Gene Essentiality Profiling Reveals Gene Networks and Synthetic Lethal Interactions with Oncogenic Ras

Differences in Gene Essentiality Reflect the Distinguishing Characteristics of AML Cell Lines

We performed CRISPR-based screens on a panel of 14 human AML cell lines selected from the Cancer Cell Line Encyclopedia (Figure 1A; Tables S1, S2, and S3; see the STAR Methods) (Barretina et al., 2012). For each gene in each line, we defined its CRISPR score (CS) as the average log₂ fold-change in the abundance of all single guide (sg)RNAs targeting the gene after 14 population doublings. Replicate screens of NB4 cells were well correlated and the CS from our screens could predict the essentiality of homologs in S. cerevisiae (Figures 1B, S1A, and S1B; see the STAR Methods). Consistent with prior work, essential genes highly overlapped between lines and were strongly enriched for roles in fundamental cellular processes (Figure 1C). Additionally, differences in gene essentiality between lines reflected known characteristics, such as cytokine dependence and developmental origin, of each of the lines (Figures S1C–S1F).

We previously demonstrated, and others recently confirmed, that Cas9-mediated cleavage of amplified genomic regions elicits a DNA damage response that causes cell death (Aguirre et al., 2016; Munoz et al., 2016; Wang et al., 2015). Thus, the low CRISPR scores of genes in amplified regions do not necessarily reflect the essentiality of the encoded products and so must be removed from datasets prior to further analyses. We devised a sliding window score (SWS) to identify contiguous stretches of the genome enriched for low scoring genes (see the STAR Methods). The SWS analysis identified several peaks across the genome of HEL cells, all of which corresponded to regions of high-level genomic amplification; the highest peak, residing in an amplicon on 9p24.1, contained JAK2, a mutationally activated driver in these cells (Figures 1D–1F) (Quentmeier et al., 2006). High SWS peaks were identified in several of the other cell lines; the genes within these peaks were also present at high copy number (Figures S1G–S1H). Thus, this simple filtering procedure, which does not rely on DNA copy number information, can be used to identify genes whose low CS are likely artifactual and thus potentially confounding to downstream analyses.

Correlated Gene Essentiality across Cell Lines Reveals Functional Gene Relationships

Genes acting in the same cellular pathway should show similar patterns of essentiality across cell lines, raising the possibility that functional gene networks can be mapped through correlation-based analysis of gene essentiality profiles (Figure 2A). To obtain biologically meaningful gene associations, comparisons must be made between genes showing significant differences in essentiality between lines. Therefore, we chose the most variably essential genes as a query set and searched for co-essential partners for each of these genes. These associations reveal known and novel gene relationships that encompass several types of functional interactions.

Many sets of highly correlated genes encoded physically interacting proteins, including heterodimers involved in transcription (LDB1 and LMO2), PI3-Kγ signaling (PIK3CG and PIK3R5), amino acid transport (SLC7A5 and SLC3A2), and components of two complexes in the mTOR pathway, mTORC2 (MAPKAP1, MLST8, and RICTOR) and GATOR1 (NPRL2, NPRL3, and DEPDC5) (Figures 2B, 2C, and S2A). This analysis also identified larger protein complexes; nearly all non-redundant components of the Fanconi anemia DNA repair machinery and the GM-CSF receptor pathway clustered tightly together (Figures 2D and S2B). Other sets of genes encoded enzymes catalyzing successive reactions in metabolic pathways (Figure 2E).

Interestingly, we identified a single case of anti-correlation between the p53 tumor suppressor gene (TP53) and its negative regulators (Figure 2F). The sgRNAs targeting TP53 provided a selective advantage (indicated by positive CS) to cells with wild-type, but not mutant, p53. In these same lines, four negative regulators of p53, TERF1, a telomere-binding factor; PPM1D, a p53-induced phosphatase; MDM2, an E3 ubiquitin ligase for p53; and MDM4, an inhibitor of p53 transactivation, were selectively required, as their loss presumably induced p53-mediated cell-cycle arrest or apoptosis (Figure S2C).

The correlation analysis also revealed several unexpected gene relationships. For example, the Furin protease cleaves and activates a diverse array of cytokines and growth factor receptors, but in our dataset the essentiality of FURIN correlated very highly with that of only one its substrates, the insulin-like growth factor receptor (IGF1R), and its adaptor IRS2 (Bassi et al., 2005) (Figure 2G). This suggests that IGF1R processing may be the only essential function of Furin in cells grown in culture.

We could also examine the opposite problem: identifying enzymes responsible for the maturation of a precursor protein. Activation of the transcription factor Nrf-1 (NFE2L1) involves retrotranslocation of Nrf-1 into the cytosol via the ER-associated degradation (ERAD) pathway, deglycosylation by PNGase (NGLY1) in the ER, and partial proteolytic digestion by an unidentified protease (Radhakrishnan et al., 2014). Our dataset showed correlated essentiality between NFE2L1, NGLY1, and the endopeptidase DDI2 (Figure 2G). These patterns of gene essentiality suggest that DDI2 may be the unknown protease that cleaves Nrf-1. Indeed, very recent work in C. elegans indicates that the homolog of DDI2 (C01G5.6) does act on the worm version Nrf-1 (Lehrbach and Ruvkun, 2016).

Our analysis also predicted associations between genes for which no functional relationship has been previously established (Figure 2H). Lastly, several genes of unknown function, such as C1orf27 and C17orf89, had correlated essentialities with genes encoding components of well-characterized pathways, suggesting that they may represent new pathway members. We performed extensive follow-up experiments to determine whether this was indeed the case

C1orf27 Interacts with UFSP2 and Is Required for deUFMylation

The essentiality of C1orf27 correlated with (1) the essentiality of several genes encoding components of the UFMylation machinery, a ubiquitin-like protein modification system that attaches UFM1 to proteins, and (2) UFM1 expression levels (Figures 3A and 3B) (Komatsu et al., 2004). Interestingly, prior work in C. elegans reported that homologs of C1orf27 and its most closely correlated partner, UFSP2, a deUFMylating enzyme, localize to the ER where they directly interact with each other (Chen et al., 2014). We confirmed these findings in human embryonic kidney (HEK)-293T cells stably expressing recombinant human C1orf27 and UFSP2 (Figures 3C and 3D).

C1orf27 is predicted to have a C-terminal transmembrane anchor, which may serve to tether it and UFSP2 to the ER surface (Figure S3A). Consistent with this model, in C1orf27-null cells, UFSP2 failed to localize to the ER and instead was dispersed throughout the cytoplasm (Figure 3E). We probed lysates of C1orf27 null cells, as well as from cells lacking UFM1, UFSP2, and the E1-like UFM1-activating enzyme, UBA5, with an antibody that recognizes free and conjugated UFM1 (Figure 3F). As expected, inactivation of UFM1 or UBA5 led to the complete loss of UFMylation activity. In contrast, loss of C1orf27 or UFSP2 led to the accumulation of UFM1-conjugated proteins, consistent with a defect in deUFMylation. These results suggest that C1orf27 is an obligate partner of UFSP2 and that this interaction is required for the proper localization and activity of UFSP2.

C17orf89 Is an Assembly Factor for Mitochondrial Complex I

C17orf89, clustered with a large group of mitochondrial genes and in HEK293T cells C17orf89 co-localized with the mitochondrial marker COX IV (Figures 3G and 3H). Mass spectrometric analysis of anti-FLAG-C17orf89 immunoprecipitates revealed an interaction with NDUFAF5, a complex I assembly factor, which we confirmed in co-immunoprecipitation experiments (Figures 3I and S3B). Strikingly, the mitochondrial gene cluster contained several complex I assembly factors, with NDUFAF5 being the top hit.

Loss of C17orf89 specifically destabilized complex I, but not other respiratory chain, complexes (Figures 3J and S3C). Consistent with a defect in OXPHOS, C17orf89-null cells consumed oxygen at a profoundly reduced rate and required the addition of pyruvate to the media to maintain optimal proliferation (Figures 3K and S3D). Importantly, expression of a sgRNA-resistant cDNA rescued these phenotypes. These findings indicate that C17orf89 encodes a component of the complex I assembly machinery and are in agreement with very recent work which characterized C17orf89 via a proteomics-based approach (Floyd et al., 2016).

Identification of Driver Oncogenes Using an Integrative Genomic Approach

Our gene essentiality catalog can also be used to determine whether a cell line carrying a specific oncogene is actually dependent on the mutated gene. A cell line might not be dependent for several reasons, including that the observed mutation does not actually activate the gene, that the oncogene was required for tumor initiation, but not maintenance, or that the cell has acquired a bypass mutation. We compiled a list of candidate oncogenes altered in each of the cytokine-independent AML cell lines based on publically available mutational and cytogenetic data and assessed the contribution of each candidate to cell fitness using our essentiality dataset. Overall, this analysis pinpointed key driver events, including common oncogenic mutations and rare translocations, in 11 of the 12 lines assessed. (As discussed above, HEL cells harbor a recurrent JAK2^V617F mutation, but it resides in an amplified region and could not be interrogated in our screens.)

Interestingly, we found that even cell lines harboring the same oncogene showed differences in dependence. Four of the cell lines carry activating mutations in FLT3, an established oncogene in AML; while FLT3 scored as an essential gene in three of these, it did not in PL-21 cells, which were also insensitive to quizartinib, a FLT3 inhibitor (Figures 4A and S4A) (Quentmeier et al., 2003).

Examination of other alterations in PL-21 revealed an uncommon mutation at codon 146 of the KRAS proto-oncogene. Residue 146 lies within the nucleotide-binding pocket of Kras and mutation of the corresponding site in Hras enhances its nucleotide exchange activity (Feig and Cooper, 1988). Consistent with Ras pathway activation in PL-21, KRAS was essential in this line. Two additional lines had mutations in KRAS and three others in NRAS. In all cases, the mutant Ras isoform was selectively essential, whereas wild-type Ras lines did not require any of the individual Ras isoforms.

Our library includes on average ten sgRNAs tiled across the body of each gene allowing for fine-scale analysis of gene fusions. EOL-1 cells harbor a recurrent FIP1L1-PDGFRA fusion gene; in these cells, only sgRNAs targeting the fused portion of PDGFRA scored, resulting in an atypical, position-dependent pattern of sgRNA depletion (Figure 4B). Translocation partners of the KMT2A (MLL) oncogene showed similar patterns in MV4;11 and THP-1 cells (Figure S4B).

We applied this “partial gene essentiality” signature to search for translocated genes in OCI-AML2, which harbors no recurrent oncogenic drivers. Remarkably, our analysis uncovered RAF1, which encodes c-Raf, a major Ras effector that regulates the MAPK signaling cascade. Consistent with a previous report, RNA sequencing revealed a chimeric transcript spanning exon 4 of MBNL1 and exon 5 of RAF1 that results in the production of a 90-kDa gene product (Figures 4C and 4D) (Klijn et al., 2015). This unique rearrangement removes the N-terminal autoinhibitory domain of c-Raf and likely leads to MAPK pathway activation.

Together, these results illustrate how functional data derived from loss-of-function screens can be integrated with genomic information to identify and validate driver oncogenes.

Two Independent Screening Approaches Reveal Common Synthetic Lethal Interactions with Oncogenic Ras

We also used our data to identify genes that are selectively essential in cell lines carrying particular driver mutations—that is, which have synthetic lethal interactions with the mutated gene. Such genes will typically not be mutated and thus cannot be reliably detected through genome sequencing. They are of significant interest because they may provide drug targets in tumors where the cancer-causing genes cannot readily be targeted, for example, in those driven by the loss of tumor suppressor genes or by oncogenes that have proven difficult to inhibit directly.

Mutations in the Ras family of GTPases (KRAS, NRAS, and, less frequently, HRAS) are commonly found in many human cancers, including AML, and are associated with poor clinical prognoses (Cox et al., 2014). Ras controls a diverse array of cellular processes through many downstream effectors. As each of these effector pathways is implicated in various aspects of Ras-driven tumorigenesis across different cellular contexts, it has been difficult to dissect the contribution of each pathway to the overall survival and proliferation of cancer cells. Furthermore, it is even less clear if Ras hyper-activation may somehow confer dependence on other, unrelated cellular pathways. Systematic screening approaches have greatly accelerated efforts to find liabilities in Ras-driven cancers (Barbie et al., 2009; Luo et al., 2009a). Here, we employed two independent screening strategies to search for synthetic lethal partners of oncogenic Ras (Figure 5A).

In our initial approach, we looked for genes that showed differential essentiality across the 12 cytokine-independent AML cell lines in our panel. Comparisons between the six Ras-dependent and six Ras-independent revealed five genes that were required only in the context of oncogenic Ras. Two genes (RCE1 and ICMT) are involved in the maturation of Ras. Two additional genes (RAF1 and SHOC2) are involved in MAPK pathway signaling. The final gene, PREX1, did not immediately fit in either category and is discussed later in its own section.

Ras is synthesized as an inactive precursor in the cytosol and converted into its mature membrane-associated form through three enzymatic steps: (1) prenylation of the CAAX box by farne-syltransferase (FTase) or geranylgeranyltransferase I (GGTase I), (2) cleavage of the terminal AAX residues by Ras converting enzyme (Rce1), and (3) methylation of the terminal cysteine residue by isoprenylcysteine carboxyl methyltransferase (Icmt). FNTB, which encodes a subunit of the FTase, was essential in all cell lines screened, suggesting that FTase acts on a universally essential protein (Figure S5A). RCE1 and ICMT, however, did show differential essentiality, suggesting that they modify a more restricted set of substrates.

Among the Ras effector genes, RAF1 and SHOC2 were the only two selectively essential in all Ras-driven lines. RAF1 encodes c-Raf, a component of the MAPK signaling cascade needed for the initiation of Kras-driven lung cancers (Blasco et al., 2011; Karreth et al., 2011). SHOC2 encodes a leucine-rich repeat-containing protein that serves as a scaffold for Ras and c-Raf (Rodriguez-Viciana et al., 2006). Similar to RAF1 and other Ras pathway members, mutations in SHOC2 have been identified in patients with Noonan-like syndromes (Cordeddu et al., 2009). Loss of SHOC2 reduced MAPK pathway activity in Ras mutant SKM-1 cells, but, importantly, not in OCI-AML2 cells in which RAF1 is constitutively active (Figure 5B).

In parallel, we devised an isogenic screening approach that did not rely on the use of genetically heterogeneous cancer cell lines. For this purpose, we screened Ba/F3 cells, a murine pro-B cell line, which we engineered to express oncogenic NRAS (CGN Ba/F3) (Figure 5A; Tables S4 and S5; see the STAR Methods). CGN Ba/F3 cells cultured in the absence of IL-3 were dependent on Ras/MAPK signaling, but, critically, this dependence was relieved by the addition of IL-3. Therefore, we could identify Ras-associated vulnerabilities by comparing gene essentiality between these two conditions. Notably, because the genetic background of the cells remains fixed in this experiment, differences in essentiality can be directly attributed to Ras dependency (Figure S5C).

Replicate screens revealed a common set of genes selectively required in the absence of IL-3. Remarkably, Shoc2, Raf1, Rce1, and Icmt all scored in the top 0.1% of all genes indicating a very high degree of overlap between the two screening approaches. Additional MAPK pathway members (Braf, Rps6ka1, and Mapk1) scored strongly as well. BRAF and MAPK1 did show a differential essentiality in the human AML lines but were dispensable in some of the mutant Ras lines presumably because they expressed redundant members of these kinase families (Figure S5A). We also identified an Nras-specific dependency. After methylation by Icmt, Nras, but not the major Kras isoform (Kras-4B), is palmitoylated. Golga7/GOLGA7, which encodes a subunit of the palmitoyltransferase, scored in CGN Ba/F3 and two of the three mutant NRAS AML lines, but not in any of the mutant KRAS or wild-type Ras lines (Swarthout et al., 2005). Other genes, such as the ubiquitin-specific peptidase Usp32/USP32, scored strongly in CGN Ba/F3 cells and a subset of the mutant Ras AML lines; the biological basis for its selective essentiality remains to be defined.

These two independent screening approaches converged on a restricted set of common dependencies required for the survival and proliferation of Ras-driven cancers. Intriguingly, the majority of these genes are involved in the maturation of Ras itself and the downstream MAPK signaling pathway (Figure 5C).

MAPK Pathway Activation Requires PREX1 in Mutant Ras AML Cells

The top scoring hit from the human AML screen was PREX1, which encodes a Dbl homology-pleckstrin homology domain-containing (DH-PH) guanine nucleotide exchange factor (GEF) for the Rac GTPases (Figure 6A) (Welch et al., 2002). Oddly, while PREX1 scored strongly in all six mutant Ras AML cell lines, it did not score highly in the CGN Ba/F3 cells. To begin to understand this difference, we designed a focused sgRNA library targeting synthetic lethal candidate and control genes and used it to pro-file: (1) the 12 cytokine-independent AML lines used in the genome-wide screens; (2) a validation set of five additional mutant Ras AML lines; and (3) 11 mutant and 14 wild-type Ras non-AML cancer cell lines derived from other hematopoietic lineages (Tables S6, S7, and S8). For all 12 of the original AML cell lines, the focused sgRNA library screen results showed the highest correlation with those from genome-wide screens conducted in the same line (Figure S6).

In all cases, the presence of an amplified or mutated allele of KRAS or NRAS correlated with dependence on KRAS and NRAS, respectively (Figure 6B). The downstream MAPK pathway members, RAF1 and SHOC2, were selectively essential in all the Ras-dependent lines as well. The requirement for PREX1, however, differed between the cancer types. Whereas PREX1 was selectively essential in both the original and validation sets of AML cell lines harboring mutant Ras, there was no difference in PREX1 essentiality between wild-type and mutant Ras lines in the other hematological cancer types.

Given the established biochemical function of PREX1, its importance in mutant Ras AML cells likely reflects a requirement for Rac pathway activity. Consistent with this possibility, Rac1/RAC1 scored as essential in the CGN Ba/F3 cells and to some degree in the human AML lines as well (Figures S5A and S5B). To test the importance of the Rac pathway, we asked whether forced activation of Rac1 could bypass the requirement for PREX1 by screening SKM-1 cells expressing a constitutively active mutant of Rac1 (Rac1^G12V) or wild-type Rac1. Consistent with our hypothesis, the dependence on PREX1 was relieved in Rac1^G12V-expressing cells (Figures 6C and 6D).

Previous studies demonstrate that PREX1 can influence MAPK signaling, suggesting that like the other screen hits, it may also act on the MAPK pathway (Ebi et al., 2013). To examine this possibility, we screened the focused library in SKM-1 cells stably expressing a constitutively active mutant of Mek1 (Mek1^DD) or wild-type Mek1. As expected, Mek1 hyper-activation relieved the dependence on the upstream MAPK pathway components, KRAS, RAF1, and SHOC2. Critically, Mek1^DD expression also bypassed the requirement for PREX1, placing it too upstream of Mek1 (Figures 6E and 6F). The Rac GTPases can induce MAPK signaling by stimulating the p21-activated kinases (PAKs), which, in turn, phosphorylate and activate c-Raf (King et al., 1998). Consistent with this model, Rac1^G12V-expressing cells had hyperactive PAK and MAPK signaling and knockdown of PREX1 inhibited these pathways in wild-type SKM-1 cells (Figures 6G–6I). Together, these data establish PREX1 as a key input for MAPK pathway activation in Ras-driven AML cells.

Lack of Paralog Expression Explains AML-Specific Dependence on PREX1

As PREX1 is highly expressed in normal myeloid cells, we reasoned that it functions as the major activator of Rac signaling in AML cells, but that perhaps other GEFs promote Rac activity in other cancers. Consistent with this notion, all of the mutant Ras AML lines examined expressed PREX1, but only three of the nine non-AML lines did (Figure 7A). Strikingly, TIAM1, another DH-PH Rac-GEF, had the opposite expression pattern—it was absent in all the AML lines but robustly expressed in all but one of the non-AML lines. Notably, the one line not expressing TIAM1, NU-DHL-1, expressed high levels of PREX1 and was the only PREX1-dependent non-AML line (Figure 6B). Though PREX1 and TIAM1 share little sequence homology (6% amino acid identity), we posited that they might nonetheless be functionally interchangeable. To test this idea, we screened the focused library in THP-1 cells stably expressing TIAM1. As compared to the parental line, THP1-TIAM1 cells have a reduced dependence on PREX1, which showed the greatest change in CS of all 132 genes screened (Figures 7B and 7C). Additionally, TIAM1 expression rescued the decrease in PAK signaling caused by PREX1 loss (Figure 7D). Thus, we conclude that Ras-driven AML cells specifically require PREX1 because it is the only active Rac-GEF expressed in this cancer subtype.

While PREX1 may not serve as an ideal target for pharmacological inhibition, our findings raise the possibility that AML and non-AML cancers driven by oncogenic Ras may be sensitive to inhibition of the group I PAKs (PAK1-3). Using FRAX-597, a small-molecule inhibitor of multiple kinases including the group I PAKs, we tested this hypothesis in two isogenic cell pairs with differential requirements for Ras signaling: SKM-1 cells expressing either Mek1^DD or the control protein Rap2A, as well as CGN Ba/F3 cells cultured in the presence or absence of IL-3. In both cases, the cells dependent on Ras signaling were more sensitive to PAK inhibition than the isogenic control cells even though FRAX-597 inhibits many other kinases besides the PAKs (Figure 7E) (Chow et al., 2012). Collectively, these results suggest a model in which all Ras-driven cancers require PAK activity in order to fully activate MAPK signaling, with each cancer subtype activating the PAKs via distinct mechanisms (Figure 7F).

An Integrative Genomic Approach Reveals Oncogene Dependency

Cancer genome sequencing efforts have provided an increasingly complete catalog of the genes altered during tumor development (Lawrence et al., 2014). Functional studies enable a direct assessment of the contribution of each of these genes to cancer cell fitness (Boehm and Hahn, 2011; Garraway and Lander, 2013). Together, these complementary approaches should accelerate the identification of novel oncogenes and potential therapeutic targets. Some cancers are driven by rare events that are difficult to distinguish from random mutations and thus require functional analysis to assess the significance of an alteration (Berger et al., 2016; Starita et al., 2015; Tsang et al., 2016). For instance, the tiled design of our libraries enabled us to identify the essentiality of translocation events including a rare inversion involving the RAF1 kinase in OCI-AML2 cells.

However, mutational information alone cannot discriminate between oncogenes required for the continued growth of cancer cells from those solely involved in tumor initiation. Even for cells harboring activating mutations in the same oncogene, we found differences in essentiality (only three of four FLT3 mutant lines required FLT3). Thus, to more accurately guide cancer treatment, functional testing of patient tumor cells, should be considered in combination with sequence analysis.

Functional Gene Network Mapping Using Correlated Gene Essentiality Analysis

The natural variability in the genetic and epigenetic makeup across human cancer cell lines leads to differences in gene essentiality and so provides a convenient means for defining functional gene networks. Even between lines of a single subtype, we found many genes with variable essentiality. Reasoning that genes in the same biological pathway should show similar patterns of essentiality, we used the CRISPR scores to cluster genes into groups with correlated essentiality. Interestingly, the scores of many gene pairs correlated linearly, with the different cell lines showing graded, rather than binary levels of requirements for the genes. Our analysis uncovered several classes of functional relationships including gene sets encoding protein complexes, metabolic pathways, and enzyme-substrate pairs and enabled us to determine the molecular functions of uncharacterized genes.

Analysis of other cancer types or across cancer types may reveal additional interactions and surveying across media conditions or in the presence of chemical compounds may also yield valuable insights. Moreover, we anticipate that more sophisticated analysis of our dataset using approaches that can detect multi-way interactions will allow for continued discovery.

With the exception of the genes involved in p53 signaling, the basis of the variable essentiality of all other gene clusters remains unclear. Such an understanding will be required in order to exploit these pathways for cancer therapy. Similar to efforts to predict cancer drug response, integrative approaches may help uncover biomarkers for gene essentiality.

Screens in Established Human AML and Engineered Mouse Cell Lines Uncover a Common Set of Ras Synthetic Lethal Interactions

We focused on a special case of co-essentiality: synthetic lethality with oncogenic Ras. In large part, our study suggests that the development of therapies that selectively impact Ras-dependent cancer cells will require re-focusing efforts on targeting select components of the Ras pathway itself.

Ras, like many small GTPases, undergoes a series of post-translational modifications to facilitate interaction with the inner leaflet of the plasma membrane. Efforts to block this process have been primarily directed toward inhibition of the initial step of the pathway catalyzed by FTase (Cox et al., 2014). However, FTase inhibitors have been ineffective in the clinic as Kras and Nras can be geranylgeranylated, an alternative prenylation pathway (Whyte et al., 1997). Additionally, our results here and from prior screens conducted in other cancer subtypes indicate that FTase is required in all cells. In contrast to FTase, the enzymes catalyzing the latter two steps of the Ras processing pathway, Rce1 and Icmt, do display synthetic lethality with oncogenic Ras and may thus serve as therapeutic targets.

Our results provide further support for the central role of MAPK signaling in Ras-driven cancers and suggest c-Raf as a therapeutic target. The unique requirement for c-Raf, but not other Raf kinases, is consistent with only c-Raf being required in lung cancer models driven by oncogenic Ras (Blasco et al., 2011; Karreth et al., 2011).

A mechanistic insight from our study is the critical role of the Rac/PAK signaling axis in promoting MAPK activity in mutant Ras cancers. Even though the Rac GTPases activate many downstream pathways, we found that forced expression of constitutively active Mek1 can bypass the requirement for PREX1. The selective essentiality of PREX1 in Ras-driven AML, but not in the other cancer types tested, likely reflects the critical role of PREX1 in normal myeloid cells. In neutrophils, where PREX1 is highly expressed, host- and pathogen-derived chemotactic factors trigger activation of the PI3-Kγ and GPCR pathways (Welch et al., 2002). This results in the generation of PIP₃ and free Gβγ subunits which recruit PREX1 and stimulate Rac-GEF activity. In AML cells, Gβγ and PIP₃ may be similarly required to activate PREX1. We note that genes encoding two Gβ subunits (GNB1/2), a Gγ subunit (GNG5), a Gβγ-modulator (PDCL), and the catalytic and regulatory subunits of PI3-Kγ (PIK3CG/PIK3R5) all showed partial Ras co-dependency (Figure S5A). We hypothesize that Ras-driven cancers originating from other cell types rely on other Rac-GEFs, such as TIAM1 and VAV1, to activate PAK signaling.

Design of Synthetic Lethal Screens and sgRNA Libraries

The combination of screening approaches employed here provides a guide for the design of robust screens for synthetic lethal interactions. As illustrated by the case PREX1 in Ras-driven AML, genetic interactions with oncogenes may occur in a cell context-dependent manner. Thus, it may be sensible to screen lines of a particular cell type or to include enough cell lines representing each cancer type. Additionally, screens across isogenic cell lines should be employed to eliminate factors that may confound analyses across genetically heterogeneous cancer cell lines. Here, we screened Ba/F3 cells expressing oncogenic NRAS in the presence and absence of IL-3. This perturbation altered oncogene dependence, but not proliferation rate (Figure S5C).

Microarray-based oligonucleotide synthesis enables the rapid generation of focused sgRNA libraries for follow-up studies. As such experiments require vastly fewer numbers of cells, many additional cell lines can be tested. By using expanded cell line panels representing more cancer types, the generality of the interactions can be assessed and with engineered panels of lines, epistatic relationships between hit genes defined. Moreover, it may be possible to conduct screens using murine cancer models and identify genes that play critical roles in vivo.

General Comments on Synthetic Lethality in Cancer

Synthetic lethal interactions in cancer cells can, in principle, occur between several classes of genes. The prototypical example is the inactivation of a so-called ‘caretaker’ gene involved in the maintenance of genomic stability that leads to dependence on a parallel maintenance pathway (Ashworth et al., 2011; Kaelin, 2005). Such interactions may arise between genes involved in distinct but functionally overlapping processes, as seen with the BRCA and PARP DNA repair pathways, or between highly related and perhaps even interchangeable paralogs, such as ARID1A and ARID1B (Farmer et al., 2005). However, this paradigm may not apply to Ras and other genes involved in signal transduction. In contrast to loss-of-function mutations in caretaker genes, oncogenic mutations in growth factor signaling pathways result in hyperactive signaling and, in most cases, render cells dependent on the altered pathway (Luo et al., 2009b). Furthermore, as these mutations act in a dominant fashion, they are typically found in the heterozygous state, leaving the wild-type allele intact.

Genes and pathways that protect cancer cells from the diverse stresses associated with the malignant state represent a second class of potential vulnerabilities. In comparison to their normal counterparts, cancer cells rely to a much greater extent on such cytoprotective pathways as they experience elevated levels of mitotic, oxidative, proteotoxic, metabolic, and DNA damage-related stress (Luo et al., 2009b). While many of these stresses can be experimentally induced by the expression of specific oncogenes, they are almost universally found in established tumors regardless of genotype (Courtois-Cox et al., 2008). Thus, it is unclear whether these liabilities can be linked to any particular oncogene per se or if they arise as a secondary consequence of the increased genomic instability and mitotic index characteristic of all cancer cells. Indeed, chaperones, such as Hsp90, act as “genetic hubs” and show epistasis with hundreds of client proteins, including several oncogenic kinases (Whitesell and Lindquist, 2005). More comprehensive studies that compare various genetically defined malignant and pre-malignant cells are needed to pinpoint the specific features of the oncogenic state that sensitize cells to inhibition of individual stress response pathways. Importantly, as full inhibition of many of these pathways is likely to be lethal, gene knockdown approaches, such as CRISPRi, may be better suited to interrogate them (Gilbert et al., 2014; Horlbeck et al., 2016).

The only consistent differences in gene essentiality between the mutant and wild-type Ras cells in our study were in genes closely connected to Ras itself (Ras post-translational processing and MAPK signaling). Extensive experimental evidence in Ras-driven cell lines and in murine cancer models supports the importance of these pathways. Our data are in general agreement with findings from our correlated essentiality analysis—as with other pathways and complexes, cells that require Ras also require other genes that act in concert with Ras to promote survival and proliferation. We anticipate that screens for synthetic lethal partners of other driver oncogenes will uncover similar networks of ancillary genes that may serve as attractive targets for therapy. More broadly, through the systematic application of CRISPR-based screens, it should be possible to comprehensively identify the acquired vulnerabilities of human cancers.

KEY RESOURCES TABLE

CONTACT FOR REAGENT AND RESOURCE SHARING

Requests for further information and resources may be directed to Lead Contact David M. Sabatini ([email protected]).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

METHOD DETAILS

QUANTIFICATION AND STATISTICAL ANALYSIS

DATA AND SOFTWARE AVAILABILITY