RNA Backbone Torsion and Pseudotorsion Angle Prediction Using Dilated Convolutional Neural Networks

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• RNA 3D structure prediction is computationally intensive.
• Current methods rely on secondary structure restraints, which are limited.
• The conformational space of RNA backbones remains a challenge.

Purpose of the Study:

• To develop a novel method for predicting RNA backbone torsion and pseudotorsion angles.
• To utilize a single RNA sequence as input for angle prediction.
• To improve restraints for RNA 3D structure prediction.

Main Methods:

• Employed a dilated convolutional neural network (SPOT-RNA-1D).
• Trained the model on a high-resolution RNA dataset.
• Tested on three independent, nonredundant, high-resolution datasets.

Main Results:

• SPOT-RNA-1D achieved substantially lower mean absolute errors compared to baseline predictors.
• Prediction accuracy ranged from 14°-44° across different angles.
• The method accurately recovered overall patterns of angle distributions.

Conclusions:

• SPOT-RNA-1D provides accurate dihedral angle predictions for RNA.
• These predictions can serve as valuable restraints for RNA structure prediction.
• The method shows potential as a model quality indicator in RNA structure modeling.