CRISPR/Cas9-mediated homology-directed repair by ssODNs in zebrafish induces complex mutational patterns resulting from genomic integration of repair-template fragments

Site-specific genome-editing technologies enable the efficient generation of knock-out model systems. However, particularly for disease modeling, there is a growing necessity to complement knock-out models with more precise knock-in models, which will accommodate two needs. First, knock-in models are relevant to study diseases caused by specific and/or recurrent point mutations with dominant-negative, hypomorphic or gain-of-function effects. Second, this approach could serve as an in vivo tool to assess the pathogenicity of newly identified variants of unknown significance (VUS), which are massively identified by whole-exome sequencing. CRISPR/Cas9 genome editing is considered the most promising and versatile technology to create knock-in disease models (Cong et al., 2013; Jinek et al., 2012; Mali et al., 2013; Ran et al., 2013a). The CRISPR/Cas9 system induces a double-stranded DNA break (DSB), carried out by the Cas9 nuclease protein, at a specific target site, recognized by the binding of a single-guide RNA (sgRNA) molecule. Following DSB formation, two main endogenous repair mechanisms can be initiated: either the error-prone non-homologous end joining (NHEJ) pathway, often leading to the introduction of indel (insertion/deletion) mutations, or the homology-directed repair (HDR) pathway. HDR is only activated in the presence of a homologous repair template, naturally provided as the sister chromatid during the G2 and S phase of the cell cycle, leading to the reconstitution of the original sequence. The knock-in modeling procedure exploits this mechanism by supplying the CRISPR/Cas9 system with an artificial repair template, homologous to the target sequence and containing base-pair substitutions or gene insertions of interest. Various types of HDR repair templates have been used so far, including both circular and linear double-stranded DNA (dsDNA) molecules, and single-stranded oligodeoxynucleotides (ssODNs) (Salsman and Dellaire, 2017). For large-scale generation of knock-in models, preference is given to the latter as the design and production of ssODNs is easier, less time-consuming and cheaper than the generation of double-stranded templates such as plasmids (Leonetti et al., 2016; Mikuni et al., 2016). Additionally, recently, methods such as ‘easi-CRISPR’ have been developed for the generation of long ssODNs consisting of more than 1 kb of sequence, enabling the insertion of longer sequences such as reporters or gene tags (Miura et al., 2018). Moreover, with the use of ssODNs, illegitimate random integration of the exogenous DNA products into the organism's genome is expected to be less frequent as opposed to the use of dsDNA templates (Won and Dawid, 2017; Würtele et al., 2003; Zorin et al., 2005). Finally, single-stranded templates have shown superiority over similarly designed dsDNA templates in terms of HDR efficiency (Beumer et al., 2013; Miura et al., 2015; Yoshimi et al., 2016).

CRISPR/Cas9-induced HDR-mediated knock-in faces two major problems. First, the mechanism can be error prone. Indeed, imprecise repair events, such as targeted base-pair substitutions combined with insertions and deletions, have been reported (Gratz et al., 2013; Hoshijima et al., 2016; Hwang et al., 2013; Paix et al., 2017; Rivera-Torres et al., 2017; Zu et al., 2013), but the origin and extent of these erroneous repair events has not been thoroughly investigated. Second, its efficiency is relatively low compared to gene disruptions through indel generation. Several cell-endogenous DSB repair mechanisms, including NHEJ and HDR pathways, exist and interact with each other, although the molecular determinants of these interactions are not yet fully known (Kakarougkas and Jeggo, 2014). The decision on which repair pathway is activated is influenced by many factors, including the phase of the cell cycle, DSB complexity, chromatin structure (Brandsma and Gent, 2012; Frank-Vaillant and Marcand, 2002), and the composition and concentration of the repair template (Lin et al., 2014; Rivera-Torres et al., 2017). Previous research explored the impact of adaptations to the length, symmetry and strand complementarity of the ssODN repair template on genome-editing efficiency, although consensus is lacking concerning the impact of these different adaptations (Richardson et al., 2016; Yang et al., 2013). Added chemicals that either block the NHEJ pathway (SCR7, NU7441 and KU0060648) (Ma et al., 2016; Maruyama et al., 2015; Singh et al., 2015; Srivastava et al., 2012) or stimulate the HDR pathway (RS-1, L755507) (Jayathilaka et al., 2008; Pinder et al., 2015; Song et al., 2016; Yu et al., 2015) improve HDR rates in cellular systems to a certain extent, but still need further evaluation in other models.

The zebrafish is frequently used for disease modeling because of its many advantages, such as a rapid development, large offspring numbers, and the ease and speed in generating mutant lines. In this study, we evaluated the efficiency of point-mutation knock-in approaches in zebrafish models. We evaluated the effect of length, symmetry and strand complementarity of ssODN repair templates on HDR-mediated knock-in rates at multiple sgRNA target sites, and assessed a further impact of multiple chemical compounds that either block NHEJ or stimulate HDR. With next-generation sequencing (NGS), we accurately determined precise and erroneous knock-in efficiencies. Our work provides new insights into the effect of repair-template composition on HDR efficiency and stresses the importance of rigorous analysis of CRISPR/Cas9-edited genomic targets to evaluate the error-free nature of the obtained knock-in.

The generation of knock-in models has been greatly simplified by the emergence of CRISPR/Cas9-mediated HDR approaches, where genomic target sequences are replaced by homologous donor repair templates. Nevertheless, the efficiency of this process is low, precluding its use for high-throughput disease modeling or therapeutic applications. This study evaluates different approaches to improve the efficiency of scarless CRISPR/Cas9-mediated knock-in of single-base-pair substitutions by using ssODN repair templates in zebrafish, and accurately assesses total, ‘erroneous’ and ‘perfect’ repair rates by deep-sequencing analysis.

In a first step, we evaluated the most effective composition of the ssODN repair template for four different sgRNA target sites in the zebrafish genome. Repair-template length significantly affects total repair rates, with 120 nt being identified as the most optimal length to promote HDR. Shorter templates performed significantly worse, while longer templates showed a moderate drop in HDR rates. These findings are comparable to results obtained in human stem cells (Yang et al., 2013), pigs (Wang et al., 2016) and mice (Shen et al., 2013), where reported optimal total lengths approximate 90, 120 and 100 nt, respectively. We obtained average total HDR-mediated knock-in rates of 4-8% depending on the sgRNA target site when using 120 nt templates. This is more or less comparable to the results of Hruscha et al. (2013), to our knowledge the only other study in zebrafish evaluating HDR using NGS on whole-embryo DNA extracts. Other studies that investigated nuclease-mediated HDR in zebrafish only report on percentages of embryos containing at least one precise editing event, or make estimations on HDR efficiency based on PCR- or restriction-assay-derived data, impeding a direct comparison with our results (Albadri et al., 2017). The inclusion of repair templates that are complementary to either the sgRNA target or non-target strand enabled us to investigate the impact of strand preference on HDR rates. Our data do not support a clear strand preference for symmetrical repair templates at four different sgRNA target sites. This is in agreement with data obtained in other organisms (Lin et al., 2014; Renaud et al., 2016), but does not rule out specific locus-dependent strand preference as suggested previously (Hwang et al., 2013).

Next, we examined the impact of repair-template symmetry on HDR rates. A recent study has suggested that use of asymmetrical templates leads to improved HDR rates; Richardson et al. (2016) showed that, in cell cultures, complete dissociation of the Cas9-DNA complex after DSB creation is preceded by an early release of the PAM-distal non-target strand (Richardson et al., 2016). This insight might contribute to rational repair-template design, particularly by constructing asymmetrical templates complementary to the non-target strand, and with homology arms of 90 nt and 30 nt, corresponding respectively to the PAM-containing and the PAM-distal side of the DSB. Exploration of the performance of this specific repair template in zebrafish, alongside a number of other asymmetrical templates complementary to either the ‘target’ or ‘non-target’ strand, did not confirm that the abovementioned optimal asymmetrical repair template (designated ‘T 120 A right’ in our study), or any other asymmetrical template, performed notably better than the symmetrical repair templates. Interestingly, though, significant differences within the set of asymmetrical repair templates were noticeable, with ‘NT 120 A left’ and ‘T 120 A right’ clearly outperforming the other two asymmetrical repair templates. This remarkable difference, together with our finding that short 60 nt templates perform significantly worse than 120 nt or 180 nt templates, adds to our understanding of the mechanisms involved in DSB repair. A currently emerging hypothesis suggests that DSB repair using single-stranded templates proceeds through the single-stranded template repair (SSTR) pathway (Jasin and Haber, 2016). This process occurs independently of RAD51 and BRCA2, the key components of homologous recombination (HR) (Bothmer et al., 2017). It is proposed that the mechanism behind SSTR is based on synthesis-dependent strand annealing (SDSA), consisting of three major steps: (1) resection of the DSB to create 3′ overhangs, (2) pairing of the overhangs with the donor DNA followed by extension of the 3′ strand by DNA-polymerase-mediated DNA synthesis (initiation), and (3) strand displacement of the extended 3′ strand from the template strand with reannealing of the extended 3′ strand and the endogenous DNA at the opposite end of the DSB (resolution) (Fig. 6A). This process is characterized by unidirectional sequence conversion (Davis and Maizels, 2016; Kan et al., 2017; Paix et al., 2017). Templates containing a short (30 nt) homology arm at the side of the repair template that does not anneal to one of 3′ overhangs (‘NT 60 S’, ‘T 60 S’, ‘T 120 A left’, ‘NT 120 A right’), but participates in the resolution step, perform poorly in our study (Fig. 6B). These findings point towards a necessity for these particular homology arms to have an optimal length (about 60 nt) for resolution, at least in the zebrafish. The difference in performance between the asymmetrical templates included in this study may thus be linked to SDSA pathway requirements, instead of being attributed to the nature of the dissociation of the Cas9-DNA complex after DSB, as suggested by Richardson et al. (2016).

The outcome of precise genome editing using ssODNs depends on the perfect replacement of the target sequence by the repair template. Extended data analysis of the obtained deep-sequencing results in this study revealed that approximately half of the sequencing reads containing the base-pair alterations of interest additionally show a complex pattern of multiple integrated fragments of the repair template that can be present in both directions. This finding was not entirely unexpected, since a number of studies have reported imprecise HDR edits (Gratz et al., 2013; Hwang et al., 2013; Paix et al., 2017; Rivera-Torres et al., 2017; Zu et al., 2013), but the extent of this phenomenon was unanticipated. It is generally believed that the use of ssODNs, unlike dsDNA templates, circumvent random and illegitimate integration of the exogenous DNA products into the organism's genome (Won and Dawid, 2017; Würtele et al., 2003; Zorin et al., 2005). Therefore, the erroneous integration events encountered in this study are most likely attributed to mechanisms other than random integration. Since ssODN-induced DSB repair most likely occurs through SDSA (Paix et al., 2017), initiating DNA replication steps may be prone to polymerase slippage and stalling (Rodgers and McVey, 2016) with subsequent dissociation of unstable intermediates and template switching (Rodgers and McVey, 2016). Multiple rounds of template switching can lead to complex mutational spectra (Guirouilh-Barbat et al., 2014). Correspondingly, template switching during CRISPR/Cas9-mediated HDR in HEK293T cells has been recently demonstrated by Paix et al. (2017). We hypothesized that the presence of microhomologies in the target sequence may explain why certain targets are prone to template switching with subsequent erroneous repair, as previously seen in genomic rearrangement events (Löytynoja and Goldman, 2017) (Fig. 6C). As a proof-of-principle, we showed, for one of the repair templates targeting the smad6 gene, that erroneous repair occurred through multiple template switching events during SDSA-based DSB repair, based on the presence of short microhomology sequences (Fig. S4↗). The sequence-specific nature of these events holds promise for future development of prediction algorithms for erroneous repair in CRISPR/Cas9-mediated HDR.

Importantly, no repair template performs unambiguously best across all sgRNA target sites, when solely considering perfect repair rates, although there is a tendency for some asymmetrical repair templates (‘NT 120 A left’ and ‘T 120 A right’) to outperform their symmetrical counterparts (‘NT 120 S’ and ‘T 120 S’), comparable to previously reported results in zebrafish (Prykhozhij et al., 2018). Therefore, it is advisable to try different types of repair templates when attempting to knock-in a point mutation of interest.

Our data clearly indicate that the outperformance of some of the templates when considering total HDR rates was mainly due to erroneous repair. Excluding the sequencing reads that contain erroneous integration events leads to a drop in HDR rates (of the base pair located closest to the DSB site) from 4-8% to 1-4%, depending on the sgRNA target site. Moreover, these perfect repair rates drop even further if the distance between the base alteration and the Cas9 cut site increases, a phenomenon that was observed previously (Elliott et al., 1998; Paquet et al., 2016; Ran et al., 2013b). This implies that the distance between the Cas9-induced DSB and the intended sequence alteration should be minimized. Despite the low rates of scarless repair in zebrafish embryos, efficient transmission of nucleotide alterations to subsequent generations was found to be relatively efficient.

Finally, we evaluated HDR rates following chemical intervention blocking the NHEJ pathway (SCR7, NU7441, KU0060648) or stimulating the HDR pathway (RS1, L755507), an approach reportedly successful in cellular systems. Two delivery methods (incubation and injection) were applied. Surprisingly, none of these compounds affected either perfect or erroneous HDR rates in zebrafish. Possible explanations include: (1) delivery through injection or compound toxicity in the zebrafish hampers the delivery of the minimal compound concentration necessary for its intended activity, (2) the absence of a response to compound administration could be attributed to species differences, although it is known that DSB repair mechanisms are generally well conserved, or (3) the RAD51 stimulator RS1 might only have an effect when using double-stranded repair templates, since single-stranded template repair of DSBs is believed to act independently from RAD51 (Bothmer et al., 2017). In view of the latter, compounds that induce the SSTR pathway could have better potential in increasing HDR rates when using single-stranded templates. Interestingly, a recent study in zebrafish has shown that SCR7 administration through incubation increased HDR rates from 5 to 13% using an ssODN donor (Zhang et al., 2018). One possible reason that our results conflict with their conclusion is the difference in analysis method. We report repair rates that match the actual frequency of altered alleles in a pool of embryos, determined by NGS, while the study of Zhang and co-workers only reports the percentage of embryos containing HDR events. Moreover, these HDR rates are determined by a method solely based on PCR amplification. The use of PCR-based techniques using primers that specifically recognize the mutant sequence introduced by HDR should be avoided since they cannot distinguish erroneous repair from perfect repair and may thus lead to an overestimation of error-free HDR events. In addition, this approach lacks sensitivity to detect small differences in HDR rates. Another study performed in Xenopus showed that, similarly to our study, SCR7 was toxic to embryos and did not improve HDR rates (Aslan et al., 2017). Administration to non-dividing oocytes, on the other hand, appeared to significantly increase HDR rates. Therefore, the procedure of oocyte treatment could be a challenging but promising approach for future experiments in zebrafish, aiming to increase HDR rates (Xie et al., 2016).

Overall, in this study we tested the performance of multiple strategies in the zebrafish model system that were previously shown to improve HDR rates in cellular systems, including optimal donor template design and chemical compound administration. We demonstrate that HDR-mediated knock-in is influenced by the design of the ssODN template, while none of the investigated chemical compounds significantly increased HDR rates in zebrafish. Importantly, we identified that imprecise complex genomic insertions of the donor template constitute a significant portion of total HDR events, and suggest that repeated template switching during SDSA is the underlying mechanism of erroneous ssODN-mediated HDR. These findings stress the need for optimal experimental design and reliable assessment of HDR-mediated knock-in in zebrafish (). Box 1