Phaseolus vulgaris cv. CIAP7247F, cv. Pinto and cv. Rosecoco were grown in the greenhouse (Brussels, Belgium) with a 16h:8h day/night cycle to obtain fresh seeds. Harvested seeds were surface sterilized in 50 ml centrifuge tubes by submerging them in 30 ml of 70% (v/v) ethanol for 10 minutes at room temperature while being shaken at 100 rpm. Subsequently, the seeds were rinsed 5 times with sterile distilled water after which the seeds were submerged in 30 ml of a 20% (v/v) dilution of commercial bleach (final sodium hypochlorite concentration of 1%) complemented with 2 drops of polysorbate 20 and incubated at room temperature for 30 minutes while being shaken at 100 rpm. The seeds were rinsed 5 times with sterile distilled water after which they were soaked in sterile distilled water for 24 hours. The sterilized seeds were grown on Steri Vent high containers (Duchefa, CAT #S1686) containing germination media (1/2 strength Murashige and Skoog (MS) salts, 1 mg/l thiamine, 30 g/l sucrose, 6 g/l agar, pH 5.7) for 3 to 5 days in the dark at 24°C after which the plants were used for the transformation with R. rhizogenes K599. Alternatively, sterilized seeds were sown directly into soil-containing pots and grown at 24°C in a culture room or greenhouse with a 16h:8h day/night cycle until they were used for transformation with R. rhizogenes K599.

Three different methods were tested for hairy root transformation of P. vulgaris. These methods were evaluated based on the percentage of transformed explants forming GFP-positive hairy roots (i.e. roots containing the T-DNAs of both the Ri plasmid and binary vector), the abundance of hairy roots emerging on one explant and the speed of hairy root appearance. An overview of the results can be found in Table 1 .

The raffinose family oligosaccharides biosynthetic genes of common bean were chosen as targets to test the efficiency of CRISPR/Cas9 in P. vulgaris (de Koning et al., 2021). Three different sgRNAs were designed to target the exon region of each gene ( Figure 4A ). The designed sgRNAs were selected based on the highest predicted efficiency scores and lowest potential off-targets. For raffinose synthase 1 (PvRS1), exon 1 was targeted by two sgRNAs (sgRNA1 and sgRNA2) and exon 2 was targeted by one sgRNA (sgRNA3). For raffinose synthase 2 (PvRS1), exon 1 was targeted by one sgRNA (sgRNA1) and exon 2 was targeted by two sgRNAs (sgRNA2 and sgRNA3). The stachyose synthase (PvSS) gene was targeted in exon 1, exon 3 and exon 4, each by one sgRNA (sgRNA1, sgRNA2 and sgRNA3, respectively). Initially, each sgRNA was first ligated into an entry vector followed by a Gateway^™ LR Clonase^™ reaction to obtain two expression clones, pMR356-GFP and pMR394-GFP both containing the same sgRNA ( Figure 4B ). The destination vectors differed in the promoter driving the A. thaliana codon-optimized Cas9 gene expression, namely PcUbi (pMR356-GFP) and 2X35S-Ω (pMR394-GFP). The pMR356-GFP and pMR394-GFP vectors were initially improved by the incorporation of an enhanced green fluorescent protein gene (eGFP) for visual distinction of transgenic hairy roots. In total, 18 different CRISPR/Cas9 constructs were developed to determine the gene editing efficiency of the designed sgRNAs in P. vulgaris.

For each sgRNA, CRISPOR, CRISPR RGEN and inDelphi generated in silico predicted sgRNA efficiency scores ranging from 0 to 100, with higher scores indicating higher efficiency. The predicted on-target activity of the sgRNAs differed substantially between sgRNAs ( Table 2 ). The efficiency scores as predicted by the Doench/Fusi 2016 model ranged from 36 to 69 whereas the CRISPRscan scores ranged from 32 to 78. The out-of-frame scores predicted by the Bae model, which only considers potential deletions, ranged from 51.0 to 79.6, while the Lindel out-of-frame scores, which includes deletions as well as insertions, were higher and ranged from 69 to 90. The inDelphi frameshift frequency also considers insertions as well as deletions in the prediction of mutation outcomes and ranged from 47.1 to 83.0. Furthermore, for the same sgRNA significant differences were observed between the predicted efficiency scores for the different test models. For instance, sgRNA1 targeting PvRS1 had the highest Lindel score of 90 but only obtained a CRISPRscan score of 43 and an inDelphi frameshift frequency of 55.7.

All gene editing analyses described herein were performed on the hairy roots derived from Method 3, as described above. To assess the efficiency of the designed sgRNAs and the influence of different promoters driving Cas9, we analyzed high-quality Sanger sequences of the target region of transgenic hairy roots for the presence of indels using TIDE. The majority of the designed sgRNAs caused mutations in the target sequence. This indicates that the Cas9 nuclease successfully cleaved DNA at the target site, triggering DNA repair and causing potential sequence alterations ( Figure 5A ). For each sgRNA, at least 15 hairy roots were analyzed for the presence of indels after which the mean INDEL score and KO score were calculated ( Figure 5B ; Supplementary Table S5 ). Our results indicate that not all designed sgRNAs had high gene editing efficiencies (mean INDEL and KO scores) and that the promoter driving Cas9 expression also influenced the cleavage efficiency. For PvRS1, sgRNA2 had the highest mean INDEL and KO scores, with respective values of 75.1% (SE = ± 4.3) and 55.3% (SE = ± 8.3) when paired with Cas9 under the PcUbi promoter (pMR356-GFP) and respective values of 84.8% ( ± 2.4) and 73.8% ( ± 6.9) when paired with Cas9 under the 2X35S-Ω promoter (pMR394-GFP). For PvRS2, sgRNA1 and sgRNA3 resulted in high gene editing efficiency scores with Cas9 under the PcUbi promoter (pMR356-GFP), with mean INDEL scores of 77.6% and KO scores ranging from 65.4 to 68.7%. The highest editing efficiency for PvSS was obtained by sgRNA3 with Cas9 under the PcUbi promoter, with a mean INDEL score of 84.7% (SE = ± 2.6) and a mean KO score of 65.5% (SE = ± 6.6). For four sgRNAs, significantly higher efficiency scores were obtained when Cas9 was driven by the PcUbi promoter, while for only two sgRNAs, Cas9 under the 2X35S-Ω promoter performed significantly better. For the remaining three sgRNAs, no significant difference was found between the two Cas9 promoters. The sgRNA3 targeting PvRS1 had very low INDEL and KO scores when paired with Cas9 under the pcUbi or the 2X35S-Ω promoter. This observation can be attributed to the presence of a single nucleotide polymorphism (SNP) in the seed sequence right next to the PAM site which caused a mismatch with the sgRNA. This SNP in the target gene is specific for P. vulgaris cv. CIAP7247F as confirmed by sequencing of the target region and was not present in the reference genome (P. vulgaris cv. G19833).

Within one hairy root, multiple dominant indels with an occurrence frequency higher than 20% could be seen. The number of dominant indels per hairy root varied for each sgRNA and ranged from one to four indicating that each sgRNA could induce different indels in one hairy root. For example, in the analyzed hairy roots edited by sgRNA1 targeting PvRS2, only one or two dominant indels were present per hairy root. Indels that occurred more than 20% in the hairy root sample pool for each sgRNA are listed in Table 3 . In general, three times more deletions than insertions occurred within the 288 analyzed hairy roots ( Figure 6 ). The most observed indels were one base pair (bp) deletions and insertions, with frequencies of 22.6 and 19.1%, respectively. The largest deletion observed was 30 bp long while the largest insertion observed was 4 bp long.

The computational models used to predict indels showed inconsistencies when compared to the observations made in planta. Specifically, the predictions of indels made by the Bae repair model were inconsistent with results observed in hairy roots, as indicated in Table 3 . Out of the nine sgRNAs studied, only two produced indels that matched the predictions made by the Bae repair model. Only for sgRNA3 targeting PvRS2 and sgRNA3 targeting PvRS3 the Bae repair model predicted respectively a 2 bp deletion and an 18 bp deletion correctly. In contrast, the inDelphi algorithm and Lindel repair model performed more effectively, accurately predicting the most occurring in planta observed indels for the majority of sgRNAs. For instance, sgRNA3 targeting PvSS resulted in 6 different mutations that occurred in more than 20% of the analyzed hairy roots of which 4 mutations were correctly predicted by the inDelphi algorithm and 3 of them by the Lindel repair model. In contrast, the Bae repair model did not predict any of the observed indels in this case.

The CRISPR/Cas9 on-target activity led to a diverse array of indels at the designated target loci. To examine the potential involvement of MMEJ in the repair process, the microhomology strength (defined by the length of local microhomologies, their GC content, and position near the cleavage site) was assessed using the inDelphi algorithm. Out of the nine sgRNAs analyzed, six target sites exhibited an average microhomology strength. Notably, the target site of sgRNA2 targeting PvRS1 and sgRNA2 targeting PvRS2 demonstrated a high microhomology strength ( Table 4 ). In contrast, the target site of sgRNA1 targeting PvRS2 exhibited a low microhomology strength.

Where MMEJ only leads to deletions, NHEJ can also result in insertions. Interestingly, the inDelphi algorithm accurately predicted the occurrence and nucleotide type of one base pair insertions in most instances. For five out of the nine sgRNAs, the in vivo outcomes aligned with the algorithm’s predictions. In these cases, the observed inserted nucleotide corresponded with the nucleotide at the -4 position upstream from the PAM site ( Table 4 ). In two cases (sgRNA2 targeting PvRS1 and sgRNA3 targeting PvSS), the predicted outcomes were partially correct. Only for sgRNA3 targeting PvRS1, the observed insertions did not correspond with the algorithm’s predictions.

Although several hairy root transformation protocols have been developed for common bean plants, we observed that the issue of genotype specificity continues to pose a challenge (Estrada-Navarrete et al., 2006; Khandual and Reddy, 2014; Voß et al., 2022). To find a preferred transformation protocol for generating transgenic hairy roots for P. vulgaris cv. CIAP7247F, we evaluated three distinct transformation protocols using R. rhizogenes K599. This strain was found to be very effective in inducing hairy roots not just in common bean plants, but also across a broad range of leguminous species (Estrada-Navarrete et al., 2006; Keyes et al., 2009; Khandual and Reddy, 2014; Xiang et al., 2016; Voß et al., 2022).

Method 1, adapted from Estrada-Navarrete et al. (2006), involved the injection of R. rhizogenes into the cotyledonary node of intact seedlings. This method resulted in a transformation efficiency of 35% and produced a low to medium number of GFP-positive hairy roots that emerged from the cotyledonary node and along the stem. This method had several limitations such as a lengthy waiting time for hairy root emergence of 35 days after infection, the emergence of a high number of GFP-negative roots at the cotyledonary node and along the stem, and the use of non-sterile growth conditions which could impact downstream processes. The transformation efficiencies reported by Estrada-Navarrete et al. (2006) using R. rhizogenes K599 ranged from 70% to 90%, which is higher than the 35% we observed for cultivar CIAP7247F. However, in the calculation for transformation efficiency, we only included explants that produced GFP-positive hairy roots that stained blue after a histochemical X-Gluc assay and did not include explants that produced only GFP-negative roots. In total 57% of our explants generated hairy roots of which 22% did not contain any GFP-positive hairy root. Explants containing GFP-positive hairy roots also showed numerous GFP-negative roots, which was also observed by Khandual and Reddy (2014). Their study reported that between 40% and 80% of the emerging roots did not contain T-DNA of the binary vector using the same method.

Method 2, adapted from Keyes et al. (2009), involved the inoculation of an incision created on the adaxial surface of the cotyledon. Although this approach was initially developed for soybean plants, its potential applicability to P. vulgaris was also considered. For common bean, this method resulted in relatively fast hairy root growth, with roots appearing just 15 days after inoculation. Additionally, the resulting hairy roots were developed in sterile conditions which could be beneficial for downstream processes. Nonetheless, only 15.2% of explants produced GFP-positive hairy roots, and the number of hairy roots per explant remained limited. A high incidence of cotyledon rot was also observed, which may pose a constraint for this technique. It is important to note that not all emerging hairy roots were GFP-positive, an observation also observed by Keyes et al. (2009), who reported that merely 55% of the developing roots contained T-DNA of the binary vector. Additionally, we noticed multiple instances where only one side of the root was stained following a histochemical X-Gluc assay ( Figure 2F ). This phenomenon might be attributed to either the early loss of transgenes (T-DNA not stably integrated in the genome) or loss of transgene expression (due to silencing) during the growth of a hairy root or the result of the inaccurate assumption that hairy roots are always derived from a single cell, rendering them chimeric, as observed in previous studies (Limpens et al., 2004; Garagounis et al., 2020). This observation is significant because it could impact the interpretation of gene editing events in hairy roots.

Method 3 was adopted from Khandual and Reddy (2014) and involved inoculating a severed radicle still attached to the rest of the seedling. For this method, the pMR356-GFP and pMR394-GFP vectors were used instead of the pGFPGUSPlus, which possess identical eGFP transcription units encompassing the same 35S promoter, eGFP, and NOS ter with polyadenylation signal. As a result, we anticipated a comparable level of eGFP expression for these three vectors. Additionally, since all experiments utilized the R. rhizogenes K599 strain for hairy root induction, uniformity in the Ri plasmid and chromosomal context is ensured. As a result, we expect consistent induction of hairy roots across all methods. The use of this method resulted in the highest transformation efficiency, with 41-48% of the explants carrying GFP-positive hairy roots. Hairy roots emerged relatively quickly (15-20 days after infection) and were obtained under sterile conditions. Furthermore, a high number of GFP-positive hairy roots were produced per explant. This technique proved effective in other P. vulgaris cultivars, such as Pinto and Rosecoco, demonstrating its capacity to induce hairy root growth in a broader range of common bean cultivars. One limitation of this method was the emergence of GFP-negative roots alongside the desired GFP-positive roots. This finding contrasts with the report by Khandual and Reddy (2014), which mentioned that all emerging hairy roots in their study contained T-DNA of the binary vector. However, the occurrence of GFP-negative roots was observed in all other tested methods and could potentially be the formation of normal roots originating from meristematic cells or callus cells at the cotyledonary node or the generation of hairy roots without incorporation of the T-DNA of the binary vector.

After evaluating the three different methods for hairy root transformation, Method 3 emerged as the most promising protocol for generating transgenic hairy roots in common bean. While our study does not involve a direct comparison of Method 3 with the other two methods, mainly due to variations in the transgenic constructs used, we are confident that these differences are minor. Consequently, we conclude that Method 3 demonstrated the highest transformation efficiency, rapid emergence of hairy roots, and the development of a significant number of GFP-positive hairy roots per explant. One drawback of all tested methods was the co-occurrence of GFP-negative alongside GFP-positive hairy roots, underscoring the need for additional screening steps to accurately identify hairy roots that contain the T-DNA of the binary vector. To address this, the incorporation of enhanced green fluorescent protein (eGFP) in our CRISPR/Cas9 vectors was necessary, allowing for a visual distinction of fluorescent hairy roots, which were subsequently used in further analysis.

The in planta effectiveness of in silico-designed sgRNAs that target genes involved in the RFO metabolic pathway of common bean were evaluated using the hairy root transformation system. Common bean has two raffinose synthase genes (PvRS1 and PvRS2) and one stachyose synthase gene (PvSS) encoding RFO biosynthetic enzymes (de Koning et al., 2021). Targeting these genes with CRISPR/Cas9 to generate knockout lines with reduced quantities of RFOs in the seed could improve the nutritional quality of these beans and alleviate consumption-related discomforts (Coon et al., 1990; Tomlin et al., 1991; Valentine et al., 2017; de Koning et al., 2021). There are important factors to consider before embarking on the more time-consuming stable transformation experiments. These factors include the fact that the effectiveness of the CRISPR/Cas9 system is influenced by the editing efficiency and specificity of the sgRNA as well as the promoter driving Cas9 expression (Ma et al., 2015; Liang et al., 2016; Ma et al., 2016; Li et al., 2022). By comparing the efficiency of sgRNAs and the promoters driving Cas9 using the hairy root transformation system, we aimed to identify the most effective combination that can be used in future experiments. Both ICE and TIDE software can be used to analyze Sanger data for potential mutations. We observed a strong correlation between the results of both programs (data not shown), consistent with findings from other studies (Conant et al., 2022). However, one major advantage of TIDE over ICE is that it provides p-values of all predicted indels and allows for specific parameter adjustments. For each sgRNA, at least 15 hairy roots were analyzed for the presence of indels, after which the mean INDEL score and KO score were calculated. This provides a general overview of the efficiency of sgRNAs and promoters within different hairy roots as CRISPR/Cas9 can potentially function differently within distinct plantlets. Furthermore, the resulting outcome of a DSB is also dependent on the repair mechanism, which relies partially on the specific target sequence. Not all designed sgRNAs showed good in planta efficiency, highlighting the importance of initially testing the sgRNA efficiency with our hairy root system. Nonetheless, we identified at least one or two highly efficient sgRNAs for all target genes. A noteworthy discovery in this study was the observation of low INDEL and KO scores for sgRNA3 targeting the PvRS1 gene, which was attributed to an SNP in the seed sequence immediately adjacent to the PAM site. This SNP led to a mismatch with the sgRNA sequence, resulting in a close to zero gene editing efficiency. This finding underscores the crucial need to screen for genetic variations within the targeted genomic regions of the chosen cultivars during sgRNA design, as SNPs or other sequence variations can have a substantial impact on gene editing outcomes (Modrzejewski et al., 2020). As shown here, only one mismatch in the seed sequence is enough to greatly reduce Cas9 cleavage, given the essential role the seed sequence plays in facilitating Cas9’s target sequence recognition (Modrzejewski et al., 2020). Furthermore, a difference in editing efficiency could be seen when Cas9 expression was driven by the PcUbi or the 2X35S-Ω promoter. In most cases, the PcUbi promoter resulted in higher INDEL and KO scores making it the preferred promoter to be used in common bean for further stable transformation experiments. On only two occasions did the 2X35S-Ω promoter result in higher scores. However, for three sgRNAs (sgRNA3 targeting PvRS2; sgRNA2 and sgRNA3 targeting PvSS), high editing efficiencies were obtained using PcUbi, while employing the same sgRNAs in combination with Cas9 driven by the 2X35S-Ω promoter resulted in significantly lower gene editing efficiencies, indicating that the 2X35S-Ω promoter may not be consistently reliable. These results demonstrate the potential of employing a hairy root transformation system as a valuable and efficient approach for evaluating the effectiveness of sgRNAs and the impact of different promoters on the editing efficiency. This system holds great promise in enhancing the success rate of CRISPR/Cas9-mediated stable transformation. Alternative methods have been developed for different plant species, primarily relying on transient expression to assess the efficacy of CRISPR/Cas9 editing. Notably, leaf infiltration has been successfully implemented in A. thaliana and N. benthamiana (Ordon et al., 2017). In the case of wheat, researchers have utilized a protoplast transfection assay and particle bombardment has been used in Barley (Arndell et al., 2019; Michalski et al., 2021).

To evaluate the accuracy of in silico prediction tools in forecasting sgRNA efficiency, we compared the predictions from CRISPOR, CRISPR RGEN, and inDelphi with the in vivo results. Generally, predicting sgRNA efficiency accurately appears to be challenging. No correlation was observed between the different in silico scores for a single sgRNA, nor between the in vivo efficiencies and most scores. For instance, sgRNA2 targeting PvRS1 performed well in vivo despite having only moderate in silico scores. Conversely, sgRNA2 targeting PvRS2 exhibited low efficiency in hairy roots but received high in silico scores. Nonetheless, the majority of designed sgRNAs demonstrated high in planta editing efficiency, which partially aligns with the high Lindel out-of-frame scores used for their initial selection. Moreover, the Lindel repair model accurately predicted indels for all nine sgRNAs, suggesting that the Lindel model can effectively predict gene editing qualities of sgRNAs in common bean. However, sgRNA2 targeting PvRS2 obtained a high Lindel out-of-frame score of 77 but did not perform as effectively in planta, with KO scores ranging from 11.3% (SE = ± 2.9) to 31.7% (SE = ± 5.7). This underscores the importance of conducting an initial sgRNA efficiency screen using our hairy root transformation system to identify efficient sgRNAs despite their predicted efficiencies in silico.

In our study we detected three times more deletions than insertions, which mirrors the indel distribution observed by Chen et al. (2019). CRISPR/Cas9 activity can result in targeted breaks in plant DNA at specific locations and plants have developed different repair mechanisms to fix the resulting DSBs. However, depending on the repair system used, different types of indels could arise at these locations. C-NHEJ is considered the most prominent repair mechanism within plants, which can result in insertions and deletions of a few base pairs (Puchta, 2005; Weterings and Chen, 2008). Within our hairy root sample pool, the most common mutations were indeed small insertions (1 or 2 bp) and deletions (1 to 5 bp) which are likely the result of C-NHEJ. Especially 1 and 2 bp insertions and deletions were present at high frequencies, an outcome that was also observed in other studies (Ordon et al., 2017; Chen et al., 2019; Zhang et al., 2020; Molla et al., 2022). While it was initially believed that SpCas9 generated blunt ends, an increasing number of studies now suggest that it can also produce 5’ staggered ends (Molla and Yang, 2019). Shou et al. (2018) demonstrated that the HNH nuclease domain of SpCas9 accurately cleaves DNA at the -3 position upstream of the PAM, while the RuvC domain exhibits greater flexibility and can cut even further upstream, generating 1-3 nt overhangs (Shou et al., 2018; Molla and Yang, 2019). These overhangs can be subsequently filled by DNA polymerase and ligated, resulting in small insertions corresponding with the template. Consequently, this leads to predictable insertions, as opposed to the unpredictable outcomes associated with blunt ends (Shou et al., 2018; Molla and Yang, 2019). Our findings are consistent with this hypothesis, as the majority of observed single nucleotide insertions matched with the expected template at the -4 position distal from the PAM. Although the inDelphi algorithm was optimized for human and mouse cell lines, it accurately predicted the insertions and most of the deletions, suggesting that the repair system may be more universally applicable than previously assumed (Shen et al., 2018). The only exception was sgRNA3 targeting PvRS1, where the observed in vivo outcome was not accurately predicted. However, this sgRNA contained an SNP immediately adjacent to the PAM site, which could have influenced the outcome and therefore should be interpreted with caution. A recent report on rice has also shown that NHEJ-mediated single nucleotide insertions can indeed be predicted based on the DNA sequence at the target loci (Molla et al., 2022). We present the first evidence that this is also applicable to common bean. These results offer promising implications for enhancing precision editing in plants, as they indicate the possibility of predicting NHEJ repair outcomes.

When micro-homologous sequences are found near the DSB, MMEJ can also lead to small deletions. Target sites with high microhomology strength typically possess local microhomologies that are lengthy, GC-rich, or situated near the cut site (Shen et al., 2018). Consequently, repair outcomes are more likely to involve microhomology-based deletions. Upon analyzing the sequences flanking the cleavage site, inDelphi indicated that six of the nine target sites exhibited an average microhomology strength, while two targets demonstrated high microhomology strength. This suggests that some of the observed longer deletions could be attributed to MMEJ in these cases. However, inDelphi suggested that the flanking region of the cleavage site of sgRNA2 targeting PvRS2 had a high microhomology strength, yet only small deletions were observed in the majority of analyzed hairy roots indicating that the predicted microhomology strength is not always a reliable parameter.

Since common bean is diploid, CRISPR/Cas9 can result in a heterozygous mutation, a biallelic mutation, or a homozygous mutation. When looking at individual hairy roots, the presence of one or two dominant indels (occurrence > 20%) was indeed observed. Additionally, hairy roots may not originate from a single cell as observed in both our and other studies, resulting in a chimeric nature of the root, which could also influence the interpretation of the observed mutations (Limpens et al., 2004; Garagounis et al., 2020). In our study, the number of dominant indels per hairy root varied depending on the sgRNA used and ranged between one and four supporting the hypothesis of the chimeric nature of each hairy root. This observation could potentially also explain the difference in indels observed by the computational programs and in vivo. Both computational programs, inDelphi and Lindel, successfully predicted the majority of observed indels, whereas the Bae repair model was less reliable in this regard, suggesting that the model may not be suitable for use in plant systems. Although these models are trained on mutational events in human and mouse cell lines, the Lindel repair model and inDelphi algorithm appear to be valuable for predicting DSB repair outcomes in plant systems as well, particularly for single nucleotide insertions. Additionally, the Lindel model seems to be the most effective tool for predicting sgRNA efficiency in common bean. These results offer promising implications for enhancing precise editing in plants.