Catalyzed by spliceosome, the process of gradually editing the precursor m-RNA (pre-mRNA) into mature mRNA through intron removal is called RNA splicing, in which 99% of the bases of 5' end (donor site) and 3' end (receptor site) of introns are almost GU and AG 25. Spliceosome is a 60S complex of RNA and protein formed during RNA splicing, which has the function of identifying 5' splice site (5' SS), 3' splice site (3' SS), and branching points of the pre-mRNA. A single pre-mRNA transcript can generate different mature mRNA isomers through alternative splicing, which is an important mechanism that leads to large diversity of genes and proteins in eukaryote 26. CircRNAs are formed by a unique alternative splicing mechanism termed backsplicing, which generate the closed loop structure via joining a splice donor to an upstream splice acceptor. After canonical splicing, the relative order of exons in mRNA matches the order in genome, and no shuffling of exons occurs. While circRNA contains scrambled exons, which means the order of exons is different from that present in the nascent transcript 3.

According to the constituent sections, circRNA can mainly be classified into three categories: one is termed exonic circRNAs (ecircRNAs), which localized in cytoplasm are mainly composed of single or several exons 232427. Another with none critical elements and little enrichment for miRNA target sites inside is termed circular intronic RNAs (ciRNAs), which only contain introns and are localized mainly in the nucleus, mainly promote gene transcription by binding to RNA polymerase II 27. The last one also localized in nucleus is termed exon-intron circRNAs (EIciRNAs), which contain both exons and introns 328.

In addition, a special type of intronic circRNAs that are generated during pre-tRNA splicing, called tRNA intronic circRNAs (tricRNAs), which have been discovered in Archaea29 and Drosophila 30. The anciently conserved tRNA sequence motifs and the tRNA splicing endonuclease complex are necessary to remove introns, which then ligated by a 3', 5'-phosphodiester to form tricRNAs 31. ElciRNAs play a monitoring role that ensures the integrity of the transcriptome by binding to the U1 element of the snRNPs. The U1 element initially binds to small nuclear ribonucleoproteins (snRNPs) to form a complex, which then combines with RNA pol II to promote transcription. U1 is a key element which can prevent the early termination during transcription 103233.

Besides the first and last exons of the pre-mRNA, all internal exons theoretically can be circularized via the splicing signals in flanking introns on both sides. However, backsplicing reactions often occur at an extremely low level 34. Their efficiency can be regulated by RNA binding proteins, exon skipping events, as well as the core spliceosomal components. At present, the theory regarding the biosynthesis of circRNAs consists of the following five models.

i) Lariat-driven circularization (Figure 1b) 35. This model is also called exon skipping, in which the formation of circRNAs based on the canonical splicing. If the pre-mRNA is partially folded, exon(s) may be skipped during the RNA splicing, then two previously non-adjacent exons are connected together to form a linear RNA. The skipped exon(s) along with the surrounding introns will form a structure of lariat, which should be degraded in normal cases. However, if the lariat further to be spliced again via backsplicing, the ecircRNA will be generated 3637.

ii) Intron pairing-driven circularization (or direct backsplicing) (Figure 1c) 38. It is demonstrated that the biogenesis of circRNAs can be promoted by complementary base pairing interactions between flanking intronic repeats 11223334394041. In details, these complementary repeats can form double-stranded RNA (dsRNA) structures, which bring the splice sites close to each other so that the backsplicing occurs 3339. In the following step, the introns are removed or retained to form ecircRNAs or EIciRNAs. The occurance of this modle relys on the Alu complementarity which is specific in primates.

iii) Circularization depends on RNA binding proteins (RBPs) (Figure 1a) 42. With the help of RBPs, the splice donor and acceptor between flanking introns are brought close to each other, thereby promoting backsplicing 4344. QKI is a alternative splicing factor that can combine with flanking intron sequences to form dimers to promote circularization which is similar to modle 2 43. Conn et al.44 have found that large amounts of circRNAs formed along with the up-regulation of QKI during human epithelial-mesenchymal transition (EMT), which indicates that the formation of circRNAs is time-space specific. Another RBP, MBL, can bind to the conserved sequence of its own introns and then regulate the circularization of exons, which derive from its own genes 43. Moreover, SP proteins and hnRNPs in Drosophila have similar effects on the production of specific circRNAs 40.

However, RBPs also inhibit the formation of circRNAs. For example, ADAR1 as a RNA editing enzyme, inhibits the synthesis of circRNAs via interacting with dsRNA and splitting it 45. On the other hand, NF90/NF11046 or the RNA helicase DHX9 could also inhibit the backsplicing by directly unwinding the dsRNA or by recruiting ADAR1 47.

iv) Formation of circular intronic RNAs (ciRNAs) (Figure 1d). A 7-nt GU-rich sequence near the 5' splice site and an 11-nt C-rich sequence near the branching point prevents the larait structure being branched and degraded, and the lariat tail downstream of the branching point is trimmed to produce a stable ciRNA 48. These special ciRNAs can act as cis-elements to interact with RNA polymerase II (Pol II) to promote transcription 27. However, it is still unknown how these functional factors escape degradation.

ⅴ) alternative splicing 49. The competitional splicing of the pre-mRNA leading the single protein-coding gene to generate mutiple transcripts along with distinct circular RNAs, this process which is known as alternative splicing 2505152. The number of repetitive elements, spatial distance, and their degree of complementarity will all affect the splicing outcome 33.

Functions for the vast majority of circular RNAs remain unknown, but recently reported that some of them play roles in regulating miRNAs, alternative splicing patterns, or can be bound with ribosomes to produce proteins. Furthermore, circRNAs have additional novel functions such as acting as sponges for RBPs, direct binding to target genes, and direct translation as templates5354

The ceRNA hypothesis refers to the combination of a series of RNAs that can competitively bind to miRNAs, such as mRNAs, pseudogenes, and lncRNAs. These molecules all contain several homologous miRNA response elements (MREs) that can adsorb miRNAs like sponges to regulate mRNA expression levels, thereby affecting their functions 55. CircRNAs can also absorb miRNAs to eliminate the inhibitory effect of miRNAs on the target gene 56. Therefore, circRNAs belong to ceRNAs, and the adsorption capacity of miRNAs is stronger than that of linear mRNAs and lncRNAs due to the molecular characteristics of its long halftime, which helps in the mining of gene functions and regulatory mechanisms. For example, CiRS-7/CDR1as has been identified as a super sponge containing more than 70 conserved binding sites for miR-7 and SRY70 and 16 binding sites for miR-138 57. It has been proven that circHIP3K contains multiple miRNA binding sites 58. Subsequently, circular RNA-ITCHs (cir-ITCHs) have been shown to adsorb miR-22-3p to upregulate CBL expression 59. Besides, cir-ITCH can also act as sponge of oncogenic miR-7 and miR-214 to enhance ITCH expression 60. Therefore, circRNAs may play an important regulatory role in disease through the interaction with disease-associated miRNAs, The interaction above could be described as circRNA-miRNA-mRNA regulation axis (Figure 2). Numerous circRNAs that have been reported to date, and thousands of circRNAs were predicted with miRNA binding sites, but it is still unknown whether they are all functional. Then by using SNP data to observe the SNP distribution at predicted miRNA target sites located on circRNAs. It is suggested that many of these predicted sites are functional sites under selective pressure due to the significant decrease of polymorphisms of circRNAs 61. Majority of circRNAs possess relatively few miRNA binding sites that may not efficiently trap miRNAs, thus failing in exhibiting the expected properties of super sponges 62. Therefore, whether circular miRNA sponges commonly occur and how the circRNA, miRNA, and ceRNA network is used in maintaining homeostasis remain unclear.

In addition to acting as miRNA sponges, circRNAs can also act as protein sponges. CircFOXO3a can interact with senescence associated transcriptional factors in the cytoplasm (e.g.,HIF1α, ID1, or E2F1), trapping these within the cytoplasm and preventing them from translocation to other organelles 63. Additionally, above mentioned blind muscle protein (MBL/MBLN1) in flies (Drosophila melanogaster) and humans can control its own levels. In detail, MBL as a kind of RBP to promote the generation of circMBLs that have multiple MBL binding sites, which in turn captures this protein when MBL is overexpressed 49. Further studies have revealed that the circRNAs corresponding to the formin gene in mice contain a translation initiation site that captures mRNAs to form non-coding linear transcripts, thereby reducing the level of Fmn protein encoded by the parental gene 64. CiRNAs and ElciRNAs localized in the nucleus can be used to cis-regulate the transcription of polymerase II by interacting with pol II 5. For example, circEIF3J and circPAIP2 can bind to U1 snRNPs to enhance the transcription of the parental gene 65. Therefore, we speculate that only exonic circRNAs play a regulatory role in the cytoplasm, whereas the circRNAs containing introns perform effective transcriptional regulation in the nucleus. In addition, studies have found that circRNAs inserted into the internal ribosome entry site (IRES) upstream of the initiation site can artificially synthesize proteins66. This suggests that perhaps some of circRNAs can also be directly converted into proteins. Subsequently, circRNAs were found with coding ability including circ-SHPRH 67, circ-FBXW7 68, Circ-ZNF609 6970, and so on. Furthermore, it is found that the N6-methyladenosine (m6A), the most abundant base modification of RNA, can promote efficient initiation of protein translation from circRNAs in human cells 71. All of these circRNA-derived proteins largely expand the coding landscape of human transcriptome and can be exploited as high specific novel targets for targeted therapies.

An increasing amount of research shows that circRNAs are involved in the pathogenesis of cardiovascular disease 72, neurodegenerative disease 73, and cancer 74, and thus may be utilized as disease biomarkers. For example, a large number of CDR1as is expressed in the brain that contain 60 bp of miR-7 binding sites, which is associated with a variety of disease pathways. It has been confirmed that CDR1as is involved in the regulation of Parkinson's and Alzheimer's disease 757677. At the same time, miR-7 is involved in carcinogenesis and tumor inhibition, and therefore, the regulatory axis of CDR1as/miR-7 is likely to be closely related to the occurrence and development of tumors 78798081. These findings show that circRNAs are closely related to the occurrence of disease and is a potential target for the future diagnosis and treatment of disease.

Although circRNAs are highly stable and has great potential as biomarkers for disease diagnosis, their use in clinical trials and in patient diagnosis remains limited. Considering the great variability in circRNAs between individuals, and even within individuals taken on different days 82, there is necessary to discuss in detail the key indicators for the assessment of the sensitivity, specificity, easy detection, and repeatability of tissue- and disease-associated circRNAs. Studying the influence of potential factors such as organization and blood collection and processing is important to ensure that standardization of reliability and reproducibility in the operational process involves the collection of data regarding conditions, equipment, applications, and sample acquisition, transportation, handling, and storage issues 838485.

It is important to find biomarkers that are highly specific and sensitive for the early diagnosis and staging of diseases to facilitate large-scale population screening because it is difficult to accurately diagnose and predict the occurrence of diseases based on multiple biomarkers 86. In addition, it is necessary to develop reagents and methods for detecting biomarkers with high sensitivity, specificity, and stability such as ELISA kits, mass spectrometry, automatic electrochemiluminescence immunoassays, and blood biomarker detection. There is a need to standardize biomarker detection, analysis of the standardization of pre-analysis factors, unified detection methods, and the use of automated analysis methods to achieve comparability of the data, and these current obstacles can provide the basis for determining the critical value of circRNAs for clinical diagnosis.

Neurodegeneration is a disease that occurs in the brain and spinal cord with the symptoms of neuronal loss. It may be destructive and irreversible for cells that are excessively damaged, because cells in this condition will not regenerate. Senile dementia is also known as Alzheimer's Disease (AD) that belongs to Chronic neurological diseases. The main histopathological traits of AD are Aβ plaques and neurofibrillary tangles (NFTs)，which often occur in the neocortex, hippocampus and other subcortical areas of the brain 87. Recent studies have revealed a potential link between AD and circRNA-associated-ceRNA networks (Figure 3).

The peptide amyloid β (Aβ) that has strong neurotoxicity are derived from the degradation of amyloid precursor protein (APP) through secretase. The cleavage of transmembrane protein APP leads to the extracellular accumulation of Aβ and thus form the Aβ plaques 87. APP has two metabolic pathways: in the normal metabolic process, through α-secretase, APP can produce soluble N-terminal fragment sAPPα, which act as neuroprotection to prevent the formation of Aβ. In another metabolic pathway, if the APP is cleaved by β-secretase, such as BACE1 (b-site app-cleaving enzyme 1), the N-terminal secretory polypeptide fragment sAPPβ will produce, however the C-terminal fragment C99 remains on the membrane. C99 is further degraded by γ-secretase to produce Aβ peptide (Aβ40 and Aβ42), in which Aβ42 will finally form senile plaques through aggregation and deposition 88. For another, the presence of microtubule associated protein tau (MAPT) may induce chronic inflammation and neuronal loss, which mainly result in NFT formation. Protein Tau will become hyperphosphorylated Tau catalyzed by glycogen synthesis kinase 3 (GSK-3). At this point, MAPT loses its ability to bind microtubules and becomes unstable, aggregating into double-stranded helical fibers (PHFs), which then formed the filaments of NFTs. In addition, Aβ can enhance the activity of GSK-3 to induce hyperphosphorylation of Tau, which cooperate with the deposition of Aβ can enhance the cytotoxicity and thus produce neuroinflammation. The imbalance between Tau phosphorylation and dephosphorylation is an early event in NFT formation and AD pathogenesis 89. Figure 4.

Studies have shown that the expression level of ciRS-7 were significantly reduced in hippocampal CA1 region samples of AD patients compared with healthy controls 75. Therefore, it is predicted that the lack of ciRS-7 may lead to decreased expression of selective miR-7 targets, and the expression of Ubiquitin protein ligase A (UBE2A) was reduced through miRNA sponge function 90. UBE2A is a miR-7 target that is essential for the clearance of AD-amyloid peptides. Due to the inhibition after miR-7 binding, UBE2A was downregulated in AD. In addition, ciRS-7 was found to promote the degradation of APP and BACE1 in a nuclear factor-κB (NF-κB)-dependent manner in SH-SY5Y cells. The network of ciRS-7-miR-7-UBE2A suggests that ciRS-7 can act as an effective therapeutic target in AD. Through backsplicing, MAPT can conjugate exon12-10 to produce circular RNA (cir12-10), which is located in the cytoplasm and contains ORF that encodes the Tau fragment 91. Studies have shown that there is a high probability to cause frontotemporal dementia (FTD) after interfering mutation to exon 10 of Tau 92. Moreover, exon 10 usage as well as cdc2-like kinase (CLK2) splicing isoforms are changed in AD 93.

In order to explore the relationship between circRNA-associated-ceRNA and AD, Zhang et al. 94 used senescence-accelerated mouse resistant 1 (SAMR1) as the control to perform deep RNA-seq analysis on the brain of senescence-accelerated mouse prone 8 (SAMP8) model. They found 235 significantly dysregulated circRNA transcripts, 30 significantly dysregulated miRNAs, and 1202 significantly dysregulated mRNAs, then constructed comprehensive circRNA-associated-ceRNA networks to conduct GO analysis. It was found that circRNA-associated-ceRNA networks can affect AD from various aspects, such as axon terminal and synapses. After further screening, it was determined that this network was involved in regulating the clearance of Aβ and the function of myelin in AD model mice.

We thank ACCDON for providing linguistic assistance during the preparation of this manuscript.

Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX20_3071) and The National Natural Science Foundation of China (31861143051) supported this study.