Comprehensive evaluation and prediction of editing outcomes for near-PAMless adenine and cytosine base editors

Zhou, Xiaoyu; Gao, Jingjing; Luo, Liheng; Huang, Changcai; Wu, Jiayu; Wang, Xiaoyue

doi:10.1038/s42003-024-07078-5

Download PDF

Article
Open access
Published: 25 October 2024

Comprehensive evaluation and prediction of editing outcomes for near-PAMless adenine and cytosine base editors

Communications Biology volume 7, Article number: 1389 (2024) Cite this article

3635 Accesses
1 Citations
1 Altmetric
Metrics details

Subjects

Abstract

Base editors enable the direct conversion of target bases without inducing double-strand breaks, showing great potential for disease modeling and gene therapy. Yet, their applicability has been constrained by the necessity for specific protospacer adjacent motif (PAM). We generate four versions of near-PAMless base editors and systematically evaluate their editing patterns and efficiencies using an sgRNA-target library of 45,747 sequences. Near-PAMless base editors significantly expanded the targeting scope, with both PAM and target flanking sequences as determinants for editing outcomes. We develop BEguider, a deep learning model, to accurately predict editing results for near-PAMless base editors. We also provide experimentally measured editing outcomes of 20,541 ClinVar sites, demonstrating that variants previously inaccessible by NGG PAM base editors can now be precisely generated or corrected. We make our predictive tool and data available online to facilitate development and application of near-PAMless base editors in both research and clinical settings.

Deep learning models to predict the editing efficiencies and outcomes of diverse base editors

Article 15 May 2023

Disease modeling by efficient genome editing using a near PAM-less base editor in vivo

Article Open access 15 June 2022

Engineered dual selection for directed evolution of SpCas9 PAM specificity

Article Open access 13 January 2021

Introduction

Base editors can directly introduce base substitutions in the genome DNA without inducing double-strand breaks¹, thus facilitating disease modeling^2,3,4 and holding the potential for correcting pathogenic mutations in vivo^5,6,7,8,9,10. Cytosine base editors (CBEs)¹¹ and adenine base editors (ABEs)¹², in which a cytosine or adenine deaminase is tethered to a nickase variant of Cas9, convert C•G base pairs to T•A base pairs, or A•T base pairs to G•C base pairs, respectively. The canonical CBEs and ABEs have an editing window of ~4–8 nucleotides and require an NGG protospacer adjacent motif (PAM) 13–17 nucleotides downstream of the target site for efficient editing¹. However, the occurrence of bystander editing within the editing window and the absence of NGG PAM sequences at specific genome sites have constrained the application of base editors¹³.

To address these limitations, various modifications have been made to base editors to enhance their precision and applicability. For example, utilizing YE1, a modified version of cytosine deaminase, led to a narrowed editing window and reduced Cas9-independent off-target effects in CBEs^{14,15,16,17,18}. Shortening the linker between Cas9 and the deaminase also narrowed the activity window, thereby improving editing precision^11,12,19,20. The use of optimized deaminases such as TadA-8e in ABEs has increased their activities^21,22,23. Moreover, replacing the original Cas9 with variants that have relaxed PAM requirements has expanded the targeting scope of base editors^24,25. Notably, the SpRY variant has enabled near-PAMless base editing in plants^26,27,28 and human cells^{29,30,31,32,33}. Compared to base editors that are restricted by NGG PAM sequences, near-PAMless base editors have the potential to modify twice as many pathogenic loci listed in the ClinVar database³⁴.

One of the challenges encountered when utilizing base editors with relaxed PAM requirements is the unpredictability of both the editing efficiency and the resulting outcomes. It is already known that editing efficiencies vary substantially among different loci for different base editors³⁵. For standard CBEs and ABEs with NGG PAM, the sequence context significantly influences the editing efficiency and the proportion of bystander editing products^35,36. While previous studies have generated datasets and developed sequence-based models to predict the outcomes and efficiency for CBEs and ABEs with NGG or NG PAM^{35,36,37,38,39,40,41,42}, and the PAM preferences for SpRY Cas9 variants were also evaluated with HT-PAMDA²⁹, a comprehensive evaluation of the relationship between target sequence and outcomes for near-PAMless base editors is still lacking. This gap highlights the need for a systematic measurement of efficiencies and outcomes, which is essential to understand the unique determinants of near-PAMless base editors and to develop reliable prediction models. Such models are crucial for enhancing the application of these editors in both research and clinical settings.

Here, we performed a comprehensive evaluation of near-PAMless base editors on 45,747 sequences using a high-throughput assay. We found both the PAM sequence and the sequences flanking target sites are key determinants of the editing efficiency and the editing window. Building on these insights, we developed a model to predict editing efficiencies and outcomes of near-PAMless base editors. Using our data and model, we investigated the ability to edit pathogenic sites in the ClinVar database that are uneditable by NGG PAM base editors, providing useful information for the guidance and design of base editing screens and therapies.

Results

Generation of near-PAMless base editors

To leverage the expanded PAM compatibility offered by SpRY, our study utilized the BE4max-SpRY and ABEmax-SpRY, which incorporate SpRY in place of SpCas9, as previously described by Walton et al.²⁹ (Fig. 1a, b). Consistent with their results²⁹, both BE4max-SpRY and ABEmax-SpRY succeeded in editing sites with non-NGG PAMs (Fig. 1c–h). As a control, the AncBE4max-NGG, which was the commonly used base editor with improved precision at the time of our experiments⁴³, seldom edited non-NGG sites (Fig. 1c–h).

To optimize the editing efficacy and precision of these PAM-flexible base editors, we introduced several modifications. We substituted the deaminase domain in BE4max-SpRY or ABEmax-SpRY with YE1 or TadA-8e, respectively. Informed by Tan et al.¹⁹, for CBEs, we utilized a shortened P(AP)₃ linker (hereafter referred to as SL), expecting it to maintain high editing efficiency and narrow the editing window. The No Linker (hereafter referred to as NL) CBE variants were reported to have negligible efficiency¹⁹, so we did not use it for SpRY version of CBEs. For ABE8e, known for its high efficiency and broad editing window, we applied both SL and NL to investigate which configuration would reduce bystander editing while maintaining high editing efficiency. A BE3-flag-tagged nuclear localization signal (FNLS), which is designed to increase the nuclear expression of base editors^16,44, was also used to replace the original nuclear localization signal (NLS). These modifications yielded four additional near-PAMless CBEs and ABEs each. We then evaluated these CBEs and ABEs in HEK293T cells at three genomic sites containing multiple Cs (Sites 1-3) and three sites containing multiple As (Sites 4–6), respectively (Fig. 1c–h; Supplementary Table 1).

In the CBE variants, YE1 substitution in BE4max-SpRY led to a reduction in bystander editing at the eighth C for site 2, where the sixth cytosine base was targeted (Fig. 1d). Similarly, at site 3 with the seventh base as the target, YE1-SpRY showed decreased editing frequencies at the fifth, ninth, and tenth Cs compared with BE4max-SpRY (Fig. 1e). When YE1-SpRY was paired with a short linker (SL or P(AP)₃), it exhibited the lowest editing frequencies among all near-PAMless base editors across sites 1–3 (Fig. 1c–e). Notably, the integration of SL with BE4max-SpRY did not reduce editing frequencies at the non-target cytosines—the fourth, fifth, and seventh—when targeting the sixth cytosine at site 1 (Fig. 1c). These results suggest that YE1 integration into near-PAMless CBEs not only maintains target base editing efficiency but also enhances specificity by minimizing unintended edits.

For the adenine base editors, TadA-8e deaminase variants consistently displayed enhanced editing frequencies across all the adenines in the tested sites, suggesting an expansion of the editing window compared to ABEmax-SpRY (Fig. 1f–h). The ABE8e-SL-SpRY and ABE8e-NL-SpRY variants, contain a shortened or absent XTEN linker respectively, exhibited reduced bystander editing frequencies at adenine position 10 A in site 4, 8 A in site 5, and 10 A in site 6 compared with ABE8e-SpRY (Fig. 1f–h). Replacing the original NLS with a FNLS sequence did not notably alter editing efficiencies of either CBE or ABE variants (Fig. 1c–h). Based on these results, we selected two near-PAMless CBEs (FNLS-YE1-SpRY and YE1-SpRY) with improved editing precision and two near-PAMless ABEs (ABE8e-SL-SpRY and ABE8e-NL-SpRY) with increased editing efficiency for a more comprehensive evaluation.

Systematic evaluation of near-PAMless base editors using a large-scale sgRNA-target library

To systematically evaluate the performance of these near-PAMless base editors, we constructed a paired sgRNA-target library containing 45,747 sequences (Supplementary Table 2). Each sgRNA-target pair consists of a 20 nt sgRNA and its corresponding target DNA sequence, plus a 4 bp PAM sequence, enabling analysis of editing efficiency and outcomes by sequencing the target sequence. This library was designed to include 24,050 randomly generated sgRNA-target pairs with NANN or NGNN PAMs²⁹, with a balanced representation of sequence contexts for mapping sequence determinants of editing efficiency; 1,023 pairs with 256 types of NNNN PAMs for evaluating PAM preferences; 20,541 sequences associated with mutations reported in the ClinVar database with their corresponding endogenous PAMs, and 133 endogenous loci with non-NGG PAMs from previous reports²⁹ (Fig. 2a). Given that SpRY variants exhibit higher activity at sequences with NRN PAMs compared to NYN PAMs²⁹, we designed the random sequences to be enriched with ten NRN PAMs known for higher activities for SpRY. For library construction, synthesized sgRNA-target pairs were PCR amplified and assembled into a lentiviral plasmid by Gibson assembly^45,46.

**Fig. 2: High-throughput evaluation of PAM compatibility and editing activities of near-PAMless base editors.**

We packaged the sgRNA-target library into lentivirus and transduced the constructs into HEK293T cells. The cells were split into 10 pools, and each pool was transfected with a different base editor, with two independent replicate experiments performed (Fig. 2b). 72 h post-transfection, we sequenced the sgRNA-target cassettes to evaluate the editing efficiency and outcomes. A total of 35,769 to 37,005 sgRNA-target pairs were recovered with sequencing reads exceeding 100 in different experiments (Supplementary Table 3). High correlation of editing rates was observed between replicates (Supplementary Fig. 1a, b), with Pearson’s correlation coefficients ranging from 0.80 to 0.84 for CBEs and 0.74 to 0.85 for ABEs. Further validation at 38 endogenous sites in HEK293T cells also revealed the reliability of our library data, with strong correlations observed between editing efficiencies at integrated-target sequences and those at the endogenous sites (Supplementary Fig. 1c, d; Supplementary Table 4, 5; Pearson’s correlation 0.82 and 0.93, respectively).

Dependence of editing efficiency on PAM sequences in near-PAMless base editors

We first compared editing efficiencies for NGG, NG, and near-PAMless base editors on sequences with different PAMs. For sequences with NGGN PAMs at positions 21–24, the median C-to-T editing efficiency at positions 4-8 was 11.5% for AncBE4max-NGG, 7.63% for AncBE4max-NG, 11.44% for BE4max-SpRY, 8.42% for FNLS-YE1-SpRY, and 10.1% for YE1-SpRY (Fig. 2c). AncBE4max-NGG and BE4max-SpRY showed comparable efficiencies on sequences with NGGN PAMs. On NGNN PAM-containing sequences, while AncBE4max-NGG had greatly reduced editing efficiencies, the other four base editors maintained similar efficiencies as on NGGN PAM sites. The three SpRY versions of CBEs could also efficiently edit sequences containing NANN PAMs. Similarly, the SpRY version of ABEs showed expanded PAM compatibility (Fig. 2d). In addition, the ABE8e variant increased editing efficiency approximately by 3.7-fold compared to ABEmax-NGG on NGGN PAM sites (Fig. 2d).

We further evaluated the PAM preferences across 256 distinct PAM sequences to reveal differences among various CBEs and ABEs. We found that base editors containing SpRY had higher editing efficiency than NG or NGG PAM base editors across a diverse array of PAM sequences (Supplementary Fig. 2a, b). The first and the fourth base of the PAM resulted in variations in the editing efficiency for the base editor at the same sgRNA (Supplementary Fig. 2a, b). To assess the adaptability of near-PAMless base editors to different PAM sequences, we categorized 256 distinct PAM sequences into 16 types of PAM motifs based on variations of the second and third bases. Among the CBEs, FNLS-YE1-SpRY showed higher editing efficiencies than AncBE4max-NGG in 13 types of NXXN PAM motifs and higher efficiencies than AncBE4max-NG in 6 kinds of NXXN PAM motifs (Fig. 2e and Supplementary Fig. 2c). Similarly, the SpRY-integrated ABEs exhibited expanded PAM compatibility compared to their non-SpRY counterparts. ABEmax-SpRY exhibited higher editing efficiency than ABEmax-NGG and ABEmax-NG across all PAM motifs, except for NGGN and NG(A/C)N, respectively (Fig. 2f and Supplementary Fig. 2d). Taken together, these results suggest that SpRY version of base editors display broad PAM compatibilities, and the variations in the PAM sequences have a significant impact on the editing activities of base editors.

Editing outcomes are differentially affected by target sequence contexts for near-PAMless base editors

To evaluate the precision of base editing, we next compared the distribution of editing activities across the protospacer in target sequences. The mean editing rate peaked at position 5 or 6 for all CBEs and ABEs (Fig. 3a, b). We define the editing window as the positions edited at a rate exceeding 50% of the target site. Compared with BE4max-SpRY, the editing window of FNLS-YE1-SpRY and YE1-SpRY narrowed from positions 4–8 to positions 5–7 (Fig. 3a). For SpRY-integrated ABEs, Tad-8e with a rigid linker displayed an editing window of positions 3–8 in the target sequence, whereas ABE8e-NL-SpRY confined it mainly to positions 3–7 (Fig. 3b).

**Fig. 3: Impact of target sequence context on editing outcomes of near-PAMless base editors.**

Furthermore, we compared the bystander editing of SpRY-integrated base editors with NGG PAM-specific base editors, using sequences containing NGGN PAMs only. We calculated the relative editing efficiency at positions 4-8 compared to the position with the highest editing (Supplementary Fig. 3a, b). We found a lowered bystander editing activity for YE1-SpRY compared to AncBE4max-NGG across most two C or three C patterns, except when three Cs occupied positions 4, 6, and 7 or two Cs at position 5 and 6 (Supplementary Fig. 3a). When sequences contained consecutive adenines within the editing window, the highest editing efficiency was typically observed at the first adenine. An exception occurred for As at position 456, where the fourth adenine lies at the edge of the editing window, did not exhibit this trend (Supplementary Fig. 3b).

Subsequently, we sought to compare the sequence determinants in the target sequence that impact editing outcomes. It was observed that the base preceding the target significantly affects editing efficiency, with different deaminases showing preferences for different preceding bases^35,36. In line with this, we found that a T base preceding the target C resulted in significantly higher mean editing rate compared to other bases, whereas a preceding G correlated with the lowest rate (Fig. 3c and Supplementary Fig. 3c). By setting a cutoff of 5% editing frequency, we found that a preceding T enabled editing from positions 2-11 for BE4max-SpRY, while only positions 5-7 were editable with a preceding G (Supplementary Fig. 3c). FNLS-YE1-SpRY showed editing activity within positions 4 to 8, even with a preceding T. (Supplementary Fig. 3c). For all ABEs, a preceding A was associated with the lowest editing frequencies, while a preceding T consistently resulted in the highest editing frequencies (Fig. 3d and Supplementary Fig. 3d). Further analysis of the local 3-bp context around the target base revealed a preference for “TCN” sequence in CBEs and a tendency towards “TAY” (Y = C or T) context for ABEmax and “TAS” (S = C or G) contexts for ABE8e variants. (Supplementary Fig. 4a-f).

These observations underscore the complex relationship between sequence context and editing outcomes, which varies significantly among different base editors. Indeed, the proportion of target outcome varied markedly between different base editors, especially between Tad-8e-containing and Tad-7.10-containing near-PAMless base editors, where the Pearson’s correlation coefficient ranged only 0.04 to 0.23 (Supplementary Fig. 5a). As a result, the predictive models trained on one type of base editors may not be universally applicable to others. When predicting base editing outcomes using BE-Hive³⁵, a model trained on data from BE4max or ABEmax, we observed Pearson’s correlation coefficients in editing proportion ranging from 0.62 to 0.71. However, other deaminase-altered and linker-altered base editors showed correlations between 0.34 to 0.51 (Supplementary Fig. 5b, c). Therefore, the development of a new model is necessary to accurately capture the sequence determinants and PAM compatibility for precise prediction of the near-PAMless base editors.

Developing BEguider for predicting base editing outcomes with near-PAMless editors

To comprehensively capture sequence-activity relationships of the near-PAMless base editors, we developed a deep learning model named BEguider (Fig. 4a; Methods, Tables 1–2). Each BEguider model comprises of two modules, one for predicting editing efficiency and the other for predicting editing outcome proportions. The two modules share the same architecture, except for their last layers. For each module, BEguider consisted of two subpaths. On one path, the one-hot encoded sgRNA sequence data are fed into a Convolutional Neural Network (CNN), which is capable for capturing local sequence features. Concurrently, the data are passed into a second path consisting of an embedding layer followed by a Bidirectional Long Short-Term Memory (Bi-LSTM) layer. This path is designed to capture the global dependencies in the sequence data, leveraging Bi-LSTM’s ability to learn complex, bidirectional dependencies and handle long-range relationships effectively. The outputs of both the CNN and Bi-LSTM are then concatenated, merging local and global features into a unified representation. This design was aimed at extracting both local and global determinants in sgRNA sequences, and a stacking strategy was employed for accurate prediction by leveraging the strengths of CNNs and Bi-LSTMs in a complementary way.

**Fig. 4: Prediction of base editing efficiencies and outcomes using BEguider.**

We used 20-nucleotide target sequences and 4-nucleotide PAM sequences as the input data for the model. We trained a unique BEguider model for each base editor using the high-throughput base editing data we have generated (Supplementary Tables 6–11). The data for each base editor were split into training and test datasets with a ratio of 9:1. There were more than two thousand unused target sequences in every test dataset to evaluate BEguider’s performance. We found that BEguider models could precisely predict editing efficiencies (Fig. 4b, Spearman’s correlation 0.76-0.80, Pearson’s correlation 0.75–0.77) and editing outcome proportions (Fig. 4c, Pearson’s correlation 0.82–0.87) for near-PAMless CBEs. Similarly, BEguider models also showed good predictive performance for ABEs (Fig. 4d, e), with Pearson’s correlation between 0.71 to 0.72 for editing efficiencies and 0.77 to 0.89 for editing proportions.

We then tested BEguider’s performance on other experimental datasets. Currently, the only reported large datasets available for SpRY version of base editors were in Kim et al.⁴⁰. They have provided high-throughput sgRNA-target editing results for 5623 sgRNAs with SpRY-ABE8e(V106W) and 750 sgRNAs with SpRY-YE1-BE4max. We first predicted editing efficiencies for positions 4–8 using BEguider models trained on our ABE8e-SL-SpRY and ABE-NL-SpRY data. The overall Spearman’s correlation between BEguider-ABE8e-SL-predicted and BEguider-ABE8e-NL-predicted editing frequency and measured data from ABE8e(V106W) were 0.74 and 0.71 (Fig. 4f, g and Supplementary Fig. 6a), respectively. Kim et al. have developed DeepBE⁴⁰, a deep learning model that takes the deaminase and PAM sequence into consideration separately. The overall Spearman correlation coefficients of DeepBE predicted datasets with our measured editing frequency for ABE8e-SL-SpRY and ABE8e-NL-SpRY were 0.54 and 0.57 (Fig. 4f, g and Supplementary Fig. 6a), respectively. This indicates our models have good generalizability. We found that, in ABEs, the preceding A base before the target site showed the lowest prediction accuracy (Supplementary Fig. 6c, d), potentially due to the lowered editing rate for sites following the A base.

For CBEs, both models exhibited moderate generalizability to other datasets (Fig. 4h, j and Supplementary Fig. 6b). For per-position editing rates with different preceding bases, the preceding G base before the target site showed the lowest prediction accuracy, worst at position 8 with preceding G (Supplementary Fig. 6e, f). Therefore, more high-quality training data should facilitate generating more accurate prediction models.

To further validate the performance and generalizability of BEguider, we compared its predictions with experimental data from a different cell line. Using editing efficiencies measured in HepG2 cells with ABE8e-SpRY for 221 adenine bases⁴⁷, we found that our ABE8e-NL-SpRY model’s predictions showed strong correlations with both endogenous editing rates (Spearman’s ρ = 0.62) and integrated-target-site editing rates (Spearman’s ρ = 0.64). Notably, these correlations approach the experimental correlation (ρ = 0.60) between integrated-target-site and endogenous editing rates reported by Ryu et al., underscoring the robust performance of our model across different cellular contexts (Fig. 4k and Supplementary Table 12). In summary, our model shows excellent predictive performance, as evidenced by the good correlation between predicted and experimental datasets.

Assessing the potential of near-PAMless base editors for targeting pathogenic variants using BEguider

An important application of near-PAMless base editors is for disease modeling and correction of pathogenic SNVs. By analyzing the ClinVar database, we identified 47,485 pathogenic or likely pathogenic SNVs that correspond to C-to-T or A-to-G conversions. Considering the possibility to design sgRNAs, we found that 40,485 of these SNVs are correctable, and 39,997 are generatable by near-PAMless base editors. In comparison, only 7.6% and 8.4% of these variants can be corrected or generated with NGG base editors (Fig. 5a). Notably, 69.8% of the identified C-to-T SNVs and 57.0% of the A-to-G SNVs contain more than one editable base within the editing window (Fig. 5b), underlining the necessity for precise prediction of editing outcomes for near-PAMless base editors.

**Fig. 5: Prediction of editing outcomes for ClinVar pathogenic variants using BEguider.**

To identify those variants that could be precisely corrected or generated, we used our computational model, BEguider, to predict the editing outcomes. We defined SNVs as “precisely editable” if they achieve a predicted desired editing outcome proportion exceeding 90% with an editing efficiency above 5%. Under this criterion, we found that near-PAMless BEs could precisely correct 14,540 pathogenic or likely pathogenic SNVs and precisely generate 17,983 SNVs for disease modeling (Fig. 5c; Supplementary Tables 13–16). The precisely editable sites are associated with 2385 genes and 4386 diseases for C-to-T near-PAMless base editors, and 1782 genes and 2538 diseases for A-to-G editors. For disease modeling, precisely editable sites span variants across 1202 genes and 1548 diseases for C-to-T, and 2844 genes and 6063 diseases for A-to-G (Fig. 5d).

Figure 5e and f showed the results of predicted proportion and efficiency when editing windows contain different numbers of target bases. Notably, when two to three target bases were present within 4–8 bp at the editing window, certain SNVs remained editable with high precision. For instance, with two Cs in the window, 902, 1648, and 1318 SNVs can be precisely edited by BE4max-SpRY, FNLS-YE1-SpRY, and YE1-SpRY, respectively (Supplementary Fig. 7a, b). These results indicate that, in order to achieve desired editing outcomes, we can select the optimal base editors based on BEguider-predicted outcome proportion and editing efficiency.

Generation of ClinVar SNVs for disease modeling and SNV correction using near-PAMless base editors

To provide guidance for future studies of pathogenic variants using near-PAMless base editors, we examined editing outcomes in our high-throughput dataset for 10,175 sites for pathogenic SNV correction and 10,366 pathogenic SNV sites for disease modeling (Supplementary Fig. 8a, b; Supplementary Tables 17–20). These SNVs, when positioned as the sixth base in the sequence context, lack an NGG PAM sequence, rendering them inaccessible to conventional Cas9 version of base editors. We found that YE1-SpRY precisely corrected 443 pathogenic or likely pathogenic SNVs and precisely generated 1596 SNVs for disease modeling. Similarly, ABE8e-NL-SpRY could precisely corrected 2632 pathogenic or likely pathogenic SNVs and precisely generated 872 SNVs for disease modeling (Fig. 6a). In comparison, NG PAM-specific CBE and ABE precisely edited only 227 and 730 SNVs for correction, and precisely generated 764 and 221 sites for disease modeling, respectively. The precisely editable sites are associated with 432 genes and 449 diseases for near-PAMless CBEs, and 1469 genes and 2055 diseases for near-PAMless ABEs. For disease modeling, precisely editable sites span variants across 1124 genes and 1380 diseases for C-to-T editing, and 708 genes and 751 diseases for A-to-G editing, representing a 2 to 4.5-fold increase compared to the outcomes generated by NG PAM-specific base editors (Fig. 6b). For instance, in congenital muscular dystrophy, the previously uneditable c.3283 C > T (p.Arg1095Ter) variant in LAMA2 was edited at 36.2% frequency in the genome of HEK293T cells by YE1-SpRY (Fig. 6c). Similarly, for the TP53 c.695 T > C (p.Ile232Thr) variant, which is inaccessible for NGG-PAM base editors, were generated by ABE8e-NL-SpRY with 82.8% editing frequency (Fig. 6d). We next compared the bystander editing outcomes for different BEs on these pathogenic sites. We analyzed the mean editing frequency of each base in the editing window and edited proportion of the sixth base for near-PAMless CBEs and ABEs, with different combinations of editable positions in the window. BE4max-SpRY exhibited a slight leftward shift in proportion of edited sixth base (Fig. 6e), indicating relatively higher bystander effects compared to YE1-integrated PAMless CBEs. Consequently, YE1-SpRY demonstrated the highest precision in editing sites containing multiple Cs. Specifically, it achieved over 90% editing proportion at the 6^th position for 321 sites when Cs were at positions 4 and 6, and 144 sites when Cs were at 6 and 8, and 38 sites for Cs at 4,6,8. (Fig. 6f). For ABEs, ABE8e-NL-SpRY owns 68.7% editable sgRNAs at the 6th position across 2832 sites where adenines were targeted, and it edited 424 sites where As were present at both positions 6 and 8. ABE8e-NL-SpRY outperformed ABE8e-SL-SpRY in editing efficiency at the sixth target base while minimizing bystander effects (Fig. 6g).

**Fig. 6: Measured outcomes at 20,541 ClinVar sites by near-PAMless CBEs and ABEs.**

In summary, we have generated an extensive dataset of experimentally measured editing outcomes for 20,541 ClinVar variants using near-PAMless base editors. This resource is now accessible through http://beguider.bmicc.org/, a website that also offers our prediction model, in an interactive online format (Supplementary Fig. 9). The website is designed to facilitate the use of near PAMless base editors. Users can input a gene name with the target sequence or chromosome position, select a BE, and BEguider will generate optimized sgRNA sequences for use with near-PAMless base editors, along with detailed predictions about editing efficiency, editing outcomes, and their proportions. Additionally, users can input a pathogenic variant from ClinVar, choose a base editor type, and specify whether they aim to correct or generate the variant. BEguider then provides the designed sgRNA along with the editing efficiency and detailed outcome predictions, indicating the potential for precise correction or generation of the variant related to diseases.

Discussion

In this study, we systematically evaluated near-PAMless base editors and developed prediction model to enable their precise application. Using a library of 45,747 sgRNAs, which include 20,541 targeting ClinVar pathogenic variants, we have shown that near-PAMless base editors can efficiently edit non-NGG sites. This significantly broadens the scope of base editors in correcting or generating disease-associated variants that were previously uneditable. For instance, in the NF1 gene, whose loss of function is associated with a group of severe genetic disorders called neurofibromatoses⁴⁸, only 986 bases could be converted using NGG-specific CBE and ABE. In contrast, near-PAMless base editors are predicted to enable editing at 16,945 sites in NF1, thereby allowing a more comprehensive evaluation of variants associated with neurofibromatoses.

Our prediction model, BEguider, enables reliable prediction of base editing efficiencies and outcomes by effectively integrating the sequence determinants of both the deaminase and the Cas9 variant-SpRY. We validated our ABE8e-NL-SpRY model using data from another cell line, demonstrating its robust performance across different cellular contexts. However, further validation of additional models would be beneficial to comprehensively assess their predictive performance across diverse datasets. The model is available through our BEguider website (http://beguider.bmicc.org/), which provides a user-friendly interface for designing sgRNA libraries and predicting editing outcomes using near-PAMless base editors. This tool, together with the editing data of ClinVar variants, will be invaluable for researchers seeking to apply base editing technology in their work, especially in therapeutic contexts where precision is critical.

The datasets and analyses from this study also provide insights to optimize near-PAMless base editors. For CBEs, replacing the deaminase with YE1 narrows the editing window despite results in slightly reduced efficiency. In contrast, modifying the linker connecting the deaminase and Cas9 in ABEs allowed more efficient editing with enhanced specificity, as evidenced by the linker-less ABE8e-NL-SpRY. Our findings thus indicate that rational modifications to the deaminase domain and linker region can fine-tune the activities of near-PAMless base editors.

Our study also reveals limitations of near-PAMless base editors. Firstly, base editing shows suboptimal efficiency at sites with consecutive adenines and low proportions of target base at sites with consecutive cytosines. As alternatives, advanced prime editors like PE5max⁴⁹ with structured epegRNAs⁵⁰ with high product purity may prove to be effective in such unfavorable local contexts. Additionally, the low editing efficiency with NCN or NTN PAM need to be addressed with other improvements. Evolved Cas9 variants may help overcome these challenges⁵¹. Lastly, off-target effects remain a concern with near-PAMless base editors. The relaxed PAM recognition of SpRY leads to expanded off-target editing compared to wildtype Cas9²⁹. Utilizing high-fidelity SpCas9 variants could potentially mitigate this issue²⁹. The ABE8e deaminase also induces increased DNA and RNA off-targets relative to ABEmax²¹, but modifications like introducing the V106W mutation or embedding the deaminase into nCas9(CE-8e-SpRY) could reduce these effects⁵². Nevertheless, more data are needed to fully address these off-target effects for clinical application.

Methods

Oligonucleotide library construction

The oligonucleotide library consisting of 45,747 paired sgRNA-target sequences was commercially synthesized as an arrayed pool by GenScript. Briefly, 37,757 and 38,215 oligonucleotides containing highly variable target sequences with editable C or A at positions 4-8 were included for ABE and CBE screening, respectively. The library was comprised of several sublibraries. The random library included 24,050 unique sequences, each consisting of a 20 nt target sequence followed by one of 10 selected ‘NANN’ or ‘NGNN’ PAMs. These PAMs were selected as the top ten PAM sequences that demonstrated highest editing efficiency in the SpRY PAM preference evaluation dataset²⁹. These target sequences were generated using an in-house Python script with the following constraints:(1) inclusion of either adenine (A) or cytosine (C) at position 6 to ensure representation of editable bases within the typical activity window, (2) GC content between 45–55%, and (3) exclusion of sequences containing five or more consecutive repeating nucleotides or BsmBI recognition sites. These constraints ensured a balanced nucleotide distribution across the target sequences, with enrichment of A/C at position 6. The ClinVar library consisted of 20,541 disease-associated loci extracted from the NCBI ClinVar database (March 2020 release)³⁴, each with its corresponding endogenous PAM, including “NANN” or “NGHN”. The PAM library contained 1023 sequences with 256 types of NNNN PAMs for evaluating PAM preferences of SpRY, along with 133 endogenous loci²⁹ (Supplementary Table 2). Each oligonucleotide contained the following elements sequentially: a 20 nt sgRNA spacer, an 8 nt randomized sequence flanked by two BsmBI sites for inserting an optimized sgRNA scaffold⁵³, a 9 nt unique barcode, a target sequence with 4 nt PAM and 2 nt random bases, and 20 nt homologous sequences on both sides. Any BsmBI site-containing sequences were excluded during library construction (Supplementary Note 1).

Plasmid library construction

LentiGuide-Puro (Addgene #52963) was used as the backbone for the paired sgRNA-target library. A two-step cloning procedure was undertaken to construct the plasmid library comprising sgRNA-target pairs. First, the synthesized oligo pool was amplified by PCR using NEBNext Ultra Q5 Master Mix (New England Biolabs) in 24 cycles (primer sequences detailed in Supplementary Table 3) followed by purification with DNA Clean & Concentrator-5 Kit (Zymo Research). Subsequently, the PCR products were assembled into a lentiviral backbone plasmid via Gibson assembly at 50 °C for 15 min. After purification, 100 ng of the constructed plasmid was transformed into 100 μl MegaX DH10B™ T1R Electrocomp™ Cells (ThermoFisher Scientific), recovered in SOC medium at 37 °C for 1 h, and spread onto 245 × 245 mm LB-agar plates containing 100 μg/ml ampicillin. The plates were incubated at 32 °C for 14 h to obtain bacterial colonies, which were then isolated and purified using the EndoFree Plasmid Mega Kit (Qiagen). To integrate an optimized sgRNA scaffold⁵³, the first-round library plasmids were linearized with BsmBI and ligated overnight at 16 °C with a PCR-amplified sgRNA cassette containing the optimized scaffold. The second-round plasmid library harboring intact sgRNA-target pairs was transformed and amplified using the same workflow. Deep sequencing was performed on the Illumina HiSeq X Ten platform to validate sgRNA-target pair diversity.

Generating near-PAMless base editors

pCAG-CBE4max-SpRY-P2A-EGFP (Addgene #139999) and pCMV-T7-ABEmax-SpRY-P2A-EGFP (Addgene #140003) served as the original plasmid backbones. To generate individual near-PAMless CBE variants, PCR-amplified fragments encoding YE1 or an FNLS peptide, and synthesized P(AP)₃ linkers were cloned to replace the corresponding components The resulting plasmids were designated as BE4max-SL-SpRY, FNLS-YE1-SpRY, YE1-SL-SpRY and YE1-SpRY. Similarly, near-PAMless ABE variants were obtained by replacing TadA-7.10 with a PCR-amplified Tad-8e fragment, substituting the XTEN linker with a P(AP)₃ polypeptide or removing XTEN linker, and replacing the NLS with an FNLS sequence. These edited plasmids were named ABE8e-SpRY, ABE8e-SL-SpRY, FNLS-ABE8e-SL-SpRY, and ABE8e-NL-SpRY. For evaluating PAM preferences, pCMV-AncBE4max (Addgene #112094), pCMV-ABEmax (Addgene #112095) were used as backbones to generate PAM-constrained editors AncBE4max-NG and ABEmax-NG via introducing mutations in the SpCas9 coding region (Supplementary Note 2).

Cell library generation

To produce lentiviral particles, plasmids encoding the sgRNA-target library (21 μg), psPAX2 (15 μg), and pMD2.G (6 μg) were transfected into HEK293T cells cultured in 15 cm dishes using standard protocols. At 16 h post-transfection, the media was replaced with viral production media: UltraCULTURE Serum-Free Medium (Lonza) supplemented with 100 mM sodium pyruvate (Gibco), 7.5% sodium bicarbonate (Gibco), 0.5 M sodium butyrate (Sigma), 2 mM L-glutamine (Gibco), and 1% penicillin-streptomycin (Gibco). After 36 h, viral supernatant was harvested, filtered through a 0.45 μm PES filter (Millipore), and used to transduce HEK293T cells at MOI 0.5 with 8 μg/ml polybrene (Sigma). Transduced cells were selected with 1.5 μg/ml puromycin for 48 h to generate cell libraries expressing the sgRNA-target constructs.

Base editor screening and sample preparation

To deliver base editors, 50 μg of plasmid encoding each base editor was transfected into the cell libraries (2 × 10⁷ cells) using FuGENE HD (Promega) according to the manufacturer’s protocol. After 48 h, GFP-positive cells were sorted by flow cytometry (SH800, Sony) and genomic DNA was extracted (Blood & Cell Culture DNA Maxi Kit, Qiagen). For each base editor, two biological replicates were performed. To prepare NGS libraries, genomic DNA was amplified in 50 parallel PCR reactions using Illumina primers containing sample indexes and adapters. PCR products were sequenced on the Illumina HiSeq X Ten platform with over 1500X average sequence read coverage for each sgRNA.

Analysis of editing efficiencies at endogenous loci

To evaluate base editing efficiencies at endogenous genomic sites, HEK293T cells were transfected in 6-well plates with a base editor plasmid (2 μg) and sgRNA plasmid (1 μg) using FuGENE HD (Promega), followed by 48 h puromycin selection to enrich for transfected cells. GFP-positive cells were sorted (SH800, Sony), and genomic DNA extracted (Blood & Cell Culture DNA Midi Kit, Qiagen). Target loci were amplified by two rounds of PCR, first using locus-specific primers followed by Illumina adapters. Amplicons were subjected to both deep sequencing (HiSeq X Ten, Illumina) and Sanger sequencing. and the editing efficiencies at each target locus were quantified from the Sanger sequencing using the base editing analysis tool BEAT⁵⁴.

Processing high-throughput sequencing data

Paired-end reads from deep sequencing were assembled using SeqPrep software⁵⁵, and then aligned to the sgRNA scaffold sequence using BWA⁵⁶. Other sequence components including sgRNAs, barcode 2, targets, and PAMs were trimmed based on their relative positions in the aligned reads. Trimmed reads were mapped to the designed sgRNA library using CALITAS⁵⁷ by leveraging the correspondence between barcode 2 s and sgRNAs. Target sequences containing base changes inconsistent with the expected editing outcomes, indels, or total read counts below 100 were excluded from downstream analysis.

Quantification of base editing efficiencies and outcomes

Base editing efficiencies and outcome proportions for each sgRNA were calculated using custom Python scripts as described in Song et al³⁶:

$${{\rm{Base}}}\,{{\rm{editing}}}\,{{\rm{efficiency}}} =\frac{{{\rm{Total}}}\,{{\rm{reads}}}\,{{\rm{of}}}\,{{\rm{all}}}\,{{\rm{sequences}}}\,{{\rm{that}}}\,{{\rm{contain}}}\,{{\rm{substrate}}} \, ({{\rm{C}}}\,{{\rm{for}}}\,{{\rm{CBE}}};\,{{\rm{A}}}\,{{\rm{for}}}\,{{\rm{ABE}}}){{\rm{transition}}}\,{{\rm{within}}}\,{{\rm{the}}}\,{{\rm{editing}}}\,{{\rm{window}}} \, ({{\rm{at}}}\,{{\rm{position}}}\,4-8)}{{{\rm{Total}}}\,{{\rm{reads}}}}$$

$${{\rm{Base}}}\,{{\rm{editing}}}\,{{\rm{outcome}}}\,{{\rm{proportion}}}=\frac{{{\rm{Reads}}}\,{{\rm{of}}}\,{{\rm{the}}}\,{{\rm{specific}}}\,{{\rm{base}}}-{{\rm{edited}}}\,{{\rm{outcome}}}\,{{\rm{sequence}}}\,({{\rm{C}}}\,{{\rm{to}}}\,{{\rm{T}}}\,{{\rm{for}}}\,{{\rm{CBE}}};\,{{\rm{A}}}\,{{\rm{to}}}\,{{\rm{G}}}\,{{\rm{for}}}\,{{\rm{ABE}}})}{\begin{array}{c}{{\rm{Total}}}\,{{\rm{reads}}}\,{{\rm{of}}}\,{{\rm{all}}}\,{{\rm{sequences}}}\,{{\rm{that}}}\,{{\rm{contain}}}\,{{\rm{substrate}}}\,({{\rm{C}}}\,{{\rm{for}}}\,{{\rm{CBE}}};\,{{\rm{A}}}\,{{\rm{for}}}\,{{\rm{ABE}}})\\ {{\rm{transition}}}\,{{\rm{within}}}\,{{\rm{editing}}}\,{{\rm{window}}}\,({{\rm{at}}}\,{{\rm{position}}}\,4-8)\end{array}}$$

$$ {{\rm{Absolute}}}\,{{\rm{frequency}}}\,{{\rm{of}}}\,{{\rm{an}}}\,{{\rm{outcome}}}\,{{\rm{of}}}\,{{\rm{base}}}-{{\rm{editing}}}\\ \quad =\frac{{{\rm{Reads}}}\,{{\rm{of}}}\,{{\rm{the}}}\,{{\rm{specific}}}\,{{\rm{base}}}-{{\rm{edited}}}\,{{\rm{outcome}}}\,{{\rm{sequence}}}}{{{\rm{Total}}}\,{{\rm{reads}}}}\\ \quad ={{\rm{Base}}}\,{{\rm{editing}}}\,{{\rm{outcome}}}\,{{\rm{proportion}}}\times {{\rm{Base}}}\,{{\rm{editing}}}\,{{\rm{efficiency}}}\qquad\qquad$$

Model training and evaluation

The experimental dataset generated for each base editor contained between 35,769 to 37,005 target sequences. To build deep learning models for predicting editing efficiency and outcomes, these sequences were randomly split into training and test sets at a 9:1 ratio for each base editor. The input data for the model were 20-nucleotide target sequences and 4-nucleotide PAM sequences, which were converted into numeric matrices using a one-hot encoding method (A → [1,0,0,0], T → [0,1,0,0], G → [0,0,1,0], C → [0,0,0,1]).

The BEguider model architecture (Table 1) consisted of two modules—a convolutional neural network (CNN) module⁵⁸ and a bidirectional long short-term memory (Bi-LSTM) module⁵⁹, based on a stacking ensemble approach⁶⁰. The CNN module, with two convolutional layers followed by max-pooling layers, was responsible for extracting local sequence features and reducing dimensions. The Bi-LSTM module included an embedding layer, which projected inputs into a continuous vector space, and a Bi-LSTM layer for learning global sequence patterns.

During training, outputs from both modules were flattened via dense layers and then concatenated. Dropout layers were implemented before and after each dense layer (except the output layer) to regularize the models and prevent overfitting. During each iteration of model training, 10% of the training data were randomly selected and utilized for evaluating the model. Hyperparameter optimization was conducted using the hyperband algorithm in KerasTuner⁶¹ during training with the different base editor datasets (Table 2).

The model performance was evaluated using the test set, which contained over 2,000 unused samples for each base editor, by calculating Pearson and Spearman correlation coefficients between the predicted and measured editing efficiencies and proportions.

Table 1 Structures of BEguider

Full size table

Table 2 Hyperparameter options of BEguider

Full size table

Base editing predictions using existing models

The performance of two previously reported models was benchmarked using our experimental data. For DeepBE⁴⁰, the online prediction tool (http://deepcrispr.info/DeepBE/) was utilized in Comparison Mode. BE-HIVE³⁵, predictions were generated using the Python package (https://github.com/maxwshen/be_predict_bystander) with parameters set to: cell type—HEK293T, base editor—BE4 or ABE. Input sequences were padded to 50 bp on both sides of the sgRNA. To enable comparison of outcome proportions, only mutations within the editing window were considered and normalized by removing results for single target bases.

Analysis of endogenous locus sequencing

Reference sequences of 241 nt were used, comprising 120 nt upstream and 120 nt downstream of the targeted chromosomal coordinates based on the hg38 human reference genome. These reference sequences were retrieved using SAMtools software⁶². Sequencing reads were aligned to the reference sequences using BWA⁵⁶ and editing efficiencies quantified as described above.

Statistics and reproducibility

Statistical analysis for Fig. 1 was performed using GraphPad Prism 5.0 for Fig. 1 and Python Scipy package (version 1.7.3) for all other figures. For Fig. 1, one-way ANOVA followed by Dunnett’s test was used to compare multiple groups to a control. For other figures, independent two-tailed t-tests were used for comparison between two groups, and one-way ANOVA with appropriate post-hoc correction was used for multiple group comparisons, as specified in the figure legends. Statistical significance was set at P < 0.05. Additional statistical details can be found in the corresponding figure legends and Supplementary Tables.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

The DNA sequencing data have been deposited in the NCBI Sequence Read Archive under accession number PRJNA1046210. Plasmids have been deposited to Addgene: ABE8e-SL-SpRY (Addgene #228516); ABE8e-NL-SpRY (Addgene #228517); YE1-SpRY (Addgene #228518); FNLS-YE1-SpRY (Addgene #228519).

Code availability

We have made the source code for BEguider and all processed data available on Github at https://github.com/Wangxiaoyue-lab/BEguider⁶³. The web application is available at http://beguider.bmicc.org.

References

Rees, H. A. & Liu, D. R. Base editing: precision chemistry on the genome and transcriptome of living cells. Nat. Rev. Genet 19, 770–788 (2018).
Article CAS PubMed PubMed Central Google Scholar
Kweon, J. et al. A CRISPR-based base-editing screen for the functional assessment of BRCA1 variants. Oncogene 39, 30–35 (2020).
Article CAS PubMed Google Scholar
Huang, C., Li, G., Wu, J., Liang, J. & Wang, X. Identification of pathogenic variants in cancer genes using base editing screens with editing efficiency correction. Genome Biol. 22, 80 (2021).
Article CAS PubMed PubMed Central Google Scholar
Hanna, R. E. et al. Massively parallel assessment of human variants with base editor screens. Cell 184, 1064–1080 e1020 (2021).
Article CAS PubMed Google Scholar
Villiger, L. et al. Treatment of a metabolic liver disease by in vivo genome base editing in adult mice. Nat. Med. 24, 1519–1525 (2018).
Article CAS PubMed Google Scholar
Ryu, S. M. et al. Adenine base editing in mouse embryos and an adult mouse model of Duchenne muscular dystrophy. Nat. Biotechnol. 36, 536–539 (2018).
Article CAS PubMed Google Scholar
Levy, J. M. et al. Cytosine and adenine base editing of the brain, liver, retina, heart and skeletal muscle of mice via adeno-associated viruses. Nat. Biomed. Eng. 4, 97–110 (2020).
Article CAS PubMed PubMed Central Google Scholar
Newby, G. A. et al. Base editing of haematopoietic stem cells rescues sickle cell disease in mice. Nature 595, 295–302 (2021).
Article CAS PubMed PubMed Central Google Scholar
Chai, A. C. et al. Base editing correction of hypertrophic cardiomyopathy in human cardiomyocytes and humanized mice. Nat. Med. 29, 401–411 (2023).
Article CAS PubMed PubMed Central Google Scholar
Lee, R. G. et al. Efficacy and safety of an investigational single-course CRISPR base-editing therapy targeting PCSK9 in nonhuman primate and mouse models. Circulation 147, 242–253 (2023).
Article CAS PubMed Google Scholar
Komor, A. C., Kim, Y. B., Packer, M. S., Zuris, J. A. & Liu, D. R. Programmable editing of a target base in genomic DNA without double-stranded DNA cleavage. Nature 533, 420–424 (2016).
Article CAS PubMed PubMed Central Google Scholar
Gaudelli, N. M. et al. Programmable base editing of A*T to G*C in genomic DNA without DNA cleavage. Nature 551, 464–471 (2017).
Article CAS PubMed PubMed Central Google Scholar
Anzalone, A. V., Koblan, L. W. & Liu, D. R. Genome editing with CRISPR-Cas nucleases, base editors, transposases and prime editors. Nat. Biotechnol. 38, 824–844 (2020).
Article CAS PubMed Google Scholar
Doman, J. L., Raguram, A., Newby, G. A. & Liu, D. R. Evaluation and minimization of Cas9-independent off-target DNA editing by cytosine base editors. Nat. Biotechnol. 38, 620–628 (2020).
Article CAS PubMed PubMed Central Google Scholar
Kim, Y. B. et al. Increasing the genome-targeting scope and precision of base editing with engineered Cas9-cytidine deaminase fusions. Nat. Biotechnol. 35, 371–376 (2017).
Article CAS PubMed PubMed Central Google Scholar
Zuo, E. et al. A rationally engineered cytosine base editor retains high on-target activity while reducing both DNA and RNA off-target effects. Nat. Methods 17, 600–604 (2020).
Article CAS PubMed Google Scholar
Zuo, E. et al. Cytosine base editor generates substantial off-target single-nucleotide variants in mouse embryos. Science 364, 289–292 (2019).
Article CAS PubMed PubMed Central Google Scholar
Grunewald, J. et al. Transcriptome-wide off-target RNA editing induced by CRISPR-guided DNA base editors. Nature 569, 433–437 (2019).
Article CAS PubMed PubMed Central Google Scholar
Tan, J., Zhang, F., Karcher, D. & Bock, R. Engineering of high-precision base editors for site-specific single nucleotide replacement. Nat. Commun. 10, 439 (2019).
Article CAS PubMed PubMed Central Google Scholar
Nishida, K. et al. Targeted nucleotide editing using hybrid prokaryotic and vertebrate adaptive immune systems. Science 353, https://doi.org/10.1126/science.aaf8729 (2016).
Richter, M. F. et al. Phage-assisted evolution of an adenine base editor with improved Cas domain compatibility and activity. Nat. Biotechnol. 38, 883–891 (2020).
Article CAS PubMed PubMed Central Google Scholar
Gaudelli, N. M. et al. Directed evolution of adenine base editors with increased activity and therapeutic application. Nat. Biotechnol. 38, 892–900 (2020).
Article CAS PubMed Google Scholar
Lapinaite, A. et al. DNA capture by a CRISPR-Cas9-guided adenine base editor. Science 369, 566–571 (2020).
Article CAS PubMed PubMed Central Google Scholar
Nishimasu, H. et al. Engineered CRISPR-Cas9 nuclease with expanded targeting space. Science 361, 1259–1262 (2018).
Article CAS PubMed PubMed Central Google Scholar
Hu, J. H. et al. Evolved Cas9 variants with broad PAM compatibility and high DNA specificity. Nature 556, 57–63 (2018).
Article CAS PubMed PubMed Central Google Scholar
Zhang, C. et al. Expanding base editing scope to near-PAMless with engineered CRISPR/Cas9 variants in plants. Mol. Plant 14, 191–194 (2021).
Article CAS PubMed Google Scholar
Ren, Q. et al. PAM-less plant genome editing using a CRISPR-SpRY toolbox. Nat. Plants 7, 25–33 (2021).
Article CAS PubMed Google Scholar
Wu, Y. et al. Genome-wide analyses of PAM-relaxed Cas9 genome editors reveal substantial off-target effects by ABE8e in rice. Plant Biotechnol. J. 20, 1670–1682 (2022).
Article CAS PubMed PubMed Central Google Scholar
Walton, R. T., Christie, K. A., Whittaker, M. N. & Kleinstiver, B. P. Unconstrained genome targeting with near-PAMless engineered CRISPR-Cas9 variants. Science 368, 290–296 (2020).
Article CAS PubMed PubMed Central Google Scholar
Liao, J. et al. Therapeutic adenine base editing of human hematopoietic stem cells. Nat. Commun. 14, 207 (2023).
Article CAS PubMed PubMed Central Google Scholar
Lebek, S. et al. Ablation of CaMKIIdelta oxidation by CRISPR-Cas9 base editing as a therapy for cardiac disease. Science 379, 179–185 (2023).
Article CAS PubMed PubMed Central Google Scholar
Yin, S. et al. Engineering of efficiency-enhanced Cas9 and base editors with improved gene therapy efficacies. Mol. Ther. 31, 744–759 (2023).
Article CAS PubMed Google Scholar
Li, J. et al. Functional phosphoproteomics in cancer chemoresistance using CRISPR-mediated base editors. Adv. Sci. 9, e2200717 (2022).
Article Google Scholar
Landrum, M. J. et al. ClinVar: public archive of interpretations of clinically relevant variants. Nucleic Acids Res. 44, D862–D868 (2016).
Article CAS PubMed Google Scholar
Arbab, M. et al. Determinants of base editing outcomes from target library analysis and machine learning. Cell 182, 463–480 e430 (2020).
Article CAS PubMed PubMed Central Google Scholar
Song, M. et al. Sequence-specific prediction of the efficiencies of adenine and cytosine base editors. Nat. Biotechnol. 38, 1037–1043 (2020).
Article CAS PubMed Google Scholar
Fan, J. et al. Annotation and evaluation of base editing outcomes in multiple cell types using CRISPRbase. Nucleic Acids Res. 51, D1249–D1256 (2023).
Article CAS PubMed Google Scholar
Pallaseni, A. et al. Predicting base editing outcomes using position-specific sequence determinants. Nucleic Acids Res. 50, 3551–3564 (2022).
Article CAS PubMed PubMed Central Google Scholar
Zhang, W. et al. In-depth assessment of the PAM compatibility and editing activities of Cas9 variants. Nucleic Acids Res. 49, 8785–8795 (2021).
Article CAS PubMed PubMed Central Google Scholar
Kim, N. et al. Deep learning models to predict the editing efficiencies and outcomes of diverse base editors. Nat. Biotechnol. https://doi.org/10.1038/s41587-023-01792-x (2023).
Marquart, K. F. et al. Predicting base editing outcomes with an attention-based deep learning algorithm trained on high-throughput target library screens. Nat. Commun. 12, 5114 (2021).
Article CAS PubMed PubMed Central Google Scholar
Park, J. & Kim, H. K. Prediction of base editing efficiencies and outcomes using DeepABE and DeepCBE. Methods Mol. Biol. 2606, 23–32 (2023).
Article CAS PubMed Google Scholar
Carrington, B., Weinstein, R. N. & Sood, R. BE4max and AncBE4max are efficient in germline conversion of C:G to T:A base pairs in zebrafish. Cells 9, https://doi.org/10.3390/cells9071690 (2020).
Zafra, M. P. et al. Optimized base editors enable efficient editing in cells, organoids and mice. Nat. Biotechnol. 36, 888–893 (2018).
Article CAS PubMed PubMed Central Google Scholar
Kim, H. K. et al. High-throughput analysis of the activities of xCas9, SpCas9-NG and SpCas9 at matched and mismatched target sequences in human cells. Nat. Biomed. Eng. 4, 111–124 (2020).
Article CAS PubMed Google Scholar
Kim, H. K. et al. SpCas9 activity prediction by DeepSpCas9, a deep learning-based model with high generalization performance. Sci. Adv. 5, eaax9249 (2019).
Article CAS PubMed PubMed Central Google Scholar
Ryu, J. et al. Joint genotypic and phenotypic outcome modeling improves base editing variant effect quantification. Nat. Genet. 56, 925–937 (2024).
Article CAS PubMed Google Scholar
Mo, J., Moye, S. L., McKay, R. M. & Le, L. Q. Neurofibromin and suppression of tumorigenesis: beyond the GAP. Oncogene 41, 1235–1251 (2022).
Article CAS PubMed PubMed Central Google Scholar
Chen, P. J. et al. Enhanced prime editing systems by manipulating cellular determinants of editing outcomes. Cell 184, 5635–5652 e5629 (2021).
Article CAS PubMed PubMed Central Google Scholar
Nelson, J. W. et al. Engineered pegRNAs improve prime editing efficiency. Nat. Biotechnol. 40, 402–410 (2022).
Article CAS PubMed Google Scholar
Huang, T. P. et al. High-throughput continuous evolution of compact Cas9 variants targeting single-nucleotide-pyrimidine PAMs. Nat. Biotechnol. 41, 96–107 (2023).
Article CAS PubMed Google Scholar
Cao, X. et al. Engineering of near-PAMless adenine base editor with enhanced editing activity and reduced off-target. Mol. Ther. Nucleic Acids 28, 732–742 (2022).
Article CAS PubMed PubMed Central Google Scholar
Dang, Y. et al. Optimizing sgRNA structure to improve CRISPR-Cas9 knockout efficiency. Genome Biol. 16, 280 (2015).
Article PubMed PubMed Central Google Scholar
Xu, L., Liu, Y. & Han, R. BEAT: A Python Program to Quantify Base Editing from Sanger Sequencing. CRISPR J. 2, 223–229 (2019).
Article CAS PubMed PubMed Central Google Scholar
John, J. S. SeqPrep, (2016) https://github.com/jstjohn/SeqPrep
Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25, 1754–1760 (2009).
Article CAS PubMed PubMed Central Google Scholar
Fennell, T. et al. CALITAS: A CRISPR-Cas-aware ALigner for In silico off-TArget Search. CRISPR J. 4, 264–274 (2021).
Article CAS PubMed Google Scholar
Derry, A., Krzywinski, M. & Altman, N. Convolutional neural networks. Nat. Methods 20, 1269–1270 (2023).
Article CAS PubMed Google Scholar
Chadha, G. S., Panambilly, A., Schwung, A. & Ding, S. X. Bidirectional deep recurrent neural networks for process fault classification. ISA Trans. 106, 330–342 (2020).
Article PubMed Google Scholar
Xiao, L. M., Wan, Y. Q. & Jiang, Z. R. AttCRISPR: a spacetime interpretable model for prediction of sgRNA on-target activity. BMC Bioinforma. 22, 589 (2021).
Article CAS Google Scholar
O’Malley, T. a. B., Elie and Long, James and Chollet, François and Jin, Haifeng and Invernizzi, Luca and others. KerasTuner. (2019) https://github.com/keras-team/keras-tuner
Li, H. et al. The sequence alignment/map format and SAMtools. Bioinformatics 25, 2078–2079 (2009).
Article PubMed PubMed Central Google Scholar
Zhou, X., et al. Wangxiaoyue-lab/BEguider: DOI created. (2024) https://doi.org/10.5281/zenodo.10893907

Download references

Acknowledgements

We thank Dr. Weiwei Zhang at the Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences for helpful discussion. We thank the Center for Bioinformatics at Institute for Basic Medical Science, and Center for Bioinformatics at the National Infrastructures for Translational Medicine at Peking Union Medical College Hospital for their invaluable support in providing high-performance computing services. This work was granted by National Natural Science Foundation of China (32122023 and 32070603 to X.W.) and National High Level Hospital Clinical Research Funding (2023-PUMCH-E-008 to X.W.).

Author information

These authors contributed equally: Xiaoyu Zhou, Jingjing Gao, Liheng Luo.

Authors and Affiliations

State Key Laboratory of Common Mechanism Research for Major Diseases; Center for Bioinformatics, National Infrastructures for Translational Medicine, Institute of Clinical Medicine & Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China
Xiaoyu Zhou, Liheng Luo, Jiayu Wu & Xiaoyue Wang
State Key Laboratory of Common Mechanism Research for Major Diseases, Department of Biochemistry and Molecular Biology, Institute of Basic Medical Sciences Chinese Academy of Medical Sciences, School of Basic Medicine, Peking Union Medical College, Beijing, China
Jingjing Gao & Changcai Huang
Department of Clinical Laboratory Medicine, Wenzhou Central Hospital, Wenzhou, China
Jingjing Gao

Authors

Xiaoyu Zhou
View author publications
Search author on:PubMed Google Scholar
Jingjing Gao
View author publications
Search author on:PubMed Google Scholar
Liheng Luo
View author publications
Search author on:PubMed Google Scholar
Changcai Huang
View author publications
Search author on:PubMed Google Scholar
Jiayu Wu
View author publications
Search author on:PubMed Google Scholar
Xiaoyue Wang
View author publications
Search author on:PubMed Google Scholar

Contributions

Xiaoyue Wang, Xiaoyu Zhou, Jingjing Gao, and Changcai Huang conceived and designed the project. Xiaoyu Zhou and Jiayu Wu performed experiments. Jingjing Gao and Liheng Luo performed bioinformatics analyses. Xiaoyue Wang supervised the project. Xiaoyu Zhou, Jingjing Gao, Liheng Luo, and Xiaoyue Wang wrote the manuscript.

Corresponding author

Correspondence to Xiaoyue Wang.

Ethics declarations

Competing interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used generative AI for language correction and editing suggestions. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the content of the published article.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Mengtan Xing. [A peer review file is available.]

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Transparent Peer Review file

Supplementary Information

Supplementary Tables 1-20

Reporting Summary

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Cite this article

Zhou, X., Gao, J., Luo, L. et al. Comprehensive evaluation and prediction of editing outcomes for near-PAMless adenine and cytosine base editors. Commun Biol 7, 1389 (2024). https://doi.org/10.1038/s42003-024-07078-5

Download citation

Received: 18 April 2024
Accepted: 14 October 2024
Published: 25 October 2024
Version of record: 25 October 2024
DOI: https://doi.org/10.1038/s42003-024-07078-5