A Kernel Method for MicroRNA Target Prediction Using Sensible Data and Position-Based Features

Abstract

MicroRNAs (miRNAs) are small endogenous RNAs of ~ 22nt that act as direct post-transcriptional regulators in animals and plants. MicroRNAs generally perform a function by binding to the complementary site on the 3’ untranslated region of its target gene and especially the 8mers on the 5’ part of miRNA seems important as a seed. Computational methods for miRNA target prediction have been focusing on this seed region, but recent researches revealed that the specificity of the seed region may be sharply decreased even by a point mutation. In this paper, we present a kernel method for miRNA target prediction in animals, which improves the prediction performance with biologically sensible data and position-based features reflecting the way of miRNA: mRNA pairing mechanism. In building a training dataset, we choose experimentally verified data only to improve the quality of dataset by excluding randomly synthesized one and consequently to make the result of learning valid. We use sensitivity, specificity, and area under ROC curve as performance measures of our algorithm and compare the results of various dataset configurations. The overall results were 92.1% in sensitivity, 83.3% in specificity, and 0.931 in area under ROC curve. With position-based features, an increase of 3.3% in sensitivity and 1.6% in specificity were observed.