RCK8 cells (DSMZ, Cat# ACC-561, RRID: CVCL_1883) and U2932 (DSMZ, Cat# ACC-633, RRID: CVCL_1896) were cultured at 37°C, 5% CO₂, in RPMI-1640 medium (Sigma-Aldrich, Cat# R8758), supplemented with 15% fetal bovine serum (PAN Biotech, Cat# P40-37100) and 1% of Penicillin/Streptomycin (Sigma-Aldrich, Cat# P4333). A549 cells (DSMZ, Cat# ACC-107, RRID: CVCL_0023) were cultured at 37°C, 5% CO₂ in high-glucose DMEM (Sigma-Aldrich, Cat# D6429) supplemented with 10% fetal bovine serum (PAN Biotech, Cat# P40-37100) and 1% penicillin/streptomycin (Sigma-Aldrich, Cat# P4333). HEK293T cells (obtained from DKFZ, ATCC, Cat# CRL-3216, RRID: CVCL_0063) were cultured at 37°C, 5% CO₂ in high-glucose DMEM (Sigma-Aldrich, Cat# D6429) supplemented with 5% FBS (PAN Biotech, Cat# P40-37100) and 1% penicillin/streptomycin (Sigma-Aldrich, Cat# P4333). Cell lines were authenticated using Multiplex Cell Authentication by Multiplexion (Heidelberg, Germany) as described recently (Castro et al., 2013). Additionally, the purity of cell lines was validated using the Multiplex cell Contamination Test by Multiplexion (Heidelberg, Germany) as described recently (Schmitt and Pawlita, 2009). No Mycoplasma, SMRV or interspecies contamination was detected.

For interferon-alpha (IFNα) stimulation, 2.5 × 10⁵ HEK293T cells were seeded in 12-well plates in a total volume of 1 ml media containing 200 U/ml of IFN-α (PBL Assay Science, Cat# 11100–1). After 16 h, cells were collected, and RNA was extracted using a Qiagen RNeasy Plus kit (Qiagen, Cat# 74134).

For doxorubicin treatment, 10⁵ HEK293T cells were seeded in 24-well plates to have around 30%–50% confluency the day after. The following day the cells were transfected with pcDNA3-hLincRNAp21-MS2 (Chillón and Pyle, 2016) using a mix of plasmid DNA and polyethyleneimine (PEI, Polysciences, Cat# 23966) in an approximately 1:1 ratio (2.5 µg DNA:2 µg of PEI). 6 h post-transfection, the media was replaced with new complete media and 2 µM doxorubicin hydrochloride (Sigma-Aldrich, Cat# D1515) or DMSO only as control (Sigma-Aldrich, Cat# D2650) were added 10–12 h post-transfection, for 12 h. RNA was then extracted using a Qiagen RNeasy Plus kit (Qiagen, Cat# 74134).

pcDNA3-hLincRNAp21-MS2 contains the 3898 nt-long LIsoE2 isoform of the human lincRNA-p21 (GenBank: KU881768.1) tagged with 24 copies of MS2 RNA hairpins, as previously described (Chillón and Pyle, 2016).

LentiCRISPRv2 was a gift from Feng Zhang (Addgene, plasmid #52961; https://addgene.org/52961; RRID: Addgene_52961) (Sanjana et al., 2014). DNA oligos #12–13 were cloned into this plasmid following the instructions of “lenti-CRISPRv2 and lentiGuide oligo cloning protocol” (Addgene plasmid #52961) to generate lenti-CRISPR-ADAR1 exon 4 [from Pestal et al. (2015); Supplementary Figure S7A]. As a control, lenti-CRISPR-NT (Lenti-NT) was cloned accordingly using oligos #14–15 based on control 800 from the GeCKO v2 library (Sanjana et al., 2014). pCMVDR8.91 (coding for HIV gag-pol) and pMD2.G (encoding the VSV-G glycoprotein) were a kind gift from Prof. Didier Trono (Lausanne, Switzerland).

pSpCas9(BB)-2A-GFP (PX458) was a gift from Feng Zhang (Addgene plasmid # 48138; https://n2t.net/addgene:48138; RRID:Addgene_48138) (Ran et al., 2013). DNA oligos #16–19 were cloned into this plasmid linearized by restriction digestion (BbsI) using NEBuilder® HiFi DNA Assembly Master Mix (NEB, Cat# E2621). We, therefore, obtained three plasmids for knocking out human DTWD1, DTWD2, or TSR3 as previously described (Takakura et al., 2019; Babaian et al., 2020) and an additional non-targeting control (NT-ctrl) based on control 800 from the GeCKO v2 library (Sanjana et al., 2014).

Ribo-depleted RNA-seq experiments from seven human tissues (Supplementary Material S1) were selected from the “RNA Atlas” project (Lorenzi et al., 2021) and downloaded from GEO under the accession GSE138734. Known annotations for antisense and protein-coding genes were obtained from Gencode (v38), downloaded in gtf format, and converted into bed format. Antisense and protein-coding annotations were intersected by means of the “intersect” function embedded in the Bedtools package (Quinlan, 2014), discarding overlapping intervals less than 300 bp. The resulting genomic coordinates of overlapping sense-antisense genes were used as input in a modified version of REDItools (Picardi and Pesole, 2013), able to split reads according to their orientation. Only editing candidates supported by more than five reads and organized in non-redundant clusters (represented by A-to-G or T-to-C mismatches according to gene strandness) were retained. All the editing sites considered in this analysis are described in Supplementary Material S1.

Circular heatmaps were generated using the R package circlize (Gu et al., 2014) and the cytoband representation of the human genome assembly hg38. Heatmaps color scale represents an RPKM-like normalization of editing events.

The entire pipeline and scripts are available at https://github.com/BioinfoUNIBA/antisenseEditing.

For editing site validation, PCRs were performed on genomic DNA (gDNA) and RNA. gDNA was extracted using the High Pure PCR Template Preparation kit (Roche, Cat# 11796828001) following manufacturer instructions. PCR amplification was performed using Q5^® High-Fidelity DNA Polymerase (NEB, Cat# M0491). RNA was extracted using the RNeasy Plus Mini kit (Qiagen, Cat# 74134) and treated with DNase (Invitrogen, Cat# AM 1907). Following RNA extraction, RT-PCRs were performed with gene-specific primers (Supplementary Table S1) and a One-step RT-PCR kit (Qiagen, Cat# 210212). All the PCR products were purified (Macherey-Nagel, Cat# 740609) and analyzed by Sanger sequencing. Quantification of editing was performed directly from the Sanger traces using MultiEditR (Kluesner et al., 2021). Alternatively, the PCR products were cloned using a CloneJET PCR cloning kit (Thermo Scientific, Cat# K1232) according to the manufacturer’s instructions and transformed into DH5a bacteria (NEB, Cat# C2987). Ten to twenty resultant bacteria colonies were sent for sequencing to determine edits and their frequency in the targeted region. All the primers used in this study were designed using Primer-BLAST (Ye et al., 2012), AmplifX 2.0.7 (by Nicolas Jullien; Aix-Marseille Univ, CNRS, INP, Inst Neurophysiopathol, Marseille, France—https://inp.univ-amu.fr/en/amplifx-manage-test-and-design-your-primers-for-pcr) or ApE (by M. Wayne Davis, https://jorgensen.biology.utah.edu/wayned/ape/). The chromosomal locations of all the editing sites analyzed in this study are listed in Supplementary Table S2.

RNA was extracted using the RNeasy Plus Mini kit (Qiagen, Cat# 74134). Before qPCR, RNA was additionally treated with DNase (Invitrogen, Cat# AM 1907). cDNA synthesis was then performed using ProtoScript M-MuLV First-Strand Synthesis Kit (NEB, Cat# E6300) using 1 µg of RNA DNAse digested. cDNA was synthesized using oligo-dT or random primers to detect DDX58 or hLincRNA-p21, respectively. Two microliters of a 1:2 diluted cDNA were used to set up a 10 µl qPCR reaction using SsoAdvanced Universal SYBR Green Supermix (Bio-rad, Cat# 1725270). Finally, fold change expression was calculated using the comparative CT method (ΔΔCT) (Livak and Schmittgen, 2001). Supplementary Table S1 lists all the primers used in this study.

Lenti-CRISPR-ADAR1 or lenti-CRISPR-NT, in combination with pCMV-DR8.91 and pMD2.G, were calcium-phosphate transfected into HEK293T cells for lentiviral particle production (ratio 3:1:3). After 48–72 h, cell-free supernatant was collected and used for transduction of HEK293T cells. The transduced cells were selected with puromycin (1 μg/ml). As soon as non-transduced cells died (∼2 days), ADAR1 knockout cells were seeded in 96-well plates in a limiting dilution (0.5 cells/well). Upon expansion of single clones, ADAR1 KO clones were validated by Western blot (Cell Signaling Technology, Cat# 14175, RRID: AB_2722520) following IFN-α stimulation using β-Actin as a loading control (Sigma-Aldrich, Cat# A5441, RRID: AB_476744). Lenti-NT control cells were kept polyclonal. After screening, clones three and four were shown to completely abolish ADAR1 (p110 and p150) expression (Supplementary Figure S7B). Therefore, clone three was used for the experiments conducted in this work.

pSpCas9(BB)-2A-GFP carrying the sgRNAs for DTWD1, DTWD2 or TSR3 were transfected into HEK293T cells using Lipofectamine 2000 (ThermoFisher, Cat# 11668019) following manufacturer instructions. 48 h after transfection, GFP-positive cells were sorted and plated (one cell per well) in 96-well plates. The clonality was validated by visual inspection with a microscope, and positive clones were screened by Sanger sequencing.

Data were analyzed and plotted using GraphPad Prism (version 9.3.1). Specific information about data presentation is provided in each figure caption throughout the manuscript. Statistical significance was calculated by unpaired, two-tailed Student’s t-test: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant.

RNA-seq data analysis represents a powerful method to detect new RNA editing sites. Unfortunately, these technologies are not error-free; thus, validation of these newly discovered RNA editing sites is still necessary. This validation is performed via PCR amplification of a specific region containing the editing sites to be validated from either DNA or cDNA (the latter represents the RNA). In a recent work from our lab, we identified RNA editing sites comparing RNA- and DNA-seq data in a cohort of Diffuse large B cell lymphoma (DLBCL) patients (Pecori et al., 2021). We used a one-step RT-PCR reaction to validate some of those sites due to its higher sensitivity in RNA editing detection for low edited or expressed transcripts (Wacker and Godard, 2005; Kluesner et al., 2021). While validating some A-to-I editing sites within the transcript DDX58 in RCK8, a B cell lymphoma-derived cell line, we also observed the presence of numerous putative U-to-C edits. In a short region (∼600 nucleotides) of the 3′ untranslated region (3′UTR) of DDX58 we could detect 11 A-to-I sites and 11 U-to-C sites (Figure 1A, upper). A-to-I and U-to-C RNA editing events are observed as A-to-G and T-to-C in cDNA. Despite all the detections and quantifications being done on cDNA, throughout this manuscript, we refer to them as A-to-I and U-to-C base changes.

All those edits are visible in Sanger sequencing following amplification of cDNA but not genomic DNA (gDNA), validating them as real RNA editing sites (Figure 1A, lower). While U-to-C editing is well described in plants (Yoshinaga et al., 1996; Knie et al., 2016; Ruchika et al., 2021), it has been rarely described in Metazoans (Villegas et al., 2002; Liu et al., 2004); thus, we decided to investigate further this preliminary observation. We then performed the same validation on another three cell lines, namely U2932, HEK293T, and A549. Except for A549, we confirmed the observation of putative U-to-C base changes at the same precise sites identified in RCK8 (Figure 1B). To check a possible functional connection between the A-to-I and U-to-C editing, we quantified the frequency of U-to-C and A-to-I at all sites for all the cell lines (Figures 1C, D). No specific trend was observed, with different cell lines showing variations in the level of both editing types. Altogether these findings demonstrate the presence of an apparent and persistent U-to-C RNA editing in DDX58 mRNA. This editing can be found at the exact locations in different cell lines, and it seems independent of A-to-I editing. Indeed, the A549 cell line shows high A-to-I editing within DDX58 but no U-to-C editing.

It is known that RNA modification in general and RNA editing specifically are crucial during the cellular response to environmental stimuli or stress (Roundtree et al., 2017; Tan et al., 2017). To test if the U-to-C editing observed in DDX58 would change after specific stimulation, we decided to treat HEK293T cells with interferon-alpha (IFN⍺). IFN⍺ treatment has two relevant consequences for this experiment: first, it induces ADAR1 p150 expression (Patterson et al., 1995), which leads to an increase in A-to-I RNA editing (Hartwig et al., 2004, 2006); second, it leads to the overexpression of DDX58, which is an interferon-stimulated gene (ISG) itself (Matsumiya and Stafforini, 2010). HEK293T cells were chosen for this experiment because of the high level of U-to-C editing observed within DDX58 and because they are responsive to IFN⍺ stimulation (Figures 1B, C and Supplementary Figure S1). Following stimulation, RNA extraction and one-step RT-PCR were performed. PCR products were introduced into bacteria, and single bacterial colonies were sequenced using Sanger sequencing. Alignment to the unedited reference genome allowed us to easily count the editing sites in the presence or absence of stimulation to assess the frequency of A-to-I and U-to-C editing for each site in the two conditions (Figure 2A and Supplementary Figure S2). Not surprisingly, IFN⍺ stimulation leads to a significant increase in A-to-I editing (∼4-fold increase of the mean, Figures 2B, D) which is expected due to the induction of ADAR1 p150 expression (Patterson et al., 1995; Hartwig et al., 2004, 2006). However, the opposite effect was observed for U-to-C editing, for which the treatment led to a significant decrease (∼5-fold decrease of the mean, Figures 2B, C). These data suggest that putative U-to-C base changes are differently regulated compared to ADAR-induced A-to-I editing.

After characterizing the U-to-C editing within DDX58 mRNA, we asked if this editing was also present in other RNA species, such as long non-coding RNAs (lncRNAs). LincRNA-p21 is a crucial molecule during the response to cellular stress driven by p53 (Huarte et al., 2010). While initially discovered in mice, LincRNA-p21 is also present in humans (hLincRNA-p21, formally known as TP53COR1). Recent work has shown that hLincRNA-p21 contains inverted-repeat Alu elements, which can fold as independent domains (Chillón and Pyle, 2016). Interestingly, putative U-to-C editing events were identified in both sense and antisense Alu elements (Chillón and Pyle, 2016).

For this reason, we decided to first transfect in HEK293T a plasmid encoding hLincRNA-p21 and then treat the transfected cells with doxorubicin, a chemotherapeutic drug that induces DNA damage. We then performed RNA extraction and one-step RT-PCR to amplify the sense Alu. Detection and quantification of editing were performed as described above for DDX58 mRNA (Figure 3A). While doxorubicin treatment was shown to upregulate hLincRNA-p21 in some cell lines (Chillón and Pyle, 2016), we did not observe any significant changes in the expression of the endogenous, or in the stability of the exogenous, hLincRNA-p21 in HEK293T upon treatment (Supplementary Figure S3). In the absence of treatment, we observed only nine U-to-C and four A-to-I edits with editing frequency lower than 0.2 within the sense Alu (Figures 3B–D and Supplementary Figure S4). However, induction of DNA damage by doxorubicin leads to a significant increase in both editing types (∼2.5- and ∼16-fold increase of the mean for U-to-C and A-to-I, respectively; Figures 3B–D). Notably, we observed a substantial increase in the number of low-frequency (<0.1) edits that were not visible in the absence of stimulation (Figure 3B and Supplementary Figure S4). The increase of A-to-I editing upon DNA damage may be explained by recent findings showing an overall change in ADAR editing in response to DNA breaks (Jimeno et al., 2021). These data confirm the previous observation that putative U-to-C editing can also be identified in lncRNAs.

We then decided to look for the enzyme responsible for generating this U-to-C RNA editing. Few RNA modifications of uridines have been described to lead to a U-to-C base change. 4-thiouridine (s4U) itself leads to low levels of U-to-C transitions after reverse transcription (Hafner et al., 2010), and this level can be increased by chemical treatments of the RNA [reviewed in (Duffy et al., 2019)]. Indeed, s4U is often used in methods to study RNA metabolism because its presence can be easily detected via sequencing (Herzog et al., 2017; Schofield et al., 2018). Unfortunately, while s4U is present in bacterial and archaeal tRNAs, it has not been described in human tRNA (Boccaletto et al., 2018). It thus is very unlikely to be related to the U-to-C editing described here. In contrast, the 3-amino-3-carboxypropylation of uridine has been recently described in humans (Takakura et al., 2019). This modification leads to the formation of a 3-(3-amino-3-carboxypropyl) uridine (acp3U), which can be observed as an apparent U-to-C conversion caused by misincorporation during cDNA synthesis (Takakura et al., 2019; Kimura et al., 2020). Additionally, amino-carboxypropylation of methylated pseudouridine (ψ) has been described in rRNA in humans (Meyer et al., 2016). This m1acp3ψ modification perturbs standard base pairing during cDNA synthesis leading to U-to-C conversion (Babaian et al., 2020). Therefore, we decided to knock out the writers of these modifications, namely DTWD1, DTWD2, and TSR3, in HEK293T cells, as previously described (Takakura et al., 2019; Babaian et al., 2020). We successfully obtained knockout cell lines for those proteins. However, we did not observe any changes in U-to-C editing within DDX58 (Supplementary Figure S5).

A-to-I RNA editing has also been reported on antisense RNA, with some studies proposing that 15% of editing originated from transcripts expressed from the antisense strand (Porath et al., 2014). Widespread antisense transcription has been reported in humans, with 5%–10% of all genomic loci expressing overlapping sense and antisense RNAs (Lehner et al., 2002; Shendure and Church, 2002; Yelin et al., 2003). Overlapping sense and antisense RNAs often form structured motifs characterized by the presence of double-stranded hairpins that can act as substrates for ADAR (Supplementary Figure S6A). Using one-step RT-PCR methods with target-specific primers, the cDNA will be synthesized from both the sense and the antisense RNA. Following Sanger sequencing, A-to-I antisense RNA editing may result in an apparent U-to-C base (Supplementary Figure S6B). Therefore, we explored the possibility that previously-uncharacterized transcripts are expressed in antisense orientation to DDX58 and hLincRNA-p21 and modified by ADAR through A-to-I editing. To answer this question, we selected RNA samples from the two experimental conditions, which showed the majority of putative U-to-C editing in DDX58 and hLincRNA-p21, namely IFN⁻ and doxorubicin⁺, respectively (Figures 2, 3). On these samples, we performed in parallel three different one-step RT-PCR, providing both forward (F) and reverse (R) primers, or only the F or only the R primer, during the cDNA synthesis step (Figure 4). In this way, we obtained strand-specific amplification, with F and R primers generating cDNA specifically from the antisense and sense RNA, respectively (Supplementary Figure S6B). Both, DDX58 and hLincRNA-p21 showed abundant amplification from the antisense RNA on an agarose gel (Figure 4B). Strikingly, antisense-specific amplification resulted in high detection of putative U-to-C and no detection of A-to-I. The opposite was observed following amplification of the sense RNA for both DDX58 and hLincRNA-p21 (Figure 4C). Additionally, standard one-step RT-PCR from a HEK293T ADAR1 KO cell line (Supplementary Figure S7) resulted in no U-to-C or A-to-I editing detected in DDX58 and only a very low residual editing in hLincRNA-p21 (Figure 4C). Our data demonstrate that the putative U-to-C editing results from A-to-I editing on the antisense RNA indicating high editing activity by ADAR1 on both sense and antisense DDX58 and hLincRNA-21.

After observing antisense A-to-I RNA editing in both an mRNA and a lincRNA, we asked what the impact of this process at the transcriptome level in different tissues is. To elucidate this point, we investigated antisense A-to-I RNA editing genome-wide by using seven ribo-depleted strand-oriented RNA-seq experiments from various human tissues of the “RNA Atlas” project (Lorenzi et al., 2021). We created a catalog of sense-antisense gene overlaps based on Gencode annotations to provide an unbiased overview of antisense editing. Known antisense transcripts were initially selected from Gencode, then overlapping regions of at least 300 bp with sense transcripts were collected to a total number of 1677 suitable overlaps. For each one of these, corresponding to a well-defined genomic interval, we called RNA editing using pre-aligned RNA-seq reads and a modified version of the REDItools software (Picardi and Pesole, 2013) able to split reads according to their orientation. A-to-I RNA editing events supported by at least five reads and organized in clusters of A-to-G or T-to-C mismatches were selected for downstream analyses.

On the whole, we observed that the number of A-to-I editing changes, normalized by the overlap length, was higher in the sense strands of overlaps than in antisense strands, and this trend was common to all analyzed samples and tissues, supporting the previous notion that antisense editing is low Figure 5 and Supplementary Figure S8; and (Neeman et al., 2005). On average, only 199 out of 1677 potential overlaps showed evidence of A-to-I RNA editing. Of these, 21 displayed obvious sense and antisense editing, 164 sense editing, and 35 antisense editing only.

However, DDX58 was not among the transcripts identified by our approach, suggesting its limitations. Namely, antisense transcripts might be less abundant, leading to lower read depth (and problems in detecting editing); alternatively, naturally poor editing on some transcripts might be reported as “no editing” (depending on cut-offs). Both of these are limitations to our approach. These limitations could be cell type-specific or disease-specific. Overall though, our work suggests that clusters of A-to-I (and U-to-C) editing might specify dually edited, convergently transcribed regions, offering a potentially simple way to identify loci that may be of disease relevance (Li et al., 2022).