Biological features between miRNAs and their targets are unveiled from deep learning models

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• MicroRNAs (miRNAs) are key gene regulators involved in cellular processes and diseases.
• Identifying miRNA targets is crucial for understanding and treating miRNA-associated conditions.
• Previous work developed the miTAR deep learning model for enhanced miRNA target prediction.

Purpose of the Study:

• To investigate the biological features learned by the miTAR deep learning model for miRNA:target interactions.
• To determine if deep learning models can reveal underlying biological mechanisms in miRNA function.

Main Methods:

• Utilized a hybrid deep learning approach (miTAR) integrating Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs).
• Analyzed the features captured by miTAR, focusing on known elements like seed regions and free energy, and exploring novel features.

Main Results:

• miTAR successfully identifies known miRNA:target interaction features, including seed region binding and free energy.
• The model also captures previously unrecognized features of miRNA:target interactions.
• CNN and RNN layers exhibit distinct capabilities in learning the free energy feature, with RNNs being more unique and CNNs more efficient.

Conclusions:

• Deep learning models like miTAR can unveil complex biological features in miRNA:target interactions, moving beyond the 'black-box' perception.
• These findings contribute to a deeper understanding of miRNA gene regulation mechanisms.
• The ability to interpret deep learning models aids in the development of targeted therapies for miRNA-related diseases.