RNA modifications, alternative splicing and circular RNA landscape in the mouse brain: inosine and beyond

Abstract

RNA modifications in the brain can be essential for regulating the transcriptome and brain function. Our study unveils the landscape of different RNA modifications associated with splicing within the mouse brain. Focusing on inosine, known for its role in alternative splicing modulation and enriched in introns, we investigated its influence using mice lacking ADAR2 alone or in combination with catalytically inactive ADAR1. While some alternative splicing-regulatory roles of inosine and ADAR enzymes are established, we observe that altering ADAR2 and ADAR1/ADAR2 is associated with changes in alternative splicing and coincides with shifts in levels of other RNA modification. Through the utilization of an innovative approach, we identified novel candidate circular RNA profiles in wild-type and mutant mice and detected potential inosine sites within circular RNAs. Collectively, our findings underscore a complex interplay among RNA modifications, alternative splicing, circular RNAs in the mouse brain.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-30030-4.

Introduction

RNA modifications, regulated by specialized enzymes, expand the diversity of the transcriptome and are crucial for maintaining cellular and organ homeostasis¹. In the brain, RNA modifications seem particularly abundant and are essential for brain development, neuronal differentiation, and synapse formation². Given their importance in key physiological functions, it is not surprising that dysregulation of RNA modifications or the enzymes that mediate these modifications can cause neurological and neurodegenerative disorders^3–6 and be associated with aging⁷. Consequently, delving into the realm of RNA modifications in the brain emerges as a critical avenue of research.

RNA modifications can alter which proteins bind to RNA, either directly or indirectly, by modifying RNA secondary structures⁸. As such, RNA modifications can affect multiple processes, including transcription, nuclear export and stabilization^9–12. Furthermore, internal mRNA modifications can potentially promote changes in the recognition of splicing factors or the assembly of the spliceosome, impacting splice site recognition and therefore altering canonical RNA splicing as well as backsplicing. Backsplicing, a mechanism responsible for generating circular RNAs (circRNAs) by joining splice sites in reverse order, adds another layer of complexity to RNA regulation¹³. Factors that bring backsplice sites into proximity by changing RNA structure promote circRNA production¹⁴. Given that any RNA modifications may influence RNA secondary structure¹⁵, they are also likely regulators of circRNA formation. circRNAs have diverse functions in cells, including acting as miRNA sponges, regulating transcription, interacting with proteins, potentially being translated into proteins, serving as scaffolds for molecular complexes, and influencing alternative splicing (AS)¹⁶. AS variants exhibit remarkable diversity, while circRNAs are notably abundant in the brain. Both play pivotal roles in synaptic plasticity and aging-related neurological diseases, contributing significantly to the intricate tapestry of gene expression regulation within neuronal cells^17,18.

While some RNA modifications have been shown to be associated with AS^19,20 and circRNA changes in cancer cell lines²¹, thereby impacting cellular functions, many RNA modifications are yet to be explored for their potential involvement in AS and circRNA formation. The interplay between RNA modifications, AS, and circRNA formation in the brain represents an important avenue for research, offering insights into mechanisms underlying neurological diseases and aging-related processes.

In our study, we aimed to identify which RNA modification might play a globally important role in AS regulation and circRNA dynamics in the brain. To achieve this, we focused on identifying modifications that are enriched in regions containing introns, as these seem more likely to be relevant to splicing processes. By screening for such modifications, we sought to narrow down RNA modification candidates that could be potentially implicated in AS and backsplicing regulation. Among these candidates, we selected one modification to study in greater detail, aiming to elucidate its association with AS and backsplicing changes. Specifically, we delve into the role of inosine: ADAR1 and ADAR2 enzymes are known to catalyze the conversion of adenosine to inosine in RNA, a process referred to as adenosine-to-inosine (A-to-I) editing²². To assess the impact, we utilized wild-type (WT) mice, ADAR2 deficient mice (Adar2^−/− Gria2^R/R; termed Adar2^KO) and mice deficient in ADAR2 and catalytically inactive ADAR1 (Adar1^E861A/E861A Ifih^−/− Adar2^−/− Gria2^R/R; termed Adar1/2^KO). The Gria2^R/R allele rescues the lethality of ADAR2 deficiency, while the Ifih^−/− background is required for viability in the absence of catalytically active ADAR1^23,24.

While the AS-regulatory functions of inosine and ADAR enzymes are well-documented^20,25, our study revealed that alterations in ADAR2 and ADAR1/ADAR2 not only affected AS processes but also led to changes in other RNA modifications. Employing an innovative approach, we identified putative circRNA profiles in both WT and mutant mice and detected inosine within circRNAs. By elucidating these complex correlations among RNA modifications, AS regulation, circRNA dynamics in the brain, our research sheds light on the intricate regulatory networks that govern brain function. Although our study does not establish a direct regulatory role, but it identifies promising candidates for further investigation, providing a foundation for future mechanistic studies.

Results

RNA modifications enriched in nuclear polyA RNA in the mouse brain

In our study, we aimed to identify an RNA modification that could play a widely important role in regulating AS regulation in the brain. To identify candidates potentially involved in AS regulation, we focused on modifications enriched in intron-containing regions of pre-mRNAs, as these are more likely to be relevant to splicing processes. By initially screening for such modifications, we sought to pinpoint candidates that could be implicated in AS and backsplicing regulation (Fig. 1a).

Fig. 1.

Enrichment of RNA Modifications in nuclear polyA RNA and Splicing-Inhibited mRNAs. (a) Schematic representation illustrating RNA modifications enriched in pre-mRNA containing introns, with isolation strategy for nuclear polyA RNA (polyA selected from nuclear enriched RNA) and total polyA RNA (polyA selected, total RNA) from WT mouse brain samples to quantify RNA modifications with Liquid Chromatography-Tandem Mass Spectrometry (LC–MS/MS). (b) Comparison of RNA modification abundance between nuclear and total mRNA samples, revealing a reduction and increase in various RNA modifications. 5-methylcytosine (m5C); 2'-O-methylcytidine (Cm); 7-methylguanosine (m7G); 1-methylguanosine (m1G); N2,N2-dimethylguanosine (Gm); Pseudouridine (Ψ); 2'-O-methyluridine (Um); inosine (I); 1-methyladenosine (Am); N6-methyladenosine (m6A); N6,N6-dimethyladenosine (m6Am); n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01; ***P < 0.001. (c) Experimental setup for analyzing RNA modifications in splicing-inhibited mRNAs in primary neurons using Pladienolide B (PlaB) and 4-thiouridine (4SU) to selectively label and isolate unspliced transcripts. (d) Analysis of RNA modification abundance in unspliced versus spliced mRNA transcripts. N2, N2,7-trimethylguanosine (m22G); n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01; ***P < 0.001.

To identify RNA modifications potentially involved in splicing regulation, we aimed to enrich for intron-containing pre-mRNAs and identify which modifications are enriched in these regions, as such modifications are likely to be associated with splicing processes. Several approaches are available to achieve intron-enrichment, each with its own advantages and limitations. For example, rRNA depletion retains both spliced and non-spliced RNAs, leading to a high background that complicates the analysis of splicing-related modifications. Cap selection, while useful for studying cap-proximal modifications, can introduce a bias toward cap-specific RNA modifications and may not capture the full spectrum of modifications involved in splicing. PolyA selection, although biased toward mature mRNAs, provides a more focused dataset by enriching for polyadenylated transcripts, including pre-mRNAs undergoing splicing²⁶. After considering these options, we opted for polyA selection, as it offered the most balanced approach for our initial screening. To mitigate the bias toward mature mRNAs, we compared nuclear polyA-selected RNA, enriched in pre-mRNAs containing introns²⁶, to total polyA-selected RNA as a control. Importantly, this approach represents a first step in identifying candidate modifications that could be enriched in introns, and further confirmation using complementary methods will be essential to validate RNA modification enrichment in introns.

Using this strategy, we isolated nuclear polyA RNA and total polyA RNA subsets from homogenized WT mouse brain samples in triplicate (n = 3 biological replicates from 3 different mice) (Fig. 1a). Nuclear RNA was successfully isolated, as demonstrated by the enrichment of intronic sequences (Supplementary Fig. 1). To detect RNA modifications, we enzymatically digested the RNA into individual nucleosides and analyzed them using Liquid Chromatography Tandem Mass Spectrometry (LC–MS/MS), a method that enables direct identification and accurate quantification of any RNA modifications²⁷. By comparing RNA modification abundance in nuclear polyA RNA versus total polyA RNA, we identified modifications specifically enriched in nuclear pre-mRNAs, providing modification candidates that could be implicated in splicing regulation. Our data revealed a reduction in several RNA modifications in nuclear polyA RNA , including 5-methylcytosine (m5C), 2′-O-methylcytosine (Cm), N7-methylguanosine (m7G), 2′-O-methylguanosine (Gm), pseudouridine (Ψ), 2′-O-methyluridine (Um), N1-methyladenosine (m1A), 2′-O-methyladenosine (Am), and N6-methyladenosine (m6A). In contrast, we observed a 3.84-fold enrichment of the RNA modification N6,2′-O-dimethyladenosine (m6Am) and a 2.77-fold enrichment of inosine (Fig. 1b). These findings suggest that inosine, consistent with previous evidence²⁸, and m6Am could be enriched in intronic regions within the mouse brain. While this screening approach does not conclusively demonstrate that inosine or m6Am are enriched in introns—due to potential biases introduced by polyA and nuclei selection—it provides valuable insights by identifying these modifications as promising candidates enriched in intron-containing regions. Further validation is necessary to confirm their association with splicing processes.

Inosine, m6Am and various other RNA Modifications are enriched in splicing-inhibited mRNAs in rat primary neurons

To determine through an enrichment-independent method whether the identified RNA modifications are localized within intronic regions of mRNA, we impeded splicing. Since directly inhibiting splicing in the mouse brain is not feasible, we turned to rat primary neurons for our investigations, employing the splicing inhibitor Pladienolide B (PlaB)²⁹. Given that splicing predominantly occurs co-transcriptionally³⁰, and that the bulk of mRNA in cells is already spliced, only newly synthesized transcripts remain unspliced after PlaB exposure. To target predominantly unspliced mRNAs for our RNA modification analyses and eliminate pre-existing spliced mRNA transcripts prior to PlaB exposure, we utilized the nucleotide analog 4-thiouridine (4SU). This strategy involved integrating 4SU into nascent RNA, allowing for subsequent biotinylation of the newly synthesized 4SU-labeled RNA to specifically isolate and examine these transcripts³¹ (Fig. 1c). Through the synergistic use of 4SU and PlaB, we selectively labeled newly transcribed RNA that retained their unspliced state, complemented by a 4SU-only control for subsequent comparative analyses. The efficacy of splicing inhibition was validated via RT-qPCR (Supplementary Fig. 2). To facilitate the analysis of RNA modifications on mRNA transcripts and exclude RNAs that typically do not undergo splicing, we enriched for RNA fractions containing polyA tails. To account for potential biases, in parallel we processed 4SU-only controls (Fig. 1c). Next, we processed these RNAs into individual nucleosides for subsequent RNA modification identifications by LC–MS/MS. Our comparisons of RNA modification abundance between unspliced (PlaB and 4SU) versus spliced (4SU) mRNA transcripts revealed a 2- to 3-fold enrichment for several RNA modifications in unspliced mRNA, including 1-methylguanosine (m1G), N2,N2-dimethylguanosine (m22G), N6-methyladenosine (m6A), m6Am and inosine (Fig. 1d).

Given that Pladienolide B specifically targets the SF3B1 component of the U2 snRNP²⁹, apparent increases in marks characteristic of snRNAs¹ under PlaB treatment should be interpreted cautiously as potentially SF3B1/U2-related effects rather than evidence for general intronic enrichment. While our observations suggest that certain modifications may be associated with intron-containing RNA, global transcriptome changes due to PlaB treatment or off-target effects could also explain the observed differences. However, these findings complement the initial observation of inosine and m6Am enrichment in nuclear polyA RNA within the mouse brain, with their prevalence in intron-enriched RNAs consistent with a potential association with splicing, although causality and precise positioning remain to be established. Together, these observations serve as supportive screening data to motivate testing inosine and m6Am as candidates potentially associated with AS and backsplicing events in the mouse brain.

Although m6Am has been implicated in splicing regulation through its presence on the small nuclear RNA U2, a component of the splicing machinery³², m6Am directly deposited in introns has so far not been explored. In contrast, inosine is a well-documented RNA modification known to influence alternative splicing^20,25, with high-throughput sequencing studies confirming its enrichment in intronic regions²⁸.

While the global correlation between inosine and the transcriptome—including its impact on the circRNA landscape—remains poorly understood, inosine’s established role in AS and its enrichment in introns make it a compelling candidate for further investigation^19,20,25. In this study, we focus on inosine to explore its global association with AS and circRNA landscape.

Abundance of Potential Inosine RNA Modification in circular RNAs

To comprehensively understand the impact of inosine on the entire transcriptome, we also investigated its association with circRNAs. Given a low expression level and therefore limited amount of circRNAs, detecting RNA modifications using LC–MS/MS is currently unfeasible. Consequently, to uncover inosine on circRNAs and estimate its prevalence, we turned to a high-throughput sequencing approach. While many sequencing methods for identifying RNA modifications are error-prone or lack quantitative accuracy due to modification-specific antibody enrichments³³, inosine can be reliably detected in a quantitative manner through high-throughput sequencing. Inosine can be identified by detecting A-to-G mutations in sequencing data, as inosine is read as guanosine during reverse transcription, leading to apparent A-to-G mutations where inosine is present²⁸ (Fig. 2a). Prior to identifying inosine in circRNAs, circRNAs are first enriched and identified by detecting backsplicing sequences in Illumina sequencing data. However, this method has limitations; the complete structure of circRNAs is often unknown due to sequencing size constraints, making it challenging to determine if certain introns are included or the exact length of the circRNAs. Additionally, abnormal splicing events can produce backsplicing sequences that do not correspond to circular RNAs. Even after depleting linear RNA using exonucleases that target linear RNA, some linear RNAs are resistant to degradation, meaning that not all circRNAs identified through Illumina sequencing are necessarily circular RNAs (Supplementary Fig. 3). To address this limitation, we have independently developed an alternative approach for circRNA identification, employing rolling circle reverse transcription³⁴. This technique specifically copies the sequence of circRNAs repeatedly in a single cDNA strand. Therefore, non-repetitive circularized RNA templates will generate sequence repeats in cDNA strands. By detecting extended RNA repeats that are typically non-repetitive, complete circRNA molecules can be reliably identified. The Oxford Nanopore Technology (ONT) sequencing platform is utilized to pinpoint these extended RNA repeats, leveraging its capability to read lengthy sequences (Fig. 2b). Applying the rolling circle reverse transcription ONT sequencing method to circRNA enriched RNA extracted from three mouse brains (n = 3 biological replicates), we acquired 4.2, 5, and 5.1 million clean sequencing reads per sample. These datasets facilitated the identification of 28,104, 25,085, and 22,347 distinct circRNAs, respectively (Supplementary Table 1). Out of these, 16,047, 12,090 and 9,959, respectively, represented previously uncharacterized putative circRNAs, underscoring the discovery of potentially novel circRNA species through this methodology (Fig. 2c). The most prevalent size range for circRNAs fell between 500 to 1000 nucleotides (Fig. 2d), predominantly comprised of exonic sequences (Fig. 2e). However, we also observed the presence of putative novel circRNAs exceeding 4 kb in length indicating the circRNA landscape might be more diverse. Notably, in all ONT identified circRNAs, we discovered 7,508, 8,697 and 6,423 potential A-to-I editing sites, respectively (Fig. 2f). We specifically utilized the ONT sequencing to estimate the abundance of potential inosines in circRNAs, as this rolling circle reverse transcription approach ensures their circular nature and completeness. This allowed us to preliminary estimate, for the first time, the prevalence of inosine within circRNAs. Since the number of potential inosines detected is low relative to the total number of circRNAs, it seems unlikely that inosine within circRNAs regulates most backsplicing events. Given that inosine is predominantly found in introns, we hypothesized that intronic inosine could be associated with backsplicing events that shape the circRNA landscape, even if the inosine itself is not retained in the mature circRNA. To determine whether and how inosine in introns influences AS and circRNA biogenesis, we modulated and removed inosine.

Fig. 2.

Identification of circRNAs with Oxford Nanopore Technology (ONT). (a) Schematic representation of circRNA biogenesis through backsplicing processes with detection strategy for inosine (I) in linear and circRNAs, with inosine identified as A-to-G mutations following reverse transcription, indicating the presence of inosine. (b) Outline of circRNA identification methodology employing enrichment for circRNAs, rolling circle reverse transcription followed by ONT sequencing to reliably identify complete circRNA molecules and characterize circRNA diversity. By detecting extended RNA repeats that are typically non-repetitive, complete circRNA molecules can be reliably identified. (c) Number of previously known and novel circRNAs identified in 3 biological replicates of WT mouse brains. Rep = biological replicate. n = 3. (d) Distribution of circRNA sizes extracted from mouse brain samples, with a predominant size range between 500 to 1,000 nucleotides. Data from 3 biological replicates is shown. n = 3. (e) Composition analysis of circRNAs, predominantly comprised of exonic sequences. (f) Estimation of inosine prevalence within circRNAs. Rep = biological replicate. n = 3.

Depletion of ADAR2 and the combined loss of ADAR2 and ADAR1 catalytic activity impacts the global RNA and DNA modification landscape in the mouse brain beyond inosine alone

To assess the impact of A-to-I editing on AS and backsplicing, we manipulated inosine levels within the brain. Specifically, we compared WT conditions with normal editing levels in pre-mRNAs (unspliced mRNAs) to conditions with reduced or completely eliminated editing (Fig. 3a). By systematically varying editing levels in pre-mRNAs, we aimed to gain a comprehensive and quantitative understanding of how inosine influences the mRNA and circRNA landscape.

Fig. 3.

Impact of ADAR Enzyme Pertrubations on the Global RNA and DNA Modification Landscape in the Mouse Brain. (a) Experimental design illustrating the manipulation of inosine levels in pre-mRNAs to investigate the influence of inosine on the mRNA and circRNA landscape in the brain using ADAR2 deficient (Adar2^KO) mice and mice lacking the enzymatic activity of ADAR1 and deficient for ADAR2 (Adar1/2^KO). (b) Quantification of inosine (I) in total RNA from Adar2^KO and Adar1/2^KO mice through Liquid Chromatography-Tandem Mass Spectrometry (LC–MS/MS). n = 3; mean ± SD; t-test; **P < 0.01. (c) Assessment of global RNA modification changes in Adar2^KO and Adar1/2^KO mice through LC–MS/MS, showing alterations in m1G, Gm, m5U, m6A, and m6Am levels compared to WT controls. 5-methylcytosine (m5C); 2'-O-methylcytidine (Cm); 7-methylguanosine (m7G); 1-methylguanosine (m1G); N2,N2-dimethylguanosine (Gm); N2,N2,7-trimethylguanosine (m22G); Pseudouridine (Ψ); 2'-O-methyluridine (Um); 5-methyluridine (m5U); 1-methyladenosine (m1A); N6-methyladenosine (m6A); N6,N6-dimethyladenosine (m6Am); n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01. (d) Analysis of long RNAs (> 200nt) from Adar2^KO and Adar1/2^KO mice through LC–MS/MS, demonstrating significant enrichment of m1G, m22G, m5U, and m6Am, with alterations in m5C, m7G, and m1A levels in Adar2^KO mice. n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01; ***P < 0.001. (e) Examination of small RNAs (< 200nt) from Adar2^KO and Adar1/2^KO mice through LC–MS/MS, indicating depletion of Am, m6A, and m6Am, along with changes in m5C, m7G, I and m22G levels compared to WT controls. n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01; ***P < 0.001. (f) DNA modifications in WT and mutant mouse brains for ADAR enzymes by LC–MS/MS, revealing a reduction in m5dC levels and an increase in hm5dC levels in Adar1/2^KO mice, deoxyinosine (dI). n = 3; mean ± SD; t-test; *P < 0.05; **P < 0.01.

To modulate A-to-I editing levels within the brain, we acquired ADAR2 deficient (Adar2^KO) mice and mice lacking the enzymatic activity of ADAR1 and deficient for ADAR2 (Adar1/2^KO) to investigate the influence of inosine on the transcriptome landscape of the mouse brain (Fig. 3a, Supplementary Fig. 4). Previous studies have indicated that Adar2^KO mice exhibit reduced levels of inosine, while Adar1/2^KO mice show absence of inosine in mRNAs or pre-mRNAs²³. By leveraging these mouse models, we aimed to assess the impact of varying A-to-I editing levels on the transcriptome. We employed LC–MS/MS to determine if global changes in A-to-I editing levels could be detected using this method. Our analysis revealed a decreased level of inosine in total RNA from Adar1/2^KO mice. However, we did not observe a significant reduction in Adar2^KO mice (Fig. 3b). This can be explained by a high molecular abundance of inosine in a small number of sites in highly abundant RNAs, such as tRNAs edited by other editing enzymes such as ADAT enzymes³⁵. ADAR1 and ADAR2 editing, in contrast, generates a large number of inosine sites in RNA but each at low stoichiometry in less abundant RNAs, and often in underrepresented regions such as introns³⁶. Thus, the loss of ADAR2 alone removes many sites but too few molecules to significantly shift the total pool, while the combined loss of ADAR2 and ADAR1 catalytic activity removes a sufficient molecular fraction to be detectable.

We subsequently investigated the effect of ADAR2 depletion, and the combined loss of ADAR2 and ADAR1 catalytic activity, on the global landscape of other RNA modifications, a phenomenon we initially hypothesized would be minimal. Using LC–MS/MS, we analyzed total RNA from Adar2^KO and Adar1/2^KO mice to evaluate how ADAR2 and ADAR2 and ADAR1 catalytic activity, along with inosine loss, affects the RNA modification landscape. To our surprise, we observed several global RNA modification changes. Specifically, we noted a modest yet significant 1.3-fold increase in m1G and m5U levels, as well as a 1.4-fold decrease in m6Am levels in Adar2^KO mice compared to WT controls. In Adar1/2^KO mice, we observed an increase in Gm and m6A levels relative to WT (Fig. 3c).

Our unexpected findings suggest a link between ADAR1 and ADAR2 with other RNA modifications, likely through indirect mechanisms or global transcriptome changes. To investigate this further, we systematically identified RNA modification alterations across various RNA types in Adar2^KO and Adar1/2^KO mice. Notably, significant differences in RNA modification levels were observed between small (< 200nt) and long (> 200nt) RNA species (Supplementary Fig. 5), prompting us to conduct separate analyses on these fractions.

Our LC–MS/MS analysis revealed that long RNAs from both Adar2^KO and Adar1/2^KO mice brains exhibited significant enrichment of m1G, m22G, m5U, and m6Am compared to WT, with slight increases in m5C and m7G. m1A was significantly altered only in Adar2^KO but not in Adar1/2^KO. Notably, inosine in long RNA was below the detection limit in all samples, indicating that inosine detected in total RNA originates predominantly from short RNAs (Fig. 3d). In contrast, small RNAs isolated from Adar2^KO and Adar1/2^KO mouse brains showed depletion of Am, m6A, and m6Am relative to the WT (Fig. 3e). Notably, a slight increase in m5C levels was detected in Adar2^KO mice, m7G was elevated in both Adar2^KO and Adar1/2^KO mice, while m22G was increased only in Adar1/2^KO mice compared to the WT. Surprisingly, inosine levels in short RNAs were significantly higher in both mutant genotypes compared to WT (Fig. 3e).

Given the numerous changes in RNA modifications observed following ADAR2 depletion as well as the combined loss of ADAR2 and ADAR1 catalytic activity, we also explored whether DNA modifications might have been altered. LC–MS/MS analysis of DNA from these mice revealed a 18% reduction in 5-methyl-2'-deoxycytidine (m5dC) levels in Adar1/2^KO mice compared to WT. Additionally, 5-hydroxymethyl-2'-deoxycytidine (hm5dC) levels were significantly elevated in both Adar2^KO and Adar1/2^KO mouse brains (Fig. 3f). No changes in deoxyinosine (dI) levels were observed upon ADAR2 depletion and the combined loss of ADAR2 and ADAR1 catalytic activity, indicating that ADAR enzymes do not catalyze dI deamination in DNA (Fig. 3f).

Our observations indicate broad shifts in measured RNA and DNA modifications under ADAR2 depletion and the combined loss of ADAR2 and ADAR1 catalytic activity. Whether these arise via inosine-dependent or independent downstream processes remains to be established. Given that DNA modifications can affect transcription³⁷, these observations motivate follow-up studies to assess potential downstream effects on the transcriptome. These observations, individually and in combination, coincide with broad shifts in measured RNA and DNA modification levels. To contextualize these associations, the subsequent analyses focus on the transcriptomic landscape—nascent RNAs, spliced RNAs, and circRNAs—comparing WT samples to those lacking ADAR2 entirely and those expressing only catalytically inactive ADAR1 in addition to lacking ADAR2.

Delving into inosine features in pre-mRNAs

To corroborate the enrichment of inosine in intronic regions, and to comprehensively understand the A-to-I editing landscape, we identified A-to-I editing events through A-to-G conversions in RNA sequencing data from WT mouse brains, which are enriched in nascent RNA. Nascent RNA was chosen for its high representation of intronic sequences, providing a robust dataset for analysis. To assess the impact of reduced A-to-I editing, we analyzed nascent RNA from Adar2^KO mice, which exhibit decreased inosine levels in mRNAs and pre-mRNAs. We also examined nascent RNA from Adar1/2^KO mouse brains, which lack inosine in mRNAs and pre-mRNAs²³ (Fig. 4a, Supplementary Table 2). While sequencing errors are inherent to high-throughput sequencing and cannot be entirely eliminated, Adar1/2^KO samples enabled us to identify false positives, serving as a background control for sequencing artifacts (see Methods). To ensure that we have identified true inosine-derived A-to-G edits and not sequencing errors, we quantified all possible nucleotide conversion types consistently detected across biological replicates in our RNA-seq datasets for comparative analysis (Supplementary Fig. 6a, see Methods). Wild-type samples showed a striking predominance of A-to-G mismatches, consistent with authentic ADAR-mediated editing. While this signature was markedly reduced in Adar2^KO, these samples still exhibited the second highest A-to-G conversion rate among all nucleotide changes (Supplementary Fig. 6a). In contrast, Adar1/2^KO samples showed A-to-G frequencies below the rate of the most frequent non-A-to-G conversions (C-to-T, G-to-A, etc.), indicating these Adar1/2^KO residual events likely represent technical noise from sequencing rather than biological editing. This multi-layered analysis confirms that the A-to-G sites identified in WT and Adar2^KO mice reflect genuine inosine editing, while effectively discriminating against stochastic artifacts. Furthermore, we considered only A-to-G conversions detected in all three WT replicates (under stringent quality control conditions) and absent in any Adar1/2^KO replicates as true editing sites (see Methods). We defined sites absent in any of the Adar2^KO replicates but present in all WT replicates as ADAR2-specific editing sites, as ADAR2 is required for their editing. Similarly, sites remaining in all Adar2^KO replicates were classified as ADAR1-specific, indicating that ADAR1 can edit these sites independently of ADAR2. Any sites found in the Adar1/2^KO samples were classified as false positives and excluded from any downstream analyses.

Fig. 4.

Exploring Inosine Features in Pre-mRNAs. (a) Experimental strategy focusing on identifying inosine sites through A-to-G conversions in nascent RNA obtained from WT, Adar2^KO, and Adar1/2^KO mouse brains to investigate the impact of reduced inosine levels on the transcriptome landscape. (b) Total number of inosine sites identified in nascent RNAs per biological replicate and their intersections in WT mouse brains. Rep = biological replicate. (c) Intersection analysis depicting the number of edited genes identified in WT mouse brain replicates. WT1-3: WT biological replicates 1–3. (d) Number of genes with editing over editing sites per gene. (e) Editing frequencies of inosine sites, demonstrating that the majority exhibit low editing rates, with a small proportion displaying higher editing frequencies. (f) Number of A-to-I editing sites attributed to ADAR1 and ADAR2. (g) Composition analysis of the identified inosine sites in nascent RNA (WT editing), as well as editing attributed solely to ADAR1 and ADAR2. Inosine is predominantly present in intronic sequences, regardless of the enzyme responsible for the editing.

We identified 36,382 A-to-G conversions that were reproducibly detected across all three WT biological replicates (WT intersection; Fig. 4b). Applying stringent quality control, we excluded any sites appearing in individual Adar1/2^KO datasets as false positives, yielding 33,534 high-confidence A-to-I editing sites. These sites mapped to 5,256 transcripts associated with pathways related to synaptic regulation and vesicle processing, underscoring the intricate role of inosine in regulating fundamental brain functions (Fig. 4c, Supplementary Fig. 6b). Most inosine-containing transcripts harbor 3–5 editing sites, with the majority of sites exhibiting low editing frequencies. Fewer than 11% of inosine sites showed editing frequencies exceeding 0.5 (Fig. 4d, e), indicating that most editing events occur at relatively low stoichiometry. However, we observed several transcripts with high number of A-to-I editing sites (> 100 editing sites per gene), and even identified some transcripts with over 1000 editing sites (Fig. 4d, Supplementary Fig. 6c). This is the case for Fgf14, a gene associated with cerebral ataxia³⁸, as well as for KCNIP4, a potassium voltage-gated channel interacting protein involved in attention-deficit and hyperactivity disorder³⁹. The majority of inosine in nascent RNA—specifically 65%—can be attributed to ADAR2, while 35% can be linked to ADAR1 (11,655 Adar1-specific and 21,879 Adar2 specific; Fig. 4f). Regardless of the enzyme responsible for the editing, the majority of inosine sites were overwhelmingly detected in introns (89.6% for WT, 90% for Adar2^KO, 90.8% for Adar1/2^KO) (Fig. 4g). This further corroborates our earlier LC–MS/MS findings. This distribution is consistent with the nuclear, intron-rich sampling strategy and the established predominance of ADAR2 activity in nuclear/intronic contexts²².

ADAR2 depletion and the combined loss of ADAR2 and ADAR1 catalytic activity impacts alternative splicing in transcripts associated with synaptic transmission

To explore whether inosine detected in nascent RNA modulates AS in the brain, we performed high-throughput sequencing analysis on mature mRNAs isolated from the brain of WT, Adar2^KO, and Adar1/2^KO mice. These datasets, coupled with nascent RNA analysis enabling the detection of inosine pre-splicing, are crucial for a comprehensive assessment of how inosine impacts AS on specific transcripts (Fig. 5a). Initially, we processed the mature mRNA datasets to identify AS events between Adar2^KO versus WT and Adar1/2^KO versus WT mouse brains. For our comparative analysis, we categorized the AS events into five distinct patterns: (i) skipped exons (SE)—entire exon omission, (ii) alternative 3’ splice site (A3SS)—partial exon omission at the 5’ end, (iii) alternative 5’ splice site (A5SS)—partial exon omission at the 3’ end, (iv) retained intron (RI)—intron retention, and (v) mutually exclusive exons (MXE)—exon retention in one condition and a different exon retention in another condition (Fig. 5b). We identified 2,223 statistically significant AS events impacting 1,583 genes in Adar2^KO mouse brains and 2,002 significant AS events affecting 1,512 genes in Adar1/2^KO mouse brains, each compared to WT controls (FDR < 0.05, dPSI > 0.1) (Supplementary Table 3, 4). SE events were the most frequently occurring AS events in Adar2^KO mice (40.53%), followed by MXE (23.08%), A3SS (15.02%), A5SS (11.07%) and RI (10.30%). In contrast, in Adar1/2^KO mice, MXE was the predominant AS event (49.05%), followed by SE (26.32%), A3SS (9.39%), A5SS (8.84%), and RI (6.39%) (Fig. 5c).

Fig. 5.

Impact of ADAR2 Depletion and the Combined Loss of ADAR2 and ADAR1 Catalytic Activity on Alternative Splicing. (a) Experimental outline conducted on mature mRNAs from WT, Adar2^KO, and Adar1/2^KO to investigate the modulation of alternative splicing (AS) influenced by inosine pre-splicing. (b) Categorization of AS events into skipped exons (SE), alternative 3’ splice site (A3SS), alternative 5’ splice site (A5SS), retained intron (RI), and mutually exclusive exons (MXE) for comparative analysis between Adar2^KO and Adar1/2^KO versus WT mouse brains. (c) Identification of AS events impacting genes in Adar2^KO and Adar1/2^KO mouse brains, with SE being the most prevalent AS event type in Adar2^KO and MXE in Adar1/2^KO mice (FDR < 0.05, dPSI > 0). (d) Evaluation of the percentage of AS events with positive delta Percent Spliced In (dPSI) values in Adar2^KO and Adar1/2^KO mice brains compared to WT, classified in AS categories. The graph also shows the effect of inosine-containing transcripts on AS processes. A dashed line has been incorporated in the figures to facilitate the visualization of the 50% value on the y-axis (FDR < 0.05, dPSI > 0). (e) Volcano plot depicting the differential splicing changes for SE and RI patterns in both Adar2^KO and Adar1/2^KO datasets, with specific transcripts of interest highlighted in blue. The x-axis denotes the dPSI (delta percent spliced in) value relative to wild-type (WT), while the y-axis represents the -log₁₀ of the false discovery rate (FDR) or q-value. (f) Gene Ontology analysis revealing enriched functions in transcripts exhibiting AS in Adar2^KO and Adar1/2^KO brains.

Subsequently, we aimed to evaluate and contrast the inclusion levels for the 5 different AS categories, while concurrently distinguishing between inosine-containing transcripts and transcripts devoid of inosine. This analysis would offer further insights into the impact of A-to-I editing within the different AS processes. Thus, we integrated mature mRNA sequencing data with previously identified inosine sites from nascent RNA sequencing and calculated the Percent Spliced In (PSI) values. PSI quantifies the inclusion level of a specific splice event. A positive delta PSI (dPSI) indicates a higher inclusion of the splice event in the mutant mice compared to WT mice. When over 50% of events within a category show positive dPSI values, this indicates an overall trend toward increased event inclusions in the mutant condition compared to WT.

In our Adar2^KO versus WT analysis, we observed prominent splicing changes in the RI category. Only 31% of the RI events show increased inclusion in Adar2^KO, a reduction from the 50% expected by chance (binomial test, P = 1 × 10^–4). This suggests an overall lower inclusion of introns compared to WT. The effect was even more pronounced in inosine-containing transcripts (23%, P = 5 × 10^–8), suggesting that inosine directly impacts intron retention (Fig. 5d). In the Adar1/2^KO versus WT conditions, MXE events were most affected, with 78% showing higher inclusion of the alternative exon (binomial test, P = 2 × 10^–8), where 77% were among inosine-containing transcripts (binomial test, P = 5 × 10^–8). This indicates a broad increase in exon switching in Adar1/2^KO mice, irrespective of inosine modifications (Fig. 5d). To corroborate some of our findings, we chose some transcripts that exhibited differential AS and employed RT-qPCR with primers designed to amplify these AS transcripts (Fig. 5e, Supplementary Fig. 7, Supplementary Table 5). Our efforts successfully confirmed several distinct patterns of AS in both Adar2^KO and Adar1/2^KO mouse brain datasets (Supplementary Fig. 7c, d).

We next asked what kind of alternatively spliced transcript are mostly impacted by ADAR2 depletion, and by the combined loss of ADAR2 and ADAR1 catalytic activity. Therefore, we carried out Gene Ontology analysis, which revealed that transcripts exhibiting AS in Adar2^KO brains were significantly enriched for functions related to mRNA splice site selection (enrichment = 5.26, P = 8.21 × 10^–3), mRNA processing (enrichment = 3.13, P = 1.81 × 10^–5), translation (enrichment = 2.04, P = 7.64 × 10^–4), synapse organization (enrichment = 2.3, P = 6.75 × 10^–5), ion transmembrane transport (enrichment = 3.82, P = 6.88 × 10^–3), and receptor internalization (enrichment = 3.13, P = 9.92 × 10^–3). In contrast, when both active ADAR enzymes were catalytically non-functional, additional pathways relevant to neuronal function emerged, including trans-synaptic signaling (enrichment = 2.05, P = 7.37 × 10^–4), neurotransmitter receptor transport (enrichment = 4.79, P = 1.56 × 10^–2), dendrite development (enrichment = 2.85, P = 6.26 × 10^–4), and learning (enrichment = 2.17, P = 3 × 10^–2) (Fig. 5f). Building on these findings, we focused our analysis on specific cases of particular biological interest. Notably, we observed altered AS in Gria4, Tra2a, Dnajc5, and Traf6 (Fig. 5e, Supplementary Fig. 7d). GRIA4, crucial in brain function, plays a key role in synaptic transmission⁴⁰. TRA2a regulates pre-mRNA splicing and is linked to Alzheimer’s disease pathology⁴¹. DNAJC5 is pivotal in membrane trafficking and protein folding, exhibiting neuroprotective properties against neurodegeneration⁴². TRAF6, a neuroinflammatory protein, contributes to central nervous system diseases including neurodegenerative diseases⁴³.

To investigate whether the differential splicing patterns observed directly result from loss of A-to-I editing, we performed rescue experiments by restoring ADAR2 function through stereotaxic delivery of adeno-associated viruses (AAVs) into the mouse brain. We compared two conditions: (1) AAVs expressing fully functional ADAR2 and (2) AAVs expressing catalytically inactive ADAR2 (Supplementary Fig. 7e). The functional ADAR2, but not the inactive mutant, successfully rescued wild-type splicing patterns, demonstrating that this regulation depends on ADAR2’s editing activity. Specifically, we observed consistent restoration of exon inclusion in Gria4 and Traf6 transcripts to WT levels. In contrast, effects on Traf2a splicing isoforms showed variability between biological replicates (Supplementary Fig. 7f.). These results show editing-dependent control of splicing for specific RNA targets, while other targets may be subject to additional regulatory influences.

Overall, these transcripts exemplify alterations in AS upon ADAR enzymes functional depletion, shedding light on critical mechanisms for overall brain function.

ADAR enzymes impact circRNA expression profiles

Given that circRNA biogenesis is mediated by the canonical splicing machinery¹⁴, we sought to investigate how inosine in pre-spliced RNA influences the circRNA landscape by comparing circRNAs in the mouse brain after functionally depleting ADAR enzymes. To systematically investigate the impact of ADAR2 depletion, and the combined loss of ADAR2 and ADAR1 catalytic activity on circRNA biogenesis and expression, we employed a dual sequencing strategy. First, we performed high-depth Illumina sequencing on Adar2^KO and Adar1/2^KO, and WT mouse brains to identify and quantify circRNAs through backsplice detection (Fig. 6a). This approach identified 573 differentially expressed circRNAs in Adar2^KO (326 upregulated, 247 downregulated) and 1487 in Adar1/2^KO (562 upregulated, 925 downregulated) compared to WT (Fig. 6b, Supplementary Table 6, Supplementary Table 7), indicating an association between ADAR perturbation and changes in circRNA abundance. Among the differentially regulated circRNAs, we identified circRNAs originating from inosine-containing pre-mRNA transcripts. Specifically, 299 circRNAs were differentially expressed from pre-mRNAs where inosine is regulated by ADAR2, while 274 circRNAs were differentially expressed from pre-mRNAs without inosine. In the context of the combined loss of ADAR2 and ADAR1 catalytic activity, 982 differentially expressed circRNAs originated from inosine-containing pre-mRNAs, whereas 505 circRNAs were differentially expressed from transcripts without inosine. Notably, the majority of these differentially expressed circRNAs stemmed from edited pre-mRNA transcripts (Fig. 6b, Supplementary Table 6, Supplementary Table 7).

Fig. 6.

Influence of Inosine-Regulated Transcripts by ADAR Enzymes on Circular RNA Profiles. (a) Investigation of circular RNAs (circRNAs) in mouse brains following loss of ADAR2 and the combination of ADAR2 loss and ADAR1 catalytic activity using ONT and Illumina sequencing to identify and predict circRNAs landscape. (b) Volcano plot illustrating the fold-change (log₂) and significance of circRNA expression changes between Adar2^KO and Adar1/2^KO versus WT mouse brains. This plot highlights circRNAs originating from edited pre-mRNA transcripts in red and circRNAs originating from non-edited pre-mRNAs in black. (c) Analysis of circRNA lengths and distributions in Adar2^KO and Adar1/2^KO mouse brains, demonstrating predominant sizes between 500 to 1,000 nucleotides. Each line represents data per biological replicate. n = 3. (d) Composition analysis of circRNAs in WT, Adar2^KO, and Adar1/2^KO mouse brains. circRNAs are predominantly composed of exonic sequences, across all samples. (e) Validation of differential AS patterns in Adar2-circ, a circRNA originating from Adar2 pre-mRNA, in WT, Adar2^KO and Adar1/2^KO mouse brains. RT-qPCR results confirm distinct AS patterns in Adar2^KO and Adar1/2^KO mouse brain datasets. Expression levels were normalized to the housekeeping gene ß-actin and to WT levels n = 3; mean ± SD; t-test; **P < 0.01; ***P < 0.001. (f) Illustration of Adar2-circ formation from the Adar2 gene where exon 6 has been deleted in Adar2^KO and Adar1/2^KO mice. Ex = exon; KO = knockout. (g) Detection and validation of differentially regulated circRNAs, Fstl5-circ-1 and Fstl5-circ-2 from the Fstl5 gene, Mettl6-circ from the Mettl6 gene, Gabra2-circ-1 and Gabra2-circ-2 from the Gabra2 gene, in Adar2^KO and Adar1/2^KO mouse brains. Expression levels were normalized to the housekeeping gene ß-actin and to WT levels. n = 3; mean ± SD; t-test; **P < 0.01; ***P < 0.001.

Complementing this, we used ONT sequencing with rolling circle amplification to validate circRNA structure across their full length. Total RNA was extracted from Adar2^KO and Adar1/2^KO mouse brains, followed by rolling circle amplification of circRNAs for ONT sequencing (Fig. 6a). This approach mirrors our previously established method for circRNA identification in WT mouse brain samples (Fig. 2b). By identifying extended RNA repeats, typically non-repetitive, we could uncover complete circRNA molecules. In the brains of Adar2^KO and Adar1/2^KO mice (n = 3 biological replicates per genotype), we generated 5.3, 6.2, and 6.1 million clean sequence reads for Adar2^KO, and 6.5, 7.6, and 9.0 million clean sequence reads for Adar1/2^KO. Subsequently, in Adar2^KO and Adar1/2^KO mice, we discovered 35,772, 47,174, 41,111 and 46,818, 33,077, 23,441 distinct circRNAs, respectively (Supplementary Table 1). Further analysis involved scrutinizing the lengths of circRNAs ONT identified in Adar2^KO and Adar1/2^KO samples, comparing these with our previous observations in WT samples. Notably, no substantial differences in circRNA sizes and distributions were observed. The predominant size range for circRNAs across all samples fell between 500 to 1000 nucleotides (Fig. 6c, Supplementary Fig. 8a), primarily composed of exonic sequences (Fig. 6d). While approximately 50% of ONT-identified circRNAs were confirmed by Illumina (Supplementary Fig. 8b), the remaining ONT-specific candidates—though clearly circular – we could not detect through Illumina sequencing. These putative novel circRNAs represent important targets for future validation through either ultra-deep ONT sequencing or emerging circRNA-specific technologies. The Illumina-first approach provided statistically robust comparative quantification, while ONT sequencing offered structural validation, together creating a comprehensive circRNA expression profile. This combined strategy overcame the limitations of ONT’s lower sequencing depth for differential expression analysis while capitalizing on its advantages for structural confirmation.

To further validate our circRNA findings from ONT and Illumina sequencing approaches, we used RT-qPCR with divergent primers that span the backsplice junction of selected circRNAs derived from inosine-containing transcripts (Supplementary Fig. 8c), confirming their presence and differential expression (Fig. 6e, g, Supplementary Fig. 8d, Supplementary Table 8). The circRNA alterations observed upon ADAR2 and ADAR1/ADAR2 perturbation support a role for A-to-I editing in circRNA biogenesis, while not excluding significant contributions from editing-independent mechanisms.

Given that circRNAs are derived from pre-mRNAs through backsplicing and considering that splicing is a co-transcriptional process^14,30, it is plausible that RNA modifications could potentially impact transcription speed, allowing the splicing machinery more time to splice. To investigate whether the observed differential expression changes in circRNAs and splicing could be attributed to alterations in transcriptional speed induced by changes in inosine modifications in pre-mRNAs, we also explored if inosine modifications on transcripts affect transcription speed. However, our analysis revealed no major differences in transcription speed between transcripts containing inosine and those without inosine in both Adar2^KO and Adar1/2^KO samples compared to WT controls (Supplemental Fig. 9). Therefore, it is unlikely that the differential changes observed in circRNAs are a result of alterations in transcriptional speed mediated through inosine modifications.

Following our investigation into the impact of inosine on transcriptional speed and circRNA expression, our study revealed compelling insights into specific circRNAs. Notably, a circRNA originating from Adar2 pre-mRNA, named Adar2-circ, displayed a striking upregulation of at least 10-fold in both Adar2^KO and Adar1/2^KO mouse brains (Fig. 6e, f). While the alteration in expression of this circRNA could be linked to structural genomic changes resulting from the removal of exon 6 of the Adar2 gene during the generation of Adar2^KO and Adar1/2^KO mice (Fig. 6f), it is also plausible that inosine changes in Adar2 transcripts contribute to the modulation of Adar2-circ. We detected two differentially regulated circRNAs, Fstl5-circ-1 and Fstl5-circ-2, derived from the Fstl5 gene (Fig. 6g). Fstl5 gene is linked to hearing loss⁴⁴. Given that Adar2^KO mice exhibit impaired hearing abilities⁴⁵, these findings suggest that hearing loss could potentially be associated with inosine changes and circRNA alterations, possibly through downstream effects of altered circRNAs Fstl5-circ-1 and Fstl5-circ-2. Additionally, we noted a decrease in the expression of Mettl6-circ, originating from the Mettl6 pre-mRNA transcript, following ADAR2 depletion, though not in Adar1/2^KO mice (Fig. 6g). The discrepancy in the impact on Mettl6-circ between the two knockout models warrants clarification (Fig. 6g). METTL6, a methyltransferase responsible for adding an RNA modification by methylating cytosine to generate 3-methylcytidine (m3C), plays a significant role in RNA regulation⁴⁶. While m3C was not detected in our LC–MS/MS analyses, it is conceivable that ADAR2 could influence m3C levels, alongside other RNA and DNA modifications observed, potentially through the regulation of METTL6 via inosine-linked Mettl6-circ modulation. Another intriguing finding was the heightened abundance of Gabra2-derived circRNAs, Gabra2-circ-1 and Gabra2-circ-2, in the brains of Adar2^KO and Adar1/2^KO mice (Fig. 6g). Gabra2 encodes an α-subunit of the GABA-A receptor in the brain, a pivotal component essential for mediating inhibitory neurotransmission and maintaining neural excitability⁴⁷. Some neurotransmitter receptors, including subunit GABRA3 of the GABA-A receptor, are edited by ADAR enzymes in protein-coding regions leading to changes in amino acid incorporation. This modifies the kinetics of these receptors, modulating their synaptic responses⁴⁸. Our data suggest another potential layer in which ADAR2 and ADAR1/ADAR2, possibly through inosine modulating circRNA expression, could affect neuron activity. These findings highlight a potential dual regulatory role for ADAR enzymes: not only do they edit protein-coding regions of neurotransmitter receptors, such as GABRA3, but they might also influence neuronal activity through wide-ranging inosine-mediated regulation of circRNA expression.

Discussion

In conclusion, our study aimed to investigate RNA modifications broadly implicated in AS in the brain, a process critical for generating the heterogeneous transcriptome required for specialized neuronal functions¹⁷. To identify RNA modifications potentially involved in AS, we focused on modifications likely enriched in intronic regions, hypothesizing that such modifications could influence both AS and the circRNA backsplicing landscape. Using LC–MS/MS, we identified several modification candidates that seem enriched in introns, including inosine, which has been previously implicated in brain function⁴⁸. Our intron-enrichment strategy, while unable to provide precise modification localization or a complete subcellular profile, effectively pinpointed candidate RNA modifications with a likely intronic bias.

We further focused on inosine, leveraging high-throughput sequencing strategies to confirm its enrichment in intronic regions and explore its association with AS and circRNA biogenesis in Adar2^KO and Adar1/2^KO mutant mice. While we classified editing sites as ADAR1- or ADAR2-specific based on their pronounced loss in the respective knockout models, we note that this represents a preference rather than an exclusive relationship. The landscape of A-to-I editing is likely more complex, encompassing sites with partial dependency on both enzymes and potential compensatory effects, particularly upon long-term loss of one ADAR enzyme. Our classification scheme robustly identifies sites with a strong primary dependence but may not capture the full spectrum of regulatory interactions that could occur in different cellular states or developmental contexts.

By using these models, we unexpectedly discovered that perturbations in ADAR2 and ADAR1/ADAR2 activities lead to changes in other RNA and even DNA modifications. This suggests that ADAR2 and ADAR1/ADAR2 may regulate other epigenetic and epitranscriptomic modifiers, most likely indirectly, through inosine-dependent or inosine-independent mechanisms. These findings highlight a broader regulatory role for ADAR enzymes beyond A-to-I editing, potentially influencing multiple layers of gene regulation in the brain.

Only Adar1/2^KO showed a measurable reduction of total RNA inosine, whereas Adar2^KO alone did not. This likely reflects that a substantial fraction of bulk inosine signal derives from a few highly abundant RNA species with heavily edited sites (for example, tRNAs modified by ADATs³⁵), while ADAR2 primarily contributes numerous low-stoichiometry edits dispersed across less abundant RNAs. Consequently, losing ADAR2 alone removes many sites but too few inosine-bearing molecules to shift the global pool, whereas combined loss of ADAR1 and ADAR2 eliminates a sufficiently large molecular fraction to be detected. Our data also highlights a critical, yet often overlooked distinction in epitranscriptomics: the difference between the frequency of modification sites and the absolute molecular abundance of a modification. Sequencing identifies editing sites but is usually insensitive to their stoichiometry, potentially over-representing low-level editing events. In contrast, mass spectrometry quantifies the total molecular flux of modified nucleosides. Our finding that global inosine levels are relatively stable despite changes in ADAR editing sites suggests that the cellular “inosine budget” may be dominated by high-stoichiometry modifications on abundant non-coding RNAs, such as tRNAs, that are not regulated by ADAR enzymes. Our mass spectrometry data thus provides a novel, quantitative perspective on the epitranscriptome that has been largely hidden by sequencing-based analyses alone.

Given that ADAR enzymes are known to influence AS^20,25,49, it is plausible that alterations in the intron landscape itself or any changes in protein translation contribute to the observed alterations in RNA modifications. This suggests that the changes in modification abundance are likely not solely due to inosine itself, although inosine may still play a significant role. Instead, these effects may arise from broader regulatory mechanisms involving ADAR enzyme-mediated splicing and transcriptome remodeling.

Previous studies have suggested interactions between epitranscriptomic modifiers^50,51. Our study extends these findings by quantitatively analyzing the global impact of ADAR2 depletion, and the combined loss of ADAR2 and ADAR1 catalytic activity on a broad spectrum of RNA and DNA modifications. Xiang et al. showed that A-to-I editing preferentially occurs on transcripts depleted of m6A. Specifically, suppression of m6A-catalyzing enzymes resulted in increased association of ADAR1 to transcripts and to their editing, while global levels of A-to-I editing remained unchanged⁵⁰. In our data, upon perturbation of ADAR enzymes, no global m6A variations can be observed in total RNA and long transcripts. However, as Xiang et al. reported, an interplay might still exist between these RNA modifications that is not reflected in the global level of m6A in total and long RNAs. In contrast, Yi Li et al. demonstrated that ADAR1 promotes the overexpression of the m6A-catalyzing enzyme METTL3 via A-to-I editing in breast cancer cells and consequently increases m6A levels⁵¹. In our data, consistent with Yi Li et al., we detected a decrease of m6A upon ADAR2 depletion and combined loss of ADAR2 and ADAR1 catalytic activity, but only affecting small RNAs. Since METTL3 can target small RNAs such as microRNAs, absence of ADARs could therefore explain our observation of a m6A decrease in small RNAs⁵². Notably, we observed a 18% reduction in m5dC levels in DNA following the combination of ADAR2 loss and ADAR1 catalytic depletion. This m5dC depletion might be attributed solely to ADAR1, as upon ADAR2 depletion m5dC levels do not change. The Adar1^E861A/E861A point mutation catalytically inactivates the ADAR1 protein while maintaining other editing-independent functions⁵³. Therefore, the global depletion of m5dC levels in DNA can be attributed to A-to-I editing mediated by active ADAR1. While this regulation could be indirect—potentially involving shifts in cell type composition or other downstream effects—this finding provides novel insight into the interplay between RNA and DNA modifications, with important implications for gene regulation.

In this study, we investigated the RNA modification inosine in the brain and identified AS events occurring upon ADAR2 loss and the combination of ADAR2 loss and ADAR1 catalytic depletion in key genes regulating normal brain functions. Previous work already implicated some RNA modifications in splicing. For example, the m6A reader protein YTHDC1 regulates splicing by recruiting splicing factors and facilitating their access to binding regions of the m6A reader¹⁹. The RNA modification Ψ was demonstrated to modulate splicing in vitro using nuclear extracts and a splicing reporter¹¹. Hsiao et al. located A-to-I editing sites in 3’ splice sites and found that editing in these specific sites affected alternative splicing in cell lines. Using minigene reporter assays with A-to-G mutations in some of these 3’ splice sites, the authors showed that A-to-G mutations changed alternative splicing²⁰. However, this approach did not demonstrate that inosine itself can regulate splicing, as A-to-I was not changed, but instead A-to-G mutations were tested. Another study using RNA-seq data from cell lines found that ADAR1 knockdown changed global alternative splicing patterns, suggesting that editing regulates splicing⁴⁹. In contrast, several studies proposed the opposite, namely that pre-mRNA splicing could control RNA modifications deposition. For example, Licht et al. discovered, through nascent RNA sequencing analysis and splicing manipulation, that lack of splicing factors NOVA1 or NOVA2 promote global changes in A-to-I editing levels. Intron retention changes correlated with editing level alterations suggesting that splicing controls editing levels.²⁸. However, splicing regulation through RNA modifications might be more complex. For example, an alternative spliced isoform of the m6A-catalyzing enzyme METTL3 found in hepatocellular carcinoma that retains introns 8 and 9 and has a shorter exon 4, leads to a global decrease in m6A modifications⁵⁴. The splicing variant retaining intron 8 and 9 has been identified in several human tissues. This variant is retained in the nucleus and therefore METTL3 protein expression is reduced⁵⁵. Similarly, inhibition of exon 10 skipping in the m6A-catalyzing enzyme METTL14 produces a global elevation of m6A in pancreatic cancer⁵⁶. These findings add another level of complexity to the intricate connection between RNA modifications and splicing related processes.

Using an innovative multi-platform sequencing approach, we characterized circRNA profiles in WT, Adar2^KO, and Adar1/2^KO mice, while simultaneously detecting potential candidate inosine modifications within these transcripts. These putative editing sites are presented as a hypothesis-generating resource to enable and guide future targeted validation in circRNAs.

Our study also revealed a cohort of putative novel circRNAs that, while demonstrating clear circularity through rolling circle amplification, will require future validation through either deeper sequencing or emerging circular RNA-specific technologies. These findings expand our understanding of ADAR enzyme-mediated regulation of the circRNA landscape in the mammalian brain. So far, research on RNA modifications on circRNA molecules has mainly focused on m6A⁵⁷, but precedents of other RNA modifications in circRNAs exist^58–60. Previous findings rely on circRNA enrichment followed by modification-specific antibodies for RNA immunoprecipitation. However, this methodology has limitations. CircRNA enrichment with RNA exonucleases still contain many resistant linear RNA molecules that can potentially become false positives⁶¹. Also, due to short-RNA sequencing methodologies, only modifications relatively close to backsplice junctions can be identified. In addition, there might be high false-positive rates as a result of non-specific antibody binding⁶². Therefore, our methodology overcomes these limitations and provides an alternative and complementary approach to identify inosine within circRNAs. Furthermore, our analysis of circRNAs offers insights into the circRNA landscape in the brain. The intricate interplay among ADAR enzymes, inosine modification in pre-mRNA, and circRNA biogenesis underscores the complex regulatory landscape shaped by RNA modifications. The regulation of neuronal activity by ADAR enzymes has traditionally been attributed to their role in editing protein-coding regions of neurotransmitter receptors. However, our study reveals an additional layer of complexity, demonstrating that ADARs, through inosine, can also modulate circRNA expression, potentially influencing synaptic function and neuronal circuits. These findings open new avenues for understanding how RNA editing shapes brain physiology and pathology. Notably, these observations delineate associations between perturbation of ADAR enzymes, alternative splicing, circRNA expression, and the broader modification landscape in the mouse brain, while explicitly not inferring direct causation from the presented data.

Instead, our collective findings highlight the intricate interplay among RNA modifications and the complex transcriptome governing brain function. This comprehensive exploration not only reveals the multifaceted regulatory network orchestrated by RNA modifications but also emphasizes their significant implications for neurological processes. While conventional research on diseases linked to inosine deregulation, such as mood disorders and schizophrenia⁶³, has predominantly concentrated on inosine at coding sites, our findings underscore the importance of intronic inosine in shaping alternative RNA processing pathways. These global changes represent a compelling landscape for future mechanistic studies, which will require targeted, modification-specific functional assays (e.g., locus-specific writer/eraser perturbations or mapping with modification-specific methods) to test whether any are direct consequences of ADAR activity versus secondary remodeling. These insights will not only enhance our understanding of the regulatory landscape governed by RNA modifications but also pave the way for novel avenues of research aimed at deciphering the complexities of brain function and disease pathology.

.

Methods

Mouse strains

C57BL/6 J WT mice were provided by the Laboratory Animal Resource Center (LARC), Chinese Institute for Brain Research, Beijing (CIBR) and were originally purchased from the Jackson Laboratory. All mice were housed in a 12:12 light–dark cycle, under controlled climate and enrichment environmental conditions with access to sterile food and water ad libitum.

Viably rescued ADAR2 enzyme deficient Adar2^KO knockout mice, were originally referred to as Adar2^−/−Gria2^R/R64. This mouse strain used for this research project, B6.129-Adarb1^tm1Phs Gria2^tm1Phs/Mmnc, RRID:MMRRC_034679-UNC, was obtained from the Mutant Mouse Resource and Research Center (MMRRC) at University of North Carolina at Chapel Hill, an NIH-funded strain repository, and was donated to the MMRRC by Kazuko Nishikura, Ph.D., The Wistar Institute. Adar2^−/− mice are viable but die around 3 weeks after birth following repeated episodes of epileptic seizures. A point mutation in the AMPA receptor subunit Gria2 gene, at the ADAR2-specific Q/R editing site, rescues lethality in mice deficient in ADAR2 enzyme, leading to the expression of the edited form of the transcript⁶⁴.

Viably rescued catalytically inactive ADAR1 and ADAR2 enzyme deficient Adar1/2^KO mice (Adar1^E861A/E861A Ifih^−/− Adar2^−/− Gria2^R/R) were described previously²³. The editing inactivating point mutation in Adar1 (Adar1^E861A/E861A) is embryonically lethal at E12.5 with an excess production of type I interferons. This lethality can be rescued by suppressing the cytosolic dsRNA immune sensor MDA5 encoded by the Ifih1gene²⁴. Due to limited natural mating capacity between Adar1/2^KO mice, initial propagation was performed via in vitro fertilization (IVF). Sperm from Adar1/2^KO males (Adar1^E861A/E861A Ifih^−/− Adar2^−/− Gria2^R/R) and oocytes from Adar1/2^KO females, either homozygous (Adar1^E861A/E861A Ifih^−/− Adar2^−/− Gria2^R/R) or heterozygous for Adar1 (Adar1^+/E861A Ifih^−/− Adar2^−/− Gria2^R/R) were used. Following IVF, the colony was maintained through natural mating of mice heterozygous for Adar1 (Adar1^+/E861A Ifih^−/− Adar2^−/−Gria2^R/R).

Genotyping

Genomic DNA was PCR amplified using the 2 × Taq PCR Mix (BGI, LH-001). PCR primers are listed in Supplementary Table 9. Positive and negative controls were always processed in parallel. Gel electrophoreses were carried in 0.5 × Tris–acetate-EDTA buffer. Gels were imaged with the Gel Doc System (Bio-Rad). Molecular weights of DNA fragments were compared to the DNA Marker II (Tiangen, MD102-02). For Adar1 genotyping, PCR bands were isolated, and DNA purified from the gel. The Adar1 genotype was determined through Sanger Sequencing.

Total RNA and DNA extraction from mouse brains

Mice were euthanized, brains were immediately isolated, quickly washed with ice-cold DPBS and fully homogenized in 3 ml TRIzol™ Reagent (Invitrogen, 15,596,026) using a Dounce tissue homogenizer. Total RNA and DNA were extracted according to the manufacturer’s instructions and quantified using Nanodrop One Spectrophotometer (Thermo Scientific). Total RNA was treated with 0.2U of DNaseI (Vazyme Biotech, EN401) per 1 μg of total RNA, and subsequently processed with phenol:chloroform:isoamyl (125:24:1, pH < 5.0) followed by isopropanol precipitation. DNA was purified with phenol:chloroform:isoamyl (125:24:1, pH > 7.8) followed by ethanol precipitation. RNA and DNA yield and purity were measured again with Nanodrop One Spectrophotometer (Thermo Scientific). RNA integrity was assessed using Fragment Analyzer™ Automated CE System (Agilent). RNA Clean & Concentrator™-5 kit (ZymoResearch, R1016) was used to separate total RNA into small (< 200nt) and long (> 200nt) RNA following the manual’s instructions.

LC–MS/MS analysis

For LC–MS/MS analysis, RNA and DNA samples were digested into single nucleosides using Nucleoside Digestion Mix (New England BioLabs, M0649S), incubating the samples at 37 °C for 2 h. After digestion, the 2’-deoxyadenosine isotope (dA-5) (Cambridge Isotope Laboratories, NLM-3895–25) was added as a spike-in to each sample at a final concentration of 10 nM. Using LC–MS/MS, nucleosides were separated based on their mass charge ratio (m/z) and the obtained peak areas were normalized with the dA-5 isotope. For each nucleoside, a synthetic standard was serial diluted and run in parallel to generate a standard curve. The peak area ratio was used to calculate the concentration of each nucleoside. Concentrations were extrapolated from standard curves based on the peak area ratio. The resulting concentration for each RNA nucleoside was normalized by the unmodified guanosine nucleoside (G) in each sample. For DNA, the concentration of each deoxynucleoside was normalized to its unmodified counterpart. To assess the statistical significance, multiple t-tests were applied, considering each nucleoside individually.

Cell nuclei isolation

Whole brain from adult C57BL/6 J WT mice (female, 6-month-old) were used to isolate nuclei. One mouse brain corresponds to one biological replicate. Mice were euthanized and brains were immediately isolated and processed. Tissue was dissociated in Nuclei EZ lysis buffer (Sigma-Aldrich, N3408) with 5U/uL Murine RNase Inhibitor (Vazyme Biotech, R301) using a 7 ml Dounce tissue homogenizer. From any tissue homogenate, a fraction was kept as a matching control for total RNA extraction. After homogenization and 5-10 min incubation in ice, cold sucrose was added to the homogenate to a final concentration of 1.15 M. The homogenate mixture was then placed above a 1.8 M sucrose solution and centrifuged at 24,000 g for 45 min at 4 °C. Pelleted nuclei from the high-density sucrose phase were resuspended and washed in a DPBS solution with 1% BSA and RNA was extracted with TRIzol™ Reagent (Invitrogen, 15,596,026). TurboCapture mRNA Kit (QIAGEN, 72,251) was used to isolate polyA RNA according to the manual’s instructions. To test if introns are enriched in nuclear RNA, we carried out Illumina sequencing following the procedure indicated in the section below (High-throughput mRNA Illumina sequencing library preparations). The generated FASTQ files were trimmed using Trimmomatic (version 0.39)⁶⁵ and reads were aligned to the GRCm39 reference genome using STAR (version 2.7.10b)⁶⁶.

RT-qPCR

DNaseI-treated total RNA processed into cDNA from mouse brains was used for RT-qPCR evaluations. PrimeScript™ RT reagent Kit (Takara Bio, RR037A) was used for reverse transcription, using 1 μg of total RNA and 100 μM random primers. cDNA was diluted 1:50 and used as a template for qPCR with SYBR® Green JumpStart™ Taq ReadyMix™ (Sigma-Aldrich, S4438) in a QuantStudio™ 3 Real-Time PCR System (Applied Biosystems™). For gene expression and circRNA expression analysis, expression values were normalized to β-actin. For alternative splicing analysis, expression values were normalized to β-actin and to the expression of the specific gene. Relative gene expression was determined using the 2^−ΔΔCt method. All primers used are listed in Supplementary Table 9.

Cortical primary neuron cultures

Cortical primary neurons were obtained from 6–8 newborn rats (SD strain, purchased from Beijing Vital River Laboratory Animal Technology Co., Ltd.). Cerebral cortices were isolated and treated with Trypsin–EDTA (Trypsin–EDTA (0.25%), phenol red, Gibco™, 25,300,054) diluted 1:2 in DPBS for 10 min at 37 °C. Trypsin was inactivated with Dulbecco’s Modified Eagle Medium (Gibco™, C11965500BT) supplemented with 10% Fetal bovine serum (VisTech™, SE100-B) and cells were dissociated through gentle pipetting. Dissociated cells were pelleted through slow speed centrifugation (1000 rpm, 3 min) and resuspended in Neurobasal™ Medium (Gibco™, 21,103,049) supplemented with 1 × B-27™ Supplement (50 × , serum free, Gibco™, 17,504,044), 1 × GlutaMAX™ Supplement (100 × , Gibco™, 35,050,061) and 1% Penicillin–Streptomycin (10,000 U/ml, Vetec™, V900929). Cells were plated on PolyD-Lysine (0.1 mg/ml, Gibco™, A3890401) precoated 100 mm dishes at a density of 10⁵ cells/cm² and kept at 37 °C in a humidified, 5% CO² atmosphere.

Splicing inhibition experiment in cortical primary neurons

Primary cortical neurons were cultured for 48 h. Next, 100 μM 4SU (Sigma-Aldrich, T4509) was added to label any new RNAs being transcribed. 1 μM PlaB (AdooQ Bioscience, A12699) was added to inhibit splicing. Cells were harvested after 16 h incubation and total RNA was extracted with TRIzol™ Reagent (Invitrogen, 15,596,026). Newly transcribed RNA was isolated as previously described³¹. In brief, 4SU-labeled RNA was biotinylated with 2ug of EZ-Link HPDP-Biotin reagent (Thermo Scintific, A35390) per 1 μg total RNA, in a final concentration of 0.2 mg/ml. After the biotinylation reaction, RNA was extracted with Phenol/Chloroform and precipitated with isopropanol. Next, the biotin-labeled RNA was isolated using streptavidin-coated magnetic beads (Dynabeads™ MyOne™ Streptavidin C1, Invitrogen, 65,001) following the manufacturer’s instructions. To release the RNA, magnetic beads were incubated with 200 µl of 100 mM DTT for 3 min and supernatant was collected into a new tube. RNA was precipitated overnight with ethanol. The TurboCapture mRNA Kit (QIAGEN, 72,251) was used to isolate polyA RNA, following the manufacturer’s instructions.

circRNA high-throughput sequencing library preparations

For circRNA high-throughput sequencing, brains from adult 3–4-month-old male Adar2^KO, Adar1-2^KO and WT mice were isolated. Brains from 3 mice were isolated for each condition, with each brain representing one biological replicate and resulting in 3 biological replicates per condition (n = 3). RNA was extracted from each sample, as described above. A small part of the RNA was used for circRNA high-throughput sequencing described in this section. The remaining RNA from the same biological sources was used for additional processing and sequencing with other approaches and for some follow-up validation tests. This provided us with matching samples for comparative downstream analyses. For circRNA sequencing library preparations, 5 μg of total RNA were used to deplete ribosomal RNA using the RiboMinus™ Eukaryote System v2 (Invitrogen™, A15026). To deplete linear RNAs, the ribosomal RNA depleted RNA was then treated with 10U Ribonuclease R (BioResearch Technologies, RNR07250) at 37 °C for 30 min and was subsequently cleaned up with the RNA Clean & Concentrator™-5 kit (ZymoResearch, R1016). The eluted RNA was reverse transcribed with SuperScript® IV Reverse Transcriptase (Invitrogen™, 18,090,010) using a random hexamer primer with an adaptor sequence (called Primer1, Supplementary Table 9). This adaptor sequence later facilitates PCR amplification of the entire cDNA molecule. The reverse transcription reaction was incubated at 23 °C for 10 min followed by 50 °C for 30 min and then inactivated at 80 °C for 10 min. Subsequently, samples were treated with 1 μl RNase Cocktail™ Enzyme Mix (Invitrogen™, AM2286) at 37 °C for 10 min to eliminate RNA fragments. Reverse transcription reaction products were cleaned up using AMPure XP beads (Beckman Coulter™, A63881) according to manufacturer’s instructions. The Terminal Transferase enzyme (New England BioLabs, M0315) was used to add a deoxycytidine (dC) tail to single stranded cDNA molecules. This dC tail enables the addition of an adaptor sequence at that cDNA end, facilitating PCR amplification of the entire cDNA molecule. For the dC addition reaction, 10U of Terminal Transferase enzyme were incubated with 100 µM dC at 37 °C for 30 min. Reactions were again cleaned up using AMPure XP beads (Beckman Coulter™, A63881). To synthesize the reverse complementary strand, we used an oligo(dG)₁₂ primer coupled with an adaptor sequence (called OligodGAdap, Supplementary Table 9). The oligo(dG)₁₂ recognizes the dC tail previously added and acts as a primer for the synthesis of the complementary strand of the cDNA. In this way, we ensure the synthesis of the complementary strand covers the entire size of the amplified cDNA. The OligodGAdap primer was incubated with the cDNA and the LongAmp® Taq 2 × Master Mix DNA polymerase (New England BioLabs, M0287) at 98 °C for 3 min, 63 °C for 1 min and 65 °C for 1 h. The reaction resulted in double-stranded cDNA. The cDNA was then purified, and size selected with AMPure XP beads (Beckman Coulter™, A63881) with a sample-to-bead ratio of 0.5 × . The purified cDNA was subsequently PCR amplified in the LongAmp® Taq 2 × Master Mix (New England BioLabs, M0287) with 14 cycles, with an elongation time of 3 min per cycle, using circDNA-F and circDNA-R-TSO primers (Supplementary Table 9). These primers contain the same sequence as the adaptor sequences from Primer1 and OligodGAdapt, respectively. Therefore, they recognize their complementary sequences at the ends of the double-stranded cDNA molecules. In this way, the amplification of the entire cDNA molecules can be ensured. PCR products were subsequently purified, and size selected with AMPure XP beads (Beckman Coulter™, A63881), using the sample-bead ratio of 0.5 × . All oligo sequences are listed in Supplementary Table 9.

The circRNA cDNA PCR products were sequenced using two sequencing platforms: Illumina sequencing ensured high sequencing depth, while the ONT sequencing platform was carried out in parallel on the same material to obtain long sequencing reads to confirm the presence of circRNAs (see results section for more information). To generate circRNA libraries for ONT sequencing, 100 ng of the circRNA cDNA PCR products were processed with the Ligation Sequencing Kit V12 (Oxford Nanopore, SQK-LS112). These libraries were sequenced on FLO-MIN112 flow cells. To generate circRNA libraries for Illumina sequencing, 5 ng of the circRNA cDNA PCR products were processed into libraries using the TruePrep DNA Library Prep Kit V2 for Illumina® (Vazyme Biotech, TD502). After cDNA tagmentation and adaptor ligation, 9 PCR cycles were carried out followed by size selection using AMPure XP beads (Beckman Coulter™, A63881) following the kit’s protocol to enrich for 350nt insert lengths. The generated libraries were sequenced using an Illumina NovaSeq 6000 instrument for 150nt paired end read sequencing.

Identification of circRNAs

For identification of circRNAs from Illumina sequencing data: FASTQ files were quality filtered using Fastp (version 0.19.7)⁶⁷. Clean FASTQ files were mapped to the mouse reference genome (GRCm39) using BWA (version 0.7.17)⁶⁸. CIRI2 (version 2.0.6) was run on the aligned data to identify circRNAs⁶⁹. To quantify the identified circRNAs, we applied CIRIquant (version 1.1.2), using the biological replicate mode⁷⁰. To identify differential expressions in circRNAs between biological samples, the backsplice junction read counts (BSJ) were used as input for DESeq2 (version 2.11.40.8)⁷¹. CircRNAs meeting the following criteria were considered as differentially expressed: ≥ twofold-change, P < 0.05, minimum of 5 reads covering backsplice junctions, present in at least two of the three replicates, and no overlapping error bars between biological groups. For identification of circRNAs from ONT long read sequencing data: FASTQ files were obtained from FAST5 files using Guppy (version 6.1.2). Adapters and barcodes were trimmed using PoreChop (version 0.2.4)⁷². Reads shorter than 500nt were discarded from subsequent analyses. circRNAs were identified using CIRIlong (version 1.0.3)⁷³, using the GRCm39 reference mouse genome. circRNAs overlapping genomic repetitive regions present in RepeatMasker, encoding ribosomal RNA, or any circRNA that were shorter than 30nt were excluded from subsequent analyses.

Discovery of novel circRNAs

To detect novel, previously unknown circRNAs among those circRNAs we identified as described above, we compared our list to those reported in the circRNA atlas, circAtlas 3.0. Specifically, we downloaded the list of known circRNAs from the mouse species from circAtlas⁷⁴. Since these circRNAs were reported for genome version mm38, we used UCSC LiftOver⁷⁵ to convert these annotations to genome version mm39. To determine which of our circRNAs were already reported and which were novel, we used bedtools v2.30.0⁷⁶.

Identification of inosine in circRNAs

Inosine in circRNAs was identified using ONT sequencing data. Using our identified circRNAs (see section ‘Identification of circRNAs’), we separated our circRNAs into two groups: one group contained circRNAs generated from the positive ( +) strand, the other group included the ones originating from the negative (-) strand. From the cand_circ.fa file generated by CIRIlong (version 1.0.3)⁷³, using seqkit (version 0.12.0)⁷⁷ and standard bash scripting, we extracted only the reads supporting these circRNAs for each group. Next, we used seqtk (https://github.com/lh3/seqtk, version 1.3) to convert the resulting FASTA files to FASTQ formats. We then aligned these files to genome version mm39 using minimap2 (version 2.17)⁷⁸. Next, we ran Reditools2⁷⁹ with the following parameters: minimum read length: 50; minimum read quality: 0; minimum base quality: 0; minimum number of reads: 10; minimum number of editing events: 3. The minimum read quality and minimum base quality were set to 0 since we commenced with FASTA files, which lack quality values. For variants identified in reads corresponding to circRNAs from the + strand, we retained only A-to-G edits. Conversely, for variants in reads corresponding to circRNAs from the—strand, we retained only T-to-C edits. We excluded variants located in homopolymeric regions (> 5nt), regions annotated in the ENCODE blacklist⁸⁰, and positions identified as SNPs in the Mouse Genome Project (mgp_REL2021_snps) and Genome Reference Consortium Mouse Build 39 (GCA_000001635.9).

Identification of inosine sites in nascent high-throughput sequencing data

To identify A-to-I RNA editing sites in the mouse brain, previously generated nascent sequencing datasets were used. FASTQ files were downloaded from the European Nucleotide Archive (ENA), project PRJEB27264. Datasets from the following samples were used: ERR2631461, ERR2631465 and ERR2631466 corresponding to WT. ERR2631475, ERR2631476 and ERR2631477 corresponding to Adar2^KO. ERR2631478, ERR2631479 and ERR2631480 corresponding to Adar1/2^KO. For all datasets, adapters were trimmed using Trimmomatic (version 0.39)⁶⁵. Reads were aligned to the GRCm39 reference genome using STAR (version 2.7.10b)⁶⁶. Reditools2 was executed (minimum read length: 50; minimum read quality: 10; minimum base quality: 20; minimum number of reads: 10; minimum number of editing events: 3)⁷⁹. Variants in homopolymeric regions (> 5nt), regions annotated in the ENCODE blacklist⁸⁰ and positions identified as SNPs in the Mouse Genome Project (mgp_REL2021_snps) and the European Variation Archive (GCA_000001635.9_current_ids.bed) were excluded. Single nucleotide conversions were comprehensively identified. All variants were annotated using Homer v4.11. A-to-G variants with an editing rate of at least 1% and present in all three WT replicates were considered to represent A-to-I editing events. From this list, editing events that also appeared in the Adar1/2^KO datasets were excluded as false positives. The remaining A-to-I events were attributed to both ADAR1 and ADAR2 enzymes. Among these high-confidence sites, those absent in the Adar2^KO datasets were attributed to the ADAR2 enzyme. In contrast, editing events still present in the Adar2^KO datasets were attributed to the ADAR1 enzyme.

High-throughput mRNA Illumina sequencing library preparations

For high-throughput Illumina sequencing, brains from adult 3–4-month-old male Adar2^KO, Adar1-2^KO and WT mice were isolated. Brains from 3 mice were isolated for each condition, with each brain representing one biological replicate and resulting in 3 biological replicates per condition (n = 3). RNA was extracted from each sample, as described above. A small part of the RNA was used for standard high-throughput Illumina sequencing described in this section. The remaining RNA from the same biological sources was used for additional processing and sequencing with other approaches and for some follow-up validation tests. This provided us with matching samples for comparative downstream analyses. To facilitate high-throughput Illumina sequencing, the NEBNext® Ultra™ II RNA Library Prep Kit for Illumina® (New England BioLabs, E7770S) was used to construct sequencing libraries, following the manufacturer’s instructions. The following conditions where used: 1 μg of total RNA was used as starting material, applying the polyA workflow. cDNA fragments of 370 ~ 420nt were purified for 150nt paired end read sequencing using the Illumina NovaSeq 6000 instrument.

Identification of AS events

To identify AS events in high-throughput sequencing data, we analyzed spliced mature mRNA sequencing data generated from Adar2^KO, Adar1/2^KO, and WT mouse brain, as described above (see methods section: High-throughput Illumina sequencing library preparations, GSE275035). First, Fastp (version 0.19.7)⁶⁷ was used for quality filtering of the FASTQ files. Second, Trimmomatic (version 0.39)⁶⁵ was used to trim polyA tails. To identify AS events between Adar2^KO and WT and between Adar1/2^KO and WT samples, rMATS (version 4.1.2)⁸¹ was run on clean FASTQ files using default parameters and allowing read clipping. Junction count output files were considered for downstream analyses, visualizations and interpretations. Statistically significant AS events were identified based on the following criteria: FDR < 0.05, dPSI > 0.1, average junction coverage > 10 and non-overlapping PSI error bars between groups.

ADAR2 rescue experiments

We delivered recombinant AAVphp.eB vectors expressing ADAR2 variants via intracranial injection to Adar2KO neonates. For the experimental group, we injected three biological replicates with AAVphp.eB -SFFV-EGFP-ADAR2^E488Q, which expresses a catalytically hyperactive ADAR2 mutant fused to EGFP⁸². For the control group, three additional replicates received AAVphp.eB-SFFV-EGFP-ADAR2^E396A, expressing a catalytically inactive ADAR2 mutant similarly tagged with EGFP⁸³. All injections were performed unilaterally in the lateral ventricle at postnatal day 0, with a volume of 2 μL per pup. Viral titers were determined by qPCR and adjusted to ensure equivalent dosing between constructs. After 14 days, brains were collected, washed with ice-cold DPBS, and green fluorescently labelled cortical tissue was harvested for RNA extraction with TRIzol™ Reagent (Invitrogen, 15,596,026) following the instructions of the manufacturer. Total RNA was treated with 0.2U of DNaseI (Vazyme Biotech, EN401) per 1 μg of total RNA, and subsequently processed with phenol:chloroform:isoamyl (125:24:1, pH < 5.0) followed by isopropanol precipitation.

Transcriptional elongation speed measurement

To estimate the average transcriptional elongation speed (RNA polymerase II speed), we quantified the distribution of read coverage along introns using a methodology previously described⁸⁴. In brief, intron coordinates were extracted from the GRCm39 gene annotation (GTF) using bedtools, and overlapping introns were merged to avoid redundancy. RNA sequencing reads were aligned to the GRCm39 reference genome using the STAR aligner⁶⁶ with the following parameters: chimSegmentMin:15; outSJfilterOverhangMin:15; alignSJoverhangMin:15; alignSJDBoverhangMin:15; seedSearchStartLmax:30; outFilterMultimapNmax:10; outFilterScoreMin:1; outFilterMatchNmin:1; outFilterMismatchNmax:2; chimScoreMin:15; chimScoreSeparation:10; chimJunctionOverhangMin:15; outWigType bedGraph; outWigNorm: None. Read coverage across introns was computed using bedtools, and coverage slopes were obtained with the scripts provided by the Beyer group⁸⁴ (https://github.com/beyergroup/ElongationRate/). Only introns supported by at least five split reads spanning both splice junctions were considered. Positive slopes (indicative of read accumulation toward the 3′ end) were excluded as non-informative. The transcriptional elongation speed for each intron was then calculated as speed = –1/slope. This analysis was applied to both nascent RNA sequencing and mature mRNA sequencing datasets.

Animal euthanasia

The employed mice for experiments were euthanized by cervical dislocation. For cortical primary neuron cultures, postnatal day 0 (P0) rats were euthanized by rapid decapitation. These methods were selected to minimize post-mortem changes in brain tissue, which is critical for downstream molecular analyses. All procedures were performed by trained personnel, in accordance with the AVMA Guidelines for the Euthanasia of Animals: 2020 Edition. All animal protocols were approved by the Institutional Animal Care and Use Committee (IACUC) at CIBR and complied with CIBR’s institutional ethical standards. All animal experiments were conducted in accordance with the ARRIVE 2.0 guidelines.

Statistics, reproducibility and plotting

Experiments were carried out in 3 biological replicates (n = 3), with each having 3 technical replicates. If mice were used in experiments, which means 3 different mice. Results were presented as mean ± standard deviation (± SD) of the mean. The two-tailed t-test was used to test statistical significance, with not significant P ≥ 0.05, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Statistical significance was determined without correction for multiple comparisons, with alpha = 0.05. Statistical analyses were performed using GraphPad (version 8.4).

Figures in this paper were generated using SRplot, Hiplot^85,86 and Graphpad (version 8.4). GO Enrichment Analyses were carried out using the GO Consortium site (https://geneontology.org/)⁸⁷. Inkscape was used for illustrations.

Author contributions

E.L. designed and performed the experiments, developed ideas, analyzed the data, generated figures, supervised all experiments, assembled all data and wrote the manuscript. M.T.A. performed bioinformatic analyses, critically evaluated data, generated figures and reviewed the manuscript. Y.P. advised on and performed all LC–MS/MS experiments and analyses. Y.Z. assisted with all animal experiments. S.F. performed some experiments, critically evaluated data and reviewed the manuscript. K.L. assisted with some experiments. X.L. facilitated and supported experiments. T.N. and Y.K. provided the Adar1/2KO mice and materials as well as valuable animal guidance and discussions. M.J.K. conceived the study and designed experiments, developed ideas, supervised all research, wrote and reviewed the paper.

Funding

This project was supported by the Beijing Natural Science Foundation (IS23091), the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (2019-I2M-5–015) and by CIBR’s core grant. M.T.A. contribution towards this work was funded by the Koziol lab. E.L. and Y.P. were supported by the Beijing Postdoctoral Research Foundation Fellowship (2020-YJ-004 and 2020-YJ-002, respectively) and by the International Postdoctoral Exchange Fellowship Program.

Data availability

All raw RNA sequencing data has been deposited in the NCBI GEO (Gene Expression Omnibus) database under the ID GSE275035. All raw Mass Spectrometry data has been deposited in Figshare: 10.6084/m9.figshare.27014107. The Adar2^KO mice (originally known as Adar2^−/− Gria2^R/R)⁶⁴ can be obtained from the MMRRC. With MMRRC’s permission, the Adar1/2^KO mice (previously referred to as Adar1^E861A/E861A Ifih^−/− Adar2^−/− Gria2^R/R)²³ can be obtained from Yukio Kawahara pending scientific review through an MTA or LOP (ykawahara@rna.med.osaka-u.ac.jp). All other data needed to evaluate the conclusions in the paper are present in the paper, and the Supplementary Materials or a referenced accordingly.

Code availability

All codes used have been deposited in the GitHub repository and are publicly available and accessible under the following link: https://github.com/KoziolLaboratory/AI-edit-circRNA. All other codes used are publicly available and have been referenced.

Declarations

Competing interests

M.J.K. is an Associate Editor for Oxford Open Neuroscience. The other authors declare no competing interests.