Abstract: In microRNA gene targets prediction there are several problems, such as the contradiction between sensitivity and specificity, contradiction between masculine threshold and test results, and the narrowing of predict scope. To solve these problems, this article discusses three new kinds of feature extraction methods: energy calibration, element ratio in seed regions, and punishment in target binding structure. These feature extraction methods cover statistics, structure and energy information. Combing the traditional and the new features, we construct classification models and test them through neural network. The results show that the new methods gather positive data and negative data respectively as far as possible, and ensure a good evaluation under higher threshold. These methods are suitable for target genes prediction of various species without increase in spending.

Key words: computer application, bioinformatics, classification of target genes, feature extraction method, machine learning, threshold of positive target genes