Extensive editing of cellular and viral double-stranded RNA structures accounts for innate immunity suppression and the proviral activity of ADAR1p150

The interferon (IFN)-mediated innate immune response is the first line of defense against viruses. However, an IFN-stimulated gene, the adenosine deaminase acting on RNA 1 (ADAR1), favors the replication of several viruses. ADAR1 binds double-stranded RNA and converts adenosine to inosine by deamination. This form of editing makes duplex RNA unstable, thereby preventing IFN induction. To better understand how ADAR1 works at the cellular level, we generated cell lines that express exclusively either the IFN-inducible, cytoplasmic isoform ADAR1p150, the constitutively expressed nuclear isoform ADAR1p110, or no isoform. By comparing the transcriptome of these cell lines, we identified more than 150 polymerase II transcripts that are extensively edited, and we attributed most editing events to ADAR1p150. Editing is focused on inverted transposable elements, located mainly within introns and untranslated regions, and predicted to form duplex RNA structures. Editing of these elements occurs also in primary human samples, and there is evidence for cross-species evolutionary conservation of editing patterns in primates and, to a lesser extent, in rodents. Whereas ADAR1p150 rarely edits tightly encapsidated standard measles virus (MeV) genomes, it efficiently edits genomes with inverted repeats accidentally generated by a mutant MeV. We also show that immune activation occurs in fully ADAR1-deficient (ADAR1KO) cells, restricting virus growth, and that complementation of these cells with ADAR1p150 rescues virus growth and suppresses innate immunity activation. Finally, by knocking out either protein kinase R (PKR) or mitochondrial antiviral signaling protein (MAVS)—another protein controlling the response to duplex RNA—in ADAR1KO cells, we show that PKR activation elicits a stronger antiviral response. Thus, ADAR1 prevents innate immunity activation by cellular transcripts that include extensive duplex RNA structures. The trade-off is that viruses take advantage of ADAR1 to elude innate immunity control.


Introduction
The innate immune response is the first line of defense against viruses [1]. This response, which must tolerate self, is based on the concerted action of interferon (IFN)-stimulated gene (ISG) products. Yet one of these, the adenosine deaminase acting on RNA 1 (ADAR1), has a key role in suppressing IFN signaling [2]. Here, we seek to characterize how ADAR1 functions. ADARs convert adenosine residues (C6 position) to inosine in double-stranded RNA (dsRNA), a process known as A-to-I editing [3,4]. There are three mammalian ADAR genes, but only ADAR1 and ADAR2 proteins edit RNA in vitro [5]. ADAR2 modifies the coding capacity of specific transcripts and the biological function of the corresponding proteins [5]. ADAR1 editing is less targeted and very extensive in many tissues, as revealed by next-generation sequencing [6,7]. However, the significance of this massive editing is still largely unexplored [8].
Indeed, ADAR1 can be proviral: it favors the replication of positive-strand RNA viruses such as yellow fever virus, Venezuelan equine encephalitis virus, and Chikungunya virus [21] and of negative-strand RNA viruses including measles virus (MeV) [22,23]. On the other hand, ADAR1 can damage viral RNA genomes by introducing large clusters of mutations, read as A-to-G (A>G) or U-to-C (U>C), depending on the strand edited [24,25].
We have recently shown that extensive ADAR1 editing occurs in defective interfering (DI) MeV RNAs generated during replication of a mutant MeV unable to express C protein (MeV-C KO ) [22,26]. These DI RNAs can form panhandle duplex RNA structures if not properly encapsidated. Since C protein controls viral polymerase fidelity, MeV-C KO generates elevated levels of DI RNA and induces strong innate immune responses [22,[26][27][28][29]. These responses involve activation of protein kinase R (PKR), which leads to translational arrest [28] and formation of stress granules [22,30].
ADAR1, on the other hand, interferes with the immune activation by viruses [23,30,31]. Here, we take advantage of two recombinant MeVs, C KO (GFP) and its isogenic parental virus vac2(GFP), and of a newly generated set of HeLa cells expressing different ADAR1 isoforms to characterize the endogenous and viral duplex RNA that activate innate immunity.

Innate immunity is activated in ADAR1-defective cells
To better characterize the mechanisms of action of the two isoforms, we targeted exon 2 of the ADAR locus (S1A Fig) and generated selectively ADAR1 p150 -deficient (p150 KO ) and fully ADAR1-deficient (ADAR1 KO ) HeLa cells. Multiple independent clones were recovered for each cell line and analyzed by western blot (S1B Fig). Genetic alterations causing the knock-out were deduced from RNA sequencing (RNAseq) data for 2 clones (p150 KO -B13 and ADAR1 KO -E7) (S1C Fig). In addition, we complemented ADAR1 KO cells with lentiviral vectors expressing wild-type ADAR1 p150 (p150wt LV ) or catalytically inactive ADAR1 p150 (p150mut LV ) [30,32]. We confirmed that both proteins had the expected cytoplasmic localization (S1D Fig). We also verified that knock-out of ADAR1 had no effect on the expression of ADAR2, which was predominantly found in nuclear extracts as expected (S1E Fig).
To assess whether ADAR1 deficiency affects cell viability or division rate, we performed a time course experiment comparing parental HeLa cells with p150 KO Fig, dashed lines). p150 KO and ADAR1 KO cells showed no difference to HeLa cells at 0 and 24 h but had more cells with lagging division rates at later time points, which was most pronounced in apoptotic and dead cell populations. ADAR1 KO cells and, to a lesser extent, p150 KO  We also asked whether innate immunity is activated in ADAR1-deficient cells. Indeed, treatment with recombinant type-I IFN-alpha (IFN A/D) resulted in stronger PKR activation in p150 KO and ADAR1 KO cells compared to parental HeLa or p150wt LV cells (Fig 1A). Since PKR is activated upon dsRNA binding, we think that endogenous transcripts forming dsRNA structures cause this activation. Catalytically inactive ADAR1 p150 did not fully suppress PKR activation ( Fig 1A). Thus, ADAR1 p150 but not ADAR1 p110 can interfere with recognition of endogenous dsRNA by PKR.

Differential editing of nuclear and cytoplasmic RNA by the two ADAR1 isoforms
To gain insights on the cellular transcripts that may activate PKR if left unedited by ADAR1, we used deep sequencing to characterize the total transcriptomes of HeLa, p150 KO , and ADAR1 KO cells. To identify potential editing sites, we adopted the Genomeindependent Identification of RNA Editing by Mutual Information (GIREMI) method (S3A Fig) [33]. As expected, we detected reduced frequencies of A>G and U>C transitions in ADAR1 KO cells (26,334 A>G sites) compared to HeLa cells (35,403 A>G sites), whereas the ratios of C>U and G>A transitions and of all transversions were unchanged (S3B Fig and S1 Data). Symmetrically, A>G and U>C transitions were increased about 1.4-and 1.15-fold, respectively, in HeLa cells compared to ADAR1 KO cells ( Fig 1B, S3C  Fig, and S1 Data). This is consistent with the enzymatic activity of ADAR1, which results in A>G and U>C transitions, depending on the strand analyzed. More than half of editing sites were located in intronic sequences, whereas exons and untranslated regions (UTRs) accounted for about 25% of A>G events (S3D Fig and S1 Data). This ratio remained unchanged in p150 KO and ADAR1 KO cells despite the overall reduction of the number of A>G events in ADAR1 KO cells, reflecting a high fraction of A>G events detected by GIREMI in the "junk DNA" genome fragments.
Since other studies indicate that ADAR1 editing preferentially occurs in Alu elements [6,34,35], we validated our approach by testing this correlation. We found that about 25% of all A>G transitions in HeLa cells are associated with retrotransposable elements. In particular, Alu elements formed the largest fraction of edited elements (over 75%), followed by long interspersed nuclear element (LINE) L1 elements and 7SL RNA (S3E Fig and S1 Data), which is consistent with previous analyses [6,34,35].
We then asked how the sites identified by GIREMI may get edited by ADAR1 within individual transcripts. We identified the most-edited transcripts based on four inclusion criteria (see Methods section) (S1 Table). In HeLa cells, within the top 156 transcripts, half of the editing sites were in introns, and the others in were in exons and UTRs (S4A Fig, left column and S1 Data). With loss of the ADAR1 p150 isoform (p150 KO ), remaining editing was more prevalent in introns (S4A Fig, middle column and S1 Data). This is consistent with intron editing by nuclear ADAR1 p110 , whereas ADAR1 150 editing occurs mainly in exons and UTRs. On the other hand, there was no preferential editing of specific transposable elements by either ADAR1 isoform (S4B Fig and S1 Data).
We also noted that the highest-ranking genes in our data set were predominantly edited by ADAR1 p150 (Fig 1C, blue dots), and editing mostly occurred in exons/UTRs (Fig 1C, light blue shading). In contrast, lower-ranking genes were equally targeted by both ADAR1 isoforms (Fig 1C, blue and red dots), and editing occurred at higher frequencies in intronic regions ( Fig  1C, orange shading). These data are consistent with ADAR1 p150 being mainly responsible for editing of cytoplasmic transcripts.

ADAR1 editing patterns of HeLa cells and primary tissues are consistent
We then refined analyses of editing within individual transcripts by constructing coverage plots and comparing their ADAR1 editing levels in HeLa, p150 KO , and ADAR1 KO cell lines. Fig 1D shows coverage plots of the 3 0 UTR of the VOPP1 (vesicular, overexpressed in cancer, prosurvival protein 1) transcript, which had the highest A>G transition differential. Whereas GIREMI detected 58 editing sites in a 6-kb region, the coverage plot was more sensitive, detecting 209 U>C transitions with >10% conversion rate ( Fig 1D, red box). In a further analytical refinement, we developed a method to compensate for sequencing mistakes, which returned a positive editing score for 388 sites in this region ( Fig 1E, top diagram). The edited region overlaps with two inverted LINE elements (Fig 1E, bottom). These inverted elements are predicted to form a nearly 3-kbp duplex secondary structure (not shown). Editing score analyses of p150 KO  To address the biological relevance of our HeLa cell-based observations, we repeated the GIREMI analyses with RNAseq data sets from human donors. Since HeLa cells are derived from cervical carcinoma [36], we repeated GIREMI analyses with data sets derived from primary cervical stromal cells (CSCs; Gene Expression Omnibus [GEO]: GSE99392) [37] (S1 Table and Table). However, the list of edited transcripts derived from primary data sets largely overlapped with the list derived from HeLa cells, and the affected regions were identical (S1 Table). Next, we analyzed primary human fibroblast RNAseq data of a healthy individual (CTRL_1) and several patients with ADAR1-sufficient AGS (GEO: GSE57353) [38]. Similarly as with the other data sets, GIREMI identified editing in the same transcripts as in HeLa cells (S1 Table and  Editing patterns from individuals were similar but not identical: more than 100 editing sites defined in HeLa cells were found in all five human samples, whereas more than 200 were found in 1-4 samples (S5K and S5L Fig and S1 Data).
We selected 8 transcripts from top, center, and bottom regions of S1 Table and compared editing frequencies in HeLa cells with those of primary human data sets (Fig 1G). These analyses indicate that, although differences between individuals exist, editing frequencies for these transcripts remain similar in the human population. Altogether, ADAR1 editing in HeLa cells generally compares well with editing in primary human samples, but individual editing patterns can differ considerably. GIREMI in HeLa, p150 KO , and ADAR1 KO cells. Non-ADAR1, percentage of editing sites remaining in ADAR1 KO cells. (D) Coverage plot of the 3 0 UTR region of VOPP1. The region in the red box shows strong enrichment of U>C transitions. Nucleotide positions with variant frequencies �10% are color-coded: A, green; C, blue; G, orange; U, red. (E) Editing score analysis of the boxed region of (D) in HeLa, p150 KO , ADAR1 KO , and HeLa cells treated with 1,000 U/ml IFN A/D (top to bottom). Gray line indicates total coverage ("Cov.") in the region. Repetitive sequences are indicated at the bottom: positive sense in blue and negative sense in red. (F) Correlation of editing scores of the VOPP1-region in HeLa + IFN A/D against untreated HeLa cells (gray dots, black line), HeLa versus p150 KO (red dots and line), and HeLa versus ADAR1 KO (blue dots and line). (G) Relative editing rates in cell lines and primary RNAseq data. Editing rates are normalized to coverage and length of the analyzed regions. Editing rate of each gene in HeLa cells is set to 100%. ADAR1, adenosine deaminase acting on RNA 1; ADAR1 KO , fully ADAR1-deficient; GIREMI, Genome-independent Identification of RNA Editing by Mutual Information; IFN A/D, recombinant type-I interferon-alpha; lin. regr., linear regression; N/A, no RNAseq data available because of low or no coverage; p150 KO , selectively ADAR1 p150 -deficient; p150mut LV , catalytically inactive ADAR1 p150 ; p150wt LV , wild-type ADAR1 p150 ; PKR, protein kinase R; pPKR, phospho-PKR; RNAseq, RNA sequencing; UTR, untranslated region; VOPP1, vesicular, overexpressed in cancer, prosurvival protein 1.

ADAR1 edits short interspersed nuclear elements (SINEs) in primates and rodents
Having identified inverted repetitive elements as the primary target within ADAR1-edited transcripts, we asked whether homologous transcripts of different species include repetitive elements potentially forming duplex RNA structures. Towards this, we analyzed data sets from rhesus macaques (Macaca mulatta) [39] and mice (Mus musculus) [14]. Fig 2A illustrates one example of evolutionarily conserved ADAR1 editing. The 3 0 UTR of the human NADH:ubiquinone oxidoreductase core subunit S1 (NDUFS1) transcript includes 20 transposable elements in a complex arrangement (Fig 2B, top half). Eleven of these elements are extensively edited by ADAR1, as indicated by the local concentration of many high editing score positions (Fig 2A). All these high-editing regions are predicted to form duplex structures (S6A Fig).
In the corresponding macaque transcript, the repetitive elements arrangement is simpler, but three groups of Alu repeats are conserved (Fig 2B, bottom half). Editing in these elements is conserved between humans (Fig 2A) and macaques ( Fig 2C). Other transcripts, including APOOL, GNL3L, TIAL1, and EXOSC2, were similarly edited in human and macaque samples.

Almost 1% of cellular transcripts are extensively ADAR1 edited
Having identified more than 150 ADAR1-edited transcripts, we sought to estimate their expression levels and thus the amount of dsRNA present in a cell. For this, we relied on RNAseq-based transcript quantification in HeLa, p150 KO , and ADAR1 KO cells, using four conditions (uninfected, vac2[GFP]-infected, C KO [GFP]-infected, and IFN A/D-treated) for each cell line ( Fig 3A and S8A Fig).
We determined values for the fragments per kilobase of transcript per million mapped reads (FPKM) of all annotated transcripts in each data set. Of over 28,000 annotated genes, 15,000 constituted more than 99% of transcript-associated fragments of each sample and were included in the downstream analysis (S8A Fig). We next ranked the genes by expression levels in uninfected HeLa cells (S8B Fig, black diamonds). Relative expression levels were elevated in both p150 KO (S8B Fig, green diamonds) and ADAR1 KO cells (S8B Fig, orange diamonds). We then assessed the expression levels of our 156 ranked genes, which mostly had FPKM values at intermediate to low levels ( Fig 3A). Infection or IFN treatment had little to no effect on the expression levels of these genes (compare lanes 1 with 2-4, 5 with 6-8, and 9 with 10-12). The differences in expression of ADAR1 isoforms also had no significant effect on the expression of these genes ( Fig 3B).
The added expression values of the 156 ADAR1-edited transcripts constituted about 1% of the total cellular transcripts ( Fig 3C). However, only a fraction of the transcribed RNA will  [39]. ADAR1, adenosine deaminase acting on RNA 1; ADAR1 KO  actually enter the cytoplasm, since editing frequently occurs in introns (S4A Fig) and UTRs at positions downstream of the annotated polyadenylation site (Fig 3D, blue dashed line). We noticed that for many of these transcripts, only about 10%-30% had elongated UTRs containing the editing sites, whereas the majority of transcripts terminated at the annotated polyadenylation site (most strikingly observed in the VOPP1 transcript). Considering these facts, we  [34] (blue, cyan, magenta, white), and Ahmad and colleagues [42] (red, magenta, yellow, white). Each transcript is represented by a single tile. The total numbers in each group (unique, shared by two or three independent studies) are indicated in the legend to the lower right. ADAR1, adenosine deaminase acting on RNA 1; ADAR1 KO  estimate that between 0.5% and 1% of cellular transcripts are ADAR1 edited, or about 1,000 to 2,000 mRNA copies per cell [43].

Not all ADAR1-edited transcripts are candidates for MDA-5 recognition
Two independent analyses of the ADAR1-edited transcripts in HEK-293T cells [34] and of Alu-dependent association of transcripts with MDA-5 in HEK-293T cells [42] were recently published. As in our analyses, lists of transcript targets were generated. We asked how much overlap there is between the three studies. For this, we compared our top 156 ADAR1-edited transcripts with 100 MDA-5-associated transcripts of Ahmad and colleagues [42] and the top 200 hits of Chung and colleagues [34].
Our analyses and those of Chung and colleagues shared 67 transcripts, whereas only 19 of our transcripts were common with those of Ahmad and colleagues ( Fig 3E). The overlap between the Chung and Ahmad studies is 33 transcripts. All three studies identified the same 16 transcripts but ranked them differently (S8C Fig and S1 Data). Among these transcripts, only the nucleolar GTPase GNL3L was consistently within the top 12 and the X-pro-aminopeptidase XPNPEP3 consistently within the top 30 (S8C Fig and S1 Data). From these 3-way analyses, we conclude that not all ADAR1-edited transcripts are strong candidates for innate immunity activation through MDA-5 recognition. Which of the transcripts identified here are responsible for PKR and IFN regulatory transcription factor 3 (IRF3) activation remains unclear.
We infected the HeLa-derived cell lines with two reporter viruses, the vaccine-equivalent strain MeV-vac2(GFP) and its isogenic mutant MeV-C KO (GFP). During the first 24 h of infection, both viruses replicated to about 10 4 TCID 50 /ml (Fig 4A and 4B and S1 Data), reaching slightly lower titers in p150 KO and ADAR1 KO cells compared to HeLa cells. MeV-vac2(GFP) continued to replicate in HeLa cells for the next 48 h, but its replication was completely inhibited in p150 KO and ADAR1 KO cells at later time points (Fig 4A and S1 Data). MeV-C KO (GFP) replication efficiency was similar to that of MeV-vac2(GFP) for the first 24 h but thereafter stopped in all three cell lines ( Fig 4B and S1 Data). These observations can be explained as follows. Because of the lack of C protein expression, MeV-C KO (GFP) stocks contain large amounts of DI genomes (S9 Fig). These are amplified to high levels already during the initial phase of replication, interfering with the replication of full-length genomes and causing innate immune activation. In contrast, MeV-vac2(GFP) stocks contain minimal amounts of DI genomes (S9 Fig). Even if DI genomes were generated at late MeV-vac2(GFP) infection stages, innate immunity activation may have limited consequences [63]. Thus, ADAR1 deficiency preferentially impacts MeV-C KO (GFP) replication. Values are average ± standard deviation of n = 5 for each time point. For p150 KO cells, 3 replicates were generated on clone B13 and 2 replicates on clone C10. For ADAR1 KO cells, 3 replicates were generated on clone E7 and 2 replicates on clone E2. Significance was determined by unpaired two-tailed Student's t test and is indicated with asterisks ( � , P < 0.05; ��� , P < 0.0001). Underlying values can be found in S1 Data. (C) Absolute number of viral reads with >5 mutations ("NM>5") in RNAseq samples. Underlying values can be found in S1 Data. (D) Frequency of these reads relative to total number of MeV-specific reads. U>C: reads with predominantly U>C mutations (red); A>G: reads with predominantly A>G mutations (green). Underlying values can be found in S1 Data. (E) Editing scores of MeV-vac2(GFP) and MeV-C KO (GFP) genomes from HeLa, p150 KO , and ADAR1 KO infections. Scores are shown for transitions (A>G, green; U>C, red; G>A, orange; C>U, blue) and a read coverage (gray) of at least 10. (F and G) Correlation of (F) MeV-vac2(GFP) and (G) MeV-C KO (GFP) genome editing between HeLa and p150 KO cells (gray dots and black line) or HeLa and ADAR1 KO cells (red dots and line). ADAR1, adenosine deaminase acting on RNA 1; ADAR1 KO , fully ADAR1-deficient; KO, knock-out; lin. regr., linear regression; MeV, measles virus; MOI, multiplicity of infection; NT, nucleotide; p150 KO , selectively ADAR1 p150 -deficient; RNAseq, RNA sequencing.

ADAR1 p150 frequently edits defective genomes
We then asked how frequently ADAR1 edits MeV genomes. For this, we amplified both MeV-vac2(GFP) and MeV-C KO (GFP) on HeLa and ADAR1-modified cells, purified ribonucleocapsids (RNPs) (S10A Fig), and analyzed them by RNAseq. We obtained purity levels ranging from 92% to 11% (S10B Fig and S1 Data), with coverages of 10 3 to 10 5 reads per nucleotide (S10C Fig, gray areas). We extracted reads with at least 5 differences from the reference sequence (S10C Fig, colored areas), which were evenly distributed over the MeV-vac2(GFP) genome but accumulated on either MeV-C KO (GFP) genome end, consistent with amplification of DI genomes in these infections [26]. Many of these reads had sudden interruptions of collinearity with the MeV genome, probably reflecting recombination artifacts during library preparation ( Fig 4C and 4D, gray color and S1 Data). Reads with predominant A>G or U>C transitions were more abundant after replication in HeLa than in p150 KO and ADAR1 KO cells, consistent with expectations ( Fig 4C and S1 Data). Only about 1 in 3,000 reads of MeV-vac2 (GFP) genomes had ADAR1 mutations, whereas 1 in 500 reads of MeV-C KO (GFP) genomes were ADAR1 edited ( Fig 4D and S1 Data). The 2:1 ratio of U>C-to A>G-mutated reads reflects the ratio of negative-strand to positive-strand MeV genomic RNA in virus preparations [64]. Thus, although coverage of MeV-vac2(GFP) genomes with mutated reads was similarly high as that of MeV-C KO (GFP) genomes, U>C and A>G transitions were predominantly introduced into the MeV-C KO (GFP) genomes.
We next calculated editing scores for each nucleotide of the two viral genomes amplified in each cell line ( Fig 4E). A>G and U>C editing scores in p150 KO cells were strongly reduced compared to HeLa cells (Fig 4F and 4G) and nearly absent in ADAR1 KO cells. Residual A>G and U>C transitions in ADAR1 KO samples may be due to edited genomes and/or DI RNAs in the virus inocula, which were generated in ADAR1-expressing Vero cells (S11A Fig).
The MeV-C KO (GFP) genome was more accessible to ADAR1 than the MeV-vac2(GFP) genome. Over 30% of A residues and nearly 60% of U residues in MeV-C KO (GFP) showed editing scores of �0.05, whereas only 8% of A and 33% of U residues were converted at equal frequencies in MeV-vac2(GFP) (S11B Fig and S1 Data). Neither virus genome accumulated significant C>U or G>A transitions, which could have been indicative for apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC) activity [65]. The nucleotide sequences surrounding the edited sites conferred to the ADAR1-specific pattern previously described [26,66] (S11C and S11D Fig and S1 Data). Altogether, these data document that ADAR1 p150 is crucial for editing viral genomes and that it more frequently edits genomes of C KO than those of standard MeV.

Catalytically active ADAR p150 counteracts immunity activation and restores virus growth
To further characterize the role of ADAR1 p150 in the antiviral response, we complemented ADAR1 KO cells with either a catalytically active (p150wt) or inactive (p150mut) cytoplasmic isoform and assessed whether these proteins rescue MeV replication. As shown in Fig 5A and  5B, MeV-vac2(GFP) and MeV-C KO (GFP) growth kinetics were almost identical in HeLa (blue lines) and p150wt LV cells (orange lines) (see also S1 Data). Consistently, green fluorescent protein (GFP) expression in p150wt LV cells was at levels similar to those of standard HeLa cells (Fig 5C), but MeV-C KO (GFP) replication was still restricted in p150wt LV cells. Expression of the catalytically defective mutant resulted in intermediate complementation levels, as measured by growth kinetics (Fig 5A and 5B, purple lines) and GFP expression analyses (Fig 5C).
MeV-vac2(GFP) infection induces minimal levels of PKR and IRF3 phosphorylation in standard HeLa cells (Fig 5D), both antiviral pathways are strongly activated upon infection in p150 KO and ADAR1 KO cells. Moreover, standard ADAR1 p150 fully suppresses this activation, whereas mutant ADAR1 p150 only partially suppresses PKR and IRF3 phosphorylation. We observed a similar effect of the expression of different ADAR1 isoforms and mutants on the antiviral response to MeV-C KO (GFP) (Fig 5E). Thus, ADAR1 dsRNA binding and catalytic activity are both required to suppress PKR activation.
To assess whether activation of intrinsic immunity is directly responsible for virus growth inhibition, we knocked out ADAR1 in IFN-incompetent Vero cells [67] (Fig 6A). If so, replication of MeV-vac2(GFP) and MeV-C KO (GFP) is expected to reach similar levels in both Vero and Vero-ADAR1 KO cells. Indeed, this is the case, as monitored by GFP expression (S12 Fig) and western blot analyses of viral N and C protein expression (Fig 6B). PKR and IRF3 activation were similar in Vero-ADAR1 KO and parental Vero cells (Fig 6B). Accordingly, MeV-vac2 (GFP) replication was not significantly reduced in Vero-ADAR1 KO cells compared to Vero cells (Fig 6C and S1 Data). Growth of MeV-C KO (GFP) in Vero-ADAR1 KO cells was even slightly enhanced ( Fig 6D and S1 Data). Thus, MeV replication depends on the immune-regulatory effect of ADAR1 p150 editing of endogenous and viral dsRNA.

The PKR-mediated stress response controls MeV infection
To determine whether MeV infection of ADAR1 KO cells is controlled by the RLR-mediated IFN response, the PKR-mediated cellular stress response, or partially by either pathway, we sought to inactivate these responses. Towards this end, we generated ADAR1 KO -MAVS KO and ADAR1 KO -PKR KO cells (S13 Fig). We designed a clustered regularly interspaced short palindromic repeat (CRISPR)/CRISPR-associated 9 (Cas9) approach targeting functional fulllength MAVS (FL-MAVS) (S13A and S13C Fig) but not ΔMAVS lacking the essential caspase activation and recruitment domain (CARD) [68]. Similarly, we inactivated PKR through CRISPR/Cas9 targeting of parts of its gene encoding the RBMs (S13B and S13C Fig). As expected, ablation of FL-MAVS expression affected phosphorylation of IRF3 upon transfection of poly(I:C), whereas it did not affect phosphorylation of PKR (S13C and S13D Fig and S1 Data). Vice versa, deletion of PKR had no impact on IRF3 activation (S13C and S13D Fig and S1 Data).
In single-cycle infections of 5 independent clones, GFP expression of either MeV-vac2 (GFP) or of MeV-C KO (GFP) in ADAR1 KO -MAVS KO cells (S14C and S14G Fig

Cellular duplex RNAs: Origin and disposal
Consistently with two recent studies [34,42], we report here that one essential function of ADAR1 is to edit duplex RNA structures located in the 3 0 UTR of pol II transcripts. These duplex structures are formed by integrated inverted retrotransposable elements, most frequently Alu elements. We characterized more than 150 highly ADAR1-edited structures, whose prevalence accounts for the massive levels of A>I editing in human cells [8,69]. The editing patterns of HeLa cell transcripts are recapitulated not only in data sets from human donors but also in those from macaques, in which A>I editing occurs in Alu-lineage repeats selectively conserved among primates. In addition to these duplex structures in pol II transcripts, similar structures in noncoding transcripts, which are also ADAR1 substrates [5], may contribute to innate immunity activation by increasing the pool of transcripts with duplex RNA.
Reexamination of mouse ADAR1 editing data [14] reveals that a few transcripts are edited based on dsRNA structures conserved across mammalian orders. This was surprising because mice lack primate-specific Alu repeats. Their genomes have accumulated rodent-specific B1 elements instead. However, like Alu elements, B1 elements derive from 7SL RNA [41]. Thus,  ADAR1 edits self and viral RNA to suppress innate immunity ADAR1 editing may have originally targeted the same transposon class. On the other hand, it was reported that retro-elements can activate innate immunity and in particular that endogenous retroviruses can trigger IFN induction [70]. Although this mechanism is plausible, we note that inverted retro-elements embedded in pol II transcripts are more prevalent than bona fide endogenous retroviruses.
Although 67 of the 156 ADAR1-edited transcripts from our study are the same as those identified by Chung and colleagues using a similar ADAR1 gene knock-out approach [34], only 19 were the same as those characterized by Ahmad and colleagues as binding MDA-5 [42]. The simplest explanation for this is that ADAR1 edits more transcript than those MDA-5 would recognize. This would be consistent with our observation that the PKR-mediated cellular stress response may operate in infected cells in addition to the RLR-mediated IFN response.
The overlap of our data with those of Chung and colleagues was more extensive but incomplete. Since the two analyses were performed on different cell lines, it is possible that ADAR1 editing activity is cell type dependent. However, different methodologies applied to evaluate gene expression may account for most differences. Our comparison of HeLa cell-derived editing with editing of data sets from human donors supports this assumption. When transcripts were highly expressed in different samples, they exhibited editing patterns nearly identical to those of HeLa cells. However, editing could not be detected in transcripts expressed at low levels in certain samples. This reflects the limited sensitivity of the assay rather than a significant difference.

Frequency of RNA editing and collateral damage of viral genomes
Here, we took advantage of a vaccine-lineage MeV and an isogenic mutant generating excess DI RNA (C KO ) to measure the ADAR1 editing frequency. These DI RNAs have the ability to form panhandle dsRNA structures, which are similar to inverted Alu repeat stem-loops that are targeted by ADAR1. The amount of DI RNA generated during standard MeV infection is low but not zero. Consistently, we detected only about 1,500 edited reads in 32,000 genome equivalents-an average of 1 read per 21 genomes. To account for contiguous reads derived from the same editing event, we correct this number to 1 in 20-100 edited genomes. In C KO genomic RNA, the ratio of edited reads was 6 times higher. Since about 1,000 MeV genomes are produced per infected cell [64], in standard infections 10-50 genomes per average cell are edited, whereas in C KO infections, 60-300 genomes per cell are edited. In contrast, more than 1,000 endogenous transcripts per cell are edited. This suggests that even during infection with a defective virus, duplex structures of cellular origin may be more abundant than those of viral origin.
Whereas editing of DI RNA accounts for the proviral activity of ADAR1, editing of regular genomes would have antiviral properties [71] by inactivating essential gene products. However, editing of regular genomes is expected to be inefficient, since these genomes are fully encapsidated. Indeed, we rarely detected ADAR1 editing in parental MeV infections. Whereas this study has focused on editing of immunostimulatory RNA by ADAR1 and the regulation of innate immune responses, it will be worthwhile to investigate the long-term effects of ADAR1 editing activity on viral genome evolution and quasi-species distribution. Our Vero-ADAR1 KO cells, which allow efficient MeV replication in the absence of ADAR1, are a valuable tool for this purpose.

Editing-dependent and editing-independent components of innate immunity activation
Partial recovery of viral replication through overexpression of a catalytically inactive ADAR1 suggests that its immunoregulatory function has editing-dependent and editing-independent components. The editing-independent function may be due to dsRNA-binding competition with immune sensors such as PKR and MDA-5. Template competition is a mechanism of action shared by many dsRNA-binding proteins, including the influenza A virus NS1 [72] and the vaccinia virus E3L [73]. ADAR1 p110 has editing-independent functions, such as protection of mRNAs from Staufen-1-mediated decay [74].
For prevention of innate immunity activation, however, simply binding to dsRNA seems insufficient. Evidence for this is found in mice, in which a homozygous mutation E861A in ADAR1, disrupting catalytic activity, exhibits the same embryonically lethal phenotype [14] as full ADAR1 knock-out mice [75]. In addition, most ADAR1 mutations associated with AGS6 are found in the deaminase domain, while the RBMs remain functional [19]. Altogether, these observations and our data indicate that only catalytically active ADAR1 has full immunoregulatory and proviral function.

A model for the regulatory function of ADAR1 in autoimmunity and infection
In Fig 8, we present a model of the ADAR1 p150 mechanism of action consistent with our data, along with evidence provided by multiple studies in mice [2,[13][14][15][16][75][76][77], human cells [17,30,34], and human participants [19]. In normal cells, ADAR1 recognizes and alters dsRNA structures in constitutively expressed transcripts and thus prevents autoimmune activation of PKR and MDA-5 (Fig 8A, ADAR1 sufficient). Thus, ADAR1 activity allows the cell to tolerate a certain amount of endogenous duplex RNA, setting a threshold for immune activation. This threshold may vary with the expression levels of ADAR1 in different cell types. In an environment in which ADAR1 is missing or lacks catalytic activity, the threshold is very low (Fig 8A, ADAR1  deficient). Unedited transcripts are recognized by innate immune receptors and induce an autoimmune response. This prevents efficient replication of viruses. Indeed, a standard MeV generating small amounts of dsRNA replicates less efficiently in p150 KO or ADAR1 KO cells than in parental HeLa cells. Reduced replication of standard MeV in ADAR1-deficient cells can be monitored already 24 h post infection and becomes more pronounced at later infection stages. Thus, MeV replication is slowed down from the beginning and completely inhibited eventually.
Our model also accounts for the immune-activating properties of the C KO virus (Fig 8B and  8C). This virus generates DI RNA from the onset of replication, which adds on top of cellular dsRNA transcripts, at some point exceeding the threshold of efficient ADAR1 editing (Fig 8B,  right column). Viral DI RNA, partially edited or unedited, then triggers innate immunity. The width of the gap between the amount of cellular duplex RNA and activation threshold determines how much viral dsRNA can be tolerated by cells before innate immunity activation occurs. For most-effective pathogen detection, the gap should be narrow. Innate immunity activation can occur by more than one mechanism: we observed parallel activation of the PKR-mediated cellular stress response and RLR-mediated IFN induction in ADAR1-deficient cells.
In summary, ADAR1 sets a threshold for intrinsic immunity activation by cellular or viral duplex RNA. By adjusting the intrinsic immune activation threshold and protecting cells from translational shutdown, ADAR1 p150 provides cover for viruses, which take advantage of enhanced tolerance to duplex RNA accidentally generated during their replication.

Generation of ADAR1 knock-out cell lines
HeLa p150 KO and ADAR1 KO cell lines were generated by CRISPR/Cas9-nickase (Cas9n). For this, pairs of Cas9n cleavage sites in Exon 2 were identified using the ATUM online tool (https://www.atum.bio/eCommerce/cas9/input; ATUM, Newark, CA, United States). gRNAs upstream of M296 (ADAR1 p110 start codon) were selected for ADAR1 p150 -specific knock-out, and gRNAs downstream of M296 were selected for general ADAR1 knock-out. gRNAs (p150 KO  . Thus, no activation of innate immunity occurs. ADAR1 expression level sets the threshold for innate immunity activation (green line). In ADAR1-deficient cells, the threshold is decreased. Levels of transcribed duplex RNA remain equal, but duplexes are not edited and innate immunity is triggered (red arrow). (B) A standard RNA virus (e.g., MeV-vac2) generates low amounts of dsRNA (blue), which is efficiently edited by ADAR1 and thus insufficient to activate innate immunity. In contrast, an RNA virus with DI genomes (e.g., MeV-C KO ) generates large amounts of duplex RNA (blue and red). ADAR1 still edits some of it (blue), but unedited dsRNA activates innate immunity (red). (C) Schematic representation of the generation of immunogenic duplex RNA (panhandle structures) during viral infection and the impact of ADAR1 on PKR-and MDA-5-mediated innate immunity activation by these RNAs. ADAR1, adenosine deaminase acting on RNA 1; DI, defective interfering; dsRNA, double-stranded RNA; IFN, interferon; MDA-5, melanoma differentiation-associated gene 5; MeV-C KO , MeV unable to express C protein; PKR, protein kinase R.

Generation of lentivirus-transduced cell populations
HeLa ADAR1 KO cells (clone E7) seeded in 6-well plates were transduced with 100 μl puromycin-selectable lentiviral vector stocks expressing either wild-type ADAR 150 (p150wt LV ) or catalytically inactive ADAR1 p150 (H910Q/E912A) (p150mut LV ). Media were replaced 24 h post transduction with fresh DMEM with FBS and Pen/Strep. Cells were expanded into 60-mm dishes 72 h post transduction, and puromycin was added at a final concentration of 1 μg/ml at this time point. Three days later, cells were again trypsinized (Cat. #25-053-CI, Corning) and seeded into T75 flasks. Expression of ADAR1 p150 was confirmed by western blot analysis. Frozen cell stocks were generated as mixed populations with heterogenous ADAR1 p150 expression levels and kept in liquid nitrogen. Puromycin was applied to each cell passage and omitted only prior to experiments.

Virus strains
Recombinant vaccine lineage MeV-vac2(GFP) and MeV-C KO (GFP) expressing enhanced GFP from an additional transcription unit were described previously [26]. Generation of recombinant viruses, stock production, and titration were described previously [79]. Infections were carried out as follows: Cells were seeded 16 to 24 h prior to infection at 50% confluency. Cells were incubated with virus inoculums at indicated multiplicities of infection (MOIs) in low volumes of Opti-MEM (Cat. #31985070; Thermo Fisher Scientific, Waltham, MA, US) for 2 h at 37˚C, after which the inoculums were replaced with fresh DMEM with FBS and Pen/Strep. Cells were processed for downstream analyses at the indicated time points post infection.

IRB and IACUC statement
No human subjects or animals were directly involved in this study. Primary RNAseq data sets of human subjects, macaques, and mice were derived from the GEO database.

Cell growth and viability assay
Cells grown to confluency in T25 flasks were washed once with PBS and incubated with Versene (Cat. #15040066; Thermo Fisher Scientific) at 37˚C until they started detaching from the surface. Cells were suspended in 10 ml PBS and counted using a Neubauer chamber. A total of 3 × 10 6 cells were stained with CellTrace Violet (Cat. #C34571; Thermo Fisher Scientific) according to the manufacturer's instructions. Then, 3 × 10 5 stained cells were seeded into several 6-well plates with 2 ml growth medium and incubated for the indicated amounts of time

Virus growth curve analysis
Cells were seeded in 6-well plates at 50% confluency 16-24 h prior to infection. Infections were carried out at an MOI of 0.1. At indicated time points, supernatants were removed and cells were scraped into 100 μl Opti-MEM per well, followed by 3 freeze/thaw cycles (liquid nitrogen/37˚C). Cell debris was removed by centrifugation (4,000g, 4˚C, 10 min). Viral titers of cleared lysates were determined by infecting monolayers of Vero-hSLAM cells [81] with 10-fold dilutions of lysates for 96 h, as described previously [79]. Viral titers were determined using the TCID 50 method.

IFN treatment
Cells were seeded in 6-well plates at 50% confluency 16-24 h prior to treatment. IFN A/D (Cat. #11200-1; PBL Assay Science, Piscataway, NJ, US) was diluted in fresh DMEM with FBS and Pen/Strep at 1,000 U/ml and added to the cells for 24 h.

RNP purification and RNA extraction
HeLa cells were seeded in 150-mm dishes at 5 × 10 6 cells per dish 24 h prior to infection, and infections were carried out at an MOI of 0.1. For a typical RNP preparation, 10 dishes were infected and cells were harvested 72 h post infection. Cells were scraped into 5 ml PBS per dish and pelleted in 50-ml conical tubes by centrifugation at 300g and 4˚C for 10 min. Cell pellets were resuspended in 3 ml ice-cold Hypotonic buffer (10 mM HEPES [pH 7.5]; 10 mM NaCl; 1.5 mM MgCl 2 ) supplemented with 0.65% (v/v) NP-40 substitute and protease inhibitor cocktail (Cat. #11836153001; Roche, Basel, Switzerland) for 30 min on ice. Cell debris was pelleted by centrifugation at 4,000g and 4˚C for 15 min. The supernatant was then supplemented with 1% (w/v) sodium deoxycholate and 10 mM EDTA and spun a second time at 20,000g and 4˚C for 15 min. The lysate was loaded on top of a discontinuous CsCl gradient in SW41 polypropylene centrifuge tubes (Beckman Coulter, Brea, CA, US). CsCl solutions of different concentrations were prepared in gradient buffer (25 mM Tris [pH 7.5]; 50 mM NaCl; 2 mM EDTA; 0.2% (w/v) sodium lauroyl sarkosinate [sarkosyl]). The discontinuous gradient was prepared as follows (from bottom to top): 1 ml of 40% (w/v) CsCl; 2.5 ml of 30% (w/v) CsCl; 1.5 ml of 25% (w/v) CsCl; 1 ml of 5% (w/v) sucrose. Ultracentrifugation was carried out in a SW41 rotor in a LE-80 ultracentrifuge (Beckman Coulter, Brea, CA, US) at 36,000 rpm and 4˚C for 22 h. RNPs banded about 2 cm above the bottom of the tube and were harvested in approximately 1 ml volume by needle aspiration using a 16-gauge needle and syringe. RNPs were diluted in 8 ml LEH buffer (10 mM HEPES [pH 7.5]; 100 mM LiCl; 1 mM EDTA), layered over 2 ml of 15% (w/v) sucrose in LEH buffer in SW41 centrifuge tubes and centrifuged a second time at 36,000 rpm and 4˚C for 6 h. Afterwards, the supernatant was discarded, and RNP pellets were carefully resuspended in 1 ml LEH buffer supplemented with 1% (w/v) SDS. Total RNA was extracted from this solution using Trizol LS (Cat. #10296010; Thermo Fisher Scientific) and precipitated with isopropanol according to the manufacturer's instructions. Precipitated RNA was resuspended in 25 μl nuclease-free H 2 O and stored at −80˚C.

RNAseq library preparation and Illumina sequencing of viral RNPs
RNP RNA (5 μl) was fragmented for 7.5 min using the Ambion Fragmentation Reagent (Cat. #AM8740; Thermo Fisher Scientific) according to the manual. The samples were then diluted with nuclease-free H 2 O to a final volume of 100 μl, mixed with an equal volume of buffered phenol/chloroform/isoamyl alcohol (Cat. #15593031; Thermo Fisher Scientific) and phase-separated by centrifugation (12,000g, 4˚C, 10 min). RNA was precipitated from the aqueous phase using sodium acetate/ethanol overnight at −20˚C followed by centrifugation (20,000g, 4˚C, 30 min). Pellets were washed with 70% (w/v) ethanol, dried, and resuspended in 11 μl nuclease-free H 2 O. RNA libraries were prepared using 220-500 ng of total RNA according to the manufacturer's instructions for the Tru-

Total cell transcriptome RNAseq library preparation and sequencing
Total RNA of HeLa, p150 KO

Gene expression quantification of RNAseq data
BAM files with mapped reads were subjected to the Cufflinks suite [85] implemented on the Galaxy platform (https://usegalaxy.org). Briefly, assembled transcripts were generated using Cufflinks, and a final transcriptome assembly was generated from this using Cuffmerge. Mapped reads were quantified on this assembly using Cuffquant, and normalized expression levels were calculated using Cuffnorm. Heatmaps were generated in Microsoft Excel 2010. For calculation of expression levels relative to GAPDH, average and 95% confidence values of the four samples derived from each cell line were calculated.

GIREMI analysis
Single nucleotide variant (SNV) calling was performed using the SAMtools (v. 1.3) and BCFtools (v. 1.3) [86]. The produced SNV list was passed to the GIREMI (v. 0.2.0) [33], which split it into two groups: RNA editing positions and SNPs, dbSNP (build 138) [87] was used for the GIREMI statistical model evaluation. The resulting tables were imported into Microsoft Excel 2010, and numbers of mutations on each chromosome for each data set were counted. Numbers of A>G mutations were also determined for contiguous chromosome segments of 1,000,000 bp, as well as for individual genes. The list of ADAR1-edited genes was generated by comparing the number of editing sites in HeLa cells with the number in ADAR1 KO cells and ranked according to the highest differential. The following inclusion criteria were applied sequentially: Genes were included, 1. if the number of editing sites was �8 in HeLa cells; 2. if the ratio of detected editing sites #ADAR1 KO /#HeLa was �0.5; 3. if the number of editing sites per 100,000-bp gene length was �10; and 4. if #HeLa/(#ADAR1 KO + 1) was �1.75.

Editing score analysis
Read base count tables of regions of interest were generated from aligned BAM files using IGVTools (igvtools count--bases-w 1) and imported to Microsoft Excel 2010. Editing scores (e [Ts]) for A>G, C>U, G>A and U>C transitions (Ts) were calculated using function (Eq 1): with n Ts being the read counts of the transition nucleotide at the analyzed position, n Tv1 and n Tv2 being the read counts of the respective transversion nucleotides, and COV being the total read count at the analyzed position (coverage). Negative editing scores occur if more transversions than transitions are reported and indicate either sequencing artifacts or single nucleotide polymorphisms at the analyzed position. Significance of the ADAR1-specific transitions was tested against transversions at the same nucleotide position using Pearson chi-squared test with one degree of freedom (Eq 2) and by the Poisson model-based Wilks log-likelihood ratio that is asymptotically distributed as chi-squared with one degree of freedom: that is N(0,1) distributed. For this interval of chromosome 7, the differentiation score S diff = 5.72 with P = 5.3E-09.

Editing site neighboring nucleotide analysis
Analysis was performed as described previously [66]. Briefly, 9-nucleotide sequences around editing sites (A>G and U>C) were extracted for sites with transition frequencies of �20% and a coverage of at least 10 reads/nucleotide. For U>C sites, reverse complementary sequences were analyzed. Relative nucleotide frequencies at each position were calculated and normalized to the U-frequency.

Quantification and statistical analysis
RNAseq analyses were performed as n = 1 of each sample. For analysis of ADAR1 editing in viral genomes, forward and reverse reads of the paired-end sequencing data were analyzed separately, and sites were confirmed to be mutated in both reads.
Cellular transcriptome RNAseq was performed as n = 1 of each sample. Editing sites were confirmed in GEO-deposited RNAseq data sets.