Improved Pre-miRNA Classification by Reducing the Effect of Class Imbalance

Abstract

MicroRNAs (miRNAs) play important roles in the diverse biological processes of animals and plants. Although the prediction methods based on machine learning can identify nonhomologous and species-specific miRNAs, they suffered from severe class imbalance on real and pseudo pre-miRNAs. We propose a pre-miRNA classification method based on cost-sensitive ensemble learning and refer to it as MiRNAClassify. Through a series of iterations, the information of all the positive and negative samples is completely exploited. In each iteration, a new classification instance is trained by the equal number of positive and negative samples. In this way, the negative effect of class imbalance is efficiently relieved. The new instance primarily focuses on those samples that are easy to be misclassified. In addition, the positive samples are assigned higher cost weight than the negative samples. MiRNAClassify is compared with several state-of-the-art methods and some well-known classification models by testing the datasets about human, animal, and plant. The result of cross validation indicates that MiRNAClassify significantly outperforms other methods and models. In addition, the newly added pre-miRNAs are used to further evaluate the ability of these methods to discover novel pre-miRNAs. MiRNAClassify still achieves consistently superior performance and can discover more pre-miRNAs.