Fast and accurate microRNA search using CNN.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are crucial regulatory molecules, and their functional annotation is an active research area.
• Classifying newly discovered miRNAs into families aids in understanding their roles.
• Deep learning models offer a promising approach for miRNA classification due to accumulated annotated data.

Purpose of the Study:

• To investigate the effectiveness of deep learning models for microRNA (miRNA) classification.
• To determine if incorporating secondary structure information improves classification accuracy.
• To assess the ability of softmax output to distinguish miRNA sequences from non-miRNA sequences and handle small miRNA families.

Main Methods:

• Developed and trained convolutional neural network (CNN) models for miRNA classification.
• Employed two encoding methods: explicit secondary structure matrix and one-hot encoding of primary sequence.
• Utilized a softmax output threshold to identify and exclude out-of-distribution (non-miRNA) sequences.

Main Results:

• CNN models achieved comparable sensitivity and accuracy to state-of-the-art tools like Infernal for miRNA classification.
• The models demonstrated significantly faster performance compared to existing methods.
• Secondary structure-based encoding and softmax output effectively improved classification and discrimination of non-miRNA sequences.

Conclusions:

• Automatic feature learning in CNNs enhances miRNA classification accuracy and sensitivity.
• The developed deep learning models provide an efficient and accurate method for miRNA functional annotation.
• The models and associated code are publicly available for research use.