RNA secondary structure prediction using deep learning with thermodynamic integration.

Area of Science:

• Computational biology
• Bioinformatics
• Molecular biology

Background:

• Accurate RNA secondary structure prediction is crucial for understanding functional non-coding RNA roles.
• Machine learning models offer high prediction accuracy but risk overfitting due to extensive parameterization.

Purpose of the Study:

• To minimize overfitting in deep learning models for RNA secondary structure prediction.
• To enhance the accuracy and robustness of RNA structure predictions using integrated thermodynamic data.

Main Methods:

• Developed MXfold2, a deep neural network algorithm integrating RNA folding scores with Turner's nearest-neighbor free energy parameters.
• Employed thermodynamic regularization during model training to align predicted folding scores with calculated free energy.
• Evaluated algorithm performance on newly discovered non-coding RNAs through computational experiments.

Main Results:

• MXfold2 demonstrated minimized overfitting compared to other algorithms.
• The algorithm achieved robust and accurate RNA secondary structure predictions.
• MXfold2 maintained computational efficiency.

Conclusions:

• Integrating thermodynamic information with deep learning improves the robustness of RNA secondary structure predictions.
• MXfold2 provides a reliable and efficient tool for predicting RNA secondary structures, aiding in the study of functional non-coding RNAs.