Machine learning approach for ab initio prediction of microRNA precursors

Although comparative genomics based methods provided important techniques to predict new miRNAs, it is unable to identify novel miRNAs for which there are no known close homologies. It is a fact that almost all pre - miRNAs have the characteristic of stem - loop hairpin structures. Therefore those hairpin structures give key clues to the ab initio prediction of pre - miRNAs. However, a large amount of pre - rniRNA - like hairpins can be folded in many genomes. It is challenging to distinguish the real pre - miRNAs from other hairpin sequences with similar stem - loops (pseudo pre - miRNAs). In this paper, to distinguish the real pre - miRNAs from other hairpin sequences with similar stem - loops (pseudo pre - miRNAs), we proposed a novel machine learning method: random forest. Coupled with a hybrid feature which consists of local contiguous structure - sequence composition, minimum of free energy (MFE) of the secondary structure and p - value of randomization test, the prediction model achieves 98.21 % specificity and 95.09% sensitivity.