Automatic learning of pre-miRNAs from different species

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNA (miRNA) and miRNA precursor (pre-miRNA) discovery relies on predictive models.
• Existing tools exhibit species-specific biases, limiting their cross-species applicability.
• Characterizing these biases is key to developing more accurate computational approaches.

Purpose of the Study:

• To evaluate the performance of 45 pre-miRNA predictive models across different species.
• To investigate the species-dependency of pre-miRNA and pseudo pre-miRNA classification.
• To develop a computational approach to mitigate species-specific biases in miRNA prediction.

Main Methods:

• Tested 45 pre-existing predictive models on datasets from 45 different species.
• Analyzed the species-dependency of feature set performance for pre-miRNA classification.
• Developed and evaluated an ensemble of classifiers to improve prediction accuracy.

Main Results:

• Pre-miRNA and pseudo pre-miRNA separability is species-dependent; no single feature set is universally effective.
• An ensemble of classifiers significantly reduced classification errors across all 45 tested species.
• The proposed ensemble approach offers a lower computational cost compared to methods using energy stability parameters.

Conclusions:

• Combining multiple feature sets and learning biases enhances pre-miRNA classifier accuracy for diverse species.
• This approach offers a promising strategy for developing more accurate and less species-dependent miRNA discovery tools.
• The study provides a method to improve computational identification of novel miRNAs.