Programmable C‐to‐U RNA editing using the human APOBEC3A deaminase

Abstract

Programmable RNA cytidine deamination has recently been achieved using a bifunctional editor (RESCUE‐S) capable of deaminating both adenine and cysteine. Here, we report the development of “CURE”, the first cytidine‐specific C‐to‐U RNA Editor. CURE comprises the cytidine deaminase enzyme APOBEC3A fused to dCas13 and acts in conjunction with unconventional guide RNAs (gRNAs) designed to induce loops at the target sites. Importantly, CURE does not deaminate adenosine, enabling the high‐specificity versions of CURE to create fewer missense mutations than RESCUE‐S at the off‐targets transcriptome‐wide. The two editing approaches exhibit overlapping editing motif preferences, with CURE and RESCUE‐S being uniquely able to edit UCC and AC motifs, respectively, while they outperform each other at different subsets of the UC targets. Finally, a nuclear‐localized version of CURE, but not that of RESCUE‐S, can efficiently edit nuclear RNAs. Thus, CURE and RESCUE are distinct in design and complementary in utility.

A new tool—the “CURE” base editor—based on an APOBEC3A‐dCas13 fusion functioning with unconventional gRNAs increases efficiency and reduces off‐target substitutions in C‐to‐U RNA editing.

Introduction

Site‐Directed RNA editing (SDRE), including A>I and C>U editing, is an essential complement to DNA base editing for both basic research and patient treatment (Rees & Liu, 2018; Mao et al, 2019; Montiel‐Gonzalez et al, 2019; Vogel & Stafforst, 2019; Reardon, 2020). In particular, whereas DNA editing is irreversible and so its use restricted to repairing the mutations in genetic diseases, RNA editing is reversible and thus theoretically fit not only for correcting mutations at the RNA level to treat genetic diseases, but also for altering wild‐type protein sequences in non‐genetic conditions such as inflammation and pain; indeed, a major advantage of SDRE over DNA editing is its applicability to non‐genetic conditions (Montiel‐Gonzalez et al, 2019). Intensive effort has led to the development of many versions of A>I RNA base editors, all exploiting the human RNA adenine deaminase ADAR proteins (Montiel‐Gonzalez et al, 2013; Montiel‐González et al, 2016; Cox et al, 2017; Sinnamon et al, 2017; Vallecillo‐Viejo et al, 2018; Vogel et al, 2018; Katrekar et al, 2019; Rauch et al, 2019). In contrast, C>U editors have remained elusive till recently. Specifically, via a combination of rationale mutagenesis and protein evolution, the human ADAR2 deaminase domain was successfully converted into a bifunctional enzyme capable of deaminating not only adenosines but also cytidines; the modified enzyme was fused to the catalytically dead RNA‐targeting CRISPR‐Cas13b (dCas13b) to create the bifunctional editor named RESCUE for programmable C> U and A>I editing (Abudayyeh et al, 2019). Although RESCUE can selectively deaminate cytidine (but not adenosine) on the bulged target nucleotides in the presence of proper gRNAs, it has massive global off‐target effects on adenosines. Introducing a point mutation (S375A) into the ADAR2 deaminase domain in RESCUE markedly suppresses the off‐target editing, producing a high‐specificity version of RESCUE dubbed RESCUE‐S (Abudayyeh et al, 2019). However, low levels of off‐target edits (both A>I and C>U) persist in RESCUE‐S. Furthermore, our reanalysis of the published data indicates that editing at endogenous transcripts by RESCUE is strongly biased against the GC and CC motifs, a situation exacerbated in RESCUE‐S which is generally weaker than RESCUE (Appendix Fig S1). Indeed, the editing efficiencies of RESCUE‐S at GC and CC on endogenous transcripts are typically below 2%, making it practically inapplicable. Finally, although RESCUE(‐S) can edit both AC and UC, UC editing is less efficient, with the rates below 4% at some targets (Appendix Fig S1A and B).

We have developed a distinct editing system named C>U RNA Editor (CURE), the first programmable C>U editor that does not deaminate adenine, with unique benefits in terms of both on‐target and off‐target editing.

Results

Development of the CURE system

In humans, C>U RNA editing is catalyzed by the cytidine deaminases of the APOBEC family, which comprises 11 members including Apobec1A (A1A), A2, A3A, and A3G (Salter et al, 2016). Among the various members, the best‐defined is A3A, which preferentially edits the target C (in the UC dinucleotide) located in particular forms of hairpins, as exemplified by the hairpin in the SDHB mRNA comprising a 5‐bp stem linked to a tetra‐loop that ends at the target C (Fig 1A; Sharma & Baysal, 2017). Stem stability, loop length, and target C position in the loop all dramatically impact editing. For example, moving the UC dinucleotide in the SDHB hairpin loop one nucleotide upstream reduces editing efficiency 10‐fold (Sharma & Baysal, 2017). Thus, there are strict structural requirements at the target sites for A3A‐mediated RNA editing.

Figure 1. Development of CURE .

• A
  Secondary RNA structure of a well‐defined A3A substrate harbored in the human SDHB mRNA. The target C is situated at the end of the tetra‐loop (UAUC).
• B
  Various versions of CURE. CURE‐C1 consists of A3A (Y132D) or A3A* fused to dCas13b via a nuclear export sequence (NES), with a nuclear localization signal (NLS) added at the N‐terminus of dCas13b. CURE‐C2 and CURE‐N were derived from CURE‐C1 by replacing the NLS with a mutant version (NLS*) and appending an extra copy of WT NLS, respectively. CURE‐X comprises A3A (Y132D) inserted into a flexible loop on dCasRx, with part of the loop (aa 559–585) deleted during plasmid construction (Zhang et al, 2018).
• C
  The GFP reporter system. We split GFP into two parts (β1–10 and β11), flanked each with a pair of interacting leucine zippers (ZIP) to prevent their spontaneous association (To et al, 2016), and inserted a 90‐bp linker (blue) between GFP (β1–10) and the downstream Zip. The linker bears CGA (preceded by UAU to mimic the SDHB tetra‐loop), where C>U editing would convert CGA (R12) to a stop codon (UGA), thus eliminating the downstream ZIP. The loss of this ZIP triggers GFP reconstitution (To et al, 2016), an event detectable by FACS. To reduce editing‐independent reconstitution (leakiness), the ZIP was flanked by a NLS and a degron (Chung et al, 2015).
• D–G
  GFP reporter editing by CURE‐C1. gRNA#1, predicted to induce a 14‐nt loop encompassing UAUC, was tested together with various controls (gRNA#2–5; Fig 1D). Plasmids expressing CURE‐C1-P2A‐mCherry, a gRNA, or the GFP reporter were co‐transfected into HEK293 cells, and editing measured 2 days later by FACS or sequencing. Representatives of FACS plot and Sanger chromatograph are shown (E and F), together with NGS analysis of editing efficiencies (G, displaying mean ± SEM, n = 3). The values in F are % editing. The blue boxes in the FACS plots indicate the gates for the cell populations analyzed, with the blue numbers being the abundance (%) of the gated cells, and the green and red numbers their mean fluorescence intensities of GFP and mCherry, respectively. Non‐targeting (NT) gRNA is identical to gRNA #1 except that the spacer sequence is irrelevant to the target transcript.
• H
  Editing of endogenous mRNAs. The target sites and gRNAs are depicted at the left, with the edited codon (auC) underlined. CURE‐C1-P2A‐mCherry and gRNAs were co‐expressed in HEK293T cells and the mCherry⁺ cells analyzed 2 days later by Sanger sequencing and NGS in parallel. The bar graph displays mean ± SEM (n = 3). The gRNAs carried 32‐nt spacers capable of inducing 14‐nt loops (designated 32‐14 hereafter). NT, non‐targeting RNA.

To harness A3A to create C>U RNA Editor (CURE), we fused it to the catalytically dead mutants of Cas13. Several key versions of CURE were created, the first being CURE‐C1 (CURE‐Cytoplasmic1), comprising A3A (Y132D) fused via a nuclear export sequence (NES) to the C‐terminus of dPspCas13b (Cox et al, 2017; Fig 1B). Y132D was found to increase the editing efficiency (see further). CURE‐C1 also carries a Nuclear Localization Signal (NLS) at the N‐terminus, intended to drive a significant portion of CURE‐C1 molecules into the nucleus, thus potentially enhancing mRNA editing. The mRNA editing rate was indeed increased (see further), but unexpectedly, CURE‐C1 proved predominantly (albeit not exclusively) cytoplasmic (Appendix Fig S2A, top), but this subcellular distribution pattern should not hamper our characterization of the basic properties of the CURE system.

We also constructed a reporter system based on split GFP, where deamination at the target C (in the context of UAUC, as in the SDHB tetra‐loop) would produce GFP fluorescence detectable by FACS (Fig 1C). GFP signal offers a convenient and fast albeit indirect and preliminary readout of editing. In contrast, sequencing of the reporter transcript enables direct and reliable quantifications of editing, with Sanger sequencing producing results generally consistent with that obtained using next‐generation sequencing (NGS; Sharma et al, 2016, 2017), especially when the chromatograms are analyzed with the EditR method (Kluesner et al, 2019).

As the UAUC in the SDHB RNA is embedded in the tetra‐loop on top of the 5‐bp stem, we reasoned that creating a similar structure (e.g., a 14‐nt loop with centrally positioned UAUC) at our reporter transcript might facilitate editing. Therefore, we designed a gRNA capable of inducing such a loop (gRNA#1, Fig 1D). Indeed, gRNA#1 markedly increased GFP fluorescence as compared with a non‐targeting gRNA carrying a random spacer sequence (Fig 1E). Sanger sequencing and NGS revealed that gRNA#1 induced efficient (~40%) C>U editing at the target C (Fig 1F and G, lane 1–2). In contrast, gRNA#2, capable of loop induction but not editor recruitment, was inactive, as were gRNA#3–5 capable of editor recruitment but not loop induction (Fig 1G, lane 3–6), indicating loop formation and CURE‐C1 recruitment were both necessary for editing. In agreement with this, A3A, alone or in combination with the free dCas13b, failed to edit the target even in the presence of gRNA#1 (Fig 1G, lane 7–8). These data demonstrate that CURE‐C1 was recruited to the target site at the reporter transcript via gRNA‐dCas13b interaction, where it could deaminate the target C if and only if embedded in an induced loop.

We then tested the ability of CURE‐C1 to edit UC at endogenous mRNA transcripts. We opted to focus on the codon AUC where the C>U edit would produce synonymous substitutions (I>I), in order to avoid potential confounding effects resulting from protein alterations. We found that CURE‐C1 edited multiple endogenous targets (ACTB, GAPDH, and TYMS) as efficiently as it did the GFP reporter (40–50%; Fig 1H).

As mentioned before, despite the presence of NLS, CURE‐C1 was predominantly cytoplasmic. Consistent with this, CURE‐C2 (Fig 1B), a CURE‐C1 variant carrying a mutant NLS and therefore exclusively cytoplasmic (Appendix Fig S2A, bottom), was as active as CURE‐C1 in editing the endogenous transcripts (Appendix Fig S2B). In contrast, deleting the NLS from CURE‐C1, removing the Y132D mutation, or replacing it with other mutations, each negatively affected CURE‐C1 activity (Appendix Fig S2C and D; it is unclear how a mutant NLS could enhance editing efficiency). Finally, fusing A3A to the N‐terminus of dCas13b as opposed to the C‐terminus also impaired editing (Appendix Figs S3 and S4A, compare A3A‐dCas13b vs. dCas13b‐A3A).

Collectively, these data indicate that A3A is tractable for SDRE. We have also tested multiple other Apobec proteins including A1, A2, and A3G, but without success (Appendix Figs S3–S4).

Creation of CURE‐N and CURE‐X, the potential high‐specificity versions of CURE

A3A overexpression in HEK293 cells is known to induce C>U editing at ~ 4,200 sites at the transcriptome (Sharma et al, 2017), suggesting that CURE‐C may also has such off‐target effects. We took two approaches to address this potential problem. First, inspired by the finding that nuclear localization of an A>I RNA editor can reduce its off‐target effects (Vallecillo‐Viejo et al, 2018), we sought to create a nuclear‐localized version of CURE. To this end, we added a copy of NLS to CURE‐C1 at the C‐terminus, finding the resulting editor (CURE‐N; Fig 1B) indeed nuclear and remained highly active at the GFP reporter (Appendix Fig S5A). The second approach of off‐target reduction involves CasRx, another member of the Cas13 family(Konermann et al, 2018). In contrast to Cas13b, CasRx is known to possess multiple flexible loops(Zhang et al, 2018). We have found, during the optimization of an A>I RNA editor comprising dCasRx fused to the ADAR2 deaminase domain, that inserting the deaminase domain into Loop 3 of dCasRx helps minimize the global off‐target effects of the editor without compromising on‐target editing (accompanying manuscript). Accordingly, we inserted A3A(Y132D) into Loop 3 of dCasRx (CURE‐X; Fig 1B). Remarkably, CURE‐X was more active than all other fusion proteins tested, including a conventional N‐terminal fusion configuration (Appendix Fig S5B).

Further characterization of the CURE system

We next systematically defined the determinants in gRNA and GFP reporter sequences that may impact editing. We tested both CURE‐C1 and CURE‐X because the two editors may behave differently, given the marked differences in both the dCas moiety and the fusion configuration. On the other hand, CURE‐C2 and CURE‐N were both highly similar in structure to CURE‐C1, and therefore not tested.

For CURE‐C1 editing, we first fixed the loop length at 14 nt at the GFP reporter but varied the spacer length (24–60 nt), finding diverse lengths (24–52 nt) capable of inducing editing, the efficiencies exceeding 20% and peaking at 38%, the latter achieved with the 32‐nt spacer (Fig 2A; see Appendix Fig S6A for a parallel FACS analysis). The same trend was observed at a few shorter (10‐nt and 6‐nt) loops tested (Appendix Fig S6B). We then fixed the spacer length at 32 nt, but systematically varied the loop length (4–20 nt) by flanking UAUC with increasing numbers of bases, adding equal numbers of bases to each side so as to maintain the central position for UAUC. Loops of all these lengths supported efficient editing (40–58%) except the tetra‐loop (16%; Fig 2B). A caveat is that when the loop length was increased, so was the distance of the target C relative to the ends of the loop, which complicated data interpretations. To isolate the potential effects of target C positions, we fixed the loop length to 14 nt and placed the C at various (5^th to 11^th) positions. Since repositioning the target C can alter its sequence context and confound the analysis, we moved it together with seven flanking bases (AUAUCGAG) as a single unit. Efficient (34–55%) editing was detected at all these positions, with a broad peak spanning the 6^th–9^th positions (Fig 2C). These data demonstrate that at the induced loops, CURE‐C1‐catalyzed deamination of UC was quite robust to the context (loop/spacer length and target C position), which was unexpected given the stringent structural requirements imposed on the natural A3A targets. Presumably, dCas13b‐enforced recruitment of A3A was able to make CURE‐C1 tolerant of suboptimal target structures, which greatly simplified gRNA design for SDRE.

Figure 2. Parameters affecting editing by CURE‐C1 and CURE‐X.

• A–D
  Basic features impacting editing: gRNA spacer length (A), loop length (B), target C position at the induced loop (C), and the bases flanking the target C (D). In (B‐D), the spacer for CURE‐C1 is 32 nt as depicted, whereas that for CURE‐X is 28 nt (not shown). Values are mean ± SEM (n = 3). NT, non‐targeting.
• E
  Sparing a bystander in a target site. The target region is a synthetic 200‐nt fragment from the human PGAP2 transcript bearing a U>C mutation (red) to be corrected. The bystander (green) was included in or excluded from the induced loops depending on the gRNAs. This transcript was co‐expressed with gRNA and CURE‐C1 in HEK293T cells, and editing analyzed 2 days later by Sanger sequencing (Fig 2E, bottom left, bar graph along with a representative chromatogram, where the values are % editing). To compare CURE‐C1 and CURE‐X, select gRNAs (32‐6, 32‐10, 32‐14, and 28‐14) were expressed and the transcript analyzed by NGS (Fig 2E, right). The bar graphs display mean ± SEM (n = 3).

Data information: For each parameter tested, the target sites and the gRNA spacers are depicted at the left, while the top and bottom bar graphs at the right display the editing rates for CURE‐C1 and CURE‐X, respectively; CURE‐C1 editing was analyzed by Sanger sequencing and while CURE‐X by NGS.

We next explored editing motif preference of CURE‐C1. CURE‐C1 proved highly active when C was preceded by U (53% editing rates) but inactive if preceded by other bases (2–5%), as predicted from the known property of A3A (Fig 2D, middle panel in the top bar graph). In contrast to the strict requirement of U preceding C, editing was relatively robust to the variations in the bases flanking UC, achieving about 20–63% of editing rates in different contexts (Fig 2D, top, position +1 and −2). Interestingly, in the auCC motif, the second C became editable once the first C was converted to U (Appendix Fig S6C). Thus, C does not have to be preceded by U for editing by CURE‐C1; CC is also editable if the dinucleotide is preceded by U. The ability to edit CC is a unique and useful feature of CURE, as it enlarges its editing scope (Appendix Fig S6D), which is particularly attractive considering the inability of RESCUE‐S to edit CC (Appendix Fig S6C). Of note, in contrast to CURE‐C1, free A3A is known to be inactive if UC is followed by U or C, or preceded by G (Sharma et al, 2017), again suggesting relaxed target site requirement for CURE‐C1 due to its forced tethering by dCas13b.

Finally, we considered a situation where two UC are located within the same loop. We found both editable by CURE‐C1, which would be undesirable if one of the Cs is a bystander (Fig 2E, gRNAs with a 14‐nt loop). However, this scenario could be readily avoided by excluding the bystander from the loop via gRNA adjustment (Fig 2E, gRNAs with 6‐nt and 10‐nt loops). On the other hand, the potential of CURE‐C1 to edit multiple UCs can be valuable for certain applications, such as stop codon induction within the editing window via C>U editing at CAA, CAG, and CGA(Billon et al, 2017); indeed, a C>U DNA base editor (BE‐Plus) has been developed to achieve analogous multiplex editing to ensure stop codon induction at the DNA level (Jiang et al, 2018).

We conclude that CURE‐C1 could be flexibly programmed to edit user‐defined UCC in a variety of contexts, and the editing was rather robust to variation in the gRNA configuration, with reasonable efficiencies achievable using gRNAs carrying 24‐ to 56‐nt spacers capable of inducing 6‐ to 20‐nt loops.

Remarkably, a similar trend was seen for CURE‐X, although, in general, CURE‐X was less active than CURE‐C1 and more sensitive to gRNA designs (Fig 2A–E, right, bottom bar graph in each panel; Appendix Fig S6C, which also shows that CURE‐N could edit UCC as did CURE‐C1).

The CURE system has complementary strengths and weaknesses to RESCUE‐S at endogenous transcripts

We next compared CUREs with RESCUE‐S in terms of on‐target editing efficiency and global off‐target effects on the transcriptome. RESCUE, the precursor to RESCUE‐S, was excluded in the comparison given the massive off‐target effects that prohibit its practical uses. For CURE, we tested CURE‐C2, CURE‐N, and CURE‐X; CURE‐C2 instead of its precursor CURE‐C1 was used here (even though the two performed indistinguishably; Appendix Fig S2A), as CURE‐C2, but not CURE‐C1, was exclusively localized to the cytoplasm just like RESCUE‐S, thus making the comparison more rigorous.

To avoid biases, we compared the editors not only at our standard four transcripts (GFP reporter, GAPDH, ACTB, TYMS) but also at the representative target sites of RESCUE(S), located at the PPIB, CTNNB1, SMARCA4, and KRAS (Appendix Fig S1B; Abudayyeh et al, 2019). For RESCUE‐S, the information on the optimized gRNAs is available for all the four targets except KRAS, and so we only optimized the gRNAs for editing KRAS and for our standard targets (Appendix Fig S7A). We also optimized the gRNAs for CURE‐X (Appendix Fig S7B). In contrast, for CURE‐C2 and CURE‐N, we simply used a generic gRNA format (32‐14) for convenience. Compared with RESCUE‐S, all three CUREs proved substantially more active at GFP reporter and PPIB (editing rates ~30–50% vs. ~10%; Fig 3A, first two target sites). At the remaining six sites, CURE‐C2 editing was moderately more active (TYMS, ACTB, and CTNNB1) or as active (KRAS, SMARCA4, and GAPDH), CURE‐N was as active except at KRAS and SMARCA4 where its editing rate was slightly lower, and CURE‐X much weaker at all the six sites except CTNNB1 where it was as active as RESCUE‐S (Fig 3A, remaining target sites). Thus, CURE outperformed RESCUE‐S at a subset of the mRNA targets, but could be weaker at other sites. Note that the efficiencies of CURE‐C2 and CURE‐N were likely underestimated, given their use of unoptimized gRNAs.

Figure 3. Benchmarking CURE against RESCUE‐S.

1. mRNA editing as measured by NGS at eight sites, arranged in the order of CURE‐C2 editing efficiency relative to RESCUE‐S. The representative transcripts for RESCUE‐S, copied from Zhang and colleagues (Abudayyeh et al, 2019), are highlighted in yellow. The three CUREs were independently tested, each side by side with RESCUE‐S. The gRNAs shown are all optimized except those for CURE‐C2 and CURE‐N. For RESCUE‐S, the gRNAs are 30/20, 30/22, or 30/24 as indicated, denoting that the gRNAs carry 30‐nt spacers with the target C specified by a mismatched U placed at 20, 22, or 24 nt upstream of the 3’ end of the spacers, respectively. The editors and their indicated gRNAs were co‐expressed and the editing analyzed as described in Fig 1H. The bar graph displays mean ± SEM (n = 3). The blue numbers above the bars are the ratios of the editing rates of CURE relative to that of RESCUE‐S.
2. Nuclear LncRNA editing as measured by NGS. CURE‐N was compared with a nuclear‐localized version of RESCUE‐S (RESCUE‐S-N) at MALAT1 and XIST in the presence of various gRNAs. The bar graph displays mean ± SEM (n = 2 and 3 for MALAT1 and XIST, respectively). NT, non‐targeting.
3. Off‐targets at the transcriptome. Various editors were co‐expressed in HEK293T cells with the GFP reporter transcript and the corresponding gRNAs (28‐14 and 30/22 for CURE and RESCUE‐S, respectively) before on‐target editing at the reporter and off‐target editing at the transcriptome quantified by NGS. The jitter plot displays total edits, with the blue dots indicating the on‐target editing rates. The Venn diagrams show the numbers of the transcripts impacted in a functionally significant way, as predicted using Ensemble Variant Effect Predictor (VEP; McLaren et al, 2016). A3A, free Apobec 3A protein. See Appendix Fig S8 for additional analysis of the edits.

Nuclear LncRNAs such as MALAT1 and XIST have emerged as crucial regulators of cellular processes. As they are located exclusively in the nucleus, we compared the editing rates of CURE‐N with a nuclear‐localized version of RESCUE‐S (RESCUE‐S‐N) that we created by flanking RESCUE‐S with NLS, which sufficed to drive RESCUE‐S into the nucleus (Appendix Fig S7C). We found CURE‐N able to edit both MALAT1 and XIST with reasonable efficiencies (40 and 28%) whereas RESCUE‐S‐N much less active (8 and 18%, respectively; Fig 3B). RESCUE‐S‐N was also much less active than RESCUE‐S at editing mRNAs (Appendix Fig S7D). Thus, CURE‐N but not RESCUE‐S‐N can effectively edit the tested nuclear and cytoplasmic RNAs.

Finally, we used RNA‐seq to profile global off‐targets. To this end, plasmids expressing A3A or various editors were co‐transfected into HEK293T cells with the GFP reporter plasmid and the corresponding gRNA expression vector, and cells analyzed 48 h later; GFP reporter and the gRNA served as positive control for editor activity, namely to exclude the possibility that a low off‐target effects seen for an editor may result from non‐specific loss of editor function. We found that A3A had 3,269 C>U edits and CURE‐C2 2x more (7,416) perhaps due to dCas13b promiscuity, whereas CURE‐N (1301) and CURE‐X (380) 6× and 20× fewer than CURE‐C2, respectively (Fig 3C, jitter plot at the top). The reductions in CURE‐N and CURE‐X off‐target edits were not due to inactivation of these proteins because their on‐target editing rates at the GFP reporter were comparable to that of CURE‐C2. Importantly, A>I edits were undetectable for free A3A or CUREs, as expected from the property of A3A. In contrast, RESCUE‐S created not only C>U but slightly more A>I edits (611 vs. 860), the total edits (1471) 5x fewer than CURE‐C2 (7416), similar to CURE‐N (1301) and 4x more than CURE‐X (380; Fig 3C, jitter plot at the top). Thus, in terms of total edits, CURE‐X was more specific than RESCUE‐S and CURE‐N as specific. Interestingly, compared with CURE‐N, the edits created by RESCUE‐S potentially impacted the function of more mRNAs than CURE‐N (608 vs. 369; Fig 3C, Venn diagram at the bottom), which is associated with the fact that RESCUE‐S created missense mutations at mRNA 3.4× more frequently and affected 2.7× more codons than CURE‐N (Appendix Fig S8A). Thus, although CURE‐N created similar numbers of global off‐target edits as RESCUE‐S, it may functionally impact fewer transcripts. We also found that the editors altered the expression of a small numbers of genes (78, 63, 55, and 39 by CURE‐C2, CURE‐N, RESCUE‐S, and CURE‐X, respectively), but these differentially expressed genes (DEGs) were either unique to each editor or shared only by variable subsets of the editors (Appendix Fig S8B) and were not preferentially edited at the RNA level (Appendix Fig S8C). The mechanisms whereby the editors affected the gene expression are unclear, as were their potential biological consequences, except that some of the genes upregulated are heat shock proteins, suggesting the induction occurred in response to cellular stress.

DNA cytidine deamination by CUREs is undetectable

APOBEC enzymes carried in the cytosine DNA base editors (CBEs) are known to have the propensity to create stochastic Cas9‐independent off‐target edits on the genome, with A3A being the most active(Zuo et al, 2019, 2020; Doman et al, 2020), raising the concern that CUREs, particularly CURE‐N and CURE‐X located in the nucleus, might produce C>T mutations on the genome. The editing rates at the off‐targets of CBEs are typically well below the ~0.1% detection limit of practical high‐throughput DNA sequencing experiments, making their detection a challenge(Doman et al, 2020). Exploiting the fact that the deaminases used in CBEs can only act on ssDNA, two groups have recently developed a fast, sensitive, and cost‐effective method (“orthogonal R‐loop assay”) to detect such rare events(Doman et al, 2020; Yu et al, 2020). In this assay, dSaCas9 together with a gRNA is used to induce a stable, ssDNA region (orthogonal R‐loop) at a specific locus, thus artificially magnifying Cas9‐independent deamination by CBEs (Fig 4A). Using this assay, Doman et al measured the off‐target edits of diverse CBEs at six genomic sites, finding that Site 5 and 6 generally offer the most sensitive readout for these editors including A3A‐BE4max. We thus used the two sites to detect stochastic DNA off‐target editing by CUREs. Site 5 carries nine Cs, only three of them (C8, C10, and C11) are known to be susceptible to deamination, with C10–11 more efficiently deaminated than C8 and A3A‐BE4max more active than BE4max (Doman et al, 2020). We confirmed this trend, although C8 was somehow more resistant to editing in our hands: we found A3A‐BE4max edited C8, C10, and C11 with 9, 20, and 19% efficiencies respectively (as opposed to ~30, ~37, and ~37% previously reported), whereas BE4max edited the three Cs with 8, 18, and 10% efficiencies (as opposed to ~15, ~25, and ~15% previously reported; Fig 4B, white bars in the left panel, compared with Fig 2C in Doman et al). We also confirmed that on‐target editing efficiencies at the EMX‐1 site were comparable for A3A‐BE4max vs. BE4max (Fig 4B, white bars in the right panel, compared with Fig S5 in Doman et al). Importantly, none of the CUREs (CURE‐C2, CURE‐N, or CURE‐X) created C>T edits above the basal level at Site 5 (Fig 4B, left panel, brown vs. gray bars; the basal level was 1–3%, perhaps due to the action of endogenous Apobec proteins). This is not surprising for CURE‐C2 given its exclusive cytoplasmic localization, or for CURE‐X given its minimal off‐target effects on RNA. But why was CURE‐N free of detectable off‐target editing, given its nuclear localization and stronger off‐target effects on RNA? We speculate that CURE‐N might in fact be able to deaminate the Cs, but the mutations may be subsequently corrected via base excision repair (BER) initiated by Uracil DNA glycosylase (UDG) which removes U from DNA. Indeed, all CBEs carry UDG inhibitor (UGI), and the first‐generation CBE (BE1) lacking UGI is virtually inactive in vivo (Komor et al, 2016). We therefore fused 2x UGI to CURE‐N to imitate A3A‐BE4max, which indeed unmasked the C>T edits created by CURE‐N, but for unknown reasons, the editing activity of CURE‐N‐UGI was substantially weaker than A3A‐BE4max (3–10 vs. 9–20%; Fig 4B, left panel). This off‐target DNA editing revealed by UGI was not an artifact resulting from generally increased editing activity conferred by UGI, as the on‐target editing (at the GAPDH mRNA) was not increased but actually reduced by UGI (from 33% in CURE‐N to 25% in CURE‐N‐UGI; Fig 4B, right panel, last two bars). The performance of the editors at Site 6 was comparable to that at Site 5, as expected (Fig 4B, middle panel).

Figure 4. Off‐target DNA editing in human cells and mouse embryos.

• A
  Principle of orthogonal R‐loop assay. dSaCas9 recruited by a gRNA generates an R‐loop at the specified genomic site (Site 5–6 in the current study; top). A subset of the Cs within the R‐loop is susceptible to stochastic deamination by A3A carried in either A3A‐BE4max (bottom left) or CUREs (bottom right). In this assay, on‐target editing (at the EMX1 locus and GAPDH mRNA for A3A‐BE4max and CURE, respectively) is also induced to serve as positive controls for the off‐target editing.
• B
  Off‐target editing in HEK293 T cells as evaluated by the R‐loop assay. dSaCas9 and its gRNA were co‐expressed in HEK293T cells with indicated editors and their respective gRNAs, or without them (NC, negative control), and editing analyzed 3 days later at off‐targets (left and middle panel) and on‐targets (right panel). On‐target editing at the EMX1 locus and GAPDH mRNA were quantified as published (Doman et al, 2020) and in Fig 3A, respectively, using the samples from the same experiment as in the left panel. The bar graphs for off‐target editing display mean ± SEM (n = 2‐3 biological replicates), while for on‐target editing, a single biological replicate was used for all the editors except CURE‐N-UGI for which three biological replicates were analyzed. The blue values above the bars are the mean editing rates.
• C–F
  Off‐target editing in mouse embryos as evaluated by GOTI. (C) SNVs (Single Nucleotide Variants) in CURE‐N-treated detected using the indicated software tools. 2–8 SNVs were detected in each of the three embryos. (D) SNVs in CURE‐N-treated embryos were dramatically fewer than induced by BE3 (˜300) and similar to that seen in Cre‐treated embryos which represents the baseline. The data are displayed as boxplots, where the central, bottom, and top lines are the median, the first quartile and the third quartile, respectively, while the whiskers extend to the minimum and maximum values. P values were calculated by two‐sided Wilcoxon rank sum tests. Sample sizes are 2, 6, and 3 biological replicates for the Cre, BE3, and CURE groups, respectively. The Cre and BE3 dataset are copied from (Zuo et al, 2019); the same dataset is also used as the control in our recent paper (Zuo et al, 2020, Fig 2A). (E) Mutation profiles of the SNVs. The number indicates the proportion of a certain mutation among all mutations. CURE‐N exhibited significantly reduced mutation bias compared with BE3. CURE‐N-treated embryos have more frequent G>A and C>T SNVs than Cre‐treated embryos, which could reflect residual editing by CURE‐N or mere random variation due to insufficient numbers of SNVs (17 total) used for quantitation. (F) Editing at GFP (β1–10) reporter mRNA analyzed 24 h after injection into the 2‐cell embryos. The mRNA was injected alone (NC) or with CURE‐N mRNA and sgRNA (CURE‐N). Values are mean ± SEM (n = 2 biological replicates).

Another highly sensitive method for DNA off‐target detection is GOTI (Genome‐wide Off‐target analysis by Two‐cell embryo Injection; Zuo et al, 2019, 2020). In this method, DNA base editor mRNA is injected into one blastomere of 2‐cell mouse embryos, and two weeks later, the cells derived from the injected blastomeres are isolated from the embryos and their off‐target edits quantified by whole genome sequencing; the cells derived from non‐injected blastomeres in the same embryos serve as the internal control. We have used GOTI to evaluate the DNA off‐target effects for CURE‐N; among the three CUREs, CURE‐N is most likely to display off‐target effects, as CURE‐C is located in the cytoplasm while CURE‐X less active and more specific than CURE‐N. We found that the off‐target edit level of CURE‐N comparable/close to the baseline (Fig 4C–E), which is not an artifact because CURE‐N was able to edit the GFP (β1–10) reporter mRNA co‐injected into the mouse embryo (Fig 4F).

We conclude that CURE‐N and presumably CURE‐X were able to weakly deaminate C, but nevertheless unable to establish C>T mutations thanks to BER.

Discussion

We have harnessed A3A to develop three CUREs, with CURE‐C being the most active version, CURE‐X the most specific, and CURE‐N the balanced. For on‐target editing, CURE offers several advantages over RESCUE‐S. Specifically, compared with RESCUE‐S, CURE was more active at some UC targets and also capable of UCC editing. Regarding off‐target editing, CURE‐N and especially CURE‐X were more specific than RESCUE‐S, affecting fewer transcripts in a functionally significant manner. The major weakness of CURE relative to RESCUE‐S is that it could not edit AC. However, the situation is remedied by the unique ability of CURE to edit UCC. As a consequence, in terms of C>U editing, the diversity and number of the codon substitutions that can be engineered using CURE are comparable to that using RESCUE‐S (Appendix Fig S9). We propose that CURE is complementary to RESCUE‐S in utility. For example, CURE‐X and CURE‐N may be the top choices for applications demanding the highest specificity and for editing nuclear RNA, respectively, and CURE the only choice for editing UCC. On the other hand, RESCUE‐S may be preferable where CURE is inefficient and for multiplex C>U and A>I editing, and is the only choice for editing AC. It should be emphasized that RESCUE‐S, just like CURE, is largely inapplicable to GC at the endogenous RNA (Appendix Fig S1), and future studies are needed to broaden the editing scope for both platforms.

Of note, the inability of the CURE or RESCUE‐S to edit GC is not a major limiting factor for correcting disease‐causing mutations, an important application of RNA editors: at the present stage, the ClinVar database has documented a total of 4,315 T>C SNVs that are (likely) pathological, among which only 23% of the Cs are preceded by G and thus refractory to editing. In contrast, 24 and ~10% of the Cs are preceded by A and TC and therefore potentially correctable by RESCUE‐S and CURE, respectively, while 13% of the Cs are preceded by T and hence potentially editable by both RESCUE‐S and CURE. Therefore, as much as 47% of the SNVs are theoretically correctable by CURE/RESCUE‐S. However, in reality, for a substantial fraction of these SNVS, efficient editing may be difficult to achieve for both CURE and RESCUE‐S, which would seriously limit the clinical applications of the C>U editors. This is especially true for CURE‐X, as it appears to require more extensive optimization of spacer and loop length for each target. Thus, an important objective of future research is to understand the rules governing the target specificity of editing efficiency. Presumably, diverse parameters can impact the editing efficiency, including the secondary structure, binding proteins, stability, abundance, and subcellular localization of RNA. Big data generated using high‐throughput assays may be essential for dissecting these variables.

We found CURE‐N able to deaminate DNA, but fortunately, this did not lead to mutations thanks to BER. If necessary, one might attempt to engineer CURE‐N variants deficient in DNA deamination; creating RNA editors lacking off‐target DNA editing may be less challenging than creating the DNA editors (such as YE1) lacking off‐target DNA editing (Doman et al, 2020; Zuo et al, 2020), especially if the structure of A3A in complex with RNA could be solved in the future.

Materials and Methods

Constructs

The editor and reporter constructs were made using standard molecular biology techniques, and key plasmids will be deposited at Addgene. Key construct sequences are provided in the Appendix.

Cell culture and transfection

HEK293T cells (from ATCC) were cultured at 37°C with 5% CO₂ in DMEM containing high glucose, sodium pyruvate, penicillin–streptomycin, and 10% fetal bovine serum. Cells were passaged three times per week and tested to exclude mycoplasma contamination. Transfections were performed with Lipofectamine 3000 in 48‐well or 24‐well plates per manufacturer's instruction. Briefly, cells were plated into 48‐well plates at 5 × 10⁴ /well and transfected the next day. DNA was mixed with 1 μl Lipofectamine P3000 (Thermo Fisher Scientific, L3000015) into 25 μl Opti‐MEM (Invitrogen) and incubated for 5 min at the room temperature. 0.75 μl of the Lipofectamine 3000 (Thermo Fisher Scientific, L3000015) was diluted into 25 μl Opti‐MEM (Invitrogen), combined with the DNA: P3000 mixture, and incubated for another 20 min at the room temperature. The DNA: P3000: Lipofectamine 3000 mixture was then added dropwise into the wells. Cells were analyzed 36–48 h post‐transfection.

Editing at the reporter plasmids

To assess CURE activities at the GFP reporters depicted in Fig 1B, the pair of the split GFP reporter plasmids (25 ng each) were co‐transfected into 48‐well plates with vectors expressing CURE‐C1‐P2A‐mCherry (300 ng), gRNA (200 ng) or, where applicable, non‐targeting (NT) gRNA bearing a random spacer sequence. Two days later, cells were analyzed by FACS, Sanger sequencing, and/or NGS. FACS was performed using BD LSR Fortessa (BD Biosciences), and the data analyzed using FlowJo, with the ratio of GFP/mCherry mean fluorescence intensities taken as a measure of editing. For Sanger sequencing, the cells were lysed and the mRNA in the lysate reverse transcribed as described (Joung et al, 2017). The amplicons were then sequenced, and editing efficiencies determined using EditR, an “accurate, fast, and low‐cost method for the identification and quantification of base editing from fluorescent Sanger sequencing data” (Kluesner et al, 2018, 2019). For NGS, the cDNA was subjected to two rounds of PCR to add Illumina adaptors and sample barcodes using NEBNext High‐Fidelity 2X PCR Master Mix (New England Biolabs). The library was then sequenced on an Illumina HiSeq. The reads were demultiplexed with Cutadapt before the C>U editing rates quantified using CRISPResso2. To analyze editing by CURE‐C1 and CURE‐X at the reporters described in Fig 2B–E, vectors expressing the editors were co‐transfected with the reporter plasmids and the cells analyzed by NGS (and occasionally also by Sanger sequencing) as described above. The target site and primers are provided in the Appendix.

Editing at endogenous transcripts

To evaluate editor activities at the endogenous transcripts, vectors expressing gRNA (400 ng) and editor‐P2A‐mCherry (600 ng) were co‐transfected into HEK293T cells grown in 24‐well plates. Two days later, mCherry⁺ cells (50–100 K) were isolated by electronic sorting using BD FACSAriaIII. The target regions were then amplified and analyzed by NGS as described above. The edited sites and the PCR primers are provided in the Appendix.

Global off‐targets at the transcriptome

To detect off‐target RNA editing sites across the transcriptome, we transfected HEK293T cells in 12‐well plates with the GFP reporter (100 ng) together with the vectors expressing the gRNA targeting the reporter (800 ng) and editor‐P2A‐mCherry (1.2 μg). mCherry⁺ cells were sorted 48 h later and their total RNA isolated using the TRIzol reagent (Invitrogen). The mRNA fraction was then enriched using a NEBNext Poly(A) mRNA Magnetic Isolation Module (NEB) before library construction using NEBNext Ultra RNA Library Prep Kit for Illumina (NEB). The libraries were sequenced on an Illumina HiseqXten‐PE150, at a depth of around 20 million reads per sample. The reads were mapped to the human reference genome (hg38) by STAR software (Version 2.7.1), using the 2‐pass mode. Duplications were removed using Picard (2.20.5), and SNV identified by GATK HaplotypeCaller (version 4.1.3), with QD (Quality by Depth) > 2, sequencing coverage > 20× and reads in the wild‐type samples supporting the reference allele > 99%. Edits shown in Fig 3C are intersection sets of two replicates.

Orthogonal R‐loop assay

HEK293T cells (~120K) were seeded in a 24‐well plate and co‐transfected at 70% confluence with plasmids expressing dSaCas9 and its gRNA, in combination either with plasmids expressing (A3A)‐BE4max and their gRNA targeting EMX1, or with plasmids expressing CUREs and their gRNA targeting GAPDH; each editor and gRNA plasmids added were 600 and 400 ng, respectively. Cells were cultured for 3 days before GFP⁺ cells were isolated by cell sorting (GFP was co‐expressed with dSaCas9 sgRNA). To analyze (A3A)‐BE4max editing, genomic DNA was extracted and the target regions at EMX1 and Site5 were amplified by PCR before NGS. To analyze editing by CUREs, half of sorted cells were used for DNA purification and the other for RNA purification, which were subsequently checked for editing at Site 5 and GAPDH mRNA by NGS, respectively.

GOTI assays

GOTI tests were carried out as described (Zuo et al, 2019). Briefly, Ai9 males (carrying the “CAG‐lox‐stop‐lox‐tdTomato” transgene) were mated with wild‐type females for embryo collection. The mixture of Cre and CURE‐N mRNAs (100 ng/μl) was injected into one blastomere of a 2‐cell embryo, which was then transferred into pseudopregnant ICR females at 0.5 dpc. The tdTomato+ and tdTomato‐ cells were isolated from the chimeric embryos at E14.5 by FACS and analyzed by WGS at 50× coverage. Sequencing reads were mapped and qualified used BWA (v0.7.12), and duplications removed by Picard tools (v2.3.0). SNVs were identified in tdTomato⁺ cells with tdTomato‐ cells as the control. Three algorithms, Mutect2 (v3.5), Lofreq (v2.1.2), and Strelka (v2.7.1), were run, and the shared SNVs were considered genuine variants. The editing efficiency at an on‐target was evaluated in the same way except that the mixture of the CURE‐N mRNA, GFP (β1–10) reporter mRNA and the targeting sgRNA (all at 100 ng/μl) was injected into both blastomeres of the 2‐cell embryos. Total RNA was isolated from the ~200 embryos 24 h later, and the editing rate at the GFP(β1–10) reporter quantified by NGS as described above for HEK293T cells.

Primers for PCR and Sanger sequencing

These are provided in the Appendix. The PCR products were also used for NGS.

Author contributions

TC conceived of the project and wrote the manuscript. Xinxin H, JL, SM, YL, ZJ, XZ, SS, MD, XW, and JG characterized the editors in HEK293 T cells under the supervision of TC and Xingxu H, with the bioinformatics analysis performed exclusively by ZJ. The GOTI assay was performed by ZL (microinjection) and YS (Bioinformatics) under the guidance of EZ.

Conflict of interest

The authors declare that they have no conflict of interest.

Supporting information

Appendix

Review Process File

The EMBO Journal (2020) 39: e104741

Data availability

The RNA‐seq data are available from: https://www.ncbi.nlm.nih.gov/bioproject/PRJNA635732.