Combined Proteomic and Genetic Interaction Mapping Reveals New RAS Effector Pathways and Susceptibilities

RAS proteins are mutated in 20% of cancers, yet few targeted therapies are available in the clinic (1). Although RAS oncogenes were discovered more than 35 years ago (2), the consequences of RAS activation remain only partly understood. KRAS, HRAS, and NRAS are small GTPases that cycle between GTP- and GDP-bound states with the aid of exchange factor (GEF) and activating (GAP) proteins. Although KRAS, HRAS, and NRAS are all known proto-oncogenes, KRAS is the most frequently mutated, notably in pancreatic, colon, and lung adenocarcinomas (1). Oncogenic mutants of RAS proteins have reduced GAP-mediated GTP hydrolysis (3) and strongly favor the GTP-bound active state in which they bind tumorigenic effectors. Some RAS effectors, such as BRAF and PI3K, are themselves proto-oncoproteins (3) and therefore are attractive therapeutic targets. RAS proteins have many other potentially important effectors and regulators with poorly understood function (3). Many of these factors are represented by multiple paralogs, which potentially compensate for one another in screening experiments, concealing their collective importance (4). The extent to which paralogous RAS effectors or regulators, such as the three RAF kinases, are functionally distinct remains incompletely understood. However, therapeutic inhibitors are often not paralog-selective and can modulate multiple paralogous targets through related binding sites. The ensemble of RAS protein effectors and regulators thus likely contains both new therapeutic targets and unexplored paralog-selective targets for unmet clinical needs (3).

The development of targeted therapies relies on the concept that genetic alterations in cancer cells cause pathway “rewiring,” making less-essential mutant and wild-type genes more essential in neoplastic cells (5, 6). Targeting such cancer-specific dependencies is a major strategy guiding efforts to treat RAS-driven cancers (3) and has been successful in other cancer types (7, 8). Although many cancer-specific susceptibilities are known, little systematic investigation of pathway rewiring has been performed. Genome-wide knockout screens (9–11) have identified new cancer pathway components, and patient-sequencing initiatives (1) have identified candidate driver mutations. However, these searches for single-gene targets address neither pathway rewiring nor the possibility of novel combination targets selective for the rewired cancer cell.

Here, we present the integrated results of two direct and orthogonal analyses of functional interactions in the RAS pathway (Fig. 1A). First, we mapped detailed protein–protein interactions (PPI) using highly optimized affinity purification/mass spectrometry (AP/MS; ref. 12). This approach enables unbiased detection of “prey” proteins copurifying with chosen “bait” proteins. Beginning with KRAS itself, using GTP- and GDP-locked (sometimes called apo) mutant forms as baits, we iteratively expanded the PPI network using preys discovered in earlier experiments as baits for later experiments, yielding a dense PPI network of RAS pathway interactions. Second, focusing on the genes identified in this PPI network and on known RAS pathway genes, we mapped genetic interactions (GI) using CRISPR dual knockout (CDKO) technology (13). CDKO screens identify synergistic or buffering GIs through interference of knockout phenotypes of two genes. GIs can link pairs of redundant or interacting genes. GI screens can therefore reveal essential activities carried out by pairs of related genes that would be maintained if only one gene were lost, and so suggest new possible combination therapy strategies (13, 14). From the identified pairwise links, we constructed networks that produce testable hypotheses regarding the biology of RAS-driven cancers. The complementary approaches of GI and PPI mapping clearly profile both known and new effectors and regulators of the major RAS oncoproteins. They also show distinctive patterns of PPIs and GIs between paralogous and nonparalogous proteins, identifying surprising new connections among RAS and other signaling pathways.

The combined GI and PPI networks connect genes important for KRAS-driven lung adenocarcinoma and together provide a physical and functional “wiring diagram” linking drivers of RAS cancer biology. This combination mapping approach provides a wealth of testable candidates for new combination therapies and demonstrates a generalizable method for mapping cell biological pathways altered by disease.

Interaction maps were assembled using an iterative process. First, we performed AP/MS of KRAS using mutations that confine it to a GTP-locked (3) or GDP-locked (or apo) state (ref. 15; Methods; Supplementary Table S1). Using high-efficiency tandem affinity purification (12, 16) in HEK293 cells, we isolated each bait with copurifying proteins, which were identified by mass spectrometry and quantified as described previously (17). We expanded the network by repeating the procedure using 5 other RAS pathway proteins (RALA, RALB, RALGDS, RAP1B, and RAP2B) as baits, creating a network of 451 proteins including both known and unknown RAS pathway members and incorporating public PPI data (Supplementary Table S2; Supplementary Fig. S1A and S1B).

We selected the most confidently identified proteins from this network and cross-referenced them with public datasets to select 119 genes for GI analysis by CDKO in two KRAS-mutant lung adenocarcinoma cell lines (A549 and H23). This analysis of 119 genes (7,021 pairs) revealed hundreds of mostly novel GIs. We then repeated and expanded the PPI mapping experiments in A549 cells with 14 baits, including most members of the RAL and RAP GTPase families (KRAS, NRAS, HRAS, RALA/B, RAP1A/B, RAP2A/B, RALGDS, RGL1, RGL2, RADIL, and RASGRF2). This produced a network of 930 proteins, to which we added relevant published interaction data (Supplementary Tables S2 and S3; Supplementary Fig. S2A). Informed by the results of both PPI maps and the initial GI maps, we created two smaller 20 × 20- and 25 × 25-gene CDKO libraries to screen a set of 9 lung adenocarcinoma lines (see Methods). These cell lines were found to have a range of dependencies on KRAS, allowing us to stratify GIs into classes positively or negatively associated with KRAS activity. Here we describe new insights into KRAS-driven biology derived from this multiomic strategy, including the identification of novel potential combinatorial susceptibilities (Fig. 1A; Supplementary Fig. S2B).

Oncogenic forms of RAS proteins cause transformation in specific tumors through enhanced binding to downstream effectors (3). Different RAS isoforms can selectively bind distinct effectors, yet no systematic comparison of RAS PPIs with their regulators and effectors has been performed. From AP/MS experiments in HEK293 cells, we identified efficient interactors of KRAS, both well known (RAF1, BRAF, ARAF) and less studied (RALGEFs RALGDS, RGL1, RGL2, RGL3, RAP1GDS1, RADIL; Supplementary Fig. S1B).

As a comparison, we conducted copurifications with GTP- and GDP-locked mutants of HRAS, KRAS, and NRAS stably expressed in A549 cells, a KRAS-transformed lung adenocarcinoma line. RAS effector proteins with known preference for the GTP-locked RAS state bound specifically to that mutant form. We observed striking differences in copurifying proteins between each RAS protein, and between their mutant forms, as apparent by silver stain (Fig. 1B) and by mass spectrometry (Fig. 1C; Supplementary Fig. S2A and S2C). Surprisingly, there were notable differences among the specific effectors copurified by activated HRAS, KRAS, and NRAS (Fig. 1C). The HRAS interactome differed most strongly from the other two. Most strikingly, RAF kinases (ARAF, BRAF, RAF1) were highly enriched by either GTP-locked KRAS or NRAS baits but not HRAS. We analyzed small-scale coimmunoprecipitation experiments in lung cancer cells by immunoblot (Fig. 1D) and confirmed that RAF1 bound GTP-locked HRAS much less efficiently than GTP-locked KRAS or NRAS, even with similar expression of RAS bait proteins. The same was true for three other RAS effectors containing canonical RAS-association (RA) domains: RIN1, RIN2, and RADIL. In contrast, three members of the RALGEF family, RGL1, RGL2, and RGL3, that also contain RA domains copurified efficiently with NRAS but poorly with HRAS or KRAS (Fig. 1C).

We considered whether observed binding specificities reflected protein-intrinsic properties that could be recapitulated in cell-free assays. Using purified GTP- and GDP-locked HRAS, KRAS, and NRAS proteins, we reconstituted binding to in vitro–translated forms of the known and new RAS-interacting proteins identified (Fig. 1E; Supplementary Fig. S2D). Although all RAS-effector interactions were strongly GTP-dependent, we found negligible differences in binding efficiency between the three GTP-locked RAS proteins and these RAS effector proteins in vitro. This and the experiments below indicate that other cellular processes (such as regulation of localization, protein signaling, and modifications or effector availability) are comparably important to RAS nucleotide state in determining the efficiency of RAS-effector engagement in tumor cells.

A number of known RAS-binding proteins were not identified by our AP/MS experiments. PPIs between GTPases and their coordinate GEFs have been historically difficult to isolate by AP/MS (see ref. 18), so the absence of signal from GEFs such as SOS in these purifications was expected. The absence of other RAS-binding proteins, such as PIK(3)CA or TIAM1, requires further study. These PPIs may occur only in specific cell lines or under stimuli not supplied by our culture method, or are simply inefficient without overexpression of preys.

It is worth noting that not all interactions efficiently observed by AP/MS are necessarily direct, because our large-scale purifications often identify complexes and secondary interactions (16). Combining our AP/MS data with data from public PPI databases including BioGRID (19), we constructed a PPI network of RAS interactors (Supplementary Fig. S2C). A selective representation of this network is shown in Fig. 1F.

Interactions between RAS family proteins and effectors are mediated by the effector binding domain of RAS proteins and RA domains in target proteins (20). In addition to the effectors of KRAS mentioned above, we found two proteins that copurified efficiently with both GTP-locked KRAS and NRAS that contain canonical RA domains: RIN1 and RADIL (Fig. 2A).

RIN1 has been shown to bind HRAS in vitro (21), and functions as a nucleotide exchange factor for RAB5 GTPases, which are central organizers of endocytosis (22). Recent work shows KRAS-transformed cancers use macropinocytosis, a specific form of endocytosis, to scavenge nutrients from their environment (23, 24). The identification of RAS–RIN PPIs suggested that KRAS might directly regulate macropinocytosis through RIN1. We assayed macropinocytosis as described previously (25). Serum-starved H23 cells take up 70-kDa fluorescent dextran by macropinocytosis, as visualized by fluorescence microscopy (Fig. 2B). Three clonal CRISPR RIN1 knockout lines (Supplementary Fig. S3A) showed decreased macropinocytosis relative to cells transduced with a negative control single guide RNA (sgRNA). Upon pretreatment with ARS853, a KRAS^G12C-specific inhibitor (26), macropinocytosis was substantially suppressed in the control cells but not in the RIN1 knockout cells. These cells appear insensitive to further inhibition of macropinocytosis by ARS853, consistent with the model that RAS-driven macropinocytosis is substantially coupled to RIN1 activity (Fig. 2C).

RADIL, which has not previously been shown to bind KRAS, also strongly copurified specifically with GTP-locked KRAS in both A549 and HEK293 lines. The N-terminal RA domain in RADIL resembles that of other RAS effectors (Supplementary Fig. S3B) and has been crystallized (27). Arginine 99 (R99) in RADIL occupies a position in the RA domain comparable to RAF1 R89 (Fig. 2D), a site required for RAS interaction (20). We find that RADIL specifically and efficiently binds in vitro to GTP-locked HRAS, KRAS, and NRAS, that the RA domain is sufficient for this PPI, and that the R99 mutant fails to bind KRAS-GTP (Fig. 2E; Supplementary Fig. S3C). Previous studies of RADIL focus on its role in RAP1A-driven cell adhesion (28, 29), and our results suggest a direct link between KRAS and RADIL, at least in lung adenocarcinoma. To validate that RADIL also affects adhesion and motility in these cells, we conducted a scratch-wound assay (30) and found that RADIL knockdown increased the migration of these cells, in agreement with other studies (28). However, RADIL knockdown did not increase the migration rate of cells treated with ARS853, consistent with a mechanism where RADIL-regulated cell migration is enhanced by activated KRAS (Fig. 2F–H).

We identified strong GTP-dependent PPIs between KRAS and the RALGEFs RALGDS, RGL1, and RGL2 in HEK293 cells. Although RALGEFs and their RAL GTPase substrates are linked to cell proliferation and carcinogenesis (31–33), they remain incompletely understood. RALGEFs also bind RAP GTPases, and different family members have distinct but unknown RAL-independent activities (33, 34). Reasoning that an unbiased survey of RALGEF PPIs would provide important clues to their function, we conducted AP/MS experiments with RALGDS, RGL1, and RGL2 as bait proteins. Each RALGEF copurified a wide variety of upstream GTPases, including RAS and RAP family members. Upregulation of many of these proteins is associated with unfavorable lung adenocarcinoma prognosis, and the genes that encode them include conditional dependencies of KRAS-dependent and KRAS-independent cancers (Fig. 3A; see Supplementary Methods). Here we identified other interactors unique to each RALGEF, implicating them in signaling pathways including the Hippo (NF2) and NFκB pathways (NKIRAS2).

To better understand RALGEF activity–specific interactions and effects, we conducted AP/MS experiments with the downstream RAL (RALA, RALB) and linked RAP (RAP1A, RAP1B, RAP2A, RAP2B) GTPases in their GTP- and GDP-locked forms. Like RAS proteins, RAP proteins engage effectors, including the RALGDS family, through effector RA domains. A comparison of common PPI partners of RAS and RAP family members in A549 cells is shown in Fig. 3B, indicating that RADIL, RALGEF, and RIN family effectors are preferentially engaged by NRAS over HRAS or KRAS, and RAP2A and RAP2B over RAP1 isoforms. Clustering analysis of RAS, RAP, and RAL GTPases showed that RAS and RAP GTPases bind surprisingly similar effectors including RALGEFs, RIN1/2, and RADIL (marked ο; Fig. 3C). In contrast, the RALA and RALB GTPases show a distinct set of interactors including all components of the exocyst complex (marked x).A subset of AP/MS experiments were conducted in both A549 and HEK293 cells. An analogous clustering analysis (Fig. 3D) shows that, generally, each GTPase interacted with a similar set of proteins in both lines. Most strikingly, clustering analysis does not distinguish RAS from RAP proteins, indicating an underappreciated intersection between RAS and RAP interactions and signaling. Both cooperative and antagonistic relationships between oncogenic RAS GTPase and the families of RAP GTPases are possible. Our genetic interactions, presented below, shed some light on this phenomenon.

These PPI maps show the benefit of systematic mapping of paralogs and similar families. The many specific PPIs of RAP, RAL, and RALGEF family proteins suggest the existence of an elaborate network of dynamic interactions controlling signaling and trafficking at plasma and organellar membranes. Curated subsets of these PPIs are diagrammed in Supplementary Figs. S1B and S2A. Complete PPI data can be found in Supplementary Table S2, and online via NDEX (https://doi.org/10.18119/N90S3V). Importantly, these PPIs predict and rationalize many genetic interactions, as we validate below.

Having established that this high-confidence PPI network contains many uncharacterized but potentially important interactors, we next sought to map GIs between them to better understand the RAS pathway and to identify combination target pairs. To map GIs, we used CDKO screening (13), in which genes or pairs of genes are systematically knocked out by a CRISPR library transduction approach. Because the number of knockout combinations scales geometrically with the number of targets, CDKO screens are limited to a smaller target gene set than single-target screens.

Beginning with the genes encoding proteins in our HEK293 PPI network, we sought to identify the ones most likely to show genetic interactions. We first constructed a single-sgRNA library targeting 223 genes based on their protein identification in AP/MS experiments (Supplementary Fig. S1). The library contained 10 sgRNAs per gene and was transduced into A549-Cas9 cells. Cells were grown for 21 days and DNA isolated and sequenced to identify depleted or enriched sgRNAs (Supplementary Fig. S4A and S4B; Supplementary Table S4). This screen identified some essential genes, which would be less likely to show GIs and thus were removed from further consideration. Similarly, we removed genes previously shown to be essential in a majority of cancer cell lines (9, 10, 35). Finally, we prioritized the remaining genes by the strength of their interactions in our AP/MS data, and whether that interaction involved KRAS. Of these, we selected 109 genes. We then added genes encoding MEK and ERK kinases, and others reflecting genetic lesions in A549 cells, for a total of 119 genes (Fig. 4A; Supplementary Table S5).

Next, we constructed a CDKO library including 3 sgRNAs per gene (13), enabling systematic deletion of 7,021 gene pairs. We included 60 “safe” sgRNAs targeting nonfunctional genomic loci, previously validated to have minimal effects on cell growth (13), to control for intrinsic effects of Cas9 activity and to allow measurement of single-gene deletion phenotypes. The library was transduced into two Cas9-expressing KRAS-mutant lung adenocarcinoma lines, A549 and H23. The abundance of each guide pair was measured before and after the 21-day growth period (Fig. 4B). We quantified GIs by comparing the abundance of targeting to safe guide pairs and double-targeting guide pairs (Supplementary Fig. S4C and S4D; Supplementary Table S5, Methods), as described previously (13).

We identified 548 GIs with FDR < 0.1 in A549 cells, and 447 in H23 cells, of which 59 overlapped (Fig. 4C; Methods). The consistently interacting pairs included paralogs such as RAP1A with RAP2A or genes in the same pathway, such as MAPK1 with RAF1. Many of the strongest synergistic GIs in either line were found between paralogous MAPK pathway components (e.g., MAPK1 with MAPK3, MAP2K2 with MAP2K4). The observed strong dependencies on this known central pathway confirm that the screening captured relevant GIs. By combining GI and PPI data, we found that roughly one in six GIs paralleled direct (first-degree) PPIs, whereas many more GI partner pairs are connected by second- or third-degree PPIs (Supplementary Fig. S5A).

Representative GI/PPI subnetworks, showing GIs common to both cell lines tested, are diagrammed in Supplementary Fig. S5B. These combined maps provide testable hypotheses about mechanisms underlying observed GIs. For example, a synergistic GI links the PI(3)K subunit–encoding gene PIK3R1 and RAF1. Both are constituents of RAS effector pathways, and this GI is supported by pharmacologic combinations (36).

GI strength is quantified by GI_T score, which describes the degree of synergy of a given GI, and accounts for assay noise (13). As one strong example, KRAS-mutant H23 cells infected with the safe/safe, RAP1GDS1/safe, and safe/RHOA sgRNA cassettes doubled once every 51, 53, and 64 hours, respectively. Those infected with a double-targeting guide doubled only every 120 hours (not shown). These changes in doubling time correspond to a GI_T score of -4.2. Strong GI_T scores may therefore indicate dramatic changes in fitness. The efficacy of this combination is confirmed in multiple assay systems, as we show below.

We included the MAPK pathway genes MAP2K1, MAP2K2, MAPK1, and MAPK3 in the CDKO library as positive controls because of known synthetic lethal GIs among these genes (37). As predicted, we observed strong synthetic lethal GIs between these paralogous pairs. Combined knockout of MAPK1 and RAF1 was also strongly synergistic, but, surprisingly, knockout of RAF1 and MAPK3 was much less so. In this screen and screens below, the specific link from MAPK1 to RAF1 is consistently seen, suggesting that MAPK1 (also called ERK2) has a specific link to RAF1 and therefore different regulatory properties from MAPK3 (or ERK1). A549 cells showed strong dependency on RAF1 in combination with other pathway members, but limited dependency on ARAF or BRAF. In H23 cells, however, GIs could be observed between all three RAF genes and other MAPK pathway components (Fig. 4D). The seven MAPK genes showed numerous but highly varied GIs with other genes in these two cell lines (Supplementary Fig. S5C). This suggests that the MAPK pathway signals to other RAS effector pathways, but does so differently in different tumor cell lines. Therefore, paralog-selective inhibitors of ERK or RAF kinases might have more specific antitumor activity in certain patients.

Although the presence of synergistic GIs common to both lines showed that CDKO screening could uncover consistent combination susceptibilities, the preponderance of cell line–specific GIs suggested that screening additional cell lines would reveal additional tumor-selective patterns of susceptibility, especially with regard to KRAS dependence.

To identify GIs specific to KRAS-driven lung adenocarcinoma, we first assayed the KRAS dependence of 9 lung adenocarcinoma cell lines by measuring their growth with and without CRISPR–Cas9-mediated KRAS knockout. We found that six lines were highly KRAS-dependent, and three were KRAS-independent (Fig. 5A; Supplementary Fig. S6A). Of the 6 KRAS-dependent lines, 5 had codon 12 mutations in KRAS, and the one that did not (H1975) had activating mutations in EGFR. None of the KRAS-independent lines had RAS mutations (Supplementary Fig. S6B and S6C).

Therapeutically useful combination target pairs should show synthetic lethal synergistic GIs in KRAS-dependent, but not KRAS-independent, lines. To identify such pairs, we constructed two smaller CDKO libraries and conducted focused screens across the cell line panel. The first, “GI-Directed” library targeted 20 genes with strong interactions identified in the first CDKO screen described above (Fig. 5B; Supplementary Tables S6 and S7). The second, “paralog” library targeted paralogous families of RAS regulator and effector proteins, focusing on those found in the A549 PPI network. The screens were conducted with both libraries in all 9 lines following the protocol described above. In both screens, we found multiple examples of gene pairs whose synergy was greater in the KRAS-dependent than KRAS-independent lines. Among the strongest hits in the “GI-Directed” library were RAP1GDS1 and RHOA (Fig. 5C, left), which had shown a strong synthetic lethal GI in the two earlier larger screens. RAP1GDS1 is a GEF for RHOA and RHOC, and also regulates RAS and RAP GTPases (38, 39). We will discuss this interaction extensively in the next section.

One of the strongest synthetic lethal pairs specific to KRAS-dependent lines in the “paralog” library was RALA–RALB (Fig. 5C, middle). This pair previously showed synthetic lethality in a genetically engineered mouse model (37). As shown in Fig. 3A, the immediate upstream regulators of RAL GTPases, the RALGEFs, link a network of proteins important for signal transduction in a variety of important pathways, as well as a variety of KRAS-related dependencies and prognostic markers.

In contrast to the RAP1GDS1–RHOA and RALA–RALB pairs, KRAS–NRAS double knockout strongly synergized specifically in KRAS-independent lines (Fig. 5C, right). This observation suggests that whereas KRAS-dependent cells are “addicted” specifically to KRAS, KRAS-independent cells may rely jointly on normal KRAS and NRAS proteins to accomplish tumor-driving signaling functions.

Many gene pairs had a greater tendency toward genetic interaction in either KRAS-dependent or KRAS-independent lines. Some GIs common to both A549 and H23 cells in the original screen, such as MAPK1–MAPK3 and MAPK1–RAF1, were recapitulated in the sublibrary screens (Fig. 5D).

GIs specific to either the KRAS-dependent or KRAS-independent cohort, and the PPI networks linking them, are plotted in Fig. 5E. Although few synthetic lethal pairs involving exactly one MAPK pathway component were consistently more severe across the KRAS-dependent lines, there were significantly more such GIs in the KRAS-dependent than KRAS-independent cohorts (Supplementary Fig. S6D). Pairs of paralogous genes were strongly enriched for GIs in both the KRAS-dependent and KRAS-independent cohorts (Supplementary Fig. S6E). Certain GI pairs, involving many different pathway members, were associated with non-KRAS mutations (Supplementary Fig. S6F).

GI targets and interrogated gene pairs are diagrammed in Fig. 5F, with nodes and edges scaled by the frequency of observed interactions across all cell lines. Interestingly, we found that RIN1 and RIN2, RADIL, and RALGEF genes, which encode central members of our PPI network, were among the most frequent interacting partners across all nine cell lines, regardless of their KRAS dependence, pointing to their central importance in RAS and RAP signaling. All GIs with FDR < 0.1 observed in all 9 cell lines and two libraries are diagrammed in Supplementary Fig. S7.

One of the strongest interactions we observed in KRAS-dependent cells was a synthetic lethal GI between RAP1GDS1 and RHOA (Fig. 5C, left). RHOA encodes a small GTPase critical for actin filament assembly and membrane organization (40). RAP1GDS1 (SmgGDS) has been described as a GEF for RHOA and RHOC (38, 39). RAP1GDS1 differs structurally from canonical RHO and RAS GEFs (41) and may also act as a chaperone for RHO, RAS, RAP, and other families of GTPases (41–43), potentially facilitating their prenylation and delivery to the membrane. RAS-driven oncogenesis also depends on RHO proteins (11). The CDKO screens point to specific GIs among RAP1GDS1, RHOA, and mutant KRAS not previously reported that raise the possibility of targeting RHOGEFs like RAP1GDS1 for cancer therapies.

To validate the specificity of this GI for KRAS-dependent cell lines, we used CRISPR–Cas9 to knock out RHOA, RAP1GDS1, or both genes together in a set of KRAS-dependent and KRAS-independent lung adenocarcinoma cell lines and in human bronchial epithelial cells (HBEC; Supplementary Fig. S8A). Double knockout of RAP1GDS1 and RHOA selectively impaired proliferation in three KRAS-dependent lung adenocarcinoma lines but showed minimal synergy in two KRAS-independent lines (Fig. 6A; Supplementary Fig. S8B). Notably, the RAP1GDS1–RHOA double knockout is comparably deleterious to KRAS knockout in the KRAS-dependent lines. Neither single-gene knockout had a strong effect on proliferation, indicating that only the loss of both genes in combination confers a vulnerability in KRAS-dependent cells. Analysis of apoptosis by flow cytometry revealed that double knockout induced cell death only in KRAS-dependent cells (Fig. 6B and C).

As we and others have shown previously, growth in spheroid culture reproduces important aspects of cancer biology not fully recapitulated by two-dimensional cultures (44). To further validate the importance of RAP1GDS1–RHOA synthetic lethality in KRAS-dependent cells, we tested the susceptibility of lung adenocarcinoma cell lines to dual RAP1GDS1 and RHOA knockout in three-dimensional culture. Individual knockout of RAP1GDS1 or RHOA had a modest and variable effect on sphere number and size in KRAS-dependent and KRAS-independent cells (Fig. 6D and E). Double knockout significantly decreased both sphere number and sphere size, but only in KRAS-dependent cells. Notably, the double knockout had a minimal effect on untransformed HBECs (Fig. 6D; Supplementary Fig. S8C).

Because RHO kinases (ROCK) are key effectors of RHOA function (45, 46) and pharmacologic inhibitors are available, we combined ROCK inhibition with RAP1GDS1 knockout to determine whether this would phenocopy dual loss of RHOA and RAP1GDS1. Treatment of two lung adenocarcinoma cell lines with different doses of the ROCK inhibitor fasudil (50, 100, and 150 μmol/L) resulted in a dose-dependent decrease in sphere growth. fasudil treatment of RAP1GDS1 knockout cells decreased sphere growth in KRAS-dependent H23 cells comparably to RHOA knockout alone (Supplementary Fig. S8D and S8E). However, we also observed decreased sphere growth from fasudil treatment in RHOA knockout and RAP1GDS1–RHOA double-knockout cells, suggesting that ROCK is not a strong determinant of the synergy of RAP1GDS1–RHOA double knockout. In addition, we cannot exclude the possibility that ROCK inhibitors may affect targets beyond ROCK1/2. In contrast to H23 cells, in KRAS-independent H1437 cells fasudil treatment had no further effects on growth in any of the knockout conditions.

Finally, to validate this synthetic lethal interaction in vivo, we performed competitive growth assays using xenografts of two KRAS-dependent lung adenocarcinoma cell lines (Fig. 6F). We injected NSG mice subcutaneously with cells transduced with a safe/safe construct expressing mCherry mixed 1:1 with one of four DKO constructs expressing GFP: safe/safe, safe/RHOA, RAP1GDS1/safe, or RAP1GDS1/RHOA. Tumors were allowed to develop for 30 days, after which the mice were sacrificed and tumor cell suspensions analyzed by flow cytometry to quantify the ratio of GFP-positive knockout to mCherry-positive control cells. We found that in RAP1GDS1–RHOA tumors of both cell lines the GFP-positive population was barely detectable, confirming that the double knockout of RAP1GDS1 and RHOA is highly deleterious in vivo (Fig. 6G and H).

Taken together, these studies identify a susceptibility of KRAS-mutant lung adenocarcinoma cells to combined suppression of both RAP1GDS1 and RHOA. Further evaluation of the mechanistic basis for this interaction may provide a new strategy for targeting KRAS-dependent cancer.

Genetic mutations cause pathway rewiring in cancer cells that both drives pathologic behaviors and introduces susceptibilities that can be exploited in the clinic. Understanding this rewiring requires both knowledge of the molecules involved and the organization of interactions between them. Large-scale tumor profiling efforts (1) and single-gene CRISPR and RNAi screening efforts (9–11) have both made valuable progress in assembling a “parts list” of cancer pathways. However, these methods have limited ability to identify the linkage of cancer-related genes in cellular pathways. Cancer pathways have complex features that make potential pharmacologic targets invisible to these methods. Such features identified in the RAS pathway, such as the RAF inhibitor paradox (47) and feedback inhibition (48), have important implications for cancer therapy development. The question of which pathway components are redundant is particularly important in RAS signaling, as many factors, including the RAS proteins themselves, are represented by multiple paralogs. Many agents in clinical use, such as pan-RAF, pan-MEK, and pan-PI(3)K inhibitors, target families of proteins rather than individual paralogs. Single-gene knockouts cannot mimic the family-wide inactivation caused by these drugs. Rewiring events may also introduce combination susceptibilities where cancer cells become sensitive to treatment by multiple agents. These possibilities, too, are not addressed by conventional, single-target screening methods, which do not directly address complex, multigene dependencies.

We reasoned that we could address these shortcomings by systematic interaction mapping. By combining AP/MS and CRISPR-based CDKO GI mapping of RAS pathway components, we both expanded our knowledge of the RAS pathway and revealed new physical and genetic relationships between pairs of components. We have identified functional redundancies and differences between RAS, RAP, RAF, and RALGEF family members. The GIs we observe point to previously undescribed cross-talk between the RAF kinases and other RAS effector proteins. Finally, we have discovered new prospects for combination therapies involving both well-known targets like the RAF kinases or RAL GTPases and newer ones like RAP1GDS1. This combined PPI and GI map both aids our understanding of the RAS pathway and serves as a model for similar approaches to elucidate other disease pathways (Fig. 7).

Protein–protein interaction mapping by AP/MS has several features ideal for the particular challenges of RAS pathway biology. Among these are the unbiased identification of interacting proteins, the ability to capture PPIs as they occur in situ, and the ability to modulate sets of observed PPIs by judicious mutation of bait proteins.

To exemplify the utility of these experiments, we focused on two proteins that interacted with GTP-locked mutant KRAS: RIN1 and RADIL. Although a RIN1–HRAS PPI was previously observed (21), RIN proteins have received little attention in RAS signaling. We demonstrate that RIN1 is a significant regulator of macropinocytosis, a process that has recently been specifically associated with RAS activation (Fig. 2B and C; refs. 23, 24). These observations provide a mechanistic link between RAS effectors and macropinocytosis and provide a context for more detailed screening (49).

We also demonstrate that KRAS binds and regulates RADIL via its RA domain similarly to known KRAS targets (Fig. 2D–H). RADIL is a new RAS pathway member previously implicated in adhesion and cell migration (28, 29), which we confirm and find linked to KRAS activity. Though important for cell migration, further work is required to determine how RADIL aids small GTPase activity.

By combining the results of several AP/MS experiments in networks, we and other groups can make powerful comparisons among proteins, especially paralogs (17, 50–52). By comparing the interactomes of HRAS, KRAS, and NRAS in A549 cells, we determined that the PPIs of HRAS with canonical RAS effector proteins are not observed in these cells (Fig. 1B–F). KRAS and NRAS, meanwhile, copurified many known RAS effector proteins, including the RAF kinase family and RALGEFs. Because the six RAS protein baits were expressed at comparable levels (Fig. 1B), the strong differences likely reflect differences in biological regulation rather than artifacts of overexpression. We note that there are some other, “known” interactions that were not recapitulated by these experiments. In some cases, such as interactions between GTPases and their upstream GEFs, we are aware that the interaction is difficult to observe with AP/MS-based methods (18). In other cases, such as the absence of TIAM1 or PIK3CA/p110α, we cannot differentiate between “false negatives” and interactions actually downregulated by cellular mechanisms. In the comparisons we make in this study, we are careful to compare PPIs involving members of the same protein families (e.g., RAS proteins and RA domain–containing effectors) in the same cell line to avoid these ambiguities as much as possible.

Virtually every RA domain–containing protein observed interacting with any GTP-locked RAS protein in vivo copurified with all GTP-locked RAS isoforms tested in cell-free systems. Judging by the PPI data, it appears that only a subset of these effectors are activated by a given RAS protein in a given cellular context. Therefore, some cellular regulation independent of the nucleotide state of each RAS protein modulates each RAS-effector PPI. Control of RAS and RAS effector localization is one likely mechanism by which different PPI patterns could be determined (53). Similarly, by mapping the different interactomes of three RALGEFs, we provide support for long-suspected differences in their cellular roles (31). By examining differences in PPIs between these different proteins, we can identify functional differences between similar proteins and more confidently map their links to different regulators, effectors, and cellular processes.

We were struck by the observation that proteins copurifying with RAS GTPases also copurify with RAP GTPases. Clustering analysis demonstrates that these two families share such strongly overlapping sets of interacting partners that they are not easily resolved (Fig. 3C). Although some links between these GTPases, and their sequence similarity, have been noted previously (34), the extent of PPI overlap suggests RAS–RAP cross-talk has hitherto been greatly underestimated. The regulation of RAP GTPase PPIs with RAS effectors, and their influences on RAS signaling, is an important problem. Our genetic interaction data (Supplementary Fig. S7) show differing patterns of synergistic and buffering interactions between RAS and RAP family members, suggesting that their functional alignment may itself be context-dependent. Further research should characterize signaling upstream of RAP GTPases and test whether RAS and RAP GTPase interactions with their common effectors are cooperative or antagonistic.

Having identified direct physical links between pathway components by PPI mapping, we conducted GI mapping to identify indirect functional links between components. The GI networks recovered predicted synthetic lethal GIs between the two MEK kinases (MAP2K1 and MAP2K2) and the two ERK kinases (MAPK1 and MAPK3). However, we found that GIs within the MAPK pathway differed between cell lines (Fig. 4D; Supplementary Figs. S5C and S7). Pan-RAF, pan-MEK, and pan-ERK inhibitors, although approved for some indications, often meet either toxicity limitations or adaptive resistance in clinical use (3, 54). The variable and specific GIs observed here support the investigation of isoform-specific RAF or ERK inhibition in RAS-driven cancers. Our subsequent comparison of GIs across multiple cell lines (Fig. 5) showed that MAPK pathway genes have more synthetic lethal GIs with non-MAPK genes in KRAS-dependent than KRAS-independent cells (Supplementary Fig. S6D). However, these GIs involve different gene pairs in different cell lines (Supplementary Fig. S7). The existence of these GIs, though, indicates a surprising degree of cross-talk between the MAPK pathway and other branches of the RAS pathway.

The heterogeneous susceptibilities we observe among cell lines agrees with other work (14). However, the lines compared in that study (A549, HEK293, and HeLa) had very different primary sources, whereas the 9 we compare are all lung adenocarcinoma. Although we might expect similar provenance of these lines to yield similar GI patterns, we find many important differences (Fig. 5E). These apparently rewired GIs present a high-dimensional and somewhat daunting problem for the discovery of combination therapies; some may be modulated by other tumor suppressors or secondary oncogenes (Supplementary Fig. S6F). Although PPIs help identify GIs, we observe most GIs between proteins distant in physical interaction space (Supplementary Fig. S5A), which could only be suggested by genetic or functional evidence. Notwithstanding, several GIs were sufficiently consistent to identify new predictions for combination therapies or the use of selective inhibitors.

For example, we found a strong synthetic lethal GI between RALA and RALB specific to KRAS-dependent lines (Fig. 5C, middle). Although RALs are well-known RAS pathway members, they have yet to gain much interest as a therapeutic target. Although a RAL inhibitor has been developed targeting RAL interactions with RALBP1 (55), our PPI and GI data suggest RAL binding of the exocyst complex may be a more relevant endpoint to assay.

Another noteworthy synthetic lethal GI between KRAS and NRAS was specific to cells not dependent on KRAS (Fig. 5C, right). Given that the PPI results indicate that KRAS and NRAS both engage common effectors that HRAS does not (Fig. 1C), we suggest that by knocking out these two paralogs we have impaired the critical aspects of RAS signaling. The fact that KRAS-dependent cells showed less synergy between KRAS and NRAS knockout suggests that these cells have indeed become specifically addicted to KRAS, rather than to RAS signaling more generally. Although we have examined only a few KRAS-independent lines here, these results show that at least some nominally KRAS-independent lines are jointly dependent on multiple RAS genes.

We have classified GIs by their association with KRAS dependence, and not their KRAS mutation, although genetic alterations are easier to identify in the clinic. We expect aberrant KRAS activity, and not mutation per se, to modulate GIs. We note, however, that in both this work and others, KRAS-mutant lines are largely a subset of KRAS-dependent lines (11). Many KRAS-dependent lines, such as the H1975 line we examine, have upstream activating mutations, notably in EGFR (1). Therefore, we expect the combination susceptibilities of this work to apply in both KRAS-mutant lung adenocarcinoma as well as some undetermined subset of KRAS–wild-type lung adenocarcinoma.

Although we and other groups have established functional links between RAS and the effectors RIN1, RADIL, and RALGDS family proteins, they were not involved in the most salient genetic interactions in this study. The genes encoding these proteins do not generally show high mutation rates in patients (1), nor are they frequently crucial dependencies in cancers (11). However, they are among the most frequently interacting genes screened in the two smaller GI libraries (Fig. 5F). How can we resolve this apparent paradox?

As we have noted before, many genes in this pathway are represented by multiple paralogs. Although there are genes, such as RAF1, that are uniquely critical in their gene family across a variety of cell lines, this seems to be an exception rather than the rule. Many GIs we observed link the same gene families in either line (Supplementary Fig. S7), but involve different members of those families in different lines. These may hint at gene family–gene family synthetic lethal interactions that would require several knockouts to observe genetically but could potentially be phenocopied by treatment with two drugs that inhibited all members of one protein family each. The numerous genetic interactions involving RADIL, RIN1 and RIN2, and RALGEFs motivate further investigation of these pathways.

The GI maps pointed to a KRAS-associated susceptibility of combined loss of RAP1GDS1 and RHOA. Mutant RAS has previously been shown to require intact RHO signaling for transformation (56), which may be required for cell-cycle progression in KRAS-mutant cancers (57). DepMap reveals RHOA and RHOC dependency in a number of KRAS-dependent lines, including lung adenocarcinoma, supporting its activity in these cancers.

Previous studies have established that RAP1GDS1 acts as a noncanonical GEF for RHOA and RHOC (38). It also promotes prenylation and membrane trafficking of RHO and RAS proteins, which is critical for their signaling (42, 43). We identified RAP1GDS1 in the PPI data with all of the RAS and RAP GTPases as baits, making it a highly central member of this signaling network (Figs. 1F and 3B; Supplementary Figs. S1B, S2A and S2C). The importance of these binding partners in cancer signaling makes RAP1GDS1 a noteworthy therapeutic target.

In vitro validation in 2-D and 3-D cultures of lung adenocarcinoma cell lines confirmed that dual RAP1GDS1 and RHOA knockout selectively impaired growth of KRAS-dependent cells (Fig. 6; Supplementary Fig. S8), an effect we find tightly linked to induction of apoptosis. The mechanism behind this GI is unclear. One parsimonious explanation is that the two knockouts are a multi-hit insult to RHOA signaling, compromising both RHOA and its RAP1GDS1-mediated prenylation and/or activation. RAP1GDS1 also serves as a GEF for RHOC, and it may be that the RAP1GDS1–RHOA double knockout effectively impairs two branches of RHO signaling (38). A RHO-focused GI network including RAP1GDS1 would greatly clarify this point. Two recent studies also point to a link between RAP1GDS1 and KRAS prenylation and trafficking to the plasma membrane, suggesting that RAP1GDS1 loss may modulate KRAS activity more directly (39, 58). Finally, disruption of RAP1GDS1 has been shown to alter global prenylation of many GTPases from the RAP, RAB, RHO, and RAS families (59). Therefore, it could be that these other GTPases synergize with RHO and KRAS signaling and that the simultaneous disruption of all of these pathways by RAP1GDS1 deletion results in the synthetic lethality we observe.

Although identifying the precise mechanism of this genetic interaction will require further study, our results establish that this GI manifests specifically in KRAS-dependent cells. In vivo analysis using lung adenocarcinoma xenografts further suggests that combined RAP1GDS1 and RHOA inhibition may provide a promising strategy to treat KRAS-driven tumors.

The concept of cancer pathway rewiring, a change of the linkage of cell biological pathways, underlies the core theories of targeted therapy development. In this work, we provide a systematic survey of RAS pathway interactions in RAS-transformed lung adenocarcinoma. Combining protein interaction mapping with genetic screens allows us to map pathway wiring directly. In doing so, we have gained new insights into RAS-driven tumor biology, discovered new susceptibilities of RAS-driven tumors, and provided a resource for other investigators to do the same.

pMCB306, pKHH30, and p293 Cas9-BFP were described previously (13). pCMV-VSV-G and pCMV-dR8.2 dvpr were gifts from Prof. Bob Weinberg of MIT (Addgene plasmid #8454 and #8455; ref. 60). pWPXLd LAP plasmids were also described previously (61). Cassettes were introduced into GST-fusion, myc-fusion, and LAP-fusion plasmids by Gateway cloning (Supplementary Methods). Point mutations in bait proteins were introduced by QuikChange mutagenesis using primers in Supplementary Table S1.

HEK293, A549, H23, H358, H1437, H1568, H1650, H1975, H2009, and NL20 cells were obtained from ATCC. HEK293-LAP cell lines were generated using HEK293-FlpIn cells (61). Culture methods are described in Supplementary Methods. Cas9- or inducible Cas9-expressing strains of all cell lines were generated using lentiviral vectors (Addgene #52962 and #83481, respectively) and tested for Cas9 cutting efficiency as described previously (13). Cell lines were screened for Mycoplasma contamination with Cell Check 9 Plus Test (IDEXX) and kept below 15 passages.

AP/MS was carried out as described in ref. 61, and quantified as in ref. 17, with changes described in the Supplementary Methods.

4 × 10⁷ cells of each line expressing a LAP–RAS-mutant construct were lysed in 300 μL of 150 mmol/L KCl, 50 mmol/L HEPES pH 7.4, 1 mmol/L MgCl₂ 10% glycerol, and 0.3% NP-40 (lysis buffer) supplemented with 3 μL phosphatase inhibitor cocktail 2 (Sigma-Aldrich) and 1 μg/mL each of leupeptin, pepstatin, and chymostatin (Sigma-Aldrich) on ice for 10 minutes. Cells were centrifuged for 15 minutes at 4°C. Supernatant concentration was measured by Bradford assay and diluted to 9 mg/mL protein in 300 μL lysate. For a 5% input sample, 9 μL were withheld and mixed with 4 × LDS sample buffer +2.5% β-mercaptoethanol. Protein A beads (4.83 μL) conjugated to rabbit αGFP antibody (Thermo Fisher Scientific) were added to the remaining supernatant. Samples were incubated 16 hours at 4°C. Beads were washed 5 times with 200 μL 200 mmol/L KCl, 50 mmol/L HEPES pH 7.4, 1 mmol/L MgCl₂ 10% glycerol, and 0.3% NP-40 (lysis buffer), then immunoblotted as described in ref. 61, probed 16 hours at 4°C with rabbit αRAF1 (Abcam) and mouse αGFP (Invitrogen).

Purifications of recombinant GST-tagged HRAS, KRAS, and NRAS with the indicated mutations were performed as described in ref. 61. Briefly, proteins were purified from Rosetta 2 E. Coli cells transformed with plasmids encoding the indicated constructs.

Cell-free coimmunoprecipitation experiments were performed as described in ref. 61. Primary antibodies used were rabbit αGST (Sigma-Aldrich) and mouse α-myc clone 9E10 (EMD Millipore).

Images were acquired on a Marianas spinning disk confocal (SDC) inverted microscope (Intelligent Imaging Innovations) using a Zeiss Plan-NEO 40x/1.3 Oil Ph3 objective, a Yokogawa csu22 confocal scanning unit, and a Photometrics Cascade 1K camera.

sgRNAs against target genes or control safe-targeting sgRNAs were cloned into the MCB306 or KH30 (MCB306 with an hU6 promoter) vectors. For double sgRNA vectors, the U6-sgRNA cassette was excised from the KH30 vector and cloned into the pMCB306 vector. All sgRNA sequences are listed in Supplementary Table S8. Lentivirus was produced in 293T cells as described previously (61), filtered, and applied directly to cells for infection at a MOI < 1. Clonal RIN1 knockout generation is described in the Supplementary Methods.

Macropinocytosis assays were conducted as described in ref. 25, and are elaborated in the Supplementary Methods.

Scratch wound assays were performed as described previously (30), and are elaborated in the Supplementary Methods.

Quantitative real-time PCR for the RADIL knockdown experiments began with reverse transfection (Supplementary Methods), except that siRNAs were pooled. Cells were harvested and frozen 24, 48, and 72 hours post-transfection. Relative RNA levels were quantified using the TaqMan RADIL Assay (Thermo Fisher Scientific) with a compatible GAPDH Assay (Life Technologies) as a control probe using the TaqMan manufacturer's protocol in 96-well MicroAmp Optical reaction plates (Applied Biosystems, N8010560)

Clustering analysis was conducted by the UPGMA (unweighted pair group method with arithmetic mean) method using the cosine distance metric, using the “linkage” method provided by the “scipy” module (62). Bait–prey interactions were binarized on the basis of an FDR of < 0.05, and this adjacency matrix was used as input.

CDKO screening was conducted as described previously (13), and is elaborated in the Supplementary Methods.

Short-term cell viability assays were performed using the IncuCyte S3 (Essen Bioscience). Five days post-infection, the cells were trypsinized and counted, and 500 to 1,000 viable cells were seeded in 200 μL of puromycin-containing medium into each well of a 96-well plate. Each condition was seeded in triplicate. Cells were imaged every 12 hours for the duration of the experiment. Cell area mask was calculated on the green channel to include only GFP-positive cells.

For anchorage-independent sphere growth, cells were seeded into 24-well ultra-low attachment (Corning, #3473) plates (20,000 viable cells per well) in 2 mL of complete medium supplemented with 0.5% methylcellulose. The spheres were allowed to form for 9 to 20 days (depending on the cell line). GFP-positive spheres were imaged using the Leica DMi8 fluorescence microscope. Sphere size and number were quantified using ImageJ (NIH, Bethesda, MD). For sphere growth assays in the IncuCyte S3 (Essen Bioscience), cells were seeded into complete medium supplemented with 1.5% methylcellulose (500–1,000 cells/well) in 96-well ultra-low attachment plates (Corning, #3474). Each condition was seeded in triplicate.

H23-Cas9 or A549-iCas9 cells were infected with the indicated double sgRNA vectors and puromycin selection was applied for 7 days. Stably infected cell populations were trypsinized and transferred to 15-cm dishes to expand. Three days prior to implantation the A549-iCas9 cells were treated with doxycycline (1 μg/mL). On the day of the implantation (D1), cells were trypsinized, counted, and analyzed by flow cytometry. Cells expressing mCherry (safe/safe sgRNA vector) and GFP (gene-targeting sgRNAs) were pooled 1:1, centrifuged, and resuspended in a mixture of serum-free DMEM (Corning, #15-017-CV) and Matrigel (Corning, #356237). Cells were injected subcutaneously into both flanks of NSG mice (2 × 10⁶ cells per flank, 3 mice per condition for A549 cells and 5 × 10⁶ cells per flank, 4 mice per condition for H23 cells). All the procedures involving mice were approved by the Institutional Animal Care and Use Committee (IACUC) at University of California, San Francisco (protocol #AN15761). Thirty days after implantation (D30), tumors were dissected, chopped, and dissociated into a single-cell suspension using PBS supplemented with Collagenase/Dispase (Roche, #11097113001) and DNAse I (Worthington, #LS002006) at 37°C for 30 minutes with agitation. Digested samples were passed through a 40-μm filter, resuspended in complete DMEM, and centrifuged. Cell pellets were incubated for 1 minute in RBC buffer (155 mmol/L NH₄Cl, 12 mmol/L NaHCO₃, 0.1 mmol/L EDTA) to remove red blood cells, resuspended in complete DMEM, and centrifuged. Resulting cell pellets were resuspended in PBS with 1% serum and analyzed by flow cytometry on Accuri C6 (BD Biosciences). The ratio of GFP to mCherry fluorescence in tumor samples at D30 was normalized to D1 cell populations and represented as log fold-change between these two time points.

Statistical significance was calculated using unpaired Student t test between the tested sample and the corresponding safe/safe control. Resulting p-values were corrected for multiple-hypothesis testing using the Benjamini–Hochberg procedure.

Analyses of statistical significance were performed using the tests described in the text, using the “scipy” library in Python. Where shown, 95% CIs were calculated using the “statsmodels” library in Python (www.statsmodels.org).

Analysis of mutation rates in Supplementary Fig. S7 relied on the provisional TCGA lung adenocarcinoma data from cbioportal.org (1). Using the EllipticalEnvelope method provided by scikit-learn (63), the mahalanobis distance of each gene from a common centroid based on relative mutation rates and copy number alteration frequencies was calculated.

No potential conflicts of interest were disclosed.

M.R. Kelly: Conceptualization, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. K. Kostyrko: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. K. Han: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. N.A. Mooney: Resources, investigation, methodology. E.E. Jeng: Conceptualization, resources, formal analysis, investigation, methodology, writing-original draft. K. Spees: Investigation, methodology. P.T. Dinh: Resources, investigation, methodology. K.L. Abbott: Resources, investigation, methodology. D.M. Gwinn: Conceptualization, resources, supervision, funding acquisition, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. E.A. Sweet-Cordero: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, writing-original draft, project administration, writing-review and editing. M.C. Bassik: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, writing-original draft, project administration, writing-review and editing. P.K. Jackson: Conceptualization, resources, supervision, funding acquisition, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing.

We thank Alex Loktev, PhD, Kevin Wright, PhD, Christopher Adams, PhD, and Ryan Leib, PhD, for their help with mass spectrometry. We thank Dr. Ji Luo for advice on growth of lung cancer cell lines in spheroid culture. P.K. Jackson, M.C. Bassik, and E.A. Sweet-Cordero were supported by grant U01CA199216 as part of the RAS Synthetic Lethal Network from the National Cancer Institute. E.A. Sweet-Cordero was also supported by 2R01CA129562-03A1 from the NIH. M.C. Bassik was also funded by the NIH Director's New Innovator Award Program (project no. 1DP2HD084069-01). K. Kostyrko was supported by SNSF Postdoc Mobility Grants P2LAP3_164922 and P300PB_174377. K. Han was supported by The Walter V. and Idun Berry Postdoctoral Fellowship. E.E. Jeng was supported by the Hubert Shaw and Sandra Lui Stanford Graduate Fellowship.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.