Accurate RNA 3D structure prediction using a language model-based deep learning approach.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Accurate prediction of RNA three-dimensional (3D) structures is a significant challenge due to RNA's inherent flexibility and limited experimental data.
• Understanding RNA 3D structures is vital for elucidating biological functions and advancing RNA-targeted therapeutics and synthetic biology.

Purpose of the Study:

• To introduce RhoFold+, a novel deep learning method for accurate, end-to-end prediction of single-chain RNA 3D structures from sequences.
• To address the challenge of data scarcity in RNA structure prediction through advanced computational techniques.

Main Methods:

• Development of RhoFold+, an RNA language model-based deep learning approach.
• Pretraining an RNA language model on a large dataset of approximately 23.7 million RNA sequences.
• Implementation of techniques to mitigate the impact of limited experimental data.

Main Results:

• RhoFold+ demonstrates superior performance in predicting RNA 3D structures compared to existing methods, including human expert groups, as validated on RNA-Puzzles and CASP15 datasets.
• The method shows strong generalizability across different RNA families and types, confirmed by cross-family, cross-type, and time-censored benchmark assessments.
• RhoFold+ also accurately predicts RNA secondary structures and interhelical angles, offering valuable features for RNA research.

Conclusions:

• RhoFold+ represents a significant advancement in computational RNA structure prediction, offering a robust and automated solution.
• The method's accuracy and generalizability provide a powerful tool for RNA structure and function studies, drug development, and synthetic biology.
• The prediction of secondary structures and interhelical angles enhances RhoFold+'s utility for empirical validation and broader applications in RNA biology.