Adenosine-to-Inosine RNA Editing of Alu Double-Stranded (ds)RNAs Is Markedly Decreased in Multiple Sclerosis and Unedited Alu dsRNAs Are Potent Activators of Proinflammatory Transcriptional Responses

A major host defense by multicellular organisms against viral infection is the expression of the IFN class of cytokines (1–3). IFNs are subdivided into type 1 IFNs (IFN-α, or leukocyte IFN, and IFN-β, or fibroblast IFN), type 2 IFN (IFN-γ, or immune IFN), and type 3 IFN (IFN-λ,also named IL-29, IL-28A, and IL-28B, respectively). IFN-α, IFN-β, and IFN-γ activate transcription factors, termed IFN regulatory factors, that bind to conserved transcriptional enhancer elements, termed IFN-stimulated response elements (ISREs) and induce similar but not entirely overlapping gene sets, termed IFN-stimulated genes (ISGs) with antiviral properties (4–6). IFN-γ also plays a critical role in orchestrating both innate and adaptive arms of the immune system (7). Induction of type 1 IFNs by viruses is mediated via pattern recognition receptors that recognize viral components such as cytosolic dsRNA (8). These include TLR3 and the DExD/H-box helicases, RIG-I, and MDA5 (9–12). A general view is that RIG-I recognizes short (<400 nucleotide [nt]) 5′ triphosphate uncapped dsRNA or ssRNA, whereas MDA5 and TLR3 recognize dsRNAs of greater nt length (10, 13, 14). Activation of these sensors induces IFN-α or IFN-β, and in turn, ISGs, as well as activation of the proinflammatory transcription factor NF-κB.

Increased expression of ISGs in the absence of apparent viral infection is also a hallmark of many human autoimmune diseases, most notably systemic lupus erythematosus but also relapsing–remitting multiple sclerosis (MS), rheumatoid arthritis, Sjogren syndrome, systemic sclerosis, myositis, and others (15–22). Sources of stimuli that lead to induction of ISGs in these human autoimmune diseases are incompletely understood. Aicardi–Goutieres syndrome is another example of a human disease associated with increased expression of type 1 IFNs and ISGs and results in severe early onset progressive encephalopathy (23, 24). Of note, Aicardi–Goutieres syndrome is a rare genetic disorder that can result from activating mutations in IFIH1, the gene that encodes the cytosolic dsRNA sensor MDA5. One activating mutation in IFIH1 found in Aicardi–Goutieres syndrome causes MDA5 loss of tolerance to endogenous cytosolic dsRNA structures and MDA5-mediated induction of IFNs and ISGs (13). A second class of mutations found in Aicardi–Goutieres syndrome is an inactivating mutation in ADAR1, the gene that encodes the enzyme dsRNA-specific adenosine deaminase ADAR (24). ADAR deaminates adenosines in RNA, resulting in conversion of adenosine to inosine (A-to-I), termed A-to-I editing, a process thought to destabilize dsRNA structures (25–28). In the case of the ADAR1 mutation in Aicardi–Goutieres syndrome, loss of ADAR function also results in increased expression of IFNs and ISGs. The general view is that loss of ADAR1 results in accumulation of endogenous cytoplasmic dsRNAs and activation of dsRNA sensors such as RIG-I, MDA5 or TLR3 (29–36). Thus, it is possible that alterations in the sensing of endogenous dsRNAs or in A-to-I editing of endogenous dsRNAs may contribute to the increased induction of ISGs observed in these more common autoimmune diseases.

The primary sources of dsRNA structures that exist in human leukocytes are RNAs produced by the retrotransposon class of genetic elements. Approximately 80% of these dsRNAs are derived from Alu elements, which are short interspersed elements of ∼300 nt in length unique to primates, and ∼10% of these dsRNAs are long interspersed elements (Line) (37) of up to ∼6000 nt in length (29, 35, 38–40). Approximately 1,000,000 Alu elements and 1,000,000 Line exist in human genomes and make up ∼30% of the human genome. Alu elements can be transcribed by RNA polymerase III as part of their normal life cycle. Alu elements are also embedded within protein-coding genes, predominantly within introns and 3′ untranslated regions (UTR) and are also transcribed by RNA polymerase II as precursor mRNAs. Importantly, studies show that most A-to-I editing sites are found within these Alu RNAs and, given their repetitive nature, can form dsRNA structures. As such, Alu dsRNAs can be recognized by endogenous dsRNA sensors, resulting in stimulation of downstream IFN and NF-κB responses. ADAR-mediated A-to-I editing of these endogenous Alu dsRNAs may decrease their dsRNA character, resulting in loss of their ability to be recognized by dsRNA sensors (12, 30, 31). Thus, A-to-I editing of these highly prevalent Alu and Line RNAs likely represents a critical cellular function of ADAR.

In leukocytes from patients with MS, induction of ISGs is mediated, at least in part, by increases in the levels of endogenous Alu dsRNAs (41). As such, increases in levels of Alu dsRNAs in leukocytes from patients with MS may result from defective A-to-I editing of Alu RNAs. To test this hypothesis, we examined levels of A-to-I editing using whole-genome RNA-sequencing (RNA-seq) data. In this study, we show that marked defects in A-to-I editing exist in leukocytes from MS patients compared with healthy controls (HC). We also find that native Alu RNAs highly expressed in leukocytes from patients with MS are recognized by RIG-I and TLR3 dsRNA sensors and induce potent IFN and NF-κB responses. Minor editing (A-to-G nt changes) of these native Alu elements results in marked loss of this activity. Thus, defects in A-to-I editing of Alu RNAs may mediate the observed increased IFN and NF-κB responses in this disease. It seems possible that similar defects in A-to-I editing of Alu RNAs may also contribute to increased IFN and NF-κB responses and pathogenesis associated with other autoimmune diseases and syndromes.

Whole-genome RNA-sequencing files employed in this study are deposited in the National Center for Biotechnology Information Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) (42). Accession numbers are GSE92472 and GSE126427. We employed SPRINT, a toolkit that identifies RNA editing sites without the need for filtering single nt polymorphisms, using whole-genome RNA-seq data (43). We used the following workflow to identify RNA editing sites from paired FASTQ sequencing files. The main identification tool is a Python-based package called the SPRINT toolkit for identifying RNA editing sites. The multithread toolkit accepts sequence files and produces a text file that contains information about each editing site in the sequence data. This information includes the following for each edit site: 1) genomic location, 2) type of edit (e.g., A-to-G or T-to-C), strand (“+” or “−“), 3) number of edits to reads at the site, and 4) total number of reads at the site. This analysis is performed on each sample of the data groups (HC, MS patients prior to therapeutic intervention [MS-N], and MS patients with established disease on therapy [MS-E]; Table I) (42). Mathematica (Wolfram Research) programs were developed to synthesize the data for each group. Mainly, we were interested in the number of samples in groups that share editing sites, mean numbers of total reads and edits for each editing site across the group, and the editing sites that are common and distinct to group pairs (e.g., HC versus MS-E). Last, the editing site information was tied to an Alu database that annotated each site with respect to nearby genes, type of location (intronic, noncoding RNA, intergenic, 3′ UTR), and whether the site was indeed an Alu or a non-Alu repetitive site.

We employed the following scheme to classify editing-enriched locations (EELs). First, genomic locations of A-to-I editing sites present in at least two HC samples or at least two MS-E samples were pooled into a single list. Second, distances between neighboring A-to-I editing sites in the genome were calculated. Third, EEL boundaries were formed if the next neighboring common A-to-I editing site in either 5′ or 3′ directions in the genome was >5000 bp in distance. Fourth, we employed a rather arbitrary cutoff of presence of at least five A-to-I editing sites within EEL boundaries for inclusion in the analysis, as this cutoff still captured the vast majority of common A-to-I editing sites across the genome. For example, the length of the CSAD gene EEL is 4576 bp and contains 51 common A-to-I editing sites for an average distance between editing sites of 90 bp (range of 1–1676 bp). The next-nearest A-to-I editing site in the 5′ direction is 1,874,544-bp away, and the next-nearest A-to-I editing site in the 3′ direction is 118,150-bp away, thus establishing the EEL boundary.

To create an EEL editing index, we identified all A-to-I editing sites within an EEL and separated them into those present in at least two HC samples or at least two MS-E samples. We summed the A-to-I edit/total read ratios for all sites for the HC class or the MS-E class and multiplied this value by the total number of HC or MS-E samples with these edits. These calculations formed an EEL editing index to help gauge whether individual EELS were preferentially edited in the HC class, the MS-E class, or exhibited similar EEL editing levels between classes.

Reference DNA sequences of Alu and Line elements were obtained from the University of California, Santa Cruz genome browser, and a SP6 bacteriophage promoter was added to the 5′ end of the sequence, and synthetic dsDNAs were obtained from Integrated DNA Technologies (44). RNA transcription was performed using the MEGAscript SP6 Kit (catalog no. AM1330; Invitrogen) per manufacturer’s instructions. Briefly, 8 μl of DNA was placed in a 1.5-ml microfuge tube and heated for 5 min at 65°C and cooled. ATP, CTP, GTP, UTP, and 10× reaction buffer and SP6 RNA polymerase were added, and the reaction was allowed to incubate overnight at 37°C. TURBO DNAse was added and incubated for 30 min, and the reaction was stopped with the addition of 30 μl of dH20 and 30 μl of lithium chloride precipitation solution and incubated for 30 min at −20°C, followed by centrifugation at 4°C at 20,000 × g for 15 min. The supernatant was removed, and the pellet washed with 70% ethanol. The centrifugation step was repeated under the same conditions. The 70% ethanol was removed, and the pellet was briefly air-dried prior to suspension in 50 μl of nuclease free water. Absorbance was determined at 260 nm to quantitate yields of RNA, and the solutions were stored at −80°C. Yields typically were 10–70 μg of RNA.

THP-1 and human embryonic kidney (HEK) 293 reporter cell lines were obtained from InvivoGen. THP-1 reporter cells contain a stably integrated luciferase gene under the control of either an ISRE or a NF-κB response element. HEK-293 reporter cells contain a stably integrated luciferase gene under the control of the ISRE element. A second line of HEK-293 reporter cells, HEK-RIG-I also contained a stably integrated DDX58 gene encoding RIG-I. Cells were cultured in RPMI 1640 media (catalog no. 21870; Life Technologies) supplemented with 10% (vol/vol) FCS (catalog no. S12450; Atlanta Biologicals), 100 U/ml penicillin, 100 μg/ml streptomycin, and 2 mM l-glutamine (final concentration, catalog no. 15140; Life Technologies). Cells were cultured at 37°C in 5% CO₂ in humidified air. TLR3, IFIH1 (MDA5), and DDX58 (RIG-I) expression plasmids were obtained from InvivoGen.

Cell transfections were performed using Lipofectamine RNAiMAX transfection reagent (catalog no. 13778150; Thermo Fisher Scientific) according to manufacturer’s instructions (45). All dilutions were performed in Opti-MEM medium (catalog no. 31985070; Thermo Fisher Scientific). Briefly, on day 0, cells were plated at 100,000 cells/ml in 96-well plates at volumes of 100 μl. On day 1, 1.5 μl of RNAiMAX was diluted into 25 μl of Opti-MEM. This was scaled, depending upon the number of samples, dilutions, and culture volumes. RNAs were diluted in Opti-MEM at various concentrations before addition to cell cultures. Culture after transfection was typically 24 h. Levels of secreted luciferase in culture fluids were determined using a luciferin substrate (InvivoGen), and light emission was determined using a TD20/20 Luminometer.

We examined our whole-genome RNA-seq data using the SPRINT software package to identify differences in A-to-I editing in leukocytes between HC and patients with MS. For these initial analyses, we used two criteria to identify A-to-I editing sites. First, we required the proportion of A-to-I edits to exceed 5% of total reads at individual nt sites to guard against possible sequencing errors and total read count >10. Second, we required that individual candidate A-to-I editing sites were edited in ≥25% of the HC cohort, ≥25% of the MS-E cohort, or ≥25% of both HC and MS-E cohorts. Overall, >90% of editing events we identified occurred within Alu elements, and these were mainly distributed among introns, 3′ UTRs of poly(A)⁺ protein-coding genes and intergenic space (Fig. 1A). Of note, using the above criteria, we identified three distinct patterns of differential A-to-I editing of Alu elements. One pattern showed extensive editing of individual Alu elements in the HC group with the absence of editing in the MS group; two examples include the AluSx1 element in the MAVS gene and the AluSz element in the TPM4 gene (Fig. 1B). The converse pattern also existed, showing editing of individual Alu elements in the MS group but not the HC group; examples include the AluSx1 element in the PLEKHG2 gene and the AluSz6 element in the COTL1 gene (Fig. 1C). The third pattern we identified showed extensive editing at identical adenosines within individual Alu elements in both HC and MS-E groups; examples include the AluSq element in the NUP43 gene and the AluSp element in the SLC12A6 gene (Fig. 1D).

Next, we examined A-to-I editing sites genome wide. Using the above criteria, we found 2724 nt sites that were A-to-I edited in genomes of the HC cohort but were not A-to-I edited in genomes from the MS-E cohort. In sharp contrast, we identified only 665 nt sites that were A-to-I edited in genomes from the MS cohort and not edited in genomes from the HC cohort (p < 0.0001, χ2 analysis) (Fig. 2A). We also found 2243 nt sites that were A-to-I edited in both HC and MS cohorts (Fig. 2B). The average proportion of edits to total reads of these shared A-to-I editing sites was similar, between 35 and 40%, and the proportion of edits to reads ranged from 5 to 100%. We next identified those A-to-I edits present in 100% of HC or 100% of MS samples or 100% of both. We identified 280 editing sites present in 100% of HC samples and absent from MS samples and only 64 editing sites present in 100% of MS samples and absent from HC samples (p < 0.0001, χ2 analysis) (Fig. 2C). As above, we identified 174 nt sites edited in all HC and all MS samples (Fig. 2D). The average proportion of edits to total reads of these A-to-I editing sites was also similar, between 35 and 40%, and the proportion of edits to reads ranged from 5 to 100%.

We next examined A-to-I editing in a second group of MS patients, using RNA-seq data. Blood samples from these patients were obtained at initial diagnosis of MS but before onset of therapies, so they were treatment naive (MS-N). We identified A-to-I edits present in 100% of HC or 100% of MS-N samples or 100% of both. We identified 271 A-to-I edits present in 100% of HC samples and absent from MS-N samples, but only 96 editing sites were present in MS-N samples and absent from HC samples (p < 0.0001, χ2 analysis) (Fig. 2E). We also identified 220 sites that were edited in all HC and all MS samples (Fig. 2F). The average proportion of edits to total reads of these editing sites was similar, between 35 and 40%, and the proportion of edits to reads ranged from 5 to 100%. Thus, we identified A-to-I editing sites that were shared between both MS-E patient samples and HC as well as A-to-I editing sites that were unique to either the MS patient groups or the HC group. At these editing sites, the average proportion of edited reads to total reads was similar across all samples. The major differences we found were in edits that were unique to either the HC cohort, which were high in number, or the MS-E and MS-N cohorts, which were much lower in number. Thus, leukocytes from MS patients showed an overall deficit in A-to-I editing.

We also enumerated numbers of A-to-I editing sites within protein-coding genes, counting those in exons, introns, 5′ UTRs and 3′ UTRs by including editing sites present in either at least two HC samples, at least two MS-E samples, or those shared between at least two HC samples and at least two MS-E samples. For this comparison, we focused on genes with the greatest number of A-to-I editing sites. Two classes of editing patterns emerged. We identified genes with approximately equivalent numbers of total A-to-I editing sites across the gene in both HC and MS-E classes (Fig. 3A), and these differences were not statistically significant. We determined expression levels of these genes from RNA-seq data, and expression levels of these genes were not significantly different between HC and MS-E classes (Fig. 3B). We also identified genes with relatively greater numbers of editing sites in the HC class compared with the MS-E class (Fig. 3C), which are similar to what we observed genome wide (Fig. 2B, 2C). We determined expression levels of these genes from the RNA-seq analysis of these same samples. We found that those genes with reduced numbers of A-to-I editing sites in the MS-E class compared with the HC class also demonstrated overall reduced levels of mRNA expression in the MS-E class compared with the HC class (Fig. 3D). Thus, genes with reduced levels of edits in MS-E compared with HC also exhibit reduced levels of expression compared with HC, whereas genes with similar numbers of edits between HC and MS-E classes exhibited similar levels of expression between the two classes. These results do not prove causality but suggest that further experimentation is warranted to better understand the relationship between A-to-I editing within protein-coding genes and mRNA expression levels of these genes.

We selected two of these genes for closer inspection, TMEM154 and CSAD, and compared locations of A-to-I editing sites to gene organization and underlying Alu elements (Fig. 4A, 4B). We used the same criteria as above and only included editing sites present in at least two HC samples, at least two MS-E samples, or shared between at least two HC and at least two MS-E samples. The inset to the left of each gene shows the overall average edit/read ratios at each editing site for the HC and MS-E samples. To the right of the inset is the genomic location of TMEM154 and CSAD genes along with positions of UTRs, exons, and introns. Below that are the positions of all the Alu elements across each gene. The green bars indicate the approximate locations of all A-to-I editing sites within each gene. We found that A-to-I editing sites were not evenly distributed across the gene but were confined to relatively small locations or EELs. Thus, the TMEM154 EEL is <1000 bp, whereas the TMEM154 gene is ∼65,000 bp. Similarly, each gene contained multiple Alu elements, and many of these were transcribed as RNA (data not shown), but only a small number of Alu elements within the EEL were A-to-I edited. Patterns of editing in the different EELs were mixed, as the CSAD EEL was predominately edited in HC samples but not MS-E samples, whereas the TMEM154 EELs were edited similarly in both HC and MS-E samples. Thus, only very small fractions of potential editing sites within Alu elements are actually recognized as dsRNA and A-to-I edited by ADAR.

To explore this idea further, we pooled all editing sites present in at least two HC samples, at least two MS-E samples, or shared between at least two HC samples and at least two MS-E samples into a single list of common A-to-I editing sites. We determined the genomic distance between one editing site and its nearest neighbor in the genome. We found that >60% of these common editing sites were within 100 bp of its nearest neighbor and almost 80% were within 1000 bp of its nearest neighbor (Fig. 5A). Next, we defined EEL length as described in 2Materials and Methods. Using these criteria, we found that the length of EELs ranged from 5 to ∼4500 bp with a mean length of ∼800 bp (Fig. 5B). Numbers of common A-to-I editing sites within an EEL ranged from 6 to ∼65 sites with a mean number of editing sites per EEL of ∼15 (Fig. 5C). Thus, as above, we found that the majority of genome-wide common editing sites are not evenly distributed but are clustered within discrete EELs.

Using the EEL definition, we next examined editing profiles of individual EELs rather than individual A-to-I editing sites. To do so, we created an EEL editing index (see 2Materials and Methods) to gauge whether all editing sites within an EEL may be preferentially edited in the HC class, the MS-E class, or whether there was no difference in the overall level of EEL editing between the HC and MS-E classes. This index showed that editing indices of individual EELs were fairly diverse, with certain EELs displaying similar editing index values between HC and MS-E classes, whereas other EELs were preferentially edited in the HC class compared with the MS-E class (Fig. 5D). We also found that differences between EEL indices in HC and MS-E classes were highly correlated with differences in expression of the genes in which these EELs resided between HC and MS-E classes (Fig. 5E).

Given the above results we wanted to determine the ability of individual Alu and Line RNAs to activate dsRNA sensors and stimulate downstream transcriptional responses as well as the effect of editing on their biologic activity. To do so, we identified several Alu elements present at high levels in leukocytes (RNA-seq analysis) with elevated levels of A-to-I editing in HC compared with MS-E samples for further analysis. In vitro, we transcribed these Alu elements and Line elements, also highly expressed as dsRNAs, in MS leukocytes and tested their ability to stimulate luciferase reporter constructs under the control of either an ISRE or NF-κB response element stably integrated into the human monocyte cell line THP-1 (41). We identified two Alu RNAs of the AluJb and AluY2 classes that were potent stimulators of both the ISRE and NF-κB transcriptional responses, whereas a third Alu RNA of the AluYK4 class was a weak stimulator of these response elements (Fig. 6A, 6B). Similarly, we identified Line dsRNA elements that were very potent inducers of both the ISRE and NF-κB responses (Fig. 6C, 6D). For comparison, we also stimulated THP-1 reporter cells with various amounts of IFN-α. We found that maximal activation of the ISRE response by IFN-α was similar in magnitude to activation by Alu or Line dsRNA elements. In contrast, activation of the NF-κB response by IFN-α was somewhat less in magnitude than that observed for the Alu or Line dsRNA elements. To actually demonstrate that these Alu and Line dsRNA elements induced changes in gene expression, we measured induction of RELB and IL6, two known NF-κB–responsive genes, and DDX58 and IFIT5, two known ISRE-responsive genes. We compared two cell lines of myeloid origin, THP-1 and HMC3, a human microglial cell line (46), and found these genes were induced >20-fold after stimulation by a representative Alu or Line dsRNA element (Fig. 6F). Gene induction in response to IFN-α was also determined. We also found that the more potent Alu and Line dsRNA elements induced ∼5-fold greater ISRE and NF-κB transcriptional responses than poly IC at amounts ∼10-fold less than poly IC. Alu and Line dsRNA elements also induced higher levels of gene expression at lower amounts of RNA than poly IC. Thus, we found that both Alu and Line dsRNA elements that were highly expressed in MS patient leukocytes were potent inducers of both ISRE and NF-κB transcriptional responses measured using either reporter assays or induction of known ISRE- and NF-κB–responsive genes in two distinct myeloid-derived cell lines.

In vitro, we transcribed additional Alu and Line RNA elements that were highly expressed as dsRNAs in the MS cohort and tested their ability to induce ISRE and NF-κB transcriptional responses using the reporter assays. We tested induction of these responses at RNA amounts ranging from 0.1 to 100 ng/culture, and data are expressed as the effective concentration required for 10× induction of either ISRE or NF-κB responses over backgrounds observed in cultures transfected without added exogenous RNA, EC_10×. We identified both potent Alu and Line RNAs that stimulated ISRE and NF-κB responses by >10-fold at amounts of ≤1 ng RNA/culture, Alu and Line RNAs with moderate activity stimulating ISRE and NF-κB responses by >10-fold at RNA amounts of 1–10 ng/culture RNA, and those that were inactive and failed to stimulate ISRE and NF-κB responses at amounts of RNA >100 ng/culture (Table II). Both Alu and Line elements are subdivided into classes (e.g., AluJ, AluY, AluS, L1-L5). We tested representatives of these different classes and found highly active Alu and Line RNAs represented by multiple classes. Also, we found representatives from identical classes that were both highly active and completely inactive. Thus, it seems that the ability to activate dsRNA sensors and ISRE and NF-κB responses was not dependent upon belonging to a specific Alu or Line class. Thus, these experiments demonstrated that multiple Alu and Line RNAs are potent activators of both ISRE and NF-κB responses.

We employed the HEK 293 cell line (HEK293) to determine which dsRNA sensors were activated by representative Alu and Line RNAs. HEK293 cells do not express dsRNA sensors RIG-I and MDA5 but do express TLR3 (37). We compared responses in HEK293 cells (HEK-TLR3), HEK293 cells with a stably integrated RIG-I (DDX58) expression construct (HEK-TLR3 plus RIG-I), and HEK293 cells transiently transfected with a MDA5 expression construct (HEK-TLR3 plus MDA5). Each of these cell lines also contained a stably integrated ISRE luciferase reporter construct. We found that transfection of poly IC activated ISRE responses almost equivalently in the three HEK lines, suggesting that under these conditions, poly IC only activates TLR3 (Fig. 7A). We found that AluJb dsRNA, one of the more potent inducers of ISRE and NF-κB responses in THP-1 cells (genomic location-chr17:76,418,582-76,418,856), was a potent inducer of ISRE responses in HEK293-TLR3 plus RIG-I cells compared with HEK-TLR3 and HEK-TLR3 plus MDA5 cells (Fig. 7B). ISRE responses were also detected in HEK-TLR3 cells indicating that AluJb is recognized by RIG-I and TLR3. We did not find an increase in ISRE responses in HEK-TLR3 plus MDA5 cells compared with HEK-TLR3 cells, suggesting that AluJb does not activate MDA5. We performed a similar analysis using one of the Line dsRNA elements, L1P1 (genomic location-chr8:12,399,322-12,400,289). We found that this L1P1 dsRNA element activated ISRE responses in HEK-TLR3, HEK-TLR3 plus RIG-I, and HEK-TLR3 plus MDA5 reporter cells (Fig. 7C). Activation of ISRE responses was strongest in the HEK-TLR3 plus RIG-I cells. ISRE responses were higher in HEK-TLR3 plus MDA5 cells compared with HEK-TLR3 cells, suggesting that the L1P1 dsRNA element also activated ISRE responses via MDA5. Thus, these results are consistent with the notion that the AluJb dsRNA element activated ISRE responses through RIG-I and TLR3, whereas the L1P1 dsRNA element activated ISRE responses via RIG-I, TLR3, and MDA5 sensors.

An Alu element is composed of a left arm of ∼100 nt and a right arm of ∼200 nt (39). Certain models predict that, if transcribed, an Alu RNA forms two dsRNA structures, one composed of the left arm and one composed of the right arm, with a linker sequence in between (47). We prepared new AluJb elements, one consisting of the right arm and one consisting of the left arm, in vitro transcribed each new element, and tested them for activation of ISRE responses in HEK-TLR3 plus RIG-I reporter cells. We found that activity of the AluJb left arm dsRNA was almost equivalent to activity of the intact AluJb dsRNA, whereas the AluJb right arm possessed almost no activity (Fig. 8A). These results seem to support the model predicting that Alu RNAs contain two independent arms, each with its own dsRNA structure.

Further, ADAR-catalyzed A-to-I editing of Alu dsRNA elements is thought to result in loss of dsRNA character and the ability of these RNA elements to be recognized by dsRNA sensors and stimulate downstream signaling cascades, resulting in activation of ISRE and NF-κB responses. Our results, as well as those of others, indicate that not all A’s are equal targets for A-to-I editing. Our results comparing editing of Alu elements in HC and MS showed that only a portion of total A’s within an Alu RNA element are edited to I’s, and the edit/read ratios at a given editing site were generally <100%. However, it is generally unknown how efficient or extensive A-to-I editing needs to be to convert an active Alu RNA recognized by dsRNA sensors into an inactive Alu RNA incapable of triggering IFN responses. To explore this idea further, we prepared new AluJb DNA elements with all A’s changed to G’s in the entire element, all A’s changed to G’s in the AluJb left arm, and all A’s changed to G’s in the AluJb right arm; we in vitro transcribed these new elements and tested them for activity in HEK-TLR3 plus RIG-I reporter cells containing the ISRE luciferase construct. We found that changing all A’s to G’s in the entire AluJb element or only changing A’s to G’s in the left arm of the intact AluJb element resulted in an almost complete loss of ability to activate ISRE responses (Fig. 8B). In contrast, changing A’s to G’s only in the AluJb right arm resulted in a small increase in activity.

Given the above results, we next determined how changes in AluJb dsRNA nt sequence altered ISRE and NF-κB responses in THP-1 reporter cells. As above, we found that the AluJb left arm dsRNA activated both ISRE responses and NF-κB responses approximately as well as the full-length AluJb dsRNA, whereas the AluJb right arm did not activate ISRE and NF-κB responses (Fig. 9A). We also found that changing either all A’s to G’s in the entire AluJb element or changing only A’s to G’s in the left arm resulted in almost complete loss of activity, whereas changing A’s to G’s only in the right arm resulted in a small increase in the ability of this modified AluJb dsRNA to activate either ISRE or NF-κB responses in THP-1 reporter cells (Fig. 9B).

The AluJb element is 274 nt in length, and changing all A’s to G’s in the left arm results in 15-nt changes. To pursue this notion further, we asked whether fewer A-to-G changes in the AluJb left arm affected the ability to activate ISRE and NF-κB responses in THP-1 reporter cells. We created three new constructs containing 10, 7, or 3 A-to-G changes in the AluJb left arm and tested the ability of these RNAs to stimulate ISRE and NF-κB responses in THP-1 reporter cells. We found that 10 A-to-G changes resulted in almost complete loss of the ability to stimulate ISRE or NF-κB responses (Fig. 9C). Making seven A-to-G changes in the AluJb left arm also resulted in substantial loss of the ability to activate ISRE and NF-κB responses, whereas even three A-to-G changes in the AluJb left arm resulted in partial loss of the ability to activate ISRE and NF-κB responses by this modified dsRNA. Thus, small changes in nt sequence can markedly affect the ability of Alu dsRNAs to stimulate dsRNA sensors and downstream signaling paths, leading to activation of ISRE and NF-κB transcriptional responses.

Our results document substantial differences in A-to-I editing between HC and MS patient leukocytes. Perhaps the most striking difference is the complete absence of editing of certain Alu RNA sequences in MS leukocytes and the presence of substantial editing at these same Alu RNA elements in HC leukocytes. Importantly, these differences are seen early in MS disease and prior to onset of treatments. In contrast, certain Alu RNA elements are A-to-I edited at approximately equivalent levels in both HC and MS leukocytes. We also find that A-to-I editing sites within Alu elements are not evenly distributed across the genome but are clustered in the genome in what we term EELs. Interestingly, loss of EEL editing in MS is associated with loss of expression of genes in which the EELs are embedded. We also find that unedited Alu and Line dsRNAs are potent activators of dsRNA sensors, resulting in induction of ISRE and NF-κB transcriptional responses and known ISRE- and NF-κB–regulated genes. The unedited Alu dsRNAs stimulate ISRE responses via both RIG-I and TLR3 but not MDA5, whereas the longer, unedited Line dsRNAs stimulate ISRE responses via all three dsRNA sensors examined (RIG-I, TLR3, and MDA5). Our results show that for the AluJb dsRNA element, one half, or arm, possesses all the ability to activate these inflammatory transcriptional responses, whereas the other half, or arm, is completely inactive, and we have replicated this finding by analyzing a second Alu RNA element (data not shown). Further, converting A’s in the active arm to G’s, as a model of A-to-I editing, results in almost complete loss of activity, whereas converting A’s to G’s in the inactive arm results in a small gain in overall activity. Finally, replacing a small number of A’s with G’s as a model of A-to-I editing has profound effects on the ability of these dsRNAs to activate ISRE and NF-κB responses. In summary, our results are consistent with a model in which deficiencies of A-to-I editing of Alu RNAs in MS patients increase the dsRNA character of these abundant Alu and Line RNAs, giving rise to increased activation of both IFN and NF-κB signaling paths.

The concept of EELs or “editing-enriched regions” (EERs) in the genome is not entirely new (48). Using similar strategies as outlined in this study, previous studies have documented presence of EERs in Caenorhabditis elegans, mice, and humans. Because our methodology is not identical, we chose to refer to these regions as EELs to avoid any confusion. Interestingly, the presence of EERs within a gene versus the absence of EERs within a gene is associated with overall higher expression levels compared with genes that lack EERs. We show in this study that loss of A-to-I editing in gene-associated EELs, which is seen in MS-E, is also associated with loss of gene expression, suggesting perhaps that one function of A-to-I editing in EELs may be to enhance expression of genes with embedded EELs. EELs probably have additional functions, and understanding the basic functions of EELs will contribute to a better understanding of the critical importance of A-to-I editing for development and survival.

One example of human disease resulting from persistent activation of the IFN system is Aicardi–Goutieres syndrome (23, 24). One form of this syndrome results from loss-of-function mutations in ADAR1 and deficiencies in A-to-I editing. A second form results from gain-of-function mutations in IFIH1, the gene that encodes MDA5, resulting in loss of tolerance to endogenous dsRNAs and elevated IFN responses. Age-related macular degeneration is another disorder in which accumulation of Alu dsRNA elements contributes to activation of IFN responses and cell death (49, 50). In this disease, accumulation of Alu dsRNAs appears to result from loss of DICER1 and subsequent failure of DICER1-mediated degradation of Alu dsRNAs. Our results show there is dramatic loss of A-to-I editing at certain editing sites in MS patient leukocytes. Several factors may contribute to selective loss of A-to-I editing, depending upon nt site. First, our studies were performed using whole blood, which contains multiple leukocyte lineages, and A-to-I editing processes in unique leukocyte lineages may be affected differently in MS. Second, besides ADAR and DICER, multiple other RNA binding proteins (RBPs) can both positively and negatively influence A-to-I editing, and these effects seem to vary, depending upon the specific nt editing site in the genome (51). For example, depletion of TROVE2 in cell lines, the gene that encodes the RBP, Ro60, results in both increases and decreases of A-to-I editing, depending upon the editing site. Ro60 also plays an important role in RNA surveillance or quality control, and we have previously shown that both TROVE2 mRNA and Ro60 protein are highly under expressed in MS leukocytes, which may influence A-to-I editing (52–54). Several other genes that encode RBPs shown to affect genome-wide A-to-I editing are also differentially expressed in MS compared with HC leukocytes, and differences in expression of these RBPs may also contribute to the observed differences in A-to-I editing in MS. Thus, there may be multiple molecular mechanisms contributing to the observed deficiencies in A-to-I editing in leukocytes in MS.

In contrast to diseases or syndromes in which increased production of type 1 IFNs is considered to be pathogenic, commonly referred to as interferonopathies, including Aicardi–Goutieres syndrome and certain autoimmune diseases, such as systemic lupus erythematosus, therapeutic versions of IFN-β are mainstays for treatment of MS, reducing both rates of disease relapse and long-term disability (55–57). Similarly, type 1 IFNs reduce severity of disease in animal models of MS. Thus, one might speculate that, in contrast to the prevailing views of interferonopathies, decreased A-to-I editing and increased accumulation of unedited Alu dsRNA elements and elevated expression of ISGs may have positive impacts on MS. A better understanding of mechanisms that regulate levels of A-to-I editing in MS may improve disease management and perhaps point to new therapeutic targets.

In mouse models, early embryonic lethality and IFN responses caused by ADAR deletion are prevented by concomitant deletion of the dsRNA sensor MDA5/IFIH1 but not by deletion of RIG-I, suggesting that MDA5 but not RIG-I is the key sensor of endogenous dsRNAs (31, 58). These interpretations are complicated by the fact that the RIG-I/DDX58 knockout is also embryonically lethal, so it may not be possible to rescue embryonic lethality observed by ADAR deletion with the DDX58 knockout because it also causes embryonic lethality and employ this experimental approach to exclude the role for RIG-I as a key sensor of endogenous dsRNAs. However, studies of inactivating human ADAR mutations in cell lines further support that MDA5 is a key sensor of endogenous dsRNAs (13, 59). In contrast, studies in certain stress models, such as ischemia/reperfusion injury, show that ADAR plays a critical role in this injury response, such that deletion of ADAR greatly exacerbates the IFN response associated with this injury model as well as the resulting pathology (60). In this case, injury caused by ADAR loss is reversed by concomitant depletion of RIG-I, suggesting that in this model RIG-I is a key sensor of endogenous dsRNAs. Thus, in one experimental setting, MDA5 appears to be the major sensor of endogenous dsRNAs, whereas in a second experimental setting, RIG-I appears to be the major sensor of endogenous dsRNAs. Two key differences between these two experimental systems are that those showing MDA5 rescue of ADAR deletion are seen early in development and result from ADAR-inactivating mutations or gene knockouts, whereas systems showing RIG-I rescue of ADAR depletion employ knockdown strategies that do not result in complete inactivation of ADAR and employ fully developed liver tissue from adult mice. Thus, it seems there may be different contributions of MDA5 and RIG-I as endogenous dsRNA sensors, depending upon either developmental status, cell lineage, or response to external or internal stimuli. In addition, we do not see total loss of A-to-I editing in MS as might result from either ADAR deletion or presence of ADAR-inactivating mutations. These differences may suggest that loss of A-to-I editing observed in MS-E may be a kind of stress response to exogenous or endogenous stimuli.

Further, loss of the ADAR p150 isoform disrupts intestinal homeostasis and B cell development, whereas loss of the ADAR p110 isoform disrupts kidney and brain function (58). These developmental defects are independent of induction of IFN responses via MDA5 activation. Thus, A-to-I editing is also essential for multiple developmental processes. Whether selective loss of A-to-I editing in EELs, as observed in MS-E, may contribute to MS pathogenesis via mechanisms independent of activation of IFN responses is unknown.

In MS leukocytes, the endogenous dsRNA fraction capable of stimulating IFN responses is composed primarily of Alu RNAs and Line RNAs (41). We in vitro transcribed and tested different Alu and Line RNAs that are highly expressed in MS leukocytes to assess their ability to activate ISRE and NF-κB responses. Although RNAs derived from Alu elements have a great deal of sequence similarity, they display very different abilities to activate ISRE and NF-κB responses. Some activate maximal responses at very low RNA amounts, 1 ng/culture, whereas others are completely inactive. Line elements exhibit similar properties, with some having very high activity and others exhibiting no activity. Sensors that recognize dsRNA include RIG-I, MDA5, and TLR3. A general view is that RIG-I activation involves sensing of a free 5′ triphosphate short dsRNA, whereas MDA5 and TLR3 are activated by longer dsRNA structures. The contribution of double-stranded structure or perhaps higher-order structure to the ability to activate these dsRNA sensors is incompletely understood. Our results show that Alu RNA elements activate both RIG-I and TLR3 dsRNA sensors but not MDA5, whereas the longer Line RNA elements activate all three dsRNA sensors to induce potent ISRE and NF-κB responses. The fact that certain Alu RNAs activate potent ISRE and NF-κB responses, whereas others are without activity and all possess a 5′ triphosphate suggests that either the double-stranded structure or perhaps higher-order structure also contribute to recognition of these Alu RNAs by dsRNA sensors.

Alu retrotransposons inserted in the genome are composed of a left arm and a right arm and a linker sequence. Certain models predict that both left and right arms may form independent dsRNA structures, whereas other models predict that the total Alu RNA sequence is involved in forming only one continuous dsRNA structure (47). Our results show that, at least for one highly active Alu RNA, all the ISRE- and NF-κB–inducing activity lies in the left arm, whereas the right arm is completely inactive. In addition, A-to-G nt editing in the inactive arm results in an increase in biologic activity, whereas A-to-G editing in the active arm completely abrogates activity. These results lend further support to this general model of a left arm–right arm structure of Alu RNAs. At this point, we do not know if all active Alu RNAs can be subdivided into an active arm–inactive arm structure or if in some cases both arms possess biologic activity.

The general view is that A-to I editing is a critical process to “de-toxify” endogenous RNAs to prevent their recognition by dsRNA sensors. Our sequencing analyses demonstrate that only a fraction of possible A-to-I editing sites in an Alu RNA are actually edited, raising the question of how many edits are actually necessary to detoxify an Alu RNA and prevent it from activating dsRNA sensors and downstream signaling paths. Our A-to-G editing model of Alu RNA editing indicates that as few as three A-to-G changes in a ∼275-nt Alu element significantly impacts activity, whereas 5–10 A-to-G changes almost completely abrogate biologic activity. This small number of editing changes is similar in magnitude to the A-to-I editing differences found in individual Alu RNAs elements when HC and MS are compared. Thus, we propose that loss of A-to-I editing observed in MS contributes to increased dsRNA character of Alu RNAs and increased activation of ISRE and NF-κB transcriptional responses and downstream inflammatory responses.

The authors have no financial conflicts of interest.