Use of RNAi Screens to Uncover Resistance Mechanisms in Cancer Cells and Identify Synthetic Lethal Interactions

Abstract

RNAi loss-of-function screens, which have proven effective to identify genes functionally responsible for cellular phenotypes, can be designed to use different genetic backgrounds or altered environmental conditions to elucidate genetic dependencies. These sorts of screening approaches can be exploited to identify genetic targets that minimize resistance to approved drugs, and provide a basis on which to develop new targeted therapies and predict the secondary targets for combinatorial treatments. Four types of pooled short hairpin RNA (shRNA) screens, in particular, have been used to look for genetic targets that work together with known drugs or other anticancer targets, either in an additive or synergistic fashion. Each method produces results that provide a useful but limited picture of the genetic elements driving oncogenesis.

Introduction

In the last 10-15 years, molecular classification of cancers has evolved to the point that it is replacing traditional tissue-based characterizations and becoming the dominant paradigm for developing new diagnoses and treatments. This molecular perspective has provided significant insights into carcinogenesis that have led to novel therapeutic approaches. However, although there have been major advances in molecular analysis over the past 20 years, our understanding of cancer genetics remains very rudimentary.

Since Golub et al. [1] first proposed classifying cancers using gene expression data, most oncological molecular characterization has been based on transcriptome profiling and DNA sequencing [2,3,4]. Although these analyses have been very useful in targeted drug development for a number of cancers, these techniques really only allow the elucidation of genetic characteristics that correlate rather than show a causal connection with oncogenesis. With these techniques, it is not possible to separate the genetic elements driving an effect from those that have been changed as a response to the effect (i.e., the genetic “passengers”).

There are two molecular technologies, however, that do actually disrupt gene functions, and so, reveal the genetic factors that produce a specific phenotype. Both gene knockout and RNA interference technologies provide causal gene-phenotype data. A major problem, however, with both of these approaches, is that most phenotypes do not result from single gene effects [5], and the lack of understanding how genetic networks are “wired” impedes understanding of the genetics of most heritable traits [6]. As a result, knockout or knockdown of a single essential gene or pathway only provides limited information about the genetic network interactions producing a phenotype. Further, cancer cells, almost by definition, are characterized by pathway rewiring, short circuits, and other signal transduction anomalies as a result of genetic lesions [7], and the high level of adaptation and heterogeneity of tumor cells endow them with a remarkable ability to work around therapeutic challenges [8]. Thus, inhibition of a single essential target or pathway will often not produce sustained therapeutic effects, and combination therapies will likely be the required treatment for most cancers.

In an attempt to expand the repertoire of combinatorial treatments, groups have run many screens to identify the effect of drug pairs on different types of cancer cells [9,10]. However, although some effective combinations have been found, for example, poly(ADP) ribose polymerase inhibitors (PARP) in combination with BRCA-positive breast cancer treatments [11], success with this approach has been limited. To expand the universe of potential combinations requires a better understanding of the genetic interactions present in oncogenic cells. As a result, there is considerable interest in unbiased approaches to uncover synthetic lethal interactions. Broad-based RNAi screening, with its ability to assay thousands of genes simultaneously to identify the small fraction driving a specific response, provides a scalable approach to screen for functionally critical genes across a range of cell types [12,13,14]. Further, since RNA interference inhibits the expression of, and hence, the function of a gene, its effect models, to some extent, that of a drug inhibiting the function of a specific gene product (protein). As a result, this technique offers an apt approach to screen for synthetic lethal interactions.

It is relatively straightforward to screen all 20,000 human protein-coding genes using RNAi to find those that are essential in a particular cell system as genome-scale screening studies like Sims et al. [15] and Marcotte et al. [16] have shown. However, a dropout viability screen of a cell line only uncovers the most obviously essential genes that by themselves produce a lethal phenotype when disrupted. Unraveling more complex genetic interactions that generate the bulk of cancers requires multiple screens in various cell models and the integration of these results with other molecular data, such as two-hybrid maps and transcriptome profiles. The particular challenge in identifying synthetic lethal interactions is their combinatorial nature. Synthetic phenotypes do not manifest unless multiple genes are simultaneously disrupted and, just testing all paired combinations of 20,000 genes, for example, produces 400 million assay combinations—and most traits likely involve the interaction of more than just two genetic elements. To make meaningful inroads into such a large assay space requires creative and shrewd use of the limited experimental tools available, and pooled shRNA screening is a pivotal tool in the toolbox.

Screening Strategies to Uncover Synthetic Interactions Using Pooled shRNA Libraries

Several curated collections are available to facilitate running RNAi screens in arrayed formats with each well containing individual or small pools of siRNA molecules or shRNA constructs. One advantage of this approach is that the oligonucleotides or expression constructs in this format can be used in combination with other biological assays, including high-content screening. For example, Laufer, et al. [17] used combinatorial RNAi screens in an arrayed format to look at the interaction of epigenetic regulatory genes on several morphological phenotypes of colon cancer cell, such as eccentricity, nuclear area, and mitotic index. However, given the expansive assay domain that needs to be mapped to identify synthetic lethal interactions, single assay screens using complex pools of shRNA expression constructs offer a more practical approach.

Complex pooled shRNA libraries are typically constructed in a lentiviral vector system so that large numbers of shRNA expression constructs can be easily introduced into cells efficiently. With a pooled-screen approach, a single library that encodes the whole set of packaged viral shRNA constructs is introduced into a population of cells via a single large-scale transduction (Figure 1). The shRNA expression cassettes integrate into the genomic DNA of transduced cells and generate stable gene-knockdown cells. The library-transduced cells are then incubated for a period of time to allow the shRNA to express and the phenotype to manifest. When running dropout viability assays to identify cytotoxic shRNAs, we typically maintain transduced cells for 6-12 doublings. After growth, surviving cells are harvested, and the relative quantity of each shRNA construct is determined by high-throughput sequencing of the expression cassettes recovered by PCR from the cells' isolated genomic DNA. The relative quantities of each shRNA in the cell population after several doublings can be compared with the relative quantities of each shRNA in the initial library to identify which shRNAs are underrepresented in the cells after several doublings. The “depleted” shRNAs either killed or inhibited growth of the cells, presumably because they interfere with genes required for survival or robust cell growth and proliferation. One important advantage of the RNAi pooled-screening technique is that it does not require major automation infrastructure or special liquid handling.

Figure 1.

Dropout Viability Screening Process. An overview of the dropout viability screening process to find genes required by cells growing in the presence of a drug or other factor. The left panel screen identifies genes generally required by the cell line for proliferation. In the right panel, screen cells are treated with a low dose of the compound after transduction. This screen finds both generally essential genes and genes essential for growth in the presence of the drug. Hits identified specifically in the right panel screen with the drug indicate genes whose expression confers some resistance to the drug since loss of the gene's function enhance sensitivity.

Pooled shRNA dropout screens have been successfully used by a number of groups for several years to identify essential genes [18,19]. Successful application of this technique simply requires the use of the well characterized libraries and careful screening methodology. Some of the initial libraries were developed about 10 years ago, and a number of advancements have been made since then. For example, to facilitate accurate measurement of the relative amount of each shRNA in a population, our libraries include short unique sequences optimized for high-throughput analysis (i.e., barcodes). PCR amplification followed by high-throughput sequencing of these optimized barcodes is more robust and less variable than direct amplification and sequencing of the hairpin sequence of the shRNA (data not shown). Also, use of libraries optimized for high-throughput sequencing ensures all shRNAs are represented in comparable numbers to accurately assess depletion or enrichment during a screen. This sort of analysis of library quality is critical since the most important factor influencing screening reproducibility is the degree to which the shRNAs in the population of transduced cells at the start of a screen reflects the full complement of shRNAs present in the original library. A sufficient number of cells need to integrate each construct in the library during the initial transduction, and the progeny of these cells must be maintained at high enough levels through screening to ensure that changes in the abundance of cells expressing a given shRNA are due to the effect of that shRNA on cell proliferation, and not simple stochastic variation. If well-characterized libraries are not used and these parameters defining the size the screen relative to the shRNA frequency and distribution in the library not observed, then it is difficult to separate true hits from random variation in he results. In such cases, simple noise can easily be mistaken for an important result, as was the case where STK33 was identified as essential in KRAS-dependent cancer cells [20]. This work was later showed to be a result of some procedural problems with the screen [21,22].

Below we provide some examples of four general approaches that use complex pooled shRNA libraries to identify different sorts of synthetic lethal interactions:

Dropout Viability Screens with a Compound

Perhaps the most direct approach to identify essential gene targets in cells resistant to a drug is simply to run a dropout viability screen in the presence of the drug. This approach uncovers genes that, when inhibited, sensitize cells to a drug's lethal effect. This sort of screen involves simply treating cells with a minimal dose of the drug, then running a dropout viability screen to identify genes whose loss of function is lethal. “Hits” (i.e., the target genes of depleted shRNA) from the screen run in the presence of the compound are, of course, then compared with a similar set of gene hits from a control screen that does not include the compound. This analysis enables identification of the interference targets that are critical for cell viability only in the presence of the drug—the ones that are synthetically lethal with the compound.

A study by researchers at the Netherlands Cancer Institute and University of Torino used a dropout viability screen with a pooled shRNA library targeting 535 kinase genes to ascertain why melanoma cancer cells harboring the BRAF (V600E) mutation are sensitive to PLX4032 (vemurafenib) whereas prostate cancers with the same mutation are not [23]. Colorectal cells containing the BRAF mutation and resistant to PLX4032 were infected with the pooled library and cultured in the absence or presence of the compound. Three of the shRNA that were depleted in this screen targeted EGFR, and it was confirmed that suppression of EGFR only in the presence of PLX4032 markedly inhibited growth of these cells. The cells also responded strongly to combination treatment of PLX4032 with EGFR-targeting antibodies (either cetuximab or gefitinib). Subsequent follow up work showed that BRAF inhibition activates EGFR.

The fact that most melanoma lines, and colorectal cell lines sensitive to PLX3032, express very low levels of BRAF seems to explain the difference in sensitivity of these two cancers to this compound.

A recent publication from the German Cancer Research Center (DKFZ) made use of this screening method to identify targets that act synthetically with the approved drug gemcitabine [24]. The researchers looked for genes that support pancreatic cancer cell resistance to gemcitabine by running dropout viability RNAi screens in the presence of the drug using Cellecta's DECIPHER shRNA libraries targeting 10,000 human genes. A pancreatic cancer cell line was transduced with the libraries, then treated with low levels of gemcitabine—or not, for the control—and then allowed to grow for an extended period. The authors identified about 70 genes with synthetic lethal effects in combination with gemcitabine. Highly represented in the hits from the gemcitabine screen were genes involved in DNA damage response and repair, which was expected since gemcitabine is a DNA damaging agent. The study authors focused specifically on genes identified in the screen that were upstream of checkpoint kinase 1 (CHK1) of the ATR/CHK1 pathway, and confirmed RAD17, HUS1, and WEE1 with three independent shRNA constructs. Further analysis revealed that these three genes appear to increase gemcitabine lethality by forcing treated cells with damaged DNA to enter mitosis.

Other examples of this sort of dropout screen to identify sensitizers in cells resistant to a drug include a screen by Qin et al. [25] to find synthetic lethal genes that work in conjunction with the anti-inflammatory/antitumor agent CDDO-Me, and Mills and colleagues' study [26] that identified lethal sensitivity of ABT-737-treated lymphoma cells to DHX9 knockdown.

It is also possible to use a similar, but “reverse,” version of this screening strategy to identify genes that are required for the cell to maintain sensitivity to a drug, sometimes called synthetic dosage lethality [27]. To find genes whose expression (rather than disruption) enhances sensitivity to a drug or other factor when a high dose of the drug, or other death-inducing factor, is used to kill the bulk of the cells, then the analysis is done with the few that survive this selection. RNAi screening to identify genes involved in an apoptotic response, for example by Dompe et al. [28], is one example of this sort of positive-selection screen. Similar sorts of synthetic dosage effects can be found for drugs, although it is less common than lethality caused by a simple disruption of two (or more) targets. This is probably because most molecularly targeted anticancer drugs work by disruption of a signaling pathway. However, there are genetic mechanisms that can ameliorate the effects produced by inhibiting an essential pathway. For example, in the lymphoma screen with ABT-737 discussed in the previous section, cells overexpressing myeloid leukemia cell differentiation protein Mcl-1 display resistance to the drug. In fact, this synthetic dosage lethal mode of resistance needed to be incorporated into the cell model through constitutive expression of Mcl-1 to enable effective selection of synthetic lethality effects in this study.

Dropout Viability Screens in Defined Genetic Backgrounds

Instead of looking for synthetic lethal gene interactions directly with a drug, an alternative is to look for genes that are only essential in cells with specific genotypes or particular genetic lesions. This sort of screen reflects the strict definition of synthetic lethality—two or more viable non-allelic genetic variants are lethal when present together in the same cell. Actually, most RNAi dropout viability screens fall into this category. For example, one of the early dropout screens by Schlabach et al. [29] using a murine stem cell virus library of 8,203 shRNAs targeting 2,924 genes found essential genes specific to two colon cancer cells they screened, but the hits did not show up in parallel breast cancer or mammary cell screens. Around the same time, in a more narrowly defined synthetic lethal screen targeting only 88 kinases, Bommi-Reddy et al. [30] found that three kinases—CDK6, MET, and MAP2K1—were lethal in renal carcinoma cells negative for the von Hippel-Lindal (VHL-) tumor suppressor activity, but not in an isogenic line without the VHL mutation.

Also, more recent studies have profiled panels of cells to identify vulnerabilities in cancers associated with specific tissues. For example, Cheung et al. [31] screened more than 11,000 genes in 102 cancer cell lines and identified a number of lineage-specific dependencies, then cross referenced the results with transcription profiling data of tumors. They found 5 genes specifically essential in ovarian cancer cells and overexpressed in a high percentage of ovarian tumors. One of these genes, PAX8, plays a critical role in female genital track development and induces apoptotic cell death when inhibited. Brough et al. [32] profiled 30 common breast cancer cell lines using arrayed siRNA screening and discovered that cells with mutated PTEN were sensitive to inhibition of TTK protein kinase which is involved in mitotic spindle assembly. Also, estrogen positive tumor lines were sensitive to ADCK2 knockdown. We also have run a tissue-specific dropout viability study in-house to identify genes specifically lethal to androgen-receptor positive prostate cancer cells using a small shRNA library targeting 1,200 prostate-cancer associated human genes. This small screen identified cyclin D1 as specifically essential in these cells but not in androgen receptor negative variants (Figure 2) [33].

Figure 2.

Dropout Viability Screen to Identify Prostate-Specific Essential Genes. Panel A is a “heat map” that depicts levels of significant depletion for 7,000 shRNA in a screen targeting 1,200 prostate-specific genes in 5 prostate and 3 other cell lines. Red or orange indicates no significant change in the shRNAs targeting the gene over time, whereas blue indicates genes with the most significantly depleted shRNAs. Panel B: Confirmation result for 17 genes identified in the screen that had two or more shRNA targeting them significantly depleted specifically in prostate cells, as well as two controls that were not depleted in any cells. For confirmation, two shRNAs targeting each selected gene were synthesized, cloned, and transduced into all cell lines. The table compares the anticipated effect of the shRNAs on the cells based on the screening data (predicted) with the actual effect seen when the shRNAs were expressed in the cells.

As a well-known oncogene, several groups have tried to identify synthetic lethal interactors with KRAS. As mentioned above, it was one of these early screens, by Scholl et al., which led to STK33 putative synthetic interaction with KRAS that attracted a lot of attention but later turned out to be a false positive. However, a later study by Vicent et al. [34] did identify several KRAS-specific lethal genes in lung carcinoma. The researchers screened two mouse KRAS-driven non-small lung cancer cell lines with a pooled lentiviral library consisting of 631 shRNA against 162 targets that had been implicated in the KRAS genetic network based on previous broader RNAi screens and gene expression analyses. Interestingly, for this analysis, the authors ran the screen using in vivo xenograft implantations of the transduced cells, as well as the standard in vitro culture. Of the 23 genes with at least two depleted shRNAs, three—Rac1, Phb2, and Wt1—were found to be specifically lethal only in cells where KRAS-signaling is active. Rac1 had been previously linked to KRAS signaling and Phb2 is a known chaperonin protein active during apoptosis. However, the function of Wt1, which appeared to have some transcriptional activation activity, was unknown. Vicent et al. went on to demonstrate that WT1 interference leads to growth arrest and senescence in the presence of oncogenic KRAS.

Dropout Viability Screens in Defined Phenotypic Models

Rather than looking for genes that produce specific lethality in cells with particular genetics as described above, a similar approach can be used to discover genes specifically lethal to cells generally displaying a common phenotype.

Lamy et al. [35] recently published a study that used this approach to identify a dependence of multiple myeloma cells on caspase-10. Myeloma cells are genetically heterogeneous. Despite this, genetically consistent dependencies across all myeloma cells have been found. For example, myelomas appear to require transcription factor IRF4 expression [36]. Based on this, the researchers sought out other essential drivers specific to myeloma cells. Using an inducible retroviral-based shRNA library targeting 2,500 human genes, the authors screened three myeloma lines and found that shRNA targeting caspase-10 was depleted significantly compared with screens in four other lymphoma lines. They pursued these results and determined that caspase-10 is a downstream effector of IRF4 and blocks autophagy-dependent death of myeloma cells.

This phenotypic-centric approach can also be used to screen for genes required for drug resistance. As outlined above, one obvious way to look for genes whose disruption enhances sensitivity to the drug is to run a dropout screen in the presence of the drug. However, an alternative approach is to look for any gene specifically essential for the viability of cells displaying the drug resistant phenotype. The premise, of course, is that cells that have developed resistance to the drug have rewired their pathways in some way to minimize the disruption caused by the compound. Dropout screens to find genes specifically essential in cells with the resistant phenotype, then, can help elucidate the mechanism of resistance and, further, may provide novel drug targets that can be used to prevent the development of resistance in the presence of the drug. We have run this sort of screen to look for required pathways in ovarian cancer cells resistant to cisplatin as shown in Figure 3.

Figure 3.

RNAi Screen to Identify Essential Genes Specific for Cisplatin Resistant Ovarian Cancer Cells. We screened ovarian cancer cell line A2780 that is sensitive to cisplatin and a cisplatin-resistant derivative A2780R line with the DECIPHER library that targets 5,000 genes with 27,500 shRNA expression constructs. Depleted shRNAs found in each cell line after the screen were cross-compared to identify viability gene targets unique to A2780R cells. About 200 putative cisplatin-resistant specific viability genes were identified by at least 2 independent shRNAs. Preliminary pathway analysis revealed that these hits were significantly associated with >20 putative viability pathways that included cell cycle checkpoints, DNA mismatch and nucleotide excision repair, as may have been expected for a cisplatin resistant phenotype. A similar experiment using Affymetrix U133+2 arrays to identify up- and down-regulated genes in cisplatin resistant vs. sensitive cells found approximately 100 differentially regulated transcripts, of which about 50% overlapped with the RNAi screen.

Combinatorial Gene-by-Gene Knockdown Analysis

A very direct RNAi-based approach to find lethal gene combinations is simply to look for lethality when two or more genes are knocked down in a single cell. One approach to do this is to run standard dropout screens using a stable knockdown cell line in the same way one would run screens using cells with a knocked out gene. However, this approach becomes more difficult when the object is to look for a broader range of lethal combinations, say, for example, if one wanted to find all synthetically interacting gene pairs in two cellular pathways.

To map out synthetic lethal interactions between large sets of genes requires a way of assessing, on a paired gene-by-gene basis, the effect of knocking down each two-gene combination at the same time. For this sort of analysis, we have developed a vector that expresses two shRNAs simultaneously and devised an approach to construct a complete shRNA library with a defined set of hairpins expressed from each position of this vector. The approach creates a library wherein each shRNA expressed in one position is individually paired with all the shRNAs in the second position, allowing all combinations of the two sets of shRNAs to be assayed in a single screen. We tested this combinatorial library using a dual shRNA expression vector with a set of four shRNAs targeting each of 40 DNA damage and repair genes. The total library contained over 25,000 constructs since each construct expressed a different paired combination of the 160 shRNAs. After running a standard dropout viability screen, dual shRNA expressing constructs targeting 13 pairs of genes were significantly depleted in hTERT-immortalized human mammary epithelial cells (HMEC), including the PARP1/BRCA1 combination that has been previously reported by Ashworth [11].

While it is not feasible to run this sort of combinatorial pair-wise RNAi screen exhaustively on the whole genome in a single screen due to the number of combinations, a pooled screen with a dual shRNA expression construct is really the only practical option for this sort of analysis with even a moderately sized set of shRNAs. A pair-wise combinatorial approach using an arrayed format to screen two targets per well rapidly becomes very costly and resource prohibitive. However, a reasonable set of genes can be selected for a combinatorial pooled screening using any number of techniques. For example, Bassik et al. [37] used this approach to characterize the interactions in a set of genes involved with ricin sensitivity. In this study, the authors used a dual shRNA expression library for a follow up screen designed around the positives identified in a broad-based standard shRNA loss-of-function screen. Also, this sort of combinatorial screen using a pooled library can be run in conjunction with one of the previously mentioned approaches to uncover interacting genes specific for cells with certain genotypes or phenotypes.

Conclusion

Cancer patients often talk about their “battle” with the disease, and the metaphor is certainly an apt fit. Cancer treatment can be viewed as an ongoing war against an evolving population of tumor cells fighting to neutralize the effect of the latest chemical attack. The disease almost always recovers from the first attack, and the result leads to an escalating arms race of therapeutic intervention.

It is clear that in many, if not most, cases eliminating the resilient carcinogenic invaders will require a combination of approaches. Determining effective combinations for different tumors, though, requires a better understanding of the tactics each employs to overcome therapeutic intervention. What cellular networks do various cancers rely on for rapid growth, what pathways can they fall back on in response to a challenge, how do different cancers suppress normal growth controls, and what novel genetic vulnerabilities do the cells expose through this rewiring? It is this knowledge that will eventually enable us to anticipate and prevent resistance mechanisms and thwart the tumor adaptations before they can be established. At present, RNAi screening is one of the few tools available that uncovers the effective particular genetic elements responsible for producing and preserving these oncogenic phenotypes.

Table 1.

A comparison of RNAi dropout viability screening methods

Screening Method	Results Yield...	Results Predict...	Results May Show...
Screen with Compound	Genes involved in resistance to drug effects	Genes that can be targeted to enhance sensitivity to known drug	Drug mechanism of action and/or compensatory pathways for resistance
Screen in Defined Genetic Background	Genes that are synthetic lethal with defined oncogenic lesions	Genetic vulnerabilities in cancers with specific genetic lesions	Genetic “addictions” produced by various genetic lesions
Screen in Defined Phenotypic Models	Genes essential in cells with particular traits	Possible phenotypic markers associated with genetic sensitivities	Genetic factors responsible for phenotypes
Combinatorial RNAi Screen	Essential paired gene combinations with additive or synergistic interactions	Most lethal gene combinations to target for therapeutic intervention	Unexpected and novel genetic interactions