Protein-coding genes

The ceRNA hypothesis represents by itself a revolution in gene expression regulation, but one of the most intriguing aspects of this new perspective is the additional function conferred to mRNAs. Before the ceRNA hypothesis, mRNAs were considered exclusively as the manual of instruction for protein synthesis, carrying the information for the aminoacidic sequence of a functional protein. In light of this new complex scenario of competing RRIs, mRNAs became regulators able to modulate the gene expression of other mRNAs at the post-transcriptional level. That means an mRNA may exert its function even if it is not translated into a protein. Intriguingly, the non-coding function of the mRNA may be opposite to the function performed by the corresponding protein [25]. In addition, new effects must be taken into consideration when an mRNA is strongly up- or downregulated: dramatically increased levels of an mRNA would massively sequester miRNAs, thus freeing other transcripts from miRNA-mediated negative regulation; on the contrary, a strong reduction of mRNA levels would make a high number of miRNA molecules available for interaction with other mRNAs [20].

Once again, the researchers have to face a very intricate landscape. The human genome contains about 20,000 protein-coding genes [9], but, because of alternative splicing patterns, the total number of mRNAs potentially transcribed is higher. As already discussed, each one of these mRNAs, expressed in a given cell type in a specific moment, includes within its sequence numerous MREs for different miRNAs, which in turn may be expressed or not expressed in the same cell or moment or specific developmental stage. The potential co-expression (or lack of co-expression) of mRNAs and miRNAs adds a new layer of complexity and tunability to the fine mechanism of regulation represented by ceRNA networks, making the system flexible and able to quickly answer to new stimuli.

The non-coding function of mRNAs was described by Tay et al. [57] in 2011 by investigating the expression of the well-known tumor suppressor gene phosphatase and tensin homolog (PTEN). The authors applied the mutually targeted MRE enrichment (MuTaME) approach, a combination of computational analysis and experimental validation, to identify mRNAs sharing MREs with PTEN mRNA; in particular, the ceRNA transcripts included at least seven out of ten MREs present in the 3’-UTR of PTEN. This approach led to the identification of several ceRNAs, among which CCR4-NOT transcription complex subunit 6 like (CNOT6L) and VAMP associated protein A (VAPA) were validated as PTEN regulators. Indeed, siRNA-mediated downregulation of both ceRNAs induced a reduction of PTEN protein expression; a less significant reduction of PTEN mRNA levels was also observed. It was also demonstrated that this decreased expression of PTEN was mediated by miRNAs by using a dicer 1, ribonuclease III (DICER1)-mutant cell line, in which miRNA processing was impaired and PTEN downregulation was consequently reduced. The effective instauration of this ceRNA circuit in vivo is sustained by the correlation of expression among PTEN and its putative ceRNAs observed in vivo. Importantly, this ceRNA mechanism resulted in the regulation of cell proliferation, thus impacting cancer growth [57]. The same group reported other ceRNAs for PTEN identified by using a similar approach in melanoma [58]. ZEB2, involved in the epithelial to mesenchymal transition (EMT), was shown to regulate PTEN expression acting as ceRNA thanks to its 3’-UTR, able to sequester and suppress the function of miRNAs targeting PTEN. This represents an interesting example of an mRNA showing two opposite coding and non-coding functions: as a protein, ZEB2 activates EMT and thus fosters metastatization, while as an mRNA it sustains the expression of a tumor suppressor gene, namely PTEN, through ceRNA interactions [58]. In the same period, similar evidence was published by Sumazin et al. [59], who also focused on PTEN regulation of expression in glioblastoma. In this study, a different approach named Hermes, based on the modulator inference by network dynamics (MINDy) algorithm, was applied to identify potential ceRNAs of PTEN. Results again showed that the 3’-UTR of these ceRNAs was implicated in the regulative mechanism [59]. Another interesting example is represented by tyrosinase related protein 1 (TYRP1) mRNA-mediated sponging of miR-16 in melanoma, which occurs with non-canonical binding [60]. miR-16 sequestration fostered tumor progression by de-repressing RAB17, member RAS oncogene family (RAB17). A therapeutic approach using ASOs to impede the interaction between miR-16 and TYRP1 by masking its MRE led to increased RAB17 levels and apoptotic rate, also decreasing cell proliferation [60]. Collectively, this evidence suggests that competing interactions play a crucial role in carcinogenesis and may be considered targets for new therapeutic approaches.

As discussed, the most important region of the mRNAs driving miRNA-mediated regulation is the 3’-UTR, which acts in cis by regulating the stability of the transcript it belongs to. Given the non-coding function of mRNAs within ceRNA networks, the 3’-UTR becomes a trans-acting regulator by titrating miRNA availability and indirectly regulating the expression of other mRNAs sharing MREs for the same miRNAs [25]. This function in trans is confirmed by the reports discussed above in addition to many others where the 3’-UTR of a given ceRNA is ectopically expressed alone, independently from the mRNA to which it belongs. In particular, Lee et al. [61] investigated the role of the 3’-UTR of versican (VCAN) in regulating miRNA function. VCAN is a component of the extracellular matrix (ECM) involved in cancer progression through the regulation of cell adhesion and migration, angiogenesis, and other processes [62]. Ectopic expression of this 3’-UTR in a breast cancer cell line reduced cell proliferation in vitro and tumor growth in vivo by sponging miR-144, miR-199a-3p, and miR-136, miRNAs also targeting the tumor suppressor genes RB transcriptional corepressor 1 (RB1) and PTEN [61]. The same group focused on the role of VCAN 3’-UTR in the onset and progression of hepatocellular carcinoma (HCC) [63]. Enforced expression of the 3’-UTR of VCAN led to HCC onset in vivo; in vitro, VCAN 3’-UTR induced increased proliferation, migration, and invasion, while decreasing apoptosis rate. In vitro assays showed that this tumor-promoting effect was due to the miRNA sponging performed by VCAN 3’-UTR on miR-133a, miR-199a*, miR-144, and miR-431; this sponging function protected from miRNA-mediated silencing VCAN mRNA itself, but also the transcripts coding fibronectin 1 (FN1) and CD34 molecule (CD34) [63]. The same group also investigated the role of the 3’-UTR of [CD44 molecule (Indian blood group)] (CD44) in breast cancer [64, 65]. Overexpression of CD44 3’-UTR reduced cell proliferation and promoted apoptosis by sequestering miR-216a, miR-330, and miR-608, thus increasing protein levels of CD44 itself and cell division cycle 42 (CDC42), targeted by the same miRNAs [64]. Shortly after, expression of CD44 3’-UTR was reported to increase cell migration and invasion in vitro and promote metastasis in vivo by miRNA-mediated de-repression of collagen type I alpha 1 chain (COL1A1) and FN1, achieved by sequestering miR-512-3p, miR-328, miR-491 and miR-671 [65]. Considering this evidence all together, FN1 may represent a link between the ceRNA networks centered on CD44 and VCAN, regulating the expression of genes crucial for cell proliferation such as CDC42, PTEN, and RB1, all participating in tumorigenesis [24].

The discovery of this sponge function performed by 3’-UTRs becomes very important in carcinogenesis. Indeed, the phenomenon of 3’-UTR shortening has been observed in several cancer models [66]. Mayr and Bartel [66] reported that aberrant production of several oncogenic proteins may occur in cancer also in absence of genetic alterations. The reason for increased protein levels often lies in the presence of a shorter 3’-UTR, because of the phenomenon of alternative cleavage and polyadenylation (APA), which abolishes several MREs. These shorter 3’-UTRs promote carcinogenesis, as demonstrated for the proto-oncogene insulin like growth factor 2 mRNA binding protein 1 (IGF2BP1). Similarly, different splicing variants frequently show distinct 3’-UTRs; that means the regulation of alternative splicing patterns may affect competing interactions and modulate phenotype [67].

Transcribed pseudogenes

Pseudogenes are copies of a parental gene that, unlike the latter, do not undergo translation. The comparison between pseudogenes and their coding counterparts shows that the abolishment of the coding ability of these gene copies is due to the loss of control regions or the accumulation of several genetic alterations acquired over time, such as the formation of premature stop codons or frameshift insertion/deletion mutations. Indeed, pseudogenes were considered “genomic fossils” [20, 39]. According to their generation, pseudogenes can be classified into: i) unprocessed pseudogenes, originated from a duplication of protein-coding genes, usually maintain 5’-UTR, 3’-UTR, and introns within their sequence; ii) processed pseudogenes, generated by retrotransposition of an mRNA transcribed from a protein-coding gene, do not include introns but may maintain 5’-UTR and 3’-UTR; iii) unitary pseudogenes, deriving from unduplicated protein-coding genes which accumulated mutations and lost their protein-coding capability, usually do not show protein-coding counterparts in the same genome but have protein-coding orthologs in other species. It was estimated that the human genome overall includes more than 14,000 pseudogenes [68].

Albeit pseudogenes have been long considered part of the so-called “junk DNA”, lacking function, next-generation sequencing experiments showed that pseudogenes are often conserved and actively transcribed (and in some cases also translated into proteins [69]), suggesting that they do exert functions and are subject to selective pressure [39, 70]. Indeed, a function for a few pseudogenes has been recently reported in the context of ceRNA networks, where they act as miRNA sponges regulating the expression of the parental gene. This function is due to the high number of MREs shared with the parental gene thanks to the sharing of the same 3’-UTR and the high sequence homology [19].

The first pseudogene reported as ceRNA is phosphatase and tensin homolog pseudogene 1 (PTENP1), a processed pseudogene able to regulate the expression of its parental gene PTEN [71]. Given the involvement of PTEN in carcinogenesis as a tumor suppressor, PTENP1 plays a crucial role in several cancer models; indeed, PTEN is frequently lost or mutationally inactivated in cancer cells. PTENP1 and PTEN show very high sequence complementarity, with only 18 mismatches, but a mutation abolishing the AUG start codon prevents the translation of the pseudogene. PTENP1 shows a 3’-UTR shorter compared to PTEN, with a difference of about 1 kilobase (kb); nevertheless, MREs for miR-17, miR-21, miR-214, miR-19, and miR-26 families are shared between the two transcripts. It has been reported that PTENP1 sponged these miRNAs, thus protecting PTEN from miRNA-mediated repression at both mRNA and protein levels. In vitro assays showed that silencing of PTENP1 promoted cell proliferation, supporting the involvement of this pseudogene in carcinogenesis as a tumor suppressor. Accordingly, PTENP1 is also frequently lost in cancer and shows a correlation of expression with PTEN, suggesting that the pseudogene regulates the expression of the parental gene also in vivo [71]. PTENP1 represents a particular example in ceRNA networks because it also encodes a lncRNA participating in PTEN regulation (see below [72]). The function of regulator of parental gene expression was also reported for other pseudogenes, including the oncogenes KRAS proto-oncogene, GTPase pseudogene 1 (KRASP1) for KRAS proto-oncogene, GTPase (KRAS) [71] and catenin alpha 1 pseudogene 1 (CTNNAP1) for catenin alpha 1 (CTNNA) [73], or the tumor suppressors TUSC2 pseudogene 1 (TUSC2P1) for tumor suppressor 2, mitochondrial calcium regulator (TUSC2) [74] and integrator complex subunit 6 pseudogene 1 (INTS6P1) for integrator complex subunit 6 (INTS6) [75]. All these pseudogenes originated from parental genes involved in carcinogenesis, supporting the hypothesis that competing interactions are crucial for cancer onset and progression.

lncRNAs

lncRNAs are non-coding RNA molecules characterized by a length comprised between 200 nts and kbs; they are similar to mRNAs concerning biogenesis, being often transcribed by RNA polymerase II, capped, polyadenylated, and spliced [76]. The total number of genes encoding lncRNAs in the human genome has been estimated between 10,000 and 100,000 [77], representing a conspicuous portion of it. Similar to pseudogenes, it has been proposed that lncRNA genes derived from duplication events occurring at both DNA and RNA levels, insertion of transposable elements, or mutational inactivation of protein-coding genes. lncRNAs are poorly conserved during evolution; accordingly, more than 80% of lncRNA families originated in primates [78]. Typically, lncRNAs are mostly localized in the nucleus and poorly expressed compared to mRNAs [79]. For these reasons, lncRNAs have been long considered non-functional, but the observation of expression and localization patterns related to specific tissues or developmental stages, at least for a fraction of them, has suggested that they do exert functions within cells [77, 78]. To date, the exact function or mechanism of action of only a few lncRNAs has been demonstrated, showing that a high heterogeneity in functions is characteristic of this class of RNA molecules. Described functions include epigenetic regulation (such as chromatin remodeling), RNA or protein localization, regulation of transcriptional and post-transcriptional events (including mRNA stability, splicing, and editing), and structural functions (scaffolding) [77, 80]. In the last decade, one of the most investigated functions performed by lncRNAs has been represented by miRNA sponging, whose involvement both in neoplastic and non-neoplastic diseases has been demonstrated [81–83].

One of the first reports describing a ceRNA function for lncRNAs was published in 2010 [84]. Highly up-regulated in liver cancer (HULC) increased levels in HCC caused the decreased availability of miR-372, resulting in the upregulation of one of its targets, namely protein kinase cAMP-activated catalytic subunit beta (PRKACB). PRKACB in turn regulated the phosphorylation of cAMP responsive element binding protein 1 (CREB1), responsible for HULC transcription. Therefore, this lncRNA regulates its expression through miRNA sponging creating an auto-regulatory loop [84]. Shortly after, Cesana et al. [85] showed that the lncRNA long intergenic non-protein coding RNA, muscle differentiation 1 (LINCMD1) participates in muscle differentiation both in mouse and human. By sponging miR-133 and miR-135 in myoblasts, LINCMD1 regulated the levels of mastermind like transcriptional coactivator 1 (MAML1) and myocyte enhancer factor 2C (MEF2C), two transcription factors (TFs) promoting the expression of muscle-specific genes [85]. HOX transcript antisense RNA (HOTAIR) is a lncRNA acting as an oncogene in several cancer models [86]. In gastric cancer, this oncogenic function may be in part exerted by the sequestration of miR-331-3p, resulting in the modulation of its target erb-b2 receptor tyrosine kinase 2 (HER2/ERBB2) [87]. To date, the list of reports about lncRNAs acting as miRNA sponges is constantly growing [34], supporting the fundamental role of such a mechanism within cells.

Besides sponging miRNAs, lncRNAs are able to regulate gene expression through other mechanisms. In particular, a specific subclass of lncRNAs includes NATs, involved in the regulation of the sense protein-coding transcript. As an example, ZEB2-AS1 regulates the splicing of the sense transcript ZEB2 in epithelial cells [30]. ZEB2 acts as a transcriptional repressor of cadherin 1 (CDH1) and is upregulated when EMT is activated. One of the TFs triggering EMT is snail family transcriptional repressor 1 (SNAI1); intriguingly, Beltran et al. [30] showed that SNAI1 did not promote ZEB2 transcription, while it indirectly modulated ZEB2 splicing through its NAT ZEB2-AS1, expressed under the control of an alternative SNAI1 promoter. ZEB2-AS1 interacted by sequence complementarity with ZEB2 mRNA, thus preventing the binding of the spliceosome and the elimination of an intron including an internal ribosome entry site (IRES), responsible for ZEB2 translation. This example shows that RRIs not involving miRNAs are also crucial for gene expression regulation and modulation of biological processes associated with cell physiology and pathology. However, NATs may also indirectly interfere with miRNA function, supporting again the high complexity and heterogeneity of ceRNA interactions. BACE1-AS is the NAT of BACE1, the enzyme producing the amyloid-β peptide whose accumulation leads to Alzheimer’s disease [88]. BACE1-AS induced upregulation of BACE1 protein by interacting with its mRNA and masking the MRE for miR-485-5p [29]. Similar results were reported in mouse myoblasts for Sirt1-AS, sponging miR-34a and protecting Sirt1 mRNA from miRNA-mediated repression [28]. In these examples, the direct interaction between sense and antisense transcripts, showing intrinsic sequence complementarity, acts as a “protector” from miRNA-mediated repression for the sense mRNA when MREs are located within a region involved in the interaction with the antisense and are thus impeded from miRNA binding, despite the miRNA is not involved in competing interactions.

As previously mentioned, the locus of the pseudogene PTENP1 also encodes a lncRNA, the NAT PTENP1 antisense RNA (PTENP1-AS), acting as a regulator of PTEN expression [72]. In particular, two isoforms of the lncRNA are transcribed, exerting different functions. The α isoform directly regulated PTEN expression acting in trans: by interacting with the promoter of PTEN, this isoform recruited on it DNA methyltransferase 3 alpha (DNMT3A) and enhancer of zeste 2 polycomb repressive complex 2 subunit (EZH2), hence inducing chromatin remodeling. On the contrary, the β isoform directly bound the sense transcript PTENP1, promoting its stability. Indeed, the lack of the polyadenylated tail reported for the PTENP1 transcript determined its instability. That means the β isoform indirectly regulated PTEN expression by stabilizing the PTENP1 transcript, thus increasing its levels and, consequently, reducing the availability of miRNAs able to repress PTEN expression [72]. This regulatory mechanism is extremely complex and fascinating, and the ceRNA network surrounding PTEN is a good illustration of how multiple ceRNA interactions may co-occur.

Another peculiar example concerning lncRNA-mRNA direct interactions reported in 2011 induced mRNA degradation [89]. The RNA-binding protein (RBP) Staufen double-stranded RNA binding protein 1 (STAU1) was reported to recognize double-stranded RNA, leading them to degradation, through a STAU1-binding site included within the 3’-UTR of the target mRNAs, in particular in a 19-base-pair stem [90]. Gong and Maquat [89] reported that not all STAU1-targeted mRNAs presented this stem structure within their 3’-UTR; on the contrary, the double-stranded RNA structure inducing STAU1 binding may be created by the imperfect interaction between an Alu element included within the mRNA 3’-UTR and a cytoplasmic polyadenylated lncRNA also including an Alu element, guiding the interaction with the mRNA.

circRNAs

circRNAs represent an unusual class of lncRNAs, characterized by a circular structure resulting from the covalent binding of the two extremities. circRNAs arise from transcripts coded by protein-coding genes which undergo a particular form of alternative splicing, named backsplicing, where the 5’-site of an upstream exon is bound with the 3’-end of the same or a downstream exon; accordingly, circRNAs are transcribed by RNA polymerase II, but they do not show 5’-cap or 3’-polyadenilation [91, 92]. Because of their structure, circRNAs are resistant to exonucleases and show higher stability compared to linear transcript [92]. circRNAs are abundantly expressed in brain, where they participate in the physiology of the central nervous system [93], albeit being ubiquitously expressed in human tissues [94]. Generally, circRNAs are mostly cytoplasmic and evolutionarily conserved [92, 93] and are transcribed also by viruses, such as SARS-CoV-2 [95]. According to the portions of protein-coding genes included within their sequences, circRNAs can be classified into: exonic circular RNAs (ecircRNAs), circular intronic RNAs (ciRNAs), and exon–intron circRNA (EIciRNA). Recent evidence suggested that ecircRNAs may prevalently act as miRNA sponges, while ciRNAs and EIciRNAs (that is, the species including introns) are localized in the nucleus, where they regulate gene expression [96]. In particular, circRNAs can regulate the expression of the parental gene, acting as cis-regulators, or of independent genes, acting in trans. Several functions have been reported: i) miRNA sponging; ii) protein binding for activation or stabilization of enzymatic activity, promotion of subcellular localization, regulation of translation, epigenetic modifications, ubiquitination; iii) translation of small peptides or proteins [97, 98]. Recently, circRNA-protein interactions are gaining attention, suggesting a role in regulating also competing interactions between RNA molecules and RBPs [99, 100]. To date, more than 140,000 annotated human circRNAs have been included in the database circBank (http://www.circbank.cn/). Several isoforms of various circRNAs deriving from the same protein-coding gene have been reported, making nomenclature very important to identify the specific transcript object of study [101].

One of the most known examples of miRNA-sponging circRNAs is represented by CDR1 antisense RNA (CDR1-AS), also known as circular RNA sponge for miR-7 (ciRS-7), which includes within its sequence 63 MREs for miR-7 [102, 103]. CDR1-AS is transcribed from the same locus of cerebellar degeneration related 1 (CDR1) (of which it represents a circular NAT) and is involved in brain physiology. Recently, the involvement of this circRNA in several diseases, including cancer, has been reported [104]. Another circRNA whose miRNA-sponging function has been widely investigated is circular RNA itchy E3 ubiquitin protein ligase (circ-ITCH), transcribed from the ITCH locus and acting as a tumor suppressor in esophageal squamous cell carcinoma (ESCC), colorectal cancer (CRC) and bladder cancer [105–107]. In ESCC, circ-ITCH sponged miR-7, miR-17, and miR-214, leading to inhibition of the Wnt/β-catenin pathway [105]; the same effect was reported in CRC through the sequestration of miR-7, miR-20a, and miR-214 [106]. In bladder cancer, circ-ITHC downregulation resulted in increased availability of miR-17 and miR-224, in turn inducing repression of the targets PTEN and p21/cyclin dependent kinase inhibitor 1A (CDKN1A) [107]. Multiple evidence about circ-ITCH acting as a sponge in several cancer models supports the hypothesis of a fundamental and conserved role of ceRNA mechanisms in both physiological and pathological conditions.

Interestingly, a study published in 2022 reported that circRNAs may also interact with mRNAs [108], similarly to lncRNAs. The study focused on circular RNA zinc finger protein 609 (circZNF609), encoded from the ZNF609 locus, an oncogenic circRNA upregulated in several cancer models and reported to act as a miRNA sponge. Rossi et al. [108] showed that circZNF609 also interacted in vivo with mRNAs, namely cytoskeleton associated protein 5 (CKAP5), UPF2 regulator of nonsense mediated mRNA decay (UPF2), and serine and arginine repetitive matrix 1 (SRRM1), recruiting on their sequences ELAV like RNA binding protein 1 (ELAVL1), an RBP, in turn, regulating stability and translation of these mRNAs [108]. This new recent finding demonstrates that little is known about the potentiality of ceRNA networks, and further investigations will be needed to increase the knowledge about their mechanisms.

Conditions for ceRNA networks

Looking at ceRNA interactions as a whole, their complexity is undeniable (Figure 1). To understand how a particular mechanism operates within cells, researchers must ignore all other phenomena and concentrate only on a single molecular event. Nonetheless, it’s important to keep in mind that the mechanisms explained here occur in the complex framework presented above. To integrate the small puzzle pieces in the comprehensive context of a living cell, several studies investigated the conditions required for the existence of a ceRNA network. First of all, each component of the ceRNA network is characterized by specific expression levels, which may differ between the single actors involved in the interactions. The abundance of a given miRNA is crucial to understand if it will be able to bind all its potential targets or only a portion of them (most likely the ones with the higher binding affinity); on the contrary, if a ceRNA molecule (such as an mRNA, a pseudogene, a lncRNA or a circRNA) is more abundant than the miRNA, it could be speculated that only a few miRNA molecules (the majority of them being sponged) will be available for target-binding. Accordingly, Denzler et al. [109] demonstrated that the miRNA: target ratio is crucial for competing interaction by investigating miR-122 function in the liver. There are numerous examples of potential scenarios occurring within a cell, but it is easy to understand that the stoichiometric equilibrium among interacting RNAs is one of the most important parameters regulating ceRNA interactions. What seems to be elucidated is that a near equimolar expression among RNA interactors is necessary for ceRNA network establishment. This would allow a small alteration in the expression levels of a given ceRNA to affect the expression of the other interactors. As an example, consider a ceRNA that is highly more abundant than a miRNA, being able to sponge virtually all miRNA molecules: a small variation in ceRNA levels would not produce a significant effect on miRNA availability in the cytoplasm. The nearly equimolar equilibrium was estimated by a mathematical model using PTEN and VAPA interplay as a model of ceRNA interactions [110]. It should be noted that a debate arose regarding the true impact of ceRNA networks on cell biology. Denzler et al. [109] concluded in their study on miR-122 in the liver that variation in expression of a single miRNA family was unlikely to impact both physiology and pathology: indeed, modulation of miR-122 expression was able to regulate the expression of miRNA targets, while ceRNA effects were induced when the concentration of the miRNA reached levels exceeding physiology. Despite this negative result, the authors did not exclude the existence of other unidentified ceRNAs highly expressed within cells nor the possibility that their method was not sensitive enough to detect perturbations of the ceRNA network [109].

Display full size

Figure 1. 

Representation of ceRNA interactions and molecular effects. (A) ceRNA interactions occurring among the involved RNA species: miRNAs bind and negatively regulate mRNAs, pseudogenes, lncRNAs, and circRNAs; lncRNAs and circRNAs bind mRNAs and modulate their splicing or stability; (B) function of miRNAs: miRNAs bind mRNAs, lncRNAs, and pseudogenes with imperfect complementarity and induce target degradation; when a miRNA interacts with its mRNA target with perfect complementarity, translation is inhibited; (C) mRNAs act as miRNA sponges by binding miRNAs through multiple MREs included throughout their sequence, thus protecting other mRNAs from miRNA-mediated repression; (D) pseudogenes can regulate the expression of parental protein-coding genes by sponging miRNAs; high sequence homology between the pseudogene and the parental gene determines the high number of shared MREs; (E) lncRNAs can protect other transcripts from miRNA-mediated repression by sponging miRNAs; direct interactions between lncRNAs and mRNAs modulate the splicing of the latter; (F) circRNAs act as miRNA sponges, impairing their function on other transcripts, or directly interact with mRNAs recruiting on the RBPs involved in splicing or mRNA stabilization

Another undeniable parameter is, as mentioned, the binding affinity between the miRNA and the MRE. The existence of MREs differing in the length of sequence complementarity with miRNA seed region (classified as 8mers, 7mers, and 6mers, see above) suggests that miRNAs may have primary targets, bound with higher affinity and effectively regulated, while other transcripts showing MREs with lower affinity may be less affected by modulation of miRNA availability, or they may act as sponges [25, 26]. Indeed, it has been reported that 7mers and 6mers were respectively 50% and 20% as effective as 8mers in regulating target abundance [111]. Therefore, it can be assumed that low levels of a given miRNA would induce its preferential binding with high-affinity MREs, while binding to low-affinity target sites would require a higher abundance of the miRNA. It could be hypothesized that this binding affinity would also mirror the efficacy of miRNA-mediated repression, suggesting that targets bearing low-affinity MREs would be negatively regulated by the miRNA only whether it is abundantly expressed. Actually, effective binding between miRNAs and low-affinity MREs (and consequential negative regulation of the target) does occur within cells, in some cases with unexpected repression efficiency. This contradiction may be explained by other MRE-related parameters influencing ceRNA networks. First of all, it is important the number of MREs included within the sequence of a given ceRNA: the higher the number of MREs, the more miRNA molecules sponged. But localization and distribution of MREs are also crucial since neighbor MREs showed to increase the binding affinity by acting in cooperation. Intriguingly, cooperation among low affinity or non-canonical MREs may contribute to target site abundance more than high affinity or canonical binding sites; experimental evidence showed that two closely spaced (58 nts apart) MREs induced a more efficient negative regulation than expected by their independent action [111].

Even in the presence of multiple MREs with comparable affinity, ceRNA interactors have to share the same subcellular localization for the physical interaction to occur. As discussed, lncRNAs (including circRNAs) showed to exert functions both in the nucleus and in the cytoplasm, while mRNAs and miRNAs typically localize in the cytoplasm where they are translated into proteins and act as negative regulators of gene expression, respectively. However, a nuclear function for miRNAs was reported, showing their involvement in regulation of transcription and alternative splicing [112]. The asymmetric distribution would impair a direct interaction. Nevertheless, this potential obstacle may be overcome by the shuttling of miRNAs and lncRNAs between nucleus and cytoplasm that has been demonstrated for some of them [113, 114], thus allowing subcellular colocalization (even if temporary) and direct interaction.

Supposing that ceRNA interactors colocalize within the same subcellular compartment, the regions of each molecule being involved in ceRNA interactions must be free to bind to each other. That means no other interactions should occur, neither intramolecular (formation of secondary structures) nor intermolecular (interactions with either RNAs or proteins). In this context, interactions with RBPs involved in physiological processes (such as splicing, miRNA-induced silencing, etc.) may represent an additional competitor for RNA binding affecting ceRNA interactions [34].

As previously discussed, alternative splicing patterns producing mRNAs differing at the level of the 3’-UTR may contribute to MRE abundance within the cell [67], as well as the phenomenon of 3’-UTR shortening observed in cancer [66]. There is another molecular process, occurring physiologically in cells, that affects RRIs, namely RNA editing. Being part of post-transcriptional processes, RNA editing consists in modifying the sequence of a transcript [including mRNAs, transfer RNAs (tRNAs), ribosomal RNAs (rRNAs), etc] by insertion, deletion, or substitution of specific nucleotides [115]. Interestingly, RNA editing is also responsible for isomiR biogenesis [41]. Adenosine-to-inosine modifications represent the most frequent sequence alterations induced by editing, modulating the function of the modified transcript [116]. Obviously, changes in a transcript sequence would affect ceRNA interactions, potentially disrupting or creating MREs. Similarly, the different patterns of expression of isomiRs, characterized by variation in length or sequence, would affect the affinity to existent MREs or induce a given isomiR to recognize new targets [41]. These observations suggest that ceRNA interactions may be finely regulated even in physiological contexts by the intricate modifications underlying the dynamics and complexity of cell processes. Hence, the dramatic twisting of molecular mechanisms inducing the onset of a disease, such as cancer, which includes accumulation of mutations, alternative splicing, and 3’-UTR shortening, would be able to extremely alter the ceRNA interactions, being simultaneously the cause and the effect of the pathological transformation of diseased cells.

Low-throughput biophysical, biochemical, and cellular methods for detecting RRIs

High-throughput targeted methods for detecting RRIs

Next-generation sequencing technologies have been developed to study RNA molecules at the transcriptome level [136].

Identification of transcriptome-wide RRIs

The development of modern biochemical methods and high-throughput sequencing technology resulted in the discovery of ncRNAs, a novel family of RNAs, and the various functions exerted through interactions with other RNAs. Several technologies using biochemical RNA cross-linking followed by high-throughput sequencing have enabled us to discover the global map of RRIs crucial for physiological activities in recent years. The transcriptome-wide RNA interactome, which has the potential to include all of the RNAs in a cell, leads to the development of several novel techniques.

Pair-wise correlation approach

The pair-wise correlation technique is based on the assumption that the expression levels of each pair of interacting ceRNAs in a network show a significant positive correlation, which can be evaluated by Pearson correlation coefficients. The positive correlation principle is based on the mechanism which evidences that a high concentration of a miRNA sponge will reduce the available miRNAs for interacting with its target RNAs [175], which, in turn, will not undergo miRNA-based repression and, accordingly, will increase their levels.

Using target prediction databases, this approach first searches a set of miRNAs for all pairs of RNAs that contain MREs. Given an RNA pair including RNA1 and RNA2, a statistical test, such as a hypergeometric test, is performed to determine the significance of the shared MREs; pairs with significant positive correlations or significant differences in correlations in two separate situations are predicted to have miRNA sponge interactions [175]. Zhou et al. [176] combined matched breast cancer miRNA and mRNA expression data with miRNA-mRNA relationships to identify breast cancer-specific miRNA sponge interactions. Using predicted miRNA target information and a hypergeometric test, the authors first identified common miRNAs for all mRNA-mRNA pairs. They then used Pearson correlation coefficients to evaluate the significance of the correlations between putative miRNA sponge interaction pairings sharing the same MREs. The miRNA sponge network was developed using the significant miRNA sponge interaction pairs that have been found. The targets in the miRNA sponge network revealed discriminative capability and were considered good biomarker candidates in breast cancer, according to the results of the survival study [176]. Xu et al. [177] applied a similar strategy to predict the landscape of mRNA-related miRNA sponge interactions across 20 major cancer types. Even though the majority of miRNA sponge connections are cancer-specific, malignancies involving the same cell type share a considerable number of common miRNA sponge interactions. In each cancer, functional analysis revealed preserved and rewired network miRNA sponge hubs, which established tense miRNA sponge connections to form conserved and cancer-specific modules [177]. Using expression profiles and miRNA-target interactions, Shao et al. [178] suggested a pair-wise correlation-based technique to uncover dysregulated miRNA sponge interactions in lung adenocarcinoma. The authors classified the dysregulated miRNA sponge interactions as “gain” or “loss”. They discovered that correlations of gain miRNA sponge interactions are significantly positive in cancer, and correlations of loss miRNA sponge interactions are significantly positive in normal conditions. It was also observed that gain miRNA sponge pairs exhibit consistent expression in cancer, whereas loss miRNA sponge pairs do not. Additionally, gain and loss miRNA sponges that participate in gain and loss miRNA sponge interaction networks as topologically important nodes are linked to cancer development. Additionally, several dysregulated miRNA sponge modules may be considered diagnostic biomarkers even in other three independent lung datasets. In the field of biomarkers for diagnosis, this study suggests the possibility of a new biological mechanism in cancer [178]. Chiu et al. [179] analyzed optimal conditions for miRNA sponge regulation by integrating expression profiles and putative miRNA-target interactions in TCGA datasets of glioblastoma multiforme (GBM), ovarian serous cystadenocarcinoma (OV), lung squamous cell carcinoma (LUSC), and large acute myeloid leukemia (LAML). They reported that miRNA sponge regulation is dynamic, differed between cancer types, and was related to the number of common miRNAs and target binding sites [179]. Their findings help us comprehend the dynamic nature and important roles of miRNA sponge regulation in cancer [175].

Partial correlation approach

Differently from pair-wise correlation methods, partial association methods consider the miRNA expression levels when examining ceRNA-ceRNA interactions. When the relationships between miRNA sponges are assumed linear, the partial correlation methods measure the correlation between two ceRNAs sharing a common MRE [175, 180]. Also, when no assumptions are made about the nature of relationships (linear or non-linear) between miRNA sponges, conditional mutual information (CMI) methods can be used rather than partial correlation [59, 180].

In their study, Sumazin et al. [59] examined the ability of coding and non-coding RNAs to interact as miRNA sponges in glioblastoma by evaluating a large set of related miRNA and gene expression profiles. They identified an extensive miRNA-mediated network of RRIs and uncovered miRNA-mediated crosstalk between different oncogenic pathways. They analyzed two types of miRNA-mediated modulators using distinctive molecular mechanisms: sponge modulators and non-sponge modulators. Owing to the prohibitive computational cost of examining a large number of ceRNA-miRNA-ceRNA triangles, they reduced the number of triangles to be evaluated by using experimentally validated or predicted miRNA-target information. Only RNA-RNA pairs with a significant overlap of common miRNAs were examined for sponge modulators, and expression data were combined with predicted miRNA target regulatory interactions to form ceRNA-miRNA-ceRNA triangles. Concerning non-sponge modulators, all RNA-RNA pairs have been analyzed to predict triangles, including the empty and non-statistically significant overlap of common miRNAs, and their expression data were merged with experimentally validated miRNA-target information. Using mutual information (MI) and CMI, Sumazin et al. [59] identified over 7,000 sponge modulators as miRNA sponges with over 248,000 pairs of miRNA-mediated RRIs, and 148 non-sponge modulators with 169 miRNA-mediated RRIs with more than 100 mRNAs involved. This wide miRNA-mediated network offers a new view on the deregulation of essential pathogenic and physiological processes. Paci et al. [181] suggested a partial correlation-based approach to investigate the role of lncRNAs as miRNA sponges. They studied the matched miRNA and gene expression profiles of 72 patients with breast invasive carcinoma (BRCA) obtained from TCGA. The tumor and normal tissue samples came from the same patients. They designed the miRNA-mediated interaction (MMI) networks in tumor and normal conditions by evaluating the difference between the Pearson correlation and partial correlation (sensitivity correlation) for each lncRNA-miRNA-RNA triangle. The tumor-MMI network appears to be clearly distinct from the normal-MMI network, demonstrating that the miRNA decoy mechanism may switch on or off its mediators depending on physiological or pathological conditions [181]. Chiu et al. [182] developed Cupid, an integrative framework for context-specific miRNA target and miRNA sponge interaction networks by combining sequence-based evidence and expression data. Cupid first applies a support vector machine (SVM) classifier to predict candidate miRNA-target interactions and then uses expression-based evidence to further determine ceRNA activities [182]. Cupid is a 3-step prediction method: firstly, it applies Fisher’s exact test [175] to estimate candidate miRNA sponge interaction pairs; secondly, it also uses MI and CMI to identify miRNA sponge interactions and finally performs Brown’s method to further infer significant miRNA sponge interactions [175]. The authors reported over 500 validated miRNA-target interactions involved in breast cancer by employing high-throughput data in breast cancer cell lines for validation. To evaluate the miRNA targets which compete for miRNA regulation, they focused on five breast cancer-associated RNA regulators, cyclin D1 (CCND1), estrogen receptor 1 (ESR1), hypoxia inducible factor 1 subunit alpha (HIF1A), platelet derived growth factor receptor alpha (PDGFRA), and nuclear receptor coactivator 3 (NCOA3). As a result, 8 out of 11 of the predicted miRNA sponge interactions showed significant up-regulation [59].

Mathematical modeling approach

As the term suggests, mathematical modeling approaches focus on predicting ceRNA interactions using different mathematical models. In addition, among the various mathematical model, the minimal model and stochastic model reported results indicating that binding free energies and repression mechanisms are actively involved in ceRNA interactions [183, 184]. Bosia et al. [183] designed a stochastic model to define the equilibrium and disequilibrium characteristic of a miRNA-ceRNA interaction network, revealing how ceRNAs and miRNAs could oscillate correlatively under complex conditions like feedback or feedforward loops. In their study, Swain et al. [185] introduced a competitive endogenous RNA score model to identify candidate ceRNA pairs and estimate their ability to sequester miRNAs. Considering that the high number of components in ceRNA networks can enhance the computational complexity of mathematical models, it could be more efficient to use this method for the construction of smaller ceRNA networks. Following ceRNA interaction prediction, Cytoscape software [186] can be used to generate and analyze the network based on the identified ceRNA interactions [185].

Ala et al. [110] and Figliuzzi et al. [184], used a similar in silico model to find the best settings for miRNA sponge activity. The model illustrates the molecular transition process for miRNA sponge interactions which are mediated by a shared pool of miRNAs. They improved the resulting network by deleting all edges that were not included in the conserved co-expression networks developed in a previous study by Piro et al. [187]. The evaluation of the complexity of the miRNA sponge interaction networks revealed that TFs and ceRNA interactions were cross-regulated, and the latter could cause upregulation or aberrant expression of the TFs in cancer. Furthermore, the results suggested that alterations of a single miRNA sponge have significant effects on the integrated transcriptional and miRNA sponge networks [110]. Yuan et al. [188] developed a coarse-grained model for a minimum miRNA sponge interaction network to quantitatively analyze the behavior of miRNA sponge regulation. They validated the computational results using a synthetic gene in cultured HEK293 cells and found that the number and binding strength of miRNAs and miRNA sponges, as well as the degree of complementarity between miRNAs and miRNA sponges, influenced the extent and strength of the miRNA sponge impact. The combined computational model and experimental validation results provided a clearer understanding of miRNA-mediated cross-regulation on miRNA sponges.

It should be mentioned that more research, optimization, and comparative study of bioinformatic approaches for creating ceRNA networks is required to gain a better understanding of ceRNA regulation.

To uncover miRNA sponge regulation in human cancers, most of the existing computational methods are proposed at the network level, and only a small number of methods are presented at the module level. Modules in this context refer to groups of miRNA sponges having mutual competition and acting as functional units in the regulation of biological processes [189]. Therefore, module identification is an important way to provide biological insights into ceRNA networks. As an example, modules could facilitate functional genome annotation by using the guilt-by-association principle [190] and help understand the origin and progression of human diseases [191]. Currently, most of the existing module discovery methods infer miRNA sponge modules by using ceRNA network data. Although these methods also use heterogeneous data (gene expression data, miRNA-target binding, and MRE information) in the process of inferring ceRNA networks, the identified miRNA sponge modules are network-level modules, each containing a set of miRNA sponges that frequently interact with each other in a ceRNA network. These module-based methods are categorized into three types: network-based clustering, matrix factorization, and step-wise evaluation, as summarized by Zhang et al. [189].

Since non-coding and protein-coding RNAs can act as miRNA sponges, RRI databases are valuable resources applied in computational approaches to develop potential networks of miRNA sponge interactions. There are two types of miRNA-target binding databases: those with experimentally validated target bindings and those with predicted miRNA-ceRNA interactions [175]. To identify potential miRNA sponge interactions, the users can query distinct databases or merge different datasets as the initial framework of computational approaches. Some of the most widely used datasets of miRNA sponge interactions are presented below (Table 1).

CeRDB (https://www.oncomir.umn.edu/cefinder/) [192] was one of the first-developed miRNA sponge interaction databases; the authors used TargetScan (https://www.targetscan.org/vert_80/) [206] to check for the co-occurrence of MREs in mRNAs across the genome. CeRDB estimates the miRNA sponges that compete with the other RNAs for the binding of a specific miRNA. CEFINDER, a web-based tool provided by CeRDB, allows users to find miRNA sponges for a specific mRNA target based on the homology of MREs located within the 3’-UTR sequence [192].

The ENCORI database (https://rnasysu.com/encori/) [172] stores 108 datasets obtained from 37 studies, including 47 high-throughput sequencing of RNA isolated by CLIP (HITS-CLIP) datasets, 51 photoactivatable-ribonucleoside-enhanced CLIP (PAR-CLIP) datasets, 9 individual-nucleotide resolution UV CLIP (ICLIP) datasets and 1 CLASH dataset. Recently, a miRNA sponge interaction database was added to ENCORI. The experiment results stored in ENCORI provide information about the physical binding between miRNA-mRNA, miRNA-lncRNA, miRNA-circRNA, miRNA-pseudogene, and miRNA-small ncRNAs. The interactions can be used to verify computational predictions [172].

LncACTdb (http://bio-bigdata.hrbmu.edu.cn/LncACTdb/) is another database created by Wang et al. [161] defining lncRNA-miRNA-gene interactions as lncRNA-associated competing triplets (lncACTs). This is a database of lncRNA-associated competing triplets designed to aid in the study of lncRNA function as miRNA sponges. The database contains 5,119 functionally active and over 530,000 computationally predicted lncACTs derived from merging heterogeneous data from multiple in silico target prediction studies, AGO-CLIP assays, and transcriptome sequencing expression profiles. The database has been recently improved in LncACTdb 3.0, an updated and significantly expanded version that provides comprehensive information on ceRNAs in different species and diseases [195].

Another database is LnCeCell (http://bio-bigdata.hrbmu.edu.cn/LnCeCell/) which provides functional lncRNA-related miRNA sponge interaction networks in single cells and sub-cellular locations curated from the published literature and high-throughput datasets across 25 human cancers [196].

LnCeVar (http://www.bio-bigdata.net/LnCeVar/) is a comprehensive database that collects genomic variations that disturb lncRNA-related miRNA sponge regulation curated from the published literature and high-throughput datasets across 33 human cancers [197].

Other useful resources are: miRTissue_ce (http://tblab.pa.icar.cnr.it/mirtissue.html), a web service for the analysis and characterization of miRNA sponge interactions in several cancer tissue types [200]; SomamiR 2.0 (http://compbio.uthsc.edu/SomamiR/), a database of cancer somatic mutations that potentially alter the interactions between miRNAs and miRNA sponges (including mRNAs, circRNAs, and lncRNAs) [205]; Pan-ceRNADB (http://bio-bigdata.hrbmu.edu.cn/pan-cernadb/), a web resource which investigates mRNA-related miRNA sponge interactions across 20 major human cancers [177].

The VECTOR (https://vectordb.it/) database is also worth mentioning; it is an interactive tool that identifies and visualizes the relationships among RNA molecules. VECTOR database contains: i) the TCGA-derived expression correlation values of miRNA-mRNA, miRNA-lncRNA, and lncRNA-mRNA pairs combined with predicted or validated RRIs; ii) data of sense-antisense sequence overlapping; iii) correlation values of TF-miRNA, TF-lncRNA, and TF-mRNA pairs associated with ChiP-seq data; iv) expression data of miRNAs, lncRNAs and mRNAs both in UM and physiological tissues [207].

Experimentally validated miRNA-mRNA binding databases mainly include miRTarBase (https://mirtarbase.cuhk.edu.cn/~miRTarBase/miRTarBase_2022/php/index.php) [199], DIANA-TarBase v.8 (https://dianalab.e-ce.uth.gr/html/diana/web/index.php?r=tarbasev8) [193], and miRWalk (http://mirwalk.umm.uni-heidelberg.de/) [201]. The commonly used predicted miRNA-mRNA binding databases are TargetScan (https://www.targetscan.org/vert_80/) [206], miRanda (https://cbio.mskcc.org/miRNA2003/miranda.html) [168], PicTar (https://pictar.mdc-berlin.de/) [202], DIANA-microT (https://dianalab.e-ce.uth.gr/microt_webserver/#/) [194], and PITA (https://genie.weizmann.ac.il/pubs/mir07/mir07_prediction.html) [203].