Circular RNA : an emerging key player in RNA world

Insights into the circular RNA (circRNA) exploration have revealed that they are abundant in eukaryotic transcriptomes. Diverse genomic regions can generate different types of RNA circles, implying their diversity. Covalently closed loop structures elevate the stability of this new type of noncoding RNA. High-throughput sequencing analyses suggest that circRNAs exhibit tissueand developmental-specific expression, indicating that they may play crucial roles in multiple cellular processes. Strikingly, several circRNAs could function as microRNA sponges and regulate gene transcription, highlighting a new class of important regulators. Here, we review the recent advances in knowledge of endogenous circRNA biogenesis, properties and functions. We further discuss the current findings about circRNAs in human diseases. In plants, the roles of circRNAs remain a mystery. Online resources and bioinformatics identification of circRNAs are essential for the analysis of circRNA biology, although different strategies yield divergent results. The understanding of circRNA functions remains limited; however, circRNAs are enriching the RNA world, acting as an emerging key player.


Introduction
Thousands of endogenous circular RNAs (circRNAs) have recently been identified and characterized in eukaryotic cells [1][2][3][4][5][6].Following microRNAs (miRNAs) and long noncoding RNAs (lncRNAs), circRNAs are enriching the RNA world.circRNAs are a distinct class of endogenous noncoding RNAs that result from a noncanonical form of alternative splicing [7,8].Although first discovered in the early 1990s as transcripts with scrambled exon order [9][10][11], circRNAs have been previously dismissed as transcriptional noise in eukaryotes [12] until quite recently.The advent of next-generation sequencing technology coupled with bioinformatics approaches has brought them back again to the forefront and allowed a comprehensive exploration of these RNA circles [4,[13][14][15].
Unlike linear mRNAs that are terminated with 5 0 caps and 3 0 tails, the 5 0 and 3 0 ends in circRNAs have been jointed together, forming covalently closed loop structures [16].Circularization enhances circRNA stability, enabling them to be resistant to RNase R that preferentially degrades linear species [17].Genome-wide analyses indicated that the majority of circRNAs are abundant, conserved across species and often exhibit celltype or tissue-specific expression, suggesting potential regulatory roles [13,14,17].circRNAs can arise from exons (exonic circRNA, ecircRNA) [9,17], introns (intronic circRNA, ciRNA) [18,19] or both (exon-intron circRNA, EIciRNA) [20], even the tRNA Xianwen Meng is a PhD student in Ming Chen's laboratory in Zhejiang University.His research focuses on biogenesis, properties and functions of circular RNAs in eukaryotic cells.Xue Li is a Master student in Ming Chen's laboratory in Zhejiang University.Her research focuses on single-cell RNA-seq transcriptome analysis of circular RNAs.Peijing Zhang is a Master student in Ming Chen's laboratory in Zhejiang University.Her research focuses on analysis of circular RNAs in plants.Jingjing Wang is a PhD student in Ming Chen's laboratory in Zhejiang University.Her research focuses on epigenetic regulation network in plants.
Yincong Zhou is a Master student in Ming Chen's laboratory in Zhejiang University.His research focuses on image analysis for high-throughput plant phenotyping.Ming Chen is a full professor in the Department of Bioinformatics, College of Life Sciences, Zhejiang University.His current research focuses on construction of noncoding RNA-mediated regulatory networks and establishing useful web servers and platforms to help biologists browse and analyze massive biological data sets.Submitted: 12 January 2016; Received (in revised form): 5 April 2016 introns could form stable circular RNAs (tricRNA) [21] (Figure 1).However, most circRNAs come from exons of protein coding genes via 'back-splicing' [22], where a 5 0 splice donor joins an upstream 3 0 splice acceptor [3].Developing evidence has revealed that complementary sequences, which can be repetitive or not, and exon skipping are responsible for circRNA biogenesis [23][24][25][26][27].In addition, RNA-binding proteins (RBPs) are involved in this process [28,29].Once generated, circRNA appears to be noncoding because of lacking start and/or stop codons.
Why do circRNAs become a research hotspot in the field of RNA?That is mainly owing to their newly discovered functions in diverse cellular processes.Strikingly, some circRNAs could serve as miRNA sponges [1,2,6] by sequestering and preventing miRNAs from binding target genes [30].For example, the human circRNA ciRS-7 contains >70 binding sites for miR-7, which could strongly suppress miR-7 activity [2], thus shedding new light on the roles of circRNAs.In addition to miRNA regulation, it has been proposed that circRNAs could regulate the intracellular transport of RBPs [3], as well as influence the parental gene expression [31].Recently, Zhang et al. [18] proposed that one ciRNA, ci-ankrd52, could act as a positive regulator of Pol II transcription and Li et al. [20] revealed that EIciRNA could participate in transcriptional regulations via specific RNA-RNA interaction with U1 snRNA, indicating a novel regulatory strategy for transcriptional control.Collectively, circRNAs have greatly expanded the complexity of transcriptomic regulation.circRNAs are not only involved in long-term processes such as early embryonic development and aging [32,33], but also in diseaserelated pathways, becoming a promising biomarker for disease diagnosis [34,35].In summary, we have undoubtedly entered an age of studying this special RNA species.
In this review, we comprehensively delineate the universe of circRNAs, from their formation to functions.Specially, we discuss the current circRNA bioinformatics.Finally, it should have come as no surprise to learn that circRNA is becoming a key player in RNA world.

Biogenesis
In eukaryotic cells, pre-mRNA splicing occurs through two-step mechanism [36].In the first step, the 5 0 -splice site is attacked by the 2 0 -OH group of a conserved RNA adenine nucleotide in the intron, producing an intron lariat-3 0 -exon as well as a free 3 0 -OH group at the 5 0 -exon.In the second step, the generated 3 0 -OH group initiates a nucleophilic attack on 3 0 splice site, resulting in the release of the intron lariat and joint of two exons (Figure 2A).
Given that pre-mRNA splicing is carried out by spliceosomal machinery, it is believed that circRNAs are also produced through spliceosomal splicing mechanisms.Researchers have discovered that canonical splice signals generally immediately flank the junction sites in circRNAs [1] and the spliceosome has been implicated in the formation of these signals [28].Inhibition of the canonical spliceosome using isoginkgetin treatment caused a reduction in both linear and circular transcript levels [26], suggesting an essential role the spliceosomal machinery plays in circRNA biogenesis.Additionally, the expression levels of circRNAs and the corresponding linear mRNAs do not always exhibit a correlation [14], suggesting that this process is regulated and the spliceosome could distinguish between backsplicing and forward splicing, whereas the detailed mechanism remains to be elucidated.
Direct back-splicing, where exons are spliced in noncanonical order, connects the downstream 5 0 splice site (splice donor) to an upstream 3 0 splice site (splice acceptor), thereby generating a circular transcript [12].The second mechanism is lariat splicing, which involves generation of a lariat intermediate containing exons, and then, the introns in the lariat are removed, generating ecircRNAs [40].Both mechanisms function in vivo and correlate with the canonical spliceosome, but direct backsplicing may occur more frequently than exon skipping in ecircRNA biogenesis [3].Recent findings also indicate that RNA pairing, which can be formed by either repetitive elements [17] or nonrepetitive but complementary sequences [23], across flanking introns is positively associated with exon circularization.Additional observations have revealed that RNA binding proteins such as ADAR, MBL and QKI function in specific ecircRNA biogenesis [25,28,29].Of note, these three proteins are known to be implicated in alternative splicing.In addition, compared with all expressed exons, the exons generating single-exon circRNA are 3-fold longer in length [13,17].Therefore, the length of a given exon may also correlate with circularization.

circRNA from exon-introns
As a special subtype of circRNA, EIciRNAs were found in mammalian cells [20].Composed of both exonic and intronic sequences of coding genes, EIciRNAs are distinct from ecircRNAs, which are composed exclusively of exonic sequences (Figure 1A).Nevertheless, like ecircRNAs, EIciRNAs also tend to be coupled with long introns as well as complementary sequences, indicating that the mechanism of EIciRNA formation should be similar with that of ecircRNAs, although the intron is retained with unknown mechanism.

circRNA from introns
Eukaryotic spliceosomal introns may form circular intronic RNAs.After the release of the 5 0 -exon, the terminal 3 0 -OH attacks the 3 0 -splice site, generating the ciRNA and releasing the 3 0 -exon (Figure 1B).A consensus motif composed of a 7 nt GUrich element near the 5 0 splice site as well as an 11 nt C-rich element close to the branchpoint site is necessary for ciRNA biogenesis.This motif should be ciRNA specific because it is not enriched in regular introns or other types of circRNAs [18].

circRNA from tRNA
As a special class of intronic circRNAs, tricRNAs are generated during pre-tRNA splicing.Conserved tRNA sequence motifs and processing enzymes are required for the formation of tricRNA [21].The tRNA splicing endonuclease complex get rid of the tRNA introns by recognizing the bulge-helix-bulge motif in the pre-tRNA and cleaving within the anticodon loop [41].Then the released ends are ligated, generating tRNA and tricRNA (Figure 1C).

Structure
The covalently closed loop structures with neither 5 0 -3 0 polarity nor a polyadenylated tail make circRNAs more stable than linear transcripts in cells, with most species exhibiting a half-life over 48 h, whereas their linear counterparts having half-lives of <20 h [17].In addition, this enhanced stability may be responsible for the phenomena that some circRNAs exhibit higher expression levels than their linear counterparts [13,14].However, circRNAs can be degraded by short interfering RNAs [1,17,42], which are useful in detecting circRNA functions.

Conservation
CircRNAs exhibit some sequence conservation.For instance, 457 of 2121 circRNAs in humans were found with circular orthologues in murine [1] and approximately 4% of humans and mice orthologous genes can generate circRNAs [14].ecircRNAs show high universal sequence conservation especially in the third codon position, whereas the other types of circRNAs show only weak conservation, but they still exhibit a remarkable enrichment of conserved nucleotides [1].The conservation of circRNAs among species also implies that circRNAs are not byproducts of pre-mRNA splicing.However, a recent study revealed that exons within the human-mouse orthologous circRNAs are no more conserved than their neighboring linear exons [43].Together with other findings, the authors considered that most circRNAs might be inconsequential by-products of imperfect pre-mRNA splicing.Thus, the circRNA sequences may not explain this issue properly.Nevertheless, circRNAs tend to be flanked by longer introns in diverse species [4,13,15,17].The long intron itself does not cause circRNA formation; however, longer introns tend to contain reverse complementary sequences, which contribute to the formation of circRNAs [23].The presence of reverse complementary sequences in the surrounding introns of circRNAs is a conserved feature in animals and plants [15,25].In addition, circRNAs harbor a deletion of polymorphisms at miRNA binding sites, indicating miRNA binding sites in circRNAs undergo similar selective pressure as those in linear mRNAs [44].

Localization
Most ecircRNAs tend to be cytoplasmic [1,13,17].The transport mechanism is currently not clear.The exon-exon junction complex may recruit mRNA export factors during ecircRNA being generated by the spliceosomal machinery [45] and then ecircRNAs are transported by nuclear export system.Alternatively, ecircRNAs may be delivered to cytoplasm during mitosis.In cytoplasm, natural ecircRNAs are not associated with ribosomes, suggesting that ecircRNAs are unable to be translated [17,43], although engineered circRNAs have been shown to be translatable [46,47].EIciRNAs and ciRNAs predominantly localize in the nucleus [18,20], suggesting that they may govern the gene transcription.For instance, knockdown of circEIF3J could significantly decrease EIF3J levels [20].

Expression
CircRNAs are expressed across the eukaryotic tree of life.In human cells, circRNAs comprise >10% of all transcripts [13].Previous analyses suggested that only weak correlation was shown between the levels of circRNAs and linear counterparts [14].A few circRNAs are more abundant than their linear counterparts while most are expressed at a low level [17].Widely expressed circRNAs across diverse cell types showed significantly higher expression level than narrowly expressed ones [48].Additionally, circRNAs often exhibit tissue/developmentalspecific expression [14,49,50].For example, in flies, circRNAs tend to have a higher expression level in neural tissues [4], while in human and mouse, circRNAs were discovered to be enriched in brain [51,52] and platelets [53]; hundreds of circRNAs appeared to be regulated when human mammary epithelial cells underwent epithelial-mesenchymal transition in response to the treatment with TGF-b [29].However, the mechanisms and related functions behind the change in circRNA abundance await further investigation.
Taken together, these properties suggest that circRNAs might play important roles in transcriptional and post-transcriptional processes.

Functions
CircRNAs are attracting more attention owing to not only their abundance in the eukaryotic transcriptomes but also their functions.So far, it has come to light that circRNAs function in multiple biological processes, such as miRNA binding, protein binding and regulation of transcription and posttranscription (Figure 3).

miRNA sponge
The most exciting finding is that circRNAs could function as miRNA sponges.For instance, ciRS-7/CDR1as harbors >70 conserved binding sites for miR-7 [1,2].miR-7 activity could be dramatically reduced through tethering of this miRNA to ciRS-7/ CDR1as, resulting in increased levels of miR-7 targets.Then, circRNA destruction could release a shower of miR-7.Sry, a testis-specific circRNA, is another validated miRNA sponge [2,11].This circRNA harbors 16 binding sites for miR-138.These findings changed the mechanistic understanding of miRNA regulatory networks and increased the complexity of competitive endogenous RNA (ceRNA) network [54].Although few circRNAs contain a majority of miRNA binding sites for a single miRNA, the depletion of polymorphisms at predicted miRNA binding sites implies that circRNAs could be efficiently targeted by miRNAs.

Protein sponge
circRNAs can also act as protein sponges.For example, circMbl, derived from the muscleblind (MBL/MBNL1) in flies and humans, harbors multiple muscleblind binding sites [28].When MBL proteins are in excess, circMbl could sponge out the redundant ones.By this way, MBL levels could be well regulated.

Alternative splicing, mRNA trap and transcriptional regulation
CircRNAs can affect their parental genes in cis-or trans-actions.Circularization and splicing compete against each other, enabling ecircRNAs to function in alternative splicing [28].Once a back-splice occurs, it removes the internal exons, causing an alternative splicing.ecircRNAs may serve as an 'mRNA trap' by sequestering the translation start site, making the truncated linear mRNA fail to translate.For instance, the mice formin (Fmn) gene can generate ecircRNAs acting as mRNA trap.Deletion of specific exons caused a failure to produce these ecircRNAs, but appeared to produce normal amounts of the linear mRNAs.The inability to generate ecircRNAs from the Fmn locus caused aberrant expression of the formin protein and changed the phenotype as a result [31].
Moreover, ciRNAs can function as a positive regulator of Pol II transcription by interacting with Pol II machinery [18].More recently, a study has revealed that EIciRNAs use a specific interaction with U1 snRNA to fine-tune the expression of their parental genes in the nucleus [20].This discovery not only reveals a role for circRNAs in transcriptional control but also highlights a regulatory strategy through specific RNA-RNA interaction between EIciRNA and U1 snRNA.
Collectively, the interplay between circRNAs and the transcriptional machinery provides new insights into the strategies to regulate gene expression in cells.

Disease
Because some circRNAs contain one or more types of miRNA binding sites, the relationship between miRNA and disease indicates that circRNAs may have a regulatory role in the occurrence of diseases [55].As mentioned above, miR-7 activity can be efficiently regulated by ciRS-7.This has significant implications for the biology of disease because miR-7 has been implicated in many diseases, such as diabetes [56] and diverse cancers [57,58].Dysregulated miR-7 will cause aberrant expression of its targets, of which many are oncogenes, and further influence the downstream pathways, of which most are cancer related.Therefore, fine-tuning of the miR-7/ciRS-7 axis plays important roles in cancer-related processes [59][60][61].Similarly, as another miRNA sponge, SRY can strongly suppress the miR-138 level [2].In cholangiocarcinoma, reduced miR-138 level will enhance proliferation, migration and invasion of cholangiocarcinoma cells [62].Another report revealed that cir-ITCH, as sponge of miR-7, miR-17 and miR-214, was down-regulated in esophageal squamous cell carcinoma, which might increase the level of ITCH [6].ITCH hyper expression promotes ubiquitination and degradation of phosphorylated Dvl2, which will further inhibit the Wnt/b-catenin pathway [6].Because Wnt pathway is frequently aberrant in cancers [63], cir-ITCH may have been implicated in tumorigenesis.Evidence is emerging that circRNAs exhibit dysregulation in atherosclerotic vascular disease [64], Alzheimer [65] and cancers [6,66], implying that circRNAs potentially contribute to pathological states.
Salivar extracellular RNAs are potential biomarkers for detection of multiple diseases, such as lung cancer [67].Interestingly, Bahn et al. [68] revealed that circRNAs even exist extracellularly in salivar and may be implicated in intercellular These observations inevitably lead to an interest in the potential roles of circRNAs in human diseases, especially in cancers.More importantly, synthetic circRNA inhibitors may become future therapeutic strategies in cancers because the circularized miRNA sponges exhibit superior anti-cancer activities.

circRNAs in plant
A knowledge gap had remained regarding circRNAs in plants until the first discovery released in 2014 [5].However, at that time, it merely demonstrated the widespread and substantial presence of circRNAs in Arabidopsis thaliana transcriptome.More recently, Ye et al. [15] performed another genome-wide identification of circRNAs in Oryza sativa and A. thaliana and revealed the common and distinct features of circRNAs between plants and animals, such as some conservation, longer flanking introns, diverse expression patterns but less reverse complementary sequences.Some circRNAs exhibited differentially expression under Pi-sufficient and Pi-starvation condition, suggesting that circRNAs may play a role in responding to Pistarvation stress in rice, although the mechanism remains unknown [15].Lu et al. [69] revealed that rice circRNAs exhibit alternative splicing circularization patterns and act as a negative regulator of their parental genes, providing new biological insights into rice circRNAs.Systematic studies of this special class of noncoding RNA have just begun in plants.Much still remains to be uncovered about circRNAs in plants especially the biological roles they play.

Online resources and bioinformatics analysis
To better understand the relationship between circRNA and disease, Ghosal et al. [55] predicted the circRNA-related diseases through bioinformatics methods and constructed the first knowledgebase of potential association of circRNAs with diseases in human, named Circ2Traits (http://gyanxet-beta.com/ circdb/).Glazar et al. [70] merged and unified previous data sets of eukaryotic circRNAs into a database, circBase (http://www.circbase.org/),while Liu et al. [71] calculated the expression of circRNAs in 464 human RNA-seq samples and constructed the CircNet database (http://circnet.mbc.nctu.edu.tw/).Notably, CircNet provides the circRNA-miRNA-gene regulatory networks, while deepBase v2.0 (http://biocenter.sysu.edu.cn/deepBase/)included 14 867 human circRNAs identified from RNAseq data [72].More recently, Dudekula et al. [73] developed CircInteractome (http://circinteractome.nia.nih.gov/) that can be used to explore the interacting miRNAs and proteins of circRNAs.The circRNA data sets used by these databases are mainly from previous studies or identified using the algorithm proposed by Memczak et al. [1].So far, based on next-generation sequencing technology, the published circRNAs tend to be identified in limited tissues or cell types (Supplementary Table S1).Because circRNAs exhibit tissue-and developmental-specific expression, the compiled circRNA list in these databases is inadequate.Although CircNet provides the expressions of circRNAs in a handful of tissues, the authors used poly(A)selected RNA-sequencing (mRNA-Seq) data to calculate circRNA expressions.This is unwise because circRNAs have no poly(A) tails.The data from rRNA-depleted total or poly(A)-depleted RNA-Seq are recommended.In addition, a better naming system for circRNAs is still needed because the naming rules of circRNAs in current databases are different.In CircNet, circRNA names include the host gene symbols while other databases number circRNAs considering the order of their locations in genome.Nevertheless, circRNA expressions and circRNArelated molecular interactions provided by these databases will facilitate experimental analysis of circRNAs.
To identify circRNAs from rRNA-depleted RNA-Seq (RibominusSeq) data, the core method is based on the presence of back-splice junction-spanning reads.First, the raw reads need to be aligned to the reference genome.The mapped reads are discarded.Twenty nucleotide anchor sequences split from both ends of the unmapped reads are further used to align to find unique anchor position.A pair of anchors that align in the reverse orientation suggests a back-splice (Figure 4).Then, to determine circRNA candidates, a series of filtering criteria should be used as follows: (i) canonical GU/AG splicing signals flanking the splice sites; (ii) at least two independent junction reads; (iii) unique anchor alignments; (iv) a limited distance between the two splice sites; (v) unambiguous breakpoint detection; (vi) allowing 1-2 nt shift within the splice sites; (vii) removal of reads with alignments on different genes or noncanonical splice; (viii) if paired end reads are available, a junction read is considered only when its paired read aligns within the region of a circRNA template; (ix) a further filter to prevent false predictions resulting from repetitive or homologous regions; (x) removal of back-splice junction-spanning reads that map to the mitochondrial genome.
Comprehensive detection and quantification of circRNA is essential for exploring their biological functions.To date, several circRNA-specific prediction tools have been developed, such as find_circ [1], UROBORUS [74] and DCC-CircTest [75].Meanwhile, some multi-functional toolkits like MapSplice2 [76], segemehl [77] and NCLscan [78] also include the circRNA detection methods (Table 1).These tools can deal with both singleand paired-end RNA-Seq data except for circRNA_finder and NCLscan, which focus on paired-end sequencing data.Most tools use backsplice reads as a key element in circRNA detection.Therefore, detection performance mainly relies on the read mappers and mapping strategies used by these tools to align the junction reads to the genome.Currently, Bowtie [79], BWA [80] and STAR [81] are frequently used owing to their higher precision and speed.Besides, the filtering criteria adopted by different tools also influence the detection results.Current detection tools can be classified into two categories: fragment-based and pseudo-reference-based strategies.A fragment-based strategy is generally depicted as above while a pseudo-reference-based strategy needs to construct all possible scrambled exon junction boundaries.Once a read mapped to the non-co-linear (NCL) junction site suggests a circRNA candidate.Because gene annotation is indispensable for the pseudo-reference-based strategy, this strategy cannot detect circRNAs with unannotated exon boundaries.Besides, this strategy seems incapable of discovering circRNAs in the incomplete or poorly annotated genomes.
The goal of MapSplice is to detect exon splice junctions from RNA-Seq data.First, short sequence tags are split into 25 nt segments and aligned to reference genome.Then, splice significance scores are calculated for splice junctions that appear in one or more tag alignments based on the quality and diversity of alignments.Last, by using a noncanonical exon ordering strategy, circRNA candidates can be found.As a circRNAspecific detection tool, find_circ extracts 20 nt anchors from both ends of the unmapped reads and aligns them against reference genome independently.Anchors that aligned in the reversed orientation and a canonical GU/AG splicing signal within the corresponding region suggest a back-splice event.After a series of filtering criteria, such as at least two independent reads, candidate circRNAs can be produced.Segemehl implements a matching strategy based on enhanced suffix arrays, which can be used for unbiased circRNA, splicing, trans-splicing and fusion detection.STAR, a fast and splice-aware read mapper, is used by DDC-CircTest and circRNA_finder to detect chimeric junction reads.Then, a sequence of filtering criteria is used to these reads to find circRNA candidates.Several tools directly use read counts over the back-splice junctions to estimate the expression of circRNAs, while DCC-CircTest uses a statistical framework based on the beta-binomial model for estimating circRNA versus host gene expression.CIRCexplorer is a TopHat-Fusion-based tool for detecting NCL events.The unmapped reads are split into 25 nt segments that are further aligned to the genome.Then it parses fusion information from mapping results.To find the donor or accepter splice site, backspliced junction reads are realigned against gene annotations.As a result, CIRCexplorer does not support de novo identification of circRNAs.CIRI proposes a novel paired chiastic clipping (PCC) signal-based algorithm combined with a systematic filtering to remove false positives.Instead of dividing an unmapped read into two segments, CIRI collects and compares mapping information for all split alignments of a read to detect PCC signals.Currently, there remain some challenges in circRNA detection.In eukaryotic genomes, other types of chromosomal events, such as genetic rearrangements, may also produce the NCL junctions, leading to false discoveries.circRNAs show a low abundance compared with their linear counterparts; however, a large amount of total RNA-Seq data lack a circRNA enrichment step, such as RNase R treatment or poly(A) depletion, causing a difficult to distinguish between noise and true reads present at low levels.In addition, currently available circRNA detection tools based on 20-nt anchors may miss the short spanning reads with low expression levels.The length of reads and depth of sequencing coverage are distinct in different data sets, which influence the unbiased detection of circRNAs.While segemehl, find_circ and CIRI showed similar levels of sensitivity, many discrepancies exist in circRNA candidates identified by different circRNA-detecting methods [43,82,83].For example, Hansen et al. [83] compared five circRNA detection tools, only 16.8% of circRNAs are predicted by all tools and the false positive rate range from 12% to 28%.Although 3/5 of the five tools use Bowtie as read mapper, these three tools use different mapping strategies, alignment parameters and filtering criteria, causing dramatic discrepancies in circRNA detection.For future circRNA detection tools, a standard filtering criteria is recommended to achieve reliable results.Current circRNA detection tools tend to sacrifice sensitivity to achieve better precision.Existing algorithms are insufficient for removal of false positives, circRNA discovery will benefit from appropriate statistical methods and unbiased strategies.Take KNIFE for an example, this tool improves detection sensitivity by assigning statistical confidence for each detected circular.Currently, to achieve reliable predictions, an efficient integration of the results from different tools should be more plausible.These tools provide the junction sites and junction reads, but they do not offer the circRNA sequences.This is mainly because multiple circRNA isoforms might originate from the same junction site [14,51].Nevertheless, the sequences and expressions of circRNA isoforms are crucial for uncovering circRNA functions.Sometimes the prediction tools may not work perfectly on local servers [83], thus, a more userfriendly interface will greatly improve their performance.
A systems biology analysis of the identified circRNAs will gain insights into circRNA functions because most biological impacts of circRNAs remain unclear.Differentially expression analysis is widely used to discover tissue-and developmentalspecific circRNAs.Construction of circRNA-miRNA coexpression network can delineate the interplays between circRNAs and miRNAs.Acting as a new class of competing endogenous RNA molecules, circRNAs can suppress miRNA functions by sequestering them efficiently.To detect ceRNAs, first, miRNA target sites on circRNAs should be identified using miRNA target prediction algorithms, such as TargetScan [84].Then, the ceRNA score [85] and p-value [86] of a circRNA-circRNA or circRNA-mRNA pair are calculated to measure the possibility of a circRNA to act as ceRNA.An index of co-expression [87] needs to be computed because the simultaneous expression of ceRNA pairs is an important indicator of their biological relevance.As a consequence, circRNAs together with mRNAs, miRNAs as well as lncRNAs form a complicated interaction network.Identifying regulatory modules from the network will help us to disclose potential functions of circRNAs.Moreover, recent advances in the analysis of RNA-protein interactions, such as PAR-CLIP, HITS-CLIP and iCLIP, have allowed precisely to find the binding sites of RBPs on RNAs.These technologies will decipher the crosstalk between circRNAs and RBPs, and further elucidate their biological roles because several circRNAs exert their effect by forming circRNA-RBP complexes.

Perspective
High-throughput sequencing analyses have revealed the prevalence of circRNAs in eukaryotic transcriptomes and brought circRNAs to the fore of the RNA world.It has been confirmed that circRNAs are stable, abundant, conserved and often exhibit tissue-and developmental-specific expression.
RNA circles can be generated from exons as well as introns, suggesting its diversity.Canonical spliceosomal machinery, ciselements and trans-factors are involved in circRNA formation.Currently, two mechanisms of exonic circRNA biogenesis are proposed, including direct back-splicing and exon skipping.Disclosure of the spliceosome structure in yeast sheds light on the structure basis of pre-mRNA splicing and circularization [88,89].However, much still remains unknown about the control of the back-splicing events that produce circRNAs and the crosstalk between RNA structure and RBPs in circRNA biogenesis.RNA modifications such as m 6 A are implicated in alternative splicing [90], whether RNA modifications correlate with the formation of circRNAs remains for further investigation.In addition, the epigenetic status of circRNA genes and the degradation strategy adopted by circRNAs remain poorly understood.More recently, Dong et al. [91] revealed that pseudogenes can be retrotransposed from circRNAs in mammalian genomes.However, the molecular mechanism of circRNA retrotransposition remains unclear.Whether this phenomenon also exists in plant genomes would be interesting to explore it.
EcircRNAs mainly accumulate in cytoplasm, while EIciRNAs and ciRNAs are enriched in nucleus, indicating different roles they may play in eukaryotic cells.However, the machinery involved in circRNA transport remains elusive.
To date, a lot of exciting findings from circRNAs have been unveiled, of which the most exciting one is that circRNAs can act as miRNA sponges.As a novel class of ceRNA regulators, circRNAs have increased the diversity of transcriptomes enormously.In addition to miRNA sponges, circRNAs are able to govern the expression of protein coding genes by participating in transcriptional control.
Surprisingly, circRNAs frequently exhibit aberrant expression in human diseases.circRNAs represent a new class of diagnostic biomarkers, although the mechanisms of these RNA circles in the development of diverse diseases are not yet fully understood.Notably, circRNAs have potential therapeutic roles.The stability makes them long-acting regulators of cellular behavior.Construction of circular miRNA sponges targeting the oncogenic miRNAs provides a potential strategy for cancer therapies because they are more effective than linear miRNA sponges in reducing the oncogenic miRNA activities in the context of cancer.Discovery of new drugs targeting risk circRNAs may improve the conditions of patients.
Although it has been known that circRNAs in plants share similar properties with those in animals, we are still at the dawn of this field in plants.The biological processes in which circRNAs are implicated during growth and stress response need to be explored.
As the first step toward understanding circRNA biology, databases and computational prediction tools are now available for researchers to retrieve and identify circRNAs.However, existing databases show some drawbacks, and current results from different tools exhibit dramatic divergences.Although integrative approaches may improve the circRNA identification, tools with higher precision are still needed.While novel mapping strategies are emerging, a recommended filtering criteria is waiting to be founded.
Except for differential expression analysis, a systematic strategy based on networks will play a role in deciphering circRNA biology.For example, construction of a comprehensive competing network comprised circRNA, mRNA and lncRNA can explain the functions of circRNAs better.
Finally, as an emerging key player in RNA world, circRNA provides novel insights into the eukaryotic transcriptomes.A significant gap still exists in our understanding of this special RNA species.However, with the rapid development of biotechnologies and bioinformatics methodologies, the whole circRNA story will be revealed soon.

Key Points
• Advances in sequencing technologies coupled with bioinformatics approaches provide a new perspective for comprehensive analyses of circular RNAs.
• CircRNAs are stable, abundant, conserved and often exhibit tissue-and developmental-specific expression.
• Given the functions of circRNA such as miRNA sponge and transcriptional regulation, they have dramatically increased the complexity of eukaryotic transcriptomes.
• Aberrant expression in diverse diseases suggests circRNAs may become diagnostic biomarkers or therapeutic targets.
• Next-generation sequencing technologies coupled with bioinformatics strategies play crucial roles in deciphering the biological functions of circRNAs.

Figure 1 .
Figure 1.Diverse types of circRNAs.(A) Exonic circRNA and exon-intron circRNA.Canonical splicing produces linear mRNAs while noncanonical splicing (back-splicing) produces ecircRNAs.Particularly, if the intron is retained, an EIciRNA will be generated.(B) Circular intronic RNA.Some lariat introns excised from pre-mRNA by canonical splicing machinery could further form stable ciRNAs.(C) tRNA intronic circRNA.tricRNAs derive from introns that are removed during pre-tRNA splicing.A colour version of this figure is available at BIB online: https://academic.oup.com/bib.

Figure 2 .
Figure 2. Models of ecircRNA formation.(A) Canonical splicing.(B) Direct back-splicing.Intron pairing or RBP pairing bridges two flanking introns close together, and then the released two exon ends are connected after the removal of the introns.(C) Exon skipping.The skipped section containing exons and introns undergoes intralariat splicing, generating an ecircRNA and a lariat as a result.A colour version of this figure is available at BIB online: https://academic.oup.com/bib.

Figure 3 .
Figure 3. CircRNA functions.EIciRNAs and ciRNAs enriched in nucleus can regulate gene transcription by interacting with Pol II machinery.Because circularization and alternative splicing compete against with each other, the production of circRNAs may change the composition of their linear counterparts.Specially, if circRNAs sequester the translation start site, it will cause a failure for the truncated mRNA to translate.ecircRNAs tend to be cytoplasmic, acting as miRNA sponges, RBP sponges and delivery intermediates.A colour version of this figure is available at BIB online: https://academic.oup.com/bib.

Figure 4 .
Figure 4. Bioinformatics identification of circRNAs.A colour version of this figure is available at BIB online: https://academic.oup.com/bib.

Table 1 .
Circular RNA prediction tools Downloaded from https://academic.oup.com/bib/article-abstract/18/4/547/2562762 by guest on 06 March 2019Junction reads with PCC signals reflect a circRNA candidate.The PCC signal is independent of annotation, making it possible for de novo back-splice detection.NCLscan constructs putative NCL references with putative NCL junctions based on alignment output and gene annotation.The unmapped reads are aligned against NCL references to detect junction reads.Because NCLscan only considers the well-known exon junction sites, it is limited to ecircRNA detection.ACF detects the head-to-tail junction reads by using a fragment-based strategy.Particularly, in addition to similar filtering processes, for each junction site, ACF calculates the strength of splicing signal.And a head-to-tail junction site whose splicing score is 10 indicates a circRNA candidate.ACF depends on neither exon annotations nor canonical splice sites, making it possible to detect circRNAs in organisms without a complete transcriptome annotation.KNIFE is a pseudo-reference-based strategy to identify circRNAs.A linear junction index and a scrambled junction index are constructed.A high-scoring alignment to scrambled junction suggests a circRNA.To discover circRNAs generated from un-annotated exon boundaries in the genome, a de novo analysis module is included.Particularly, alignment score, mapping quality and offset position are used to fit a logistic generalized linear model.Consequently, a statistical score is assigned for each detected circRNA.Finally, it exhibits a reduction of false-positive results compared with other approaches.UROBORUS divides the noncanonical junction reads into balanced mapped junction reads and unbalanced mapped junction reads based on whether the both anchors are mapped to the back-spliced exons.Of note, by taking into consideration short spanning reads, UROBORUS can find those circRNAs expressed at lower levels.However, current version can only detect ecircRNAs and miss other types of circRNAs.