Evaluating native-like structures of RNA-protein complexes through the deep learning method.

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• RNA-protein complexes are crucial for cellular functions like translation and gene regulation.
• Determining the high-resolution structures of these complexes is vital for understanding their roles.
• Existing computational methods struggle with accurately predicting native-like RNA-protein structures, especially those involving conformational changes.

Purpose of the Study:

• To develop a novel deep-learning-based computational method for identifying native-like RNA-protein structures.
• To evaluate the performance of this new method across a range of RNA-protein complex scenarios, from rigid to flexible docking.
• To improve the accuracy of predicting native interfaces in RNA-protein complexes.

Main Methods:

• Development of DRPScore, a deep-learning approach for RNA-protein structure identification.
• Testing DRPScore on diverse sets of RNA-protein complexes, including bound-bound (rigid) and unbound-unbound (flexible) docking scenarios.
• Comparative analysis of DRPScore's performance against existing computational methods.

Main Results:

• DRPScore achieved a 91.67% success rate in identifying native-like structures within the top 20 predictions for bound complexes.
• For unbound complexes, DRPScore showed a 56.14% success rate, outperforming existing methods.
• DRPScore demonstrated a significant improvement of 10.53-15.79% over current methods, even for challenging unbound cases.
• Enhanced accuracy in predicting native interface interactions within RNA-protein complexes.

Conclusions:

• DRPScore is an effective deep-learning tool for identifying native-like RNA-protein structures.
• The method shows superior performance compared to existing approaches, particularly for flexible complexes.
• DRPScore has broad applicability for the modeling and rational design of RNA-protein complexes.