Loading…
RNA co-design model generating sequence and 3D backbone together, with SE(3) flow matching over coupled base-centered and sugar-centered frames.
No providers recorded yet. Browse all providers
A nucleotide is not one rigid object. Frame-based RNA generators inherited their geometry from protein backbone design, where one rigid frame per residue is standard, but an RNA residue does two distinct jobs at once. The nucleobase mediates pairing and stacking, carrying the tertiary organization; the ribose-phosphate backbone holds most of the conformational flexibility and is what must be rebuilt atom by atom. A frame anchored to one is a poor proxy for the other.
DuetRNA is a joint sequence-structure generative model for RNA co-design that gives every nucleotide two coupled frames instead of one. It was developed by the Wave Intelligence Lab and the Department of Computer Science and Engineering at the Chinese University of Hong Kong, and released as a preprint in August 2026. The motivating audit compares seven frame constructions across 11,497 static RNA chains and 31,432 multi-state relation groups. A base-anchored plane frame holds a canonical Watson-Crick pair to 7.78° and 0.61 Å of drift across conformers of the same molecule, against 13.35° and 1.70 Å for the sugar-centered Gram-Schmidt frame earlier RNA flow models use — yet that sugar frame rebuilds backbone atoms to 1.88 Å where the base frame manages only 2.98 Å.
Rather than choose, DuetRNA carries both. Each residue holds a base-plane frame anchored at the glycosidic connection atom and a sugar frame anchored at C4', the two evolve together under a single SE(3) flow, and a relative-pose term supervises the transform between them so the pair stays chemically coherent. Nucleotide identity and torsions are read off the coupled state, so sequence and structure emerge from one trajectory rather than being assigned in turn.
The network is a dual-frame trunk extending AlphaFold2's invariant point attention, running base and sugar streams in parallel over base-view, sugar-view and bridge pair channels fused into a shared pair tensor, with periodic object-token mixing through a joint transformer encoder and self-conditioning. Training combines endpoint-parameterized flow matching with nucleotide cross-entropy and torsion/atom reconstruction terms, on the RNAsolo corpus of cleaned PDB-derived RNA structures; RNA3DB supplies a development split for ablations. Sampling uses 100 ODE integration steps, and the paper reports no parameter count.
On the RNA-FrameFlow length grid (40–150 nt), DuetRNA reaches 48.67% scTM-validity under the inverse-folded protocol, which assigns sequences with gRNAde and forward-folds them with RhoFold+ — 7.67 percentage points above RNA-FrameFlow's published 41.00% (p = 0.0076). On the RiboFlow grid it reaches 44.33% against 34.70%. Under the generated-sequence protocol, which folds the model's own sequence, it reaches 38.50% on the RiboGen grid with Boltz-1 as evaluator against 34.17%, a difference whose confidence interval includes zero. Mean inverse-folded validity across four independently trained runs is 45.68% ± 2.33%. Ablations are decisive: removing the base frame costs 14.00 points of validity, and dropping the relative-pose supervision costs 18.67.
RNA co-design is the generative step in aptamer discovery, ribozyme engineering, and structured RNA nanotechnology, where a designer needs a sequence and a plausible fold together. DuetRNA targets that step for unconditional generation of 40–150 nucleotide RNAs. Its supporting evidence is computational — folding-based self-consistency plus direct coordinate analysis — and does not establish biochemical function or experimental viability, so generated candidates remain hypotheses for wet-lab screening.
The contribution is an argument that frame choice is a modeling decision, not a coordinate convention, backed by an audit separating the two roles a nucleotide frame plays. The geometry supports it: against a capacity-matched sugar-only model, DuetRNA cuts steric clashes from 45.58 to 26.55 per 100 heavy atoms and produces an 89%/11% North/South sugar pucker split close to the 80–90% North seen in experimental RNA. The trade-offs are stated plainly — rebuilding base and sugar atoms from two separately predicted frames loosens the glycosidic linkage, raising C1'–N9 RMSD from 0.003 Å to 0.417 Å, and generated structures sit closer to the training corpus (reference-set pdbTM 0.822 versus 0.742), indicating reduced novel-fold coverage. Chain breaks remain the dominant failure mode away from the favorable length regime. Code and weights remain unreleased as of the preprint; the repository is a placeholder alongside a project page with interactive viewers of generated samples.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.