RNA3DClust: unsupervised segmentation of RNA 3D structures using density-based clustering

A growing body of evidence shows that the biological activity of RNA molecules is not only due to their primary and secondary structures, but also to their spatial conformation. As the experimental determination of the three-dimensional (3D) structure is a costly and uncertain process, the development of computational methods for predicting RNA fold is a necessity. A critical task for such prediction of the 3D structure consists of finding substructures that can be predicted independently, before being assembled into a global fold. In protein structures, these subparts are the “structural domains” and, to the best of our knowledge, no equivalent concept has been defined for RNA macromolecules.

In this work, we present RNA3DClust, an adaptation of the Mean Shift clustering algorithm to the RNA 3D structure partitioning problem. This approach allowed us to delimit compact and separate regions in RNA conformations, analogously to the seminal definition of domains in proteins. Benchmarking RNA3DClust required us to create a reference dataset of RNA 3D domain annotations and to devise a new scoring function for assessing the segmentation quality. Importantly, we show that the domain decompositions produced by RNA3DClust may be consistent with those based on RNA biological function and evolution. Finally, the emerging interest in long non-coding RNAs (lncRNAs) and their likeliness of containing folded regions has motivated us to generate an additional reference dataset of lncRNA predicted conformations. The resulting delineations of 3D domains by RNA3DClust illustrate the potential of our method for analyzing lncRNA structures. Source code and datasets are freely available for download on the EvryRNA platform at: https://evryrna.ibisc.univ-evry.fr .