combining functional RNAi analysis with gene expression, gene mutation and genomic analysis provides a new strategy for identifying essential determinants of specific breast cancer subtypes, which are potential novel drug targets.

RESULTS

Generation of functional viability profiles of breast cancer

To generate functional profiles of breast cancer, we used an approach that involves high-throughput RNA interference/short interfering RNA (siRNA) viability screening of multiple cell lines and the integration of viability profiles with gene expression, genetic, genomic and histological analysis (3). We reasoned that siRNAs causing significant loss of cell viability in all of the cell lines assayed likely targeted genes that had an essential ubiquitous function in both normal and tumor cells. Similarly, siRNAs that had no significant effect on viability in any of the cell lines were either not functional or targeted non-essential genes. Finally, siRNAs that caused significant lethality in only some but not all cell lines likely identified genes that represented tumour-specific dependencies and candidate therapeutic targets (Supplementary Fig. 1a).

To generate functional viability profiles for breast cancer we selected a panel of 34 breast cancer-derived cell lines and optimised these for high-throughput siRNA screening. Subsequently, each cell line was transfected with a 96 well-plate arrayed siRNA library targeting 714 kinases and kinase-related genes (see Methods and summarised in Fig. 1 and Supplementary Fig. 1b). After five population doublings, cell viability in each well was estimated by use of a highly sensitive luminescent assay measuring cellular ATP levels. To identify loss of viability/inhibition/failure to proliferate effects in each cell line, luminescence readings from each well were log transformed and then centred by the plate median, to account for plate-to-plate variation (Fig. 1 and Supplementary Fig. 1b). To allow data to be compared between different cell lines, plate-centred data from each screen was standardised by the use of a Z score statistic, where Z=0 represents no effect on viability and negative Z scores represent loss of viability. RNA interference screens were carried out in triplicate and comparison of Z score data from replica screens of each cell line showed the screening process to be highly robust (Supplementary Fig. 1c-e and Supplementary Table 1).

Figure 1. Functional profiling of a breast cancer tumour cell line panel.

34 breast cancer cell lines were selected for study and characterised at the molecular level using techniques such as aCGH, transcript microarray, and immunochemistry. Functional profiling of the panel was performed using an siRNA viability screen targeting 714 kinases and kinase-related genes. Each cell line was screened in triplicate. After processing data (see materials and methods and Supplementary Fig. 1a) the quality of each screen was assessed and data from 20 breast cancer lines were selected for subsequent interrogation.

Having screened and processed data from 34 breast cancer lines, we selected 20 for further analysis, based on stringent QC thresholds (such as Z prime-factors; Supplementary Table 2). This panel encapsulated each of the major breast cancer subtypes (4) and included hormone receptor (Estrogen Receptor (ER), Progesterone Receptor (PR)), ERBB2 positive as well as ER, PR and ERBB2 negative models (triple negative), as characterised by parallel immunohistochemical, fluorescent in situ hybridisation (FISH) and western blot analysis (Supplementary Table 3 and Supplementary Fig. 1f). To complement this functional viability profiling, we also characterised the breast tumor panel using transcript microarrays, array-based comparative genomic hybridisation (aCGH), drug sensitivity for a small number of targeted agents, as well as gene mutation data (summarised in Supplementary Tables 3-8). In the siRNA screens we used a stringent threshold of Z<−2 for defining loss of viability effects (approximately equal to Q=0.02275 in one screen replica, and Q= 1.21 ×10⁻⁵ in triplicate screens (5)) and identified 330 genes whose depletion by siRNA caused loss of viability in at least one cell line, and 180 genes that caused loss of viability in two or more cell lines. This analysis enabled us to identify the predominant dependencies in each cell line (Supplementary Tables 3 and 4). All 20 cell lines were reliant upon Polo-like Kinase 1 (PLK1), thus providing an ideal internal control to monitor successful transfection. Furthermore significant fractions of the breast tumour cell panel were sensitive to silencing of the mitotic kinases AURKA (19/20; 95%), GUCY2D (19/20; 95%) and WEE1 (15/20; 75%).

Identification of de facto genetic dependencies using functional viability profiling

The delineation of functional viability profiles for the breast cell line panel provided a framework for identifying dependencies in particular genetically-defined subgroups of the disease. As proof of this principle, we used supervised hierarchical clustering of the viability data and integration with DNA sequence, aCGH, FISH and western blot data to identify additional gene dependencies in PIK3CA mutant or ERBB2 amplified cell lines (summarised in Supplementary Tables 5 and 6). Somatic mutations of PIK3CA, which encodes the p110α catalytic subunit of PI3-kinase, have been shown to induce oncogenic transformation in vitro and in vivo (6). PIK3CA mutations are found in 8% - 35% of human breast cancers making them one of the most common genetic aberrations in this disease (7). Supervised hierarchical clustering of functional viability data according to PIK3CA mutation status identified 30 siRNAs that caused loss of viability preferentially in the PIK3CA mutant subgroup (Fig. 2a and Supplementary Fig. 2a). Of 714 genes profiled in the cell line panel, the highest ranked PIK3CA mutant selective effect was targeting of PIK3CA itself (p=0.014; Fig. 2b and Supplementary Fig. 2a-c). In six of seven PIK3CA mutant lines, PIK3CA siRNA gives effects of Z<−2, suggesting that PIK3CA addiction is one of the most significant effects in PIK3CA mutant tumor cells. Statistically, the chance probability, Q, of a PIK3CA siRNA effect of <−2 in six independently-derived PIK3CA mutant lines is approximately Q=1.39×10⁻¹⁰ (where Z −2 in one cell line is approximately equal to, a Q = 0.02275 effect, or 1 in 44; in six cell lines this is a Q = (0.02275)⁶ chance effect), demonstrating a strong dependency on PIK3CA in breast cancers with PIK3CA mutations despite their diverse genetic backgrounds. We also noted that PIK3CA mutant breast cancer lines were preferentially sensitive to AKT2 and AKT3 siRNAs (Fig. 2c and Supplementary Fig. 2d). Although the mechanisms by which mutant PIK3CA mediates cellular transformation are not completely understood, it is likely that part of this effect is mediated by signalling through AKT (8). The preferential sensitivity of PIK3CA mutant cells to AKT targeting supported the hypothesis that the viability profiles had the ability to illuminate true addiction pathways.

Figure 2. PIK3CA oncogene addiction effects in breast cancer confirmed by functional viability profiling.

a. Heat map showing the results of a supervised clustering of siRNA Z scores. Breast tumor cell lines were clustered according to PIK3CA gene mutation status and differential effects between PIK3CA mutant and wild type groups identified using the median permutation test. Statistically significant effects (p<0.05) are shown. siRNA targeting PIK3CA is marked by the arrow. b. Waterfall and box/whiskers plots of PIK3CA siRNA Z scores across the breast tumor cell line panel. In the box/whiskers plot, *p = 0.014 between PIK3CA mutant and wild type groups using the median permutation test. c. Box/whiskers plots of AKT2 and AKT3 siRNA Z scores in PIK3CA mutant and wild type models (*p= 0.047 and p 0.003 respectively using the median permutation test). d. Box/whiskers plots of individual and pooled PIK3CA siRNA in breast tumour cell lines with kinase catalytic domain or non catalytic domain mutations (where p<0.05 when comparing survival of *helical mutant and wildtype; **kinase mutant and wildtype; ***helical mutant and kinase mutant using Student’s t-test).

Our analysis of the functional viability profiles based on PIK3CA mutation status also suggested that cell lines with mutations in the kinase domain of PIK3CA tend to be more sensitive to PIK3CA siRNA targeting than cell lines harbouring non-kinase domain PIK3CA mutations (Fig. 2d). Whilst a wide variety of tumour-specific PIK3CA mutations have been identified, the vast majority occur in two hotspots, either in the kinase domain (for example p.H1047R) or in the helical domain (e.g. p.E542K or p.E545K) (Supplementary Table 5). It has been suggested that helical and kinase domain mutants have distinct physiologic phenotypes in human cells (9, 10), and the differential effects of PIK3CA targeting in helical vs. kinase domain mutants could also suggest differences in PIK3CA addiction.

Interrogation of the functional profiles also identified other well-established oncogene addiction effects. For example, approximately 15% of breast cancers exhibit amplification and overexpression of the ERBB2 gene and the addiction of some tumor types to ERBB2 signalling is the basis for the clinical success of ERBB2-targeting agents such as trastuzumab and lapatinib (2). Supervised clustering of viability data based on ERBB2 amplification status (based upon an integration with aCGH and FISH data (see Supplementary Table 3 and Supplementary Table 6)) identified a number of genes that when silenced were selectively lethal in ERBB2 amplified breast cancer lines (Fig. 3a and Supplementary Fig. 2e). One of the most dominant effects was targeting of ERBB2 itself (p=0.003), supporting the hypothesis that ERBB2 amplification can lead to an oncogene addiction effect (Fig. 3a,b and Supplementary Fig. 2e-g). This reaffirms that our approach may have merit in identifying genes functionally relevant to tumor cell viability in clinically defined subgroups. Moreover many ERBB2 amplified cell lines are highly sensitive to PIK3CA siRNA (Fig. 3a,c) supporting the clinical development of PI3K inhibitors in ERBB2-dependent breast cancer regardless of PIK3CA mutation status.

Figure 3. Candidate genetic dependencies for breast cancer subtypes.

a. Heat map showing the results of a supervised clustering of siRNA Z scores according to ERBB2 amplification status in the breast tumor cell line panel. Statistically significant effects (p<0.05, median permutation test) are shown. ERBB2 and PIK3CA siRNA are marked by arrows. b. Waterfall and box/whiskers plots of ERBB2 siRNA Z scores across the breast tumor cell line panel. In the box/whiskers plot, *p = 0.003 between ERBB2 amplified and non-amplified groups using the median permutation test. Waterfall plots of kinases that selectively kill breast cancer cell lines containing c. a ERBB2 amplification. d. TN/Basal breast cancer cell lines. e. luminal breast cancer cell lines. p<0.05 for a-c, using the Student’s t-test.

In associated work examining gene addiction effects in melanoma, we performed a similar analysis on COLO-829 cells, a melanoma cell line whose entire genome sequence has recently been determined (11). Deep sequencing of COLO-829 and a normal B cell line from the same patient identified over 33,000 tumour-associated mutations, including a BRAF V600E mutation recurrently found in melanoma (11). Using siRNA viability profiling of COLO-829, we demonstrated that of 779 genes targeted, siRNA targeting of BRAF had the most significant effect (Supplementary Fig. 3a), supporting the hypothesis that functional viability profiling of tumour cell lines has the ability to identify real addiction effects. Similarly, functional viability profiling of the BRAF V600E mutant breast cancer cell line MDAMB231, also identified the addiction to BRAF (Supplementary Table 3 and Supplementary Fig. 3b).

Candidate genetic dependencies for breast cancer subgroups

On the basis of the proof-of-concept examples, we assessed whether functional viability profiles could be used to identify novel addictions/candidate therapeutic targets that could be used in specific genetic or phenotypic backgrounds. To do this, we annotated the viability data set with gene mutation, transcript expression and phenotypic data and then used hierarchical clustering of viability data to identify candidate gene dependencies in different genetic backgrounds. For example, integrating functional viability profiles with genomic profiles generated by aCGH profiling enabled the identification of candidate genetic dependencies for breast tumor cells carrying amplification events commonly found in breast cancer (12, 13). Similar to the ERBB2 paradigm, we identified candidate genetic dependencies associated with chromosome (chr.) 11q13.2-q13.3 amplification; the chr. 8q23.3-q24.3 amplification encompassing the MYC oncogene; amplification of FGFR1 on chr. 8p12; amplification of AURKA on chr. 20q13; amplification of CDK4 on chr. 12q14.1; the chr. 17q21-q23 amplification encompassing the PPM1D oncogene; and amplification of MDM2 on chr. 12q14.3-q15 (Supplementary Fig. 4a-g). A similar analysis allowed us to identify candidate dependencies in models with common tumor suppressor or oncogene mutation events such as those for PTEN mutation, CDKN2A (p16) mutation, KRAS mutation, p53 mutation, BRCA1 mutation or RB mutation (Supplementary Fig. 5a-f). Finally, we compared the functional viability profiles to transcript microarray defined subtypes of breast cancer (e.g. ERBB2, basal/triple negative and luminal status (14)) to identify candidate dependencies in each group (Fig. 3c-e). Further candidate dependencies identified in the ERBB2 amplified models included the calcium/calmodulin-dependent protein kinase, CAMK1 and a member of the mitogen-activated protein kinase family, MAP2K3 (Fig 3c). Key potential targets for the triple negative (estrogen receptor, progesterone receptor, ERBB2 amplification, negative) subgroup included the ribosomal S6 kinase, RPS6KA3, the protein kinase-C related protein kinase 2, PRKCL2, and the pyruvate kinase PKLR (Fig 3d). Finally candidate luminal subtype dependencies included PIK3CA, the casein kinase, CSNK1E and the serine/threonine kinase, MINK (Fig. 3e).

PTEN deficient breast tumor cells are dependent upon the mitotic checkpoint kinase TTK

To validate our approach we selected two genetic dependencies for more detailed analysis. By comparing candidate genetic dependencies between PTEN mutant models and those with wild type PTEN expression (Supplementary Fig. 6a,b and Supplementary Table 7), we observed the apparent dependency of PTEN mutant breast cancer lines for a series of genes controlling the mitotic spindle assembly checkpoint (Supplementary Fig. 6c). The most significant amongst these was dependency on the TTK protein kinase gene (15) (aka MPS1) (p=0.0012; Fig. 4a,b and Supplementary Fig. 6d,e). To directly address the hypothesis that TTK inhibition was selective for PTEN mutant cancer cells, we used isogenic cancer lines in which both copies of the PTEN gene had been rendered dysfunctional by gene targeting (16). We showed that multiple siRNAs silencing TTK (Fig. 4c,d and Supplementary Fig. 6f) or chemical inhibition of TTK (15) (using two distinct TTK inhibitors) was selective for PTEN deficiency (Fig. 4e,f), confirming the observations made in the more genetically diverse breast cancer line panel. This indicated that TTK inhibition may be a novel therapeutic strategy for treating PTEN mutant tumours. Aneuploidy is frequently observed in both human breast carcinomas with low expression of PTEN and prostatic intraepithelial neoplasia from Pten mutant mice (17). This latter phenomenon is perhaps explained by the centromeric dysfunction in PTEN mutant tumor cells, most likely mediated by a loss of the interaction between PTEN and CENP-C, a key kinetochore component (18). TTK is required for normal function of the mitotic spindle checkpoint and it is established that TTK inhibition drives early exit from mitosis and chromosomal aneuploidy. In tumor cells with an aneuplpoid phenotype TTK inhibition further exacerbates aneupliody and is particularly lethal (19). By examining metaphases from PTEN deficient tumor cells exposed to a small molecule TTK inhibitor, we showed that the frequency of abnormal metaphases was substantially increased in PTEN deficient cells (Supplementary Fig. 6g,h), which may explain the PTEN deficient-selective effect of inhibiting this mitotic checkpoint kinase.

Figure 4. Inhibition of TTK is PTEN selective.

a. Heat map showing the results of a supervised clustering of siRNA Z scores. Breast tumor cell lines were clustered according to PTEN gene mutation status and differential effects between PTEN mutant and wild type groups identified using the median permutation test. Statistically significant effects (p<0.05) are shown. siRNA targeting TTK is marked by the arrow. b. Waterfall and box/whiskers plots of TTK siRNA Z scores across the breast tumor cell line panel. In the box/whiskers plot, *p = 0.0012 between PTEN mutant and wild type groups using the median permutation test. c,d. Surviving fraction data from two independently-derived isogenic PTEN mutant/wild-type models transfected with siRNA targeting TTK. Surviving fractions were determined seven days after transfection. *p<0.05 (Student’s t test) e,f. Surviving fraction data from two independently-derived isogenic PTEN mutant/wild-type models exposed to two chemically distinct TTK inhibitors (AZ3146 or CCT132774). Surviving fraction was determined after seven days exposure to inhibitor. In each case the two-way ANOVA statistic of differences in survival curves is p<0.0001).

Genetic dependency on ADCK2 in ER+ve breast tumor models

As a second example of the potential of functional viability profiling we interrogated our datasets to identify candidate genes whose targeting was selective for ER+ve versus ER−ve subtypes. The most significant effect we identified was the sensitivity of ER+ve breast tumor cell line models to ADCK2 siRNA (p=0.004; Fig. 5a,b and Supplementary Fig. 7a,b), an effect validated in an additional four ER+ve breast tumor cell line models not examined in the original screen (Supplementary Fig. 7c,d). This suggested that the genetic dependency of ER+ve breast tumor cells to ADCK2 is a relatively common phenomenon. Integration of viability data with transcript and protein profiling also identified a correlation between sensitivity to ADCK2 silencing and ADCK2 mRNA/protein levels (Supplementary Table 4 and Supplementary Fig. 7e-g) perhaps suggesting that elevated expression of this protein could be essential for the survival of particular tumor cells. ADCK2 (Uncharacterized aarF domain-containing protein kinase 2) is a member of a family of aarF-domain containing proteins. Other members of this protein family are localised to the mitochondria and mitochondrial membranes and have been implicated in ubiquinone biosynthesis (20). Given that the most dominant driver in ER+ve breast cancers is estrogen signalling itself, we hypothesised that ADCK2 could target ER+ve breast tumor cells by modulating estrogen signalling. Supporting this hypothesis we noted that sensitivity to siRNA targeting of ESR1, the gene encoding ERα, closely correlated with the inhibitory effect of ADCK2 siRNA (Supplementary Fig 7c,h and Supplementary Table 8) suggesting that ADCK2 is an important dependency in breast tumor cells with a strong addiction to estrogen signalling. More specifically we noted that ADCK2 silencing reduced estrogen signalling, as measured by the expression of well established ER target genes, including ESR1 itself (Fig. 5c), estrogen stimulation stimulated ADCK2 expression, suppression of estrogen signalling (with faslodex/ICI182780) inhibited ADCK2 expression (Fig. 5d), and finally that ADCK2 co-immunoprecipitated with ERα (Fig. 5e), indicating that the estrogen receptor and ADCK2 interact, either directly or indirectly. Taken together, these observations suggest that the genetic dependency of ER+ve breast tumor cells for ADCK2 is explained by ADCK2 playing a role in estrogen signalling itself. To extend this observation we examined ADCK2 expression in publically available breast cancer gene expression datasets. ER+ve breast tumours divide into two broad and overlapping subgroups termed luminal A and B, based predominantly upon proliferative rate and strength of estrogen signalling (luminal A being low proliferative with strong estrogen signalling (21)). We noted in two independent studies that ADCK2 expression was elevated in the luminal A subgroup when compared to luminal B (Fig. 5f; p=0.0135 and 0.0054 in the two studies, respectively (12, 22)), in concordance with our in vitro observations.

Figure 5. Genetic dependency of ER+ve breast tumor cells upon ADCK2.

a. Heat map showing the results of a supervised clustering of siRNA Z scores. Breast tumor cell lines were clustered according to ER status and differential effects between ER positive and negative groups identified using the median permutation test. Statistically significant effects (p<0.05) are shown. siRNA targeting ADCK2 is marked by the arrow. b. Waterfall and box/whiskers plots of ADCK2 siRNA Z scores across the breast tumor cell line panel. In the box/whiskers plot, *p = 0.004 between ER+ve and ER−ve groups using the median permutation test. c. ADCK2 silencing by multiple ADCK2 siRNA species reduces the expression of ER-target genes as shown by western blotting of MCF7 total cell lysates generated after siRNA transfection. d. Western blot of total cell lysates from MCF7 cells treated with either estradiol (E2) or faslodex (FAS, ICI182780). An E2-dependent increase in ADCK2 expression is shown. e. Western blot of ERa immunoprecitated material from MCF7 cells, indicating an ERa- ADCK2 interaction. f. Tumor ADCK2 mRNA expression in ER+ve breast tumours classified according to luminal A or B status. Transcript expression data from two independent studies was used (12, 22) (12, 22)and luminal A or B status determined according to Sorlie et al (14). p values were calculated using Student’s t test.

Functional viability profiling identifies candidate functional taxonomies

Gene expression and genomic profiles have been used to sub-classify breast cancers. We investigated whether the patterns of response to gene silencing by siRNA could also be used for this purpose to define a “functional taxonomy”. Hierarchical cluster analysis of transcript profiles classifies breast tumours and tumor cell lines into luminal and basal-like molecular subtypes (Supplementary Fig. 8a,b). Hierarchical clustering of genes that when silenced caused loss of viability (Z<−2) in two or more cell lines revealed two distinct groups, distinct from those formed by the clustering of the expression data (Fig. 6 and Supplementary Fig. 8c,d). Group 1 was enriched for breast cancer lines with PTEN mutations (5/5 PTEN mutant lines in group 1, p=0.0016), whilst Group 2 was enriched for PIK3CA mutant lines (5/7 PIK3CA mutant lines, p=0.038). CAL51 and MDAMB453, which carry both PTEN and PIK3CA mutations were classified into Group 1. Interestingly, although the ERBB2 amplified cell lines were distributed evenly between the two groups, the cell lines resistant to ERBB2 silencing and also lapatinib treatment (Supplementary Table 6) were all contained within Group 1 (JIMT1, MDAMB453 and VP229) and those sensitive to ERBB2 silencing/lapatinib (HCC202, BT474 and SKBR3) fell into Group 2 (Fig. 6b). In general, the distinction between Groups 1 and 2 implies that our panel of breast cancer cell lines functionally divide into two groups according to their dependency on well-established essential cancer networks, and is independent from the currently used clinical and transcriptomically-defined breast cancer subgroups.

Figure 6. Hierarchical clustering of breast tumor cell lines according to functional viability profiles.

a. Heat map of breast tumor cell line functional viability data. Z score data from siRNAs that caused Z<−2 in two or more cell lines was used. pvclust approximately unbiased (au) p values are 0.92 and 0.91 for Group 1 and 2 respectively. b. Summary of hierarchical clustering. Group 1 was significantly enriched with PTEN mutant breast tumor cell lines (p = 0.0016), whereas Group 2 was significantly enriched with PIK3CA mutant breast tumor cell lines (p = 0.038). p values were calculated using a Chi-squared test.

DISCUSSION

One of the major limiting factors in cancer drug discovery is identifying targets for specific subtypes of the disease. One of the major approaches to address this question has been to extensively profile the molecular changes that specifically occur in tumour cells. Whilst this has proven productive and has led in part to the development of drugs such as trastuzumab and imatinib, it is often difficult with molecular profiling alone, to identify the critical proteins that when targeted cause inhibition of tumor cell growth or induce tumor cell death. Here we show that multidimensional datasets that include molecular profiling data as well as functional viability profiles can be used to identify genetic dependencies and candidate targets in tumor cells. Our re-identification of well-validated targets such as ERBB2 and PIK3CA using this approach demonstrates the potential of this method and suggests that a study of the clinical potential of the other genetic dependencies identified here may be worthwhile.

In this study, we have focused upon integrating functional viability profiles for breast tumor cell line models with accompanying molecular profiling datasets such as gene mutation status, aCGH, transcriptomic data and hormone receptor status. One of the obvious extensions of this approach is to apply a similar method to other cancer types but also to add new molecular and phenotypic annotation to the breast cancer dataset defined here. Already, low depth, whole genome sequences have been generated for a number of breast tumor cell line models (23) and it is expected that hundreds of complete cancer genomes and their normal counterparts will be available within a year or two (24). Somatic mutations in cancer genomes include both “driver” mutations, which confer clonal growth advantage on the cancer cell and have therefore been selected during development of the cancer, and “passenger” mutations that are not implicated in cancer development. As an example of this, the recent comprehensive identification of somatic DNA mutations in a melanoma cell line COLO829 identified 33,345 substitutions, 66 small insertions/deletions and 37 rearrangements (11). Based on sequencing data alone it may be difficult to interpret the significance of most of these mutations and therefore to identify what drives survival in this particular tumour cell. While computational approaches may indeed direct the discrimination of driver from passenger mutations, this may require many thousands of samples to be profiled. Therefore, functional assessments of cancer cell dependencies, as described here, will no doubt complement these large-scale sequencing efforts and aid the identification of therapeutic targets (25).

The selectivity of TTK inhibition for tumor cell models with PTEN mutations we describe here is a clear example of synthetic lethality/sickness, where a combination or synthesis of effects (i.e. PTEN mutation and TTK inhibition) has a greater effect on cellular fitness than either defect alone. One of the standard approaches to identifying such synthetic lethalities and candidate therapeutic targets in cancer is to use isogenic pairs of cell lines where the only engineered genetic difference between isogenic comparitors is the tumorigenic mutation in question (in this case PTEN) (26). Whilst these genetically simple isogenic systems can be very powerful in identifying the prime determinants of response to a particular therapeutic approach (27), when used in isolation they do not address the impact that changes in additional genes may have on any therapeutic response (28). Although the true extent of intra- or intertumoral genetic heterogeneity is not yet fully known, it seems likely that it is a key determinant of the response to treatment. On this basis alone, identifying synthetic lethal effects that are relatively unaffected by changes in additional genes (i.e. hard synthetic lethalities (28)) is critical. With these issues in mind, using genetically heterogeneous, non-isogenic panels of tumor cell lines to first identify candidate genetic dependencies (as we have done here), and then where available, using isogenic systems to validate effects is one rational approach to this problem. Using such an approach goes some way towards assessing the effects of genetic heterogeneity upon candidate therapeutic approaches and could even be complimented with additional methods, such as the use of synthetic rescue screens (29). Given the cost and time involved in the processes of pre-clinical and clinical drug development, it seems reasonable to propose that similar or complementary approaches to address the impact of genetic heterogeneity become commonplace prior to and during the drug development process.

Our identification of TTK as a candidate therapeutic target for PTEN deficient tumors is particularly interesting given the current interest in the development of clinical inhibitors of this mitotic kinase. At present a number of TTK inhibitors have been developed including relatively potent and orally bioavailable small molecules (19, 30). Primarily, the utility of TTK inhibitors has been proposed to be in the targeting of tumors defined by high levels of chromosome instability (CIN) (31) and/or those reliant upon the function of the spindle assembly check point (SAC) (32). Here, we define PTEN deficiency as an additional biomarker that could be used to direct the use of these agents. As well as defining PTEN gene defects by conventional Sanger sequencing or Fluorescence In Situ Hybridisation (FISH), tumor-specific PTEN deficiency can also be effectively defined by immunohistochemisty (33). Given this, there is the real potential that once developed into clinically useable agents, TTK inhibitors could be assessed in patient cohorts defined by tumor PTEN status. Moreover, the PTEN selectivity of agents such as PARP inhibitors (26) also suggests that combinatorial approaches using these agents together with TTK inhibitors in PTEN null tumors should also be investigated. In terms of understanding the mechanistic basis underlying the PTEN/TTK synthetic lethality, we show that PTEN null cells possess an underlying aneuploidy or genomic instability that is exacerbated by the effect of TTK inhibition. This paradigm has some precedence. For example, in the study by Gray and colleagues (19), a small molecule TTK inhibitor caused cells to exhibit abnormal numbers of chromosomes and was particularly lethal in the U2OS cell line that has an underlying aneuploidy phenotype. Aneuploidy is frequently observed in both human breast carcinomas that have low expression of PTEN and prostatic intraepithelial neoplasias from Pten mutant mice (17, 34). It is reasonable to conclude that inhibition of mitotic checkpoint kinases such as TTK exploit these underlying defects and thus elicit tumor selective cell death.

We have also demonstrated that ADCK2 silencing is selectively lethal in ER positive breast tumor cell lines. We also noted that ADCK2 inhibition abrogates estrogen signaling. The ADCK2 protein encompasses a predicted protein kinase catalytic domain, an aarF domain and is highly conserved (amino acid identity between H.sapiens and P. troglodytes being 99% and 54% between H.sapiens and D. rerio). Whilst the function of ADCK2 is little understood, other aarF-domain proteins, such as ADCK3, have been implicated in the metabolism of coenzyme Q (ubiquinone), a lipid electron and proton carrier in the electron transport chain (35). In addition, ADCK2 has been identified as a genetic dependency in a glioblastoma tumor cell line, a phenotype that also correlates with an increase in ADCK2 copy number (36). At present it is unclear whether ADCK2 itself has a coenzyme Q metabolism role or whether this plays a part in impairing estrogen signaling. There are currently no ADCK2 inhibitors available, however our results implicate ADCK2 as a possible candidate target for future drug development as a means to target ER positive breast cancer.

In summary, we describe how functional viability profiles of breast tumor cell lines may be used to identify novel candidate therapeutic targets as well as increasing our understanding of the fundamental dependencies that breast tumor cells carry. As far as we are aware, this is the first attempt to comprehensively identify genetic dependencies for a set of potentially “druggable” genes in breast cancer cell lines. In a system akin to the free availability of transcriptomic datasets, we believe that this functional data, and subsequent expansions of this approach, will be of significant use to the community. Importantly, the addition of further annotation to our dataset (for example data describing the phenotypic properties of each model and the whole genome sequence of each model) will hopefully enable the wider utility of this data. Given this, we have deposited all raw and processed data on the ROCK database (37) as well as the Stand Up To Cancer Breast Cancer Browser (due for release 2012), along with tools that facilitate data analysis. We envisage that expanding this data set to other cancer types should enable the identification of novel candidate targets in specific genetic and phenotypic subsets of cancer, and ultimately speed the route towards the development novel therapeutic approaches.

METHODS

Cell lines, compounds and siRNA

All cell lines were obtained from ATCC (USA) and maintained according to the supplier’s instructions. Cell lines were grown and transfected with SMARTpool siRNAs using Dharmafect 3 (DF3), Dharmafect 4 (DF4) (Dharmacon), Oligofectamine, Lipofectamine 2000 or RNAiMAX (Invitrogen) transfection reagents as indicated in Supplementary Table 9. The siRNA library (siARRAY – targeting 714 known and putative human protein kinase genes) was obtained in nine 96 well plates from Dharmacon (USA). Each well in this library contained a SMARTpool of four distinct siRNA species targeting different sequences of the target transcript. Each plate was supplemented with siCONTROL (ten wells, Dharmacon (USA)).

Antibodies

Antibodies targeting the following were used as per manufacturers instructions: ACTIN and C-MYC (Santa Cruz Biotech), TTK, ER, PR, ERBB2, CYCLIN D1, TFF1, FOXO1, C-JUN, PTEN C-terminus (Cell Signalling, Danvers, USA) and ADCK2 (Ab Cam). All secondary antibodies used for western blot analysis were HRP conjugated.

RNA interference screening

We transfected cells with the SMARTpool library, where each well of the 96 well-plate contained a pool of four different siRNAs (a SMARTpool) targeting one gene. After five population doublings (Supplementary Table 9), cell viability in each well was estimated by use of a highly sensitive luminescent assay measuring cellular ATP levels (Cell Titre Glo, Promega). To identify loss of viability/failure to proliferate effects in each cell line, luminescence readings from each well were log transformed and then centred by the plate median, to account for plate-to-plate variation common in high-throughput screens. Well position effects were identified and eliminated and the quality of data from each plate estimated by calculating Z’-factors (38) based on positive (siRNA targeting PLK1) and negative (non-targeting siRNA) controls in each plate. The dynamic range of each screen was determined by calculating Z prime values (38); we use a threshold of Z prime >0.3 to define acceptable screens, based upon previous screens where Z prime >0.3 is predictive of reproducible data (39, 40). To allow data to be compared between different cell lines and to minimise the impact of moderate variations in transfection efficiency between cell lines, plate-centred data from each screen were standardised by the use of a Z score statistic, where Z=0 represents no effect on viability and negative Z scores represent loss of viability. Here, Z scores were calculated using the Median Absolute Deviation (MAD) of all effects in each cell line (5). In each screen the Z score data approximated a normal distribution allowing comparison of the individual siRNA effects across cell lines (Supplementary Figures 1d-f). siRNA screens were carried out in triplicate and comparison of Z score data from replica screens of each cell line showed the screening process to be highly robust, as demonstrated by Spearman’s r² values approaching 1 for all comparisons (Supplementary Table 1). Quality metrics for each screen are shown in Supplementary Table 2. Raw luminescence values and processed data are now deposited on the ROCK Breast Cancer Functional Genomics Database (37) (rock.icr.ac.uk) as a community resource. This rigorous quality control method reduced our cell line panel to 20 cell lines with robust viability profiles (see Supplementary Table 3). Heatmaps of Z scores that separated different phenotypes as defined by a median difference permutation test were created to show the significant genes ranked row-wise according to their median difference and ordered column-wise by phenotype.

Western blots

Protein lysates were prepared using RIPA lysis buffer (50 nM Tris pH 8.0, 150 mM NaCl, 0.1% SDS, 0.1% DOC, 1% TritonX-100, 50 mM NaF, 1 mM Na3VO4and protease inhibitors). 100mg of total cell lysate was loaded onto prefabricated 4–12% Bis-Tris gels (Invitrogen), with full range rainbow molecular weight marker (GE Healthcare, UK) as a size reference, and resolved by SDS-PAGE electrophoresis. Proteins were transferred to nitrocellulose membrane (Bio-rad, USA), blocked and probed with primary antibody diluted 1 in 1000 in 5% milk overnight at 4°C. Secondary antibodies were diluted 1 in 5000 in 5% milk and incubated for one hour at room temperature. Protein bands were visualised using ECL (GE Healthcare, UK) and MR or XAR film (Kodak).

Validation of gene silencing by siRNA

Validation of RNAi gene silencing was determined by western blotting and by viability assays of silencing effects with individual oligos. Cells were transfected with individual ERBB2, ADCK2, ESR1, PIK3CA or TTK siGenome oligos (Dharmacon). Protein lysates were collected 48 hours following transfection for western blot analysis. Cell viability was measured using CellTiter Glo (Promega, USA) after five populations doublings.

Survival assays

For measurement of sensitivity to TTK inhibitor treatment, cells were plated in 96 well plates and exposed to the drug at the indicated concentrations. Cells were dosed at 24 and 96 hours. After 7 days, cell viability was measured using CellTiter Glo Luminescent Cell Viability Assay (Promega, USA). Surviving fractions were calculated and drug sensitivity curves plotted as previously described (27).

Metaphase Spreads

After 24hours of treatment with 2uM AZ3146, HCT116 PTEN wt or PTEN null cells were treated with colcemide (10ng/ml, Sigma) and MG132 (20uM, Sigma) for 1 hour. Cells were then lysed in hypotonic solution (0.03M Sodium Citrate) for 20 minutes at 37°C and fixed in methanol/acetic acid (3:1). Two or three drops of suspended cells were applied to glass slides and chromosomes were stained with DAPI.

PTEN FISH

Cell pellets of CAl51 and Hec-1b were washed twice with PBS and fixated with methanol/acetic acid solution (3:1) onto slides. LSI PTEN (10q23, red/orange) / chromosome 10 centromere (CEP 10, green) Dual Colour Probe (Abbott Molecular, IL, US) was hybridised to representative slides of the cell lines according to the manufacturer’s instructions. Signals were counted in 100 non-overlapping nuclei using the Leica TCS SP2 confocal microscope (Leica, Milton Keynes, UK).

Transcript profiling

RNA was extracted from cell lines with Trizol and phenol/chloroform extraction followed by isopropanol precipitation. For each cell line, triplicate extractions and profiles were performed. Biotin-labelled cRNA was produced by means of a linear amplification kit (IL1791; Ambion, Austin, TX) using 250 ng of quality-checked total RNA as input. Chip hybridisations, washing, Cy3-streptavidin (Amersham Bio-sciences) staining, and scanning were performed on an Illumina BeadStation 500 (Illumina, San Diego, CA) platform using reagents and following protocols supplied by the manufacturer as previously described (12). cRNA samples were hybridised on Illumina human-6 v2 BeadChips, covering approximately 47,000 RefSeq transcripts. The random distribution of large populations of oligonucleotide-coated beads across the available positions within the human-6 v2 chip enables, on average, 30 intensity measurements per RefSeq, yielding quantitative assessments of gene expression. All basic expression data analysis was carried out using the manufacturer’s software BeadStudio 3.1. Illumina expression profiles were performed in triplicate, the raw data were then variance-stabilizing transformed and robust spine normalised using the lumi package in the Bioconductor software. Expression values for each sample were median scaled and the mean expression value was established over the three replicates. Genes with significant difference in expression between cell lines were identified by one-way analysis of variance (ANOVA).

Array CGH (aCGH) analysis method

Genomic DNA was extracted from cell lines using the QIAamp DNA Blood Mini Kit (51104, Qiagen), according to manufacturer’s instructions. Microarray-based CGH analysis was performed on an in-house 32K tiling path BAC array platform as previously described (41). For copy number correlations, the circular binary segmentation (CBS) ratios of BACs containing the gene of interest were used for copy number correlations, and copy number assigned as previously described (41). Briefly, AWS smoothed log2 ratio values <−0.12 were categorised as losses, those >0.12 as gains, and those in between as unchanged. Amplifications were defined as smoothed log2 ratio values >0.4 (41). Data processing and analysis were carried out in R 2.9.0 (42) and BioConductor 1.5 (43), making extensive use of modified versions of the packages aCGH, marray and aws in particular.

Correlation of gene expression and aCGH data

aCGH and gene expression were compared by direct Pearson correlation of gene expression log intensity values with smoothed log ratio values for every probe in the gene expression data. Correlation p values were adjusted for multiple comparison testing using the local False Discovery Rate (FDR) method of Benjamini and Hochberg (44) as previously described (12). Mann Whitney U tests were preformed for each gene to compare gene expression values in groups defined as assigned as amplified or not, gained or not, lost or not and deleted or not using the thresholded aCGH calls. Wilcox test p-values were corrected for multiple comparison testing within contiguously altered regions. For each gene, cases in which genes were amplified, gained, lost or deleted were recorded along with the fold changes between samples carrying a copy number change and those which did not. Total counts of gains, losses, amplifications and deletions were also recorded.

Correlation of siRNA Z score with gene expression and aCGH data

The correlation between siRNA Z score and normalised gene expression was examined for genes where siRNA caused significant loss of viability (Z<−2). Z score was compared to normalized gene expression using Pearson correlation coefficient. A gene was taken as being significantly correlated if the Pearson correlation coefficient was significantly different to the null hypothesis, the correlation was inverse, and the variation in gene expression between cells lines were significantly different as assessed by one-way ANOVA.

For aCGH overlay, Z scores for the common 714 genes in the gene expression and the kinase library were correlated directly using Pearson correlation and p-values adjusted for multiple comparison testing using the FDR (12). Gene expression values were subsequently compared in those cell lines that displayed Z score hits for each gene and those that did not. Fold changes between these groups for each gene were also calculated. Array CGH copy number changes for each gene were used to divide the cell lines into those carrying amplifications, gains, losses or deletions, from those which did not and Mann Whitney U tests were performed to compare gene expression between these groups for each gene.

Clustering Methods

Supervised analysis was performed by calculating the absolute difference in the medians of the two groups, followed by estimating a p-value by permuting the labels to create a distribution of median differences against which to compare the actual effects. We used a significance cut-off of p=0.05 with no corrections for multiple comparisons.

The combined gene expression, Z scores and aCGH matrices were subjected to hierarchichal clustering using a Pearson correlation distance measure and Wards clustering algorithm. Equivalent heatmaps were drawn which clustered the samples and the genes in the combined overlay using the corresponding Z scores. Heatmaps of gene expression and copy number changes were then displayed under the same dendrograms derived by clustering by Z score. Stability of the clustering was established by pvclust giving an approximately unbiased (AU) p-value greater than 0.9. All analysis was performed in R 2.9.0 using in-house scripts.

Cell Line Verification

We used STR (short tandem repeat) profiling to verify authenticity of the panel of cell lines. We simultaneously amplified 8 STR loci, in a multiplex PCR reaction (Promega PowerPlex 1.2 System) and used the ATCC database for comparison, where possible. In addition, gain and loss of expression of particular subtype markers, were confirmed by western blot as well as microarray analysis.

SIGNIFICANCE.

Despite the wealth of molecular profiling data describing breast tumors and breast tumor cell models, our understanding of the fundamental genetic dependencies in this disease is relatively poor. Using high-throughput RNA interference screening of a series of potentially druggable genes, we have generated comprehensive functional viability profiles for a wide panel of commonly used breast tumor cell models. Analysis of these profiles identifies a series of novel genetic dependencies including the dependency of PTEN null breast tumor cells upon mitotic checkpoint kinases and provides a framework upon which additional dependencies and candidate therapeutic targets may be identified.