Improving classification of mature microRNA by solving class imbalance problem

Area of Science:

• Bioinformatics
• Molecular Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are crucial non-coding RNAs regulating gene expression post-transcriptionally.
• Accurate identification of mature miRNA start sites from precursor miRNAs (pre-miRNAs) is challenging.
• Mature miRNA prediction is often hindered by class-imbalanced datasets, leading to suboptimal performance.

Purpose of the Study:

• To enhance the accuracy of mature miRNA identification from pre-miRNAs.
• To address the class-imbalance problem in mature miRNA prediction.
• To present MatFind, an effective computational tool for identifying 5' mature miRNA candidates.

Main Methods:

• Utilized the K-nearest neighbor algorithm to extract balanced datasets.
• Trained multiple support vector machine (SVM) classifiers sequentially on balanced datasets using representative features.
• Developed an ensemble classifier by combining individual SVM classifiers, incorporating the AdaBoost concept.

Main Results:

• The proposed method, MatFind, demonstrated improved efficiency compared to approaches not addressing class imbalance.
• MatFind achieved significantly higher classification accuracy on an independent testing dataset than three other methods.
• Ensemble SVM classifiers and balanced datasets effectively resolved the class-imbalance issue, boosting classifier performance.

Conclusions:

• Ensemble SVM classifiers and balanced datasets are effective strategies for improving mature miRNA identification.
• MatFind provides an accurate and rapid method for identifying 5' mature miRNA candidates.
• The study highlights the importance of addressing class imbalance in miRNA prediction tasks.