MicroRNA Precursor Prediction Using SVM with RNA Pairing Continuity Feature

Background

MicroRNAs (miRNAs), one of the non-coding RNA families, are a class of endogenous, single-stranded, small (19-27nt) nucleic acids that have an extremely conserved structure. In the nucleus, a gene encoding miRNA is first transcribed into a pri-microRNA, which is cut into pre-miRNA with a hairpin (stem-loop) structure (Bartel, 2004) of about 70nt in length using an enzyme called “Drosha RNase” (Lee et al., 2003). The mature miRNAs are derived from cleavage of pre-miRNA by the Dicer enzyme out of the nucleus. miRNAs are essential in animal and plant development (Bartel, 2004), stress response in plants (Bari et al., 2006), and various diseases including cancers (Blenkiron et al., 2007). They also play key roles in the regulation and control of a variety of the metabolic processes of different organisms (Hua & Xiao, 2005).

Since 2003, thousands of miRNAs have been experimentally identified. At the same time, a variety of prediction methods have been developed too. For examples, miRScan relies on the observation that the known miRNAs are derived from phylogenetically conserved stem-loop precursor RNAs (Lim et al., 2003). Xue et al. (2005) proposed an SVM-based method for classification of real and pseudo pre-miRNAs based on local contiguous structure-sequence composition feature. Nam et al. (2005; 2006) constructed hidden Markov models (HMM) to search for distant homologs of miRNA families. Yousef et al. (2006) used a Naıve Bayes classifier along with the integration of data from multiple species to predict miRNA genes. Jiang et al. (2007) tried to construct a classifier, called MiPred, between per-miRNA and pseudo miRNA using Random forest and P-value. As suggested by Helvik et al. (2007), the miRNA gene prediction methods can also be improved by reliable predictions of Drosha-processing sites. Aiming at identifying miRNAs from genomes with a few known miRNA, Xu et al. (2008) proposed and developed a novel miRNA prediction method, miRank, with a novel ranking algorithm based on random walks. RNAmicro (Hertel & Stadler, 2006) was designed to classify the surveys of large-scale comparative genomics for predicting putative RNAs. Recently, Ahmed et al. (2009) demonstrated that guide and passenger strands of miRNA precursors can be distinguished using nucleotide sequence and secondary structures.

However, new problems also occurred: the features extracted using the existing forecasting methods are limited, and the range of species that can be predicted is small. Some methods can only predict one type of pre-miRNAs from either animal, plant or bacteria; some can predict pre-miRNAs with only a single loop or with only multi-loops; some can only predict a few designated species. Therefore, to overcome these existing difficulties, this chapter presents a new feature for pre-miRNA classification. We have implemented this feature in the Triplet-SVM model with satisfactory results.