In vitro protection of Cas9 off-targets with CRISPR GUARD

We tested the concept of CRISPR GUARD by measuring the binding kinetics of a perfectly complementary 15-nt GUARD RNA versus a mismatched gRNA to an immobilised DNA off-target template for a gRNA targeting VEGFA with four mismatches21,22,23. Bio-layer interferometry (BLI) revealed comparable off-target association kinetics for catalytically inactive Cas9 complexed with GUARD RNA or mismatched gRNA, with 29 ± 4% slower binding for the GUARD RNA (Fig. 2a), suggesting that competition at off-target loci is feasible. To test the possibility of DNA protection by CRISPR GUARD, we selected an array of potential GUARD RNA designs for in vitro Cas9 DNA cleavage assays. We considered GUARD RNA lengths of 14-nt or 15-nt, or a full length 20-nt design with only 15-nt of target complementarity (15-nt+ spacer). GUARD RNAs were designed as competitive molecules that are truncated versions of the on-target gRNA but incorporate mismatches found at the off-target site (Fig. 2b). Alternatively, we designed proximal GUARD RNAs that bind flanking regions of the off-target site, speculating that the reduced sequence homology would reduce competition at the on-target site, but still block the protospacer and protospacer adjacent motif (PAM). We termed these designs as competitive and non-competitive GUARD RNAs, respectively (Fig. 2b). Firstly, we confirmed that GUARD RNAs cannot direct nuclease activity by in vitro assessment of on-target and off-target cleavage using short purified DNA targets. None of the GUARD RNAs directed DNA cleavage when complexed with Cas9 alone (Fig. 2c). Cas9 complexed with VEGFA gRNA showed robust cleavage of both on-target, and off-target CAVIN4 DNA. However, addition of non-competitive GUARD RNAs failed to protect the off-target site. Significant protection was only observed using a 15-nt competitive GUARD RNA, but on-target cleavage was also impaired suggesting competition at both sites (Fig. 2c). Competition with the on-target gRNA may be particularly pertinent when there are very few mismatches, or when the mismatches of the gRNA cannot be incorporated in the truncated GUARD RNA design.

Fig. 2: GUARD RNAs reduce Cas9 off-target cleavage activity without affecting on-target editing.


a Bio-layer interferometry (BLI) analysis of binding kinetics for dead Cas9 ribonucleoprotein complex to an immobilised biotinylated DNA substrate. Cas9 was precomplexed with tracrRNA and (NT) non-targeting control gRNA; (gRNA) 20-nt VEGFA gRNA with 4 mismatches; (GUARD RNA) a 15-nt GUARD RNA targeting the off-target site or no guide RNA. b CRISPR GUARD design. Schematic showing GUARD RNAs designs. Variation in length (14-nt versus 15-nt) and protospacer positioning likely influence binding energy and competition with the on-target gRNA (i.e., competitive/overlapping, non-competitive/proximal). Mismatches in the gRNA are shown in red. RNA bases overlapping the gRNA PAM are shown in bold. c Competition between on-target gRNA and GUARD RNA can be reduced by proximal positioning. On-target (VEGFA) and off-target (CAVIN4) DNA abundance was measured by qPCR following an in vitro Cas9 DNA cleavage assay. The corresponding GUARD RNA was added at a molar ratio of 5:1 to the VEGFA on-target gRNA (100 nM to 20 nM). d GUARD RNAs are more effective at blocking Cas9-mediated DNA cutting with pre-incubation in vitro. On-target (VEGFA) and off-target (CAVIN4) DNA abundance was measured by qPCR following an in vitro Cas9 DNA cleavage assay, with a 30 min pre-incubation with GUARD RNAs and Cas9, before addition of the on-target gRNA and Cas9. Increasing molar ratios of GUARD RNA to the VEGFA on-target gRNA were used (10:10, 20:10, 50:10, 100:10 nM). e GUARD RNAs are effective in blocking Cas9-mediated DNA cleavage of off-target sites within a 10 bp window of the gRNA PAM. Synthetic DNA fragments containing the VEGFA gRNA off-target site CAVIN4 positioned progressively further away from the GUARD RNA binding site were quantified by qPCR following an in vitro Cas9 DNA cleavage assay (50:10 nM GUARD RNA to gRNA ratio). f CRISPR GUARD is effective at blocking off-target editing in cells. Indel rates from NGS of amplicons from Cas9-expressing HEK293 cells transfected with VEGFA gRNA and multiple GUARD RNA designs for protection of the CAVIN4 proximal off-target site (25:25 nM GUARD RNA to gRNA ratio). Data represent the mean of two (f) independent experiments, or the mean ± SD of three (c, e) or five (d) independent experiments with symbols representing each replicate. For a, data are representative of two independent experiments. NT non-targeting gRNA, UTC untransfected control. Unpaired, two-tailed student’s t-test. For c, *P = 0.0208, **P = 0.0083. For d, *P = 0.0164, **P = 0.0021 (VEGFA) and **P = 0.007 (CAVIN4), ***P = 0.0003, ***P < 0.0001. For e, **P = 0.0041 and ***P = 0.0001. Source data are provided as a Source Data file.

Cas9-gRNA complexes show rapid target binding and cleavage in vitro24. The failure of GUARD RNAs to protect off-target sites could therefore be due to these rapid kinetics. In line with this, a 30-min pre-incubation of Cas9 and a 14-nt non-competitive GUARD RNA with the DNA substrate resulted in robust protection of the off-target DNA, even at a 1:1 ratio of GUARD RNA to gRNA (Fig. 2d). Further increasing the ratio of GUARD RNA to a five-fold excess led to complete protection from cleavage. In addition, on-target cleavage could be completely prevented using a competitive 14-nt GUARD RNA against VEGFA, demonstrating that GUARD RNAs can compete against perfectly complementary gRNAs (Fig. 2d). On-target protection of VEGFA required a higher ratio of GUARD RNA to gRNA, presumably due to the higher binding affinity of the on-target gRNA.

To better understand GUARD RNA positioning rules, we designed an in vitro Cas9 DNA cleavage assay for the VEGFA off-target site CAVIN4, whereby the same non-competitive GUARD RNA binding site was positioned incrementally further away within a synthetic DNA fragment (Supplementary Fig. 1). After incubation with Cas9 complexes, the remaining uncleaved DNA was quantified by qPCR. In this way, we could determine the optimal distance between the GUARD RNA and the off-target region for protection from Cas9 nuclease. As expected, the 14-nt non-competitive GUARD RNA against the CAVIN4 off-target site provided significant protection when it was overlapping the off-target site, but also proved similarly effective when placed 10 bp away from the gRNA PAM (Fig. 2e). However, GUARD RNAs positioned 25 and 50 bp distal to the off-target gRNA PAM on either the 5’ or 3’ flank provided no protection (Fig. 2e), implying that GUARD RNAs are effective in rendering an off-target region inaccessible for editing if positioned ≤10 bp away from the gRNA PAM.

Reduced off-target editing without affecting on-target editing

Next, we tested the cellular activity of CRISPR GUARD. Using a Cas9-expressing HEK293 cell line25, we co-transfected a VEGFA gRNA and multiple GUARD RNA designs (Fig. 2b) at equimolar ratios for protection of the CAVIN4 proximal off-target site and performed next-generation sequencing (NGS) of amplicons to detect insertion and deletions (indels). Strikingly, all tested GUARD RNA designs significantly protected the off-target site (Fig. 2f). Notably, none of the GUARD RNA designs significantly interfered with on-target cutting efficiency, including the competitive GUARD RNA. Furthermore, unlike the in vitro scenario, phased delivery of the GUARD RNA before the gRNA provided no additional benefit (Supplementary Fig. 2). We reasoned that, in contrast to the in vitro setting, the kinetics of Cas9 binding to the off-target region is slower due to scanning of the mammalian genome, thus concomitant delivery of GUARD RNA is sufficient to allow for effective competition at off-target sites and may also reduce the effects of direct competition at the on-target locus with competitive GUARD RNAs. Due to the apparent efficacy of the 14-nt and 15-nt GUARDs in vitro and in cells, we disregarded the 15-nt GUARD RNA with a 5-nt mismatched spacer sequence (Fig. 2b), as this provided the least protection (Fig. 2f) and could potentially give rise to a catalytically active Cas9 complex if the GUARD RNA binds to a DNA site where the mismatched spacer has >2-nt of base pairing.

Next, we investigated the optimal dosing for CRISPR GUARD. We co-transfected a gRNA against VEGFA and increasing concentrations of GUARD RNA to protect the CAVIN4 proximal off-target site. Although an equimolar ratio of GUARD RNA to gRNA was effective in reducing off-target indel rates (Fig. 2f), higher concentrations of GUARD RNA provided additional reduction of off-target editing (Fig. 3a). Notably, the highest concentration of GUARD RNA (10:1 ratio) reduced off-target indel rates to background levels (Fig. 3a). We noted that indel rates in untransfected control cells for the CAVIN4 proximal region are approximately 0.3%, owing to intrinsic error in the NGS of a polymeric cytosine tract at this locus. We assume that Cas9 concentration is in excess in the cell as increasing the ratio of GUARD RNA to gRNA did not have a negative impact on on-target editing, so in this context we recommend routinely adopting a 5:1 ratio of GUARD RNA to gRNA.

Fig. 3: Protection of multiple, endogenous off-target sites with CRISPR GUARD.


a Increasing GUARD RNA concentration leads to complete protection of Cas9 off-target sites in cells. Indel rates from NGS of amplicons from Cas9-expressing HEK293 cells transfected with VEGFA gRNA and increasing concentrations of 14-nt GUARD RNA for protection of the CAVIN4 off-target site (10:10, 20:10, 50:10, 100:10 nM). The total concentration of delivered RNA in each case was kept constant by co-delivery of NT gRNA (first condition is 100 nM NT). b Indel rates from amplicons sequencing of Cas9-expressing HEK293 cells transfected with EMX1 or VEGFA on-target gRNAs, and a single GUARD RNA (25:25 nM ratio). GUARD RNA positioning relative to the off-target protospacer sequence is shown schematically in each panel. Data represent the mean of two independent experiments with symbols representing each replicate. Source data are provided as a Source Data file.

To determine if CRISPR GUARD was also effective in the therapeutically relevant setting of Cas9 ribonucleoprotein (RNP) delivery, we separately precomplexed catalytically inactive Cas9 (dCas9) protein with a GUARD RNA against VEGFA off-target site TENT4A, and wild-type Cas9 protein with the VEGFA on-target guide RNA, to co-deliver these RNPs for protection and cutting, respectively. In principle, it is an option to use a full-length 20-nt gRNA for protection with dCas9. However, we continued to use 14-nt GUARD RNAs in this scenario in order to prevent guide swapping in the cell26, whereby wild-type Cas9 acquires full-length gRNA against the off-target site. Encouragingly, CRISPR GUARD using RNPs successfully protected VEGFA off-target TENT4A, and there was no detectable impact of protection of the TENT4A off-target site on the other known off-target sites by redistribution of active Cas9 (Supplementary Fig. 3).

To begin to understand GUARD RNA design rules, we assessed the activity of an array of GUARD RNAs protecting known endogenous off-target sites of EMX1 and VEGFA gRNAs21,22,23,27, assessing nucleotide composition and positioning. We used equimolar GUARD RNA and gRNA in order discern off-target protection that would be masked at higher GUARD RNA concentrations (Fig. 3a). Notably, both MYC and CAVIN4 GUARDs worked exceptionally well, reducing indel rates from 2.6 ± 0.3% to 0.08 ± 0.08%, and 19 ± 1.9% to 3.3 ± 0.4%, respectively (Fig. 3b). All GUARD RNAs tested reduced off-target indel rates except for one (VEGFA off-target HDLBP). Upon inspection of the non-functional HDLBP GUARD RNA, we noted that it had relatively low GC-content compared to the cognate on-target gRNA, potentially leading to low binding affinity and reduced ability to compete at the off-target locus. Introduction of a revised GUARD RNA with higher GC-content, length and increased overlap of the gRNA seed region, achieved only modestly improved protection (Supplementary Fig. 4). We only observed protection at a 5:1 molar ratio of GUARD RNA to gRNA, and no protection at a 1:1 ratio. Taken together, these data suggest that some off-target loci are more amenable to protection by CRISPR GUARD than others, indicating more systematic investigation is warranted. This is likely contingent on the relative affinities of the GUARD RNA and the cognate gRNA for the off-target region.

As protection of one off-target with CRISPR GUARD apparently does not increase editing at other off-target sites (Supplementary Fig. 3), we attempted multiplexing of three GUARD RNAs together to assess if it was possible to protect multiple off-target sites simultaneously. We performed multiplexed GUARD RNA experiments by transfecting Cas9 expressing HEK293 cells with on-target gRNA and three independent GUARD RNAs all at equimolar concentrations. For both EMX1 and VEGFA, multiplex GUARD RNA delivery was generally effective in protecting multiple off-target sites from indel formation (Supplementary Fig. 5). Consistent with the in vitro data (Fig. 2b), the non-competitive GUARD RNAs used in these experiments showed no effect at the on-target site (Supplementary Fig. 5), thus allowing robust protection of several off-targets whilst maintaining efficient on-target editing.

To further validate the safety of CRISPR GUARD, we transfected various functional GUARDs (14-nt and 15-nt in length) at the highest effective concentration used in this study (125 nM) and assessed if they could support Cas9-mediated DNA cleavage in cells by deep sequencing of amplicons. In concordance with our in vitro data (Fig. 2b, c), none of the GUARD RNAs could generate Cas9-induced indels in isolation, in contrast to a 20-nt control gRNA at the same concentration (~82% indel formation), affirming that GUARD RNAs form nuclease-dead complexes with Cas9 in the cell (Supplementary Table 1).

CRISPR GUARD for base editing

Next, we applied CRISPR GUARD to base editing. Due to the strict positioning requirements for base editing activity, gRNA design possibilities are more limited and therefore off-targets are harder to avoid28,29. We reasoned that CRISPR GUARD could reduce Base Editor 3 (BE3) activity at off-target sites, since the activity window of BE3 is mostly absent from GUARD RNAs (Supplementary Fig. 6a), and optimal activity of BE3 is dependent on nickase Cas9 nuclease acitivity28, which is compromised with short gRNAs18,19,20. Using a HEK293 cell line expressing BE3, we tested GUARD RNA designs for protection of EMX1 and VEGFA off-target sites (Supplementary Fig. 6b). We demonstrated significant protection of the EMX1 off-target cytosines proximal to MYC (a reduction from 6.03 ± 0.6% to 0.98 ± 0.09% editing), and the VEGFA off-target cytosines proximal to CAVIN4 (a reduction of 23.33 ± 0.45% to 11.37 ± 2.72%; Fig. 4a). These GUARD RNAs also performed well in the CRISPR/Cas9 system (Fig. 3b). As with Cas9, introduction of GUARD RNAs did not compromise on-target editing efficiencies (Fig. 4a).

Fig. 4: GUARD RNAs can reduce off-target base editing if designed to avoid cytosine exposure.


a Protection from off-target cytosine deamination with CRISPR GUARD. Base editing rates from amplicon sequencing of BE3-expressing HEK293 cells transfected with EMX1 or VEGFA on-target gRNA (25 nM) and the indicated GUARD RNA (125 nM). Shown is the cumulative editing rate for the cytosines highlighted. The predicted cytosine deamination activity window is indicated. b Increased base editing frequencies with GUARD RNAs that expose cytosines. Base editing rates from amplicon sequencing of BE3-expressing HEK293 cells transfected with EMX1 (left) or VEGFA (right) on-target gRNA (25 nM) and the indicated GUARD RNA (125 nM). c GUARD RNAs can facilitate base editing without full-length on-target gRNAs. Base editing rates from amplicon sequencing of BE3-expressing HEK293 cells transfected with MFAP1 GUARD RNA only (125 nM). Data represent the mean of two independent experiments with symbols representing each replicate. Source data are provided as a Source Data file.

To date, the activity of short gRNAs for base editing applications has not been systematically investigated. The HDBLP and MFAP1 GUARD RNAs are predicted to expose cytosines as single-stranded DNA near the BE3 deaminase activity window (Supplementary Fig. 6a and 6b). Interestingly, when BE3 expressing cells were transfected with both on-target gRNA and GUARD RNA, an overall increase in deamination rates was observed (Fig. 4b). Strikingly, the HDBLP GUARD RNA introduced two distinct G-to-A mutations that were not detected in the absence of GUARD RNA (Fig. 4b). These mutations were found both linked and unlinked to those generated by the on-target gRNA, suggesting they could be occurring in the same editing event (Supplementary Fig. 6c). Moreover, when we transfected the MFAP1 GUARD RNA alone (without EMX1 on-target gRNA), we could detect a low but significant number of C-to-T deamination events (Fig. 4c; 3.44 ± 0.40%), suggesting that short GUARD RNAs are sufficient to independently support base editing in cells. It is likely that the low level of base editing observed with short GUARD RNAs alone is due to the lack of nickase activity, which is analogous to editing with base editor 2 versions28.

A high-throughput screen identifies functional GUARD RNAs

Thus far, we have analysed the performance of a relatively small number of GUARD RNAs targeting endogenous loci. To systematically screen all possible Cas9 or BE3 GUARD RNAs for a given target, we adapted a high-throughput approach30,31 using a library of lentiviral constructs that express both gRNA and GUARD RNA and harbour the off-target sequence within 79 bp of its genomic context. Using this system, we screened the performance of ~600 GUARD RNAs including NT GUARD RNA controls, obviating amplification of individual endogenous loci to analyse editing events (Fig. 5a). We tested four gRNAs based on their therapeutic relevance (HBB; a therapeutic target for correcting sickle-cell disease14) or because of an extensive knowledge of experimentally validated off-target sites (EMX1, FANCF, and HEKsite1). Each gRNA expression vector was linked to one of 30 experimentally validated off-target sites2. For each of the off-target sites, we tested all possible 15-nt GUARD RNAs within a 10 bp window either side of the mismatched gRNA protospacer with an NRG PAM (where R is a purine). We introduced the lentiviral pooled library into HEK293 cells expressing doxycycline-inducible Cas9 or BE3 and measured the number of indels or base edits for each GUARD RNA by NGS (Fig. 5a).

Fig. 5: High-throughput screening identifies functional GUARD RNAs.


a Schematic of the lentiviral pooled screening approach to analyse off-target editing frequencies in cells expressing different GUARD RNA species. Numbers indicate the length of the plasmid segment in base pairs. b Scatter plots of indels (Cas9) or SNPs (base edits) for two independent replicate experiments with or without induction of Cas9 or BE3 expression with doxycycline for 48 h. c Box and whiskers plot of off-target indel or base editing frequency for coding and non-targeting (NT) GUARD RNAs in the Cas9 or base editor screen, respectively. NB: some outliers fall outside of the y-axis limit. d GUARD RNAs binding to the same DNA strand as the gRNA are more effective at blocking Cas9 off-target editing but can increase base editing efficiency. Box and whiskers plot of off-target indel or SNP frequencies in the screens. NB: some outliers fall outside of the y-axis limit. Box and whiskers plot: centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; points, outliers. Data were compared using unpaired, two-tailed student’s t-tests.

As expected, we observed a significant induction of editing at off-target sites by Cas9 and BE3 in the presence of doxycycline with a high correlation of editing frequencies between biological replicates (Fig. 5b and Supplementary Fig. 7). Editing profiles were compared to those derived from deep sequencing of the original plasmid library, such that we focussed on mutations caused by Cas9 or BE3 expression in the cell in downstream analyses. Low frequency editing was observed without doxycycline, likely reflecting a degree of leakiness in the inducible system. To further validate the screen, we compared the Cas9-driven indel rates in the screen with those detected with GUIDE-seq methodologies on endogenous genomic loci2 and found that the observed mutation rates correlated well (Supplementary Fig. 8), verifying the physiological relevance of the system30,31.

Comparing the abundance of Cas9-driven off-target indels, we observed that expression of GUARD RNAs tended to reduce off-target mutations compared to cells expressing NT GUARD molecules (Fig. 5c). It was also clear that many of the screened GUARD RNAs have minimal effects, which is perhaps to be expected as we did not pre-select GUARD RNAs based on sequence features and many of these GUARD molecules may perform better at higher concentrations relative to the gRNA (Fig. 3a).

Conversely, we discovered that GUARD RNAs significantly increased base editing at off-target sites on average when compared to NT GUARD RNA controls (Fig. 5c), consistent with some GUARD RNAs being able to support base editing (Fig. 4). BE3 had a higher rate of off-target editing than Cas9 (Fig. 5c), with most SNPs detected being C->T and G->A mutations, consistent with cytosine deamination (Supplementary Fig. 9). Notably, base editing with GUARD RNAs was promoted by binding to the same DNA strand as the gRNA (Fig. 5d), presumably because the full-length gRNA-BE3 complex nicks the opposite strand, thus encouraging repair of the unedited strand and retention of the mutation28. In contrast, Cas9 GUARD RNAs binding on the same DNA strand as the gRNA significantly reduced indel rates (Fig. 5d), implying that this may be sterically more obstructive.

Sequence features of functional GUARD RNAs

To delineate features of protective GUARD RNAs, we designated GUARD RNAs that reduced off-target editing by more than two standard deviations from the mean of the NT GUARD RNA controls as functional molecules (Figs. 6a, 6b and Supplementary Fig. 10). We selected a pair of reproducibly functional and non-functional Cas9 GUARD RNAs and performed validation experiments on the endogenous genomic locus. Results from these experiments were in direct concordance with the screening data supporting the validity of this categorisation (Fig. 6c).

Fig. 6: GUARD RNA features that significantly reduce off-target editing.


a GUARD RNA performance for Cas9 and, (b) BE3. Base editing rates for BE (% of NGS reads with SNPs) or indel rates for Cas9 (% of NGS reads with indels) for a subset of off-target sites. Off-targets are labelled with an identifier, gene name (if genic), gRNA name, and the number of mismatches between the gRNA and the off-target site. Data are pooled from two independent experiments from cells treated with doxycycline. Each point represents a GUARD RNA and are compared to three non-targeting (NT) control RNAs. GUARD RNAs are classed as functional if they reduce off-target editing by at least two standard deviations from the mean of the NT controls. Data shown here is a subset of that presented in Supplementary Fig. 10. c Lentiviral plasmid screening reveals GUARD RNAs that are effective at endogenous genomic loci. (Left panel) indel rates from GUARD RNA screening of EMX1 gRNA off-target site MFAP1, highlighting non-functional GUARD RNA (v1) and functional GUARD RNA (v2). Screening data shown here is a subset of that presented in Supplementary Fig. 10. (Right panels) indel rates from amplicon sequencing of on-target and off-target endogenous loci from Cas9-expressing HEK293 cells transfected with EMX1 gRNA and GUARD RNA v1, GUARD RNA v2 or a NT control (all at 25 nM). Data represent mean of two independent experiments with symbols representing each replicate. d GUARD RNAs can mediate base editing at position 1 and 2 within the 15-nt GUARD RNA. GUARD RNAs were grouped according to whether they had a cytosine at each nucleotide position or not, and P values were generated by comparing the % of NGS reads with SNPs between GUARD RNA groups with an unpaired, two-tailed student’s t-tests. e Functional GUARD RNAs predominantly have NGG PAMs. The proportion of functional GUARD RNAs from the Cas9 screen that had NGG PAMs versus NAG PAMs was compared using a two-sided Fisher’s exact test. f Model depicting important parameters for GUARD RNA design. For GUARD RNAs that reduce Cas9 off-target editing, reducing distance from the gRNA, having an NGG PAM, binding to the same DNA strand as the gRNA, increased GC-content and high concentration in the cell will all increase the effectiveness of CRISPR GUARD. For BE3, the presence of cytosines at position 1 or 2 is to be avoided for 15-nt GUARD RNAs to prevent editing. Source data are provided as a Source Data file.

Half of the analysed off-target sites (15 of 30) had functional GUARD RNAs for Cas9, and 80% (24 of 30) had functional GUARD RNAs for BE3 (Supplementary Fig. 10). Functional GUARD RNAs were evenly distributed between off-targets with low and high mutation rates and for off-targets with different numbers of gRNA mismatches (Fig. 6a). Of the 510 targeting GUARD RNA designs tested, 18 ± 1% showed functional protection from Cas9 off-target activity, with 23 ± 0.3% of these reducing off-target editing to below 0.5%. For BE3, fewer GUARD RNAs showed functionality, with 8 ± 1% showing reproducible protection of off-target sites (Fig. 6b). Only one GUARD RNA was able to reduce BE3 editing to below 0.5%, consistent with the finding that some GUARD RNAs can also support base editing. Specifically, the presence of a cytosine within the first 2-nt of the 15-nt GUARD RNA gave rise to significantly more base editing events (Fig. 6d). Concordant with GUARD RNAs with high affinity having superior protective effects, NGG PAMs were significantly enriched in functional GUARD RNAs over NAG (Fig. 6e). Selecting only Cas9 GUARD RNAs with an NGG PAM significantly increased the percentage of functional Cas9 GUARD RNAs to 26 ± 1%. Moreover, the proportion of functional GUARD RNAs tended to increase with higher GC-content (Supplementary Fig. 11a). Finally, GUARD RNAs with a higher degree of spatial overlap with the gRNA, especially at the seed and PAM region, led to superior off-target protection (Supplementary Fig. 11b). Taken together, these data support a model of off-target protection by CRISPR GUARD through direct competition with the mismatched gRNA and highlight important parameters for GUARD RNA design (Fig. 6f).

Finally, we generated a publicly available tool for automated GUARD RNA design called CRISPR GUARD Finder ( and, which predicts potential off-target sites for a given gRNA, and generates a list of possible GUARD RNAs with relevant sequence features. An example is given in Supplementary Fig. 12.