Rich parameterization improves RNA structure prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Current RNA structure prediction methods utilize physics-based or machine learning (ML) approaches.
• While ML is adopted for parameter estimation, model parameterizations have remained largely unchanged.
• There's a need to enhance RNA folding prediction models by incorporating more information.

Purpose of the Study:

• To investigate the impact of increased information utilization on RNA folding prediction model quality.
• To develop novel, fine-grained RNA folding prediction models.
• To create a new, high-performance RNA folding prediction model with improved accuracy.

Main Methods:

• Proposed novel ML models that examine more structural element types and larger sequential contexts.
• Leveraged large training sets, ML advancements, and accelerated RNA folding algorithms.
• Developed and evaluated a new RNA folding prediction model with approximately 70,000 parameters.

Main Results:

• The fine-grained models demonstrated improved prediction quality.
• The running time of the folding algorithm remained efficient.
• The new model achieved an 84% F₁-measure score for correctly-predicted base-pairs, reducing the error rate by approximately 50% compared to previous models (70% F₁-score).

Conclusions:

• Increasing the information processed by RNA folding prediction models significantly enhances prediction accuracy.
• The developed model offers a substantial improvement in RNA structure prediction performance.
• Trained models and source code are publicly available for broader research use.