Next-Generation Forward Genetic Screens: Uniting High-Throughput Perturbations with Single-Cell Analysis

Abstract

Programmable genome engineering technologies, such as CRISPR nucleases, and massively-parallel CRISPR screens that capitalize on this programmability have transformed biomedical science. These screens connect genes and noncoding genome elements to disease-relevant phenotypes but, until recently, have been limited to singular phenotypes such as growth or fluorescent reporters of gene expression. By pairing massively-parallel screens with high-dimensional profiling of single cell types/states, we can now measure how individual genetic perturbations or combinations of perturbations impact the cellular transcriptome, proteome and epigenome. Here, we review technologies that pair CRISPR screens with single-cell multiomics and the unique opportunities afforded by extending pooled screens using deep multimodal phenotyping.

CRISPR-BASED TOOLS FOR POOLED SCREENS

Over the last decade, diverse CRISPR (clustered regularly interspaced short palindromic repeats) Cas (CRISPR-associated) systems have been harnessed for targeted genome modification [1–4]. A major feature of these systems is their programmability, enabling high-throughput functional genomic screens such as genome-wide knock-out screens [5,6].

The best known CRISPR nuclease — S. pyogenes Cas9 — uses a 20-nucleotide guide RNA (gRNA) to induce genetic alterations at specific genomic locations that are complementary to the gRNA sequence [7–9]. Cas9 creates double-stranded breaks [7], but engineered Cas9 mutants can create strand-specific nicks (nCas9) [10], modulate gene expression with a catalytically inactive enzyme (dCas9) tethered to other functional domains [11–14], and bind regions with different protospacer adjacent motif (PAM) target requirements [15–17]. Another CRISPR nuclease that has been used in pooled screens is Cas12a (also known as Cpf1), which recognizes a different (T-rich) PAM and allows for easier multiplexing of multiple gRNAs [18]. Certain CRISPR systems, such as with Cas13 [19] and Cas7–11 [20], target RNA for cleavage instead of DNA. Modifications to Cas9 include fusing domains which enable gene expression control through repression (CRISPRi) [12,21], activation (CRISPRa) [22,23], or other types of chromatin remodeling [24]. Cas9 and other DNA-targeting CRISPRs have also been engineered to create single-nucleotide changes via cytosine [25] and adenine base editors [26], or via prime editors [27] (Fig. 1).

Figure 1. CRISPR systems for targeted DNA, epigenome, and RNA editing.

DNA-targeting CRISPR systems can induce double-stranded breaks (Cas9 or Cas12a) to inhibit gene function or induce strand-specific nicks to perform targeted nucleotide mutagenesis (Cas9 nickases, Cas9n) through base editing or prime editing and specialized prime editing gRNAs (pegRNAs). Epigenome-targeting CRISPR systems use a nuclease-dead Cas protein (dCas9) to recruit transcriptional activators or repressors at specific loci. RNA-targeting CRISPR systems use single-strand cutting (Cas13 or Cas7–11) or nuclease-dead Cas proteins (dCas13) to knockdown or perform targeted RNA nucleotide mutagenesis, respectively.

For functional genomic screens, CRISPR systems and associated gRNA libraries can be delivered in: an arrayed fashion, for studies of the effects of typically a few perturbations individually; or a pooled fashion, for genome-scale studies with thousands of perturbations and/or combinations of perturbations [28]. High-content CRISPR screening has been discussed at length by Bock et al. [29], and in this review, we discuss recent advances in single-cell readouts for CRISPR-based pooled screens, matching the large scale of genetic perturbations with similarly high-dimensional multiomic profiling.

TECHNOLOGIES FOR CRISPR GUIDE CAPTURE IN SINGLE CELLS

The first pooled CRISPR screens relied on cell survival to reveal genes required by organisms under certain environmental conditions through measuring the depletion or enrichment of gRNAs targeting specific genes [5,30]. Pooled CRISPR screens can also use cell sorting for phenotypic selection prior to measuring gRNA abundances to further inform cellular processes. For example, genome-wide CRISPR knockout screens have identified genes important for T cell activation, polarization, or differentiation using fluorescence-activated cell sorting (FACS) [31,32]. After CRISPR knockout, T cells are sorted based on markers for activation or differentiation of T cells in order to identify key genes that regulate these aspects of T cell biology. For example, Shifrut et al. [31] found that loss of FAM105A increased resistance of cytotoxic T cells to adenosine receptor-mediated immunosuppression, a key mechanism of immune evasion in cancer. Using naive CD4+ T cells, Henriksson et al. [32] identified multiple genes, Pparg and Bhlhe40, with broad effects on helper T cell activation and differentiation. Cell sorting is not limited to FACS-based approaches either, as genome-wide CRISPR knockout screens have also been used to identify genes important for phagocytosis by measuring gRNA abundances after magnetic selection for cells that had phagocytosed magnetic substrates [33]. One particularly powerful extension of sorting-based screens called Flow-FISH uses fluorescent labelling of RNA with in situ hybridization for phenotypic selection based on gene expression [34]. Flow-FISH with tiling CRISPRi libraries can identify regulatory elements, such as enhancer-gene pairs, but throughput is somewhat limited as each locus requires its own screen [35]. Importantly, these methods rely of bulk sequencing of gRNA perturbations after selection via growth, cell sorting or other means. These methods do not provide explicitly the transcriptomes of individual cells and the perturbations they received and typically require follow-on experiments with individual perturbations to quantify changes in gene expression. Single-cell RNA sequencing (scRNA-seq) and related technologies provide an integrated approach to directly connect genetic perturbations to key molecular phenotypes, such as gene expression changes.

CRISPR screens with scRNA-seq readouts (i.e., Perturb-seq [36], CRISP-seq [37], Mosaic-seq [38], CROP-seq [39], and ECCITE-seq [40]) facilitate exploration of gene function and systematic delineation of gene regulatory networks (Table 1). These single-cell CRISPR screening approaches require lentiviral delivery of pooled gRNAs to single cells, where functional gRNAs are expressed with RNA polymerase III promoters and lack polyadenylated (poly-A) tails [9]. Methods such as Perturb-seq, CRISP-seq, and Mosaic-seq utilize a separate barcode sequence with a poly-A tail to indirectly capture the gRNA, whereas ECCITE-seq [40] and later, direct Perturb-seq [41], use direct gRNA capture (Fig. 2). We will briefly describe the differences between indirect and direct gRNA capture.

Table 1.

Single-cell perturbation screens. Each method includes a description of the modalities captured, the CRISPR gRNA or perturbation capture method, the types of pooled screens they enable, and the single-cell partitioning and chemistry.

Single-cell perturbation screen method	Modalities captured	Guide RNA or perturbation capture	Pooled perturbations	Single-cell capture and chemistry	Reference
ECCITE-seq	Transcriptome and cell surface markers	Does not require a specialized gRNA plasmid. Requires a direct capture spike-in oligo.	CRISPR Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Droplet-based single-cell experiments relying on 5’ capture of transcripts.	40
CROP-seq	Transcriptome	Requires a specialized CROP-seq plasmid to capture poly-adenylated gRNA barcodes.	CRISPR Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Combinatorial indexing and droplet-based single-cell experiments relying on 3’ poly-A tail capture.	39
Direct Perturb-seq / Perturb-CITE-seq	Transcriptome and cell surface markers*	Requires specialized gRNA plasmids with encoded capture sequences. Requires a direct capture spike-in oligo.	CRISPR Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Droplet-based single-cell experiments relying on 3’ or 5’ capture of transcripts.	41, 62
TAP-seq	Select transcripts	Can be coupled with gRNA capture method of choice.	Can be coupled with CRISPR screening method of choice. Requires nested primers designed to enrich single-cell sequencing libraries for transcripts of choice.	Droplet-based single-cell experiments relying on 3’ or 5’ capture of transcripts.	52
CaRPool-seq	Transcriptome and cell surface markers	Requires specialized gRNA plasmids with encoded capture sequences in a cleavable gRNA array.	CRISPR Cas13-based screens (e.g. RNA knockout, inhibition, base editing).	Droplet-based single-cell experiments relying on 3’ capture of transcripts.	103
OverCITE-seq	Transcriptome and open reading frames**	Requires a direct capture spike-in oligo to capture open reading frames.	Alternate screening approach for CRISPR activation.	Droplet-based single-cell experiments relying on 5’ capture of transcripts.	104
CRISPR-sciATAC	Open chromatin	Does not require a specialized gRNA plasmid. Requires tagging integrated gRNAs with reverse transcription and PCR.	CRISPR Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Combinatorial indexing-based single-cell experiments relying on DNA tagmentation.	47
Perturb-ATAC	Open chromatin	Does not require a specialized gRNA plasmid. Requires a direct capture spike-in oligo.	Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Physically isolated single cells relying on DNA tagmentation.	70
Spear-ATAC	Open chromatin	Requires a specialized gRNA plasmid with Nextera read adapters flanking the gRNA and a direct capture spike-in oligo.	CRISPR Cas9-based screens (e.g. gene knockout, activation, inhibition, base editing).	Droplet-based single-cell experiments relying on DNA tagmentation.	71

^*Direct Perturb-seq captures the transcriptome only, Perturb-CITE-seq captures both transcriptome and cell surface markers

^**Uses lentivirally transduced open-reading frames as an alternative to CRISPR activation screening

Figure 2. Lentiviral vectors for single-cell pooled CRISPR screens.

Direct Pol3 capture vectors can be used to directly recover the functional gRNA in a pooled CRISPR screen. Direct Pol2 capture allows for the direct recovery of the gRNA sequence with a polyA tail matching the functional gRNA. Indirect Pol2 capture allows for the inference of gRNA sequence by capturing a barcoded capture sequence with a polyA tail.

Indirect gRNA capture methods were designed to be compatible with massively parallel droplet-based scRNA-seq assays that capture mRNA by their poly-A tails, using poly-T sequences [42,43]. For droplet-based scRNA-seq assays, cells are loaded into a partitioner device alongside gel beads containing barcoded reverse transcription (RT) oligo-dT primers, droplets encapsulate the beads and cells along with RT enzymes, cellular transcriptomes are amplified within each droplet, and sequencing libraries are prepared. Droplet-based methods attach droplet-barcodes and unique molecular identifiers (UMIs) to each mRNA captured so they can be assigned to specific droplets (i.e., cells) and counted. Indirect gRNA capture methods use this poly-A capture mechanism to their advantage, by designing lentiviral plasmids with polyadenylated barcodes that would be captured within droplets and sequenced, allowing for the identification of gRNAs per cell (Fig. 2). However, a major caveat of the plasmid design for indirect gRNA capture is that the polyadenylated barcode (transcribed via RNA Pol II) and gRNA (transcribed via RNA Pol III) were 2.5 kb apart, resulting in high barcode-swapping frequencies (1 event per kb) due to lentiviral recombination, causing a ~4.8-fold decrease in gRNA representation [44–46]. This problem is especially insidious because only the barcode is captured making it impossible to know whether this barcode comes from the correct gRNA or not. A rigorous analysis of this problem showed that approximately 50% of all gRNAs with Perturb-seq have barcode swapping between gRNAs [45].

An improved CRISPR droplet sequencing protocol developed by Datlinger et al. [39] (CROP-seq) addressed the barcode-swapping challenge by clever use of the molecular processes underlying lentiviral integration. Datlinger et al. engineered a lentiviral vector that used a EF-1a promoter to transcribe the antibiotic resistance gene and the gRNA. Here, the U6 promoter and its gRNA are placed downstream of the EF-1a promoter in the vector’s 3’ long terminal repeat (LTR) region. Upon integration, the 3’ LTR is copied, thus creating an independent U6-driven gRNA in the 5’ region of the genomically-integrated provirus. The resulting RNA Pol II transcript (driven by EF-1a) encodes both the antibiotic resistance gene and a polyadenylated gRNA detectable as a barcode (Fig. 2, Table 1). Although CROP-seq is a highly versatile approach given its compatibility with single-cell technologies that relies on poly-A tail capture and avoidance of barcode recombination issues seen with Perturb-seq, the CROP-seq approach results in a significantly lower titer virus due to the modified 3’ LTR which may require optimization for some applications [47].

Indirect gRNA capture restricted the scale of previous studies and can be incompatible with the delivery of multiple gRNAs; therefore, direct gRNA capture alongside single-cell transcriptomes offers a more versatile single-cell CRISPR screen [40,41] (Fig. 2, Table 1). Mimitou et al. [40] performed the first direct gRNA capture through the development of Enhanced CRISPR-Compatible Cellular Indexing of Transcriptomes and Epitopes (ECCITE-seq), utilizing gRNA RT primers to directly capture gRNAs by leveraging a template switching oligo (TSO) with a 5’ scRNA-seq droplet-based approach. Replogle et al. [41] later developed Direct Perturb-seq, where appending a capture sequence to the stem loop of gRNA oligos does not affect their function in CRISPRi screens and that separate gel beads containing RT oligo-dT primers complementary to the capture sequence can capture gRNAs. Therefore, for direct gRNA capture and depending on the transcriptome capture strategy, the gRNA capture sequence can be linked to the oligo-dT beads (3’) or spiked in alongside the oligo-dT beads (5’) (Fig. 3a). As with CROP-seq, these direct capture strategies avoid the barcode swapping issues with Perturb-seq.

Figure 3. Modalities beyond the transcriptome for single-cell pooled screens.

a) Droplet-based single-cell capture methods currently allow for CRISPR gRNA, cell surface antibodies/tags/receptors, transcriptome, and open chromatin modalities. b) Combinatorial indexing-based single-cell capture methods currently allow for CRISPR gRNA, transcriptome, and open chromatin modalities.

These recent innovations in single-cell CRISPR screens have also been paired with multiplexed CRISPR technologies to simultaneously activate and repress multiple genes in the same single cells [41] and enabling linked profiling of perturbations with the transcriptome, proteome, and clonotype, all at single-cell resolution [40]. ECCITE-seq is particularly notable as it uses existing CRISPR lentiviral gRNA vectors without any special modification for scRNA-seq studies. We will next explore how these technologies have been applied to drive novel insights into regulatory and disease biology.

APPLICATIONS OF SINGLE-CELL CRISPR SCREENS

Single-cell pooled CRISPR screens have yielded tremendous insights into gene function in vivo and in vitro. By capturing both gRNAs and the transcriptome in a single cell, researchers have identified causal links between perturbations and gene expression patterns at-scale. In an early Perturb-seq study, Dixit et al. [36] performed single-cell pooled CRISPR knockout (CRISPRko) screens to study the consequences of perturbing transcription factors (TF) in mouse bone marrow-derived dendritic cells (BMDC), discovering altered gene expression networks in response to lipopolysaccharide (LPS) stimulation. Through dissecting these networks, the authors identified TFs with known target genes for anti-viral response and were able to nominate new gene functions by examining causal pathways of gene activation (e.g., Stat2 controls guanylate binding protein (GBP) genes through Irf8). They also used Perturb-seq to study temporal regulation of proliferation in K562, a human erythroid progenitor cell model, correctly predicting TF function for GABPA repressing mitochondrial function and YY1 reducing oxidative phosphorylation. Jaitin et al. [37] performed CRISP-seq single-cell pooled CRISPRko screens in mouse immune cells to study innate immunity genes. Jaitin et al. mapped regulatory networks of differentiation and immune responses and, for example, identified known roles of Rela in regulating monocyte inflammatory response and a novel role for Cebpa for regulating differentiation into dendritic or monocytic cells. Shifrut et al. [31] also performed a single-cell pooled CRISPRko screen in primary human T cells to better understand mechanisms underlying sustained T cell proliferation. Single-cell pooled CRISPRko screens continue to be applied to great effect today and, for example, have made novel insights into host immunity in response to Covid-19 infection. Daniloski et al. [48] performed a genome-scale pooled CRISPRko screen in human A549 alveolar cells to identify required host factors for SARS-CoV-2 infection. ECCITE-seq of the top-ranked genes identified six genes (ATP6AP1, ATP6V1A, CCDC22, NPC1, PIK3C3, and RAB7A) that led to upregulation of pathways affecting lipid and cholesterol homeostasis, finding that increases in cholesterol led to SARS-CoV-2 resistance and that cholesterol modulation using existing FDA-approved therapies could block viral infection.

With transcriptional modulation instead of nuclease-driven knock-out, Adamson et al. [49] performed Perturb-seq using CRISPRi (knock-down) to identify genes required for endoplasmic reticulum (ER) homeostasis in the mammalian unfolded protein response (UPR) pathway [49]. Adamson et al. also developed a triple gRNA vector, testing single, double, and triple combinations of gRNAs targeting UPR genes, enabling the delineation of IRE1a-, PERK-, and ATF6-controlled transcriptional programs. Schmidt et al. [50] performed Perturb-seq using CRISPRa to validate 14 genes from their genome-scale CRISPRi and CRISPRa screens in T cells. Schmidt et al. analyzed CD4+ and CD8+ T cells clustered by expression patterns and found that regulators of cytokine production tune T cell activation and program cells into different stimulation responsive states. Using these insights, they were able to nominate genes to enhance the efficacy of adoptive T cell therapies [50].

Until recently, single-cell pooled CRISPR screens had not been applied at a comparable scale to pooled screen with bulk readouts. Given the 100s of single cells needed to interrogate perturbations effects the sequencing depth required to detect significant effects, genome-scale scRNA-seq perturbation screens are limited primarily by their high cost. Replogle et al. [51] performed the first genome-scale CRISPR screens using single-cell readouts. They targeted all expressed genes in K562 (~10,000), and all essential genes shared between K562 and the retinal pigment epithelial cell line RPE1 (~2,000). In all, they sequenced ~2.5 million cells to achieve ~100 cells per perturbation, with 800–3000 reads per cell. This study identified functions for several poorly annotated genes (e.g. ribosome biogenesis, transcription, and mitochondrial respiration) by observing that their transcriptional profiles clustered tightly with genes with known functions. Given the enormous costs for library preparation and sequencing, it can be helpful to develop methods with similar perturbation scale but more focused readout: Targeted Perturb-seq (TAP-seq) represents one potential solution for overcoming sequencing depth requirements for single-cell studies (Table 1) [52]. In TAP-seq, the sequencing library is first enriched for transcripts of interest (e.g. 100 – 1000 specific transcripts), which provides a balance between multidimensional phenotyping and cost, as it can reduce reads-per-cell requirements by up to 50-fold.

Rather than directly targeting genes, single-cell pooled CRISPR screens can also target noncoding cis-regulatory elements (CREs) or enhancers. CREs typically regulate gene transcription by binding to enhancer elements and TFs. By interrogating CREs with single-cell screens, we can identify putative gene targets and understand epigenetic changes that precede gene regulation in development and disease. Xie et al. [38] developed Mosaic-seq, one of the first indirect gRNA capture methods for droplet-based approaches, by targeting gRNAs to enhancers, quantifying enhancer penetrance on target gene expression. Later studies by Xie et al. (2019) [53] and Gasperini et al. [54] used CRISPRi coupled with scRNA-seq and indirect gRNA capture to further identify hundreds of enhancer-gene pairs. These studies found that, while most enhancers target a single gene, some enhancers targeted multiple genes. Also, multiple enhancers can target the same gene and enhancer dosage is a major determinant of gene expression. Xie et al. (2019) explicitly investigated putative enhancers near TFs and found that enhancers for the same TF can modulate different sub-modules within a regulatory network. Therefore, by targeting enhancers, researchers can flexibly control entire regulatory networks. Recently, Morris et al. [55] applied CRISPRi with scRNA-seq and direct gRNA capture to inhibit noncoding loci identified from genome-wide association studies of blood traits. This approach (Systematic Targeted Inhibition of Noncoding GWAS loci, or STING-seq) starts by targeting common genetic variants in CREs and, using ECCITE-seq, can identify CRE target genes in cis and entire regulatory networks in trans. This study underscores an important utility for single-cell pooled CRISPR screens, where targeted perturbations can reveal dynamic regulatory networks for human complex traits and common diseases and help dissect noncoding loci with human genetic evidence.

The aforementioned studies were conducted in in vitro systems, which are often not representative of intact organisms. In 2020, Jin et al. [56] performed the first in vivo single-cell CRISPR screen. They knocked out 35 autism spectrum disorder (ASD) and neurodevelopmental delay (ND) genes in mouse embryos to study developmental processes of the early postnatal brain. ASD/ND genes are predicted to act in multiple brain cell types. The authors linked specific gene knockouts with perturbed regulatory networks, revealing novel functions of ASD/ND genes in specific types of neurons. A recent similar study uses a different viral vector (adeno-associated virus) to study a broader set of cell types and timepoints in 22q11.2 deletion syndrome [57]. Taken together, these studies demonstrate the power of single-cell pooled CRISPR screens to functionally dissect genes involved in disorders of mammalian brain development.

BEYOND THE TRANSCRIPTOME: CRISPR SCREENS WITH SINGLE-CELL PROTEOMICS AND GENOME ACCESSIBILITY

Although single-cell transcriptomics provides insight into regulatory networks, cell states, and differentiation, multimodal single-cell CRISPR pooled screens can enhance our understanding of the underlying biology by measuring protein or epigenetic modifications. For example, Cellular Indexing of Transcriptomes and Epitopes (CITE-seq) [58] combines protein measurements with droplet-based scRNA-seq (Fig. 3a). CITE-seq leverages single-cell barcoding to capture information encoded on the cell surface, and using antibodies tagged with an oligonucleotide sequence, marks specific samples with a unique barcode. This allows for pooling of multiple samples in a single assay and sequencing run, reducing potential batch effects and technical confounders. Recently, CITE-seq and similar approaches have been expanded in several directions. Wroblewska et al. [59] developed Pro-Codes, which use lentiviral vectors encoding epitope barcodes that can be expressed on the cell surface for cell indexing. McGinnis et al. [60] developed MULTI-seq, which uses lipid-tagged oligos, to increase sample multiplexing and single cell recovery. Mylka et al. [61] benchmarked CITE-seq and MULTI-seq, finding that both performed well for cell line, nuclei, and tissue staining, but given the experimental design of the study, one method may outperform the other under specific circumstances. Mimitou et al. [40] developed ECCITE-seq, which combines direct gRNA capture, whole transcriptomes, and cell surface protein quantification. More recently, a similar method called Perturb-CITE-seq was developed using Direct Perturb-seq gRNA capture sequences [62].

Given that CITE-seq was initially designed to target cell surface proteins, further innovation has been required to target intranuclear or intracellular proteins. For example, inCITE-seq [63], NEAT-seq [64], and INs-seq [65] are able to capture intranuclear proteins in single cells, and Phospho-seq [66] can capture phosphorylated intranuclear and intracellular proteins in single cells. While these methods have enabled measurement of protein levels at both the cell surface and within the cell, they do not yet include gRNA capture. Altogether, advances to the CITE-seq toolkit have enabled simultaneous measurement of the effects of CRISPR perturbations on the transcriptome and cell surface proteins, with intranuclear and intracellular proteins representing the next major advancements on the horizon.

Another key cellular modality is accessible (open) chromatin, which indicates functionally-active coding and noncoding genomic regions [67]. One such method for identifying open chromatin is ATAC-seq, which uses transposases to fragment and attach sequencing primers to accessible DNA [68]. ATAC-seq has been adapted for single-cell sequencing (scATAC-seq) [69], and combining CRISPR perturbations with CRISPR gRNA capture and scATAC-seq can reveal the impact of individual genes or noncoding regulatory elements on chromatin accessibility (Fig. 3). Perturb-ATAC [70] relies on physical isolation of single cells, whereas CRISPR-sciATAC [47] and Spear-ATAC [71] capitalize on more scalable methods (combinatorial indexing and microfluidic droplets, respectively) to analyze larger numbers of single cells, albeit with less sequencing depth (Table 1). These studies profiled chromatin accessibility in response to transcription factor perturbations in GM12878 (immortalized B lymphoblasts), primary human keratinocytes, and K562 (blood progenitors). Briefly, single-cell combinatorial indexing (sci) approaches rely on split-pooling cells for unique barcodes (Fig. 3b), often using easily obtained 96-well plates [72], and have been applied to a range of modalities, such as sciRNA-seq [73] and sciATAC-seq [74].

Liscovitch-Brauer et al. [47] developed CRISPR-sciATAC, which combines single-cell combinatorial indexing with CRISPR gRNA capture. In the first step of CRISPR-sciATAC, they perform barcoded tagmentation on open chromatin from nuclei and use reverse transcription to apply the same barcode to gRNAs, allowing for the recovery of both modalities in multi-well plates. After pooling and re-splitting, a second set of barcodes is applied via PCR using primers specific for each molecular species (ATAC or gRNA). Using this method, they profile the impact of knock-out of every chromatin modifier in the human genome on chromatin accessibility, mapping how each chromatin-modifying complex and individual subunit modulates chromatin accessibility of specific TF binding sites across the genome. For example, the loss of EZH2 increased accessibility of genes found in the HOX cluster, highlighting the key role played by EZH2, a histone methyltransferase, in human development.

Several other combinational indexing approaches combining CRISPR gRNA capture and additional modalities have been developed recently. For example, Xu et al. [75] developed PerturbSci-Kinetics, a pooled CRISPRi screening approach enabling the capture of CRISPR gRNAs, whole transcriptomes, and nascent transcriptomes. RNA kinetics, such as RNA synthesis and degradation, can be studied by labeling nascent transcripts in cells with a metabolic label, 4sU [76]. After inducing C>T conversions in 4sU-labelled transcripts via thiol(SH)-alkylation [77], the difference in C>T conversions between nascent transcripts and unlabeled transcripts enable per-transcript kinetic estimates. PerturbSci-Kinetics was used to identify a novel role for AGO2, a known post-transcriptional silencer functioning in RNA interference [78], as a regulator of RNA synthesis as well [75].

Single-cell pooled CRISPR screens that combine all of the aforementioned modalities – RNA, protein, chromatin accessibility, and CRISPR gRNA capture – do not currently exist, but recent advances suggest they are possible. Mimitou et al. (2021) [79] developed DOGMA-seq, a variant of CITE-seq, which simultaneously profiles RNA, cell surface proteins, and chromatin accessibility in single cells. Given that other groups have interrogated these modalities individually alongside CRISPR gRNA capture, it should be possible to combine them in a single assay — an exciting area for future technology development.

EMERGING TECHNOLOGIES FOR CRISPR SCREENS WITH MULTIMODAL READOUTS

Most single-cell CRISPR screens are performed with short-read sequencing, which captures either the 5’ or 3’ transcript end; however, emerging studies leverage long-read sequencing to capture entire RNA molecules for a more informative output. Long-read sequencing enables the detection of transcript isoforms, which can result from perturbations. Yet, none of these long-read sequencing technologies have been combined with CRISPR gRNA capture. Instead, single-cell arrayed CRISPR screens implement long-read sequencing in cells where the CRISPR gRNA identity is already known [80]. While powerful for examining the effects of specific targets of interest, arrayed screens are low-throughput given the requirement to perform a CRISPR screen for each individual target. Future studies should aim to combine long-read sequencing with single-cell pooled CRISPR screens and indirect or direct gRNA capture to enable scalability.

For the genome, the 3D structure and folding of DNA is crucial to understand gene regulatory compartments and which promoters might be regulated by specific distal enhancers [81–83]. Chromosome conformation capture (3C) and related sequencing-based methods like Hi-C capture DNA fragments which physically interact via cross-linking, restriction digest and then ligation [84,85]. The earliest single-cell study adapting 3C/Hi-C technology were relatively low-throughput by current standards, as they required manual isolation of cells [86,87]. The development of combinatorial indexing-based sci-Hi-C approaches enabled the simultaneous profiling of thousands of single cells [88,89]. Recently, Liu et al. [90] developed a multimodal combinatorial indexing-based approach, termed Hi-C and RNA-seq Employed Simultaneously (HiRES), to simultaneously capture Hi-C and RNA-seq in single nuclei from thousands of cells. Single-cell Hi-C approaches continue to advance and incorporate additional modalities, but they do not yet incorporate CRISPR gRNA capture, restricting the scale of perturbation studies and their consequences for 3D chromosome interactions. For example, Guo et al. [91] performed arrayed CRISPR perturbations of CTCF binding sites with 3D conformation capture, finding that inverting the orientation of these binding sites can reconfigure chromatin looping and alter enhancer-gene interactions. Adding 3C/Hi-C as a modality in single-cell pooled CRISPR screens promises to functionally profile TF binding domains, 3D compartmentalization junctions, and pathogenic variants.

While not strictly a single-cell method, Repair-seq quantifies double-stranded break (DSB) outcomes using an innovative approach with UMIs and bulk sequencing [92]. By introducing a gRNA together with a CRISPR target site and flanking restriction enzyme cut sites, the sites of gRNA and target region integration can be extracted from genomic DNA and tagged with a UMI prior to PCR. Deep sequencing then allows for the assignment of gRNAs to specific DSB repair outcome in a single cell labeled by a single UMI, and enables the study of genetic modulators of DNA repair [92].

Imaging-based approaches can be used to select cells with distinct cellular and subcellular features, such as nuclear size, enabling CRISPR screens for genes that regulate cell morphology, cellular dynamics, or cell-to-cell interactions [93–95]. For example, Feldman et al. [93] developed an optical pooled CRISPR screening approach where cells are lentivirally transduced with gRNAs that include 12-nt barcodes. Upon selecting cells by physical features under a microscope, single cells can then be matched to their perturbation by in situ sequencing of the 12-nt barcode. Briefly, dyes are annealed, imaged, and cleaved to identify nucleotides individually (in situ sequencing-by-synthesis). Feldman et al. [93] examined p65 translocation dynamics upon knocking out 952 genes involved in NF-kB signalling, finding a new role for Mediator complex subunits. Another imaging-based pooled screen is BARC-FISH (Barcode Amplification by Rolling Circle and Fluorescence In Situ Hybridization), which uses FISH to image pooled, barcoded CRISPR gRNAs in single cells coupled with DNA FISH chromatin tracing to examine 3D chromatin changes [96]. BARC-FISH was used to identify several new modifiers of 3D genome organization, including CHD7, a gene where de novo mutations lead to specific birth defects and CHARGE syndrome.

Recent single-cell CRISPR screens with multimodal readouts have expanded beyond targeting DNA by capitalizing on the tremendous metagenomic diversity of CRISPR systems. For example, the class II type VI CRISPR-Cas13 family can be used to directly target RNA for degradation [19,97,98] or base editing [99] (Fig. 1). Wessels et al. [100,101] and Méndez-Mancilla et al. [102] determined Cas13 RNA-targeting rules, optimizing Cas13 gRNA sequence and chemical modifications, respectively. Compared to DNA-targeting CRISPRs, the Cas13d nuclease has several unique advantages: It does not require a specific protospacer adjacent motif (PAM) sequence (e.g., ‘NGG’ for Cas9) for target recognition and acts in a strand-specific manner [97], allowing Cas13d to target any sequence in the transcriptome. To combine Cas13d with droplet-based single-cell methods, Wessels et al. developed CaRPool-seq for Cas13d-based single-cell pooled RNA-targeting CRISPR screens (Table 1) [103]. Since Cas13d can cleave gRNA arrays into independent perturbations, this enables combinatorial perturbations with multimodal single-cell profiling. Arrays were cloned with up to three distinct gRNAs and a single barcode, not only resulting in separate targets encoded in the same array, but enabling consistent and efficient gRNA capture. The authors found that CaRPool-seq outperformed the Cas9-based direct-capture Perturb-seq with respect to combinatorial perturbation detection. CaRPool-seq uses a single barcode for up to three gRNAs, and in cells with a barcode detected there was a 99% concordance rate of gRNAs assigned and the barcode label. Direct-capture Perturb-seq with dual-gRNA delivery requires the detection of individual gRNAs per pair, and in cells with at least one gRNA detected there was a 67% detection rate for the expected paired gRNA. Given these differences in capture efficiency, Cas13-based screens may be a preferred approach for combinatorial gene perturbations over Cas9-based screens.

Other notable single-cell pooled screens include other, non-CRISPR-based genetic perturbations. For example, Legut et al. [104] developed OverCITE-seq, building upon the existing CITE-seq toolkit by enabling the direct capture of lentivirally-delivered gene overexpression via open reading frames (ORFs). Rather than using CRISPRa to constitutively express genes within a cell, OverCITE-seq lentivirally delivers ORFs driven by a constitutive promoter (CMV) to express genes directly (Table 1). In addition to driving higher levels of gene expression that CRISPRa, a major benefit of OverCITE-seq is the reduced size of the viral payload, making it easier to transduce cells that have been traditionally difficult to study, such as primary human T cells. Demonstrating OverCITE-seq’s utility, Legut et al. identified LTBR, a gene not canonically expressed in CD4⁺ and CD8⁺ T cells, as a driver for T cell effector functions, making them resistant to exhaustion in chronic stimulation settings. Using OverCITE-seq, Legut et al. also captured single-cell TCR clonotypes, enabling them to show that modifier genes like LTBR were truly drivers of the phenotype, rather than a clonal effect due to lentiviral integration site (e.g. due to disruption of the gene at the integration site [105]). Similarly, orthogonal perturbation methods can be combined with CRISPR-based approaches. For example, Li et al. [106] developed a combinatorial indexing-based approach to capture lentivirally delivered short-hairpin RNAs (shRNAs) for gene knockdown flanked by a T7 promoter and a reverse transcription primer binding site to study modifiers of CRISPR prime editing efficiency. This dual-perturbation strategy enabled the identification of trans-acting regulators of prime editing efficiency; for example, the authors found that HLTF inhibition improves prime editing efficiency [106].

ANALYSIS METHODS AND TOOLS FOR SINGLE-CELL POOLED SCREENS

The analysis of pooled CRISPR screens is a developing field, with new methodologies to tackle the challenges inherent in distilling key phenotypes from high-dimensional, noisy measurements of single cells. The analysis workflow for single-cell pooled CRISPR screens is conceptually straightforward: 1) Sequence CRISPR gRNA libraries, these can be either the direct sequence or their barcodes, which can be referred to as feature barcodes; 2) Align sequencing data to a reference of gRNA sequences or feature barcodes; 3) Perform quality control, such as only keeping those cells with a single gRNA detected and with a minimum number of gRNA reads; 4) Differential expression (or other modality of interest) analysis (Fig. 4).

Figure 4. Analyzing single-cell pooled perturbation screens.

Single-cell perturbations (e.g. guide RNA [gRNA] or barcoded overexpression libraries) can be represented with a n x m matrix, whereby n is the number of single cells and m is the number of perturbations, and the matrix is populated with the number of UMIs per perturbation per cell. After defining UMI thresholds and assigning perturbations to cells for follow-up analyses, changes in different modalities of interest (e.g., open chromatin or gene expression changes) are measured. First, non-targeting (negative) controls should be inspected for inflation of observed test statistics. This can be performed by comparing observed significance p-values with p-values that would be expected by random chance given the number of tests performed and examining results for systematic deviations from the null expectation. Upon verifying that there is no systematic inflation, targeting gRNAs are examined for their effects on measured molecular phenotypes (e.g. gene expression changes in cis and trans or open chromatin).

There are multiple decisions to be made at each step, for example, the gRNA vector and single-cell method will determine whether gRNAs are captured directly and the specific capture strategy. After sequencing, there are multiple tools that can be used for alignments, each with their own strengths and weaknesses. The goal is to use a list of gRNA sequences or barcodes as a reference genome for mapping sequenced reads and generating UMI count matrices, of UMIs per gRNA per cell. 10X Genomics provides a proprietary software suite to perform read alignments and generate unique molecule identifier (UMI) count matrices for the transcriptome and additional modalities, Cell Ranger [43]. Cell Ranger uses the STAR aligner [107] for antibody or gRNA capture analysis, allowing up to one mismatch in sequence alignments. CITE-seq-Count, originally designed for processing CITE-seq antibody capture data, can be easily adapted to gRNA sequencing reads and allows for the user can specify the number of maximum mismatches allowed. Alternatively, kallisto-bustools [108] and alevin-fry [109], which can perform computationally rapid and low-memory scRNA-seq read alignment via pseudoalignment [110], can be repurposed for gRNA read alignment and generating UMI count matrices. Once UMI count matrices have been generated, quality control checks should be performed for determining UMI thresholds per gRNA per cell. Cell Ranger will perform a secondary analysis upon generating UMI count matrices, where for each gRNA, it will calculate its UMI threshold to call cells bearing that gRNA using a Gaussian mixture model. Barry et al. [111] developed a model for determining gRNA UMI thresholds, GLM-based errors in variables (GLM-EIV), and observed through simulation studies and applications to real data, that in the absence of modeling a UMI threshold, a minimum of 3 UMIs can be used to reliably call a cell bearing a gRNA.

Lastly, after generating quality controlled UMI count matrices, and assigning gRNAs to cells, pairwise differential tests can be performed. For low multiplicity-of-infection (MOI) experiments, where most cells express a single gRNA, differential expression testing generally involves dividing cells into a non-perturbed set (e.g., cells bearing non-targeting control gRNAs) and a perturbed set (e.g., cells targeting a specific gene), and performing differential expression tests [36,112,113] on a selected gene for differences between the sets. Given that CRISPR knockout may have incomplete penetrance, methods are also available for identifying populations of cells bearing gRNAs that remained unperturbed, such as Mixscape [114] or MELD [115]. For high MOI experiments, where most cells express multiple gRNAs, these approaches can suffer from test statistic inflation [116]. For these cases, SCEPTRE [116,117] or Normalisr [118] approaches outperform other methods that are not explicitly designed for high MOI experiments [53,54,112]. SCEPTRE couples a conditional resampling test with a negative binomial approach for differential expression testing, whereas Normalisr uses a linear regression approach. Regardless of the differential expression testing method, a good practice is to first examine negative controls and verify a null distribution of test statistics (e.g., examining p-values on the -log₁₀ scale with a quantile-quantile plot). This demonstrates that the results do not suffer from model miscalibration or systematic test statistic inflation or deflation (Fig. 4). It is important to verify that non-targeting gRNAs are not systematically inflated or deflated compared to a null distribution; otherwise, significant observations may be due to a high false positive rate. After these steps, perturbations can then be confidently tested for their effects on gene expression in cis and in trans, or for other single-cell modalities such as open chromatin (Fig. 4).

CONCLUSIONS AND PERSPECTIVES

Over the past decade since we and others developed the first CRISPR pooled screens [5,30], there has been a widespread adoption of these methods throughout academia and biopharma for genetic discovery on a genome-wide scale. Similarly, we expect increased adoption of single-cell CRISPR screens over the next 5 – 10 years and the development of new technologies that will drive higher throughput, lower cost, and simplified experimental and analytic workflows. For example, current droplet-based approaches, while allowing for more modalities than combinatorial indexing approaches, are limited by the number of single cells that can be generated in a single assay. Datlinger et al. [119] developed single-cell combinatorial fluidic indexing RNA-seq (scifi-RNA-seq), by performing combinatorial indexing barcoding prior to droplet-based barcoding, generating 100-fold more single cells than standard droplet-based scRNA-seq approaches. The field is rapidly evolving to a place where we will no longer be limited by the number of single cells we can generate. Future hurdles will be the modalities we can capture within the same cells and the depth to which we can sequence them (see Outstanding Questions).

The next few years will likely see multiple genome-scale and combinatorial single-cell CRISPR screens, yielding new insights into gene function and a deeper understanding of how combinations of genes function and interact at-scale. We expect these genome-scale datasets to include millions of single cells, and to include not only transcriptome and CRISPR gRNA capture modalities, but also cell surface and intracellular proteins, chromatin accessibility, DNA and RNA modifications, histone post-translational modifications, TF activity and binding, and chromosome conformation. For example, recent advances in applying nanobody-tethered transposases can allow for simultaneous profiling of multiple chromatin states within a single cell [120].

Similarly, on the perturbation side, new CRISPR platforms will broaden the set of genome elements that can be reliably perturbed. For example, RNA-targeting Cas13 can be combined with single-cell pooled screens to profile targeted knockdowns of specific transcript isoforms or RNAs that do not code for proteins (e.g. long noncoding RNAs, enhancer RNAs, micro RNAs, etc.). All of these applications depend, crucially, on sequencing: As the number of single cells increases in an experiment, the amount of sequencing required scales linearly (e.g., if 10,000 reads are required per cell, each additional cell requires an additional 10,000 reads) until reaching saturation. The last few years have seen many exciting developments to reduce sequencing cost with the latest advances in short-read sequencing chemistry (e.g., the Ultima Genomics $100 genome [121]), and increased throughput and accuracy of long-read sequencing (e.g., Oxford Nanopore’s low-cost devices [122]) (see Outstanding Questions). Taken together, these innovations in multimodal capture, new perturbation capabilities, sequencing technologies and analytic frameworks will enable new kinds of cutting-edge pooled single-cell screens, bringing us closer to understanding the functions of all genes and noncoding elements in the genome [123].