CircRNAs can arise from different regions of gene loci in eukaryotes, such as from exons (called exonic circRNA, or ecircRNA), from introns (called intronic circRNA, or ciRNA) or from both exons and introns (called exon–intron circRNA, or ElciRNA) (Li Z. et al., 2015; Zhang Y. et al., 2016) (Figure 1A). Nevertheless, most circRNAs are derived from exons of protein-coding genes through a non-canonical splicing process called ‘back-splicing,’ in which a downstream splice donor joins an upstream splice acceptor leading to exon circularization (Jeck et al., 2013; Salzman, 2016). The canonical splice machinery, however, is required for exon circularization as was shown by mutagenesis experiments (Chen and Yang, 2015; Starke et al., 2015). Since mRNAs that are generated from linear splicing is also mediated by the canonical splicesome, how circRNAs and mRNAs can both be generated from the same pre-mRNA remains to be an important question (Hamid and Makeyev, 2014). Similar to mRNA splicing, the presence of canonical splice sites of bracketing exons can determine, to a large degree, the circularization rates (Barrett et al., 2015; Wang and Wang, 2015; Zhang Y. et al., 2016). On the other hand, functional studies have elucidated that, during circular vs. linear splicing, different splice sites can be activated and the biogenesis of circRNAs and mRNAs may compete with each other (Tay et al., 2014; Barrett and Salzman, 2016). Moreover, there are other differences between canonical spliceosome mediated circRNA and mRNA biogenesis, which might largely determine which molecule will be predominantly generated from the pre-mRNA, leading to distinct expression patterns of circRNA and mRNA (Ashwal-Fluss et al., 2014; Starke et al., 2015). For example, the efficiency of back-splicing is lower than those of the corresponding linear counterparts, perhaps because spliceosomes are inefficiently aggregated at back-splicing sites at the steady-state of circRNA production (Ashwal-Fluss et al., 2014; Guo et al., 2014). Although the exact details on how the splicesome carry out exon circularization is not known, the canonical splice signals and canonical spliceosomal machinery are important in the processing of circRNAs biogenesis (Chen and Yang, 2015; Petkovic and Müller, 2015).

In addition to the splicing machinery, both cis-regulatory elements and trans-acting factors are needed to bring the two splice sites into close proximity to facilitate back-splicing (Jeck and Sharpless, 2014; Kramer et al., 2015). For example, reverse and complementary sequences resided in flanking introns of circulated exons are able to form RNA duplexes, which enhance the efficiency of back-splicing (Ivanov et al., 2015; Zhang et al., 2014). RNA duplexes can form by repetitive sequences such as Alu elements (Kim et al., 2004; Jeck et al., 2013), as well as from non-repetitive but complementary sequences (Zhang et al., 2014). Longer complementary sequences in general have stronger RNA pairing strengths and enhanced circRNA biogenesis than shorter sequences; however, sequences as short as 30 nts also possess the ability to stimulate circRNA production (Liang and Wilusz, 2014). In addition, distinct ways of RNA pairing can lead to generation of different circRNAs. For example, repetitive sequences such as Alu elements can form different combinations of Alu pairs, which results in production of diverse circRNAs from a single gene – a process known as alternative circularization (Jeck et al., 2013) (Figure 1A). The types and expression levels of circRNAs are much higher in humans than in mice, and the increased circRNA complexity during species evolution is correlated with an increased number and RNA-pairing capability of those complementary sequences (Dong et al., 2016). Importantly, in addition to RNA base-pairing formed across flanking introns, RNA base-pairing can also form within an individual intron in linear RNA production (Hossain et al., 2011; Zhang et al., 2014). Therefore, there are competitions between non-canonical splicing and canonical splicing, based on choice of RNA base-pairing in flanking introns or within an individual intron (Jeck et al., 2013; Hamid and Makeyev, 2014).

Trans-acting factors such as RNA-binding proteins (RBPs) have been reported to be capable of regulating circRNA biogenesis (Castello et al., 2012; Conn et al., 2015; Dudekula et al., 2016). Three factors, muscleblind (MBL), Adenosine deaminae 1 acting on RNA (ADAR1) and quaking (QKI) have been shown so far to regulate circRNA production by different mechanisms (Ashwal-Fluss et al., 2014; Ivanov et al., 2015; Conn et al., 2015). MBL is a conserved regulator of RNA splicing. The second exon of MBL can be circularized to produce a circular transcript (called circMbl) and circMbl and its flanking introns have abundant MBL binding sites (Ashwal-Fluss et al., 2014). These sites are required for circMbl biogenesis and therefore, MBL protein can control the production of the circRNA of its own (Ashwal-Fluss et al., 2014) (Figure 1B). ADAR1 expression is negatively correlated with production of certain circRNAs (Szabo et al., 2015). ADAR1 converts Adenosine-to-Inosine (A-to-I), which predominantly occurs in double-stranded Alu repeats (Athanasiadis et al., 2004). Therefore, it is thought that ADAR1 inhibits circRNA biosynthesis by breaking the RNA stem of RNA-RNA interactions that form during RNA base-pairing (Ivanov et al., 2015) (Figure 1C). QKI regulates the production of many circRNAs and the presence of QKI binding sites in the intron of genes that are normally spliced in a linear fashion is sufficient to trigger circRNA production (Conn et al., 2015) (Figure 1D).

It was reported that circRNA production occurs post-transcriptionally, since a 3′ end processing signal is necessary for back-splicing and the 3′ polyadenylation is required for splicing of some introns (Liang and Wilusz, 2014). Nevertheless, other studies point out that circRNA biogenesis can occur both co-transcriptionally and post-transcriptionally, and the strength of base-pairing of complementary sequences in the flanking intronic repeats determines whether back-splicing occurs co- or post-transcriptionally (Wilusz and Sharp, 2013; Ashwal-Fluss et al., 2014; Kramer et al., 2015). While back-splicing mediated by minimal intronic repeats [<40 nucleotides (nt)] requires 3′ end processing and occur post-transcriptionally, long flanking repeats (∼400 nt) are able to produce circRNAs co-transcriptionally (Kramer et al., 2015).

Proliferation and differentiation of neurons from stem cell precursors and progenitors are landmarks of neural development. NcRNAs such as miRNAs and lncRNAs have been shown to be important in these processes (Rani et al., 2016). Similarly, circRNAs have also become increasingly recognized to play crucial roles. During neural development, the expression pattern of neuronal circRNAs show conservation between mouse, pig, and human, and the highly expressed circRNAs is more conserved than average, indicating that these highly expressed circRNAs are potentially more important (Westholm et al., 2014; Rybak-Wolf et al., 2015) than others. In cell culture models of neural differentiation, the expression of the majority of circRNAs was shown to be significantly up-regulated (Rybak-Wolf et al., 2015). The host genes of these upregulated circRNAs are known to play essential roles in neuronal activities such as dendritic mRNA transport and synaptic membrane exocytosis (Ng et al., 2013; Heraud-Farlow and Kiebler, 2014; Jung et al., 2015). In addition, during porcine brain development, different anatomic regions were found to have different circRNA expression levels, with the minimum amount in the brain stem and the highest amount in the cortex and cerebellum (Venø et al., 2015). Moreover, at different stages of neural differentiation, there are distinct levels and types of up-regulated circRNAs (Rybak-Wolf et al., 2015). Taken together, circRNA expression is spatio-temporally regulated during brain development, suggesting that they may play an essential role in the process of neural proliferation and differentiation (Rybak-Wolf et al., 2015; Dang et al., 2016).

How do circRNAs control progenitor cell proliferation and neural differentiation? The origin of circRNAs may reflect their functions to some extent, assuming circRNAs have similar biological function to their host genes (Rybak-Wolf et al., 2015; Venø et al., 2015). In a study investigating spatio-temporal expression patterns of circRNAs during porcine embryonic brain development, three host gene associated pathways, Wnt signaling, axon guidance and the TGFβ signaling pathway, were found to be overrepresented at a developmental stage corresponding to major neurogenesis (Venø et al., 2015). These over-represented signaling pathways play a major role in neural stem cell proliferation, migration and differentiation (Rosso and Inestrosa, 2013; Mattar and Cayouette, 2014). How circRNAs impact these pathways remains to be an important question in understanding their contributions to neural proliferation and differentiation.

The expression change, localization and origin of circRNAs signify the importance of circRNAs in synaptogenesis to a large degree. During porcine and mouse brain maturation, the amount of circRNAs produced was found to be significantly increased, especially when synaptic connections and neural network forms (Venø et al., 2015; You et al., 2015). For example, in mice, the expression pattern of most circRNAs changes abruptly at postnatal day (P)10 when synaptogenesis begins (You et al., 2015). Moreover, circRNAs originated from synapse-related genes are consistently up-regulated during the establishment of mature neural networks (You et al., 2015). In addition, consistent with a fourfold increase of synaptic density in the human cortex compared to mice (Herculano-Houzel, 2009), higher levels of circRNAs were detected in human brains compared to mouse brains. Moreover, instead of being equally distributed in all neuronal compartments, circRNAs are highly enriched in synaptoneurosomes (You et al., 2015). Although both circRNAs and mRNAs can be detected in the cell body and dendrites, they are not co-localized. For example, circStau2a is mainly presented in synapses; however, the linear Stau2a mRNA transcripts primarily locate in the cytoplasm (Venø et al., 2015). Furthermore, neural circRNAs are disproportionally originated from host genes encoding proteins with synaptic functions. Genes related to synapse, synapse part, presynaptic active zone, presynaptic membrane and postsynaptic density can give rise to abundant circRNAs (Venø et al., 2015; You et al., 2015). For example, circHomer1_a, a circRNA originated from Homer1 that encodes a key protein in postsynaptic density regulation (Meyer et al., 2014), has the highest expression level during synaptic plasticity (You et al., 2015).

The function of circRNAs in mammalian brain remains elusive. Because of their primary localization to synapses, it has been hypothesized that circRNAs may function as topologically complex platforms which recruit ribonucleoprotein (RNP) granules and transport required RNAs and proteins (Rybak-Wolf et al., 2015) that are needed for synaptogenesis when synaptic connections forms (You et al., 2015). Moreover, due to their great stability and specificity, circRNAs might be used as synaptic tags to keep a molecular memory (Rybak-Wolf et al., 2015). Another possible function for circRNAs is to serve as miRNA sponges (Hansen et al., 2013). Certain circRNAs have been shown to contain multiple binding sites for core regions of miRNAs (Hansen et al., 2013; Zheng et al., 2016). It is possible that circRNAs can serve as competitive inhibitors for miRNA activities by inhibiting binding of miRNAs to their target transcripts. Multiple studies have shown that miRNAs play an important role in synapse formation, stabilization and plasticity (Rajasethupathy et al., 2009; McNeill and Van Vactor, 2012). Therefore, circRNAs may indirectly control synaptogenesis through regulating activities of synapse-related miRNAs and the expression of miRNA targeted mRNAs (Hansen et al., 2013; Zheng et al., 2016).

Several circRNAs were shown to function as miRNA sponges that sequester miRNAs and affect their interaction with downstream mRNAs (Floris et al., 2017; Piwecka et al., 2017) and an abnormal circRNA-miRNA-mRNA system has been indicated in the pathogenesis of CNS diseases (Li et al., 2017; Zhou et al., 2017). For example, Alzheimer’s disease (AD), a progressive neurodegenerative disorder, is caused by damage to synapses, neuronal cell bodies and axons due to the accumulation of pathological amyloid-β or hyper-phosphorylated Tau in the brain (Cohen et al., 2015). miRNA-7 is abundant in the brain and can directly regulate the expression of AD-related genes such Ubiquitin protein ligase A, which mediates the clearance of amyloid peptides in AD (Zhao et al., 2016). A circRNA, ciRS-7 (circRNA for miRNA-7), also termed as cerebellar degeneration-related protein 1 antisense RNA (Cdr1as), contains more than 70 selectively conserved anti-miRNA-7 sequences, thus acting as competing endogenous sponges to efficiently quench normal miRNA-7 activities (Hansen et al., 2013). miRNA-7 and ciRS-7 were found to be co-expressed in cultured primary murine neurons and in mouse brain sections (Hansen et al., 2013). In addition, expression of ciRS-7 is significantly reduced in hippocampus of AD patients (Lukiw, 2013), suggesting an abnormal miRNA-circRNA system may underline the disease mechanism of AD (Figure 2). Because of the importance of miRNA-7 in other diseases such as cancer and Parkinson’s disease (Junn et al., 2009; Saydam et al., 2011; Liu et al., 2014), it is possible that ciRS-7 may exert a regulatory role in brain cancer and other neurodegenerative diseases as well. Deleting the ciRS-7 locus in mice leads to dysfunctional synaptic transmission and abnormal neuropsychiatric-like behavior in mice (Piwecka et al., 2017). Other than miR-7, ciRS-7 also has a binding site to miR-671. Interestingly, the expression of miR-7 and miR-671 were both deregulated in all brain regions in ciRS-7 deficient mice; however, the direction of changes was opposite. It was thought that the binding site on ciRS-7 is completely complimentary to miR-671, and the interaction of these two molecules could lead to AGO mediated ciRS-7 slicing and miR-671 degradation. In contrast, the binding sites for miR-7 are partially complimentary which cannot trigger ciRS-7 cleavage, leading to stabilized miR-7. Therefore, it is possible that circRNAs can serve as a platform to store and transport certain miRNAs. Similarly, another circRNA generated from the gene Sry, has 16 putative binding sites to miR-138 and can attenuate activity of miR-138 in culture (Hansen et al., 2013). Expression of miR-138 was increased in cell culture models of AD suggesting the role of circRNA Sry in the disease pathology (Wang et al., 2015).

Other functions of circRNA in the CNS may be indicated by their host genes. For example, expression and complexity of circRNAs peak at embryonic day (E)60 during porcine brain development when dramatic morphological changes occur in the cortex. The top three highly expressed circRNAs at E60 are circCSPP1, circHDAC2 and circRIMS2 (Venø et al., 2015). The host genes of these circRNAs are crucial for CNS development and synaptic formation (Akizu et al., 2014). For instance, CSPP1 plays a part in the normal physiological functions of primary cilia and mutations in this gene lead to a developmental brain disorder called Joubert syndrome (Tuz et al., 2014). HDAC2 is important for neuronal survival and forebrain-specific deletion of this gene can lead to excessive repetitive behaviors in mice (Mahgoub et al., 2016). RIMS2 are involved in synaptic vesicles release and presynaptic plasticity and deletion of this gene can result in decreased activation of the presynaptic Ca²⁺-channel (Kaeser et al., 2012). CircRNAs might also play a role in nerve damage. Biosynthesis of circHomer1 may compete with synthesis of Homer1b/c mRNA (You et al., 2015). In an in vitro model inducing synaptic plasticity using primary hippocampal neurons, circHomer1a expression was remarkably increased, which could inhibit potential overexpression of Homer1b/c that is detrimental to synaptic plasticity (You et al., 2015). Down-regulation of Homer1b/c was also shown to improve neuronal survival in an in vitro traumatic nerve injury model (Fei et al., 2014). Therefore, circHomer1a may have great potential as a therapeutic target on repair and regeneration of injured nerves.

Due to lack of 3′ termini, circRNAs are generally endowed with a strong resistance to exonuclease RNase R, thus being more stable than linear RNAs (Memczak et al., 2013; Enuka et al., 2016). The great stability and abundance of circRNAs make them as great candidates for molecular biomarkers in neurological diseases. Recently, several studies have reported the use of circRNAs as potential diagnostic biomarkers for atherosclerotic vascular disease, pre-diabetes and varieties of tumors including gastric cancer, hepatocellular carcinoma, lung cancer, colon carcinoma, laryngeal cancer and leukemia, et al (Ren et al., 2016; Xie et al., 2016; Zhao Z. et al., 2017). For example, circPVT1 could act as independent prognostic marker in gastric cancer (Chen J. et al., 2017). The expression of circPVT1 was up-regulated in gastric cancer tissues and it could promote cell proliferation through serving as a sponge for miR-125 family members (Chen J. et al., 2017). Also, similarly, circular RNA hsa_circ_0000190 was down-regulated in both human gastric cancer tissues and plasma samples and had higher sensitivity and specificity compared with traditional gastric cancer biomarkers, carcinoembryonic antigen (CEA) and CA19-9 (Chen S. et al., 2017). So far, multiple circRNAs have been explored as a biomarker for a variety of diseases especially in cancer (Meng et al., 2017).

Neural circRNAs show high conservation during evolution and are abundantly expressed in the brain (Venø et al., 2015). In Drosophila, circRNA expression increased substantially during CNS aging and could be regarded as a class of aging biomarkers (Westholm et al., 2014). Accumulation of circRNAs in aging brains has also been described in mice (Gruner et al., 2016) suggesting it could be a general theme for other species including humans. Therefore, although it has not been reported in CNS diseases, because circRNAs are stable, specific, conserved and sensitive, distinct circRNA profiles may be used to evaluate the physiologic and pathological circumstances in the brain (Memczak et al., 2013; Rybak-Wolf et al., 2015; Song et al., 2016). On the other hand, biopsies for neuronal tissues may be difficult to access. However, the pathological information of CNS tissues could be acquired through detecting the levels of certain components in the plasma and cerebrospinal fluid, etc. that are easily accessible (Luo et al., 2013). Therefore, detecting circRNAs from these sources can be possible and using circRNAs as biomarkers for neurological diseases should be further investigated.

Furthermore, in CNS diseases, the blood–brain barrier (BBB) can be compromised and small molecules including circRNAs and exosomes can be transported out of BBB (Grapp et al., 2013). Exosomes are membranous small vesicles that can transfer circRNA, miRNA, mRNA and protein into extracellular fluid or cells (Cocucci and Meldolesi, 2015). Recent studies revealed that, circRNAs contained in exosomes (exo-circRNAs) were about twice more than those in their producer cells (Li Y. et al., 2015). CircRNAs tend to be more accessible to exosomes than linear RNAs, and the ratio of exo-circRNAs to their linear transcripts in exosomes is about six times higher than that in their producer cells (Li Y. et al., 2015) (Figure 3). In addition, exo-circRNAs showed high stability, perhaps because of their protein partners and the protection of exosomes (Li Y. et al., 2015). Taking together, the specific circRNAs profiles carried in exosomes might serve as potential biomarkers in CNS disorders (Lu and Xu, 2016).