Following the initial discovery of CRISPR/Cas9 system, major adaptations were made to enable the system to work in human cells. These adaptations included: (1) the codon-optimized sequences of Cas9 that ensured preferable expression by the codon table used in the organism of interest (Cong et al., 2013; Jinek et al., 2013) (2) the attached nuclear localization signals (NLSs) to Cas9 to ensure the nuclear localization of Cas9 in human cells (Cong et al., 2013; Jinek et al., 2013), and (3) a single guide RNA (sgRNA), termed gRNA, constructed to possess both the guiding portion in crRNA and an RNA scaffold derived from tracrRNA (Jinek et al., 2012). These adaptations have enabled the CRISPR/Cas9 system to be programmable to any gene region by changing the protospacer sequence and flexible for use in any organism of interest (Hsu et al., 2014).

Evidence of high specificity using nuclease-based genome editing systems is critical for genetic screening in preclinical studies and corresponding transitional research. Since functional DNA is not comprised of random sequence due to evolutionary constraint, identical copies or highly homologous sequences to a designated target could exist in the same genome. Unwanted off-target editing and consequential toxicity has been demonstrated in the use of ZFNs and TALENs (Miller et al., 2007; Szczepek et al., 2007; Guo et al., 2010; Doyon et al., 2011). Soon after the engineered CRISPR/Cas9 was shown to work in human cells, off-target edits induced by CRISPR/Cas9 were addressed using systematic screening approaches (Fu et al., 2013; Hsu et al., 2013; Pattanayak et al., 2013; Qi et al., 2020). Using the Surveyor assay, Cong et al. (2013) showed that some gRNAs bearing up to five mismatches with target sequences induced CRISPR-mediated cleavage. Further experimentation showed that selected gRNAs could induce cleavage events at undesired off-target sites with up to 6 mismatches using the T7E1 assay in three human cell lines. In addition, they did not detect any off-target edits using two selected gRNAs individually at ~50 tested potential sites each. Another study used synthetic oligomers to generate sequence libraries that contained 10¹² potential off-target sites derived from the sequence of 4 gRNAs (Pattanayak et al., 2013). The results showed that the cleavage events occurred at synthetic off-target sequences with up to 7 mismatches against treated gRNAs, in agreement with previous studies, showing that incomplete complementarity still induced CRISPR-mediated edits. Together these studies suggested several important concepts: (1) the positions of mismatches affected off-target activity; the mismatches distal to PAM site were better tolerated than those proximal to PAM, (2) off-target edits could occur even with more than six mismatches between gRNA and off-target DNA, and (3) design of gRNAs without detectable off-target events is possible; RNF2 and FACNF gRNA caused no off-target mutations.

The use of computational predictive tools in gRNA design has developed rapidly to accommodate the increasing needs of CRISPR/Cas9 applications. In addition to identifying potential targets, computational tools for gRNA design must rate the exclusivity of those targets in order to avoid the use of gRNAs with off-target proclivity. The search can be as simple as mismatch counts between guide and target. However, recent approaches have adapted sophisticated algorithms into search tools. BLAST serves as the most accessible way to identify off-target sites on the basis of sequence similarity. However, the uniform penalty matrix in BLAST is not sufficient to describe guide-target interaction in the CRISPR/Cas9 system.

Two initial studies utilized similar strategies to characterize off-target activity due to sequence mismatches in the 20-bp complementary target region. One generated a set of gRNA variants that possessed gRNA variants that contained one mismatch against a fixed on-target DNA sequence in the human genome. Hsu et al. quantified the effect of mismatches by high-throughput sequencing of PCR amplicons from the on-target site after treatment of CRISPR/Cas9 (Hsu et al., 2013). Given a 20-bp target site, a set of gRNAs that covered all possible single-mismatch guide sequences were generated systematically such that 3 possible mismatch mutations at each complementary position were synthesized to acquire the contribution of CRISPR/Cas9 activity of each position. The modification frequency at the 20-bp complementary region was used to describe the CRISPR/Cas9 activity at the target site, which was determined by the number of reads that contained either mutations or indels from deep sequencing. The result indicated that the base pairing at the PAM-proximal region tolerated less mismatches than the PAM-distal region. The authors aggregated sequence modification efficiency over 400 gRNA variants from 15 target sites within EMX1 gene regions, which created the pairwise penalty matrix for each type of mismatch spanning the guide-target binding region. A simplified 20-element matrix for a 20 bp guide-target pair was then used as the basis of the scoring algorithm for a gRNA design tool despite the type of mismatches. This score matrix is referred to as the MIT matrix.

Another experimental test using a larger number of gRNA variants demonstrated the improved prediction of potential off-target loci (Doench et al., 2016). The concept remained the same; given a fixed target DNA sequence, the reduction of CRISPR/Cas9 activity due to guide sequence mutations including 1-nucleotide mismatch, 1-nucleotide deletion, and 1-nucleotide insertion was measured. Over 27,000 unique gRNAs were generated to target the coding sequence of human CD33 regardless of PAM alternatives, along with the perfect match gRNAs for each selected target locus. This set of gRNAs gave a high coverage of every mutation type on each of the 20 guide-target base paring as well as every possible PAM. The goal of this experiment was to understand how CRISPR/Cas9 actively disrupts the expression of an easy-to-detect coding gene with or without guide sequence mutations against target DNA. Therefore, reduction of CD33 expression level on the plasma membrane was used to determine the CRISPR/Cas9 activity instead of deep sequencing. The percent activity was calculated by the mean differences of CD33 detected by phycoerythrin-conjugated anti-CD33 antibody between perfect gRNA and variant gRNA. A table of percent activity for every type of mutation (12 mismatch types × 20 positions and 64 possible PAMs) was used to generate the cutting frequency determination (CFD) score. The CFD score of a guide-target pair with multiple mutations is the multiplication of percent activity for specific mutations.

This data along with subsequent data generated from methods discussed below has led to vast increase in the number of computational techniques for predicting the likelihood of off-target cleavage. The range of computational tools for gRNA design includes E-CRISP (Heigwer et al., 2014), CRISPick (Doench et al., 2014), CHOPCHOP (Montague et al., 2014), CRISPR-ERA (Liu et al., 2015), CRISPOR (Haeussler et al., 2016), GUIDES (Meier et al., 2017), GeneArt (Liang et al., 2017), and uCRISPR (Zhang et al., 2019). More recently published tools tend to use the CFD matrix to evaluate penalty scores (i.e., CRISPOR, GUIDES, CRISPick, and GuideScan) and are therefore more reliable tools than those published before the development of the CDF matrix (i.e., E-CRISP, CHOPCHOP, CRISPR-ERA). More recently, the method uCRISPR has been shown to outperform methods using the MIT and CFD matrices (Zhang et al., 2019).

These tools have been reviewed previously and several publications offer a more in-depth review of this topic (Bolukbasi et al., 2016; Tycko et al., 2016; Manghwar et al., 2020; Wang et al., 2020). While the tools are useful for an initial discounting of egregious target choices, in silico predictions should always be confirmed by additional techniques.

Whole genome sequencing (WGS) is a straightforward approach to unbiased survey of the full genome for off-target nuclease activity. Endogenous repair mechanisms leave sequence-based evidence of nuclease activity on genomic DNA. Non-homologous end-joining (NHEJ) has been shown to introduce indels during the repair of double-strand breaks induced by nucleases (Cradick et al., 2013; Fu et al., 2013; Hsu et al., 2013; Pattanayak et al., 2013). Other repair outcomes for nuclease-induced DSBs include inversions, translocations, and large deletions (Frock et al., 2015; Hu et al., 2016; Giannoukos et al., 2018). Deep sequencing allows the identification of those repaired sites (Figure 1A). WGS ensures a survey of the full genome. There are several advantages to WGS as an off-target detection method. WGS allows an unbiased look at all sites across the genome and has been used to detect unpredicted off-target CRISPR/Cas9 cleavage in clonal cell populations and animal models (Smith et al., 2014; Veres et al., 2014; Dash et al., 2019). WGS detects the behavior of the nucleases in a cellular environment. The signatures of nuclease activity detected by WGS are introduced to the genomic DNA during endogenous repair processes. This is important because cellular features such as chromatin structure have been shown to impact the off-target profile of the CRISPR/Cas9 system (Kuscu et al., 2014; Wu et al., 2014; Chari et al., 2015; Chen et al., 2016, 2017c; Daer et al., 2017; Jensen et al., 2017; Kim and Kim, 2018; Chung et al., 2020). Furthermore, in vitro techniques for unbiased off-target detection have demonstrated that CRISPR/Cas9 cleaves more targets in vitro compared to targeting within the cellular environment thereby requiring further experimentation to validate the biological relevance of detected targets (Kim et al., 2015; Cameron et al., 2017; Tsai et al., 2017).

WGS has drawbacks though. It is considered inefficient due to a low signal to noise ratio. The vast majority of sequence data collected during WGS represents unedited genomic DNA and the depth-of-coverage for sequence locations of interest is sacrificed to the undisturbed regions. Thus, WGS is limited by throughput, cost, and efficiency compared to whole-genome methods which incorporate target enrichment strategies (e.g., GUIDE-seq) which are discussed in detail later. Nonetheless, the current efficiency of next-generation sequencing does enable this approach. In a study to detect off-target mutations in mice altered with Cas9, a reported 20–25 × depth of coverage was achieved for each sample as a single sequencing library using an Illumina HiSeq 2500 platform (Iyer et al., 2015). Results indicated that a sequencing depth of 10–13X was sufficient to detect 95% of homozygous variants. Other studies report between 33X and 50X coverage as necessary to detect single-nucleotide polymorphisms in human genomes (Bentley et al., 2008; Ajay et al., 2011). Exome sequencing has also been used to assess the targeting specificity of genome editing nucleases (Cho et al., 2014). In a study comparing whole genome sequencing to exome sequencing, the authors conclude that there is no difference in cost effectiveness between the two approaches with respect to detection of known variants across the exome and that WGS produces better uniformity of read coverage. The results of that study show a mean on-target depth of coverage of 14 × to capture 95% of single-nucleotide variants (SNVs) (Meynert et al., 2014).

Modern methods of off-target detection deliver sensitivity on the order of 0.1% meaning that cleavage events which occur in 1 out of 1,000 cells are detectable (Frock et al., 2015; Kim et al., 2015, 2016; Tsai et al., 2015; Cameron et al., 2017; Yan et al., 2017; Kim and Kim, 2018; Wienert et al., 2019). The studies described above do not pinpoint the depth of coverage in WGS necessary to match genome-wide off-target detection methods which incorporate target enrichment strategies. Furthermore, the metrics reported are not directly comparable to off-target detection sensitivity. What those studies indicate is that WGS sensitivity can be variable depending on experimental conditions and sequencing platform and that exome sequencing does not confer an advantage in this strategy.

A recent study which applied WGS for detection of gene-editing outcomes has implemented a technique termed genome-wide off-target analysis by two-cell embryo injection (GOTI) (Zuo et al., 2019). To implement this method single blastomeres of two-cell mouse embryos were edited with CRISPR/cas9 or a base editor and progeny cells were examined by WGS for SNVs. The results of this study showed that CRISRP/Cas9-induced mutations were not carried through cell division, an important characterization of CRISPR/Cas9 effects. GOTI underscores that WGS still plays an important role for off-target detection in some experimental paradigms. Other off-target detection methods would have been unsuitable for collecting these results.

Despite genome-wide surveillance which made WGS a potential choice for off-target detection, recently published methods for unbiased survey of the whole genome offer greater sensitivity, fewer false-positives, and a better signal-to-noise ratio (Frock et al., 2015; Kim et al., 2015; Tsai et al., 2015, 2017; Cameron et al., 2017; Yan et al., 2017; Kim and Kim, 2018; Wienert et al., 2019). For example, the whole genome sequencing approach has been improved in the form of in vitro nuclease-digested genome sequencing (Digenome-seq). Digenome-seq enhances WGS performance as an unbiased off-target detection method (Kim et al., 2015, 2016; Kim and Kim, 2018).

Digenome-seq is an unbiased, in vitro off-target cleavage detection technique (Kim et al., 2015). It introduces a change to the WGS approach by implementing nuclease cleavage outside of the cellular environment. Digenome-seq involves the in vitro digestion of genomic DNA using CRISPR/Cas9 and the gRNA to be evaluated. The digested genome is then prepared as an ordinary next-generation sequencing library. The alignment of fragment reads from nuclease cleavage sites is distinct from the staggered reads of other fragments because of the absence of sequence repair after nuclease cleavage. This is because endogenous DSBs occur at random locations while the targeted DSBs induced by nuclease cleavage occur at precise sequence locations. Nuclease cleavage sites are distinctly characterized therefore by repeated detection of DSBs at the same sequence location. Digenome-seq achieves target enrichment by introducing a distinct signature to nuclease cleavage targets which improves the resolution of cleavage detection to single-nucleotide precision (Figure 1B). This is not achievable using a WGS approach without the in vitro digestion of the genome due to the non-specific nature of the indels relied upon for detection (Kim et al., 2015).

There are several published improvements to the Digenome-seq technique. A multiplex version of Digenome-seq has been published, allowing the testing of multiple gRNAs on the same sample simultaneously (Kim et al., 2016). The multiplex method has several modifications. The algorithm used for analysis was modified to allow the identification of cleavage events that leave different end moieties, specifically one or two nucleotide overhangs; the original algorithm only detected blunt ends. This modification reduced false-negatives and identified targets missed by the original Digenome-seq algorithm. False positives were also reduced compared to Digenome-seq by transcribing gRNAs with a plasmid template rather than an oligonucleotide. Plasmid transcripts were reportedly less heterogenous than oligonucleotide transcripts leading to higher fidelity in target recognition. Multiplex analysis of gRNAs was achievable in the Digenome-seq methodology by choosing gRNAs with target sequences differing by at least 11 nucleotides and thus avoiding ambiguity in target detection within the same sample. Multiplex Digenome-seq results were achieved without an increase in depth of coverage. These results demonstrate not only the ability of the technique to detect off-target cleavage from multiple gRNAs simultaneously but also the ability of Cas9 to be directed to multiple targets in vitro by multiplexed gRNAs.

Measures of improvement to off-target detection techniques can depend on the specific measurement goals. The unfettered nature of Digenome-seq with respect to chromatin architecture can be viewed as an advantage compared to WGS or techniques such as genome-wide, unbiased identification of DSBs enabled by sequencing (GUIDE-seq) and high-throughput genome-wide translocation sequencing (HTGTS) (Frock et al., 2015; Tsai et al., 2015). GUIDE-seq and HTGTS are mentioned here to make a point of contrast compared to Digenome-seq; both will be discussed in detail in later sections. The distinction allows for the detection of otherwise obscured gRNA off-target affinities. However, DIG-seq, another Digenome-seq modification, can also be considered an improvement of the Digenome-seq method for the opposite reason (Kim and Kim, 2018). DIG-seq is a Digenome-seq based method applied to DNA with chromatin architecture in place. Native chromatin is isolated via nuclei extraction and put through the Digenome-seq protocol (Figure 1C). DIG-seq is considered an improvement over Digenome-seq under the assumption that the cleavage targets that will be detected under these conditions are of keener interest and greater relevance than the full palate of in vitro detected cleavage targets outside of the chromatin architecture. This assumption is upheld by the performance of DIG-seq. DIG-seq performance was compared to two other in vitro off-target detection methods: selective enrichment and identification of tagged genomic DNA ends by sequencing (SITE-Seq) and circularization for in vitro reporting of cleavage effects by sequencing (CIRCLE-seq), discussed in detail below. Although identifying fewer off-target cleavage sites than CIRCLE-seq or SITE-Seq for the same VEGFA target, DIG-seq had a 62% deep sequencing validation rate compared to 29 and 10% validation rate for the other two techniques, respectively.

SITE-Seq is an unbiased, in vitro detection technique for nuclease-induced DSBs (Cameron et al., 2017). SITE-Seq involves the in vitro digestion of genomic DNA with CRISPR/Cas9, similar to Digenome-seq. Following 3′ adenylation, DSBs are ligated with biotinylated Illumina-compatible adapters. This leaves a pool of labeled DSBs in genomic DNA, predominately induced by gRNA-guided nuclease cleavage, which allows the selective enrichment of sequence surrounding cleavage sites. Following the initial labeling of the strand break sites, the genomic DNA is fragmented, end-repaired, and 3′ adenylated allowing for another round of Illumina-compatible adapter ligation. Thus, fragments containing sequence from one side of a DSB are exclusively bookended by the P5 and P7 binding-sites necessary for Illumina sequencing. Biotin selection and PCR amplification then lead to a selectively enriched deep-sequencing library comprised predominately of sequences surrounding nuclease cleavage targets.

Similar to Digenome-seq, the technique relies on in vitro digestion of genomic DNA and nuclease cleavage targets are distinguished from randomly induced DSBs during sequence analysis by aligned read pileups. SITE-Seq differentiates itself from Digenome-seq particularly by the selective enrichment of nuclease-cleavage targets (Figure 1D). This aspect considerably increases the signal-to-noise ratio of the readout compared to Digenome-seq.

SITE-Seq is highly sensitive, around 0.1%. SITE-Seq analysis of the commonly used controls VEGFA and FANCF detected nearly all of the sites identified by Digenome-seq, GUIDE-seq, and (HTGTS). SITE-Seq reportedly detected all previously identified cellular off-targets from preassembled Cas9-gRNA ribonucleoprotein (RNP). Although Digenome-seq sensitivity is equivalent to SITE-Seq, the signal to noise ratio of SITE-Seq is far greater due to the process of enrichment, which allows sequencing of cleavage sites while excluding the remainder of the genomic DNA. However, SITE-Seq shares the problem of a high false-discovery rate with CIRCLE-seq and Digenome-seq. Cellular factors play a role in the off-target activity of nucleases. In vitro techniques identify potential off-target sites in the absence of such factors and the sheer quantity of potential sites can inhibit validation of relevant bona fide sites. For example, SITE-Seq identified nine novel off-target sites for VEGFA and two for FANCF in spite of limiting cellular validation to a subset of identified sites. This is touted as a feature in this instance, and it is a good demonstration of the sensitivity of the method. But if the identified potential off-targets for a particular gRNA are too numerous to be efficiently screened for cellular activity, validation of off-target sites becomes a biased technique in spite of the unbiased nature of the assay. A further complication is that the effect of cellular factors and nuclease concentration on off-target cleavage may limit the relevance of validation to the experimental conditions under which it is carried out.

CIRCLE-seq is an unbiased method for detection of off-target CRISPR/Cas9 cleavage (Tsai et al., 2017; Lazzarotto et al., 2018). The method entails fragmentation of genomic DNA via sonication, end-repair, and self-ligation of fragments for intramolecular circularization. After circularization, remaining linear DNA is digested using a plasmid-safe ATP-dependent DNase. What remains is a library of circularized fragments of genomic DNA which is then digested using CRISPR/Cas9 and an gRNA to be profiled for off-target affinity. During Cas9 digestion, circles containing on-target and off-target sequence are linearized and are then prepared for next-generation sequencing.

CIRCLE-seq was adapted from earlier published in vitro methods for characterizing the off-target profiles of genome-editing nucleases (Pattanayak et al., 2011, 2013). An in vitro selection method was introduced to characterize the performance of two zinc-finger nucleases (ZFNs) on a library of 10¹¹ sequences. ZFNs targeting human genes for CCR5 and VEGFA were used. VEGFA has become a standard control for evaluation of genome-editing nucleases (Frock et al., 2015; Kim et al., 2015, 2016; Tsai et al., 2015, 2017; Cameron et al., 2017; Yan et al., 2017; Kim and Kim, 2018; Wienert et al., 2019). Both ZFNs were able to cleave target numbers on the order of 10⁵ sequences, the majority of which do not arise in the human genome. CCR5-224 also cleaved 37 in vitro human sequence targets 10 of which were validated in human K562 cells. The VEGFA-targeting ZFN, VEGFA2468, cleaved 2652 human sequence targets in vitro, 32 of which were validated in human K562 cells (Pattanayak et al., 2011).

In a subsequent study, the previous in vitro library method for ZFNs was modified to measure CRISPR/Cas9 off-target capacity on an in vitro library of 10¹² sequences (Figure 2A) (Pattanayak et al., 2013). Between two gRNAs tested, five off-target human sequences were validated in HEK293T cells. Both ZFNs and the CRISPR/Cas9 system, were shown to exhibit off-target specificity dependent on enzyme concentration with some rare off-target cleavage events occurring only at higher enzyme concentrations (Pattanayak et al., 2011, 2013).

CIRCLE-seq is a further adaptation of the in vitro library off-target cleavage detection method (Figure 2B). Generating the in vitro sequence library from genomic DNA increases the relevance of the library of identified cleavage targets. Additionally, because of the mechanism of cleavage-detection in CIRCLE-seq, each readable fragment contains the sequence from both sides of a given cleavage-site allowing for reference-genome-free off-target sequence identification with single nucleotide resolution. Earlier in vitro library methods detected significant background sequence noise, with hundreds of thousands of in vitro cleavage targets that are not relevant to the human genome. CIRCLE-seq by contrast, finds only human-genome sequence targets.

At the time of initial publication, CIRCLE-seq was the only unbiased, in vitro alternative to Digenome-seq and in some facets of performance CIRCLE-seq exceeds Digenome-seq. In particular, CIRCLE-seq has 180,000-fold higher signal-to-noise ratio than Digenome-seq. CIRCLE-seq owes this increase to the process of enrichment which ensures that only cleavage-target sequences are prepared for deep sequencing. There is however a trade-off between the CIRCLE-seq and Digenome-seq techniques in terms of resource consumption as each CIRCLE-seq sample requires 25 μg of genomic DNA while each Digenome-seq sample requires 1 μg. The high background noise in Digenome-seq can make the identification of rare targets difficult, and it has been suggested that some valid off-target cleavage sites are missed by Digenome-seq because of the filtering thresholds necessary to process excessive background signal (Kim et al., 2015; Tsai et al., 2017). CIRCLE-seq is reportedly more sensitive than Digenome-seq. The error rate of current next-generation sequencing (~0.1%) is the limiting factor in the detection of rare off-target cleavage events. Both techniques directly detect cleavage events with single nucleotide resolution which is not common to all off-target detection methods (Frock et al., 2015; Tsai et al., 2015).

Recently an updated version of the CIRCLE-seq methodology has been published (Lazzarotto et al., 2020). The modified technique is called circularization for high-throughput analysis of nuclease genome-wide effects by sequencing (CHANGE-seq). CHANGE-seq utilizes a tagmentation reaction in early steps of the protocol which drastically reduces the labor and preparation time for this methodology. Compared to CIRCLE-seq, CHANGE-seq allows more rapid sample processing for higher-throughput experiments and will likely be the preferred method for any experiment using in vitro library digestion in the future.

In comparison to cell-based methods for unbiased off-target detection, in vitro methods boast some attractive features. In vitro methods avoid the need for transfection, which can complicate both inter- and intra-experimental comparisons. Also in vitro detection does not rely on endogenous repair pathways like WGS, GUIDE-seq, and HTGTS (Bentley et al., 2008; Ajay et al., 2011; Meynert et al., 2014; Veres et al., 2014; Frock et al., 2015; Iyer et al., 2015; Tsai et al., 2015). GUIDE-seq and HTGTS are mentioned here to make a point of contrast compared to CIRCLE-seq; both will be discussed in detail in later sections. However, in vitro techniques also do not give insight into the behavior of gene-editing nucleases in cells. The false positive rate for CIRCLE-seq is reportedly low enough that the sensitivity limits of deep sequencing inhibit its estimation. However, the false discovery rate is high in CIRCLE-seq meaning that CIRCLE-seq frequently identifies off-target sites in vitro that are not validated in cellular experiments.

The standard for validation of bona fide off-target sites is targeted deep sequencing. A method has been published that is termed verification of in vivo off-targets (VIVO) which consists of CIRCLE-seq to identify off-target candidate sites followed by targeted deep sequencing to validate those sites (Figure 2C) (Akcakaya et al., 2018). This hybrid technique constitutes a method for validating off-target sites in vivo in an animal model. Candidate sites were examined which were identified by CIRCLE-seq in the livers of mice treated with CRISPR/Cas9 in adenoviral vectors using targeted deep sequencing. To do so, they chose a subset of sites from three classes of off-target sequences that they delineate by high, moderate, or low CIRCLE-seq read counts. Results indicate that the probability of validating off-target sites is higher amongst sites that return higher CIRCLE-seq read counts. This agrees with the findings of the originally published CIRCLE-seq method which show that sites with higher CIRCLE-seq read-counts are more likely to be detected by the cell-based method GUIDE-seq (Tsai et al., 2015). Although CIRCLE-seq data sets provide an unbiased genome-wide survey of off-target proclivity for CRISPR/Cas9 gRNAs, the sheer volume of potential off-target sites limited the validation of sites in the VIVO study to a subset of candidates, essentially a biased analysis. Importantly though, off-target sites were validated across all classes in the VIVO study, i.e., high, moderate, and low CIRCLE-seq read counts, underscoring the need for comprehensive analysis of gene-editing nuclease targeting particularly with respect to therapeutic development.

GUIDE-seq is a method for tagging and enriching the sequence surrounding DSBs for deep sequencing (Tsai et al., 2015). Originally published in 2015, the technique remains an important methodology for assessing the targeting fidelity of genome-editing nucleases (Chaudhari et al., 2020). Briefly, cells are transfected with a plasmid coding for Cas9 and a gRNA and co-transfected with a blunt, double-stranded oligodeoxynucleotide (dsODN). The dsODN is then incorporated into DSBs during NHEJ, thus tagging DSB sites with a short, known sequence. Extracted genomic DNA is then fragmented enzymatically or via sonication and the resulting fragments undergo end-repair, dA-tailing, and ligation of a universal adapter sequence which is added to both ends of all fragments. Target enrichment is achieved by two rounds of PCR which amplify only fragments containing the dsODN. Thus, the amplified library consists of strands which each contain one half of the sequence surrounding a DSB repaired by NHEJ. GUIDE-seq is conceptually derived from earlier methods. Precursors to GUIDE-seq include anchored multiplex PCR (AMP) and integrase-defective lentiviral vector (IDLV) integration (Gabriel et al., 2011; Zheng et al., 2014; Wang et al., 2015).

AMP is a target enrichment method for deep sequencing applications. Early target enrichment methods include AmpliSeq, TruSeq Amplicon, HaloPlex, and Nested Patch PCR (Varley and Mitra, 2008; Johansson et al., 2011; Do et al., 2013; Yousem et al., 2013). AMP improves on these techniques by enriching targets with only one known primer binding site rather than two (Figure 3A). In principle, AMP resembles a much earlier method called rapid amplification of cDNA ends (RACE) which utilizes known DNA sequence to determine the sequence of an adjacent region (Frohman et al., 1988). AMP involves preparation of double-stranded cDNA or sheared genomic DNA using earlier published methods (Zheng et al., 2010, 2011; Neiman et al., 2012). Following end-repair and dA-tailing, sequencing adapters, called universal half-functional adapters, are ligated randomly to the ends of all fragments. Enrichment is accomplished by PCR amplification using anchored primers for known targets. Primers for a second round of PCR are 5′-tagged with sequencing adapters. The resulting libraries have a fully functional pair of adapters for deep sequencing. This results in the selective amplification of targets with only one known primer binding site. Unknown adjacent sequence is then captured, and genomic rearrangements can be identified following deep sequencing.

Detection of IDLV integration has been used to identify on- and off-target cleavage of ZFNs, TALENs, and CRISPR/Cas9 (Gabriel et al., 2011; Wang et al., 2015) IDLV detection takes advantage of the IDLV capability to integrate into DSBs during NHEJ. Integration tags break-sites with known sequence which can be exploited for target enrichment (Figure 3B). Targets are amplified for sequencing by linear amplification-mediated (LAM) PCR or non-restrictive LAM (nrLAM) PCR (Schmidt et al., 2007; Gabriel et al., 2009; Paruzynski et al., 2010). IDLV has shortcomings including a low rate of integration and the tendency of IDLVs to sometimes integrate at sites up to 120 bp from the target DSB site (Gabriel et al., 2011; Tsai et al., 2015).

GUIDE-seq technology is a significant advancement over its predecessors. AMP allows the selective amplification of sequence with one side known which was an important step forward from earlier PCR techniques requiring two known primer sites. GUIDE-seq allows selective amplification of a target sequence in which no portion is known by placing the anchor primer on the dsODN (Figure 3C). This is essentially the principle behind IDLV detection but the more reliable rate of uptake of the dsODN into DSBs and the precise integration between the two ends of the DSB mark GUIDE-seq as a significant advance over IDLV.

At the time of publication GUIDE-seq set a new benchmark for off-target detection of nuclease-induced DSBs by filling a methodological gap for unbiased survey of the full genome with an effective target enrichment strategy that greatly improved the signal to noise ratio of off-target detection methods utilizing deep sequencing. GUIDE-seq has a detection sensitivity of ~0.12%, equivalent to that of other current methods (Kim et al., 2015, 2016; Tsai et al., 2015; Cameron et al., 2017; Kim and Kim, 2018). Furthermore, the biological relevance of GUIDE-seq data tends to be more robust than other methods because DSBs are tagged in the context of a cellular environment, not requiring targeted sequence validation for recognition as a bona fide editing site.

However, there are several limitations to the GUIDE-seq method. The dsODN, the key component to the effectiveness of the method, has not been adapted to be administered in an animal model, limiting the range of GUIDE-seq application. In addition, the dsODN has shown cytotoxicity in some primary cells (Wienert et al., 2019). Another limitation of GUIDE-seq is its dependence on the endogenous process of NHEJ to detect and tag cleavage events. DSBs not processed by NHEJ will be missed by the GUIDE-seq method.

A recent update to the GUIDE-seq approach is the iGUIDE method which deals with the problem of mispriming in GUIDE-seq experiments (Nobles et al., 2019). During library preparation, GSP primers can anneal to fragments which lack the dsODN. Amplification can then yield false positive library fragments containing human DNA sequence that were not the sites for nuclease cleavage and dsODN incorporation but functionally resemble true positive library fragments. The iGUIDE method involves the use of a 46 bp dsODN in place of the 34 bp version in the original method. The additional sequence allows filtering of misprimed library fragments during analysis. Use of the iGUIDE method reportedly reveals features of DSB distribution, such as the stronger tendency for spontaneous DSBs to occur near active genes, which are obfuscated by the noise generated by unfiltered mispriming events (Nobles et al., 2019). To date, the iGUIDE method has gained very little traction and is cited by only a single data paper in the literature. Further discussion in this manuscript will be focused on GUIDE-seq in its originally published form.

Tagmentation-based tag integration site sequencing (TTISS) is a recently published technique which enables a multiplex examination of nucleases and nuclease targets (Schmid-Burgk et al., 2020). The technique is based on GUIDE-seq with some modifications. The protocol is streamlined by utilizing the previously published Tn5 transposase for tagmentation (Picelli et al., 2014). DNA is then purified by spin column and target enrichment is accomplished via two nested PCR reactions. TTISS was used to examine the balance between specificity and activity in nine SpCas9 variants including wild-type SpCas9, seven previously published variants, and one novel variant (Kleinstiver et al., 2016; Slaymaker et al., 2016; Chen et al., 2017a; Casini et al., 2018; Hu et al., 2018; Lee et al., 2018; Vakulskas et al., 2018; Schmid-Burgk et al., 2020). The results indicate a trade-off between specificity and activity in general with the precise ratio differing between Cas9 variants. Sequenced targets are attributed to a given Cas9-gRNA pair on the basis of sequence homology. This was effective in the published experiment but could conceivably confound interpretation of some results, limiting the usefulness of TTISS in some contexts. TTISS can reportedly be scaled to accommodate 60 gRNAs per transfection in HEK293T cells. But there is a trade-off in efficiency with 28% fewer off-target sites detected in a multiplexed experiment. The technique is effective for a large-scale screen of Cas9 variants but for a comprehensive look at the full off-target profile of a given Cas9 variant and gRNA-target, the reduced detection efficiency would dictate the use of another technique, e.g., GUIDE-seq or discovery of in situ cas off-targets and verification by sequencing (DISCOVER-Seq) (Tsai et al., 2015; Wienert et al., 2019).

GUIDE-seq is not the only relevant modification to the AMP methodology. Uni-directional targeted sequencing (UDiTaS) is also a useful DSB detection technique which utilizes universal adapters and anchored primers to characterize the repair outcomes following engineered nuclease cleavage (Figure 3D) (Giannoukos et al., 2018). The modifications introduced in UDiTaS increase the robustness and utility of the AMP approach. In particular, UDiTaS introduces enzymatic fragmentation known as tagmentation, for genomic DNA rather than shearing by sonication. This modification addresses the tendency for shearing by sonication to introduce damage to genomic DNA that leads to base miscalling during deep sequencing (Costello et al., 2013; Chen et al., 2017b, 2018). UDiTaS introduces a novel Tn5 transposon which contains an Illumina forward adapter (i5), a barcode, and a UMI. Tagmentation yields a fragmented genomic library with adapters on either end of each fragment. Sequence-specific primers are then used to PCR amplify sites targeted by engineered nucleases. A second round of PCR adds an Illumina reverse adapter (i7), similar to the GUIDE-seq protocol. Not only does tagmentation drastically improve efficiency in hands-on time for library preparation protocols, but it also reportedly showed increased library complexity and increased linearity between expected and measured editing outcomes compared to AMP (Giannoukos et al., 2018).

As an off-target detection technique UDiTaS has limited utility due to its biased nature. Sequence-specific primers target sites of interest which require a priori knowledge to design. However, UDiTaS has significant utility in its ability to characterize repair outcomes for nuclease-induced cleavage. This is due to the structure of constructed library segments and the use of site-specific primers. Deep sequencing of UDiTaS will capture the junctions of repaired DSBs and thus structural rearrangements can be identified. These include translocations, inversions and large deletions. GUIDE-seq, by its nature does not detect those repair outcomes. The inserted oligonucleotide, which allows anchored priming without sequence knowledge for cleavage sites in GUIDE-seq, allows the capture of only one half of any repaired DSB junction. Reconstruction of complete cleavage sites is accomplished by mapping during analysis (Tsai et al., 2015). Thus, UDiTaS fills an important gap for data relating to repair outcomes for nuclease induced DSBs. Importantly, one approach to the problem of detecting large deletions is to use long read sequencing technologies (Amarasinghe et al., 2020). However, the accuracy and affordability of short-read sequencing platforms by comparison often make short read next-generation sequencing methods preferable and more accessible. An advantage of UDiTas is that it allows the capture and sequencing of large deletions on short read sequencing platforms. Notably, WGS could also be used to detect translocations, inversions, and large deletions but without targeted enrichment the signal to noise ratio of WGS would be markedly lower. Targeted deep sequencing on the other hand cannot capture translocations and efficient capture of inversions and large deletions would require more a priori knowledge for targeted deep sequencing than UDiTaS.

HTGTS is a method to detect and sequence translocations resulting from DSBs. Originally it was published as a method to study the mechanism of translocation (Chiarle et al., 2011). It has since been adapted as a method to detect off-target cleavage events caused by gene-editing nucleases (Frock et al., 2015). The original published HTGTS method utilized the I-SceI meganuclease to introduce targeted DSBs to specific c-myc and IgH loci. The sites were selected for their frequent involvement in B cell lymphoma oncogenic translocations (Chiarle et al., 2011). DSBs induced at these known locations were then subsequently fused to other DSBs across the genome by endogenous processes (Figure 4A). By exploiting the known sequence of one side of the translocation junction, the sequence of fused sites involved in translocation can then be identified. The original study presented two enrichment chemistries for library preparation to capture the sequence surrounding translocations. Starting with genomic DNA containing translocation fusions with known sequence on one half of the translocation junction, the genomic DNA samples are sheared via restriction enzyme digestion. End-repair and adapter ligation are then carried out for all fragments in a sample (Figure 4B).

The HTGTS method was repurposed for detection of nuclease off-target activity and protocol modifications were introduced that enhance the adapter-PCR target-enrichment methodology of the original method (Figure 4C) (Chiarle et al., 2011; Frock et al., 2015). The modified method is called linear amplification mediated (LAM) high throughput genome-wide translocation sequencing (LAM HTGTS). Applying the HTGTS method, introduced previously, as a nuclease off-target detection method is effectively a function of choosing applicable nucleases to induce desired bait and prey cleavage events. Using the original published method of HTGTS, previously unidentified off-target sites for the I-SceI nuclease were reported. In the updated LAM HTGTS, protocol modifications contribute to the performance of HTGTS as an off-target detection method enabling sensitivity and throughput comparable to other contemporary methods (Frock et al., 2015; Kim et al., 2015, 2016; Tsai et al., 2015; Hu et al., 2016; Cameron et al., 2017; Yan et al., 2017; Kim and Kim, 2018; Wienert et al., 2019).

The two key modifications introduced in the LAM HTGTS protocol are LAM PCR and bridge adapter ligation. LAM PCR is a method of target enrichment for sequences with a single known primer site (Schmidt et al., 2007; Paruzynski et al., 2010). LAM PCR utilizes a 5′ biotinylated primer targeting the known half of each captured junction i.e., one of the two sides of the DSB at the bait site, to linearly amplify across junction sites. Streptavidin selection is then used to magnetically isolate target sequences from genomic DNA. Bridge adapter ligation uses a double-stranded linker with a nucleotide-variable 3′ overhang to facilitate the attachment of adapters to the single-stranded library resulting from linear PCR (Figure 4C) (Zhou et al., 2013; Frock et al., 2015; Hu et al., 2016). Implementing these modifications yields 10–50 times more junctions for sequencing compared to the unmodified HTGTS method (Hu et al., 2016).

Performance of LAM HTGTS is comparable to other methods. For gRNAs targeting VEGFA and EMX1, LAM HTGTS identified the same major off-target sites as GUIDE-seq, although the two methods each identified unique subsets of low frequency off-target cleavage sites. This could be due to the cell lines tested but also to differences in the detection methods, which, by nature, may not be able to identify the same low-abundance cleavage sites (Hu et al., 2016). In particular, HTGTS can capture DSBs containing overhang ends, due to the endogenously repaired nature of translocation junctions, while GUIDE-seq only detects blunt-ended cleavage sites, due to the nature of uptake for oligonucleotide linkers (Tsai et al., 2015; Hu et al., 2016).

One drawback to the LAM HTGTS method is the substantial requirement of starting material. Translocations are rare compared to local rejoining events. They occur in 0.1–0.5% of cells in HTGTS libraries. The authors recommend a starting DNA mass between 20 and 100 μg for a single HTGTS library to achieve a 0.5–1.0 × 10⁶ read depth on an Illumina MiSeq (Hu et al., 2016). GUIDE-seq, by contrast requires 800 ng of genomic DNA to achieve comparable detection sensitivity. Although the authors state that the sensitivity of LAM HTGTS could be increased by starting with even more DNA, the input requirements could be prohibitive for this technique on samples of limited abundance.

There is an additional point worth noting, which is made clear by the results presented in the HTGTS publications (Chiarle et al., 2011; Frock et al., 2015; Hu et al., 2016). Even on-target cleavage events can have undesirable consequences. Translocations contribute to genomic instability (Elliott and Jasin, 2002; Ramiro et al., 2006; Kosicki et al., 2018). Also, translocations can result from on-target cleavage events as readily as off-target cleavage events (Chiarle et al., 2011; Frock et al., 2015; Hu et al., 2016; Kosicki et al., 2018). This point highlights the need for detailed characterization of genome-editing systems.

DISCOVER-Seq (described below) is an off-target detection method which selectively amplifies CRISPR/Cas9 cleavage sites by detecting the signature of endogenous DNA repair processes (Wienert et al., 2019). The basis of DISCOVER-Seq is ChIP-seq which entails chromatin immunoprecipitation (ChIP) and subsequent deep sequencing of captured DNA fragments (ChIP-seq). Briefly, ChIP begins with formaldehyde crosslinking of a single-cell suspension (Hoffman et al., 2015). Nuclei are then extracted and fragmented via sonication. Fragments of interest can then be isolated—pulled down—using bead-bound antibodies allowing the study of protein-DNA interactions (Kim and Ren, 2006; Wienert et al., 2019). In the ChIP-seq methodology, the pulled-down DNA fragments are then prepared for deep sequencing (Wienert et al., 2019).

ChIP has been extensively employed to capture the sequence surrounding DSBs and characterize the genomic landscape of DSBs. Early studies utilized tiled microarrays with DNA pulled down by ChIP in a method dubbed ChIP-chip (Iacovoni et al., 2010; Szilard et al., 2010; Staszewski et al., 2011). More recent studies have moved to ChIP-seq, utilizing contemporary sequencing methods coupled with ChIP (Kim and Ren, 2006; Frietze and Farnham, 2011; Rodriguez et al., 2012; Barlow et al., 2013; Yamane et al., 2013; Zhou et al., 2013; Duan et al., 2014; Kuscu et al., 2014; Wu et al., 2014; Khair et al., 2015; Knight et al., 2015; Madabhushi et al., 2015; O'Geen et al., 2015). γH2AX has been used as a marker for DSBs in ChIP experiments (Iacovoni et al., 2010; Szilard et al., 2010; Rodriguez et al., 2012) DSBs trigger expansive γH2AX binding domains however, and γH2AX can bind kilobases away from the site of a DSB, yielding poor resolution for DSB mapping (Bonner et al., 2008; Iacovoni et al., 2010).

Studies using ChIP-seq to characterize CRISPR/Cas9 off-target proclivity represent early attempts at unbiased survey of Cas9 activity on a genome-wide scale. Multiple studies used ChIP-seq with catalytically inactive Cas9 (dCas9) to pull down Cas9 binding sites (Duan et al., 2014; Kuscu et al., 2014; Wu et al., 2014; Knight et al., 2015; O'Geen et al., 2015). However, ChIP-seq using dCas9 is limited with respect to off-target detection; it has been shown to yield abundant false positives (Kuscu et al., 2014; Wu et al., 2014; Knight et al., 2015; Tsai et al., 2015). For example, only one out of 295 dCas9 binding sites identified by ChIP-seq in mouse embryonic stem cells (mESCs) was identified by targeted sequencing as a bona fide cleavage-target (Wu et al., 2014).

DISCOVER-Seq advances the ChIP-seq method by utilizing meiotic recombination 11 homolog 1 (MRE11), a DNA repair protein that is part of the MRE11-RAD50-NBS1 (MRN) complex (Figure 5). The MRN complex is involved in DNA damage responses (DDRs) in general, including DSB repair (Connelly and Leach, 2002; Moreno-Herrero et al., 2005; Borde, 2007; Oh and Symington, 2018; Syed and Tainer, 2018; Bian et al., 2019). It also has roles in replication stress, handling of dysfunctional telomeres, cellular response to viral infection, and tumorigenesis (Spehalski et al., 2017; Syed and Tainer, 2018; Bian et al., 2019). Notably, the way that the MRE11 subunit in particular handles different DSB end-moieties may dictate whether DSBs are repaired by HR or NHEJ (Shibata et al., 2014; Liao et al., 2016).

MRE11 is optimal for nuclease-cleavage detection because the MRN complex localizes to DSBs, including those created by CRISPR/Cas9, before ends are joined by repair (Syed and Tainer, 2018; Bian et al., 2019; Wienert et al., 2019). MRN is recruited to DSBs by γH2AX. In addition, MRE-11 is ubiquitous and conserved across all taxonomic kingdoms (Connelly and Leach, 2002; van den Bosch et al., 2003; Wienert et al., 2019). Disruption of each individual component of the MRN complex has been shown to be embryonically lethal in mice (Luo et al., 1999; Zhu et al., 2001; Buis et al., 2008) and mutations in the genes of each individual component have been linked to genomic instability in humans (van den Bosch et al., 2003). Expression of MRE11 across a range of tissues in mice has been demonstrated and following induction of DSBs, MRE11-detection peaks in cells before indels are formed (Wienert et al., 2019).

DISCOVER-Seq detects DSBs with single-nucleotide resolution and compares favorably to other off-target detection methods. However, DISCOVER-Seq reportedly has a sensitivity threshold of 0.3%, slightly higher than other contemporary techniques. A VEGFA target was examined in human K562 cells using both DISCOVER-Seq and GUIDE-seq. They identified 49 off-target sites in common between the techniques but also 41 off-targets sites unique to GUIDE-seq and eight off-target sites unique to DISCOVER-Seq (Wienert et al., 2019). This head-to-head comparison suggests that capture of the entirety of the off-target landscape for at least some gRNAs will require multiple methods. Another favorable feature of DISCOVER-Seq compared to GUIDE-seq is that DISCOVER-Seq works in primary induced pluripotent stem cells (iPSCs). DISCOVER-Seq was shown to detect off-target sites in iPSCs and to differentially detect an allelic specificity in primary cells from a Charcot-Marie-Tooth (CMT) patient with a heterozygous mutation. Data was also shown demonstrating that transfection of the dsODN necessary for GUIDE-seq was toxic to iPSCs (Wienert et al., 2019).

Although other techniques may boast greater sensitivity, DISCOVER-Seq is currently one of only two techniques shown to detect off-target events in vivo in an animal model (Wienert et al., 2019); VIVO is the other (Akcakaya et al., 2018). DISCOVER-Seq was tested on the same system as VIVO for comparison. A Pcsk9-gP gRNA was delivered via adenoviral infection in a murine model. Mice were then sacrificed at 24-, 26-, and 48-h time points. Twenty-seven off-target sites identified by DISCOVER-Seq were validated by amplicon sequencing and had indel rates between 0.9 and 78.1%. An important point of comparison is that 17 of the 27 sites identified by DISCOVER-Seq were identified by the in vitro CIRCLE-seq stage of the VIVO method but were not validated due to the high volume of potential sites generated by the CIRCLE-seq method. This is an important point with respect to the efficiency of in vitro techniques and the differential utility of currently available off-target detection methods. Unbiased full-genome survey of the off-target landscape is critical for translation of gene-editing to clinical application. And in vitro methods are sensitive and thorough means to characterize the activity of targeted nucleases with respect to sequence homology alone. But the need to validate the high volume of targets detected with in vitro methods can lead to a biased survey of high-priority or high-probability sites and bona fide off-target loci can be lost among the false positives.