Base-and-Sugar Dual-Frame Flow Matching for RNA Co-Design

Current frame-based RNA generators typically represent each nucleotide with a single coordinate frame. However, our representation analysis across $11{,}497$ static chains and $31{,}432$ multi-state relation groups shows that a single frame is insufficient to simultaneously capture base-mediated interactions and atom-level reconstruction. Motivated by this observation, we introduce \textbf{DuetRNA}, a dual-frame model for joint RNA sequence-structure generation. Each nucleotide carries two coupled $\mathrm{SE}(3)$ frames: a base-centered frame anchored at the glycosidic nitrogen for base mediated relationship, and a Gram-Schmidt sugar-centered frame for sugar-backbone reconstruction. DuetRNA learns their joint distribution with $\mathrm{SE}(3)$ flow matching and co-generates nucleotide identities in the same forward pass, avoiding post-hoc sequence recovery by an external inverse-folding model. Our experiments across two de novo generation protocols show that DuetRNA produces RNA structures with strong folding-based self-consistency. It achieves $44.00\%$ scTM-validity under inverse-folded sequence evaluation and $38.50\%$ Boltz-1-based / $37.83\%$ RhoFold-based TM-validity under direct joint generation. These results demonstrate that explicitly factorizing RNA residues into base-centered and sugar-backbone-centered frames is a strong representation choice for RNA sequence--structure generation.