Smarca4 Inactivation Promotes Lineage-Specific Transformation and Early Metastatic Features in the Lung

Genes encoding for components of the mammalian ATP-dependent chromatin remodeling complex SWI/SNF (also known as BAF) are among the most commonly mutated targets in cancer (1). However, their exact contributions to tumorigenesis are not well understood in many cancer types. Such lack of understanding reflects the complexity of SWI/SNF function due to its cell type–specific roles and the heterogeneity of SWI/SNF complexes within a cell at a given time (2). Previous studies have shown highly context-specific roles for SWI/SNF in tumor progression (3, 4). These studies strongly emphasize the need for a deeper mechanistic understanding of the impact of precise SWI/SNF mutations on the complex's function and tumor cell biology to devise effective therapeutic strategies tailored to specific SWI/SNF mutations and tumor types.

SMARCA4 (BRG1) encodes for one of two mutually exclusive ATPases of SWI/SNF complexes (5) and is among the most frequently mutated genes (6) in non–small cell lung cancer (NSCLC), occurring at a frequency of 10% (7, 8). NSCLCs harboring SMARCA4 mutations are predominantly of the lung adenocarcinoma (LUAD) subtype (8). Among SMARCA4 mutations, truncating and missense mutations are the most prevalent, and these can be monoallelic or biallelic (7–9). Of these, inactivating SMARCA4 alterations that result in the complete absence of SMARCA4 protein expression, such as truncating mutations, are associated with the poorest outcomes in patient survival (8, 9).

Previous studies in mouse models have examined the functional consequences of SMARCA4 inactivation on lung cancer progression. In a carcinogen-induced model of lung cancer, loss of one allele of Smarca4 at tumor initiation promoted tumorigenesis, whereas loss of both alleles had no detectable effect (10). Interestingly, two independent studies using a genetically engineered mouse model (GEMM) of Kras-driven LUAD produced conflicting findings, with loss of Smarca4 resulting in a negative or positive effect on tumorigenesis (11, 12). These results raise questions about the tumor-suppressive functions of SMARCA4 in the lung, and whether SMARCA4-inactivating mutations and the loss of its protein expression observed in human patients confer a functional advantage in tumor initiation or progression. Furthermore, despite SMARCA4's well-described role as a core catalytic component of the SWI/SNF chromatin remodeling complex, studies investigating its functions in chromatin regulation in lung cancers have been limited to NSCLC cell lines (13–15). As such, the direct consequences of SMARCA4 inactivation on SWI/SNF function on chromatin regulation during lung cancer evolution are unknown.

Here, we address the impact of SMARCA4 inactivation on tumor initiation and progression, chromatin accessibility, and SWI/SNF function in LUAD using a combination of GEMMs, patient-derived xenograft (PDX) models, and epigenomic profiling. We demonstrate a tumor-suppressive function for Smarca4 that is dependent on the cell type in which the mutation occurs. We further identify transcription factor (TF) programs altered in the context of SMARCA4 loss. In particular, our studies reveal that SMARCA4-deficient tumors effectively lose lung lineage TF activities and harbor features of dedifferentiation, reminiscent of a metastatic cell state. We further determine the underlying mechanism behind these aberrant cell states as a consequence of altered SWI/SNF function upon Smarca4 inactivation. Collectively, this work provides key insights into SMARCA4 function in tumor initiation and progression, chromatin state, and SWI/SNF function in lung cancer. More broadly, our data have implications for understanding the patterns of cancer-associated mutations in subtypes of human cancer.

To determine the impact of SMARCA4 loss on tumor initiation and progression, we first sought to model SMARCA4 inactivation in a defined and relevant genetic system. KRAS is the most frequent oncogene comutated with SMARCA4 (35%), in contrast to EGFR, which tends to be mutated in SMARCA4 wild-type lung cancers (7, 8). Mutations in the tumor suppressor TP53 also occur at a high frequency (56%) among SMARCA4-mutant tumors (8). Given the spectrum of these co-occurring mutations, we crossed a floxed allele of Smarca4 (16) into a well-characterized mouse model (17) of LUAD [Kras^LSL-G12D/+; Trp53^fl/fl (KP); Fig. 1A]. In this model, concomitant activation of oncogenic Kras, deletion of the tumor suppressor Trp53, and deletion of exons encoding for the ATPase domain of Smarca4 occur in the lungs of mice upon intratracheal delivery of adenoviral Cre recombinase. For these experiments, we used adenoviruses in which Cre expression is driven by the Sftpc [surfactant-associated protein C (SPC)] promoter (18), the activity of which is observed in alveolar type II (AT2) cells—one of the presumed cells-of-origin of LUADs (19).

The KP mouse model recapitulates the full cascade of LUAD development (17). Tumor-initiating cells infected with adenoviral Cre undergo hyperplasia, progression to adenomas, and finally progression to adenocarcinomas, which have the ability to metastasize to local and distal sites. We assessed the impact of Smarca4 loss on tumorigenesis 17 weeks after tumor initiation using various metrics: tumor number, tumor burden, tumor grade, and metastatic incidence.

We observed no differences in the number of tumors among the three genotypes (Fig. 1B). A selection against full Smarca4 loss was evident, as shown by a decrease in overall tumor burden in Kras^LSL-G12D/+; Trp53^fl/fl; Smarca4^fl/fl (KPS) mice compared with those with wild-type Smarca4 (KP) or reduced Smarca4 gene dosage [Kras^LSL-G12D/+; Trp53^fl/fl; Smarca4^fl/+ (KPS-HET); Fig. 1C and D]. Histologic examination of SMARCA4 protein expression in the lungs of KPS mice revealed that a considerable fraction of tumors (14%–35%) across all animals retained expression (Supplementary Fig. S1A and S1B). We used laser-capture microdissection to isolate SMARCA4-positive tumors from KPS mice, performed genotyping PCR, and detected both the floxed and recombined alleles in all cases (Supplementary Fig. S1C). Tumors with retained SMARCA4 expression as a result were classified as recombination escapers and were heterozygous for Smarca4 loss. Despite a substantial fraction of tumors in KPS animals retaining SMARCA4 expression, we did detect tumors clearly absent of SMARCA4 staining. These tended to be smaller in size and were associated with decreased proliferation (measured by Ki-67 staining) compared with their SMARCA4-positive counterparts (Supplementary Fig. S1D and S1E). To estimate the distribution of tumor grades in these animals in an unbiased fashion, we applied a deep-learning algorithm based on well-established criteria (17, 20) that histologically classify KP tumors by tumor grade to the lesions in the lungs of these animals. We found an increase in the fraction of early lesions (grades 1–2) and a decrease in more advanced grade 3 tumors in KPS mice compared with KP and KPS-HET animals (Fig. 1E). Altogether, these data indicate that full Smarca4 inactivation restrains tumor progression in the vast majority of tumors initiated from SPC⁺ cells.

Closer inspection of tumor-bearing lungs of KPS mice, however, revealed that these animals had the highest fraction of the most advanced type of lesions (grade 4), despite their markedly decreased overall tumor burden (Fig. 1E). Furthermore, we observed increased frequency of metastases to the thymus and lymph node among KPS and KPS-HET mice that was higher than in KP animals (Fig. 1F). Strikingly, grade 4 lesions and metastases from KPS animals universally lacked SMARCA4 protein expression (Fig. 1G and H)—indicating a strong selection for full SMARCA4 loss in these highly advanced tumors and metastases.

Collectively, these results point toward a paradoxical role for Smarca4 in tumor suppression. Although Smarca4 loss inhibits tumor progression in a large fraction of tumors initiated by SPC-Cre, a subset of SMARCA4-deficient transformed cells can give rise to highly advanced and metastatic tumors.

Given the role of SMARCA4 in chromatin remodeling, we hypothesized that Smarca4 inactivation directly alters distinct TF programs—perhaps in a cell type–specific fashion—to affect the final tumorigenic outcome. As a catalytic subunit of SWI/SNF, SMARCA4 has a key role in the ability of the complex to regulate nucleosome positioning and chromatin accessibility. Such accessibility is crucial for TF binding to regulatory elements in order to specify gene-expression programs that dictate cell state. To address this hypothesis and investigate the heterogeneity among Smarca4-deficient tumors in vivo, we performed the single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) on isolated cancer cells from KP, KPS-HET, and KPS animals (Fig. 2A). We performed these experiments on moribund animals to better capture the cell states spanning tumor progression—including those of high-grade tumors and metastasis cells—from all three SMARCA4 genotypes. In total, we generated chromatin accessibility profiles from 25,229 cells. These include 21,780 cells from the tumor-bearing lungs of three animals per genotype and 3,449 cells from metastatic sites (thymi and lymph nodes) of at least two animals per genotype (Supplementary Fig. S2A–S2C).

As shown in Fig. 2B and C, chromatin accessibility profiles from KP and KPS-HET primary tumors formed a continuum of states, whereas metastasis cells clustered separately, as previously reported and characterized (20). By contrast, cancer cells isolated from the lungs of KPS animals generated strikingly distinct clusters composed almost exclusively of cells of this genotype, indicating unique epigenetic states in these cells (Fig. 2B and C). Importantly, KPS-specific clusters were clearly reproduced using two distinct dimensionality reduction methods (Fig. 2B and C; Supplementary Fig. S2D and S2E). A fraction of KPS cells belonged to clusters predominantly composed of KP and KPS-HET cells, likely reflecting the presence of SMARCA4-positive tumors in the lungs of KPS animals. A small number of KPS-HET cells were also found in KPS clusters, which we attribute to a fraction of KPS-HET tumors harboring cancer cells absent of SMARCA4 protein expression (Supplementary Fig. S2F). We identified TF motifs uniquely marking 21 clusters identified through the Louvain modularity method (21) and performed hierarchical clustering (Fig. 2D). We annotated the clusters based on their sample composition and chromatin accessibility profile in relation to features of early and advanced tumors previously described in this model (20, 22): SPC KPS (composed almost exclusively of KPS cells; clusters 1–4), early SPC KP (Nkx2-1 high; clusters 5–11), late SPC KP (Nkx2-1 low; clusters 12–15), and SPC metastases (composed of metastasis cells from all three genotypes; clusters 16–21).

Tumor progression in the KP model is characterized by key epigenetic state transitions and loss of cell identity as cancer cells evolve toward an advanced state (20, 22–24). Strikingly, SPC KPS clusters displayed characteristic features of SPC metastasis clusters, despite being composed almost exclusively of primary tumor cells (Fig. 2D–F). Activities for the lung lineage TFs, Nkx2-1/Ttf1 and Gata6, were markedly absent in SPC KPS clusters, demonstrating a lack of lung lineage cell identity among these primary tumor cells. A fraction of cells from SPC KPS clusters were also highly enriched for peaks associated with TFs marking late tumors or metastatic cells, such as Runx2, Sox2, and Sox9 (20, 25). Importantly, the activities of these TF programs were markedly higher in KPS cells compared with advanced KP or KPS-HET cells within late SPC KP clusters. Furthermore, decreased accessibility in motifs of the repressor Zeb1 indicated that a subset of KPS cells were undergoing epithelial–mesenchymal transition (26)—consistent with a metastatic-like phenotype. In contrast to cells isolated from KPS primary tumors, we did not detect clear TF activities that distinguished cancer cells isolated from metastases in KPS mice from those isolated from metastases in KP and KPS-HET animals (Fig. 2D). Instead, KPS metastasis cells generated multiple clusters and displayed variable accessibilities for prometastatic TF programs (Runx2, Onecut2, Sox2, and Sox9), suggesting heterogeneous routes to metastases in the context of Smarca4 deficiency (Fig. 2D and E; Supplementary Fig. S2G).

Although SPC KPS clusters had features consistent with an advanced cancer cell state, cells from these clusters were depleted for peaks associated with AP-1 TF family motifs (Fos and Jun, among others), in stark contrast to late SPC KP and SPC metastasis clusters (Fig. 2D; Supplementary Fig. S2G). SWI/SNF complexes bind directly to AP-1 TF motifs, and members of this TF family have been shown to be important modulators of enhancer selection (27). Depletion of AP-1 motif accessibility may be attributed to the abrogation of a direct interaction between the AP-1 TF family member JUNB and SWI/SNF complexes upon Smarca4 deficiency, which we have previously shown through quantitative mass spectrometry (28). Interestingly, metastasis-derived cells isolated from KPS animals were enriched for peaks with AP-1 TF motifs, suggesting that a gain in the activities of these TF programs may be a key event selected for in the transition of KPS primary tumors to a fully metastatic cell state. Given that the absence of AP-1 activity distinguishes KPS clusters from SPC metastasis clusters, the gain in AP-1 motif accessibility as cancer cells from KPS primary tumors transition into a premetastatic/metastatic cell state results in KPS metastasis cells that cluster with those isolated from KP and KPS-HET animals (Fig. 2D).

Altogether, these data show that Smarca4 inactivation in tumor-initiating cells leads to cancer cells with distinct cell states that largely recapitulate metastatic cell states. In contrast to cancer cells from KP and KPS-HET primary tumors, which undergo similar epigenetic state transitions and a gradual loss of lineage fidelity throughout tumor evolution, those from KPS primary tumors are characterized by a general lack of lung lineage specificity and the robust activation of TF programs associated with metastases in a subset of cells. Thus, we hypothesized that Smarca4 inactivation in certain tumor-initiating cells facilitates the rapid acquisition of a metastatic-like cell state.

We next sought to understand the heterogeneity of SMARCA4-deficient tumors in KPS animals. Notably, SMARCA4-negative tumors in this model tended to be either of low tumor grade or highly advanced. In particular, we sought to identify determinants of high-grade tumors in this context. These tended to be located by the airways of the lungs of KPS animals—in sharp contrast to high-grade tumors from KP and KPS-HET mice, which were predominantly located in the alveolar spaces (Supplementary Fig. S2H). We hypothesized that these highly advanced tumors arose from an atypical cell-of-origin within the bronchioles or the bronchioalveolar duct junction that is uniquely sensitive to transformation upon SMARCA4 loss.

Although independent studies have demonstrated that SPC⁺ AT2 cells are the predominant cell-of-origin in the Kras-driven GEMM of LUAD (19, 29), other cell types, including the club cell secretory protein-positive (CCSP⁺) bronchiolar epithelial club cell (30) and the SPC⁺CCSP⁺ double-positive bronchioalveolar stem cell (BASC; ref. 31), may have this ability given the proper context. Importantly, BASCs are able to regenerate multiple cell lineages, including AT2 and club cells upon lung injury (32, 33). Among human LUAD patients, both AT2 and club cells are hypothesized to be the cell-of-origin of LUAD, as these tumors tend to express markers of these lineages (34).

To investigate the role of the cell-of-origin in our data, we scored each cell in the scATAC-seq data set for AT2 and club cell identities using the promoter and gene body accessibilities (also referred to as gene scores; ref. 20) of a set of marker genes for each cell type identified from single-cell RNA sequencing (scRNA-seq) of the developing lung (35). The AT2 signature was enriched in early SPC KP clusters and depleted in late SPC KP, SPC KPS, and SPC metastasis clusters (P < 2.2 × 10⁻¹⁶; Fig. 2G). Consistent with a distinct cell-of-origin giving rise to advanced SPC KPS tumors, KPS clusters displayed a strikingly specific enrichment of the club cell signature, which was completely absent in clusters largely comprised of primary tumors from KP and KPS-HET animals (P < 2.2 × 10⁻¹⁶; Fig. 2H).

Cells within SPC KPS clusters, therefore, have a club cell state. These results point toward the club cell state potentially being highly sensitive to malignant transformation and rapid tumor progression in the absence of Smarca4. Such a state may be achieved through Kras-driven transformation of SPC⁺CCSP⁺ BASCs in this model, since these cells could adopt a club cell state (32, 33) upon tumor initiation by SPC-Cre.

To formally test whether the club cell state is sensitive to transformation in the setting of Smarca4 inactivation, we initiated tumors in the lungs of KP, KPS-HET, and KPS animals using adenoviruses in which Cre expression was driven by the Scgb1a1 (CCSP) promoter (Fig. 3A), which is predominantly active in the club cell population (18).

In striking contrast to KPS animals in which tumors were initiated using SPC-Cre, these KPS animals displayed a significant increase in the number of tumors per mouse compared with KP (1.86×) and KPS-HET (1.72×) animals (Fig. 3B). They also exhibited a trend toward increased tumor burden compared with their Smarca4 wild-type and heterozygous counterparts (Fig. 3C and D). Histologic examination of these tumors for SMARCA4 expression revealed that the vast majority of tumors (86%–96%) within the lungs of these KPS animals were SMARCA4-negative (Supplementary Fig. S3A and S3B), in contrast to the tumors arising from SPC-expressing cells in KPS animals (Supplementary Fig. S1A and S1B). These KPS animals were also consistently enriched for higher-grade tumors (grades 3–4; Fig. 3E), and all metastatic lesions in this cohort were identified exclusively in KPS animals (Fig. 3F), some of which displayed multiple metastatic lesions within a single tissue (Supplementary Fig. S3C). Importantly, all high-grade tumors and metastases identified in KPS animals lacked SMARCA4 protein expression (Supplementary Fig. S3D and S3E). Of note, we detected SMARCA4-negative cells within some tumors in the majority of KPS-HET animals infected with the CCSP-Cre virus (Supplementary Fig. S3F and S3G). Consistent with a potent tumor-suppressive role for SMARCA4 in LUAD, these KPS animals rapidly succumbed to disease and displayed decreased overall survival compared with KP animals (Fig. 3G). Thus, Smarca4 inactivation in CCSP⁺ tumor-initiating cells promotes tumor progression at multiple stages of lung tumorigenesis.

Collectively, these data demonstrate a cell type–specific role for SMARCA4 in tumor initiation and progression in the lung. These models show that the tumor-suppressive function of SMARCA4 is influenced by the tumor cell-of-origin. Although SMARCA4 loss inhibits tumor progression in the vast majority of transformed SPC⁺ cells, predominantly AT2 cells, it promotes malignant transformation and consistently accelerates tumor progression in transformed CCSP⁺ cells, predominantly club cells.

We next sought to determine whether the epigenetic states of KPS clusters we previously identified by scATAC-seq are driven by an alternative cell-of-origin for KPS tumors in the SPC KP model or are a direct effect of Smarca4 loss-of-function. To address this, we performed scATAC-seq on sorted cancer cells arising from CCSP-expressing cells from KP, KPS-HET, and KPS animals (Fig. 4A).

We generated chromatin accessibility profiles from 16,321 cells isolated from the tumor-bearing lungs of two animals per genotype (Supplementary Fig. S4A–S4C). Similar to SPC-Cre–initiated tumors, KP and KPS-HET primary tumors formed a continuum of epigenetic states, whereas KPS primary tumors displayed distinct states represented by clusters composed exclusively of cells from KPS animals (Fig. 4B and C).

We uncovered 15 distinct clusters, identified differential TF motifs across the data set, and performed hierarchical clustering (Fig. 4C and D). We annotated these clusters as early CCSP KP (Nkx2-1 high; clusters 10–15), late CCSP KP (Nkx2-1 low; clusters 1–3), and CCSP KPS (composed almost exclusively of KPS cells; clusters 4, 5, 7–9) based on their chromatin accessibility profile and sample composition (Fig. 4C and D). Of note, we could detect enrichment of Fox motifs (Foxa1 and Foxc1, among others) in cluster 10, an early CCSP KP cluster, representing normal club cells and early transformed club cells. Fox motifs are among the most enriched motifs distinguishing normal club cells from normal AT2 cells (Supplementary Fig. S4D).

We also detected epigenetic state transitions and loss of cell identity as KP and KPS-HET tumor cells from this model progressed toward an advanced state, consistent with those observed in tumors arising from SPC-expressing cells in KP and KPS-HET animals. For example, a gain in the activity of the AP-1 TF family accompanied the loss of Nkx2-1 activity that demarcates early CCSP KP from late CCSP KP (Fig. 4D and E; Supplementary Fig. S4E). Additionally, we could detect reduced lung lineage TF activities (Nkx, Cebp, and Gata families) in cells belonging to late CCSP KP clusters compared with those in early CCSP KP clusters.

By contrast, cells in CCSP KPS clusters were almost entirely depleted for peaks harboring these lung lineage TF motifs (Fig. 4D and E; Supplementary Fig. S4E). We detected low activity for Nkx2-1 and Cebpa in cluster 9, representing a KPS cell state early in tumor progression, whereas programs associated with more advanced tumors (Runx2, Sox2, and Sox9) were activated in another subset of KPS cells and are reflective of a KPS cell state late in tumor progression (Fig. 4D–F). Thus, the marked reduction of lung lineage TF activities and activation of protumorigenic programs in KPS cells are consistent features of Smarca4-deficient primary tumors.

Similar to KPS primary tumors initiated from SPC-expressing cells, KPS cancer cells initiated from CCSP-expressing cells were also depleted for AP-1 TF family motifs (Fig. 4D; Supplementary Fig. S4E). However, a subset of cells in late CCSP KPS clusters had high AP-1 TF activity and were also enriched for programs defining a metastatic cell state, most notably Runx2 (Fig. 4D and E; Supplementary Fig. S4E). Thus, a gain in AP-1 TF activity in Smarca4-deficient cells occurs in primary tumors transitioning into a metastatic cell state. Gain in AP-1 TF activity in Smarca4-deficient cells was also observed in metastases from SPC-Cre–infected KPS animals (Supplementary Fig. S2G). Of note, Ctcf and Irf/Stat motifs displayed high accessibility in KPS clusters (Fig. 2D and 4D; Supplementary Fig. S4E).

Our results, thus far, indicate that complete loss of Smarca4 during tumor initiation results in distinct epigenetic states in KPS primary tumors throughout tumor evolution. To determine whether this distinction is maintained at the transcriptional level, we performed bulk RNA-seq on sorted cancer cells from the lungs of moribund tumor-bearing KP, KPS-HET, and KPS animals. Both unsupervised hierarchical clustering and principal component analysis (PCA) revealed that KPS samples clearly separated from KP and KPS-HET samples (Supplementary Fig. S4F; Fig. 4G), indicating a unique transcriptional profile associated specifically with complete Smarca4 loss.

To understand the gene-expression programs characterizing KPS primary tumors, we next performed pairwise analysis to determine the most significant differentially expressed (DE) genes between KPS and KP primary tumors (Fig. 4H). Top genes increased (adjusted P < 0.05, |FC| > 1.5) in KPS primary tumors were enriched for genes characterizing the inflammatory response, whereas those decreased (adjusted P < 0.05, |FC| > 1.5) were enriched for genes characterizing xenobiotic metabolism (Supplementary Fig. S4G). Genes constituting a classic epithelial–mesenchymal transition signature were represented in both directions (Supplementary Fig. S4G). Interestingly, the oxidative phosphorylation signature, which was previously found to be the most significant gene set increased in KPS primary tumors (12), was not enriched in our analysis of KPS primary tumors (P = 0.56; FDR = 0.67).

We next sought to understand the extent to which the transcriptional profiles of KPS primary tumors correlated with the chromatin states of KPS primary tumors. To this end, we scored both CCSP and SPC scATAC-seq data sets using the mean gene scores of the most significantly increased (KPS UP) and decreased (KPS DOWN) genes in KPS primary tumors compared with those from KP by RNA-seq (Fig. 4I; Supplementary Fig. S4H). Subsets of cells within KPS clusters scored highly for the KPS UP signature compared with cells in other clusters. Likewise, cells within KPS clusters tended to have lower gene scores for genes significantly decreased in KPS primary tumors compared with KP cells in the data set. Furthermore, top genes in either direction (P < 0.00001) displayed similar directionalities in mean gene scores in CCSP KPS samples (Supplementary Fig. S4I). Of note, we could detect one KPS-HET sample that displayed reduced gene scores in top genes decreased in KPS primary tumors. These analyses indicate a degree of correlation between bulk transcript levels and mean gene scores in these data sets.

We next sought to determine the underlying mechanism behind the altered cell states observed in KPS primary tumors. We chose to focus on the absence of lung lineage TF activities in KPS clusters. Primary tumor cells within KPS clusters, including those representing a KPS cell state early in tumor progression, consistently showed reduced accessibilities of lung lineage TF motifs. Loss of lineage specificity in cancer cells can facilitate the acquisition of cell states that support tumor progression. This particular feature of KPS primary tumor cells may explain the acceleration of lung tumorigenesis resulting in the increased incidence of highly advanced tumors and metastases observed in KPS animals upon Smarca4 inactivation.

We hypothesized that the lack of lung lineage TF activities signified by a decrease in their motif accessibilities in Smarca4-deficient cells may be caused by loss of expression of the TFs themselves or altered SWI/SNF function at their binding sites as a consequence of Smarca4 loss. To discriminate between these possibilities, we first compared Nkx2-1 and Gata6 transcript levels of sorted cancer cells from KP, KPS-HET, and KPS primary tumors. We observed a significant decrease in the expression levels of both TFs in KPS primary tumors by bulk RNA-seq (Fig. 5A and B). Reduced transcript levels may indicate a general reduction in expression of these TFs across all KPS tumors or may reflect the increased frequency of grade 4 tumors—which typically lose expression of both proteins—in KPS animals. We next quantified NKX2-1– and GATA6-positive cells in KP, KPS-HET, and KPS primary tumors by IHC. Among lung tumors initiated from SPC⁺ and CCSP⁺ cells in KPS animals, we could detect a striking reduction in NKX2-1 and GATA6 protein expression in SMARCA4-negative grade 4 tumors, as expected. By contrast, these proteins were readily detected in SMARCA4-negative tumors of lower grade (Fig. 5C and D; Supplementary Fig. S5A and S5B). Altogether, these data suggest that loss of NKX2-1 and GATA6 protein expression does not primarily drive the general absence of their activities in KPS clusters.

We next sought to determine the consequences of SMARCA4 loss on the ability of SWI/SNF complexes to bind and open chromatin. SWI/SNF gene products assemble in a combinatorial fashion resulting in three classes of complexes in mammalian cells (36, 37): canonical BAF (cBAF), polybromo-associated BAF (PBAF), and noncanonical or GLTSCR1/GLTSCR1L-associated BAF (ncBAF/GBAF; Fig. 5E). These complexes have both overlapping and unique subunits, as well as binding sites on chromatin (37, 38). As one of two mutually exclusive ATPases that can assemble into all three complexes, SMARCA4 has a central role in SWI/SNF function and directly regulating chromatin accessibility.

To determine the direct effects of Smarca4 loss on chromatin accessibility and SWI/SNF binding in Kras-driven LUAD, we took advantage of isogenic pairs of Smarca4 wild-type and knockout KP cell lines. These lines were generated by transiently expressing Cas9 and a guide RNA for Smarca4 (or control guide) in KP tumor–derived cell lines and screening single-cell clones for complete loss of SMARCA4 expression (28). We performed bulk ATAC-seq and CUT&RUN (C&R) epigenomic profiling of SWI/SNF components in two pairs of isogenic Smarca4 wild-type (SMARCA4-WT; n = 2) and knockout (SMARCA4-KO; n = 2) cell lines generated from two independently derived parental KP cell lines (Fig. 5F). Additionally, we generated genome-wide maps of chromatin features characterizing promoters and enhancers in these cells, as these are the sites predominantly bound by SWI/SNF complexes.

We identified 9,497 differential bulk ATAC-seq peaks between SMARCA4-WT and SMARCA4-KO lines (q < 0.05, LFC > 1; Fig. 5G). Statistically significant differential peaks were overrepresented (1.75×) in the down direction in SMARCA4-KO cells (P = 3.8e−199, hypergeometric test), indicating a general compaction of chromatin upon Smarca4 inactivation. Motifs enriched among differential bulk ATAC-seq peaks between SMARCA4-WT and SMARCA4-KO lines were reminiscent of the clearest motif changes we observed in vivo upon Smarca4 loss (AP-1, Ctcf, among others; Supplementary Fig. S5C). This suggests that these changes are direct effects of Smarca4 inactivation, rather than an indirect effect of transformation or tumor progression. To understand the extent to which these cell lines captured the chromatin states observed in vivo, we next scored each cell in both SPC and CCSP scATAC-seq data sets for correlation to the chromatin accessibility profiles of the four single-cell clones (Supplementary Fig. S5D). Cells from late KP and metastasis clusters generally scored highly for these profiles compared with early KP clusters, consistent with the notion that 2-D cell lines resemble late-stage tumors and metastases. Cells from KPS clusters also scored highly for these profiles, supporting the idea that Smarca4 loss results in an advanced tumor cell phenotype. Importantly, chromatin states of KPS clusters best correlated with the chromatin accessibility profiles of SMARCA4-KO cells, while those of late KP clusters best matched the chromatin accessibility profiles of SMARCA4-WT cells. These analyses suggest that these cell lines are reasonable models to study the direct impact of Smarca4 loss on SWI/SNF binding to chromatin.

We mapped the genome-wide binding profile of pan-SWI/SNF components SMARCA4 and SMARCC1, as well as subunits distinguishing the three complex classes: ARID1A (cBAF), PBRM1 (PBAF), and BRD9 (ncBAF/GBAF) in SMARCA4-WT and SMARCA4-KO cell lines. Additionally, we mapped the binding profile of SMARCA2, the only other catalytic subunit of SWI/SNF, in order to understand any compensatory effects mediated by SMARCA2 that may occur in the absence of SMARCA4. Peaks of the various SWI/SNF components correlated well with one another in varying degrees, suggesting overlapping binding sites among them in Kras-driven Smarca4-WT LUAD (Fig. 5H). Importantly, we detected robust ATAC-seq and SWI/SNF C&R peaks at transcription start sites (TSS) of stably expressed genes in both SMARCA4-WT and SMARCA4-KO cell lines except for SMARCA4 C&R peaks in SMARCA4-KO cells as expected (Supplementary Fig. S5E and S5F), indicating that these are high-quality data sets to examine the direct consequences of Smarca4 loss on SWI/SNF function in chromatin regulation.

Differential ATAC-seq peaks that were reduced upon SMARCA4 loss displayed robust SMARCA4 occupancy in SMARCA4-WT cells, demonstrating that these peaks are direct SMARCA4 binding sites (Fig. 5I). There was a clear loss of binding at these sites of the pan-SWI/SNF component SMARCC1, as well as all three class-specific subunits ARID1A, PBRM1, and BRD9 in SMARCA4-KO cells (Fig. 5J). Of these, ARID1A occupancy displayed the greatest reduction. These changes were accompanied by a loss of primed sites (H3K4me1), as well as enhancer (H3K27ac, H3K4me1) and promoter (H3K27ac, H3K4me3) activities (Fig. 5K). These results are consistent with a model in which SMARCA4 loss results in a defect in the ability of all three major classes of SWI/SNF complexes to bind and open chromatin. Furthermore, the reduction in chromatin accessibility, primed sites, and active regulatory regions appears to be largely a direct effect of the loss of SWI/SNF binding.

We hypothesized that the reduction in chromatin accessibility caused by SMARCA4 inactivation and SWI/SNF loss-of-function occurs at lung lineage motifs, thereby rendering these sites inaccessible to their associated TFs. To test this, we chose to examine previously identified GATA6 binding sites in KP LUAD (39), as these tumor-derived cell lines maintain GATA6 expression, in contrast to NKX2-1 (Fig. 5L; Supplementary Fig. S5G). Consistent with this hypothesis, Smarca4 inactivation resulted in a decrease in chromatin accessibility at GATA6 binding sites, demonstrated by a reduction of ATAC-seq peak strength at these sites in SMARCA4-KO cells compared with SMARCA4-WT cells (Fig. 5M). We observed robust SMARCA4 occupancy at GATA6 binding sites in SMARCA4-WT cells, but not in SMARCA4-KO cells, in line with a direct role for SMARCA4 in remodeling chromatin at these sites (Fig. 5M). Furthermore, occupancies of the pan-SWI/SNF subunit SMARCC1 as well as class-specific subunits ARID1A, PBRM1, and BRD9 likewise were significantly reduced at GATA6 binding sites in the absence of SMARCA4 (Fig. 5N). Thus, we conclude that the loss of GATA6 activity in Smarca4-deficient cells is largely caused by the inability of SWI/SNF complexes to bind and open chromatin at GATA6-binding sites upon Smarca4 inactivation. The reduction in chromatin accessibility at these sites would, in turn, prohibit GATA6 from dictating transcriptional programs that maintain cell identity.

We next examined ATAC-seq peaks increased upon SMARCA4 loss to determine whether these changes are also a result of altered SWI/SNF binding (Supplementary Fig. S5H). Gain of SWI/SNF function has been described as a key driver of malignancy in other SWI/SNF-mutant cancers, most notably synovial sarcoma (40). However, in our data sets, the majority of gained peaks did not display increased SMARCC1, ARID1A, PBRM1, or BRD9 occupancy (Supplementary Fig. S5I), indicating that increased accessibility in SMARCA4-KO cells is not a direct result of increased SWI/SNF activity at these sites. Instead, we observed a slight reduction in the occupancies of these components at these sites. Thus, increased accessibility upon SMARCA4 loss may be a secondary effect or due to loss of SWI/SNF-mediated chromatin compaction.

Next, we examined the binding profile of SMARCA2, a paralog of SMARCA4 and the only other SWI/SNF subunit with catalytic activity. SMARCA2 has been previously identified to be a synthetic lethal target in SMARCA4-deficient cancers (41, 42), and SMARCA2 inhibitors have been developed as a potential targeted therapy for SMARCA4-mutant cancers (43). We detected modest changes in SMARCA2 occupancy among differential ATAC-seq peaks (Supplementary Fig. S5J). However, we did detect 138 differential SMARCA2 peaks between SMARCA4-WT and SMARCA4-KO cell lines, 137 of which were increased in SMARCA4-KO cells (Supplementary Fig. S5K). The increase in SMARCA2 peaks occurred at direct SMARCA4 binding sites, was accompanied by ATAC-seq peaks, and was strongly enriched for AP-1 TF family motifs (Supplementary Fig. S5L and S5M). SMARCA4 loss, therefore, results in some compensatory chromatin remodeling activity by SMARCA2 at certain TF binding sites.

We next sought to determine whether the TF-directed programs we observed in our murine models recapitulated those in patients with LUAD harboring loss-of-function SMARCA4 mutations. We profiled the chromatin accessibility of single cells from PDX models of human KRAS-mutant LUADs with intact SMARCA4 (n = 2), and those with biallelic inactivating SMARCA4 alterations (n = 3) identified through MSK-IMPACT (Supplementary Table S1; ref. 44). We generated chromatin accessibility profiles from 30,992 single cells following murine cell depletion and cell sorting for viability (Fig. 6A and B; Supplementary Fig. S6A–S6C). Each PDX model clustered independently, suggesting that these patient samples have evolved and selected for distinct epigenetic states (Fig. 6C). Despite these differences among patients, we identified programs broadly characterizing these clusters that were associated with SMARCA4 status (Fig. 6D). We therefore grouped these clusters into three categories: SMARCA4-MUT^AP-1-lo (clusters 1–4), SMARCA4-MUT^AP-1-hi (clusters 5–7), and SMARCA4-WT (clusters 8–10; Fig. 6C and D). Importantly, all three SMARCA4-mutant samples were represented in the SMARCA4-MUT^AP-1-lo group, but not in the SMARCA4-MUT^AP-1-hi group.

Marker motifs for these clusters revealed that a subset of SMARCA4-mutant PDXs, in particular those with p53 pathway inactivation (SMARCA4-MUT^AP-1-lo), recapitulated key features of murine KPS clusters (Fig. 6D and E; Supplementary Fig. S6D). These had low activity for the AP-1 TF family and increased accessibilities for RUNX2 and IRF1 motifs. They were also depleted for peaks harboring the FOX TF family motifs and enriched for activities of POU TFs, indicating a highly undifferentiated cell state (Fig. 6D) consistent with our findings in murine models of SMARCA4 loss. Changes in FOX and POU TF motif accessibilities in SMARCA4-MUT^AP-1-lo clusters were accompanied by a loss and gain, respectively, in peaks of the TFs themselves (Fig. 6F). Notably, individual SMARCA4-mutant samples belonging to the SMARCA4-MUT-AP-1^lo group tended to generate multiple clusters, demonstrating a substantial level of heterogeneity of epigenetic states within individual SMARCA4-deficient patient samples.

We next turned to The Cancer Genome Atlas (TCGA; ref. 6) to explore the relevance of our models to human LUAD by examining a larger set of SMARCA4-mutant tumors (Fig. 6G). We first investigated whether we could detect indications of a club cell-of-origin specifically in human SMARCA4-mutant LUAD. We examined the expression levels of SCGB1A1 and SFTPC, two frequently used markers that distinguish the club cell and AT2 lineages, respectively, in TCGA LUAD grouped by SMARCA4 mutation status (SMARCA4-WT, SMARCA4 missense mutant, and SMARCA4 truncating mutant), and observed no differences in their expression levels among the groups (Supplementary Fig. S7A and S7B). We next scored the transcriptional profiles of these tumors for club cell and AT2 signatures derived from an extensive scRNA-seq study of the human lung (45). In these analyses, we included signatures associated with canonical AT2s, as well as signaling AT2s, a distinct subset of AT2s that express genes involved in WNT signaling defined by this study (45).

Transcriptional profiles of SMARCA4-truncating mutant tumors had significantly decreased correlation with all three cell type–specific signatures derived from both 10× Chromium and SmartSeq2 (SS2) platforms compared with the transcriptional profiles of tumors with intact SMARCA4 (Fig. 6H–J; Supplementary Fig. S7C–S7E). Importantly, these differences were maintained when comparing SMARCA4-WT and SMARCA4-mutant TCGA LUAD samples matched by tumor grade (Supplementary Fig. S7F–S7N). Altogether, these analyses indicate that SMARCA4-mutant LUADs, specifically those with SMARCA4 truncating mutations, correlate poorly with gene sets that are associated with either putative cell-of-origin.

We reasoned that these results may reflect SMARCA4-mutant tumors being more undifferentiated than their SMARCA4-WT counterparts—in line with observations from phenotypic analyses and chromatin profiling of the murine models of Smarca4-deficient LUAD. We therefore scored these tumors for a core embryonic stem cell (ESC)–like gene module, which consists of genes upregulated in both mouse and human ESCs (46). Interestingly, the ESC-like gene-expression signature exhibited strikingly increased correlation with the transcriptional profiles of SMARCA4 truncating mutant tumors compared with those with SMARCA4 missense mutations or intact SMARCA4 (Fig. 6K), a trend not observed when examining a cell proliferation signature (Supplementary Fig. S7O). When tumors were stratified by their correlation with the ESC-like gene module, a substantial fraction of top-scoring tumors (z > 1) had significantly reduced SMARCA4 expression (Supplementary Fig. S7P). Furthermore, top-scoring tumors were specifically enriched for SMARCA4 truncating mutations (P = 4.75e−03), but not for missense mutations (P = 0.38). Collectively, these data show that the transcriptional profiles of SMARCA4 truncating mutant tumors are not only significantly less associated with lineage signatures, but also enriched for an embryonic signature—pointing toward a poorly differentiated state in SMARCA4-deficient LUAD.

Finally, we sought to determine whether the KPS transcriptional signature (q < 0.05, |FC| > 1.5) derived from our murine model captured the transcriptional profiles of human SMARCA4-mutant LUAD. The KPS signature was strongly associated with the transcriptional profiles of TCGA LUAD harboring SMARCA4 truncating mutations (Fig. 6L), indicating that this signature consistently characterizes both SMARCA4-deficient murine and human LUAD. When we stratified patients according to the correlation of the transcriptional profiles of their tumors to the KPS signature, top-scoring patients (z > 1) showed no differences in five-year or overall survival compared with the rest of the cohort (Supplementary Fig. S7Q and S7R). However, these patients had significantly reduced SMARCA4 expression (Fig. 6M) and were strikingly enriched for SMARCA4 truncating mutations (P = 3.18e−03), but not missense mutations (P = 0.29). These analyses demonstrate that the murine models and data sets we have generated are relevant to human SMARCA4-deficient LUAD.

SMARCA4, a catalytic component of the SWI/SNF chromatin remodeling complex, is among the top mutated genes in LUAD, and its mutation is a major predictor of poor patient survival. However, the functional impact of SMARCA4 loss-of-function on tumor initiation, progression, and the chromatin landscape in lung cancer has been unclear to date.

Our experiments in autochthonous mouse models of LUAD identify CCSP⁺ lung cells to be uniquely sensitive to malignant transformation upon Smarca4 loss. SMARCA4-negative tumors initiated from SPC⁺ lung cells in KPS animals were either low grade or highly advanced, which can be explained by a heterogeneous population of SPC-expressing cells—including AT2 and BASCs—transformed at tumor initiation. Our results suggest that these cell types have differential inherent sensitivities to Smarca4 perturbation. Analysis of our scATAC-seq data set supports a club cell state for high-grade SMARCA4-deficient tumors in this model. Indeed, when we initiated tumors in KPS animals from CCSP-expressing cells, which are predominantly club cells, we observed a consistent increase in tumor number, grade, and metastatic incidence, as well as a shorter overall survival for KPS animals. Collectively, these results show that the club cell state is uniquely sensitive to malignant transformation and tumor progression upon Kras activation and Trp53 loss in the absence of SMARCA4 function.

This particular cell state can be adopted by BASCs upon differentiation. These cells have been shown through elegant lineage-tracing experiments to have the ability to repopulate the club cell population upon injury (32, 33). Alternatively, transformed AT2 cells in KPS animals may also transdifferentiate into a club cell state following tumor initiation in the context of Smarca4 loss. This possibility can be directly addressed by lineage-tracing experiments; however, our data suggest that this is predominantly not the case. SMARCA4-negative high-grade tumors arising from SPC-expressing cells in KPS animals were typically found by the airways where BASCs reside, whereas SMARCA4-negative low-grade tumors were located in the alveolar space. This supports a cell-of-origin switch and not an AT2 to club cell transdifferentiation event that gives rise to SMARCA4-negative high-grade tumors initiated from SPC⁺ cells in KPS animals.

The phenotypes we describe in the SPC and CCSP models corroborate aspects of previous studies that investigated the effects of Smarca4 inactivation on tumor progression in Kras-driven GEMMs of LUAD (11, 12). Cell-type specificity in the tumor-suppressive function of Smarca4 in the lung provides a potential explanation for seemingly contradictory results in these studies, which used ubiquitous promoters to drive Cre expression and initiate tumors. The effects of Smarca4 loss on overall tumor progression would be dependent on the fraction of transformed cells that were singly SPC⁺ or CCSP⁺. These likely vary depending on the relative efficiencies of the method of viral transduction in these cell types, as well as the relative activities of the promoter used to drive Cre expression in SPC⁺ and CCSP⁺ cells. Our results in the CCSP model are consistent with the accelerated tumorigenesis observed by others in KPS animals in which tumors were initiated using adenoviral delivery of Cre (12). Although we do not observe an enrichment for the oxidative phosphorylation signature described in this study in KPS tumors from our model, we do observe an association between this signature and SMARCA4 truncating mutant tumors in TCGA LUAD (Supplementary Fig. S7S). Interestingly, our results in the SPC model are also in line with the restrained tumor progression observed in Kras-driven GEMMs in which Smarca4 loss-of-function mutations were generated by CRISPR/Cas9 (11). We speculate that the majority of transformed cells harboring Smarca4 mutations in this case were AT2 cells, which would be consistent with previous work demonstrating this cell type to be the predominant cell-of-origin in this model (19, 29).

The mutation spectrum of SWI/SNF subunits is distinct across cancer types, indicating that the requirements for subunit function are highly context-specific. Our results indicate that within murine LUAD, a distinct cell-of-origin underlies Smarca4 mutants. Our findings demonstrate cell-type specificity in the tumor-suppressive function of SMARCA4 and identify the requirements for cell state that are permissive for transformation upon Smarca4 mutation in the lung. Investigation of SMARCA4 and SWI/SNF function in AT2 and club cells under normal physiologic conditions and in response to oncogenic stress will be key to understanding the differences underlying the sensitivities of these cell types to undergo malignant transformation upon Smarca4 inactivation. Interestingly, Kras-driven LUADs harboring Keap1 loss also display a bronchiolar cell-of-origin (47). KEAP1 mutations are strongly associated with SMARCA4 mutations in patients with LUAD (8), leading us to speculate that most human tumors harboring both SMARCA4 and KEAP1 mutations arise from cells of the club cell lineage. Though we were unable to detect a specific enrichment of a club cell signature in SMARCA4-deficient human LUAD—presumably due to these tumors being highly undifferentiated—we anticipate that determining the cell-of-origin of distinct molecular subtypes of the human disease will be increasingly addressable as more extensive and sophisticated molecular profiling and analyses of both normal cell types and tumors throughout tumor evolution are performed.

Importantly, our studies also provide mechanistic insights into the tumor-suppressive functions of SMARCA4 in the lung, particularly in chromatin regulation. We show that Smarca4-deficient cancer cells have a chromatin state resembling that of metastatic cells. In contrast to Smarca4-intact primary tumors, which undergo a gradual loss of lung epithelial cell identity and lineage fidelity (20, 22–24), Smarca4-deficient primary tumors largely lack activities of lung lineage TFs, similar to metastatic cells. SMARCA4 loss directly results in a defect in the ability of all three classes of SWI/SNF complexes (cBAF, PBAF, and ncBAF/GBAF) to bind and open chromatin in regulatory regions, including target sites of lung lineage TFs. Among these, the most profound change occurs at cBAF-binding sites. Inaccessible chromatin as a direct consequence of Smarca4 inactivation and SWI/SNF loss-of-function would directly hamper the ability of lung lineage TFs, such as GATA6, to maintain cell identity. Additionally, TF programs known to be active in metastatic-like and metastatic cells (20, 25) are highly enriched in subsets of primary Smarca4-deficient tumor cells.

Primary LUADs display aberrant expression of TFs specifying epithelial lineages, whereas metastases exhibit transcriptional programs characteristic of a stem-like or progenitor state (25). Loss of the final differentiation state of the tumor cell-of-origin and the acquisition of plasticity in cancer cells enables the adoption of progenitor-like states or alternative differentiation states. Such phenotypic plasticity is thought to support malignancy throughout tumor evolution (48). That Smarca4-deficient primary tumors efficiently lose cell identity is a potential explanation for the rapid acquisition and selection of a cell state associated with highly advanced tumors and metastases and the acceleration of tumor progression observed in KPS animals.

Taken together, these data demonstrate a direct role for SMARCA4 loss in driving an aggressive malignant phenotype that underlies the poor prognosis of this molecular subtype of LUAD. More broadly, these data are in line with other highly undifferentiated SMARCA4-mutant malignancies observed in patients, including small cell carcinomas of the ovary, hypercalcemic type (49–51), and thoracic sarcomatoid tumors, which are thought to represent undifferentiated or dedifferentiated lung carcinomas (52, 53).

Our work has focused on complete Smarca4 loss in conjunction with Kras activation and Trp53 loss in LUAD. Indeed, the KPS signature derived from our model best captures the tumor expression profiles of TCGA LUAD patients harboring SMARCA4 truncating mutations. Future studies modeling recurrent SMARCA4 missense mutations, which have been shown to have dominant-negative and gain-of-function effects (54, 55), will be crucial to understand their specific effects on lung cancer evolution. Additionally, modeling Smarca4 mutations in combination with other frequently co-occurring genetic alterations, such as Keap1 and Stk11 mutations, will be critical to assess the impact of SMARCA4 mutations in other contexts and to expand our repertoire of relevant Smarca4-mutant preclinical models that can be used to test therapeutic strategies.

Although we have focused on the consequences of Smarca4 loss on tumor cell state and SWI/SNF function in chromatin regulation in this work, these models and data sets are also poised to assess other functions of SWI/SNF, such as maintenance of genomic stability upon Smarca4 loss, and to examine potentially altered interactions among Smarca4-deficient lung cancers and the tumor microenvironment during tumor evolution. Furthermore, the CCSP KPS model is a useful preclinical platform to evaluate various therapeutic approaches that have been proposed for SMARCA4-mutant LUAD.

In sum, this work puts forth a model wherein SMARCA4 loss in transformed CCSP⁺ cells directly results in the inability of SWI/SNF complexes to bind to chromatin, and eject and mobilize nucleosomes, which prohibits lung lineage TFs from exerting lineage-specifying gene-expression programs (Fig. 7). Absence of lineage specificity in turn promotes phenotypic plasticity of Smarca4-deficient cells and accelerates the sampling and selection of protumorigenic states throughout tumor evolution. Ultimately, this drives the increased incidence of high-grade tumors and metastases in Smarca4-deficient murine LUAD, and highly undifferentiated tumors and poor overall survival in patients with NSCLC harboring SMARCA4-inactivating alterations. Collectively, this work provides a global view of SMARCA4-mediated tumor suppression in the lung.

Mouse strains used in this study were previously published: Kras^LSL-G12D (56), Trp53^fl (57), Smarca4^fl (16), and Rosa26^LSL-tdTomato (58). Mice were maintained in a mixed Sv129/C57BL/6 genetic background. Tumors were initiated using 1.0 or 2.5 × 10⁸ plaque-forming units (PFU) of Ad-SPC-Cre or 1.0 × 10⁸ PFU of Ad-CCSP-Cre (18) from the Viral Vector Core of the University of Iowa through intratracheal instillation as previously described (59) in age-matched (∼8–12 weeks of age) and sex-matched littermate cohorts. Animal health was monitored daily by the investigators and/or veterinary staff at the Department of Comparative Medicine at MIT. Cohorts of mice were euthanized by CO₂ inhalation or cervical dislocation. Mice were euthanized at defined time points (17 weeks after infection for Ad-SPC-Cre—infected animals in which 1.0 × 10⁸ PFU per mouse was used, 14 weeks after infection for Ad-SPC-Cre—infected animals in which 2.5 × 10⁸ PFU per mouse was used, and 16 weeks after infection for Ad-CCSP-Cre—infected animals in which 1.0 × 10⁸ PFU per mouse was used) or upon reaching a body condition score under 2 for long-term studies. Animal studies were approved by the MIT Committee for Animal Care.

Lung tissues were perfused with PBS through the heart and inflated with zinc formalin through the trachea. Tissues were then fixed overnight in zinc formalin, transferred to 70% ethanol, and embedded in paraffin. Tissues in the chest cavity of each mouse were also fixed and paraffin-embedded to identify micrometastases. Sections were cut at 4-μm thickness and stained for hematoxylin and eosin (H&E) for histologic examination. For IHC, slides were dewaxed, and antigen retrieval was performed using citrate buffer (pH 6.0). Endogenous peroxidase was blocked using DAKO Dual Endogenous Enzyme Block, and endogenous species protein was blocked using the appropriate species serum depending on the secondary antibody. Tissues were incubated with primary antibodies overnight. Primary antibodies used were: anti-SMARCA4 (Abcam; catalog no. ab110641, RRID:AB_10861578, 1:500), anti–Ki-67 (Cell Signaling Technology, catalog no. 12202, 1:200), anti–NKX2-1 (Abcam; catalog no. ab76013, RRID:AB_1310784, 1:1,000), and anti-GATA6 (Cell Signaling Technology, catalog no. 5851, 1:400). ImmPRESS horseradish peroxidase secondary antibodies and the DAB Peroxidase Substrate Kit (Vector Laboratories) were used for signal detection. Tissues were counterstained with hematoxylin. Histologic quantification of tumor area, tumor grade, and lung area was performed in H&E-stained sections using an automated deep neural network developed by Aiforia (nsclc_V25 or nsclc_V37) in collaboration with the Jacks and Tammela labs (20) under the guidance of R.T. Bronson. Quantification of the number of tumors per mouse was performed in H&E-stained sections in a blinded manner using QuPath (60). Identification of metastatic lesions in animals was performed by microscopic examination of H&E-stained sections of paraffin-embedded tissues of the chest cavity including thymus and lymph node. All H&E slides were independently examined by an expert in mouse pathology (R.T. Bronson), who identified slides with grade 4 tumors and metastases in a blinded manner. Classification of SMARCA4 protein expression status in tumors in KPS animals, measurements of tumor size, and percentages of Ki-67–positive cells in Ad-SPC-Cre–infected KPS animals were determined using Aperio ImageScope (v12.3.2.8013). Measurements of the percentages of NKX2-1– and GATA6-positive cells were performed using QuPath (60).

SMARCA4-positive tumors in KPS animals identified by IHC were laser-capture microdissected from paraffin sections using the Veritas Laser-Capture Microdissection microscope. DNA was extracted from individual tumors using the Arcturus PicoPure DNA Extraction Kit (Applied Biosystems). Dissected tumor sections were incubated in extraction solution containing proteinase K at 65°C overnight, spun down, incubated at 95°C for 10 minutes, and cooled. The samples were directly subjected to published genotyping protocols to identify Smarca4 recombined and floxed alleles (16).

Tumor-bearing lungs and macrometastases from moribund KP; Rosa26^LSL-tdTomato, KPS-HET; Rosa26^LSL-tdTomato, and KPS; Rosa26^LSL-tdTomato animals were dissociated using the Miltenyi Biotec Lung Dissociation Kit (130-095-927). Tissues were submerged in enzymes A and D diluted in 1X Buffer S, mechanically dissociated using dissecting scissors, and incubated at 37°C for 25 minutes with rotation. The dissociated cells were then filtered using a 100-μm strainer. Red blood cells were lysed using ACK (Thermo Scientific), and stained with APC-conjugated anti-CD31 (BioLegend; catalog no. 102510, RRID:AB_312917, 1:500), anti-CD45 (BD Biosciences; catalog no. 559864, RRID:AB_398672, 1:500), anti-CD11b (eBioscience, catalog no. 17-0112-82, 1:500), and anti-TER119 (BD Biosciences; catalog no. 557909, RRID:AB_398635, 1:500). DAPI was used for live/dead staining. FACS was performed using a FACSAria sorter (BD) to isolate DAPI⁻/tdTomato⁺/APC⁻ cancer cells. Isolated cancer cells were then subjected to sciATAC or Chromium Single-Cell ATAC protocols. Alternatively, single-cell suspensions were frozen in freezing media (DMEM supplemented with penicillin–streptomycin, 20% FBS, and 10% dimethyl sulfoxide) prior to antibody staining, sorting, and scATAC-seq or RNA extraction for bulk RNA-seq at a later date.

SMARCA4-intact and SMARCA4-mutant LUAD PDXs, identified through MSK-IMPACT (44), were maintained in NSG (NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ) mice (Jackson Laboratories Stock #5557; ref. 61). Tumor cells were thawed from frozen stocks and washed with PBS. Human cells were enriched using the Miltenyi Biotec Mouse Cell Depletion Kit (130-104-694). Cells were washed with column buffer, incubated with mouse cell depletion cocktail, and subjected to magnetic separation using an LS column. Human cells from the flow through were collected and stained with DAPI. Live cells were sorted using the FACSAria sorter (BD) and subjected to the chromium single-cell ATAC protocol.

Nuclei from sorted cells were isolated using the low cell input nuclei isolation protocol from 10X Genomics. Cells were spun at 300 relative centrifugal force (rcf) for 5 minutes at 4°C and resuspended in 50 μL PBS + 0.04% BSA. Cells were spun at 300 rcf for 5 minutes at 4°C, and 45 μL supernatant was removed prior to addition of 45 μL chilled lysis buffer (10 mmol/L Tris-HCl pH 7.4, 10 mmol/L NaCl, 3 mmol/L MgCl₂, 0.1% Tween-20, 0.1% NP-40, 0.01% Digitonin, 1% BSA), gentle pipetting, and incubation on ice for 6 minutes. 50 μL of chilled wash buffer (10 mmol/L Tris-HCl pH 7.4, 10 mmol/L NaCl, 3 mmol/L MgCl₂, 1% BSA, 0.1% Tween-20) was added, and the sample was spun at 500 rcf for 5 minutes at 4°C prior to removal of 95 μL of supernatant. Nuclei were washed with 45 μL chilled diluted nuclei buffer, spun at 500 rcf for 5 minutes at 4°C, and resuspended in 7 μL chilled diluted nuclei buffer prior to counting. After counting, nuclei were diluted to capture 2,000 to 6,000 nuclei per sample. Five microliters of nuclei was mixed with the ATAC Buffer and the Tn5 transposase and incubated for 60 minutes at 37°C. Further processing of the sample and library generation was performed as described in the 10x Genomics single-cell ATAC regent kits user guide.

Sorted cancer cells were transferred to Eppendorf tubes precoated with 7.5% BSA and pelleted by centrifugation at 300 rcf for 3 minutes. The cell pellets were resuspended in 100 μL cold PBS and counted. Samples with a cell number more than 60,000 were performed with fixation, and samples with a lower cell number were not fixed and immediately proceeded to the transposition step. Cells were fixed by adding 6.7 μL 1.6% formaldehyde to a final concentration of 0.1% and incubated at room temperature for 5 minutes. Fixation was stopped by adding 5.6 μL 2.5M glycine, 5 μL 1M pH 8.0 Tris, and 1.3 μL 7.5% BSA followed by incubation on ice for 10 minutes. The cells were pelleted by centrifugation at 500 rcf for 3 minutes at room temperature. Cells were gently washed twice with 0.5 mL PBS by pipetting against the side of the tube without resuspending the pellet followed by centrifugation at 500 rcf for 3 minutes.

Fixed cells (1 μL) were distributed across a 96-well plate and combined with 7 μL transposition buffer (41.25 mmol/L Tris-acetate, 82.5 mmol/L K-acetate, 12.5 mmol/L Mg-acetate, 20% DMF, 0.125% NP-40, 0.5% protease inhibitor cocktail) and incubated at room temperature for 10 minutes. The assembled Tn5 was diluted 1:1 by adding 8 μL transposition buffer to 8 μL of the assembled oligo containing Tn5 as previously described (20). One microliter of the Tn5 containing Ad1 and 1 μL of the Tn5 containing Ad2 were added to each well. The transposition reaction was carried out at 37°C at 300 rpm for 30 minutes. The reaction was stopped by adding 1 μL 0.5M EDTA to each well, mixing well, followed by incubation at 37°C for 15 minutes at 300 rpm. All 96 reactions were pooled and combined with 38.4 μL MgCl₂. The sample was transferred to an Eppendorf tube precoated with 7.5% BSA. The sample was pelleted by centrifugation at 500 rcf for 2 minutes, then washed in 1 mL nuclei isolation buffer (10 mmol/L Tris-HCl, 10 mmol/L NaCl, 3 mmol/L MgCl₂, 0.1% NP-40). The pellet was resuspended in 0.5 mL nuclei isolation buffer and passed through a 40-μm filter and then diluted to 13.3 cells/μL.

The sample (1.5 μL) was distributed across a 96-well plate and was combined with 2.5 μL reverse cross-link buffer (100 mmol/L Tris pH 8.0, 400 mmol/L NaCl, 2 mmol/L EDTA pH 8.0, 2% SDS, and 40 mg/mL proteinase K), 0.5 μL of 10 μm P1 oligo and 0.5 μL of 10 μm P2 oligo. Reverse cross-linking was carried out at 55°C for 16 hours in a thermal cycler. Five microliters of 10% Tween20 was added to quench SDS. PCR reactions were carried out by adding 12.5 μL 2× NEBNext PCR mix and 2.5 μL water to each well, then using the following conditions: 72°C for 5 minutes (extension), 98°C for 5 minutes, and thermocycling at 98°C for 10 seconds, 70°C for 30 seconds, and 72°C for 1 minute. After thermocycling for five cycles, 5 μL from four randomly chosen wells was used to perform qPCR. 10 μL PCR mix with 0.6× SYBRgreen was added and cycled through the following conditions: 98°C for 30 seconds (initial extension) and thermocycling at 98°C for 10 seconds, 70°C for 30 seconds, and 72°C for 1 minute for 25 cycles to determine the number of additional cycles required for the remaining samples on the plate. Libraries were amplified for 13 to 14 cycles in total. The libraries from each plate were pooled and purified using the Qiagen MinElute PCR purification column. Libraries were quantified using the KAPA library quantification kit and then sequenced on the Next-seq platform (Illumina) using a 150-cycle kit.

scATAC-seq data were preprocessed using Cell Ranger ATAC to generate fragment files after removing duplicates. The reads were aligned to either mm10 or hg19 genome. The fragment files for each sample were used as input for peak calling with MACS v2.1.2 (62). All default options were used, with the following flags explicitly set: –nomodel, –nolambda, –keep-dup all, and –call-summits. The peak summits were merged and padded with 150 base pairs (bp) at either end to obtain fixed-width peak windows. Only peaks with smallest P values were kept if peaks overlapped. Using the generated peak region list, the number of reads overlapping a given peak window was determined for each unique cell barcode tag. This generated a peak-by-cell count matrix corresponding to ATAC reads in peaks for each cell profiled. The peak × cell count matrix was used to generate TF motif score × cell matrix using chromVAR (63).

The dimension reduction of scATAC data was performed using cisTopics (64). The uniform manifold approximation and projection (UMAP) algorithm (65) was then applied to project single cells in two dimensions using the topics from cisTopics. To further cluster cell populations, the Louvain method for network community detection, a heuristic method based on modularity optimization (21), was then applied on a k-nearest neighbor graph built using the topics and visualized in the original UMAP space.

We calculated the gene scores by summing the reads intersecting the gene body and promoter region (2 kb upstream of TSSs). To reduce sequencing depth bias, we then normalized gene scores by the total gene score per cell. We then selected the top 30 marker genes of each cell type based on a previous scRNA-seq data set (35) and defined the gene module score by averaging gene scores of selected genes for each single cell. P values are from the Student t test.

Previously identified normal lung scATAC-seq profiles were utilized to identify club and AT2-specific gene scores and motif signatures (20). To match each cell to a meta-cluster, scATAC-seq profiles were filtered for highly variable gene scores. The coefficient of variation (CV) for each gene was determined and filtered for genes with a CV > 1. Then, the most correlated (Pearson) meta-cell was used for cell type matching. Absolute differential gene scores were computed between club and AT2 cell types and determined as significant with a P < 0.01 and an absolute difference on > 1.5 between club and AT2 genes scores or motifs.

Peaks were first called on each bulk ATAC sample using MACS v2.1.2 (62) with the following flags explicitly set: –nomodel, –nolambda, and –keep-dup all. We then used the getAnnotations function from chromVAR package (66) to find the overlap between bulk ATAC peaks and scATAC peaks. Next, we derived z score for each bulk ATAC peak set using the computeDeviations function from chromVAR, which compares the peak counts to GC-matched background peak counts. The z score was further smoothed over 50 nearest neighbors and painted on UMAP.

We scored scATAC cells with gene scores (see Signature Score section) of differential genes from bulk RNA-seq. To do this, we first identified significantly differential genes (DE genes) between KP and KPS using DESeq2 (adjusted P < 0.05, |FC| > 1.5; ref. 67). Then we calculated the mean gene scores for each cell for the DE genes that are up in either KP or KPS. To visualize the score on scATAC-seq UMAP, we further smoothed the mean gene score over 50 nearest neighbors and painted on UMAP.

Isogenic pairs of Smarca4-deficient and Smarca4-intact cell lines were generated in the Jacks Laboratory from two independent KP tumor–derived cell lines and were previously described (28). The T2 (SMARCA4-WT) and M (SMARCA4-KO) cell lines were generated from the parental line LG1233. The 36 (SMARCA4-WT) and 23 (SMARCA4-KO) cell lines were generated from the parental line LG1234. Cell lines were authenticated by performing Western blots for SMARCA4 expression using both N- and C-terminal antibodies for SMARCA4 and by RNA-seq. Cells were maintained in DMEM (Corning; 10-013-CV) supplemented with 10% FBS and 1× penicillin–streptomycin solution (VWR 45000-652). All experiments were performed within four to five days from the time of thawing. Cell lines tested negative for Mycoplasma by the MycoAltertTM Kit from Lonza (October 17, 2017).

RNA from isogenic cell lines was extracted from cell lines using TRIzol as per the manufacturer's instructions. RNA from sorted primary tumors was extracted using the Qiagen RNeasy Microkit (catalog no. 74004) according to the manufacturer's instructions.

Bulk ATAC-seq was performed on cell lines as previously published (68).

CUT&RUN for SWI/SNF components and histone marks was performed as previously described (69).

Concanavalin A–coated magnetic beads (Polysciences 86057-3) were activated by washing twice with 1 mL binding buffer (20 mmol/L HEPES pH 7.9, 10 mmol/L KCl, 1 mmol/L CaCl₂, 1 mmol/L MnCl₂). Bead suspension (10 μL) was used per condition.

Cells (5 × 10⁵ per condition) were washed twice with 1.5 mL of wash buffer (20 mmol/L HEPES pH 7.5, 150 mmol/L NaCl, 0.5 mmol/L Spermidine, 1× protease inhibitor), and resuspended in 1 mL of wash buffer. Activated bead suspension was added to each sample. Samples were then rotated at room temperature for 10 minutes.

Samples were subjected to a quick spin and placed on a magnet stand. The liquid was removed and discarded. Fifty microliters of the antibody solution (1:50 dilution of antibody in 0.1% digitonin wash buffer and 2 mmol/L EDTA) was then added to each sample and mixed. Samples were then rotated at 4°C overnight. The following day, the tubes were subjected to a quick spin and placed on a magnet stand. The liquid was removed and discarded, and the beads were washed twice with 1 mL of 0.1% digitonin wash buffer (wash buffer + 0.1% digitonin) and resuspended in 50 μL of 0.1% digitonin wash buffer. The following antibodies were used: anti-SMARCA4 (Abcam; catalog no. ab110641, RRID:AB_10861578), anti-SMARCA2 (Abcam; catalog no. ab15597, RRID:AB_443214), anti-ARID1A (Abcam; catalog no. ab217154; Cell Signaling Technology; catalog no. 12354, RRID:AB_2637010), anti-SMARCC1 (Cell Signaling Technology; catalog no. 11956, RRID:AB_2797776), anti-BRD9 (Abcam; catalog no. ab155039), anti-PBRM1 (Active Motif; catalog no. 61381, RRID:AB_2793612), anti-H3K27ac (Active Motif; catalog no. 39133, RRID:AB_2561016), anti-H3K4me1 (Abcam; catalog no. ab8895, RRID:AB_306847), and anti-H3K4me3 (Millipore; catalog no. 07-473, RRID:AB_1977252).

EpiCypher CUTANA pAG-MNase (2.5 μL) was added to each sample. Samples were then rotated at 4°C for 1 hour, subjected to a quick spin, and placed on a magnet stand. The liquid was removed and discarded. The samples were washed then with twice with 1 mL of 0.1% digitonin wash buffer and resuspended in 150 μL of 0.1% digitonin wash buffer.

Three microliters of 100 mmol/L CaCl₂ was then added to each tube with gentle vortexing and placed on a chilled metal block (4°C) for 2 hours with periodic gentle shaking throughout the incubation period. 100 μL 2X stop buffer (340 mmol/L NaCl, 20 mmol/L EDTA, 4 mmol/L EGTA, 0.02% digitonin, 50 μg/mL RNase A, 50 μg/mL glycogen, and 2 pg/mL heterologous spike-in DNA) was then added to each sample and mixed. The tubes were incubated at 37°C for 10 minutes at 500 rpm, and then spun at 16,000 × g at 4°C for 5 minutes. The samples were then placed on a magnet stand and the liquid was transferred to DNA low-bind tubes.

Two microliters of 10% (w/v) SDS and 1.5 μL of proteinase K (20 mg/mL) was added to each tube and mixed by inversion. Samples were incubated at 70°C for 10 minutes, after which 200 μL of phenol–chloroform–isoamyl alcohol 25:24:1 was added. Samples were vortexed and then transferred to a phase-lock tube, and centrifuged at 16,000 × g at room temperature for 5 minutes. Chloroform (200 μL) was then added, and tubes were inverted to mix the solution and centrifuged at 16,000 × g at room temperature for 5 minutes. The sample was then transferred to a 1.5-mL tube containing 2 μL of 2 mg/mL glycogen. Five-hundred microliters of 100% ethanol was added to each sample and mixed by inversion. Samples were chilled on ice for 10 minutes, and subsequently centrifuged at 16,000 × g at 4°C for 10 minutes. The liquid was decanted, and the pellets were washed with 1 mL of 100% ethanol. Samples were then centrifuged at 16,000 × g at 4°C for 1 minute. The liquid was decanted and drained on a paper towel. The pellets were air-dried for 5 minutes and dissolved in 40 μL of 1 mmol/L Tris-HCl at pH 8 with 0.1 mmol/L EDTA.

Libraries were prepared by NEB Ultra II. Paired-end Illumina sequencing was then performed.

Single-ended 50mer RNA-seq reads for SMARCA4-WT (T2, 36) and SMARCA4-KO (M, 23) samples were mapped to the UCSC mm9 mouse genome build (genome.ucsc.edu) using Bowtie (70) v1.0.1 and gene counts were quantified using RSEM (71) v1.2.12. Estimated expression counts generated by RSEM were upper quartile normalized to a count of 1,000 (72). Genes with low expression across all samples (upper quartile of normalized counts < 10) were dropped from downstream analyses. Genes with normalized expression standard deviation less than 50 across all four samples were classified as stably expressed genes. These were used in downstream bulk ATAC-seq and CUT&RUN analyses where intervals of ATAC peaks overlapping TSS of stably expressed genes were analyzed for chromatin accessibility and occupancy profiles across SMARCA4-WT and SMARCA4-KO conditions. Raw counts from RSEM were used to detect DE genes between SMARCA4-WT and SMARCA4-KO conditions using DESeq2 (67) v1.26.0 with WT as the baseline condition. These results were used to assess DE status of Gata6 and Nkx2-1. Processed data for RNA-seq are included in Supplementary Information (RNAseq_supp_material.xlsx).

Single-ended 50 mer RNA-seq reads for KP, KPS-HET, and KPS primary tumor samples were trimmed to 35 mers in order to drop lower quality 3′ read positions. 35 mer reads were then mapped to the USCC mm9 mouse genome build (genome.ucsc.edu) using Bowtie (73) v1.2.3, and gene counts were quantified using RSEM (71) v1.3.1. Estimated expression counts generated by RSEM were used to detect DE genes between pairwise conditions using DESeq2 (67) v1.26.0. Processed data for RNA-seq are included in Supplementary Information (RNAseq_invivo_supp_tbl.xlsx).

Paired-end 25 mer CUT&RUN reads were mapped to the UCSC mm9 mouse genome build (genome.ucsc.edu) using Bowtie2 (73) v2.2.6. Read-alignment BAM files were processed with Picard MarkDuplicates v2.17.0 (broadinstitute.github.io/picard/) to drop duplicate alignments and sorted by read-name using Samtools (74) v1.5. Alignments were converted to BED format using BEDTools (75) bamtobed v2.29.2 and subsequently to bedgraph format using genomecov. Peaks were called with the SEACR (76) v1.2 pipeline by selecting the top 0.5% of regions enriched using area under the curve statistic. Peaks were annotated by genomic feature (promoter, distal intergenic, etc.) using ChIPseeker v1.22.1 with UCSC mm9 genome annotation. Differential peak analyses between WT and KO conditions per factor were conducted using DiffBind (77) v2.4.8. DiffBind normalized read-count correlation plots were generated using the default normalization scheme (sequencing depth based), and reads in consensus peaks across samples (peaks that overlapped in at least two samples) were used. DiffBind results for SMARCA2 peaks (WT: T2,36 vs. KO: M,23 samples) were visualized using a volcano plot. Motif analyses for differentially represented SMARCA2 peaks (FDR < 0.05) were conducted using HOMER (78) v4.10 with a background set consisting of nondifferentially enriched peaks (FDR = 1; |log₂ fold change| < 0.01). The known-motif result set from HOMER was reviewed for significantly enriched motifs. Read density metaplots and heat maps were generated using deepTools (79) v3.0.1 on 1× normalized BAM files with ± 5 kb flanks upstream and downstream of peaks. Representative heat maps and metaplots shown in Fig. 5 and Supplementary Fig. S5 show data generated from the T2 (SMARCA4-WT) and M (SMARCA4-KO) isogenic pair.

Gata6-binding sites were obtained from supplementary data for ChIP-seq peaks in Gene Expression Omnibus (GEO) accession GSE124601 (39). Mouse genome build mm10 coordinates were translated to mm9 coordinates using the UCSC-tools liftOver utility. Occupancy metaplot profiles were generated using deepTools (79) v3.0.1 with ± 2 kb upstream and downstream flanks. Processed data for CUT&RUN are included in Supplementary Information (CnR_supp_material.xlsx).

Paired-end 40 mer bulk ATAC-seq reads for SMARCA4-WT (T2, 36) and SMARCA4-KO (M, 23) samples (two replicates each for T2, 36, M, 23) were mapped to the UCSC mm9 mouse genome build (genome.ucsc.edu) using Bowtie (70) v1.0.1. Read-alignment BAM files were processed with Picard MarkDuplicates v2.17.0 (broadinstitute.github.io/picard/) to drop duplicate alignments. BAMs for all samples were merged using Samtools (74) v0.1.13. Peaks were called on the merged BAM using MACS2 (80) v2.2.1. Read counts per peak per sample were quantified using BEDTools multicov (75) v2.26. Differential analysis for peaks was performed using DESeq2 (67) v1.16.1 using default “median of ratios” normalization. Significant peaks (q < 0.05, |log₂ fold change| > 1) were selected for downstream analyses. Peaks were annotated by genomic feature using ChIPseeker v1.22.1 with UCSC mm9 genome annotation. Per-feature motif analyses for differentially enriched peaks were conducted using HOMER (78) v4.10 against a background set of common nonenriched peaks for a given genomic feature (FDR > 0.5, |log₂ fold change| < 1.1). The known-motif result set from HOMER was reviewed for significantly enriched motifs. Read density metaplots and heatmaps were generated using deepTools (79) v3.0.1 on 1× normalized BAM files with ± 5kb flanks upstream and downstream of peaks. Metaplots of chromatin accessibility for Gata6-binding sites were generated as described earlier (see CUT&RUN data analysis). Processed data for bulk ATAC-seq are included in Supplementary Information (ATACseq_supp_material.xlsx).

RNA-seq gene-expression profiles of primary tumors and relevant clinical data of 515 patients with LUAD were obtained from TCGA (gdac.broadinstitute.org; ref. 6). SMARCA4 and KRAS mutational status of TCGA tumor samples was retrieved from cBioPortal (81, 82) using the TCGA PanCancer Atlas collection (gdc.cancer.gov/about-data/publications/pancanatlas) wherein 510 of 515 TCGA tumors had mutational status available. Within this data set of 510 samples, counts were as follows: 465 SMARCA4 WT; 45 SMARCA4 mutant (21 truncating mutations; 22 missense mutations; 2 fusions); 154 KRAS mutant (9 SMARCA4 mutant, 145 SMARCA4 WT). Individual tumor transcriptomes were scored with signatures (marker genes with adjusted P < 0.05 for club, AT2, and signaling AT2 cells from published scRNA-seq of the lung; ref. 45) using single-sample gene set enrichment analysis (ssGSEA; ref. 83). Patients were stratified based on standardized ssGSEA scores, and Kaplan–Meier five-year and overall survival analyses were conducted to compare high-scoring patients with the rest of the cohort, and significance was assessed using the log-rank test. All survival analyses were conducted using the survival package in R. Patients were also grouped by mutational status, as described in relevant figure legends, and the distribution of standardized signature ssGSEA scores across groups was illustrated using an empirical cumulative distribution function (ECDF) plot where significance was assessed using a Kolmogorov–Smirnov test. Standardized SMARCA4 expression counts were similarly illustrated using an ECDF plot. All statistical analyses were conducted in the R statistical programming language (R-project.org).

KPS signature genes were derived by taking the top differential genes (adjusted P < 0.05; |FC| > 1.5) in KPS primary tumors versus KP primary tumors for both directions. Mouse identifiers were then mapped to human orthologs using the biomaRt R package. These mapped gene signatures were then tested for enrichment of pathways and annotated gene sets derived from the Molecular Signature Database (MSigDB; ref. 84) using the hypeR R package (85).

Statistical analyses were performed as indicated in the figure legends, Supplementary figure legends, and Methods for each experiment. GraphPad Prism software version 8.3.0 or R (R-project.org) was used. No statistical method was used to determine sample size prior to experimentation. Mice with no detectable tumor burden at endpoint were excluded (SPC cohort from Fig. 1: 1 KPS; CCSP cohort from Fig. 3: 1 KP, 2 KPS-HET).

The data discussed in this article have been deposited in NCBI's GEO (86) and are accessible through GEO Series accession number GSE164867.

C.P. Concepcion reports grants from American Cancer Society and Koch Institute during the conduct of the study. A.J. Schoenfeld reports personal fees from Johnson and Johnson and Perceptive Advisors outside the submitted work. P.M.K. Westcott reports nonfinancial support from Aiforia Technologies Oy during the conduct of the study. G.J. Riely reports grants from Mirati, Pfizer, Roche, Novartis, Lilly, and Takeda outside the submitted work. C.M. Rudin reports personal fees from AbbVie, Amgen, AstraZeneca, Epizyme, Genentech/Roche, Ipsen, Jazz, Lilly, Syros, Bridge Medicines, Harpoon Therapeutics, and Earli outside the submitted work. C.F. Kim reports consortium associations with Bristol-Myers Squibb (formerly known as Celgene Corporation) and Longfonds Stichting. The projects of those consortiums are not related to the work in this publication. C.F. Kim was supported by a Mission Boost Grant, MBG-18-204-01-COUN, from the American Cancer Society and by the NCI of the NIH under Award Number R01CA216188. A. Regev reports personal fees from Genentech, Celsius Therapeutics, Immunitas, Thermo Fisher Scientific, Syros Pharmaceuticals, Neogene Therapeutics, and Asimov outside the submitted work. J.D. Buenrostro reports grants from Harvard University during the conduct of the study; in addition, J.D. Buenrostro has patents related to the invention of ATAC-seq and scATAC-seq issued, licensed, and with royalties paid. T. Jacks reports grants from Howard Hughes Medical Institute, NIH P01 CA42063, NIH CCSG P30 CA14051, Ludwig Center at MIT, Koch Institute Bridge Project, and Koch Institute Frontier Grant during the conduct of the study. No disclosures were reported by the other authors.

C.P. Concepcion: Conceptualization, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. S. Ma: Data curation, formal analysis, writing–review and editing. L.M. LaFave: Validation, investigation, writing–review and editing. A. Bhutkar: Data curation, software, writing–review and editing. M. Liu: Investigation. L.P. DeAngelo: Investigation, visualization, writing–review and editing. J.Y. Kim: Investigation. I. Del Priore: Investigation. A.J. Schoenfeld: Resources, writing–review and editing. M. Miller: Investigation. V.K. Kartha: Visualization. P.M.K. Westcott: Resources, writing–review and editing. F.J. Sánchez-Rivera: Conceptualization, writing–review and editing. K. Meli: Investigation. M. Gupta: Writing–review and editing. R.T. Bronson: Validation. G.J. Riely: Resources, funding acquisition. N. Rekhtman: Resources, writing–review and editing. C.M. Rudin: Resources, supervision, funding acquisition, writing–review and editing. C.F. Kim: Supervision, funding acquisition, writing–review and editing. A. Regev: Supervision. J.D. Buenrostro: Resources, writing–review and editing. T. Jacks: Conceptualization, resources, supervision, funding acquisition, writing–review and editing.

This work was supported by the Howard Hughes Medical Institute, the Virginia and D.K. Ludwig Center at MIT, NIH P01 Jacks P01-CA42063, The Bridge Project, a partnership between the Koch Institute for Integrative Cancer Research at MIT and the Dana-Farber/Harvard Cancer Center (DF/HCC), and a Koch Institute Frontier grant, and in part by a Koch Institute Cancer Center Support Grant P30-CA14051 from the NCI. C.P. Concepcion was supported by a Koch Institute Quinquennial Postdoctoral Fellowship and an American Cancer Society Postdoctoral Fellowship (PF-17-009-01-CDD). This work was supported, in part, by a grant from John and Georgia DallePezze to Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center Support Grant/Core Grant (P30-CA008748), and the Druckenmiller Center for Lung Cancer Research. C.F. Kim was supported by a Mission Boost Grant, MBG-18-204-01-COUN, from the American Cancer Society and by the NCI of the NIH under Award Number R01CA216188. P.M.K. Westcott is a Damon Runyon Fellow. We thank the Jacks lab, particularly S. Naranjo and R. Romero, for helpful discussions; A. Berns from the Netherlands Cancer Institute for Ad-SPC-Cre and Ad-CCSP-Cre; S. Henikoff for CUT&RUN reagents; T. Tammela and T. Westerling for development of the Aiforia deep neural network for murine NSCLC; J. Teixeira, K. Yee, and K. Anderson for administrative support; K. Mercer and M. Magendantz for laboratory and technical support; the Swanson Biotechnology Center, particularly the Flow Cytometry and Histology core facilities, for technical support; the MIT BioMicro Center, particularly S. Levine, and the Harvard Bauer Core Facility, particularly N. El-Ali, for scATAC and sequencing support.

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