Wave Intelligence Lab at CUHK

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

Junzhe Li1,2,*   Lijian Peng2,*   Yuhao Li1,2   Yize Zhou1,2
Hanqun Cao2   Cheng Tan2   Shengchao Liu1,2
1Wave Intelligence Lab    2CSE, The Chinese University of Hong Kong
*Equal contribution

A 120-nt generated backbone (rainbow, N-to-C) overlaid with the RhoFold prediction of its designed sequence (grey). scTM 0.959, scRMSD 1.70 Å.

Overview

DuetRNA is a generative model for de novo RNA backbone design. Each nucleotide carries two coupled rigid frames — a base-anchored frame for inter-residue geometry and a sugar-anchored frame for backbone reconstruction — supervised with a relative-pose loss that keeps the pair chemically coherent. Trained on RNAsolo, DuetRNA improves designability over RNA-FrameFlow on matched protocols, and yields more distinct designable structural modes per generation attempt.

Evaluation protocols

DuetRNA is evaluated under two complementary de novo settings. The inverse-folded (IF) protocol measures backbone designability: gRNAde proposes sequences for each generated backbone, and an independent predictor (RhoFold) tests whether those sequences fold back to it, with a sample counted as valid when scTM ≥ 0.45. The generated-sequence (GS) protocol instead folds the sequence proposed by the model itself, directly measuring sequence-structure compatibility. On the RNA-FrameFlow grid, DuetRNA reaches 48.67% IF validity (vs 41.00% published for RNA-FrameFlow), and on RiboFlow's native grid 44.33% (vs 34.70%); under the matched GS comparison with Boltz-1 it reaches 38.50% (vs RiboGen's 34.17%). Across four independently trained runs, IF validity is 45.68% ± 2.33%, with every run above the 41.00% baseline. Direct atom-level analyses further show improved steric packing and sugar-ring closure over the single-frame ablation.

Why two frames per nucleotide?

RNA is unusually flexible for structure generation: no single rigid frame can capture both how bases pair and stack and how the phosphate backbone connects. Where a frame is anchored determines what it represents well — a frame anchored at the sugar follows the backbone but drifts on base-pair geometry, while a frame anchored at the base plane tracks base organization instead. We systematically compared candidate anchor placements in a representation analysis covering drift on canonical pairs and atom-level reconstruction, and got the outcome below: two specialized frames per nucleotide, a base-anchored frame carrying inter-residue relations and a sugar-anchored frame carrying backbone reconstruction, each doing what it is best at.

Dual-frame representation: Sugar-GS and Base-Plane frames on a nucleotide

Two ways to anchor a rigid frame on a nucleotide. (a) Sugar-GS, built from (O4$'$, C4$'$, C3$'$). (b) DuetRNA's dual-frame: Sugar-GS together with a Base-Plane frame anchored at the glycosidic nitrogen (N9 for purines, N1 for pyrimidines).

Coupled, not just parallel

DuetRNA predicts both frames jointly and supervises their relative pose with a dedicated constraint, keeping the pair physically consistent. The coupling is what makes it work: removing the base frame drops inverse-folded validity by 14 points, and removing the relative-pose constraint drops it by 18.7 points. A reconstruction audit from ground-truth frames confirms the division of labor — bases rebuilt through the base channel and backbone through the sugar channel reach 1.58 Å and 1.88 Å RMSD respectively.

DuetRNA dual-frame flow matching overview

Overview of the DuetRNA pipeline. Each nucleotide carries a base frame and a sugar frame. The model jointly evolves both from noise to a structured state; a dual-frame constraint on the predicted base-to-sugar relative pose keeps the two frames chemically coupled through generation.

Interactive 3D structures

Here we present six unconditionally generated backbones spanning the 40–140 nt range. You can drag to rotate and scroll to zoom — clicking a residue zooms into it.

40 nt · scTM 0.586

70 nt · scTM 0.712

90 nt · scTM 0.765

110 nt · scTM 0.863

120 nt · scTM 0.974

140 nt · scTM 0.679

Generated samples

Citation

@article{Li2026.08.18.745543,
  title   = {{Base-and-Sugar Dual-Frame Flow Matching for RNA Co-Design}},
  author  = {Li, Junzhe and Peng, Lijian and Li, Yuhao and Zhou, Yize
             and Cao, Hanqun and Tan, Cheng and Liu, Shengchao},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.1101/2026.08.18.745543},
  URL     = {https://www.biorxiv.org/content/10.1101/2026.08.18.745543}
}