Secondary-Structure-Informed RNA Inverse Design via Relational Graph Neural Networks

RNA inverse design is an essential part of many RNA therapeutics strategies. To date, there have been great advances in the computationally-driven RNA design. Current machine-learning approaches can predict the sequence of an RNA given its 3D structure with acceptable accuracy and tremendous speed. Design and engineering of RNA regulators such as riboswitches, however, is often more difficult, partly due to their inherent conformational switching abilities. Although recent state-of-the-art models do incorporate information about the multiple structures a sequence can fold into, there is great room for improvement in modeling structural switching. In this work, a relational geometric graph neural network is proposed that explicitly incorporates alternative structures to predict the RNA sequence. The proposed model uses edge types to distinguish primary-, secondary-, and tertiary-structure pairing/positioning of nucleotides for its training. Results show higher native sequence recovery rates over gRNAde across different test sets (eg. 72% vs 66%) and a benchmark from the literature (60% vs 57%). The impact of secondary-structure on prediction accuracy was more significant than that of tertiary-structure dependencies as defined here. The code for the above relational GNN can be found here.