Sensitive and unbiased genome-wide profiling of base-editor-induced off-target activity using CHANGE-seq-BE

Base editors^1,2 are a transformative genome-editing technology that can precisely install single-base alterations into cellular genomic DNA (gDNA) without requirements for DNA double-strand breaks (DSBs) or exogenous DNA donor templates or reliance on cellular homology-directed repair pathways. For example, adenine base editors (ABEs) and cytosine base editors (CBEs) are capable of reverting G•C to A•T and T•A to C•G mutations, respectively, which represent the most common pathogenic single-nucleotide variants reported in the ClinVar database (~47% and 14%)³, highlighting the enormous potential of this class of genome editors for therapeutic application.

However, like other genome editors, base editors can bind and modify off-target genomic loci with sequence homology to the target protospacer^4,5,6. As base editors rapidly advance toward clinical applications, defining guide RNA (gRNA)-dependent off-target activity of base editors in a sensitive and unbiased way is important for assessing the safety of these new therapeutic approaches⁷. A method that is unbiased is one that does not preselect candidate off-target sequences for experimental analysis.

Current methods for profiling the genome-wide activity of base editors^8,9,10 have certain limitations. Digenome-seq⁸ and EndoV-seq⁹ are based on whole-genome sequencing (WGS) of base-editor-treated gDNA in vitro, followed by processing with end repair enzymes to generate DSBs detectable by high-throughput sequencing. As these methods do not enrich for base-editor-modified gDNA, they require hundreds of millions of sequencing reads, reducing sensitivity and scalability and increasing costs. The apparent validation rate of these methods for identifying bona fide cellular off-target base-editing mutations is relatively low^8,9 and the high sequencing depth requirements limit their application to many targets. Another technique, ONE-seq, relies on biased computational preselection of candidate off-target sites and, therefore, cannot detect genome-wide off-target activity beyond those sites selected a priori.

Sensitive and unbiased approaches offer the unique advantage of detecting unintended on-target genome editor off-target activity caused by ‘unknown unknowns’ such as contaminant gRNAs.

In earlier studies, we and others applied indirect methods to identify base editor off-target activity^{11,12,13,14,15}, relying on the simplifying but not necessarily true assumption that base editor off-target effects are a subset of nuclease off-target effects. For example, to experimentally nominate candidate off-target sites for an ABE therapeutic strategy to treat sickle cell disease, we performed CIRCLE-seq^16,17 (an earlier method we developed for selective sequencing of Cas9-cleaved gDNA) using Cas9 nuclease to nominate candidate off-target sites. We then measured editing activity at CIRCLE-seq Cas9-nominated off-target sites in a base-editing context in CD34⁺ hematopoietic stem and progenitor cells (HSPCs) treated with ABE8e-NRCH:HBB^S-gRNA using multiplex targeted sequencing. We identified 54 Cas9-nuclease-nominated off-target sites with unintended off-target base-editing activity as high as 82%¹¹.

Although gRNA-dependent base-editing off-target sites often overlap with Cas9 nuclease off-target sites^8,9,18, productive base editing requires additional criteria, such as the presence of the target base within the editing window, base edited nucleotide sequence context and R-loop accessibility by deaminases, which may not be satisfied for all Cas-nuclease-dependent off-target sites¹⁹. Thus, limitations of our initial nuclease-centric approach are that base editors may have a different spectrum of off-target activity than cognate Cas9 nucleases, highlighting an unmet need for a simultaneously sensitive and unbiased method for specifically defining base editor genome-wide off-target activity. Unbiased methods to characterize gene therapy products that use base editing to gain regulatory approval for clinical trials should be performed with the relevant base editor and not a similar nuclease.

Previously, we developed CHANGE-seq²⁰, a sensitive, unbiased and high-throughput method to identify Cas9 nuclease off-target activity in vitro. CHANGE-seq leverages a Tn5 tagmentation-based workflow for efficient generation of circularized gDNA libraries, coupled with stringent selection of circular DNA molecules using a cocktail of exonucleases, thereby generating a population of circular DNA molecules with minimal free DNA ends. Covalently closed circular DNA molecules are treated with CRISPR–Cas9 ribonucleoprotein (RNP) complexes and nuclease-cleaved DNA fragments are then ligated to adaptors and sequenced, enabling the selective sequencing of Cas9-cleaved DNA molecules.

Here, we present CHANGE-seq-BE, a method for both sensitive and unbiased genome-wide discovery of gRNA-dependent base editor on-target and off-target sites, adapted from CHANGE-seq. After extensive molecular biology optimization, we overcame substantial technical challenges in enriching for editor-modified DNA reported by us and others^10,20. CHANGE-seq-BE now enables the selective and efficient sequencing of base-editor-modified gDNA for both ABEs and CBEs.

We applied CHANGE-seq-BE to define the gRNA-dependent genome-wide off-target activity of ABE8e and CBE eA3A-BE3 at therapeutically relevant genes in primary human cells. We found that CHANGE-seq-BE sensitively identifies not only off-target activity but also unintended on-target activity of contaminating synthetic gRNAs, highlighting one of the advantages of using unbiased sensitive methods for genome-wide assessment of genome-editing activity. In direct comparisons of ABEs and Cas9 nuclease at the same targets, we found substantially higher off-target editing activity with ABE8e base editors. Lastly, we used CHANGE-seq-BE to support genotoxicity studies in an emergency investigational new drug (IND) application for an adenine-base-editing treatment for a person with CD40L-deficient X-linked hyper IgM (X-HIGM) syndrome.

Development of a sensitive and unbiased biochemical method to measure base editor genome-wide activity

Motivated by a critical need for better methods to directly characterize the gRNA-dependent genome-wide off-target activity of base editors, we sought to develop a method that would be simultaneously sensitive, unbiased, sequencing efficient and straightforward to practice. We reasoned that CHANGE-seq could be adapted for base editors by treating purified, circularized gDNA with base editor RNP complexes and selectively sequencing base-editor-modified gDNA (Fig. 1a). First, linear gDNA could be circularized using our previously described CHANGE-seq Tn5 tagmentation-mediated approach²⁰. Second, for ABEs, enzymatically purified gDNA circles could be treated with ABE RNP complexes to nick the target DNA strand and deaminate adenine bases to inosine (within the base-editing window) on the nontarget strand at on-target and off-target sites. Third, nicked and inosine-containing gDNA circles could be further processed with endonuclease V (ref. ²¹), which cleaves DNA adjacent to inosines to produce linear DNA with 5′ staggered ends. Similarly, for CBEs, circularized gDNA could be treated with CBE RNP complexes to nick the target DNA strand and deaminate cytosine bases to uracil on the complementary DNA strand. Uracil-containing gDNA circles could be further processed with USER enzyme (uracil-specific excision reagent), a mixture of uracil DNA glycosylase and the DNA glycosylase–lyase endonuclease VIII, which cleaves and removes uracil in the nontarget DNA strand to generate linearized DNA. The base-editor-modified linear DNA processed by endonuclease V or USER enzyme could be end-repaired and ligated to sequencing adaptors for selective high-throughput sequencing of base-editor-modified gDNA molecules.

a, Schematic of CHANGE-seq-BE workflow. Purified gDNA is tagmented with custom Tn5 transposome containing circularization adaptor DNA. Tagmented gDNA is circularized by intramolecular ligation. Residual linear DNA is enzymatically degraded with an exonuclease cocktail, leaving highly pure gDNA circles. Circularized gDNA is treated in vitro with ABE or CBE RNP complexes to nick the target DNA strand and deaminate adenine bases to inosine or cytosine bases to uracil (within the base-editing window) on the nontarget strand at on-target and off-target sites (red). For ABEs, nicked inosine-containing gDNA circles are then processed with endonuclease V, which cleaves DNA adjacent to inosines. For CBEs, uracil-containing DNA circles are treated with USER, which excises the uracil base to produce linear DNA with 5′ staggered ends. The 5′ overhangs are filled in with Klenow exo-, end-repaired, ligated to sequencing adaptors and amplified for selective high-throughput sequencing. b, Visualization of detected off-target sites identified by CHANGE-seq-BE aligned against the intended target site for ABE:gRNA (ABE8e-NRCH:HBB^S-gRNA) complexes targeted against HBB. The intended target sequence is shown in the top line and off-target sites are ordered from top to bottom by CHANGE-seq-BE read count. Mismatches to the intended target sequence are indicated by colored nucleotides; matches are shown as dots and read counts and A-to-I (G) deamination frequencies (%) are shown at the end of each line (only the top off-target sites are listed). Representative alignment of CHANGE-seq-BE reads of the ABE8e-NRCH:HBB^S-gRNA target, as visualized by the Integrative Genomics Viewer (IGV) for an off-target site on chr2. The target adenine base is highlighted in red, the protospacer-adjacent motif (NRCH) is shown in bold and the single-stranded DNA breaks are marked with colored arrows by nCas9 (purple) and endonuclease V (red). c, Visualization plot of off-target sites aligned against the intended target site for selected eA3A-BE3:RNF2 complexes. The intended target sequence is shown in the top line and off-target sites are ordered by descending CHANGE-seq-BE read count. Mismatches to the intended target sequence are indicated by colored nucleotides and the arrow indicates the on-target site. Alignment of CHANGE-seq-BE reads of the eA3A-BE3-NGG:RNF2 gRNA off-target site on chr17, as visualized by the IGV. The target cytosine base in the editing window is highlighted in red, the protospacer-adjacent motif (NGG) is depicted in bold and single-stranded breaks made by USER (red) and nCas9 (purple) are indicated with respective colored arrows.

Full size image

However, a major challenge we initially encountered was that the addition of the required end-repair step substantially decreased the enrichment of editor-modified DNA²⁰. To gauge the performance of our initial CHANGE-seq for base editor approach, we quantified the number of validated off-target sites detected for an ABE8e RNP targeting the HBB gene (ABE8e-NRCH:HBB^S-gRNA) that we previously characterized extensively and validated 54 off-target sites¹¹. At first, we found that addition of an end-repair step to fill in 5′ overhangs and detect base-editor-modified gDNA resulted in high background and detection of only a small fraction (20%) of previously validated off-target sites (Extended Data Fig. 1a). We hypothesized that residual linear DNA molecules could become competent for adaptor ligation after end repair. After systematic optimization of the entire workflow, where we varied circularized gDNA exonuclease treatments, base-editing reaction conditions and end-repair protocols, we identified an optimized protocol (Supplementary Note 1 and Supplementary Protocol) that consistently detected all 54 previously validated ABE8e-NRCH:HBB^S-gRNA off-target sites, which we called CHANGE-seq-BE (Extended Data Fig. 1a).

For ABEs, CHANGE-seq-BE reads capture a distinct molecular signature of base editing in vitro, evidence of A-to-I deamination as inosine bases are converted to guanine after PCR amplification, increasing confidence in sites nominated by our method. We observed a range of deamination frequencies of 25–48% in top ABE8e-NRCH:HBB^S-gRNA off-target sites nominated by CHANGE-seq-BE (Fig. 1b).

CHANGE-seq-BE is also readily adaptable to CBEs, as long as high-quality, purified CBE proteins with efficient biochemical activity are available. To evaluate the activity of CBEs, we first performed in vitro cleavage profiling of a PCR amplicon using AnBE4max, A3A-BE3, eA3A-BE3, Td-CGBE and Td-CBEmax^22,23,24 (Extended Data Fig. 1b) RNP complexes, followed by USER treatment to create DSBs. On the basis of protein activity, quality and purity, we selected eA3A-BE3 that showed 95% cleavage activity to further test with CHANGE-seq-BE. In contrast to ABEs, the USER enzyme recognizes and excises deaminated uracil bases from the noncomplementary strand and, thus, C-to-T base conversion is not detectable in CHANGE-seq-BE reads, although the breakpoint positions remain clear (Fig. 1c).

CHANGE-seq-BE identifies new bona fide off-target sites for ABE8e-NRCH:HBB ^S target

To initially evaluate technical reproducibility, we performed independent CHANGE-seq-BE library preparations and found that read counts from technical replicates were strongly correlated (Pearson’s correlation coefficient, r = 0.9814) (Fig. 2a), with sites of off-target activity distributed broadly throughout the genome (Fig. 2b and Supplementary Table 1). For this gRNA target, the on-target site did not have the highest CHANGE-seq-BE read counts, consistent with the lower specificity profile we observed.

a, Scatter plot showing read count correlation for two CHANGE-seq-BE library preparations of ABE8e-NRCH targeting HBB^S. b, Manhattan plot showing the genome-wide distribution of CHANGE-seq-BE-detected sites for ABE8e-NRCH:HBB^S-gRNA. The pink arrow indicates the on-target site. c, Venn diagram showing overlap between candidate off-target sites nominated by CHANGE-seq-BE for ABE8e-NRCH:HBB^S-gRNA and nominated off-target sites for which off-target editing was observed by multiplex targeted sequencing in CD34⁺ cells (from a person with sickle cell disease) nucleofected with ABE8e-NRCH mRNA. d, Dot plot showing the percentage of sequencing reads containing A-to-G base editing within the protospacer positions 4–10 at off-target sites in gDNA samples derived from CD34⁺ HSPCs treated with ABE8e-NRCH mRNA (red, n = 2), or untreated controls (gray, n = 2). χ² tests were performed between mRNA-treated and control samples. The FDR was calculated using the Benjamini–Hochberg method. Reported off-target sites were considered statistically significant with FDR ≤ 0.05 and difference ≥0.5% between treated and untreated controls. e, Dot plot showing the enrichment ratio for sites detected by CHANGE-seq-BE using ABE8e or for CIRCLE-seq using cognate Cas9 nuclease. Statistical analysis was conducted using an unpaired t-test (two-tailed) with P = 0.0098.

Full size image

To determine whether CHANGE-seq-BE identifies base editor-specific off-target activity not detected by nuclease-specific methods, we compared the sites identified by CHANGE-seq-BE and CIRCLE-seq using ABE8e-NRCH base editor and Cas9-NRCH nuclease, respectively. We noted that CHANGE-seq-BE and CIRCLE-seq read counts were only weakly correlated (r = 0.34) (Extended Data Fig. 2a) and only 40% of the off-target sites detected by both methods overlapped (Extended Data Fig. 2b), suggesting that ABE8e genome-wide off-target activity quantitatively differs from Cas9 nuclease off-target activity. One explanation for some of these differences is the number and location of editable adenine bases within on-target and off-target sites, as off-target sites with mismatches of the editable adenine bases could eliminate base-editing activity without affecting nuclease activity.

To confirm that new off-target sites detected exclusively by CHANGE-seq-BE are bona fide off-target sites in cells, we performed multiplex targeted sequencing on 131 sites detected exclusively by CHANGE-seq-BE (Supplementary Table 2) using the same gDNA of CD34⁺ HSPCs treated with ABE8e-NRCH mRNA evaluated in our previous study¹¹. We found that CHANGE-seq-BE identified all 54 known cellular off-target sites, as well as an additional 29 previously unknown bona fide off-target sites, 53% more than our initial screen with CIRCLE-seq (Fig. 2c), with ABE activity ranging from 0.55% to 13.9% (Fig. 2d). For the HBB^S-gRNA bona fide off-target sites identified by CIRCLE-seq using Cas9 nuclease and CHANGE-seq-BE using ABE8e, we noted that the CHANGE-seq-BE enrichment ratio was significantly higher than that of CIRCLE-seq, by an average of 2.1-fold (Fig. 2e). HBB^S-gRNA bona fide off-target activity occurred primarily in intergenic and intronic regions and six off-target sites were in exons (Extended Data Fig. 2c). Taken together, our results demonstrate the importance of using sensitive and unbiased base-editor-specific off-target discovery methods to comprehensively identify ABE genome-wide off-target activity.

CHANGE-seq-BE sensitively identifies bona fide off-target sites in human primary cells

To systematically assess the genome-wide off-target activity of ABEs and CBEs in vitro, we performed CHANGE-seq-BE on five therapeutically relevant target sites in human primary T cells (B2M, CBLB, CD7, CIITA and PDCD1)²⁵ and in human primary hepatocytes (PCSK9)²⁶ with ABE8e (Fig. 3a and Extended Data Fig. 3a). Overall, CHANGE-seq-BE nominated a range of 236 to 1,604 on-target and off-target sites (total of 4,199) for B2M, CBLB, CD7, CIITA, PDCD1 and PCSK9 across the genome in two experimental replicates (Fig. 3b, Extended Data Fig. 3b and Supplementary Table 3).

a, Schematic of CHANGE-seq-BE off-target nomination at five targets in primary human T cells (B2M, CBLB, CD7, CIITA and PDCD1). b, Manhattan plot showing the genome-wide distribution of CHANGE-seq-BE-detected sites. The arrow (pink) indicates the on-target site. c, Venn diagrams depicting the overlap of off-target sites nominated by CHANGE-seq-BE and Digenome-seq and validated sites by hybrid capture sequencing. d, Bona fide cellular off-target sites detected by hybrid capture sequencing for the sites nominated by CHANGE-seq-BE and Digenome-seq, including CHANGE-seq-BE read counts (columns). Dot plot showing the percentage of sequencing reads containing A-to-G mutations within the protospacer positions 1–10 at off-target sites in gDNA samples from primary human T cells treated with ABE8e-NGG mRNA (red, n = 3) or untreated controls (gray, n = 3). χ² tests were performed between mRNA-treated and control samples. The FDR was calculated using the Benjamini–Hochberg method. Reported off-target sites were considered statistically significant with FDR ≤ 0.05 and difference ≥0.5% between treated and untreated controls. e, Stacked bar plot representing the number of validated off-target sites exclusively identified by CHANGE-seq-BE (yellow), Digenome-seq (purple) or both (blue). f, Genomic features of validated off-target sites for six ABE targets (B2M, CBLB, CD7, CIITA, PDCD1 and PCSK9). TTS, transcription termination site; UTR, untranslated region. The proportion of transcribed regions in the validated sites was statistically significant compared to CHANGE-seq-BE-nominated sites using Fischer’s exact test (two-sided) with P = 0.0264. Panel a created with BioRender.com.

Full size image

To compare the sensitivity of CHANGE-seq-BE to another leading off-target discovery method, we performed Digenome-seq on the same gRNA targets and evaluated them with matching analysis parameters. We identified a total of 796 on-target and off-target sites (ranging from 36 to 162 sites) for B2M, CBLB, CD7, CIITA and PDCD1. Next, we comprehensively evaluated 4,291 sites by hybrid capture sequencing for the sites nominated by either of the methods on gDNA obtained from edited cells compared to unedited controls. At these sites that were successfully captured and amplified, 96 sites showed significant off-target alterations (Supplementary Table 4), with cellular activity ranging from 0.44% to 93.1% (Fig. 3c,d and Extended Data Fig. 3c,d). At 83 (of 96) bona fide T cell off-target sites, 33.3–100% of sites were identified by both methods, 6.9–16.7% of sites were identified by Digenome-seq and 29.3–55.6% of sites were identified exclusively by CHANGE-seq-BE (Fig. 3e). Additional analysis of off-target sites with relaxed parameters found an additional six validated sites (Supplementary Note 2 and Supplementary Table 5).

To better understand the functional consequences of off-target activity for ABE targets, we annotated and analyzed the genomic locations of off-target sites by CHANGE-seq-BE and Digenome-seq. The proportion of nominated off-target sites found across various genomic regions (intronic, intergenic, transcribed and promoter) was similar between CHANGE-seq-BE and Digenome-seq (Extended Data Fig. 3e). Both nominated and validated off-target sites occurred predominantly in intergenic and intronic regions, consistent with their abundant representation within the human genome. Compared to the proportion of CHANGE-seq-BE-nominated off-target sites (10.3%), validated cellular off-target sites (17.7%) were significantly enriched in transcribed regions (Fig. 3f).

Similarly, we performed CHANGE-seq-BE and Digenome-seq using eA3A-BE3 to target B2M, CBLB, CD7, CIITA and PDCD1 regions at different loci. We identified a total of 850 on-target and off-target sites (ranging from 61 to 198) by CHANGE-seq-BE and 25 sites (ranging from 3 to 11) by Digenome-seq for B2M, CBLB, CD7, CIITA and PDCD1 targets, respectively (Fig. 4a,b and Supplementary Table 6). By hybrid capture sequencing, we evaluated 691 sites nominated by CHANGE-seq-BE and Digenome-seq for all five gRNA targets. Both methods identified on-target editing for all targets with cellular activity ranging from 52.3% to 97.6%. Both methods also detected a high-frequency off-target site (OT01) for CIITA that has a wobble mismatch to the intended target with 17% off-target editing. However, many low-frequency off-target sites with editing rates ranging from 1.7% to 7.5% were not identified by Digenome-seq and were only discovered by CHANGE-seq-BE (Fig. 4c and Supplementary Table 7).

a, Manhattan plot showing the genome-wide distribution of off-target sites detected by CHANGE-seq-BE for the targets B2M, CBLB, CD7, CIITA and PDCD1. The arrow (pink) indicates the on-target site. b, Venn diagram depicting the overlap of off-target sites detected by CHANGE-seq-BE, Digenome-seq and validated off-target sites. c, Bona fide cellular off-target sites by hybrid capture sequencing for the sites nominated by CHANGE-seq-BE and Digenome-seq, including CHANGE-seq-BE read counts (columns) with sequences. Dot plot showing the percentage of sequencing reads containing C-to-any mutations within the protospacer positions 1–10 at off-target sites in gDNA derived from primary human T cells treated with eA3A-BE3 mRNA (red) or untreated controls (gray) (n = 2). χ² tests were performed between mRNA-treated and control samples. The FDR was calculated using the Benjamini–Hochberg method. Reported off-target sites were considered statistically significant with FDR ≤ 0.05 and difference ≥0.5% between treated and controls.

Full size image

In sum, our data show that CHANGE-seq-BE is more sensitive than Digenome-seq while requiring 20-fold fewer sequencing reads. CHANGE-seq-BE is a strong predictor of base editor cellular off-target activity in human primary cells and can be used to select candidate target sites for both routine and therapeutic applications.

CHANGE-seq-BE identified unexpected synthetic gRNA contaminants

Safety risks specifically associated with human genome-editing products include unintended on-target and off-target activity and their long-term consequences. One less obvious but key concern is that contamination in a critical genome-editing component such as the gRNA could direct genome editing to unintended on-target sites. Chemically modified synthetic gRNAs are now a commonly used product in genome-editing experiments, particularly in primary cells, because of enhanced activity and stability²⁷.

In initial pilots of CHANGE-seq-BE experiments using research-grade, chemically synthesized and high-performance liquid chromatography (HPLC)-purified gRNAs, we observed CHANGE-seq-BE read counts across multiple targets in base-editor-treated samples but not in untreated controls, suggesting activity of contaminating synthetic gRNAs (Fig. 5a). We detected strong, unintended activity at PDCD1, CD7 and CIITA target sites in gDNA samples treated with the CBLB gRNA target site; similarly, PDCD1 and CD7 base editing mutation-containing reads were found in gDNA samples treated with CIITA gRNA and B2M and CD7 unintended on-target edited reads were found in gDNA samples treated with PDCD1 gRNA. In hindsight, according to manufacturer discussions, the most likely source of this cross-contamination during manufacturing was the sequential purification of these gRNAs over the same HPLC column. Detectable cross-contaminating activity was mitigated by a more stringent, column clearing-in-place procedure used in a second batch of gRNAs from the same manufacturer.

a, Bar plot showing CHANGE-seq-BE read counts for three synthetic chemically modified gRNAs (CBLB, CIITA and PDCD1) from the manufacturer supplier A (batch 1) and the unexpected detection of gRNA contaminants (red). Data are shown as the mean (n = 2). b, SMARTer gRNA-sequencing workflow. c, Alignment plot for CBLB gRNA target displaying top 20 mapped reads. Mismatched nucleotides in the target sequence are highlighted in colors (blue, green, red and yellow) representing potential contaminants CIITA and PDCD1. The arrow highlighted in the scaffold region (green) shows the scaffold specific primer used for first cDNA synthesis. d, Bar plot showing the percentage of sequencing reads mapping to target gRNAs, for one specific synthetic gRNA supplier A (batch 1). The activity of gRNA contaminants is shown in red. e, Bar plot showing the percentage of sequencing reads containing A-to-G mutations within the protospacer positions 4–10 at off-target sites in gDNA samples from primary human T cells treated with ABE8e in the mRNA format and gRNAs targeting CBLB or CIITA, as well as the detection of contaminant gRNAs (red) during the manufacture process. Data are shown as the mean. Unpaired t-tests (two-sided) were performed using the Holm–Šídák multiple-comparisons method (n = 3); adjusted P = 0.0013 (CBLB on-target), 0.0013 (PDCD1), 0.00002 (CIITA), <0.000001 (CIITA on-target), 0.0014 (PDCD1) and 0.0187 (CD7).

Full size image

To identify and measure the frequency of contaminating gRNAs, we used a gRNA-sequencing assay²⁸ based on an established small RNA-sequencing technology (Fig. 5b). We then evaluated five research-grade gRNA sequences from three synthetic gRNA manufacturer suppliers (A, B and C) for the presence of gRNA contaminants (Extended Data Fig. 4a). We detected the presence of contaminating gRNAs in CBLB and CIITA gRNAs from supplier A (batch 1), with frequency ranging from 0.09% to 2.71% (Fig. 5c,d).

To determine whether the relatively minute amounts of gRNA contaminants detected in vitro by CHANGE-seq-BE activity and gRNA sequencing would be active in cells, we performed targeted sequencing of gDNA from primary human T cells edited with ABE8e, with gRNAs manufactured by supplier A targeting CBLB and CIITA loci. For cells with intended CBLB edits, we additionally evaluated editing at PDCD1 and CIITA on-target sites and, for cells with intended CIITA editing, we also evaluated PDCD1 and CD7 on-target sites by targeted sequencing. We detected point mutations consistent with adenine base editing in all unintended on-target sites associated with contaminating gRNAs evaluated, with activity ranging from 0.83% to 5.24% (Fig. 5e).

Our results highlight one additional feature of using sensitive and unbiased genome-wide methods for base editor off-target discovery is identification of unintended ‘on-target’ activity of contaminant gRNAs. This may be a pervasive issue when multiple chemically synthesized gRNAs are purified using the same chromatography supplies and caution is warranted to minimize gRNA synthesis contamination risk, particularly for therapeutic applications.

CHANGE-seq-BE reveals biological activity differences between nucleases and base editors

To evaluate biochemical activity differences between base editors (ABE8e) and nucleases (Cas9), we additionally performed CHANGE-seq (with Cas9 nuclease) on gDNA from human primary T cells for the same five gRNAs (B2M, CBLB, CD7, CIITA and PDCD1) we previously evaluated with ABEs. A total of 7,430 on-target and off-target sites were detected by CHANGE-seq for all five targets (in contrast to 3,734 sites by CHANGE-seq-BE; Supplementary Table 3). Read counts from experimental replicates were strongly correlated for CHANGE-seq (r > 0.86) and CHANGE-seq-BE (r > 0.96) (Extended Data Fig. 5a,b). However, across methods, CHANGE-seq and CHANGE-seq-BE read counts were less correlated and varied depending on the gRNA target (r = 0.2–0.73) (Fig. 6a). Our results show that the biochemical activities of ABE8e base editor and Cas9 are both highly reproducible and distinct.

a, Scatter plot showing read counts correlation between CHANGE-seq-BE and CHANGE-seq for five ABE targets (B2M, CBLB, CD7, CIITA and PDCD1). b, Sequences of bona fide validated off-target sites for five gRNA targets in hybrid capture sequencing. The heat map represents the percentage of A-to-G editing by ABE8e base editor or indels by Cas9 nuclease (hybrid capture sequencing) or normalized off-target reads relative to on-target reads (GUIDE-seq-2). Cas9-mediated cellular off-target sites are highlighted in red box. χ² tests were performed between mRNA-treated and control samples. The FDR was calculated using the Benjamini–Hochberg method. Reported off-target sites were considered statistically significant with FDR ≤ 0.05 and difference ≥0.5% between treated (red box) and controls. c, Stacked bar plot representing the total of number of validated off-target sites for ABE8e and Cas9 mRNA edited T cells.

Full size image

To directly assess the cellular off-target activity differences between Cas9 nuclease and ABEs, we edited T cells for all five targets with Cas9 mRNA and performed hybrid capture targeted sequencing on a total of 2,226 off-target sites (including 1,436 off-target sites nominated only by Cas9 CHANGE-seq). High-frequency on-target editing was achieved for both ABE8e (96.5–99.4% A-to-G base editing; Fig. 3d) and Cas9 (70.7–92% indels) for all five targets. A total of 83 bona fide T cell off-target sites with base editing frequencies ranging from 0.44% to 93.1% were identified for ABE8e across all five targets. In contrast, only one bona fide off-target site with an average of 2.4% indels was confirmed for Cas9 targeting PDCD1 (Fig. 6b,c).

To independently confirm the editing activity measured by hybrid capture sequencing, we performed GUIDE-seq-2 (ref. ²⁹) with Cas9 delivered as a mRNA for all five targets (Extended Data Fig. 6a). Under matching delivery conditions, we observed GUIDE-seq-2 off-target activity profiles that were consistent with our hybrid capture targeted sequencing results (Fig. 6b), where only OT01 for PDCD1 was clearly above the lower limits of detection of GUIDE-seq-2 (Extended Data Fig. 6b). GUIDE-seq-2 with Cas9 RNP resulted in slightly more but still fairly low off-target activity, ranging from 0.01% to 3.4%.

Overall, base editors have different off-target editing profiles when directly compared to cognate nucleases. For example, our comparisons of Cas9 and ABE8e under controlled delivery conditions demonstrate that ABE8e has significantly increased off-target activity compared to Cas9, with 98.8% of validated off-target sites unique to ABE8e. Therefore, it is essential to use direct methods to evaluate the off-target activity of base editors rather than relying on nuclease-based methods only.

CHANGE-seq-BE supports emergency IND for personalized genome-editing treatment

We applied the sensitive off-target detection capability of CHANGE-seq-BE to quantitatively characterize the genome-wide gRNA-dependent off-target profile of an ABE base-editing strategy to treat a person with X-HIGM syndrome by correcting the CD40L mutation (c.658C>T; p.Q220X). Using CHANGE-seq-BE, we identified a total of 81 potential off-target sites across the human genome and experiments were well correlated across two technical replicates (Extended Data Fig. 7a,b). Most of these off-target sites occurred in intronic and intergenic regions and in two exonic regions (Extended Data Fig. 7c). With confirmatory testing by multiplex targeted sequencing using rhAmpSeq, we observed on-target editing of 95.4% with no detectable off-target effects above the limit of threshold compared to controls (Extended Data Fig. 7d). This successfully supported an emergency personalized base editor hematopoietic stem cell and T cell treatment of a person with CD40L HIGM syndrome (NCT06959771), attesting to the IND-enabling role of CHANGE-seq-BE.

Here, we described CHANGE-seq-BE, which enables a simultaneously sensitive and unbiased approach for comprehensive profiling of the genome-wide activity of base editors. CHANGE-seq-BE is a way of directly identifying gRNA-dependent base editor off-target effects that combines the desirable qualities of high sensitivity with minimal experimental bias within a single protocol and works well with both ABEs and CBEs. In this study, we used CHANGE-seq-BE to support an emergency IND application.

Our approach has some limitations. Firstly, CHANGE-seq-BE libraries currently require higher sequencing depth to achieve similar on-target read counts to CHANGE-seq libraries for nucleases. This may be associated with gRNA-independent deamination on both DNA strands or other technical factors, which we plan to explore and optimize further. Secondly, another limitation shared with other biochemical methods is a low apparent validation rate. Future studies, perhaps leveraging artificial intelligence and machine learning, may help to understand and predict the subset of candidates nominated by biochemical methods most likely to be validated in cells. Thirdly, an intrinsic limitation of all biochemical approaches to characterize the genome-wide activity of editors is the requirement for purified proteins. More base-editing proteins are becoming available commercially and, with optimization, miniaturized, high-throughput protein purification approaches may enable characterization of more engineered base editor variants or orthologs. Lastly, CHANGE-seq-BE is not designed to detect gRNA-independent off-target effects, which are more suitably measured in relevant cellular contexts. This may be important to consider for evolved variants that also exhibit higher gRNA-independent deamination activity³⁰.

CHANGE-seq-BE has key advantages over existing methods in terms of sensitivity and sequencing efficiency. In contrast to biochemical methods such as Digenome-seq that rely on WGS of base-editor-modified gDNA, CHANGE-seq-BE has a substantially reduced rate of observed background reads, enabling it to be ~20-fold more sequencing efficient while simultaneously achieving higher sensitivity. The relative sequencing efficiency, requiring ~25 million reads per sample, makes CHANGE-seq-BE accessible to most labs with access to short-read high-throughput sequencers, while enabling the evaluation of more targets in a single sequencing run and the rapid identification of highly active and specific targets. We previously automated CHANGE-seq²⁰ and expect that CHANGE-seq-BE should be similarly automatable.

Another advantage is that CHANGE-seq-BE enables direct profiling of base editor activity, rather than relying on the false assumption that nuclease and base editor activities are the same. It is critical to use proteins with the same amino acid sequence to that used during the clinical editing process to assess off-target risks in therapeutics.

Our findings clearly show that nuclease and base-editing off-target activity differ significantly in both biochemical and cellular contexts. We found that ABE8e induced more detected off-target sites and considerably higher frequencies of off-target editing activity compared to Cas9 under controlled cellular delivery conditions. This increased off-target activity is likely associated with the 540–1,000-fold increase in DNA deaminase activity in comparison to earlier-generation ABEs^30,31 and updates the dogma in the field that base editors are generally more specific than nucleases.

CHANGE-seq-BE is relevant for IND-enabling characterization of clinical genome-editing approaches. One example in this study is that we successfully used CHANGE-seq-BE to nominate candidate off-target sites, clear an IND application and treat a person with X-HIGM syndrome with a base editor designed to correct their disease-causing mutation.

One interesting observation of our study was that gRNAs with an intended target sequence could be contaminated with others during routine chromatography steps during synthesis and these low-frequency contaminants resulted in measurable activity in biochemical assays such as CHANGE-seq-BE, as well as activity in cells. This unintended ‘on-target’ activity of contaminant gRNAs will be important to consider and measure whenever synthetic gRNAs are used. For research purposes, it may be advisable to avoid sequential HPLC purification of synthesized gRNAs unless stringent procedures to avoid cross-contamination are applied.

A general challenge is the functional interpretation of potential base editor off-target sites. For example, it is not technically straightforward to perturb off-target sites individually without affecting other closely related sites and, thus, difficult to derisk them. Thus, although CHANGE-seq-BE represents an important advance to nominating base editor off-target sites, complementary approaches will be required to dissect their function.

Once better understood, several strategies can be applied to mitigate deleterious off-target effects and improve the safety and precision of base editors. Previous studies reported that high-fidelity versions of base editors can minimize off-target mutations^30,32,33. Alternatively, chimeric DNA–RNA hybrid gRNAs³⁴ offer a promising strategy for mitigating off-target activity without changing the editor protein, although the rules for designing them are not well understood. A recent study by Whittaker et al. demonstrated that hybrid gRNAs can significantly reduce off-target effects including bystander edits while maintaining high on-target editing efficiency³⁵.

Overall, with favorable properties of high sensitivity and sequencing efficiency, we anticipate that CHANGE-seq-BE will become increasingly broadly applied³⁶ as an unbiased discovery method to characterize the genome-wide gRNA-dependent off-target activity of base editors for life sciences research and clinical applications.

Isolation of human primary T cells

Research-consented and deidentified peripheral blood mononuclear cells (PBMCs) were obtained commercially (Key Biologics); CD4⁺CD8⁺ T cells were purified using magnetic separation on a CliniMACS Plus instrument. Cells were washed in CliniMACS buffer with 0.5% HSA and resuspended in 190 ml of intravenous immunoglobulin, followed by incubation for 15 min. Subsequently, cells were incubated and labeled with CD4 and CD8 microbeads (Miltenyi Biotec). Next, two washes were performed using CliniMACS buffer with 0.5% HSA, followed by CD4⁺CD8⁺ cell selection. The percentage of CD3⁺, CD4⁺, CD8⁺ and CD19⁺ cells in the selected population was determined by flow cytometry as quality control.

Isolation of human PBMCs, CD34⁺ cells and gDNA

PBMCs were collected from a person with X-HIGM after informed consent was provided (approved by the National Institutes of Health Institutional Review Board, protocol 94-I-0073). CD34⁺ cells were enriched by immunomagnetic bead selection using a CliniMACS Plus or AutoMACS instrument (Miltenyi Biotec). Human CD34⁺ cells were cultured in vitro using complete medium (CM) consisting of StemSpan II (StemCell Technologies) medium supplemented with 100 ng μl⁻¹ human stem cell factor (R&D systems, 255-SC/CF), 100 ng μl⁻¹ human thrombopoietin (R&D systems, 288-TP/CF) and 100 ng μl⁻¹ human FLT-3 ligand (R&D systems, 308-FK/CF). Cells were seeded and maintained at a density of 0.5–1 × 10⁶ cells per ml.

For rhAmpSeq for the X-HIGM studies, gDNA was isolated from bulk edited or naive cell samples. For editing, cells were thawed and cultured for 2 days in CM before electroporation (MaxCyte) with editing reagent. gDNA was isolated at 2–5 days after electroporation as per the manufacturer’s instructions³². In MKSR studies, we used gDNA isolated from CD34⁺ cells for rhAmpSeq according to the previously reported protocol¹¹.

Protein production and purification

Briefly, the expression plasmids (pRha-ABE8e-NRCH, Addgene, 165417; pABE8e-protein, Addgene, 161788) were transformed into BL21Start DE3 competent cells (Thermo). ABE8-NRCH and ABE8e-NGG were expressed and purified as previously described^11,37 at the Protein Production Core Facility at St. Jude. The eA3A-BE3-encoding plasmid was transformed into Rosetta 2(DE3) competent cells (MilliporeSigma) and grown in 2× tryptone yeast extract at 37 °C until reaching an optical density at 600 nm of ~2.0. Bacterial cells were induced with 0.3 mM IPTG for 20 h at 20 °C and harvested by centrifugation. Pellets were then suspended in 50 mM Tris pH 7.8, 500 mM NaCl, 10 % glycerol, 1 mM TCEP and 25 mM imidazole and lysed by passage twice through a homogenizer. The lysate was clarified by centrifugation at 20,000g for 1 h at 4 °C and the eA3A-BE3 protein was purified by Ni-NTA chromatography. The eluted protein was concentrated and treated with tobacco etch virus (TEV) protease (1 mg of lab-made TEV per 40 mg of protein) and benzonase (100 U per ml; Novagen) overnight at 4 °C. The cleaved protein was further purified by size-exclusion chromatography on a HiLoad 26/60 Superdex 200 (Amersham Biosciences) isocratically in 20 mM HEPES pH 6.8, 10% glycerol and 1 mM TCEP. Fractions containing the eA3A-BE3 were pooled and loaded onto a 5-ml SP HP column (GE) in the same buffer and eluted with a linear gradient of NaCl. Fractions were pooled and dialyzed into 20 mM HEPES pH 7.5 containing 300 mM NaCl, 10% glycerol and 1 mM TCEP. Endotoxin was removed by passage over 4 ml of Detoxigel (Pierce) and then quantitated using a toxin sensor chromogenic LAL endotoxin assay kit (GenScript). Purified proteins were concentrated using Amicon Ultra filter units (MilliporeSigma) and filtered with an ultrafree MC centrifugal filter (MilliporeSigma), followed by a 0.5 ml of 100-kDa cutoff Pierce concentrator. Protein purity was assessed on TGX stain-free 4–20% SDS–PAGE (Biorad) and quantified by rapid gold BCA assay (Thermo Fisher).

Cell culture

Human primary T cells were cultured in X-Vivo 15 medium (Lonza) supplemented with 10% human heat-inactivated serum (Fisher), 10 ng ml⁻¹ interleukin (IL)-7 (Miltenyi) and 10 ng ml⁻¹ IL-15 (Miltenyi) at 37 °C with 5% CO₂. T cells were stimulated with MACS GMP T cell TransAct polymeric nanomatrix (Miltenyi) for 3 days according to the manufacturer’s instructions before transfection. Human primary hepatocytes were cultured in Upcyte hepatocyte high-performance medium (BioIVT) with supplement A (BioIVT) and L-glutamine (BioIVT) in plates coated with collagen type I (Sigma) at 37 °C with 5% CO₂.

CHANGE-seq-BE and analysis

CHANGE-seq-BE was performed on gDNA isolated from human CD4⁺CD8⁺ T cells or human primary hepatocytes (BioIVT). gDNA was isolated using a PureGene tissue kit (Qiagen) and quantified by Qubit fluorimetry (Invitrogen). Purified gDNA was tagmented with a custom Tn5 transposome, gap-repaired with HiFi HotStart Uracil+ ready mix (Kapa) and Taq DNA ligase (New England Biolabs (NEB)) and treated with a mixture of USER enzyme and T4 polynucleotide kinase (NEB). DNA was circularized at a concentration of 5 ng µl⁻¹ with T4 DNA ligase (NEB) and treated with a cocktail of exonucleases, lambda exonuclease (NEB), exonuclease I (NEB), plasmid-safe ATP-dependent DNase (Lucigen) and exonuclease III (NEB), to enzymatically degrade remaining linear DNA molecules, followed by dephosphorylation with Quick CIP (NEB). gRNAs were refolded before ABE8e:gRNA or CBE:gRNA complexing at a ratio of 1:3 to ensure full complexation of RNP. ABE in vitro deamination reactions were performed in a 50-µl volume with deamination buffer (50 mM Tris-HCl pH 8.0, 25 mM KCl, 2.5 mM MgSO₄, 0.1 mM EDTA, 10% glycerol, 2 mM DTT and 10 µM ZnCl₂). The final concentrations of ABE and gRNAs were 300 nM and 900 nM (sequences are listed in Supplementary Table 9), along with 125 ng of circularized DNA. In the case of in vitro deamination with CBEs, NEBuffer r3.1 was used in a 50-µl reaction with final concentrations of 300 nM CBE and 900 nM gRNA, along with 125 ng of circularized DNA. ABE8e-deaminated DNA products were treated with proteinase K (NEB), followed by endonuclease V (NEB) treatment. CBE-deaminated DNA products were treated with proteinase K (NEB), followed by USER and T4-PNK (NEB) treatment. Linearized DNA fragments were end-repaired with Klenow fragment (3′→5′ exo-) (NEB), dA-tailed, ligated with a hairpin adaptor (NEB), treated with USER enzyme and amplified by PCR using HiFi HotStart Uracil+ ready mix (Kapa/Roche). Completed libraries were quantified by qPCR using the Kapa library quantification kit (Kapa Biosystems) and sequenced with 151-bp paired-end reads on an Illumina NextSeq instrument. A detailed user protocol for CHANGE-seq-BE is provided (Supplementary Protocol). For data analyses, adaptor trimming was performed using cutadapt (version 1.18) with parameters of ‘–overlap 40 -e 0.15’ to remove the Tn5 adaptor (CTGTCTCTTATACACATCTACGTAGATGTGTATAAGAGACAG). Then, paired-end reads were mapped using bwa mem with default parameters. Reads with zero mapping quality were filtered out (for example, requiring MAPQ ≥ 1). CHANGE-seq-BE paired-end reads should be in an outward orientation with some degree of self-overlapping bases. For ABE, assuming an editing window from position 1 to position 10, we used min_overlap = 5 and max_overlap = 15 because the deamination event occurred at the +2 position. Similarly, for CBE, we used min_overlap = 6 and max_overlap = 16 because the deamination event occurred at the +1 position. Reads were filtered on the basis of the above criteria. The start positions of the valid reads were then tabulated and genomic intervals enriched in nuclease-treated samples were identified. The interval and 30 bp of flanking reference sequence on either side were searched for potential base editor off-target sites with up to six mismatches, one insertion or one deletion. The alignment distance was the sum of number of mismatches, where insertions and deletions were counted as the equivalent of three mismatches. Off-target sites with alignment distance > 6 were further selected with read count ≥ 20. For off-target sites with alignment distance ≤ 6, no read count filter was applied except for CBE, where read_count_cutoff = 6 was used. Deamination events were A-to-G conversions if the off-target strand was forward or C-to-T conversions if the off-target strand was reverse. Deamination-containing reads were counted if the deamination events occurred close to +2 of the read start position (allowing for ±2 bp of flanking). The percentage of deamination was calculated as the ratio between number of deamination-containing reads and the total number of valid reads assigned to the off-target site. CHANGE-seq-BE data analyses were performed using open-source software (https://github.com/tsailabSJ/changeseq/tree/BE).

CHANGE-seq

CHANGE-seq was performed as previously described²⁰. Briefly, purified gDNA was tagmented with a custom Tn5 transposome, followed by gap repair with Kapa HiFi HotStart Uracil+ DNA polymerase (KAPA Biosystems) and Taq DNA ligase (NEB). Gap-repaired tagmented DNA was treated with USER enzyme (NEB) and T4 polynucleotide kinase (NEB). Intramolecular circularization of the DNA was performed with T4 DNA ligase (NEB) and residual linear DNA was degraded by a cocktail of exonucleases containing plasmid-safe ATP-dependent DNase (Lucigen), lambda exonuclease (NEB) and exonuclease I (NEB). In vitro cleavage reactions were performed with 125 ng of exonuclease-treated circularized DNA, 90 nM SpCas9 protein (NEB), NEB buffer 3.1 (NEB) and 270 nM gRNA in a 50-μl volume. Cleaved products were dA-tailed, ligated with a hairpin adaptor (NEB), treated with USER enzyme (NEB) and amplified by PCR with barcoded universal primers NEBNext Multiplex Oligos for Illumina (NEB), using Kapa HiFi Polymerase (KAPA Biosystems). Libraries were quantified by qPCR (KAPA Biosystems) and sequenced with 151-bp paired-end reads on an Illumina NextSeq instrument. CHANGE-seq-BE data analyses were performed using open-source software (https://github.com/tsailabSJ/changeseq).

Digenome-seq and analysis

Digenome-seq performed as previously described for ABE and CBEs³⁸. RNP complexation of ABEs and CBEs to gRNA occurred at a ratio of 1:3 with final concentrations of 150 nM ABEs or CBEs and 450 nM gRNAs (refolded before complexation), followed by incubation at 37 °C for 10 min. A total of 5 μg of gDNA was treated with ABE8e-NGG-complexed and eA3A-BE3-complexed RNPs for in vitro deamination with respective buffers at 37 °C for 8 h. Monarch RNAse A (NEB) was added before incubating at 37 °C for 15 min, followed by proteinase K (NEB) treatment, incubation at 37 °C for 15 min, purification with 1× volume of AMPure XP beads and elution in 1× Tris–EDTA (Integrated DNA Technologies (IDT)). Next, deaminated gDNA was treated with endonuclease V for ABE-treated samples and USER for CBE-treated samples with respective conditions. A total of 3 μg of purified in vitro ABE-deaminated DNA was mixed with 4 μl of endonuclease V (10 U per μl) and 20 μl of NEB buffer 4 (10×) in a final volume of 200 μl with nuclease-free water and incubated at 37 °C for 2 h. On the other hand, 2 μg of purified in vitro CBE-deaminated DNA was treated with 6 μl of USER (1 U per μl) and 10 μl of NEB cut smart buffer (10×) in a final volume of 100 μl with nuclease-free water and incubated at 37 °C for 3 h. Next, the final reaction was treated with proteinase K (NEB) at 37 °C for 15 min, followed by 1× volume of bead purification, and quantified by Qubit HS. A total of 1 μg of in vitro deaminated gDNA was fragmented to the 550–750 bp using Covaris (Life Technologies) with the following settings: duty factor, 20%; peak displayed power, 50 W; cycles per burst, 200; duration, 65 s; temperature, 20 °C. Later, sheared DNA was end-repaired and dA-tailed using Kapa HyperPrep PCR free (Kapa), ligated with adaptor stock (UDI-UMI, IDT), purified twice with 0.8× beads and eluted in 1× Tris–EDTA. Libraries were subjected to WGS to sequence with 20–30× coverage on Novaseq 6000 with 150–17–8–150 cycles. All samples were normalized to 400 million reads and analyzed using Digenome-Seq standalone software (version 1.0) with default parameters: MAPQ score ≥ 1, a minimum of five forward reads with identical 5′ ends, a minimum of five reverse reads with identical 3′ ends, a minimum read depth of 10 at each position, a minimum cleavage ratio of 0.2 and a minimum cleavage score of 2.5 (ref. ³⁹). Editing windows and read overlap lengths were consistent with those used in CHANGE-seq-BE: 1–10-long editing window, 5–15-long read overlap for ABE and 6–16-long read overlap for CBE. Potential off-target sites containing up to six mismatches, one insertion or one deletion that overlapped the Digenome cut sites were identified as nuclease-induced candidates. Digenome-seq data analyses were performed using open-source code (http://www.rgenome.net/digenome-js/standalone).

Cell transfection

Transfections of human primary T cells were performed with 2 µg of ABE8e, eA3A-BE3 or Cas9 mRNA and 1 µg of synthetic chemically modified gRNA. ABE8e in vitro transcribed mRNA was commercially synthesized by TriLink Biotechnologies. eA3A-BE3 in vitro transcribed mRNA was synthesized using cotranscriptional capping using CleanCap reagent AG from TriLink and a HiScribe T7 high-yield RNA synthesis kit from NEB according to the manufacturer’s instructions. Synthetic gRNAs ordered from Synthego (same gRNAs applied in CHANGE-seq and CHANGE-seq-BE) were used for electroporation to perform hybrid capture sequencing. mRNA:gRNA was added directly to 1 × 10⁶ cells resuspended in 20 µl of P3 solution and nucleofected with preprogrammed pulse EO-115 in a 4D-Nucleofector system (Lonza). After nucleofection, cells were recovered in X-vivo 15 medium with 20% human heat-inactivated serum (Fisher), 10 ng ml⁻¹ IL-7 (Miltenyi) and 10 ng ml⁻¹ IL-15 (Miltenyi). After 5 days, cells were harvested for gDNA purification using Agencourt DNAdvance (Beckman Coulter). Transfection of human primary hepatocytes was performed using Lipofectamine MessengerMAX transfection reagent (Thermo Fisher). Briefly, 1.5 × 10⁵ hepatocytes were seeded in a six-well plate coated with collagen type I. After 24 h, cells were transfected with 2.5 µg of ABE8e mRNA, 2.5 µg of synthetic chemically modified gRNA (PCSK9) and 7.5 µl of Lipofectamine MessengerMAX, following the manufacturer’s recommendations. After 3 days, cells were harvested for gDNA purification using Agencourt DNAdvance (Beckman Coulter).

Multiplex targeted amplicon sequencing

To determine the A-to-G conversion frequency at CHANGE-seq-BE-identified sites, on-target and off-target sites were amplified from gDNA from edited cells or unedited controls using the rhAmpSeq system (IDT), in triplicate, with primers listed in Supplementary Table 2. Note that, for the validation of the newly CHANGE-seq-BE-detected sites for MKSR targeting, we used the same gDNA of CD34⁺ cells treated with ABE8e-NRCH:HBB^S-gRNA mRNA described in our previous study. Sequencing libraries were generated according to the manufacturer’s instructions. Completed libraries were quantified by qPCR using the Kapa library quantification kit (Kapa Biosystems) and sequenced with 151-bp paired-end reads on an Illumina MiSeq or MiniSeq instrument.

GUIDE-seq-2 and analysis

GUIDE-seq-2 was performed as previously reported²⁹. Briefly, 100 µM assembled double-stranded oligodeoxynucleotide (dsODN) tag was prepared by annealing forward and reverse dsODN oligos in 1x STE buffer (10 mM Tris-HCl, 50 mM NaCl, 1 mM EDTA), heated at 95 °C for 5 min and gradually cooled down to 25 °C in a thermocycler. For RNP, Cas9-3xNLS (15 µM), gRNA (45 µM), 1 µl assembled dsODN (100 µM) were used to nucleofect 0.6 × 10⁶ human primary T cells. For mRNA, 1 × 10⁶ cells were nucleofected with Cas9 (2 µg), gRNA (1 µg) and 1 µl assembled dsODN (100 µM), and cells were recovered in supplemented X-vivo 15 media and harvested after 3 days (RNP) and 5 days (mRNA) of nucleofection. gDNA from edited and control cells containing the dsODN tag was isolated using the DNAdvance kit (Beckman Coulter) according to the manufacturer’s instructions. Purified gDNA was quantified by Qubit BR and 100 ng of gDNA was tagmented to an average length of 600 bp with a Tn5 transposome, purified from pTXB1-Tn5 (Addgene, 60240), assembled with oligos containing a 10-nt unique molecular identifier and then purified using solid-phase reverse immobilization (SPRI) guanidine magnetic beads. Tagmented gDNA was PCR-amplified using dsODN sense-specific and antisense-specific primers in separate reactions for each sample for target enrichment, which were further purified with a limited number of SPRI magnetic beads for one-step purification and normalization. Libraries for the replicates were pooled and size selected for 250–450 bp by Lightbench (Yourgene Health), quantified by a qPCR library quantification kit (TakaraBio) and sequenced with 146-bp paired-end reads on an Illumina NextSeq 2000. GUIDE-seq-2 data analyses were performed using open-source software (https://github.com/tsailabSJ/guideseq).

Hybrid capture sequencing

First, 50 ng of gDNA per sample was used for enzymatic fragmentation at 4 °C (∞), 37 °C (20 min), 65 °C (30 min) and 4 °C (∞) using the Twist library preparation EF kit 2.0. Twist UMI adaptors were ligated to the dA-tailed DNA fragments, incubated at 20 °C for 15 min and later purified with DNA purification beads. Adaptor-ligated gDNA libraries were amplified with Twist UDI primers with the following PCR conditions: 98 °C for 45 s (denaturation), followed by eight cycles of 98 °C for 15 s, 60 °C for 30 s and 72 °C for 30 s (annealing) with a final extension at 72 °C for 1 min, before holding at 4 °C. PCR-indexed gDNA libraries were then purified with beads, quality-controlled on an Agilent D1000 Tapestation with the expected size of library ~350–425 bp and quantified by Qubit BR. The indexed samples were pooled at a concentration of 250 ng and the total mass of the each pool was ~2,000 ng. The eight multiplexed libraries were dried with a speed vacuum and hybridized with respective ABE and CBE panels of custom probes (Twist) at 70 °C for 17 h (lid at 85 °C). Next, hybridized targets were captured with streptavidin beads and PCR was used to amplify postcaptured products according to the manufacturer’s instructions. The ABE panel PCR conditions were as follows: 98 °C for 45 s (denaturation), followed by nine cycles of 98 °C for 15 s, 60 °C for 30 s and 72 °C for 30 s (annealing) with a final extension at 72 °C for 1 min (final extension), before holding at 4 °C. The CBE panel PCR conditions were as follows: 98 °C for 45 s (denaturation), followed by 11 cycles of 98 °C for 15 s, 60 °C for 30 s and 72 °C for 30 s (annealing) with a final extension at 72 °C for 1 min, before holding at 4 °C. Postcaptured PCR products were purified using DNA purification beads and enriched libraries were quantified by Qubit HS. The library size was determined on an Agilent D1000 HS Tapestation to be ~350–450 bp. CBE and ABE libraries were pooled and loaded separately on NextSeq 2000 and NovaSeq X paired-end sequencing with 146–15–15–146 cycles according to the manufacturer’s instructions. Paired-end sequencing reads were trimmed using cutadapt with ‘-a AGATCGGAAGAGC -A AGATCGGAAGAGC’. Then, FLASH was used to merge R1 and R2 reads. Unmerged reads were treated as individual molecules. Merged and unmerged reads R1/R2 were concatenated. Unique molecular identifier deduplication was performed using open-source software (https://github.com/aryeelab/umi). Unique molecular identifier deduplicated reads were mapped to hg38 using bwa mem with default parameters. Lastly, CRISPRessoWGS was used to quantify indel frequency and base-editing frequency. A description of the analysis is available online (https://hemtools.readthedocs.io/en/latest/content/NGS_pipelines/hybrid_capture.html). The source code is available from GitHub (https://github.com/YichaoOU/HemTools/blob/master/share/lsf/hybrid_capture.lsf).

Targeted sequencing

To determine the A-to-G conversion frequency at intended and unintended on-target sites, primary human T cells were nucleofected with ABE8e mRNA and CBLB or CIITA gRNAs. On-target sites (CBLB, CIITA, B2M, CD7 or PDCD1) were amplified from T cell gDNA using 2× Phusion HotStart flex master mix (NEB) (primers ordered from IDT listed in Supplementary Table 9) and 100 ng of gDNA as the input for each PCR. PCR products were purified with SPRI magnetic beads and normalized. Then, 10 ng of each amplicon were used for the second PCR to add complete Illumina adaptors and indices. Completed libraries were quantified by qPCR using the Kapa library quantification kit (Kapa Biosystems) and sequenced with 151-bp paired-end reads on an Illumina MiSeq or MiniSeq instrument.

gRNA synthesis, sequencing and analysis

Synthetic gRNAs ordered from BioSpring, Synthego and IDT were used for the experiments. Two sets of synthetic gRNAs from BioSpring were tested. Both sets were HPLC-purified and one set used an HPLC column clearing-in-place procedure between gRNA purifications. Synthetic gRNAs from Synthego and IDT were purified using a standard desalting method. These gRNAs were chemically modified with 2′-O-methyl analogs and 3′ phosphorothiate internucleotide linkages at the first three 5′ and 3′ terminal residues (Supplementary Table 9). To generate gRNA-seq libraries, we repurposed the SMARTer smRNA-seq kit (Takara). First, cDNA synthesis was performed with 1–2 ng of gRNA template using 1 μl of custom-tailed scaffold specific primer (10 μM) 5′ GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTTGCCACTTTTTCAAGTTG 3′. The reaction was incubated at 72 °C for 3 min and immediately placed on ice for 2 min to anneal a custom-tailed scaffold primer at the 3′ end of the scaffold gRNA. cDNA synthesis, library enrichment and sequencing were performed according to the manufacturer’s instructions, as previously reported²⁸. gRNA sequencing data analyses were performed using open-source software (https://github.com/tsailabSJ/gRNA_sequencing).

Quantification analysis of base-editing activity, annotation and statistical analyses

The base-editing frequency for each position in the protospacer was quantified using CRISPRessoPooled (version 2.0.41). For the statistical analysis shown in dot blots between mRNA treatments and control samples (n = 2 or 3), a χ² test was performed, followed by false discovery rate (FDR) calculation using the Benjamini–Hochberg method. The reported off-target sites with difference ≥ 0.5% for at least one treatment and with FDR ≤ 0.05 were determined as significant. Custom code was used to conduct off-target quantification and the statistical analysis method is available from GitHub (https://github.com/tsailabSJ/MKSR_off_targets). The genomic features of all nominated and bona fide off-target sites were annotated using HOMER (version 4.10)⁴⁰. Standard statistical analyses were performed using GraphPad Prism version 10.5.0. The statistical t-tests, exact tests, methods, P values and numbers of biological replicates are provided in the figure legends. Venn diagrams and Manhattan plots were created using R version 4.5.1.

Reporting summary

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

1. 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 

2. 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 

3. 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 

4. Tsai, S. Q. & Joung, J. K. Defining and improving the genome-wide specificities of CRISPR-Cas9 nucleases. Nat. Rev. Genet. 17, 300–312 (2016).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Bao, X. R., Pan, Y., Lee, C. M., Davis, T. H. & Bao, G. Tools for experimental and computational analyses of off-target editing by programmable nucleases. Nat. Protoc. 16, 10–26 (2021).

   Article  CAS  PubMed  Google Scholar 

6. Wienert, B. & Cromer, M. K. CRISPR nuclease off-target activity and mitigation strategies. Front. Genome Ed. 4, 1050507 (2022).

   Article  PubMed  PubMed Central  Google Scholar 

7. Center for Biologics Evaluation and Research. Human Gene Therapy Products Incorporating Human Genome Editing (US Food and Drug Administration, 2024); https://www.fda.gov/regulatory-information/search-fda-guidance-documents/human-gene-therapy-products-incorporating-human-genome-editing

8. Kim, D., Kim, D.-E., Lee, G., Cho, S.-I. & Kim, J.-S. Genome-wide target specificity of CRISPR RNA-guided adenine base editors. Nat. Biotechnol. 37, 430–435 (2019).

   Article  CAS  PubMed  Google Scholar 

9. Liang, P. et al. Genome-wide profiling of adenine base editor specificity by EndoV-seq. Nat. Commun. 10, 67 (2019).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

10. Petri, K. et al. Global-scale CRISPR gene editor specificity profiling by ONE-seq identifies population-specific, variant off-target effects. Preprint at bioRxiv https://doi.org/10.1101/2021.04.05.438458 (2021).

11. 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 

12. Alves, C. R. R. et al. Optimization of base editors for the functional correction of SMN2 as a treatment for spinal muscular atrophy. Nat. Biomed. Eng. 8, 118–131 (2024).

    Article  CAS  PubMed  Google Scholar 

13. Mayuranathan, T. et al. Potent and uniform fetal hemoglobin induction via base editing. Nat. Genet. 55, 1210–1220 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

14. Arbab, M. et al. Base editing rescue of spinal muscular atrophy in cells and in mice. Science 380, eadg6518 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

15. McAuley, G. E. et al. Human T cell generation is restored in CD3δ severe combined immunodeficiency through adenine base editing. Cell 186, 1398–1416 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Tsai, S. Q. et al. CIRCLE-seq: a highly sensitive in vitro screen for genome-wide CRISPR–Cas9 nuclease off-targets. Nat. Methods 14, 607–614 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. Lazzarotto, C. R. et al. Defining CRISPR–Cas9 genome-wide nuclease activities with CIRCLE-seq. Nat. Protoc. 13, 2615–2642 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

18. Kim, D. et al. Genome-wide target specificities of CRISPR RNA-guided programmable deaminases. Nat. Biotechnol. 35, 475–480 (2017).

    Article  CAS  PubMed  Google Scholar 

19. 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 

20. Lazzarotto, C. R. et al. CHANGE-seq reveals genetic and epigenetic effects on CRISPR–Cas9 genome-wide activity. Nat. Biotechnol. 38, 1317–1327 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

21. Yao, M., Hatahet, Z., Melamede, R. J. & Kow, Y. W. Purification and characterization of a novel deoxyinosine-specific enzyme, deoxyinosine 3′ endonuclease, from Escherichia coli. J. Biol. Chem. 269, 16260–16268 (1994).

    Article  CAS  PubMed  Google Scholar 

22. Koblan, L. W. et al. Improving cytidine and adenine base editors by expression optimization and ancestral reconstruction. Nat. Biotechnol. 36, 843–846 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

23. Gehrke, J. M. et al. An APOBEC3A-Cas9 base editor with minimized bystander and off-target activities. Nat. Biotechnol. 36, 977–982 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

24. Chen, L. et al. Re-engineering the adenine deaminase TadA-8e for efficient and specific CRISPR-based cytosine base editing. Nat. Biotechnol. 41, 663–672 (2022).

    Article  PubMed  Google Scholar 

25. 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 

26. Musunuru, K. et al. In vivo CRISPR base editing of PCSK9 durably lowers cholesterol in primates. Nature 593, 429–434 (2021).

    Article  CAS  PubMed  Google Scholar 

27. Hendel, A. et al. Chemically modified guide RNAs enhance CRISPR–Cas genome editing in human primary cells. Nat. Biotechnol. 33, 985–989 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

28. Katta, V. et al. Development and IND-enabling studies of a novel Cas9 genome-edited autologous CD34⁺ cell therapy to induce fetal hemoglobin for sickle cell disease. Mol. Ther. 32, 3433–3452 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

29. Lazzarotto, C. R. et al. Population-scale cellular GUIDE-seq-2 and biochemical CHANGE-seq-R profiles reveal human genetic variation frequently affects Cas9 off-target activity. Preprint at bioRxiv https://doi.org/10.1101/2025.02.10.637517 (2025).

30. 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 

31. 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 

32. Bzhilyanskaya, V. et al. High-fidelity PAMless base editing of hematopoietic stem cells to treat chronic granulomatous disease. Sci. Transl. Med. 16, eadj6779 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

33. Wu, L. et al. Adenine base editors induce off-target structure variations in mouse embryos and primary human T cells. Genome Biol. 25, 291 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

34. Donohoue, P. D. et al. Conformational control of Cas9 by CRISPR hybrid RNA–DNA guides mitigates off-target activity in T cells. Mol. Cell 81, 3637–3649 (2021).

    Article  CAS  PubMed  Google Scholar 

35. Whittaker, M. N. et al. Improved specificity and efficiency of in vivo base editing therapies with hybrid guide RNAs. Nat. Biomed. Eng. https://doi.org/10.1038/s41551-025-01545-y (2025).

    Article  PubMed  Google Scholar 

36. Musunuru, K. et al. Patient-specific in vivo gene editing to treat a rare genetic disease. N. Engl. J. Med. 392, 2235–2243 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

37. Huang, T. P., Newby, G. A. & Liu, D. R. Precision genome editing using cytosine and adenine base editors in mammalian cells. Nat. Protoc. 16, 1089–1128 (2021).

    Article  CAS  PubMed  Google Scholar 

38. Kim, D., Kang, B.-C. & Kim, J.-S. Identifying genome-wide off-target sites of CRISPR RNA-guided nucleases and deaminases with Digenome-seq. Nat. Protoc. 16, 1170–1192 (2021).

    Article  CAS  PubMed  Google Scholar 

39. Park, J. et al. Digenome-seq web tool for profiling CRISPR specificity. Nat. Methods 14, 548–549 (2017).

    Article  CAS  PubMed  Google Scholar 

40. Heinz, S. et al. Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol. Cell 38, 576–589 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references

Author notes

1. Elizabeth Urbina

   Present address: Vanderbilt University, Nashville, TN, USA

2. GaHyun Lee

   Present address: Northwestern University Feinberg School of Medicine, Chicago, IL, USA

3. These authors contributed equally: Cicera R. Lazzarotto, Varun Katta.

Authors and Affiliations

1. Department of Hematology, St. Jude Children’s Research Hospital, Memphis, TN, USA

   Cicera R. Lazzarotto, Varun Katta, Yichao Li, Garret Manquen, Rachael K. Wood, Jacqueline Chyr, Elizabeth Urbina, Azusa Matsubara, GaHyun Lee & Shengdar Q. Tsai

2. Cancer Research Technology Program, Frederick National Laboratory of Cancer Research, Frederick, MD, USA

   Xiaolin Wu

3. Genetic Immunotherapy Section, LCIM, NIAID, National Institutes of Health, Bethesda, MD, USA

   Suk See De Ravin

Contributions

C.R.L., V.K. and S.Q.T conceptualized and designed the study. S.Q.T. supervised the research. C.R.L., V.K., G.M., R.K.W., E.U., A.M., G.L. and X.W. performed the experiments. S.S.D.R. provided conceptual assistance. Y.L. and J.C. performed the computational analysis. C.R.L., V.K. and S.Q.T. wrote the paper with input from all authors.

Corresponding author

Correspondence to Shengdar Q. Tsai.

Competing interests

C.R.L. and S.Q.T. are inventors on a patent covering CHANGE-seq. S.Q.T. is an inventor on patents covering CIRCLE-seq and GUIDE-seq. S.Q.T. is a member of the scientific advisory board of Prime Medicine and Ensoma. The other authors declare no competing interests.

Peer review information

Nature Biotechnology thanks Samuele Ferrari, Luigi Naldini and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Reporting Summary

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