Accurate de novo prediction of RNA 3D structure with transformer network

Abstract

RNA 3D structure prediction remains challenging though after years of efforts. Inspired by the recent breakthrough in protein structure prediction, we developed trRosettaRNA, a novel deep learning-based approach to de novo prediction of RNA 3D structure. Like trRosetta, the trRosettaRNA pipeline comprises two major steps: 1D and 2D geometries prediction by a transformer network; and full-atom 3D structure folding by energy minimization with constraints from the predicted geometries. We benchmarked trRosettaRNA on two independent datasets. The results show that trRosettaRNA outperforms other conventional methods by a large margin. For example, on 25 targets from the RNA-Puzzles experiments, the mean RMSD of the models predicted by trRosettaRNA is 5.5 [A], compared with 10.5 [A] from the state-of-the-art human group (i.e., Das). Further comparisons with two recently released deep learning-based methods (i.e., DeepFoldRNA and RoseTTAFoldNA) show that all three methods have similar accuracy. However, trRosettaRNA yields more accurate and physically more realistic side-chain atoms than DeepFoldRNA and RoseTTAFoldNA. Finally, we apply trRosettaRNA to predict the structures for the Rfam families that do not have known structures. Analysis shows that for 263 families, the predicted structure models are estimated to be accurate with RMSD < 4 [A]. The trRosettaRNA server and the package are available at: https://yanglab.nankai.edu.cn/trRosettaRNA/.