Machine learning a model for RNA structure prediction.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• RNA structure is critical for its function.
• Current thermodynamic models for RNA secondary structure prediction can be inaccurate and often require experimental data.
• Existing models struggle to identify native RNA structures without auxiliary information.

Purpose of the Study:

• To develop an improved computational model for RNA secondary structure prediction.
• To integrate diverse data types including thermodynamic parameters, chemical probing (DMS, SHAPE), and co-evolutionary data (DCA) into a unified network.
• To enhance the accuracy of predicting native RNA structures and their ensemble populations.

Main Methods:

• A novel network model was constructed to combine thermodynamic parameters with chemical probing and co-evolutionary data.
• The network outputs perturbations to the ensemble free energy, trained to increase populations of known native RNA structures.
• Convolutional windows were used in chemical probing nodes to capture local conformational ensemble information; regularization and cross-validation were employed for model selection and transferability.

Main Results:

• The developed network model successfully increased ensemble populations for known native RNA structures.
• The model achieved more accurate RNA structure predictions on an independent validation set compared to previous methods.
• The approach demonstrated good transferability across different RNA systems.

Conclusions:

• Integrating multiple data sources via a network model significantly improves RNA secondary structure prediction accuracy.
• The model's flexibility allows for adaptation and retraining with new experimental data.
• This approach offers a powerful tool for understanding RNA structure-function relationships.