Aiming at identifying miRNAs from genomes with a few known miRNA and/or little annotation, we propose and develop a novel miRNA prediction method, miRank, based on our new random walks- based ranking algorithm. We first tested our method on Homo sapiens genome; using a very few known human miRNAs as samples, our method achieved a prediction accuracy greater than 95%. We then applied our method to predict 200 miRNAs in Anopheles gambiae, which is the most important vector of malaria in Africa. Our further study showed that 78 out of the 200 putative miRNA precursors encode mature miRNAs that are conserved in at least one other animal species. These conserved putative miRNAs are good candidates for further experimental study to understand malaria infection.

Availability: MiRank is programmed in Matlab on Windows platform. The source code is available upon request.

Identification of novel miRNA genes is an eminent and challenging problem towards the understanding of post-transcriptional gene regulation. Two major strategies for identifying novel miRNAs are experimental cloning and in silico prediction (Bartel, 2004; Berezikov et al., 2006; Jones-Rhoades et al., 2006). In a cloning-based approach, distinct ~22 nt RNA transcripts were first isolated and then intensively cloned and sequenced. However, these methods are highly biased towards abundantly and/or ubiquitously expressed miRNAs; only abundant miRNA genes can be easily detected (Bartel, 2004; Berezikov et al., 2006; Jones-Rhoades et al., 2006). Note that not all miRNAs are well expressed in many tissues, cell types and developmental stages that have been tested (Bartel, 2004). Some miRNAs are expressed constitutively at low abundance, and some miRNAs have preferential or specific temporal and spatial expression patterns. Computational approaches have been developed to overcome some of the technical difficulties of experimental methods. First, in silico methods have been shown to be efficient for finding miRNAs that are expressed at constitutively low levels or in specific tissues (Bartel, 2004). Further, breakdown products of mRNA transcripts, other endogenous ncRNAs (e.g. tRNAs, rRNAs, nat-siRNAs) as well as exogenous siRNAs comprise a large portion of the non-coding small RNA population isolated from the cytoplasmic total RNA extracts (Bartel 2004; Berezikov et al. 2006; Jones-Rhoades et al. 2006; Lagos-Quintana et al., 2001). To avoid erroneously designation of other non-coding small RNAs and even broken mRNA fragments as novel miRNAs, the secondary structures of flanking genomic sequences of cloned small RNAs are further assessed computationally. Only those small RNA sequences reside on arms of hairpin-structured precursors are likely to be miRNAs. However, due to their short length, cloned small RNAs may match to many genome regions that can potentially fold into hairpin structures which may not be unique to miRNAs. Thus, genome-wide screening of novel miRNA precursors is technically involved.

Based on particular strategies adopted, the existing computational methods for miRNA prediction can be grouped into two categories.

Methods in the first category were developed to find homologous miRNAs in closely related species. In these methods, known miRNA precursors were first folded into typical hairpin structures, local features in the hairpins were extracted, and extreme values of these featured were obtained from all known miRNAs. A filter was then constructed to screen novel hairpinned sequences. Those hairpinned sequences that passed the filter were further analyzed in related species to assess their evolutionary conservation (Berezikov et al., 2005; Bonnet et al., 2004; Grad et al., 2003; Lai et al., 2003). As they were designed for, these conservation-based methods built upon this framework were successful in detecting evolutionarilly conserved miRNAs. In order to further identify distantly homologous miRNAs, a probabilistic co-learning method using a paired hidden Markov model (HMM) was recently developed as a more general miRNA prediction method (Nam et al., 2005). All these methods relied mainly on evolutionary conservation to eliminate a large number of false positive predictions. However, a substantial number of lineage- or species-specific miRNA genes do exist (Fahlgren et al., 2007; Kasschau et al., 2007; Lindow and Krogh, 2005; Molnar et al., 2007; Rajagopalan et al., 2006; Zhang et al., 2006), which escape the detection of a conservation-based approach.

In the second category, supervised classification methods, e.g. Support Vector Machines (SVMs), were adopted to train models based on positive sets of genuine miRNA precursors and negative sets of hairpins obtained from exon regions of protein coding genes (Hertel and Stadler, 2006; Ng and Mishra, 2007; Xue et al., 2005). Thanks to the sequence features extracted from the training examples, such classification models were expected to preform well in predicting novel miRNAs from unseen sequences. However, many sequenced genomes are poorly annotated. Hence, it is difficult to obtain a good set of negative training samples for these methods. Moreover, miRNAs are located in intergenic regions and introns. A major task for miRNA prediction is to distinguish miRNA hairpins from other hairpin structures originated from intergenic or intronic sequences. In summary, classification-based methods are not effective on not well-annotated genomes.

The methods in both categories discussed above all require a non-trivial set of positive samples of known miRNAs. Unfortunately, except well-studied model species such as human and Caenorhabditis elegans, most sequenced genomes have a small number of miRNAs reported. For instance, only 38 miRNAs have been identified in the genome of A.gambiae. This is particularly true for viral genomes for which not many miRNAs have been studied or reported (Stern-Ginossar et al., 2007). Furthermore, the small number of known viral miRNAs rarely share any sequence homology (Stern-Ginossar et al., 2007). Yet, viral miRNAs have been shown to play important roles in the pathology of viral infection by targeting host immune systems (Stern-Ginossar et al., 2007).

In short, the importance of miRNAs in post-transcriptional regulation, the lack of a sufficient number of known miRNAs and poorly annotated genomes collectively call for novel effective computational approaches to miRNA prediction.

For a given species, the known miRNAs, putative candidate miRNAs and their relationship can be modeled by a weighted graph G=V, E. Each vertex vV in G represents a known miRNA precursor or a putative candidate, an edge eEV×V captures the relation between two vertices linked by the edge, and the weight w of edge e quantifies the relation. In general, edge weights are determined by pairwise distances. For example, two closely related samples may share an edge with a large weight. The degree of vertex v_i is d_i=∑_j w_ij, i.e. the total weight of all edges that are connected to v_i. Note that the graph is fully connected, so that there always exists a path between each pair of nodes in the graph.

The miRank method has the following remarkable properties. First, it does not require information of genome annotation. This is particularly important because many sequenced genomes have not been well annotated, and their closely related species are yet to be sequenced. Thus, a large number of false positive candidates with hairpinned secondary structures cannot be filtered out with genome annotation or by phylogenetic conservation. miRank can be applied to such newly sequenced genomes with little annotation. Second, it does not rely on cross-species conservation so that it can identify species-specific miRNAs. Third, miRank is able to accommodate a small number of known miRNAs while enjoys a high-prediction accuracy. Hence, miRank is a useful tool for many species including most viruses that have a very few reported miRNAs. Identification of novel miRNAs in various viral genomes will be our future research topic.