RNA Backbone Torsion and Pseudotorsion Angle Prediction Using Dilated Convolutional Neural Networks

RNA three-dimensional structure prediction has been relied on using a predicted or experimentally determined secondary as restraint to reduce the conformational sampling space. However, secondary-structure restraints are limited paired bases, and space of ribose-phosphate backbone is still too large be sampled efficiently. Here, we employed dilated convolutional neural network predict torsion pseudotorsion angles single sequence input. The method called SPOT-RNA-1D was trained high-resolution training data set tested three independent, nonredundant, test sets. proposed yields substantially smaller mean absolute errors than baseline predictors based random predictions helix conformations according actual angle distributions. for sets range from 14°-44° different angles, compared 17°-62° by 14°-58° prediction. also accurately recovers overall patterns pairwise In general, further away bases associated with unpaired involved in tertiary interactions more difficult predict. Compared best models RNA-puzzles experiments, yielded accurate dihedral and, thus, potentially useful model quality indicators protein