REPAIRx, a specific yet highly efficient programmable A > I RNA base editor

Abstract

Programmable A > I RNA editing is a valuable tool for basic research and medicine. A variety of editors have been created, but a genetically encoded editor that is both precise and efficient has not been described to date. The trade‐off between precision and efficiency is exemplified in the state of the art editor REPAIR, which comprises the ADAR2 deaminase domain fused to dCas13b. REPAIR is highly efficient, but also causes significant off‐target effects. Mutations that weaken the deaminase domain can minimize the undesirable effects, but this comes at the expense of on‐target editing efficiency. We have now overcome this dilemma by using a multipronged approach: We have chosen an alternative Cas protein (CasRx), inserted the deaminase domain into the middle of CasRx, and redirected the editor to the nucleus. The new editor created, dubbed REPAIRx, is precise yet highly efficient, outperforming various previous versions on both mRNA and nuclear RNA targets. Thus, REPAIRx markedly expands the RNA editing toolkit and illustrates a novel strategy for base editor optimization.

A multipronged approach allowed engineering of an optimized A > I RNA base editor facilitating specific and efficient base editing.

Introduction

Site‐directed RNA editing (SDRE) enables the recoding of genetic information at the RNA level and can conceivably complement DNA editing in multiple ways (Mao et al, 2019; Montiel‐Gonzalez et al, 2019; Rees & Liu, 2018; Vogel & Stafforst, 2019). First, being reversible, SDRE is applicable to both genetic and non‐genetic diseases just as conventional medications, whereas the use of DNA editing would be limited to genetic diseases for ethical and safety reasons. Second, SDRE is tunable and thus potentially applicable when, for example, DNA editing (an all‐or‐none event) proves lethal to the cells.

Site‐directed RNA editing has been achieved for adenine via recruitment of adenosine deaminase acting on RNA (ADAR), which converts adenosines to inosines (functional equivalent of guanines) within duplex RNAs. The human ADAR family has 2 members (ADAR1 and ADAR2) that are functional, each carrying a deaminase domain (ADAR_DD) in addition to several dsRNA binding domains. Multiple ADAR‐based SDRE platforms have been designed, but each with significant limitations (Rees & Liu, 2018; Vogel & Stafforst, 2019). These platforms fall into two major categories, each relying on an antisense RNA oligo to specify the target. The first platform, developed by Stafforst and colleagues, requires chemically modified oligos and is efficient and extremely specific (Stafforst & Schneider, 2012; Vogel et al, 2014, 2018; Merkle et al, 2019; Vogel & Stafforst, 2019). The latest version of system comprises a 15‐nt antisense fused to an imperfect hairpin adapted from an ADAR2 target site in the GluR2 mRNA, which can recruit endogenous ADARs to user‐defined sites specified by the anti‐sense RNA; the chemical modifications (2′ ‐O‐methylations, phosphorothioate), introduced into both the antisense RNA and the ADAR recruiting domain, stabilize the RNA and inhibit bystander editing, which is essential for the efficient and specific editing (Merkle et al, 2019). Unfortunately, the RNA must be synthesized in vitro, which hampers its application. In the second category, the RNAs are genetically encoded, thus bypassing the above limitation. However, despite much effort, genetically encoded editors that are both specific and efficient have remained elusive, as outlined below.

There are two types of genetically encodable editors. The first uses very long antisense oligos to recruit endogenous ADARs to target sites(Katrekar et al, 2019; Qu et al, 2019). This method lacks global off‐target effects (due to the specificity of the antisense oligos and lack of exogenous editors), but produces frequent local off‐target editing at the bystander adenosines flanking the on‐target adenosine, and furthermore, the on‐target editing efficiency is highly variable in different cell types partly due to variable ADAR expression (Qu et al, 2019). The second type of strategy, pioneered by Rosenthal and colleagues, uses exogenous editors to liberate the editing from dependence on endogenous ADARs; an exogenous editor typically consists of ADAR_DD fused to a programmable RNA binding moiety which recognizes its cognate RNA partner fused to the antisense oligonucleotide in the gRNA (Montiel‐Gonzalez et al, 2013; Montiel‐González et al, 2016; Cox et al, 2017; Fukuda et al, 2017; Heep et al, 2017; Sinnamon et al, 2017; Wettengel et al, 2017; Vallecillo‐Viejo et al, 2018; Katrekar et al, 2019; Rauch et al, 2019). A frequent problem with this method is the strong off‐target effects (both global and local) induced by overexpressed exogenous editors. Two countermeasures of the off‐target effects have been described, each with its limitations. The first, developed by Rosenthal and colleagues, is to relocate the editor to the nucleus, which reduces the off‐target effects without impacting on‐target editing for an editor (Vallecillo‐Viejo et al, 2018). However, the effect is relatively mild (2×), and furthermore, this method is not generally applicable: Nuclear localization does not reduce off‐target effects but instead compromises on‐target editing for another editor [MCP‐ADAR2_DD (E488Q)] (Katrekar et al, 2019). The second countermeasure is to weaken the deaminase domain with point mutations, as exemplified in REPAIR (Cox et al, 2017). Two versions of REPAIR were created. V1 comprises the hyperactive ADAR2_DD (E488Q) fused to the C‐terminus of dCas13b of the type VI CRISPR/Cas family, and V2 is identical to V1 except for a point mutation (T375G) that weakens the deaminase domain (Cox et al, 2017). V1 is highly efficient, but also produces massive off‐target edits both locally and globally. The point mutation (T375G) in V2 dramatically reduces the off‐target editing, but also jeopardizes its on‐target editing (Cox et al, 2017); indeed, when tested under stringent condition (expressed from a single copy transgene stably integrated into the genome, where limits the protein expression level), ADAR2_DD (E488Q/T375G), in the context of SNAP‐tagged protein, proves largely inactive (Vogel et al, 2018). Thus, additional methods are needed to produce an editor that is both specific and efficient.

Using an multipronged approach, we have succeeded in creating REPAIRx (Vx), which is specific as V2 but active as V1, outcompeting all its rival editors at both mRNA and nuclear RNA.

Results

Vx outperformed other CasRx‐ADAR2_dd (E488Q) fusion proteins in N2a cells

We sought to optimize V1 by replacing dCas13b with dCasRx (another member of the Cas13 family) (Konermann et al, 2018) for the following reasons. First, CasRx functions well in the nucleus in RNA knockdown assays (whereas Cas13b prefers the cytoplasm) (Cox et al, 2017; Konermann et al, 2018), suggesting that dCasRx‐based editors may remain active in the nucleus, but its off‐target effect might be reduced. Second, CasRx contains multiple flexible loops potentially tolerant of ADAR2_DD (E488Q) insertion, and such a fusion protein configuration might create steric hindrance, which might reduce the global off‐target editing but still allow on‐target editing thanks to forced tethering of the editor to the on‐targets. Finally, compared with Cas13b, CasRx is small (967 aa vs 1,090 aa) and shorter gRNA spacers (~ 22 nt vs ~ 40 nt) are sufficient to induce near‐maximal cleavage of specific target RNAs; the more compact sizes of the CasRx system would facilitate AAV packaging, and the short spacers are also desirable because of the concern that long spacers bound to the mRNA might potentially interfere with translation(Vogel et al, 2018).

ADAR2dd was fused to dCasRx in various ways to produce Vx and six other fusion proteins (Fig 1A), which were then compared at a reporter plasmid bearing a target sequence from the human PRKN mRNA inserted between BFP and GFP (Fig 1B). The PRKN mRNA carried a G > A point mutation that prevented GFP expression until A was edited back to G; this mutant mRNA was selected because it is known to be editable by REPAIR (Cox et al, 2017). Thus, the reporter constitutively expressed BFP, but additionally expressed GFP following successful editing at the target A. The change in GFP fluorescence was detectable by FACS, and the ratio of GFP fluorescence intensity over that of BFP taken as a measure of editing efficiency.

Figure 1. Development of REPAIRvx (Vx).

1. ADAR2_DD (E488Q)‐dCasRx fusion proteins tested. CasRx possesses several loops at the external protein surface that can be removed without affecting protein function. Vx consists of hADAR2_DD (E488Q) inserted into a dCasRx loop (Δ3, aa T558 to G587) between T558/N559 (aa L560‐G586 deleted during the construction). Vx outperformed other fusions (Vx‐1 to Vx‐6). ADAR, hADAR2_DD (E488Q). NLS, nuclear localization sequence; NES, nuclear export sequence; XTEN, a linker (see Plasmid Construction in SI).
2. The reporter and gRNA. A 224‐bp PRKN cDNA fragment (black line) carrying a premature stop codon (UGG → UAG, W55X) (Cox et al, 2017) was inserted between BFP and GFP in the reporter. The gRNA carried a mismatched cytidine (cyanine) that specified the target adenosine (red; depicted is a representative “30‐17” gRNA whose spacer length and mismatch distance were 30 and 17 nt, respectively). The full sequence of the 224‐bp mutant PRKN cDNA fragment cloned into the reporter is shown at the bottom. It contains 54 adenosines, only some clearly editable (capitalized), with A45 being the on‐target while others bystanders. DR, direct repeat.
3. FACS analysis of editing by the fusion proteins. The gRNAs used are shown at the top, representative FACS plots in the middle and relative editing efficiencies at the bottom. The reporter plasmid (20 ng) was co‐transfected into N2a cells with the plasmids expressing the editors (150 ng) and the gRNA (300 ng), and cells analyzed 36 h later by FACS; select samples were subjected to NGS in parallel for validation. (Appendix Fig S1A). The GFP MFI under various conditions (normalized to that of BFP) provided an indirect measure of editing and was plotted relative to that induced by Vx in the presence of the 30–17 gRNA, the latter set as 1. Green and blue numbers within the FACS plots, MFI of GFP and BFP, respectively, within the GFP⁺ population (square). The bar graph displays mean ± SEM from triplicate transfections. X, Vx; 1–6, Vx‐1 to Vx‐6.
4. NGS analysis of editing by various editors in the presence of 30‐17 gRNA. The experiment was done as in (C) except that the vectors expressing 30‐17 gRNAs carrying or lacking DR were used, and editing quantified by sequencing instead of FACS. Furthermore, puromycin was added to select for cells expressing gRNA, which enhanced the detected editing efficiencies due to killing of cells expressing the reporter transcripts without gRNA. Values are mean ± SEM (n = 3).
5. Systematic dissection of the roles of various components of the Vx editing system. N2a cells were transfected with various plasmids as indicated. Transfection was done as in (C). gRNA, intact gRNA with the 50‐17 spacer; the 50‐nt spacer was used here instead of the 30‐nt spacer to better reveal the spacer‐dependent but Vx‐independent editing. NT‐gRNA, Non‐Targeting gRNA of a random 50‐17 spacer sequence. Values are mean ± SD (n = 3).
6. RNA‐Seq analysis of global off‐target edits. N2a cells were transfected with the indicated expressing vectors before the transfected cells were sorted and analyzed 36 h later. The jitter plots display the off‐target edits, where the pink numbers and blue circles denote the off‐target numbers and on‐target editing rate at the Ppib site 1, respectively, and where the box spans the interquartile range (1^st to 3^rd quartiles), the band inside the box indicating the median (2^nd quartile), and the whiskers extend to the ± 1.5× interquartile range. gRNA, a plasmid expressing 30‐17 gRNA for Ppib‐1. Vx‐2/3 was not evaluated because they are not essential for dissecting the contributors to the high specificity of Vx.

As we are interested in applying SDRE to mouse models, we compared the seven fusion proteins in the mouse line N2a. Specifically, the reporter plasmid was co‐transfected into N2a cells with the plasmids expressing the editors and gRNAs, and the cells were analyzed by FACS 36 h thereafter (Fig 1C). In cells transfected with a positive control reporter plasmid where the mutant PRKN fragment had been replaced with the WT fragment, GFP was constitutively co‐expressed with BFP, the mean fluorescent intensity (MFI) of the two proteins being similar (Fig 1C, FACS plot 2). In contrast, in cells transfected with the reporter bearing the mutant PRKN fragment, GFP was hardly detectable, but strongly induced upon co‐expression of Vx and gRNA (Fig 1C, middle, plot 3–4; the gRNA spacer in this plot was 30‐nt in length, with a A:C mismatch at the 17^th nt, which will be termed 30‐17 hereafter). Overall, among the seven editors, Vx was the most active one at some gRNAs (30‐17, 50‐17 and 50‐36), and above or near the average at the remaining (30‐8, 30‐24, 50‐26; Fig 1C, bottom). Thus, Vx was highly robust to the variation in RNA designs. Of note, the 30‐24 gRNA, bearing the mismatch only 6 nt away from the 3′ end of the spacer, was the least active configuration for all the editors, consistent with previous observation that mismatches close to the ends of spacers perform poorly for V1 (Cox et al, 2017). These data suggest that Vx outperformed other variants in terms of on‐target editing.

Targeted next‐generation sequencing (NGS) of the PRKN reporter transcript confirmed the FACS results described above and additionally revealed that Vx was more specific than some of the variants, in that it edited the bystanders less efficiently (Appendix Fig S1A). To corroborate and extend these observations, we determined the roles of dCasRX in on‐target editing by including a control gRNA lacking the direct repeat (DR); this control was essential, because ADAR2_DD has an intrinsic ability to recognize and edit the RNA duplex (formed by the spacer bound to the target site) independently of dCasRx. This control was not included in the previous study on V1/2 (Cox et al, 2017), which complicates data interpretation (Vogel et al, 2018). Note that in this experiment, to better compare the bystander effects of Vx with the variants, we selected the 30‐17 spacer, because with this spacer, Vx was far more active at on‐target than other fusion proteins (Fig 1C and Appendix Fig S1A), thus ensuring that its weaker bystander effect was not an artifact resulting from nonspecific protein inactivation. Finally, since our focus on the effects of fusion strategies on editor behavior, we compared Vx only with the other nuclear‐localized fusion proteins (Vx‐1 to Vx‐5). Our results indicate that Vx achieved 78% editing at A45, which was dramatically reduced to only 6% following DR deletion; the residual editing presumably resulted from a weak ability of ADAR2_DD (E488Q) in Vx to edit the gRNA: target duplex (Fig 1D, top). Thus, Vx editing was critically dependent on dCasRx. Furthermore, Vx was far more active than other fusion proteins at A45 as expected from the data in Fig 1C. Remarkably, Vx was less active than Vx‐1/2/3 at the flanking off‐targets in the presence of DR except at A47, where Vx was among the most active (Fig 1D, middle). A similar trend was observed in the presence of 30‐24, 50‐26 and 50‐36 gRNAs (Appendix Fig S1A). On the other hand, in the presence of the 30‐17 gRNA lacking DR, all the editors but Vx‐2 were hardly active except at A40 (Fig 1D, bottom). The data in Fig 1C and D collectively demonstrate that compared with other fusion proteins, Vx was more robust to gRNA designs, being highly active at multiple gRNAs tested, whereas its off‐target editing was not proportionally increased; on the contrary, it was decreased as compared with some fusion proteins. We also found that Vx could efficiently edit the PRKN reporter in the presence of 22‐nt (but not 18‐nt) spacer, with the mismatches engineered at various positions all capable of supporting substantial editing except those near the spacer ends (Appendix Fig S1B). As expected, Vx could repair additional human mutant transcripts beside the mutant PRKN reporter transcript and displayed a codon preference reflecting the property of hADAR2_DD (Appendix Fig S1C and D).

gRNA consisted of a spacer and the DR, and Vx comprised dCasRx and ADAR2_dd (E488Q). All these four components were required for highly efficient editing, as expected: The 50‐17 gRNA lacking DR could only support low‐level editing by Vx as mentioned above, whereas a gRNA bearing a non‐targeting spacer (NT‐gRNA) was totally inactive, and dCas13Rx and ADAR2_dd (E488Q) separately were likewise hardly active even in the presence of the gRNA (Fig 1E). Curiously, coexpressing gRNA with dCasRx or ADAR2_DD (E488Q) led to lower editing than expressing only gRNA or gRNA no DR. The reason is unclear, but as the editing by gRNA/gRNA no DR alone is very low (< 10%), such weak editing may be susceptible to (random) influences by poorly understood or unpredictable variables.

Finally, we compared global off‐target edits for select editors. To this end, plasmids expressing the editors were transfected into the N2a cells and the transfected cells sorted for RNA‐seq 36 h later. A vector expressing gRNA (30‐17 format) targeting Ppib site1 was co‐transfected. This site was editable by Vx (see further) and thus served as a positive control for editing. The key results of the RNA‐seq experiments comparing Vx with other Vx variants are shown in Fig 1F, and additional data presented in Appendix Fig S2A and B. We found that expressing ADAR2_DD (E488Q), gRNA or Vx alone yielded 441, 123, and 315 global off‐target edits, respectively, without detectable effects on Ppib (Fig 1F, jitter plots); the fact that Vx off‐targets (315) were only slightly above the background level (123) demonstrates that Vx was highly specific. When gRNA and Vx were co‐expressed, the Ppib transcript was edited robustly (45%) whereas the global off‐targets were only mildly increased (524 edits; Fig 1F). On the other hand, Vx‐4 (Loop 1 insertion) was virtually inactive, whereas Vx‐5 (Loop 2 insertion) and Vx‐1 (C‐terminal fusion) each hit more off‐targets (919–967) than Vx (Fig 1F), suggesting Loop3 insertion reduced the off‐target effects up to ~ 2× (524 vs 967). Finally, Vx‐6, which was similar to Vx‐1 but located in the cytoplasm (as it contained NES instead of NLS; Fig 1A), was far more promiscuous than Vx‐1 (6x, 6,194 vs 967; Fig 1F), indicating the nuclear localization reduced the off‐target effects 6x.

We conclude that Vx was both robust and specific, outperforming other ADAR2_DD (E488Q)‐dCasRx fusion proteins, with its high specificity resulting in part from Loop 3 insertion (2× effect) and nuclear localization (6× effect).

Vx outperformed V1 and V2 in N2a cells

We next benchmarked Vx against V1/2 (Fig 2A; Appendix Fig S3A and B), with the gRNA configurations (spacer lengths and mismatch distances) systemically varied to ensure fair comparison.

Figure 2. Vx outperformed V1 and V2 in N2a cells.

1. Three versions of REPAIR. V1 we used consists of ADAR2_DD (E488Q) fused to the deletion mutant of PspCas13b lacking the C‐terminal 107 aa (Δ985–1,090); this deletion does not affect V1 function (Cox et al, 2017). V2 is identical to V1 except for the additional T375G mutation. Vx and V1 were expressed at similar levels and, as expected, localized to the nucleus and cytoplasm, respectively (Appendix Fig S3A and B). ADAR, ADAR2_DD (E488Q); ADAR*, ADAR2_DD (E488Q/T375G).
2. NGS analysis of editing at the GFP reporter in the presence of gRNAs varying in the spacer length (18–50 nt) but fixed in the mismatch distance (17 nt). Cells were transfected in triplicates and analyzed as in Fig 1D, with puromycin used for selection. Note that gRNA alone (with or without DR) were of very low activity (gray lines), which were slightly enhanced by ADAR2_DD (E488Q) (pink line), while the reporter plasmid transfected with the editors in the absence of gRNA was the least edited (< 1.4%; Appendix Fig S3C). The values are mean ± SEM from triplicate transfections. Asterisks indicate significant differences between Vx vs V1 (*P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, calculated using two‐tailed Student's t‐test).
3. NGS analysis of editing at endogenous transcripts. Cells were transfected with editors‐P2A-mCherry, and mCherry⁺ cells sorted 36 h later. The gRNAs used for Vx and V1/2 were 30‐17 and 50‐17, respectively, as depicted at the left, where the targeted adenosines are highlighted in red. The heatmaps display data from one of the triplicate transfections, with the editing rates at the bystanders also displayed if the values are ≥ 5%. The numbers at the bottoms of the heatmaps denote the positions of the adenosines at the target sequences. For each gRNA, the region of duplex RNA is outlined in red; the sizes of the red boxes will vary with the adenosine numbers in the duplex. The bar graph display mean ± SEM from triplicate transfections, with the asterisks indicate significant differences between V2 vs Vx (***P ≤ 0.001, two‐tailed Student's t‐test). Red arrowheads, on‐targets.
4. RNA‐Seq profiling of global off‐target edits. The data are from the same experiment as Fig 1F. The pink numbers and blue circles denote the off‐target numbers and on‐target editing rates at the Ppib site 1, respectively, as in Fig 1F. The box and whiskers are also as defined in Fig 1F.

We first tested the editors at the PRKN reporter, using gRNAs with a fixed mismatch distance (17 nt) but varying spacer lengths (18–50 nt). We found that while V1/2 required long (40–50 nt) spacers for the optimal activity at the on‐target (A45), a spacer as short as 25 nt sufficed to support near‐maximal editing by Vx (Fig 2B, left), as expected from the known properties of Cas13b and CasRx (Cox et al, 2017; Konermann et al, 2018). Importantly, Vx proved more active than V1–V2 at all spacer lengths tested, except the 18‐nt spacer where all editors were inactive (Fig 2B, left). Furthermore, consistent with Fig 1D, DR deletion markedly compromised Vx editing, which tended to be more obvious at shorter spacers but remained pronounced even at the longest: At 50 nt, DR deletion reduced the editing rate 3.6× (from 86 to 24%), indicating that the majority (~ two‐thirds) of the editing induced by the intact gRNA was dCasRx‐dependent (Fig 2B, left). DR deletion also compromised V1 and V2 editing, but to less extents. For example, at 50 nt, DR deletion reduced the V1 editing rate only 1.6× (from 70 and 43%), indicating that dCas13b contributed to only a minority (~ one‐third) of the editing (Fig 2B, left). Finally, we examined editing at the off‐targets flanking A45, finding V1 much more promiscuous than V2 as expected and importantly Vx as specific as V2 except at A47 and to some extent A40 (Fig 2B, right). We then varied the mismatch distances in the context of 30‐ and 50‐nt spacers, observing that although the mismatch distances could impact the editing, the same trend was maintained across the various gRNA designs: Vx was more efficient than V1, and almost as specific as V2 (Appendix Fig S3D).

To benchmark Vx against V1/2 at the endogenous transcripts in N2a cells, we targeted the mouse Ppib, Gusb, and Actb transcripts. Ideally, we should test a variety of gRNA for both Vx and V1/2 at each of the target transcripts to reveal the maximal editing potentials for each editor. For convenience, we simply compared their performance using generic non‐optimized formats based on their performance at the PRKN reporter. Specifically, we used 30‐17 for Vx and 50‐17 for V1/2, both supportive of (near) maximal editing at the PRKN reporter (Figs 1C and 2B); 50‐17 is also the standard configuration used by Cox et al (2017) for V1/2 editing at endogenous transcripts. We found that at the Ppib transcript, Vx achieved 36 and 46% editing rates at two sites tested, which was comparable to V1 (36 and 42%) but significantly higher than V2 (25 and 19%; bar graph, Fig 2C). Likewise, at Gusb, the efficiencies of Vx at two sites tested (29 and 33%) were comparable to V1 (31 and 37%), but again much higher than V2 (10%), and at Actb, Vx (58%) and V1 (69%) were again far more active than V2 (24%; Fig 2C). On the other hand, the bystander effects of Vx were as low as Vx (heatmap, Fig).

Finally, we compared global off‐target effects of the three REPAIRs in the presence of gRNA. To avoid any potential bias in gRNA design that might affect off‐target edits, we used the 50‐17 spacer format for all three editors. We found Vx hit 721 off‐target edits, which slightly exceeded V2 (239) but dramatically reduced as compared with V1 (54,526), whereas Vx was as active as V1 and far more active than V2 at the on‐target (Ppib site 1) as expected (Fig 2D). As expected, V1 edited far more transcripts than Vx or V2 (but all editors altered the expression of similar numbers of genes; Appendix Fig S2C). Importantly, V1 created 9× more off‐target edits than Vx‐6 (54,526 vs 6,197). Since the key difference between the two is that Vx‐6 used dCasRx while V1 used dCas13b, it can be inferred that the promiscuity of dCasRx was 9× lower than dCas13b, which therefore made the largest contribution to the high specificity of Vx relative to V1.

We conclude that in N2a cells, Vx was efficient as V1 and specific as V2, with the specificity rooted in part in the low promiscuity of dCas13Rx (9× effect).

Benchmarking Vx against V1/2 and other rivals in HEK293T cells

In addition to V1/2, three other editing platforms (based on MCP, exogenous full‐length hADAR and long antisense oligos, respectively) have recently been optimized and a fourth one (CIRTS) developed, totaling eight different versions of genetically encoded editors including V1/2 (Appendix Fig S4A). We sought to benchmark Vx against all these rivals. Editors developed in different labs have been optimized using different targets, although all in the human HEK293T cells. To ensure fairness, we did the comparison not only at the PRKN reporter we have used to optimize Vx, but also under the optimal conditions specified for these rivals in the original studies, namely at their representative targets and in the presence of their optimized gRNAs. Theoretically, we should also optimize gRNA for Vx at each of these sites, but for convenience and to make our conclusions conservative, we simply used a single generic gRNA configuration (50‐17) for Vx, and tested a few more only if 50‐17 proved insufficient for Vx to outcompete the rivals; this additional testing is justifiable given that the gRNAs for the rivals have been optimized. As will be shown later, only in rare cases was the additional testing necessary (Fig 3D, F and G); for the most part, the generic, unoptimized 50‐17 format sufficed to enable Vx to outcompete the rivals. In these experiments, 50‐17 was selected as the default spacer format instead of 30‐17 (as in Fig 2C), because the former seemed slightly more active at endogenous transcripts (for example, compare Fig 1F with Fig 2D).

Figure 3. Vx vs all other major editors in HEK293T cells.

• A–I
  Vx was compared with eight other editors at a total of nine sites (editor and site information detailed in Appendix Fig S4A and B). For comparison at the PRKM reporter (A), the reporter plasmid (20 ng) was co‐transfected into cells in 48‐well plates with plasmids expressing gRNA (300 ng) and editor (150 ng), or with the plasmids expressing the 151‐nt oligo (300 ng) and mCherry (150 ng). Forty‐eight hours later, cells were harvested and editing analyzed by NGS. For comparison at endogenous transcripts, the amounts of the transfected plasmids were scaled up 2× and transfection done in 24‐well plates, with the mCherry⁺cells sorted 48 h later for NGS analysis. Black asterisks in front of the editor names indicate that the gRNAs used for editing the relevant target sites have been optimized. The bystander editing rates which are ≥ 5% are denoted using black numbers in the heatmaps, but for the mutant long oligos in (E, H, and I), red numbers are used instead to denote the editing rates at the bystander As opposite the mismatched Gs; the remaining mismatched Gs, whose editing rates are below 5%, are indicated by red asterisks. Data are represented as mean ± SEM (n = 3). For statistical analysis, we compared Vx with each of the other editors at a target. For clarity, we group together the editors significantly (*P ≤ 0.05) different from Vx with a bracket, and only display the P value that is the largest within the group (***P ≤ 0.001; **, P ≤ 0.01, calculated using two‐tailed Student's t‐test). Red arrowheads, on‐targets.

A caveat of the experimental design above is that only one to a few well‐defined representative sites is known for the rival editors (Appendix Fig S4A). To address this problem, we tested the editors not only at their respective representative targets but also at those of others, and we also added MALAT1, a representative nuclear lncRNA (Sun et al, 2018). We did not optimize gRNA for Vx or any other editor when comparing their performance at their non‐representative sites, but simply used the same format shown to be optimal at the representative sites (for example, the 50‐17 format optimized for Vx at the representative PRKN reporter was also used at the non‐representative sites). Overall, nine sites were tested, only one of them (PRKN) a representative site of Vx, while all others (except MALAT1) representative sites of the rivals, thus avoiding any bias for Vx in target site selection. However, we did not test each editor indiscriminately at all the nine sites, but each at its own representative site together with an informative subset of non‐representative sites selected out of the nine sites on a case‐by‐case basis, which enabled meaningful yet cost‐effective editor characterization (Material and Methods). For the most part, we compared Vx with each rival at four–six sites (Appendix Fig S4A), finding Vx generally more active across all the sites tested, as detailed below.

We first compared Vx with V1/2 at our standard PRKN reporter, using the 50‐17 gRNA format for all three as optimized in Fig 2B. While V2 was much weaker than V1 in N2a cells (editing rate 27 vs 70%; Fig 2B), it unexpectedly proved far more active in HEK293T cells, achieving an editing rate only slightly below V1 (31 vs 43%; Fig 3A). Importantly, V2 remained markedly (~ 2×) weaker than Vx at the PRKN reporter. Likewise, at PPIB and KRAS, the only endogenous transcripts tested by Cox et al, Vx efficiency was comparable to V1 but again higher than V2, exceeding it by up to two‐fold (62 vs 29% at PPIB site 1; Fig 3B–C and E). Finally, MALAT1 site 1 could be efficiently edited by Vx (60%) but not V1 (10%) or V2 (1%), as expected from their respective subcellular localizations (Fig 3H). Thus, Vx markedly outperformed V2 at all the five sites tested and V1 at one of these sites (MALAT1 site 1).

We next compared Vx with MCP‐based editors. The MCP platform was pioneered by Tsukahara et al (Azad et al, 2017) and recently optimized by Mali et al (Katrekar et al, 2019). The optimized editors are MCP‐ADAR2_dd‐NES and MCP‐ADAR2_dd‐NLS (located in the cytoplasm and nucleus, respectively), which show the best balance of specificity and efficiency. As shown in Fig 3F, at the representative target (RAB7A), where the gRNA for the MCP editors has been optimized, the cytoplasmic version was ~ 2‐fold more active than the nuclear version as reported (Katrekar et al, 2019). Importantly, the default 50‐17 gRNA was sufficient for Vx to outcompete MCP‐ADAR2_dd‐NLS, albeit not MCP‐ADAR2_dd‐NES. We next tested three more gRNAs (50‐20, 50‐26, 50‐30), which is justifiable because the gRNA for MCP‐ADAR2_dd‐NES has been optimized, finding all three enabled Vx to outcompete MCP‐ADAR2_dd‐NES (Fig 3F). Vx also proved more active than both MCP‐based editors at all the 4 non‐representative target sites tested (Fig 3A, B, C and H). In contrast to the on‐target editing, the local off‐target effects seemed comparable between Vx and MCP editors (Fig 3A–C and H). Thus, Vx outperformed MCP editors as it did V1/2.

RNA editing has also been achieved using the gRNAs carrying the GluR2 hairpin in cells overexpressing exogenous, full‐length hADAR2 (exoADAR). This strategy was pioneered by Stafforst, Fukuda and colleagues (Fukuda et al, 2017; Wettengel et al, 2017) and the gRNA design recently optimized by Mali and colleagues at RAB7A (Katrekar et al, 2019). We therefore compared Vx with ExoADAR at RAB7A, finding Vx markedly more active at all the Vx gRNAs tried (Fig 3F). Vx also proved far more active at all 3 other target sites tested, whether using optimized (Fig 3A) or non‐optimized (Fig 3B and H) gRNAs. In contrast, the local off‐target effects seemed comparable between Vx and ExoADAR (Fig 3A, B, F and H), as was the global off‐target effects (Appendix Fig S4C). Of note, the global off‐target effect of ExoADAR is also similar to the MCP‐based system (Katrekar et al, 2019). Collectively, the data demonstrate that Vx was more efficient but equally specific as compared with the exogenous ADAR2‐ and MCP‐based tools.

An alternative method of SDRE is based on long antisense oligos, which are able to recruit endogenous ADARs to edit target mRNA, as initially demonstrated by Woolf and colleagues (Woolf et al, 1995) and recently optimized by Wei and colleagues (Qu et al, 2019). The optimal gRNAs are 111–151 nt in length carrying a mismatched C (opposite the on‐target) at the center, with the 151‐nt oligos displaying similar or stronger activities as compared with the 111‐nt oligos. As mentioned in the Introduction, such oligos frequently induce off‐target editing within the RNA duplex, which can be reduced at the expense of on‐target editing using mutant antisense oligos bearing mismatched G opposite the off‐targets. Multiple target sites were tested in the original paper, as represented by PPIB 5′ UTR (the most efficiently edited site by the 151‐nt oligo) and KRAS site 1 (used to demonstrate successful elimination of off‐target editing by a mutant 111‐nt oligo). We therefore compared Vx with the long oligos at these two representative sites and five other sites.

Compared with the WT 151‐nt oligo, at its representative PPIB 5′ UTR, Vx was as active with the default 50‐17 gRNA, but ~ 3× as active with the 50‐26 gRNA (Fig 3D) and also markedly more active at all five other targets tested using the default 50‐17 gRNA (Fig 3A, B, C, H and I). In particular, the 151‐nt oligo was only weakly active at PPIB sites 1–2, its editing rates 10× and 3× lower than Vx (Fig 3B and C). Indeed, the editing rates of the 151‐nt oligo were generally unimpressive (below 20%) at all the six sites tested except the two MALAT1 sites (~ 50%). Unfortunately, at both MALAT1 sites, extensive local off‐target edits at the flanking adenines were also created. The off‐target effects at site 1 could be suppressed using the mutant oligo (to a level comparable to that induced by Vx), but on‐target editing was also severely reduced (to 29%; Fig 3H), whereas at site 2, a mutant oligo spared on‐target editing but also failed to suppress multiple off‐target edits (A53, 63, 106,119, 139), resulting in far more collateral damage than Vx which only had 1 off‐target edit (A89; Fig 3I). We conclude that Vx tended to be more active than WT 151‐nt oligo (let alone the mutants) and more specific than the mutants (in terms of collateral editing at the target RNA). The same trend was seen for the 111‐nt oligo tested at its representative site (Fig 3E). Of note, these experiments were done in HEK293T cells, where the long oligos are known to be most active among eight human cell lines tested presumably because of high endogenous ADAR protein expression in HEK293T cells (Qu et al, 2019), raising the possibility that Vx may outperform the oligos to even greater extents in most other cell lines, given that Vx editing is independent of the endogenous ADAR protein.

CIRTS‐8 is an elegant innovation developed by Dickinson et al, where TBP6.7, a small RNA binding protein, is harnessed to deliver ADAR2_DD‐(E488Q) (Rauch et al, 2019). CIRTS‐8 is known to edit the Luciferase reporter in the presence of 40‐21 gRNA based on the luciferase assay, which is the only target transcript described in the original paper, and furthermore, only a single gRNA was tested, making it unclear if the 40‐21 gRNA was optimal (Rauch et al, 2019). We therefore tested additional gRNAs for CIRTS‐8. Using NGS, we confirmed the editing at the 40‐21 gRNA, but found the efficiency quite modest (23%), which could not be improved with all three other gRNAs tested (Fig 3G). We tested CIRTS‐8 at two more targets, and again found it modest (26 and 41% at the PRKN reporter and PPIB site 2, respectively; Fig 3A and B). Importantly, although Vx also displayed only modest activity at the Luciferase reporter (26%; Fig 3G), it proved markedly (~ 2×) more efficient at the PRKN reporter (60%) and PPIB site 2 (77%; Fig 3A and B), suggesting that Vx is overall more active than CIRTS‐8.

In summary, Vx outperformed all the rivals when tested at non‐representative sites of the rivals, where a single unoptimized gRNA was used for each editor. Vx also outperformed the rivals at their representative sites even though the latter (but not Vx) used optimized gRNAs, with only three exceptions: At PPIB 5′UTR, RAB7A and Luciferase reporter, the representative sites for 151‐nt oligo, MCP‐ADAR2_dd‐NES and CIRTS‐8, respectively, Vx using the generic 50‐17 gRNA was not more active than the rivals using optimized gRNAs (Fig 3D, F and G). However, at PPIB 5′UTR and RAB7A, Vx efficiencies can be readily raised above the rivals by adjusting the gRNA designs (Fig 3D and F). Thus, Vx is generally more active than the rivals, at least when the gRNAs are optimized. In contrast to the on‐target editing described above, the bystander effects of Vx are generally similar to or lower than the rivals (heatmaps in Fig 3), reminiscent of the relatively low bystander effects at the PRKN reporter (Fig 1D and Appendix Fig S1A). Taken together, these data show that Vx is highly efficient yet specific, comparing favorably with all the rivals.

Discussion

Vx, a critical addition to the base‐editing toolbox

We have shown that Vx is not only specific, but also highly active, outcompeting all its rivals at the targets (including nuclear RNA) tested. Vx could be the top choice in cases demanding maximal editing rates, such as the installing of loss‐of‐function mutations or correction of gain‐of‐function mutations. However, the long oligos are free of global off‐target edits, MCP‐based editors and CIRTS‐8 are small, and CIRTS‐8 is also unique in being made exclusively of human proteins. Thus, these rival editors have their own merits, and are complementary to Vx in RNA editing applications. We should also remind the readers that V1/2 used in this study bear truncated dPspCas13b lacking the C‐terminal aa 984–1,090 to facilitate the packaging into AAV. Although this mutation does not seem to affect editing, the mutant has not been extensively characterized (Cox et al, 2017). Finally, we wish to emphasize that all high‐specificity overexpressed editors, including Vx and V2, still possess detectable levels of global off‐target edits, with a few percent of the edits created by Vx and V2 (and presumably by other editors as well) occurring at mRNAs encoding tumor suppressors, whose transient editing might suffice to start transformation of a cell toward malignancy (Appendix Fig S2C, middle). Thus, further optimization is needed before clinical applications for these editors could be launched.

A multipronged approach for base editor optimization

Specific and efficient A > I editors have been created using chemically modified oligos. Given the challenges in the synthesis and delivery of such oligos, there has been much interest in creating genetically encodable editors with similar outstanding performance. However, such editors have remained elusive, because efficient editors are associated with extensive off‐target editing, with its countermeasures scanty and unsatisfactory (Mao et al, 2019). For example, nuclear localization can at best only mildly reduce the off‐target effects (Vallecillo‐Viejo et al, 2018), and mutagenesis can also compromise on‐target editing (Cox et al, 2017).

We have successfully taken a multipronged approach to selectively counter the off‐target effects in V1. This strategy has three components: use of the non‐promiscuous dCasRx, embedding the deaminase domain in dCasRx and localization of the editor to the nucleus, which reduced the off‐target effects 8×, 2× and 6×, respectively. These reductions are each mild, but the effects are presumably multiplicative, thus explaining the striking ~ 100× increase in the specificity of Vx relative to V1 (582 vs 54,523 off‐target edits). Among the three components, the promiscuity issue of Cas proteins seems largely ignored in the literature, and our study brings attention to this critical issue. Nuclear localization is known to selectively reduce off‐targets 2× for an editor (λN‐ADAR2_dd) (Vallecillo‐Viejo et al, 2018), but has no effect on another (MCP‐ADAR2_dd(E488Q)) (Katrekar et al, 2019). The six‐fold reduction we achieved using nuclear localization highlights the potential of this strategy, and is of interest in light of conflicting reports in the literature. Finally, the ability of Loop3 insertion to reduce off‐target effects is also impressive. Although the effect was only 2×, it was achieved without compromising on‐target editing and therefore significant, just as the 2× effect achieved by Rosenthal and colleagues using nuclear localization (Vallecillo‐Viejo et al, 2018).

Embedding ADAR2_DD(E488Q) in dCasRx could presumably create steric hindrance thus reducing the global off‐target editing, whereas forced tethering of Vx to the target sites might conceivably overcome the effect of steric hindrance to allow on‐target editing. However, this scenario cannot readily explain why, compared with (some) other fusion proteins, Vx was more robust to RNA designs at editing the on‐target while being less active at the bystanders flanking the on‐target (Fig 1C and D). Deciphering the puzzle would perhaps require the elucidation of the structure of the gRNA: editor: target transcript ternary complex. Whatever the mechanism, it is clear that inserting a catalytic domain into a recruitment module is worth trying when optimizing RNA base editors.

Materials and Methods

Constructs

The editor and reporter constructs were made using standard molecular biology techniques and key plasmids will be deposited at Addgene. Details of the constructs are provided as Supplemental Information.

Cell culture and transfection

The mouse neuroblastoma line N2a and human embryonic kidney line (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 the first day and transfected 1 day later. DNA was mixed with 1 μl Lipofectamine P3000 (Thermo Fisher Scientific, L3000015) into 25 μl Opti‐MEM (Invitrogen) and incubated for 5 min at room temperature. 0.75 μl of the Lipofectamine 3000 (Thermo Fisher Scientific, L3000015) was diluted into 25 μl Opti‐MEM (Invitrogen) and combined with the DNA: P3000 mixture, incubated for another 20 min at room temperature. The DNA: P3000: Lipofectamine 3000 mixture was added dropwise into the wells. Cells were analyzed 36–48 h post‐transfection.

Analysis of editing at the reporter by next‐generation sequencing and by FACS

To evaluate REPAIR activity at the GFP reporters bearing mutant human cDNA fragments, the reporter plasmid (20 ng) was co‐transfected in triplicates with editor (150 ng) and gRNA (300 ng) expression vectors into 48‐well plates. The GFP reporter constitutively expressed BFP, whereas the editors and gRNAs were co‐expressed with mCherry and puromycin resistance gene, respectively. For experiments described in Figs 1D and 2B, Appendix Fig S3D, puromycin (InvivoGen) was added to 3 μg/ml 12 h after transfection to select for the cells transfected with the gRNA expressing vector, which led to higher apparent editing yields. In all other transfection experiments in this study, including those focused on endogenous transcripts, puromycin was skipped. Cells were analyzed 36–48 h post‐transfection by targeted NGS of the reporter mRNA and sometimes also by FACS, as described below. First, for targeted NGS, total RNA was isolated using the TRIzol reagent (Invitrogen) and reverse transcribed with Oligo dT using HiScript II Q RT SuperMix kit (Vazyme, R223‐01). The cDNA in the reaction mixture was then amplified by PCR targeting the regions encompassing the mutant human cDNA fragments using a high‐fidelity enzyme; the PCR primers are listed in Supplemental Information. The amplicons were then subjected to two rounds of PCR to add Illumina adaptors and sample barcodes before being sequenced on the Illumina HiseqXten‐PE150 platform. The adapter pair of the pair‐end reads was removed using AdapterRemoval version 2.2.2, and pair‐end read alignments of 11 bp or more bases were combined into a single consensus read. All processed reads were then mapped to the target sequences using the BWA‐MEM algorithm (BWA v0.7.16). For each site, the mutation rate was calculated using bam‐readcount with parameters ‐q 20 ‐b 30. Second, for FACS, we collected the data using BD LSR Fortessa (BD Biosciences), and analyzed it with FlowJo. Briefly, we gated on the live cell population (identified based on their characteristic forward/side scatter pattern) and then displayed its GFP vs BFP expression as a bivariate contour diagram. We then gated on the BFP⁺GFP⁺ subset, quantified its BFP and GFP mean fluorescence intensities (MFI), and calculated the ratio of GFP MFI/BFP MFI; this ratio is taken a convenient but preliminary and relative measure of the editing rate.

Editing at endogenous transcripts

Cells in 24‐well plates were co‐transfected with plasmids expressing editors‐P2A‐mCherry (300 ng) and gRNA (600 ng). For editing by the 151‐ and 111‐nt oligos, 600 ng of the plasmids expressing the long oligos was co‐transfected with 300 ng of a marker plasmid expressing mCherry. 36–48 h later, transfected cells were enriched by sorting mCherry‐expressing cells. Editing at the target sites at the endogenous transcripts was then analyzed using targeted NGS, as described in the reporter assay. The PCR primers are listed in Supplemental Information.

Target selection rationale and gRNA design strategy when benchmarking Vx against other editors in HEK293T cells

Vx was compared with each of the eight competing editors at its representative site and a subset of non‐representative sites, the latter selected on a case‐by‐case basis in an effort to reach meaningful conclusions cost‐effectively (Fig 3, Appendix Fig S3A). For example, we compared Vx with V1/2 at four representative sites (PRKN, PPIB site1, PPIB site 2, KRAS site1) in combination with 1 non‐representative site (MALAT1 site 1), but not at the remaining four non‐representative sites (PPIB 5′ UTR, RAB7A, Luc reporter, MALAT1 site 2). This is because the results at the four representative sites (all on mRNAs) proved highly consistent, making additional testing at mRNAs hardly intriguing. Besides, PPIB 5′ UTR was difficult to amplify with PCR, while Luc reporter is not an endogenous transcript; indeed, the two sites were not used to test any editor except their cognate editors. Neither did we test V1/2 at MALAT1 site 2, because we found them barely active at MALAT1 site 1 as expected, and their performance at site 2 would be fully predictable. In general, Vx was compared with its rivals at 3–6 sites, with more sites tested for stronger or more interesting rivals. For example, we tested the 151‐nt oligo at six sites but 111‐nt only one site, because the 151‐nt is known to be more active than the 111‐nt, and is a major rival of Vx. However, we did not test the 151‐nt at all nine sites because the six sites already sufficed, as in the case of V1/2.

Regarding the gRNA design strategy, we used the 50‐17 gRNA for Vx by default because it performed more consistently than 30‐17 (see main text). For the eight rivals, the gRNAs at their representative sites in HEK293 cells have already been optimized (Appendix Fig S4A), and so were used as such; optimized gRNAs are indicated by asterisks in Fig 3. The gRNAs at the non‐representative sites were all designed in this study using the same configurations as the optimal gRNA at the representative sites. For example, for the long oligo method, we designed 151‐nt oligos carrying a mismatched C⁷⁶ as does the oligo targeting the representative PPIB 5′ UTR, and where indicated (Fig 3H and I), we also engineered mismatched Gs opposite 5–6 heavily edited bystanders to reduce their editing, per instruction described in the original paper(Qu et al, 2019). Of note, the mutant 111‐nt oligo in Fig 3E is already published (Appendix Fig S4).

Analysis of global off‐target effects at the transcriptome

To detect off‐target RNA editing sites across the transcriptome, we transfected the cells in 24‐well plate as described in the section above (300 ng editor and 600 ng gRNA plasmids) before the transfected 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 21–38 million reads per sample. In order to combat the effects of variable sequencing depths, all RNA sequence files were uniformly downsampled to 20 million reads per library (Appendix Fig S3A). The reads were mapped to the mouse reference genome (mm10) by STAR software (Version 2.5.1), using annotation from GENCODE version M21. After removing duplications, variants were identified by GATK HaplotypeCaller (version 4.1.2) and filtered with QD (Quality by Depth) < 2. All variants were verified and quantified by bam‐readcount with parameters ‐q 20 ‐b 30. The depth for a given edit should be at least 10× and these edits were required to have at least 99% of reads supporting the reference allele in the wild‐type samples. Finally, only A‐to‐G edits in transcribed strand were considered for downstream analysis. Edits were annotated using snpEff; proto‐oncogenes and tumor‐suppressor genes were downloaded from the UniprotKB/Swiss‐Prot database. Edits shown in the jitter plots are a union set of two replicates; the intersection set of edits is shown in Appendix Fig S3A. Transcript abundance was quantified using Salmon (v0.14.0) and DESeq2 (v1.22.2) with default parameters. To find genes whose expression was altered by the editors and to determine the variability between the replicates, we compared the expression profile of each replicate with the averaged profile of both replicates of the control sample expressing gRNA alone. Only the genes with FPKM (fragments per kilobase per million) > 10 in at least one replicate and with log₂(fold change) > 2 were considered.

Of note, our experimental procedures differ substantially from that in the original study on REPAIR(Cox et al, 2017): We transfected 300 ng editor plasmids into N2a cells in 24–well plate and sorted the transfected cells before NGS, while they transfected 10 ng into HEK293 in 96‐well plate and used the crude cells for NGS. Besides, we used union dataset while they counted the edits detected in at least two out of three replicates. Furthermore, we used V1–2 bearing the truncated dCas13b, whereas they used the intact version. Finally, we analyzed the off‐target edits in N2a instead of HEK293. These discrepancies may have contributed to much higher numbers of off‐target edits reported in this study (54,523 vs 18,385 for V1 and 239 vs 20 for V2).

Statistics

Statistical significance throughout the paper was calculated using two‐tailed Student's t‐test (*P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001). Data are represented as mean ± SEM, unless otherwise noted.

Author contributions

TC conceived of and supervised the project, with inputs from XingH and DL. SH analyzed the RNA‐seq data. YaL, SM, and YoL performed experiments with help from YC, MD, XinxH, JL, XW, JG, SS, and XZ.

Conflict of interest

The authors declare that they have no conflict of interest.

Supporting information

Appendix

Review Process File

The EMBO Journal (2020) 39: e104748

Data availability

Raw RNA‐seq and targeted NGS data are available at: BioProject PRJNA629461 (https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA629461).