Profiling the genetic determinants of chromatin accessibility with scalable single-cell CRISPR screens

Abstract

Pooled CRISPR screens have been used to connect genetic perturbations with changes in gene expression and phenotypes. Here, we describe a CRISPR-based single-cell combinatorial indexing assay for transposase-accessible chromatin (CRISPR-sciATAC) to link genetic perturbations to genome-wide chromatin accessibility in a large number of cells. In human myelogenous leukemia cells, we apply CRISPR-sciATAC to target 105 chromatin-related genes, generating chromatin accessibility data for ~30,000 single cells. We correlate loss of specific chromatin remodelers with changes in accessibility globally and at the binding sites of individual transcription factors. For example, we show that loss of the H3K27 methyltransferase EZH2 increases accessibility at heterochromatic regions involved in embryonic development and triggers expression of genes in the HOXA and HOXD clusters. At a subset of regulatory sites, we also analyze changes in nucleosome spacing upon loss of chromatin remodelers. CRISPR-sciATAC is a high-throughput single-cell method for studying the role of genetic perturbations on chromatin in normal and disease states.

Editorial summary:

The effects of gene knockouts on chromatin accessibility are measured with single-cell CRISPR screens.

Chromatin accessibility orchestrates cis- and trans-regulatory interactions to control gene expression and is dynamically regulated in cell differentiation and homeostasis. Alterations in chromatin state have been associated with many diseases including several cancers¹. In recent years, CRISPR screens have been combined with single-cell RNA-sequencing to measure the effects of genetic perturbations on gene expression across the transcriptome^2–5. However, methods to capture changes in the epigenome following CRISPR perturbations have limited throughput⁶. To study how genetic perturbations affect chromatin states, we developed a platform for scalable pooled CRISPR screens with single-cell ATAC-seq profiles called CRISPR-sciATAC. In CRISPR-sciATAC, we simultaneously capture transcripts encoding a Cas9 guide RNA (gRNA) and perform single-cell combinatorial indexing ATAC-seq⁷ (Fig. 1a, Supplementary Fig. 1). Following cell fixation and lysis, nuclei are recovered and the open chromatin regions of the genomic DNA undergo barcoded tagmentation in a 96-well plate using a unique, easy-to purify transposase from Vibrio parahemolyticus (Fig. 1b, Supplementary Fig. 2). Next, gRNA sequences are barcoded with the same barcode sequence as the ATAC fragments, using in situ reverse transcription. The nuclei are pooled together and split again to a new 96-well plate and both the ATAC fragments and gRNA sequences are tagged with a second barcode in two consecutive PCR steps. At the end of this process, a unique combination of barcodes (“cell barcode”) tag both the gRNA and the ATAC fragments from each cell (Fig. 1a, Supplementary Fig. 1, Supplementary Table 1).

Fig. 1 |. CRISPR screens with single-cell combinatorial indexing assay of transposable and accessible chromatin sequencing (CRISPR-sciATAC) enables the joint capture of chromatin accessibility profiles and CRISPR perturbations.

(a) CRISPR-sciATAC workflow with initial barcoding, nuclei pooling and re-splitting, and then second round barcoding. (b) Comparison of bulk ATAC-seq chromatin accessibility profiles from K562 cells using Tn5 and TnY transposases and aggregated CRISPR-sciATAC single cell profiles from 11,104 cells. (c) Guide RNA (gRNA) reads mapping to human or mouse CRISPR libraries (n = 1986 cells). (d) ATAC reads mapping to human or mouse genomes (n = 721 cells). For display purposes, we removed one cell that had >10-fold the average number of ATAC reads. (e) Concordance between the percent of ATAC and gRNA reads mapping to the human and mouse genomes and human and mouse gRNA libraries, respectively, for each cell (n = 496 cells). (f) ATAC-seq fragment size distribution from K562 cells of bulk ATAC-seq data, aggregated CRISPR-sciATAC single cell profiles from 11,104 cells and one representative single cell from CRISPR-sciATAC. (g) Number of CRISPR gRNAs detected per cell. (h) Proportion of cells with 1, 2, or more than 2 gRNAs.

Results

CRISPR-sciATAC captures both chromatin accessibility profiles and CRISPR perturbations in single cells

To quantify capture and barcoding of single cells, we performed CRISPR-sciATAC on a mix of human (HEK293) and mouse (NIH3T3) cells. Human and mouse cells were each transduced with a small library of 10 distinct non-targeting gRNAs with no overlapping gRNAs between the two pools. The gRNA sequences were cloned into a lentiviral vector that includes the Pol3-driven gRNA within a longer Pol2 transcript for perturbation readout (CROP-seq vector)². We found that 93% of cell barcodes had gRNA-containing reads that could uniquely be assigned to either human or mouse gRNAs (Fig. 1c) and 96% of cell barcodes had ATAC-seq reads mapping to either the human or mouse genome, indicating that the majority of cell barcodes were correctly assigned to single cells (Fig. 1d). As an additional verification of single-cell separation, we also measured the species concordance between the ATAC-seq and gRNA reads. We found that for 92% of the captured cell barcodes both ATAC-seq and gRNA reads aligned either to human or mouse reference genomic and gRNA sequences, respectively. In 4.4% of cells, the ATAC-seq and/or gRNA reads could not be exclusively assigned to one species. ATAC-seq and gRNA reads were assigned to different species (species collision) in 3.6% of cells (Fig. 1e). The low rates of these two failure modes suggest that CRISPR-sciATAC can simultaneously identify accessible chromatin and CRISPR gRNAs in single cells.

To test the ability of CRISPR-sciATAC to capture biologically meaningful changes in chromatin accessibility, we targeted 21 chromatin modifiers that are highly mutated in cancer (Supplementary Fig. 3a, b). Using the Catalog of Somatic Mutations in Cancer (COSMIC) database⁸, we selected 21 chromatin-related genes that carry the highest mutational load across all cancers, including 9 chromatin remodelers (ARID1A, ATRX, CHD4, CHD5, CHD8, MBD1, PBRM1, SMARCA4, and SMARCB1), 2 DNA methyltransferases (DNMT3A and TET2), 3 histone methyltransferases (EZH2, PRDM9, and SETD2), 1 histone demethylase (KDM6A), 1 histone deacetylase (HDAC9), 3 histone subunits (H3F3A, H3F3B, and HIST1H3B), and 2 readers (ING1 and PHF6). We designed 3 gRNAs per gene and also included 3 non-targeting gRNAs in our library (Supplementary Table 2). We transduced Cas9-expressing human myelogenous leukemia K562 cells with this lentiviral gRNA library at a low multiplicity of infection and selected with puromycin for transduced cells. After 1 week of selection, we collected single-cell paired-end ATAC-seq data. After filtering for cells with ≥500 unique ATAC-seq fragments and ≥100 gRNA reads (Supplementary Fig. 3c–h), we obtained 11,104 cells with a median of 1,977 unique ATAC-seq fragments mapping to the human genome. Aggregated ATAC-seq profiles for these cells correlate well with bulk data from K562 cells (Fig. 1b, Supplementary Fig. 3i). Single cells retained an ATAC fragment length distribution similar to cells tagmented in bulk (Fig. 1f). The majority of cell barcodes (83%) had one gRNA (Fig. 1g, h), and for 90% of cell barcodes, a single gRNA represented ≥99% of the reads (Supplementary Fig. 3j).

We recovered all 66 gRNAs with a median of 148 single cells per gRNA and 468 single cells per gene. Upon closer examination, we noticed that not all gene targets resulted in the same number of single cells captured, suggesting that some of our targets might be essential genes whose targeting leads to drop-out of those cells. To distinguish gRNA depletion of essential genes from inability to capture gRNAs using CRISPR-sciATAC, we separately amplified gRNAs from the bulk population at an early time point and at 1 and 2 weeks post-selection (Supplementary Fig. 4a). We found high correlation between all samples across 3 independent transduction replicates (Supplementary Fig. 4b, c). For several genes, multiple, distinct gRNAs targeting the same gene were consistently depleted or enriched: H3F3A, CHD4, SMARCA4, and SMARCB1 were depleted, while targeting KDM6A accelerated cell growth (Supplementary Fig. 4d). Using robust rank aggregation to measure consistent enrichment across multiple gRNAs⁹, we computed gene-level enrichment scores (Supplementary Fig. 4e, Supplementary Table 3), which were highly correlated with a previous genome-wide CRISPR screen in K562 cells¹⁰ (r = 0.85) (Supplementary Fig. 4f). Reassuringly, enrichment of individual gRNAs was positively correlated with cell numbers estimated from CRISPR-sciATAC cell barcodes (r = 0.73, Supplementary Fig. 4g). Different gRNAs targeting the same gene tend to result in similar numbers of single cells, highlighting consistent proliferation phenotypes between different genetic perturbations targeting the same gene (Supplementary Fig. 4h, i). We did not observe changes in the number of ATAC fragments per cell between the different perturbed genes (Supplementary Fig. 5a, b) and gene enrichment was not correlated with the number of ATAC fragments, peaks, or differential peaks obtained from gRNAs targeting the same gene (Supplementary Fig. 5c–e).

Loss of chromatin modifiers alters chromatin accessibility at regulatory elements in single cells

We examined how loss of these chromatin modifiers impacts accessibility within known chromatin marks (primarily histone post-translation modifications) using ENCODE ChIP-seq data from K562 cells (Fig. 2a, Supplementary Tables 4, 5). We found similar accessibility changes between different gRNAs targeting the same genes, further highlighting the consistency between distinct genetic perturbations targeting the same gene (Fig. 2b). Targeting the Polycomb repressive complex 2 (PRC2) subunit EZH2 resulted in an increase in chromatin accessibility at regions with H3K27me3, a marker of heterochromatin (Fig. 2a). EZH2 catalyzes nucleosome compaction via H3K27 trimethylation¹¹ and thus loss of EZH2 increases accessibility in these regions. Differential accessibility at known chromatin marks can be highly specific: A downsampling analysis reveals that for some target genes, like EZH2 and ARID1A, a small number of cells correlates well (r_p ≥ 0.75) to an aggregated (pseudo-bulk) cell population (5 single cells for EZH2, 25 single cells for ARID1A) (Fig. 2c, Supplementary Fig. 6a, b). For cells receiving a non-targeting gRNA, we find that 75 cells correlate well with the respective pseudo-bulk populations (Supplementary Fig. 6c). Over all perturbations, we find that the median cell number to represent the pseudo-bulk is also 75 cells. For some CRISPR perturbations more cells are needed to accurately represent the pseudo-bulk (e.g. 225 cells for TET2) (Supplementary Fig. 6d), indicating that disruption of these genes creates more variable chromatin accessibility than non-targeting controls.

Fig. 2 |. CRISPR-sciATAC reveals changes in accessibility at HOX genes following loss of EZH2.

(a) Heatmap of chromatin accessibility Z-scores at histone and DNA modifications for different CRISPR perturbations (n = 3 gRNA per gene). We converted the fraction of accessible regions for each modification into Z-scores (using all cells in the screen). For visualization, we show the average Z-score for all cells receiving a particular gRNA. (b) Distances in the histone and DNA modifications accessibility profiles shown in panel a between gRNAs targeting different genes and gRNAs targeting the same gene. The distance metric is 1-(Pearson correlation of the Z-scores). (c) Pearson correlation between averaged accessibility Z-scores at histone and DNA modifications of the indicated number of single cells and the average profile of 400 single cells, for cells with either EZH2-targeting or non-targeting (NT) gRNAs. (d) UMAP representation of chromatin accessibility Z-scores at histone and DNA modifications from single cells receiving either EZH2 or NT gRNAs. Also shown is the same UMAP representation with single cells colored by TFBS accessibility enrichment scores for CBX2, CBX8, EZH2, POLR2B, and SIRT6. (e) H3K27me3 ChIP-seq coverage at the HOXA-D loci (top). Changes in accessibility at the HOXA-D loci in cells transduced with EZH2-targeting or NT gRNAs (bottom). *** denotes p < 0.001. (f) CRISPR-sciATAC fragments mapping to the HOXA locus in cells transduced with EZH2-targeting or NT gRNAs (n = 510 cells per condition). The sum of all ATAC fragments over the entire HOXA locus in cells transduced with EZH2-targeting and NT gRNAs is shown on the right. K562 H3K27me3 ChIP-seq coverage is shown at the bottom. (g) Gene expression (qPCR) of EZH2, HOXA3, HOXA5, HOXA11A, HOXA13 and HOXD9 for cells transduced with either EZH2-targeting or NT gRNAs. HOX gene expression in cells targeted by the two more effective EZH2-targeting gRNAs (g1 and g3, as defined by decrease in EZH2 expression compared to non-targeting gRNAs) is greater than in cells targeted by the less effective gRNA (g2) (Student’s t-test, p < 0.05 for HOXA3, HOXA5, HOXA11 and HOXA13, p = 0.09 for HOXD9).

Hierarchical clustering of single cells transduced with EZH2-targeting gRNAs or non-targeting gRNAs reveals a clear separation (Supplementary Fig. 7a, b), that can also be observed in a uniform manifold projection (UMAP) (Fig. 2d). We verified this separation is not due to differences in library complexity in cells with EZH2-targeting gRNAs (Supplementary Fig. 7c), and found that increased accessibility in Polycomb repressive complex 1 (PRC1) components CBX2 and CBX8 binding sites has the highest predictive power in differentiating EZH2-targeted cells. Similarly, differential accessibility of POLR2B and SIRT6 binding sites can also be used to differentiate between cells with EZH2-targeting and non-targeting gRNAs, however in the opposite direction, where loss of EZH2 leads to decreased accessibility in their binding sites. As expected, we find an increase in accessibility at EZH2 binding sites, which is expected given EZH2’s role in repression through heterochromatin formation¹².

Using Gene Ontology (GO) analysis of differentially accessible regions in EZH2-targeted cells, we found an enrichment in genes involved in embryonic development and cell differentiation (Supplementary Fig. 8, Supplementary Table 6). Indeed, EZH2 is known to play important roles in embryonic development and cell- and tissue-specific differentiation¹¹ and we found large changes in chromatin accessibility at several of the homeobox (HOX) genes (Fig. 2e). In K562 cells, the HOXA and HOXD gene clusters contain the highest amount of H3K27me3 repressive heterochromatin. In the HOXA gene cluster, we found that there was a nearly 3-fold increase in accessibility (Fig. 2f). A similar increase in accessibility was also seen at the HOXD gene cluster (Supplementary Fig. 8d). To understand the functional consequences of these changes, we measured the expression of EZH2 and several HOX genes (HOXA3, HOXA5, HOXA11, HOXA13, and HOXD9) (Fig. 2g). After EZH2 loss, we found that these previously-silenced genes become highly expressed. Among the 3 gRNAs targeting EZH2, we noticed that the least effective gRNA resulted in less HOX gene expression, further reinforcing the role of EZH2 in maintaining HOX gene repression (Fig. 2g, Supplementary Fig. 8e). Taken together, these results suggest that loss-of-function mutations in EZH2 lead to aberrant expression of genes from HOXA and HOXD clusters.

Beyond EZH2, we found that changes in accessibility in single cells at transcription factor binding site (TFBS) are consistent between gRNAs targeting the same gene (Supplementary Fig. 9a, b). We also noticed that large changes in TF accessibility correlate with decreased cell proliferation (Supplementary Fig. 9c), suggesting that perturbing chromatin modifiers which broadly disrupt TFBS can impact cell viability. We found similar changes in TFBS accessibility using either ENCODE ChIP-seq data from K562 or predicted TF binding sites (JASPAR TF motifs¹³ and chromVAR¹⁴) (Supplementary Fig. 9d, e). To determine if chromatin accessibility is modified at single nucleotide polymorphisms (SNPs) that regulate gene expression, we measured overlap with cis-regulatory expression quantitative trait loci (cis-eQTLs). For two of our targets — KDM6A and ARID1A — we found a reduction in accessibility at tissue-matched (blood) cis-eQTLs in cells after perturbation of these genes (Supplementary Fig. 10a). KDM6A-targeted cells had the largest reduction of cis-eQTL accessibility with eQTL genes (eGenes) involved in DNA condensation and chemokine receptor activity (Supplementary Fig. 10b–e).

A loss-of-function screen of chromatin remodeling complexes using CRISPR-sciATAC

To further demonstrate the scalability of CRISPR-sciATAC, we designed a CRISPR library to target all human chromatin remodeling complexes in the EpiFactors database¹⁵ (Fig. 3a). In total, we targeted 17 chromatin remodeling complexes that each include between 2 and 14 subunits. As before, we targeted the coding exons of each subunit with 3 gRNAs and also included gRNAs designed not to target anywhere in the human genome. Over the 17 chromatin remodeling complexes, we captured paired CRISPR perturbation and single-cell ATAC-seq data from 16,676 cells. As in the previous screen, the number of cells recovered for each CRISPR perturbation correlated with gene essentiality scores¹⁰ (Supplementary Fig. 11a). We recovered particularly low numbers of cells for the two subunits of the FACT complex, which are known to be highly essential¹⁶ (Supplementary Fig. 11a, b).

Fig. 3 |. A CRISPR-sciATAC screen targeting 17 chromatin remodeling complexes uncovers widespread disruptions in accessibility upon SWI-SNF disruption.

(a) Chromatin remodeling complex subunits and cofactors targeted in the CRISPR library. (b) Heatmap of chromatin accessibility Z-scores at transcription factor binding sites (TFBSs) for the different chromatin remodeling complexes targeted in the screen. We converted the fraction of accessible regions for each TFBS into Z-scores (using all cells in the screen). For visualization, we first average over all cells for a particular target gene and then average over all genes in the complex. The histograms (left) show the distribution of Z-scores for each complex. The FACT complex is not shown due to a low number of single cells (n = 75 cells). (c) UMAP representation of the genes perturbed in the screen based on the TFBS differential accessibility Z-score profiles. Subunits of the SWI-SNF pBAF complex are labeled with filled circles and gene names. (d) The number of transcription factor binding sites with significant differential accessibility for cells that receive a specific gene-targeting CRISPR perturbation, as compared to cells that receive a non-targeting (NT) control gRNA (FDR q ≤ 0.1). SWI/SNF components and co-factors are highlighted in red. (e) The percent of ATAC fragments in enhancers and promoters in cells transduced with ARID1A-targeting and NT gRNAs. Each point is a single cell. K562 enhancer and promoter genome segmentation is from ENCODE (see Methods). (f) CRISPR-targeted chromatin complex genes with significant differential accessibility at enhancers and/or promoters. (g) Volcano plots showing significant changes in accessibility at TFBSs in cells transduced with ARID1A (left), SMARCA5 (middle) and RCOR1 (right) -targeting gRNAs. Standardized Z-scores are averaged over single cells. Points in red represent TFBSs with a significant change in accessibility (FDR q ≤ 0.1 and |Z-score| > 0.25).

Given the larger scale of this CRISPR-sciATAC screen, we initially analyzed changes in accessibility at the level of different chromatin remodeling complexes instead of individual proteins/subunits (Fig. 3b). Examination of differential accessibility in TFBSs revealed two major groups: Complexes where loss of subunits generally results in increased accessibility, such as the CoRepressor for Element-1-Silencing Transcription factor (CoREST) and the Nucleosome Remodeling Factor (NuRF) complexes, and another group where loss of subunits leads to decreased accessibility, such as the Calcium RESponsive Transactivator-BRG1 (CREST-BRG1) and SWI/SNF-B (pBAF) complexes. However, loss of individual subunits within these complexes displays tremendous heterogeneity: Nearly all complexes have subunits where loss triggers increased accessibility and other subunits with the opposite effect (Supplementary Fig. 12, Supplementary Table 7). A two-dimensional UMAP projection of the TFBS accessibility profiles reveals a cluster enriched in SWI/SNF components and, in particular, pBAF components (hypergeometric p = 4 × 10⁻⁴) (Fig. 3c). Loss of SWI/SNF subunits tends to alter accessibility at many TFBS, with the greatest number of disrupted TFBSs from ARID1A loss (Fig. 3d). Previously, ARID1A loss has been shown to impair enhancer-mediated gene regulation¹⁷, and indeed we find that loss of ARID1A dramatically reduced accessibility at enhancers, but not at promoters (Fig. 3e).

Combining data from both CRISPR-sciATAC experiments, we found that the chromatin modifiers targeted in our two screens resulted in a greater number of accessibility changes at enhancers than at promoters (Fig. 3f), supporting a gene regulatory model with more dynamic chromatin accessibility at distal regulatory elements compared to promoters¹⁸. Loss of SWI/SNF-ATPase subunit ARID1A or ISWI-ATPase subunit SMARCA5 results in many changes in TFBS accessibility (Fig. 3g). For ARID1A, some of these changes include a reduction in accessibility at JUN and FOS binding sites, which are subunits of the AP-1 transcription factor that cooperate with the SWI/SNF complex to regulate enhancer activity¹⁹. Loss of SMARCA5, which helps load cohesin onto chromosomes²⁰, triggered a reduction in accessibility in binding sites of cohesin subunits RAD21 and SMC3 along with cohesin cofactor ZNF143²¹. In contrast to these gene perturbations that affect a wide range of TFBSs, other perturbations result in accessibility changes at only one or a few TFBSs. For example, we observe an increase in accessibility at only PU.1 binding sites upon loss of RCOR1 (Fig. 3g). RCOR1 has previously been shown to promote erythroid differentiation via repression of myeloid genes such as PU.1 and thus may have a focused role in lineage specification²².

Quantifying nucleosome movement at transcription factor binding sites following chromatin remodeler loss

In addition to changes in accessibility, chromatin remodeling complexes can regulate gene expression by changing specific nucleosome positions around regulatory sequences²³. We developed a computational framework to measure changes in nucleosome position in CRISPR-sciATAC, focusing on 7 TFs with a bimodal coverage profile around their binding sites, suggesting a symmetric positioning of nucleosomes around them²⁴ (Fig. 4a, b). Using this pipeline, we found that loss of chromatin remodelers generally results in expansion of nucleosomes around TFBSs (Fig. 4c), with the exception of BAF/pBAF (SWI/SNF) subunits ARID1A and PBRM1 where knock-out leads to compaction of nucleosomes around the TFBSs studied (Fig. 4b). At specific TFBS, loss of different chromatin remodelers can have opposing effects: For example, ARID1A loss results in a 20 nt nucleosome compaction at AP-1 binding sites (p = 0.03), which has also been demonstrated in a recent study suggesting that the BAF complex controls occupancy of AP-1²⁵. In contrast, loss of EP400, which is part of the Sick With Rat8ts (SWR) complex, causes a large, 56 nt expansion of nucleosomes around AP-1 binding sites (p = 10⁻⁴) (Fig. 4d).

Fig. 4 |. Nucleosome dynamics around transcription factor binding sites (TFBSs) following CRISPR targeting of chromatin remodelers.

(a) Schematic depicting the computational approach to identify changes in nucleosome positions around TFBSs. (b) The difference in nucleosomal distances in gene-targeted cells and nucleosomal distances in non-targeting cells (“peak shift”) across 7 TFBS following CRISPR targeting of chromatin remodelers (top). Bubble-plot of the peak shifts for individual TFBS (bottom). The color of the bubble corresponds to the peak shift (nt) and the size of the bubble represents the empirical p-value calculated by a label permutation test. (c) The number of nucleosome expansion and compaction events around TFBSs following CRISPR targeting of chromatin remodelers. (d) Coverage profiles of mononucleosomal fragments around AP1 binding sites in cells transduced with ARID1A-targeting (blue) and non-targeting (NT) (grey) gRNAs (top) and in cells transduced with EP400-targeting (blue) and NT (grey) gRNAs (bottom). Dashed lines represent the most highly covered base in each peak. (e) Peak shifts in TFBSs located in enhancers and promoters. Each point is a CRISPR-perturbed gene (average of all gRNAs for that gene). (f) Peak shifts in TFBSs located in enhancers and promoters in SFMBT1-targeted cells (left). Coverage profiles of mononucleosomal fragments in cells transduced with SFMBT1-targeting (blue) and NT (grey) gRNAs around AP1 binding sites in promoters (top) and in enhancers (bottom). (g) Peak shifts in TFBSs located in enhancers and promoter in SMARCB1 targeted cells (left). Coverage profiles of mononucleosomal fragments in cells transduced with SMARCB1-targeting (blue) and NT (grey) gRNAs around RAD21 binding sites in promoters (top) and in enhancers (bottom). For panels d, f, and g, the shaded regions represent s.e.m. (n = 3 gRNAs).

We further asked if there are specific differences in nucleosome dynamics surrounding TFBSs residing in enhancers versus promoters. We found that changes in nucleosome peak positions occur typically in either enhancers or promoters, depending on the specific TFBS. For example, across all CRISPR perturbations, the expansion of nucleosome spacing around AP-1 binding sites occurs mostly in sites that are located in promoters (Fig. 4e). In contrast, expansion of nucleosome spacing around ZNF143 binding sites occurs mostly in sites that are located in enhancers. An exception to this trend is ATF1: Knock-out of chromatin remodelers results in nucleosome expansion around ATF1 binding sites in promoters, but compaction in ATF1 binding sites in enhancers (Supplementary Fig. 13a, b). For specific chromatin modifiers, we often observed more expansion in either enhancers or promoters (Supplementary Fig. 13c). Knock-out of CoREST subunit SFMBT1 tends to cause nucleosome expansion around TFBSs in promoters but not in enhancers, for the 7 TFs analyzed: for example, an 85 nt expansion around AP-1 binding sites in promoters and no change in nucleosomal positions around AP-1 binding sites in enhancers (Fig. 4f). In contrast, knock-out of BAF/pBAF subunit SMARCB1 tends to cause nucleosome expansion around TFBSs in enhancers but not in promoters (e.g. at RAD21 sites) (Fig. 4g).

Discussion

In this work, we develop CRISPR-sciATAC, a platform for pooled forward genetic screens that jointly captures CRISPR perturbations and ATAC profiles in single cells. Pooled CRISPR screens have been used extensively to identify genes responsible for therapeutic resistance, cell proliferation, and Mendelian disorders²⁶. While several methods that combine CRISPR screens with single-cell RNA-sequencing have been developed^2–5, the ability to capture changes in chromatin accessibility following CRISPR perturbations have been limited. Rubin and collaborators⁶ published a related method (Perturb-ATAC) which uses a programmable microfluidic device to physically isolate single cells into small chambers. This method delivers high-depth single-cell ATAC-seq data (~10⁴ fragments per cell), but the throughput per experiment is limited to the 96 chambers of the microfluidic device. CRISPR-sciATAC offers an alternative approach that takes advantage of two-step combinatorial indexing to label DNA molecules with unique cell barcodes and requires no specialized equipment. When compared with Perturb-ATAC, CRISPR-sciATAC can generate thousands of single cells at ~20x less reagent cost and requires ~14x less time (Supplementary Tables 8 and 9). In this work, we analyzed 28,510 cells, which is ~7-fold more cells than in the Perturb-ATAC dataset (Supplementary Fig. 14a). Using a library of 318 gRNAs targeting 105 genes, we investigated differential accessibility at histone and DNA modifications and at TFBSs following loss of chromatin modifiers. By perturbing chromatin remodeling complexes in a high-throughput and uniform setting, we reduce batch effects and generate data for a large number of different chromatin complexes. Since it is based on combinatorial indexing ATAC-seq, CRISPR-sciATAC shows comparable yield to published sciATAC datasets, which do not have the additional modality of gRNA capture (Supplementary Fig. 14b). As we demonstrate with gRNAs targeting EZH2, one important caveat of CRISPR nuclease-driven perturbation is that knock-out can be incomplete due to in-frame repair and efficiency can vary depending on the gRNA. However, CRISPR-sciATAC does not depend critically on any specific gRNA but rather looks for consistent effects between gRNAs. To more completely address these issues, future computational methods to discern perturbed cells from unperturbed cells, as has been recently developed for single-cell RNA-sequencing²⁷, will also be useful for single-cell ATAC-seq.

Future CRISPR-sciATAC studies could profile the effects of genetic perturbations on chromatin accessibility in diverse tissues and organoids, and in biological processes where dramatic changes in chromatin accessibility occur such as differentiation or oncogenic transformation. Although we have targeted primarily chromatin modifiers whose loss might be expected to result in large changes in accessibility, perturbations that result in more subtle accessibility changes can also be detected but may require greater cell numbers. Cell throughput could be increased by uniting the combinatorial approach of CRISPR-sciATAC with droplet-based single-cell methods. This hybrid strategy can yield high complexity single-cell ATAC-seq libraries with the ability to profile many more perturbations, including potentially genome-scale CRISPR libraries. Overall, CRISPR-sciATAC can be applied to study diverse phenotypes and diseases and to understand the interaction between genetic changes and genome-wide chromatin accessibility.

Methods

Cell culture and monoclonal K562-Cas9 cell line

NIH-3T3 and K562 cells were acquired from ATCC (CRL-1658 and CCL-243). HEK293FT cells were acquired from Thermo Fisher (R70007). NIH-3T3 (mouse) and HEK293FT (human) cells were maintained at 37°C with 5% CO₂ in D10 media: DMEM with high glucose and stabilized L-glutamine (Caisson DML23) supplemented with 10% fetal bovine serum (Thermo Fisher 16000044). K562 cells were maintained at 37°C with 5% CO₂ in R10 media: RPMI with stabilized L-glutamine (Thermo Fisher 11875119) supplemented with 10% fetal bovine serum. To generate monoclonal K562 cells expressing Cas9, K562 cells were transduced with lentiCas9-Blast (Addgene 52962) at a multiplicity of infection (MOI) of 0.1 and selected and maintained in R10 with 5 μg/ml blasticidin. Monoclonal K562-Cas9 cells were isolated and expanded through limiting dilution. Expression of Cas9 was confirmed by Western blot using an anti-2A peptide antibody (Millipore Sigma MABS2005).

Lentiviral CRISPR libraries

To generate NIH-3T3 and HEK293FT cells expressing gRNAs for the human/mouse experiment, 10 human non-targeting gRNAs and 10 mouse non-targeting gRNAs (Supplementary Table 2) were individually synthesized and cloned into the lentiviral transfer vector CROPseq-Guide-Puro² (Addgene 86708), which leads to the synthesis of an RNA Pol3 transcript of the Cas9 gRNA and an RNA Pol2 polyadenylated transcript containing the puromycin resistance gene, a U6 promoter, and the gRNA. The RNA Pol2 transcript allows for the selection of transduced cells (via puromycin) and detection of the gRNA targeting sequence via reverse-transcription and PCR (Supplementary Fig. 1c). Equal amounts of each gRNA plasmid were mixed and then, with packaging plasmids pMD2.G (Addgene 12259) and psPAX2 (Addgene 12260), transfected into HEK293FT cells²⁸. NIH-3T3 and HEK293FT cells were transduced at MOI ~ 0.1 and selected and maintained in D10 with 1 μg/ml puromycin. The gRNA library coverage was 1,500x on average for the species-mixing experiment, the chromatin modifier screen and the chromatin remodeling complex subunit screen.

For the chromatin modifier pooled CRISPR screen, we identified 21 frequently mutated chromatin modifiers across all cancers in the Catalogue of Somatic Mutations in Cancer (COSMIC) database⁸ (Supplementary Fig. 3a, b) and designed three targeting gRNAs per gene. An important issue to consider when designing gRNAs is that not all CRISPR nuclease-driven modifications will result in loss-of-function, as some genome modifications (non-homologous end-joining) will result in in-frame repair that may preserve gene function. However, it has been previously demonstrated that even in-frame mutations can be disruptive when targeting functional domains of proteins²⁹. To capitalize on this discovery, we chose a CRISPR library design algorithm that uses protein functional domains (from the Pfam database) to target our gRNAs³⁰. The final library was composed of 63 targeting and 3 non-targeting gRNAs that were individually synthesized (IDT) and annealed (Supplementary Table 2). For the chromatin remodeling complex subunit pooled CRISPR screen, we designed a CRISPR library to target all chromatin remodeling complexes in the human genome, as defined by the EpiFactors database¹⁵ (Fig. 3a). The library was composed of 252 targeting and 3 non-targeting gRNAs that were individually synthesized (IDT) and annealed (Supplementary Table 2). Annealed oligos were pooled in equimolar ratio and cloned as a pool into the CROPseq-Guide-Puro lentiviral transfer vector. K562-Cas9 cells were transduced at a MOI of ~0.1 and selected and maintained in 1 μg/ml puromycin and 5 μg/ml blasticidin. The CRISPR-sciATAC protocol was performed on these cells at one week post-selection.

Transposase identification and isolation

We were motivated to use a different transposase than Tn5 due to the difficulty of obtaining sufficient yields of Tn5³¹. In order to identify new transposases, sequences were aligned using ClustalW³² (version 2.1). We found a range of transposon sequences that were related to the Tn5 sequence and selected a transposon from Vibrio parahemolyticus (ViPar) for further analysis. The inside and outside ends (IE and OE) of the ViPar transposon utilize the same sequence as the IE and OE of the Tn5 transposon, giving us confidence the ViPar transposon would be compatible with existing Tn5-based workflows (Supplementary Fig. 2a, b). The identified ViPar transposase was synthesized (Twist Bioscience) and cloned into the vector pTXB1 (NEB, N6707S). Two mutations were introduced: (1) P50K, equivalent to the mutation E54K in Tn5, which is predicted to make the transposon hyperactive³³ and (2) M53Q, which changes the residue that interacts with nucleotide 9 (a thymine) on the non-transferred strand of the mosaic end (ME) similar to Tn5 Q57, predicted to increase binding to the Tn5 ME. The ViPar transposase with P50K and M53Q mutations, henceforth referred to as TnY, showed Tn5 ME loading and tagmentation activity (Supplementary Fig. 2c–f). Finally, we characterized the insertion site preference of TnY by performing tagmentation on NA12878 DNA and sequencing on a MiSeq Instrument (Illumina); we found that TnY has insertion site preferences distinct from, but of a similar magnitude to those of Tn5 (Supplementary Fig. 2g, h). The chromatin accessibility profiles resulting from TnY and Tn5 are highly correlated (Supplementary Fig. 2i, j).

TnY transposase production

The pTXB1-TnY vector was transformed into BL21(DE3) competent E. coli cells (NEB C2527) and TnY was produced via intein purification with an affinity chitin-binding tag³¹. One liter of LB culture was grown at 37°C to OD600 = 0.6. TnY expression was then induced with IPTG 0.5 mM at 18°C overnight. After induction, cells were pelleted and then frozen at −80°C overnight. Cells were then lysed by sonication in 100 ml HEGX (20 mM HEPES-KOH at pH 7.5, 0.8 M NaCl, 1 mM EDTA, 10% glycerol, 0.2% Triton X-100) with a protease inhibitor cocktail (Roche 04693132001). The lysate was pelleted at 30,000 x g for 20 min at 4°C. Supernatant was transferred to a new tube, 3 μl of neutralized PEI 8.5% (Sigma Aldrich P3143) was added dropwise to each 100 μl of bacteria extract, gently mixed and centrifuged at 30,000 x g for 30 minutes at 4°C to precipitate DNA. The supernatant was loaded on four 1-ml chitin columns (NEB S6651S). Columns were washed with 10 ml HEGX; 1.5 ml HEGX containing 100 mM DTT was added to the column and incubated for 48 h at 4°C to allow cleavage of TnY from the intein tag. TnY was eluted directly into two 30 kDa MWCO spin columns (Millipore UFC903008) by adding 2 ml of HEGX. Protein was dialyzed in five dialysis steps using 15 ml 2x Dialysis Buffer (100 HEPES-KOH at pH 7.2, 0.2 M NaCl, 0.2 mM EDTA, 2 mM DTT, 20% glycerol) and concentrated to 1 ml by centrifuging at 5,000 x g. The protein concentrate was transferred to a new tube and mixed with an equal volume of glycerol 100%. Then, we added Triton X-100 (0.04% final concentration). TnY aliquots were stored at −80°C.

Transposome assembly

To produce mosaic-end double-stranded (MEDS) oligos, we annealed the single T5 tagmentation oligo with the pMENT common oligo (100 μM each) (Supplementary Table 2) as follows in TE buffer: 95°C for 5 minutes, then cooled at a rate of 0.2°C /s down to 4°C (“MEDS A”). The same process was used to anneal each barcoded T7 tagment sciATAC oligo with the pMENT common oligo (“MEDS B”) (Supplementary Table 2). MEDS A and MEDS B were mixed together, diluted 1:6 in TE buffer and 2 μl were transferred into a new tube and mixed with 3 μl of TnY enzyme. After 30 minutes at room temperature to allow for transposome assembly, we added 45 μl Dilution Buffer, mixed by pipetting up and down and stored at −20°C until ready for tagmentation. Dilution Buffer consists of 2x Dialysis Buffer (see TnY transposase production above) diluted 1:1 by volume with 100% glycerol. We observed optimal tagmentation when transposome assembly was carried out on the same day as the CRISPR-sciATAC tagmentation.

PfuX7 polymerase production

For CRISPR-sciATAC, we used a purified PfuX7 DNA polymerase³⁴. First, we transformed BL21(DE3) competent E. coli cells (NEB C2527) with pET-PfuX7 and grew them in 1 L of LB culture at 37°C to OD₆₀₀ = 0.6. PfuX7 expression was then induced with IPTG (0.5 mM final concentration) at 30°C overnight. After induction, cells were pelleted and resuspended in 20 ml Lysis Buffer (50 mM Tris-HCl pH8, 150 mM NaCl, 1 mM EDTA, 1 mM PMSF, 10 μg/ml EDTA-free protease inhibitor (Sigma 11873580001)) and sonicated in an ice slurry. Sonication was at 20% amplitude for ten cycles of 1 minute duration with a 30 second pause between cycles (Branson Ultrasonics, Model 450 Digital Sonifier). The lysate was pelleted at 30,000 x g for 15 min at 4°C. Supernatant was transferred to a new tube and incubated with DNA Digestion Buffer (20 μl DNaseI (NEB M0303), 0.5 mM CaCl2, 2.5 mM MgCl2) for 30 minutes at 37°C. DNaseI was then inactivated by incubating for 30 minutes at 85°C. After inactivation, the lysate was placed on ice for 20 minutes. Lysate was then centrifuged at 50,000 x g for 20 minutes at 4°C. Supernatant was loaded on two 1-ml Ni-NTA (Qiagen 30210) columns, washed twice with Wash Buffer (50 mM Tris-HCl pH 8, 150 mM NaCl). PfuX7 enzyme was eluted in 5 ml Elution Buffer (50 mM Tris-HCl pH 8, 150 mM NaCl, 0.25 M imidazole) and desalted in Storage Buffer (100 mM Tris-HCl pH 8, 0.2 mM EDTA, 2 mM DTT) by performing buffer exchange three times using one Amicon 30 kDa MWCO spin column (Millipore UFC903008). The purified protein was then transferred to a new tube, combined with equal volume of 100% glycerol and adjusted with Tween-20 (0.1% final concentration) and IGEPAL CA630 (0.1% final concentration). Aliquots were stored at −20°C.

Bulk ATAC-seq

For bulk ATAC-seq³⁵, we resuspended 500,000 cells in 1 ml PBS and gently lysed them by adding 10 ml Resuspension Buffer (10 mM Tris-HCl at pH 7.5, 10 mM NaCl, 3 mM MgCl₂) with 0.1% Tween-20. Cells were then centrifuged at 500 xg for 10 min at 4°C to pellet the nuclei. Pelleted nuclei were resuspended in 600 μl 1x Tagmentation Buffer (10 mM TAPS-NaOH at pH 8.5, 5 mM MgCl₂, 10% DMF), 30 μl (~25,000 nuclei) were then transferred into 1.5 ml tubes and 20 μl TnY transposomes were added. Tagmentation was performed at 37°C for 30 min. Samples were then purified using the DNA Clean & Concentrator kit (Zymo Research D4014) and eluted in 10 μl TE. Eluted DNA was thermocycled with PfuX7 in Phusion GC Buffer (Thermo Fisher F519L) as follows: 72°C 5 min, 98°C 30 s, (98°C 10 s, 63°C 30 s, 72°C 3 min) x 10 cycles, 4°C hold. Samples were purified using the DNA Clean & Concentrator kit, eluted in 6 μl TE and size-selected using a 0.9X volume of Ampure XP Beads (Beckman Coulter A63882) to remove excess oligos.

CRISPR-sciATAC species mixing experiment in HEK293 and NIH3T3 cells

HEK293FT (human) and NIH-3T3 (mouse) transduced with non-targeting gRNAs libraries were grown separately. On the day of the experiment, cells were counted, and 500,000 cells were resuspended in 1 ml PBS per cell line (1:1 ratio of human and mouse cells). Cells were then pelleted, resuspended in Fixation Buffer and fixed for 7 min at room temperature. Fixation Buffer consists of 2.8 ml H₂O, 790 μl 100% ethanol, 310 μl 40% glyoxal (Sigma 128465), 30 μl glacial acetic acid (Sigma A6283); after preparing Fixation Buffer, adjust the pH to 5.0 by adding NaOH and keep ice-cold until immediately before use. In line with a previous study³⁶, we found that glyoxal fixation resulted in better preservation of intact nuclei than the more commonly used paraformaldehyde fixative.

After fixation, cells were then washed three times with 1 ml PBS and gently lysed by adding and resuspending in 10 ml Resuspension Buffer (see Bulk ATAC-seq above) with 0.1% Tween-20 and 0.1% Igepal CA630. Cells were then incubated on ice for 3 minutes and then pelleted at 500 xg for 10 min at 4°C to obtain nuclei. Nuclei were washed in 1 ml Tagmentation Buffer (see Bulk ATAC-seq above) with 5 μl RiboLock RNase Inhibitor (ThermoFisher EO0381) and centrifuged at 500 xg for 5 min at 4°C. Human and mouse nuclei were resuspended and mixed together in a final volume of 3.2 ml Tagmentation Buffer with 28 μl RiboLock RNase Inhibitor. Nuclei (30 μl, ~20,000) were distributed into each well of a 96-well plate containing 20 μl of TnY assembled with MEDS A and 96 barcoded MEDS B (see Supplementary Table 2 for MEDS sequences). Tagmentation was performed for 30 minutes at 37°C and then stopped by adding 2 μl EDTA 500 mM into each well. After incubating for 15 minutes at 37°C, EDTA was quenched prior to reverse transcription by adding 2 μl of 50 mM MgCl₂ into each well.

For reverse transcription, 5 μl of the nuclei solution (~2,000 nuclei) were transferred into a new 96-well plate containing barcoded reverse transcription primers. Reverse transcription primers contain the same barcode as the MEDS B oligos (see Supplementary Table 2 for RT oligos). Nuclei were transferred keeping plate orientation to match tagmentation and reverse transcription barcodes. The reverse transcription master mix (RTMM) consisted of 1 mL 5x RT buffer, 270 μl dNTPs, 1.6 mL water, 262 μl RevertAid reverse transcriptase, 27 μl RiboLock RNase Inhibitor (all components: Thermo Fisher, EP0442). We distributed 15 μl of RTMM into each well, mixed, and incubated for 30 min at 37°C.

Reverse transcription was stopped by adding 2 μl of Stop and Stain buffer (1 mL 500 mM EDTA, 2 μl of 5 mg/ml DAPI) and incubated for 5 minutes on ice. Nuclei were pooled together and pelleted at 500 xg for 5 min at 4°C. Supernatant was carefully removed taking care to not disturb the pellet. The nuclei were gently resuspended in 250 μl PBS and counted using a hemocytometer. PBS was added in order to obtain a final concentration of 10 nuclei/μl. 2 μl of the nuclei solution (~20 nuclei) were transferred into a new 96-well plate with DNA extraction and digestion buffer in each well. Specifically, each well contained 24.5 μl of DNA Rapid Extract Buffer (1 mM CaCl₂, 3 mM MgCl₂, 1% Triton X-100, 10 mM Tris-HCl at pH 7.5) and 2 μl of Digestion Buffer (1 μl H2O, 0.5 μl SDS 5.8%, 0.5 μl Proteinase K 20 mg/ml (Sigma P2308)). Nuclei were digested for 5 min at 65°C; digestion was stopped by adding 3 μl PMSF (Sigma 93482) and incubating for 30 min at room temperature.

For the first PCR, ATAC-seq primers and gRNA-PCR1 primers were added at a final concentration of 0.5 μM and 0.1 μM, respectively. Amplification for ATAC-seq/gRNA-PCR1 was performed with PfuX7 in Phusion GC Buffer as follows: 72°C 5 min, 98°C 30 s, (98°C 10 s, 63°C 30 s, 72°C 3 min) x 14–18 cycles, 4°C hold. For the second PCR, 2 μl of PCR product were transferred into a new 96-well plate keeping plate orientation to match ATAC-seq and gRNA barcodes. gRNA-PCR2 primers were added to a final concentration of 0.5 μM. Amplification for gRNA-PCR2 using PfuX7 in Phusion GC buffer was: 98°C 30 s, (98°C 10 s, 55°C 10 s, 72°C 20 s) x 20 cycles, 72°C 5 min, 4°C hold.

We then purified ATAC-seq and gRNA amplicons. The ATAC-seq/gRNA-PCR1 PCR plate was purified using four columns of the DNA Clean & Concentrator kit, eluted in 10 μl elution buffer and size-selected using 0.9X volume of Ampure XP Beads. The gRNA-PCR2 PCR plate was purified using ten columns of the DNA Clean & Concentrator kit, eluted in 20 μl elution buffer. Eluted samples were run on E-gel 2% (Thermo Fisher G402002) and the expected band (~250 bp) gel extracted, purified using 1 column of Zymoclean Gel DNA Recovery Kit (Zymo Research D4008) and eluted in 20μl. Libraries were separately sequenced on the MiSeq Sequencer (Illumina) using the read lengths shown in Supplementary Fig. 1d,e^37,38.

CRISPR-sciATAC for chromatin modifiers in K562 cells

The CRISPR-sciATAC protocol for the chromatin modifier library in K562 cells was performed similarly to the human/mouse experiment described above. K562-Cas9 cells transduced with the pool of 63 chromatin modifiers gRNAs and 3 non-targeting gRNAs (library 1) or with the pool of 252 chromatin modifiers gRNAs and 3 non-targeting gRNAs (library 2) and were cultured for one week after selection. We prepared between either 12 (library 1) or 41 (library 2) 96-well plates and pooled amplicons. The ATAC-seq amplicons were sequenced on a HiSeq 2500 (Illumina) and the gRNA amplicons were sequenced on a MiSeq.

Gene essentiality screen and analyses

K562-Cas9 cells were transduced with the chromatin modifiers pooled CRISPR screen at MOI ~ 0.1 and selected and maintained in 1 μg/ml puromycin and 5 μg/ml blasticidin. Genomic DNA was extracted at three days (early time point), one week and two weeks post-selection. The gRNA cassette was PCR amplified². Libraries were sequenced on a MiSeq sequencer (Illumina). In addition to the CRISPR-sciATAC experiment, two independent transduction replicates were also analyzed. To identify essential genes, a p-value per gRNA was calculated using the MAGeCK algorithm and p-values for the three gRNAs targeting one gene were aggregated into a gene-level p-value using a Robust Rank Aggregation approach followed by a Bonferroni correction^9,39.

Read alignment

CRISPR-sciATAC gRNA and ATAC datasets were demultiplexed based on cellular barcodes using the snATAC_mat.py script in an established sci-ATAC-seq pipeline (https://github.com/r3fang/snATAC)⁴⁰. The processed gRNA sequences were aligned to a custom guide reference using bowtie⁴¹ (version 1.1.2) using the command bowtie -v 1 -m 1. For the human-mouse experiment, we show data for cells with at least 20 mapped reads. Cells with over 90% of gRNA reads that mapped exclusively to human or mouse gRNAs were considered species-specific cells. Cells where one gRNA represented at least 90% of the total reads were kept for further analyses. The remaining cells were considered collisions and/or the result of multiple infections. For downstream analysis of the K562 data, we required each cell to have at least 100 aligned gRNA reads with ≥ 99% of the reads assigned to one gRNA sequence for the chromatin modifier screen and at least 10 aligned gRNA reads with ≥90% of the reads assigned to one gRNA sequence for the chromatin remodeling complex subunit screen. Gene knock-outs with at least 50 identified single cells were considered for further analysis (98/105 targeted genes).

The processed ATAC sequences were aligned to the reference genome using bowtie2⁴² (version 2.2.8) using the command bowtie2 -D 15 -R 2 -L 22 -i S,1,1.15 -p 5 -t -X2000 -e 75 --nomixed --no-discordant. The reference genome was a chimeric human hg19 and mouse mm10 genome for the human-mouse experiment and a human hg19 for the K562 datasets. Improperly paired and non-uniquely mapped alignments and reads mapping to mitochondrial DNA were removed. Reads overlapping ENCODE blacklist regions were removed (https://www.encodeproject.org/annotations/ENCSR636HFF/). Reads were then deduplicated using Picard (version 2.16.0) (http://broadinstitute.github.io/picard). For the human-mouse experiment, we show data for cells with at least 20 unique ATAC-seq reads. For the K562 datasets, we require at least 500 unique ATAC-seq reads.

Differential accessibility at genomic regions with specific chromatin and DNA modifications

To assess changes in accessibility, we downloaded from ENCODE ChIP-seq files covering post-translational histone modifications and DNA methylation (Supplementary Table 4). For each ChIP-seq track, we considered the fraction of fragments in each single cell that overlap ChIP-seq peaks. We standardized the averaged fractions over all single cells into Z-scores and then averaged the Z-scores obtained for each ChIP-seq file over cells that received the same gRNA for the visualization in Fig. 2a. To find significant deviations in accessibility per gene-KO and per modification, we performed a two-tailed t-test on the Z-scores, of all cells for one gene knock-out and all the non-targeting cells, for each modification. The p-values were adjusted for multiple hypothesis testing using a Benjamini-Hochberg false-discovery rate correction (q ≤ 0.1).

For correlation of downsampled cell populations with the aggregated (pseudo-bulk) data, we randomly sampled cells (from a total of 400 single cells) without replacement. We performed this resampling procedure 200 times for each cell number. For each cell sample, we average the accessibility Z-scores and then compute the Pearson correlation with the pseudo-bulk. For this analysis, we only included target genes with at least 400 single cells.

Differential accessibility in TF binding sites using ENCODE ChIP-seq

To identify enrichment or depletion in accessibility of TF binding sites following chromatin modifier knock-out, we downloaded 116 TF K562 ChIP-seq peak files from ENCODE (Supplementary Table 4) and considered the fraction of fragments in each single cell that overlap ChIP-seq peaks. We standardized the averaged fractions over all single cells into Z-scores and then averaged the Z-scores obtained for each ChIP-seq file over cells that received the same gRNA for the visualization in Supplementary Fig. 9a. For dimensionality reduction, we used the function umap (from the R package umap) and, to predict cell perturbation, we fit TFBS Z-scores with a generalized linear model using the function glm (from the R package stats). To find significant deviations in accessibility per gene-KO and per TF, we performed a two-tailed t-test on the Z-scores, of all cells for one gene knock-out and all the non-targeting cells, for each TF. The p-values were adjusted for multiple hypothesis testing using a Benjamini-Hochberg false-discovery rate correction (q ≤ 0.1). For genes with multiple ENCODE ChIP-seq datasets, we denote with (1) ENCODE ChIP-seq profiles obtained using an antibody that directly recognizes the protein of interest; we denote with (2) ENCODE ChIP-seq profiles obtained using an antibody directed against an EGFP-tag.

Differential accessibility in TF binding sites using JASPAR motifs

As an orthogonal method to ENCODE ChIP data, we also utilized predicted TF binding sites from the JASPAR database (386 motifs from JASPAR 2016, human CORE dataset)¹³. Transcription factor motif enrichment and depletion scores were calculated using chromVAR¹⁴. Briefly, Z-scores quantifying deviations in the frequency of each motif in each of the single cells were calculated based on the frequency of the motif in the collection of peaks that exist in each cell, out of all 358,028 peaks called on the aggregated single cell alignment files (pseudo-bulk). This frequency was compared to the frequency of the motif in peaks found in the entire aggregated single cell dataset¹⁴. We considered cells with a minimum of 2000 fragments per cell and a minimum of 10% of total fragments in peaks. To avoid biases from recovery of different numbers of cells for each gRNA, we subsampled all gRNA cell populations to 12 cells (the lowest number of cells for a single gRNA in our K562 dataset), calculated the deviation Z-scores, and repeated this resampling process 1000 times to obtain deviation Z-scores for each gRNA.

Gene ontology analysis of differential EZH2 chromatin accessibility sites

In order to identify and annotate genomic regions that are differentially accessible in cells with EZH2-targeting gRNAs, we aggregated equal numbers of single cells (n = 170 cells per gRNA) for each of the three EZH2 and non-targeting gRNAs. We next binned the genome into 150 nt regions and identified all bins covered by all three EZH2 gRNAs and not covered by any of the three non-targeting gRNAs. These bins were then mapped to the transcription start site of the closest genes. We used this (unranked) gene list (n = 3,740) as input for Gene Ontology enrichment analysis, with all human genes as a background set⁴³.

Differential accessibility at HOX loci and gene expression

To measure accessibility at HOX loci, EZH2-targeted and non-targeting single cells were downsampled to 100 cells, aggregated and fragments overlapping the HOXA-D loci were counted. Empirical p-values were calculated over 1000 bootstrap iterations. To select HOX genes for expression profiling, we compared CRISPR-sciATAC coverage in EZH2 KO cells and NT cells. We computed the number of reads in each HOX gene body (including 500 nt flanking sequence on each side). We then selected the top 5 HOX genes with the most significant change in CRISPR-sciATAC coverage (Student’s t-test). Gene expression of HOX genes (HOXA3, HOXA5, HOXA11, HOXA13, HOXD9) and EZH2 following EZH2 knock-out was quantified using quantitative qRT-PCR. Briefly, 1 million K562-Cas9 cells were infected with EZH2 gRNA 1–3 or NT gRNA 1–3 at an MOI of ~0.1 for each of the 6 gRNAs and grown in 6-well plates. At 24h post-infection, cells were selected in 1 μg/ml puromycin. Cells were harvested 10 days after transduction and lysed using TRIzol (Life Technologies), RNA was purified using Direct-zol (Zymo Research). We reverse-transcribed 1 αg of total RNA using random hexamer primers and RevertAid Reverse Transcriptase (Thermo Fisher) at 25°C for 10 min, 37°C for 60 min, and 95°C for 5 min. After cDNA synthesis qPCR reactions were performed using Luna Universal Probe qPCR Master Mix (NEB), custom primers and probes (IDT) were designed to detect each target gene and normalized to β-actin (ACTB) (see Supplementary Table 2 for primer and probe sequences). All qPCRs were thermocycled on a ViiA 7 Real-Time PCR System (Applied Biosystems) as follows: initial denaturation at 95 °C for 1 min, then 40 cycles at 95 °C for 15 s, 60 °C for 30 s. Quantification was performed via the ΔΔC_t using 3 biological replicates and 4 technical (qPCR) replicates for each biological replicates.

eQTL enrichment

To test if targeting chromatin modifiers resulted in changes in accessibility at SNPs associated with regulatory function through expression quantitative trait locus (eQTL) association testing, we utilized cis-eQTLs (SNP-gene combinations within 1 Mbp) from the eQTLGen consortium. The consortium performed association testing for 19,960 genes expressed in blood in 31,684 samples⁴⁴. We considered the fraction of fragments in each single cell that overlap cis-eQTLs and compared these fractions for each population of single cells that received gRNAs targeting a gene to the fractions in non-targeting cells using a Wilcoxon signed-rank test followed by a Benjamini-Hochberg multiple hypothesis correction. To identify specific cis-eQTLs with altered accessibility, we downsampled KDM6A single cells and non-targeting single cells to the same number of cells (n = 737 cells) and focused on a subset of 7829 highly covered (≥ 50 reads) cis-eQTLs in the two cell populations combined. For each of these 7829 cis-eQTLs we considered the proportion of cells with a read covering the cis-eQTL in the KDM6A cell population (n = 921 cells) and in the non-targeting cell population (n = 737 cells) and performed a χ² test of proportion. Allelic effects are from the Genotype-Tissue Expression (GTEx) database⁴⁵. For each cis-eQTL, we show allele specific expression for the two closest genes in whole blood samples. In cases where allele specific expression was not available in whole blood samples, we show the most significant association.

Differential accessibility in enhancers and promoters

For each single cell, we calculated the fraction of reads that intersect with promoters and with enhancers, as defined by ENCODE (wgEncodeAwgSegmentationCombinedK562.bed, http://hgdownload.cse.ucsc.edu/goldenpath/hg19/encodeDCC/wgEncodeAwgSegmentation/). We then compared the fractions in each gene knock-out cell population with the fractions in the non-targeting cell population. To find significant differences in reads in promoters/enhancers (versus non-targeting), we performed a two-sample Wilcoxon test and the p-values were adjusted for multiple hypothesis testing using a Benjamini-Hochberg false-discovery rate (q ≤ 0.1).

Nucleosome dynamics at TFBS, promoters and enhancers

To investigate nucleosome dynamics around TFBSs, we first subset the ATAC fragments into fragments putatively spanning one nucleosome (mono-nucleosome fragments, 147 – 280 bp³⁵). We next calculated coverage profiles around TFBSs sites with BEDTools⁴⁶ (version 2.25.0). We focused on TFBSs that had two nucleosomes spanning them symmetrically as seen in our data. The TFBS selected for this analysis thus reflect those whose nucleosome positions are strongly bimodal. We chose these sites by calculating a nucleosome-free region (NFR) score for each, a metric to assess bi-modality⁴⁷. Specifically, we take the difference in average base-pair coverage between flanking regions (50 to 150 bp upstream and downstream of site) and the central region (50 bp across site center). We focused on a small subset of 7 TFs with NFR score of more than zero, and indeed, for the binding sites of TFs, we observe strong bimodality. These TFs have also been previously shown to have a bimodal profile²⁴.

ATAC-seq fragment coverage plots were smoothed using the smooth.spline function (from the R package stats), with smoothing parameter spar = 0.8. Next, positions of maximum coverage upstream and downstream of motif centers were used to estimate nucleosome location. To determine expansion, we first calculated the distance between the upstream and downstream nucleosomes at a particular TFBS. Then, this distance was compared to non-targeting cells, to obtain a positive (expansion) or negative (compaction) score. Empirical p-values for each score were generated using a label-permutation test, where non-targeting and knock-out labels were randomly shuffled while keeping group size constant to avoid biases that stem from different numbers of cells in the non-targeting and knock-out cell populations. Labels were shuffled 10,000 times and in each iteration the distance between the upstream and downstream nucleosomes was measured, to create a null distribution to which the true distance was compared.

For each TF, we calculate mononucleosomal coverage profiles separately for sites located in promoters or enhancers, as defined by UCSC (wgEncodeAwgSegmentationCombinedK562.bed, http://hgdownload.cse.ucsc.edu/goldenpath/hg19/encodeDCC/wgEncodeAwgSegmentation/).

Statistical analysis

Data between two groups were analyzed using a two-tailed unpaired t-test or a non-parametric Wilcoxon signed-rank test. The p values and statistical significance were estimated for all analyses. Corrections for multiple-hypothesis testing was performed using the Benjamini-Hochberg approach⁴⁸. In all the box plots, the central rectangle in the plot covers the first to the third quartile (the interquartile range, or IQR) and the bold line is the median. The whiskers are defined as: whisker_upper = min(max(x), Q₃ + 1.5 × IQR) and whisker_lower = max(min(x), Q₁ − 1.5 × IQR). All statistical analyses were performed in R/RStudio.

Data availability

Processed and raw data can be downloaded from NCBI GEO (PRJNA674902, GSE161002).

Code availability

The scripts and pipeline for the analysis can be found at https://gitlab.com/sanjanalab/crispr-sciatac.