Circular RNAs (circRNAs) (Salzman et al., 2012; Hansen et al., 2013a; Jeck et al., 2013; Memczak et al., 2013) are covalently closed RNA molecules characterized by a non-linear 3′-5′ junction resulting from an unusual splicing event denominated backsplice junction (Ashwal-Fluss et al., 2014; Lasda and Parker, 2014; Conn et al., 2015; Ivanov et al., 2015; Starke et al., 2015; Wang and Wang, 2015; Zhang et al., 2016). This particular splicing event causes circRNAs to lack both 3′ poly(A) tail and 5′ capping, conferring resistance to exonuclease activity [e.g., RNaseR (Suzuki et al., 2006; Vincent and Deutscher, 2006)] and therefore determining a general longer half-life as compared to linear RNAs (Jeck et al., 2013). Despite the great interest towards this class of non-coding RNA especially for their potential as disease biomarkers (Li et al., 2015; Memczak et al., 2015; Lyu and Huang, 2017; Jahani et al., 2020; Wang et al., 2021), only few of them have been functionally characterized to date [a few examples are reported in (Memczak et al., 2013; Ashwal-Fluss et al., 2014; Barrett and Salzman, 2016; Suenkel et al., 2020)]. Several biological functions have been proposed for circRNAs, ranging from the regulation of pluripotency to early lineage differentiation (Yu et al., 2017; Izuogu et al., 2018) to the control of unique functions in specialized cells (Berta et al., 1990; Capel et al., 1993). However, in terms of research interest, the most prominent function attributed to circRNAs is their action as miRNA sponge to regulate target gene expression by inhibiting miRNA activity through competition with the RISC complex, a function commonly defined as miRNA binding (Hansen et al., 2013a; Hansen et al., 2013b; Memczak et al., 2013; Huang et al., 2017; Panda, 2018; Verduci et al., 2019).

The computational resources that have been recently developed to investigate circRNA function and expression substantially consist of databases collecting all available information relative to already known circRNAs in different species [e.g., circBase (Glažar et al., 2014)]. Although these databases have been expanded to account for circRNA expression patterns, functional predictions (e.g., miRNAs binding sites) and disease associations (Yang et al., 2011; Ghosal et al., 2013; Li et al., 2014; Dudekula et al., 2016; Liu et al., 2016), still they are limited to previously identified circRNAs in human and mouse species only. Moreover, most of the available tools to predict miRNA:circRNA interactions fall short when dealing with previously uncharacterized circRNAs. For instance, some resources, requiring a Gene ID as input and not permitting sequence browsing, can predict miRNA targets only for annotated genes and are therefore informative only if circRNAs include a 3′ UTR. On the other hand, tools that allow the analysis of sequences are restricted to a fixed nucleotide sequence length limiting the number of sequences that can be investigated at once (Kertesz et al., 2007; Betel et al., 2008; Rennie et al., 2014; Agarwal et al., 2015; Rennie et al., 2016; Karagkouni et al., 2018). In an attempt to bypass these limitations, some prediction tools provide offline stand-alone versions of their algorithms (Enright et al., 2003; Rehmsmeier et al., 2004; Thadani and Tammi, 2006; Bandyopadhyay and Mitra, 2009; Hsu et al., 2011; Coronnello and Benos, 2013; Agarwal et al., 2015). This solution, while allowing a full customization of the analysis, comes with the drawback of a high rate of false positives as these approaches easily return thousands of putative miRNA:RNA interactions. A common solution to reduce these lists is to perform a systematic validation of the targets (e.g., with RIP assays) or the integration with known interactions. In particular, since the binding of the RISC complex is mediated by the interaction of the miRNA with members of the Argonaute protein family (Hammond et al., 2000; Hutvágner and Zamore, 2002), it is fundamental, for a predicted miRNA:circRNA site to be functional, that also Ago proteins are binding in the same positions. Currently, the development of various CLIP-Seq (Cross-linking and Immunoprecipitation followed by sequencing) protocols provides an extremely valuable source of high-throughput data for Ago binding sites (Chi et al., 2009; Hafner et al., 2010; Jungkamp et al., 2011; Leung et al., 2011; Helwak et al., 2013; Clark et al., 2014; Grosswendt et al., 2014). In addition, the availability of already validated miRNA:RNA sites obtained through RIP assays (or other experimental approaches) provides a further layer of valuable information for filtering candidate miRNA:circRNA pairs. Following the approach proposed in (Dori and Bicciato, 2019), we developed Circr, a tool that integrates miRNA:target predictions from three commonly used algorithms, i.e., miRanda (Agarwal et al., 2015), TargetScan (Enright et al., 2003), and RNAhybrid (Rehmsmeier et al., 2004) with validated miRNA:RNA interactions and AGO peaks data. The combination of multiple tools for the identification of a single miRNA recognition site together with experimental data allows to efficiently reduce the pool of candidate miRNA:circRNA interactions for functional studies.

To test the performances of Circr, we analyzed a set of circRNAs predicted to be expressed in the developing mouse brain (Dori et al., 2019). Data were obtained from RNAseR treated RNA of the three main cell population of the lateral cortex (proliferating and differentiating progenitors and newborn neurons). From this cohort of sequences, we selected 100 circRNAs of different lengths overlapping genes in the sense and antisense strand, and intergenic ones to cover all possible genomic features of circRNAs. As described in the Methods section, it is possible to provide in input the coordinates of the full sequence of the circRNA and let Circr calculate the exon coordinates or a final set of coordinates with all the features included in each circRNA. To test Circr, we used the 100 mouse sequences and analyzed them using both types of input (i.e., full sequence and final set of coordinates), with the default set of support files relative to the mm9 mouse genome. When providing the genomic coordinates of circRNAs full sequences, Circr completed the analysis in 46 min and retrieved more than 190,000 miRNA:circRNA binding sites (Figure 4). Conversely, providing the input file as a set of final exon/intron coordinates, Circr took 227 min for the prediction of 263,581 sites (Figure 4). Starting from Circr output, we can either focus on interactions identified by all 3 algorithms, resulting in 12,142 hits for the full sequence analysis and 53,539 for the feature coordinates one, or we can select only those overlapping either validated seed regions or an AGO peak, obtaining a list of 25,529 sites (32 if we account for the presence of both an AGO peak and a validated interaction) from the full sequence analysis and 16,991 (20 with both AGO/validated interaction) for the feature coordinates set up. By combining all criteria (3 tools and either an AGO peak or a validated interaction), the list goes down to 1,198 and 1,413 interaction sites for the full sequence and feature coordinates respectively, providing the user with a more manageable number of putative interacting miRNA:circRNA pairs for further validation. The analyses were carried out on an Ubuntu server using 8 threads. The same analyses performed on a personal computer with MacOS 10.15 and 8 threads required 132 min for the full sequence and 384 min for the final set input.

Circular RNAs have emerged as a significant player in both physiological and pathological conditions and have become an invaluable source of information as prognostic markers in cancer and several other diseases. However, despite a large number of circRNA databases is currently available, their limitation to already known and annotated sequences in specific organisms hampers the identification of novel circRNAs and the investigation of their potential role as miRNA regulators. As an example, CircNet 2.0 (Liu et al., 2016) focuses on circRNAs found in human cancers whereas ENCORI (Li et al., 2014) collects circRNAs only from human and mouse. Other limitations reside in the nature of miRNA:circRNA interactions reported and their annotation: CircNet 2.0 reports only interaction predicted by at least one tool among PITA, miRanda and TargetScan, but provides only the miRNA ID without specifying the number and genomic location of binding sites. The reported interactions are all supported by miRTarBase (Huang et al., 2020) but these validations involve the miRNA and the gene hosting the circRNA, irrespectively of the site of the interaction. ENCORI reports interactions predicted by just a single tool (miRanda) and only if supported by Ago CLIP-seq Data. From a usability point of view, ENCORI is miRNA-centered and does not allow searching by circRNA ID or coordinates. To cope with these needs, we developed Circr, a tool that combines the output of different miRNA binding site prediction algorithms with validated miRNA:RNA interactions (through direct approaches as luciferase assays or RNA-IP) and Argonaute binding site data. The output of Circr returns a comprehensive view of the circRNAs interactome, it has been structured to be easily browsed and investigated, thus helping researchers to narrow down the list of potential targets for further validation and functional characterization. We applied Circr to the analysis of a set of circRNAs predicted to be expressed in the developing mouse brain and, among the interactions supported by a validated seed region or an AGO peak, we focused our attention to those involving miRNAs known to be expressed in the same tissue (as, for instance, miR-9, miR-124, miR92a, miR92b). As a confirmation, we found a binding site identified by all 3 algorithms for miR-9 on CiCo_mm9_circ_000349, a circRNA already annotated in circBase (mmu_circ_0000044) and identified in mouse brain by (Memczak et al., 2013; Rybak-Wolf et al., 2015). In addition, we also found a binding site, predicted by 2 tools, for miR-92a/miR-92b on CiCo_mm9_circ_002328 (mmu_circ_0006166). To compare Circr performance on a circRNA reported in existing databases, we interrogated CircNet 2.0 and ENCORI for miRNA bindings of hsa_circ_0001946 (CDR1as), a well characterized human circRNA. Circr results reproduced the findings of the two databases and was able to identify two validated interactions with mir-7 and mir-671 (Hansen et al., 2011; Memczak et al., 2013).

One of the major features of Circr resides in its flexibility, as users can virtually investigate any given sequence in any given organism, as long as all the necessary files are available and provided. A possible limitation of our method could be represented by the fact that, in its current implementation, Circr adopts exclusively prediction tools based on seed matching, a feature that might affect the completeness of the information provided. However, we foresee that the growing interests toward this class of RNA will lead to the development of more refined and complete databases coupled with dedicated computational approaches that could be easily incorporated in Circr. In conclusion, although exploiting a previously described approach [e.g., (Chen et al., 2021)], to date Circr represents the first standalone tool that combines multiple prediction algorithms together with experimental data and we believe that it can provide a relevant contribution to the discovery of new and potentially functionally relevant miRNA:circRNA interactions.