Graph Learning-Based Scoring of RNA-Protein Complex Structures.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Structural Biology

Background:

• Accurate scoring functions are crucial for predicting RNA-protein complex structures.
• Traditional methods struggle with conformational flexibility.
• Convolutional Neural Networks (CNNs) show promise but graph deep learning offers superior performance for biomolecular tasks.

Purpose of the Study:

• To develop a novel graph learning-based scoring function for RNA-protein complex structures.
• To enhance the evaluation of intermolecular and intramolecular interactions within these complexes.
• To improve the accuracy and robustness of RNA-protein structure prediction.

Main Methods:

• Proposed EGARPS+, a graph learning algorithm utilizing equivariant graph neural networks and attention mechanisms.
• Incorporated novel sequence, structural, and interaction features for interface representation.
• Developed separate intermolecular and intramolecular modules for comprehensive evaluation.

Main Results:

• EGARPS+ consistently outperformed CNN-based methods and statistical potentials on both bound and unbound datasets.
• The model demonstrated superior performance on complexes with significant conformational changes, small interfaces, and low structural similarity.
• EGARPS+ enhanced de novo RNA-protein complex prediction when integrated with RoseTTAFoldNA and AlphaFold3.

Conclusions:

• Graph deep learning, specifically EGARPS+, offers a powerful approach for scoring RNA-protein complex structures.
• The model's ability to handle complex cases and improve existing prediction tools highlights its significance.
• Interpretability analysis revealed the importance of conserved motifs and hydrogen bonding in RNA-protein interactions.