To explore the outcome of targeted DNA ADP-ribosylation, we selected the previously characterized DarT2 from enteropathogenic E. coli (EPEC) O127:H6 str. E2348/69 (ref. ). EPEC DarT2 was shown to ADP-ribosylate single-stranded DNA (ssDNA) at the third position in a 5'-TYTN-3' motif (Y = C/T), with the fourth position biased against a G. Paralleling its growth-inhibitory effects in vivo, this DarT2 blocked extension by the large fragment of E. coli's DNA polymerase I in vitro from a ssDNA template with the recognition motif (5'-TCTC-3'), whereas extension was unhindered with a mutated motif (5'-ACTC-3') or with DarT2 containing the inactivating E170A substitution (dDarT2) (Fig. 1c,d and Extended Data Fig. 1).

To direct DNA ADP-ribosylation, we fused DarT2 to the N terminus of the protospacer-adjacent motif (PAM)-flexible (5'-NNG-3') Streptococcus canis Cas9 (ScCas9) (Fig. 1e). Directing the DarT2-Cas9 fusion to a target sequence through a designed single guide (sg)RNA would localize DarT2 to the nontarget strand displaced during R-loop formation (Fig. 1e). If the non-target strand contains a 5'-TYTN-3' motif accessible to DarT2, then the target thymine within the motif would be ADP-ribosylated and serve as a block to DNA replication. As wild-type (WT) DarT2 would arrest cell growth through genome-wide ADP-ribosylation, we included a previously reported spontaneous G49D substitution in the NAD-binding loop helix (DarT2) exhibiting reduced cytotoxicity. To promote repair through a provided DNA template rather than the sister chromatid, we used a nickase version of Cas9 (D10A) that only cleaves the target strand and provided a plasmid-encoded repair template with ~500-bp homology arms flanking the intended edits.

As a simple readout of homologous recombination, we introduced a premature stop codon into a chromosomally integrated kanamycin resistance gene in E. coli strain MG1655 (Fig. 1f). The premature stop codon overlaps with an ScCas9 target containing the 5'-TTTC-3' DarT2 motif and a PAM sequence, while a provided repair template with ~500-bp homology arms introduces mutations that revert the premature stop codon and remove the DarT2 motif. As part of an editing assay, plasmids encoding the editor, sgRNA and repair template are transformed into E. coli and colony counts are compared following editor induction and plating with or without kanamycin.

To set a baseline, we applied dsDNA cleavage with Cas9, which is commonly used for genome editing in bacteria. As dsDNA cleavage principally removes cells that did not undergo recombination, using Cas9 resulted in an average of 76% kanamycin-resistant colonies and a 153-fold colony reduction compared to the nontargeting (NT) control (P = 0.0002, n = 3) (Fig. 1g). The nickase version of Cas9 did not deplete colony counts (3.7-fold increase relative to the NT control, P = 0.02, n = 3) but at the expense of fewer kanamycin-resistant colonies (5.1%), in line with nicking being less cytotoxic but a poor driver of homologous recombination. Binding DNA alone with a catalytically dead Cas9 (dCas9) exhibited similar colony counts to nCas9 (P = 0.07, n = 3) and did not drive any measurable editing.

Turning to append editing with DarT2, the DarT2-nCas9 fusion yielded an average of 97% kanamycin-resistant colonies and negligible depletion in colony counts compared to its NT control (1.7-fold increase; P = 0.25, n = 3) (Fig. 1g). Both DNA ADP-ribosylation and opposite-strand nicking were important, as conferring kanamycin resistance was less effective with nicking alone (dDarT2-nCas9, 0.18%; P = 0.003, n = 3) or ADP-ribosylation alone (DarT2-dCas9, 43%; P = 0.029, n = 3) when compared to DarT2-nCas9. All screened kanamycin-resistant colonies contained the intended edit (Extended Data Fig. 2). DarT2 still conferred cytotoxicity, as cell counts were low even for the NT controls and increased upon deactivation of DarT2 (Fig. 1g), creating an opportunity to further attenuate the toxin. Collectively, append editing with DarT2 drives homologous recombination with a provided template in E. coli, yielding editing that outperforms traditional Cas9-based approaches but with target-independent cytotoxicity.

Our reporter assay requires homologous recombination to confer kanamycin resistance. However, chemically modifying DNA bases can lead to single-nucleotide edits as demonstrated by BEs. We, therefore, asked whether append editing could drive editing without antibiotic selection but also induce base mutagenesis. First, we repeated the kanR reporter assay in the absence of kanamycin selection and performed amplicon sequencing on the target site from liquid culture (Fig. 1h). Under targeting conditions, append editing yielded 82% of total reads with the desired edit that drastically dropped with nicking alone (0.9%), paralleling the fraction of kanamycin-resistant colonies (Fig. 1g). Of the remaining reads, the few detected substitutions of the ADP-ribosylated thymine were not significantly elevated in any particular sample (F = 1.03, P = 0.39, df = 3) (Extended Data Fig. 3). As homologous recombination could overshadow base editing, we performed the assay in the absence of the repair template. However, the 16 screened colonies only yielded the original sequence (Supplementary Fig. 1). Therefore, append editing with DarT2 did not result in detectable base edits in E. coli, further supporting sole triggering of homologous recombination.

Base editing can also occur at genomic sites unrelated to the target sequence presumably through the DNA modification domain acting on temporary ssDNA. Given the lack of obvious substitutions at the target site with append editing, we hypothesized that DarT2 expression would not lead to such edits associated with BEs. Culturing editor-expressing cells and performing whole-genome sequencing of three individual clones (Fig. 1i and Supplementary Table 1), a cytosine BE (CBE) yielded the expected C-to-T edits, with either three or eight edits in each clone. In contrast, the ADPr-TA editor yielded no T-to-G edits and few T-to-C edits similarly to the CBE or no editor. One of the three clones with the ADPr-TA editor yielded a single T-to-A edit, whereas none were observed with the CBE or no editor. This one edit was associated with the 5'-TYTN-3' motif, suggesting that base mutagenesis is possible but rare (Supplementary Table 1). Thus, even a highly active DarT2 that reduces cell viability (Fig. 1g) does not inherently drive base edits across the E. coli genome.

ADPr-TAE yielded high editing efficiencies, although the expressed DarT2 exhibited strong cytotoxicity (Fig. 1g). As the cytotoxicity was likely because of ADP-ribosylation of ssDNA across the genome, we aimed to attenuate DarT2 without compromising localized ADP-ribosylation and subsequent initiation of homologous recombination using structural insights and sequence conservation (Fig. 2a). While the structure of EPEC DarT2 remains to be experimentally determined, a crystal structure is available for the Thermus sp. 2.9 DarT2 that shares 34% amino acid identity with EPEC DarT2 (ref. ). Aligning this structure with the AlphaFold-predicted structure of EPEC DarT2 (ref. ), we selected a subset of residues potentially involved in binding the DNA recognition motif (M84, M86, R57, R92 and R166) or potentially flanking regions of the DNA strand not captured in the crystal structure (R193). The positively charged arginines were substituted to uncharged alanine, while the methionines were substituted to leucine to disrupt the coordinating sulfur while preserving the residue's hydrophobicity and chain length. Testing these substitutions in combination with G49D as part of the kanamycin resistance reversion assay (Fig. 1f), we found that all improved cell viability (Fig. 2b). At the same time, three of the substitutions (M86L, R92A and R193A) maintained the fraction of kanamycin-resistant colonies comparable to the original G49D (P = 0.77, 0.51 and 0.27, respectively, n = 3) (Fig. 2b), representing candidates for further use with append editing.

Viability was greatly enhanced across the single-substitution variants, yet DarT2 may still exert target-independent ADP-ribosylation that could have more subtle effects on cell growth and behavior. We, therefore, generated cells hypersensitive to ADP-ribosylation by deleting the core repair gene recA to disable homologous recombination and assessed cell growth when expressing each ADPr-TAE variant under non-targeting conditions (Fig. 2c and Supplementary Fig. 2). While growth rates in the exponential phase were similar (Supplementary Fig. 2), we observed marked differences upon entry into the stationary phase. In particular, amino acid substitutions that previously compromised editing (M84L, R57A and R166A) yielded final turbidities paralleling the inactivating E170A (P = 0.35, 0.65 and 0.22, respectively, n = 3) (Fig. 2d and Supplementary Fig. 2). In contrast, substitutions that previously showed high editing efficiencies (M86L, R92A and R193A) exhibited a final turbidity similar to G49D alone (P = 0.99, 0.05 and 0.17, respectively, n = 3) and lower than E170A. We, therefore, combined the high-editing-efficiency substitutions (M86L, R92A and R193A) into a four-substitution version of DarT2, DarT2. This version maintained cell viability and a high frequency of kanamycin-resistant colonies (49%) in E. coli MG1655 (Fig. 2b). Moreover, in the recA-deletion (ΔrecA) strain, the append editor with DarT2 restored final turbidity to approach that of the editor lacking ADP-ribosylation (E170A; P = 0.09, n = 3) (Fig. 2d).

By improving cell viability and growth in a strain in which homologous recombination was fully disabled, the append editor with DarT2 afforded the opportunity to probe the genetic basis of templated-mediated editing. Prior work on the cytotoxicity of DarT2 in E. coli revealed a key role by RecF and possibly nucleotide excision repair. However, the involved DNA repair pathways as part of targeted ADP-ribosylation with opposite-strand nicking could differ. Within the kanamycin reversion assay (Fig. 1f), recA was essential for editing and even showed some reduction in colony counts under non-targeting conditions (Fig. 2e). Disrupting the RecBCD branch of recombination (ΔrecB) reduced viability but also increased the frequency of kanamycin-resistant colonies, suggesting a role in survival in the absence of recombination with the provided repair template. In contrast, disrupting the alternative RecFOR recombination pathway (ΔrecF and ΔrecO) reduced editing relative to the WT (one-sided Welch's t-test, P = 0.048 and 0.001, respectively, n = 3) but not viability for recF (one-sided Welch's t-test, P = 0.40, n = 3), suggesting involvement in templated recombination. Disrupting RecA-independent RecT recombination (ΔrecT) significantly reduced both viability and editing (one-sided Welch's t-test, P = 0.002 and 0.003, respectively, n = 3), suggesting involvement in both survival and templated recombination. Lastly, disruption of the DNA repair exonuclease RecJ (ΔrecJ), mismatch repair (ΔmutS), base excision repair (ΔxthA) and nucleotide excision repair (ΔuvrA) did not impact editing (one-sided Welch's t-test, P = 0.89, 0.68 and 0.81, respectively, n = 3) or viability (one-sided Welch's t-test, P = 0.87, 0.24 and 0.93, respectively, n = 3) relative to the WT. These findings implicate multiple recombination pathways as part of ADPr-TAE in E. coli.

Append editing with DarT2 efficiently reverted the premature stop codon in the kanamycin reversion assay. However, the reliance on homologous recombination lends to a much broader range of edits in different genes and bacteria. We, therefore, explored the bounds of ADPr-TA editing. For simplicity, editing was performed around the premature stop codon in the kanamycin reversion assay. When testing edits beyond reversion of the stop codon, editing efficiency was determined without kanamycin selection by assessing the size of the target site or sequence of individual colonies.

Beginning with the homology arms, condensing their length from ~500 to 100 bp reduced the frequency of kanamycin resistance from 86% to 28%, whereas arm lengths of 50 bp and below exhibited virtually no kanamycin resistance (Supplementary Fig. 3). Continuing with ~500-bp homology arms, we tested increasingly larger replacements, deletions and insertions (Fig. 2f-h). Replacements extending up to 60 bp upstream or downstream of the target site or 91 bp spanning the target site were present in 80-100% and 50-75% of screened colonies, respectively, either as complete or partial conversions (Fig. 2g and Extended Data Fig. 4). Separately, deletions up to 91 bp were present in 90-100% of screened colonies, albeit with a high fraction of partial conversion with the largest deletion. Lastly, insertions of 10 bp and 100 bp were present in 100% and 50-90% of screened colonies, respectively. No colonies contained an insertion of 500 bp (Supplementary Fig. 4), indicating an upper limit to recombination. Editing was not limited to this target site in E. coli, as we could introduce substitutions at four additional targeted genes in E. coli (Extended Data Fig. 5a) and one targeted gene in the pathogen Salmonella enterica (Extended Data Fig. 5b). Collectively, ADPr-TAE can introduce ranging replacements, insertions and deletions in bacteria without sacrificing viability.

Given that append editing drove templated recombination in bacteria, we asked whether eukaryotes would undergo similar editing outcomes. Beginning with the baker's yeast Saccharomyces cerevisiae cultured as a haploid, we transformed plasmids encoding the DarT2 append editor, an sgRNA and a repair template with ~250-bp homology arms to introduce a premature stop codon as part of six substitutions in the FCY1 gene. Individual colonies were then screened on the basis of Sanger sequencing of the target site (Fig. 3a and Extended Data Fig. 6a,b). Append editing with DarT2-nScCas9 yielded templated edits in only 17% of the screened colonies, a reduced frequency compared to 50% generated via dsDNA breaks with ScCas9 (Fig. 3b). No edited colonies were obtained under non-targeting conditions or with DNA nicking alone, affirming the necessity of either dsDNA breaks or targeted ADP-ribosylation for templated editing.

Beyond templated edits achieved with targeted ADP-ribosylation, we also observed a distinct set of edits in 25% of the screened colonies: conversion of the ADP-ribosylated thymine into a different base (Fig. 3c and Extended Data Fig. 6a). These base substitutions principally occurred at the thymine expected to undergo ADP-ribosylation by DarT2, with the modified base becoming an A (67%) or a C (33%) (Fig. 3c). Such edits were absent with any of the other tested editors (Fig. 3c). Homologous recombination and base mutagenesis represented mutually exclusive repair outcomes, as removing the repair template enhanced the mutagenesis frequency without altering the location and distribution of mutations (Fig. 3c,d and Extended Data Fig. 6b). Base mutation was also observed when targeting sites within the genes ALP1 and JSN1, albeit at lower frequencies (Supplementary Fig. 5a,b). Thus, in yeast, append editing drives either homology-directed repair (HDR) or mutagenesis of the ADP-ribosylated thymine.

The outcomes of append editing in yeast represented a major deviation from what we observed in tested bacteria and could reflect distinct editing outcomes in eukaryotes at large. However, in contrast to higher eukaryotes, S. cerevisiae engages in nonhomologous end joining less frequently and lacks poly(ADPr) polymerases involved in dsDNA break repair that add and extend ADPr groups on DNA ends. We, therefore, assessed the impact of ADPr-TAE in the model plant Nicotiana benthamiana. As a simple and fast assay, Agrobacterium constructs encoding the append editor were injected into N. benthamiana leaves, after which the type and frequency of edits were assessed by targeted amplicon sequencing of transfected tissues (Fig. 3e). In this setup, no repair template was included given the generally low frequencies of homologous recombination in this type of transfection assay in plants. Additionally, the ScCas9 component of the append editor was exchanged for Streptococcus pyogenes Cas9 (SpCas9) to use available constructs.

Despite expectedly low transfection efficiencies, we could measure substitution of the ADP-ribosylated thymine as the dominant outcome in 1.4% of reads targeting the PDS1 gene (Fig. 3f,g). This thymine was converted to the three other bases but with a bias toward A (59%) over C (19%) and G (22%). Testing two other target sites within PDS1, including one containing multiple DarT2 motifs, resulted in similar mutagenesis of the ADP-ribosylated T, with a bias toward A (Fig. 3g and Supplementary Fig. 6). Indels were observed under targeting conditions but at frequencies 6-80-fold lower than base mutagenesis (Supplementary Fig. 7). Thus, append editing can drive mutagenesis of the ADP-ribosylated base in both yeast and plants, reflecting distinct editing outcomes from those we observed in bacteria.

As a final but important branch of eukaryotes, we sought to explore append editing in human cells. Unlike S. cerevisiae and N. benthamiana, human cells possess an o-acyl-ADPr deacylase (OARD1), also known as TARG1, that was previously shown to reversibly remove the ADPr moiety appended to thymines by DarT2 (Fig. 4a). We, therefore, began by assessing ADPr-TAE in human cells with an intact or disrupted TARG1 gene (Supplementary Fig. 8) using SpCas9 on the basis of available constructs. Plasmid constructs encoding an SpCas9-based editor and an sgRNA were transiently transfected into HEK293T cells and editing was assessed through next-generation sequencing of the target site in EMX1 without sorting or selection of transfected cells (Fig. 4b). An oligonucleotide repair template specifying a nine-base substitution and four-base deletion was included to evaluate both homologous recombination and base mutagenesis in parallel.

Using SpCas9 in HEK293T cells as a baseline, we observed matching extents of templated edits (22%) and small indels (32%), with no significant difference in the absence of TARG1 (P = 0.99 and 0.94, respectively, n = 3) (Fig. 4c). Nicking similarly generated a high level of templated edits whether or not TARG1 was intact (18%) but with minimal small indels (0.4%) because of the lack of dsDNA breaks. The append editor with DarT2 also yielded templated edits, with the editing frequency increasing from 7% to 10% by disrupting TARG1. However, no significant differences were observed for append editors with the attenuated DarT2 or with dDarT2 (P = 0.32 and 0.33, respectively, n = 3), suggesting that the templated edits were driven primarily through DNA nicking rather than DNA ADP-ribosylation.

At the same time, the ADPr-TA editor with DarT2 yielded 9% base substitutions specifically at the ADP-ribosylated thymine within two overlapping DarT2 recognition motifs, but only with TARG1 disrupted (Fig. 4c). Base substitutions were negligible with DarT2 (0.2%) or dDarT2 (0.3%), suggesting that higher levels of ADP-ribosylation were necessary to drive editing (Fig. 4c). Indel frequencies for ADPr-TAE were slightly elevated over nCas9 with TARG1 disrupted (1.5% versus 0.9%; P = 0.03, n = 3) but still 22-fold lower than that observed with Cas9 (33%) (Fig. 4c), indicating that the principal repair outcome of ADP-ribosylation and opposite-strand nicking is base mutagenesis. Thus, ADPr-TAE in HEK293T cells drives base mutagenesis similarly to what is observed in plants and yeast, but only in the absence of TARG1.

As different oligonucleotide templates revealed reduced templated repair with increased base mutagenesis (Supplementary Fig. 9), we repeated the editing assay without the oligonucleotide template. Base mutagenesis at both modified thymines increased to 16% (Fig. 4d), with conversion to either A or C at similar frequencies. Additionally, base mutagenesis was reduced 20-fold to 0.8% in the absence of DNA nicking, indicating the importance of the nick (Fig. 4d). We also observed a low frequency of deletions up to ~25 bp that were elevated with DNA nicking (Supplementary Fig. 10), paralleling observations with BEs. Probing base mutagenesis beyond this target site, we performed transient transfections without the oligonucleotide template at 16 additional target sites in five genes containing one or more DarT2 recognition motifs (Fig. 4e and Supplementary Fig. 11). We observed measurable editing at all but two of these sites, with editing frequencies reaching up to 39% (Fig. 4e and Supplementary Fig. 11). Similar trends were observed in U2OS ∆TARG1 cells, with generally lower editing frequencies (up to 5.0%) likely because of lower transfection efficiencies (Extended Data Fig. 7a,b). We could also couple DarT2 with the nearly PAM-less SpRY variant of SpCas9 (ref. ) to drive base substitutions through non-NGG PAMs (Extended Data Fig. 8).

Given the need to delete TARG1 to observe editing, we assessed the ability to transiently silence TARG1 expression to promote editing in WT HEK293T cells using RNA interference (Fig. 4f). Of three tested small interfering RNAs (siRNAs) that each reduced TARG1 transcripts by at least 75% (Supplementary Fig. 12a), one siRNA yielded significant editing across all four tested target sites (Fig. 4g and Supplementary Fig. 12b). Editing was greatly diminished compared to that in cells lacking TARG1 (for example, 39% in HEK293T ∆TARG1 versus 0.8% in WT cells with TARG1-siRNA at PTEN site 2), suggesting that residual TARG1 blocks editing with DarT2. Overall, these results show that append editing with DarT2 principally drives substitution of the ADP-ribosylated base in human cells, with TARG1 posing a barrier to editing.

The expanded set of target sites allowed us to explore unique features of base mutagenesis. Across these sites, editing principally occurred at the modified thymine falling between positions 3 and 9 of sgRNA guide (Fig. 5a). For targets with multiple DarT2 recognition motifs, co-occurring mutations were observed 1.1-fold to 5.1-fold more frequently than expected if the motifs could be edited independently (Supplementary Fig. 13). Across these sites, we noticed distinct mutagenesis distributions that strongly depended on the DarT2 recognition motif (Fig. 5b). Specifically, 5'-TCTN-3' motifs were associated with similar conversion frequencies to A and C. In contrast, 5'-TTTN-3' were associated with a strong bias toward A, with secondary edits biased toward C (5'-TTTA-3') or equally split between C and G (5'-TTTC-3'). We further assessed the frequency of small indels at selected target sites. Compared to Cas9, the append editor resulted in 6-110-fold lower indel frequencies (Fig. 5c). Indel frequencies measured by next-generation sequencing or predicted using the Rule Set 2 scoring method at each target site with Cas9 correlated with base mutagenesis frequencies (Spearman correlation, ρ = 0.80 and 0.58, respectively) (Extended Data Fig. 9), indicating that indel formation with Cas9 offers a starting point to identify efficient sites for append editing. Additionally, the append editor resulted in 3-21-fold fewer kilobase-scale deletions compared to Cas9, as detected through long-read sequencing (Fig. 5d and Supplementary Fig. 14).

Beyond unintended on-target edits, we also investigated guide-independent off-target effects knowing that errant DNA ADP-ribosylation could drive base substitution (Fig. 5e). As a point of comparison, we used the CBE BE4 previously demonstrated to introduce cytosine base edits at orthogonal R-loops. Creating orthogonal R-loops at each of five sites using the dead Cas9 from Staphylococcus aureus (dSaCas9), BE4 generated cytosine base substitutions significantly more often than a no-editor control at two of the sites (Fig. 5e and Extended Data Fig. 10). Intriguingly, base substitutions in the R-loop were highest when BE4 was targeted ~650 bp upstream of the R-loop (Fig. 5e and Extended Data Fig. 10), suggesting enhanced frequencies of sgRNA-independent editing in the vicinity of the target site. In contrast, the append editor did not result in any increase in base mutations at the thymine within the DarT2 motif across all five sites compared to the no-editor control. These results are in line with the need for opposite-strand nicking to drive append editing at the target site (Fig. 4d).

BEs using thymine glycosylases were recently reported, raising the question how editing of the modified thymine compares between base excision and ADP-ribosylation. We, therefore, assessed editing outcomes at two target sites in HEK293T ∆TARG1 cells with our append editor and the deaminase-free (DAF) thymine BE (TBE) as a representative example (Fig. 5f). DAF-TBE edited multiple thymines within the target, with the most efficient editing at position 5 of the sgRNA guide. The append editor edited only the thymine in the recognized motif, with higher editing than any single thymine with DAF-TBE at both sites (P = 0.0001 and 0.052). Interestingly, the editing profiles were distinct, with the DAF-TBE predominantly yielding T-to-C or T-to-S edits compared to predominantly T-to-A or T-to-M edits using append editing. Both editors exhibited similarly low levels of indel formation and large deletions (Supplementary Figs. 15 and 16). These results show that ADP-ribosylation and base excision of thymine drive distinct editing outcomes.