CRISPR-free base editors with enhanced activity and expanded targeting scope in mitochondrial and nuclear DNA

Adapting BE-PACE to evolve TALE-based DdCBEs

PACE uses an M13 phage that is modified to contain an evolving gene in place of gene III (gIII)²⁷. gIII encodes a capsid protein pIII that is essential for producing infectious phage progeny. To establish a selection circuit, gIII is encoded in an accessory plasmid (AP) within the Escherichia coli host cell such that gIII expression is dependent on the evolving activity. We previously reported a BE-PACE system to evolve CRISPR cytosine base editors²⁵. In this system, the AP encodes gIII under the control of a T7 promoter. A complementary plasmid (CP) encodes T7 RNA polymerase (T7 RNAP) fused to a degron through a 2-amino-acid linker (Fig. 1a). In the absence of C•G-to-T•A editing of the linker sequence, the degron triggers constitutive proteolysis of T7 RNAP, preventing gIII expression (Fig. 1b). The target cytosines for DdCBE-mediated editing in this selection are C₆ and C₇, where the subscripted numbers refer to their positions in the spacing region, counting the DNA nucleotide immediately after the binding site of the left-side TALE (TALE3) as position 1 (Fig. 1c). Successful C•G-to-T•A editing of either or both C₆ and C₇ targets introduces a stop codon within the linker to prevent translation of the degron tag. Active T7 RNAP then initiates gIII expression (Fig. 1b). The nucleotide at position 8 may be modified to A, T, C or G to enable selection against TC and non-TC contexts (Fig. 1b).

To enhance phage propagation in the selection circuit, we hypothesized that a DdCBE architecture with maximal editing efficiency would provide a favorable starting point to evolve activity against TC and non-TC targets. We designed a DdCBE that consisted of a left-side TALE (TALE3) and a right-side TALE (TALE4) flanking a 15-base pair (bp) spacing region, with targets C₆ and C₇ within the transcription template strand (Fig. 1c). We fused one copy of UGI to the N-terminus of the TALE protein and split DddA at G1397 to maximize editing of cytosine targets in the transcription template strand¹⁸. The resulting UGI–TALE3–DddA-G1397-N and UGI–TALE4–DddA-G1397-C fusions, which we refer to hereafter as T7-DdCBE, were encoded in the selection phage (SP) to co-evolve both halves of DdCBE (Fig. 1a). The phage genome is continuously mutagenized by an arabinose-inducible mutagenesis plasmid (MP6)²⁸ (Fig. 1a).

To modulate selection stringency, we generated host strains 1–4. Each host strain contained combinations of AP and CP with different ribosome-binding-site strengths, such that strain 1 resulted in the lowest selection stringency and strain 4 provided the highest stringency. All tested CPs encoded the TCC linker sequence (Extended Data Fig. 1a). We then tested overnight SP propagation in these host strains. At the highest stringency, we observed ~100-fold overnight phage propagation of an SP containing an active T7-DdCBE, consistent with DdCBE’s ability to edit 5′-TC targets. Notably, phage containing an inactivating E1347A DddA mutation (dead T7-DdCBE phage) did not propagate (Extended Data Fig. 1b). These results establish the dependence of phage propagation on DdCBE activity and that BE-PACE can be successfully adapted to select TALE-based DdCBEs.

Phage-assisted evolution of DdCBE toward higher editing efficiency at 5′-TC

We reasoned that beginning evolution with PANCE may be useful to increase activity and phage propagation before moving into PACE²⁶. PANCE is less stringent because fresh host cells are manually infected with SP from a preceding passage, so no phage is lost to continuous dilution.

To evolve DdCBEs for higher activity at TC targets, we initiated PANCE of canonical T7-DdCBE by infecting SP into high-stringency strain 4 transformed with MP6 (Extended Data Fig. 1a). After seven passages, phage populations from all four replicates propagated approximately 10,000-fold overnight (Extended Data Fig. 1c). Isolated clonal phages from two or more independent replicates were enriched for the mutations T1372I, M1379I and T1380I within the DddA gene (Supplementary Table 1).

To validate the editing activity associated with these DddA genotypes, we incorporated each mutation into our previously published G1397 split DdCBEs that targeted human MT-ATP8, MT-ND4 and MT-ND5 (Supplementary Note 1)¹⁸. We plasmid-transfected HEK293T cells with canonical versions of ATP8-DdCBE, ND4-DdCBE or ND5.2-DdCBE and compared their editing efficiencies to those produced from the corresponding mutant DdCBEs. Although T1372I and M1379I impaired editing, T1380I increased C•G-to-T•A conversions by an average of 1.2-fold to 2.0-fold across the three mtDNA sites (Extended Data Fig. 1d). It is possible that the benefit of T1372I and M1379I may require additional mutations evolved during PANCE but not tested in mammalian cells. These results indicate that PANCE of canonical T7-DdCBE was able to yield a DddA variant that modestly improved TC editing. We refer to the DddA (T1380I) mutant as DddA1 (Fig. 2a).

To further increase selection stringency, we conducted PACE using an SP encoding the DddA1 variant of T7-DdCBE (T7-DdCBE-DddA1). After 140 hours of continuous propagation at a flow rate of 1.5–3.0 lagoon volumes per hour, distinct mutations enriched across the four replicates, with the starting T1380I mutation maintained in all lagoons (Supplementary Table 2). We selected the most enriched genotype in each of the four replicates (DddA2, DddA3, DddA4 and DddA5) and tested their mtDNA editing efficiencies (Fig. 2a,b). DddA2, DddA3, DddA4 and DddA5 improved average TC editing efficiencies from 7.6 ± 2.4% with starting DddA to 14 ± 5.8%, 22 ± 6.1%, 21 ± 7.9% and 24 ± 4.4%, respectively, within MT-ND5 and MT-ATP8 (Fig. 2c,d).

The T1413I mutation in DddA4, which is in the C-terminal half of split DddA, improved base editing efficiency of DddA4 by an average of 1.6-fold compared to DddA1. Given that T1413I is positioned along the interface between the two split DddA halves (Fig. 2b), we hypothesized that this mutation might promote the reconstitution of split DddA halves. Incorporating T1413I into DddA5 to form DddA6 (Q1310R + S1330I + T1380I + T1413I) resulted in a modest editing efficiency improvement to 26 ± 3.7%, a 3.4-fold average improvement in TC editing activity compared to wild-type DddA (Fig. 2c d). Close to 90% of the edited alleles produced from DddA6 contained a TCC-to-TTT conversion, suggesting that consecutive cytosines are likely targets for processive base editing (Fig. 2e,f). These results establish DddA6 as a dsDNA cytidine deaminase variant with enhanced editing activity at TC sequences.

We evolved DddA6 from DddA split at G1397. To check if DddA6 is compatible with the G1333 split, we tested DddA6 at three mtDNA sites using DdCBEs split at G1333 and observed a 1.3-fold to 3.6-fold improvement in editing efficiencies compared to wild-type DddA (Extended Data Fig. 2a–c). These data indicate that mutations in DddA6 can enhance mtDNA editing efficiencies of the G1333 split variant, but the extent of improvement is lower than with the G1397 split. We noted that editing improvements mediated by DddA6 were modest at sites that exhibit efficient editing even with wild-type DddA, such as MT-ND1 and MT-ND4 (Extended Data Fig. 2a,d). For sites already efficiently edited with canonical DdCBEs, other deaminase-independent factors, such as mtDNA repair, could limit editing efficiency more than deaminase activity.

Evolving DddA variants with expanded sequence context compatibility

To assess if the enhanced activity of DddA6 would enable base editing at target cytosines not in the native TC sequence context, bacteria expressing the evolved T7-DdCBE were transformed with a plasmid library encoding NC₇N targets, where N = A, T, C or G. After overnight incubation, the plasmid library was isolated and subjected to high-throughput sequencing to measure the C•G-to-T•A conversion at each of the 16 NC₇N targets (Fig. 3a).

Consistent with earlier human mtDNA editing results, DddA6 improved the average editing efficiencies of bacterial plasmids containing TC₇N substrates by approximately 1.3-fold. DddA6-mediated editing at non-TC sequences, however, remained negligible (<0.2%) (Fig. 3b), suggesting the need to further evolve DddA to deaminate non-TC targets.

We modified the linker sequence in the CP to contain ACC, CCC or GCC (Fig. 1b). We co-transformed cells with AP1 and one of three CP plasmids (CP2-ACC, CP2-CCC or CP2-GCC) to generate three high-stringency E. coli strains (5, 6 and 7) for infection with SP encoding T7-DdCBE-DddA1 (Fig. 2 and Extended Data Fig. 3a). A large drop in overnight phage titers across strains 5, 6 and 7 suggested that the starting T7-DdCBE-DddA1 activity against non-TC sequences was too low to support PACE, so we initiated evolution with PANCE (Extended Data Fig. 3b). We first conducted a round of mutagenic drift to diversify the phage genome in the absence of selection pressure²⁹. Next, we initiated three parallel PANCE campaigns of T7-DdCBE-DddA1 (PANCE-ACC, PANCE-CCC and PANCE-GCC). Each campaign was challenged with a non-TC linker and was conducted in four replicates (Extended Data Fig. 3c–e).

We isolated phage that propagated more than 10,000-fold overnight after nine passages of PANCE. The surviving DddA genotypes were strongly enriched for N1342S and E1370K mutations across all linker targets. Positions A1341 and G1344 were replaced with different amino acids depending on the linker target (Supplementary Table 3).

Given the substantial increase in phage propagation strength after nine PANCE passages (surviving, in total, ~10^16–19-fold dilution), we increased selection stringency by challenging three surviving phage populations (PANCE-CCC-B, PANCE-GCC-A and PANCE-GCC-D) to 138 hours of PACE at a flow rate of 1.5–3.5 lagoon volumes per hour. For PACE, we used the same MP6-transformed strains 6 and 7 that had been applied earlier in PANCE (Extended Data Fig. 3a). The resulting PACE-evolved DddA variants acquired the additional mutations T1314A, E1396K and T1413I (Supplementary Table 4).

Characterizing sequence context preferences of DddA variants

From each phage population that survived selection on a CCC or GCC linker target, we sequenced 6–8 clones. We chose five PACE-evolved DddA variants (DddA7, DddA8, DddA9, DddA10 and DddA11) for further characterization based on their genotypes (Fig. 3c). We profiled their sequence context preferences using the same bacterial NC₇N plasmid assay used to characterize DddA6 (Fig. 3a).

All variants, except DddA8, maintained or improved editing efficiencies at TC (Fig. 3b). DddA9 and DddA10 resulted in approximately 2.0-fold higher TC editing than canonical T7-DdCBE but very low CC editing (<3.0%) (Fig. 3b). Although the average AC and CC editing levels by canonical T7-DdCBE were negligible (<0.66%), DddA7, DddA8 and DddA11 yielded an average of 4.3% editing at these contexts (Fig. 3b). These results demonstrate that PACE can be successfully applied to evolve for DddA variants that show expanded targeting activity beyond TC.

To validate the activity of these DddA variants in human mtDNA, we replaced wild-type DddA in ND5.2-DdCBE and ATP8-DdCBE with DddA7, DddA8, Ddd9, DddA10 or DddA11 (Supplementary Note 1). Consistent with bacterial plasmid editing results, DddA9 and DddA10 resulted in similar improvements in TC editing as DddA6 but did not exhibit consistent non-TC editing across multiple mtDNA sites (Fig. 3d,e).

Among variants tested, DddA11 supported the highest mtDNA editing efficiencies at AC (4.3–5.0%) and CC (7.6–16%) (Fig. 3d,e). Processive editing of consecutive cytosines in the spacing region could edit a preceding cytosine to a thymine, thus changing the starting ACC target into ATC₁₀ in MT-ND5 and ATC₉ in MT-ATP8 (Fig. 3d,e). To clarify if C₁₀ and C₉ are edited as ACC or ATC targets, we compared the percentage of edited alleles that contained an ACT or ATT product. We noted that most of the on-target edited alleles retained the preceding 5′-C (48% for ND5.2-DdCBE and 27% for ATP8-DdCBE) (Extended Data Fig. 4a,b). We thus classified C₁₀-to-T₁₀ and C₉-to-T₉ conversions mediated by ND5.2-DdCBE and ATP8-DdCBE, respectively, as editing of CC sequences (Fig. 3d,e).

Given that DddA11 resulted in the highest non-TC editing efficiencies, we generated eight reversion mutants of ATP8-DdCBE containing DddA11 (11a–h) to identify the contributions of individual mutations. Variants 11f, 11g and 11h resulted in detectable AC and CC conversions averaging 3.0%, indicating that a combination of at least two of the three mutations A1341V, N1342S and E1370K is sufficient to enable editing of AC and CC. To maximize non-TC editing, S13330I should be incorporated (Fig. 3f). These results collectively suggest that the combination of the six mutations—S1330I, A1341V, N1342S, E1370K, T1380I and T1413I—enable DddA11 to catalyze efficient base editing at AC and CC contexts that are poorly edited by canonical DdCBE.

Next, we tested mitochondrial base editing by DddA6 and DddA11 in three other human cell lines. We fused fluorescent markers eGFP and mCherry to the right and left halves of ND5.2-DdCBE, respectively, by a self-cleaving P2A sequence³⁰ to enable fluorescence-activated cell sorting (FACS) of nucleofected cells that express both halves of the DdCBE. For poorly transfected cells, such as HeLa, enriching for cells expressing eGFP and mCherry (eGFP⁺mCherry⁺) substantially increased TC and non-TC editing levels from less than 1% to 4–31% (Extended Data Fig. 5a and Supplementary Note 2). In K562 cells and U2OS cells, this enrichment strategy improved average editing by 11-fold and 1.5-fold, respectively, to efficiencies ranging from 20% to 60% (Extended Data Fig. 5b,c). These results indicate that cell lines other than HEK293T support improved mitochondrial base editing by evolved DddA variants.

Mitochondrial off-target activity of evolved DddA variants

To profile mitochondrial off-target editing activities of DdCBEs containing DddA6 and DddA11, we performed assay for transposase-accessible chromatin using sequencing (ATAC-seq) of whole mitochondrial genomes from HEK293T cells transfected with plasmids encoding canonical or evolved variants of ND5.2-DdCBE or ATP8-DdCBE. A sequencing depth of approximately 3,000–8,000× was obtained per sample (Supplementary Table 5).

Consistent with previous results¹⁸, the average frequencies of mtDNA-wide off-target editing arising from canonical DdCBEs (0.033 ± 0.002%) were similar to those of the untreated control or DdCBEs containing dead DddA6 (0.028 ± 0.001%) (Fig. 3g). Off-target frequencies associated with ND5.2-DdCBE were 1.5-fold higher for DddA6 and DddA11 compared to wild-type DddA and 3.0-fold to 4.8-fold higher for ATP8-DdCBE (Fig. 3g). We analyzed the average frequencies of all unique off-target single-nucleotide variants (SNVs) containing a C•G-to-T•A conversion in cells treated with ATP8-DdCBE or ND5.2-DdCBE variants (Extended Data Fig. 6a–f). Although all 27 SNVs in cells treated with canonical ATP8-DdCBE were less than 1.0% (Extended Data Fig. 6a), 76 and 159 SNVs with more than 1% editing were detected for cells treated with DddA6 and DddA11, respectively (Extended Data Fig. 6b,c). As expected, these results suggest that deaminase-dependent off-target editing increases in the presence of DddA variants with higher activity and expanded targeting scope, although the frequency of off-target editing per on-target editing event remains similar to that of canonical DdCBE (Extended Data Fig. 6g).

In addition to deaminase-dependent off-target editing, the TALE repeats also contribute to overall off-target activity. For ND5.2-DdCBEs containing DddA6 or DddA11, we observed fewer than four SNVs with more than 1% frequency—far lower than those observed in ATP8-DdCBE containing the same DddA6 or DddA11 (compare Extended Data Fig. 6b,c to 6e,f). We hypothesize that TALE repeats that bind promiscuously to multiple DNA bases are more likely to result in higher off-target editing when fused to the evolved DddA variants^31,32.

These results collectively indicate that, although mtDNA off-target editing increases for DdCBEs that use DddA6 and DddA11, consistent with their higher activity and expanded targeting scope, ratios of off-target:on-target editing remain similar to canonical DdCBE (Extended Data Fig. 6g).

Profiling the editing window of evolved DddA variants

We wondered if the improved activity of the evolved DddA6 and DddA11 variants might influence the editing window for DdCBE editing. Using the bacterial plasmid assay for context profiling, we generated a separate library of 14 target plasmids for editing by canonical and evolved T7-DdCBEs containing the G1397 split. The spacing region in each target plasmid contained repeats of TC sequences ranging from 12 bp to 18 bp. These repeats were positioned on the top or bottom DNA strand (Fig. 4a).

Across 12-bp to 18-bp spacing regions, DddA11 resulted in the highest editing efficiencies and widest editing window compared to wild-type DddA and DddA6 (Fig. 4b–h). Within 12-bp to 15-bp spacing regions, the canonical and evolved DdCBEs preferentially edited cytosines positioned 4–6 nucleotides upstream of the 3′ end of the bottom strand and 3–6 nucleotides upstream of the 3′ end of the top strand (Fig. 4b–e,i).

At 16-bp to 18-bp spacing regions, canonical and DddA6-containing DdCBEs maintained editing of target cytosines positioned six nucleotides upstream of the 3′ end of the bottom strand but failed to efficiently edit cytosines in the top strand (Fig. 4f–h). In contrast, DddA11 retained activity for top-strand cytosines positioned 6–7 nucleotides upstream of the 3′ end, but efficiencies were substantially lower compared to shorter spacing lengths (compare Fig. 4b–e to 4f–h). These results indicate that the editing windows of evolved variants split at G1397 were generally similar to those of canonical DdCBE, with DddA11 exhibiting a larger editing window for spacing lengths more than 15 bp (Fig. 4i).

Evolved DdCBE edits nuclear DNA

We previously showed that nuclear-localized DdCBE can mediate base editing at nuclear TC targets, which may provide useful alternatives to CRISPR CBEs when guide RNA or PAM requirements are limiting¹⁸ or when targeting heterochromatin sites³³ (Supplementary Note 3). To test if DddA11 also expands the targeting scope of nuclear base editing, we transfected HEK293T cells with DdCBEs that targeted nuclear SIRT6 or JAK2 (ref. ³⁴). When localized to the nucleus in the G1397 split orientation, DddA11 substantially improved AC, CC and GC editing from a typical range of 0–14% to 17–35% (Fig. 5a,b; see Extended Data Fig. 4c,d for frequencies of edited alleles). These results collectively show that DddA11 enhances non-TC editing efficiencies for all-protein base editing of both mitochondrial and nuclear DNA.

To assess for nuclear off-target editing, we used the off-target prediction tool PROGNOS³⁵ to rank human nuclear DNA sequences that were predicted to be targeted by the TALE repeats in SIRT6- and JAK2-DdCBE. We treated HEK293T cells with the canonical or evolved DdCBEs and performed amplicon sequencing of the top 9–10 predicted off-target sites for each base editor (Supplementary Table 6). The average frequencies in which C•G base pairs within the predicted off-target spacing region were converted to T•A base pairs were very similar between the canonical and evolved DdCBEs (Fig. 5c,d). These results suggest that DddA6 and DddA11 did not increase nuclear off-target editing within a subset of computationally predicted off-target sites for a given pair of TALE repeats.

Attempts to further increase activity at GC sequences

We noted that DddA11 was active mostly at GC₇C₆ and not GC₇C₆ (Fig. 3b). Given that a single C₆-to-T₆ conversion in a CC context is sufficient to generate a stop codon (Fig. 1b), the selection pressure to evolve acceptance of GC substrates was likely attenuated. To increase selection stringency, we modified the linker to encode either GC₈A or GC₈G such that only DddA variants that show activity at GC were able to restore active T7 RNAP (Extended Data Fig. 7a). We generated host strains 9 and 10 to contain the GCA and GCG linker, respectively (Extended Data Fig. 7b). Consistent with the weak GC activity of DddA11, we observed a drop in overnight phage titers in strains 9 and 10, suggesting that the activity of T7-DdCBE-DddA11 against GCA and GCG was too low to support PACE (Extended Data Fig. 7c and Supplementary Discussion).

We initiated PANCE of T7-DdCBE-DddA11 in MP6-transformed host strains 9 and 10. After 12 passages, overnight phage propagation increased to approximately 100-fold to 1,000-fold (Extended Data Fig. 7d and Supplementary Discussion). We sequenced the surviving phage isolates from round 9 and round 12 to derive four consensus DddA genotypes (Extended Data Fig. 8a and Supplementary Table 7). The evolved variants did not improve editing efficiencies or targeting scope consistently across four mtDNA sites when compared with DddA11, although we noted that variant 7.9.1 showed higher editing efficiencies at TC targets compared to DddA6 and DddA11 (Extended Data Fig. 8b–e and Supplementary Discussion). These results suggest that DddA11 variants that can process GC substrates with improved efficiency are very rare.

Installing previously inaccessible pathogenic mutations in mtDNA

To demonstrate the utility of evolved DddA variants with broadened sequence context compatibility, we designed three DdCBEs to install disease-associated C•G-to-T•A mutations at non-TC positions in human mtDNA. ND4.2-DdCBE installs the missense m.11696 G > A mutation in an ACT context. This mutation is associated with Leber’s hereditary optic neuropath³⁶. ND4.3-DdCBE installs the missense m.11642 G > A mutation in a GCT context, and ND5.4-DdCBE installs the nonsense m.13297 G > A mutation in a CCA context. Both mutations were previously implicated in renal oncocytoma³⁷ (Fig. 6a).

We compared the editing efficiencies among DdCBEs containing wild-type DddA, DddA6 or DddA11 split at G1397. Although canonical DdCBEs resulted in negligible editing at non-TC sites, DddA11 edited the on-target cytosines at efficiencies ranging from 7.1% to 29% in bulk HEK293T cell populations (Fig. 6b–d). Despite the very low levels of bacterial GC editing mediated by DddA11 (Fig. 3b), DddA11 yielded 7.1 ± 0.69% on-target GC₆ editing when tested in ND4.3-DdCBE (Fig. 6b). Most of the alleles containing the on-target edit, however, also harbored bystander edits that resulted in unintended changes to the protein-coding sequence (Extended Data Fig. 9a).

Unlike ND4.3-DdCBE, ND4.2-DdCBE and ND5.4-DdCBE resulted in higher bulk editing efficiencies ranging from 17% to 29% (Fig. 6c,d). More than 57% of the edited alleles contained the desired C•G-to-T•A on-target edit and a silent bystander edit (Extended Data Fig. 9b,c). DddA11 tested in the G1333 orientation resulted in lower on-target editing compared to the G1397 orientation (Fig. 6c). No on-target editing was detected with DddA11 when the target cytosine falls outside the preferred editing window of the G1333 split¹⁸ (Fig. 6d). These results collectively indicate that DddA11 tested in the G1397 split orientation enables much higher levels of base editing at AC, CC and GC targets than wild-type DddA, even though absolute editing levels at GC targets are modest.

Next, we assessed the phenotypic consequences of installing the missense m.11696 G > A mutation and the nonsense m.13297 G > A mutation in human mtDNA. We sorted for eGFP⁺mCherry⁺ cells to enrich those that expressed both halves of DdCBE. This enrichment increased on-target editing from 17% to 29% in unsorted cells to 35–43% in sorted cells (Fig. 6c,d). Compared to sorted cells containing dead DddA, sorted cells treated with DddA11-containing ND4.2-DdCBE and ND5.4-DdCBE exhibited reduced rates of basal and uncoupled respiration (Fig. 6e,f). These results establish that DddA11 can install candidate pathogenic mutations that canonical DdCBEs are unable to access and that these edits can occur at levels sufficient to result in altered mitochondrial function. These capabilities could broaden disease-modeling efforts using mitochondrial base editing.