Multi-view Co-training for microRNA Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are key regulators of gene expression.
• Current computational miRNA prediction methods rely heavily on large labeled datasets, which are scarce for many species.
• Unlabeled high-throughput genomic and RNA expression data are increasingly available but underutilized by existing supervised methods.

Purpose of the Study:

• To develop a novel multi-view co-training approach for miRNA classification.
• To enhance the utility of unlabeled data in miRNA prediction.
• To improve miRNA classification accuracy, particularly for understudied species.

Main Methods:

• A multi-view co-training framework was designed to integrate sequence- and expression-based features.
• The approach iteratively trains classifiers using both labeled and unlabeled data.
• A final confidence-based classifier was developed by integrating sequence and expression views.

Main Results:

• Co-training significantly improved miRNA classification accuracy (p < 0.01) with minimal labeled data.
• Expression-based classification saw an average AUPRC increase of 15.81% via co-training, outperforming self-training and passive learning.
• Sequence-based classification showed even greater gains with co-training (46.47% increase).
• The integrated classifier outperformed individual sequence and expression classifiers.

Conclusions:

• Multi-view co-training is a promising method for miRNA prediction, especially when labeled data is limited.
• This approach effectively harnesses large unlabeled datasets for improved accuracy.
• The study demonstrates the potential of co-training for advancing miRNA research in diverse species.