Improving RNA Secondary Structure Prediction Through Expanded Training Data.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Biology

Background:

• Deep learning has advanced protein structure prediction but RNA structure prediction lags.
• Current RNA structure models lack the accuracy of protein models, potentially due to limited training data.
• Existing methods show moderate success but haven't achieved high accuracy for RNA secondary and tertiary structures.

Purpose of the Study:

• To develop a large, diverse dataset for RNA structure prediction to address data limitations.
• To evaluate the impact of this enhanced dataset on deep learning models for RNA structure prediction.
• To improve the accuracy and generalization of de novo RNA secondary structure prediction.

Main Methods:

• Development of the RNASSTR dataset, comprising paired RNA sequences and secondary structures.
• Retraining of two deep learning models, SincFold and MXfold2, using the RNASSTR dataset.
• Assessment of model performance and generalization on unseen RNA families.

Main Results:

• SincFold demonstrated improved generalization to new RNA families after retraining on RNASSTR.
• The enhanced dataset improved SincFold's capability for accurate de novo RNA secondary structure prediction.
• Retraining MXfold2 was computationally expensive and did not yield high performance on the test set.

Conclusions:

• The RNASSTR dataset represents a significant advancement for RNA structure modeling.
• This dataset provides a strong foundation for developing future RNA secondary structure prediction algorithms.
• Addressing data limitations is crucial for advancing RNA structure prediction accuracy.