The clustered regularly interspaced short palindromic repeat (CRISPR) system has been widely used to mediate genome editing in a variety of organisms and cell lines3. CRISPR/Cas9 protein–RNA complexes localize to a target DNA sequence through base pairing with a guide RNA, and natively create a dsDNA break (DSB) at the locus specified by the guide RNA. In response to DSBs, cellular DNA repair processes mostly result in random insertions or deletions (indels) at the site of DNA cleavage through non-homologous end joining (NHEJ). In the presence of a homologous DNA template, the DNA surrounding the cleavage site can be replaced through homology-directed repair (HDR). HDR competes with NHEJ during the resolution of DSBs, and indels are generally more abundant outcomes than gene replacement. For most known genetic diseases, however, correction of a point mutation in the target locus, rather than stochastic disruption of the gene, is needed to study or address the underlying cause of the disease4.

Motivated by this need, researchers have sought to increase the efficiency of HDR and suppress NHEJ. Despite recent progress (see Supplementary Information for a detailed discussion), current strategies to correct point mutations using HDR under therapeutically relevant conditions remain inefficient (typically ~0.1 to 5%)5,6, especially in unmodified, non-dividing cells. These observations highlight the need to develop alternative approaches to correct point mutations in genomic DNA that do not require DSBs.

We envisioned that direct conversion of one DNA base to another at a programmable target locus without requiring DSBs could increase the efficiency of gene correction relative to HDR without introducing an excess of random indels. Catalytically dead Cas9 (dCas9), which contains Asp10Ala and His840Ala mutations that inactivate its nuclease activity, retains its ability to bind DNA in a guide RNA-programmed manner, but does not cleave the DNA backbone7. In principle, conjugation of dCas9 with an enzymatic or chemical catalyst that mediates the direct conversion of one base to another could enable RNA-programmed DNA base editing.

The deamination of cytosine (C) is catalysed by cytidine deaminases8 and results in uracil (U), which has the base-pairing properties of thymine (T). Most known cytidine deaminases operate on RNA, and the few examples that are known to accept DNA require single-stranded (ss) DNA9. Recent studies on the dCas9–target DNA complex reveal that at least nine nucleotides (nt) of the displaced DNA strand are unpaired upon formation of the Cas9–guide RNA–DNA ‘R-loop’ complex10. Indeed, in the structure of the Cas9 R-loop complex, the first 11 nt of the protospacer on the displaced DNA strand are disordered, suggesting that their movement is not highly restricted11. It has also been speculated that Cas9 nickase-induced mutations at cytosines in the non-template strand might arise from their accessibility by cellular cytosine deaminase enzymes12. We reasoned that a subset of this stretch of ssDNA in the R-loop might serve as an efficient substrate for a dCas9-tethered cytidine deaminase to effect direct, programmable conversion of C to U in DNA (Fig. 1a).

Figure 1: First-generation base editor (BE1) mediates specific, guide RNA-programmed C→U conversion in vitro. a, Base editing strategy. DNA with a target C (red) at a locus specified by a guide RNA (green) is bound by dCas9 (blue), which mediates local DNA strand separation. Cytidine deamination by a tethered APOBEC1 enzyme (red) converts the single-stranded target C→U. The resulting G:U heteroduplex can be permanently converted to an A:T base pair following DNA replication or DNA repair. b, Deamination assay showing a BE1 activity window of approximately five nucleotides. Samples were prepared as described in the Methods. Each lane is labelled according to the position of the target C within the protospace, or with ‘–’ if no target C is present, counting the base distal from the PAM as position 1. c, Deamination assay showing the sequence specificity and sgRNA-dependence of BE1. The DNA substrate with C at position 7 in b was incubated with BE1 and the correct sgRNA, a mismatched sgRNA or no sgRNA. The positive control sample used a synthetic DNA substrate with a U at position 7. For uncropped gel data, see Supplementary Figure 1. Full size image Download PowerPoint slide

Four different cytidine deaminase enzymes (human AID, human APOBEC3G, rat APOBEC1, and lamprey CDA1) were evaluated for ssDNA deamination. Of the four enzymes, rat APOBEC1 showed the highest deaminase activity under the conditions tested (Extended Data Fig. 1a). Fusing rat APOBEC1 to the amino terminus, but not the carboxy terminus, of dCas9 preserves deaminase activity (Extended Data Fig. 1a). We expressed and purified four rat APOBEC1–dCas9 fusions with linkers of different length and composition (Extended Data Fig. 1b), and evaluated each fusion for single guide RNA (sgRNA)-programmed dsDNA deamination in vitro (Fig. 1b and Extended Data Fig. 1c–f).

We observed efficient, sequence-specific, sgRNA-dependent C to U conversion in vitro (Fig. 1c). Conversion efficiency was greatest using rat APOBEC1–dCas9 linkers over nine amino acids in length. The number of positions susceptible to deamination (the ‘activity window’) increases from approximately 3 to 6 nt as the linker length was extended from 3 to 21 amino acids (Extended Data Fig. 1c–f). The 16-residue XTEN linker13 offered a promising balance between these two characteristics, with an efficient deamination window of approximately 5 nt, typically from positions 4 to 8 within the protospacer, counting the end distal to the protospacer-adjacent motif (PAM) as position 1. The rat APOBEC1–XTEN–dCas9 protein served as the first-generation base editor (BE1).

We assessed the ability of BE1 in vitro to correct seven T→C mutations relevant to human disease (Extended Data Fig. 2). BE1 yielded products consistent with efficient editing of the target C, or of at least one C within the activity window when multiple Cs were present, in six of these seven targets in vitro, with an average apparent editing efficiency of 44% (Extended Data Fig. 2).

Although the preferred sequence context for APOBEC1 substrates is TC or CC14, we anticipated that the increased effective molarity of the tethered deaminase and its ssDNA substrate upon dCas9 binding might relax this preference. To illuminate the context dependence of BE1, we assayed its ability to edit a dsDNA 60-mer containing a single fixed C at position 7 within the protospacer, as well as all 36 single-mutant variants in which protospacer bases 1–6 and 8–13 were individually varied to each of the other three bases. High-throughput DNA sequencing revealed 50–80% C to U conversion of substrate strands (25–40% of sequence reads from both DNA strands, one of which is not a substrate for BE1) (Fig. 2a). Editing efficiency was independent of sequence context, unless the base immediately 5′ of the target C was a G, in which case editing efficiency was substantially lower (Fig. 2a). Next we assessed BE1 activity in vitro on all four NC motifs at positions 1 through 8 within the protospacer (Fig. 2b). BE1 activity followed the order TC ≥ CC ≥ AC > GC (the second nucleotide (C) is the target nucleotide), with maximum editing efficiency achieved when the target C is at or near position 7 (see Supplementary Information). In addition, we observed that the base editor is processive, and will efficiently convert most or all Cs to Us on the same DNA strand within the five-base activity window (Extended Data Fig. 3).

Figure 2: Effects of sequence context and target C position on base editing efficiency in vitro. a, Effect of changing the sequence surrounding the target C on editing efficiency in vitro. The deamination yield of 80% of targeted strands (40% of total sequencing reads from both strands) for C 7 in the protospacer sequence 5′-TTATTT(C 7 )GTGGATTTATTTA-3′ was defined as 1.0, and the relative deamination efficiencies of substrates containing all possible single-base mutations at positions 1–6 and 8–13 are shown. b, Positional effect of each NC motif on editing efficiency in vitro. Each NC target motif was varied from positions 1 to 8 within the protospacer as indicated in the sequences shown on the right. The PAM is shown in blue. The graph shows the percentage of total DNA sequencing reads containing T at each of the numbered target C positions following incubation with BE1. Note that the maximum possible deamination yield in vitro is 50% of total sequencing reads (100% of targeted strands). Values and error bars reflect the mean and standard deviation of three (a) or two (b) independent biological replicates performed on different days. Full size image Download PowerPoint slide

While BE1 efficiently processes substrates in a test tube, in cells, a variety of possible DNA repair outcomes determines the fate of the initial U:G product of base editing (Fig. 3a). We tested the ability of BE1 to convert C→T in human cells on 14 Cs in six well-studied target sites in the human genome (see Supplementary Information and Extended Data Fig. 4a)15. Although C→T editing in cells was observed for all cases, the efficiency of base editing was 0.8% to 7.7% of total DNA sequences, a large 5- to 36-fold decrease in efficiency compared to that of in vitro base editing (Fig. 3b and Extended Data Fig. 4).

Figure 3: Base editing in human cells. a, Possible base editing outcomes in mammalian cells. Initial editing results in a U:G mismatch. Recognition and excision of the U by uracil DNA glycosylase (UDG) initiates base excision repair (BER), which leads to reversion to the C:G starting state. BER is impeded by BE2 and BE3, which inhibit UDG. The U:G mismatch is also processed by mismatch repair (MMR), which preferentially repairs the nicked strand of a mismatch. BE3 nicks the non-edited strand containing the G, favouring resolution of the U:G mismatch to the desired U:A or T:A outcome. b, HEK293T cells were treated as described in the Methods. The percentage of total DNA sequencing reads with Ts at the target positions indicated are shown for treatment with BE1, BE2, or BE3, or for treatment with wild-type Cas9 with a 200-nt donor HDR template. c, Frequency of indel formation (see Methods) is shown following the treatment in b. Values are listed in Extended Data Table 1. For b and c, values and error bars reflect the mean and s.d. of three independent biological replicates performed on different days. Full size image Download PowerPoint slide

We hypothesized that the cellular DNA repair response to U:G heteroduplex DNA was responsible for the large decrease in base editing efficiency in cells (Fig. 3a). Uracil DNA glycosylase (UDG) catalyses removal of U from DNA in cells and initiates base-excision repair (BER), with reversion of the U:G pair to a C:G pair as the most common outcome (Fig. 3a)16. Uracil DNA glycosylase inhibitor (UGI), an 83-residue protein from Bacillus subtilis bacteriophage PBS1, potently blocks human UDG activity (IC 50 = 12 pM)17. In an effort to subvert BER at the site of base editing, we fused UGI to the C terminus of BE1 to create a second-generation base editor (BE2, APOBEC–XTEN–dCas9–UGI) and repeated editing assays on all six genomic loci. Editing efficiencies in human cells were on average threefold higher with BE2 than BE1, resulting in gene conversion efficiencies of up to 20% of total DNA sequenced (Fig. 3b).

Importantly, BE1 and BE2 resulted in indel formation rates ≤ 0.1% (Fig. 3c and Extended Data Table 1), consistent with the known mechanistic dependence of NHEJ on DSBs (see Supplementary Information)18. We assessed BE2-mediated base editing efficiencies on the same genomic targets in U2OS cells, and observed results similar to those in HEK293T cells (Extended Data Fig. 5). Together, these results indicate that conjugating UGI to BE1 can increase the efficiency of base editing in human cells.

Converting and protecting the substrate strand of a C:G base pair (bp) results in a maximum base editing yield of 50%. To augment base editing efficiency beyond this limit, we sought to further manipulate cellular DNA repair to induce correction of the non-edited strand containing the G. Eukaryotic mismatch repair (MMR) uses nicks present in newly synthesized DNA to direct removal and resynthesis of the newly synthesized strand (Fig. 3a)19,20. We reasoned that nicking the DNA strand containing the unedited G would simulate newly synthesized DNA, inducing MMR, or simulate damaged DNA, inducing long-patch BER, to preferentially resolve the U:G mismatch into desired U:A and T:A products (Fig. 3a). We therefore restored the catalytic His residue at position 840 in the Cas9 HNH domain of BE2 (ref. 7), resulting in the third-generation base editor (BE3, APOBEC–XTEN–dCas9(A840H)–UGI) that nicks the non-edited strand containing a G opposite the edited U. BE3 retains the Asp10Ala mutation in Cas9 that prevents dsDNA cleavage, and also retains UGI to suppress UDG-initiated BER of the editing strand.

Nicking the non-edited strand augmented base editing efficiency in human cells treated with BE3 by an additional two- to sixfold relative to BE2, resulting in up to 37% of total DNA sequences containing the targeted C→T conversion (Fig. 3b). Importantly, only a small frequency of indels, averaging 1.1% for the six tested loci, was observed from BE3 treatment (Fig. 3c and Extended Data Table 1). In contrast, when we treated cells with wild-type Cas9, sgRNA to target each of three loci, and a 200 nt ssDNA donor template to mediate HDR, we observed C→T conversion efficiencies averaging only 0.5%, with much higher indel formation averaging 4.3% (Fig. 3c). The ratio of allele conversion to NHEJ outcomes averaged >1,000 for BE2, 23 for BE3, and 0.17 for wild-type Cas9 (Fig. 3c). We confirmed the permanence of base editing in human cells by monitoring editing efficiencies over multiple cell divisions in HEK293T cells at the HEK293 site 3 and 4 genomic loci (Extended Data Fig. 6 and Supplementary Information). These results collectively establish that base editing can induce much more efficient targeted single-base editing in human cells than Cas9-mediated HDR, and with substantially less (BE3) or almost no (BE2) indel formation.

Next, we examined the off-target activity of BE1, BE2, and BE3 in human cells for five previously studied loci (see Supplementary Information and Supplementary Tables 1–5). Because the sequence preference of rat APOBEC1 is known to be independent of bases more than one nucleotide from the target C (ref. 21), consistent with Fig. 2a, we assumed that off-target base editing arises from off-target Cas9 binding. Therefore, we sequenced the top 34 known Cas9 off-target sites in human cells15, and the top 12 known dCas9 off-target binding sites for the six genomic loci studied in Fig. 3 (Supplementary Tables 1–5)22. We observed detectable off-target base editing at a subset of known Cas9 off-target sites (16/34 for BE1 and BE2; and 17/34 for BE3), but no detectable base editing at the known dCas9 off-target sites. All detected off-target base-editing substrates contained a C within the five-base activity window (see Supplementary Information). We also monitored C→T mutations at 3,200 cytosines surrounding the six on-target and 44 off-target loci tested and observed no evident increase in C→T conversions outside the protospacer upon BE1, BE2, or BE3 treatment compared to that of untreated cells (Extended Data Fig. 7). Taken together, these findings suggest that off-target substrates of base editors include a subset of Cas9 off-target substrates, and that base editors in human cells do not induce untargeted C→T conversion throughout the genome.

Finally, we tested the potential of base editing to correct two disease-relevant mutations in mammalian cells. The apolipoprotein E gene variant APOE4 encodes two arginine residues at amino acid positions 112 and 158, and is the largest and most common genetic risk factor for late-onset Alzheimer’s disease23. ApoE variants with Cys residues at these positions, including APOE2 (Cys112 and Cys158), APOE3 (Cys112 and Arg158), and APOE3r (Arg112 and Cys158) have been shown or are presumed24 to confer lower Alzheimer’s disease risk than APOE4. We attempted to convert APOE4 into APOE3r in immortalized mouse astrocytes in which the endogenous Apoe gene was replaced by human APOE4. We delivered into these astrocytes by nucleofection DNA encoding BE3 and an appropriate sgRNA placing the target C at position 5 relative to a downstream PAM. After two days, we isolated nucleofected cells and measured editing efficiency by HTS of genomic DNA. We observed conversion of Arg158 to Cys158 in 58–75% of total DNA sequencing reads (Fig. 4a and Extended Data Fig. 8a). We also observed 36–50% editing of total DNA at the third position of codon 158 and 38–55% editing of total DNA at the first position of Leu159, as expected since all three of these Cs are within the base editing window. Neither of the other two C→T conversions, however, alters the amino acid sequence of the ApoE3r protein, as both TGC and TGT encode Cys (all C→T changes at the third position of a codon are silent), and both CTG and TTG encode Leu.

Figure 4: BE3-mediated correction of two disease-relevant mutations in mammalian cells. The sequence of the protospacer is shown to the right of the mutation, with the PAM in blue and the target base in red with a subscripted number indicating its position within the protospacer. Underneath each sequence are the percentages of total sequencing reads with the corresponding base. Cells were treated as described in the Methods. a, The Alzheimer’s disease-associated APOE4 allele was converted to APOE3r in mouse astrocytes by BE3 in 74.9% of total reads. Two nearby Cs are also converted to Ts, but with no change to the predicted sequence of the resulting protein. Identical treatment of these cells with wild-type Cas9 and a 200-nt ssDNA donor results in 0.3% correction, with 26.1% indel formation. b, The cancer-associated p53 Y163C mutation is corrected by BE3 in 7.6% of nucleofected human breast cancer cells with 0.7% indel formation. Identical treatment of these cells with wild-type Cas9 and donor ssDNA results in no mutation correction with 6.1% indel formation. Full size image Download PowerPoint slide

The efficiency of BE3-mediated editing of APOE4 demonstrates that a combination of suppressing BER and guiding MMR to repair the unedited strand enables base editing efficiencies to exceed the 50% maximum yield that would result from DNA replication alone. We observed no evident increase in mutations within 50 bp of either end of the protospacer compared with untreated controls (Supplementary Table 6). We observed 4.6–6.1% indels at the targeted locus following BE3 treatment. In contrast, identical treatment of astrocytes with wild-type Cas9 and donor ssDNA resulted in 0.1–0.3% APOE4 correction and 26–40% indels at the targeted locus, efficiencies consistent with previous reports of single-base correction using Cas9 and HDR5,6 (Fig. 4a and Extended Data Fig. 8a). Astrocytes treated identically but with an sgRNA targeting the VEGFA locus displayed no evidence of APOE4 base editing (Supplementary Table 6 and Extended Data Fig. 8a). These results demonstrate that base editors can mediate highly efficient and precise single amino acid changes in the coding sequence of a protein, even when their processivity results in >1 nucleotide change in genomic DNA.

The dominant-negative p53 mutation Tyr163Cys is strongly associated with several types of cancer25 and can be corrected by a C→T conversion on the template strand (Extended Data Fig. 2), resulting in the translation of corrected protein even before the edited base is made permanent by DNA replication or DNA repair. We nucleofected a human breast cancer cell line homozygous for the p53 Tyr163Cys mutation (HCC1954 cells) with DNA encoding BE3 and an sgRNA programmed to correct Tyr163Cys. We observed correction of the Tyr163Cys mutation in 3.3–7.6% of nucleofected HCC1954 cells (Fig. 4b, Extended Data Fig. 8b, and Supplementary Table 7), with ≤ 0.7% indel formation. In contrast, treatment of cells with wild-type Cas9 and donor ssDNA resulted in no detectable TP53 correction (<0.1%) with 6.1–8.0% indels at the target locus (Fig. 4b and Extended Data Fig. 8b). These results collectively represent the correction of disease-associated point mutations in mammalian cell lines with an efficiency and lack of other genome modification events that may not be achievable using previously described methods. An additional 300–900 clinically relevant known human genetic diseases that in principle are correctable by the base editors described in this work are shown in Extended Data Fig. 9 and Supplementary Table 8 (see Supplementary Information).

The development of base editing advances both the scope and effectiveness of genome editing. The base editors described here offer researchers a choice of editing with very little (<0.1%) indel formation (BE2), or more efficient editing with ≤1% indel formation (BE3). That the product of base editing is, by definition, no longer a substrate likely contributes to editing efficiency by preventing subsequent product transformation, which can hamper traditional Cas9 applications. By removing the reliance on dsDNA cleavage, donor templates, and stochastic DNA repair processes that vary by cell state and cell type, base editing has the potential to expand the type of genome modifications that can be cleanly installed, the efficiency of these modifications, and the type of cells that are amenable to editing. It is likely that engineered Cas9 variants26,27,28 or delivery methods29 that offer improved DNA specificity or altered PAM specificities30 can provide additional base editors with improved properties. These results also suggest architectures for the fusion of other DNA-modifying enzymes, including methylases and demethylases, that may enable additional types of programmable genome and epigenome base editing.