Graph Learning-Based Scoring of RNA-Protein Complex Structures

Development of suitable scoring functions is essential for the prediction of RNA-protein complex structures. Conventional statistical potential-based scoring functions suffered from deficiencies in handling conformational flexibility. The recent application of convolutional neural network (CNN) to this field has shown the potential to address the problem. Compared to CNN, however, the graph deep learning generally exhibited better performance for biomolecule-related structural and functional prediction tasks. Herein, we propose EGARPS+, a novel attempt to apply the graph learning theory to evaluate RNA-protein complex structures. This algorithm comprised the intermolecular and intramolecular modules, which were established on the equivariant graph neural networks and specifically designed attention mechanisms. Additionally, we adopted previously unexplored sequence, structural and interaction features to fully represent interface regions. Our algorithm consistently outperformed the CNN-based approach and traditional statistical potentials on both bound and unbound data sets. The proposed model excelled in processing complexes with larger conformational changes, smaller interface sizes, and lower structural similarities. EGARPS+ could also improve the de novo RNA-protein complex prediction by RoseTTAFoldNA and AlphaFold3. Finally, interpretability analyses underscored the importance of conserved motifs and hydrogen bonding in RNA-protein interactions.