Fig. 1: Overview of NuFold.

a The architecture of NuFold. NuFold is an end-to-end architecture for RNA tertiary structure prediction, taking target sequence information and generating corresponding full-atom tertiary structures. The query sequence is initially used to construct an MSA, which, along with predicted secondary structure information, serves as input to the NuFold network. The NuFold network comprises three components: Preprocessing, EvoFormer, a transformer model that extracts co-evolutional information from MSA and embeds it into both single and pair representations; and the structure module, which further processes the embedded information into 3D structures. These processes are iteratively performed in a recycling process to refine predictions. b The network predicts two key components for full-atom structure prediction: the translation and rotation of the base frame, along with a set of torsion angles derived from the base frame. These torsion angles are used to extend new atoms. c Representing two puckering conformations, C3’-endo and C2’-endo conformations. Green and magenta structures represent ground truth and structures built by NuFold, respectively. NuFold predicts torsion angles not only for the main chain or chi angles but also for the ribose ring, allowing for the reproduction of sugar-puckering formations. RMSD of the structures built were 0.04 Å and 0.03 Å, respectively, for C3’-endo and C2’-endo conformations.