Deep Learning of Ligand-bound RNA Tertiary Structures Diverges from Learning Unbound Ones: A Case Study Using The gRNAde Software

Modeling the relationship between the native RNA sequence and its in-vivo structure is challenging, partly due to the flexible nature of the RNA molecular structure. In addition, the RNA structure can take on different conformations in the presence of specific molecules, metabolites, temperatures or other signaling and environmental factors, making it difficult to construct a universal statistical model for the sequence-structure relationship of the RNA. Using a Geometric-Vector-Perceptron Graph Neural Network architecture, Joshi, et al. predict the RNA sequence from its given 3D structure with good performance and on a dataset including RNA structures of different type and length, namely RNAsolo. In this work, using the Authors open-source software package, gRNAde, we confirm their results on a more updated version of RNAsolo and for structure of different resolution, confirming the ability of the algorithm to capture RNA structural features and generalize to sequences of different lengths. We did observe, however, that performance on riboswitches is lower than expected that RNAs whose structure has been resolved while being bound to a ligand, such as riboswitches, may require a statistical model that diverges from those of native structures.