Selective enrichment of A-to-I edited transcripts from cellular RNA using Endonuclease V

Abstract

Creating accurate maps of A-to-I RNA editing activity is vital to improving our understanding of the biological role of this process and harnessing it as a signal for disease diagnosis. Current RNA sequencing techniques are susceptible to random sampling limitations due to the complexity of the transcriptome, and require large amounts of RNA material, specialized instrumentation, and high read counts to accurately interrogate A-to-I editing sites. To address these challenges, we show that Escherichia coli Endonuclease V (eEndoV), an inosine-cleaving enzyme, can be repurposed to bind and isolate A-to-I edited transcripts from cellular RNA. While Mg²⁺ enables eEndoV to catalyze RNA cleavage, we show that similar levels of Ca²⁺ instead promote binding of inosine without cleavage and thus enable high affinity capture of inosine in RNA. We leverage this capability to demonstrate EndoVIPER-seq (Endonuclease V inosine precipitation enrichment sequencing) as a facile and effective method to enrich A-to-I edited transcripts prior to RNA-seq, producing significant increases in the coverage and detection of identified editing sites. We envision the use of this approach as a straightforward and cost-effective strategy to improve the epitranscriptomic informational density of RNA samples, facilitating a deeper understanding of the functional roles of A-to-I editing.

Graphical Abstract

Introduction

Adenosine-to-inosine (A-to-I) RNA editing is an abundant post-transcriptional modification found in animals. Catalyzed by adenosine deaminases acting on RNAs (ADARs), this reaction alters both the chemical structure and hydrogen bonding patterns of the nucleobase (Fig 1a).¹ Inosines preferentially base pair with cytidine, effectively recoding these sites as guanosine. A-to-I editing is widespread across the transcriptome and present in most types of RNA. In mRNA, these sites are primarily found in repetitive and untranslated regions, affecting transcript stability, localization, and interactions with cellular pathways. mRNA editing sites can also augment transcript splicing and directly alter amino acid sequences in open reading frames.² Additionally, A-to-I editing modulates the target specificities and biogenesis of small-interfering RNAs (siRNAs) and microRNAs (miRNAs), in turn affecting global gene expression patterns and overall cellular behavior.³ A-to-I editing continues to be implicated in a variety of critical biological processes including embryogenesis, stem cell differentiation, and innate cellular immunity.^{2, 4} Dysfunctional A-to-I editing has also been linked with numerous disease processes such as autoimmune disorders and several types of cancer.^5-6 Recent work has also demonstrated A-to-I editing as a vital driver of human brain development and overall nervous system function, and dysregulated activity has similarly been implicated in a variety of neurological disorders including epilepsy, amyotrophic lateral sclerosis, glioblastoma, schizophrenia, autism, and Alzheimer’s disease.^7-13

Figure 1. eEndoV recognizes inosine in ssRNA. Supplementation with Ca²⁺ enables high affinity binding and selective immunoprecipitation of inosine-containing ssRNAs.

a) Chemical alterations of adenosine-to-inosine RNA editing catalyzed by ADAR enzymes. b) Crystal structure (PDB 2W35) of eEndoV (green) complexed with ssDNA (purple), illustrating recognition of inosine (red) in a nucleic acid substrate and Mg²⁺ (cyan) positioned adjacent to cleavage site. c) Oligoribonucleotide test sequences with putative cleavage site (arrow) and PAGE analysis of digestion reactions with eEndoV illustrating specificity toward RNA I and confirming Mg²⁺ requirement for cleavage. d) Mg²⁺ or Ca²⁺ supplementation modulates eEndoV activity towards inosine-containing RNA substrates between cleavage and binding. e) EndoVIPER schematic targeting a Cy5-labeled ssRNA using recombinant eEndoV-MBP fusion protein and anti-MBP magnetic beads. f) Representative PAGE analysis of initial (I), flowthrough (FT) and eluate (E) EndoVIPER fractions, illustrating the effects of Ca²⁺ supplementation on pulldown efficiency. g-h) Densitometric analysis of pulldown efficiency for A- and I-containing RNA. i) Quantification of eEndoV binding affinity towards ssRNA I (red) and ssRNA A (blue) using MST. Values represent mean with standard deviation, and Kd denotes mean with 95% confidence interval. (n = 3).

Robust identification and detection of A-to-I sites is vital to understanding these broader biological roles, regulation dynamics, and relationships with disease. Because inosine is decoded as guanosine during reverse transcription, most contemporary methods utilize high-throughput RNA sequencing (RNA-seq) to identify editing sites from A-G transitions.¹⁴ While seemingly simple, the natural complexity of cellular RNA and large dynamic ranges between individual transcripts renders RNA-seq inherently susceptible to random sampling and technical variability, making it challenging to consistently capture and detect RNA editing events, especially in light of the relative scarcity of A-to-I editing sites. Although ~5 million sites have been identified across the transcriptome,^15-17 inosine content is low in the context of total cellular RNA, appearing in relatively few actual reads in RNA-seq datasets. This can be attributed to the fact that many key edited transcripts are expressed at low copy number. Moreover, the editing rates at individual sites can be very low or only conditionally active, and can differ significantly across cell and tissue types, individual organisms, developmental stages, and disease states.^18-20 Because of these technical challenges in RNA-seq, stringent bioinformatic analyses are also crucial for accurate detection, and extensive computational screening is needed to separate true A-to-I sites from sequencing errors, single-nucleotide polymorphisms (SNPs), somatic mutations, or spurious chemical alterations in RNA.²¹

These limitations can be overcome in part by using significant quantities of starting RNA material and/or collecting very large numbers of sequencing reads to achieve sufficient depth and coverage for accurate A-to-I calling. Alternatively, microfluidic or droplet-based PCR methods have been developed to specifically amplify regions of interest prior to RNA-seq, achieving greater sensitivity in detecting editing activity at focused A-to-I sites.²² However, amplification-based enrichment is also significantly lower in throughput, susceptible to PCR bias, and requires both specialized instrumentation and prior knowledge of the target transcripts. While current approaches enable characterization of A-to-I editing and have yielded substantial insights into the “inosinome” in a variety of different species and tissues,¹⁸ these methods remain impractical, expensive, and time consuming. Together, present technical limitations have made it challenging to both characterize existing A-to-I editing activity as well as increasingly difficult to discover new editing sites, restricting our overall understanding of these epitranscriptomic dynamics.

Enriching A-to-I edited transcripts prior to sequencing would largely address these challenges by depleting RNAs that otherwise lead to “wasted” sequencing reads while also helping to validate the editing sites that are observed. Despite the simplicity of this idea, effective methods to specifically target and isolate inosine in RNA have remained elusive. While a previous report detailed the generation of inosine-targeting polyclonal antibodies for isolating modified tRNAs, these were also found to cross-react with several other nucleobases, and this research has not been reproduced.²³ We and others have also explored inosine chemical labeling strategies using acrylamide and acrylonitrile derivatives, and while these approaches are feasible for labeling and capturing inosine-containing RNAs, these reagents also irreversibly modify transcripts with adducts that inhibit reverse transcription, and inherently display off-target reactivity with pseudouridine and uridine, limiting enrichment efficiency.^24-25 Taken together, these attempts to improve A-to-I editing detection through enrichment remain limited and do not address existing technical challenges. As a result, the most widely used approach for detecting A-to-I sites remains a standard RNA-seq workflow followed by bioinformatic detection. In pursuit of alternative enrichment methods, we identified EndonucleaseV (EndoV), a conserved nucleic acid repair enzyme capable of recognizing and binding to inosine. In prokaryotes, EndoV cleaves downstream of inosine lesions resulting from oxidative damage in DNA to promote base excision repair.²⁶ In humans and other metazoans, EndoV has now been implicated in the metabolism of A-to-I edited RNAs.^27-28 We hypothesized that if cleavage activity could be selectively suppressed without compromising recognition and binding, then EndoV could be leveraged for enriching A-to-I edited RNAs. While human EndoV (hEndoV) appears to be a good candidate toward this goal, its biological functions and substrate preferences are still not entirely known. Recent studies have identified possible affinity toward both unedited double-stranded RNA (dsRNA) and ribosomal RNA (rRNA), properties which could be problematic for use in cellular RNA samples.²⁹ Interestingly, these reports also showed that Escherichia coli EndoV (eEndoV) was both specific and highly active toward inosine in single-stranded RNA (ssRNA) and exhibited minimal sequence bias.^27-28 These observations, as well as the commercial availability of a purified recombinant enzyme, encouraged us to explore eEndoV for the pulldown and enrichment of A-to-I edited transcripts. Herein we demonstrate EndoVIPER-seq (Endonuclease V inosine precipitation enrichment sequencing) as a novel and effective approach to bind and isolate inosine-containing transcripts prior to RNA-seq, producing significantly improved coverage and detection of A-to-I editing sites in cellular RNA.

Results

Structural analyses have revealed that EndoV requires Mg²⁺ as a cofactor for inosine recognition and strand scission (Fig. 1b).³⁰ Similar studies have shown that replacing Mg²⁺ with Ca²⁺ facilitates binding of EndoV to inosine-containing substrates without supporting catalysis.³¹ Thus, we hypothesized that supplementing eEndoV with Ca²⁺ would enable enrichment of inosine-containing RNAs from cellular RNA. As an initial test of feasibility, we synthesized a pair of Cy5-labeled oligoribonucleotides having either A or I in a defined position and evaluated eEndoV activity in the presence of both cations. Consistent with previous reports, we observed not only specific cleavage activity towards inosine in ssRNA (RNA I) when benchmarked against a non-edited control (ssRNA A), but also an obligate Mg²⁺ requirement for cleavage (Fig. 1c). After verifying that EndoV was unable to cleave target strands in the presence of increasing amounts of Ca²⁺ (Fig. S1), we next evaluated the effect of Ca²⁺ supplementation on the ability of eEndoV to bind and isolate inosine-containing ssRNA. The recombinant enzyme is fused to a maltosebinding protein (MBP) tag, conveniently enabling us to implement a magnetic IP workflow using anti-MBP functionalized beads, which we term EndoVIPER (Endonuclease V inosine precipitation enrichment, Fig. 1e). We used this method to attempt pulldown of both ssRNA A and ssRNA I in the presence of variable amounts of Ca²⁺, while monitoring the initial, unbound (flowthrough), and elution fractions after washing (Fig. 1f). Not surprisingly, omitting Ca²⁺ produced little binding of either oligonucleotide, supporting the idea that both recognition and cleavage of inosine is mediated through divalent cations. Increasing amounts of Ca²⁺ from 0-10 mM improved binding efficiency substantially, approaching ~80% recovery with excellent selectivity (~350-fold over pulldown of RNA A). Additional supplementation beyond 10 mM Ca²⁺ quickly decreased pulldown efficiency and selectivity (Fig. 1g,h), and while unconfirmed, these results likely arise from electrostatic shielding of the negative charge on the RNA phosphodiester backbone, disrupting interactions with key amino acid residues on the protein. In any case, we selected 5 mM Ca²⁺ as a suitable concentration for maximizing both recovery and selectivity. We then applied these conditions to measure the binding affinity of eEndoV for each RNA substrate using microscale thermophoresis (MST) and observed low nanomolar affinity for ssRNA I and no measurable binding to the ssRNA A control (Fig. 1i).

While these results were encouraging, we also recognized that ADAR primarily targets structured duplexes,^1-2 and thus the majority of inosine likely resides in the context of dsRNA. We were concerned that eEndoV may have difficulty interacting with inosine in these substrates in our native binding conditions, so we synthesized several complementary RNA strands to both ssRNA A and ssRNA I targets with differing bases opposite the A/I position (supplementary table 1). After annealing these strands together (Figs. 2a-b), we assessed eEndoV affinity and EndoVIPER performance with each of the duplex constructs (Figs. 2c-f). The enzyme exhibited no detectable binding with any unedited dsRNA A substrates, yet binding affinity towards dsRNA I combinations was highly variable and dependent on the identity of the opposing base in the complementary strand. In particular, a fully complementary duplex (dsRNA I:C) showed virtually no detectable binding by both MST and EndoVIPER (Figs. 2d-f), while mismatches ranging from I:U to I:G demonstrated increased binding in both assays. These results are also intriguing in that they are consistent with prior studies of eEndoV on DNA repair,³⁰ together indicating an approximate substrate preference of ssI>>> dsI:G > dsI:U > dsI:C. While interesting, these results posed a challenge to our ultimate goal of designing an unbiased approach to enriching A-to-I edited transcripts from cellular RNA, and we recognized the need to reduce or eliminate RNA secondary structure in order to mitigate the effect of these affinity biases.

Figure 2. eEndoV binding favors ssRNA over dsRNA substrates.

a) Schematic of dsRNA target annealing and b) duplex verification by 10% native PAGE. c) MST analysis of eEndoV binding affinity towards dsRNA A and d) dsRNA I targets using MST. Values represent mean with standard deviation. (n = 3) e) Representative PAGE analysis of initial (I), flowthrough (FT) and eluate (E) EndoVIPER fractions when tested with various dsRNA targets. f) Densitometric analysis of EndoVIPER efficiency for dsRNA targets. Values represent mean with standard deviation (n = 2). Unpaired t-test was performed for all samples against ssRNA I pulldowns (*** denotes p = 0.0003 and **** denotes p < 0.0001).

Our first attempt involved reducing the ionic strength of our buffer conditions, as duplex formation is highly dependent on the presence of cations. While we initially chose 5 mM Ca²⁺for the pulldown step, our results indicate that ~1-10 mM Ca²⁺ produce similar pulldown efficiencies (Fig. 1g). These tests also employed a standard Tris-buffered saline (19 mM Tris, 137 mM NaCl, 2.7 mM KCl, pH 7.4), and we recognized that lower concentrations of monovalent cations may be tolerated. To explore these options, we assayed conditions having varying concentrations of each cation and found that removing KCl altogether and reducing CaCl₂ to 1 mM resulted in highly similar binding affinity and EndoVIPER performance (Fig. S2). However, we also found that NaCl concentrations below 100 mM resulted in a significant increase in non-specific binding (Fig. S2). Despite some promising results, both EndoVIPER and MST analyses indicated that this approach remained insufficient for opening RNA duplexes in our system, and that binding remained highly dependent on structure (Fig. S3).

We next investigated stronger chemical methods to fully denature potential dsRNA targets. While several non-covalent denaturants, including formamide and urea, are effective in unfolding stable RNA structures, these also act on proteins, and we doubted it would be possible to denature RNA structure while maintaining native eEndoV activity. Due to these concerns, we searched for covalent methods to reversibly denature RNA prior to EndoVIPER. We required a reagent that 1) rapidly reacts with RNA under non-degrading conditions, 2) stably maintains RNA in a single-stranded state, 3) does not interfere with eEndoV binding, and 4) can be fully removed for downstream sequencing. We were inspired by previous reports of glyoxal modification of RNA, as this reagent reacts readily with amines on the Watson-Crick-Franklin face to form stable adducts that interfere with basepairing and RNA secondary structure.³² While glyoxal can react with A, C, and G, the N¹,N²-dihydroxyguanosine adduct is by far the most stable (Fig. 3a).³³ Importantly, glyoxal does not react with inosine, an observation that has been leveraged to study A-to-I locations through RNase T1-mediated cleavage assays.^34-35 While this appeared promising, we were uncertain if RNA glyoxalation would be compatible with eEndoV binding. To assess this, we first subjected our ssRNA I and ssRNA A oligoribonucleotides to glyoxal treatment using previously reported conditions, and observed the expected upward shift in molecular weight when analyzed via 20% PAGE (Fig. S4a). We then analyzed binding affinity of eEndoV towards each of the treated RNAs. Surprisingly, we observed a slight improvement in affinity toward glyoxalated ssRNA I, as well as some increased non-specific response towards ssRNA A at higher concentrations of eEndoV (Fig. S5a). We hypothesized that installation of hydrophilic groups capable of hydrogen bonding with the protein might be responsible for this non-specific activity. We also theorized that lower concentrations of eEndoV would likely confer improved specificity, so we titrated the amount of eEndoV used in the pulldown step and observed a clear optimum for both selectivity and efficiency at 100 nM enzyme (Figs S5b-d). Next, we repeated our full performance assay on dsRNA A and I duplex combinations. We first treated the target and complementary strands with glyoxal and unsurprisingly observed no duplex formation between glyoxalated RNAs and their complementary strands via 10% native PAGE (Fig. 3d). We then tested binding affinity (Figs. 3e,f) and EndoVIPER efficiency (Figs. 3g,h) on the denatured RNA duplexes and observed equivalent performance across all RNA I combinations, indicating successful elimination of structural biases in eEndoV binding. While we were encouraged by these results, intermolecular duplexes are relatively easy to disrupt, so we also wanted to ensure that glyoxal treatment prior to EndoVIPER was similarly robust in RNAs having a highly stable internal secondary structure. To test this, we designed a hairpin substrate representing a “worst case” RNA target due to its high melting temperature (Figs. S6a,b). When we chemically denatured this hairpin with glyoxal, we observed almost identical EndoVIPER performance compared to previous experiments (Figs. S6d-f). Together, these data demonstrated that we could overcome even strong secondary structure to enable pulldown with little to no effect on selectivity or enrichment of edited RNAs. However, due to the preferential reaction of glyoxal with guanosine,³³ we were also concerned about the possibility that G bases adjacent to or near an inosine site could inhibit eEndoV binding. To address this concern, we synthesized a “G heavy” RNA strand as an additional “worst case” test substrate (Fig. S7a), and we again observed nearly identical pulldown and binding affinity towards this substrate (Fig. S7b-d). While there was a slight increase in overall binding affinity when measured by MST (Fig. S7d), there was no detectable difference in pulldown performance (Fig. S7c). Together, these experiments demonstrated that our optimized EndoVIPER protocol is robust and displays minimal bias in vitro, and thus we were ready to test our method in a high-throughput sequencing workflow using cellular RNA.

Figure 3. Glyoxal treatment disrupts RNA secondary structure and enables unbiased pulldown of inosine in both ssRNA and dsRNA.

a) Schematic of glyoxal addition to the Watson-Crick-Franklin face on guanosine residues, forming a N¹,N²-dihydroxyguanosine adduct. b) General reaction conditions for installation and removal of glyoxal adducts on test RNA strands. c) Disruption of dsRNA target annealing by glyoxal treatment and d) verification by 10% native PAGE. e) MST analysis of eEndoV binding affinity towards glyoxal-treated dsRNA A and f) dsRNA I targets using MST. Values represent mean with standard deviation. (n = 3) g) Representative PAGE analysis of initial (I), flowthrough (FT) and eluate (E) EndoVIPER fractions when tested with various glyoxal-treated dsRNA targets. h) Densitometric analysis of EndoVIPER efficiency for glyoxal-treated dsRNA targets. Values represent mean with standard deviation (n = 2). Unpaired t-test was performed for all samples against ssRNA I pulldowns (** denotes p = 0.0045 and “ns” indicates no significant difference).

We selected human brain mRNA to quantify EndoVIPER-seq performance, as this tissue is known to have high A-to-I editing activity and would thus provide ample editing sites to validate our method. Additionally, nervous system tissue is a biologically interesting setting for exploring the enrichment and clinical detection of RNA editing sites crucial for neurological function or indicative of disease. To prepare for the ultimate step of high-throughput sequencing, we needed to randomly fragment our starting RNA material into smaller strand lengths. We recognized that this step would also fortuitously decrease any remaining likelihood of secondary structure formation, enhancing the resolution and performance of our pulldown. Inspired by the approaches in other RNA pulldown workflows,³⁶ we targeted fragment sizes of ~200-500 nt, and found that ~1 minute treatment time with Mg²⁺ at 94 °C was sufficient to yield the desired size distribution (Figs. S8a,b). While glyoxal removal is well-characterized and has been used previously in both Sanger and RNA-seq applications,^34,35 we also wanted to confirm that this step was compatible with EndoVIPER performance. We first subjected fragmented mRNA to full glyoxalation and deprotection, while maintaining an identical untreated sample as a control. We then reverse transcribed cDNA for two neuronal ionotropic receptor mRNAs having known A-to-I editing sites, GRIA2 and KCNA1 (Fig S9a).^37,38 Using real-time PCR (RT-PCR), we monitored the amplification of both transcripts and observed no kinetic difference between untreated and glyoxalated/deprotected mRNA samples, indicative of complete glyoxal removal (Fig S9b). We also confirmed the compatibility of glyoxalation/deprotection with Sanger sequencing, as we not only observed identical electropherogram traces between samples, but also detected the known A-to-I editing sites in both transcripts for both control and treatment samples (Figs S10, S11). Together with previous studies using glyoxal in sequencing workflows,^34-35 these experiments confirm that glyoxal denaturation is fully reversible and does not interfere with critical EndoVIPER library preparation steps. With these conditions established, we next sought to directly benchmark EndoVIPER to the currently used RNA-seq methodology, and so we fragmented 2 μg of mRNA and divided this material into duplicate “RNA-seq” and “EndoVIPER-seq” groups (500 ng each), and EndoVIPER samples were subjected to the enrichment workflow (Fig. 4a). After all samples underwent deprotection using heat, all samples were analyzed for size distribution and integrity, confirming that our full workflow could be completed without appreciable RNA degradation (Fig. S8c). We then prepared libraries using ~4 ng of each respective RNA-seq and EndoVIPER-seq mRNA and proceeded to sequencing. To assess and measure A-to-I editing across samples, we employed a read aligner optimized for RNA editing (RASER³⁹) as well as the specialized REDITools script package and associated filtering steps.⁴⁰ From these analyses, it was immediately apparent that the total number of identified sites was significantly higher in EndoVIPER samples (mean 34,084 sites), achieving ~1.8-fold more called A-to-I editing sites compared to RNA-seq without enrichment (mean 19,308 sites, supplementary tables 2-5, Fig 4b). We also merged grouped data and screened these sites against the RADAR,¹⁶ REDIPortal,¹⁵ and DARNED¹⁷ databases, observing a large increase in both existing and novel A-to-I locations in EndoVIPER samples (supplementary tables 6-7, Fig 4c). Although the number of newly identified sites was larger than we expected in both sample groups (RNA-seq 19,515 novel positions out of 31,310 total called sites versus EndoVIPER 27,429 novel positions out of 56,744 total called sites; supplementary tables 6 and 7) it is worth noting that these databases catalog sites only when detected in several genome-matched donors across many RNA-seq experiments. Our experiment utilized commercially available brain mRNA (Takara Bio) isolated and pooled from a small number of donors, and larger scale verification studies are needed to further characterize these candidate sites. In any case, we applied consistent computational assessment between RNA-seq and EndoVIPER-seq samples and reliably observed a large increase in the detection of both known and novel editing sites, demonstrating the effectiveness of our method for increasing the sensitivity of detecting A-to-I editing. As a further measure to functionally validate our method, we merged and aligned all RNA-seq and EndoVIPER datasets (73,578 sites) and compared both coverage and editing rate at each detected A-to-I location. We observed a significant increase in both metrics across paired sites, indicating that EndoVIPER-seq selectively enriched A-to-I edited RNAs (supplementary tables 8-9, Figs. 4d,e). We also observed, on average, ~38-fold enrichment from read coverage values across all sites, with >75% of these sites displaying equivalent or significantly increased sequencing depth (supplementary table 8, Fig. 4f). Because of the inherent complexity of cellular RNA, we were curious whether EndoVIPER enrichment was affected by individual transcript abundance. When we plotted read coverage against enrichment scores for all sites detected in RNA-seq samples, we expectedly observed an overall decrease in fold enrichment with increased abundance (Fig. S12). However, the overall correlation between these two variables was poor (R² = < 0.005), indicating that EndoVIPER is capable of enriching edited RNA transcripts across a large dynamic range of relative abundance. To ensure that eEndoV did not display a sequence context bias, we compiled the top 200 most enriched A-to-I sites and performed a sequence motif analysis (supplementary table 10). We observed no discernable consensus surrounding the editing site in highly enriched transcripts, suggesting minimal EndoVIPER sequence bias (Fig. 4g). While certainly desirable towards our goal, this activity is also somewhat expected given the canonical role of eEndoV in agnostic, genome-wide surveillance and repair of oxidative lesions in DNA.^{26, 30}

Figure 4. EndoVIPER-seq enables enrichment and high-throughput analysis of A-to-I RNA editing sites.

a) Schematic of EndoVIPER-seq workflow. Cellular RNA is first randomly hydrolyzed into ~200-500 nt fragments, followed by glyoxal denaturation. A-to-I edited RNA is then enriched by eEndoV pulldown, followed by glyoxal removal, library preparation and high-throughput sequencing. b) Mean number of sites between duplicate RNA-seq and EndoVIPER-seq samples shows significantly increased (unpaired t-test, p = 0.03) detection of called A-to-I positions. c) Merged datasets cross-referenced against known databases show that detection of both novel and existing A-to-I sites is enhanced by EndoVIPER. Box and whisker plots show that d) read coverages and e) editing rate at all A-to-I editing sites (n = 73,578) are significantly increased by EndoVIPER (paired t-tests, p = < 0.0001). Means are denoted by black crosses. f) Box and whisker plot of calculated fold enrichment at all sites (mean, black cross = ~38-fold, n = 73,578 sites). g) Sequence motif analysis compiled from the top 200 most enriched transcripts. Red arrow denotes A/I site.

A-to-I editing is critical for normal brain development and function, and editing activity has now been identified as a reliable, differential biomarker in a number of neurological disorders. Detection of these pathological editing events is likely to be a vital component of future RNA-based diagnostic applications, and thus we sought to employ EndoVIPER-seq for monitoring specific editing sites of interest to demonstrate its utility for improving such epitranscriptomic characterization. In particular, we applied RNA-seq and EndoVIPER datasets toward four specific editing site panels, assessing read coverage at 462 editing sites upregulated in postnatal brain development (supplementary table 11),⁸ 403 increased editing events found in autism spectrum disorder (supplementary table 12),⁹ 115 sites with increased editing activity in schizophrenic patients (supplementary table 13),¹² and 31 hyperedited protein recoding events implicated in glioblastoma carcinogenesis (supplementary table 14).^{6, 13} We directly compared read coverage at these sites in both RNA-seq and EndoVIPER-seq samples, and saw a consistent overall increase in total read coverage at these positions (Figs. 5a-d). We also expressed these data as the number of “edited reads” containing inosine by multiplying coverage with respective calculated editing rate at each site, and this trend was expectedly similar (Fig. S13). Together, these data indicate that EndoVIPER-seq both increased coverage at sites of interest as well as improved specific detection of pathological, edited transcript isoforms, positioning this method as a valuable tool for future clinical epitranscriptomics applications.

Figure 5. EndoVIPER-seq enhances detection of clinically relevant A-to-I editing sites.

Read coverage heatmaps in upregulated RNA editing sites of interest (supplementary tables 11-14), in a) brain development (462 sites), b) autism spectrum disorder (403 sites), c) schizophrenia (115 sites) and d) protein recoding events in glioblastoma (31 sites), demonstrating increased coverage of important editing sites in EndoVIPER treated samples. Heatmap columns display both replicate RNA-seq datasets for RNA-seq and EndoVIPER-seq samples, and each row denotes an individual editing site scaled to illustrate low (blue) and high (red) read coverage between groups.

Discussion

As a scientific community, we now understand that A-to-I RNA editing is a vital post-transcriptional change affecting a variety of essential cellular pathways. Additionally, dysregulated editing underlies the molecular pathogenesis of many diseases, and is particularly important in the human nervous system. However, the true landscape and prevalence of A-to-I editing in the transcriptome remains unknown, and the total amount of A-to-I sites in the transcriptome is estimated to be significantly greater than those currently found in existing databases. Mapping and determining the precise function of individual editing sites is difficult due to current technical limitations in RNA-seq experiments, which display high sampling variability and remain challenging to specifically apply toward A-to-I editing. These studies also require specialized instrumentation for focused library preparation on a small number of sites, or use costly “brute-force” approaches that require large amounts of input RNA material and very high sequencing depth. Together, these challenges make it increasingly difficult to access the epitranscriptomic data landscape in RNA-seq experiments.

Herein we present EndoVIPER as a new method for the affinity pulldown of inosine-containing transcripts from cellular RNA, overcoming many of the limitations outlined above and significantly improving detection and characterization of A-to-I RNA editing in complex samples. We first assess and verify that eEndoV displays high affinity and selectivity for inosine in RNA, and we chemically optimize our conditions to eliminate bias arising from different structural motifs. We then validate and demonstrate EndoVIPER-seq with brain mRNA, and show a significant increase in the ability to detect and discover A-to-I sites. In addition, we show the utility of this method for focused characterization in four panels of biologically relevant A-to-I sites, illustrating the power of our method in detecting critical RNA editing events in brain development, autism, schizophrenia, and glioblastoma. EndoVIPER is simple, straightforward, and flexible, and is easily implemented in standard library preparation workflows for RNA-seq experiments in different biological contexts. Additionally, our approach utilizes low-cost, commercially available reagents with little to no modification, enabling researchers to obtain significantly more epitranscriptomic data with smaller amounts of input RNA material. There remains opportunity for further improvement regarding EndoVIPER efficiency and selectivity, and we look forward to evaluating the performance of different EndoV orthologs, as well as exploring directed evolution strategies to further enhance binding affinity and selectivity. We also plan to extensively apply our overall strategy toward much larger scale studies of A-to-I editing across multiple individuals, tissues, and disease states, in turn providing a more detailed understanding of the overall inosine landscape in humans. To our knowledge, EndoVIPER is also the first demonstrated repurposing of an enzyme “reader” toward binding and enriching edited RNA transcripts, and our results provide strong evidence for the versatility of this strategy for isolating other epitranscriptomic or epigenetic modifications using their cognate readers. Extending this approach to other modified nucleotides would generate a new toolbox for characterizing the corresponding transcriptional changes, and we plan to explore this in the immediate future.

Together, this report details a simple yet powerful new tool to complement existing epitranscriptomic sequencing technologies. The overall ease of use and accessibility of this method create potential for broad utility in many research disciplines. We anticipate that this will significantly improve our understanding of the dynamics and global regulation of A-to-I RNA editing across a multitude of biological contexts, further probing the potential of this epitranscriptomic mark to reveal critical information about biological function and disease progression.

Methods

RNA Oligoribonucleotides

All oligonucleotides used in this study were custom designed and purchased from Integrated DNA Technologies. Complete sequences used in this study can be found in supplementary table 1.

RNA Cleavage Assays

10 pmol of either ssRNA I or ssRNA A was incubated in the presence or absence of both Mg²⁺ at a 10 mM final concentration and/or 9 pmol recombinant eEndoV (New England Biolabs) in a total volume of 10 μL. Final buffer conditions in all reactions were 10 mM Tris, 125 mM NaCl, 15 μM EDTA, 150 μM DTT, 0.025% Triton X-100, 30 μg/ml BSA, 7% glycerol, pH 7.4. Reactions were incubated for 1 hour at 25 °C, followed by a 10 min heat inactivation at 85 °C. Reaction products were separated using 10% denaturing PAGE, and gels were imaged with a GE Amersham Typhoon RGB scanner using 635 nm excitation laser and the Cy5 670BP30 emission filter. To test cleavage in the presence of Ca²⁺, 10 pmol of ssRNA I was incubated with 840 nM eEndoV with either 10 mM MgCl₂ or variable amounts of CaCl₂ (0, 0.1, 0.5, 1, 2.5, 5, 10 and 20 mM) in a total volume of 50 μL. Final buffer conditions in all reactions were 19 mM Tris, 137 mM NaCl, 3 mM KCl, 15 μM EDTA, 150 μM DTT, 0.025% Triton X-100, 30 μg/ml BSA, 7% glycerol, pH 7.4. Reactions were incubated at room temperature for 3 hours, after which a 3 μL sample was taken for 10 % denaturing PAGE analysis as described above.

EndoVIPER Magnetic IP Assays

For our initial binding tests (Fig. 1e), 10 pmol of either RNA I or RNA A was combined with 840 nM eEndoV and variable amounts of CaCl₂ (0, 0.1, 0.5, 1, 2.5, 5, 10 and 20 mM) in a total volume of 50 μL. Final buffer conditions were 19 mM Tris, 137 mM NaCl, 3 mM KCl, 15 μM EDTA, 150 μM DTT, 0.025% Triton X-100, 30 μg/ml BSA, 7% glycerol, pH 7.4. Reactions were incubated at room temperature for 30 min, after which a 3 μL sample (initial, I) was taken and set aside for later analysis. Separately, 70 pL of anti-MBP magnetic bead slurry (New England Biolabs) was washed extensively with a buffer containing 19 mM Tris, 137 mM NaCl, 3 mM KCl, 7% glycerol, and variable amounts of CaCl₂ (0, 0.1, 0.5, 1, 2.5, 5, 10 and 20 mM), pH 7.4. After washing, beads were resuspended in eEndoV-RNA samples and incubated at 25 °C for two hours with end-over-end rotation. Magnetic field was applied to the beads and a 3 μL sample (unbound, UB) of the supernatant was saved for later analysis. Beads were washed extensively with respective buffer containing variable amounts of Ca²⁺, and resuspended in 50 μL 19 mM Tris, 137 mM NaCl, 3 mM KCl, 47.5% formamide 0.01% SDS, pH 7.4 and heated to 95 °C for 10 min. Magnetic field was applied and a 3 μL final sample (eluate, E) of the supernatant was taken of each reaction. Collected fractions were analyzed using 10% denaturing PAGE, and gels were imaged using a GE Amersham Typhoon RGB scanner. Densitometric quantification of bands was performed using ImageJ software. % Bound is expressed as a band intensity ratio of unbound versus initial fractions. % Recovered was defined as the intensity ratio of eluate versus initial fractions. Fold-selectivity was calculated as the ratio of ssRNA I versus ssRNA A recovery percentages. For experiments utilizing RNA duplexes (Fig. 2e), stock constructs were first annealed as described in the later section and 10 pmol of this duplex was used for pulldown using the same protocol as outlined above. For buffer optimization experiments (Figs. S1,2), this pulldown procedure was identical to our initial studies above while altering the components of the buffer as outlined in the figure. We will refer to these optimal formulations as 1X EndoVIPER (EV) binding buffer (19 mM Tris, 100 mM NaCl, 1 mM CaCl₂, 15 μM EDTA, 150 μM DTT, 0.025% Triton X-100, 30 μg/ml BSA, 7% glycerol, pH 7.4.) and 1X EV wash buffer (19 mM Tris, 100 mM NaCl, 1 mM CaCl₂, 7% glycerol, pH 7.4). To identify optimal eEndoV concentrations (Figs. S4b-d), the pulldown procedure was performed by combining 10 pmol of glyoxalated ssRNA I or ssRNA A with 25 nM, 50 nM, 75 nM, 100 nM, 150 nM 200 nM, 400 nM, or 840 nM eEndoV in 1X EV binding buffer and bead-purified with 1X EV wash buffer as described above. Final elution was performed in 50 μL 0.5 M triethylammonium acetate (TEAA) pH 8.6, 47.5% formamide 0.01% SDS (“1X EV elution buffer”) and heated to 95 °C for 10 min, after which samples were analyzed and imaged using 10% denaturing PAGE as described earlier. For pulldown analysis of the hairpin RNA I substrate (hRNA I, Fig. S6d), 10 pmol of glyoxalated and untreated RNA was incubated with 100 nM eEndoV in 1X EV binding buffer and purified, eluted and analyzed as described earlier using 1X EV wash and EV elution buffers respectively. 10 pmol of “G heavy” RNA strand (G ss RNA I, Fig S6b), was tested in an identical manner using 1X EV buffers.

Microscale Thermophoresis (MST)

For our initial binding studies (Fig 1h, Figs. 2c,d), varying amounts of eEndoV were combined with 6 fmol of respective ssRNA or dsRNA targets in a final volume of 20 μL and allowed to incubate for 30 min at room temperature. Final buffer conditions in these samples were 19 mM Tris, 137 mM NaCl, 3 mM KCl, 5 mM CaCl₂, 15 μM EDTA, 150 μM DTT, 0.025% Triton X-100, 30 μg/ml BSA, 7% glycerol, pH 7.4. After incubating, samples were loaded into NT.115 standard glass capillaries. MST experiments were performed using a Nanotemper Monolith NT.115 Pico instrument. All measurements were analyzed using the Pico-RED filter with 12% LED intensity and 40% laser power. Data were fitted using GraphPad Prism 8 analysis software to determine K_d values. Binding tests were performed in triplicate in separate trials. For subsequent experiments (Figs. 3e,f, S4a, S5f, S6d), RNAs were treated with glyoxal and purified as described below and incubated in 1X EV binding buffer and analyzed with the Nanotemper instrument using the same settings as our initial studies above.

RNA Duplex annealing

To assess duplex formation, 100 pmol of each RNA pair (untreated or glyoxalated) were mixed together in 19 mM Tris, 137 mM NaCl, 3 mM KCl, pH 7.4. Mixtures were heated to 95 °C for 5 minutes and slowly cooled to room temperature over the course of approximately 1 hour. 10 pmol of annealed construct was then loaded onto a 10% native non-denaturing polyacrylamide gel and imaged with a GE Amersham Typhoon RGB scanner.

Glyoxal treatment and deprotection

For our initial tests of RNA glyoxalation (Fig. S4a), 5 ug of ssRNA A or ssRNA I was added to 100 μL of 50% DmSo, 6% glyoxal (Sigma Aldrich) in nuclease-free water. Samples were reacted for 1 hour at 50 °C and ethanol precipitated. 10 pmol of treated and purified RNA was then analyzed by 10% denaturing PAGE and imaged using a Typhoon RGB scanner. To remove glyoxal adducts (Fig. S3b), 10 pmol of treated and purified RNA was added to 50 μL 0.5 M TEAA pH 8.6, 47.5% formamide, 0.01% SDS and heated to 95 °C for 0, 0.5, 1, 2, 5, 10, 15, and 20 minutes. 5 μL of these reactions were directly analyzed by 20% denaturing page and imaged as described earlier.

mRNA glyoxal deprotection, Real-Time PCR, and Sanger sequencing

2 μg human brain mRNA (Takara bio, whole brain tissue pooled from 8 Caucasian males, ages: 43-66) was fragmented for 1 minute at 94 °C using the NEBNext® Magnesium RNA Fragmentation Module (New England Biolabs) and ethanol precipitated. Purified pellet was then dissolved in nuclease-free water and quantified using a NanoDrop™ spectrophotometer (Thermo Fisher Scientific). 1 μg of fragmented mRNA was then reacted for 1 hour at 50 °C in 100 μL of 50% DmSo, 6% glyoxal (Sigma Aldrich) in nuclease-free water, followed by ethanol precipitation. Purified pellet was then dissolved in 200 μL of 1X EV elution buffer and glyoxal was removed by heating to 95 °C for 10 min. RNA was then purified with the Monarch® RNA Cleanup Kit and eluted in nuclease-free water. To ensure full removal of glyoxal adducts, RNA was then incubated at 65 °C for 2 hours in 100 μL 50% DMSO in 137 mM NaCl, 2.7 mM KCl, 8 mM Na₂HPO₄, and 2 mM KH₂PO₄, pH 7.4 followed by ethanol precipitation, resuspension in nuclease-free water, and quantification using a NanoDrop™ spectrophotometer. In separate tubes, 100 ng of untreated or deprotected mRNA fragments were combined with 20 pmol of gene specific reverse primer (GRIA2 reverse primer 5’ CCACACACCTCCAACAATGCG 3’ and KCNA1 reverse primer 5’ CTCGGTGGTAGAAATAGTTGAAATTGGACAC 3’) and heated to 70 °C for 10 minutes and then placed on ice. cDNA was then synthesized at 42 °C for 1 hour using OneTaq® M-MuLV reverse transcriptase (New England Biolabs). 10 μL of each cDNA reaction was then mixed with 10 μL of 2X iTaq Universal SYBR Green Supermix (BioRad) and 10 pmols of gene specific forward and reverse primers (GRIA2 forward 5’ GAGAACTTGTATATGGGAAAGCTGATATTGC 3’, GRIA2 reverse 5’ CCACACACCTCCAACAATGCG 3’, KCNA1 forward 5’ GAATCTTCAAGCTCTCCCGCCAC 3’, KCNA1 reverse 5’ CTCGGTGGTAGAAATAGTTGAAATTGGACAC 3’). PCR reactions were monitored in real-time using a LightCycler® 96 instrument (Roche) using the following thermal cycling program: 94 °C for 3 min, followed by 45 cycles of (94 °C for 15 s, 60 °C for 30 s, 68 °C for 30 s), 68 °C for 5 min. Amplification traces were analyzed using the LightCycler® software, and cycle thresholds (C_t) were determined using a default fluorescence value setting of 0.2 RFUs. PCR reactions were then purified using the Monarch PCR & DNA Cleanup Kit (New England Biolabs). 80 ng of each purified amplicon was then analyzed on a 1 % agarose gel and compared to a GeneRuler 50 bp DNA ladder (Thermo Fisher Scientific). 50 ng of each purified amplicon was also submitted for Sanger sequencing (Genscript). Sequencing traces were analyzed using SnapGene viewer.

EndoVIPER-seq

2 μg human brain mRNA (Takara bio, whole brain tissue pooled from 8 Caucasian males, ages: 43-66) was fragmented for 1 minute at 94 °C using the NEBNext® Magnesium RNA Fragmentation Module (New England Biolabs) and ethanol precipitated. Fragmented mRNA was then reacted for 1 hour at 50 °C in 100 μL of 50% DMSO, 6% glyoxal (Sigma Aldrich) in nuclease-free water, followed by ethanol precipitation. Purified pellet was then dissolved in nuclease-free water and quantified using a NanoDrop™ spectrophotometer (Thermo Fisher Scientific). 500 ng of fragmented and glyoxalated mRNA was then added to each of two tubes (duplicate “RNA-seq” samples) containing 30 μL nuclease-free water and frozen at −80 °C for later use. For EndoVIPER samples, 500 ng of fragmented, glyoxalated mRNA was added to each of two tubes containing a 250 μL solution of 100 nM eEndoV and 120 units RNaisin Plus inhibitor (Promega) in 1X EV binding buffer, and was incubated at room temperature for 30 minutes. Separately, 300 μL anti-MBP magnetic bead slurry (New England Biolabs) was added to a new microfuge tube and washed extensively with 1X EV wash buffer. After washing, beads were resuspended in the eEndoV-mRNA samples and incubated at room temperature for two hours with end-over-end rotation. Magnetic field was applied and the supernatant was discarded. Beads were then washed three times with 500 μL 1X EV wash buffer and then resuspended in 200 μL of 1X EV elution buffer. Bound mRNA was then eluted by heating to 95 °C for 10 min. Residual magnetic beads were removed from the collected supernatant using 0.22 μM microfuge spin filters (Corning® Costar®), and RNA was purified further with the Monarch® RNA Cleanup Kit and eluted in nuclease-free water. To ensure full removal of glyoxal adducts, RNA was incubated at 65 °C for 2 hours in 100 μL 50% DMSO in 137 mM NaCl, 2.7 mM KCl, 8 mM Na2HPO4, and 2 mM KH2PO4, pH 7.4 followed by ethanol precipitation and resuspension in nuclease-free water. Starting mRNA material, fragmented RNA-seq mRNA, and enriched EndoVIPER mRNA were quantified and assessed for size distribution (Fig S6) using an Agilent 2100 Bioanalyzer instrument and the Agilent 6000 RNA Pico kit. 8 ng of each RNA-seq and EndoVIPER-seq RNA replicate was then used to prepare sequencing libraries with the SMARTer Stranded Total RNA-Seq Kit v2 - Pico Input kit (Takara Bio), standard 8-bp i5 and i7 Illumina index barcodes and adapters were added to each library. All libraries were then sequenced using a NextSeq 550 (Illumina) to produce paired end 150-bp reads.

Read trimming and mapping

Adapter and barcode sequences were removed using Trimmomatic⁴¹ and processed with the following additional parameters (HEADCROP:3 LEADING:31 TRAILING:31 SLIDINGWINDOW:6:31, ILLUMINACLIP using a custom list of known Illumina adapter sequences, supplementary table 15). HEADCROP was used to remove the first three nucleotides of the second paired-end sequencing read (read 2), which originate from the Pico v2 SMART template switching oligonucleotide adapter utilized in the SMARTer Stranded Total RNA-Seq Kit v2 - Pico Input kit (Takara Bio). Trimmed reads were aligned to the human reference genome assembly GRCh37 (hg19) using reads aligner for SNPs and editing sites of RNA (RASER)³⁹ with previously optimized parameters m = 0.05 and b = 0.03, and uniquely mapped reads were retained. PCR duplicates were removed with the Genome Analysis Toolkit program MarkDuplicates (Picard).

Identification of RNA Editing Sites

To call editing sites, we used the REDItools⁴⁰ python package, filtering for sites with mapping quality score ≥ 10, variant call quality ≥ 20, minimum read coverage ≥ 3, minimum number of reads supporting variation ≥ 3, and minimum editing frequency ≥ 0.1. We also removed substitutions in homopolymeric regions of ≥ 5 nt, discarded any hits corresponding to intronic regions 4 nt next to known splice sites, and retained only AG and TC transitions. We also removed known common SNPs (dbSNP151) obtained from the UCSC genome browser.

Statistics

Total called editing sites were compiled for each sample, and averaged across groups. Means were graphed and significance (unpaired t-test) was calculated using Prism 8. Coverage and editing rate were compared by merging datasets from each group and aligning by editing site. Box and whisker plots and statistical significance (paired t-test) were graphed and determined in Prism. To enable calculation of fold-enrichment, all datasets were merged and aligned by site, and a pseudocount of 0.1 was added to all raw coverage values.⁴² Box and whisker plots were graphed and determined in Prism. To analyze enrichment against transcript abundance, sites detected in RNA-seq samples were plotted against fold-enrichment scores and a semi-log regression analysis was calculated using Prism.

Motif analysis

The top 200 most enriched sites were compiled (supplementary table 10), and neighboring sequence information (4 nucleotides up and downstream of editing site) was obtaining using the Integrative Genomics Viewer (Broad Institute).⁴³ Sequence logo graph representing consensus frequency was created using WebLogo (Berkeley, CA).⁴⁴

Heatmaps

Heatmap images were generated from compiled read coverages and “edited” reads (supplementary tables 11-14) using the web-based Heatmapper tool.⁴⁵ “Edited” read counts were calculated by multiplying read coverage with editing frequency at each site.

• Additional information, experimental data and analysis (PDF)

• Supplementary tables 1-14 (excel spreadsheet)