DeepRMSF: a deep learning-based automated approach for predicting atomic-level flexibility in RNA structure.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Understanding RNA conformational dynamics is crucial for deciphering its biological functions.
• Predicting RNA local flexibility from static structures remains a significant computational challenge.
• Existing methods struggle to efficiently predict dynamic properties of RNA.

Purpose of the Study:

• To develop a deep learning-based method, DeepRMSF, for predicting RNA vibrational flexibility.
• To provide a computationally efficient tool for assessing RNA local dynamics.
• To facilitate large-scale analysis of RNA flexibility across transcriptomes.

Main Methods:

• Developed DeepRMSF, a deep learning model utilizing atomic-level RNA descriptions.
• Trained the model on root-mean-square fluctuations (RMSF) derived from molecular dynamics (MD) simulations.
• Benchmarked DeepRMSF on 371 nonredundant RNA structures using rigorous cross-validation and an independent test set.

Main Results:

• DeepRMSF accurately predicts RNA vibrational flexibility with high correlation (PCC ~0.73-0.75) on independent test sets.
• Achieved a >3000-fold speed-up compared to traditional MD simulations for flexibility prediction.
• Demonstrated strong extrapolative accuracy for medium-sized RNAs (~75 nucleotides), predicting flexibility in ~8.2 seconds.

Conclusions:

• DeepRMSF provides a scalable and practical approach for transcriptome-wide RNA flexibility screening.
• The method complements MD simulations, offering a faster alternative for analyzing RNA dynamics.
• Facilitates deeper understanding of RNA structure-dynamics-function relationships and aids computational RNA biology.