Author response: Fitness effects of CRISPR endonucleases in Drosophila melanogaster populations

Article Figures and data Abstract Editor's evaluation Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)/Cas9 provides a highly efficient and flexible genome editing technology with numerous potential applications ranging from gene therapy to population control. Some proposed applications involve the integration of CRISPR/Cas9 endonucleases into an organism’s genome, which raises questions about potentially harmful effects to the transgenic individuals. One example for which this is particularly relevant are CRISPR-based gene drives conceived for the genetic alteration of entire populations. The performance of such drives can strongly depend on fitness costs experienced by drive carriers, yet relatively little is known about the magnitude and causes of these costs. Here, we assess the fitness effects of genomic CRISPR/Cas9 expression in Drosophila melanogaster cage populations by tracking allele frequencies of four different transgenic constructs that allow us to disentangle ‘direct’ fitness costs due to the integration, expression, and target-site activity of Cas9, from fitness costs due to potential off-target cleavage. Using a maximum likelihood framework, we find that a model with no direct fitness costs but moderate costs due to off-target effects fits our cage data best. Consistent with this, we do not observe fitness costs for a construct with Cas9HF1, a high-fidelity version of Cas9. We further demonstrate that using Cas9HF1 instead of standard Cas9 in a homing drive achieves similar drive conversion efficiency. These results suggest that gene drives should be designed with high-fidelity endonucleases and may have implications for other applications that involve genomic integration of CRISPR endonucleases. Editor's evaluation The manuscript describes an attempt to assess fitness costs of CRISPR/Cas9 endonucleases in the context of gene drive in Drosophila melanogaster by looking at direct fitness costs of the transgene and indirect fitness costs due to off-target cleavage. The authors performed experimental cage population studies and a maximum-likelihood approach to disentangle the contribution of direct and off-target-related fitness costs. The combined experimental and mathematical approach allows the authors to conclude that off-target cleavage is largely responsible for the observed fitness costs, although no mutated alleles were detected at the most likely computational predicted off-target sites. The authors also use a high-fidelity Cas9 nuclease (Cas9HF) to confirm reduced fitness costs probably due to increased cleavage specificity. The data are of interest for CRISPR/Cas9 applications in general and for gene drive applications in particular and the manuscript is of interest to a wide range of readers. https://doi.org/10.7554/eLife.71809.sa0 Decision letter Reviews on Sciety eLife's review process Introduction The ability to make specific edits of genetic material has been a long-standing goal in molecular biology. Until recently, such DNA engineering was cumbersome, expensive, and difficult since it relied on site-specific nucleases or random insertions. CRISPR technology represents a milestone in genome editing because it makes DNA engineering highly efficient, relatively simple to use, and cost-effective through the use of endonucleases that can be flexibly programmed to cut specific sequences dictated by a guide RNA (gRNA; Moon et al., 2019; Mali et al., 2013). The programmability of CRISPR/Cas9 systems allows for numerous potential applications (Moon et al., 2019), including cancer and disease treatment (Chen et al., 2019; Hodges and Conlon, 2019; Jo et al., 2019; Zhang et al., 2020a; Yang et al., 2017), stimuli tracking in living cells (Tang and Liu, 2018), and crop improvement (Zhang et al., 2020b). While most applications of CRISPR use this technology to engineer specific modifications in a given DNA sequence, some proposed applications take the idea one step further by integrating the CRISPR machinery itself into an organism’s genome. In that case, endonuclease activity can continue to produce genetic changes in the cells of the living organism. When present in the germline, these genetic changes might even be passed on to future generations, such as in CRISPR-based gene drives—‘selfish’ genetic elements that are engineered to rapidly spread a desired genetic trait through a population (Esvelt et al., 2014; Champer et al., 2016; Unckless et al., 2017; Noble et al., 2017; Burt, 2014). However, major questions loom large about the technical feasibility of these proposed applications. For example, it remains unclear whether activity of CRISPR endonucleases could entail unintended and potentially harmful consequences in the transgenic organisms, for instance due to the tendency to produce non-specific DNA modifications (so-called ‘off-target effects’; Zhang et al., 2015). Such off-target cleavage could be substantially higher when Cas9 is continuously expressed from a genome and inherited by offspring, where further off-target cleavage can occur. In this study, we seek to address this question in the context of CRISPR gene drive, an emerging technology that could be used for applications ranging from the control of vector-borne diseases to the suppression of invasive species (Esvelt et al., 2014; Unckless et al., 2017; Burt, 2014; Alphey, 2014). One class of CRISPR-based gene drives is the so-called ‘homing drives’. These genetic constructs are programmed to cleave a wild-type sister chromatid and get copied to the target site through homology-directed repair. Since ‘homing’ occurs in the germline, the drive allele will be inherited at a super-Mendelian rate and can thereby spread quickly through the population. The effectiveness of such systems has now been demonstrated in various organisms, including yeast (Roggenkamp et al., 2018; Shapiro et al., 2018; DiCarlo et al., 2015; Basgall et al., 2018), mosquitoes (Hammond et al., 2017; Gantz et al., 2015; Kyrou et al., 2018; Hammond et al., 2016), fruit flies (Oberhofer et al., 2018; Carrami et al., 2018; Gantz and Bier, 2015; Champer et al., 2017; Champer et al., 2018; Champer et al., 2019a; Champer et al., 2019b; Champer et al., 2020b), and mice (Grunwald et al., 2019). Another class of CRISPR gene drives operates by the ‘toxin-antidote’ principle (Champer et al., 2020a). Here, the drive allele serves as the ‘toxin’ by carrying a CRISPR endonuclease programmed to target and disrupt an essential wild-type gene. At the same time, the construct also contains a recoded version of that gene (the ‘antidote’), which is immune to cleavage by the drive. Over time, the drive will continuously increase in frequency and remove wild-type alleles from the population (Burt and Crisanti, 2018). Both homing and toxin-antidote drives can be ‘modification drives’, intended to spread a desired genetic payload through the population (e.g. a gene that prevents mosquitoes from transmitting malaria), or ‘suppression drives’ that seek to diminish or outright eliminate the target population (Champer et al., 2020a; Champer et al., 2021a). A key factor in determining the expected population dynamics of any type of gene drive is the fitness cost imposed by the drive (Wedell et al., 2019). Such fitness costs could come in the form of reduced viability, fecundity, or mating success of the individuals that carry drive alleles. In suppression drives, some fitness costs are typically an intended feature of the drive, necessary to ultimately achieve population suppression. However, these costs are usually recessive to allow the drive to spread to high frequency, and there is generally a limit as to how high other costs can be before the drive will lose its ability to spread effectively (Champer et al., 2020a; Champer et al., 2021a; Champer et al., 2021b; Deredec et al., 2008). For modification drives, fitness costs tend to slow the spread of the drive and can thereby increase the chance that resistance alleles evolve, which could ultimately defeat the drive (Unckless et al., 2017). For such applications, it is therefore desirable to minimize any fitness costs. In drives with frequency-dependent invasion dynamics, such as most CRISPR toxin-antidote systems (Champer et al., 2020b; Oberhofer et al., 2019), fitness costs typically determine the frequency threshold required for the drive to spread through the population (Champer et al., 2020a; Champer et al., 2021a; Champer et al., 2021b). We believe it is useful to distinguish between two types of fitness costs of a gene drive. The first class comprises any costs resulting from the genomic integration of the drive construct (e.g. when this disrupts a functionally important region), costs of potential ‘payload’ genes included in the drive construct, costs resulting directly from the expression of the endonuclease or other drive elements such as gRNAs, and costs due to cleavage of the intended target site. We will call these ‘direct’ fitness costs. By contrast, the second class comprises any potential fitness costs resulting from cleavage and disruption of unintended sites in the genome, so-called ‘off-target’ effects. Despite their critical importance, we still know surprisingly little about the specific types of fitness costs imposed by gene drives. Furthermore, it remains unclear whether there are certain baseline fitness costs that would be difficult to avoid in any gene drive construct, for instance, because they are inherent to the expression and activity of the CRISPR endonuclease. In this study, we conduct a comprehensive assessment of the fitness effects resulting from the genomic expression of CRISPR/Cas9 in experimental Drosophila melanogaster populations. We specifically investigate four different transgenic constructs that allow us to disentangle direct fitness costs from those due to off-target effects. We estimate these fitness costs both through statistical inference from allele frequency trajectories in cage populations using a maximum likelihood approach, and a direct evaluation of individual fitness components using viability, fecundity, and mate choice assays. Results Construct design We designed four constructs to assess the fitness costs of in vivo CRISPR/Cas9 expression in D. melanogaster. As a starting point for our transgenic fly lines, we engineered an Enhanced Green Fluorescent Protein (EGFP) marker driven by the 3xP3 promoter into a gene-free, nonheterochromatic position on chromosome 2 L (region targeted by gRNA: 20,368,542–20,368,561; Figure 1A). This EGFP marker was then used as an insertion point for the four constructs we tested. Our first construct, ‘Cas9_gRNAs’, contains Cas9 expressed by the nanos promoter, the fluorescence marker Discosoma sp Red (DsRed) driven by the 3xP3 promoter, and four gRNAs driven by the U6:3 promoter (Figure 1B), which are separated by tRNAs that are removed after transcription (Champer et al., 2018). The gRNAs of the Cas9_gRNAs construct target a gene-free, nonheterochromatic position on a different chromosome (3 L, region targeted by gRNAs: 18,297,270–18,297,466), preventing any homing activity. In addition to Cas9_gRNAs, three other constructs were designed: ‘Cas9_no-gRNAs’ has a similar architecture as Cas9_gRNAs but lacks the four gRNAs driven by the U6:3 promoter (Figure 1C) ‘no-Cas9_no-gRNAs’ contains neither Cas9 nor the gRNAs but only the fluorescence marker DsRed driven by the 3xP3 promoter (Figure 1D) the last construct, ‘Cas9HF1_gRNAs’ (Figure 1E), has the same architecture as Cas9_gRNAs, except that Cas9 is replaced by a high-fidelity version (Cas9HF1), which has been reported to largely eliminate off-target cleavage (Kleinstiver et al., 2016). We confirmed with PCR-based genotyping that—as expected—all progeny of individuals with the Cas9_gRNAs and Cas9HF1_gRNAs alleles had at least one of their gRNA target sites mutated and that all four gRNAs were similarly active in both these constructs. Figure 1 Download asset Open asset Overview of constructs and the potential types of fitness costs in the four constructs. (A) The starting point for our constructs is an EGFP marker inserted into chromosome 2 L (~20.4 Mb). The four constructs are then inserted into this EGFP locus (thereby disrupting EGFP). (B) The Cas9_gRNAs construct contains Cas9, DsRed, and gRNAs. The gRNAs target chromosome 3 L (~18.3 Mb), instead of the sister chromatid. (C) The Cas9_no-gRNAs construct carries Cas9 and DsRed, but no gRNAs are expressed. (D) The no-Cas9_no-gRNAs construct carries only the fluorescent marker DsRed. (E) The Cas9HF1_gRNAs construct has the same structure as Cas9_gRNAs but carries Cas9HF1 instead of Cas9. The specific designs of these four constructs allow us to identify and disentangle different types of Cas9-related fitness costs. If double-strand breaks at the target site impose fitness costs, such costs should be present for the Cas9_gRNAs and Cas9HF1_gRNAs constructs, but not for the Cas9_no-gRNAs and no-Cas9_no-gRNAs constructs. Cas9_no-gRNAs have no gRNAs expressed to guide Cas9 to the target site and without gRNAs, Cas9 does not cleave DNA (Jinek et al., 2012; Cong et al., 2013). The no-Cas9_no-gRNAs construct neither expresses Cas9 nor the gRNAs. If the expression of Cas9 imposes a fitness cost, all constructs except for no-Cas9_no-gRNAs should incur such a cost, because only this construct does not express Cas9. If off-target effects of Cas9 impose fitness costs, only the Cas9_gRNAs construct should incur them, because the designs of Cas9_no-gRNAs and no-Cas9_no-gRNAs prevent cutting events, and Cas9HF1_gRNAs reportedly have a much lower rate of off-target cleavage (Kleinstiver et al., 2016). Figure 1 summarizes the designs and different potential fitness costs for our four constructs. Population cage experiments To assess the fitness effects of the four constructs, we tracked their population frequencies relative to the baseline EGFP construct (Figure 1A) over several generations in large cage populations by phenotyping the whole population for both dominant fluorescent markers (DsRed and EGFP). Overall, we assessed 13 cages: seven with the Cas9_gRNAs construct, and two each with the Cas9_no-gRNAs, no-Cas9_no-gRNAs, and Cas9HF1_gRNAs constructs (Figure 2). In each cage population, the construct frequency was tracked for at least eight consecutive, non-overlapping generations. The median population size across all experiments was 3602 (Figure 2—figure supplement 1). To avoid potentially confounding maternal fitness effects on the construct frequency dynamics (which could arise based on minor differences in health or age between the initial batches of flies mixed together), we excluded the first generation of five cage populations (Cas9_gRNAs construct: replicates 1, 2, 5, 6, and 7) from the analysis, because their founding individuals (construct homozygotes and EGFP homozygotes) were raised in potentially different environments. Figure 2 with 1 supplement see all Download asset Open asset Construct frequency trajectories in the cage populations. Each line is one cage experiment. To obtain construct frequencies, we screened all adult flies for each generation in the respective cage experiments (see Figure 2—figure supplement 1 for population sizes). The Cas9_gRNAs construct was the only one that systematically decreased in frequency over the course of the experiment (average allele frequency change = –0.11, SEM = 0.03, Figure 2). This provides a first indication that Cas9 off-target effects could be the primary driver of fitness costs. However, the frequency dynamics of the Cas9_gRNAs construct varied widely between individual cage populations (Figure 2). For example, in the two replicates where the construct had the highest starting frequency, its frequency remained approximately constant, whereas it clearly decreased in the other replicates. Maximum likelihood analysis To provide a more quantitative analysis of the fitness costs of the different constructs in our cage populations, we employed a maximum likelihood framework developed for the estimation of selection parameters based on genotype frequency time series data (Liu et al., 2019). We specifically modified the method to support two unlinked autosomal loci, representing the construct and a single idealized off-target site (see the section on ‘Maximum likelihood framework for fitness cost estimation’ in the Methods for a more detailed description of the underlying model). This model can estimate fitness costs with CI while fully accounting for stochastic allele frequency fluctuations due to random genetic drift. Furthermore, we can perform statistical model selection and goodness-of-fit analyses on different selection scenarios to disentangle different types of fitness costs for each construct. General model assumptions Each of the two loci in our model is biallelic (EGFP/construct; uncut/cut off-target site). In individuals that carry a construct, all uncut off-target alleles are assumed to be cut in the germline (i.e. germline cut rate was set to 1), which are then passed on to offspring that could suffer negative fitness consequences. In the early embryo, all uncut off-target alleles are assumed to be cut by maternally deposited Cas9/gRNAs if the mother carries at least one construct allele (i.e. embryo cut rate was also set to 1), changing the individual’s genotype at the off-target site and exposing it to the potential fitness costs associated with this new genotype. Because individuals could carry numerous off-target sites, and the fitness of cleaved alleles could differ vastly between off-target sites, our model of a single off-target site is highly idealized. However, modeling a more complex off-target landscape would require numerous parameters (fitness costs, cut rates, epistatic interactions, etc.) that would be difficult if not impossible to disentangle given our limited number of data points. To reduce model complexity, we therefore limited the model to one off-target locus being always cut in the presence of Cas9. Fitness costs due to carrying the construct and/or the presence of cut off-target sites are assumed to be multiplicative across the two loci, as well as for the two alleles at each locus. We studied models where fitness costs affect only viability, and models where they affect only mate choice and fecundity (both equally). Overall, our maximum likelihood model infers three parameters: the effective population size (Ne) of the cage, the ‘direct fitness estimate’ of the construct (defined as the relative fitness of construct/EGFP heterozygotes versus EGFP/EGFP homozygotes), and the ‘off-target fitness estimate’ (defined as the relative fitness of cut/uncut heterozygotes versus uncut/uncut homozygotes). Note that in our idealized model with a single cleavage site, this site could in principle also represent ‘on-target’ cleavage. However, due to the intergenic location of all gRNA target sites in our constructs, we do not expect such fitness costs to be present. Furthermore, if on-target cleavage had a measurable negative fitness effect, this should have been apparent in the frequency trajectories of the Cas9HF1_gRNAs construct. Since this construct had no apparent reduction in fitness (Figure 2, Table 2), we refer to this fitness parameter exclusively as ‘off-target’. Model evaluation For each construct, five different models with different selection scenarios were studied (Table 1): in the ‘full inference model’, both the construct and cut off-target alleles can impose fitness costs. In the ‘construct’ model, only construct alleles impose a fitness cost. In the ‘off-target’ model, only cut off-target alleles impose a fitness cost. In the ‘initial off-target model’, we assumed that fitness costs originated before the experiment (e.g. through the injection process or perhaps transient maternal effects in the ancestral generation). For the ‘initial off-target model’, the construct homozygotes in the ancestral generation all had cut off-target alleles, but no additional off-target cutting occurred during the experiment (i.e. the germline and embryo cut rate were set to 0). Finally, in the ‘neutral’ model, no fitness costs were present at all. Table 1 Fitness cost model overview. The table shows which types of fitness costs are contained in each model. ModelConstruct alleleCut off-target alleleFull inference++Construct+−Off-target−+Initial off-target−+*Neutral−− * No additional cutting events at off-target sites during the experiment. Inferences were performed on the combined data of the replicated experimental populations for each construct. The individual models were compared using the corrected Akaike information criterion (AICc; Akaike, 1998)—a goodness-of-fit measure that also penalizes for complexity (i.e. number of parameters) in a given model. A lower AICc value indicates a higher quality model. Construct frequency dynamics match a model with moderate off-target fitness costs For the Cas9_gRNAs construct, we found that the full inference model with viability selection yielded the highest quality, with a ‘direct fitness estimate’ of 0.98 and an ‘off-target fitness estimate’ of 0.84 (Table 2). Note, however, that the 95% CI of the direct fitness estimate includes a value of 1. The simpler ‘off-target’ model (where only cut off-target alleles impose a fitness cost) with viability selection and direct fitness estimate set to 1 in fact had an equal AICc value to the ‘full’ model, which further supports that direct fitness costs in construct/EGFP heterozygotes are likely small. Models with fecundity/mate choice selection generally had lower quality than models with viability selection. The ‘initial off-target’ and ‘neutral’ models yielded the highest AICc values. Taken together, these results suggest that among the five different models we tested (Table 1), the observed frequency trajectories of the Cas9_gRNAs construct in our cage populations are best explained by a model where direct effects are less than a few percent and off-target effects impose moderate fitness costs of ~30% (= 1–0.842) in cut/cut homozygotes in our idealized single off-target site model (Table 2). Table 2 Model comparison and parameter estimates for Cas9_gRNAs. ModelSelectionN^eDirect fitness estimateOff-target fitness estimatelnL^PAICcFullViability175[140–215]0.98[0.95–1.00]0.84[0.77–0.91]384.73–763FullMate choice = fecundity163[131–200]0.96[0.94–0.98]1.00[0.95–1.06]378.83–751ConstructViability164[131–201]0.96[0.93–0.98]1*378.92–754ConstructMate choice = fecundity163[131–200]0.96[0.94–0.98]1*378.82–754Off-targetViability173[139–212]1*0.80[0.74–0.88]383.62–763Off-targetMate choice = fecundity157[126–192]1*0.95[0.90–1.01]375.12–746Initial off-targetViability156[125–191]1*0.92[0.82–1.02]374.82–745Initial off-targetMate choice = fecundity156[125–191]1*0.96[0.91–1.01]374.82–745NeutralNone154[123–189]1*1*373.61–745 Each row shows the parameter estimates for effective population size (N^e), maximum log likelihood (lnL^) , number of free parameters in the maximum likelihood framework (P), and corrected Akaike information criterion value (AICc=2p−2lnL^+(2p2+2p)/(n−p−1)) where n=87 is the number of generation transitions for a specific model and selection type. 1* entries indicate that a parameter was fixed at 1 (= no fitness effect is estimated). Values in squared brackets in the parameter estimate columns represent the 95% CI estimated from a likelihood ratio test with one degree of freedom. The model with the lowest AICc (i.e. the best fit) is highlighted in bold. A scenario in which fitness costs are primarily due to off-target effects also suggests a possible mechanism for why the decline in the frequency of the Cas9_gRNA construct could be greater in cages of lower construct starting frequencies, which appears to be the case in our experiment (Figure 2). This mechanism would work due to the accumulation of previously cut off-target sites that should typically be protected from future cutting due to sequence mutations caused by the repair process, similar to the creation of resistance alleles in a homing drive (Champer et al., 2017). Early in the experiment, cut off-target sites should be found primarily in individuals that also carry a construct allele. Fitness costs resulting from such cuts will therefore also impose negative selection against construct alleles. However, as mutated off-target sites accumulate over the course of an experiment, they will increasingly segregate independently from construct alleles, thereby reducing selection against these alleles. By the time all potential off-target sites in the population have been mutated, construct alleles would no longer experience any negative selection if off-target effects are indeed the only cause of fitness costs. Importantly, cages where the construct is introduced at a higher frequency (e.g. Cas9_gRNAs replicate 6 in Figure 2) should experience this effect faster than cages where it is introduced at a lower frequency (e.g. Cas9_gRNAs replicate 3 in Figure 2) due to the higher overall rate of cleavage events in the population. To test how well our best-fitting model from Table 1 (full inference model with viability selection) can resemble the observed frequency-dependent construct dynamics of the Cas9_gRNAs construct, we simulated construct trajectories under its maximum likelihood parameter estimates (N^e=175, direct fitness estimate = 0.98, off-target fitness estimate = 0.84). We found that the simulated genotype frequencies not only closely resemble the observed decrease in construct frequency but also capture the heterogeneity in frequency trajectories observed among individual replicates (Figure 3). Additionally, we compared simulated trajectories for this model with simulated trajectories from the ‘construct’ model that only considers direct fitness costs (Figure 3—figure supplement 1). We found that the full inference model captures the observed frequency-dependent construct dynamics better than this model, with most of the improvement due to better matching trajectories from cages with low starting frequencies, where off-target effects would be expected to have a more drastic impact on the relative fitness of construct-carrying individuals. Figure 3 with 2 supplements see all Download asset Open asset Comparison of observed Cas9_gRNAs construct frequencies with simulated trajectories of the full model with viability selection under its maximum likelihood parameter estimates (N^e=175, direct fitness estimate = 0.98, off-target fitness estimate = 0.84). Solid red lines present observed construct frequencies, black lines show 10 simulated trajectories for each cage, and the shaded area represents the range between the 2.5 and 97.5 percentile of the simulated trajectories (10,000 simulations per cage). To further support our hypothesis that off-target effects may be the primary driver of fitness costs, we applied our maximum likelihood inference framework to the experimental cage data of the three other constructs (Cas9_no-gRNAs, no-Cas9_no-gRNAs, and Cas9HF1_gRNAs). Because none of these three constructs, by design, should be capable of producing substantial amounts of off-target cuts, we set the germline and embryo cut rate to 0 (i.e. no off-target alleles are cut in the presence of a construct or maternally deposited Cas9/gRNAs) and inferred viability fitness effects for the construct. Except for the ‘initial off-target’ model (i.e. the model in which fitness costs originated before the experiment), construct homozygotes of the ancestral population were assumed to not carry any cut off-target alleles. For Cas9_no-gRNAs, and no-Cas9_no-gRNAs, the ‘neutral’ model without any fitness costs explained the observed construct frequency trajectories best (Table 3, Figure 3—figure supplement 2), further corroborating the notion that Cas9 fitness costs in our experimental populations may be primarily due to off-target cuts (Figure 1). However, the construct frequency dynamics of Cas9HF1_gRNAs are best explained by an ‘initial off-target’ model, where cut off-target alleles are beneficial, closely followed by the neutral model (Table 3). While we cannot completely rule out that the initial construct homozygotes of Cas9HF1_gRNAs may have had some fitness advantage due to cut off-target alleles or transgenerational beneficial effects, the 95% CI for the off-target fitness parameter still includes a fitness value of 1 (i.e. no fitness effects). A putative short-term fitness advantage could also be explained by maternal effects that persisted for the first 2–3 generations. Although we do not anticipate that any other construct than Cas9_gRNAs can produce substantial off-target effects (Figure 1), we repeated the analysis of the three other constructs with cut rates set to 1 (i.e. off-target alleles are always cut in the presence of a construct or maternally deposited Cas9/gRNAs) and inferred viability selection, which yielded similar results (Supplementary file 1). Table 3 Model comparison and parameter estimates for Cas9_no-gRNAs, no-Cas9_no-gRNAs, and Cas9HF1_gRNAs. ConstructModelSelectionN^eDirect fitness estimateOff-target fitness estimatelnL^PAICcCas9_no-gRNAsConstructViability243[152–366]1.0[0.96–1.04]1*88.62–173Cas9_no-gRNAsInitial off-targetViability250[156–377]1*0.84[0.65–1.18]89.22–174Cas9_no-gRNAsNeutralNone243[152–366]1*1*88.61–175no-Cas9_no-gRNAsConstructViability162[101–243]1.0[0.97–1.10]1*81.52–158no-Cas9_no-gRNAsInitial off-targetViability162[101–243