Multiplexed screens identify RAS paralogues HRAS and NRAS as suppressors of KRAS-driven lung cancer growth

Oncogenic KRAS mutations occur in approximately 30% of lung adenocarcinoma. Despite several decades of effort, oncogenic KRAS-driven lung cancer remains difficult to treat, and our understanding of the regulators of RAS signalling is incomplete. Here to uncover the impact of diverse KRAS-interacting proteins on lung cancer growth, we combined multiplexed somatic CRISPR/Cas9-based genome editing in genetically engineered mouse models with tumour barcoding and high-throughput barcode sequencing. Through a series of CRISPR/Cas9 screens in autochthonous lung cancer models, we show that HRAS and NRAS are suppressors of KRASG12D-driven tumour growth in vivo and confirm these effects in oncogenic KRAS-driven human lung cancer cell lines. Mechanistically, RAS paralogues interact with oncogenic KRAS, suppress KRAS–KRAS interactions, and reduce downstream ERK signalling. Furthermore, HRAS and NRAS mutations identified in oncogenic KRAS-driven human tumours partially abolished this effect. By comparing the tumour-suppressive effects of HRAS and NRAS in oncogenic KRAS- and oncogenic BRAF-driven lung cancer models, we confirm that RAS paralogues are specific suppressors of KRAS-driven lung cancer in vivo. Our study outlines a technological avenue to uncover positive and negative regulators of oncogenic KRAS-driven cancer in a multiplexed manner in vivo and highlights the role RAS paralogue imbalance in oncogenic KRAS-driven lung cancer. Using somatic genome editing and Tuba-seq, Tang et al. uncover a previously uncharacterized role for HRAS and NRAS in impairing KRAS–KRAS interaction to suppress lung tumour growth.


Validation of HRAS and NRAS as tumour suppressors
To further validate the effect of inactivating six top candidate genes (Hras, Nras, Cand1, Aldh1a, Fnta and Nme2) on oncogenic KRASdriven tumour growth in vivo and confirm that these results are driven by ontarget effects, we generated three barcoded LentisgRNA/Cre vec tors targeting each gene. To contextualize the effects of inactivating these genes we also included vectors targeting three established tumour suppressors (Lkb1, Rbm10 and Rb1) in this pool (LentisgValidation/Cre; Fig. 2a) 18,20,34 . We initiated tumours with the LentisgValidation/Cre pool in KT;H11 LSL-Cas9 and KT mice and assessed metrics of tumour initia tion and growth 12 weeks after tumour initiation (Fig. 2b,c). Targeting Fnta consistently reduced growth across all three sgRNAs, while the impact of inactivating Aldh1a and Nme2 was more variable (Fig. 2d and Extended Data Fig. 4). Most importantly, all sgRNAs targeting Hras and Nras significantly increased tumour growth (Fig. 2d,e and Extended Data Fig. 4b). Notably, Hras inactivation increased tumour growth to a similar extent as inactivation of the Rb1 and Rbm10 tumour sup pressors ( Fig. 2d and Extended Data Fig. 4b). These results suggest a potentially pivotal role for wildtype HRAS and NRAS in suppressing oncogenic KRASdriven lung tumour growth in vivo.
We also validated the tumoursuppressive functions of HRAS and NRAS by initiating tumours in KT;H11 LSL-Cas9 mice with individual sgInert, sgHras or sgNrascontaining LentisgRNA/Cre vectors Genetic and proteomic mapping has revealed that KRAS interacts with a large network of proteins 10,11 . These KRASinteracting proteins include canonical regulators and effectors, as well as many proteins that remain poorly understood in the context of oncogenic KRASdriven lung cancer. Much of our understanding of RAS signalling stems from diverse cellular and cellfree systems [12][13][14] . Thus, while recent studies have mapped KRAS protein-protein interaction networks 10,11,15,16 , it remains difficult to assess the relevance of these interactions to cancer growth in vivo. Genetically engineered mouse models of oncogenic KRASdriven cancer uniquely recapitulate autochthonous tumour growth and have contributed to our understanding of KRAS signalling 17 . However, the development and use of such models has traditionally been insufficiently scalable to broadly assess modifiers of KRASdriven tumour growth. The ability to uncover functional components of RAS signalling that affect lung cancer growth in vivo in a multiplexed man ner would accelerate our understanding of RAS biology and could aid in the development of pharmacological strategies to counteract hyperactivated KRAS.
To enable the analysis of genetic modifiers of lung tumour growth in vivo, we recently integrated somatic clustered regularly interspaced short palindromic repeats (CRISPR)/Cas9based genome editing with tumour barcoding and highthroughput barcode sequencing (Tubaseq) [18][19][20] . This approach allows precise quantification of the effects of inactivating panels of genes of interest on lung tumour ini tiation and growth in a multiplexed manner. By employing Tubaseq to assess the functions of KRASinteracting proteins nominated by unbiased affinity purification/mass spectrometry (AP/MS), we show that wildtype HRAS and NRAS suppress the growth of oncogenic KRASdriven lung adenocarcinoma. Competition between oncogenic KRAS and wildtype HRAS and NRAS diminishes KRAS-KRAS interac tion and suppresses downstream signalling. In vivo screening across multiple oncogenic contexts revealed that HRAS and NRAS specifically suppress the growth of tumours driven by oncogenic KRAS. Our study reveals that changes in the ratio of RAS paralogues (which we term 'RAS paralogue imbalance') is a driver of oncogenic KRASdriven lung cancer.

Selection of candidate KRAS-interacting proteins
To identify KRASinteracting proteins that could affect oncogenic KRASdriven lung tumour growth in vivo, we integrated preexisting proteomic data from AP/MS studies with gene expression data from cancer cells from autochthonous mouse models ( Fig. 1a) 10,21 . We pri oritized a list of candidate genes according to the probability of their protein products interacting with KRAS and other RAS GTPases, as well as their messenger RNA expression in mouse models of oncogenic KRAS G12D driven lung cancer (Fig. 1b,c and Extended Data Fig. 1a-d) 10,21 . We selected 13 proteins that represent diverse aspects of RAS biology, including RAS paralogues (HRAS and NRAS, which were supported by the identification of paraloguespecific peptides), RAS regulators (RAS GRF2 and RAP1GDS1) (refs. 22,23), a RAS farnesyltransferase (FNTA) 24,25 and RAS effectors (RAF1, RGL2) (refs. 26,27), as well as several proteins whose functions in RAS signalling are understudied. While the major ity of these candidate genes trend towards amplification in oncogenic KRASdriven lung adenocarcinoma, NRAS, HRAS and ALDH1A1 have deep genomic deletions (Extended Data Fig. 1e) 28 . Interestingly, some of these proteins bound preferentially to either GTP or GDPbound KRAS, while others interact with KRAS independently of its nucleotide state (Fig. 1c).

KRAS-interacting proteins impact lung tumour growth in vivo
Given that KRASinteracting proteins could have either positive or negative effects on tumour growth, we first assessed whether we could detect genetargeting events that have deleteri ous effects on tumour fitness using Tubaseq. We initiated tumours in Kras LSL-G12D/+ ;Rosa26 LSL-tdTomato ;H11 LSL-Cas9 (KT;H11 LSL-Cas9 ) and control KT Article https://doi.org/10.1038/s41556-022-01049-w ( Fig. 2f). Inactivation of either Hras or Nras increased tumour growth as assessed by direct fluorescence and histological analyses (Fig. 2g-k). Collectively, these results suggest that RAS paralogues constrain the growth of oncogenic KRAS G12D driven lung cancer.

HRAS and NRAS suppress growth of human lung cancer cells
To assess the relevance of HRAS and NRAS as tumour suppressors in human lung cancer, we tested the function of these proteins in oncogenic KRASdriven human lung adenocarcinoma cell lines. Previ ous genomescale CRISPR/Cas9 screens revealed that inactivating these genes is most often either detrimental or of no consequence to cancer cell line growth under standard culture conditions (Extended Data Fig. 5a) 10,35 . Interestingly, HRAS and NRAS suppressed the growth of oncogenic KRAS G12S driven A549 cells grown in 2D culture conditions, and were growthsuppressive in several oncogenic KRASdriven lung cancer cell lines grown in 3D culture conditions, suggesting that these   Fig. 5d). Inactivation of HRAS or NRAS in oncogenic KRASdriven cells increased proliferation when cells were grown with limited serum and increased clonal growth potential in anchorageindependent con ditions (Fig. 3a,c,d). Reexpression of HRAS in these HRASnull cells impaired proliferation and clonal growth (Fig. 3b,e,f and Extended Data Fig. 5e). H23 cells with inactivated HRAS or NRAS also formed larger and more proliferative tumours after intravenous and subcutaneous KT;H11  Percentiles of the tumour size distribution 50 60 70  80  90  95   b   sgAldh1a#2  sgAldh1a#3  sgCand1#1  sgCand1#2  sgCand1#3  sgFnta#1  sgFnta#2  sgFnta#3  sgNeo1  sgNeo2  sgNeo3  sgNT  sgHras#1  sgHras#2  sgHras#3  sgNras#1  sgNras#2  sgNras#3  sgNme2#1  sgNme2#2  sgNme2#3  sgNeo1  sgNeo2  sgNeo3  sgNT  sgHras#1  sgHras#2  sgHras#3  sgNras#1  sgNras#2  sgNras#3 Relative LN mean tumour size   Fig. 5f-i). These results demonstrate that wildtype HRAS and NRAS can suppress the growth of oncogenic KRASdriven human lung cancer cells in vitro and in vivo, further suggesting that HRAS and NRAS are tumour suppressors in oncogenic KRASdriven lung adenocarcinoma.

RAS paralogue inactivation increases signalling downstream of oncogenic KRAS
Wildtype KRAS has been shown to be tumoursuppressive in multiple experimental models of oncogenic KRASdriven cancer, probably due to its ability to interact and compete with oncogenic KRAS [36][37][38] . We have demonstrated that wildtype HRAS and NRAS suppress onco genic KRAS G12D driven lung cancer growth in vivo. We first assessed the expression of KRAS, HRAS and NRAS in human and mouse lung cancer cells. HRAS and NRAS are more highly expressed than KRAS in KRASdriven lung cancer cells, supporting their roles in regulating KRAS signalling (Extended Data Fig. 6a,b). To further explore the molecular mechanism driving this effect, we assessed whether HRAS and NRAS alter signalling downstream of oncogenic KRAS. We performed pERK immunohistochemistry on lung tumours initiated with LentisgRNA/ Cre vectors containing sgInert, sgHras or sgNras in KT;H11 LSL-Cas9 mice. Inactivation of HRAS or NRAS increased the number of pERKpositive cells in KRAS G12D driven lung cancer ( Fig. 4a and Extended Data Fig. 6c). Subcutaneous tumours from H23 cells with inactivated HRAS or NRAS also contained more pERKpositive cells compared with tumours from wildtype (sgSAFE) H23 cells ( Fig. 4b and Extended Data Fig. 6d). Finally, sorted cancer cells from KT;H11 LSL-Cas9 mice with lung tumours initiated with LentisgHras/Cre also had greater pERK and pAKT compared with those from tumours initiated with LentisgInert/Cre (Fig. 4c). Inactivation of either Hras or Nras in mouse (HC494) and human (H23 and HOP62) oncogenic KRASdriven cell lines increased ERK phosphorylation, while their effects on AKT phosphorylation were more cell context dependent (Fig. 4d,e). Reexpression of wildtype HRAS in HRASnull H23 and HOP62 human lung cancer cells reduced ERK phosphorylation while again having a cellcontextdependent effect on AKT phosphorylation ( Fig. 4f and Extended Data Fig. 6e). Furthermore, reexpression of either HRAS or NRAS in HRAS/NRAS doubleknockout HOP62 cells reduced pERK. Previous publications have shown that inactivating wildtype KRAS increases sensitivity to MEK inhibitors 37,39 . Consistent with these studies, inactivation of HRAS in H23 cells modestly increased sensitivity to the MEK inhibi tor trametinib while reexpression of HRAS made cells more resist ant (Fig. 4g,h). These data suggest that inactivation of HRAS or NRAS hyperactivates MAPKERK signalling in KRAS mutant cancer cells [40][41][42] .

RAS paralogues suppress oncogenic KRAS-KRAS interaction
RAS proteins interact and form functional clusters on membranes to efficiently recruit downstream effectors [43][44][45] . Whether RAS proteins form dimers or oligomers through direct interactions or through close physical proximity is debated within the field 16,46,47 . We assessed whether HRAS and NRAS interact with KRAS. Both AP/MS data and coimmunoprecipitation experiments suggest that HRAS and NRAS interact with KRAS G12D , supporting the existence of heterotypic RAS-RAS interactions, possibly through a domain containing the α4/α5 interface ( Fig. 5a and Extended Data Fig. 7a,b). To assess the ability of RAS paralogues to interact with oncogenic KRAS G12D , we adapted a splitluciferase reporter system, which relies on luciferase complementation to quantify RAS-RAS interactions in living cells (Fig. 5b) 16 . We first used this splitluciferase reporter system to confirm the interaction between HRAS and NRAS with KRAS G12D (Extended Data Fig. 7c-e) 16 . Through expression of wildtype KRAS, HRAS or NRAS in KRAS G12D -KRAS G12D interaction reporter cells and control reporter cells, we found that all wildtype RAS paralogues can disrupt KRAS G12D -KRAS G12D interactions. While the other RAS family mem bers RAC1 or RALA did not impact KRAS G12D -KRAS G12D interactions, we validated the RAP1A-KRAS G12D interaction that was predicted from the initial AP/MS data ( Fig. 5c and Extended Data Fig. 7f-i). Lastly, we overexpressed HRAS in KRAS G12D expressing 293T cells and found HRAS-KRAS G12D interaction reduced BRAF-KRAS G12D interactions (Extended Data Fig. 7j).

Patient-derived HRAS and NRAS mutations impair interaction with oncogenic KRAS
Our findings suggest that the tumoursuppressive function of wildtype HRAS and NRAS are mediated, at least in part by competitive interac tions with oncogenic KRAS. We therefore hypothesized that there could be HRAS and NRAS mutations in human tumours with oncogenic KRAS that impair this interaction. To evaluate this possibility, we analysed data from AACR Project GENIE 48 . Mutations in HRAS and NRAS were rare (pancancer frequency of nonsynonymous mutations was 0.83% and 2.87%, respectively). The majority of these were oncogenic muta tions in codons 12, 13 or 61 that occurred in samples lacking oncogenic KRAS (Extended Data Fig. 8a,b). We did, however, identify multiple rare nononcogenic HRAS and NRAS mutations ( Fig. 5d and Extended Data Fig. 8c,d). We next assessed the ability of these mutants to interact with oncogenic KRAS. We measured the ability of four HRAS mutants and five NRAS mutants, as well as a control Y64A HRAS mutant that has been suggested to reduce HRAS-HRAS dimerization 47 , to inhibit KRAS G12D -KRAS G12D interactions. This identified two HRAS mutants (T50M and R123C) and one NRAS mutant (R102Q) that are unable to reduce KRAS G12D -KRAS G12D interactions ( Fig. 5e and Supplementary  Fig. 8e,f). Interestingly, both HRAS T50 and HRAS R123 are located close to the predicted HRASKRAS G12D interface involving the α4 and α5 helices ( Fig. 5f and Extended Data Fig. 9). These findings are consistent with a model in which interaction of wildtype RAS paralogues with oncogenic KRAS suppresses tumour growth, such that mutations that impair this interaction are beneficial to tumour growth.
Previous publications have shown that RAS proteins differen tially bind to RAS effectors and thus could function differently in their downstream signalling 10,49 . Reanalysis of HRAS and NRAS AP/ MS datasets shows that the binding affinity of GTPbound HRAS to RAF is more similar to GDPbound KRAS than to its activated, GTPbound form, suggesting that RAS heterodimers containing HRAS may be less able to activate downstream oncogenic signalling (Fig. 5g) 10 . To test this hypothesis, we reexpressed wildtype HRAS, HRAS Y64A or the two patientderived HRAS T50M and HRAS R123C mutants in HRASnull lung cancer cells. Reexpression of wildtype HRAS, but not any of the three mutants, reduced ERK phosphorylation and cell proliferation ( Fig. 5h,i). Similarly, reexpression of wildtype NRAS, but not NRAS R102Q , suppressed ERK phosphorylation and proliferation in NRASnull lung cancer cells (Extended Data Fig. 8g,h). These results further suggest that RAS paralogue imbalance alters oncogenic KRAS signalling via oncogenic KRAS-wildtype RAS paralogue interactions and is thus a driver of lung cancer growth.

HRAS and NRAS are specific suppressors of oncogenic KRAS-driven lung cancer growth
Our in vivo data demonstrate that HRAS and NRAS function as tumour suppressors, and our in vitro results suggest that these suppressive effects are mediated through the interaction of these RAS paralogues with oncogenic KRAS. If the mechanism by which HRAS and NRAS suppress tumour growth is mediated by interactions with oncogenic KRAS, then these proteins should not be tumour suppressors in lung adenocarcinomas in which activation of the RAS/RAF/MEK signalling pathway occurs downstream of KRAS. To test this directly in autoch thonous tumours, we initiated tumours in mouse models of oncogenic KRASdriven and oncogenic BRAFdriven lung cancer using a subpool of barcoded LentisgRNA/Cre vectors (LentisgMultiGEMM/Cre; Fig. 6a). In addition to vectors targeting Hras and Nras, this pool contained vectors targeting several known tumour suppressors (Apc, Rbm10 Article https://doi.org/10.1038/s41556-022-01049-w and Cdkn2a) and other KRASinteracting proteins (Aldh1a and Nme2), as well as control vectors (Fig. 6a). We initiated tumours with the LentisgMultiGEMM/Cre pool in KT and KT;H11 LSL-Cas9 mice as well as in BrafT;H11 LSL-Cas9 mice that contain a Creregulated allele of oncogenic BRAF V618E (the mouse equivalent of human BRAF V600E ) (Fig. 6b) 50 . Fifteen weeks after tumour initiation, the two models had similar maximum tumour sizes, but BrafT;H11 LSL-Cas9 mice had larger tumours of relatively uniform size, which is consistent with previous results (Fig. 6c-f) 50 .
Our Tubaseq data also allowed us to compare the impact of the CRISRP/Cas9inactivated genes across oncogenic contexts. Importantly, while inactivation of Hras or Nras increased the growth of oncogenic KRASdriven lung tumours, inactivation of Hras or Nras had no effect on the growth of oncogenic BRAFdriven lung cancer ( Fig. 6g and Extended Data Fig. 10). These results were consistent for both LentisgRNA/Cre vec tors targeting each gene. The known tumour suppressor genes assayed (Apc, Cdkn2a and Rbm10) generally retained their growthsuppressive  Article https://doi.org/10.1038/s41556-022-01049-w effects in the BRAFdriven model, suggesting that the abrogation of effect observed for Hras and Nras is not due to a generic inability of addi tional alterations to increase BRAFdriven lung tumour growth ( Fig. 6h and Extended Data Fig. 10). Thus, HRAS and NRAS specifically suppress oncogenic KRASdriven tumour growth in vivo. We also identified several other instances of oncogenetumour suppressor epistasis. For instance, Apc inactivation has a greater effect on BRAFdriven lung cancer, whereas Rbm10 inactivation has a greater effect on KRASdriven lung cancer ( Fig. 6h and Extended Data Fig. 10). In contrast, inactivation of Nme2, Fnta and Aldh1a reduced initiation and growth of oncogenic KRASdriven and oncogenic BRAFdriven lung cancer, suggesting that they are generally required for optimal lung cancer growth in vivo (Extended Data Fig. 10). Thus, our paired screens not only localized the effect of Hras and Nras inactivation, but also highlighted the value of this approach in uncovering alterations that have effects within or across oncogenic contexts.          Interestingly, our data suggest that HRAS is a more potent sup pressor of tumour growth than NRAS in mouse models of oncogenic KRASdriven lung adenocarcinoma, while NRAS appears to have stronger effects on KRAS-KRAS dimerization and downstream signal ling in human cells. Our data suggest that NRAS may be slightly more potent than HRAS at inhibiting oncogenic KRAS-KRAS interactions (Extended Data Fig. 8c-g) but that the protein expression of the RAS paralogues can vary among cell types and cancer types (Extended Data Fig. 7a,b). Thus, permolecule ability to disrupt KRAS-KRAS interac tion, preferential interactions with downstream effectors, and the stoichiometry of KRAS, HRAS and NRAS proteins probably integrate to drive the cellular and in vivo phenotypes.

Discussion
The impact of RAS paralogue imbalance may extend beyond lung cancer and KRAS codon 12 mutations. Germline HRAS deletion increases the development of KRASdriven pancreatic cancer, skin pap illomas and carcinogeninduced KRAS Q61 lung cancer [52][53][54] . However, the impact of inactivating RAS paralogues in cancers with different driver oncogenes (for example, oncogenic EGFRdriven lung cancer) is likely to be different since RAS proteins also serve as important components in growth factor signalling pathways. Whether RAS paralogue heter odimerization also impacts signalling during normal development, homeostasis or other diseases states remains unknown. Collectively, these findings suggest that modulating RAS protein interactions, such as by skewing the stoichiometry of oncogenic to wildtype RAS or by forcing interparalogue competition, could lead to therapeutic strategies.
Given the complexity of RAS signalling, other nonmutually exclu sive mechanisms by which RAS paralogues could reduce oncogenic KRASdriven cancer growth should be considered. For example, it has been reported that upstream regulators, such as SOS1, could bridge the interaction between oncogenic and wildtype RAS 55 . GDPbound wildtype HRAS and NRAS could also compete with oncogenic KRAS for upstream guanine nucleotide exchange factors and thus reduce RAS signalling 56 . Additionally, it is possible that HRAS and NRAS compete with oncogenic KRAS for downstream effectors. Whether HRAS and NRAS also function through these alternative routes, and how differ ent mechanisms are integrated to execute their tumoursuppressive functions, will require additional investigation.
The National Cancer Institute 'RAS Pathway V2.0', contains more than 200 proteins known or suspected to be involved in RAS signalling. Characterizing the role of these proteins in tractable in vivo mod els of RASdriven cancer remains a challenge. Our study outlines a technological avenue to study KRASspecific signalling components in a multiplexed manner. By harnessing the power of Tubaseq, we quantified the effects of more than a dozen putative RAS pathway genes on tumour growth simultaneously. Furthermore, by per forming paired screens in oncogenic KRASdriven and oncogenic BRAFdriven cancer models, we localized the growthsuppressive effects of these RAS paralogues. Our study demonstrates the fea sibility of performing in vivo genetic interaction screens, and the power of such approaches to provide insight into the mechanisms of tumour suppression. Future studies of this type should enable a more quantitative understanding of the role of RAS pathway components in oncogenesis.

Online content
Any methods, additional references, Nature Portfolio reporting sum maries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contri butions and competing interests; and statements of data and code avail ability are available at https://doi.org/10.1038/s4155602201049w. Article https://doi.org/10.1038/s41556-022-01049-w

Mouse research
The use of mice for the current study has been approved by the Institutional Animal Care and Use Committee at Stanford Univer sity, protocol number 26696. Kras LSL-G12D/+ (RRID:IMSR_JAX:008179), R26 LSL-tdTomato (RRID:IMSR_JAX:007909) and H11 LSL-Cas9 (RRID:IMSR_ JAX:027632) mice have been previously described. They were on a C57BL/6:129 mixed background for the experiments in Figs. 1, 2 and 4, and were on a C57BL/6 background for the experiments in Fig. 6. The B6.129P2(Cg)Braf tm1Mmcm /J (BRAF CA/+ ) mice were initially generated by Dankort et al. and obtained from the Jackson Laboratory (RRID:IMSR_ JAX: 017837). We used balanced sex of animals with age ranging from 8 to 15 weeks at the time of tumour initiation. Mice were housed at Stanford SIM1 barrier facility under a 12 h-12 h light-dark cycle with dark hours between 18:30 and 6:30. Housing temperature at 68-73 °F under 40-60% humidity.
Trametinib was purchased from MedChemExpress (HY10999); 5bromo2′deoxyuridine (10280879001) and dluciferin (L95045MG) were purchased from SigmaAldrich. All plasmids used in this study are listed in Supplementary Table 1 and are available from our laboratory (key plasmids will be donated to Addgene).

Design, generation, barcoding and production of lentiviral vectors
sgRNA sequences targeting the putative tumour suppressor genes were designed using CRISPick (https://portals.broadinstitute.org/ gppx/crispick/public). All sgRNA sequences are presented in Supple mentary Table 2. Each desired sgRNA vector was modified from our previously published pll3U6sgRNAPgkCre vector via sitedirected mutagenesis (New England Biolabs, E0554S). The generation of the barcode (BC) fragment containing the 8nucleotide sgID sequence and 20nucleotide degenerate BC, and subsequent ligation into the vectors, were performed as previously described 18,19 .
Lentiviral vectors were produced using polyethyleniminebased transfection of 293T cells with delta8.2 and VSVG packaging plasmids in 150 mm cell culture plates. Sodium butyrate (SigmaAldrich, B5887) was added 8 h after transfection to achieve a final concentration of 20 mM. Medium was refreshed 24 h after transfection. Twenty millili tres of viruscontaining supernatant was collected 36, 48 and 60 h after transfection. The three collections were then pooled and concentrated by ultracentrifugation (112,000 g for 1.5 h) and resuspended overnight in 100 μl PBS, then frozen at −80 °C and thawed and pooled at equal ratios immediately before delivery to mice.
For the individual sgRNA tumour initiation experiments in Fig. 3, tumours were allowed to develop for 12 weeks after delivery of indi vidual sgRNAexpressing lentiviral vectors targeting Neo2, Hras or Nras. Tumours were initiated in KT;H11 LSL-Cas9 mice with 1 × 10 5 ifu per mouse.
For the paired screen experiments in Fig. 6, tumours were allowed to develop for 15 weeks after delivery of a lentiviral pool that con

Tuba-seq library generation
Genomic DNA was isolated from bulk tumourbearing lung tissue from each mouse as previously described. Briefly, benchmark control cell lines were generated from LSLYFP mouse embryonic fibroblasts transduced with a barcoded LentisgNT3/Cre vector (NT3: an inert sgRNA with a distinct sgID) and purified by sorting YFP pos cells. Three benchmark control 'spikein' cell lines (500,000 cells each) were added to each mouse lung sample before lysis to enable the calculation of the absolute number of neoplastic cells in each tumour from the number of sgIDBC reads. Following homogenization and overnight protease K digestion, genomic DNA was extracted from the lung lysates using standard phenol-chloroform and ethanol precipitation methods. Subsequently, Q5 HighFidelity 2x Master Mix (New England Bio labs, M0494X) was used to amplify the sgIDBC region from 32 μg of genomic DNA in a total reaction volume of 800 μl per sample. The unique dualindexed primers used were Forward: AAT GAT ACG GCG ACC ACC GAG ATC TAC AC8 nucleotides for i5 indexACA CTC TTT CCC TAC ACG ACG CTC TTC CGA TCT6 to 9 random nucleotides for increased diversityGCG CAC GTC TGC CGC GCT G and Reverse: CAA GCA GAA GAC GGC ATA CGA GAT6 nucleotides for i7 index GTG ACT GGA GTT CAG ACG TGT GCT CTT CCG ATC T9 to 6 random nucleotides for increased diversityCAG GTT CTT GCG AAC CTC AT. The PCR prod ucts were purified with Agencourt AMPure XP beads (Beckman Coulter, A63881) using a double size selection protocol. The concentration and quality of the purified libraries were determined using the Agilent High Sensitivity DNA kit (Agilent Technologies, 50674626) on the Agi lent 2100 Bioanalyzer (Agilent Technologies, G2939BA). The libraries were pooled on the basis of lung weight to ensure even sequencing depth, cleaned up again using AMPure XP beads, and sequenced (read length 2× 150 bp) on the Illumina HiSeq 2500 or NextSeq 500 platform (Admera Health Biopharma Services).

Western blot
Cells were lysed in RIPA buffer (50 mM Tris-HCl (pH 7.4), 150 mM NaCl, 1% Nonidet P40 and 0.1% SDS) and incubated at 4 °C with continu ous rotation for 30 min, followed by centrifugation at 12,000 rcf for 10 min. The supernatant was collected, and the protein concentra tion was determined by BCA assay (Thermo Fisher Scientific, 23250). Protein extracts (10-50 μg) were separated on 4-12% SDS-PAGE and transferred onto polyvinylidene fluoride membranes. The mem branes were blocked with 5% nonfat milk in TBS with 0.1% Tween 20 (TBST) at room temperature for 1 h, cut according to the molecular weight of the target protein (with at least two flanking protein mark ers), followed by incubation with primary antibodies diluted in TBST (1:1,000) at 4 °C overnight. After three 10 min washes with TBST, the membranes were incubated with the appropriate secondary anti body conjugated to HRP diluted in TBST (1:10,000) at room tempera ture for 1 h. After three 10 min washes with TBST, protein expression was quantified with enhanced chemiluminescence reagents (Fisher Scientific, PI80196). Antibodies

Co-immunoprecipitation assay
293T cells with stable expression of mycKRAS G12D were transfected with plasmids expressing HAtagged HRAS and Flagtagged NRAS for 24 h before lysed with icecold immunoprecipitation lysis buffer (Thermo Scientific, 87788) containing protease inhibitor cocktail (Thermo Fisher Scientific 78442). The lysates were precleared with magnetic beads (Thermo Fisher Scientific, 88802) at 4 °C for 2 h. Then protein concentration was determined by BCA assay (Thermo Fisher Scientific, 23250) and equal amount of protein lysis were incubated with antiMyc (Thermo Fisher Scientific, 88842) or IgG (Cell Signaling, 5873S) magnetic beads at 4 °C for 12 h. The immunoprecipitates were collected using a MACSiMAG Separator (Miltenyi Biotec, 130092168), washed for three times with immunoprecipitation lysis buffer and three times with TBST. The immunoprecipitates were eluted via incubating in 1× NonReducing Sample Buffer (Thermo Fisher Scientific, 39001) at 95 °C for 10 min before subjected to immunoblotting.

Histology and immunohistochemistry
Lung lobes were fixed in 4% formalin and paraffin embedded. Hae matoxylin and eosin (H&E) staining was performed using standard methods. Immunohistochemistry was performed on 4 μm sections using the Avidin/Biotin Blocking Kit (Vector Laboratories, SP2001), AvidinBiotin Complex kit (Vector Laboratories, PK4001) and DAB Peroxidase Substrate Kit (Vector Laboratories, SK4100) following standard protocols.

Cell proliferation assay (CCK8)
For cell proliferation assays, cells were seeded in 96well plates at a density of 5,000 cells per well and allowed to adhere overnight in regular growth medium (day 0). Cells were then cultured in medium as indicated on each figure panel for 7 days. Relative cell number were measured every other day using Cell Counting Kit8 (Bimake, B34304) according to the manufacturer's instructions.

Colony formation assay
For clonogenic assays, cells were seeded in sixwell plates at a den sity of 500 cells per well and allowed to adhere overnight in regular growth medium. Cells were then cultured in medium as indicated on each figure panel for 14 days. Growth medium with or without drugs was replaced every 2 days. At the end point, cells were stained with 0.5% crystal violet in 20% methanol. Colony numbers were calculated using ImageJ.

Xenograft studies in immunocompromised mice
For intravenous transplants into immunocompromised NSG mice, 5 × 10 5 H23 cells were injected into one of the lateral tail veins. Mice were killed 28 days postinjection and lung lobes were fixed in 4% formalin and paraffin embedded. For subcutaneous transplants into immunocompromised NSG mice, 2 × 10 6 each of H23 cells (sgSAFE, sgHRAS and sgNRAS) were resuspended in 200 μl Matrigel ® Basement Membrane Matrix (Corning, 354234) and injected into three parallel sites per mouse. Mice were killed 28 days postinjection. Tumours were dissected, and the weight, height, width and length of each tumour was measured. Tumour volume was roughly calculated via the follow ing formula: Tumor volume = (4/3) × π × (Tumor length/2) × (Tumor depth/2) × (Tumor depth/2).
Maximal tumour size/burden permitted by Institute of Medicine Animal Care and Use Committee is 1.75 cm 3 , the maximal tumour size/ burden was not exceeded in our study. Institute of Medicine Animal Care and Use Committee approved all animal studies and procedures.
Upon termination of the ReBiL assay, (1) to measure raw luciferase activity, 300 μM dluciferin was added to 96well plate cultures and incubated at 37 °C for 30 min, and raw luminescent data for both Renilla and firefly luciferase were collected by a Tecan microplate reader; (2) Nature Cell Biology Article https://doi.org/10.1038/s41556-022-01049-w to quantify the expression of 1/2luc fusion proteins, ReBiL cells from 6well plate cultures were collected with RIPA lysis buffer for protein extraction, and western blots were performed for HAtag, Myctag and HSP90 expression. Then the ReBiL2.0 score was calculated via the following formula:

Analysis of human lung adenocarcinoma cancer genome sequencing data (for rare HRAS and NRAS mutations)
To assess evidence that HRAS and NRAS function as KRASspecific tumour suppressors in human cancer, we queried publicly available cancer genomic datasets. GENIE Release 9.1public was accessed through the Synapse platform and data on somatic mutations (data_mutations_extended.txt), sample and patientlevel clinical data (data_clinical_sample.txt and data_clinical_patient.txt), and genotyping panel information (genomic_information.txt) were downloaded. While it is unclear how our findings may extrapolate to cancer types beyond lung adenocarcinoma, HRAS and NRAS muta tions are rare (occurring at frequencies of just 0.83% and 2.87% in GENIE samples, respectively), so we performed a pancancer anal ysis. Each sample was assigned to its patient of origin and anno tated for the presence of oncogenic KRAS mutations (defined as missense mutations in KRAS exons 12, 13 or 61) and for the presence of potentially functional HRAS or NRAS mutations (variants that were silent, intergenic or intronic, or fell in the 3′ or 5′ untranslated regions were excluded from this analysis). When multiple samples were derived from the same patient, the patient in question was annotated as having a mutation if it occurred in at least one of their associated samples. From this information we produced a list of the frequency of all HRAS and NRAS variants in patients with and without oncogenic KRAS in both datasets. The genotyping panel information was used to identify GENIE patients who were not geno typed at HRAS and/or NRAS and exclude these from the frequency calculation.

Analysis of DepMap data
Cancer cell line dependency data (DepMap Public 19Q4) and mutation data (Cancer Cell Line Encyclopedia) were acquired from the Broad Institute DepMap Portal (https://depmap.org/portal/). Lung adenocar cinoma cell lines were identified by their Project Achilles identification code and partitioned into KRAS mutant, if they contained a hotspot mutation, or wildtype groups. Subsequently, dependency scores for NRAS or KRAS were calculated for each cell line within the two groups. Finally, the distributions of dependency scores were plotted using GraphPad Prism 9.

Processing of paired-end reads to identify the sgID and BC
Sequencing of Tubaseq libraries produces reads that are expected to contain an 8nucleotide sgID followed by a 30nucleotide BC of the form GCNNNNNTANNNNNGCNNNNNTANNNNNGC, where each of the 20 Ns represents a random nucleotide. Each sgID has a onetoone cor respondence with an sgRNA in the viral pool; thus, the sgID sequence identifies the gene targeted in a given tumour. Note that all sgID sequences in the viral pool differ from each other by at least three nucleotides such that incorrect sgID assignment (and thus, inference of tumour genotype) due to PCR or sequencing error is extremely unlikely. The random 20nucleotide portion of the BC is expected to be unique to each lentiviral integration event and, thus, tags all cells in a single clonal expansion. Note that the length of the BC ensures a high theoretical potential diversity (~4 20 > 10 12 BCs per vector), so while the actual diversity of each LentisgRNA/Cre vector is dictated by the number of colonies generated during the plasmid barcoding step, it is very unlikely that we will observe the same BC in multiple clonal expansions.
FASTQ files were parsed using regular expressions to identify the sgID and BC for each read. To minimize the effects of sequencing error on BC identification, we required the forward and reverse reads to agree completely within the 30nucleotide sequence to be further processed. We also screened for BCs that were likely to have arisen due to errors in sequencing the BCs of genuine tumours. Given the low rate of sequenc ing error, we expect these spurious 'tumours' to have read counts that are far lower than the read counts of the genuine tumours from which they arise. While it is impossible to eliminate these spurious tumours, we sought to minimize their effect by identifying small 'tumours' with BCs that are highly similar to the BCs of larger tumours. Specifically, if a pair of 'tumours' had BCs that were within a Hamming distance of two, and if one of the tumours had fewer than 5% as many reads as the other, then the reads associated with the smaller tumour were attrib uted to the larger tumour. After these filtering steps, the read counts associated with each BC were converted to absolute neoplastic cell numbers by normalizing to the number of reads from the 'spikein' cell lines added to each sample before lung lysis and DNA extrac tion. The median sequencing depth across experiments was ~1 read per 6.4 cells.
For statistical comparisons of tumour genotypes, we applied a minimum tumour size cutoff of 100 cells. In selecting a cutoff, we sought to include tumours that are large enough to be consistently detected despite differences in sequencing depth among mice, while using as many tumours as possible to maximize the statistical power. Importantly, we analysed each Tubaseq dataset with multiple mini mum tumour size cutoffs (50, 100, 200 and 500 cells) and found that our findings were robust.

Summary statistics for overall growth rate
To assess the extent to which a given gene (X) affects tumour growth, we compared the distribution of tumour sizes produced by vectors targeting that gene (sgX tumours) to the distribution produced by our negative control vectors (sgInert tumours). We relied on two statistics to characterize these distributions: the size of tumours at defined percentiles of the distribution (specifically the 50th, 60th, 70th, 80th, 90th and 95th percentile tumour sizes), and the lognormal mean size (LN mean). The percentile sizes are nonparametric summary statistics of the tumour size distribution. In considering percentiles corresponding to the right tail of the distribution, we focus on the growth of larger tumours, thereby avoiding issues stemming from potential variation in cutting efficiency among guides. The LN mean is the maximumlikelihood estimate of mean tumour size assuming a lognormal distribution. Previous work found that this statistic represents the best parametric summary of tumour growth based on the maximum likelihood quality of fit of various common parametric distributions.
To quantify the extent to which each gene suppressed or promoted tumour growth, we normalized statistics calculated on tumours of each genotype to the corresponding statistic. The resulting ratios reflect the growth advantage (or disadvantage) associated with each tumour genotype relative to the growth of sgInert tumours.
For example, the relative ith percentile size for tumours of geno type X was calculated as: Relative size at i th percentile sgX = i th percentile of sgX distribution i th percentile of sgInert distribution Article https://doi.org/10.1038/s41556-022-01049-w Likewise, the relative LN mean size for tumours of genotype X was calculated as: Relative LNmean sgX = LNmean of sgX distribution LNmean of sgInert distribution

Summary statistics for relative tumour number and relative tumour burden
In addition to the tumour size metrics described above, we character ized the effects of gene inactivation on tumourigenesis in terms of the number of tumours and total neoplastic cell number ('tumour burden') associated with each genotype. Unlike the aforementioned metrics of tumour size, tumour number and burden are linearly affected by lentiviral titre and are thus sensitive to underlying differences in the representation of each LentisgRNA/Cre vector in the viral pool. Criti cally, each Tubaseq experiment included a cohort of KT control mice. As KT mice lack expression of Cas9, all LentisgRNA/Cre vectors are functionally equivalent in these mice, and the observed tumour num ber and burden associated with each sgRNA reflects the makeup of the viral pool.
To assess the extent to which a given gene (X) affects tumour num ber, we first normalized the number of sgX tumours in KT;H11 LSL-Cas9 mice (also KT;p53 flox/flox ;H11 LSL-Cas9 and Braf LSL-V600E/+ T; H11 LSL-Cas9 mice in the initial Krasinteracting protein screen and the paired screen, respectively) to the number of sgX tumours in the KT mice: Tumour number sgX = Number of sgX tumours in KT; H11 LSL−Cas9 mice

Number of sgX tumours in KT mice
As with the tumour size metrics, we then calculated a relative tumour number by normalizing this statistic to the corresponding statistic calculated using sgInert tumours: Relative tumour number sgX = Tumour number sgX

Tumour number sgInert
Genes that influence relative tumour number modify the probabil ity of tumour initiation and/or the very early stages of oncogenedriven epithelial expansion, which prior work suggests are imperfectly cor related with tumour growth at later stages. Relative tumour number thus captures an additional and potentially important aspect of tumour suppressor gene function.
Analogous to the calculation of relative tumour number, we char acterized the effect of each gene on tumour burden by first normal izing the sgX tumour burden in Cas9expressing mice to the burden in KT mice: We then calculated relative tumour burden by normalizing this number to the corresponding statistic calculated using sgInert tumours: Relative tumour burden sgX = Tumour burden sgX

Tumour burden sgInert
Tumour burden is an integration of tumour size and number, and thus reflects the total neoplastic load in each mouse. Tumour burden is thus more strongly related to morbidity than are our metrics of tumour size and is closely related to traditional measurements of tumour progression such as duration of survival and tumour area. While intuitively appealing, tumour burden is notably noisier than our metrics of tumour size as it is strongly determined by the size of the largest tumours.

Calculation of confidence intervals and P values for tumour growth and number metrics
Confidence intervals and P values were calculated using bootstrap resampling to estimate the sampling distribution of each statistic. To account for both mousetomouse variability and variability in tumour size and number within mice, we adopted a twostep, nested bootstrap approach where we first resampled mice, and then resampled tumours within each mouse in the pseudodataset. A total of 10,000 bootstrap samples were drawn for all reported P values. The 95% confidence intervals were calculated using the 2.5th and 97.5th percentile of the bootstrapped statistics. As we calculate metrics of tumour growth that are normalized to the same metrics in sgInert tumours, under the null model where genotype does not affect tumour growth, the test statistic is equal to 1. Twosided P values were thus calculated as follows: where T is the test statistic and Pr(T > 1) and Pr(T < 1) were calculated empirically as the proportion of bootstrapped statistics that were more extreme than the baseline of 1. To account for multiple hypothesis testing, P values were false discovery rate (FDR)adjusted using the Benjamini-Hochberg procedure as implemented in the Python pack age stats models. Summarized statistics of all Tubaseq experiments in this study can be found in Supplementary Tables 3-6.

AP/MS data visualization
AP/MS data were analysed as described 57 . Briefly, protein spectral matches 10 were normalized by protein length and total spectral matches per experiment. These normalized spectral abundance factors (NSAFs) were then normalized to NSAFs of matched prey proteins from a large cohort of unrelated AP/MS experiments to produce a Zscore. Zscores are proportional to the areas of circles in bubble plots. In clus ter diagrams, NSAFs are binarized by statistical significance (FDR >0.5), similarities between interactome profiles were determined by cosine distance, and dendrogram topology was determined by unweighted pair group method with arithmetic mean.

Modelling RAS-RAS dimers
Potential templates for modelling RAS heterodimers were obtained from the ProtCID database. ProtCID is built from clustering interfaces of homologous proteins obtained from domain-domain contacts within protein crystals in the Protein Data Bank. Hierarchical cluster ing of interfaces is performed with a Jaccardindex similarity metric based on the contacts shared between different interfaces. Models for the structure of the HRAS/KRAS heterodimer were built by super imposing a structure of KRAS G12D (PDB: 5USJ) onto a monomer of the HRAS homodimer in PDB entry 3K8Y. All structural data files gener ated in this study can be accessed via Zenodo (https://zenodo.org/ record/7104280).

Statistical analysis for non-Tuba-seq experiments
Sample or experiment sizes were estimated on the basis of similar experiments previously performed in our laboratory, as well as in the literature. Biological replications (more than five mice for each cohort, more than ten wells per culture condition) of the experiments were as detailed in the figures. All values are presented as mean ± standard deviation (s.d.), with individual data points shown in the figure when possible. Comparisons of parameters between two groups were made by twotailed Student's ttests. The differences among several groups were evaluated by oneway analysis of variance (ANOVA) with Tukey-Kramer post hoc evaluation. P values less than 0.05 and 0.01 were considered significant (*) or very significant (**), respectively. Article https://doi.org/10.1038/s41556-022-01049-w

Statistics and reproducibility
The statistical tests used for each analysis are described in detail in the sections above. All analyses of BC sequencing data were performed in Python (3.4), and visualizations of data were performed in R (3.6.0). Sample sizes were determined on the basis of our previous experi ence conducting similar experiments and, in the case of Tubaseq experiments, on the basis of previously published power analyses 18 . For experiments using western blot as a readout, at least three indepen dent experiments were repeated with similar results. In all the experiments reported in this study, no data points were excluded. No randomization was used in this study. Data collection and analysis were not performed blind to the conditions of the experiments. Analyses of BC sequencing data used nonparametric statistics; therefore, no assumptions about the distribution of data were made. For all other ana lyses, data distributions were assumed to be normal, but this was not for mally tested, and individual data points are plotted to show distribution.

Reporting summary
Further information on research design is available in the Nature Port folio Reporting Summary linked to this article.

Data availability
The human cancer genomic data analysed for the presence of rare HRAS and NRAS variants in this manuscript were derived from the AACR's Project GENIE (https://www.aacr.org/professionals/research/ aacrprojectgenie/) Release 9.1public dataset. All data files that sup port the findings of this study are available through the Synapse plat form (https://www.synapse.org/#!Synapse:syn24179657). Human cancer genomic data analysed for alterations in KRASinteracting proteins were derived from the TCGA PanCancer Atlas dataset, which is publicly available through cBioPortal at https://www.cbioportal. org/study/summary?id=luad_tcga_pan_can_atlas_2018. The protein templates used to model RAS heterodimers in this study are avail able through the ProtCID database (KRAS G12D : PDB entry 5USJ, HRAS homodimer: PDB entry 3K8Y), and resulting structural data files can be accessed through Zenodo (https://zenodo.org/record/7104280). AP/ MS data were derived from ref. 10 Fig. 2 | Tumor barcoding coupled with barcode sequencing (Tuba-seq) can uncover engineered alterations that reduce tumor number and growth. a-b. Schematic of the Tubaseq approach to measure the effects of essential gene inactivation on tumor growth. Tumors were initiated with pool of barcoded lentiviralsgRNA/Cre vectors targeting known essential genes and tumor suppressor Apc (LentisgEssential/Cre) in KT and KT;H11 LSL-Cas9 mice (a). Tubaseq was performed on each tumorbearing lung 12 weeks after initiation (b). c. Points denote tumor sizes at indicated percentiles for each sgRNA relative to the size of sgInertcontaining tumors at the corresponding percentiles. Percentiles that are significantly different from sgInert (twosided FDRcorrected p < 0.05) are in color. d. The impact of each sgRNA on mean tumor size relative to sgInerts, assuming a lognormal distribution of tumor sizes (LNmean). sgRNAs with twosided FDRcorrected P < 0.05 are in bold. e. Points denote the impact of each sgRNA on tumor burden relative to sgInerts and normalized to the same statistic in KT mice. Relative burdens significantly different from sgInert (twosided FDRcorrected p < 0.05) are in color. f. Points denote the impact of each sgRNA on tumor number relative to sgInerts and normalized to the same statistic in KT mice. Relative tumor numbers significantly different from sgInert (twosided FDRcorrected p < 0.05) are in color. g. Points denote the impact of each sgRNA on tumor number plotted against its impact on LNmean tumor size. The lines at y = 1 and x = 1 indicate no effect relative to sgInert on tumor number and size, respectively. For panels c and e-g: Error bars indicate 95% confidence intervals around point estimates of the test statistic. Confidence intervals and Pvalues were calculated using a nested bootstrap resampling approach across 9 KT;H11 LSL-Cas9 mice and 2 KT mice. sgInerts are in gray and the line at y = 1 indicates no effect. The α4α5 HRAS dimer from PDB entry 3K8Y was used as a template. KRAS G12D from PDB entry 5USJ was superposed with the program PyMol on one or both monomers of 3K8Y to form the heterodimers and the homodimer respectively.

Extended Data
Residues T50 and R123 were mutated with PyMol. R123 is involved in an intrachain salt bridge with residue E143, which also participates in the RASRAS interface. Mutation to cysteine results in an uncompensated charge on E143, which may destabilize the RASRAS interaction. All four structures were relaxed with the program Rosetta using the FastRelax protocol with the Ref2015 scoring function 59 . Rosetta uses the backbonedependent rotamer library of Shapovalov and Dunbrack to repack side chains around the mutated sites 60 . The resulting energies were: KRAS G12D KRAS G12D , 1122.8 kcal/mol; HRASKRAS G12D , 1144.8 kcal/ mol; HRAS T50M KRAS G12D , 1135.5 kcal/mol; HRAS R123C KRAS G12D , 1130.9 kcal/mol. Residues T50 (magenta) and R123 (orange) are indicated in sticks.