A Root Mean Square Deviation Estimation Algorithm (REA) and its use for improved RNA Structure Prediction

Abstract

The 3D structure of RNA is crucial for biotechnological applications and to comprehend its biological function. Recent developments using AlphaFold-inspired deep neural networks improved the prediction of 3D structure from RNA sequence, but evaluation of the accuracy of these predictions is still necessary. We present the RMSD Estimation Algorithm (REA), a feed-forward neural network to predict the root-mean-square deviation (RMSD) of a 3D RNA structure from its experimentally determined counterpart using its Molprobity [1] stereochemical validation features. It is trained on structures predicted by the DeepFoldRNA [2] and trRosettaRNA [3] transformer-based deep neural networks on a set of 182 models of RNA structures with pseudoknots. We compare REA with ARES [4], a deep learning algorithm that predicts the RMSD by extracting geometric patterns with equivariant convolution, assessing the prediction accuracy on RNAs with and without pseudoknots. REA outperformed ARES on both test sets with smaller absolute difference between the true and the predicted RMSD. Using a combination of REA and a Support Vector Regression (SVR) trained on the same data as REA, we can select RNA structures predicted with DeepFoldRNA, trRosettaRNA and Rhofold [5] to achieve a significantly higher prediction accuracy than any of the prediction methods used alone. This was shown on a validation set with 261 novel RNA chains extracted from the Nonredundant 3D Structure Dataset [5] and a test set with 55 novel RNA chains from RNA-Puzzles [5]. Our selection based prediction method can easily incorporate additional prediction algorithms.