The roles of microRNAs (miRNAs) in regulating biological processes through reshaping cellular transcriptome and proteome have been an exciting field since they were discovered in metazoan in 1993 (Lee et al., 1993). These small RNA molecules act through negatively regulating transcript levels post-transcriptionally, although positive regulation has been described in some cases (Vasudevan et al., 2007; Orom et al., 2008). MiRNAs also interact with protein translation machinery to attenuate the protein synthesis. It is estimated that miRNAs may regulate over 60% of transcripts in humans (Friedman et al., 2009), where each miRNA might target several messenger RNAs (mRNAs) and a single mRNA can also be targeted by several miRNAs (Brennecke et al., 2005; Lewis et al., 2005).

Technological advances such as bioinformatics and next-generation sequencing (NGS) allowed the identification of a great number of additional putative miRNAs in different organisms in recent years. Although it is difficult to estimate the total number of miRNAs in humans and other species (Berezikov et al., 2006), to date, over 2,000 human miRNAs have been identified and deposited in miRBase (release 19¹) (Griffiths-Jones et al., 2008). However, the process of identifying miRNAs is still a complex and difficult task requiring interdisciplinary strategies, including the integration of experimental approaches with computational methods. A number of miRNA prediction algorithms have been developed, nonetheless the results merely provide a list of miRNA candidates, which requires extensive experimental works to identify real miRNAs in cells (Liu et al., 2012). In order to increase the accuracy of miRNA prediction, several characteristics of these regulatory RNA molecules have been taken into account.

Furthermore, since miRNAs differ between kingdoms, distinct approaches that take such characteristics into account have been applied (Mendes et al., 2009). For example, when compared to animal miRNAs, the stem-loop structures in plant miRNAs are more variable in size, being usually larger, they show more pairing between the miRNA and the other arm of the stem-loop, and tend to have around 21 nucleotides in length (Lau et al., 2001; Reinhart et al., 2002; Bartel, 2004). Despite applying similar principles to plant and animal miRNA prediction, fewer methods have been proposed to find new miRNAs in plants, in part, due to the heterogeneous nature of plant miRNA stem-loops (Mendes et al., 2009). On the other hand, several tools can be used to predict miRNAs both in animals and plants. In this review we focused on the strengths and weaknesses of tools used in animal miRNA discovery.

Understanding the miRNA biogenesis is essential to build effective predictive models. Hence, some key points are important to briefly mention here (thoroughly reviewed in Krol et al., 2010; MacFarlane and Murphy, 2010; Starega-Roslan et al., 2011). Mature miRNAs are usually around 22 nucleotides long and a number of them are highly conserved throughout evolution (Bartel, 2004). They are located in either the introns of protein-coding genes or in the non-coding region of genes or intragenic regions of the genome (Lagos-Quintana et al., 2001; Lau et al., 2001; Lee and Ambros, 2001). MiRNAs derive from longer transcripts called primary miRNAs (pri-miRNAs), which are transcribed and processed similarly to protein-coding genes, with a 5′ cap and a 3′ poly-adenosine tail, but with hairpin structure(s). Such hairpin is characteristic of pri-miRNAs and is essential to be recognized and cleaved in the nucleus by the RNase III enzyme, Drosha. Its product is an approximately 70 nucleotides long hairpin precursor miRNA (pre-miRNA), which is then exported to the cytoplasm where it is further processed by Dicer, another RNase III enzyme, resulting in a double stranded ∼22 bp miRNA. Usually one of these strands produces the mature miRNA and will be associated to the RNA-induced silencing complex (RISC) to interact with its mRNA targets (Lee et al., 2002; Gregory et al., 2004).

In animals, the miRNA generally interacts with mRNAs through partial sequence complementation at the 3′-untranslated region (3′-UTR) of the transcripts, however, it has been documented that the 5′-UTR may also be targeted (Lee et al., 2009). It is due to this imperfect base pairing that one miRNA is able to have multiple mRNA targets and one transcript can be targeted by multiple miRNAs, thus forming a complex miRNA mediated gene regulatory network (Friedman et al., 2009). The down regulation of specific transcript levels is achieved through destabilization and degradation after miRNA-mRNA pairing (Filipowicz et al., 2008; Guo et al., 2010). Thus, generally a miRNA expression profile is negatively correlated with its targets expression profiles.

The computational prediction of novel miRNAs revolves around miRNA gene identification. Different approaches have been adapted to identify putative miRNAs over the years. The traditional experimental method used to discover miRNAs consisted of cloning size-fractionated RNAs followed by Sanger sequencing and experimental validation. However, this procedure is time consuming, data is noisy (may clone and sequence a large number of degraded RNA fragments from the samples), and may not detect miRNAs that have low expression levels (Ambros et al., 2003; Mendes et al., 2009). Once these miRNAs were cloned, bioinformatic tools were required to locate their origin in the genome; sometimes this may not be a trivial task especially because eukaryotic genomes commonly present hairpin structures that are not necessarily related to miRNAs (Berezikov et al., 2006). Other experimental approaches have also been used to investigate new miRNAs such as Northern blot, microarray, and in situ hybridization, but these approaches are also tedious and time consuming (Table 1). The advent of NGS technology reduced the cost for discovery and offers significant advantage to identify lowly abundant miRNAs. It also provides a more reliable and sensitive method to quantify known miRNAs. However, to discover new miRNAs from NGS data, some kinds of miRNA prediction algorithms with proper computational infrastructure are required (Friedlander et al., 2008). Several computational tools have been developed to complement experimental approaches to identify and validate novel miRNAs from high throughput platforms such as NGS (Huang et al., 2011).

Among the main miRNA characteristics used by different computational tools are their length, high sequence conservation among species, and structural features like hairpin and minimal folding free energy (Li et al., 2010). Several algorithms to obtain putative secondary structure based on minimum free energy, like RNAfold and Mfold, have been used in miRNA identification (Hofacker, 2003; Zuker, 2003). Although cross-species homology is an effective and simpler criterion to discover new miRNAs, it misses the ability of finding non-conserved and species-specific miRNAs, such as the miR-466 cluster in mouse (Hertel and Stadler, 2006). Consequently, computational miRNA identification methods can be divided into two main strategies: comparative and non-comparative algorithms (Hertel and Stadler, 2006; Batuwita and Palade, 2009), although more specific and sub-classifications may be applied, once juxtapositions occur. A comparison between selected tools is summarized in Table 2.

With the advent of novel and powerful informatics infrastructure as well as biocomputational tools, the possibility to discover novel biomolecules and interactions in complex datasets became feasible. Recent studies have shed some light over the function of miRNA, which may lead to numerous potential applications from infectious diseases control, cancer development decrease, and inhibition of protein syntheses to improvement of plant production in the agribusiness (as revised by Pillai, 2005). Gaining some knowledge on the complete spectrum of miRNA is critical in understanding its function and developing miRNA based applications.

Even though methods, as summarized in this review, have been developed in an attempt to identify miRNAs, new algorithms are still in need to improve the ability to find new miRNAs and relate them with their respective functions. For instance, an accurate practical comparison between the methods is difficult because they use different datasets to generate the results and are not always based on the same parameters to evaluate their performance. Therefore, the selection of a methodology to use for a study probably depends on the informaiton available. Many tools still generate a great number of false positives and don’t provide insights into the function or regulatory role of the predicted candidates. The lack of a clear and simpler pipeline to predict and validate miRNA candidates also makes the task of predicting miRNA transcripts and its encoded miRNAs complicated.

Taking in consideration the steps along the process of miRNA maturation could improve such miRNA prediction; however we do not completely understand various steps involved, like how the strand selection of the mature miRNA is decided. Thus, the improvement of computational tools is tightly linked to experimental biological research and vice versa. For example, the comprehension that isomiRs are not artifacts from experimental approaches should be incorporated into prediction tools. With improvement in precision and accuracy of miRNA prediction, the number of miRNAs might possibly be increased significantly, which may provide more comprehensive understanding on miRNA mediated gene/protein regulation.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.