MBSTAR: multiple instance learning for predicting specific functional binding sites in microRNA targets.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) regulate gene expression by binding to mRNA targets.
• Current machine learning methods for miRNA target prediction suffer from high false positive rates due to unvalidated binding sites.

Purpose of the Study:

• To develop an accurate machine learning approach, MBSTAR, for predicting functional miRNA binding sites.
• To improve the reliability of miRNA target prediction and reduce experimental validation costs.

Main Methods:

• Utilized a multiple instance learning framework to address the challenge of identifying actual binding sites in target mRNAs.
• Trained the model on 9531 biologically validated interacting and 973 non-interacting miRNA-mRNA pairs from Tarbase 6.0.
• Validated predictions using PAR-CLIP data.

Main Results:

• MBSTAR demonstrated the highest overlap with PAR-CLIP data, achieving an F-Score of 0.337.
• MBSTAR outperformed existing methods in predicting target mRNAs with superior accuracy.
• The tool and genome-wide predictions are publicly available online.

Conclusions:

• MBSTAR offers a significant advancement in accurately identifying functional miRNA binding sites.
• The developed method enhances the precision of miRNA target prediction, facilitating more effective experimental validation.