Genes encoding MYB transcription factors in tomato were obtained from the Plant Transcription Factor Database Expression (Jin et al., 2017). A total of 29 MYB genes was selected for sgRNA design (Supplementary Dataset S1). In the selected genes, sgRNAs and potential off-target sites with a maximum of 4 mismatches were predicted using CRISPOR (Concordet and Haeussler, 2018). Potential off-target sites with DNA or RNA bulges compared to the target site are not predicted by CRISPOR and were therefore predicted using Cas-OFFinder (Bae, Park, and Kim, 2014).

The backbone of the vectors for plasmid transfections was constructed using Golden Gate cloning and parts from the MoClo Toolkit, a gift from Sylveste Marillonnet (AddGene kit #1000000044, and the MoClo Plant Parts kit, a gift from Nicola Patron (Addgene kit #1000000047) (Engler, Kandzia, and Marillonnet, 2008; Weber et al., 2011; Engler et al., 2014). A Level 1 vector for easy cloning of sgRNAs behind an Arabidopsis thaliana U6-26 promoter was a gift from Marc Youles (The Sainsbury Laboratory). This vector contains an operon expressing an RFP protein that can be replaced by a spacer in a BsmBI-mediated Golden Gate cut/ligate reaction, allowing pink/white screening of colonies that have successfully integrated the sgRNA (“CRISPRpink.”) To allow its use in a Level 2 vector, the BsmBI sites were replaced by BsaI sites using restriction/ligation. Level 1 constructs pICH4773::proNOS-NPTII-tOCS, pICH4772::2xproCaMV35-Cas9-tNOS, pICH47751::2xproCaMV35S-tGFP-tCaMV35S, pICH54044, pICH54055, the adapted pICH47781::pAtU6-26-CRISPRpink sgRNA expression vector and end-linker pICH41822 were then combined into Level 2 vector pICSL4723.

Subsequently, sgRNA transcription units were completed by replacing the CRISPRpink operon with annealed oligos encoding the spacer in a Golden Gate cut/ligate reaction using BsaI. Oligonucleotide sequences used can be found in Supplementary Dataset S2.

Highly pure DNA for transfection was prepared from 3 mL of overnight E. coli culture in LB medium using the PureYield Plasmid MiniPrep System (Promega), with the following adaptations: bacterial pellets were frozen at −20°C before processing to increase DNA yield, the column was washed twice with the endotoxin removal wash to acquire the desired purity, and plasmid DNA was eluted with 30 µL elution buffer preheated at 60°C.

Tomato (Solanum lycopersicum) cv. “Moneyberg” plants were grown in vitro on hormone-free ½ concentration Murashige and Skoog medium (Duchefa) with 10 g/L sucrose and 8 g/L plant agar (Duchefa), pH 5.8, at 25°C, under a 16 h photoperiod. In vitro grown plants were used for protoplast isolation at 10–12 weeks after sowing. Reagents for protoplast isolation and transfection were prepared according to Sheen (Sheen, 2002). Young, fully developed leaves were cut in a feather-like pattern and incubated overnight in the dark with the abaxial side down in 10 mL digestion solution (0.4 M mannitol, 20 mM MES, 20 mM KCl, 10 mM CaCl₂, 1% cellulase Onozuka R10 (Duchefa) and 0.25% macerozyme R10 (Duchefa); pH 5.7). The next day, the suspension was gently swirled to release protoplasts. The suspension was sieved through a 100 μm cell sieve into a 50 mL tube. The remaining leaf material was washed with 10 mL W5 solution (154 mM NaCl, 125 mM CaCl₂, 5 mM KCl, 2 mM MES, pH 5.7) to release additional protoplasts and the suspensions were pooled. The pooled suspension was subsequently centrifuged for 3 min at ×100 g and the pellet was resuspended in 10 mL W5. This washing step was repeated twice, and finally pelleted protoplasts were resuspended in 10 mL magnesium-mannitol solution (0.4 M mannitol, 15 mM MgCl₂, 4 mM MES, pH 5.7) to prepare for transfection. Protoplast density was determined by counting in a haemocytometer and adjusted to 5.0 × 10⁵/mL.

Protoplasts were transfected in 96 well plates, as described by Wehner et al. (2011), with the following modifications. For each replicate, 5 µL of plasmid DNA (2–3 µg per construct) was transferred to a well of a V-bottom microtiter plate and 30 µL protoplast suspension (containing 1.5 × 104 protoplasts) was added. Transfections were started by adding 35 µL polyethylene glycol solution (40% PEG-4000 (Fluka), 0.2 M mannitol, 0.1 M CaCl2) and mixing by pipetting up and down. After a 10 min incubation, 120 µL W5 solution was added and mixed to stop transfection. Protoplasts were pelleted by centrifuging for 3 min at ×200 g. The pellet was resuspended in 100 µL WI solution (0.5 M mannitol, 20 mM KCl, 4 mM MES, pH 5.7). Protoplasts were incubated in the dark for 24 h at 25°C and GFP expression was imaged 24 h after transfection by confocal microscopy (Leica DM5500) to determine transfection efficiency. Protoplasts were then pelleted by centrifuging at ×200 g for 3 min, supernatant was removed, and pellets were kept at −80°C until DNA extraction. Three biological replicates were performed on separate days.

sgRNAs were transcribed in vitro using the EnGen sgRNA Synthesis Kit (NEB), followed by DNAse treatment to remove the template. The sgRNAs were subsequently purified using the RNA Clean and Concentrator kit (Zymo Research). RNP complexes were formed by combining 1.5 µg (0.5 µL) EnGen SpCas9 NLS (NEB) with 1.5 µg sgRNA in Buffer 3.1 (NEB) in a total volume of 5 μL, and incubated for 20 min at room temperature. The mixture containing RNP complexes was used for the transfection of freshly isolated protoplasts as described previously. Protoplasts were incubated in the dark for 24 h at 25°C, pelleted and kept at −80°C until genomic DNA extraction.

Protoplast DNA was purified 24 h after transfection in 96-well format using magnetic beads (NucleoMag Plant, Macherey-Nagel), following manufacturer’s instructions. DNA was eluted in 50 μL, of which 6 µL was subsequently used as a template in 25 µL PCR reactions using PHUSION HotStart Flex DNA polymerase (NEB) to amplify genomic DNA fragments containing target or predicted off-target sites. For the PCR, an initial denaturation for 30 s at 98°C was followed by 38 cycles of denaturation for 10 s at 98°C, annealing for 20 s at 58°C, extension for 20 s at 72°C, and a final extension step of 3 min at 72°C. Primer sequences are listed in Supplementary Dataset S3. The resulting PCR products were visualized on a 2% agarose gel. Equal amounts of PCR products were pooled to obtain sequencing libraries. Libraries were subsequently column-purified with the NucleoSpin Gel and PCR Clean-up kit (Macherey-Nagel), following manufacturer’s instructions. Illumina HiSeq sequencing (paired-end, 2x150 bp reads) was performed commercially by Eurofins Genomics Europe Sequencing GmbH, Constance, Germany.

Paired sequencing reads were uploaded to the CLC Genomics Workbench v20, trimmed, merged and demultiplexed using standard settings. Mutation frequencies at target and predicted off-target sites were determined using Amplican (Labun et al., 2019). Amplicons amplified from the genomic DNA of mock-transfected protoplasts were used as controls to normalize on- and off-target mutation frequencies. Mutagenic spectra were determined using a custom Python script for filtering the AmpliCan output and is available from the authors upon request. Large insertions were aligned to the tomato genome and plasmid used for transfection using HiSat2 (Kim et al., 2019).

We selected 29 members of the MYB (MYELOBLASTOMA) transcription factor gene family as targets. For the selected genes, sgRNAs and potential off-target sites with a maximum of 4 mismatches were predicted using CRISPOR (Concordet and Haeussler, 2018). For each gene, two to six perfectly matching sgRNAs were selected, including at least one with off-target sequences having three or more mismatches (“specific”) and one with one or two mismatches (and 0 mismatches in case of an alternative PAM, “non-specific”). A total of 89 sgRNAs as well as 213 of their respective predicted off-target sites (of which 68 in another MYB gene) were selected for the screen. Additionally, we considered potential off-target sites forming DNA or RNA bulges when hybridizing with sgRNAs. These sites are not predicted by CRISPOR and were therefore predicted using Cas-OFFinder (Bae, Park, and Kim, 2014). Several predicted off-target sites were selected where a bulge could not be resolved by an annealing alternative that would result in only a mismatch. This added 11 potential off-target sites to our screen, resulting in a total of 224. An overview of all selected off-target sites can be found in Supplementary Dataset S1.

Potential off-target sites with 0 (with an alternative PAM) to 4 mismatches to the sgRNA were selected. The alternative NGA or NAG Cas9 PAMs are functional at a low frequency in mammalian cells (Hsu et al., 2013; Cho et al., 2014; Tsai et al., 2015) and in rice (Meng et al., 2017). To further assess to which extent these non-canonical PAMs allow DNA cleavage in plants, we selected several sites that had either an NGA or NAG PAM (Figure 1A). Additionally, the position of the mismatch within the spacer might influence the likelihood of cleavage, resulting in an off-target mutation. Therefore, off-target site sequences with mismatches distributed over the length of the spacer were selected (Figure 1B). Notably, 20 sites were selected that had only a single mismatch in the 12 bp most proximal to the PAM (the so-called “seed sequence”).

Annealed oligonucleotides representing the spacers of the selected sgRNAs were directly ligated in a Golden Gate Level 2 vector containing SpCas9 and turboGFP expression cassettes. This arrangement ensured simultaneous delivery of all components in equimolar amounts (Figure 1C). Thus, the percentages of protoplasts to which Cas9 was delivered could be determined by fluorescence microscopy and these were found to be similar for the three replicates (41% ± 16%, 41% ± 19% and 42% ± 10%). A graphical overview of the subsequent experimental steps is shown in Figure 1D. Genomic DNA was isolated after 24 h, and fragments containing the target sites and predicted off-target sites were amplified with barcoded primers and sequenced after pooling. Per biological replicate, approximately 7 million pooled paired-end reads were obtained, with a median of approximately 18,200 per sample (average approximately 21,500). The reads were trimmed, merged and demultiplexed and the frequency and type of CRISPR-induced mutations at target and predicted off-target sites were determined.

We did not a priori select sgRNAs for high predicted efficiency, and a wide range of mutation frequencies was obtained (Figure 2A). For only 8 targets no mutations could be detected at frequencies at least 0.1% above their wildtype control. The experimentally obtained mutation frequencies were subsequently compared to the predictions as expressed by the Doench score, currently also known as the Azimuth score (Fusi et al., 2015; Doench et al., 2016) and by the Moreno-Mateos score, also known as CrisprScan (Moreno-Mateos et al., 2015). We divided the sgRNAs in quartiles based on their prediction score. For the Azimuth score, experimentally obtained mutation frequencies of sgRNAs in the third and fourth quartile (high score) were significantly increased compared to those in the first and second quartile (Figure 2B). This indicates that the Azimuth score has some, albeit weak, predictive power. No correlation was found between the Moreno-Mateos score and experimentally determined mutation frequencies (Figure 2C).

We quantified the types of mutations found at the (on-) target sites (Figure 3A). As seen previously, deletions occurred at a slightly higher frequency than insertions (62% ± 5% and 38% ± 5%, respectively). Most insertions were only 1 bp in size, making that the most common single event, although larger insertions did occur at much lower frequencies. Most deletions were small-sized, and frequencies declined with increasing deletion size. However, larger deletions up to 145 bp were found.

As the 1 bp insertion was the single most favoured mutation induced by Cas9, we investigated this mutation type further. Although it is still often believed that Cas9 introduces breaks with blunt ends, multiple reports state that staggered cuts with 1–3 bp 5’ overhangs frequently occur (Zuo and Liu, 2016; Shou et al., 2018). These overhangs can be filled in by templated DNA synthesis, leading to two blunt ends that can be ligated, resulting in a 1 bp insertion, or more if the overhang is larger (Figure 3B). Thus, in the case of templated insertion, the inserted nucleotide can be predicted from the identity of the −4 position relative to the PAM. This has been shown for yeast (Lemos et al., 2018) and mammalian cells (Shen et al., 2018; Allen et al., 2019; Shi et al., 2019), but not yet for plants.

We investigated all on-target sequencing reads containing a 1 bp insertion. Strikingly, we found a bias for A and T insertions, as opposed to C and G insertions (Figure 3C). Additionally, for every sgRNA, we determined whether the most frequently found insertion at the target site corresponded to the predicted insertion based on the identity of the base at the −4 position. For this analysis, we disregarded sgRNAs that had no or very few (<100) sequencing reads with an identified 1 bp insertion. For only 38 out of the analysed 72 sgRNAs, the most frequently found insertion was equal to the predicted insertion, i.e., a duplication of the −4 base (Supplementary Dataset S4).

Next, all sequencing reads belonging to sgRNAs with the same nucleotide at position −4 were grouped. We then compared the frequency of A, T, C, and G insertions within these groups to the frequencies of the same insertion in the other groups (e.g., the frequency of an A insertion when A is at the −4 position versus the frequency of an A insertion when T, C or G is at the −4 position) (Figure 3D). This revealed a bias towards duplication of the −4 nucleotide for each of the four nucleotides. Interestingly, only for gRNAs that were predicted to result in an A or T insertion was the most frequently observed insertion equal to the predicted insertion. Although C and G insertions occur more frequently for gRNAs with that base at the −4 position compared to the complete set of targets, the most frequently occurring inserted nucleotide was still A or T. We investigated if the bias could be caused by a higher likelihood of Cas9 inducing staggered cuts if the nucleotide at the −4 position is an A or T. This would lead to a higher apparent percentage of 1 bp insertions out of total mutations for sgRNAs with an A or T at the −4 position. However, upon closer inspection of our data, no such significant difference could be found (Supplementary Figure S1).

In our screen of merged paired reads we could also identify larger insertions of up to 133 bp at target sites. For the three biological replicates combined, we considered all identified insertions ≥11 bp and aligned these to both the tomato genome as well as to the sequence of the plasmid used for transfection (Figure 3E). In total, we found that 2.07% of the reads (8,761 reads out of 422,323) with a mutation at an on-target site contained an insertion of ≥11 bp. These reads belonged to 1,345 individual distinct events. Of these identifiable insertion events, 673 aligned to parts of the vector used for transfection and another 62 aligned to the tomato genome. For 610 insertion events, the origin of the insertion remains unknown (Figure 3E). An overview of all large insertions can be found in Supplementary Dataset S5.

To identify off-target mutations, 224 predicted off-target sites were amplified and analyzed for the presence of InDels using AmpliCan (Labun et al., 2019). Amplicons produced from mock-transfected protoplasts were used as controls for normalizing mutation frequencies. To reliably identify true off-target mutations, excluding any sequences that might result from PCR or sequencing errors, we disregarded off-target sites for which no homogeneous control (no Cas9 added) sequences were obtained. Because it might interfere with the identification of genuine mutations, all sites for which the control showed an apparent mutation frequency over 0.1% were disregarded, which was the case for 30 sites. The remaining sites were considered to be a confirmed off-target site if the average mutation frequency was 0.1% or higher and if the mutagenic spectrum showed a pattern expected for CRISPR mutations (e.g., InDels instead of substitutions, as we considered the latter to be more likely due to polymerase or sequencing errors). Mutation patterns at the identified off-target sites can be found in Supplementary Figure S2.

At 18 of the remaining 194 analyzed sites, corresponding to 13 sgRNAs, mutations above this threshold were identified (Figure 4A; Supplementary Figure S2 for mutagenic spectrum). Off-target activity mostly occurred at sites that had only one mismatch compared to the target (13 out of 42 tested sites), but was also found for three sites (out of 39) that had a 2 bp mismatch with the guide (Figure 4B). Interestingly, some off-target activity was identified for one site that did not contain any mismatches in the spacer, but had an alternative PAM (NGA instead of NGG, Figures 4A, B). No off-target activity was found at 89 and 49 sites that had 3 or 4 mismatches compared to the target site, respectively.

The two sites with the most frequently occurring off-target mutations (sites 54–2 and 40–3) had a single mismatch which occurred at the nucleotide most distal from the PAM (Figure 4B). Interestingly, off-target effects were only found for sites with two mismatches when one was located at the nucleotide most distal from the PAM (Figure 4B). This indicates that this nucleotide in particular adds very little to specificity and a mismatch here will most likely result in off-target activity. Apart from the two off-target sites mentioned above, off-target mutation frequencies were generally low, never reaching more than 5% of total reads, and in most cases no more than 1% of total reads for the respective sites (Figure 4B).

Rather than the absolute mutation frequencies at off-target sites, the ratio of off-target mutation frequency relative to the on-target mutation frequency is likely to be more relevant for practical situations. Relative off-target frequencies varied widely (Figure 4B, note the different scales). Again, these were highest for the two off-target sites that had a single mismatch most distal from the PAM, even reaching a mutation frequency similar to the on-target site for sgRNA40.

It is sometimes proposed that the use of ribonucleoproteins (RNPs) might decrease the frequency of off-target mutations. For RNP transfection, purified Cas9 protein preloaded with in vitro transcribed or synthetic sgRNAs are introduced in the protoplast.

To compare the (relative) off-target mutation frequencies for the two approaches, we synthesized the 13 sgRNAs for which we found off-target activity in our first screen (Figure 4B) by in vitro transcription and loaded them onto purified Cas9 protein. We again transfected protoplasts in a 96 wells format with either plasmid DNA or the corresponding RNPs. Transfection efficiencies for the three biological replicates of plasmid transfections were similar (56% ± 7%, 64% ± 3% and 62% ± 1%). We amplified the on-target sites as well as the previously identified off-target sites and subjected the resulting amplicons to next-generation sequencing. After processing of the reads, the absolute and on-target and absolute and relative off-target mutation frequencies were determined.

Overall, the on-target mutation frequency of plasmid-transfected protoplast was approximately 50% higher than in the first experiment (Figures 2–4), matching the higher transfection efficiency in the latter experiment. The on-target mutation frequencies for plasmid- and RNP-transfected protoplasts were comparable for each target, with less active guides tending to give higher frequencies with plasmid transfection (Figure 5A). Overall, the on-target mutation frequencies for the two methods correlated well (Figure 5B). For most of the tested off-target sites, the use of RNPs either did not significantly affect the mutation frequency or decreased the mutation frequency (Figure 5C). There were two notable exceptions: off-target sites 36–1 and 40–2. At these sites, the RNP complex induced significantly more InDels than Cas9 and sgRNA expressed by the plasmid.

Differences in on- and off-target mutation frequencies between these two types of transfection might simply be explained by varying final amounts of Cas9 protein and sgRNA present in the cells. Therefore, as discussed above for between-sgRNA comparisons, a fairer comparison between plasmid- and RNP-based transfections is the comparison of the ratio of on/off-target mutation frequencies (Figure 5C). The use of RNPs significantly reduced the relative off-target mutation frequency for 6 out of 17 sites. For the two previously mentioned sites 36–1, and 40–2, however, RNPs led to significantly higher relative frequencies. For these two sites, the off-target mutation frequency induced by RNPs was significantly higher than for plasmids (Figure 5C, left panel). Although the on-target efficiency was also higher for RNPs than for plasmids, this difference was much smaller (Figure 5B). For 9 sites, no significant difference in the ratio of on/off-target mutation frequencies was found.

It was expected that the use of RNPs instead of plasmid DNA for mutagenesis would reduce the risk of integration of non-host DNA fragments in CRISPR-induced DSBs, as no plasmid DNA is introduced in the cells. To test this, we first determined the spectrum of CRISPR-mutations induced by plasmid-encoded Cas9 and RNPs at on-target sites (Figure 6A). As expected, RNP-induced mutations were very similar to those from plasmid-encoded Cas9, but the percentage of reads that contained insertions larger than 5 bp, although overall low, differed substantially (Figure 6B). A significantly higher percentage of reads from plasmid-transfected protoplasts contained larger insertions, while only very few reads from RNP-transfected protoplasts contained an insertion of 32 bp or larger. In this experiment, we found that for plasmid-transfected protoplasts, 0.86% of on-target reads contained an insertion larger than or equal to 11 bp, pertaining to 181 individual events. For RNP-transfected protoplasts, this number was much lower, at 0.14% and 41 insertion events.

As these large insertion events still occurred, even for RNP-transfected protoplasts albeit a lower frequency, we again attempted to identify their origin. For the plasmid-transfected protoplasts, 94 insertion events out of 181 could be aligned to vector DNA. Another 6 could be aligned to the tomato genome–for the remaining 81 large insertion events, the origin remains unknown (Figure 6C). The relative frequencies for these origins are very similar to those observed in the larger initial experiment (Figure 3E). Interestingly, for the RNP-transfected protoplasts, 13 out of 41 insertion events could be traced back to the dsDNA template that was used to synthesize the sgRNA in vitro–even though the resulting sgRNA was treated with DNase. Another 21 events could be aligned to the tomato genome. For 7 events, the origin could not be determined (Figure 6C).