Learning from positive examples when the negative class is undetermined--microRNA gene identification

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Machine learning for classification using only positive examples is emerging in computational biology.
• Traditional two-class methods for microRNA discovery require artificial negative classes, which can be problematic and bias performance.
• One-class machine learning offers an alternative by not requiring a negative class.

Purpose of the Study:

• To investigate the utility of one-class machine learning for microRNA discovery.
• To compare the performance of one-class methods against two-class approaches (naïve Bayes, Support Vector Machines).
• To evaluate the ability of both methods to identify microRNAs in newly sequenced species.

Main Methods:

• Implemented and evaluated one-class machine learning algorithms.
• Compared one-class performance with two-class naïve Bayes and Support Vector Machines using selected features.
• Validated methods on the EBV genome for identifying known and predicting novel microRNAs.

Main Results:

• Two-class methods (naïve Bayes, SVM) achieved higher accuracy (approx. 90%) with optimal negative examples.
• One-class methods demonstrated average accuracies of 70-80% on the same feature sets.
• Some one-class methods outperformed published two-class approaches with different features and performed comparably in external validation.

Conclusions:

• Both one-class and two-class methods yield useful accuracies when negative classes are well-defined.
• One-class methods excel when optimal negative class features are not well-defined, avoiding performance bias.
• One-class approaches can be superior when positive class features are clearly defined, simplifying microRNA discovery.