A framework for improving microRNA prediction in non-human genomes

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Predicting pre-microRNA (miRNA) from genomic sequences is crucial but largely focused on human genomes.
• Existing methods show high sensitivity but drastically low specificity when applied to non-human genomes due to a high ratio of pseudo-miRNA sequences.
• This low specificity limits the practical application of current tools for non-human miRNA discovery.

Purpose of the Study:

• To introduce a novel framework (SMIRP) for developing species-specific miRNA prediction systems.
• To address the significant drop in specificity observed when applying human-trained models to non-human genomes.
• To improve the precision and specificity of miRNA prediction while maintaining sensitivity across diverse species.

Main Methods:

• Developed the SMIRP framework incorporating sequence conservation and phylogenetic distance information.
• Applied the SMIRP framework to three distinct prediction systems, including support vector machine and Random Forest classifiers.
• Utilized three different feature sets and both human-specific and taxon-wide training data for evaluation.

Main Results:

• Achieved substantial improvements in specificity and precision for four non-human test species.
• Demonstrated the effectiveness of SMIRP across different prediction systems, feature sets, and training data types.
• Showcased the framework's ability to significantly enhance prediction accuracy for non-human genomic sequences.

Conclusions:

• The SMIRP framework provides a robust solution for species-specific miRNA prediction.
• It significantly boosts specificity and precision, overcoming limitations of existing methods in non-human genomes.
• SMIRP is broadly applicable to various miRNA prediction systems and machine learning techniques, promising improved accuracy for genomic research.