Technologies and Computational Analysis Strategies for CRISPR Applications

Abstract

The CRISPR-Cas system offers a programmable platform for eukaryotic genome and epigenome editing. The ability to perform targeted genetic and epigenetic perturbations enables researchers to investigate questions in basic biology and potentially develop novel therapeutics for the treatment of disease. While CRISPR systems have been engineered to target DNA and RNA with increased precision, efficiency, and flexibility, assays to identify off-target editing are becoming more comprehensive and sensitive. Furthermore, techniques to perform high-throughput genome and epigenome editing can be paired with a variety of readouts and are uncovering important cellular functions and mechanisms. These technological advances drive and are driven by accompanying computational approaches. Here, we briefly present available CRISPR technologies and review key computational advances and considerations for various CRISPR applications. In particular, we focus on the analysis of on- and off-target editing and CRISPR pooled screen data.

The CRISPR-Cas system has accelerated the development of tools for genome and epigenome editing. Since the initial adaptation of Streptococcus pyogenes Cas9 (SpCas9) for eukaryotic genome editing, the CRISPR toolbox has rapidly expanded to include new CRISPR-Cas orthologs and variants with various target sequence requirements, fidelities, and perturbation mechanisms (Jiang and Doudna, 2017). CRISPR-Cas systems have also been repurposed in various ways for gene activation (CRISPR activation [CRISPRa]), gene repression (CRISPR interference [CRISPRi]), epigenome editing via fusion to epigenetic modifiers (Thakore et al., 2016), and DNA sequence alteration in the absence of a double strand break (DSB) (base editing and prime editing) (Anzalone et al., 2019; Rees and Liu, 2018).

Characterizing and quantifying products of genome editing is essential for the development of new tools and for bridging the knowledge gap between genome sequence and function. Biochemical assays for measuring editing events with simple—sometimes binary—readouts are being replaced by next-generation sequencing (NGS) approaches that improve accuracy and sensitivity, while also providing a more comprehensive view of genome editing outcomes. However, data from NGS-based experiments require several steps of downstream processing, which has led to the development of various computational tools for analysis.

In addition, it is well established that CRISPR-Cas editing may also occur at unintended genomic loci, also known as off-target sites. Locating and quantifying editing at these off-target sites is critical for interpretation of genome editing experiments as well as for assessing the safety of therapeutic genome editing programs. In silico methods for predicting off-targets have evolved from the simple enumeration of sites based on on-target sequence similarity to more advanced tools that incorporate common or personal genomic variants. At the same time, experimental techniques have also become more sophisticated, leveraging downstream computational analysis to identify sites of off-target editing with increased sensitivity.

The programmability of CRISPR-Cas systems has enabled the association of genomic perturbations to phenotypes at scale. High-throughput CRISPR perturbations can be performed on many cells in parallel using a pool of guide RNAs (gRNAs) targeting many genes or tiled across a region of interest. The perturbation response in each cell can be measured using a variety of readouts. This approach has enabled the dissection of critical functional elements within a region of interest and the identification of critical genes and gene networks associated with a phenotype of interest. An enormous amount of data is generated with each screen, and computational analysis methods have been developed to efficiently and properly interpret the experimental results.

In this review, we outline biological questions that can be investigated using current tools from the CRISPR-Cas toolbox and provide computational perspectives into the advantages, disadvantages, and analytical challenges of these technologies. Table 1 summarizes these technologies, associated computational methods, and key references with example applications and tools to help address these questions.

Table 1.

Summary of Key Technologies and Computational Strategies

Purpose	Approach	Technology	Comments	Examples
Quantify/validate editing at a region	Mismatch assays	Heteroduplex cleavage assays A locus of interest is amplified from genomic DNA using PCR, followed by denaturation and annealing to form heteroduplex complexes between wildtype (non-edited) and nuclease-modified DNA strands. DNA mismatch enzymes cleave sites of sequence mismatch. The editing proportion is indicated by measuring the proportion of cleaved products.	• Poor ability to reliably detect low-frequency mutations • Heterogeneous sensitivities for different variant types (e.g., limited detection of single nucleotide polymorphisms [SNPs]) • Under-estimation of editing frequencies including inability to detect high editing frequencies with an editing frequency of 30–40% possibly representing an upper limit that can be reliably detected (Mashal et al., 1995) • False positives in the setting of heterozygous germline mutations (Qiu et al., 2004), which can be falsely attributed to CRISPR-mediated mutagenesis if proper controls are not performed	T7E1 (Mashal et al., 1995), Surveyor (Qiu et al., 2004)
	Amplicon sequencing	Conventional chain termination (Sanger) sequencing Amplified DNA is sequenced using Conventional chain termination (Sanger) sequencing, producing a quantification of the proportion of each base at each position.	• Inexpensive sequencing costs • Useful for screening clones or homogeneous populations • Deconvolution of alleles in heterogeneous populations may be inaccurate	Brinkman et al., 2014, 2018; Hsiau et al., 2019
		Next-generation sequencing Amplified DNA is sequenced using next-generation sequencing, producing the DNA sequence of alleles from different cells	• Deep coverage at known loci, with the ability to detect and quantify rare editing events • Precise quantification of alleles and editing rates on heterogenous populations • Moderate sequencing costs	Boel et al., 2016; Clement et al., 2019; Güell et al., 2014; Hwang et al., 2018; Lindsay et al., 2016; Park et al., 2017; Pinello et al., 2016
Measure (single) guide specificity	In silico	Sequence homology to reference genome A reference genome sequence is scanned for similarity to a given guide sequence. Similarity is measured by the number of mismatches or gaps between the guide sequence and the match in the reference genome.	• Quickly determine number and location of putative off-targets without any experimental labor • Useful in designing guides or selecting guides with few off-targets • Cellular conditions may affect CRISPR activity to prefer or avoid sites with few mismatches to the guide • Genomic sequence in target cells may vary from reference sequences	Bae et al., 2014; Cancellieri et al., 2019; Haeussler et al., 2016; Heigwer et al., 2014; Klein et al., 2018; Lei et al., 2014; McKenna and Shendure, 2018; Montague et al., 2014; Xiao et al., 2014
		Machine learning approaches Machine learning models are trained on datasets of genome editing outcomes to predict the activity of untested guides or the activity of guides at untested regions.	• Predicts activity rates at sites with sequence homology • Models are based on experimental data and may be biased based on cell type, guide sequence, or other variables	Abadi et al., 2017; Allen et al., 2018; Lin and Wong, 2018; Listgarten et al., 2018; Peng et al., 2018
	Experimental	In vitro Genomic DNA is tested for the ability to be cleaved by a guide, outside of the cellular context, and not in the presence of native DSB repair machinery.	• Edited genomic fragments can be selected, resulting in reduced sequencing costs • Double-stranded break repair mechanisms are not present in the in vitro context, so adaptor ligation or other biochemical tagging can be performed • Sensitivity of method may overestimate editing in cellular context • Some methods allow the detection of off-targets for individuals/samples for which limited genomic material is available	SITE-seq (Cameron et al., 2017), Digenome-seq (Kim et al., 2015, 2016), CIRCLE-seq (Tsai et al., 2017)
		In cellula The occurrence of genome editing in cellular context is observed by observing events associated with DSB repair, or the introduction of molecular tags to edited regions.	• Epigenetic context of the off-target is taken into consideration • Many cell types do not tolerate genomic manipulation required for readout	Cas-ChIP (Duan et al., 2014; Kuscu et al., 2014; Wu et al., 2014), DISCOVER-seq (Wienert et al., 2019), IDLV insertion at DSB (Wang et al., 2015b), GUIDE-seq (Tsai et al., 2015), iGUIDE (Nobles et al., 2019), BLESS (Crosetto et al., 2013), BLISS (Yan et al., 2017)
Characterize gene function using a single perturbation	Single gene knockout	Single guide edit resulting in frameshift mutation or early stop codon A single guide is designed to target Cas9 or another editing enzyme to an early exon in the gene. The frameshift mutation resulting from the edit renders the gene non-functional.	• Targeting region is well defined by exon boundaries • Genomic sequencing can validate success of edit • Edit knocks out genes and is propagated across cell divisions • Variety of CRISPR editing nucleases and editors can be used—Cas9, Cas12, BE, PE • Early exons may be skipped in edited cells (Mou et al., 2017) • Allelic diversity of edits may affect phenotype—can be overcome by growing and selecting edited clones	Cong et al., 2013
	Non-coding perturbation	CRISPRi/a targeting promoter A single guide targets deactivated Cas9 to the promoter of a gene of interest. The deactivated Cas9 is tethered to an activating or inhibiting enzyme which activates or inhibits transcription of the gene.	• Target is defined by a region of putative regulatory activity, e.g., an enhancer identified with epigenetic marks • Perturbations are transient (can be overcome by genomic integration of CRISPR system) • Requires annotated (single) promoter • Results in overall reduction in number of transcripts of the target gene • qPCR or RNA-seq (or ChIP of CRISPR binding at target) to validate changes in expression of target gene • Targeting boundaries are less well defined (as opposed to exon boundaries in genetic targeting)	Gilbert et al., 2013; Qi et al., 2013
		Disruption of binding site sequence A single guide targets Cas9 or another editing enzyme to a binding site that controls expression of a gene. After gene editing, the binding site is disrupted, and the gene is not transcribed.	• Target sequence must be known (e.g., ChIP-seq or TF models) and well-described	Zhou et al., 2014a
	Deletion of large element of DNA	Editing with a flanking pair of guides A pair of guides target a cleaving enzyme to the flanking sites of a region of interest. After cleavage, the region of interest is excluded upon DSB repair, resulting in a large deletion.	• Genetic element is completely removed or disrupted • Inversions, NHEJ, or other complex outcomes are likely	Chen et al., 2014; Mali et al., 2013
Discover putative genes associated with a phenotype	Pooled genome-wide genetic screen	Phenotype readout Many guides are designed to target genes across the genome. Cells are sorted by phenotype, and the guides producing high versus low phenotype are compared.	• Measure selectable phenotypes • Potentially sequence alleles resulting in strongest phenotype	Allen et al., 2019; Hart and Moffat, 2016; Li et al., 2014; Lindsay et al., 2016; Love et al., 2014; Shalem et al., 2014; Spahn et al., 2017; Wang et al., 2014, 2015a; Yu et al., 2016; Zhou et al., 2014b)
		Single-cell RNA-seq readout Many guides are designed to target genes across the genome. Cells are treated with one guide per cell, and the transcription levels of all genes are read out using single-cell RNA-sequencing.	• Granular results at single-gene expression level • Measure effect on genes or sets of genes • Readout of heterogeneity of response • Identify gene networks with similar response profiles • Limited number of cells profiled (per perturbation) • Requires high analysis resources	Adamson et al., 2016; Datlinger et al., 2017; Dixit et al., 2016; Duan et al., 2019; Jaitin et al., 2016; Xie et al., 2017; Yang et al., 2020
	Pooled genome-wide epigenetic screen	CRISPRi/a targeting of gene promoters Many guides are designed to target genes across the genome with deactivated Cas and an enhancer/inhibitor. Phenotypic or transcriptomic changes associated with activation/repression of genes are measured.	• CRISPRi/a must be targeted precisely to achieve consistent gene knockout, as compared to genetic mutations which can disable translation with a single mutation	Gilbert et al., 2014; Konermann et al., 2015
Discover critical elements within an annotated gene/enhancer	Tiling screen profiling a single locus	Genetic screen of coding region Many guides are designed to target a region of the genome with coding significance. Cas9 or another editing enzyme create genetic modifications to the coding region, altering the physical properties of the resulting protein.	• Can be used to discover functional domains of a protein	He et al., 2019; Neggers et al., 2018; Shi et al., 2015
		Genetic screen of noncoding region Many guides are designed to target a region of the genome with putative regulatory significance (e.g., enhancer). DNA edits by Cas9 or another editing enzyme affect the regulatory properties near the guide. Guides associated with a phenotype change indicate specific locations critical for function.	• Effect of single-base-pair changes can be measured • Critical regions may be missed if not near perturbation window of guide • Low probability of creating activating mutations	Canver et al., 2015; Sanjana et al., 2016
		Epigenetic screen of noncoding region Many guides are designed to target a region of the genome with putative regulatory significance. Epigenetic effectors are targeted using these guides, and guides associated with phenotype change indicate specific locations critical for function.	• Larger regions can be interrogated using fewer guides • Can be used to identify silenced enhancers	Fulco et al., 2016

CRISPR-Cas Perturbation Technologies

The CRISPR-Cas system is a versatile perturbation platform that can introduce different forms of genetic and epigenetic modifications. CRISPR-Cas enzymes were first discovered as components of an adaptive immunity system used by bacteria and archaea to counter foreign DNA and have been repurposed and engineered to perform genome editing in other organisms. The CRISPR-Cas-induced DSBs are repaired using endogenous cellular machinery involved in the non-homologous end joining (NHEJ), microhomology-mediated end joining (MMEJ), and homology-directed repair (HDR) pathways. Repair by NHEJ or MMEJ typically results in a spectrum of variable-length insertion or deletion (indel) mutations (Allen et al., 2018; Shen et al., 2018) introduced with relatively high frequencies. The induction of indels is particularly useful in introducing frameshift mutations or disrupting non-coding regulatory elements. By contrast, exogenous DNA templates can be provided to take advantage of HDR pathways to introduce precise genomic edits, albeit at lower efficiencies (though recent studies suggest that efficiency may be enhanced with small molecules [Riesenberg and Maricic, 2018; Song et al., 2016; Vartak and Raghavan, 2015]).

Recent technological advancements have led to newer forms of CRISPR-based genetic editors that modify DNA sequences without the requirement for DSBs. Base editors (BEs), a fusion of nickase Cas9 and a cytidine or adenosine deaminase enzyme, introduce targeted substitution mutations within a defined editing window at target DNA sequences (Rees and Liu, 2018). Prime editors (PEs), a fusion of nickase Cas9 and a reverse transcriptase, can be used with an associated RNA template contained on the prime editing guide RNA (pegRNA) to introduce a wide variety of genetic modifications ranging from single-base substitutions to short indel mutations (Anzalone et al., 2019). BEs and PEs are useful tools for sequence mutagenesis because they can introduce precise mutations and differ from CRISPR-Cas nucleases in that they significantly reduce the frequency of repair outcomes containing unwanted insertion or deletion mutations.

CRISPRa and CRISPRi technologies can be used to induce robust gene activation and repression, respectively (Gilbert et al., 2013; Kampmann, 2018; Qi et al., 2013; Thakore et al., 2016). For CRISPRa/i, the CRISPR-Cas enzyme is catalytically disabled and utilized to target effector domains such as transcriptional activators (e.g., p65, VPR), repressors (e.g., KRAB), or other epigenetic modifiers (e.g., DNMT3A) (Qi et al., 2013) to a specific loci and has also been used to block transcriptional initiation or elongation through steric hindrance. CRISPRa and CRISPRi have been useful in studying the effects of gene activation and knockdown in single gene or pooled settings (Adli, 2018; Thakore et al., 2016).

Characterization of DNA Editing at Defined Loci

A variety of methods exist to quantify on- and off-target genome editing at individual loci (Figure 1 and Table 1). These methods include sequencing and non-sequencing-based techniques. In general, non-sequencing-based assays offer cost-effective and rapid solutions to semiquantitatively detect the presence of nuclease-mediated sequence modification, while sequencing-based methods offer the ability to accurately quantify editing frequencies and define mutation alleles induced by genome editing.

Figure 1. Overview of Strategies to Detect Editing at Known Loci Including Heteroduplex DNA Analysis, Loss of Binding Site Analysis, and Sequencing-Based Approaches.

Heteroduplex DNA (left panel) is formed by denaturing and annealing of PCR amplicons generated from a bulk population of edited cells. The mismatches in the heteroduplex are detected and cleaved by enzymes such as T7E1 or Surveyor. Loss of binding site analysis (middle panel) relies on the ability of a PCR primer (green DNA sequence) to bind based on Watson-Crick complementarity or a transcription factor or restriction enzyme to identify its recognition sequence. Editing can be identified by binding site modification that results in the loss of PCR primer binding or loss of restriction enzyme mediated cleavage. In sequencing-based assays (right panel), Sanger sequencing or Next Generation Sequencing (NGS) are used to analyze a given site(s).

Assessment of Editing at Defined Loci Using Non- Sequencing-Based Methods

Mismatch cleavage and heteroduplex mobility assays are examples of non-sequencing-based techniques for indel detection and rely on similar principles to detect editing frequencies within bulk-edited cell populations. Specifically, a locus of interest (e.g., an on-target or predicted off-target site) is amplified from genomic DNA using polymerase chain reaction (PCR), followed by denaturation and annealing steps to form heteroduplex complexes between wild-type (non-edited) and nuclease-edited DNA strands (Vouillot et al., 2015; Zhu et al., 2014). This heteroduplex DNA can be analyzed in several ways: heteroduplexes can be treated with DNA mismatch endonucleases, such as Surveyor or T7E1, that cleave at sites of sequence mismatch. The sizes and relative intensities of resulting cleavage products can be quantified by either gel or capillary electrophoresis, and these values can be used to estimate editing frequencies (Mashal et al., 1995; Qiu et al., 2004; Vouillot et al., 2015). Alternatively, heteroduplex DNA products can be resolved by polyacrylamide gel electrophoresis (PAGE), which relies on heteroduplex (edited) DNA running with slower mobility than homoduplex (non-edited) DNA (Yu et al., 2014; Zhu et al., 2014). Another analysis strategy is high-resolution melting analysis, which relies on differences in melting temperature (T_m) between homoduplex and heteroduplex DNA to identify nuclease-induced mutations (Thomas et al., 2014). Importantly, these non-sequencing-based methods resolve single nucleotide substitutions poorly and therefore have minimal utility for base and prime editing experiments yielding substitution edits.

Other approaches to detect genome editing include methods that measure disruption of a PCR primer binding site (Yu et al., 2014) or restriction endonuclease sites (Kim et al., 2014). Insertions and deletions as small as 1 base pair (bp) can be detected by fluorescent capillary electrophoresis (Cho et al., 2014; Ramlee et al., 2015; Yang et al., 2015), but these methods are unable to detect substitution edits (Yang et al., 2015). In addition, digital droplet PCR (ddPCR) offers a quantitative method to evaluate NHEJ/MMEJ and HDR outcomes. One strategy using ddPCR relies on the usage of two fluorescent probes with one probe at the predicted cleavage site and the other at a distant site. This and other ddPCR-based assays offer precise quantification of editing outcomes including the ability to distinguish mono-allelic and bi-allelic modifications. Of note, ddPCR cannot resolve sequence information from NHEJ- and MMEJ-mediated out- comes (Findlay et al., 2016).

The methods described above are predominantly focused on directly detecting DNA sequence alterations. However, surrogate outcomes can also be evaluated as functional readouts. For example, changes in gene expression or protein abundance can indicate that a genomic change has occurred to knock out a gene or otherwise influence the production of the protein of interest. While many approaches exist, quantitative reverse-transcription PCR (RT-qPCR) and western blotting are common methods for detection of RNA and protein expression changes, respectively. Surrogate outcomes are particularly useful in the context of CRISPR-based screens, which seek to identify candidate functional elements or regions through changes in gene expression, protein abundance, or cell viability.

Assessment of Editing at Defined Loci Using Sequencing-Based Methods

Analysis of genome editing outcomes based on sequencing can take several forms. First, Sanger sequencing trace decomposition can be performed by tools such as Tracking Indels by Decomposition (TIDE) or Inference of CRISPR Edits (ICE) to calculate editing and allelic frequencies from bulk edited cells. These tools can be used for the quantification of both NHEJ/ MMEJ and HDR outcomes and are able to infer individual allelic frequencies (Brinkman et al., 2014, 2018; Hsiau et al., 2019). Sanger trace decomposition offers a rapid and low-cost methodology to assess editing outcomes and is particularly useful in screening clones of edited cells. Additionally, PCR amplicons for a given locus can be cloned into a plasmid backbone (e.g., TA, TOPO, or blunt end cloning), transformed into bacteria, and sequenced by Sanger methods to identify specific alleles (Canver et al., 2014).

Next-generation sequencing (NGS) represents the gold standard for both determination of editing frequency and characterization of the resulting alleles. A recent comparison of T7E1, Indel Detection by Amplicon Analysis (IDAA) (Yang et al., 2015), and TIDE-based Sanger decomposition to NGS revealed that T7E1 exhibited poor performance when compared to NGS. In contrast, both TIDE and IDAA offered comparable results to NGS for indel identification and indel frequency calculation (assuming indels with a frequency of >5%); however, both TIDE and IDAA approaches overestimated the presence of wild-type alleles and were less sensitive compared to NGS (Sentmanat et al., 2018). Depending on access to NGS technology, NGS assays may be limited by high cost, laborious preparation, and time delay to the data. However, as sequencing costs continue to decline, the analysis of genetic editing by NGS is becoming increasingly common, facilitating the development of multiplexed NGS readouts and more sensitive assays for the detection of rare alleles.

Several computational tools have been developed to analyze NGS data from genetic editing experiments with CRISPR-Cas nucleases (Boel et al., 2016; Güell et al., 2014; Lindsay et al., 2016; Park et al., 2017; Pinello et al., 2016) and base editors (Clement et al., 2019; Hwang et al., 2018). A variety of alignment and analysis methods are used to separate sequencing errors from true genome edits. One example is the use of an “editing window,” whereby only mutations overlapping the predicted target of activity are quantified, and mutations distal to the predicted target of activity are ignored (Clement et al., 2019; Hwang et al., 2018). This editing window approach on simulated sequencing data demonstrated a significant reduction in the quantification of false-positive mutations (e.g., sequencing error at the ends of reads) (Pinello et al., 2016).

Technology Outlook

Current efforts to quantify on-target editing events are focused on expanding the ability to measure all types of genome editing events. For example, existing methods can measure short insertions and deletions, but large insertions or deletions (>100 bp) or translocations are not detectable with standard amplicon sequencing approaches. For example, these editing outcomes may not be discovered if one or both primer binding sites are deleted by a large deletion or if a large insertion occurs, since it may not be efficiently amplified (Cullot et al., 2019; Kosicki et al., 2018). More specialized strategies such as anchored multiplex PCR (AMP) (Zheng et al., 2014) and UDiTaS (Giannoukos et al., 2018) utilize single-anchor amplicon sequencing to detect translocations and insertions at on-target regions, but would also not discover deletions if the PCR primer site was deleted. Long-read sequencing may be a useful alternative to capture these larger-scale events (Dastidar et al., 2018; Gasperini et al., 2017; Kosicki et al., 2018).

To increase the quantification accuracy of amplicon sequencing, the use of unique molecular identifiers (UMIs) has been proposed to reduce the impact of PCR amplification bias or other artifacts that may skew quantification of paired (Kennedy et al., 2014; Kinde et al., 2011) or single-anchor amplicon sequencing (Tsai et al., 2015). These strategies introduce a unique DNA barcode during library preparation or in the first PCR cycle that can be read out via NGS. Reads with the same barcode likely originate from the same molecule and can be considered as duplicates in the downstream analysis. Computational frameworks for incorporating UMIs into genome editing quantification will be useful to increase quantification accuracy (Clement et al., 2018).

Another challenge facing the field is lowering the limit of detection for rare editing events. Currently, the limit of detection is bounded by PCR amplification and sequencing error rates—with current NGS and amplification technologies, it can be difficult to determine whether a deviation from the expected reference sequence is due to sequencing error, PCR error, or a rare genome editing event. One solution is to use UMIs coupled with high sequencing coverage, so that reads from each UMI are sequenced several times and can be used to error-correct sequencing and/or PCR errors. Several additional experimental and computational strategies could lower this detection limit, including machine learning approaches to distinguish sequencing errors from genome edits (Poplin et al., 2018).

Assessing CRISPR-Cas Targeting Specificity

Many techniques have been developed to identify sites of potential off-target editing activity, i.e., editing at unwanted locations that may confound the interpretation of on-target perturbations or potentially complicate therapeutic applications (Figure 2; Table 1). The identification of off-target editing represents an important technical challenge for all genome editing platforms, including CRISPR. Thus, it is of great importance and interest to the scientific community to reliably detect and accurately quantify CRISPR off-target editing.

Figure 2. Overview of In Silico, In Vitro, and In Cellula Strategies to Nominate Off-Target Editing at Known/Unknown Loci.

In silico approaches (left panel) utilize sequence homology to identify genomic loci (top strand) with similarity to the gRNA sequence (bottom strand) up to a particular number of mismatches, 2019In silico approaches (left panel) utilize sequence homology to identify genomic loci (top strand) with similarity to the gRNA sequence (bottom strand) up to a particular number of mismatches. Other approaches use enzymatic modeling to predict gRNA binding specificity at putative off-target sites. Machine learning approaches have also been developed to identify off-target sites. In vitro approaches (middle panel) first extract DNA from cells, and in CIRCLE-seq, DNA is circularized. Next DNA is exposed to editing reagents and cleaved fragments can be selected and sequenced to identify off-target cleavage or subjected to whole genome re-sequencing with identification of cleavage sites by identifying reads that start or end at the same base position. In cellula strategies (right panel) introduce CRISPR reagents into cells in the native cellular context. Cleavage events can be detected through a variety of methods such as ligation of known sequences to double strand breaks, biochemical tagging with biotinylated primers, or immunoprecipitation of DNA repair factors recruited to sites of cleavage.

A two-step approach of nomination followed by validation has been widely used to identify off-targets. First, a superset of potential off-target editing sites are nominated using one or more in silico, in vitro, or in cellula approaches. Second, these sites are individually validated in the target cell type using the assays for characterizing editing at a single locus, with a strong preference for NGS given its greater sensitivity. Together, these steps have been used to comprehensively and sensitively identify off-targets in vivo (Akcakaya et al., 2018).

In Silico Nomination of Sites with On-Target Homology

Off-target genomic cleavage known to occur at certain sequences with high similarity to the on-target site (Cho et al., 2014; Pattanayak et al., 2013). However, the rules governing whether or not certain high-homology sequences are cleaved remain incompletely understood (Tycko et al., 2016). Given the sequence-dependence of off-target activity, there have been extensive efforts to computationally predict off-target sites at the gRNA design stage (Hanna and Doench, 2020). These efforts have focused on the identification of sites with sequence homology up to a specified number of mismatches and/or RNA/DNA bulges; some of these methods also attempt to incorporate predictions about the effects of mismatches on cleavage activities (Doench et al., 2014, 2016b; Sanson et al., 2018).

Putative off-target sites with high homology to the on-target site can be identified using standard sequence aligners. Ontarget sequence alignment to the reference genome of interest is performed, and all genomic loci up to a desired number of mismatches are reported (e.g., Blast [CRISPR-P (Lei et al., 2014)], Bowtie [Langmead et al., 2009] [CHOPCHOP (Montague et al., 2014), GT-Scan (O’Brien and Bailey, 2014), Bowtie2 (Langmead and Salzberg, 2012) (E-CRISP [Heigwer et al., 2014]), or BWA (Li and Durbin, 2009) (CRISPOR [Haeussler et al., 2016])]). Sequence aligners offer rapid and scalable in silico enumeration of closely matched sites but are limited because they do not perform an exhaustive genomic search, which can oftentimes lead to an incomplete list of such sites (Naito et al., 2004).

Newer search algorithms have been developed to overcome limitations of using general sequence aligners for finding potential off-target sites. Off-target sites can be enumerated rapidly if only mismatches to the on-target sequence are considered (McKenna and Shendure, 2018). Other search algorithms further include the ability to incorporate biological information about sequence requirements for different PAM sites (Xiao et al., 2014; Zhu et al., 2014), to identify sites with RNA and/or DNA bulges (Bae et al., 2014; Cancellieri et al., 2019), and to perform biochemical modeling of the enzymatic binding and cleaving process (Klein et al., 2018). Genetic sequence variants can also alter the off-target profiles of a genetic editor (Canver et al., 2018b; Scott and Zhang, 2017), which requires additional capabilities for identifying sites affected by these variants (Cancellieri et al., 2019). Of note, many of these search algorithms offer user-friendly web interfaces with a variety of customizable parameters to facilitate user-specific analysis (Bae et al., 2014; Heigwer et al., 2014; Montague et al., 2014; Naito et al., 2015). After enumerating sites from in silico analysis, the potential editing activity at these sites can also be predicted (Doench et al., 2016a; Hsu et al., 2013; Singh et al., 2015; Stemmer et al., 2015; Xu et al., 2017). The gRNAs can be chosen to minimize the overall number of closely matched sites in a genome of interest and to avoid off-targets that lie in annotated functional genomic regions such as coding sequences, promoters, putative enhancers, or insulators (Cancellieri et al., 2019).

In general, the presence of off-target editing decreases with an increasing number of mismatches (Cho et al., 2014; Haeussler et al., 2016). However, various experiments have shown that off-target mutations can occur in sites with as many as six mismatches relative to the on-target site (Tsai et al., 2015), but why some more closely matched sites are not mutated while other less closely matched sites are mutated remains poorly understood. (Tsai et al., 2015). As a result, the number of mismatches that should be tolerated during in silico off-target site enumeration remains an open question. Further insights may be derived from the increasing availability of larger experimental datasets cataloguing genome editing outcomes. Models have been trained on experimental data using a variety of computational approaches including logistic models (Allen et al., 2018), two-layer regression models (Elevation [Listgarten et al., 2018]), neural nets (Lin and Wong, 2018), or random forest regression models (CRISTA [Abadi et al., 2017]). However, generating sufficient data for model training has been a challenge. For example, previous work primarily utilized a published dataset of only 30 gRNAs for model training (Lin and Wong, 2018). In addition, these models were trained only for the CRISPR-Cas9 nuclease and are not directly applicable to other editors (e.g., Cas12a nuclease or base editors).

In Vitro Off-Target Nomination Assays

In vitro approaches measure the biochemical cleavage activities of CRISPR-Cas nucleases on DNA substrates in a cell-free or in vitro environment (as compared to in cellula assays in which the CRISPR-Cas nucleases are exposed to genomic DNA in a cellular context, see below). Multiple different in vitro off-target cleavage assays that were originally developed for zinc finger nucleases and TAL effector nucleases have been adapted for assessment of CRISPR-mediated off-target cleavage (Pattanayak et al., 2011). For example, one such strategy uses the generation of concatemeric DNA oligonucleotide libraries containing all possible variants with up to 8 mismatches relative to a gRNA sequence of interest (~10¹² distinct variations). After exposure to a CRISPR-Cas nuclease, flanking adapters are ligated at DSB positions to allow for PCR amplification and NGS to identify the cleaved sites (Pattanayak et al., 2013).

Whole genome sequencing (WGS)-based strategies have also been used to identify off-target cleavage events. The Digenome-seq method detects cleavage activity by fragmenting genomic DNA, treating the sample with a given CRISPR-Cas nuclease and gRNA of interest, and then performing WGS on the sample. Nuclease-mediated cleavage events are then identified by a pileup of reads that consistently terminate at a particular base position. Digenome-seq can be used to analyze multiple gRNAs in a single sequencing run, although reduced sequencing depth can lead to lower sensitivity for identifying low-frequency off-target events (Kim et al., 2015, 2016). WGS-based approaches are also limited by high cost, limited access to specialized high-throughput sequencing platforms (HiSeq x10 or NovaSeq), and highly inefficient yield of information due to the lack of enrichment for reads cleaved by the nuclease.

One strategy for improving the yield of information from an in vitro off-target identification strategy is to perform an enrichment for those sequencing reads that are informative for nuclease-mediated cleavage. The SITE-seq assay accomplishes this by ligating biotinylated adapters at DSB sites in genomic DNA, which can then be selectively enriched prior to sequencing (Cameron et al., 2017). Similarly, the CIRCLE-seq method accomplishes enrichment for desired cleavage events by first circularizing sheared genomic DNA fragments and then treating with a nuclease of interest. Following nuclease cleavage, the resulting free DNA ends can then be substrates for sequencing adaptor ligation followed by NGS. This strategy greatly enriches for nuclease cleavage events and enables NGS to be more efficient, requiring the use of only a MiSeq run to detect rare off-target events. In addition, CIRCLE-seq can be performed without the need for a reference genome and therefore can be utilized in cases where the reference genome sequence is unknown or when the genomic sequence of interest deviates significantly from the reference (Tsai et al., 2017).

In cellula Off-Target Nomination Assays

Cell-based off-target nomination assays take the endogenous chromatin and cell-type DNA repair preferences into account for the identification of off-target CRISPR-Cas nuclease activity (Iyama and Wilson, 2013; Riesenberg and Maricic, 2018; Sun et al., 2019).This can be an advantage if the assay can be performed in the actual cell type of interest. However, it can also be a disadvantage if one needs to perform the assay in a different cell type because cell-type-specific effects can lead to different off-target profiles. Initial in cellula nomination assays for CRISPR-Cas nucleases used chromatin immunoprecipitation-sequencing (ChIP-seq) to identify genomic loci bound by a catalytically inactive Cas9 (dCas9) as a surrogate for off-target cleavage sites (Duan et al., 2014; Kuscu et al., 2014; Wu et al., 2014). However, little correlation was found between dCas9 binding and Cas9 cleavage with many false positive sites bound by dCas9 but not cleaved by Cas9 (Tsai et al., 2015). DISCOVER-seq offers a ChIP-seq-based in cellula (and in vivo) detection method that identifies DNA associated with the MRE11 DNA repair protein, which is recruited to genomic loci that have DSBs (Wienert et al., 2019). However, the recruitment of MRE11 to DSB sites may be unrelated to nuclease activity, the method may introduce false positives intrinsic to the use of ChIP (e.g., antibody specificity, non-specific binding), and the assay exhibits lower sensitivity overall compared to other in cellula methods.

Another general strategy for in cellula nomination of off-target activity leverages NHEJ-mediated insertion of known sequences into sites following DSBs. For example, capture of integration-deficient lentiviral vectors (IDLVs) (Wang et al., 2015b), protected double-stranded oligonucleotides (GUIDE-seq) (Nobles et al., 2019; Tsai et al., 2015), or AAV genomes (Hanlon et al., 2019) have been used to identify off-target sites. With these strategies, the sites of insertion are selectively amplified by using a primer that is designed to be complementary to the inserted sequence, followed by NGS and read mapping to a reference genome. An alternative approach called BLESS labels the two ends of a DSB with a biotinylated linker followed by streptavidin capture (Crosetto et al., 2013). Other methods blunt DSBs and then ligate an adaptor that harbors sequences such as barcodes, UMIs, sequencing adapters, and/or a T7 promoter (for T7-mediated in vitro transcription as demonstrated by BLISS) (Yan et al., 2017). Yet another nomination strategy, high-throughput genome-wide translocation sequencing (HTGTS), exploits the formation of translocations between off-target DSBs and ontarget DSBs. HTGTS uses a sequencing primer targeting one side of the on-target editing location to sequence across the predicted cleavage position to unbiasedly identify off-target loci following translocation events (Frock et al., 2015; Hu et al., 2016).

Validation of Nominated Sites

Candidate off-target sites nominated by in silico, in vitro, or cell-based assays can be validated in cellula or in vivo using targeted amplicon sequencing, AMP, or UDiTaS. The latter two approaches have the advantage of being able to capture events other than simple indels (e.g., larger deletions, inversions, trans-locations). A major limitation of all existing validation approaches is their inability to detect indel mutations at frequencies lower than the error rate of NGS (typically 0.1 to 0.01%). In addition, because no gold standard currently exists for identifying off-targets, the use of sensitive nomination assays to identify sites for validation is strongly recommended (Akcakaya et al., 2018).

Technological Outlook

One of the key challenges for improving off-target nomination assays is to increase their specificity without sacrificing sensitivity to detect rare off-targets. Recent experimental approaches to improve double-stranded oligonucleotide (dsODN)-based integration methods for the detection of in cellula off-targets have incorporated additional sequences into the dsODN tag to reduce mis-priming events during amplification (Nobles et al., 2019). Computational frameworks have shown progress in modeling editing at on-targets and could be extended to predict whether editing will occur at off-targets as well (Allen et al., 2018; Leenay et al., 2019; Shen et al., 2018). Models could also potentially incorporate epigenetic information related to off-target sites, which has been shown to alter cleavage efficiency (Verkuijl and Rots, 2019).

WGS has been proposed as an alternate method to identify off-target editing activity by comparing the whole genome sequence of an edited sample against an unedited sample (Iyer et al., 2015; Smith et al., 2014; Veres et al., 2014). Multiple technological challenges exist with this type of approach, but one major challenge is finding a suitable control sample so that one can distinguish editing events from pre-existing genetic variation, DNA replication errors, or other non-editing sources of mutation (Lareau et al., 2018; Lescarbeau et al., 2018; Nutter et al., 2018; Wilson et al., 2018). Genetic heterogeneity may be overcome to some degree by performing CRISPR editing in one cell of an organism at the 2-cell state (Zuo et al., 2019).

Another challenge in the field of off-target nomination assays is identifying potential off-target activity of non-cleaving enzymes such as base editors. WGS has been used with some success (Jin et al., 2019; Zuo et al., 2019), but the method is very low-throughput and expensive because most of the DNA fragments sequenced do not contain useful information about off-target editing. Methods such as GUIDE-seq (Tsai et al., 2015) that select for edited DNA using integration of known sequence tags at DSBs cannot be used to identify off-targets of non-cleaving enzymes. Some methods have been suggested to selectively measure off-target editing of base-editors (Doman et al., 2020), but as novel CRISPR technologies emerge, new assays and methods will be required to assess their specificity. This challenge is compounded by the fact that unintended editing may affect other cellular components such as RNA (Grünewald et al., 2019).

Through the further development of off-target assays, the field can better compare specificity across different editing enzymes, more sensitively identify accurate gRNAs, and drive the development of highly specific editing tools to approach safe and effective therapeutic editing.

Characterization of DNA Elements Using a Single Perturbation

CRISPR-Cas systems can be used to target a perturbation to a single locus (Bauer et al., 2015). For example, it is possible to characterize the function of a gene using a CRISPR-Cas nuclease whereby a gRNA is targeted to a gene exon (usually one of the first exons) to perform gene knockout. Phenotypic measurements such as cell viability, protein staining, or gene expression can be used to determine the perturbation effect. When using a CRISPR-Cas nuclease (Cebrian-Serrano and Davies, 2017), genetic editing can introduce frameshift mutations or premature stop codons, resulting in nonsense-mediated decay of the transcript or nonfunctional protein. Alternately, base editors can be used to introduce a novel stop codon in the coding sequence (Billon et al., 2017). The non-coding region controlling gene expression can also be targeted using CRISPR technology to disrupt—for example—a transcription factor binding site. Notably, genetic changes are heritable across cell division and edits can easily be verified using approaches described above.

Epigenetic editing can be used to characterize the function of genes or regulatory elements. CRISPRa uses dCas9 fused to an activating domain (e.g., VP64) which can increase gene expression when targeted to promoter regions. When CRISPRa is targeted to inactive enhancers, these enhancers can become activated and result in an increase in target gene expression. CRISPRi can be used to repress active regulatory elements and silence target genes in a similar manner (Kampmann, 2018; Simeonov et al., 2017; Thakore et al., 2016).

Large deletions can be mediated using a pair of gRNAs flanking the region of interest (Chen et al., 2014; Mali et al., 2013). This approach can be useful for deleting entire enhancers or even genes (Moorthy and Mitchell, 2016). The robust downstream functional changes resulting from enhancer or gene deletion can be read out using a variety of phenotypic assays. However, deletion rates may be low, and inversion or translocation events and incomplete editing are also possible outcomes when using this approach (Canver et al., 2014).

Technology Outlook

Prediction of editing effects by Cas9 has been approached with some success using computational frameworks (Allen et al., 2018; Doench et al., 2016b; Leenay et al., 2019; Listgarten et al., 2018; Moreno-Mateos et al., 2015; Shen et al., 2018). Forecasting the functional effect of mutations has also shown promise in other contexts (Gallion et al., 2017; Lyon and Wang, 2012; Ng and Henikoff, 2001), and combining predicted phenotype effects with predicted genome editing mutations could aid in gRNA design. However, modeling the phenotypic effect of epigenome editors remains an important and unsolved problem, especially in the context of the endogenous epigenetic environment of the intended target.

Discovery of Functional DNA Elements Using Pooled Screens

The utility of the CRISPR-Cas system at a single locus can be applied to multiple loci using a pooled screen design where individual cells are edited using one or more gRNAs to enable high-throughput functional interrogation of the genome (Doench, 2018) (Figure 3; Table 1). Broadly, pooled screens introduce a single gRNA into each cell within a large pool of cells. Each gRNA targets a particular element (e.g., gene, non-coding sequence), and the element function is assessed following a CRISPR perturbation (genetic or epigenetic). Cells can be sorted or selected based on phenotypes of interest, and gRNAs associated with the phenotypic changes can be read out to discover associated genes. Alternately, single-cell assays can be used to characterize the effect of each perturbation in each cell.

Figure 3.

Overview of Analysis Strategies for Tiling and Gene-Targeted Pooled Screens followed by Screen Validation Genome-wide screening approaches (left panel) utilize gene-targeted gRNA libraries in viral vectors. gRNA abundance is determined before and after phenotypic selection or enrichment. Scores are generated by comparing the relative gRNA abundance in pre-and post-selection populations, and identification of critical genes is performed using mean/variance modeling to address overdispersion or hierarchical mixture models to accommodate gRNA- and gene-specific variation. Alternately, single-cell readouts such as scRNA-seq can be applied to populations of cells to link phenotypes to specific perturbations. Tiling screens (middle panel) are performed by targeting gRNAs across a genomic interval. gRNA abundance is determined before and after phenotypic selection/enrichment. Scores are generated by comparing the relative gRNA abundance in pre-and post-selection populations. The effect of each gRNA can be computed using simple moving averages, hidden Markov models, or deconvolution frameworks. Pooled screen validation (right panel) often involves re-testing gRNAs in an arrayed format in bulk cell populations or individual clones using the techniques for measuring editing at known loci. For example, next-generation sequencing of an individual gRNA target site can be followed by computational analysis to identify generated alleles and calculate a per-base activity score.

Pooled genome-wide screens allow for flexibility in design and can be adapted to study many different phenotypes of interest. After the design of an gRNA library, gRNAs can be synthesized on a large-scale to construct pooled libraries with up to hundreds of thousands of unique library members, typically involving the use of viral vectors such as lentiviruses or adeno-associated viruses. Pooled libraries are transduced at low multiplicity of infection (MOI) into a population of cells, whereby individual cells receive less than one gRNA, on average. An antibiotic selection can then be applied to select for cells that were successfully transduced. The subpopulation of cells that receive a given gRNA represent an individual experiment assessing the functional effects of that gRNA (Canver et al., 2018a).

Screen Design

Genome-wide, gene-targeted screens typically include 4–10 gRNAs per gene to reduce gRNA-specific effects and increase confidence in screen hits (Sanson et al., 2018). Several pre-designed genome-wide gene-targeted libraries exist with gRNAs that have been selected based on low off-target potential and high on-target potency (Doench et al., 2016b; Horlbeck et al., 2016; Sanson et al., 2018). While genome-wide gene-targeted pooled screens are most common, genome-wide transcription factor binding sites or other regulatory regions can be targeted as well (Fei et al., 2019; Seruggia et al., 2019; Zhou et al., 2014b).

Phenotype Readouts

Pooled gRNA libraries can be analyzed at the population or single-cell level. The gRNA incorporated into each cell is typically read out using amplicon sequencing. Population-level readouts require a selective enrichment step (e.g., cell sorting) for a phenotype of interest (e.g., cell viability, gene expression, differentiation). The functional effects of gRNAs are then inferred based on comparing gRNA frequencies between pre- and post-enrichment populations or across multiple sorted populations (Shalem et al., 2014; Wang et al., 2014; Zhou et al., 2014b). Quantification of the perturbation effect for each guide can be improved with the use of UMIs (Michlits et al., 2017).

Single-Cell Readouts

Pooled genome-wide screens have been complemented by the development of single-cell assays that can be used to measure the effects of gene knockout on complex cellular phenotypes. Single cell RNA sequencing (scRNA-seq) has been shown to be effective in measuring transcriptomic changes resulting from CRISPR-based genetic and epigenetic perturbations (Adamson et al., 2016; Datlinger et al., 2017; Dixit et al., 2016; Genga et al., 2019; Jaitin et al., 2016; Mimitou et al., 2019; Replogle et al., 2020; Xie et al., 2017) and can be used to probe regulatory circuits and identify gene interactions. These screens are designed such that the gRNA sequence (or a barcode that can be linked to the gRNA sequence) are read out in the single-cell RNA-sequencing and link the global transcriptional changes to the gRNA present in that cell.

The epigenomic changes resulting from gene knockout can also be uncovered by combining single-cell ATAC-seq with targeted amplification of the gRNAs after the pooled screen (Rubin et al., 2019). Spatially resolved cellular characteristics such as protein localization can be linked to genetic perturbations in single cells using pooled screens that combine an optical readout of a given phenotype with in situ sequencing of the gRNA present in each cell (Feldman et al., 2019).

Analysis

Several analytical approaches applicable to CRISPR genome-wide knockout screens have been developed previously in other contexts, particularly in the field of RNA interference (RNAi). RNAi uses synthetic anti-sense oligonucleotides to degrade mRNA transcripts in a targeted manner, leading to gene silencing. Pooled RNAi screens in which small-interfering RNA (siRNA) or short-hairpin (shRNA) libraries targeting many genes are applied to a pool of cells to identify critical genes for a given phenotype (Boettcher and Hoheisel, 2010; König et al., 2007; Luo et al., 2008). Both RNAi and CRISPR approaches take into account the differences in knockdown/knockout efficiencies of different library members, so certain methods originally developed for RNAi analysis have been adapted for genome-wide CRISPR knockout screens (König et al., 2007; Luo et al., 2008). The following sections introduce the key analytical considerations in analyzing genome-wide pooled screens.

Fold Change Analysis

One challenge in analyzing pooled screens is that the number of gRNAs targeting each gene is not necessarily uniform in the pre-enrichment gRNA pool, causing some gRNAs to be apparently more or less abundant in the post-enrichment readout. This can be addressed by using the fold-change analysis method, which is performed by comparing the abundance of each gRNA in the post-enrichment pool to the abundance in the pre-enrichment pool. This simple processing step mitigates the effect of nonuniform distribution of gRNA counts and is employed as a first step in most processing pipelines.

Modeling gRNA Effects

Next-generation sequencing of gRNAs relies on several rounds of PCR to create sequencing libraries. Because of this, gRNAs that are highly represented in the population tend to have greater variability in NGS readouts as compared to gRNAs that are lowly represented, a phenomenon called overdispersion. For this reason, many analysis tools account for over-dispersion in pooled screen gRNA counts using negative binomial distributions (e.g., PinAPL-Py [Spahn et al., 2017], RSA [König et al., 2007], RIGER [Luo et al., 2008]) or beta binomial distributions (e.g., CRISPRBetaBinomial [Jeong et al., 2019]). Using these distributions, p values can be calculated for each gRNA. Permutation-based non-parametric analysis (PBNPA) permutes gRNA labels to assign p values to genes without assumptions on the underlying distributions (Jia et al., 2017).

Aggregating gRNAs

For pooled screens in which multiple gRNAs target each gene, p values representing the effect of each gRNA are aggregated to derive a gene-level statistic. Each gRNA may have varying effects on its target due to induced allelic diversity, perturbation range, or cleavage efficiency and may even have unintended off-target effects that may affect the phenotype readout. MAGeCK ranks gRNAs by p value, then calculates gene-level p values using modified Robust Ranking Aggregation (Li et al., 2014). Maximum likelihood estimation methods (e.g., MAGeCK-MLE [Li et al., 2015]) and hierarchical mixture models (e.g., CRISPhieRmix [Daley et al., 2018], ScreenBEAM [Yu et al., 2016]) that account for variable gRNA efficiencies have been proposed to calculate gene-level statistics. If multiple experiments are performed using the same gRNA library, the software package Jacks can model individual gRNA effects using a Bayesian approach (Allen, Behan et al., 2019).

Quality Control and Data Visualization

Assessing data quality and visualizing results from CRISPR genome-wide screens is addressed by several methods. The MAGeCK algorithm is supported by MAGeCK-VISPR and MAGeCKFlute, which provide comprehensive quality control (QC) analysis and visualizations when working with the MAGeCK pipeline (Li et al., 2015; Wang et al., 2019). Other analysis platforms such as caRpools (R package), CRISPRcloud (cloudbased platform), and CRISPRBetaBinomial (R package) provide user-friendly environments to explore CRISPR genome-wide screen data (Jeong et al., 2017, 2019; Winter et al., 2016). These tools have extensive QC metrics, intuitive data visualizations, and a wide selection of popular statistical methods (including several described above) to analyze CRISPR screen data. QC and other data visualization techniques can help ensure appropriate analysis of screen data by investigators.

Single-Cell Analysis

Computational analysis of scRNA-seq readouts can be performed to measure transcriptional effects of perturbations in a pool of cells. In order to overcome the sparsity of scRNA-seq data, cells with poor transcriptional signal are discarded, or imputation of missing values is applied in some cases. scMAGeCK (Yang et al., 2020) and MIMOSCA (Dixit et al., 2016) model the effect of gRNAs on gene expression using a regularized linear model. MUSIC (Duan et al., 2019) uses topic modeling to discover biological functions induced by perturbations. It should be noted that the methods discussed in this section are far from comprehensive, and this is still a rapidly developing field.

Technological Outlook

Technical developments in the area of pooled genome screens include innovations in characterizing multiplexed perturbation effects as well as perturbation readouts. Screens in which multiple elements (e.g., genes) are perturbed in the same cell in a controlled manner can give additional insights into gene networks and interactions, including the identification of synthetic lethal or buffering gene interactions (Du et al., 2017; Han et al., 2017; Horlbeck et al., 2018; Najm et al., 2018). These gene-interaction screens may also incorporate single-cell readouts to discover complex responses to gene network perturbation (Norman et al., 2019; Replogle et al., 2020).

Pooled screens can be coupled with many single-cell analyses to read the perturbation effects to understand complex phenotypic consequences of gene perturbation. Multi-omic single-cell approaches may yield richer data to more comprehensively describe perturbation responses. However, interpreting the results of gene perturbations remains a computational problem because the data collected in a single experiment is very large and because many single-cell assays still exhibit low resolution and sensitivity (e.g., the transcription of only a subset of genes can be reliably measured using single-cell RNA-sequencing).

Discovery of Critical Elements within an Annotated Region of Interest Using Pooled Screens

CRISPR tiling screens allow for scanning of genomic regions to uncover functional sequences related to a phenotype of interest. Typically, gRNAs are designed in an unbiased fashion to target a genomic interval, which can include an entire locus or specific annotated elements within a given locus. Unlike genome-wide screens where targets are annotated, tiling screens are performed on coding or noncoding sequences to discover functional protein domains and critical regulatory elements, respectively (Canver et al., 2015, 2017; Fulco et al., 2016; Klann et al., 2017; Sanjana et al., 2016; Schoonenberg et al., 2018; Shi et al., 2015; Simeonov et al., 2017).

The analysis of CRISPR tiling screens depend on the type of CRISPR perturbation introduced, largely due to the differences in how these perturbations affect endogenous DNA. While epigenome editors (CRISPRa/i) remodel chromatin across hundreds of bp, CRISPR-Cas nucleases typically introduce narrow indels (oftentimes 1–10 bp). In general, simple moving averages have been used for CRISPRa and CRISPRi tiling screens (Fulco et al., 2016; Simeonov et al., 2017), whereas hidden Markov models (HMMs) and deconvolution frameworks have been used for CRISPR-Cas nuclease tiling screens (Canver et al., 2015, 2017; Hsu et al., 2018). The following sections introduce commonly adopted methods to analyze tiling pooled screens.

Fold Change Analysis

Similar to data from a genome-wide pooled screen, the data generated from a tiling screen is represented as gRNA counts in a pre- and post-enrichment population. Fold change analysis between the pre- and post-enrichment populations generally measures the gRNA effect sizes reasonably if the gRNA library is sampled sufficiently. The gRNA fold change values serve as input for most tiling screen analyses described below.

Simple Moving Averages

Simple moving averages (SMAs) smooth signal from CRISPR tiling screens by averaging across gRNA fold change values within fixed windows. The number of gRNAs that go into each “averaging window” is the only parameter for this strategy. SMAs have been successfully used in the analysis of CRISPRa/i tiling screens primarily because of the extent of shared information between neighboring gRNAs, meaning that the CRISPRa/i effect produced by two closely neighboring gRNAs will likely be similar (Fulco et al., 2016; Simeonov et al., 2017). Although smoothing is desirable to reduce noise of individual gRNAs, the averaging window size must be carefully selected or else short functional elements may be missed. Furthermore, the gRNA spacing in a tiling screen is not uniform, which can present problems in using a constant-length averaging window, especially if both densely and sparsely targeted regions exist within a screen. Following SMA analysis, statistical tests (e.g., t test) can be used to identify significant regions.

Hidden Markov Models

Hidden Markov models (HMMs) infer underlying DNA regulatory states (i.e., neutral, active, and repressive) from observations in the form of gRNA fold change values derived from CRISPR tiling screen data. The HMM uses perturbation effects of gRNAs to predict the underlying regulatory state at each position across the perturbation locus. The HMM has been successfully applied to CRISPR-Cas9 tiling screens to uncover critical regulatory DNA sequences within enhancer elements (Canver et al., 2015, 2017).

Deconvolution Framework

Deconvolution frameworks have recently been proposed for the analysis of CRISPR tiling screen data. This framework models the observed gRNA fold change values by means of a convolution operation between an underlying genomic regulatory signal and a CRISPR perturbation profile. The CRISPR perturbation profile takes the form of a parameterized Gaussian window or is constructed based on empirical data. This type of framework attempts to leverage biological knowledge of how different CRISPR technologies perturb endogenous DNA and models shared information between neighboring gRNAs based on this knowledge. Importantly, the deconvolution framework explicitly parameterizes the exact targeting coordinates of all tiled gRNAs and only models shared information between neighboring gRNAs if perturbation profiles overlap. In contrast, SMAs and HMMs force information to be shared based on the qualitative ordering of gRNAs and do not account for the details of non-uniform gRNA spacing. The deconvolution framework has been successfully applied to CRISPRa/i and CRISPR-Cas nuclease tiling screens and implemented in a software package called CRISPR-SURF (Hsu et al., 2018).

Validation

Validation is typically required after computational analysis of gene-targeted or tiling pooled screens (Figure 3). At present, screen analysis is reliant on gRNA enumeration or single-cell phenotype readouts (e.g, scRNA-seq) as opposed to direct assessment of perturbations (i.e., DNA sequence or epigenetic modifications). Therefore, validation often involves re-testing gRNAs in an arrayed format in bulk cell populations or individual clones to confidently connect gRNAs or specific genetic alterations to the phenotype of interest. Upon re-testing of gRNAs implicated by a given screen, analysis can be performed using the full spectrum of techniques described above for defined loci to characterize resulting DNA modifications as well as assess changes in gene expression, protein expression, or epigenetic marks.

Technological Outlook

Current analysis approaches for CRISPR-based pooled screens measure the change in gRNA abundances pre- and post- selection or their effects on gene expression through scRNA-seq. However, the activity of gRNAs and the resulting mutational spectrum from CRISPR-Cas targeting is lost in this analysis approach. Efforts to directly observe the introduced perturbations may enable more sensitive and higher-resolution screening capabilities. Screens targeting coding sequences may benefit from the in silico separation of frameshift and in-frame mutations to enable more accurate identification of critical residues within the target protein (Shi et al., 2015). Screens targeting non-coding sequences could also benefit from the identification of the precise alleles that result in a given phenotype of interest, with the potential to map functional transcription factor binding sites with higher resolution. Similarly, barcodes or gRNAs themselves have been used as surrogate readouts for targeted deletion screens (Gasperini et al., 2017; Zhu et al., 2016).

New CRISPR-based editing mechanisms provide new avenues for pooled screens moving forward. Base editing introduces specific substitutions with low indel rates and may provide an alternative method to study coding sequences without introducing frameshift mutations (Rees and Liu, 2018). Prime editing offers the unique ability to introduce a wide range of mutations at a target site, ranging from all substitution mutations to small insertion and deletion mutations. Though still in its technological infancy, prime editing opens the possibility of introducing precise genomic perturbations to study DNA sequence at single-base resolution (Anzalone et al., 2019).

Conclusions

The rapid innovations in genome editing technologies and their applications have been substantially fueled by both experimental and computational efforts. For the assessment of ontarget editing, methods that were primarily experimental with simple binary or semiquantitative readouts have transitioned toward NGS and tailored computational analyses that are able to comprehensively characterize genome editing outcomes at the allelic level. Advances in assays for detecting and quantifying off-targets have enabled researchers to build nucleases with improved genome-wide specificities. At the same time, specialized alignment algorithms and machine learning models are being leveraged to better design gRNAs and predict their off-target sites, which may accelerate the development of safer and more predictable editing tools. In the applications of using CRISPR technology to study different cellular phenotypes, initial studies focused on characterizing gene function using a single perturbation and have expanded into genome-wide or tiling pooled screens with a variety of single-cell readouts, creating a rich, comprehensive, and high- throughput view of many perturbation effects. The availability of these tools will further enable the elucidation of the biological underpinnings, regulatory structures, and non-linear interactions of gene networks through the use of big data approaches. Moving forward, both experimental and computational efforts will continue to foster the improvement and creation of accurate and precise CRISPR technologies for a variety of research and clinical applications.