Synthetic lethal interactions enable a novel approach for discovering specific genetic vulnerabilities in cancer cells that can be exploited for the development of therapeutics. Despite successes in model organisms such as yeast, discovering synthetic lethal interactions on a large scale in human cells remains a significant challenge. We describe a comparative genomic strategy for identifying cancer-relevant synthetic lethal interactions whereby candidate interactions are prioritized on the basis of genetic interaction data available in yeast, followed by targeted testing of candidate interactions in human cell lines. As a proof of principle, we describe two novel synthetic lethal interactions in human cells discovered by this approach, one between the tumor suppressor gene SMARCB1 and PSMA4, and another between alveolar soft-part sarcoma-associated ASPSCR1 and PSMC2. These results suggest therapeutic targets for cancers harboring mutations in SMARCB1 or ASPSCR1 and highlight the potential of a targeted, cross-species strategy for identifying synthetic lethal interactions relevant to human cancer. Cancer Res; 73(20); 6128–36. ©2013 AACR.
Synthetic lethality is an exciting new avenue to disrupt cancer cells for targeted treatment. Two genes are said to be synthetic lethal if mutations in both genes cause cell death, but a mutation in either of them alone is not lethal. In applying synthetic lethality to the discovery of cancer drugs, the goal would be to identify a target gene that when mutated or chemically inhibited, kills cells that harbor a specific cancer-related alteration, but spares otherwise identical cells lacking the cancer-related alteration (1). This concept has recently been exploited in the development of PARP inhibitors as novel chemotherapeutics for breast cancer. Although PARP is not an essential gene in normal cells, BRCA mutant cells are dependent on PARP for their survival. The described efficacy of an oral PARP inhibitor, olaparib (AZD2281), in early-phase clinical trials for treating BRCA mutant tumors is a remarkable success story for translational cancer therapeutics (2). Importantly, strategies based on synthetic lethal interactions enable drug targeting of cancer-specific alterations in tumor suppressors that might otherwise be untreatable by drugs. Several recent studies have reported large-scale assays on the basis of on RNA-interference (RNAi) technology to discover synthetic lethal interactions with common cancer mutations, including BRCA1/2 and RAS genes (3–9). These studies typically target cells with a well-defined genetic background using a library of short hairpin RNAs (shRNA) to identify combinations that result in cell death or growth inhibition. Although such approaches have the potential to rapidly discover genetic interactions at a full genome scale, a number of technologic challenges remain to be solved, and the number of independently validated interactions produced by these efforts has been relatively limited to date (10).
One complementary strategy to whole-genome screens in cancer cell lines is motivated by the wealth of potentially relevant interaction data in model organisms. Publication of the first eukaryotic genome-scale genetic interaction map in yeast (Saccharomyces cerevisiae; ref. 11), where approximately 30% of all possible gene pairs were tested for interactions, provides a unique opportunity for discovering potentially therapeutic synthetic lethal interactions. For example, putative synthetic lethal interactions in human could be inferred on the basis of yeast synthetic lethal interactions between conserved genes in yeast and human. These predicted pairs of human genes provide a rich database of possible candidates for further study in the context of human disease. In fact, several interactions related to chromosome stability have already been mapped from yeast to worm to human (12), suggesting that such a strategy has the potential to yield promising new drug targets. In combination with the exponentially accumulating volume of data regarding the landscape of genomic alterations in human cancer, such an approach has the potential to become increasingly powerful going forward.
We describe a combined computational and experimental approach whereby yeast interactions between human orthologs are filtered by cancer association and interaction strength in yeast, and candidates from the prioritized list are then validated in human cell lines. Using this approach, we discovered two previously unknown synthetic sick interactions: one between SMARCB1 (yeast SNF5) and PSMA4 (yeast PRE9), and another between ASPSCR1 (yeast UBX4) and PSMC2 (yeast RPT1). The predicted synthetic sick/lethal interactions between these genes were validated with shRNA double knockdown in multiple cell lines and single knockdown of PSMA4 in two cancer cell lines containing endogenous SMARCB1 mutations. These interactions suggest potentially new therapeutic targets for SMARCB1 and ASPSCR1 mutated cancers and, more broadly, illustrate the potential of this cross-species approach.
Materials and Methods
Cell culture and shRNAs
IMR90, 293TN, A-204, G-401, and 293 cell lines were obtained from the American Type Culture Collection. IMR90, 293TN, and 293 cells were maintained in Dulbecco's Modified Eagle's Medium (DMEM) supplemented with 10% FBS, penicillin, and streptomycin. A-204 and G-401 cells were cultured in McCoy's 5A media containing 10% FBS, penicillin, and streptomycin. Bacterial stocks of control and validated gene-specific shRNA-expressing vectors, including PSMA4 and SMARCB1 shRNAs, were selected from the RNAi consortium database and purchased from Sigma-Aldrich.
Preparation of viral particles
Bacterial stocks of validated shRNAs clones were amplified and DNA extracted using the HiSpeed Plasmid purification kit (Qiagen). 293TN cells were then transfected with shRNA vector clones mixed with viral package vectors pMD2 and psPAX2 using Lipofectamine 2000 transfection reagent (Invitrogen). After 48 hours, culture media containing viral particles were mixed with polybrene and centrifuged at 10,000 rpm to precipitate and concentrate the viral particles.
RNAi-mediated gene knockdown
IMR90 cells were seeded into 96-well plates and transduced with predetermined pairs of shRNAs to generate four conditions with six replicates each: (i) control shRNAs; (ii) control shRNA/PSMA4; (iii) control shRNA/SMARCB1; and (iv) PSMA4/SMARCB1. A similar set-up of four conditions were used for other pairs of interactions tested. Cells from each treatment were cultured for eight and 10 days, and the number of viable cells determined by the CellTiter-Glo Luminescence Cell Viability Assay (Promega). This assay determines the number of viable cells in culture based on the amount of ATP produced by the living cells and is designed for use with multiwell plate formats and high-throughput screening for cell proliferation and cytotoxicity assays. The addition of the assay reagent results in cell lysis and generation of a luminescent signal proportional to the amount of ATP present, which is directly proportional to the number of living cells present in each well. The intensity of the luminescent signal was measured in relative luminescence units (RLU) using the Beckman Coulter DTX 880 multimode plate reader.
Cell lysates from 293, A-204, G-401, and IMR90 control cells, and from shRNA-infected cells were extracted after five days of incubation and quantified using Bio-Rad Protein Assay reagent. An equal amount of protein (50 μg) was subjected to SDS–PAGE, transferred onto a polyvinylidene difluoride membrane and blocked with nonfat milk. The membranes were then incubated in primary antibody overnight at 4°C and then with anti-mouse (1:6,000) or anti-rabbit (1:1,000) secondary antibody at room temperature for one hour. Primary antibodies rabbit anti-SNF5 (SMARCB1) and rabbit anti-TUG (ASPSCR1) were purchased from Cell Signaling, whereas rabbit anti-PSMC2, 20S Proteosome α-4 (PSMA4), and anti-β-actin-HRP were obtained from Santa Cruz Biotechnology. Protein expression was detected using enhanced chemoluminiscence substrate (Pierce).
Estimation of the significance of genetic interactions in human shRNA experiments
Growth rate of a single or double shRNA knockdown relative to the empty shRNA vector control were calculated using
where is relative growth rate for a single or double knockdown experiment (A, B, or AB), and is the intensity of the luminescent signal measured in relative luminescence units (RLU). Because is a ratio of two quantities that has error associated with it, error for is given by
where is the SD in the n (6 for our experiment) observations of .
Expected double mutant fitness and error associated with it is given by
To assess the significance of the interaction, we assumed a normal distribution for and , and compared with using the Welch t test (13). The significance of the difference between and can be calculated using a one-tailed t test, which requires the t test score t (13) and degree of freedom ν given by the Welch–Satterthwaite equation (14), as follows:
Here, because we have six replicate observations for control, single, and double knockdown experiments.
InParanoid7 (15) was used to map yeast genes to human genes. Only 1:1 orthologs were used for our study (Supplementary Table S1).
Collection and processing of yeast genetic interaction data
Yeast genetic interaction data was taken from the report by Costanzo and colleagues in 2010 (11), which reported data for interactions between 1,711 query genes and 3,885 array genes. We applied a P value cutoff of less than 0.05 on all interactions. Furthermore, we applied an interaction cutoff in two ways: first, we considered stringent negative genetic interactions (ε < −0.2) and, second, we allowed intermediate interactions (ε < −0.08), which were reported in reciprocal screens. Specifically, in the Costanzo and colleagues network, query genes were screened against the entire nonessential deletion array and, in some cases, genes present on the array were also screened as queries. For these cases, an interaction between genes A and B was tested in both screens: A (query) × B (array) and B (query) × A (array). In such cases, we applied an intermediate cutoff because an interaction appearing in both of these screens is of high confidence.
Eleven new SGA screens were also used to generate candidate gene pairs, including screens for the following queries (human/yeast orthologs): XPC/RAD4, VTI1A/VTI1, NOP56/NOP56, POLD2/POL31, MLH1/MLH1, XPO1/CRM1, UBA3/UBA3, ERCC4/RAD1, XPA/RAD14, PSMC2/RPT1, and PSMB1/PRE7. A screen involving a temperature-sensitive (TS) allele of yeast RPT1 (human PSMC2) was the basis for testing the human interaction PSMC2–ASPSCR1; therefore, the yeast interaction data supporting that inference are included here (Supplementary Table 2). A genome-wide screen for the RPT1 TS allele's genetic interactions was conducted as described by Baryshnikova and colleagues in 2010 (16). Briefly, a rpt1-1 mutant strain marked with a nourseothricin (NatMX4) resistance cassette and harboring the SGA haploid-specific markers and reporter (16) was mated to an array of approximately 4,000 viable S. cerevisiae deletion mutants. Nourseothricin- and Geneticin-resistant heterozygous diploid mutants were selected and sporulated and MATa rpt1-1 double mutants were subsequently selected (16). To confirm the SGA results, all gene deletions were constructed in a SSL204 MATa strain and crossed with an isogenic rpt1-1 MATα strain. Diploid cells were sporulated at 25°C and dissected. Plates were incubated for 3 to 5 days at either 25°C or 30°C.
Yeast tetrad dissection
Confirmations by tetrad analyses were conducted as described previously (17).
To discover cancer-associated genetic interactions in human cells, we first selected a set of highly significant interactions between yeast genes from the large network of synthetic genetic interactions that has recently been mapped in yeast. A recent study reported testing genetic interactions for 5.4 million yeast gene pairs, consisting of instances where two nonessential genes were deleted in combination, or a temperature-sensitive mutation of an essential gene was used together with a deletion of a nonessential gene (11). In total, approximately 116,000 pairs were reported as having a detectable synthetic sick or lethal interaction, of which approximately 24,000 interactions connect two genes that both have human orthologs (Fig. 1). More than 500 of these latter interactions involve at least one gene that has been previously associated with mutations in cancer (Sanger Institute Cancer Gene Census; Fig. 1B; ref. 18), suggesting a large number of candidate pairs can be generated by this approach (Supplementary Table S3).
To narrow the candidate list for testing in human cells, we first applied a very stringent cutoff on interactions in yeast, either requiring a high-magnitude effect, high-confidence interaction to be reported (ε < −0.2; P < 0.05) or selecting gene pairs for which interactions were reproduced in two reciprocal screens (see Materials and Methods for details). Furthermore, we restricted our search to genes with one-to-one orthologs in human to increase the likelihood of functional conservation between yeast and human, and to avoid potentially buffering effects of paralog functional redundancy (Supplementary Table S1; ref. 19). Applying these relatively stringent criteria, we obtained 1,522 putative synthetic sick/lethal interactions between human orthologs of yeast genes, of which 70 interactions involved a gene that has been previously implicated in some form of human cancer (Fig. 1B; Supplementary Table S4). In addition to these published interactions, we applied the same criteria to 11 previously unpublished yeast screens involving human orthologs (see Materials and Methods; Supplementary Table S2). Candidate interaction pairs involving cancer-associated mutations (Sanger Institute Cancer Gene Census) were ranked on the basis of the strength of the yeast interactions, and were selected in order up to a maximum of three interactions per gene. In total, 21 pairs of genes representing mutations associated with a diverse set of cancers were selected for further experiments in human cell lines (Fig. 2).
The candidate synthetic sick or lethal pairs derived from the yeast genetic interaction network were screened in normal human IMR90 fibroblast cells using an RNAi approach. IMR90 cells were chosen because the cell line was established from the lungs of a 16-week female fetus and have the advantage of early passage and a low likelihood of accumulated genetic alterations. This stable genetic background allowed us to assess the validity of candidate interactions with the lowest possibility of unknown, confounding genetic alterations. We screened the selected 21 pairs of potential interactions using a CellTiter-Glo luminescence viability assay. We found evidence for significant synthetic sick or lethal interactions for 6 of the 21 tested pairs (see Supplementary Fig. S1 for data and Supplementary Fig. S2 for significant interactions). We focused further validation efforts on the strongest 2 of the 6 significant interactions: SMARCB1/PSMA4 and ASPSCR1/PSMC2 (panels 1 and 2 for Fig. 2B).
To further validate these two interactions, we first retested them in yeast cells by dissecting tetrads (Fig. 3), which indeed confirmed a strong synthetic sick effect between the pairs of yeast orthologs, SNF5/PRE9 (human SMARCB1/PSMA4) and UBX4/RPT1 (human ASPSCR1/PSMC2; Fig. 3). In human cells, we repeated the same viability assay and additionally conducted knockdowns with independent targeting shRNAs for both pairs of genes. After simultaneous depletion of the targeted gene pairs, the number of cells that survived was significantly reduced in all cases (Fig. 4B, C, E, and F). Importantly, the extent of survival was significantly lower than the expected survival of double knockdowns estimated from the single shRNA effects (Fig. 4B of PSMA4/SMARCB1; Welch t test; Score = 8.11 at day 8, and 8.90 at day 10; Fig. 4E for ASPSCR1-PSMC2; Welch t test score = 14.86 at day 7 and 20.95 at day 10; P < 0.0001 in all cases; ref. 13). Expected double knockdown effects were calculated assuming a multiplicative null model, which has been widely used in the genetic interaction community (see Materials and Methods for details; ref. 20). Similar results were observed when different shRNA clones for PSMA4/SMARCB1 and ASPSCR1/PSMC2 knockdown were used (Fig. 4C and F). In addition, we confirmed the effectiveness of shRNA silencing of the targeted genes by conducting protein expression analyses using Western blots (Fig. 4A and D), which showed greatly reduced protein levels in the shRNA-infected cells.
The discovery of cancer-related synthetic lethal interactions can directly impact therapeutic potential, as the synthetic lethal interactor of a cancer related gene can be targeted selectively to kill cancer cells. To test the clinical relevance of the PSMA4 and SMARCB1 interactions, we identified an epithelial muscle rhabdosarcoma cell line (A-204) and a renal rhabdoid sarcoma cell line (G-401), each harboring SMARCB1 mutations, and used embryonic kidney HEK-293 cells expressing wild-type SMARCB1 as a control (Fig. 5). We observed that PSMA4 knockdown almost completely kills the cell lines harboring SMARCB1 mutations and that this effect was exaggerated compared with controls when following the cells to later time points (day 7, Fig. 5A–D). In addition, we showed the complete absence of SMARCB1 protein in cell lines A-204 and G-401 by Western blot analysis (Fig. 5E). A-204 carries a TC deletion of codons 181 and 182 in exon 5 whereas G-401 harbors a homozygous deletion of exons 1 to 9 (21). In both cell lines harboring SMARCB1 mutation, the decrease in growth is greater than expected by the multiplicative combination of the individual SMARCB1 mutation and PSMA4 knockdown effects, as estimated from the control cell line (Fig. 5; P < 2.5 × 10−6 for all days, all replicates, and both cell lines).
We describe an experimental pipeline where we prioritized synthetic genetic interactions from the global map of yeast interactions to test candidate synthetic sick/lethal pairs involving cancer-associated mutations in human cells. We propose this general approach, involving computational prioritization followed by experimental validation, as a complementary strategy to large-scale RNAi screens that are in progress by several other groups.
On the basis of the synthetic sick/lethal interaction, we discovered between SMARCB1 and PSMA4, we hypothesize that targeting PSMA4 in therapeutic approaches could selectively inhibit the growth of cancer cells harboring SMARCB1 mutations. Human PSMA4 is a proteasome subunit component expressed across numerous tissues. PSMA4 mRNA levels are increased in lung tumors compared with normal lung tissues, and downregulation of PSMA4 expression in lung cancer cell lines decreases proteasome activity and induces apoptosis (22). Human SMARCB1 is a core component of the BAF ATP-dependent chromatin-remodeling complex, known to play important roles in cell proliferation and differentiation, and inhibition of tumor formation. Deletions in SMARCB1 are associated with epitheliod sarcomas (23), and are a known cause of rhabdoid tumor predisposition syndrome (RTPS), a highly malignant group of neoplasms that usually occur in early childhood (24, 25). No described direct protein interaction exists between PSMA4 and SMARCB1. Although the clinical implications of this synthetic lethal interaction await further study, one potential application could be in the use of existing proteasome inhibitors such as bortezomib for the treatment of tumors harboring SMARCB1 mutations. Interestingly, interactions between other SWI/SNF subunits and the proteasome were also observed in yeast (11), suggesting the possibility that perturbations in multiple combinations of subunits across these complexes could have the same effect. Whether interactions exist in human between other genes encoding the SWI–SNF complex and the proteasome remains to be determined, but this merits further study because mutations in other subunits of SWI/SNF have been observed in many other types of cancer (26, 27).
The direct clinical implications of the ASPSCR1–PSMC2 synthetic sick interaction are less clear, but this case also merits further study. ASPSCR1 is a relatively uncharacterized gene that has been associated with alveolar soft-part sarcoma (ASPS), a rare class of tumors that typically occur in younger patients (28). Most cases of this cancer are associated with an unbalanced translocation der(17)t(X;17) (p11;q25) that results in an ASPSCR1–TFE3 fusion protein. The fusion protein appears to act as an aberrant transcription factor, inducing unregulated transcription of TFE3-regulated genes (28, 29). This fusion truncates one ASPSCR1 allele, leaving the other allele intact in most cases (28, 29). How this genetic interaction could be leveraged for therapeutic purposes awaits further investigation, but one possibility is the potential combined effect of reduced expression of ASPSCR1 in conjunction with proteasome inhibition.
Interestingly, the two strongest synthetic sick/lethal interactions we observed involved components of the proteasome, although we tested a variety of genes from multiple pathways that were produced by our approach. These data suggest the proteasome may be a rich target for synthetic lethal approaches in human cancer therapy, and indeed, successful cancer treatment involving proteasome inhibitors has been reported recently in a number of different contexts (30). The availability of several approved proteasome inhibitors may make such interactions between cancer-associated mutations and the proteasome immediately translatable to several clinical settings. These results also highlight the potential for discovering interactions within core biologic pathways with strong yeast/human homology. Importantly however, one of the limitations of our approach is its dependence on genes and proteins with such homology, and an inability to reflect many known oncogenic pathways where yeast/human homology does not exist. In addition, we note that a recent study identified both PSMA4 and PSMC2 as two of a set of 56 genes (and the only proteasomal components) for which gene knockdown inhibited the growth of cells with partial copy number loss in the same gene (31). Our independent finding of synthetic sick/lethal interactions for these same proteasomal subunits is intriguing and suggests that perturbations of these subunits have a relatively unique effect on proteasome function that may not be replicated by manipulation of its other components.
The appeal of a targeted approach for identifying synthetic sick/lethal interaction candidates is strengthened by the fact that there are currently large numbers of tumor genome sequencing efforts in progress that will produce new, potentially lengthy lists of mutations associated with various types of cancers. As we gain richer knowledge of the spectrum of mutations present in cancer, we can continue to directly screen the most promising candidate synthetic sick/lethal interactions involving these genes. The identification of specific cancer subtypes harboring specific mutations may provide therapeutic opportunities for synthetic lethal approaches that are not currently appreciated. Furthermore, in future studies, we intend to leverage data beyond sequence-similarity and literature-derived functional information to prioritize interactions for testing across species. For example, the large collections of functional genomic data in both yeast and human could allow for a more robust and unbiased assessment of the likelihood of functional conservation of genes and conserved synthetic lethal interactions between them. Our initial results highlight the feasibility of this comparative genomic approach, and suggest its potential use for rapid translation of novel sequence variants into new therapeutic targets. We believe this approach has the potential to provide a dramatic increase in the number of therapeutic targets beyond those currently available for drug development.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
The funding entities had no role in study design, data collection and analysis, decision to publish, or preparation of the article.
Conception and design: R. Deshpande, M.K. Asiedu, M. Klebig, M. Yoshida, D.A. Wigle, C.L. Myers
Development of methodology: R. Deshpande, M.K. Asiedu, M. Klebig, S. Sutor, D.A. Wigle, C.L. Myers
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M.K. Asiedu, E. Kuzmin, C. Boone, D.A. Wigle
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R. Deshpande, M.K. Asiedu, E. Kuzmin, J. Nelson, J. Piotrowski, C. Boone, D.A. Wigle, C.L. Myers
Writing, review, and/or revision of the manuscript: R. Deshpande, M. Asiedu, M. Klebig, J. Piotrowski, S.H. Shin, M. Costanzo, C. Boone, D.A. Wigle, C.L. Myers
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M.K. Asiedu, M. Klebig, S. Sutor, D.A. Wigle, C.L. Myers
Study supervision: M. Yoshida, D.A. Wigle, C.L. Myers
This work was financially supported by a grant from the Minnesota Partnership for Biotechnology and Medical Genomics program to C.L. Myers and D.A. Wigle. R. Deshpande was funded by a University of Minnesota Doctoral Dissertation Fellowship and Biomedical Informatics and Computational Biology (BICB) traineeship. C.L. Myers and R. Deshpande are also partially supported by a grant from the NIH (1R01HG005084-01A1) and a grant from the National Science Foundation (DBI 0953881). C. Boone is supported by the Canadian Institutes of Health Research (grant nos. MOP-102629, MOP-97939, and MOP-57830), the Ontario Research Fund (grant no. GL2-01-22), and the NIH (grant no. 1R01HG005853-01). S.H. Shin was supported by a BICB fellowship. C. Boone, M. Yoshida, and J. Piotrowski are supported by the RIKEN President's Discretionary Fund. C.L. Myers and C. Boone are partially supported by the Canadian Institute for Advanced Research (CIFAR) Genetic Networks Program.
Note: Supplementary data for this article are available at Cancer Research Online (http://cancerres.aacrjournals.org/).
- Received October 19, 2012.
- Revision received July 14, 2013.
- Accepted August 5, 2013.
- ©2013 American Association for Cancer Research.