High precision in microRNA prediction: A novel genome-wide approach with convolutional deep residual networks

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• MicroRNAs (miRNAs) are crucial regulators of gene expression.
• Accurate computational prediction of miRNA precursors (pre-miRNAs) is essential.
• Existing homology-based and machine learning methods have limitations in precision and handling secondary structures.

Purpose of the Study:

• To develop a novel computational approach for precise pre-miRNA prediction.
• To address the limitations of existing methods, particularly in real-world genomic data.
• To improve the accuracy and reliability of identifying novel miRNA candidates.

Main Methods:

• Development of a convolutional deep residual neural network named mirDNN.
• Testing mirDNN on full genomes of various animal and plant species.
• Implementation of a novel validation methodology for assessing performance on new species.

Main Results:

• mirDNN achieved up to 5 times higher precision than other approaches at similar recall rates.
• The model demonstrated robust performance across diverse animal and plant genomes.
• The novel validation methodology confirmed the practical applicability of mirDNN.

Conclusions:

• mirDNN offers a significant advancement in the computational prediction of pre-miRNAs.
• The approach overcomes limitations of traditional methods by considering secondary structures and improving robustness.
• A web demo and source code are available for accessibility and further research.