Pan-RNA editing analysis of the bovine genome

ABSTRACT

RNA editing is an essential process for modifying nucleotides at specific RNA sites during post-transcription in many species. However, its genomic landscape and characters have not been systematically explored in the bovine genome. In the present study, we characterized global RNA editing profiles from 50 samples of cattle and revealed a range of RNA editing profiles in different tissues. Most editing sites were significantly enriched in specific BovB-derived SINEs, especially the dispersed Bov-tAs, which likely forms dsRNA structures similar to the primate-specific Alu elements. Interestingly, ADARB1 (ADAR2) was observed to be predominant in determining global editing in the bovine genome. Common RNA editing sites among similar tissues were associated with tissue-specific biological functions. Taken together, the wide distribution of RNA editing sites and their tissue-specific characters implied the bovine RNA editome should be further explored.

Introduction

RNA editing, a widespread post-transcriptional modification process that confers specific nucleotide changes in RNA transcripts, not only expands the number of functions encoded by genomes but also provides intricate mechanisms of gene regulation [1,2]. Editing mRNAs effectively alters the amino acid sequence of an encoded protein, resulting in missense codon changes [3], alternative splicing, or regulatory RNA [4–6], and modifying miRNA seed sequences or their targeting sites [7,8]. In the post transcription, the most common RNA editing mechanisms comprise adenosine (A) to inosine (I) (A-to-I) deamination, which is catalysed by a family of adenosine deaminases that act on RNA (ADARs) [1]. ADAR enzymes can bind to double-stranded RNAs (dsRNAs) through their RNA-binding domains and deaminate A-to-I. Inosine is recognized as guanosine (G) by the cell translational apparatus. Previous research has found that most of the A-to-I editing sites in the human transcriptome are clustered within primate-specific Alu elements, which is attributed to their capacity to induce dsRNAs [9,10]. The primate ADAR gene family includes three members: ADAR1, ADAR2, and ADAR3. ADAR1 and ADAR2 are the primary editors of repetitive sites and non-repetitive coding sites, respectively, while ADAR3 is expressed primarily in the brain as an editing inhibitor [11,12]. In previous validation experiments, phenotypic changes induced by ADAR1 and ADAR2 confirmed their roles in mice [13,14]. Numerous previous studies have reported RNA editing in genes associated with cancer-relevant activities [15–17], embryogenesis [13], modification of regulatory RNAs [18,19], and brain development or neurological disorders [20–22]. Despite the apparent widespread importance of these mechanisms, RNA editing functionality in the bovine genome has yet to study. Furthermore, although three ADAR genes, ADAR, ADARB1 (ADAR2), and ADARB2 (ADAR3), are present in the bovine genome, their roles in the A-to-I editing process are not fully understood.

With the advent of next-generation sequencing, RNA-seq data have been accumulated, drawing attention to the analysis of RNA editing. However, most previous research on cattle focused on sporadic RNA editing events. The critical issue in those studies concerned the large number of A-to-I editing sites being clustered together due to promiscuous simultaneous editing of multiple adenosines by ADAR proteins [23–25]. Consequently, reads with editing clusters are not easily aligned to the genome. To overcome this problem, we adopted a clustering strategy that was developed to efficiently detect continuous RNA editing sites without genome sequencing data from the same sample [26,27].

Hitherto, several million RNA editing sites have been confidently detected in the human genome [26,28–32]. In contrast, few RNA editing sites have been reported in the bovine genome [2,33–35], and RNA editing functionality has not been systematically characterized. Thus, the specificity of RNA editing in different tissues, and its association with ADAR genes, remain unclear. Unlike primates, bovine does not have Alu elements; the detailed distribution of RNA editing in the bovine genome is uncovered. To address some of these gaps in our understanding, we detected RNA editing sites in the dairy cattle genome based on RNA-seq data. Notably, discoveries of RNA editing sites were further validated using new samples from dairy cattle by Sanger sequencing.

Results

Identification of hyper-editing events in bovine

To investigate the RNA editing landscape and its variability across different bovine tissues, we analysed 50 RNA-seq samples collected from 26 kinds of tissue in three datasets, which comprised 3.04 billion clean reads after quality control (Table S1). For de novo identification of RNA editing sites, we analysed only the RNA-seq data without a priori knowledge of genomic data from the same individual (see Methods). Briefly, we used the Burrows-Wheeler Aligner (BWA) package to align reads to the bovine genome, resulting in 461.7 million unmapped reads. Then we transformed all of the As to Gs (or other 11 types of editing events) in both the unmapped reads and reference genome. Finally, we mapped these transformed reads onto the transformed reference genome. A total of 4,077,440 hyper-edited reads were detected containing 1,117,831 RNA editing sites at 125,510 cluster regions (Table 1 and S2). We defined the cluster regions as part of the edited read starting at the first A-to-G mismatch and ending at the last one with a distance of ≤50 bp. The number of RNA editing sites detected in the present study far exceeded those bovine editing sites identified in previous studies (2,002 [34], 673 [35], 795 [33], and 200,770 [2]). As expected, most of the previously identified editing sites (77.8% of the 2,002 sites in [33], 85.2% of the 673 sites in [34], and 67.2% of the 795 sites in [32]) were also detected in our study, indicating that our method is highly accurate and efficient (Fig. 1A). The majority of the sites identified in our study (99.78%) were novel RNA editing sites.

Table 1.

RNA editing number and A-to-G rates

	Sites		Clusters
	Total number	A > G rate	Total number	A > G rate
All	1,117,831	92.29%	125,510	89.31%
Repetitive Sites	986,451	97.72%	109,965	97.08%
BovB-derived SINEs	857,041	99.41%	95,752	98.93%
Bov-tA family	567,537	99.64%	64,904	99.34%
Nonrepetitive sites	131,383	51.52%	15,545	34.35%

Figure 1.

The RNA editome of multiple tissues. (A) Our RNA editing results compared with previous studies. (B) The A-to-G rate, the ratio of G-to-A to A-to-G mismatches, and the accuracy of detected editing sites in different tissues. (C) The chromatogram of predicted ED3 edited region by Sanger sequencing

To validate the newly identified editing sites, eight editing cluster regions from three Chinese Holstein dairy cattle, including five repetitive cluster regions and three non-repetitive cluster regions (also located in protein coding regions) were randomly selected for Sanger sequencing, which containing 228 predicted A-to-G editing sites. Of the 27 sequenced editing regions, 23 (85.2%) exhibited a clear editing signal (Table S3, Figure S1). Our results confirmed 62 RNA editing sites in these editing regions (Table S3). For example, we successfully validated 12 sites in the ED3 editing region in the spleen tissue of cattle III (Fig. 1B). More results are shown in Figure S1. Meanwhile, 18 sites that were not predicted in the RNA-seq data were detected by Sanger sequencing, which may be attributed to individual variations or the stringent filtering criteria of our protocol.

Accuracy of identified RNA editing sites

To evaluate the accuracy of the strategy, we further used three measures to analyse the newly discovered RNA editing sites. Firstly, we calculated the proportion of mismatches attributed to A-to-G editing as 92.29% of the unique RNA editing sites (Table 1, Fig. 1C, Table S4). Then we calculated the ratio of G-to-A to A-to-G mismatches as 6.11 × $10^{-3}$ overall (4.51 × $10^{-4}$ in the cerebellum to 0.033 in the spleen; Fig. 1C, Table S4). Finally, single nucleotide variants in whole-genome sequencing data from the same animal were utilized to evaluate the accuracy of editing sites in 18 tissue datasets (see Methods). As expected, our protocol estimated that our results were highly accurate (98.0–98.8%) across the different tissues tested (Fig. 1C, Table S4).

The characters of RNA editing sites and clusters

The length of the RNA editing cluster regions and the average number of edited sites in each region were 64.9 and 8.9, respectively (Fig.s 2A and S2A, B). We found that most of the editing clusters contained ~4–7 editing sites (55.2%). The average ratio of the edited adenosines to all adenosines in the A-to-G editing cluster was 0.47 (Figure S2C). We checked the flank sequence contexts of the A-to-G editing sites and found that Gs were depleted one base upstream and enriched one base downstream of the editing sites, which agreed with the ADAR preference target sequences. These characters are prevalent in other animals (Fig. 2B) [2,36,37]. For non-repetitive editing sites, we found that Gs were depleted one base upstream of these editing sites (Figure S2D). Overall, we found that the conservation of editing sites was low, being nearly the same as that of randomly selected regions (Figure S2E). However, editing sites presented in multiple tissues were more likely to be evolutionarily conserved, especially those editing sites detected in at least ten tissues, which were conserved at a rate much closer to that of exon regions (Fig. 2C). Interestingly, the editing sites themselves were less conserved than their neighbouring bases (Fig. 2C).

Figure 2.

The characters of RNA editing sites and clusters. (A) The length of the cluster and the number of contained sites in different clusters. (B) RNA editing motif. The sequence near editing sites was depleted of Gs upstream and enriched with Gs downstream as expected from ADAR targets. (C) Mean conservation score of RNA editing sites. Position 0 indicates the RNA editing site. Green lines represent the conversation score of RNA editing detected by the different number of tissues. Purple and red lines represent background and exon conservation, respectively. (D) The conserved editing sites across three species

Comparative analysis of RNA editing among species

To evaluate the conservation of RNA editing sites between bovine and other species, we analysed ~5.95 million and 9,097 RNA editing sites for human and mouse, respectively, from DARNED, RADAR, and REDIportal databases. Using the UCSC LiftOver tool, 2,249,960 human and 834 mouse editing sites were detected in the bovine genome (Table S5). We found 920 editing sites in 429 cluster regions that were conserved between human and bovine genomes (Fig. 2D, Table S6A). Of these, 330 sites were detected in multiple tissues. Furthermore, 21 RNA editing sites in nine cluster regions were conserved between mouse and bovine genomes, including 13 sites that were detected in multiple tissues (Fig. 2D, Table S6B). The editing sites detected in multiple tissues were found to be significantly more highly conserved compared with those detected in one tissue (human, P < 9.55 × $10^{-16}$, mouse, P < 2.27 × $10^{-4}$). This result was consistent with that from the last conservation analysis by the PhastCon score.

We found that 13 RNA editing sites in six clusters were conserved among all three species. Interestingly, genes including BLCAP, NNAT, TMUB1, ENSBTAG00000045928, FLNB, FASTK, IGFBP7, SLC4A2, and FLNA could be edited through these conserved editing sites. Of these, two sites in BLCAP and one in IGFBP7 may lead to missense mutations (Table 2). For the 928 editing sites that were conserved either between bovine and human or bovine and mouse genomes, we found 388 genes that could be edited. Based on functional annotation, these 388 genes were associated with a wide range of critical pathways, such as calcium signalling, neuron-specific functions, D-myoinositol (1, 4, 5)-trisphosphate, and salvage pathways, implying that these conserved editing sites function in crucial biological processes (Table S7).

Table 2.

Evolutionarily conserved sites among human, mouse, and bovine genomes

Chromosome	Position	Strand	Region	Gene	Amino acid change
Chr4	114454773	+	Downstream	SLC4A2	-
		-	Intron	FASTK	-
		-	Downstream	TMUB1	-
Chr4	114454790	+	Downstream	SLC4A2	-
		-	Intron	FASTK	-
		-	Downstream	TMUB1	-
Chr4	114454791	+	Downstream	SLC4A2	-
		-	Intron	FASTK	-
		-	Intron	TMUB1	-
Chr6	74150139	-	CDS	IGFBP7	M/T
Chr13	67116439	-	CDS	BLCAP	L/P
		+	Upstream	NNAT
Chr13	67116448	-	CDS	BLCAP	F/S
Chr13	81802824		Intergenic
Chr13	81802829		Intergenic
Chr13	81802837		Intergenic
Chr13	81802852		Intergenic
Chr22	43688232	-	Intron	FLNB	-
ChrX	40313486	-	Intron	FLNA	-
ChrX	40313487	-	Intron	FLNA	-

The locations of RNA editing site in the bovine genome

To ascertain the distribution of the most frequent RNA editing events, we checked the distribution of RNA editing sites over the whole bovine genome. As shown in Fig. 3A, the editing sites were mainly located in intergenic regions and introns further downstream/upstream regions. However, the editing sites in 5′ UTR/3′ UTR and coding regions were relatively limited. Furthermore, we detected 25,475 (2.28%) RNA editing sites in the protein-coding areas overlapped with 2,603 genes. Among those editing sites, 15,897 (1.42%) were associated with missense variants and may modify 2,412 gene structures. After normalization based on length bias, RNA editing sites were highly enriched at 3′ UTR regions (Fig. 3B). Since the edited 3′ UTRs may alter gene expression by changing the binding sites of miRNA target sequences [7,8], we exhaustively checked the 14,700 editing sites that were located in miRNA targets, which could prevent miRNAs or other molecules from binding (Table S8A). We observed 40 editing sites occurring in nine known miRNAs, including 12 sites in three mature miRNA regions (miR-631, miR-2417, and miR-2432) as listed in Tables S8B, C.

Figure 3.

The RNA editing occurrences in the repeat region. (A) The distribution of RNA editing sites throughout the genome. (B) The enrichment of RNA editing numbers in different genome regions. Green colour represents that RNA editing sites were depleted. The red colour represents that RNA editing sites were enriched. (C) The circle packing plot of repeats in the bovine genome. The size of each circle is proportional to the number of editing sites in each repeat type. The Bov-tA family contains Bov-tA1, Bov-tA2, and Bov-tA3. The Bov-tA family and BOV-A2, ART2A belong to BovB-derived SINEs. (D) The dsRNA rate, the number of RNA editing clusters, and their enrichment factors of different kinds of repeat elements

Relations between RNA editing and BovB-derived SINE elements

Interestingly, RNA editing sites mainly reside within repetitive elements, such as the primate-specific Alu elements [38]. Similarly, we found that the vast majority of RNA editing sites (88.25%) and cluster regions (87.61%) resided in repetitive element regions (Table 1). Of which, 97.73% of editing sites and 97.08% of cluster regions were the A-to-G editing type, which agrees with the results of previous studies [9,39,40]. In the non-repetitive areas, the proportion of A-to-G editing sites and cluster regions was 51.52% and 34.34%, respectively. We also found that these non-repetitive editing sites contained more changes, which were closer to those of repeat elements compared with those of randomly selected regions, especially for the non-repetitive A-to-G editing sites (Figure S3). In the human genome, nearly all A-to-I editing sites are exclusively clustered in Alu elements [26]. In contrast, we found that most RNA editing sites (857,041, 79.46%) and cluster regions (95,702, 77.23%) in the bovine genome resided in BovB and BovB-derived SINEs (containing Bov-tA, BOV-A2, and ART2A) with a higher A-to-G rate (editing sites, 99.41%; cluster regions, 98.93%; Fig. 3D).

Subsequently, we proposed an enrichment factor to normalize their length bias in the genome. Interestingly, we found the RNA editing events were highly enriched in the Bov-tA repeat family, especially for Bov-tA2, in which the enrichment factor for sites and cluster regions reached 20.64 and 21.04, respectively (Fig. 3C, Table 3). To ascertain their distinct properties, we further examined their complementary RNA strands within 2 kb up- or downstream of the edited sites. Interestingly, we found that 55.1% of the A-to-G editing in repetitive regions contained hairpin structures, which was higher than that of other RNA editing types (23.17%), non-repeat editing clusters (17.3%), and randomly selected non-editing sequences (5.7%) as shown in Fig. 3C. It has been reported that editing sites are often located in double-stranded structures, and ADAR binding and enzymatic activities are specific to these regions [41,42]. In the Bov-tA repeat regions, 62.8% of the hyper-editing clusters can form a dsRNA structure, indicating that the Bov-tA repeats harboured large amounts of dsRNA for combining with ADAR proteins (Table 3). An example of a dsRNA structure predicted in a Bov-tA region in our study is shown in Figure S4. The two repetitive elements were complementary and formed a dsRNA structure, which harboured 37 RNA editing sites. The large proportion of dsRNA distributed in BovB and BovB-derived SINEs may explain why RNA editing events are more likely to occur in these regions.

Table 3.

RNA editing enriched in repetitive regions

Category	Proportion of repeats in genome	Number of editing sites	Enrichment factors	DsRNA rate
Bov-tA2	0.57%	352,772	20.64	62.7%
Bov-tA1	1.53%	167,295	4.89	62.2%
ART2A	3.06%	175,223	2.39	44.0%
BOV-A2	6.56%	50,198	2.09	72.9%
Bov-tA3	2.15%	47,501	7.43	65.4%
BovB	7.29%	70,354	0.86	26.5%
L1-2_BT	1.38%	19,025	1.23	21.9%
L1_BT	3.69%	15,218	0.37	21.8%
Other L1	7.91%	28,901	0.33	29.3%
Other repeats	15.97%	67,545	0.38	39.0%
No repeat	49.89%	131,383	0.24	5.7%

DsRNA, double strand RNA.

RNA editing number and level among tissues and organs

To evaluate variations in RNA editing among tissues within the same animal, we defined an enrichment factor, which is the number of hyper-edited RNAs divided by the expected number in the tissue. We found that brain tissue, such as the cerebellum and lobes, exhibited the highest number of enrichment factors for editing sites and cluster regions (Fig.s 4A and S5A–C). In contrast, muscle tissue, such as skeletal and heart, exhibited the lowest number of enrichment factors. Our results are consistent with those of previous studies on mammals [26]. Meanwhile, large numbers of RNA editing sites were also observed in other tissues such as bone marrow, anterior pituitary, uterine, and lung. Immune tissues exhibited moderate enrichment levels compared with other tissues. Interestingly, for some tissues, such as liver collected from lactating cows, we found a high abundance of enrichment factors but only a moderate number of editing sites (Figure S5B).

Figure 4.

RNA editing in multiple tissues. (A) The enrichment factors of editing sites in different tissues, which is based on the number of hyper-edited RNAs in the tissue divided by the expected number. (B) The PCA plot of 50 tissues. (C) The heatmap and dendrogram of editing level using k-mean (k = 6), which divided 50 tissues into six clusters. (D) The editing sites were divided into five groups based on the local sequence context near the editing sites. The points in the boxplot represent the 50 tissues. The size of the points indicates the enrichment factor in different tissues. The colour of points represents that they belong to different tissue clusters

To investigate whether there were shared patterns of RNA editing among different tissues, we applied a principal component analysis (PCA) and k-mean cluster analysis to 4,749 common editing sites in at least 20 tissues. As shown in Fig. 4, the PCA results suggested that the muscle, brain, digestive, and immune organs could be clearly distinguished from each other (Fig. 4B). The K-mean cluster analysis revealed that these 50 tissues could be divided into six distinct clusters (Fig. 4C). Notably, the editing profiles within each cluster were highly correlated. All types of brain tissue were clustered. Similar patterns were found between digestive and immune-associated organs, and between heart tissue and skeletal muscle. Although the overall level of all editing sites in each tissue was generally similar, the number of edited reads in muscle tissue was significantly lower than that of other tissues (P < 2.2 × $10^{-16}$, Wilcoxon rank-sum test; Figure S6). Twenty-eight RNA editing sites located in seven cluster regions were identified in all 50 tissues (Table S9). Additionally, we specifically explored tissue-specific editing sites and identified 4686 sites that were edited preferentially in only one tissue type (Figure S7).

To examine the ADAR binding motif of editing sites, we checked whether those sites with matched ADAR recognition exhibited distinct patterns. We found that RNA editing levels were significantly higher in sites with matched ADAR recognition compared with those with partially-matched or non-matched recognition motifs (P < 2.2 × $10^{-16}$, Wilcoxon rank-sum test). Notably, editing sites with a 5′ matched motif usually showed higher editing levels than sites with a 3′ matched motif. These results are consistent with those of a previous study on rhesus macaques (Fig. 4D) [37].

RNA editing and ADAR gene family

ADAR enzymes catalysing A-to-G (I) editing have been reported in humans and mice but remain poorly understood in bovines [1,11]. From our datasets, we found ADAR and ADARB1 to be dynamically expressed in 50 tissues (Fig. 5A). ADAR was relatively highly expressed in immune tissues, while ADARB1 expression was the highest in brain tissues. Both were lowly expressed in muscle and heart tissues (Fig. 5A). Subsequently, we assessed expression in the brain, spleen, heart, liver, longissimus muscle, colon, and kidney tissues from three Chinese Holstein cattle. As shown in Fig.s 5B and S8, the relative expression of these two genes detected by qRT-PCR was highly correlated with that from the RNA-seq data ($R^2$= 0.74 and 0.81, respectively). Besides, ADARB2 (ADAR3) was only expressed in brain tissues as confirmed by conventional PCR (Fig. 5C).

Figure 5.

The ADAR genes dynamically expressed among tissues. (A) The expression of ADAR, ADARB1, and ADARB2 in 50 tissues from RNA-seq data. (B) The relative expression of ADAR and ADARB1 detected by qRT-PCR in seven tissues (relative to the kidney). (C) The PCR product analysis for ADARB2 using conventional PCR. All PCR products were electrophoresed on 1.5% agarose gels. The samples from left to right are heart, liver, spleen, brain, kidney, muscle, and colon

Subsequently, we explored the association between editing variation and the expression of different ADAR enzymes in various tissues. Using standardized regression analysis, we found that ADARB1 expression accounted for ~38.3% of the variations in unique editing numbers, while ADAR expression only explained 7.5% (Fig. 6A). The standardized regression coefficients (beta values) showed the contribution of each ADAR family member to RNA editing numbers. We found that ADARB1 predominantly contributed to editing numbers (P < 9.17 × $10^{-6})$ (Table 4). Moreover, ADARB2 expression, exclusively found in brain tissue, was positively correlated with the editing number of brain tissue (P < 0.084). Next, we checked whether the ADAR gene family could affect the global RNA editing level in different tissues. We found that ADARB1 had a significant $R^2$, indicating that ADARB1 expression may be a significant determining factor for global editing levels (P < 0.022). The expressions of three ADAR genes were more positively correlated with A-to-G editing sites than other types of editing sites (Fig. 6B, ADAR, P < 4.81 × ${10}^{-5}$; ADARB1, P < 0.027; ADARB2, P < 0.14). We also conducted a regression analysis between the expression of ADAR enzymes and each editing site level. In which, 273 sites were significantly correlated with ADARB1, 33 sites with ADARB2, and one site with ADAR (Table S10).

Figure 6.

The ADAR gene family is affecting RNA editing events. (A) Correlations between the expression level of ADAR family genes and editing number (corrected by sequence depth, the top three figures). Relationships between the expression level of ADAR family genes and overall editing level (the bottom three figures). (B) The distribution of correlation between the expression level of ADAR family genes and the editing level of each site. The editing sites were divided into two groups (A-to-G type and non-A-to-G type). (C) The RNA editing levels in different tissues

Table 4.

The ADAR gene family affecting RNA editing events

	RNA editing number			RNA editing level
	Adjusted $R^2$	Beta	P-value	Adjusted $R^2$	Beta	P-value	Significant sites
ADAR	0.075	2.20	0.052	−0.021	−0.048	0.734	1
ADARB1	0.383	13.79	9.17 × $10^{-6}$	0.091	0.342	0.022	273
ADARB2	0.103	1.73	0.084	−0.017	−0.022	0.88	33

RNA editing and biological functions

The massive RNA editing events in the bovine transcriptome raised questions regarding their roles in A-to-I changes. To explore this further, we assumed that the editing sites, specifically recurring in similar tissues were functionally related. We found 72, 31, 133, and 5 edited genes specifically occurring in all brain, immune, digestion, and muscle tissues, respectively (Table S11). To eliminate the effect of tissue-specifically expressed genes, we removed the edited genes that common with tissue-specifically expressed genes. We obtained 63, 24, 128, and 5 convincedly edited genes for the brain, immune, digestion, and muscle tissue, respectively. Functional enrichment analysis confirms that the specially edited genes in brains were involved in oxidative phosphorylation, metabolic pathways, peroxisome, and neurological diseases, such as Parkinson’s disease, Alzheimer’s disease, Huntington’s disease. The specially edited genes of immune organs were associated with phosphatidylinositol phosphate phosphatase activity, protein activation cascade, complement activation, and humoral immune response. For digestion tissues, the specifically edited genes were enriched in drug metabolism, arachidonic acid metabolism, and metabolic pathways, such as fatty acid metabolic process, unsaturated fatty acid metabolic process, and peptide biosynthetic process. The specially edited genes of muscle tissues were observed to be involved in oxidoreductase activity, cofactor binding, and copper ion binding. Interestingly, we found that 12 edited genes shared by all four tissue types. Their functions were associated with apoptosis, cytokine-cytokine receptor interaction, and phospholipase activity (Table S12).

Discussion

Although RNA editing has been systematically studied in many mammals, accurate definition of the editome by using RNA-seq data alone remains technically challenging [43–45]. To our knowledge, this is the first study of RNA editing patterns across different bovine tissues to characterize A-to-G editing sites and clusters. We identified 1,117,831 high-confidence candidate editing sites at 125,510 RNA editing cluster regions across diverse bovine tissues with high A-to-G rates (>92%), high accuracy (>98%), and low G-to-A/A-to-G ratios (<3.3%). Some of the editing sites we detected have been reported in previous studies; however, most of them are newly discovered.

Large-scale sequencing data on a broad range of tissue samples from the same or different animals allowed for a comparative editome analysis. We found that >92% of RNA editing sites were A-to-G type, meaning that RNA editing by adenosine deamination was conspicuous in bovine transcriptome. Although most editing sites were not highly conserved, those detected in multiple tissues were. The editing sites themselves were less conserved than their neighbouring bases, which is consistent with previous findings [46,47]. Only a small proportion of RNA editing sites was conserved between the bovine and human or mouse genomes, which is similar to the situation between human and mouse genomes [48]. This may not be surprising given that human coding RNA editing is generally non-adaptive [49]. Interestingly, genes harbouring conserved A-to-G editing sites were enriched for metabolic functions. In our study, 13 RNA editing sites were conserved among all three species, including distinct edited genes, such as FLNA, FLNB, BLCAP, and IGFBP7, which have been reported in human and mouse genomes [48,50]. We also found that these genes could be edited at the transcription level in bovines. These results further indicate essential functions and common metabolic activities among mammals. These highly conserved sites are considered to function similarly.

RNA editing sites are primarily located in non-coding regions [2,,^35,36]. Similarly, we found a large proportion of non-coding editing sites in the bovine genome. RNA editing events in protein-coding regions are rare, which has been reported in previous bovine studies [33,,³⁵]. However, our results based on multiple datasets revealed that 2,412 genes could be edited at 15,897 sites at the translation level. Besides, our work intensively delves into the often-ignored non-coding RNA editing events. We found that non-coding RNA editing sites were highly enriched in 3′ UTR, highlighting the importance of future research to determine whether these 3′ UTR editing sites could be associated with miRNA binding sequences. Tens of thousands of editing sites are located in 3′ UTR regions containing miRNA binding sites, which may prevent miRNAs or other molecules from binding. Therefore, further research should also ascertain whether 3′ UTR editing causes gene expression changes.

The vast majority of RNA editing sites in mammals reside in repetitive elements [2,39,51,52]. These regions potentially hybridize with nearby oppositely oriented repeats, forming the dsRNA structures required for ADAR binding. This is the reason for most of the A-to-G editing sites being clustered within the primate-specific Alu elements [9,10]. Similarly, we found that editing sites and cluster regions in the bovine genome mainly resided in repetitive elements, especially BovB-derived SINEs. These retrotransposons are prevalent in all ruminants [53,54]. Further study revealed that the BovB-derived SINE regions had more opportunities to form dsRNA compared with other regions, which may explain the relatively high number of RNA editing sites in BovB-derived SINEs. Besides, we noted that the proportion of A-to-G transitions increased from 93.39% to 98.83% when focusing on BovB-derived SINEs, providing further evidence to indicate that the ADAR enzymes preferentially recognize these BovB-derived SINEs.

The number of unique RNA editing sites dynamically changed across tissues. There were more editing sites in brain tissues and fewer in muscle tissues, which is consistent with previous studies on other species [11,26,55]. Although some tissues have a small number of unique editing sites, their edited reads were abundant. Besides, we found that global editing activity exhibited organ-preferential clusters. The brain, muscle, immune, and digestion tissues could be separated by their common editing profiles, implying that the specificity of pan-RNA editing profiles may be used as unique biomarkers to distinguish tissues and help us trace the inosinome fingerprint to characterize A-to-I events in biological and physiological phenotypes.

We successfully validated ADAR gene expression using samples from three additional cattle. ADAR was highly expressed in immune tissues, while ADARB1 expression was the highest in brain tissues. The ADARB2 gene was exclusively expressed in brain tissues. Similar results have been reported for humans, rhesus macaques, and mice [36,37,55]. ADAR enzymes may be related to RNA editing activities [11,56]. To determine if this was the case in bovine, RNA editing was evaluated based on the overall level and the number of RNA editing sites. All three ADAR genes were associated with RNA editing numbers, especially ADARB1. Regression analysis between ADAR genes and RNA editing has been previously reported in humans [11]. Only ADARB1 was related to the overall level of RNA editing, which may be attributed to the limited number of samples compared with the 8,551 samples used in the human study [11]. Compared with ADAR and ADARB2, ADARB1 expression was more significantly correlated with both editing number and overall level, implying that ADARB1 may be responsible for A-to-I editing. ADARB2 expression was negatively correlated with overall levels of RNA editing in humans [11], whereas we found that ADARB2 expression was positively correlated with RNA editing numbers in bovine. Both ADARB1 and ADARB2 were highly expressed in brain tissues, which may explain the relatively high editing numbers compared with other tissues.

The specific editing sites recurring in different tissues may have essential biological functions. Aberrant editing is correlated with several various human disorders, most of which are neurological [21,38,57]. In our study, there were shared RNA editing sites within similar tissue categories. We found that the specially edited genes in different tissue categories were associated with tissue-specific functions and aetiologies. For example, the specific editing sites of brain tissues were related to neurological disorders, such as Alzheimer’s disease, Huntington’s disease and Parkinson’s disease [58]. The specific editing sites of immune tissues were associated with phosphatidylinositol phosphate phosphatase activity, complement activation, and humoral immune response.

In summary, our study provides a comprehensive analysis of the landscape of RNA editing in different bovine tissues, highlighting editome variations and characters. We further ascertained that RNA-editing sites are preferentially located in BovB-derived SINEs and explored the relationship between RNA-editing and ADAR genes. The conservation of editing sites in different species suggests that RNA editing could play essential roles in the regulation of organismal development, metabolism, and speciation. However, their biological functions and regulatory mechanisms need to be explored further. Given the broad distributions of RNA editing sites and that most of those discoveries have not been previously reported, future research may lead to their use as biomarkers. Most importantly, our results provide valuable resources for further investigation of bovine RNA editing with potential applications in bovine genetics and breeding, health, and animal welfare.

Materials and Methods

Ethics statement

All procedures pertaining to the handling of experimental animals were conducted in accordance with and approved by the Animal Welfare Committee of China Agricultural University (Permit Number: DK996). All efforts were made to minimize discomfort and suffering.

Datasets

The 50 RNA-seq samples were collected from 26 tissue types from three datasets containing 9, 23, and 18 tissues, respectively (Table S1). Briefly, the 9-tissue dataset consisted of ~8.74 × ${10}^8$ reads generated from nine tissues from one Holstein bull. Each tissue was sequenced at both 80-bp PE reads using strand-specific libraries [59]. The 23-tissue dataset included ~1.14 × ${10}^9$ reads with 75-bp PE from the newly assembled bovine reference genome annotation project. The 18-tissue dataset contained ~1.03 × ${10}^9$ reads with 100-bp PE from the Functional Annotation of Animal Genomes (FAANG) project [60,61].

Identification of RNA editing sites

To identify RNA editing sites, we adopted a clustering strategy to align and examine the unmapped reads carefully [26]. Firstly, we collected all unmapped reads from the initial alignment using BWA ‘aln’ and ‘mem’ modules [62]. Secondly, we transformed all As to Gs in both the unmapped reads and the reference genome and realigned the transformed RNA reads to the transformed reference genome using BWA ‘aln’. We obtained the mapped reads and recovered them to the original sequences, which were treated as candidate RNA editing reads. To improve the accuracy of identifying RNA editing clusters, we used the number of A-to-G mismatches that occupied at least 5% of the read length (at least three A-to-G mismatches for the read length ≤60 bp) and >80% of the total number of mismatches as cut off. To avoid false positives caused by technical artefacts, further filters were applied: (1) required average Phred quality score >25; (2) removed reads with >10% ambivalent nucleotides, >10 simple repeats, or >20 for a successively single nucleotide. Then RNA editing sites overlapped with the SNP database (dbSNP), and splicing regions from Ensemble annotation (version 93) were removed. We repeated this procedure for searching the other 11 types of editing events (e.g., A-to-C and G-to-A). It is worth noting that the 18-tissue dataset was not stranded, and thus, the edited sites may appear as T-to-C mismatches. Based on the other two datasets, T-to-C editing seldomly occurred in bovines. The expected error rate (T-to-C editing rate) was 0.014. We treated all T-to-C sites as A-to-G sites in the 18-tissue dataset. The cluster regions of RNA editing were defined as the part of the edited read starting at the first A-to-G mismatch and ending at the last one with a distance ≤50 bp. The pipeline scripts are available at https://github.com/RNA-editing/Scripts.

Accuracy of identified RNA editing sites

The A-to-G rate was defined as the proportion of mismatches attributed to A-to-G editing because non-A-to-G editing events are infrequent [26,63]. These previous studies assumed that all G-to-A differences reflect sequencing errors. Thus, the ratio of G-to-A mismatches to A-to-G mismatches was used to evaluate the false positive rate in our study [47,64]. To determine the accuracy of our detected editing sites, we first aligned the 18-tissue DNA-seq and RNA-seq datasets to the bovine genome using the default parameters of BWA. Alignments were improved by Picard (http://picard.sourceforge.net) and GATK tools [65]. Single nucleotides variant (SNV) calling was performed using REDItool DnaRna module with -v 1 -n 0.01 -N 0.01 -c 1,1 [66]. To determine the accuracy of our clustering strategy, we focused on the detected SNVs by the clustering strategy. Then we estimated accuracy as follows: accuracy = (number of DNA–RNA SNVs)/total number of SNVs.

Analysing the conservation of A-to-I editing

To determine the conserved RNA editing sites between bovine and other species, we downloaded human and mouse RNA editing site data from DARNED [30], RADAR [32], and REDIportal [31] databases and generated ~5.95 million and ~9,097 RNA editing sites for humans and mice, respectively. Then we performed the UCSC LiftOver tool to convert the genome coordinates of human and mouse RNA editing sites into those of bovine. We obtained the mappable sites of other species using the following criteria, (1) human editing sites that could be mapped to the bovine genome; (2) the resulting sites could be mapped back from the bovine to the human genome; (3) the genome coordinates of the twice mapped sites should be the same as those of the original editing sites. The conserved editing sites were defined as the common sites between bovine editing sites and the mappable sites of other species. The PhastCon scores were downloaded from the UCSC genome browser. Pathway and functional enrichment analysis of edited genes were performed using IPA with log2(p-value) >1.3 as a cut-off.

Genome and repeat annotations


We annotated genomic regions using the Variant Effect Predictor for known genes on the ensemble website. Bovine miRNA coordinates were taken from miRBase v21. The coordinates of the conserved mammalian miRNA regulatory targets of conserved miRNA families in the 3′UTR were predicted by miRanda with scores > 160 and energy < −15. The known repeats were annotated using the RepeatMasker tables, which were downloaded from the UCSC genome browser. To detect potential dsRNA structure formed by RNA editing, the sequences of the RNA editing regions were aligned to the sequences 2 kb up- and downstream of the regions. We used bl2seq with parameters -F -W 7 -r 2 -D 1 and considered a match only for alignment with 90% identity along 80% of the editing region length.

Computing editing levels

To calculate the editing levels of the identified sites, we aligned the RNA-seq datasets to the bovine genome by BWA with default parameters. Alignments were improved by Picard and GATK tools and analysed for editing levels using the REDItools [66] with parameters: -v 1 -n 0.01 -c 1 -T 6–6. Then we analysed editing level results to compute RNA editing further using A-to-I as an example as described below:

$$\begin{array}{l}A-to-Ieditinglevel\\ \text{}\text{}\text{}\text{}\text{}\text{}=\frac{NumberofG+NumberofhypereditedG}{NumberofAandG+NumberofhypereditedG}\end{array}$$ (1)

The level of editing was determined in a clustering manner as:

$$\begin{array}{rl}& A-to-Ieditinglevel\\ & =\frac{{\sum }_i\left(NumberofG+NumberofhypereditedG\right)}{{\sum }_i\left(NumberofAandG+NumberofhypereditedG\right)}\end{array}$$ (2)

with i = 1, …, N, where N is the total number of editing sites in each designated cluster region.

The global level of editing in each tissue was determined using formula 2 with i = 1, …, N, where N is the total number of editing sites in each tissue.

The editing levels were estimated according to RNA-Seq data, where estimations might be less accurate for sites with lower sequencing coverage. Therefore, we defined coverage ≥10 as the cut-off. To identify the major sources of variation in 50 tissue samples, we performed PCA and heatmap. We removed all sites with missing editing data in >30 tissues. The missing values of the remaining sites were imputed using the missForest R package [11,67]. Then the ‘prcomp’ function in R was used to determine the principal components of the complete dataset. The heatmap of different clusters was performed using Euclidean distance in the k-mean method with k = 6 [68]. Using gap statistics, we determined that k = 6 was the optimal choice for distinguishing tissue categories. To identified tissue-specific editing sites, we focused on sites in which the editing level can be detected in at least 5 tissues. We then applied the ROKU R package to rank the sites by their overall tissue specificity using Shannon entropy and detected tissues specific to each site. We required Shannon entropy lower than 0.4 to generate a list of tissue-specific editing sites [69]. The ROKU R package was also conducted to identify tissue-specifically expressed genes. Pathway and functional enrichment analysis of tissue-specific edited genes in brain, immune, digestion, and muscle were performed using GOstats package with p-value < 0.05 as a cut-off [70].

ADAR, ADARB1 and ADARB2 expression

To estimate levels of gene expressions, all RNA-seq data were aligned to the bovine genome using STAR software with default parameters [71]. The expression of known genes (i.e., expected fragments per kilobase of transcript per million fragments mapped (FPKM)) was quantified using StringTie -e -B [72]. The enrichment factor of the editing number was equal to the number of RNA editing sites (reads or clusters) in each tissue divided by the expected number, which was calculated by multiplying the total number of RNA editing sites in all tissues by the ratio of mapped reads in each tissue to total mapped reads in all tissues [36].

To identify the correlations between the expression levels of ADAR family genes and the editing numbers or editing levels in different tissues, linear regression was applied to calculate R² values and beta coefficients on z-score normalized gene expression and z-score normalized enrichment factors or editing levels as follows:

$$Y=\mu +{\beta }_1{X}_1+{\beta }_2{X}_2+{\beta }_3{X}_3+\mathrm{e}$$ (3)

where Y is the z-score normalized enrichment factors of editing numbers or the editing levels, $\mu$ is the mean of the enrichment factor of editing numbers or editing levels, $X_1$, $X_2$, and$X_3$indicate the expression levels of ADAR, ADARB1, and ADARB2, respectively. ${\beta }_1$, ${\beta }_2$, and ${\beta }_3$ the corresponding regression coefficients, and e the normally distributed random error. ${\mathrm{R}}^2$ was used as a quantitative indicator of the variation in the editing level explained by the ADAR expression profile. Finally, according to normalized ${\beta }_1$, ${\beta }_2$, and ${\beta }_3$ values, we determined which ADAR was predominantly correlated with the editing number or editing level. The Spearman’s rank correlation method was used to calculate the correlations between the expressions of ADAR family genes and RNA editing level of each site. Statistical analysis was performed using the R package.

Quantitative real-time PCR (qRT-PCR) and RNA editing validation

Three multiparous and healthy Chinese Holstein cows in the lactating period were selected from the Beijing Sanyuan Lvhe Dairy Farm. The cows were killed by electroshock, bled, skinned, and dismembered in the same slaughterhouse. The seven tissue samples, including the brain cortex, heart, liver, spleen, longissimus muscle, colon, and kidney, were collected within 30 min after slaughter. Total RNA was extracted from each tissue sample using TRIzol (Invitrogen, Carlsbad, CA, USA) according to the manufacturer’s protocols. RNA samples were reverse transcribed to cDNA using the PrimeScript RT reagent Kit with gDNA Eraser (TaKaRa) according to the manufacturer’s instructions. Primers for qRT-PCR were designed using Primer 5.0, and qRT-PCR was run in triplicate using the LightCycler® 480 SYBR Green I Master Kit (Roche). The results were normalized to GAPDH expression to obtain ΔCt values. The relative expression of ADAR, ADARB1, and ADARB2 was calculated using the $2^{-\mathrm{\Delta }\mathrm{\Delta }Ct}$ method. The resulting cDNA and genomic DNA from the same samples were amplified by PCR. PCR amplifications were performed in a reaction volume of 20 μL comprising 2 μL of 50 ng/μL gDNA (cDNA), 1 μL of each primer, 10 μL of premix (containing dNTPs and DNA polymerase) (Tiangen, Beijing, China), and 6 μL of ddH2O. The amplification procedures were as follows: 10 min at 95°C for initial denaturing; followed by 35 cycles at 95°C for 30 s, 60°C for 30 s, and 72°C for 30 s; and a final extension at 72°C for 10 min. Amplified PCR products were cleaned of agarose gels and then subjected to Sanger sequencing. Compared with those observed for products amplified from the control (gDNA), the predicted A-to-G conversions exhibited clear signals with double peaks in the chromatograms at the predicted RNA editing sites for the products amplified from the corresponding cDNA, confirming the occurrence of RNA editing at these sites. The primer sequences used for qPCR and Sanger validation are listed in Table S13.

Funding Statement

The Jorgensen Fund supported this study. Transcriptome and Angus beef quality, 2015; Maryland Agriculture Experimental Station (MAES), 2014.

Availability of data

All sequencing data used in this study are publicly available. RNA-seq reads for this analysis were obtained from the NCBI short read archive (PRJNA379574, PRJEB25677, and PRJNA177791). The bovine editing data is available at: https://github.com/RNA-editing/Bovine.

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary material

Supplemental data for this article can be accessed here.