Noncoding regions of C. elegans mRNA undergo selective adenosine to inosine deamination and contain a small number of editing sites per transcript

Abstract

ADARs (Adenosine deaminases that act on RNA) “edit” RNA by converting adenosines to inosines within double-stranded regions. The primary targets of ADARs are long duplexes present within noncoding regions of mRNAs, such as introns and 3′ untranslated regions (UTRs). Because adenosine and inosine have different base-pairing properties, editing within these regions can alter splicing and recognition by small RNAs. However, despite numerous studies identifying multiple editing sites in these genomic regions, little is known about the extent to which editing sites co-occur on individual transcripts or the functional output of these combinatorial editing events. To begin to address these questions, we performed an ultra-deep sequencing analysis of 4 Caenorhabditis elegans 3′ UTRs that are known ADAR targets. Synchronous editing events were determined for the long duplexes in vivo. Furthermore, the validity of each editing event was confirmed by sequencing the same regions of mRNA from worms that lack A-to-I editing. This analysis identified a large number of editing sites that can occur within each 3′ UTR, but interestingly, each individual transcript contained only a small fraction of these A-to-I editing events. In addition, editing patterns were not random, indicating that an editing event can affect the efficiency of editing at subsequent adenosines. Furthermore, we identified specific sites that can be both positively and negatively correlated with additional sites leading to mutually exclusive editing patterns. These results suggest that editing in noncoding regions is selective and hyper-editing of cellular RNAs is rare.

Abbreviations

A

adenosine

I

inosine

ADAR

adenosine deaminase that act on RNA

nts

nucleotides

UTR

untranslated region

bp

base-pair

Introduction

Adenosine (A) to inosine (I) deamination of RNAs is a conserved RNA editing mechanism catalyzed by the adenosine deaminase that act on double-stranded RNA (ADAR) family.^1,2 In animals, A-to-I editing is the most prevalent type of nucleotide modification, with current transcriptome-wide studies predicting over one hundred million A-to-I editing events in human mRNAs.^3-5 Inosine preferentially base pairs with cytosine. Therefore, depending upon the location within the mRNA, A-to-I editing can alter the amino acid encoded by a codon, modify splice sites or affect the interaction of the RNA molecule with itself or other RNAs, such as miRNAs and siRNAs.^6-8 Thus, A-to-I editing in both coding and noncoding regions of mRNA can have biological significance. Consistent with an important biological role, loss of ADARs results in lethality in mice^9,10 and neurological phenotypes in Caenorhabditis elegans¹¹ and Drosophila melanogaster model systems.^12,13

Editing at individual sites rarely reaches 100% deamination and can vary both during development¹⁴ and between cell types.¹⁵ In addition, multiple A-to-I editing events have been identified within one local genomic region. Therefore, RNA editing generates multiple transcripts from a single genomic locus and is thought to contribute to protein and RNA diversity.^16,17 However, very little is known about the extent to which editing sites co-occur on target RNAs or mechanistically how ADARs deaminate multiple sites on a single RNA.

ADARs bind their target mRNAs via double-stranded RNA binding domains (dsRBDs).¹⁸ Recent structural studies of the dsRBDs of mammalian ADARs and short hairpin double-stranded RNA (dsRNA) have revealed that ADARs recognize both the shape of the dsRNA and some specific sequences in the minor groove.¹⁹ In addition, the nucleotides both adjacent to and opposing a target adenosine are known to influence the efficiency of editing.²⁰ However, recent studies of human ADAR2 have suggested that these neighboring nucleotides affect the ability of the target adenosine to flip out of the duplex and undergo deamination, rather than serve as a recognition sequence for ADAR to target specific adenosines.²¹

A number of factors are known to contribute to ADAR specificity in vitro, including intra- and intermolecular base-pairing,²² length of the dsRNA,²³ and the positioning of bulges, loops and mismatches within a duplex.^24,25 Consistent with this, imperfect base-pairing of short exonic regions with downstream intronic regions promotes editing of select adenosines within the coding regions of many ADAR target mRNAs in vivo.²⁶ This specificity is thought to arise from helical disruptions of imperfect base-pairing that limit the number of binding modes available to the dsRBDs of ADARs.²⁷ Accordingly, this binding limitation should not only restrict the specific adenosines that undergo deamination, but also increase the frequency at which a specific site is deaminated; both of which have been demonstrated in vitro.^25,28 Long, nearly perfect RNA duplexes would not have this limitation, and in vitro studies indicate that many adenosines are deaminated across perfect duplexes.^22,23 At the most extreme, over 50% of the adenosines on each strand of 100 bp duplex RNA could be deaminated in vitro,^22,29 a process commonly referred to as hyper-editing. For these reasons, A-to-I editing of long completely base-paired substrates is often referred to as promiscuous deamination.²⁰ However, editing in long duplexes is not random in vivo. Some adenosines are highly edited, while others are not edited. For example, editing of specific adenosines within Alu repeats located in long double-stranded regions of human mRNAs occurs to a similar extent among different human individuals.³⁰ In addition, despite the presence of 2 homologous editing enzymes in mammals, certain adenosines in repetitive elements are specifically edited by only one of the ADAR enzymes.³¹ Furthermore, editing site selection within noncoding regions appears to be regulated. For example, in C. elegans, the double-stranded RNA binding protein, ADR-1, can promote editing at specific adenosines within long double-stranded regions of mRNA.³²

Despite the known importance of editing in gene regulation, the mechanisms that regulate editing site choice in long RNA duplexes are still not understood. Furthermore, although editing studies in many organisms have identified multiple editing sites in long double-stranded regions of mRNAs, it is unknown the extent to which these editing sites co-occur on individual transcripts in vivo. This gap in knowledge is primarily because most next-generation sequencing studies to date have focused on identifying transcriptome-wide editing events using short read lengths. Therefore, only small portions of ADAR target mRNAs are analyzed at one time, which prevents identification of the editing pattern for the entire transcript. Furthermore, most mapping strategies utilized in transcriptome-wide sequencing studies cannot properly align an extensively edited transcript to the reference genome. One study searching for hyper-edited RNAs utilized an approach of a 3-letter genome, where all adenosines are replaced with guanosines before alignment, and identified nearly 700 genomic regions that had greater than 12 editing sites.³³ However, the vast majority of these regions (>92%) had only one sequencing read that was hyper-edited, suggesting that this phenomenon is rare in vivo.

To begin to understand how multiple editing sites in noncoding regions are edited together on individual mRNA molecules, we performed an ultra-deep sequencing analysis (>10,000X read coverage per biological replicate) of 4 known C. elegans mRNAs. In contrast to mammals, ADR-2, the sole A-to-I editing protein in C. elegans, is not essential.¹¹ The use of adr(−) worms allowed for the confident identification of even rare editing events (less than 1.0%) in endogenous mRNAs. This highly sensitive analysis detected editing at 95 sites (out of 279 adenosines in double-stranded regions) across 4 transcripts: syntaxin, Y67D2.7, Y75B8A.8 and laminin. Most importantly, 3 of the 3′UTRs were sequenced in their entirety to allow for the first detailed analysis of synchronous editing events in long, noncoding ADAR targets. Although more than 50% of adenosines in some genes were identified as editing sites, the average amount of editing per transcript was found to be very low, ranging from 2.0–6.3% deamination among the 3 genes (1.1 to 4.4 editing events per transcript). Furthermore, the majority of reads (68%) contained less than 5% deamination, indicating that hyper-editing is rare among these genes in vivo. Correlation analyses revealed the existence of coupling among select edited sites and a strong negative correlation between other editing sites, which impacted the overall pattern of editing events on a given transcript. This data can be used in future studies to determine the functional role of specific combinatorial editing events in 3′ UTRs, as well as to predict noncoding editing patterns in vivo.

Results

Identification of A-to-I editing sites in the 3′ UTRs of endogenous C. elegans mRNAs

To determine the transcript complexity and overall pattern of noncoding editing by ADARs, we performed high-coverage, next-generation sequencing on the 3′ UTRs of 4 previously identified C. elegans ADAR substrates.^32,34 Three of the genes, Y67D2.7, Y75B8A.8 and syntaxin (unc-64) were selected because the 3′ UTR region containing the known editing sites could be sequenced in their entirety on the Illumina MiSeq platform. The fourth substrate, laminin (lam-2), was used as a positive control as its editing landscape has been extensively characterized using Sanger sequencing.³² However, due to the limitations of sequencing read length on the next generation sequencing platform, only half of the known edited region within the laminin 3′UTR was sequenced.

For each gene, multiple biological replicates of RNA extracted from wild type and adr(−) adult worms were reverse transcribed with gene-specific primers and PCR amplified for the target double-stranded region. As adr(−) worms lack A-to-I editing,¹¹ RNA-seq data from this strain was used as a negative control to distinguish true A-to-I editing events from single nucleotide polymorphisms present in the C. elegans strains, as well as sequencing, reverse transcription and low-frequency PCR errors. The pools of cDNAs were combined, barcoded and subjected to next generation sequencing. The resulting 250 nucleotide paired end reads were overlapped to generate a single long read and then aligned with the published genomic reference sequences (WS220, ce10) (Fig. S1A–D). Any reads that did not contain high quality base-calling across the entire reference sequence or contained any changes from the reference other than A-to-G were eliminated from the final pool (see Materials and Methods for in-depth description of bioinformatics analysis). Depending upon the gene and sample, read depths (per replicate) varied from a minimum of 6,693 reads to a maximum of 25,230 reads, with an average of 12,530 reads per gene, per strain (Table S1).

Candidate A-to-I editing sites were identified by the presence of guanosine residues in the RNA-seq reads where adenosine existed in the genomic reference. Inosine has similar base-pairing properties as guanosine, therefore A-to-I editing changes in RNA will appear as A-to-G changes between the reference DNA and the cDNA.³⁵ The frequency of reads containing G in place of the genomically encoded A in the negative control adr(-) RNA-seq dataset averaged 0.06% and ranged from 0–5.43% across all transcripts (Table S2). This level of noise was similar to the A-C (0.07%) and A-T (0.025%) mismatches detected in the adr(-) RNA-Seq data set and likely represents the transcription, reverse-transcription, PCR and sequencing noise (Fig. S2). Due to the variability in the frequency of reads containing G in place of A in the negative control, a universal minimum editing percentage was not employed. Instead, both a 2-tailed Welch's t-test (p-value <0.025) and fold change (>10X) of the percent G occurring at specific sites between wild-type and the adr(-) datasets were used to identify high confidence A-to-I editing sites. In sum, 95 sites were identified, which range in editing frequency from 0.09–91.19% across the 4 transcripts in the wild-type RNA-seq data set (Fig. 1A–D). All of the editing sites in the laminin region previously identified by Sanger sequencing were also detected in this RNA-seq dataset. In addition, our approach identified 52 novel editing sites across the 4 genes (noted by asterisk in Figs. 1 and 2). All of the novel sites are edited less than 15%, with an average editing of 2.3%, suggesting that this approach results in a more sensitive detection of RNA editing sites than other previously used techniques.

Figure 1.

Wild-type editing levels in 3′ UTRs of endogenous mRNAs. (A–D) The percent A-to-I editing at individual nucleotides within the 3′ UTRs of the indicated genes was measured for multiple biological replicates of young adult worms (n = 3 for Y67D2.7, syntaxin, laminin and n = 2 for Y75B8A.8). Error bars represent standard error of the mean (SEM). Novel editing sites identified are marked with an asterisk.

Figure 2.

Predicted location of edited adenosines across the 3′ UTRs. MC-Fold was used to generate the structural predictions for the main double-stranded region present in the 3′ UTRs of the indicated genes. Edited adenosines are marked with a colored line. The color of the line indicates the average editing level of each adenosine in the wild-type RNA-Seq data set; <1% editing (blue), between 1% and 10% editing (green), between 10% and 50% editing (yellow) and >50% editing (red). Novel editing sites are marked with an asterisk.

Biological properties of C. elegans editing sites and ADR-2 preferences

ADAR substrates are completely, or largely double-stranded RNAs and editing sites are only known to localize in double-stranded regions.³⁶ To map the location of the newly identified editing sites within these substrates, the secondary structure of the double-stranded regions of each gene were predicted using both MC-fold³⁷ (Fig. 2) and m-fold³⁸ (data not shown). As the double-stranded structure of laminin extended beyond the sequenced region, the secondary structure was predicted based on the entire known edited region.³⁴ Regardless of the secondary structure prediction program, mapping of all the identified editing sites revealed that 92 of the 95 sites localize within the major double-stranded region predicted for each 3′ UTR, while the other 3 exist within smaller double-stranded regions of the syntaxin 3′ UTR (Fig. 2). The number of editing sites identified for each 3′ UTR does not directly correlate with the length of the double-stranded region. This is most obvious in the fact that the gene with the shortest double-stranded region, syntaxin, has the most editing sites. Furthermore, the number of editing sites in the Y75B8A.8 3′ UTR (8) is almost four-fold lower than the number of editing sites present in a long double-stranded region of syntaxin (28), yet the 2 regions differ in length by only one nucleotide and contain a similar number of adenosine residues (55 and 53, respectively, Table 1). Therefore, the length of the double-stranded region is not a good predictor for the number of editing events that will occur within a particular substrate in vivo.

Table 1.

Characteristics of editing across the double-stranded 3' UTRs

Gene	Length of ds Region (nt)	# of A's in ds Region	# of A's Edited	Maximum # editing events per read	Average # editing events per read	# of Patterns Observed	# Patterns Expected
Y67D2.7	260	69	22	15	4.35	3537	8174
Y75B8A.8	178	55	8	4	1.09	34	58
syntaxin*	175	53	28	16	2.96	4146	10924

Legend: The double-stranded region was defined by the MC-Fold secondary structure output to determine the length and number of adenosines. Number of patterns represents those seen in the wild-type strain or were expected to appear given the read depth and editing site percentages of all editing sites. *For syntaxin, only the major double-stranded region is being considered in the first 3 columns.

The efficiency of editing at a given adenosine is thought to be influenced by nucleotides immediately surrounding the editing site.²⁰ The vast majority (88.4%) of the editing sites in all 4 3′ UTRs are opposite of a uridine (U) (Fig. 2), suggesting that editing in C. elegans 3′ UTRs most often disrupts the double-stranded structures by generating I-U mismatches. To determine other nearest neighbor nucleotide preferences for C. elegans ADR-2, the percent occurrence for the 5′ and 3′ neighbors were weighted according to the editing percentage of each edited adenosine in the wild-type RNA-seq data set. As the dataset contains editing sites that vary from 0.1%–91% editing, this approach gives a larger representation to nucleotides adjacent to sites that are edited more frequently compared to those with low editing percentages. To determine whether or not the appearance of particular nucleotides in the nearest neighbor positions were statistically significant (p < 0.005), a 2-sample logo analysis³⁹ was used to compare the weighted, edited adenosines to the sequence context of all adenosines in the double-stranded region. This analysis revealed that uridine (U) is over-represented 5′ to the edited adenosine, while guanosine (G) is over-represented 3′ to the edited adenosine. Additionally, cytidine (C) and G are under-represented 5′ and C and U are under-represented 3′ to the edited adenosine (Fig. 3A). These preferences are similar to those found for human ADARs;⁴⁰ however, the 3′ position has overall higher enrichment values compared to the 5′ neighbor suggesting the 3′ neighbor may have a more significant role in adenosine selection by the C. elegans editing enzyme. To test this possibility, the triplet sequences centered on each edited adenosine were analyzed for enrichment or depletion over the triplet sequences that existed around every adenosine in the double-stranded regions (Fig. 3B). The most highly over-represented triplet for all genes contains a U 5′ and G 3′ to the edited adenosine as expected from the independent nearest neighbor analysis. Furthermore, the 2 most over-represented and 3 most under-represented triplets contain the most over- or under-represented 3′ nucleotide, supporting the idea that, in C. elegans, the 3′ nucleotide has a stronger influence on deamination efficiency than the nucleotide 5′ to the edited adenosine.

Figure 3.

C. elegans editing sites exhibit nearest neighbor preferences as other organisms. (A) A 2 logo analysis was used to determine the nearest neighbor preferences of all editing sites weighted by their editing percentage compared to all adenosine residues present in the double-stranded region. Enriched and depleted nucleotides are shown above and below the axis, respectively. The level of conservation is represented by letter height. Logos were generated using a Student's t test with p < 0.005 and no Bonferroni correction. (B) Triplet sequences containing each editing site (weighted based on their editing percentage) were compared to triplet sequences containing all adenosine residues in the double-stranded region. The frequency of occurrence of a given edited triplet was normalized to the frequency of occurrence of the triplet in the double-stranded region. The log₂ ratio of these normalized values (referred to as fold change) is plotted on the Y-axis. Enriched and depleted triplets are shown above and below the axis, respectively.

Hyper-edited mRNAs are rare transcripts in C. elegans

The major advantage of sequencing the entire double-stranded region of a 3′ UTR on the next generation platform is that, for the first time, the pattern of editing events that co-occur within a noncoding region of mRNA can be observed. The number and location of the editing sites identified above were determined for each individual sequencing read of syntaxin, Y75B8A.8 and Y67D2.7 to generate a summary of editing patterns for the wild-type RNA-seq data set (Table S3). Editing patterns of laminin were not included due to the fact that only half of the double-stranded region was sequenced. The number of editing events per read varied for each gene and did not directly correlate to the overall number of edited adenosines in a given gene (Fig. 4A, Table 1). Overall, the average amount of editing per transcript was very low and ranged from 2.0–6.3% deamination among the 3 genes (1.1 to 4.4 editing events per transcript) (Table 1). While the maximum deamination observed among all transcripts was 22% (15 editing events in Y67D2.7), this only occurred in 0.003% of wild-type sequencing reads. In fact, for all 3 genes combined, only 0.8% of the reads exhibited more than 10% deamination, and the majority of reads (68%) contained less than 5% deamination.

Figure 4.

C. elegans 3' UTRs exhibit low levels of editing per transcript. (A) The number of editing sites per sequencing read was determined for each gene from the sum of all wild-type biological replicates. The number of reads containing a given number of editing sites was divided by the total number of reads in the wild-type RNA-Seq dataset and plotted as percent reads. In addition to the plotted values, 15 editing sites were observed in the Y67D2.7 3′ UTR in 0.003% of reads and 16 editing sites were observed in the syntaxin 3′ UTR in 0.0027% of reads. (B) Relative mRNA expression for the indicated genes was quantified by qRT-PCR in the wild-type (gray) and adr(-) (white) worms. Graph illustrates the average fold change of each mRNA relative to gpd-3 mRNA and normalized to the wild-type level. Error bars represent standard error of the mean (SEM) for 3 biological replicates.

It is possible that highly edited RNAs are degraded in vivo and are therefore largely absent from our wild-type RNA-seq dataset. In fact, dsRNAs hyper-edited in vitro (>50% deamination) or synthetic hyper-edited RNAs can be specifically cleaved by Tudor staphylococcal nuclease (Tudor-SN).^41,42 To directly test this possibility, the mRNA levels of the 3 genes were measured in wild-type and adr null worms using quantitative real-time PCR (qRT-PCR). No significant difference in transcript levels was detected between wild-type and adr(-) worms despite the absence of editing in the adr(-) worms (Fig. 4B). Therefore, while upwards of 50% of the total number of adenosines in a double-stranded region can be edited in vivo, the number of deamination events that co-occur on an individual mRNA is much lower, suggesting that hyper-editing of C. elegans transcripts rarely occurs in vivo.

Patterns of editing events in 3′ UTRs are not random

As very few editing sites are edited at 100% efficiency, the co-occurrence of multiple editing sites has been proposed to contribute to transcriptome diversity.^43,44 Although, in theory, the total number of editing sites across the 3′ UTR (N) could generate a total of 2^N unique transcripts, the differential editing levels observed for each editing site limits the realistic probability of all possible combinations occurring in our data set. Using the information regarding the number and location of editing sites per sequencing read for syntaxin, Y75B8A.8 and Y67D2.7 (Table S3), the number of patterns observed for each gene was determined by counting all different patterns that occurred in 1 or more read (Table 1). In addition, the expected number of reads for each pattern was calculated based on the probability that multiple adenosines would co-occur given their individual editing site percentages, and the assumption that editing at multiple sites is independent of one another. Based on the read depth for each gene, all editing patterns that had a probability of occurring at least one time in the wild-type dataset were counted to generate the expected number of patterns (Table 1). For all 3 genes, the number of different transcripts observed was approximately two-fold lower than the expected number of combinations (Table 1). These data suggest that certain editing sites are more likely to co-occur on a given transcript than others, and that editing of multiple sites in a 3′ UTR does not occur randomly.

Editing of adjacent adenosines are coupled, but not a result of sequential deamination

A dice correlation co-efficient was used to identify pairs of sites that co-occur on individual RNA-seq reads more often than they occur independently. The dice correlation co-efficients were calculated for the co-occurrence of any 2 identified editing sites across all observed patterns for the 3 genes. To test whether these frequently co-occurring sites exhibit interdependence for editing or occur at the frequency one would expect given their individual editing site percentages, the observed dice correlation coefficient was normalized to the expected dice value by taking the log₂ ratio of the observed/expected value. All dice values reported throughout the manuscript indicate this normalized ratio (see Materials and Methods for a description of this calculation). For all 3 genes, the combination of sites with the most positive dice co-efficient was comprised of 2 directly adjacent adenosines (Table S4). Furthermore, even adjacent sites which exhibit very low editing frequencies (such as sites 265 and 266 in syntaxin, 3.3 and 1.6% respectively) show strong correlation values (3^rd highest) indicating that positively correlated sites are not simply a result of the combined probability of 2 highly edited sites. These data suggest that ADR-2 has a strong preference to edit an adjacent adenosine before editing other sites along the 3′ UTR.

The strong positive correlation for editing of adjacent sites is further exemplified by the fact that, across the 3 genes, all adenosines edited in greater than 10% of reads that also contain an adjacent adenosine exhibit some level of editing at the adjacent site. However, it should also be noted that there are other appearances of multiple adjacent adenosines in the double stranded regions that are completely unedited, such as a stretch of 5 adenosines in Y75B8A.8 (Fig. 2). Among the 3 transcripts, editing of multiple adjacent adenosines ranged in complexity from 2 adjacent editing sites (8 sets) to a stretch of 4 adjacent sites (one occurrence). The majority of editing events within this context have one site that has the highest editing frequency among the group. However, the location of the highest frequency editing site appears nearly equally between the 5′ and 3′ position (Fig. 2).

Coupling of editing at adjacent adenosines has also been observed in coding regions of human ADAR targets.⁴⁵ Editing of these sites was proposed to occur by attraction of the ADAR to a primary site of deamination and then sequential editing of adjacent sites due to slipping of the enzyme along the dsRNA. However, it is also possible that specific sets of adjacent adenosines are located near high affinity ADAR binding sites and multiple editing reactions occur due to flexibility in the catalytic domain rather than movement from the primary binding site or dissociation of the enzyme from dsRNA after the initial editing event. To directly test these hypotheses, we designed a reporter expressing the double-stranded region of the Y75B8A.8 3′ UTR fused to the coding region of green fluorescent protein. Despite being over 150 bp in length, the Y75B8A.8 3′ UTR contains only 2 adenosines that are edited with greater than 10% efficiency. These two adenosines are located adjacent to one another and exhibit a strong positive correlation (Table 2). Using the transgenic worms, editing of the reporter mRNA was analyzed by Sanger sequencing of a reporter-specific reverse transcription and PCR product. Editing of the wild-type reporter exhibited a similar pattern as endogenous Y75B8A.8 mRNA, although both sites were edited at a lower frequency than the endogenous gene (Fig. 5). In addition, a reporter containing an adenosine to guanosine mutation at nucleotide 228 (the highest edited site in the Y75B8A.8 3′ UTR) was generated. This mutant reporter mRNA was edited at the adjacent adenosine (site 227) slightly more frequently (8.4%, p-value = 0.06) than the wild-type reporter (Fig. 5). This result was confirmed with 3 biological replicates each from 3 different transgenic worm lines. Together, these data suggest that editing of adjacent adenosines occurs due to a preferential ADR-2 binding site allowing deamination at multiple adenosines and not a result of sequential deamination from enzyme slippage.

Table 2.

Characteristics of anti-correlated editing sites

Y67D2.7 Sites (Editing Percentage)	Observed	Expected	Dice Value	Distance (bp)	Same Side of Helix?
136 (18.5%)–287 (20.4%)	384	1034	−0.15	5	Yes
133 (12.2%)–293 (15.3%)	49	468	−0.15	1	No
136 (18.5%)–281 (15.3%)	377	754	−0.10	10	Yes
136 (18.5%)–293 (15.3%)	518	757	−0.07	2	Yes
113 (41.5%)–293 (15.3%)	1547	1805	−0.04	21	Yes
133 (12.2%)–287 (20.4%)	615	669	−0.02	8	Yes
145 (78.8%)–281 (15.3%)	3335	3497	−0.01	3	No
syntaxinSites (Editing Percentage)	Observed	Expected	Dice Value	Distance (bp)	Same Side of Helix?
136 (55.6%)–188 (30.4%)	4397	5968	−0.11	10	Yes
145 (8.6%)–178 (79.5%)	1940	2276	−0.03	9	Yes
122 (8.1%)–208 (6.3%)	128	157	−0.02	3	No

Legend: Pairs of negatively correlated sites are ranked by the Dice value. Dice values reported are the log₂ ratio of the observed/expected value (see materials and methods for calculation). Editing sites are reported with their corresponding editing site percentage in parenthesis. Distance was calculated from the secondary structural prediction of the MC-Fold algorithm and the side of the helix was determined from the teriary prediction using MC-Sym.

Figure 5.

Editing of adjacent adenosines does not require sequential deamination. Total RNA from worms expressing reporters for either the wild-type double-stranded regions of Y75B8A.8 or one containing an A-to-G change in the DNA at nucleotide 228 of the 3' UTR was reverse transcribed, PCR amplified with reporter or endogenous specific primers and Sanger sequenced. In addition, wild-type genomic DNA was PCR amplified and sequenced as a negative control. Sequencing chromatograms of each PCR product are shown. The bold A indicates adjacent editing sites 227 and 228.

Anti-correlated editing sites affect the overall pattern of editing across the 3′ UTR

The observation that all 3 genes express fewer editing patterns than expected suggests that in addition to a positive correlation for select adenosines, the co-occurrence of certain editing sites may be negatively correlated. In order to accurately determine the correlation between editing sites across the 3′ UTRs, all adjacent editing sites were combined and considered as one single event in a given pattern, denoted by the 3′ adenosine of the group (Table S5, see Materials and Methods for a detailed list of combined sites). This method allows correlation between non-adjacent sites in a particular pattern to be measured without influence of the high correlation rate observed for adjacent sites. The dice correlation co-efficient was then recalculated for both the observed and expected co-occurrence of all identified editing sites that exhibit greater than 5% editing in the combined patterns. Dice correlations were not calculated for Y75B8A.8, as it only contains 3 editing sites above this threshold; 2 of which are positively correlated, adjacent adenosines. For syntaxin, 3 pairs of editing sites exhibited anti-correlation (Table 2), including 2 very highly (>55%) edited sites (nucleotides 136 and 178). There are 7 pairs of adenosines that are anti-correlated in Y67D2.7, of which contain 2 of the most highly (>40%) edited sites (nucleotides 113 and 145). Of course, one straightforward explanation of these anti-correlated sites is that they occur on transcripts expressed in different cell types. However, there are a substantial number of sequencing reads containing both of the anti-correlated sites (ranging from tens to thousands, Table 2), but the number of observed reads is significantly lower than expected based on the editing frequencies of the individual sites.

To better understand why these editing sites are anti-correlated, the tertiary structure of the edited region for syntaxin and Y67D2.7 was predicted using the MC-Fold and MC-Sym algorithm³⁷ (http://www.major.iric.ca/Web/mctools). These assign and score small RNA building blocks observed in NMR and crystallographic data that best accommodate the sequence, which is then assembled into secondary and tertiary structures. The location of the editing sites on the predicted tertiary structures suggest that there is no apparent relationship between negatively correlated pairs and their orientation to one another on the side of the helix (Table 2). However, the majority of pairs were within 10 bp of one another, suggesting perhaps some kind of interference at this spacing (Table 2). However, it should be noted that there were also positively correlated pairs of editing events that were within 10 bp of one another (ex. nucleotides 178 and 188 in syntaxin). Future biochemical studies need to be done to address the driving factors behind these different outcomes.

To determine whether the pairs of anti-correlated sites impact the overall editing patterns across the 3′ UTR, a binomial distribution was employed to compare the observed and expected read counts for each pattern. As the dice correlation is only valid for 2 sites, the binomial distribution was used for more complex patterns to determine whether or not each pattern occurred significantly more or less than expected if the editing events were independent. Due to the large number of patterns observed for each gene, the binomial distribution values were Bonferroni corrected. It should be noted that usage of a Bonferroni corrected p-value is considered to be quite conservative, therefore, we have high confidence in the patterns identified as significantly observed more than expected (referred to as selected for) or observed less than expected (selected against). An editing site tree was generated to map positively and negatively selected editing patterns (Fig. 6). For simplicity purposes, any patterns containing sites edited less than 5% were not included on the tree. Additionally, once a pattern was mapped as selected against, all subsequent patterns containing those sites and also selected against were not mapped. Importantly, for syntaxin there was never an instance in which a pattern that was selected for could not be logically mapped from a path of previously selected for patterns as the number of sites increased. The editing patterns of Y67D2.7 were less straightforward and for this reason were not mapped in the tree format.

Figure 6.

Successive editing events within a 3' UTR are affected by previous editing events. Binomial distribution values were calculated for editing site patterns comprised of individual sites edited greater than 5%. Green boxes indicate patterns that occur more than expected (selected for) and red boxes indicate patterns that are selected against. The tan box indicates a pattern that was neither selected for nor against, but necessary to include for mapping of successive patterns. Numbers inside the box indicate the nucleotide edited followed by the number of reads were the pattern was observed in parentheses. Asterisks indicate patterns that are mapped in more than one location because they could logically follow from more than one path of previously mapped patterns.

By examining the paths of the editing events, it is clear that there are certain sites that are selected to occur in a specific combination. For example, editing at nucleotide 122 is selected for in the branch containing editing at nucleotides 136 and 178, but not in the branch containing editing at nucleotides 178 and 188. Similarly, editing at nucleotides 191 or 233 is selected for when nucleotides 178 and 188 have been edited, but selected against when editing occurs at nucleotides 178 and 136. As nucleotides 136 and 188 are anti-correlated (Table 2), these data suggest that not only are these events occurring together less than expected, but they also influence the selection of additional editing events. Similarly, certain editing events, such as those at nucleotide 95 are selected against when present as the only editing event or even a combination of 2 editing sites, but are selected for in patterns that contain 3 or more editing events (Fig. 6). Thus, there are certain preferences to the order and combination in which multiple editing events occur within the 3′ UTRs.

Discussion

In this study, we performed an ultra-deep sequencing analysis of the double-stranded regions of 4 endogenous C. elegans 3′ UTRs that are known ADAR targets. Many current transcriptome-wide studies employ a universal minimum editing site percentage to identify A-to-I editing events. However, we found that sequencing endogenous mRNAs isolated from wild-type worms and those that lack A-to-I editing allowed for a more accurate, site-specific determination of editing sites based on the variation measured in the negative control. Our in-depth analysis detected editing at 95 sites, 52 of which were novel editing sites (Fig. 1). The adenosines range in editing frequency from 0.09–91.19%, and all of the novel sites are edited less than 15%, with an average editing of 2.3%, suggesting that this approach provides the most sensitive method for identifying RNA editing sites to date.

Our analysis of over- and under-represented nucleotides surrounding the A-to-I editing sites revealed similar preferences as those observed in both in vitro and in vivo studies of human ADARs.^4,40 This is particularly intriguing as C. elegans ADR-2 is unique among the ADAR family members in that it contains only one double-stranded RNA binding domain (dsRBD) and the sequence of it's catalytic domain is the most distantly related to human ADARs when compared to other invertebrates, namely squid, fly and hydra ADARs.² It is important to note that over- and under-represented nucleotides reported here differ from the transcriptome-wide preferences previously reported for ADR-2.³² However, those preferences were obtained from a comparison to a random list of sequences and therefore likely reflect properties of double-stranded regions containing editing sites, rather than specifically the edited adenosines. Interestingly, when all editing sites were treated equally, no significant enrichment or depletion of nucleotides in either position was found. However, when nucleotides were weighted based on editing frequency, we identified clear nearest neighbor preferences. This indicates that neighboring nucleotides do no strictly preclude editing, but rather influence editing frequency. This finding is consistent with a recent ultra-deep sequencing of small segments of human mRNAs, which reported that nearly every adenosine in a double-stranded region was capable of deamination, but over- and under-represented nucleotides appear when focusing on the most highly edited sites.⁵ In addition, similar to that study, we observed the appearance of over- and under-represented nucleotides when focusing on the most highly edited sites. These preferences are consistent with recent biochemical data from human ADARs indicating that nucleotides adjacent to edited adenosines influence the ability of the adenosine to flip out of the duplex, a key step in the deamination reaction.²¹

In addition to determining the characteristics of the individual editing sites, our ability to sequence the entire double-stranded region of 3 genes allows examination of synchronous editing across an entire 3′ UTR. One of the most surprising results from this study is that the number of editing events occurring together on an individual transcript is significantly lower than the total number of editing sites that were identified for each 3′UTR. The average number of editing sites per transcript ranged from 1.1 to 4.4 sites, while the total number of editing events identified in the genes ranged from 8 to 31. For all 3 genes combined, the majority of reads (68%) contained less than 5% deamination and only 0.8% of the reads exhibited more than 10% deamination. Therefore, unlike the hyper-editing (>50% deamination) that occurs on long duplexes in vitro,^22,29 the editing of long duplexes in C. elegans is more restrictive.

Hyper-editing of synthetic dsRNAs has previously been implicated in promoting degradation by the nuclease Tudor-SN or nuclear retention by binding to p54^nrb.^41,42,46 However, whether these 2 pathways function to regulate edited RNAs in vivo is unknown. As we rarely detect hyper-edited transcripts in this study, one possibility is that Tudor-SN degrades these mRNAs in wild-type worms. However, a lack of editing did not increase expression of the mRNAs in this study (Fig. 4B) or mRNAs analyzed in a previous C. elegans study,⁴⁷ it is unlikely that Tudor-SN targets a significant number of C. elegans mRNAs for degradation. An intriguing possibility is that Tudor-SN regulates expression of hyper-edited mRNAs in distinct cell types within the worm. Since our approach utilizes mRNA isolated from whole worms, effects on mRNA expression in a small subset of cells would be undetected. In regards to nuclear retention, while a handful of genes that express hyper-edited transcripts have been observed in mammalian nuclei,^48,49 the proportion of transcripts expressed from these genomic loci that undergo hyper-editing is unknown. Furthermore, our finding that hyper-editing is rare is consistent with studies demonstrating that both human and worm mRNAs containing long duplexes are exported from the nucleus and localize to translating ribosomes, irrespective of their editing status.^47,50 In addition, recent high-throughput sequencing studies of cytoplasmic and nuclear RNA from human cell lines show no evidence for editing-dependent subcellular localization.⁵¹ Hyper-editing is rarely detected in the transcripts we analyzed in this study; therefore, we propose that editing in 3′ UTRs leads to specific mechanisms of gene regulation, rather than a global repression of gene expression. Consistent with this hypothesis, a recent study demonstrated that one specific editing event in the 3′ UTR of the human ARHGAP26 mRNA prevents miRNA binding and translational repression in normal brain tissue and the lack of editing in brain cancers leads to increased protein expression.⁵² Although a similar finding is lacking in C. elegans, as our current analysis revealed that over 95% of the edited transcripts of Y75B8A.8 exhibit editing of one or 2 specific adjacent adenosines in the long (>70 base-pair) double-stranded 3′ UTR, future work will be aimed at understanding the function of these editing events in regulating Y75B8A.8 expression in vivo.

Our analysis of synchronous editing events demonstrated that certain editing sites are more likely to co-occur on a given transcript than others, and that editing of multiple sites in a 3′ UTR does not occur randomly. By far, the strongest positive correlations were between immediately adjacent adenosines. Consistent with this, a preference for an editing site to be preceded or followed by another editing site was observed in one of the first transcriptome-wide studies of human ADAR substrates.⁵³ Furthermore, coupling of editing at adjacent adenosines has also been observed in coding regions of human ADAR targets.⁴⁵ Our data not only extends this observation to noncoding regions, but our mutational analysis demonstrates that editing of adjacent adenosines does not result from the enzyme slipping to the adjacent adenosine after one deamination event.

Sequencing the entire double-stranded region from each 3′ UTR allowed us to not only determine correlations between nearby sites, but also determine, for the first time, whether specific patterns of editing sites occur across the entire 3′ UTR more frequently than would be expected by random chance. With use of the binomial distribution and an editing tree to depict the positive and negative correlations among editing sites, we postulate that multiple editing events arise due to multiple ADAR molecules binding the dsRNA rather than processive deamination along the dsRNA. This hypothesis is in agreement with previous studies on the coupling of editing sites within short imperfect duplexes present in coding regions of mRNA.^45,54 However, in contrast to those studies, the positively correlated sites we observe in 3′ UTRs do not occur in regular intervals along the duplex. Therefore, it is unlikely that cooperative interactions occur between ADAR molecules binding sequentially along the RNA. Furthermore, the fact that extensively edited transcripts are rare suggests that cooperativity between ADAR molecules does not occur. In addition, the anti-correlated pairs of sites suggest that short distances between sites (<10 bp) can negatively affect deamination events. Based on structural and biochemical data of human ADARs,^19,55 we predict that binding of 2 ADR-2 molecules is prohibited if the minor groove binding site for the β1-β2 loop of one dsRBD overlaps the second deamination site. However, it is important to note that negatively correlated sites do occur together on a handful of transcripts, suggesting that once a deamination event occurs the ADAR molecule could fall off the helix and a second molecule could bind and deaminate the other adenosine of the anti-correlated pair. Consistent with this idea, recent work in Drosophila has demonstrated that editing events located in close proximity (∼10 nucleotides apart) are not functionally coupled and require binding of ADARs to 2 distinct locations on the dsRNA structure.⁵⁶ Furthermore, early biochemical work supports the hypothesis of re-binding and editing by ADARs, as partially edited RNAs re-introduced to ADARs in vitro were capable of undergoing further deamination.²³

In addition to the pair-wise editing events, the correlation analysis between the overall co-occurrence of 2 adenosines, revealed that distant editing events separated by multiple helical turns, such as nucleotides 95 and 178 in syntaxin, were positively correlated (dice coefficient of 0.05). Multiple studies characterizing editing events in the coding and noncoding regions of human mRNAs have also previously detected these long-range positive correlations, however, the mechanism of ADAR recognition that resulted in these coupled editing events was unclear.^44,54,57 Our ability to analyze the editing pattern across the 3′ UTR revealed that although the overall co-occurrence of editing sites that are separated by multiple helical turns is positively correlated, when focusing on transcripts that contain only 2 editing events, these distant adenosines co-occur less than expected (dice coefficient of −0.09 for nucleotides 95 and 178 in syntaxin). This data suggest that there are certain preferences to the order in which multiple editing events occur within the 3′ UTRs. In support of this, the patterns of editing across the syntaxin 3′ UTR (Fig. 6) revealed that there are certain sites that are selected to occur in a specific combination of editing events. However, how ADARs choose these specific patterns and the cellular factors that influence the selection remains to be determined.

Materials and Methods

Worm strains and maintenance

The wildtype Bristol strain N2 and adr(-) strain BB21 adr-1(tm668);adr-2(ok735) were used in this study. Transgenic worm lines were generated by microinjection into the gonads of wild-type, young adult worms. The injection mix used for generating transgenic lines contained the following: 20ng/μl of the transgene of interest, 20ng/μl of the co-injection marker (rol-6), and 60ng/μl of 1kb DNA ladder (NEB). Transgenic strains were maintained by passaging only worms expressing GFP and the co-injection marker. The markers used in this study were rab3::gfp::unc-54 and the rol-6(su1006) plasmid, both of which were previously described.^47,58,59 Worm strains were maintained by growth on NGM plates seeded with Escherichia coli OP50.

Reporter constructs

The double-stranded region of the Y75B8A.8 3′ UTR was cloned downstream of green fluorescent protein driven by the pan-neuronal rab-3 promoter. Mutations to the 3′ UTR of Y75B8A.8 were generated by PCR and confirmed by Sanger sequencing.

Flow cytometry

To prepare for flow sorting, worms were starved for 1 week on 5 cm NGM plates, chunked onto 15 cm NGM plates and harvested 48 hours later. Flow cytometry was conducted at the IUB Flow Cytometry Core Facility by a dedicated technician using the COPAS Select (Union Biometrica) large particle sorter. Parameters were adjusted manually to select either only adult worms for non-transgenic strains or adult worms expressing GFP for transgenic lines. Prior to sorting, worms were washed thoroughly with 1X M9 buffer to remove bacteria and debris. Sorted worms were collected onto OP50 seeded NGM plates. After sorting, worms were immediately washed twice with 1X M9 buffer, resuspended in Trizol and frozen in liquid Nitrogen for RNA extraction.

RNA isolation, reverse transcription

Total RNA was isolated using Trizol (Life Technologies). RNA was further treated with Turbo DNase (Ambion) and then isolated using the RNA Easy Extraction kit (QIAGEN). Reverse transcription was performed with Thermoscript (Life Technologies) using gene-specific primers. Incubating at 85°C for 10 minutes stopped the reverse-transcription reaction and remaining DNA/RNA hybrids were removed by treatment with RNaseH (NEB). cDNAs were amplified by 2 rounds of PCR using nested primer sets and PFX Platinum DNA Polymerase (Life Technologies), a high fidelity DNA Polymerase. A small amount of PCR product was run on an agarose gel to ensure the RT-PCR was successful and the rest of the sample was purified with a PCR-purification kit (QIAGEN). All reverse-transcriptions were done concurrently with a negative control that did not contain Thermoscript to ensure all DNA came from cDNA amplification only.

Real-time PCR

RNA isolated as outlined above was reverse transcribed with Superscript III (Life Technologies) using gene-specific primers spanning an exon-exon junction of each gene. Incubation at 70°C for 15 minutes stopped the reverse-transcription reaction and, then, remaining DNA/RNA hybrids were removed by treatment with RNaseH (NEB). 2 μl of dilute cDNA was amplified by 40 cycles of PCR and quantified relative to a standard curve for each gene. A melting curve was measured for each sample to ensure the existence of only one product from the PCR reaction. The cDNA levels were normalized to gpd-3, a non-ADR target gene.

Library preparation and RNA sequencing

Equimolar amounts of PCR product from each gene were combined for a total of 1.2 μg of DNA. Input DNA was end-repaired with T4 DNA Polymerase (NEB), Klenow DNA Polymerase (NEB), and T4 PNK (NEB) at 20°C for 120 minutes. After heat inactivation, cDNA was purified from the reaction mixture with a PCR purification kit (QIAGEN). Pure cDNA was then 3′ adenylated with Klenow Fragment (3′ to 5′ exo-) (NEB) at 37°C for 120 minutes. The reaction was purified with a MinElute PCR purification kit (QIAGEN) before ligation of sequencing adapters. Sequencing adapters were prepared by annealing a unique indexed adapter to the universal adapter (40 μM) in a 3:1 ratio with Tris (1M, pH 8). Annealed oligos were run on a native page gel to ensure the adapters were annealed. Annealed adapters were ligated to input cDNA with T4 DNA Ligase (NEB). Agencourt Ampure Beads (Beckman Coulter) were used for size-selection to remove unligated adapter dimers. Clean DNA libraries were sequenced at the IU Center for Genomics and Bioinformatics on the Illumina MiSeq platform (250 nt paired-end reads).

Bioinformatics and data analysis

Output reads were cleaned with Trimmomatic (v3.0). A four nucleotide sliding window was used to remove bases if the average quality fell below 20. Flash (v1.2.2) was used to overlap paired-end reads to generate a single long read. Bowtie2 (v2.1.0) run in local mode was used to align the long read to the reference transcripts. Any mapped read that contained mismatches other than A-to-G, contained one or more indels, or that failed to span the entire target portion of the transcript was removed from consideration. The pattern of A-to-G edits for every mapped read was enumerated allowing for an explicit count of every observed pattern. Simultaneously this information was used to determine the proportion of editing that occurred at any given A. To determine the expected count for any given pattern, the total number of mapped reads was multiplied times the joint probability of all editing sites of any particular pattern assuming that editing occurs independently at any given site with the likelihood that the specific base is edited being based on the proportion of the edited/not-edited events observed across all the reads belonging to that particular sequence. Log-ratios of observed over expected were calculated by adding one to both the numerator and denominator to avoid infinite or undefined values. The p-value of a given result was determined using the binomial distribution. Briefly, the observed difference from expectation was used to determine the likelihood that such a deviation would be seen assuming the probability of the expected frequency of the pattern out of all patterns. Many of the simpler patterns were embedded within more complex patterns. To examine the explicit behavior of these simpler patterns we enumerated and counted all the sub-patterns for every pattern seen for both the observed and expected set of full patterns. The binomial for sub-patterns could also be calculated based on the expected number of sub-patterns out of all observed patterns (as opposed to the sum of all sub-patterns). The observed and expected dice coefficients were calculated using the formula 2a/(2a + b + c) as previously described.⁴⁵ Here, a indicates the number of reads that contain editing events at both sites, b is the number of reads that contain editing at the first site but not the second, and c is the number of reads that contain editing at the second site but not the first. Dice values were calculated in this manner for both the observed and expected read counts and the log₂ ratio (observed/expected) was reported as the final dice value.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Acknowledgments

We thank Stephane Bentolila and Maureen Hanson for sharing unpublished methods for ultra-deep sequencing of specific transcripts, James Ford of the Indiana University Center for Genomics and Bioinformatics for providing guidance about library preparation and next generation sequencing and Christiane Hassel of the Indiana University Flow Cytometry core for assistance with worm sorting.

Funding

This work was supported by start-up funds from Indiana University School of Medicine to HAH, a NIH predoctoral training grant to MCW (T32 GM007757) and an Indiana University Hutton Honors College summer research fellowship to ECW.

Supplemental Material

Supplemental data for this article can be accessed on the publisher's website.