Combinatorial In Vivo Genome Editing Identifies Widespread Epistasis and an Accessible Fitness Landscape During Lung Tumorigenesis

Abstract

Lung adenocarcinoma, the most common subtype of lung cancer, is genomically complex, with tumors containing tens to hundreds of non-synonymous mutations. However, little is understood about how genes interact with each other to enable the evolution of cancer in vivo, largely due to a lack of methods for investigating genetic interactions in a high-throughput and quantitative manner. Here, we employed a novel platform to generate tumors with inactivation of pairs of ten diverse tumor suppressor genes within an autochthonous mouse model of oncogenic KRAS-driven lung cancer. By quantifying the fitness of tumors with every single and double mutant genotype, we show that most tumor suppressor genetic interactions exhibited negative epistasis, with diminishing returns on tumor fitness. In contrast, Apc inactivation showed positive epistasis with the inactivation of several other genes, including synergistic effects on tumor fitness in combination with Lkb1 or Nf1 inactivation. Sign epistasis was extremely rare, suggesting a surprisingly accessible fitness landscape during lung tumorigenesis. These findings expand our understanding of the interactions that drive tumorigenesis in vivo.

Introduction

Genetic interactions, in which the phenotype driven by a genetic alteration depends on the underlying genotype, indicate functional relationships between genes that drive evolution. Genetic interactions underlie all aspects of development, homeostasis, and disease, with most human traits being under complex genetic control (Civelek and Lusis 2014; Boyle et al. 2017; Wray et al. 2018; Seldin et al. 2019). Nevertheless, most mechanistic studies within relevant in vivo contexts alter only one or a few genes at a time, limiting their ability to explore how gene products and pathways interact within complex biological settings. Systematic genetic interaction maps in yeast and in mammalian cell lines have revealed a multitude of insights into many aspects of biology (Ehrenreich et al. 2009; Costanzo et al. 2010; VanderSluis et al. 2010; Jahchan et al. 2013; Jameson et al. 2013; Costanzo et al. 2016). For instance, maps of these interactions have allowed rapid identification of new functional complexes, predicted roles for uncharacterized genes, revealed unexpected network rewiring in response to environmental changes, and demonstrated functional repurposing of complexes and interactions during evolution (Costanzo et al. 2010; VanderSluis et al. 2010; Winslow et al. 2011; Xue et al. 2011; Jahchan et al. 2013; Jameson et al. 2013; Mazur et al. 2014; Costanzo et al. 2016).

Cancer genome sequencing has cataloged somatic alterations at the genome-wide level and identified many putative driver genes (Aaltonen et al. 2020; Nguyen et al. 2022). However, identifying recurrent genomic alterations does not necessarily reveal their functional importance to cancer growth, and the impact of gene inactivation remains difficult to glean from cancer genome sequencing data alone (Birkbak and McGranahan 2020; Hebert et al. 2023; Lee et al. 2023). Tumor suppressor gene alterations in particular are diverse and have nonrandom patterns of co-occurrence, suggesting underlying genetic interactions that drive selection in vivo (Kandoth et al. 2013; Sanchez-Vega et al. 2018; van de Haar et al. 2019; Aaltonen et al. 2020). Moreover, the frequency of co-occurrence cannot predict how biologically important inactivation of a gene is within the context of the other alterations within a tumor (van de Haar et al. 2019).

The impact of combinatorial tumor suppressor gene inactivation on neoplastic growth has been investigated using high-throughput studies on cell lines and in much lower throughput using mouse models (Rogers et al. 2018; Foggetti et al. 2021; Parrish et al. 2021; Zhao et al. 2021; Heitink et al. 2022). Cancer cell lines have limited ability to provide insight into how tumor suppressor genes interact to constrain the expansion of tumors in vivo due to their near-optimal growth in culture, their widespread genetic and epigenetic changes, and the lack of an autochthonous microenvironment (Hebert et al. 2023). Conversely, genetically engineered mouse models enable the introduction of defined alterations in oncogenes and tumor suppressor genes in normal somatic cells. The resulting tumors grow entirely within their natural in vivo setting (Chiou et al. 2014; Platt et al. 2014; Xue et al. 2014; Yin et al. 2014). The analysis of combinatorial genetic alterations on tumorigenesis in vivo using conventional Cre/Lox-based systems of controlled gene inactivation has uncovered important genetic interactions between tumor suppressor genes (Walter et al. 2019). However, conventional in vivo cancer models are not sufficiently scalable to generate a broad understanding of genetic interactions between tumor suppressor genes that drive the evolution of tumors (Caswell et al. 2014; Chiou et al. 2014; Xue et al. 2014; Hebert et al. 2023).

CRISPR/Cas9-mediated genome editing in somatic cells has increased the scale of in vivo functional analyses, and integration with tumor barcoding has enabled precise and multiplexed quantification of tumor initiation and growth (Rogers et al. 2017; Cai et al. 2021; Foggetti et al. 2021; Tang et al. 2023). We have generated moderate-scale maps of gene function within autochthonous cancer models by inactivating many genes in parallel in mouse models of lung cancer using pools of barcoded sgRNA-containing lentiviral vectors. Tumor barcoding coupled with high-throughput barcode sequencing (Tuba-seq) can quantify the size of each barcoded clonal tumor and the number of tumors of each genotype, enabling precise measurement of the impact of each gene on tumorigenesis (Rogers et al. 2017). Epistatic interactions between tumor suppressor genes during cancer growth in vivo can be assessed by combining Cre/Lox and CRISPR/Cas9 genome editing (Rogers et al. 2018). However, the requirement for floxed alleles, the considerable time and effort to generate the required compound mouse strains, and cross-cohort comparisons make it unlikely that an accurate and comprehensive analysis of pairwise interactions will be achieved in this manner.

Here, we establish and validate an efficient platform, as well as associated analytical and statistical methods, to integrate Tuba-seq with combinatorial CRISPR/Cas9-mediated gene inactivation within autochthonous mouse cancer models. To uncover the broad epistatic interactions that drive lung tumor initiation and growth in vivo, we initiated lung tumors with a pool of barcoded dual-sgRNA vectors and quantified the effect of pairwise inactivation of every combination of ten tumor suppressor genes. Quantitative and multiplexed analysis of the combinatorial effects of cancer drivers within autochthonous models can uncover the epistatic interactions of genes and pathways, illuminate novel aspects of tumorigenesis, and enable a better understanding of cancer evolution.

Results

Generation of Barcoded Lentiviral Libraries for Dual-sgRNA Screening With Clonal Resolution

Cellular barcoding has facilitated the quantification of clonal heterogeneity and the investigation of experimental evolution (Ben-David et al. 2018; Rovira-Clavé et al. 2022; Sankaran et al. 2022; Serrano et al. 2022). The integration of clonal barcoding into CRISPR screens has increased resolution both in vitro and in vivo (Michlits et al. 2017; Rogers et al. 2017). Viral vectors with pairs of sgRNAs have been used for combinatorial genome editing of cells in culture, but these have lacked the ability to provide simultaneous clonal information (Han et al. 2017; Park et al. 2022). Conversely, while barcoded dual-sgRNA vectors have led to key insights into the tumorigenic potential of complex tumor genotypes in vivo, they have been very limited in scale, required laborious individual cloning (Murray et al. 2019; Murray et al. 2022; Yousefi et al. 2022; Yousefi et al. 2023) and/or have suffered from lentiviral template switching that decouples the barcode from the sgRNA (Hill et al. 2018). To generate barcoded lentiviral-sgRNA/Cre vector pools for tumor barcoding and high-throughput barcode sequencing (Tuba-seq), we recently integrated a diverse barcode within the 21 nucleotides of the U6 promoter directly adjacent to a single sgRNA (U6 barcode Labeling with per-Tumor Resolution Analysis; Tuba-seq^Ultra) (Tang et al. 2024). To generate analogous Tuba-seq^Ultra lentiviral libraries with pairs of sgRNAs, we synthesized dual-sgRNA oligo libraries and cloned them through a circular intermediate prior to ligation into a barcoded lentiviral-U6^BC/Cre backbone (Fig. 1a and supplementary fig. S1a, Supplementary Material online). PCR amplification and high-throughput sequencing of the BC-sgRNA1–sgRNA2 region from bulk tumor-bearing lungs from mice with tumors initiated with these vectors should capture information on tumor genotype and size.

Fig. 1.

A dual-sgRNA system integrated with barcoding of clonal cell lineages and application to tumor suppressor gene epistasis in lung cancer. a) Generation of a library of dual-sgRNA vectors with an integrated diverse barcode to enable clonal tracking. BbsI and BsmBI, type IIS restriction enzyme sites; mU6 and bU6, mouse and bovine RNA Pol III promoters; BC, barcode; PGK, phosphoglycerate kinase promoter; LTR, long terminal repeat. b) sgRNA pairs in the Lenti-U6^BC-sgDbl/Cre pool. The sgRNA in position 1 (sgRNA1) and position 2 (sgRNA2) is indicated, and the number of sgRNAs targeting each gene is shown in parentheses. This pool should create tumors with 45 distinct double knockout combinations. c) Initiation of KRAS^G12D-driven lung tumors with Lenti-U6^BC-sgDbl/Cre. Genotype, lentiviral titer (ifu, infectious units), and mouse numbers are indicated. Note that Kras^LSL−G12D;R26^LSL−Tomato (KT) mice received 25-fold more virus than KT;H11^LSL−Cas9 and KT;H11^Cas9 mice. Bulk tumor-bearing lungs were analyzed after 12 weeks of tumor growth. d) Tomato fluorescent images of lung lobes from the indicated genotypes of mice. Dashed lines outline tissue. Mean ± SD lung weights are indicated. Scale bars, 2 mm. e) Overview of BC-sgRNA1–sgRNA2 read processing, filtering, and analysis. f) Box plot of tumor burden (total number of neoplastic cells of any genotype) normalized to lentiviral titer (ifu) across mouse genotype. g) Box plot of total number of tumors (any genotype of tumor estimated to have >50 neoplastic cells) normalized to lentiviral titer (ifu) across mouse genotype.

Efficient Inactivation of Genes and Gene Pairs in Somatic Cells

To make use of this approach to study epistasis during tumorigenesis, we selected ten tumor suppressor genes that are commonly altered in human lung adenocarcinoma and generated a dual-sgRNA oligo library targeting every pairwise combination of these genes with two sgRNAs each (Lenti-U6^BC-sgDbl/Cre; Fig. 1b, supplementary fig. S1b, Supplementary Material online, and supplementary table S1, Supplementary Material online). For comparison with the effects of inactivation of individual genes, three Inert sgRNAs were paired with each tumor suppressor gene-targeting sgRNA (Fig. 1b). To assess Cas9 cutting efficiency from the first and second sgRNA positions within the vector, every sgRNA pair was created in both arrangements. Finally, the Lenti-U6^BC-sgDbl/Cre pool also contained an sgRNA targeting the essential gene Pcna paired with five Inert sgRNAs, as well as an sgRNA pair targeting the synthetic lethal genes Ccln1 and Ccln2 (Fig. 1b) (Parrish et al. 2021).

We delivered the Lenti-U6^BC-sgDbl/Cre pool through intratracheal intubation to the lungs of Kras^LSL−G12D/+;Rosa26^LSL−Tomato (KT; n = 11), KT;H11^LSL−Cas9 (n = 20), and KT;H11^Cas9 (n = 45) mice (Fig. 1c). In these mice, the integrated lentiviral vector uniquely tags each transduced cell and the resulting tumors with a unique BC-sgRNA1–sgRNA2 element. Cre removes the Lox-Stop-Lox cassettes to enable expression of oncogenic KRAS and tdTomato, which initiates neoplastic growth and labels cancer cells, respectively. Cre-mediated (H11^LSL−Cas9) or constitutive (H11^Cas9) Cas9 targets the genes defined by the sgRNAs. Twelve weeks after lentiviral transduction, KT;H11^LSL−Cas9 and KT;H11^Cas9 mice had similar lung weights (indicating similar overall tumor weights) to KT mice that lack Cas9 despite receiving 25-fold less virus (supplementary fig. S1c, Supplementary Material online). KT;H11^LSL−Cas9 and KT;H11^Cas9 mice also had visibly larger tumors, consistent with dramatically increased growth of tumors that have combinations of tumor suppressor genes inactivated (Fig. 1d).

The BC-sgRNA1–sgRNA2 region was PCR-amplified from genomic DNA extracted from tumor-bearing lungs followed by high-throughput sequencing (Fig. 1e and supplementary fig. S1d, Supplementary Material online). Stringent filtering allowed us to quantify the number of reads from each clonal tumor with each pair of sgRNAs, while accurately excluding spurious reads generated by PCR uncoupling and sequencing errors (Fig. 1e, supplementary fig. S1d, Supplementary Material online, and Methods) (Hegde et al. 2018). We used the number of reads from each tumor to calculate the number of neoplastic (cancer) cells in each tumor and the number of tumors of each genotype, as well as several metrics of tumor growth (supplementary fig. S1e, Supplementary Material online). Analysis of tumors from KT mice showed that all sgRNA1–sgRNA2 pairs were well-represented in the pool, and all genotypes were well-represented in KT;H11^LSL−Cas9 and KT;H11^Cas9 mice (supplementary fig. S2, Supplementary Material online).

The total tumor burden (defined as the total number of neoplastic cells of any genotype) normalized to lentiviral titer was about 10-fold higher in KT;H11^LSL−Cas9 and KT;H11^Cas9 mice relative to KT mice, consistent with increased tumor growth enabled by some genotypes (Fig. 1f). KT;H11^LSL−Cas9 and KT;H11^Cas9 mice also had more tumors than KT mice when normalized to lentiviral titer, suggesting that inactivation of certain genes or gene pairs promotes tumor initiation and/or very early tumor growth and survival (Fig. 1g).

Homozygous inactivation of the essential gene Pcna and homozygous inactivation of both Ccln1 and Ccln2 should reduce tumor number. We initially evaluated the determinants of efficient Cas9-mediated gene inactivation by quantifying the number of tumors with sgPcna relative to sgInert. Targeting Pcna reduced the number of tumors more in KT;H11^Cas9 mice than in KT;H11^LSL−Cas9 mice (Fig. 2a). sgPcna encoded in position 1 reduced tumor number modestly but consistently more than when encoded in position 2 in KT;H11^LSL−Cas9 mice, but both positions were similarly effective in KT;H11^Cas9 mice (Fig. 2a). sgPcna reduced tumor number similarly regardless of whether the sgInert partner was Safe-targeting, non-targeting (NT and eGFP), or not expressed (NE), suggesting that sgRNA competition for Cas9 is not a major factor in driving target gene inactivation in this system (Fig. 2b) (Najm et al. 2018). Coincident targeting of Ccnl1 and Ccnl2, but not targeting either gene individually, reduced tumor number specifically in the more efficient KT;H11^Cas9 background (Fig. 2c). These data from our control vectors document highly efficient gene inactivation.

Fig. 2.

Efficient inactivation of an essential gene, a synthetic lethal pair, and tumor suppressor genes. a) Impact of targeting an essential gene (Pcna) on tumor number in KT;H11^LSL−Cas9 and KT;H11^Cas9 mice with sgPcna as sgRNA1 or sgRNA2 in the vector. Tumor number is normalized to the representation of each vector in KT mice (which lack Cas9) as well as to all sgInert–sgInert vectors. Means ± 95% confidence intervals are shown. b) Impact of targeting Pcna on tumor number in KT;H11^Cas9 mice across all vector configurations with sgPcna as sgRNA1 or sgRNA2 in the vector and different non-targeting (NT), not expressed (NE), and Safe-targeting sgInerts. sgeGFP is a non-targeting sgInert that was previously suggested to be highly competitive for Cas9. Means ± 95% confidence intervals are shown. c) Impact of targeting Ccln1 and/or Ccln2 (synthetic lethal pair) on tumor number in KT;H11^Cas9 mice across all vector configurations. Means ± 95% confidence intervals are shown. d) Tumor sizes at the Xth percentile within the tumor size distribution for tumors with inactivation of each single tumor suppressor gene (TSG) normalized to sgInert–sgInert tumors. Data for tumors with inactivation of each TSG are an aggregate of all sgTSG–sgInert, sgInert–sgTSG, and sgTSG–sgTSG vectors targeting that gene. 95% confidence intervals are shown. e) Comparison of effects on tumor size with TSG-targeting sgRNAs as sgRNA1 or sgRNA2 in a vector. Each dot represents a gene. Means ± 95% confidence intervals are shown.

We also evaluated the efficiency of Cas9-mediated gene inactivation within our dual-sgRNA system by examining the effects of targeting tumor suppressor genes with one sgRNA in either position 1 or position 2, or with two sgRNAs. Inactivation of individual tumor suppressor genes largely recapitulated previous data, with excellent resolution to identify effects on tumor size (Fig. 2d and supplementary fig. S3a, Supplementary Material online) and results consistent between each sgRNA targeting each gene (supplementary fig. S3b and c, Supplementary Material online). In most cases, tumor suppressor gene-targeting sgRNAs did not have markedly different effects when encoded in position 1 compared to position 2 (Fig. 2e and supplementary fig. S3d and e, Supplementary Material online). Targeting the same tumor suppressor gene with two sgRNAs instead of one modestly increased the effect in some cases, but for most tumor suppressor genes had no additional effect on tumor growth (supplementary fig. S3f, Supplementary Material online). Collectively, these data indicate that this barcoded dual-sgRNA system efficiently inactivates genes in somatic cells in vivo, enabling a quantitative dissection of genetic interactions that influence tumor fitness.

Inactivation of Most Tumor Suppressor Gene Pairs Shows Diminishing Returns Epistasis

To identify epistatic interactions between pairs of tumor suppressor genes, we calculated the fitness of each tumor genotype (single and double mutants) relative to tumors with inert sgRNAs (relative fitness; Fig. 3a). We compared the expected double mutant relative fitness (the product of the relative fitness of both single mutants) to the actual double mutant relative fitness to calculate an epistasis score for each gene pair (Fig. 3a). Positive epistasis scores indicate synergistic effects, and negative epistasis scores indicate buffering effects on tumor fitness (Fig. 3a and b). To maximize the number of tumors supporting the interactions between each pair of tumor suppressor genes, we combined the data from double mutant tumors with any sgRNA targeting each gene, as well as pooled tumors with sgRNA pairs in both possible configurations (i.e. sgGeneA–sgGeneB and sgGeneB–sgGeneA). While the actual and expected fitness values for double mutants were highly correlated (supplementary fig. S4a, Supplementary Material online), most epistasis scores were negative/buffering (Fig. 3b and c), suggesting that combined tumor suppressor gene inactivation generally has diminishing returns on tumor fitness. As expected, no epistatic effects were observed in KT mice lacking Cas9, in which no genes were inactivated (supplementary fig. S4b, Supplementary Material online). Moreover, driven largely by interactions involving Pten, epistasis scores showed a negative correlation with the expected fitness of a double mutant (Fig. 3c), consistent with greater diminishing returns when single mutant genotypes already had high tumor fitness.

Fig. 3.

Broad epistasis between tumor suppressor genes impacts lung cancer growth. a) Definitions and calculations of fitness and epistasis metrics. Expected number of tumors is calculated from the observed number of tumors of each genotype in Cas9-negative KT mice, scaled to the lentiviral titer used for each experiment. b) Heatmap of epistasis scores for all pairwise comparisons in KT;H11^Cas9 mice. The color indicates positive or negative epistasis, and the box size indicates significance. Pairs with significant (P < 0.05) and trending (P < 0.1) positive and negative epistasis scores are indicated. c) Expected double mutant relative fitness and epistasis scores negatively correlate. Each dot is a tumor suppressor gene pair. Spearman's rho and associated P value are indicated. d) Fitness effect of a second mutation (columns) on the background of a first mutation (rows) for all genes that increase fitness when inactivated as single mutants in KT;H11^Cas9 mice. The color indicates whether each second mutation increases or decreases fitness, and the box size indicates significance. Mutations with significant (P < 0.05) and trending (P < 0.1) positive and negative changes to fitness are indicated. Keap1 inactivation increases mean tumor size but decreases tumor number, resulting in an overall negative effect on fitness, and so interactions involving Keap1 are shown in supplementary fig. S4c and d, Supplementary  Material online. The fitness effects of second mutations are also depicted as bar plots with 95% confidence intervals in supplementary fig. S5, Supplementary Material online.

While many double mutants had negative epistasis scores and diminishing returns on fitness (Fig. 3b and c and supplementary fig. S4, Supplementary Material online), nearly all of these interactions represented magnitude epistasis, such that the double mutant still had higher fitness than either single mutant. When comparing epistatic interactions between mutants that increased fitness, only Apc–Rnf43 showed sign epistasis, such that the double mutant had lower fitness than either the Apc or Rnf43 single mutants (Fig. 3d and supplementary fig. S5, Supplementary Material online). Keap1 inactivation slightly increased tumor size but greatly reduced tumor number, resulting in an overall decrease in fitness for Keap1 single and double mutant genotypes, though not sign epistasis (supplementary fig. S4c and d, Supplementary Material online). Thus, the overall fitness landscape for oncogenic KRAS-driven tumors is highly accessible, with the overwhelming majority of additional mutations increasing tumor fitness.

Apc Genetic Interactions Are Enriched for Synergistic Effects While Pten Genetic Interactions Most Often Have Buffering Effects

To explore gene-specific trends in epistasis, we calculated average epistasis scores for every pair involving each gene (supplementary fig. S4e, Supplementary Material online). While most genes had slightly negative epistasis scores on average (consistent with overall trends), Apc stood out for having a positive average epistasis score (supplementary fig. S4e, Supplementary Material online). This was driven by synergistic interactions between inactivation of Apc and Keap1, Lkb1, Nf1, and p53 (Fig. 4a). These genetic interactions had the highest epistasis scores across all pairs (supplementary fig. S4f, Supplementary Material online). In contrast, Pten exhibited strongly negative epistasis overall (Fig. 4b), in keeping with the strong effect of Pten inactivation alone on tumor fitness (supplementary fig. S4e, Supplementary Material online).

Fig. 4.

Tumor suppressor gene-specific genetic interactions influence tumor fitness. a and b) Epistasis scores ± 95% confidence intervals for all pairs with Apc a) and Pten b). c to f) The size of tumors with single and double mutants of p53/Rb1 c), Apc/Lkb1 d), Pten/Nf1 e), and Apc/Rnf43 f) at the given percentiles of the tumor size distribution normalized to sgInert–sgInert tumors. p53–Rb1 and Apc–Lkb1 genotypes represent positive (synergistic) epistasis because the double mutants have significantly greater fitness than would be expected if the fitness effects of the composite single mutations were merely multiplicative (additive in log space). Pten–Nf1 represents negative (buffering) epistasis because the double mutant genotype has significantly lower fitness than would be expected if the fitness effects of the composite single mutations were multiplicative. However, as the Pten–Nf1 double mutant still has greater tumor sizes than either single mutant, this represents magnitude epistasis. Conversely, Apc–Rnf43 is an instance of negative epistasis and sign epistasis, as the double mutant has lower tumor sizes compared to Apc single mutant tumors. 95% confidence intervals are shown.

To contextualize genetic interactions, we rank-ordered epistasis scores and assessed their statistical significance. A positive genetic interaction between p53 and Rb1 is well-established across cancer types, including being supported by mutational co-occurrence rates in human lung adenocarcinoma and by data from analogous mouse models of oncogenic Kras-driven lung adenocarcinoma (Xu et al. 1996; Dosaka-Akita et al. 1997; Xing et al. 1999; Zhou et al. 2006; Yamaguchi et al. 2014; Rogers et al. 2018). Consistent with this, inactivation of p53 and Rb had a positive epistasis score and a synergistic effect on tumor growth (Fig. 4c and supplementary fig. S4f, Supplementary Material online). This confirms our ability to detect these types of synergistic genetic interactions and underscores the relative strength of the relationship between the P53 and RB pathways. Apc–Lkb1 and Apc–Nf1 had even higher positive epistasis scores, with marked synergistic effects on tumor growth. Inactivation of these pairs resulted in the growth of exceptionally large tumors not created by the individual gene mutants (Fig. 4d and supplementary fig. S6a, Supplementary Material online). As anticipated, no changes in tumor size were observed in KT mice, which lack Cas9 (supplementary fig. S6b to f, Supplementary Material online).

Combined loss of Pten and Nf1 had the most negative epistasis score among all gene pairs, with Pten–Nf1 double mutant tumors being of a similar size to Pten or Nf1 single mutant tumors (supplementary fig. S4f, Supplementary Material online and Fig. 4e). Similarly, tumors with combined loss of Apc and Rnf43 were smaller than tumors with inactivation of either gene individually and had low tumor number (Fig. 4f), thereby displaying the only instance of sign epistasis across all pairs (Fig. 3d and supplementary fig. S5, Supplementary Material online). Thus, our quantitative approach for examining the effects of tumor genotype on fitness uncovered both strongly synergistic and buffering genetic interactions with broad implications for cancer evolution.

Discussion

Given the genomic complexity of human tumors, understanding genetic interactions is critical to making sense of how tumor genotypes drive tumor evolution and progression. In this study, we combined the pairwise inactivation of ten commonly mutated tumor suppressor genes with quantitative tumor barcoding and sequencing methods to deconvolute how these genes cooperate to promote tumor growth. While we observed several surprising synergistic interactions involving Apc, many gene pairs showed diminishing returns, which could be the result of several factors. There is likely some degree of overlap in the pathways through which these tumor suppressor genes constrain tumor growth. For instance, Nf1 and Pten, which had the strongest negative epistatic interaction in our data, are involved in RAS and PI3K signaling, respectively, two very closely linked pathways regulating cell proliferation and survival (Shaw and Cantley 2006; Krygowska and Castellano 2018; Cuesta et al. 2021; Nussinov et al. 2021). Similarly, Pten and the RAS suppressor Rasa1 show negative epistasis, as do Rasa1 and Nf1. Redundancy between or within pathways could therefore result in diminishing returns on tumor fitness. Furthermore, there is likely an upper limit to the degree to which cell proliferation can be increased through gene inactivation within the native environment. These tumors express oncogenic Kras^G12D and thus have sufficient RAS pathway activation to drive proliferation. As a result, they could have relatively little room to increase their proliferation even further compared to tumors without such a strong oncogenic driver. Concordantly, we previously showed that the combined inactivation of Nf1, Rasa1, and Pten was highly synergistic in oncogene-negative lung adenocarcinoma, showing that these interactions can have positive or negative epistasis depending on the underlying oncogenic context (Yousefi et al. 2022, 2023).

Apc inactivation stands out for its ability to synergize with inactivation of other tumor suppressor genes to promote tumor growth. Apc gene pairs represent the four highest observed positive epistasis scores, including the only significantly synergistic combinations aside from the well-established interaction between p53 and Rb1. The only significantly buffering Apc interaction was Apc–Rnf43, which is understandable given that Apc and Rnf43 both regulate the Wnt pathway and likely have redundant effects on tumorigenesis (van de Wetering et al. 2015; Bond et al. 2016; Yaeger et al. 2018). It is unknown why Apc in particular is a great cooperator, which underscores the importance of in vivo interrogation of genetic interactions.

It is worth noting that the negative epistasis we observed between tumor suppressor genes during tumorigenesis is largely not sign epistasis, in which an additional mutation reduces absolute tumor fitness compared to a single mutant, as was previously observed for inactivation of Lkb1 or Setd2 on an oncogenic EGFR background (Foggetti et al. 2021). The double mutants that we investigated nearly all increased overall tumor fitness, and thus mutations in these genes would still be expected to co-occur to varying degrees in human lung adenocarcinomas. The only exception was Apc–Rnf43, which was the only occurrence of sign epistasis. Observing one instance of sign epistasis out of the 45 pairs tested suggests a surprisingly accessible fitness landscape, as opposed to the rugged landscapes found in other evolutionary systems (Weinreich et al. 2006; Karageorgi et al. 2019). It remains to be seen if this pattern would change if we were to examine more complex genotypes (e.g. triple instead of double mutants), mutations with more modest fitness effects, mutations less frequently observed in human tumors, or different oncogenic driver backgrounds (Yousefi et al. 2023).

Our results imply that the functional impact of genetic interactions on tumor fitness cannot be inferred solely from tumor mutational data. Indeed, there are a number of confounding factors in uncovering genetic interactions from human data alone, including tumor subtype-specific mutations, variable tumor mutational burden across samples, as well as differing therapies and therapy resistance mechanisms (van de Haar et al. 2019). Even our in vivo system requires high gene inactivation efficiency and a sufficient number of tumors of each genotype to tease out the fitness impact of inactivating each gene pair.

The potential future applications of this platform in investigating tumorigenesis are multifold. It can be immediately applied to understand the effect of larger, more complete, pairwise combinations of tumor suppressors in lung adenocarcinoma. Lentiviral vectors with pairs of sgRNAs could also be combined with floxed alleles to create tumors with combinatorial inactivation of three genes, which may even better approximate the genetic makeup of human lung tumors. These quantitative and combinatorial approaches could be combined with CRISPRa, CRIPSRi, or Cas9-based methods for epigenetic modulation to increase or decrease the expression of pairs of genes in vivo (Winters et al. 2018; Nuñez et al. 2021). Cas12a-based somatic genome editing could also facilitate the generation of large numbers of complex genotypes with cancer models (Campa et al. 2019; Dong et al. 2023; Hebert et al. 2024). Broadly, these approaches can identify critical genetic interactions that could not have been predicted a priori.

Moreover, this platform is not limited to lung adenocarcinoma and could be adapted to study other cancer types that can be initiated with viral vectors, including pancreatic cancer, bladder cancer, sarcoma, breast cancer, and prostate cancer (Kersten et al. 2017; Hebert et al. 2023). This method could also contribute to the investigation of other aspects of tumor growth, progression, and metastasis. We defined epistasis effects based on tumor fitness, which is a metric derived largely from tumor size. However, given our data-rich readout, it should be possible to describe genotypes based on other phenotypes, such as tumor initiation rate and the frequency of rare large tumors (Cai et al. 2021). Creating more sophisticated mathematical models for tumor growth and epistasis across these measures might offer new insights. Finally, combinatorial inactivation of tumor suppressor genes and drug targets could uncover genotype-specific therapeutic effects, which could ultimately lead to more personalized patient treatments (Han et al. 2017).

These experiments have established and validated a platform for the rapid and quantitative analysis of the effect of combinatorial genetic alterations on tumor growth in vivo. Our results suggest a paradigm in which loss of many tumor suppressor genes has diminishing returns on tumor fitness, but that the fitness landscape of lung tumorigenesis is highly accessible overall. These approaches have the potential to accelerate our understanding of the combinatorial driving forces of cancer evolution.

Methods

Generation of Barcoded Dual-sgRNA Lentiviral Vector Libraries

To generate barcoded vector backbones, a 98 bp oligo containing a 16-nucleotide degenerate barcode (BC) and two BsmBI Type IIS restriction enzyme sites was cloned into the 3′ end of the bovine U6 promoter using Gibson Assembly (NEBuilder HiFi, NEB) (see supplementary fig. S1a, Supplementary Material online), in a vector also containing a tracrRNA (for guide 2) and PGK-Cre recombinase (Tang et al. 2024). To create pooled dual-sgRNA inserts, designed combinations of guides were ordered as a single-stranded DNA pool (Twist Biosciences) of 138 bp oligos, each containing two guide spacers flanked by BsmBI and BbsI Type IIS restriction enzyme sites (all oligo sequences in supplementary table S1, Supplementary Material online). Following low-cycle PCR amplification, these double guide oligos were ligated via Golden Gate Assembly (using BbsI) with a donor sequence containing a tracrRNA (for guide 1) and the mouse U6 promoter, resulting in a 528 bp circular intermediate. After PCR amplification to create linearized insert sequences, the inserts were cloned via Golden Gate Assembly (using BsmBI) into the barcoded vector backbone to yield the final pool of barcoded dual-sgRNA vectors (Lenti-U6^BC-sgDbl/Cre). This pooled product was then electroporated into competent Escherichia coli (C3020K, New England Biosciences) and plated onto LB-ampicillin plates. To ensure sufficient barcode diversity for Tuba-seq^Ultra sequencing and delineation of tumors, ∼10⁶ colonies were collected for the Lenti-U6^BC-sgDbl/Cre pool. This cloning procedure thereby enables the generation of lentiviral vectors containing any desired combinations of two sgRNAs in a highly multiplexed manner.

We observed some uncoupling of guides during vector library generation, resulting in the formation of vectors with unintended guide combinations. This uncoupling most likely occurred during PCR amplification of the circular intermediate and could thereby be averted by eliminating this PCR step, instead proceeding directly to BsmBI-mediated Golden Gate Assembly using a sufficient amount of the circular intermediate itself.

Lentiviral Packaging

To generate lentivirus, Lenti-U6^BC-sgDbl/Cre vectors were transfected as a pool into 293T cells with pCMV-VSV-G (Addgene #8454) envelope plasmid and pCMV-dR8.2 dvpr (Addgene #8455) packaging plasmid using polyethylenimine (Polysciences). Sodium butyrate (Sigma Aldrich, B5887) was added 8 h after transfection to a final concentration of 20 mM to inhibit silencing of viral expression, and medium was changed after 24 h. Supernatants were collected 36 and 48 h after transfection, filtered with 0.45 µm syringe filters (Millipore, SLHP033RB) to remove cells and cell debris, concentrated by ultracentrifugation (25,000 × g for 1.5 h at 4 °C) and resuspended in phosphate-buffered saline (PBS) overnight. Each virus was titered against a known standard using LSL-YFP Mouse Embryonic Fibroblasts (MEFs) (a gift from Dr. Alejandro Sweet-Cordero/UCSF). Lentiviral vector aliquots were stored at −80 °C and were thawed and pooled immediately prior to delivery to mice.

Animal Studies

The use of mice for this study was approved by the Institutional Animal Care and Use Committee at Stanford University, protocol number 26696. Kras^LSL−G12D/+ (RRID:IMSR_JAX:008179) (Jackson et al. 2001), R26^{LSL−tdTomato} (RRID:IMSR_JAX:007914) (Madisen et al. 2010), H11^LSL−Cas9 (RRID:IMSR_JAX:027632) (Chiou et al. 2015), and H11^Cas9 (RRID:IMSR_JAX:006054) (Chiou et al. 2015) mice have been previously described. All mice were on a pure C57BL/6 background.

Tumors were initiated by intratracheal delivery of 60 μl of resuspended lentiviral vectors in PBS. Lentiviral titer and time of tumor development are indicated in each figure.

Tumor Barcode Sequencing and Analysis

Tuba-seq^Ultra libraries were generated by isolating genomic DNA from bulk tumor-bearing lung tissue, followed by PCR amplification of the BC-sgRNA1–sgRNA2 region from 32 μg of bulk lung genomic DNA using Q5 Ultra II High-Fidelity 2× Master Mix (New England Biolabs, M0494X). Unique dual-indexed primers were used to amplify each sample followed by purification using Agencourt AMPure XP beads (Beckman Coulter, A63881). The libraries were pooled based on lung weights to ensure even read depth and sequenced (read length 2 × 150 bp) on the Illumina NextSeq 6000 platform. Tuba-seq analysis of tumor barcode reads was performed as previously described (Rogers et al. 2017; Cai et al. 2021), except that sgRNAs were first directly sequenced and their sequences matched with our intended target genes and guide versions (e.g. Apc-V1). At library preparation, a template switching event might happen between the two sgRNA regions. This can create chimeric sequences, groups of BC-sgRNA1–sgRNA2 regions that are created from two distinct tumors via PCR template switching. The resulting sequencing reads of groups of BC-sgRNA1–sgRNA2 do not correspond to any real tumor. With the diverse barcode region and a matching sgRNA1 sequence close to the DNA-barcode, we can easily remove these PCR artifacts by simply removing the “tumors” with the lower read count from all such pairs.

Calculation of Relative Fitness of Each Genotype of Tumors

Fitness of each tumor genotype was approximated by the growth rate of tumors. Fitness was calculated for each tumor genotype in a given mouse strain by considering the expected number of cells with each lentiviral vector at the start of the experiment (#start_{genotype_mouse-strain}) and number of neoplastic cells with each lentiviral vector after tumor growth at the end of the experiment (#end_{genotype_mouse-strain}).

$$\mathrm{Fitnes}{\mathrm{s}}_{\mathrm{genotyp}{\mathrm{e}}_{\mathrm{mouse}\mathrm{strain}}}=\mathrm{lo}{\mathrm{g}}_2\left(\frac{\mathrm{\#}\mathrm{en}{\mathrm{d}}_{\mathrm{genotyp}{\mathrm{e}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}}{\mathrm{\#}\mathrm{star}{\mathrm{t}}_{\mathrm{genotyp}{\mathrm{e}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}}\right).$$

#end_{genotype_mouse-strain} is the sum of all neoplastic cells in all tumors of that genotype at the end of the experiment from all mice of a given strain.

$$\mathrm{\#}\mathrm{en}{\mathrm{d}}_{\mathrm{genotyp}{\mathrm{e}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}=\mathrm{sum}(\mathrm{neoplastic}\mathrm{cell}{\mathrm{\#}}_{\mathrm{genotyp}{\mathrm{e}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}).$$

#start_{genotype_mouse-strain} is the expected number of cells of that genotype at the start of the experiment. #start_{genotype_mouse-strain} was determined from the number of tumors with each viral vector (titer_genotype) from KT Cas9-negative control mice. In KT mice, the number of tumors with each vector represents the exact titer of each vector (titer_genotype) in the lentiviral pool. Data from KT mice were used to calculate the relative titer across genotypes. To put #start_genotype in the context of the most potent genotype (resulting in the most tumors per titer units) in the given mouse strain, we normalized the number of tumors with each lentiviral vector in KT mice (titer_genotype) to the number of Nf1;Pten double mutant tumors in KT mice (titer_Nf1;Pten) and then multiplied this by the total number of Nf1;Pten double mutant tumors across all mice of a given strain (TumorNumber_{Nf1;Pten_mouse-strain}). Thus, #start_{genotype_mouse-strain} represents the titer-corrected expected number of tumors of that genotype if that genotype were as potent as double Nf1;Pten mutation in that strain. As an example, #start for single mutant genotype Apc in KTC mice (#start_{Apc_KTC}) was calculated as:

$$\mathrm{\#}\mathrm{star}{\mathrm{t}}_{Ap{c}_{KTC}}={\displaystyle \frac{\mathrm{tite}{\mathrm{r}}_{Apc}}{\mathrm{tite}{\mathrm{r}}_{Nf1;Pten}}\times \mathrm{TumorNumbe}{\mathrm{r}}_{Nf1;Pte{n}_{KTC}}.}$$

Finally, we calculated “Relative Fitness” within each mouse strain by normalizing the fitness of each genotype to the fitness of tumors with the inert sgRNAs in the lentiviral vector.

$$\mathrm{relative}\mathrm{fitnes}{\mathrm{s}}_{{\mathrm{genotype}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}={\displaystyle \frac{\mathrm{lo}{\mathrm{g}}_2(\mathrm{\#}\mathrm{en}{\mathrm{d}}_{{\mathrm{genotype}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}/\mathrm{\#}\mathrm{star}{\mathrm{t}}_{{\mathrm{genotype}}_{\mathrm{mouse}\text{-}\mathrm{strain}}})}{\mathrm{lo}{\mathrm{g}}_2(\mathrm{\#}\mathrm{en}{\mathrm{d}}_{{\mathrm{inert}}_{\mathrm{mouse}\text{-}\mathrm{strain}}}/\mathrm{\#}\mathrm{star}{\mathrm{t}}_{{\mathrm{inert}}_{\mathrm{mouse}\text{-}\mathrm{strain}}})}.}$$

The above defined measure is a commonly used fitness definition (Lenski et al. 1991; Remold and Lenski 2001; Melnyk et al. 2015; Han et al. 2017; Alonso-del Valle et al. 2021).

Statistical Analysis

All statistical analyses were performed using the R software environment. P values and 95% confidence intervals (represented by whiskers) were calculated using bootstrap resampling (10,000 repetitions). Bootstrapping was done by random resampling with replacement of all the tumors in all of the mice of a given strain. For Figs. 2 and 4c to f and supplementary figs. S3 and S6, Supplementary Material online, a nested bootstrap approach was used where first all the mice from a given strain were resampled with replacement, and then, all tumors in this virtual mouse cohort were resampled again with replacement.

Supplementary Material

Supplementary material is available at Molecular Biology and Evolution online.

Data Availability

All data generated from this study have been deposited in the NCBI's Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE281604).