DeepFoldRNA: A Graph Neural Network for RNA 3D Structure Prediction

Abstract

DeepFoldRNA presents a novel graph neural network architecture for RNA 3D structure prediction, integrating evolutionary information with physics-based refinement. By representing RNA molecules as graphs and incorporating co-evolutionary data through multiple sequence alignments, the method naturally captures complex structural interactions. Benchmarking on RNA-Puzzles dataset demonstrated average RMSD of 3.21 Å and TMscore of 0.82, achieving 28% improvement over existing methods. Performance gains were particularly notable for riboswitches and ribozymes with 35% RMSD reduction. Rosetta framework integration enables iterative refinement with 3-fold speedup while maintaining superior accuracy. Multi-head attention mechanisms successfully identified long-range interactions spanning over 200 nucleotides. The open-source implementation provides comprehensive interfaces, requiring only 12 minutes for 200-nucleotide RNAs on single GPU. This advancement establishes new standards for RNA structure prediction with significant implications for drug discovery and synthetic biology.