Predicting 3D RNA structure from solely the nucleotide sequence using Euclidean distance neural networks.

Abstract

Fast and accurate 3D RNA structure prediction remains a major challenge in structural biology, primarily due to the size and flexibility of RNA molecules, as well as lack of diverse experimentally determined structures of RNA molecules. Unlike DNA structure, RNA structure is far less constrained by base pair hydrogen bonding, resulting in an explosion of potential stable states. Here, we propose a convolutional neural network which predicts all pairwise distances between residues in an RNA, using a recently described smooth parametrization of Euclidean distance matrices. We achieve high accuracy predictions on RNAs up to 100 nucleotides in length in fractions of a second, a factor of 107 faster than existing molecular dynamics-based methods. We also convert our coarse-grained machine learning output into an all-atom model using discrete molecular dynamics with constraints. Our proposed computational pipeline accurately predicts all-atom RNA models solely from the nucleotide sequence.