Structural and functional properties of mSWI/SNF chromatin remodeling complexes revealed through single-cell perturbation screens

Summary

The mammalian SWI/SNF (mSWI/SNF or BAF) family of chromatin remodeling complexes play critical roles in regulating DNA accessibility and gene expression. The three final-form subcomplexes – cBAF, PBAF, and ncBAF— are distinct in biochemical componentry, chromatin targeting, and roles in disease, however the contributions of their constituent subunits to gene expression remain incompletely defined. Here, we performed Perturb-seq-based CRISPR-Cas9 knockout screens targeting mSWI/SNF subunits individually and in select combinations, followed by single-cell RNA-seq and SHARE-Seq. We uncovered complex-, module-, and subunit-specific contributions to distinct regulatory networks, and defined paralog subunit relationships and shifted subcomplex functions upon perturbations. Synergistic, intra-complex genetic interactions between subunits reveal functional redundancy and modularity. Importantly, single-cell subunit perturbation signatures mapped across bulk primary human tumor expression profiles both mirror and predict cBAF loss-of-function status in cancer. Our findings highlight the utility of Perturb-Seq to dissect disease-relevant gene regulatory impacts of heterogeneous, multi-component master regulatory complexes.

eTOC Blurb

Otto et al. use Perturb-seq-based single-cell profiling to reveal that chromatin accessibility and gene expression is driven by unique subunits, modules, and subassemblies within the heterogeneous mammalian SWI/SNF family of ATP-dependent chromatin remodeling complexes.

Graphical Abstract

Introduction

Gene regulation is coordinated via intricate regulatory circuits, involving transcription factors (TFs), histone post-translational modifications, large chromatin remodeling complexes and general transcriptional machinery. Deciphering the molecular logic of such circuits informs our fundamental understanding of both normal cellular function and human diseases in which circuit disruptions alter chromatin topology and gene expression. Understanding gene regulatory wiring and function has been limited by the large number of participating components, including those that function in complexes, and their combinatorial activities which may be non-additive.

This is especially the case for key regulatory complexes such as the mammalian SWI/SNF (mSWI/SNF or BAF) complexes, multimeric ATP-dependent chromatin remodeling machines that alter chromatin accessibility and modulate gene expression ^,2. These complexes are composed of 10–15 subunits assembled combinatorically from 29 genes and exist in three distinct forms: canonical BAF (cBAF), polybromo-associated BAF (PBAF) and non-canonical BAF (ncBAF)^3,4 (Figure 1A). While mSWI/SNF subunits exhibit high-affinity interactions with one another and work in concert to alter chromatin accessibility, each subunit has unique roles⁵. This is reflected by (i) tissue-specific expression of certain mSWI/SNF subunits^6,7,8; (ii) coordinated subunit switching during development, such as in neuronal cell fate specification⁹; and (3) specific subunit mutations found in highly specific subtypes of human cancers and neurodevelopmental disorders^10,11,12.

Figure 1. An integrative, single-cell Perturb-seq approach to define mSWI/SNF regulatory function.

A. Experimental strategy for single mSWI/SNF gene Perturb-Seq in MOLM-13 cells. B-C. Distribution of (B) number of cells with 0–4 distinct guides captured per cell and (C) number of cells recovered per guide across n= 30,891 cells. D. Distribution of expression (x axis) of each gene (y-axis) in cells targeted by its guide (blue) or control cells (gray). Median (+/− 25% quartiles) with whiskers representing additional 1.5 interquartile ranges, and dots as outliers. **** P< 10⁻⁴, ns – non significant, two-sided Mann-Whitney-test with Bonferroni correction. E. Extent (dot color) and significance (dot size, -lop₁₀(B-H FDR), Kolmogorov-Smirnov (KS) test) of differential RNA expression vs. control cells (color bar) for each mSWI/SNF subunit for each subunit gene (columns) in all cells carrying guides targeting a specific subunit gene (rows). See also Figure S1, Table S1, S2.

Human genetic studies have unmasked extensive mutational frequency in individual mSWI/SNF subunit genes across several diseases, most notably, cancer and neurodevelopmental disorders. Such mutations are found in over 20% of tumors profiled, including some cancer types in which the mSWI/SNF perturbation represents the driver event. These have motivated efforts to define the mechanisms by which mSWI/SNF subunits contribute to disease. We and others have leveraged primary tumor genetics and cancer cell line models to identify the biochemical and gene regulatory impacts of individual mSWI/SNF subunit mutations in their relevant disease contexts^{13,14,15,16,17,10,18,19}. Furthermore, synthetic lethality screens have identified factors supporting tumor maintenance in mSWI/SNF-perturbed disease contexts^3,20,21,22. However, the gene regulatory mechanisms underpinning such dependencies remain largely unknown.

While such efforts have advanced our understanding of mSWI/SNF biology, a comprehensive evaluation of mSWI/SNF perturbations alone and in combination, in a unified context, has not been performed to date. Challenges include labor-intensive arrayed assays to assess the impact on gene expression for numerous perturbations, reduced long-term viability of cells with critical mSWI/SNF subunits deleted, compensatory processes enabling cell viability following deleterious perturbations, extensive clonal variation in single-cell clone-derived populations, and batch effects across assays^23,24,25.

Results

A Perturb-seq-based strategy to identify mSWI/SNF subunit contributions to gene expression

To define the roles of individual mSWI/SNF subunits, we used Perturb-seq to knock out 28 mSWI/SNF genes in an acute myeloid leukemia cell line, MOLM-13, chosen for its low mutational burden, lack of any mSWI/SNF gene mutations, and sensitivity to mSWI/SNF perturbations²³ (Figure 1A, Methods). We individually lentivirally packaged four guides targeting each of 28 mSWI/SNF genes, together with five negative control guides and pooled them as a library for transduction into Cas9-expressing MOLM-13 cells^26,27 (Figure 1A, S1A-B and Table S1; STAR Methods). We did not perturb ACTB due to its pan-essential nature and membership in multiple chromatin remodeling complexes²⁸. We sorted for BFP-fluorescent guide-containing cells two days after transduction, and following another 5 days of recovery, profiled them using droplet-based 3’ scRNA-Seq (Figure 1A, STAR Methods), using dial-out PCR to assign cells to perturbations. We recovered 52,152 high quality single-cell expression profiles (STAR Methods), with 30,891 (59%) annotated with a single guide per cell, resulting in a median of 1,093 cells per perturbed gene (258 per guide) (Figure 1B-C, S1C). We then normalized counts, regressed out technical covariates (STAR Methods), and standardized expression for each gene across all cells.

We confirmed the high quality of our data by multiple metrics. First, we detected only minimal batch effects (Figure S1D), which we regressed out as covariates (Methods). Second, for most guides, the expression of the targeted gene was significantly reduced (64 of 111, Benjamini-Hochberg (BH) FDR<1%, two-sided t-test) (Figure 1D-E, S1E), except for: guides that were ineffective (9 guides with poor sequence targeting, whose profiles differ from the majority of guides for a given gene; Figure S1E), guides that induced mutations not triggering nonsense-mediated decay^26,29,30 but showing consistent effects with other guides for the same gene (11 guides; Figure S1E) and guides targeting lowly-expressed genes (6 genes: DPF1, DPF3, ACTL6B, and SMARCD3 are not expressed in this cell line; GLTSCR1 and GLTSCR1L are lowly expressed). Third, cells with guides targeting the same gene had similar mean expression profiles (Figure S1E), with few exceptions, which we filtered out in downstream analyses (STAR Methods). As expected, guides targeting lowly-expressed subunits produced profiles most similar to control guides (Figure S1E). Finally, in certain cases, expression of mSWI/SNF genes themselves were altered upon knockdown of another subunit (i.e., ARID1A mRNA is elevated in the SMARCB1 knockout conditions), likely reflecting either indirect downstream effects or cellular feedback in response to perturbation, and presenting an additional layer of complexity in disentangling specific effects of a given mSWI/SNF perturbation, which we discuss below (Figure S1E). Taken together, these data establish a robust Perturb-Seq approach for the comprehensive evaluation of mSWI/SNF genes.

Perturbation of subcomplex-defining mSWI/SNF subunits impacts distinct expression programs

Correlation of mSWI/SNF perturbations by their mean expression profiles revealed three groupings of perturbations, consistent with specificity to mSWI/SNF subcomplexes (Figure 2A): (1) the cBAF-specific subunit, ARID1A, with functional core subunits SMARCC1, SMARCB1, SMARCE1, SMARCD2 and SMARCA4 (average Spearman $\text{ρ}=0.52$); ACTL6A and the cBAF-specific subunit DPF2 also correlated with this group, albeit to a lesser extent; (2) the ncBAF subcomplex, composed of the ncBAF-specific subunit BRD9 and SMARCD1, the only SMARCD paralog that can nucleate ncBAF subcomplexes ($\text{ρ}=0.51$); as expected, SMARCD1 perturbation also correlated weakly with the cBAF set, consistent with its ability to be incorporated into both complexes⁵; and (3) perturbations with weaker expression changes (or lower outlier scores, STAR Methods), including PBAF subunits BRD7, PBRM1 and ARID2 ($\text{ρ}=0.31$), control guides, and guides targeting lowly-expressed subunits. This suggests that PBAF complexes have a minimal impact on gene expression in this cell context (Figure 2A).

Figure 2. Distinct single-cell gene expression outcomes following differential mSWI/SNF subcomplex and paralog subunit perturbations.

A. Spearman correlation coefficient (color bar) between mean profiles of cells with each perturbation. Color boxes, bars: PBAF (red), ncBAF (green), and cBAF/functional core (blue) subunits. B. UMAP embedding of cell profiles (dots), colored by the density of cells with a perturbation specific to each of cBAF/Core (ARID1A, SMARCA4, SMARCB1, SMARCC1, SMARCD2), PBAF (PBRM1, BRD7, ARID2), ncBAF (SMARCD1, BRD9, GLTSCR1), or controls. C. UMAP embedding of cell profiles (dots, as in B) colored by clusters and labeled with perturbations whose cells are enriched in the respective cluster. D. Significance (signed -log₁₀(q-value), Fisher’s exact test) of enrichment (red) or depletion (blue) of cells with each perturbation (columns) in each cluster (rows) from C. E. Subunits (nodes) from the Core BAF (blue), ncBAF (green), PBAF (red) complexes, and peripheral BAF subunits (purple) connected by significant similarity in viability across perturbed cell lines based on Project Achilles (edges, left)²³ or correlation (Spearman’s $\text{ρ}$) in mean expression profiles in Perturb-seq (right, edges). Edge width: rank-normalized Pearson correlation for fitness network, rank-normalized Spearman correlation for Perturb-seq network. F. Top: UMAP embedding of all cell profiles (dots, as in B) colored by the density of cells with each perturbation (red and blue). Top: Significance (-log₁₀(P-value), x axis, Hotelling’s T² test) of difference in mean profiles between cells with perturbation of each paralog and control cells or cells with the other paralog perturbed (y axis). See also Figure S2.

Perturbations to different subcomplexes spanned different portions of the phenotypic space (Figure 2B, STAR Methods), with specific cell states enriched with cBAF-specific, ncBAF-specific and control guides (Figure 2B). cBAF perturbations deviated most from controls, while PBAF-perturbed cells deviated least (Figure 2B, S2C, STAR Methods). Partitioning the cell profiles into clusters (STAR Methods), cBAF perturbations were enriched in cell states 13 and 14, while ncBAF and PBAF loss biased towards cluster 2 and 4, respectively (Figure 2C-D). Notably, ACTL6A perturbed cells formed a distinct cluster 6, consistent with the fact that this is the only subunit (aside from beta-actin) that is also a member of other chromatin remodeler complexes (i.e., NuRD and INO80 complexes)³¹ (Figure 2C-D).

Interestingly, functional similarity between perturbations based on Perturb-seq generally agreed with fitness-based similarity derived from CRISPR-Cas9- and shRNA-based screens across hundreds of cancer cell lines^23,24 (Figure 2E, STAR Methods). Exceptions may possibly arise from expression changes not affecting cell fitness, a single cell context (in Perturb-Seq) versus multiple genetic and cell-of-origin contexts (in the viability assays), different perturbation timescales (7 days in Perturb-Seq vs. 40 days / 16 doublings in viability screens), or other factors³². Taken together, these findings highlight subunit- and complex-specific impacts on gene expression.

Subunits within the same paralog family coordinate gene expression changes of distinct cell state space and magnitude

We next compared the impacts of perturbing individual subunits within a paralog family (Figure 2F, S2D-E). Intriguingly, comparing ARID1A and ARID1B, subunits that nucleate cBAF complex assembly, only ARID1A-perturbed cells were enriched in cBAF-associated cluster 14, whereas ARID1B-perturbed cells distributed largely like control cells (Figure 2F). We observed a similar pattern with SMARCA4/SMARCA2 and SMARCC1/SMARCC2 paralogs (Figure 2D-F, S2D-E). Conversely, SMARCD1 and SMARCD2 paralogs had distinct distributions across clusters, with SMARCD1 predominantly in ncBAF-enriched cluster 2 (consistent with the key role of SMARCD1 in nucleating ncBAF complexes), and SMARCD2 in cBAF-enriched cluster 14 (Figure 2B-D, F). Collectively, these data demonstrate the differential impact of subunit paralogs on gene expression, with differences between subunits that fit the same position in the complex (ARID1A/B, SMARCA4/2, GLTSCR1/1L) and paralogs that nucleate different complexes (SMARCD1/2, SMARCC1/C2).

Convergent and divergent gene regulatory pathway impacts identified via individual mSWI/SNF complex subunit perturbation

Perturbation of mSWI/SNF subunits impacted diverse biological processes, including immune activation, signaling, proliferation, and developmental and tissue specific genes of neural, renal, and cardiac tissues (Figure 3A-B, S3A-C). We estimated the effect of each perturbation on each of 1,246 variable genes using a regularized linear model (STAR Methods), obtaining a regulatory matrix that we clustered to obtain 13 gene programs (P0-P12) (Figure 3A).

Figure 3. mSWI/SNF subcomplexes and their constituent subunits regulate distinct gene regulatory programs.

A. Left: Regulatory model coefficients (color bar) for the regulatory impact of each perturbation (row) on the expression of each gene (columns), clustered into programs of genes with similar patterns across perturbations. Right: Mean magnitude (dot color, color bar) and significance (dot size, -(log₁₀ (B-H FDR), KS-test) of changes in activity of each program (columns) in cells with each perturbation (rows) vs. control cells. B. Schematic effect (edges) of sets of perturbations (left) on gene programs labeled by their enriched GO terms (right). C. Shared and distinct genes (C) or biological processes (GO terms) regulated by perturbation in the indicated subcomplex subunits (C) or subcomplexes (D). See also Figure S3, S1.

Perturbation of complex-defining subunits affected gene programs concordantly regulated by all mSWI/SNF subunits and those that were subunit- or subcomplex-specific. First, perturbation of cBAF down-regulated three programs: “response to metal ions” (program 0), “DNA replication and repair” (program 1) and “cell cycle” (program 2). The role of mSWI/SNF in regulating cell cycle and associated processes such as DNA replication and repair has been well documented^33,34,35. Moreover, when considering the distributions of cells across the cell cycle, cBAF/core perturbations had the strongest impact, increasing the proportion of non-cycling cells from 12% in controls to 28–44% upon perturbation, with a more modest impact for ncBAF (9–18%) and PBAF (12–13%) (Figure S3D), consistent with the paucity of cell fitness effects observed across cancer cell lines for ncBAF and PBAF genes²³. The effects on cellular fitness mirror these results, with cBAF/core perturbations reducing fitness the most (Figure S1C). Finally, programs up-regulated upon loss of cBAF/core subunits were enriched for “leukocyte activation and adhesion” (program 12).

cBAF/core perturbations also uniquely upregulated an immune, leukocyte activation, immune development and differentiation, motility, and chemotaxis program (#11), potentially consistent with cBAF as a critical determinant in development-centered gene regulation, by establishing and maintaining DNA accessibility at distal tissue-specific enhancers^3,16,36,37 (Figure 3A-B). ncBAF perturbation uniquely up-regulated program 3, enriched for cytoskeleton, cell motility, adhesion, muscle differentiation, and regeneration genes. Finally, ACTL6A perturbation uniquely activated a program enriched for response to stimuli, metabolism and developmental genes (#10), perhaps owing to its additional roles in both NuRD and INO80 complexes (Figure 3A-B)³¹.

We next compared impacted genes between perturbations of each subcomplex and perturbation of the ATPase subunit, SMARCA4, which nucleates all three subcomplexes. Although SMARCA4-regulated genes accounted for a significant fraction of the genes affected by perturbing cBAF/ARID1A (42%, OR 23.3, p < 10⁻⁵, Fisher’s exact test), ncBAF/BRD9 (57%, OR 34.7, p < 10⁻⁵), or PBAF/ARID2 (67.4%, OR 32.3, p < 10⁻⁵) (Figure 3C), others were not impacted by SMARCA4 perturbation. These findings may be due to potential non-enzymatic mSWI/SNF functions, such as the interaction of subcomplex-specific subunits with TFs; compensation by SMARCA2, the SMARCA4 paralog, as has been previously suggested in studies implicating SMARCA2 as a top dependency in SMARCA4-mutant cancers^38,21,39; or false negatives (owing to limited power). In addition, there was a significant overlap (369 genes, OR 7.9, p<10⁻⁵) between genes impacted by cBAF and ncBAF complex perturbations, pathways which may underlie the unique synthetic lethal relationship between these complexes (strong ncBAF dependencies observed in all cBAF-perturbed cancer types studied to date)^3,20,40 (Figure 3D). In addition, 29 genes were regulated by perturbation in every subcomplexes and were enriched for IL-8 cytokine production. Finally, distinct gene signatures were associated with perturbation of subunits such as SS18, DPF2, and SS18L1, all of which had subtle effects in our dataset (Figure S3E-F). For example, while no genes are significantly affected by perturbation of SS18L1, genes with the largest effect sizes were involved in dendrite extension, corresponding to SS18L1’s known role in calcium-dependent dendritic outgrowth in cortical neurons (Figure S3E)⁴¹.

The knockout of some subunits (SMARCB1, SMARCE1, SMARCD2, SMARCC1) led to changes in the expression of genes encoding other subunits (Figure S1E), such that the observed effects of perturbations could be either due to the direct loss of the targeted subunit, or due to the combined direct and indirect changes in expression across multiple complex subunits. Case-by-case analyses revealed that minimal effects are attributable to these indirect effects. For example, indirect effects on the ARID1A and BRD9 expression appear in only 4 of 6 cBAF/core-specific perturbations, and thus cannot explain the shared signatures of cBAF/core loss in all six cases (Figure 3A, S1E). Moreover, indirect effects that decrease the expression of specific subunits are not associated with the signatures of the respective knockouts (Figure 3A, S1E).

Taken together, these analyses show that (1) perturbation of cBAF has the largest impact, especially via downregulation of developmental and differentiation-related genes; (2) the SMARCA4 ATPase supports the activity of all complexes, albeit to different extents; (3) while ncBAF and cBAF subcomplexes impact shared gene targets, their perturbation also leads to markedly distinct upregulated expression programs; (4) cBAF perturbation has the strongest impact on the cell cycle; and (5) perturbation of ACTL6A and cBAF are similar, except for a set of up-regulated genes in the ACTL6A perturbation, consistent with its roles in other chromatin remodeler families.

Loss of cBAF is associated with increased relative accessibility at ncBAF and PBAF target sites

Next, to relate the changes in gene expression upon mSWI/SNF complex perturbations to their underlying effects on chromatin, we performed SHARE-seq⁴², obtaining concomitant gene expression and chromatin accessibility profiles in the same perturbed cells (STAR Methods). We performed CRISPR-mediated knockout of ARID1A (cBAF), SMARCD2 (cBAF and PBAF), BRD9 (ncBAF) and SMARCA4 (ATPase, pan-BAF), along with a control guide, each with two biological replicates assayed in arrayed format, followed by pooling for SHARE-Seq. We recovered 1,044 cells with both high-quality ATAC-seq and RNA-seq (Figure 4A) and clustered them based on latent semantic indexing within the ATAC-seq dataset (Figure S4A, STAR Methods), recovering five cell clusters (Figure 4B-C).

Figure 4. Chromatin states and TF regulatory events coupled with mSWI/SNF perturbation-induced expression changes.

A. Schematic for SHARE-seq experiment. B. UMAP embedding of single-cell ATAC-Seq profiles (from SHARE-Seq, dots) colored by density of cells perturbed for different mSWI/SNF subunits or controls (color bar), or by Louvain-based clusters. C. Proportion of single cells (color bar) from each perturbation (rows, clustered) in each chromatin accessibility cluster (columns, as defined in B). D. Change in fragment length distribution between perturbed and control cells (y axis) for reads within (top, proximal) or beyond (bottom, distal) 1kb upstream of TSS. E. Schematic of strategy for identifying subcomplex-specific target sites by combining ChIP-seq and accessibility scores. F. Top: UMAP embedding of scATAC-Seq profiles (from SHARE-Seq, as in B) colored by cBAF- (left), ncBAF- (middle) or PBAF- (right) specific ChromVAR accessibility scores (color bar). Bottom: Distribution of complex-specific accessibility scores in individual cells, computed via ChromVAR (y axis, median +−25% quartiles, with whiskers representing 1.5 interquartile ranges and outliers shown as dots) for cells with different perturbations (x axis). *** P<10⁻³, **** P<10⁻⁴, ns – not significant. Mann-Whitney-Wilcoxon test, Bonferroni adjusted p-values. G. cBAF (x axis), ncBAF (y axis, left) and PBAF (y axis right) accessibility scores (ChromVAR) in individual cells (dots) perturbed by different perturbations (color legend). Top: Spearman correlation coefficient. H. Mean ChromVAR accessibility scores for different genomic regions (left, columns) or motifs significantly differing between any perturbation and control cells (middle, columns) and mean program scores in the SHARE-seq experiment (right, columns) in cells perturbed with different perturbations (rows). I. Spearman correlation coefficient (color bar) across single cells between gene program score profiles (columns) and different TF and chromatin feature accessibility score profiles (rows). CTCF motifs in loop anchors as defined in Rao et al., 2015). J. Summary of changes caused by perturbations to cBAF. K. Principal Components Analysis (PCA) of chromatin feature accessibility (acc.) score profiles and gene program score profiles across single cells. L. Significance (Bonferroni corrected signed -log₁₀(P-value), t-test) of chromatin accessibility changes at TF binding sites (x axis) and of changes in the expression level of the respective TF upon ARID1A knockout (y axis). See also Figure S4.

The perturbations differentially impacted single-cell chromatin accessibility (Figure 4B-C). Clusters 2 and 3 were assumed upon cBAF perturbations (ARID1A and SMARCD2), cluster 1 by ncBAF (BRD9) disruption, and SMARCA4 disruption spanned all three clusters (Figure 4B-C). Further, perturbations to ARID1A/SMARCD2 cBAF/core and SMARCA4 decreased the fractions of sub-nucleosomal fragments and increased mononucleosomal fragments at distal sites (Figure 4D), consistent with the preferential distal binding of cBAF. Instead, at promoters, both sub-nucleosomal and mononucleosome-associated fractions increased upon cBAF/core and SMARCA4 loss. Notably, ncBAF loss showed minimal changes at both distal and proximal sites, with moderate reduction in accessibility near promoters, consistent with its promoter-proximal localization³.

Accessibility changes were concordant with ChIP-seq derived chromatin binding profiles in MOLM-13 cells for cBAF (DPF2), ncBAF (BRD9), PBAF (BRD7) and SMARCA4 (Figure 4E-G). Perturbation of cBAF components or SMARCA4, but not the BRD9 ncBAF component, were associated with decreased accessibility at loci uniquely bound by cBAF (Bonferroni adjusted P-value =1.5*10⁻²¹, 9.3*10⁻¹⁰ (cBAF, ARID1A and SMARCD2 respectively), 5.1*10⁻⁶ (SMARCA4), and 1 (BRD9), two-sided Mann-Whitney test) (Figure 4E-G, S4B). cBAF perturbations specifically reduced accessibility at motifs for AP-1 factors (FOS, JUN), BATF, JDP2, BACH1, BACH2, NFE2, CEBPA, CEBPD, and CEBPG (Figure 4H, Figure S4B), consistent with known relationships between cBAF complexes and TFs ^43,44,45.

Interestingly, perturbation of cBAF components not only decreased accessibility over cBAF target sites, but also increased the relative accessibility over both ncBAF and PBAF target loci (Bonferroni adjusted P-value = 7.9*10⁻²⁰ and 1.2*10⁻²³ for ARID1A and SMARCD2 perturbations respectively at ncBAF sites, 7.43*10⁻¹⁴ and 1.17*10⁻¹⁵ at PBAF sites) (Figure 4F, S4C). Furthermore, CTCF motifs and loop anchors were, respectively, the top motifs and chromatin features with gains in accessibility upon cBAF perturbation (Figure 4H, S4C), suggesting that cBAF loss increases ncBAF-mediated accessibility over its top-targeted motif^46,47. Finally, cBAF- and ncBAF-dependent accessibility were significantly anti-correlated across single cells (Spearman’s $\text{ρ}=-0.32$, p=4.98*10⁻²⁷, Figure 4G, top); a weaker anticorrelation was identified for cBAF- and PBAF-dependent accessibility (Figure 4G, bottom). These data support a model in which compromising cBAF alters the assembly and/or function of ncBAF and PBAF complexes, increasing their resulting relative accessibility over their target sites.

Accessibility changes over specific TF motifs underlie impact on distinct expression programs in single cells

While Perturb-Seq screens can systematically identify expression changes governed by transcriptional and other regulators²⁶, they cannot readily distinguish direct from indirect effects or identify the mediating mechanisms⁴⁸. We thus leveraged the SHARE-Seq data to integrate changes in mSWI/SNF-mediated accessibility with target gene expression in the same single cells, to directly link their impact on or relation to underlying mechanisms.

First, we tested the Perturb-SHARE-seq data for co-variation between the single-cell score of each of 13 expression programs defined in the initial Perturb-Seq analysis (Figure 3A, S4D) and TF accessibility scores in the same single cells from ATAC-seq. Reduced accessibility at cBAF-specific binding sites under cBAF perturbations involves loss of accessibility at distal elements and at motifs for AP-1 factors, CEBP, BATF, and BACH, and reduced scores for expression programs enriched for response to metal ions, DNA replication and repair and cell cycle genes (programs #0, 1 and 2) (Figure 4H-K). Concomitantly, there was increased accessibility at ncBAF-, PBAF-, and non-BAF sites, at TSS and loop anchors, and at motifs for CTCF, TCF, ZEB1, SNAI, NFATC3, ELF1 factors, and increased expression of programs enriched for leukocyte activation, adhesion, immune development and differentiation, motility and chemotaxis (programs #11 and 12) (Figure 4I-K). Perturbation of ncBAF (BRD9) specifically enriched for gene program #3, involving cytoskeleton, cell motility, adhesion, muscle differentiation, and regeneration, which was coupled with reduced CTCF accessibility (Figure 4H-J).

We next related changes in the accessibility of TF motifs with expression of the cognate TF gene at the single cell level, across perturbations (Figure 4L). Following cBAF perturbation, mRNA levels of BATF, BATF3, CEBPA, ATF4, ETV4 and GATA1 were reduced, as was accessibility at their corresponding motifs, whereas expression and motif accessibility of CTCFL and SNAI2 were increased (Figure 4L). For other TFs (AP-1 factors, JUN, FOSL2, JUND and NFE2 as well as RUNX2, ZBTB18 and NFIL3), decreased motif accessibility was associated with increased TF mRNA expression, possibly owing to compensatory mechanisms.

Combinatorial mSWI/SNF perturbations identify paralog family functions and genetic interactions

To study genetic interactions, we measured a series of arrayed combination perturbations focusing on four categories of subunits: (1) all subunits of a given paralog family; (2) all subunits unique to a specific subcomplex; (3) subunits present in two of three subcomplexes; or (4) core subunits present in more than one complex coupled with depletion of a unique complex component (Figure 5A). We leveraged the initial screen to identify top-performing guides for the Combo-Perturb-Seq experiments, and targeted recovery of a smaller number of cells per combination (~100 cells per perturbation), based on power analysis (Figure S5A-C). As before, there were two strong classes of gene program effects, mapping to perturbation of either cBAF or ncBAF (Figure 5B,C, S5D,E).

Figure 5. Combinatorial Perturb-Seq reveals additive and synergistic roles for mSWI/SNF paralogs.

A. Overview of combinatorial Perturb-seq experiment. B. UMAP embedding of all cell profiles from the combinatorial Perturb-Seq experiment colored by density of cells with single perturbations from each subcomplex/functional group or controls. C. Magnitude (dot color) and significance (dot size, -log₁₀(B-H FDR), KS test) for score changes in cells from each perturbation condition (rows) vs. control cells in each of 18 gene programs (columns) derived from Combo-Perturb-seq. D-F. Top: UMAP embedding of all cell profiles from a combinatorial Perturb-Seq experiment colored by density of cells with specific perturbations. Bottom left: Mean expression of each gene (dot) as observed in the combinatorial perturbation data (y axis) or predicted under an additive linear model from the constituent single perturbations experiment data (x axis). Bottom Right: Linear model coefficients (color bar) for the overall regulatory effect (sum) and its portion attributed to single perturbations or their combination (interaction term) (columns) for each gene (rows). Side color bar: orange: synergistic interactions; gray: additive effect. Black horizontal lines separate interaction classes²⁶. G. Spearman correlation coefficient (color bar) between ncBAF single and combinatorial perturbation profiles. See also Figure S5, S6.

Next, we distinguished between additive and non-linear effects of different subunit pairs both at the global level and at the level of individual gene expression. To identify non-linear genetic interactions on the expression of the 1,761 variable genes in the combinatorial experiment, we fit linear models with interaction terms²⁶ (Methods). To identify global patterns of genetic interactions (using the entire profile as a phenotype), we tested the extent to which the mean expression profile of a combinatorial perturbation is predicted from a linear combination of the mean profiles of the constituent individual perturbations; additive combinations will be well predicted, while synergistic ones will deviate from the predictions of an additive model⁴⁹ (Figure 5D-F, scatterplots). While the global test uses the expression profile as a cellular phenotype, the same combination of perturbations can have different effects on different individual genes^26,50 and thus the analyses are complementary.

Focusing first on subunits of a given paralog family (category (1) above), Combo-Perturb-Seq revealed functional features of paralogous subunits, including those relevant in cancers driven by dual loss of both paralogs, such as small cell carcinoma of the ovary hypercalcemic type (SCCOHT; SMARCA4/2 loss) and dedifferentiated endometrial cancers (ARID1A/B loss)⁵¹. For example, perturbation of both paralog ATPases, SMARCA4 and SMARCA2, showed substantial interaction terms (Figure S6A, bottom left), with one third of the impacted genes (33.3%, 35/105 genes) observed only upon dual perturbation (Figure S6A, right). These genes were enriched for protein and RNA localization to telomeres and Cajal bodies (Figure S6A, right). Thus, the likely complete depletion of ATPases halts catalytic activity from all three mSWI/SNF subcomplexes, as is observed in cancers such as SCCOHT and SMARCA4-deficient thoracic sarcoma (SA4DTS), the only settings driven by dual deletion of mSWI/SNF ATPases³⁶.

There were also extensive genetic interactions following dual perturbation of SMARCC1 and SMARCC2 (54% of the 300 genes with an effect in either SMARCC1, SMARCC2 or combination loss, Figure 5E), impacting immune response, response to metal ions, cell-cell adhesion, telomere maintenance and protein folding (Figure 5E right). SMARCC1 and SMARCC2 assemble as either heterodimers or homodimers within complexes (except ncBAF, which only carries SMARCC1 homodimers), and their joint perturbation thus compromises biochemical integrity of all mSWI/SNF complexes⁵. However, as in the single perturbation experiment, loss of SMARCC1/2 subunits did not activate ncBAF loss-associated gene program 4 (Figure 5C). Similarly, joint perturbation of SMARCD1 and SMARCD2, each part of the initial structural core of each of the mSWI/SNF complexes, also led to substantial genetic interactions (Figure S6B), including for many individual genes (23.6% genes with interaction terms of 403 genes impacted by either perturbation alone or in combination), affecting DNA repair, DNA metabolic processes, protein folding and chemotaxis, (Figure S6B, right).

Conversely, perturbations of ARID1A and ARID1B had few interaction terms and were largely additive (Figure 5D). However, because the effect of ARID1B perturbation alone is modest (possibly due to its lower expression; Figure 2F, S2E), dual perturbations resembled ARID1A perturbation alone, making it difficult to determine if the combinatorial effects are truly additive (Figure 5D, top left). Interestingly, the dual perturbation of ncBAF-specific subunits GLTSCR1 and GLTSCR1L, which alone had modest effects (Figure 5B, 3B), led to an expression profile more similar to that of BRD9-based perturbation of the ncBAF subcomplex (Figure 5F-G).

Collectively, across all pairs evaluated, interaction terms were typically consistent with the direction of the individual perturbation terms, suggesting lesser antagonism between paralog family members. We find ARID1A/ARID1B and GLTSCR1/GLTSCR1L to act additively on most genes, with all other pairs having a substantial number of positive (synergistic) interactions (Figure 5D-F, S6A-D). A subset of combinations resulted in low numbers of cells, making it difficult to conclusively distinguish between additivity and synergy, specifically for the SS18/SS18L1, DPF2/PHF10 and ARID1A/ARID1B/ARID2 conditions (Figure S6C-E).

Combinations simultaneously targeting one or two subcomplexes reveal strong modularity with intra-complex but not inter-complex genetic interactions

We further leveraged combinatorial perturbations to assess (i) how subcomplex-specific subunits contribute to complex function (by perturbing all unique subunits of a complex type (category (2) above)); (ii) how core subunits impact the functions of each complex (by perturbing subcomplex-specific subunits along with a core subunit (category (4) above)); and (iii) how individual complexes interact (by perturbing subcomplex-specific subunits belonging to two different subcomplexes; category (3) above)) (Figures 5A, 6A).

Figure 6. Intra-complex combinatorial perturbations yield assembly-related genetic interactions, but inter-complex combinations are mostly additive.

A. Chart depicting combination perturbations examined. B. PCA of intra-complex combinatorial perturbation pseudo-bulk expression profiles, with vectors representing difference between each single or combinatorial perturbations and control, colored by complex (as in A). C,D.. Top: UMAP embedding of all cell profiles from a combinatorial Perturb-Seq experiment colored by density of cells with specific perturbations. Bottom left: Mean expression of each gene (dot) as observed in the combinatorial perturbation data (y axis) or predicted under an additive linear model from the constituent single perturbations experiment data (x axis). Bottom Right: Linear model coefficients (color bar) for the overall regulatory effect (sum) and its portion attributed to single perturbations or their combination (interaction term) (columns) for each gene (rows). Side color bar: orange: synergistic interactions; gray: additive effect, purple: buffering effect. Black horizontal lines separate interaction classes²⁶. E. Magnitude (dot color) and significance (dot size, -log₁₀(B-H FDR), KS-test) of difference in activity scores between perturbed and control cells (columns) for gene sets up- or down-regulated by different subcomplexes (rows). See also Figure S6.

We first assessed the role of subcomplex-specific subunit combinations (Figure 6B). For cBAF, both dual ARID1A/B and triple ARIDA/B/DPF2 perturbations resulted in highly similar profiles (Figure 6B), with few interaction terms for individual genes between the DFP2 perturbation and the dual ARID1A/B perturbation (Figure 6C), suggesting limited additional contribution from DPF2 loss in this setting. This agrees with biochemical studies suggesting that DPF2 requires an ARID1 member (ARID1A/B) to assemble into cBAF complexes⁵. In addition, whereas perturbation of PBAF subunits ARID2, PBRM1, or BRD7 alone resulted in similar gene expression profiles, their combined loss, which results in a completely absent PBAF complex, yielded a distinct ‘opposing’ profile that also differed from that of other subcomplexes (Figure 6B). Together, these findings suggest that the effects of perturbing mSWI/SNF subunit combinations are tightly linked to mSWI/SNF modular assembly⁵.

We next studied subunits that are part of multiple subcomplex types. We defined subcomplex-specific gene signatures as genes affected by perturbing only subcomplex-defining subunits (STAR Methods) and then scored the impact of each of the combinatorial perturbations on these genes (Figure 6E). Intriguingly, perturbing subunits shared between subcomplexes often impacted signatures associated with all of their constituent subcomplexes, albeit to differing extents (Figure 6E). For example, SMARCD1 predominantly impacted ncBAF signatures, consistent with it being the only paralog that can incorporate into ncBAF, but one of two paralogs that can integrate into cBAF and PBAF complexes³. SMARCC1/SMARCC2, SMARCA4/SMARCA2, and GLTSCR1/GLTSCR1L each showed stronger effects on the corresponding complex signatures when perturbed in combination. SMARCA2/4 dual perturbation affected the signatures associated with all three complexes: while cBAF signatures were impacted by SMARCA2 alone, ncBAF and PBAF signatures were affected only in the combination. SMARCB1 and SMARCE1, as expected, affected both cBAF and PBAF signatures (Figure 6E).

Finally, perturbations of subunits from different subcomplexes led to largely additive effects (albeit based on a small number of doubly perturbed cells) (Figure 6D). This was the case when perturbing both cBAF and ncBAF (Figure 6E), or cBAF and PBAF (Figure 6D, S6E,F). Thus, despite the major cellular role of these complexes, combined perturbation of multiple subcomplexes led to largely additive phenotypes highlighting their modularity. Our analysis helped determine the impact of specific subunit losses, and their selective, modular effects on the gene regulatory activities of each mSWI/SNF subcomplex.

Perturb-seq signatures identify and predict mSWI/SNF-mutant human tumor features

mSWI/SNF genes are mutated in ~20% human cancers (TCGA), as well as in ~100% of some rare cancer types where they represent driving, hallmark features^35,52. However, comprehensive identification of signatures hallmark to mSWI/SNF-mutant tumors has been challenging, given the multiplicity of other mutations present in most tumors as well as challenges in analyzing bulk RNA-seq data, that is a mix of malignant and microenvironment cells. We thus set to (1) test if the well-defined programs from our screens correspond to expression signatures in cBAF-mutant primary tumors; and (2) identify cBAF-mutant cancers based on their bulk RNA-seq expression profiles (Figure 7A). We trained random forest classifiers on our Perturb-seq dataset to predict the genetic perturbation from single cell expression profiles, using NMF-defined programs as features. We defined cBAF loss signatures using the cBAF perturbation predictions (STAR Methods) (auROC on held-out cells = 0.69, Figure S7A).

Figure 7. mSWI/SNF subunit perturbation signatures classify human cancers with subunit disruptions or convergent mechanisms.

A. Schematic of classification strategy. B. Distribution of cBAF loss Perturb-seq signatures scores (y axis, median (+/− 25% quartiles), whiskers represent additional 1.5 interquartile ranges, and dots represent outliers) in cBAF KO or control single cells (x axis, blue) or in bulk RNA-seq profiles of mSWI/SNF perturbation-driven rare cancers (x axis, orange). *** P<10⁻³, one-tailed Welch’s t-test. C. True (y axis) and false (x axis) positive rate for classifying cBAF mutant tumors across all TCGA tumors based on cBAF loss Perturb-seq signatures. D. Distribution of cBAF loss Perturb-Seq signature scores (x-axis) for all (blue) and cBAF mutant (yellow) TCGA tumors. Dashed line: threshold for calling tumors with cBAF loss signatures. E. Significance of enrichment (Benjamini-Hochberg FDR, Fisher’s exact test) of mutation rate (vs. mutational background within tumor class) and proportion (green/white color bar) of mutations in different categories (columns) for genes (rows) with increased mutation rate in tumors with high cBAF loss signature scores but no BAF mutations. F. Significance (y axis, -log₁₀(P-value), one-sided Wilcoxon rank-sum test) of association of TF regulators with tumors with high cBAF loss signature scores but no BAF mutations (x axis, rank ordered by significance). G. Log fold-change in viability (y axis) in response to drugs that specifically reduce viability in BAF mutant cell lines for cell lines with BAF mutations (red), BAF loss-like mutations (pink) and all other cell lines (grey) (* * * q < 0.001, * * 0.001 < q < 0.01, * 0.01 < q < 0.05, t-test). See also Figure S7, Table S3-S5.

We first confirmed the validity of our screen-based signatures, by correctly classifying a compendium of rare cancers driven by cBAF/mSWI/SNF perturbations, including malignant rhabdoid tumors (MRT; SMARCB1 loss), epithelioid sarcomas (EpS; SMARCB1 loss), renal medullary carcinoma (RMC; SMARCB1 loss), SCCOHT (dual SMARCA4/2 loss), and SA4-DTS (dual SMARCA4/2 loss)¹⁸. These cancers scored strongly for cBAF loss signatures (p < 0.001, one-tailed Welch’s t-test), consistent with their driver mutations (Figure 7B). In addition, the strongest cBAF loss signatures were in SA4-DTS and SCCOHT, cancers lacking activity of all three mSWI/SNF complexes due to loss of both SMARCA4 and SMARCA2 ATPases ^36,53.

Next, we applied our classifier to bulk RNA-seq profiles from over 6,000 tumors in TCGA. The cBAF loss signatures were highly predictive of the 2,305 cBAF-mutant tumors found in TCGA (Figure 7C, auROC 0.73; random auROC=0.5; Table S3). Another 70 tumors spanning multiple tissues (Figure S7B) were classified with strong cBAF loss signature scores (score > 0.46, top 10% of all cBAF loss signatures), comparable to those of the rare BAF-driven cancers (Figure 7B), but did not contain any mSWI/SNF gene mutations (Figure 7D, Table S3), suggesting convergent mechanisms of cBAF loss-of-function.

These 70 tumors included mutations in sequence-specific TFs, kinases, phosphatases, RNA-binding proteins, and GTPases, and were strongly enriched for mutations in CHD2, EHMT1, and RAPGEF6 (FDR<0.01, Fisher’s exact test, Figure 7E, Table S4, Methods), suggesting there may be previously unidentified interactions between these factors and BAF activity, converging to similar expression effects. Furthermore, 67.4% of the mutations in these genes were missense in nature, suggesting potential opportunities to understand the domain-level contributions of these factors, as they relate to potential modification of the genomic targeting, activity or binding interactions of cBAF complexes in either a direct or an indirect manner (Figure 7E, Table S4).

Tumors with cBAF-loss-like signatures also had reduced expression of 275 genes compared to other tumors of the same class (Benjamini-Hochberg FDR q < 0.05 Welch’s t-test, Table S5). Using Lisa⁵⁴ we identified several regulators whose targets defined from published histone mark and TF ChIP-seq data are enriched in down-regulated genes (Figure 7F, Table S5). Intriguingly, these TFs include ERG, which has been shown to hijack mSWI/SNF activities to achieve cancer-associated gene expression in prostate cancer⁴³ and MAX, which has been proposed to act as a mediator between MYC – SWI/SNF gene programs in small cell lung cancer⁵⁵, as well as GATA3 (Figure 7F, Table S4 and S5). Notably, recent evidence suggests that methylation-mediated silencing of mSWI/SNF subunits can also contribute to oncogenic signatures, which could also explain correlation to mSWI/SNF loss signatures in our dataset^56,57. Taken together, the enriched mutations and expression patterns in cBAF loss-like tumors present a valuable foundation upon which to define disease mechanisms in human tumors that may converge on mSWI/SNF complex functional alteration.

Finally, we studied the relevance of these possible convergent disease mechanisms to cancer therapeutics using the PRISM drug repurposing dataset⁵⁸. We compared viability following exposure to the library of cancer therapeutics for cell lines harboring deleterious cBAF mutations, cell lines with deleterious mutations enriched in tumors with cBAF loss-like signatures (“cBAF loss-like mutations”, FDR < 10% from Figure 7E, Table S4), and cell lines without either of these sets of mutations (Figure 7G, S7C). For the 18 drugs that specifically reduced the viability of BAF-mutant cell lines relative to cell lines without BAF mutations (p < 0.01, one-tailed Welch’s t-tests), there was a correlation between the differential viability of these drugs in the cBAF-mutant and in cell lines with cBAF loss-like mutations (Pearson’s r = 0.55, Figure S7C). Furthermore, 9 of the 18 drugs also specifically inhibited the growth of cell lines with BAF loss-like mutations (Figure S7C, p < 0.05 one-tailed Welch’s t-test). For example, there was reduced proliferation of cBAF-mutant and cBAF loss-like mutant cell lines upon treatment with hydroxyfasudil and the natural anti-cancer agent sodium tanshinone IIA sulfonate (Figure 7G). These drugs inhibit the RhoA/ROCK and YAP/TAZ pathways, respectively, which have been shown to be functionally dependent on SMARCA4/SMARCA2 mSWI/SNF ATPases ^59,60. Such findings highlight the importance of defining convergent mechanisms leading to mSWI/SNF disruption that may be targetable by existing small molecules.

Discussion

We present a large-scale effort to dissect the contributions of each subunit, functional module, and paralog family within the mSWI/SNF complex in one cell context. Our findings highlight the utility of single-cell expression signatures derived from comprehensive Perturb-seq experiments spanning an entire family of chromatin remodeling complexes as a resource for mining disease-associated expression programs in large tumor sequencing databases, enabling the discovery of new or converging mechanisms and therapeutic vulnerabilities.

Combining single-cell expression data with chromatin accessibility, we found that cBAF depletion increased relative accessibility over ncBAF and PBAF sites, suggesting that complex function (via biochemical abundance or activity on chromatin) can be “shifted” upon depletion of a specific subcomplex type, consistent with our^40,3 and others’²⁰ recent results. Intriguingly, the effects of BRD9 perturbation were subtle at the chromatin level, albeit in the expected direction, involving loss of accessibility at ncBAF-specific sites and CTCF binding sites. Future work is needed to further investigate the chromatin accessibility changes induced by ncBAF perturbation, especially given the low cellular abundance of this complex type.

We characterized additive and synergistic effects of mSWI/SNF paralog family perturbations with ARID1A/B, GLTSCR1/L1, and SS18/SSL1 acting largely additively and most other paralog pairs displaying synergistic effects, where each individual perturbation had little effect but the joint perturbation had a strong effect (Figure 5E, S6A,B). This is consistent with a model of paralog structural and/or functional redundancy (at least in a given cell line context). We observed a unique behavior of the GLTSCR1/1L subunits, where the mild effects of each single perturbations did not lead to a distinctive profile, but their combined loss yielded a profile more similar to that of ncBAF perturbed cells. These results are consistent with functional ncBAF complexes forming in WT MOLM-13 cells (also supported by ChIP-seq and similarity between ncBAF-specific BRD9 and SMARCD1 perturbations), and requiring loss of both GLTSCR1/1L subunits for loss of activity.

Notably, combinations of perturbations of subunits unique to different complexes led to largely additive effects, in contrast to the non-additive (synergistic) interactions for submits from the same complex. This suggests the prevalence of intra- over inter-complex interactions and a largely modular organization for genetic interactions, such that interactions (when present) are restricted within modules, also observed in other systems⁶¹. If this pattern generalizes to other complexes and regulatory modules, it should greatly simplify the screening for genetic interactions that are often too cumbersome to screen comprehensively by experiments alone. Such constraints may also help test for genetics interactions in human genetic data. In addition, future work is needed to characterize the complex relationship between gene-expression-based phenotyping and fitness: for example, cBAF and ncBAF have consistent genetic interactions, but show additive effects on expression levels.

Finally, our “mSWI/SNF perturbational atlas” identified tumors among the 10,000 tumors in TCGA that contain mutations in pathways that may impinge upon mSWI/SNF (specifically, cBAF) activity. While exploratory, these analyses revealed that mutation or loss in expression of TFs that have been shown to bind to and hijack BAF complexes in cancer settings (such as FLI1, which forms a fusion oncoprotein in Ewing sarcoma⁶², or the ETS factor, ERG, which is overexpressed in prostate cancers⁴³) can give rise to cBAF loss-of-function expression signatures, highlighting the important cooperativity between this mSWI/SNF family subcomplex and TFs shown in several contexts⁴⁵. Further, upon cBAF disruption, we observed increased accessibility over motifs such as CTCF, SPI1, and CEBPB, which also exhibited increased expression in tumors with cBAF loss-like signatures but lacking mutations in cBAF genes, suggesting alternative convergent pathways. These highlighted mutations and TFs will need to be further studied to determine whether they interact with or somehow intersect with mSWI/SNF activities. One potential mechanism contributing to “cBAF loss-like” signatures could be DNA methylation, which has been identified as a mechanism of silencing of mSWI/SNF subunits^56,63. This is especially important for understanding the contribution of mSWI/SNF function to the pathology of mSWI/SNF-driven in vivo tumors, relative to cellular models such as the one studied here, in which loss of BAF activity decreases fitness, pointing to potential addiction to these complexes.

Taken together, our datasets and findings provide a foundation for understanding the differential functions of and relationships between mSWI/SNF complexes and their constituent subunits, the logic of mSWI/SNF subunit and complex perturbations, and gene sets that mirror associated disease biology and the impacts of therapeutic disruption. We hope that this resource will facilitate new hypotheses in gene regulation and translational medicine.

Limitations of the Study

All Perturb-Seq and SHARE-seq experiments were carried out in just one cell line of AML origin, MOLM-13, which has known dependencies on cBAF and ncBAF complexes; further, SHARE-seq experiments were only performed with a subset of mSWI/SNF perturbations. While our data suggests that gene regulatory profiles derived from our experiments have meaningful pan-cancer/cell line utility, specific findings may be unique to this cell line and linkage between gene expression and chromatin landscape features may be incomplete. As an example, while PBAF-specific subunit perturbations had more subtle effects in this study, its role in regulating gene expression in different contexts or during dynamic processes may be highly context-dependent and thus remain to be uncovered. Indeed, in a recent Perturb-seq study performed in the K562 cell line, more pronounced effects were reported for PBAF subunit disruption (Figure S2F)⁶⁴. Finally, additional chromatin-level and single-molecule studies will be needed to better define direct ncBAF-mediated changes to chromatin architecture and gene expression.

STAR Methods

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Cigall Kadoch (Cigall_kadoch@dfci.harvard.edu).

Materials Availability

This study did not generate unique new reagents.

Data and Code Availability

Sequencing data were deposited at GEO under accession number GSE200201 (link: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE200201. Accession numbers are listed in the Key Resources Table. Original Western blot images have been deposited at Mendeley and are publicly available as of the date of publication. The DOIs are listed in the Key Resources Table.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Mouse monoclonal anti-TBP	Abcam	Cat# Ab51841
Rabbit monoclonal anti-ARID1A	Cell Signaling Technologies	Cat# 12354
Mouse monoclonal anti-SMARCA4	Santa Cruz Biotechnology	Cat# sc-17796
Rabbit monoclonal anti-SMARCC1	Cell Signaling Technologies	Cat# 11956
Mouse monoclonal anti-SMARCD1	Santa Cruz Biotechnology	Cat# sc-135843
Mouse monoclonal anti-FLAG M2	Sigma Aldrich	Cat# F1804-50ug
Bacterial and virus strains
One Shot Stabl3 Chemically Competent E-coli	Invitrogen	Cat# C737303
Chemicals, peptides, and recombinant proteins
T4 DNA ligase	New England Biolabs	Cat# M0202L
Polyethylenimine (PEI)	Thermo Fisher Scientific	Cat# NC1014320; CAS: 9002-98-6
T4 PNK	New England Biolabs	Cat #M0201L
Polybrene	Fisher Scientific	Cat #TR1003G
Blasticidin S HCL (10mg/mL)	Fisher Scientific	Cat #A1113903
Puromycin Dihydrochloride (10mg/mL)	Fisher Scientific	Cat #A1113803
DMSO	Fisher Scientific	Cat #BP231-100
EDTA	Fisher Scientific	Cat #BP2482
BstXI	New England Biolabs	Cat# R0113S
BlpI	New England Biolabs	Cat# R0585S
Critical Commercial Assays
QIAquick Gel Extraction Kit	Qiagen	Cat #2876X4
MinElute Reaction Cleanup Kit	Qiagen	Cat# 28206
QIAprep Spin Miniprep Kit	Qiagen	Cat# 27104
NextSeq™ 500/550 High output flow cell kit, 150 cycles	Illumina	Cat# 20024907
MiSeq Reagent Kit v2	Illumina	Cat# MS-102-2001
10x Genomics 3’ gene expression kit v3	10x Genomics
Deposited data
Perturb-seq (single and combinatorial) scRNA-seq and SHARE-seq	This study	GSE200201
ChIP-seq data	Michel et al., 2018	GSE113042
Immunoblot data	This study	Mendeley, 10.17632/yjypfbb5xf.1
Experimental models: Cell Lines
HEK 293T LentiX	Clontech	Fisher Cat #NC9834960
MOLM-13	DSMZ	ACC 554
Recombinant DNA
Perturb-seq vector, pBA439	Dixit et al. 2019	Addgene #85967
18 nt Perturb-seq barcoded vector, pBA571	Dixit et al. 2019	Addgene #85968
mSWI/SNF KO CRISPR Perturb-seq Library	This study	N/A
lentiCas9-Blast	Sanjana et al. 2014	Addgene #52962
psPAX2 plasmid	Didier Trono Lab	Addgene #12260
pMD2.G plasmid	Didier Trono Lab	Addgene #11259
Software and algorithms
Code for analyses in this work (perturbseq_baf version 1.0)	This study	https://doi.org/10.5281/zenodo.7682087
Cellranger version 3.02	Zheng et al., 2017 67	https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/installation
Scanpy version 1.7.2	Wolf et al., 2018 79	https://scanpy.readthedocs.io/en/stable/
MIMOSCA	Dixit et al., 2016 26	https://github.com/asncd/MIMOSCA
Scikit learn version 0.24.1	Pedregosa et al., 2011 68	https://scikit-learn.org/stable/index.html
Goatools 1.1.6	Klopfenstein et al., 2018 70	https://github.com/tanghaibao/goatools
SHARE-seq-alignment	Ma et al., 202042	https://github.com/masai1116/SHARE-seq-alignment
MACS2 version 2.2.7.1	Zhang et al., 2008 73	https://github.com/macs3-project/MACS
Seurat version 4.0.4	Hao et al., 2021 72	https://satijalab.org/seurat/
Signac version 1.3.0	Stuart et al., 202171	https://stuartlab.org/signac/
ChromVAR version 1.14.0	Schep et al., 2017 76	https://greenleaflab.github.io/chromVAR/index.html
Geosketch version 1.2	Hie et al, 2019 77	https://github.com/brianhie/geosketch
Other
GlutaMAX Supplement	Fisher Scientific	Cat# 35-050-061
Non-essential Amino Acids Solution (100X)	Fisher Scientific	Cat# 11-140-050
Sodium Pyruvate (100 mM)	Fisher Scientific	Cat# 11-360-070
HEPES (1M)	Fisher Scientific	Cat# 15630080
Penicillin-Streptomycin (10,000 U/mL)	Fisher Scientific	Cat# 15-140-122
RPMI 1640	Thermo Fisher Scientific	Cat# 61870036
DMEM high glucose, no glutamine	Thermo Fisher Scientific	Cat# 11960044
4%–12% Bis-Tris gels	Invitrogen	Cat# NP0321
Immobilon PVDF Transfer Membrane	Millipore	Cat# IPVH00010

The code used to generate all analysis and Figures in this paper can be found at https://doi.org/10.5281/zenodo.7682087.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cell lines

HEK293T and HEK293T LentiX cells used for virus generation were grown in DMEM medium (high glucose, no glutamine) supplemented with 10% FBS, 1% Penicilin-Streptomycin, 1% GlutaMax, 1% sodium pyruvate, 1% HEPES, and 1% non-essential amino acids (NEAA) (All TC reagents from Gibco). MOLM-13 cells were grown in RMPI 1640 medium (no glutamine) supplemented with 10% FBS (Omega), 1% Penicillin-Streptomycin (Gibco), and 1% GlutaMax (Gibco). Cells were maintained in an incubator at 37°C with 5% CO₂. HEK293T were obtained from ATCC, HEK293T LentiX cells were obtained from Clontech, and MOLM-13 AML cells from DSMZ.

METHOD DETAILS

Plasmid construction and cloning

The Perturb-seq vector (pBA439, Addgene #85967) with 18 nucleotide barcode (pBA571, Addgene #85968) was used to generate the BAF subunit-targeting and control guides for the Perturb-seq experiments. Barcoded vectors were digested with BstXI (NEB) and BlpI (NEB) to create an insertion site for the 20 nucleotide guide RNA sequences (Table S1). Guides were designed using the Broad Institute sgRNA Designer for CRISPR knockout⁶⁵. Sense and antisense guide oligonucleotides were annealed, phosphorylated (T4 PNK, NEB), diluted, and ligated into the digested perturb-seq vector (T4 DNA ligase, NEB). The lentiCas9-Blast plasmid (Addgene #52962) was used for Cas9 expression.

Lentivirus generation

Lentivirus was produced using second generation virus packaging constructs psPAX2 and pMD2.G. PEI was used to transfect HEK293T LentiX cells with packaging and expression constructs (Cas9 or barcoded Perturb-seq vector). Perturb-seq vector viruses were produced in an arrayed fashion to avoid guide and barcode swapping between vectors during packaging. Viral supernatant was harvested 48 hours after transfection and was concentrated by ultracentrifugation at 20,000 rpm for 2.5 hours at 4°C. Viral pellets were resuspended in PBS and frozen at −80°C.

Lentiviral infection

MOLM13 cells were spinfected with virus for 1.5 hours at 2000 RPM at 25°C using the concentrated lentivirus and 5 mg/mL polybrene. Cas9 spinfection was performed first with blasticidin selection beginning 2 days after spinfection and Western blot confirmation of expression at one week post-infection. Subsequent spinfection with the Perturb-seq vector was then performed. HEK293T cells were infected by adding concentrated virus dropwise to the media, adding 10 μg/μL polybrene, and allowing cells to incubate for 48 hours before removing the viral media and replacing with fresh DMEM.

Fluorescence-activated cell sorting (FACS)

MOLM13 or 293T cells were resuspended in FACS buffer containing 10mM EDTA and 2% fetal bovine serum in PBS. Cells kept were sorted for live populations and for BFP expression using either the violet laser of the BD FACSAria II or BD FACSAria II UV. Cells were sorted into media containing 20% fetal bovine serum and allowed to recover from sorting for 2 days before media were replaced with fresh media containing 10% fetal bovine serum.

Single cell RNA-seq

MOLM13 cells were prepared by diluting in RPMI with 10% FBS to a concentration of approximately 500 cells/μL, and 6,000 cells were loaded onto each channel of the 3’ Gene expression chip (10x Genomics, with version 3 chemistry). Fifteen channels were run for the single gene experiment, and 3 channels for the combination perturbation experiment. After generation of cDNA libraries from the harvested mRNA, libraries were amplified on a standard thermocycler, as specified per manufacturer instructions.

Perturb-seq libraries were sequenced on an Illumina Hi-Seq at a ratio of 3 channels to 1 lane of sequencing (5 for the single perturbation experiment, and 1 for the combination perturbation experiment), using paired end reads as follows. Read 1 is 28 bases long (16 for cell barcode, 12 for UMI), read 2 is 91 bases long and sample index read is 8 bases long.

Dial-out PCR and targeted sequencing for assigning perturbations to cells

An aliquot of the cDNA library from each channel was used for dial-out PCR using primers designed to amplify the segment between the sgRNA constant region and the BFP expression cassette in the Perturb-seq vector in order to obtain amplified GBC-CBC-UMI (guide barcode-cell barcode-UMI) combinations²⁶. The sequences of the primers used are given in Table S2.

Dial-out PCR products were sequenced on an Illumina MiSeq, with paired-end reads (for both the single guide and combination experiments, read 1 is 28 bases and read 2 is 60 bases long), with read 1 having the same structure as in the high-throughput sequencing of the libraries. Each experiment required only one MiSeq run for the dial-out libraries.

Perturb-SHARE-seq experimental methods

Perturb-SHARE-seq was performed for five conditions: NTC (guide 1), ARID1A (guide 3), SMARCD2 (guide 3), SMARCA4 (guide 4) and BRD9 (guide 2), each with two replicate wells in an arrayed format. Cells were spinfected with lentiviral guides, and were then selected with puromycin before harvesting at the 7 day time point.

Harvested cells were fixed and profiled by SHARE-Seq previously described to generate linked RNA and ATAC libraries⁴². Briefly, cells were fixed with 0.1% formaldehyde and quenched with glycine, Tris-HCl, and BSA. Cells underwent transposition with homemade Tn5¹⁰ loaded with Read1 and Read2 oligos. Reverse transcription was performed with RT primers with poly-T, UMI, ligation overhang, and biotin tag sequences. Cells were split amongst barcoding plates during each round of ligation and hybridization, similar to SpLIT-seq⁶⁶. Cells were reverse crosslinked, and ATAC fragments were purified and amplified during library preparation. RNA was harvested with streptavidin beads and amplified during library preparation. ATAC and RNA libraries were pooled 3:2 for sequencing. Libraries were sequenced using a 150 cycle high output kit (Illumina) on a NextSeq500 (Illumina)

QUANTIFICATION AND STATISTICAL ANALYSIS

Perturb-seq data pre-processing

Cellranger software (version 3.02)⁶⁷ was used to align reads to the GRCh38 human transcriptome version GRCh38–1.2.0, obtaining a matrix of counts for each gene in each cell. The feature barcoding option from cellranger was used to process dial-out PCR data together with expression data, to assign guides to cells.

Scanpy⁶⁵ was used for quality-control filtering, normalization and scaling for downstream analyses. For each dataset, cells with less than 200 detected genes were removed, and then genes detected in less than 3 cells were removed. Remaining cells were removed if their counts were below 1,000 or if their percent mitochondrial reads were above 20%. Counts were normalized per cell such that their sum per cell was 10,000 (TP10K). Normalized counts were transformed to log(normalized count +1), to yield “raw expression values”. To account for technical variation, batch (a {binary {0,1} variable for each batch}, total UMI counts per cell and the percent mitochondrial reads per cell were regressed.

Differential gene expression analysis

Differential expression was computed for each gene between cells with each specific perturbation vs. all control cells (all cells with either of the 5 control guides) on raw log(normalized count + 1) data, on the batch-corrected z-scores, using a two-sided t-test, with Benjamini-Hochberg FDR reported.

Linear model for regulatory matrix inference in the single gene perturbation data

A linear model was used as described previously²⁶. The model was trained on only variable genes defined as genes with variance exceeding 0.5, when scaled relative to other genes with similar expression level, and subsetting to genes with absolute log(1+TP10K) between 0.0125 and 4. A regression model was trained with coefficients $\text{β}$ for each individual perturbation with both L1 and L2 regularization (elastic net), as below.

$$\frac{1}{2N}{‖y-\left(p_1{\beta }_1+p_2{\beta }_2+\dots +p_n{\beta }_n\right)‖}^2+\alpha L_1\Vert \beta \Vert +0.5\alpha \left(1-L_1\right)\Vert \beta {\Vert }^2$$

where perturbation ${\text{p}}_i=1$ in cells that have perturbation $i$ and 0 otherwise, y is the raw expression (log(1+TP10K)) of each gene across the cells in the dataset, $\text{β}$ is the regulatory coefficient for each perturbation on each gene, $\text{α}$ is a multiplier for the penalty terms (relative to the error in the objective function) and L₁ is the L1 ratio specifying the relative weights of L1 and L2 penalties in the model.

The model was trained using the ElasticNet implementation in the python scikit learn package⁶⁸ using 80% of the cells, and model parameters were optimized based on a held-out validation set of 10% of the data. The remaining 10% of cells were used for assessing final model performance. The best performing model parameters were $\text{α}=0.0005$, L₁=0.5.

Functional similarity networks

Data from genome-wide CRISPR/Cas9-based dropout screens were analyzed to assess functional similarity between subunits of the mSWI/SNF across 789 cancer cell lines in the DepMap Achilles dataset from the Broad Institute (version 20Q4 downloaded from https://depmap.org/portal/download/all/)²³. The Pearson correlation coefficient was calculated between each pair of fitness profiles and similarly between each pair of mean Perturb-Seq profiles, for each of the CRISPR dropout and Perturb-Seq screens. The correlation vector of each perturbation was rank-order normalized, the gene-gene correlation matrix was row-wise rank-transformed, and plotted as a network with nodes (subunits) and edges (correlation), as described previously (Pan et al., 2018), using Fruchterman and Reingold force-directed layout algorithm (igraph R package)⁶⁹.

Comparison with published genome-scale K562 screen

We compared our Perturb-seq experiment with a published genome-scale CRISPRi screen in K562 cells⁶⁴. We downloaded data that were normalized and converted to pseudobulk for each perturbation condition, resulting in a matrix of perturbations x genes (link: https://plus.figshare.com/ndownloader/files/35773217). Then, we computed the Spearman correlation between each pair of perturbations, and clustered the resulting correlation matrix using Euclidean distance and single linkage and optimally ordered rows and columns.

Definition of gene programs

Gene programs were extracted by clustering genes in the regulatory matrix (perturbations x genes) obtained by linear regression as described above. The regulatory coefficients for each gene across perturbations were input to Louvain clustering, using a nearest-neighbor graph with k=5 neighbors. Gene programs were defined separately for the single gene perturbation and combinatorial experiments, resulting in 13 and 18 gene programs respectively. Note that gene programs computed in the combinatorial experiment were derived from a linear model trained with each perturbation condition as input, not considering any interaction terms.

Gene Ontology annotation of gene programs

To identify biological processes enriched in each gene program, genes in each program were subjected to Gene Ontology (GO) enrichment analysis, using the python package goatools⁷⁰ to perform a Fisher’s Exact Test computing the enrichment of each GO term in the genes present in each gene program, reported as Benjamini-Hochberg FDR. As a background, we used all genes.

SHARE-seq data pre-processing

Data was processed as previously described⁴⁰, using the hg19/GrCh37 human genome assembly (using the pipeline at https://github.com/masai1116/-alignment). Signac R package⁷¹, version 1.3.0 and Seurat 4.0.4⁷² were used for scATAC-Seq data analysis.

For scATAC-Seq data, peaks were called across all cells in the experiment using MACS2, version 2.2.7.1⁷³, parameters “effective genome size = human (2.7e9), extsize = 200, shift = -extsize/2” (default parameters when running MACS2 from Signac). Cells were removed if they had fewer than 1,000 fragments/cell, TSS enrichment < 1 or FRiP < 0.2. Counts of reads falling in each peak were extracted for each cell and used as input to latent semantic indexing (LSI), obtaining a low-dimensional representation of accessibility for each cell, scored across 50 LSI dimensions. As typically done⁷⁴, the first LSI was highly correlated with sequencing depth and ignored in downstream analyses.

For scRNA-seq data, counts were normalized to sum to 10,000 in each cell (TP10K) and cells were removed if they had less than 2,000 UMIs/cell, and then normalized reads were converted to log(1+normalized counts).

Association of perturbations to cell accessibility clusters

Cell accessibility profiles were clustered with Louvain clustering based on a k-nearest neighbor (k-NN, k=20, default) graph computed on LSI components 2 to 30. The fraction of cells with each perturbation in each cell cluster was computed for each replicate of each of the 5 conditions.

Integration of cell accessibility for each complex perturbation with complex binding

The impact of subcomplex perturbation was integrated with binding sites for each subcomplex obtained from three bulk ChIP-seq experiments in the same cell line without perturbations for DPF2 (cBAF), BRD9 (ncBAF) and BRD7 (PBAF)³. Binding peaks were called in each ChIP-Seq sample using MACS2⁷³ and each SHARE-seq peak was annotated by its overlap with each ChIP-seq peak. A finer grained annotation was generated by further stratifying peaks as TSS-proximal (1kb upstream of TSS, as defined from GENCODE v19⁷⁵) or distal (otherwise). For each cell and peak type, an accessibility score was computed with ChromVAR version 1.14.0⁷⁶, as the number of reads falling in the set of peaks of interest, scaled to a z-score by the reads falling in a background set of peaks matched for GC content and average accessibility. Scores were compared in two ways across perturbations for each set of peak annotations (cBAF, ncBAF, PBAF peaks, cBAF-only, ncBAF- only, PBAF-only peaks etc.). First, the distributions of ChromVAR scores at cBAF-, ncBAF- and PBAF-specific sites was compared between perturbed cells and control cells using a Mann-Whitney test, with Bonferroni adjustment for multiple testing. Second, reads with ends falling within 100 bp of summits of the cBAF-, ncBAF- and PBAF-specific peaks were aggregated (normalized to the total number of fragments), as a function of the distance between the midpoint of the read and the peak summit, and these read distributions were compared between cells with perturbations vs. control cells.

Association of changes in expression and chromatin accessibility

Thirteen gene programs were retrieved from the analysis of the single perturbation experiment. Each program was scored in the SHARE-seq scRNA-seq profiles using the scanpy function scanpy.tl.score_genes(), which averages in each cell the expression of genes in the program relative to a group of 50 control genes, matched in expression level to the genes in the program. Spearman correlation coefficient across single cells was calculated between gene program scores and various ChromVAR scores profiles.

Combinatorial perturbation linear model and synergistic gene sets

A linear model with interaction terms was used as described previously²². Specifically, for a given combination of perturbations, a regression model with interactions was trained with coefficients, $\text{β}$, for both the individual perturbations and pairs of perturbations and both L1 and L2 regularization (elastic net). For example, for two perturbations, the model is fit to minimize the following objective function:

$$\frac{1}{2N}{‖y-\left(p_1{\beta }_1+p_2{\beta }_2+p_{12}{\beta }_{12}\right)‖}^2+\alpha L_1\Vert \beta \Vert +0.5\alpha \left(1-L_1\right)\Vert \beta {\Vert }^2$$

where perturbation ${\text{p}}_i=1$ in cells that have perturbation $i$ and 0 otherwise, y is the raw expression (log(1+TP10K)) of each gene across the cells in the dataset, $\text{β}$ is the regulatory coefficient for each perturbation on each gene, $\text{α}$ is a multiplier for the penalty terms (relative to the error in the objective function) and L1 is the L1 ratio specifying the relative weights of L1 and L2 penalties in the model. (In all cases with three perturbations, one pair was modeled jointly as p1 and the third perturbation was p2.)

The model was trained using the ElasticNet implementation in the python scikit learn package using 80% of the cells, and model parameters were optimized based on a held-out validation set of 10% of the data. The remaining 10% of cells were used for assessing final model performance. The best performing model parameters were $\text{α}=0.005$, L₁=0.5.

The significance of linear model coefficients was estimated as previously described²². Briefly, the assignment of guides to cells was permuted (while preserving the co-occurrence of guides per cell), and shuffled data were used to create an empirical distribution of coefficients. Coefficients were grouped into bins by gene expression and variance and the number of cells per guide. For each bin, an empirical p-value was computed (defined as the fraction of empirical coefficients equal to or less extreme than the observed one), and the Benjamini-Hochberg FDR was calculated. Significant coefficients are reported at an FDR of 1%.

Genes are annotated as additive if none of their interaction terms was significant, and non-additive if at least one interaction term is significant. For pairs of perturbations, we further categorize genes into specific interaction categories, as described²².

BAF activity model

To estimate how each perturbation affects each subcomplex type, complex-specific impacted gene sets were defined from the single perturbation experiment, and the effect of each perturbation combination on these in the combo experiment was estimated. Complex-specific gene sets were defined as the union of genes differentially expressed between cells with guides targeting complex-defining subunits (ARID1A, ARID1B, DPF2 for cBAF, BRD9 for ncBAF and ARID2, PBRM1, BRD7 and PHF10 for PBAF) vs. cells with control guides. This was done separately for up- and down-regulated genes for each complex type. Genes that were shared between multiple complexes-defined sets were removed. The impact of perturbations in on each set of complex-specific impacted gene sets was defined by scoring the mean expression of each set of subcomplex up- and down-regulated genes defined above in each perturbed cell in each perturbation category, and comparing to the distribution of scores in control cells using a Kolmogorov-Smirnov test, and reporting Benjamini-Hochberg FDR.

Classifying TCGA tumors by Perturb-seq signatures

To define gene signatures, the Perturb-Seq data were first transformed by z-normalizing the expression of each gene with respect to the mean and standard deviation of the gene in the control cells. Cells with each perturbation condition were partitioned into equally-sized training and test sets, with the training set sampled using Geosketch⁷⁷. NMF was applied to the training set using 5 to 20 factors, with the number of factors chosen based on the minimal reconstruction error on the test set of cells, as calculated by the mean-squared error.

For each tumor’s bulk gene expression profile, raw counts were first normalized per sample such that the counts in each sample sum to 10,000, after which a pseudocount was added and a log transformation was applied. A similar z-normalization step was then performed with respect to the gene expression of corresponding normal tissue. The mean expression of the 10 genes with the highest factor loadings for each of the NMF factors was calculated to define a gene signature score for each NMF factor in each perturbation condition or tumor. The result is a gene signature profile for each perturbation condition and each tumor. A softmax transformation was also applied to this gene signature profile. To determine the similarity of a given tumor to a perturbation signature, random forest classifiers were first trained on the Perturb-seq dataset’s gene signature profiles to predict the perturbation condition of each cell. The training set was composed of 250 randomly selected cells from each of the perturbation conditions, and a classifier was trained for each perturbation condition to predict whether a cell is from that perturbation condition. Each random forest classifier was trained using 500 decision trees with otherwise default parameters from the scikit-learn package. The predicted class probability for each tumor with respect to a perturbation condition was used to score its similarity to the perturbation condition. The cBAF loss signature score for a gene expression profile was calculated as the average of the prediction scores for the cBAF perturbation conditions.

To identify genes that are mutated at higher frequencies in cBAF loss-like tumors, cBAF loss-like tumors were grouped by tissue type and the of tumors that featured a particular gene mutation was compared to a background set of tumors from the same tissue type. Enrichment of gene mutations was calculated by Fisher’s exact test (FDR < 10%).

To identify regulators of genes that are differentially expressed in cBAF loss-like tumors, genes that are differentially expressed in the cBAF loss-like tumors relative to other tumors from the same type were identified (Welch’s t-test, q < 0.05, Benjamini-Hochberg FDR). Genes that were differentially expressed in at least half of the tissue types represented in the set of cBAF loss-like tumors were retained, yielding 275 downregulated genes and no up-regulated genes. Lisa⁵³ was applied to the set of 275 downregulated genes, using default parameters, to predict transcriptional regulators implicated in the regulation of these genes.

Non-negative matrix factorization (NMF)

To identify major axes of variability in cell states implicated in the perturbed BAF subunits in an unsupervised manner, we used NMF for both the single and combinatorial perturbation datasets (separately). For each dataset, gene expression values were transformed such that each gene has unit variance, and Geosketch⁶⁷ was used to sample 50 cells from each perturbation condition to obtain an equal and representative sample of cells across and within each perturbation condition. Samples were combined and randomly split into a training and test set, consisting of 80% and 20% of the total cells, respectively. The scikit-learn implementation of NMF with $l_1$-regularization was applied to the training set. The number of factors $k$ and $l_2$-regularization coefficient were selected using a grid search, with the reconstruction error on the test set as the model selection criterion.

Cell cycle-related NMF factors were identified as individual factors that are predictive (AUROC > 0.8) of cell cycle stages assigned to cells using marker genes⁷⁸. Differentially activated NMF factors for each perturbation condition were identified by performing a $t$-test on the NMF factor scores in perturbed vs. NTC cells ($q<0.01$, Benjamini-Hocbherg FDR). The program scores of differentially activated gene programs (aside from those associated with cell cycle) were used as cell profiles for UMAP embedding and for clustering with Louvain community detection using the default parameters for each method in Scanpy⁷⁹.

Outlier score calculation

The NMF factor scores of the set of differentially activated NMF factors were calculated for each cell. A one-class support vector machine (SVM) as implemented in scikit-learn was trained on these NMF factor scores for the NTC cells using default parameters. The outlier score for each perturbed cell was then computed as its negative signed distance to the learned separating hyperplane, with more positive scores representing greater deviations from the distribution of NTC cells.

Supplementary Material

2. Table S1. CRISPR screen guide information and experimental design, Related to Figure 1, S1.

Information about each guide RNA used in this work, specifically the guide name (“Guide Name”), the gene that the guide is targeting (“Targeted subunit”), the number assigned to each of multiple guides targeting the same gene (“Guide number”), the guide sense sequence (“Guide Sense Sequence”), the barcode associated with each guide (“Guide barcode”), and whether the guide was used in the single perturbation experiment (“In Single Perturbation Experiment”), the combination experiment (“In Combination Experiment”) and in SHARE-seq (“In SHARE-seq experiment”).

3. Table S2. Primer sequences used for the dial-out PCR, Related to Figure 1,S1.

Information about dial-out PCR primer sequences, namely the TruSeq Universal and barcoded primers which consists of a P7 homology for sequencer compatibility as well as BFP homology complementary to the guide plasmid. 8-nucleotide index sequences are given and their location in the barcoded primer is represented by NNNNNNNN.

4. Table S3. cBAF loss signature scores for TCGA tumors, Related to Figure 7, S7.

Metadata for TCGA tumors with cBAF mutations and without BAF mutations, including the TCGA file ID, TCGA barcode, TCGA tumor class, and cBAF loss signature score for each tumor.

5. Table S4. Genes with enriched mutation frequencies in cBAF loss-like tumors, Related to Figure 7, S7.

List of genes with significantly higher mutation frequencies in cBAF loss-like tumors compared to the background gene mutation frequencies of the corresponding tumor class (FDR < 10%, Fisher’s exact test), including the significance of enrichment, gene function, and distribution of mutation types for each gene.

6. Table S5. Differentially down-regulated genes and their predicted transcriptional regulators in cBAF loss-like tumors, Related to Figure 7, S7.

Lists of differentially down-regulated genes shared across different tumors classes in tumors with strong cBAF loss or ncBAF loss signatures, including gene function and the significance of differential expression across different tumor classes. Lisa⁵⁴ predictions of transcriptional regulators for these sets of genes.

Highlights.

mSWI/SNF complex-, module-, and subunit-specific impacts defined by Perturb-seq

Perturb-, SHARE-seq define paralog subunit relationships, shifted complex functions

Intra-complex mSWI/SNF genetic interactions are largely synergistic

Single-cell perturbation signatures mirror and predict cBAF loss-of-function in cancer