miRNAss

miRNAss predicts pre-miRNA hairpins from stem-loop sequences using a semi-supervised learning approach to improve genome-wide microRNA discovery.

Key Features:

• Semi-Supervised Learning Approach: miRNAss leverages both labeled and unlabeled stem-loop sequences to improve prediction accuracy when known pre-miRNA examples are limited.
• Handling of Negative Examples: The method includes an automatic procedure for identifying negative examples to differentiate miRNA hairpins from non-miRNA sequences without manual curation.
• Class Imbalance Management: miRNAss is designed to address extreme class imbalance in genomic datasets, where pre-miRNAs can be outnumbered by non-pre-miRNA sequences by approximately 1:10,000.
• Efficiency and Speed: The approach achieves higher prediction rates and reduced execution times compared to existing supervised methods, enabling large-scale analyses.

Scientific Applications:

• Genome-wide pre-miRNA prediction and validation: Validated on genome-wide datasets from three model species, each containing over one million hairpin sequences, for microRNA discovery and characterization.

Methodology:

Semi-supervised learning that integrates information from unlabeled stem-loop sequences to refine model predictions; the method includes an automatic search for negative examples and provisions to handle severe class imbalance (~1:10,000).