Precise in vivo RNA base editing with a wobble-enhanced circular CLUSTER guide RNA

Wobble bases modulate editing depending on their orientation

To characterize the effect of wobble base pairs in the nearest neighbor context, we first studied all 12 possible triplets that contained either a U or a G either 5′ and/or 3′ next to the target A. While the target A was always placed opposite of a U base, similar to a regular bystander site, the nearest neighboring nucleotides were either conventionally Watson–Crick base-paired or wobble base-paired (Fig. 1a). Under these circumstances, four types of wobble base pairs can occur. When the U base in a 5′-UAN or 5′-NAU triplet (N = any base) is base-paired with a G, we refer to this as a 5′-G•U or 3′-G•U wobble, respectively; when the G base in a 5′-GAN or 5′-NAG triplet is base-paired with a U, we refer to this as a 5′-U•G or 3′-U•G wobble, respectively. As a benchmark to previous studies²¹, we also included experiments where the target A was mismatched with a G base (G•A mismatch), to suppress editing.

The experiment used an editing reporter construct, based on the earlier R/G-guide RNA approach^36,37. While the trans-acting guide RNA comprised only a double-stranded ADAR-recruiting domain and a single-stranded 20-nt specificity domain (SD), the cis-acting reporter additionally contained its own 20-nt target sequence (TS) and was located in the 3′ untranslated region (UTR) of an eGFP transcript to enable convenient Sanger sequencing readout (Fig. 1a). Editing was performed by transfecting the plasmid-borne editing reporter into Flp-In T-REx cells, overexpressing ADAR1 p110.

Importantly, we found that both, 5′-G•U and 3′-G•U wobble base pairs strongly suppress editing in the five triplets, 5′-UAN and 5′-AAU (Fig. 1b and Supplementary Fig. 1), which are highly editable under normal Watson–Crick base-pairing conditions and are, thus, a major source of bystander off-target events for trans-acting guide RNAs. In the 5′-UAU triplet, both G•U wobbles seemed to cooperate. For four of the five triplet contexts, the suppressive effect of the G•U wobble on editing significantly outcompeted the suppressive effect of the G•A mismatch (Fig. 1b). Interestingly, we found the opposite effect for the U•G wobble base pair. In particular, for the three triplets 5′-UAG, 5′-AAG and 5′-CAG, a clear editing-enhancing effect on the A directly adjacent to the 3′-U•G wobble was apparent (Fig. 1c). Because of the inability of ADARs to achieve sufficient editing at 5′-GAN triplets, the effect of 5′-U•G wobbles could not be verified (Supplementary Fig. 1).

The 5′-UAG triplet can simultaneously accommodate both a suppressive 5′-G•U and an enhancing 3′-U•G wobble base pair and, thus, allows studying their interplay. Our data suggest that the suppressive effect of the G•U wobble entirely dominates the activating effect of the U•G wobble (Fig. 1b).

It is intriguing to speculate that the enhancing effect of the 3′-U•G wobble could be combined with or replace the activating effect of the commonly used C•A mismatch at an on-target site such as 5′-UAG. We tested this idea using the CLUSTER guide RNA system on three endogenous targets (Supplementary Fig. 2a,b), with a mixed outcome. Only in one example (ACTB) was the U•G wobble a promising alternative to the C•A mismatch, although it could be useful in certain other sequence contexts for guide RNA structure optimization.

A guide RNA might find and bind to near-cognate sequences within the transcriptome to induce off-target editing. In this context, we aimed to understand whether a randomly occurring G•U wobble suppresses editing only in the context of a U-paired A or also in the rare context of a C-mismatched A, which is known to be more prone to editing. We evaluated the potential of 5′-G•U wobbles for suppressing editing at 5′-UAG sites where the A was mismatched with C. In two of three examples (GUSB and NUP43), the suppressive effect of the 5′-G•U wobble was strong enough to suppress editing even at the mismatched A (Supplementary Fig. 2c).

To see whether the enhancing and suppressing effects of wobble base pairs also apply for ADAR1 p150 and ADAR2, we selected the 5′-UAG triplet, which is a particularly frequent site of bystander editing, and transfected the corresponding plasmid-borne editing reporters into Flp-In T-REx cells, overexpressing ADAR1 p150 or ADAR2, respectively. As expected, the underlying mechanisms were not ADAR isoform dependent and worked equally well (Extended Data Fig. 1).

G•U wobbles improve precision and efficiency of LEAPER guides

We next applied G•U wobble base pairs in the context of unstructured 111-nt-long LEAPER guide RNAs¹⁶ (Fig. 2a and Supplementary Fig. 3), which are highly prone to bystander editing. Our data (Fig. 1b) indicated that G•U wobble base pairs would be convenient to suppress bystander editing in the five highly editable triplets 5′-UAN (N = A, U, G or C) and 5′-AAU and could be combined with the G•A mismatch at all other editable triplets, such as 5′-AAG and 5′-CAG. To test this concept, trans-acting LEAPER guide RNAs (Fig. 2a) encoded on plasmids were cotransfected into HeLa cells with plasmids carrying the full-length complementary DNA (cDNA) of one of three different target genes (AHI1 (ref. ³⁸), COL3A1 (ref. ³⁹) and BMPR2 (ref. ⁴⁰)) each carrying a disease-relevant W>amber STOP mutation (5′-UAG) (Fig. 2b–d). To also evaluate endogenous targets, LEAPER guide RNAs targeting a 5′-UAG within the 3′ UTR of NUP43 and RAB7A were transfected into HEK293FT cells expressing these genes (Extended Data Fig. 2a,b). As seen in both settings, LEAPER guide RNAs recruited endogenous ADAR to induce significant on-target editing in each of the five targets (54–80%); however, this was contaminated with massive bystander editing (>10 sites per target), as previously reported¹⁵.

a, Schematic of 111-nt-long unstructured linear LEAPER guide RNA. b–d, Editing heat maps of the LEAPER guide RNA-binding sites within the indicated transcripts: AHI1 (b), COL3A1 (c) and BMPR2 (d). The basic design column, LEAPER, lacks G•A mismatches and wobble base pairs. The other guide RNAs contain either G•A mismatches or G•U wobbles at G•U-amenable sites or a combination of both solutions at all bystander sites. In the latter case, G•A mismatches are placed at sites not amenable to G•U wobbles. The triplet context for each listed editing event is given with the target A highlighted in green and all off-target A bases in blue. The position of each site is given relative to the transcript and the target A (±0 position). Editing was performed with plasmid-borne guide RNA and target in HeLa cells (endogenous ADAR). Data are shown as the mean editing percentage ± s.d. of n = 3 (AHI1 and COL3A1) or n = 5 (BMPR2) biological replicates.

First, we tested for all five targets whether G•U wobble base pairs outcompete G•A mismatches to suppress bystander editing at such sites, where G•U wobbles are amenable (G•U at G•U-amenable sites) and compared such guide RNAs with guide RNAs that apply G•A mismatches at the same sites (G•A at G•U-amenable sites) (Fig. 2b–d and Extended Data Fig. 2a,b). Notably, we found that the G•U wobble strategy was very potent to suppress bystander editing, often but not always outcompeting the G•A mismatch approach. Moreover, the on-target editing yield was higher for four of the five targets (AHI1, 74% versus 51%, and COL3A1, 47% versus 29%, Fig. 2; endogenous NUP43, 36% versus 17%, and endogenous RAB7A, 45% versus 25%, Extended Data Fig. 2).

Second, we aimed to suppress bystander editing entirely by either fully relying on the prior-art G•A mismatch approach (G•A all off-target sites) or by combining G•A mismatches with G•U wobble base pairs (G•A and G•U at all off-target sites). For AHI1 (Fig. 2b), only the combination of G•A mismatches with G•U wobble base pairs achieved the entire suppression of bystander editing and a good on-target editing yield (45%), while the guide RNA using only G•A mismatches suffered from residual bystander editing (position −38, 43%; position −35, 15%) and a reduced on-target efficiency (26%). These results highlight the power of G•U wobbles to improve the precision of LEAPER guide RNAs. For COL3A1 (Fig. 2c) and NUP43 (Extended Data Fig. 2a), the on-target yields were again clearly better when complementing G•A mismatches with wobble base pairs (41% versus 21% and 23% versus 12%, respectively). However, because of the lower yields at all edited sites, the G•A mismatch approach appeared to give slightly better precision. The BMPR2 target transcript (Fig. 2d) gave equal on-target yields for both strategies but better precision when G•U wobble base pairs were included. In the case of RAB7A (Extended Data Fig. 2b), both designs performed similarly.

According to the literature, LEAPER guide RNAs that carry numerous G•A mismatches can suffer from a loss of editing efficiency¹⁶. The reason for this might be that G•A mismatches have the lowest duplex stability among all known nucleotide mismatches⁴¹. In contrast, the thermodynamic stability of the G•U wobble is comparable to the A-U base pair³². This might partly explain why on-target efficiency often benefitted when G•A mismatches were complemented with G•U wobble base pairs. Notably, the number of suppressive G•U wobble base pairs required to improve precision can be smaller than that of G•A mismatches because one G•U wobble acts simultaneously at its 5′ and 3′ nearest neighbor positions. Examples can be found in Fig. 2b,c (AHI1, positions +29 and +31; COL3A1, positions −5 and −3). Furthermore, we found several cases where G•A mismatching failed to fully suppress bystander editing, while the G•U wobble strategy succeeded (for example, AHI1 at positions −38, −35 and −31 and BMPR2 at positions −32, +32 and +35; Figure 2b,d).

After targeting both exogenous and endogenous transcripts ranging from high expression levels (AHI1, BMPR2 and COL3A1 all as cDNAs), over medium (RAB7A, normalized transcripts per million (nTPM) = 105) to low (NUP43, nTPM = 36) expression levels, our data suggest that the G•U wobble strategy is unaffected by target transcript abundance and, thus, widely applicable. Overall, the wobble strategy complements the prior-art G•A mismatch strategy very well and regularly improves both editing efficiency and precision.

G•U wobbles are widely applicable to RNA base-editing tools

Bystander editing is a common problem of all RNA base-editing systems, particularly those such as the λN-ADAR⁸ and Cas13-ADAR¹⁰ approaches (Fig. 3a–d and Supplementary Fig. 3) that also use genetically encoded guide RNAs. Thus, we applied the G•U wobble strategy with boxB guide RNAs (Fig. 3a) and direct repeat (DR) guide RNAs (Fig. 3c) to see whether it improves the precision of the λN-ADAR and the Cas13-ADAR approaches, respectively. For this, triple-plasmid protocols were used where plasmids encoding the guide RNA, the editase (λN-ADAR2Q or Cas13-ADAR2Q) and the target (AHI1 W725X) were cotransfected into HeLa cells. In contrast to the LEAPER approach, the λN-ADAR and Cas13-ADAR approaches apply a hyperactive ADAR mutant and use guide RNAs with comparably short antisense part (boxB, 49 nt; DR, 59 nt). Because of the shorter duplex, the number of bystander sites is overall smaller. Nevertheless, the λN-ADAR approach induced significant bystander editing (Fig. 3b). Notably, amenable sites were readily controlled by G•U wobbles but not by G•A mismatches alone. A combination of G•U wobble and G•A mismatches was able to fully suppress bystander editing at a minor cost of editing efficiency (65% ± 7% to 49% ± 2%; Fig. 3b). For Cas13-ADAR, the combination of G•U wobble and G•A mismatches gave the best results in terms of editing efficiency (25% ± 3% versus 20% ± 4%) and allowed for complete bystander suppression (Fig. 3d). However, the Cas13-ADAR system itself gave dramatically lower editing yields compared to the λN-ADAR system (26% ± 2% versus 65% ± 7%) and showed little specificity for the Cas13b protein, as the editing yield of the DR guide RNA with and without overexpression of the editase differed only by ~9% (Fig. 3d).

a, Schematic of the λN-ADAR editing system. The 2× boxB motif-containing guide RNA (84 nt) binds its target mRNA (through 49 bp) to recruit the engineered hyperactive λN-ADAR2Q editase. b, Editing heat map of exogenous AHI1 W725>amber mRNA targeted by the λN-ADAR system. c, Schematic of the Cas13b-ADAR system. The DR motif-containing guide RNA (87 nt) binds its target mRNA (through 51 bp) to recruit the engineered PspCas13b-ADAR editase carrying a hyperactive ADAR2 E488Q deaminase domain. d, Editing heat map of exogenous AHI1 W725>amber mRNA targeted by the Cas13b-ADAR system. e, Schematic of a symmetric ASO bound to its target mRNA (through 58 nt). The ASO is end-blocked by 2′-OMe, contains phosphorothioate linkages and recruits endogenous ADARs. f, Editing heat map of the ASO binding site within exogenous PEX1^G843D transcript. In b,d,f, the basic designs (columns 2× boxB, DR guide RNA and ASO) do not contain G•A mismatches, wobble base pairs or 2′-OMe modifications beyond the end-blocks. The other guide RNAs contain additional 2′-OMe modifications, G•A mismatches, G•U wobbles at G•U-amenable sites or a combination of these solutions at all bystander sites. In the case of a combined solution, G•A mismatches or 2′-OMe modifications are placed at sites not amenable to G•U wobbles. The triplet context for each listed editing event is given with the target A highlighted in green and all off-target A bases in blue. The position of each site is given relative to the transcript and the target A (±0 position). Editing was performed in HeLa cells using plasmid-borne guide RNAs (2× boxB and DR guide RNA) and editase (λN-ADAR2Q and Cas13b-ADAR2Q) or ASOs recruiting endogenous ADAR. Data are shown as the mean editing percentage ± s.d. of n = 3 biological replicates.

In ADAR-recruiting ASOs, the strategic placement of chemical modifications allows to control bystander events^13,14. However, dense chemical modification, for example with 2′-O-methylated ribose (2′-OMe), can interfere strongly with editing. Thus, we evaluated G•U wobble base pairs in a case where additional chemical modifications diminished the on-target efficiency. Using a chemically modified (phosphorothioate linkage and 2′-OMe end-blocked) 59-nt-long symmetric ASO (Fig. 3e), we targeted the PEX1 transcript, specifically the G843D substitution causative for the peroxisome biogenesis disorder Zellweger syndrome⁴². While placement of additional 2′-OMe modifications at the −25, −6 and +7 positions controlled bystander editing, they also reduced the on-target yield drastically (34% ± 7% to 12% ± 3%; Fig. 3f). By contrast, a combination of G•U wobble base pairs and 2′-OMe modifications enabled the control of bystander editing while preserving the on-target yield (28% ± 8%; Fig. 3f). The latter example shows that ASO-based approaches can also potentially benefit from wobble base pairs to maintain high on-target yields.

Superior off-target control in A-rich target sites

The suppression of bystander editing in closest proximity to an on-target A is a common problem for all fully encoded RNA base-editing systems. Strategies such as G•A mismatching¹⁶ or U depletion¹⁷ often lead to a substantial loss of editing yield when they are applied too close to the on-target site. To assess the G•U wobble strategy for such a setting, we systematically tested how far the suppressive effect of a single G•U wobble base pair extends in the 5′ and 3′ directions. For this purpose, we again used cis-acting constructs that placed an editable duplex in direct extension of an ADAR-recruiting domain (R/G helix; Extended Data Fig. 3a and Supplementary Fig. 4) into the 3′ UTR of an eGFP reporter. We then evaluated the editing yields after transfection of these constructs into ADAR1 p110-expressing Flp-In T-Rex 293 cells (Supplementary Fig. 4). To study the suppressive effect of the 3′-G•U wobble in the 5′ direction, we studied a series of three 5′-UA(A)_iU base-paring motifs (i = 1–3 A bases) with increasing distance between the 3′-G•U wobble and the on-target A (Supplementary Fig. 4a–c). Accordingly, we also designed a series of three 5′-U(A)_iAG base-paring motifs (i = 1–3 A bases) to test the effect of a 5′-G•U wobble in the 3′ direction (Supplementary Fig. 4d–f). The target triplet was either 5′-UAA or 5′-AAG, hereinafter indicated by square brackets. In both series, we also benchmarked the effect of the G•A mismatch for the same bystander off-target A site. Neither the 5′-G•U nor the 3′-G•U wobble affected the on-target editing yield negatively at any distance tested. Instead, the suppressive effect was almost entirely focused on the direct 3′ and 5′ neighboring base. This was in clear contrast to the G•A mismatch where not only the mismatched base but also the first and sometimes even the second neighboring base in both directions were negatively affected (Supplementary Fig. 4). Thus, the G•U wobble strategy should be particularly strong to precisely suppress bystander editing close to an on-target A. We show exemplary data on how the 5′-G•U wobble controls editing precision in a 5′-U[AAG] base-pairing motif (Extended Data Fig. 3a,c,e) and how the 3′-G•U wobble acts in the 5′-[UAA]U base-pairing motif (Extended Data Fig. 3a,d,f), always in comparison to the G•A mismatch. In both cases, the G•U wobble clearly gave a better balance of editing efficiency over editing precision.

Next, we transferred the concept to trans-acting CLUSTER guide RNAs (linear design), which targeted a 5′-U[AAG] site in the BMPR2 transcript (K984, on cDNA) by harnessing endogenous ADAR in HeLa cells (Extended Data Fig. 3b,c,g). As expected, the reference guide RNA showed good on-target yields (A, 59%; Extended Data Fig. 3g) but also a strong bystander editing at the 5′ neighboring A (30%). The strategically placed 5′-G•U wobble base pair was able to fully suppress this bystander editing. This was not the case with the G•A mismatch where only a partial suppression of bystander editing was achieved. Notably, the U depletion strategy¹⁷ even increased bystander editing to 38% (Extended Data Fig. 3g). The 5′-G•U wobble also gave the best on-target efficiency of the compared bystander solutions with 47% yield, whereas a G•A mismatch reduced the yield and U depletion almost fully blocked editing, highlighting the power of the G•U wobble strategy to achieve high efficiency and high precision in very A-rich triplet contexts where G•A mismatch and U depletion fail (Extended Data Fig. 3g and Supplementary Fig. 5a,b).

To show that this finding is highly generalizable, we performed a meta-analysis over three different target transcripts (AHI1, BMPR2 and COL3A1) representing three different A-rich target triplets (5′-[UAA]U, 5′-U[AAG] and 5′-U[AAA]U) while using two different editing approaches, the CLUSTER guide RNA with endogenous ADAR and the boxB/λN-ADAR system with engineered ADAR. Notably, not only bystander control but also on-target efficiency was significantly better with G•U wobble base pairs, demonstrating the strength of the strategy to suppress bystander editing precisely within A-rich triplets (Extended Data Fig. 3h and Supplementary Fig. 5).

Wobble bases improve the engineering of CLUSTER guide RNAs

CLUSTER guide RNAs represent a recent strategy to harness endogenous ADAR for precise and efficient RNA base editing¹⁵. The basic concept combines an ADAR recruitment motif, a 20-nt SD that binds the target site and three or more additional RSs 15–20 nt in length that bind to the target mRNA over a larger stretch of sequence space in a multivalent fashion (Supplementary Fig. 3). In silico optimization of the guide RNA sequence is applied to choose RSs in such a way that highly editable A bases are avoided, enabling high control over bystander editing. Furthermore, guide RNAs with a high tendency to form inhibitory secondary structure are sorted out automatically, which helps to improve editing efficiency. Similar to the report for the LEAPER system¹⁷, we herein established the ribozyme-based Tornado expression system⁴³ for circularization and, thus, stabilization of guide RNAs (Extended Data Fig. 4). Starting with a simple LEAPER design, we could verify the formation of cleanly circularized guide RNAs (Supplementary Fig. 6), which gave an improved editing efficiency on the endogenous RAB7A transcript (Supplementary Fig. 7). Particularly notable was the positive effect of circularization on the editing yield after stable integration of the LEAPER guide RNA cassette by the PiggyBac transposase into the genome of HeLa cells, which gave moderate and stable editing yields over several weeks even with the weaker Pol2 promoter, elongation factor 1α (EF1α) (Supplementary Fig. 8).

Compared to simple LEAPER guide RNAs, the flexible design of CLUSTER guide RNAs represented a considerable engineering challenge for circularization. The order of elements within the CLUSTER guide RNA, such as the SD, the RSs, the ADAR recruitment motif and the target mRNA exit points, can all be placed individually and relative to each other. On the guide RNA side, RSs can be placed 5′ and/or 3′ to the SD, while the ADAR recruitment motif can be flexibly placed anywhere in between. On the mRNA side, binding sites that correspond to the RS within the guide RNA can be located 5′ and/or 3′ of the TS, which corresponds to the SD within the guide RNA. Binding-site placement ultimately defines whether the exit points of the mRNA are close or distant relative to the ligation stem. Furthermore, the number and length of all antisense elements and linkers can be varied, which may individually affect the torsion within the guide RNA circle and, thus, its interaction with the target transcript.

To identify general design rules for circular CLUSTER guide RNAs, we tested various constructs on a luciferase reporter construct (Supplementary Figs. 9 and 10). We found that the mRNA exits should be placed approximately opposite of the SD and that the ADAR recruitment motif can be combined with the ligation stem into a larger RNA structure placed adjacent to the SD. Notably, a 5-nt bulge that separates ADAR recruitment motif and ligation stem achieved particularly high editing efficiency (Supplementary Fig. 10a–c, design L13). This ADAR recruitment motif was called the split-R/G motif and was used in subsequent designs.

After identifying general design principles for circular CLUSTER guide RNAs, we aimed to design an optimal guide RNA for a proof-of-concept in vivo study. Specifically, we aimed to target a premature STOP codon in the murine Mecp2 transcript (W104>amber), which causes severe Rett syndrome-like symptoms in mice carrying this patient mutation⁴⁴.

CLUSTER guide RNAs avoid bystander editing by choosing RSs that minimize the presence of editable A bases in the guide RNA–target RNA duplex. However, in highly A-rich target RNAs, such as Mecp2, individual binding regions for RSs can be separated by long distances within the target transcript (for example, spread over several exons) and the available sequence space for secondary structure optimization can be limited. Both effects can negatively impact the editing efficiency of CLUSTER guide RNAs¹⁵. However, by applying the wobble base-pairing strategy to suppress bystander editing, the highly limiting filter set that defined eligible A bases within the RS-binding regions could now be considerably expanded; 5′-UAB, 5′-BAU triplets (B = C, G or U) and 5′-KAAU (K = G or U) sequence motifs are now included, while only 5′-GAB triplets were previously allowed¹⁵. Furthermore, 5′-CAC triplets and all A bases located at either end of a binding region (edge A bases), which we identified as being resistant to off-target editing, could be used to expand the sequence space. With the old filter settings (Fig. 4a), a circular CLUSTER guide RNA with four RSs and a split-R/G ADAR recruitment motif needs substantial space on the Mecp2 target RNA—specifically, 1,470 nt for guide RNA V1 and 1,208 nt for guide RNA V2 (Fig. 4c). In contrast, with the new filter settings (Fig. 4b), the entire guide RNA (V3) covers only 127 nt on the murine Mecp2 transcript (Fig. 4c). This guide RNA V3 gave significantly better on-target editing than V1 and V2 (87% versus 63% and 65%), which used the old filter settings (Fig. 4d). Partially, this might also be attributed to the larger available sequence space, which allows selecting guide RNAs with a much lower level of inhibitory secondary structure (Fig. 4e). As expected, the expansion of eligible A bases led to the appearance of bystander editing at binding sites of the RSs, which were albeit well suppressed with the G•U wobble strategy (Fig. 5).

a, Results from an in silico search for RS-binding sites within the murine Mecp2 W104>amber ORF when applying the previous filter settings to the GuideRNA-Forge tool. The binding site for the SD is displayed in turquoise and outlined in black. It contains the target A in green and underlined. All potential RS-binding regions ≥ 20 nt length, in which binding sites for RSs can be selected, are displayed in light blue. The previous filter excludes all A bases except for those within a 5′-GAB (B = U, C or G) triplet context. b, Results for the same in silico search with the latest filter that included A bases in a 5′-GAB, 5′-UAB, 5′-BAU, 5′-KAAU and 5′-CAC triplet context, in addition to allowing A bases independent of their sequence context at the edges of a detected binding region. c, Illustration of binding regions in which specific binding sites of RSs of three circular CLUSTER guide RNAs are located within the Mecp2 ORF, generated using the previous filter (V1 and V2) or the latest filter (V3). d, Editing was performed with plasmid-borne guide RNA and murine Mecp2 transcript in HeLa cells (endogenous ADAR). e, Secondary structure prediction of guide RNA V1–V3 generated using the ViennaRNA Package 2.0 (ref. ⁵⁴). The mean free energy (MFE) of the antisense part (SD and CLUSTER of RS) is given in kcal per mol. The dot–bracket ratio (DBR) indicates the number of dots (unpaired bases) per bracket (paired bases) within the dot–bracket annotation of the antisense part. Data are shown as the mean editing percentage ± s.d. of n = 3 biological replicates. For statistical analysis, a Student’s t-test (two-tailed, parametric) was applied.

Editing heat map of guide RNA-binding sites within the murine Mecp2 W104>amber transcript. The triplet context for each listed editing event is given with the target A highlighted in green and all off-target A bases in blue. The position of each site is given relative to the transcript and the target A (±0 position). Editing was performed with plasmid-borne guide RNA and murine Mecp2 transcript in HeLa cells (endogenous ADAR). a, Circular CLUSTER guide RNA design. b, Circular LEAPER guide RNA design. c, Yields achieved using circular unstructured LEAPER guide RNAs containing a 111-nt-long antisense part. The basic design (column circular LEAPER) does not contain bystander solutions. The other guide RNAs either contain G•A mismatches or apply U depletion at all bystander sites, G•U wobbles at G•U-amenable sites or a combination of G•A mismatching or U depletion with G•U wobbles at amenable sites. d, Yields achieved using circular CLUSTER guide RNAs containing a 100-nt-long antisense part split into a targeting sequence (20 nt) and four RSs (each 20 nt). The basic design (column circular CLUSTER) does not contain bystander solutions. The other guide RNAs contain G•U wobbles at positions −49, −27 and +48, as well as no bystander solution, G•A mismatch or depleted U at position −5. Data in a,b are shown as the mean editing percentage ± s.d. of n = 3 biological replicates.

Very recently, circular LEAPER guide RNAs were demonstrated to recruit endogenous ADARs with moderate editing efficiency and precision in cell culture and in vivo^17,18. We benchmarked our best circular CLUSTER guide RNAs (Fig. 5a) for the Mecp2 W104>amber transcript against 111-nt-long symmetric LEAPER guide RNAs (Fig. 5b) and assessed various means of bystander editing suppression (Fig. 5 and Extended Data Fig. 5). Linear (Extended Data Fig. 5) and circularized LEAPER guide RNAs (Fig. 5b,c), using the Tornado expression system were tested by transfection into HeLa cells. As expected, the circular LEAPER-based design gave massive bystander editing at >10 sites, with yields of 20–39% at seven such sites (Fig. 5c, circular LEAPER). Again we first compared G•A mismatching¹⁶, U depletion¹⁷ and G•U wobbles to suppress bystander editing at the four bystander sites, which are amenable to G•U wobble base pairing. Given the low number of amenable sites (Fig. 5c) the effects on on-target editing efficiency and bystander editing were comparably low. However, when we aimed to suppress off-target editing at the major ten bystander sites, the on-target yield of a pure G•A mismatch (17% ± 2%) or pure U depletion (16% ± 1%) solution dropped considerably, while a combination of G•U wobble and G•A mismatch (25% ± 2%) outperformed the combination of G•U wobble and U depletion (18% ± 1%) for the best-performing circular LEAPER guide RNA (Fig. 5c). In contrast, a linear CLUSTER guide RNA applying the G•U wobble strategy already achieved bystander-free 38% ± 7% on-target editing (Extended Data Fig. 5f). Furthermore, all optimized circular CLUSTER guide RNAs gave an exceptionally good editing efficiency of >84% on target with very good precision (Fig. 5d), clearly outcompeting all tested LEAPER guide RNAs (Extended Data Fig. 5). G•U wobbles entirely suppressed bystander editing at all three bystander sites where RSs bound the Mecp2 mRNA. It might be possible that G•A mismatches or U depletion would work similarly well to suppress bystander editing at such sites but this was not tested. There was one remaining site (5′-AAG) at position −5 with a moderate editing yield of 8% that was not amenable for wobble base pairing. Editing at this site would not change the amino acid sequence of MeCP2 yet its editing was reduced by U depletion or G•A mismatching (Fig. 5d).

In vivo proof-of-concept in a murine Rett syndrome model

Before applying the optimized circular CLUSTER guide RNA in vivo, we verified its successful circularization in cell culture with two sets of reverse transcription (RT)–qPCR primer pairs (Supplementary Fig. 11). The use of an outward primer pair (Extended Data Fig. 6a) allowed us to verify circularization with high confidence (Extended Data Fig. 6b,c), while the use of an inward primer pair (Extended Data Fig. 6d) allowed us to quantify the strong effect (235-fold increase) of circularization on the total guide RNA abundance (Extended Data Fig. 6e), suggesting that the majority of guide RNAs are fully processed (Extended Data Fig. 6f). We chose the PHP.eB serotype for adeno-associated virus (AAV) encapsulation as it allows cargo delivery to the mouse brain after systemic administration⁴⁵. Indeed, this serotype was successfully used by us before to deliver the boxB/λ-ADAR tool into the same Rett syndrome mouse model for mutation correction⁴⁴. The targeting virus encoded the circular CLUSTER guide RNA as displayed in Fig. 5a,d (G•U at G•U-amenable sites). For the nontargeting virus control, the guide RNA’s antisense parts were scrambled.

Mice were treated with 4 × 10¹² viral genomes by retro-orbital injection and killed 4 weeks later; brain regions were analyzed separately for editing efficiency by Sanger sequencing. Editing levels differed among the seven brain regions, with clearly detectable editing in the midbrain, brainstem and thalamus, with an editing efficiency up to 19% (Fig. 6a). To better understand the key factors for successful editing, we analyzed all seven brain regions for the expression of the guide RNA (Fig. 6c), the AAV episome abundance (Fig. 6e) and the expression levels of all catalytically active murine Adar isoforms: total Adar1 (Fig. 6g), Adar1 p150 (Fig. 6i) and Adar2 (Fig. 6k). The relative expression of guide RNA and Adars were directly compared through RT–qPCR by normalization to the geometric mean of the same three housekeeping genes Actb, Rps29 and Rnu6 (Supplementary Figs. 12 and 13), while the AAV episome abundance was determined as the number of copies per cell. Unexpectedly, editing yield correlated the least with the Adar levels (Fig. 6h,j,l; R² = 0.29–0.51), even though Adar expression differed among brain regions, particularly for Adar2 (for example, thalamus versus other brain regions). This indicates that Adar abundance did not limit the editing outcome. The strongest correlation was found between guide RNA expression level and editing yield (Fig. 6d; R² = 0.87), suggesting that, even under circularization, guide RNA levels still limit on-target editing. Editing also correlated well with AAV abundance (Fig. 6f; R² = 0.84) and, in most brain regions, guide RNA expression seemed to also correlate well with AAV abundance. However, there were a few exceptions, such as in the cortex, where guide RNA expression was low even though AAV abundance was comparably high. This indicates, in agreement with our previous findings⁴⁴, that the strength of the guide RNA promoter (for example, U6 promoter) or guide RNA stability differs among brain regions, thereby limiting editing even after successful viral delivery. In summary, these data show that the guide RNA’s abundance (determined by its delivery, expression and stability) is the most important factor to achieve high on-target editing in the brain, while Adar levels seem less important. This may instruct future designs of expression cassettes for CNS applications.

a, Editing yields in different brain regions, after delivery of the AAV-PHP.eB-encoded guide RNA through retro-orbital injection into Rett mice carrying the Mecp2 W104>amber mutation and quantification by Sanger sequencing 4 weeks later. b, Correlation between the median editing in a and the geometric mean of the RT–qPCR targets in c,g,i,k. Olfactory bulb, Ob; cerebellum, Cb; hippocampus, Hi; cortex, Cx; midbrain, Mb; thalamus, Th; brainstem, Bs. c, Guide RNA expression quantified by RT–qPCR (normalized to the geometric mean of Actb, Rps29 and Rnu6. d, As in b but correlating a,c. e, Absolute quantification of AAV episomes per cell by standard curve qPCR (normalized to Actb). f, As in b but correlating a,e. g, As in c but for Adar1. h, As in b but correlating a,g. i, As in c but for Adar1 p150. j, As in b but correlating a,i. k, As in c but for Adar2. l, As in b but correlating a,k. m, Bystander off-target events at the guide RNA-binding sites and the TS. n, All amplicon reads with on-target editing binned according to their number of bystander events (based on Supplementary Fig. 14). o, Global RNA editing at 2,533 endogenous sites (coverage ≥ 50 reads, REDIportal⁵⁵). p, Thalamus sections stained for MeCP2 and DAPI (nuclei). For a, data are shown as the median editing percentage ± 95% CI determined in n = 3 mice (nontargeting virus) and n = 5 mice (targeting virus). For statistical analysis, a Mann–Whitney U-test (two-tailed, nonparametric) was applied. For c,e,g,i,k,m,n, data are shown as the median fold change ± 95% CI or median number of copies per cell ± 95% CI determined in n = 2 mice (three technical replicates each). The NGS analysis in o is based on results from n = 1 (nontargeting virus) or n = 2 (targeting virus) mice. The results in p are derived from n = 1 mouse per group. For b,d,f,h,j,l, the R² values were determined by simple linear regression.

Particularly in a clinical setting, editing must be efficient and precise, represented by cleanly edited transcripts devoid of unintended recoding events. To evaluate bystander off-target events in the Rett mouse model, we performed deep amplicon sequencing (average read depth of 47,009 and average coverage of 45,991) of the Mecp2 target transcript in all seven brain tissues. To ensure extensive detection of bystander events, we selected the two mice from the targeting virus group that had given the highest on-target editing yield in the thalamus. The on-target editing results matched very well with the Sanger sequencing (Fig. 6a,m and Extended Data Fig. 7). As controls, two nontargeting virus-treated mice were used. They showed considerably lower background (~0.2%) compared to Sanger sequencing (~5%) (Fig. 6a,m and Extended Data Fig. 7). Consequently, this now enabled us to measure on-target editing yields of 1.7–3.0% with high confidence in the olfactory bulb, cerebellum, hippocampus and cortex (Fig. 6m). Notably, bystander editing was hardly detected in any brain region. At three positions (−49, −27 and +48), G•U wobble base pairs were applied to suppress bystander editing. At positions −49 and −27, this was very successful; no bystander editing was detectable. At position +48, there may have been up to 0.12% bystander editing in the brainstem but 100-fold below the on-target editing yield in that tissue. Only one bystander was detected with high confidence. This was the unresolved, silent bystander site at position −5 that was already discovered in cell culture (Fig. 5d; 8% ± 4%). Here we detected up to 0.37% bystander editing in vivo in the brainstem (Fig. 6m). If they appear at many sites, even bystander events with low editing yield may sum up to interfere with on-target editing. To address this potential issue, we studied how often an on-target edited read is damaged by an additional bystander edit. We found that 98.3% of the ~10,000 detected on-target edited reads were completely bystander free (Fig. 6n and Supplementary Fig. 14), highlighting the impressive degree of editing precision achieved in the in vivo proof of concept.

Next, we evaluated the global editing precision by interrogating transcriptome-wide RNA sequencing (RNA-seq) data collected from the thalamus of targeting and nontargeting virus-treated Rett mice (Fig. 6o and Extended Data Fig. 8a). As before, we selected the two mice from the targeting virus group that gave the highest on-target editing yield in the thalamus with Sanger sequencing. First, we applied the RNA-editing index method that monitors changes in global RNA-editing levels in a highly unbiased manner and focused the analysis on the particularly critical coding sequence space⁴⁶. The A-to-G RNA-editing indices were nearly identical in both groups (Extended Data Fig. 8a), indicating that the global editing activity was overall not affected by the presence of the targeting guide RNA. Second, we tried to detect differentially edited sites between the two conditions (Fig. 6o), similar to a previous study¹⁵. However, compared to the same analysis performed with untreated Rett mice (Extended Data Fig. 8b), we detected no clear off-target sites. Only three sites fell slightly outside of the ±25% Δ-editing margin but this was likely because of normal variability among mice. None of the three sites were located in the coding region and none of them were complementary to the guide RNA. Third, we tried to detect off-target events in a candidate approach and searched in silico throughout the whole murine genome for transcripts with 20-nt similarity (with up to one mismatch) to the guide RNA-binding regions. We identified 65 potential sites; however, only four of these sites were sufficiently expressed (≥20 reads of coverage) to be evaluated and no off-target editing was detected at any of them. Overall, we were unable to identify any global off-target events, excluding mouse-to-mouse variability, which suggests a very high precision of our approach on the transcriptome-wide level.

RNA base-editing approaches that apply the overexpression of an engineered editase, such as λN-ADAR2, typically suffer from substantial, editase-dependent global off-target events^46,47,48. In our recent study—from which the noninjected mice data were derived (Extended Data Fig. 8a), we also detected numerous off-target events even though the native ADAR2 deaminase domain was fused to the λN peptide. We revisited this published RNA-seq data⁴⁴ to compare with the CLUSTER generated data. The boxB/λN-ADAR2 data were generated using the same experimental setup (mouse model, AAV serotype and application route) as in this CLUSTER study. We ran the data in parallel through the same next-generation sequencing (NGS) analysis pipelines. First, we evaluated the same 2,533 endogenous RNA-editing sites (with coverage ≥ 50 reads) that we evaluated for the CLUSTER approach, as shown in Fig. 6o. Again, we selected the tissue with the highest on-target editing yield, which was the brainstem in this study⁴⁴. We found a notable number (25) of potential off-target sites exhibiting a Δ-editing margin above 25% (Extended Data Fig. 8c). Importantly, the identified events included two evolutionary highly conserved, Adar2-specific editing sites in glutamate metabotropic receptor 4 (GRM4) and neuro-oncological ventral antigen 1 (NOVA1)⁴⁹, for which the change in editing level could have functional impact. Second, we analyzed the editing index in the coding sequence space and found a considerable increase in the index in the presence of the λN-ADAR2 deaminase (Extended Data Fig. 8c). Together, this shows, in accordance with the literature⁴⁷, that the harnessing of endogenous ADAR is more precise on the transcriptome-wide level than the ectopic expression of engineered ADAR effectors, whether hyperactive or native.

Lastly, we studied the restoration of MeCP2 protein expression and function upon treatment using the circular CLUSTER guide RNA (targeting virus). In the noninjected Rett mice, MeCP2 expression was not detectable by single-cell immunohistochemistry⁴⁴ in the thalamus, as the mutation resulted in an unstable protein (Fig. 6p and Extended Data Fig. 9). In clear contrast, MeCP2 was restored in ~33.3% ± 4% (median ± 95% confidence interval (CI)) of the cells in the thalamus of the treated brain. Importantly, the restored MeCP2 protein was localized in heterochromatic foci inside the nucleus, which is an accepted proxy for its in vivo binding ability to methylated DNA^50,51 and very similar to the positive control mouse where up to 100% of the cells showed MeCP2 protein associated with heterochromatin (Fig. 6p and Extended Data Fig. 9a,b). The abundance of foci increased with increasing editing yields over the evaluated brain tissues (Extended Data Fig. 9c). The results suggest that MeCP2 protein and its function are restored in cells where guide RNA is delivered and expressed.