BackgroundMachine learning techniques are known to be a powerful way of distinguishing microRNA hairpins from pseudo hairpins and have been applied in a number of recognised miRNA search tools. However, many current methods based on machine learning suffer from some drawbacks, including not addressing the class imbalance problem properly. It may lead to overlearning the majority class and/or incorrect assessment of classification performance. Moreover, those tools are effective for a narrow range of species, usually the model ones. This study aims at improving performance of miRNA classification procedure, extending its usability and reducing computational time.ResultsWe present HuntMi, a stand-alone machine learning miRNA classification tool. We developed a novel method of dealing with the class imbalance problem called ROC-select, which is based on thresholding score function produced by traditional classifiers. We also introduced new features to the data representation. Several classification algorithms in combination with ROC-select were tested and random forest was selected for the best balance between sensitivity and specificity. Reliable assessment of classification performance is guaranteed by using large, strongly imbalanced, and taxon-specific datasets in 10-fold cross-validation procedure. As a result, HuntMi achieves a considerably better performance than any other miRNA classification tool and can be applied in miRNA search experiments in a wide range of species.ConclusionsOur results indicate that HuntMi represents an effective and flexible tool for identification of new microRNAs in animals, plants and viruses. ROC-select strategy proves to be superior to other methods of dealing with class imbalance problem and can possibly be used in other machine learning classification tasks. The HuntMi software as well as datasets used in the research are freely available at http://lemur.amu.edu.pl/share/HuntMi/.

This paper presents a proposal of a rule induction algorithm selecting a rule quality measure adaptively. The quality measure plays the role of an optimization criterion of the generated rules. Nine quality measures applied by the algorithm are presented and discussed in the paper. It is shown experimentally that the proposed algorithm provides us with obtaining a classifier of the best quality. During experiments, three criteria of the classifier quality were considered: overall accuracy, balanced accuracy (average accuracy of decision classes), and complexity of the classifier (understood to mean the number of induced rules). The experiments were carried out on 34 data sets coming from the UCI machine learning repository. Moreover, a proposal of four-rule filtration algorithms is presented in the paper. Their task is to limit the number of rules in the classifier. In particular, filtration influence on the classifier quality is studied.