Context-aware geometric deep learning for RNA sequence design

RNA design has emerged to play a crucial role in synthetic biology and therapeutics. Although tertiary structure-based RNA design methods have been developed recently, they still overlook the broader molecular context, such as interactions with proteins, ligands, DNA, or ions, limiting the accuracy and functionality of designed sequences. To address this challenge, we present RISoTTo (RIbonucleic acid Sequence design from TerTiary structure), a parameter-free geometric deep learning approach that generates RNA sequences conditioned on both their backbone scaffolds and the surrounding molecular context. We evaluate the designed sequences based on their native sequence recovery rate and further validate them by predicting their secondary structures in silico and comparing them to the corresponding native structures. RISoTTo performs well on both metrics, demonstrating its ability to generate accurate and structurally consistent RNA sequences. Additionally, we present an in silico design study of domain 1 of the NAD+ riboswitch, where RISoTTo-generated sequences are predicted to exhibit enhanced binding affinity for both the U1A protein and the NAD+ ligand.