The SWI/SNF complex member SMARCB1 supports lineage fidelity in kidney cancer

Summary Lineage switching can induce therapy resistance in cancer. Yet, how lineage fidelity is maintained and how it can be lost remain poorly understood. Here, we have used CRISPR-Cas9-based genetic screening to demonstrate that loss of SMARCB1, a member of the SWI/SNF chromatin remodeling complex, can confer an advantage to clear cell renal cell carcinoma (ccRCC) cells upon inhibition of the renal lineage factor PAX8. Lineage factor inhibition-resistant ccRCC cells formed tumors with morphological features, but not molecular markers, of neuroendocrine differentiation. SMARCB1 inactivation led to large-scale loss of kidney-specific epigenetic programs and restoration of proliferative capacity through the adoption of new dependencies on factors that represent rare essential genes across different cancers. We further developed an analytical approach to systematically characterize lineage fidelity using large-scale CRISPR-Cas9 data. An understanding of the rules that govern lineage switching could aid the development of more durable lineage factor-targeted and other cancer therapies.


INTRODUCTION
Lineage-specific transcription factors (TFs), such as SOX10 and MITF in melanoma, have emerged as a common class of essential genes in large-scale functional cancer cell line fitness screens. 1,2 Clinically relevant examples include estrogen and androgen receptors, which are well-established therapeutic targets in breast and prostate cancer, respectively. 3 The success of hormone therapies suggests that lineage factor dependencies could be exploitable for clinical benefit also in non-hormone receptor-driven cancers. However, the cancer-relevant biology of lineage factors and the mechanisms that maintain lineage fidelity, i.e., the dependency of a cancer on the transcriptional lineage factor programs of its tissue of origin, in advanced cancer clones remain poorly understood. It is also unclear what the long-term consequences of lineage factor inhibition are, how lineage factor independence may arise, and how lineage switching, or loss of lineage fidelity, an emerging mechanism of therapy resistance, 4 is facilitated.
Clear cell renal cell carcinoma (ccRCC) is the most common form of kidney cancer with $300,000 diagnoses and $100,000 deaths annually worldwide. 5 Inactivation of the von Hippel-Lindau tumor suppressor gene (VHL), seen in $90% of ccRCCs, is the only clonal genetic driver alteration in most ccRCCs. 6 VHL loss leads to stabilization of the hypoxia-inducible factors HIF1A and HIF2A, of which HIF2A is critical for ccRCC development. 7 Interestingly, HIF2A inhibitors have demonstrated efficacy against ccRCC in some patients, but de novo and acquired resistance are common. 8,9 The widespread and uniform expression of PAX8, 10 a well-established example of a lineage-specific transcription factor dependency, 1,2,11-14 make PAX8 an attractive alternative target for ccRCC therapy, especially given the redundancy between Pax8 and Pax2 in normal renal development in mice. 12 Recent evidence suggests that in ccRCC cells PAX8 maintains the expression of CCND1 and MYC, two canonical oncogenes that are required for ccRCC proliferation. 14 PAX8 regulates CCND1 expression through a distal enhancer, the activity of which also depends on HIF2A, whereas PAX8-dependent MYC expression involves the downstream mediator HNF1B, another renal lineage factor. 14 iScience Article While ligand-independent transcription factors lack active pockets where small-molecule inhibitors could bind, emerging modalities, such as proteolysis-targeting chimeras or protein-protein interaction inhibitors, could expand the clinically druggable target space to include a broader set of transcription factors in the future. 15,16 Meanwhile, detailed functional genetic modeling in experimental systems could help evaluate the potential of lineage factors as therapeutic targets. To understand the consequences of lineage factor inhibition, we have used genetic screening to identify mechanisms that maintain lineage fidelity and PAX8 dependency in ccRCC cells. We find that loss of the SWI/SNF complex member SMARCB1 can facilitate the development of resistance to PAX8 inhibition, giving rise to a strongly altered histological appearance which displays morphological features of neuroendocrine differentiation but no molecular neuroendocrine markers. The dedifferentiated phenotype relies on newly acquired dependencies on IRF2, BHLHE40, and ZNFX1, which represent rare pan-cancer dependencies. Moreover, systematic analysis of hundreds of cell lines revealed evidence of molecular mechanisms that may promote lineage factor inhibition resistance across several different cancer lineages. We conclude that resistance to lineage factor inhibition follows molecular logic that could be exploited for the prevention of lineage switching, potentially leading to more sustained therapy responses to various anticancer agents.

SMARCB1 loss facilitates resistance to lineage factor inhibition in ccRCC cells
To identify genes that maintain the dependency of ccRCC cells on PAX8, we performed loss-of-function screening on a PAX8 knockdown (KD) background in 786-M1A cells, a metastatic VHL mutant ccRCC cell line, 17 using a single guide RNA (sgRNA) library targeting chromatin regulators, 18 key factors involved in the maintenance of cellular identity ( Figure 1A). To establish the optimal conditions for screening, we derived a doxycycline-inducible Cas9 clone with high editing efficiency and validated that two short hairpin RNAs (shRNAs) targeting PAX8 could effectively suppress PAX8 expression within 24 h and induce a negative proliferative phenotype (Figures S1A-S1E). The shRNAs were co-expressed with a fluorescent reporter, which allowed us to monitor expression and select for cells that maintained PAX8 KD throughout the experiment ( Figure S1F). All genes and 98.5% of sgRNA were captured at the start of the screen and there was a negative selection against known essential genes but not non-essential genes (Figures S1G-S1I). The most enriched PAX8 KD condition (P8 1/2 )-specific hit from the screen was SMARCB1, a member of the SWI/SNF chromatin remodeling complex, which has a key role in chromatin landscape maintenance in development and tissue homeostasis ( Figures 1B and 1C). 19 Three additional SWI/SNF complex members were also specifically enriched in the PAX8 KD condition, of which ARID1A showed the strongest enrichment ( Figure S1J). ARID1A and SMARCB1 inactivation have recently been shown to reduce the sensitivity of breast cancer cells to estrogen receptor inhibition, 20,21 giving support to our findings in a second cancer type. Interestingly, unlike ARID1A, and in contrast to its effect in our screen, SMARCB1 is a prototypical pan-cancer essential gene ( Figures 1D, S1K, and S1L).
We validated the results from the screen by competitive cellular proliferation assays using two additional SMARCB1 sgRNAs (S1 1/2 ) ( Figures 1E and S1M). Acute PAX8 KD/SMARCB1 knockout (KO) (P8 1/2 S1 1/2 (A)) resulted in a proliferative rescue but the cells grew less stably and were sensitive to passaging. Given time ($1 month, mid-term, MT), they began to grow more robustly ( Figures 1F and S2A). SMARCB1 loss thus provides an immediate proliferative advantage to PAX8 KD cells but with a significant stability trade-off that can be selected against over time. Over a period of $2-3 months (long-term, LT), the P8 1 S1 1/2 (LT) cells maintained a similar growth phenotype but cells without SMARCB1 KO (P8 1 Ctrl) also adapted to PAX8 suppression ( Figures S2B and S2C). Critically, despite a partial proliferative rescue, a selective pressure to regain PAX8 suppression was maintained in P8 1 Ctrl(LT/MT) cells but not in their SMARCB1 KO counterparts, as evidenced by the gradual loss of PAX8 shRNA expression in P8 1 Ctrl (LT/MT) cells ( Figures 1G and S2D). Tumor growth in vivo was totally abrogated by PAX8 KD (P8 1 Ctrl(A)), iScience Article partially rescued for P8 1 Ctrl(LT) cells and completely rescued for P8 1 S1 1 (LT) cells, which were transduced with the more efficient SMARCB1 sgRNA (Figures S1M, S2A, and S2H). P8 1 S1 2 (LT) cells, on the other hand, showed only weak tumorigenicity ( Figure S2E). Histological analysis revealed a high-grade ccRCC phenotype with sarcomatoid dedifferentiation in the control cells, characteristic of the 786-M1A cells. 17 On the other hand, tumors arising from the PAX8 KD background presented as high-grade undifferentiated carcinomas, with extensive areas of necrosis and morphological appearance reminiscent of neuroendocrine differentiation, in keeping with a large-cell phenotype, but no morphological evidence of rhabdoid dedifferentiation ( Figures 1H and S2F). In agreement with the in vitro data, P8 1 Ctrl(LT) tumors displayed also areas of the original histology, possibly indicating the presence of escapers from PAX8 KD. However, immunohistochemical and RNA analyses did not detect expression of typical neuroendocrine markers specifically associated with the PAX8 KD-resistant phenotype ( Figure S2G). Interestingly, a similar phenotype, morphological neuroendocrine features without molecular neuroendocrine markers, has recently been described in an experimental mouse-derived renal carcinoma model that displays molecular features of an aggressive ccRCC subtype. 22 To expand our study to additional ccRCC models, we took a systematic approach and identified VHL mutant ccRCC cell models which have a non-synonymous and predicted damaging or TCGA/COSMIC hotspot SMARCB1 (n = 3) or ARID1A (n = 1) mutation from the cell line encyclopedia (CCLE) 23 and compared their PAX8 inhibition sensitivity to their SMARCB1/ARID1A wild-type (WT) counterparts (n = 13) using lossof-function data from the cancer dependency map project (DepMap). 24 In line with the findings from our screen, SMARCB1/ARID1A mutant lines showed strong resistance to PAX8 KO compared to their WT counterparts ( Figure 1I). However, there were also two VHL mutant cell lines that showed similar resistance to PAX8 KO but did not have a SMARCB1 or ARID1A mutation, indicating that PAX8 inhibition resistance can arise through several mechanisms. In line with this, PAX8 depletion in UOK101 cells, another VHL mutant ccRCC cell line, resulted in quick emergence of a resistant population which maintained the essential status of SMARCB1 ( Figures S2H and S2I). In sum, inactivation of SMARCB1, a generally essential gene in cancer cells, is associated with lineage factor independence in ccRCC, but other mechanisms of lineage factor independence also exist.

Large-scale alterations in enhancer activation states upon SMARCB1 loss
To understand the role of SMARCB1 in PAX8 inhibition resistance, we performed RNA sequencing (RNAseq) and assay for transposase-accessible chromatin using sequencing (ATAC-seq) to measure changes in the transcriptome and chromatin accessibility upon SMARCB1 loss. We detected differentially expressed genes across conditions, with more gene expression changes in the P8 1 S1 1/2 (MT/LT) conditions compared to P8 1 Ctrl(A) (Figure 2A and Tables S1, S2, S3, S4, S5, and S6). In accordance with the proliferative phenotype, we found that SMARCB1 loss triggered an increase in proliferative gene signatures (MYC_V1, MYC_V2, G2M, E2F), which increased over time ( Figure 2B). The hallmark apoptosis signature was also reduced in the comparison between acute and long-term SMARCB1 KO, supporting our observation that SMARCB1 simultaneously triggers heightened proliferation and instability, and over time clones which can tolerate SMARCB1 loss are selected for ( Figure 2B). From our ATAC-seq experiment, we detected $72,000 high confidence peaks in total across conditions, with control and P8 1 Ctrl(A) having a similar number of peaks and P8 1 S1 1/2 (LT) having substantially less ( Figure S3A). There was a large proportion of differentially accessible regions in the SMARCB1 KO conditions compared to the control, 18,414 in total, 13,892 of which had lower accessibility (LA) and 4,522 higher accessibility (HA) ( Figures 2C, S3B, and S3C). P8 1 S1 1 (LT) and P8 1 S1 2 (LT) showed a similar overall pattern of altered DNA accessibility ( Figure 2C), but the changes were more pronounced in P8 1 S1 1 (LT) cells ( Figure S3B). As expected, increased chromatin accessibility was associated with increased gene expression whereas reduced chromatin accessibility was associated with reduced gene expression ( Figures 2D, 2E, S3D, and S3E). We annotated the differentially accessible regions based on their location in the genome and found that the majority were intronic and intergenic and that there was a statistically significant underrepresentation of promoter annotations ( Figures 2F and 2G). To test whether these regions were enhancers, we looked for an overlap with markers of active chromatin. A re-analysis of H3K27ac and H3K4me1 chromatin immunoprecipitation sequencing data in 786-M1A cells 25 showed a typical bimodal distribution of average signal flanking the center of the ATAC-seq peaks for both LA and HA regions ( Figures S3F and S3G). In summary, SMARCB1 KO triggers large-scale enhancer re-programming in association with resistance to PAX8 suppression. vs P81/S11/2(LT) P81/S11/2(A) vs P81/S11/2(LT) Figure 2. Large-scale alterations in enhancer activation states upon SMARCB1 loss (A) MA plots of RNA-seq differential expression analysis from 786-M1A-C6 cells. Gene expression fold change was calculated relative to shRen/sgNTC (Ctrl.Ctrl) cells. Highlighted points have an FC > 1.5 or <(-1.5) and p.adjust <0.05. Adjusted p values calculated with DEseq2. (B) Gene set enrichment analysis (GSEA) using the hallmarks collection from mSigDB, for different comparisons as indicated on the left. Highlighted points (purple/cyan) have a p.adjust <0.05. (C) Heatmaps showing normalized ATAC-seq signal +/À 2kb centered on summits of differentially accessible (DA) regions, defined by Ctr.Ctrl vs. P8 1 S1 1/2 (FC > 2 or <(-2), p.adjust <0.001). Top panels show the average signal for higher accessible and lower accessible regions. (D and E) Correlation of ATAC-seq and transcriptional changes, for Ctrl.Ctrl(A) vs. P8 1 S1 1/2 (LT). Downregulated genes near lower accessibility regions in (D). Upregulated genes near higher accessibility regions in (E). Left y axis, the ratio of the number of downregulated/upregulated genes found within windows created around lower/higher accessible regions compared to the number of expressed genes (universe) also found within the windows. Right y axis, p value, one-tailed hypergeometric test. Matched Ctrl peaks for LA and HA regions were generated from the consensus list of all peaks merged across conditions. (F) Stacked bar plots of genomic annotations for LA and HA regions from comparisons Ctrl.Ctrl(A) vs. P8 1 Ctrl(A), P8 1 S1 1 (LT), and P8 1 S1 2 (LT) (FC > 2 or <(-2), p.adjust <0.001). (G) Percentage of regions annotated as a promoter in the consensus list of all peaks merged across conditions (dark gray) and lower and higher accessible regions from (F). Two-tailed hypergeometric test. See also Figure S3 and Tables S1, S2, S3, S4, S5, and S6. Enhancers are key mediators of lineage specification and the SWI/SNF complexes have been demonstrated to maintain tissue-specific enhancers, 26,27 suggesting the possibility that lineage re-programming or dedifferentiation could underlie PAX8 inhibition resistance following SMARCB1 KO. To test this hypothesis, we first performed motif analysis for our differentially accessible peak sets. As expected, the PAX motif was highly enriched in the LA set using both known and de novo motif analysis for P8 1 S1 1/2 (LT) and P8 1 Ctrl(A) ( Figures 3A and S4A-S4D). Interestingly, the most specifically enriched motif for the P8 1 S1 1/2 HA peak set was CTCF/BORIS, which has been linked to the formation of SMARCB1 mutant rhabdoid tumors and the maintenance of a naive pluripotent stem cell state 28  iScience Article To functionally annotate the differentially accessible enhancers in PAX8 inhibition-resistant cells, we downloaded 504 DNAase open chromatin profiles from the ENCODE project, spanning a range of adult and developmental cell types, and clustered the samples into tissue-specific clusters ( Figure S5A). We derived sets of peaks that showed specificity for each cluster and ran an overlap analysis with our differentially accessible regions. The kidney-specific clusters were most enriched for the lower accessibility peaks sets for both P8 1 Ctrl and P8 1 S1 1/2 ( Figures 3C and S5B). However, the global loss of signal at these peak sets was substantially greater for P8 1 S1 1/2 compared to P8 1 Ctrl, suggesting that SMARCB1 loss triggers a widespread loss of renal epithelial epigenetic identity ( Figure 3E). This was supported by specific genomic loci harboring known proximal tubule marker genes as defined by single-cell RNA-seq experiments, for example, CDH6 and SLC16A7 30 ( Figures 3F-3I). The higher accessibility peaks for P8 1 S1 1/2 overlapped most strongly with an IPS/progenitor cluster, which was not significantly enriched in the P8 1 Ctrl higher accessibility regions ( Figures 3D and S5C-S5F).
The global loss of the renal epithelial signal in conjunction with the gain of IPS/progenitor features at the chromatin accessibility level supports the notion that SMARCB1 may maintain a lineage-differentiated cellular state, the loss of which promotes PAX8 inhibition resistance. To test this at the level of gene expression, we used the mSigDB cell-type-specific signature collection (C8), supplemented with a signature that we derived from SMARCB1 re-introduction experiments in rhabdoid tumor cell lines. 31 The two most significantly downregulated signatures in the SMARCB1 KO lines were from renal proximal tubules, the proposed origin of ccRCC ( Figure 4A). The loss of renal transcriptional identity followed a similar pattern to the chromatin accessibility changes: PAX8 KD alone showed a negative enrichment for the proximal epithelial signature C4 but failed to reach significance (p < 0.05) and P8 1 S1 1/2 (A) showed significant downregulation of the signature which reduced further over time (P8 1 S1 1/2 (LT)) ( Figures 4B and 4C). Similarly, PAX8 KD alone induced a positive enrichment of the rhabdoid SMARCB1 signature, but significance was only reached when SMARCB1 was also knocked out ( Figures 4A, 4D, and 4E). Upregulated and downregulated signatures derived from our RNA-seq data were also significantly positively and negatively enriched, respectively, in the SMARCB1 mutant ccRCC cell lines in the CCLE dataset, suggesting that a similar mechanism accounts for the PAX8 inhibition insensitivity in these models ( Figure 4F). In summary, PAX8 inhibition-resistant ccRCC cells display a global reduction in the kidney-specific cis-regulatory and transcriptional programs in favor of a dedifferentiated state which shares molecular features of SMARCB1 loss in pediatric rhabdoid tumors.
Acquired requirement of rare transcriptional dependencies in lineage factor inhibitionresistant ccRCC cells PAX8 inhibition resistance was associated with changes in the cellular lineage state, suggesting the possibility that the role of PAX8 in supporting ccRCC growth had been replaced by alternative transcriptional lineage factors. We therefore performed a second CRISPR-Cas9 screen targeting known and predicted TFs using the P8 1 S1 1/2 (LT) cell lines ( Figure 5A). As expected, constructs targeting essential genes were depleted and those targeting non-essential genes were neither enriched nor depleted ( Figure S6A). We identified three new dependencies which had no phenotype in the control cells, IRF2, BHLHE40, and ZNFX1 ( Figures 5B, S6B, and S6C), all of which were expressed in cells prior to SMARCB1 loss ( Figure S6D). IRF2 is a member of a TF family which regulates Toll-like receptor signaling, hematopoietic differentiation, and the expression of interferons (IFNs) and their target genes. 32,33 Similar to the role of PAX8 in renal development and ccRCC, IRF2 plays an important role in cancers originating from the plasma cell lineage (Figures 5C and S6E). In line with IRF2's role in regulating IFNs, compared to P8 1 Ctrl cells, there is a strong increase in both interferon-alpha and gamma gene sets from the hallmarks collection in P8 1 S1 1/2 cells (Figures 5D and S6F). BHLHE40 is a ubiquitously expressed stress-responsive transcription factor that is important in several physiological responses including differentiation, tumorigenesis, and response to hypoxia. 34 The mutation of VHL and the stabilization of HIF2A protein is a key tumorigenic event in ccRCC, and HIF2A perturbation RNA-seq has placed BHLHE40 downstream of HIF2A signaling. 35 In line with this, BHLHE40 dependency shows tissue specificity for RCC, and P8 1 S1 1/2 cells maintain strong HIF2A signaling when compared to P8 1 Ctrl cells (Figures 5E, S6G, and S6H). In the DepMap cohort, approximately half of the VHL mutant ccRCC lines are sensitive to BHLHE40 KO, and interestingly, this includes all the SMARCB1 mutant lines ( Figure 5F). Furthermore, the dependency of ccRCC cells on BHLHE40 anti-correlates with PAX8 dependency ( Figure S6I). ZNFX1 is a ubiquitously expressed, IFN-stimulated SF1 helicase capable of detecting viral dsRNA. Unlike IRF2 or BHLHE40, dependency on ZNFX1 is not associated with a particular lineage. Instead, there are a small number of cell lines across multiple lineages which show a strong dependency on ZNFX1, including TUHR10TKB, one of the three SMARCB1 mutant ccRCC lines ( Figure 5G). iScience Article PAX8 maintains ccRCC proliferative capacity by supporting the expression of MYC. 14 Furthermore, compared to P8 1 Ctrl(A) cells, P8 1 S1 1/2 (LT) cells showed increased MYC expression levels based on our RNA-seq data ( Figure 5H). This suggested that IRF2, BHLHE40, and ZNFX1 could contribute to P8 1 S1 1/2 (LT) proliferative fitness by maintaining optimal levels of MYC expression. We tested this possibility by targeting IRF2 and BHLHE40, the two strongest P8 1 S1 1/2 (LT)-specific hits from our TF screen ( Figure 5B) using two sgRNA constructs in P8 1 S1 2 (LT) cells and a wild-type Cas9-expressing 786-M1A-derived clone ( Figures S6J and S6K). As expected, IRF2 and BHLHE40 inhibition reduced the proliferation of P8 1 S1 2 (LT), but not the wild-type control cells ( Figure S6L). However, MYC expression was activated by inhibition of IRF2 and BHLHE40 in P8 1 S1 2 (LT) cells, but not in control cells ( Figure 5I). These results are in line with the observations that IRF2 and BHLHE40 can function as transcriptional repressors 36,37 and that cancer cells require an optimal level of MYC activity for maximal proliferative capacity. 38 De novo resistance to lineage factor inhibition across cancer types The finding that some ccRCC cell lines were insensitive to PAX8 inhibition without known inhibitory challenge on PAX8 ( Figure 1I) suggested that lineage factor independence could emerge naturally during tumor evolution and that this could be associated with specific molecular features. To test this possibility systematically, we developed an analytical approach to evaluate the prevalence of lineage factor independence and its molecular determinants across different cancer lineages in large-scale CRISPR-Cas9 loss-of-function data from the cancer DepMap dataset ( Figure 6A). Briefly, we identified lineage-specific  , P8 1 S1 1/2 (A), and P8 1 S1 1/2 (LT). (C) GSEA plot of Kidney proximal tubule C4 signature for Ctrl.Ctrl(A) vs. P8 1 S1 1/2 (LT). (D) SMARCB1 signature NES from GSEA, for Ctrl.Ctrl(A) vs. P8 1 Ctrl(A), P8 1 S1 1/2 (A), and P8 1 S1 1/2 (LT). (E) GSEA plot of SMARCB1 signature for Ctrl.Ctrl(A) vs. P8 1 S1 1/2 (LT). (F) Ridge plot of GSEA result from the comparison of ccRCC CCLE lines, SMARCB1 wild type vs. mutant from Figure 1I, using the SMARCB1, proximal tubule C3/C4 and upregulated and downregulated genes from Ctrl.Ctrl(A) vs. P8 1 S1 1/2 (LT  iScience Article TF dependencies by comparing the dependency score (CERES score) for each TF in a particular lineage against the CERES score for the same TF in cell lines pooled from all other lineages, creating a lineage dependency (LD) score for each TF in each lineage context. The distribution of the LD scores showed that for most TFs there was no specific dependency in a particular lineage, but there was a rare set of TFs which showed very strong specificity ( Figure 6B). We identified specific TF dependencies in 17 of 25 lineages (LD score < (À1.2), p < 0.05) ( Figure S7A).
The relative cellular dependency on individual TFs varied within lineages. In some instances, cells depended strongly on the lineage factor, for example PAX8 in RCC, MITF in melanoma, and IRF4 in multiple myeloma (MM) ( Figure S7A). In contrast, the dependency on TCF3 in acute lymphoblastic leukemia and non-Hodgkin lymphoma was considerably weaker ( Figure S7A). In addition, some of the TFs with low LD scores were pan-lineage dependencies that were particularly depleted in certain lineages. These genes had a low median CERES score across all cell lines, the clearest example being RELA ( Figure S7B); RELA was depleted in all 17 lineages but preferentially in MM ( Figure S7C). To focus the analyses on the strongest lineage dependencies while accounting for the possibility that there are de novo resistant cell lines, we included genes with an overall low CERES score (R50% of cell lines within a lineage with CERES score of % À0.5, Figure S7D) and excluded pan-cancer dependencies (median CERES % À0.2 across all cell lines, Figure S7E), leaving ten different lineages including hematological, epithelial, and neuroectodermal malignancies, sarcomas, and melanomas ( Figure S7F and Table S7).
We then sought to uncover examples of de novo resistance to lineage factor inhibition. The LD scores revealed a strong bimodal distribution ( Figure 6C), which was maintained at the cell line level with LD scores averaged for each cell line ( Figure S7G), supporting the idea that a subset of cell lines was resistant to lineage factor inhibition. Using a distribution-informed cutoff of average CERES score > À0. 45, we identified examples of lineage-resistant cell lines in all ten lineages ( Figure 6D), allowing the comparison between resistant and sensitive cell lines using permutation-based statistics ( Figure 6E). Interestingly, we detected instances of both acquired and lost dependencies in the lineage factor inhibition-resistant cell lines (Figure 6E). For example, in melanoma, TP53 KO cells are enriched among MITF-dependent cells, in line with its known tumor suppressive role, but in MITF-independent cells the enrichment is less strong (Figure 6F). The reduced CERES score for TP53 in MITF-independent melanoma cells suggests it may already be downregulated or inactivated, thereby facilitating a transition to an MITF-independent state. In line with this, there is an increase in the proportion of TP53 mutations in MITF-independent melanoma cell lines (47% vs. 28%) and a reduction in TP53 mRNA expression (Figures 6G and S7H).
Other examples include rhabdomyosarcoma cells resistant to PAX3 or both PAX3 and MYOD1 KO that had an enhanced dependency on MYC compared to lineage sensitive lines ( Figure 6H). The enhanced dependency on MYC raises the possibility that PAX3/MYOD1-independent cell lines have enhanced MYC activity which promotes lineage factor resistance. In line with this, MYC trends toward higher expression in lineage-resistant cell lines ( Figure S7I). In addition to changes in the genetic dependency profiles between resistant and sensitive cell lines, using mRNA expression data, it was possible to measure changes in the transcriptional programs. For example, differential expression analysis between lineage-resistant and sensitive neuroblastoma cell lines, coupled with gene set enrichment analysis using the hallmarks collection, revealed a very strong upregulation in the epithelial-mesenchymal transition signature ( Figure S7J), paralleling the observation that a mesenchymal-like primary neuroblastoma subgroup with features of highly aggressive mesenchymal glioblastoma exists in humans. 39 In summary, the prevalent lineage factor inhibition resistance across different cancer lineages is associated with specific molecular features and acquired genetic dependencies.

DISCUSSION
Transcriptional lineage factor dependencies are observed across a range of malignancies, making them an attractive target class for therapy development, but what maintains lineage fidelity in advanced iScience Article cancers, and how cancer cells react to long-term lineage factor inhibition have remained unclear. We demonstrate that ccRCC cells can overcome their dependency on the renal lineage factor PAX8 through a dedifferentiation process that can be enhanced by SMARCB1 loss. SMARCB1 maintains the kidneyspecific enhancer program and its inactivation results in the loss of renal transcriptional and epigenetic identity, altering the cellular context and reducing the requirement for PAX8. Two additional SWI/SNF complex members were enriched in our screen, including ARID1A, the loss of which can facilitate hormone therapy resistance in breast cancer. 20,21 SMARCB1 loss can also promote hormone independence in breast cancer cells, 20 and alterations in the SWI/SNF complex have been linked to androgen independence in prostate cancer. 40 This suggests that resistance mechanisms to lineage-targeted therapy may converge on SWI/SNF complex members and that our findings in ccRCC may be generalizable to other cancers. The SWI/SNF ATP-dependent chromatin remodeling complexes interact with various transcription and chromatin factors to regulate chromatin architecture and gene activation. 19 Three distinct SWI/SNF complex subtypes with characteristic subunit complements have been described: the BRG1/BRM-associated factor complexes (BAFs), the polybromo-associated BAF complexes (PBAFs), and the non-canonical BAF complexes (ncBAFs). 41 SMARCB1 is a member of the BAF and PBAF complexes but not of the ncBAF complex and it regulates enhancer activation states in various cell types. 31,41 Interestingly, SMARCB1 loss does not destabilize the BAF and PBAF complexes, but it changes their chromatin distribution through altering their interaction with nucleosomes. 42 Enhanced lineage switching upon SMARCB1 loss could reflect re-distribution of SWI/SNF complexes across the chromatin. More than 20% of human cancers harbor mutations in SWI/SNF complex members, but the mutation frequencies vary widely between different tumor types, 43 highlighting the relevance of SWI/SNF complex subtype-specific mechanisms in different cancers. We find that ARID1A and SMARCB1 mutations are associated with reduced PAX8 dependency in ccRCC cell lines. As our work focused on SMARCB1, additional experimental analysis would be needed to test whether ARID1A or other SWI/SNF complex members have similar functions in lineage fidelity maintenance in ccRCC. Mutations in ARID1A and/or SMARCB1 are also present in $5% in human ccRCCs, 44,45 indicating that reduced lineage factor dependency as described by our results may develop naturally in some ccRCCs. Moreover, PBRM1, another SWI/SNF complex member, is inactivated clonally in $40% of ccRCCs. 6 Possible effects of PBRM1 loss on lineage factor activity warrants further investigation in models that recapitulate the earliest stages of ccRCC development.
Biallelic SMARCB1 mutations are frequently observed in malignant rhabdoid tumors (MRTs) and atypical teratoid rhabdoid tumors, which are aggressive and poorly differentiated pediatric tumors that occur predominately in the kidney or soft tissue and central nervous system, respectively. 46 However, MRTs in the kidney originate from a different cell type than ccRCCs suggesting that the molecular similarities of PAX8 inhibition-resistant ccRCC cells and rhabdoid tumors is likely to reflect the shared SMARCB1 mutation status and general dedifferentiation rather than the acquisition of a rhabdoid ccRCC phenotype. 47,48 In line with this, the xenograft tumors formed by PAX8 inhibition-resistant ccRCC cells did not display histological features of rhabdoid dedifferentiation.
We find that PAX8 inhibition resistance in ccRCC cells is associated with a dramatic change in tumor histology with acquired features of morphological neuroendocrine differentiation, a phenotype not commonly seen in adult renal tumors. 49 However, even though some reports have described the expression of neuroendocrine markers in renal cancer, 50 the tumors formed by PAX8 inhibition-resistant cells did not specifically express molecular neuroendocrine markers. Interestingly, a similar phenotype has recently been described in a mouse-derived experimental Vhl-mutant renal cancer model. 22 Neuroendocrine differentiation is associated with androgen deprivation resistance in prostate cancer 51 and EGFR inhibition resistance in lung cancer. 52 Molecular features of neuroendocrine differentiation have also been more generally detected in a subset of different cancer types and in association with advanced disease. 53 Neuroendocrine differentiation and morphologically similar but molecularly distinct dedifferentiation processes as described here may therefore represent a broadly shared mechanism of resistance toward different growth inhibitory insults, ranging from inhibition of hormone and oncogene signaling to direct lineage factor inhibition.
Analogous to the identification of newly acquired dependencies in lineage factor-resistant ccRCC cells, the androgen-independent state in prostate cancer and neuroendocrine differentiation more generally can result in the acquisition of new transcription factor dependencies. 53,54 The newly acquired dependencies are unlikely to be selected purely stochastically. Rather, the high expression of IRF2, BHLHE40, and ZNFX1 already before PAX8 inhibition indicates that the cells were primed to become dependent on these factors upon lineage factor inhibition. Akin to the tissue-specific patterns of mutations in cancer, tissue-specific mechanisms are also likely to determine which acquired dependencies are most likely to emerge from lineage factor inhibition in different tissues. Our results suggest that IRF2 and BHLHE40 may contribute to the optimization of MYC levels in PAX8 inhibition-resistant cells, although other explanations for their role in maintaining proliferative fitness remain possible at this stage. An understanding of the underlying mechanisms could help predict and prevent resistance to lineage factor-targeted therapies. An alternative approach would be to target directly the possibly shared pro-tumorigenic pathways downstream of lineage factors and acquired dependencies.
Our systematic pan-cancer analysis detected frequent occurrence of de novo lineage factor resistance in the absence of direct lineage factor inhibition in most cancer lineages. This is supported by reports of lineage ll OPEN ACCESS iScience 26, 107360, August 18, 2023 iScience Article plasticity in response to multiple treatment modalities including chemotherapy, MAPK-targeted therapy, and immunotherapy, as well as environmental cues such as hypoxia and inflammation. 4,[55][56][57][58][59][60][61] We also identify shared molecular features among lineage-evaded cancer clones in several cancer types, suggesting that the development of lineage factor inhibition resistance can follow pre-determined molecular logic. However, in line with the detailed analysis of PAX8 inhibition resistance in ccRCC, resistance to lineage factor inhibition seems to develop via multiple molecular routes even within a specific lineage.
In conclusion, our study demonstrates that SMARCB1 is a key regulator of the renal enhancer program, which defines the context in which PAX8 is required for tumor growth. Resistance to PAX8 suppression can be achieved through multiple routes and it is linked to dedifferentiation, a dramatically altered tumor morphology and acquired dependencies on previously dispensable transcriptional regulators. The association between neuroendocrine differentiation and resistance to different genetic and pharmacological anticancer approaches indicates that therapeutic enhancement of lineage fidelity could be helpful in combatting acquired drug resistance in several different cancer contexts.

Limitations of the study
Our results are based on the analysis of human cancer cell lines and xenografts. Even though the cell lines carry genetic alterations that are also commonly seen in human tumors, we cannot exclude the possibility that our observations are specific to the cell lines studied. Detailed analyses on a larger set of cell lines and human tumors would be necessary to understand how broadly applicable our results are. Our analysis of the large cancer DepMap dataset that contains hundreds of cell lines, the largest currently available dataset, revealed putative mechanisms of lineage factor independence in cell lines derived from multiple tumor types. However, larger CRISPR-Cas9 datasets would be needed for more robust interrogation of lineage factor independence across cancers. Finally, the strongly dedifferentiated histological phenotype observed in the tumors formed by PAX8 inhibition-resistant cells was not associated with the expression of neuroendocrine markers, even though it exhibited morphological features of neuroendocrine differentiation. It therefore remains unclear whether the phenotype is truly related to neuroendocrine differentiation as described in other tumor types. Current therapies do not target renal lineage factors in the clinic. The molecular consequences of lineage factor inhibition in human patients thus remain unclear at this point.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following: Cells were transduced with a lentiviral library at a low MOI (<0.3) to ensure 1000x sgRNA representation. An MOI of <0.3 was used so that >85% of cells had a single sgRNA integration. After 48h following transduction, the cells expressing the integrated library were selected for with puromycin or hygromycin for 5 days. For doxycycline naive cells, the screen was initiated after antibiotic selection by supplementing the medium with 0.6mg/ml doxycycline to induce the expression of Cas9, otherwise, the screen was considered to have started 24h post-transduction. Cells were cultured for 17-21 days after screen initiation and two replicates at various time points were collected for each condition. For time points that required FACS, enough cells to ensure >130x coverage were harvested, otherwise, >500x coverage was maintained. Day 17 of the ll OPEN ACCESS iScience Article putative LDs for which the majority (>50%) of cell lines in their respective lineage had a CERES score of % -0.5. (3) Putative LDs which were pan-cancer dependencies but were more strongly depleted in a particular lineage were also removed. This was accomplished by using a plot of the distribution of the median CERES score across all cell lines for each putative LD. Based on the bimodal distribution of the data, a cut-off medianCERES >-0.2 was identified ( Figure S7E). After filtering, CRC predictions were available for 10/25 lineages. To identify lineage-resistant cell lines within each of the 10 lineages, a distribution of the averaged CERES scores of LDs in each cell line of their respective lineage was plotted ( Figure 6C). Based on the bimodal distribution of the data, a cut-off of average CERES score > -0.45 was used to identify lineage-resistant cell lines.

QUANTIFICATION AND STATISTICAL ANALYSIS
Statistical analyses were performed in R. The Kruskal-Wallis test was used for competitive proliferation assays and comparison of dependency data between subtypes. For Kaplan-Meier curves of tumour free progression, the logrank test was used. The hypergeometric distribution (phyper) test was used to measure significance of ATAC/DNAse I, gene set, and genomic region overlaps. Pearson correlation was used for correlation analysis. The Wilcoxon test was used for gene expression comparison. For all tests, a p-value of <0.05 was considered significant.