Programmable RNA base editing with a single gRNA-free enzyme

Abstract

Programmable RNA editing enables rewriting gene expression without changing genome sequences. Current tools for specific RNA editing dependent on the assembly of guide RNA into an RNA/protein complex, causing delivery barrier and low editing efficiency. We report a new gRNA-free system, RNA editing with individual RNA-binding enzyme (REWIRE), to perform precise base editing with a single engineered protein. This artificial enzyme contains a human-originated programmable PUF domain to specifically recognize RNAs and different deaminase domains to achieve efficient A-to-I or C-to-U editing, which achieved 60–80% editing rate in human cells, with a few non-specific editing sites in the targeted region and a low level off-target effect globally. The RNA-binding domain in REWIREs was further optimized to improve editing efficiency and minimize off-target effects. We applied the REWIREs to correct disease-associated mutations and achieve both types of base editing in mice. As a single-component system originated from human proteins, REWIRE presents a precise and efficient RNA editing platform with broad applicability.

INTRODUCTION

Base editing is a new genetic modification strategy that directly installs point mutations into the DNAs or RNAs without making pernicious breaks in nucleic acid chain (1–4). Directed base editing has provided new potentials in treating genetic diseases caused by point mutations, such as Duchenne muscular dystrophy (5) or Hutchinson–Gilford progeria syndrome (6). Compared to DNA manipulation, editing on RNAs is non-permanent and reversible, making it relative safer for in vivo application as a potential therapy (7–10). Previously several groups had used guided antisense RNA oligonucleotide to recruit adenosine deaminases from the ADAR (adenosine deaminase acting on RNA) family to transform the RNA base from adenosine to inosine (recognized as guanosine) (3,11–16). However, the formation of dsRNA region and the recruitment of ADARs to the target RNAs are hard to control. Recently, Zhang's group developed two CRISPR-based systems using RNA-targeting Cas13 protein and ADAR2 to achieve base editing of the specific RNAs: the REPAIR used the active domain of ADAR2 to achieve programmable adenosine to inosine (A-to-I) editing, and the RESCUE employed a modified ADAR2 to enable additional cytidine to uridine (C-to-U) editing (4,17). More recently, the Chi's group used similar design to engineer two CRISPR-based editing systems, REPAIRx and CURE, that improve the target recognition with dCasRx (18,19). To overcome the difficulty for in vivo delivery of a large CRISPR-Cas system, a comparable but more concise system, CRISPR-Cas-inspired RNA targeting system (CIRTS), has also been developed by Dickinson's group to execute A-to-I editing on specific RNA targets (20).

All these engineered RNA base editing systems require the assembly of guide RNA (gRNA) and protein effector to form an RNA–protein complex for recognizing target RNAs. However, the cellular RNAs have much higher copy numbers than the genomic DNAs, the assembly and co-folding of the RNA–protein complex can be the rate-limiting step that reduce editing efficiency and hinder its application (21). Pairing of the gRNA and target RNA also leads to a relatively large editing window with potential ‘bystander’ editing sites (4,17,18). Finally, because Cas proteins are encoded by bacterial genome, the pre-existing immunity in a large fraction of human population presents a looming concern for future applications of CRISPR-based systems (22,23). Alternatively, cellular RNAs can be directly manipulated using engineered proteins that contain a programmable RNA-binding module and different functional modules (24). For example, the RNA binding scaffold of Pumilio and FBF homology (PUF) proteins contains eight repeat motifs, each can be reprogrammed to specifically bind any RNA base through interaction with the Watson–Crick edge (25–27). The PUF-based systems with different functional domains have been developed by several groups to detect cellular RNAs (28) or manipulate RNA splicing (29), translation (30), degradation (31) or methylation/demethylation (32).

Here we engineered a single component RNA editing tool, RNA editing with individual RNA-binding enzyme (REWIRE), which contains a programmable PUF domain (25,33) to specifically bind RNAs and a functional domain from human ADARs to edit adenosine or APOBEC3A (apolipoprotein B mRNA editing enzyme catalytic subunit 3A, or A3A) to edit cytidine (2,34) (Supplementary Figure S1A and S1B). Since the PUF domain can be reprogrammed to recognize almost any short 8-nucleotide (8-nt) RNA sequences (35,36), the REWIRE system can theoretically be designed to install two key transitions (A-to-I or C-to-U) in any RNA. We further redesigned the PUF domain in REWIREs to enhance the specificity and the editing efficiency. As a simple human-originated protein without any RNA components, this system provides an easier and more practical system for basic research and gene therapy.

MATERIALS AND METHODS

Plasmid construction

DNA fragments encoding the deaminase domains of the human ADAR1 and ADAR2 were amplified from the cDNA of HEK 293T cells. The hyperactive variants of hADAR1/2 were generated with primers containing corresponding mutations (37,38). All hADAR1/2 fragments were cloned into the pCI-Neo vector (Promega) to generate the pCI-ADAR vectors. Different versions of codon optimized PUF repeats from human PUM1 that recognize all four RNA bases were chemically synthesized (GENEWIZ). The eight or ten PUF repeats were assembled by PCR amplification and inserted to N-terminal of ADARs to generate various AI-REWIRE expression vectors. The same strategy was applied to generate the CU-REWIRE expression vectors, where we first cloned human APOBEC3A gene into pCI-Neo and then inserted different versions of PUF domains. NLS peptide sequences were synthesized (GENEWIZ) and inserted into the CU-REWIRE1.0 vector via Gibson Assembly (New England Biolabs).

For the upgraded REWIREs, we assembled ten synthesized PUF repeats using PCR to generate PUF-10R (35), and replaced the PUF-8R in the original REWIRE expression vectors, producing the AI-REWIRE3.0, AI-REWIRE3.1 and CU-REWIRE2.0. For AI-REWIRE4.0, AI-REWIRE4.1 and CU-REWIRE3.0, we further modified the PUF-10R by inserting a short peptide loop (MNDGPHS) between the ninth and the tenth repeat, forming PUF-10R*.

For the wide type and mutant EGFP reporters, the coding sequences were amplified from pEGFP-C1 (Clontech) with PCR and cloned into the pCDH-CMV-MCS-EF1-Puro vector (Promega). For the reporters containing disease-relevant mutations, the coding sequences of target genes bearing pathogenic G > A or T > C mutations (as defined in ClinVar Database) were synthesized and cloned into the pCDH-CMV-MCS-EF1-Puro vector. To compare different RNA editing tools, we cloned all the editing tools according to the published sequences and include a P2A-EGFP fragment at the C-terminus for Fluorescence Activated Cell Sorting (FACS). To carry on the orthogonal R-loop assay, we added the P2A-mCheery fragment at the C-terminal of the dSaCas9 to sort the cells with higher protein expression.

To clone various PUF domains that targeting different mRNAs into the REWIRE vectors, we used 2X MultiF Seamless Assembly Mix kit (ABclonal). DNA fragments for assembly were amplified with Q5 Hot Start High-Fidelity DNA Polymerase (NEB). All vector sequences were confirmed by Sanger sequencing.

All targeted sites and primers were listed in the Supplementary Table S1, and all sequences of REWIRE proteins were listed in the Supplementary Note 1.

Mammalian cell culture and transfection

HEK 293T (ATCC CRL-3216) and SH-SY5Y (ATCC CRL-2266) cells were maintained in Dulbecco's Modified Eagle's Medium with high glucose (HyClone) supplemented with 10% fetal bovine serum (FBS) at 37°C with 5% CO₂. HCT 116 (ATCC CCL-247) cells were maintained in McCoy's 5a Medium (Gibco) supplemented with 10% FBS at 37°C with 5% CO₂.

To assess the editing efficiency on the exogenous reporters, the cells were plated into a 12-well plate and transfected 12 h later (approximate 70% confluency) with different ratios of REWIRE expression vectors and reporter plasmids at a total DNA amount of 1 μg per well. The resulting cells were harvested 48 hours after transfection for further analyses. To edit the endogenous mRNAs, HEK 293T cells were transfected with 1μg REWIRE expression plasmids and harvested 48 hours post-transfection for subsequent analyses. All transfections were performed with Lipofectamine 3000 (Thermo Fisher Scientific) in 12-well plates (or 6-well plates with adjusted plasmid amount). To control for nonspecific transfection artifacts, we also treated the cells with only transfection reagent as a mock group.

Flp-In cell line construction

DNA fragment encoding dead EGFP (K163X) was amplified from pCDH-EGFP(K163X), and then inserted into modified pcDNA5/FRT/TO (without CMV enhancer) by Gibson Assembly. The resulting plasmid, pcDNA5/FRT/TO-EGFP(K163X), were co-transfected with pOG44 into Flp-In T-Rex 293 cells at a 1:9 ratio. 48 h after the transfection, the cells were washed and split into fresh medium (<25% confluency), further cultured at 37°C overnight. Subsequently the cells were feed with fresh medium containing 100μg/ml Hygromycin every 3 days until the resistant colonies were visible. 3–4 individual colonies were picked for further use. The inserted sequences were also confirmed by Sanger sequencing. The expression of the target gene could be induced by adding doxycycline to a final concentration of 1 μg/ml. After 24 h induction, the cells could be used to perform transfection and then be collected 2 days later for measurement of RNA editing.

Vector construction and production of AAV

To construct the AAV vectors of REWIREs, the backbone plasmid pAV-FH vector (Vigenebio, AV88001) was linearized with Kpn1 and EcoRV, and the PCR-amplified AI-REWIRE 4.1 and CU-REWIRE 3.0 were then cloned into linearized vector by Gibson Assembly. The resulting plasmid was named as pAV-AI-REWIRE 4.1 or pAV-CU-REWIRE 3.0. The AAV9 particles were produced by Vigenebio using HEK 293T cells using the triple-transfection protocols and purified with iodixanol gradient with standard protocols.

Flow cytometry

To assess the correction of K163X mutation in EGFP, the transfected cells were analyzed using flow cytometry at indicated time point with CytoFLEX S (Beckman Coulter). The percent of EGFP-positive cells and fluorescence intensity were calculated to reflect editing efficiency. The data was analyzed and visualized by CytExpert software (Beckman Coulter).

To compare the different RNA editing tools, the cells with high fluorescence intensity were sorted by the BD FACSAria Fusion Flow Cytometer. The data was analyzed and visualized by BD FACSDiva software.

Western blot

The HEK 293T cells were lysed in 1 × SDS-PAGE loading buffer (Beyotime) and heated at 95°C for 10 min. The mixtures were subsequently separated by 4–20% SDS-PAGE Gel (GenScript) and transferred to poly-vinylidene fluoride (PVDF) membrane. The primary antibodies, anti-FLAG antibody (F1084, Sigma-Aldrich) and anti-GAPDH antibody (14C10, CST), were used at a 1:2000 dilution as per manufacturer's instructions. For the secondary antibody, the HRP-linked goat anti-mouse IgG (CST#7076) and goat anti-rabbit IgG (CST#7074) were used at a 1:5000 dilution. The HRP conjugated secondary antibodies were visualized by using an enhanced chemiluminescence detection kit and ChemiDoc Touch Imaging System (Image Lab, Bio-Rad).

RNA extraction and reverse transcription

The RNAs were extracted using the TRIzol (Thermo Fisher) following the manufacturer's instructions. 1 μg RNA was reverse transcribed using the PrimeScript RT reagent Kit with gDNA Eraser (Takara). The cDNA was then amplified for further analyses using deep sequencing on Illumina NextSeq platform. Extraction of genomic DNA was performed using the Mammalian Genomic DNA Extraction Kit (Cat#D0061, Beyotime).

Animal experiments

All experiments with live animals were approved by the Institutional Animal Care and Use Committee (IACUC) of the Shanghai Institute of Nutrition and Health, Chinese Academy of Sciences. B6-EGFP mice (C57BL6-Tg (CAG-EGFP)10sb/J) were purchased from Jackson Laboratory. The male and female mice were evenly assigned to groups of predetermined sample size. AAV particles encoding AI-REWIRE 4.1 or CU-REWIRE 3.0 were injected into the B6-EGFP mice through tail intravenous injection (2×10¹² vector genomes per mouse in 4-week-old B6-EGFP mice (5,6)). Four weeks after injection, we collected muscles and heart samples from B6-EGFP mice and evaluated corresponding editing efficiency via RNA-Seq and Sanger-seq. No visible symptoms, such as a rough hair coat and moved slowly were observed in all mice after injection.

Sanger sequencing and deep sequencing of RNA amplicons

The cDNA was amplified with gene-specific primers flanking the target sequence (all primers and next-generation sequencing amplicons are listed in Supplementary Table S1). The A-to-I or C-to-U editing rates were evaluated from Sanger sequencing using the percentage of peak area at the target adenosine site according to the sequencing chromatograms. For deep sequencing of RNA amplicons, the cDNA was amplified with gene-specific primers with appropriate forward and reverse adaptor sequences. The PCR products were purified, and the amplicon libraries were constructed using the 300-cycle Mi-Seq Reagent Kit v2 or Micro Kit v2 according to the manufacturer's protocols for paired-end sequencing (2 × 150 bp) on the Illumina Hi-Seq X Ten machine by the Omics core facility of PICB. The RNA editing rates were evaluated by the percentage of the edited reads at the target sites (see below).

RNA-Seq experiments

The cells were harvested 48 h after transfection, and total RNAs were purified with TRIzol Reagent (Thermo Fisher). For each sample, 1 μg of total RNAs were used to prepared ribo-minus RNA-Seq libraries using the KAPA Stranded RNA-Seq Kit with RiboErase (Roche, KK8484), and the depletion of rRNAs confirmed by fluorometric quantification using the Qubit RNA HS Assay kit (Invitrogen). The sequencing was performed on an Illumina Hi-Seq X Ten machine by the Omics core facility of PICB (2 × 150-base pair paired end; 50M reads for each sample). The summary of RNA-Seq data was listed in Supplementary Table S2.

RNA editing analysis using RNA-Seq data

For deep sequencing analysis, we first conducted quality control and trimming using TrimGalore (0.6.0). To analyze sequences near the targeted sites, the sequence index was generated using the targeted site sequence (upstream and downstream 200-nt) of each transcript. The RNA-Seq reads were aligned to the index and quantified using BWA(0.7.17-r1188) (39). Alignment BAMs were then sorted by Samtools (1.9) (40), and the RNA editing sites were analyzed using REDItools (1.2) (41) with the following parameters: −t 20 −p −u -m20 −T6-6 −W −v 1 −n 0.0 −e −d −l. AWK scripts were used to filter out all the A-to-I or C-to-U conversions within the REWIRE targeting region, which were considered significant if the Fisher's exact test yielded a p-value less than 0.05 after multiple hypothesis correction by Benjamini Hochberg correction and at least two of three biological replicates identified the edit site. The mutations that appeared in both control and experimental groups were considered as single nucleotide polymorphisms.

To analyze entire transcriptome, AWK scripts were used to filter out the overexpressed REWIRE gene. After trimming, the reads longer than 90-nt were mapped to the reference genome (GRCh38/hg38 for human and GRCm38/mm10 for mouse) by STAR(2.4.2) (42). We then used Samtools (1.9) to sort alignment BAMs and remove duplicated reads by using Picard MarkDuplicates.jar(2.20.1–0). RNA base-editing variants were called using REDItools(1.2) with the following parameters: −t 20 −p −P −u −U −a6-6 −A6-6 −v 1 −n 0.0 −N 0.0 −e −E −d −D −l −L. With consideration of strand information, all the A-to-I or C-to-U conversions were selected using AWK scripts. From all the called variants, downstream analyses focused solely on canonical chromosomes (1–22, X, Y and M for human, 1–19, X, Y and M for mouse).

An additional layer of filter for known variants was performed using SNPs data from dbSNP version 154 and Mouse Genome Project version 7 for human and mouse respectively. The variants shared by at least two out of three biological replicates in each sample were identified as the RNA editing sites. The RNA editing level of the mock group was treated as the background, and the global off-targets of REWIRE were calculated by subtracting the background variants. The thresholds we used to filter the SNVs are based on minimum coverage (20 reads), number of supporting reads (at least five mutated reads), allelic fraction (≥0.5%), quality of the mapped reads (>30), and base quality (>30). For editing rate plot we set a higher depth cutoff at 100. The sequence motifs were generated with R package ggseqlogo(0.1) (43). Off-target sites were considered when the edit sites were identified in at least two out of three biological replicates. The editing rate of off-target edits were calculated as the average number of edits reads divided by the average coverage of those sites (mean expression level) of the overlapping samples, multiplied by 100.

To analyze the number of on- and off-target editing sites (Figure 1C, Figure 3C, Figure 5E and Figure 5F), we used reads-per-million (RPM) (a.k.a., counts-per-million, CPM) to normalize the sequencing depth among samples, which was also the raw data used by other groups when they quantify the editing rates.

Figure 1.

Engineering AI-REWIREs for A-to-I transformation at specific positions. (A) Engineering AI-REWIREs for specifically A-to-I editing. Domain configuration of AI-REWIRE (left); schematic of AI-REWIRE mediated specific A-to-I base editing on RNAs (right). (B) Efficient editing of EGFP mRNA by different versions of AI-REWIREs. The PUF binding site in EGFP is highlighted in blue, with the nearby adenosines marked in red (top). Heatmap showed the editing rates of all adenosines near the on-target site A437 (position 1) for each AI-REWIRE and control. The on-target editing rate of A437 was shown at right. The editing rates presented were measured by RNA-Seq with triplicates (see methods). (C) Scatter plots of transcriptome-wide A-to-I RNA editing in the samples treated with AI-REWIREs. For each edited adenosine, the numbers of unedited reads (x-axis) and edited reads (y-axis) were plotted. The on-target site A437 is highlighted in orange, and the adenosines near the targeted site (in a 400-nt window) are labelled in blue. n, numbers of all edited adenosines detected. The 5% editing rate was indicated with dashed lines. Three independent experiments were performed (see methods). (D) Editing rates of transcriptome-wide A-to-I RNA editing in the samples treated with AI-REWIREs (data from Figure 1C with minimum coverage cutoff at 100). On-target sites are represented by red rhombuses; average editing rates of off-targets in each sample are indicated with orange dots and values are nearby; the 5% editing rate was indicated with a dashed line. (E) Sequence logos derived from edited adenosines. The analyses were conducted using RNA-Seq data from samples treated with different AI-REWIREs (data from C). (F) Editing of exogenously expressed GRIA2 mRNA by AI-REWIRE1.0 that binds at different distances from the intended editing site (A1820). Left, the editing rates of all adenosines near the binding site were plotted against the distance from the editing site to the 5′ end of PUF binding site. Right, the editing rates of two major sites were plotted against the binding distance (mean ± s.e.m. n = 3).

Figure 3.

Engineering CU-REWIREs for targeted C-to-U conversion. (A) Engineering CU-REWIRE for specific C-to-U editing. Left, Domain configuration of CU-REWIRE; right, schematic of CU-REWIRE mediated specific C-to-U base editing on RNAs. (B) Efficient editing of EGFP mRNA by CU-REWIRE. The PUF binding site in EGFP and the nearby cytidines were marked (The experiments and analyses are similar to Figure 1B), with the C504 as the on-target editing site. CU-REWIRE with mutated deaminase domain was included as negative control. (C) Scatter plots of transcriptome-wide C-to-U RNA editing in the samples treated with overexpressed free APOBEC3A, CU-REWIRE and its inactive version. The data with free wt APOBEC3A was adopted from (Huang et al, 2020) (19). All these editors were co-expressed in HEK293T cells with the GFP reporter transcript. The on-target site C504 is highlighted in orange, and the cytidines near the targeted site (in a 400-nt window) are labelled in blue. n, numbers of all edited cytidines detected. The 5% editing rate was indicated with dashed lines. Off-targets in each sample are intersection sets of 2 biological replicates at a depth of 30M reads. (D) Jitter plots showing editing rate of total off-targets (data from C with additional coverage cutoff at 100). The number of off-targets are display at the top and the orange dots showing the average editing rates with numbers listed. (E) On-target editing rates of CU-REWIRE with different editing distances from binding sites, which represent the number of bases between the editing site and 3′ end of the PUF binding site. (F) Sequence logos derived from edited cytidines. The analyses were conducted using RNA-Seq data from samples treated with different CU-REWIRE (data from C). (G) Editing activity of CU-REWIRE on all combinations of 3-nt flanking bases around the candidate cytidine on the EGFP transcript. (H) Schematic diagram of orthogonal R-loop assay. Upper, the dSaCas9 could generate an R-loop at the selected genomic site by recruiting a gRNA (Site 5 in this study). Several C sites within the R-loop are anticipated to be deaminated by the deaminase domains (UGI in the BE4max (low left) and APOBEC3A in the CU-REWIRE (low right)). In this study, on-target editing (at the EMX1 locus and CTNNB1 mRNA for BE4max and CU-REWIRE, respectively) is also analyzed to serve as positive controls for the off-target editing assay. (I) Off-target editing in HEK293T cells quantified with the R-loop assay. dSaCas9 and its gRNA were co-expressed with the indicated editors or the controls. The cells with high expression of the editors were sorted by the flow 3 days later. The off-target was measured by the deep sequencing. On-target editing at the EMX1 locus and CTNNB1 mRNA were quantified by deep sequencing. The samples were from the same experiment as shown in (I). Values represent mean ± s.e.m. (n = 2). (J) Editing of CTNNB1 mRNA with different doses of CU-REWIREs expression vectors. The editing rates of all cytidines nearby the PUF binding site were measured by Sanger sequencing. Values represent mean ± s.e.m. n = 3. (K) Editing rates of on-target C1549 at CTNNB1 with different amounts of CU-REWIREs. The rates were measured by Sanger sequencing (top). Values represent mean ± s.e.m. n = 3. The protein expression level of CU-REWIREs in different co-expression experiments with GAPDH as a loading control (bottom). The western blot was done in technical duplicate.

Figure 5.

In vivo RNA editing in the B6-EGFP mouse model. (A) Schematic for AI-REWIRE 4.1 delivered to B6-EGFP mice through tail vein IV injection of AAV9. The total insert size between the two inverted terminal repeats (ITR) was 4.4 kb. (B) Efficient Editing of EGFP mRNA by AI-REWIRE 4.1 in the tibialis anterior muscles. The PUF_10R binding site and the on-target editing site A437 (position 1) were highlighted. The editing rates were measured by RNA-Seq with triplicates as described in Figure 1B. Values represent mean ± s.e.m. n = 3. (C) Schematic for CU-REWIRE 3.0 delivered to B6-EGFP mice using AAV9. The total insert size between the two ITR was 3.7 kb. (D) Efficient editing of EGFP mRNA by CU-REWIRE 3.0 in the tibialis anterior muscles. The experiments were similar to (B). Values represent mean ± s.e.m. n = 3. (E) Scatter plots of transcriptome-wide A-to-I RNA editing in the muscle samples treated with AI-REWIRE 4.1 and PUF control in mouse muscle cells. The labels are similar with Figure 1C and Figure 3C. (F) Scatter plots of transcriptome-wide C-to-U RNA editing in the muscle samples treated with CU-REWIRE 3.0 and PUF control in mouse muscle cells. The labels are similar with Figure 1C and Figure 3C. (G) Editing rates of transcriptome-wide RNA editing in samples treated with different REWIREs (data from E, labels and cutoff are same as Figure 1D and Figure 3D). (H) Efficient editing of EGFP mRNA in muscle and heart tissues. The editing rate was evaluated by Sanger sequencing. Values represent mean ± s.e.m. n = 3, respectively.

For analyzing whether the REWIRE constructs perturbed the RNA expression levels, the transcript per million (TPM) values output from the RSEM analysis were used for expression counts and transformed to log-space by taking the log₁₀(TPM). To find differentially regulated genes, a Student's t-test was performed on three REWIRE edited or PUF-ctrl replicates versus three Mock replicates. The statistical analysis was performed on genes with TPM values greater than 0.5, and genes were considered differentially regulated if |log₂(fold change)| >1. Genes were reported if they had a false discovery rate (Benjamini–Hochberg correction) of less than 0.01.

The effect of identified variants was determined using the Variant Effect Predictor tool from Ensemble with default parameters and option ‘-pick’ to filter for one consequence per variant.

Whole-genome sequencing and data analysis

The cells were harvested 48 hours after transfection, and total genomic DNA were extracted and purified with TIANamp Genomic DNA Kit (Tiangen, DP304) for subsequent sequencing analysis. Indexed DNA libraries were prepared by using NEBNext Ultra II FS DNA Library Prep Kit for Illumina. A total of 12 Tb WGS data were obtained by using Illumina Hiseq X Ten (2 × 150) at Omics core of Bio-Med Big Data Center, CAS Shanghai Institute of Nutrition and Health.

For genome-wide mutation profiling, we adopted sequencing data for hA3A-BE3 sample from PRJNA692189 (GSE164837) (44) and sequenced the others. We scaled sequencing depth to 30× for each sample. Reads were trimmed using Trimmomatic (v.0.38; parameters: ILLUMINACLIP:TruSeq3-PE-2.fa: 2:30:10 LEADING:3 TRAILING:3 SLIDINGWINDOW:4:15 MINLEN:36) to remove low-quality reads. The BWA-MEM algorithm (v.0.7.17-r1188; default parameters) was used to map clean reads to the human reference genome (GRCh38/ hg38). Samtools (v.1.9; parameters: −bh −F 4 −q 30) was performed to select reads with a mapping quality score ≥30 and to sort BAM files. Duplicated reads were marked by Picard (v.2.21.2; parameters: REMOVE_DUPLICATES = false) in the BAM file. GATK (v.4.1.3; default parameters) was next used to correct systematic bias using a two-stage process (BaseRecalibrator and ApplyBQSR; default parameters). To identify genome-wide de novo variants with high confidence, we used three algorithms: GATK (v.4.1.3), LoFreq (v.2.1.3.1, default parameters) and Strelka2 (v.2.9.10; default parameters), separately. The overlapped SNVs were considered to be the reliable variants. To further obtain de novo SNVs, we further filtered out the background variants in human 293T cells, including: (i) SNVs identified in the non-transfected 293T cells of this study and that existed in the NCBI dbSNP (v.151; http://www.ncbi.nlm.nih.gov/SNP/) database; (ii) the overlapped SNVs with mutation rate less than 10%; (iii) variants overlapped with the UCSC repeat regions and (iv) further analyses focused solely on canonical (chromosomes 1–22, X, Y and M) chromosomes.

Statistical analysis

To assess whether REWIRE constructs perturbs natural editing homeostasis, we analyzed the global editing sites shared by the mock group and PUF-Ctrl group. The differential RNA editing rates at native A-to-I or C-to-U editing sites were assessed using Pearson's correlation coefficient analysis. Pearson correlations of the editing rate between the mock group and PUF-Ctrl group were calculated and annotated in Supplementary Figure S2C and D.

RESULTS

Engineering AI-REWIREs for specific A-to-I transformation

To engineer the artificial enzymes that specifically catalyze A-to-I editing, we fused the PUF domain with the deaminase domain of ADAR1 or ADAR2 in different orders, generating PUF-ADAR and ADAR-PUF fusion proteins (Supplementary Figure S1C). We engineered the PUF domain to target EGFP (Enhanced Green Fluorescent Protein) mRNA and co-expressed them in HEK293T cells, where the A-to-I editing events were measured by sequencing the RT-PCR products. Only the PUF-ADAR configuration showed detectable base editing activity near the PUF binding site (Supplementary Figure S1C), and thus this configuration was used for the future design of all the AI-REWIREs (Figure 1A).

We engineered various versions of AI-REWIREs using wild type ADARs or their hyperactive mutants (ADAR1 in AI-REWIRE1.0, ADAR1-E1008Q in AI-REWIRE1.1, ADAR2 in AI-REWIRE2.0, ADAR2-E488Q in AI-REWIRE2.1) (37,38), and found that all AI-REWIREs showed efficient base editing at the on-targeted site A437 of the EGFP mRNA with highest editing rate at ∼81% (Figure 1B and Supplementary Figure S1D and S1E). As expected, the editing sites occur at the upstream of AI-REWIRE binding region, consistent with that the PUF domain binds RNA in an antiparallel fashion (25). As a control, the expression of PUF or ADAR alone did not induce detectable editing near the targeted region (Figure 1B). Moreover, the inactive AI-REWIREs (with inactive ADAR1-E912A/ADAR2-E396A mutant) also had no A-to-I editing capability, indicating that the editing carried by the AI-REWIREs was indeed cause by the deaminase activity of ADARs (Figure 1B and Supplementary Figure S1E). In addition, the editing efficiency was also affected by the expression levels of AI-REWIREs, as judged by co-expression of AI-REWIREs and the target mRNAs in different ratios (Supplementary Figure S1F). The estimated editing rates are highly correlated between the measurements with Sanger sequencing and RNA-Seq (Supplementary Figure S2A and B). While the Sanger sequencing provides a more accurate picture at the targeted site and the bystander editing site near the REWIRE binding window, the RNA-Seq can give a global picture for the transcriptome-wide off-target effect.

Transcriptome-wide RNA-Seq revealed that the major off-target A-to-I editing sites by AI-REWIREs are in the targeted EGFP mRNA rather than endogenous RNAs (Figure 1C, blue versus grey). As expected, the AI-REWIRE1.1 and AI-REWIRE2.1 have higher editing activities at both the target site and the off-target sites compared to those with wild-type domains (version 1.0 and 2.0) (Figure 1C and D). In addition, we found no significant difference in global editing of endogenous sites between cells transfected with PUF domain only and mock vector (Supplementary Figure S2C and D), suggesting PUF domain had little impact on normal editing by endogenous ADARs. Moreover, the analysis on the cellular transcriptome also indicated the overexpression of the PUF domain does not make a significant change on the transcriptome compared to the mock transfection control (Supplementary Figure S2E–G). Consistent with sequence preferences of ADARs (45), the AI-REWIREs were found to have a slight preference of guanine next to the edited adenosines (Figure 1E).

To determine the optimal distance between the PUF binding site and the editing site, we constructed several AI-REWIREs that specifically bind to the shifting positions at downstream of a natural edited site in GRIA2 mRNA (Figure 1F and Supplementary Figure S3A). We observed an optimal editing window between the position 6–8 from the PUF binding site to the on-target editing site and with a low bystander editing (Figure 1F). When the binding sites of AI-REWIREs were shifted toward 3′ end, editing of a downstream adenosine (A1824) was increased (Figure 1F and Supplementary Figure S3A). Our subsequent results suggest that AI-REWIREs had a fairly narrow window for optimal editing (∼3 nt) (Supplementary Figure S3B and C).

Application of AI-REWIREs in base correction

We next applied this system to repair a nonsense mutation (K163X) in EGFP gene. Initially the dead EGFP was co-expressed with two versions of AI-REWIREs designed to recognize a downstream sequence CUUCAAGA, and we found that the adenosine was efficiently edited to restore the green fluorescence (Supplementary Figure S4A), enabling a functional rescue of EGFP in 40–60% transfected cells as judged by flow cytometry (Supplementary Figure S4B). Furthermore, we generated a stable cell line containing a single copy of dead EGFP stably integrated into the genome by Flp-In system (Figure 2A). The dead EGFP was expressed from a moderate CMV promoter without the proximal enhancer to mimic the expression level of a typical endogenous gene. We expressed the two versions of AI-REWIREs in different amounts (Figure 2B), and found a dose-dependent base editing of K163X site, reaching 20% editing rate with the AI-REWIRE2.1 (Figure 2C–E). We also monitored the phenotypic change for several days after AI-REWIRE transfection and found that the EGFP signals appeared at day one after transfection, reached the peak at day 2, and faded out after day 4 (Figure 2F).

Figure 2.

Application of AI-REWIREs. (A) Using AI-REWIREs to rescue dead EGFP. The K163X mutation was introduced to create a premature stop codon in the EGFP gene that was stably integrated into 293 cells with Flp-In system. The A-to-I editing on this site can restore the translation of full-length EGFP. (B) Western blot of expression level of AI-REWIREs and the rescued EGFP, with detection of GAPDH as loading controls. (C) Flow cytometry of dead EGFP cell lines transfected with different amounts of AI-REWIRE2.1 using a 12-well plate. The cells were analyzed 48 h post transfection. (D) Editing rates of the K163X (measured by Sanger sequencing) in dead EGFP cells. The experiment conditions were same as panel C. (E) Fluorescence intensity of the rescued dead-EGFP cells. The experiment conditions were same as panel C. (F) Duration of RNA editing by the AI-REWIRE2.0/2.1, judged by the fluorescence intensity at each day after a single transfection of 1 μg plasmid. (G) AI-REWIREs edited the TP53 A553 site in a dose-dependent manner (see methods). The editing rates of all adenosines nearby were measured by Sanger sequencing (Left), with the editing rate of the on-target A553 site shown in right. Values represent mean ± s.e.m. n = 3. (H) Western blot of expression level of AI-REWIREs in different co-expression experiments (samples are consistent with G). The AI-REWIERs were Flag-tagged and the GAPDH were used as a loading control. The western blot was done in technical duplicate. (I) A-to-I editing of DMD 1682G > A mutation that is associated with Duchenne muscular dystrophy disease, the Deep-seq data showing the 1682G > A in DMD is corrected to varying levels with different AI-REWIREs. (J) Editing of exogenous disease-related G > A mutations and endogenous sites using AI-REWIREs specifically designed to recognize the target mRNAs. The mutations were selected from ClinVar. (mean ± s.e.m. n = 3).

We further designed AI-REWIREs to edit several endogenous genes, including the cancer-associated TP53 gene that was previously edited by other RNA editing tools (4,16) (Figure 2G–I). Using different amounts of plasmids in transfection, we again found a dose-dependent editing of the A553 in TP53 gene with efficiency ranging from 13% to 58% with two versions of AI-REWIREs (Figure 2G and H). For the five endogenous editing sites tested, we achieved 15–59% editing using different AI-REWIREs (Figure 2J and Supplementary Figure S4C), suggesting that AI-REWIREs are capable of editing endogenous mRNAs although the editing efficiency remains to be improved.

We further examined whether AI-REWIREs could effectively edit disease-related mutations in human cells. Different versions of AI-REWIREs were engineered to target several mutations, including two DMD mutations that cause Duchenne muscular dystrophy. We co-expressed the mutated genes with corresponding AI-REWIREs, and found that the mutated base can be corrected efficiently, with AI-REWIREs yielding higher editing rates ranging from 20% to 60% (Figure 2I–J and Supplementary Figure S4D–E).

Engineering CU-REWIREs for specific C-to-U conversion

The successful engineering of AI-REWIREs prompted us to employ similar design for C-to-U editing. We combined the PUF domain and the cytosine deaminase APOBEC3A with a peptide linker to generate the CU-REWIRE (Figure 3A). As a proof of concept, we tested the activity of CU-REWIREs on EGFP mRNA and found that the CU-REWIRE could enable precise C-to-U editing with efficiency up to ∼60% (Figure 3B and Supplementary Figure S5A). Additionally, the CU-REWIRE(Mut) containing an inactive APOBEC3A mutation (E72A) did not show any detectable editing (46), suggesting that the base editing is indeed catalyzed by APOBEC3A (Figure 3B).

We further used RNA-Seq to examine the off-target C-to-U editing induced by CU-REWIRE, and found some off-target editing sites in both the targeted EGFP mRNA and other endogenous transcripts (Figure 3C and D). Consistent with earlier reports (19), expression of wide-type APOBEC3A alone also induced many off-target RNA editing (average editing rate ∼20%), suggesting an intrinsic limitation for this deaminase domain that should be further diminished with mutational screen. Interestingly, the endogenous off-target sites with high editing rates are mostly expressed in a very low level, which might be contributed by an ‘editing noise’ (see discussion for more details). Compared to AI-REWIREs, the CU-REWIRE appeared to have a tighter editing window and performed C-to-U editing at position 2 after the PUF binding site (Figure 3E). Moreover, the edited cytidines were predominantly found within the short consensus motif UC (Figure 3F). Such preference was experimentally validated using all base combinations around the targeted cytidine, where the UCN sequence was strongly edited by the CU-REWIRE (50.4–67.9%) and the CCN motif was also tolerated to a less extend (2.8–11.6%) (Figure 3G). This finding is consistent with the target preference of APOBEC3A (46). Interestingly, we found that the guanosine at immediate downstream of the cytidine can reduce editing efficiency, reflecting an unreported feature for APOBEC3A target preference (Figure 3G).

Since APOBEC3A was reported to catalyze the cytosine deamination in single stranded DNA (2), it is also possible for CU-REWIRE to cause unwanted editing on genomic DNA. To address this concern, we use orthogonal R-loop assay (47,48) to directly test if CU-REWIRE can induce editing on the ssDNA region (Figure 3H). We found a low level DNA editing (0.6–1.42%; Figure 3I) at the genomic ssDNA region (site 5) by the CU-REWIRE that efficiently edited CTNNB1 mRNA (∼68%). In comparison, the DNA editor BE4max showed a much higher rate of off-target editing (5–15%) in this region, which is similar to what has been reported (19). Furthermore, we performed the whole genome sequencing and used three different methods to analyze potential mutational effect induced by CU-REWIRE (Supplementary Figure S5B). As a control, the DNA base editor hA3A-BE3 (2) containing the human APOBEC3A was analyzed in parallel. We found that the genome-wide DNA editing by CU-REWIRE was not more prominent than the background (Supplementary Figure S5C) and not biased toward C-to-T or G-to-A substitution (Supplementary Figure S5D), suggesting that observed ‘mutations caused by CU-REWIRE’ are more likely to be the background sequence errors.

In addition, to test if CU-REWIREs are capable of editing endogenous mRNAs, we designed a CU-REWIRE to target an endogenous gene CTNNB1, and achieved 20–45% editing efficiency at the targeted C1549 site in a dose-dependent fashion (Figure 3J and K). To further test the general applicability in RNA base correction, we designed CU-REWIREs targeting two disease-related T > C mutations in human genes (EZH2 in Weaver syndrome, SCN1A in Dravet syndrome). We found that the deleterious mutations were indeed corrected via cytidine editing with efficiency ranging from 21% to 29% (Supplementary Figure S5E and S5F). Furthermore, to test if CU-REWIRE is capable to editing a viral RNA that is usually expressed in a high level, we also targeted the C246 in SARS-Cov-2 membrane protein, and achieved high editing efficiency (∼62%) (Supplementary Figure S5G).

Improvement of the REWIREs

To optimize the REWIREs, we redesigned a PUF domain with 10 tandem repeats (PUF_10R) that can specifically recognize RNAs by a 10-nt binding site rather than an 8-nt site of wild type PUF_8R, thus should significantly increase its specificity and reduce off-targets (27,35). We applied this design strategy to generate AI-REWIRE3.0 and AI-REWIRE4.0 using two versions of PUF_10R (Figure 4A, see Materials and Methods), and further engineered AI-REWIRE3.1 and AI-REWIRE4.1 by upgrading deaminase domain with hyperactive ADAR2. As expected, the AI-REWIREs with PUF_10R substantially reduced both the editing rates and the numbers of off-target sites without sacrificing the on-target editing efficiency (Figure 4B and C). Compared to the most active AI-REWIRE2.1 (on-target editing rate = 81%, off-target counts = 31804), the updated AI-REWIRE3.1 had much less off-target counts (1908) and a decent editing rate (∼62%) while the AI-REWIRE4.1 showed similar tendency (on-target editing rate = 72%, off-target counts = 6664) (Figure 4C and G). Interestingly, such improvement was achieved with a lower expression level than the early versions (Supplementary Figure S6A and B), which is a desirable feature for in vivo application where the high expression of engineered editase may be difficult to achieve and/or have adverse effects. We further used the upgraded AI-REWIRE4.0 to edit an endogenous gene (PPIB), and found that it achieved a comparable editing rate to the AI-REWIRE2.0 targeting the same site (Supplementary Figure S6C), but with much less off-targeting effect (Supplementary Figure S6D), which is consistent with its longer recognition sequence (10nt vs. 8nt). Since the PUF domain recognizes a short RNA sequence, this specific AI-REWIREs has other recognition sites in the transcriptome, some of which also being edited (brown dots in Supplementary Figure S6D). This is an inherent limitation of all engineered factors using PUF domain, which could be managed through target design and additional controls in protein expression (see discussion). As expected, both AI-REWIREs function in a dose-dependent manner in editing endogenous targets (Supplementary Figure S6E).

Figure 4.

Optimization of the REWIREs. (A) Schematic of AI-REWIREs with upgraded PUF variants. Top, the original AI-REWIRE with the PUF-8R that recognize targets by 8-nt sequence. Bottom, upgraded AI-REWIRE with the PUF-10R that recognize targets by 10-nt sequence. (B) Efficient editing of EGFP mRNA by upgraded versions of AI-REWIRE. The PUF binding site in EGFP is underlined in red. The experiments and analyses were similar to Figure 1B. The AI-REWIRE 3.0, 3.1, 4.0 and 4.1 all used PUF_10R scaffold as target recognition domain, with 4.0 and 4.1 containing a short loop between PUF repeat 9 and 10 (see method). (C) Scatter plots of transcriptome-wide A-to-I RNA editing with edited/unedited reads (labelled as Figure 1C) for the samples treated with AI-REWIREs. (D) Schematic of CU-REWIREs with upgraded PUF variants. (E) Efficient editing of EGFP mRNA by upgraded versions of CU-REWIRE. The PUF binding site in EGFP is underlined in blue, with the nearby cytosines marked in red (on-target site, C459). The experiments and analyses were similar to Figure 3B. The CU-REWIRE 2.0 and 3.0 all use PUF_10R as target recognition domain (see methods). (F) Scatter plots of transcriptome-wide C-to-U RNA editing in the samples treated with CU-REWIREs (labeled as Figure 3C). (G) Editing rates of transcriptome-wide RNA editing in samples treated with different REWIREs (data from (C) and (F), labels and cutoff are same as Figure 1D). (H) Successive editing of adenosine and cytidine with the AI-REWIRE 4.1 and CU-REWIRE 3.0 that bind to the same site on EGFP (the binding site is underlined in red and blue respectively). The experimental condition is similar with (B) and (E) except that half amounts of each vector were transfected. The experiments and sequencing were carried out in duplicates. Left, representative Deep-seq reads surrounding the target site; right, the editing rates calculated using the number of reads supporting single and concurrent editing divided by the total reads.

We further compared the on- and off-targets of REWIRE system with two other RNA editing tools recently published (4,18). We expressed the AI-REWIREs, REPAIRE-V1 and REPAIRE-Vx to target the same endogenous PPIB target site (c.325A), and sequenced the total RNAs. For this comparison, we used FACS to sort the cells with high expression of the editors, which was routinely performed by other groups. We found that the AI-REWIREs showed higher editing efficiency (AI-REWIRE4.0, 46.54%; AI-REWIRE4.1, 86.78%) than the CIRSPR-Cas based tools (REPAIRE-V1-sgRNA1, 67.22%; REPAIRE-V1-sgRNA2, 44.90% and REPAIRE-Vx, 2.24%) (Supplementary Figure S6F). Meanwhile, we also analyzed the off-target effect of each tool, and found that REWIREs had similar (or slightly lower) number of off-target editing sites compared to REPAIR-V1. The REPAIR-Vx had the lowest off-target editing, however its on-target editing rate was also low (Supplementary Figure S6G).

With the same strategy, we upgraded the original CU-REWIRE (CU-REWIRE1.0) into CU-REWIRE2.0 and CU-REWIRE3.0 using different versions of PUF_10R (Figure 4D and E), and measured their editing efficiency on EGFP mRNA (Figure 4E and Supplementary Figure S6B). While all the CU-REWIREs showed precise editing at the targeted cytidine (C459), the CU-REWIRE3.0 achieved the highest editing rate (up to 65%) and relatively low off-target effects across the entire transcriptome (Figure 4F and G), indicating a comprehensive improvement. We designed new CU-REWIRE3.0s targeting two disease-related T > C mutations in EZH2 and SCN1A genes, and found that these mutations were indeed corrected via cytidine editing with modestly increased efficiency while the target specificity was improved (Supplementary Figure S6H and I).

To take advantage of REWIREs that can introduce multiple variants in RNAs, we applied AI-REWIRE and CU-REWIRE together for concurrent editing on a single transcript, producing efficient A-to-I and C-to-U editing simultaneously (Figure 4H and Supplementary Figure S7). We also noticed that using two REWIREs in combination showed comparable editing rates to the AI-REWIRE-4.1 or CU-REWIRE-3.0 alone at the single base (compare Figure 4H to Figure 4B and E), suggesting they may function independently. Importantly, the overlap of the binding sites for two editases did not interfere with the concurrent editing (Figure 4H), suggesting they undergo dynamic turn over on the target RNA like what would be expected for bona fide enzymes. Within each targeted mRNA molecule, both AI-REWIRE 4.1 and CU-REWIRE3.0 preferably edited on the single intended site as judged by reanalyzing the RNA-Seq data (Supplementary Figure S8). In addition, we also tested the REWIREs in additional two human cell lines (SH-SY5Y and HCT116), and found that different versions of REWIREs achieved comparable editing efficiency on the targeted sites (60%-86% for A-to-I editing, and 25–80% for C-to-U, Supplementary Figure S9).

RNA editing in mouse model with REWIREs

We further tested the in vivo editing efficacy of the REWIRE system using a transgenic mouse model (B6-EGFP) that stably expresses EGFP. The EGFP-targeting AI-REWIRE4.1 or CU-REWIRE3.0 were packed in AAV9 vectors, which were systematically delivered to 4-week-old mice using intravenous injections at the tail vein (Figure 5A and C). Because the AAV9 vectors has a high tropism for muscle and heart (49,50), we collected the muscle and heart samples at 4 weeks post-injection, and evaluated the editing efficiency via Sanger sequencing and RNA-Seq. At the tibialis anterior muscle of the three mice, the AI-REWIRE 4.1 achieved 27%-34% editing efficiency at the targeted A437 site of EGFP (Figure 5B), whereas the CU-REWIRE 3.0 targeting the C459 of the EGFP showed 44%-51% editing efficiency (Figure 5D). As a control, the expression of the PUF domain alone through same AAV9 vectors did not induce detectable editing (Figure 5B and D).

We further measured the low off-target editing on the mouse transcriptome using RNA-Seq profiling, and found that most off-target sites were edited with low efficiency (∼6.8% for AI-REWIRE4.1 and ∼4.6% for CU-REWIRE3.0) (Figure 5E–G). In addition, we observed effective RNA editing of the intended sites in both heart and muscles, which is consistent with the tissue preference of AAV9 (Figure 5H). These results suggested that REWIREs can effectively edit RNAs in animal, however the editing efficiency remains to be optimized.

DISCUSSION

Compared to DNA, targeted RNA editing is reversible and more flexible, presenting a more tractable route for gene manipulation. Here we engineered a series of programmable RNA editases for specific base editing. With different functional domains, the REWIREs can effectively achieve both A-to-I and C-to-U editing, and can be used in correcting pathogenic mutations of endogenous genes. The protein sequences of this system are entirely originated from human proteome, which should avoid the innate immunity from CRISPR-based system using bacterial proteins. Importantly, the REWIRE system achieved in vivo base editing in mice through systematic delivery with AAV (Figure 5), suggesting its potential in personalized gene therapy.

The REWIREs consist of a single protein that recognizes target RNAs without the assembly of gRNA/protein complex, thus harboring high efficiency while avoiding assembly intermediates or misassembled by-products. The RNA-binding affinity is similar for all PUF_8Rs with similar GC content in the targets (K_d in the order of nM) (25,29), and PUF_10R can increase the binding affinity by ∼10 fold (35). A single component editase can also be transported to different subcellular compartments (e.g. mitochondria or chloroplast), which may be used in specific editing of extranuclear genes. As an RNA editing tool, the optimized REWIRE has several desirable features. First, the REWIREs with PUF_10R have a small size (873 aa and 644aa for AI- and CU- REWIREs) suitable for AAV delivery in clinical applications. Second, the REWIRE sequences are entirely originated from human genes, making it less likely to induce immune response for in vivo applications. Third, the REWIRE directly converts the RNA base without using endogenous repair pathways, thus may be able to edit RNA in post-mitotic cells including neurons. Finally, different REWIREs can be independently applied to achieve simultaneous A-to-I and C-to-U editing in the same transcript, allowing concurrent editing of multiple mutations that synergistically cause human diseases (Supplementary Figure S8).

As a new technology, the REWIRE system also has limitations that should be improved with future optimization. First, PUF domains target a relatively short sequence (8–10-nt in this study), which is shorter than the CRISPR-based system but is higher than the 7-nt seed match of siRNAs. This may contribute to some unintended targets with similar recognition sites in the transcriptome. Indeed, we found that the editing rates on random off-target sites are significantly lower than the sites with a nearby PUF-binding motif, suggesting that the sequences containing similar PUF recognition sites were more likely be edited by REWIREs (Supplementary Figure S10). This problem could be managed by optimizing target designs. For example, the specificity could be improved by using tandem PUFs or a PUF with even more repeats (a PUF with 16 repeats has been proved feasible (35)); and computational analyses could be applied during target selection to minimize unwanted targets in transcriptome. However, the implementation of additional PUF repeats may require careful engineering, as the increase of PUF repeats can affect the structural curvature and reduce the RNA binding affinity (35). Additional number of repeats will affect the binding affinity to the substrates, but may reduce the dynamics of product releasing, thus it is hard to predict how the editing efficiency will be affected. We have introduced some loop region to increase the structural flexibility of PUF-10R (Figure 4) with the hypothesis that it may help product release. As a result, we observed a small improvement in the editing efficiency, suggesting that the idea to further optimize REWIRE is possible. Thus, we are exploring further optimization in engineering REWIREs with more PUF repeats. Second, REWIREs still have bystander editing sites within the editing window, although this problem is less severe than the CRISPR-based editing tools (4,17). It is possible to optimize the deaminase domain and the linker region to increase editing accuracy. Theoretically, we should be able to introduce mutations in the deaminase domain to increase the selectivity of editing sites, and/or modify the linker region to refine the editing window and reduce the bystander editing.

As reported previously, the CRISPR-based editors can impact the expression of various genes that not preferentially edited at the RNA level (19). Therefore, we examined the changes of gene expression using RNA-Seq, and found that the REWIREs altered expression of a small number of genes (Supplementary Figure S11A and S11B). These differentially expressed genes (DEGs) were not preferentially edited at the RNA level and were not enriched in specific functional pathway (Supplementary Figure S11C). We also predicted the effect and location of off-targeted edited sites identified in AI-REWIRE-4.1 and CU-REWIRE-3.0, and found some alterations in coding sequence (with ∼50% of all edits causing synonymous or non-synonymous mutations) and non-coding sequence (Supplementary Figure S11D).

The off-target effect has been a major concern for gene editing, especially for the genomic DNA because the deleterious changes will be permanent. However, the off-target editing on RNAs may be less detrimental since the cellular RNAs constantly undergo synthesis and degradation cycles. Both the edited and unedited versions of transcripts are co-expressed in the same cells, probably damping the side effect of the off-targeting editing. The off-target effects are mainly depended on the choice of functional domains. For example, the APOBEC1 were shown to induce more off-target editing sites in both DNA and RNA than the APOBEC3A, which can be reduced by more specific deaminase (7–10,51). We found that REWIREs with ADAR2 showed less off-target effects than ADAR1, and the hyperactive mutants increased off-target effects while improving editing rates (Figure 1C and D). Further optimization on the PUF domain also remarkably reduced the off-target effects while maintaining the high editing efficiency (Figure 4).

The activities of REWIREs may also be affected by the structure of target RNAs. The ADARs have several N-terminal dsRNA binding domains, and thus preferably catalyze the A-to-I editing in a dsRNA region (52). A detailed study of the ADAR2 deaminase domain (ADAR2DD) structure (53) showed that ADRA2DD prefer the unpaired Adenosine in a flip-out conformation, suggesting that ADRA2 prefers unpaired bases in dsRNA region. In the CRISPR-based editing tools, the gRNA pairs with the substrates forms a dsRNA region with substrate with the bulged out ‘A’, where the ADAR2DD can independently catalyze the editing reaction without dsRBMs. We found that the deamination activities of REWIREs do not require a formation of dsRNAs, suggesting that ADAR2DD can edit the single stranded RNA when recruited through a ssRNA binding domain (like PUF). However, it is possible that the activity of REWIREs may still be affected by local structure around the Adenosine. The effect of RNA structure to APOBAC3A activity is different from ADARs. Although it was reported that APOBAC3A generally prefer to edit C located in the end of a four-base hairpin loop (54), we found that the local sequence plays a more important role in determine the substrate preference. The CU-REWIRE preferably editing the C in the UCN motif and, to a less extend, CCN motif (Figure 3F). Such preference is similar to a previous report (19). We speculate that the binding of PUF may affect the local conformation of RNA bases, which in turn may affect the substrate preference. It is also worth mention that the RNA structures are generally quite flexible, and thus the substrate preference based on the RNA structure prediction from thousands of editing site should be considered with caution.

In addition, the common procedures to measure editing sites using RNA-Seq tend to over-estimate off-target sites, probably due to the sequencing errors associated with NGS technologies and underlying genomic variants that may confound the analyses (55). It has been reported that the accuracy for quantifying the RNA editing events can be generally improved by a higher sequencing depth and more uniform read coverage (56). Therefore, for a more reliable estimation, we sequenced the total RNAs (rather than mRNAs) with a deep coverage (∼50 million 150 bp paired-end reads for each sample) and conducted technical triplicates to reduce sequencing errors. We found that many ‘editing’ sites covered only by a small number of reads have surprisingly high editing rates. However, such estimation may not be reliable, as increasing the cut-off for the read numbers will eliminate these ‘editing’ sites whereas the true positive editing events remain unchanged (Supplementary Figure S12). This ‘editing noise’ is similar to a previous report that virtually all adenosines within human Alu regions could be edited as judged by an ultra-deep sequencing of selected Alu elements (>5000 reads/site) (57).

Compared to other gRNA-based RNA editing tools, including the new CRISPR-based REPAIRv2 (4) and REPAIRx (18) systems, the AI-REWIREs showed a higher editing efficiency at the targeted sites and a slightly higher off-target effect (Supplementary Figure S6F and S6G), meanwhile the CU-REWIREs presented comparable properties on editing efficiency and off-target effect to the latest C-to-U editing tools, CURE systems in HEK 293T cells (19). Our results also suggested that the editing efficiency could be affected by the choice of different targeted sites, and RNA local structure may play an important role in this processing (Supplementary Figure S6F and G). This is a rough comparison because of the technical challenges in estimating off-target effect (Supplementary Figure S12). Nevertheless, REWIRE proteins present a simple and human originated system with comparable performance. Multiple technologies may complement to each other in editing different targets, where the users should consider a tradeoff between editing efficiency and the off-target effects to find a sweet spot.

Although we have improved the REWIREs with different versions of PUFs, additional research is certainly needed to further optimize this system. Compared to previous tools using adenosine deaminase domain of ADARs for A-to-I editing and modified ADAR2 for C-to-U editing (4,17), we achieved C-to-U editing with cytidine deaminase APOBEC3A that is smaller than ADARs. The APOBAC family has many members with various specificities and strengths, therefore further exploration on the APOBAC family with possible modifications may expand the functional modules of REWIREs to edit additional bases. In addition, adjusting the composition and length of the internal linker in REWIRE may also improve its performance. Collectively, the REWIRE represents a simple and practical alternative to CRISPR-based approaches, expanding the current RNA editing toolbox.

DATA AVAILABILITY

The original RNA and DNA sequencing datasets have been deposited in the NCBI Gene Expression Omnibus (accession code GSE155734) and the National Omics Data Encyclopedia (accession code OEP001013) for public downloading. Custom computer code is available upon request.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

FUNDING

National Natural Science Foundation of China [32030064, 91940303, 31730110, 31570823, 31661143031 to Z.W., 31971367 to M.M.]; National Key Research and Development Program of China [2018YFA0107602, 2021YFA1300503]; Strategic Priority Research Program of Chinese Academy of Sciences [XDB38040100 to Z.W.]; StarryNight Science Fund at Shanghai Institute for Advanced Study of Zhejiang University [SN-ZJU-SIAS-009 to Z.W.]; Science and Technology Commission of Shanghai Municipality [17JC1404900, 18XD1404400 to Z.W.]; M.M. is supported by the National Postdoctoral Program for Innovative Talents [BX20180336] and Shanghai Super Postdoctoral Program; Z.W. is supported by the CAS Pioneer Hundred Talents program (type A). Funding for open access charge: National Natural Science Foundation of China.

Conflict of interest statement. The authors declare that a patent application related to this technology has been filed. Z.W. is a founder and scientific adviser for Enzerna Biosciences Inc that seeks commercial applications for engineered PUF factors.