Integrating end-to-end learning with deep geometrical potentials for ab initio RNA structure prediction

Area of Science:

• Computational Biology
• Structural Biology
• Molecular Biology

Background:

• Ribonucleic acid (RNA) molecules are essential for cellular functions.
• Their biological activity is dictated by complex three-dimensional (3D) structures.
• Accurate prediction of RNA tertiary structures is a significant challenge in molecular biology.

Purpose of the Study:

• To develop a novel computational method for predicting RNA tertiary structures.
• To improve the accuracy and efficiency of RNA 3D structure modeling.

Main Methods:

• The DRfold method was developed, integrating simultaneous learning of local frame rotations and geometric restraints from experimental RNA structures.
• A hybrid energy potential was created to guide RNA structure assembly.
• Deep end-to-end learning supervised with atom coordinates was employed.

Main Results:

• DRfold significantly outperforms previous methods, achieving over 73.3% improvement in TM-score on a non-redundant dataset.
• The enhanced performance is attributed to deep learning and a composite energy function combining geometric restraints and learned models.
• The DRfold program offers a fast training protocol for large-scale, high-resolution RNA structure modeling.

Conclusions:

• DRfold represents a significant advancement in RNA tertiary structure prediction.
• The method's effectiveness stems from its hybrid approach combining deep learning and geometric principles.
• The open-source DRfold program facilitates broader applications in structural biology and can be further enhanced with larger datasets.