TANGO: RNA Topology and Geometry Co-Design

Tianmeng Hu; Biao Luo; Ke Li

cs.LGOct 14, 2026

TANGO: RNA Topology and Geometry Co-Design

Authors: Tianmeng Hu, Biao Luo, Ke Li

Organizations: University of Exeter · Earlham Institute · Central South University

Abstract

Designing RNA sequences that satisfy target tertiary-structure objectives is an important problem in RNA engineering, with broad applications in synthetic biology and therapeutics. Existing 3D RNA design methods have made substantial progress, but they largely treat the target fold as a geometry-only objective. This is incomplete for RNA, whose functional folds are specified not only by tertiary geometry but also by secondary base-pairing topology. We therefore formulate RNA inverse design as a topology-and-geometry co-design problem and propose TANGO, a multi-objective multi-agent reinforcement learning framework for this task. TANGO adopts a staged curriculum: it first learns a topology-aware policy from secondary-structure targets, then transfers this policy to tertiary-structure targets to train a general co-design model, and finally applies frontier-guided fine-tuning to refine target-specific trade-offs among secondary-structure fidelity, tertiary-structure fidelity, and sequence diversity. Experiments show that TANGO generates diverse and novel designs, improving tertiary- and secondary-structure performance by 18.3% and 52.7%, respectively, over the strongest baseline.

Figures & tables

Figure 1: RNA

\mathrm{3D}

inverse design requires joint topology and geometry fidelity. Each point denotes a designed sequence evaluated by fold-back self-consistency, with tertiary-structure fidelity measured by

\mathrm{sc\text{-}TM}

and secondary-topology fidelity measured by

\mathrm{sc\text{-}BP\text{-}F1}

. The upper-right region contains designs that satisfy both objectives. Existing

\mathrm{3D}

RNA design methods can produce geometry-consistent but topology-inconsistent designs, showing that tertiary similarity alone does not guarantee recovery of the target base-pairing topology. Conversely, topology preservation alone is insufficient for specifying the final tertiary fold.

Figure 2: Overview of TANGO . TANGO formulates RNA design as cooperative sequence generation over structural units and follows a three-stage training pipeline.

Method	$\mathrm{sc\text{-}TM}$ $\uparrow$	$\mathrm{sc\text{-}BP\text{-}F1}$ $\uparrow$	$\mathrm{sc\text{-}RMSD}$ $\downarrow$	Success rate $\uparrow$	Solved targets $\uparrow$
gRNAde	$0.32\pm 0.12$	$0.36\pm 0.26$	$3.54\pm 1.00$	$0.17$	$38.00\%$
RDesign	$0.31\pm 0.10$	$0.31\pm 0.21$	$3.61\pm 0.91$	$0.13$	$42.00\%$
R3Design	$0.35\pm 0.14$	$0.33\pm 0.27$	$3.47\pm 1.03$	$0.21$	$50.00\%$
RiboDiffusion	$0.32\pm 0.13$	$0.30\pm 0.25$	$3.58\pm 1.14$	$0.17$	$42.00\%$
RIDER	$0.60\pm 0.13$	$0.55\pm 0.22$	$2.94\pm 1.26$	$0.49$	$60.00\%$
TANGO	$\mathbf{0.71\pm 0.17}$	$\mathbf{0.84\pm 0.15}$	$\mathbf{2.16\pm 1.07}$	$\mathbf{0.86}$	$\mathbf{88.00\%}$

Table 1: Structural design performance on the test set. Each method generates

K=10

candidates per target. Continuous metrics report mean

\pm

standard deviation over target-level aggregated scores. Success denotes co-design success , defined as

\mathrm{sc\text{-}TM}>0.45

and

\mathrm{sc\text{-}BP\text{-}F1}>0.50

. Solved targets denotes the fraction of targets for which at least one candidate satisfies both thresholds. Threshold sensitivity is reported in Appendix D.1 .

Figure 3: Representative co-design successes for PDB targets 3IQP , 6WJS , 3UCU , and 3ADB . Gray denotes the target reference structure, and color denotes the fold-back structure of the designed sequence. All shown designs satisfy the joint co-design success criterion.

Figure 4: Target-wise frontier-guided refinement. (a) Structural quality and sequence diversity for the stage-II policy ( TANGO-Base ), the bi-objective frontier-refined model ( TANGO ), and the diversity-aware tri-objective variant ( TANGO-Div ). (b) Multi-objective trade-off visualization between design quality and diversity. HV denotes the dominated hypervolume of each method’s Pareto frontier in the normalized objective space. Significance markers denote * for

p<0.05

, ** for

p<0.01

, *** for

p<0.001

, and ns for not significant.

Figure 5: Effect of topology-aware initialization. The ablated variant removes Stage-I topology pretraining and trains the geometry-conditioned policy from scratch, while keeping the structural-unit action space and model architecture unchanged. Significance is computed by paired target-level tests.

Figure 6: Sequence novelty and design quality across methods. Novelty is measured as normalized Hamming distance to the native target sequence. Design quality is reported by

\mathrm{sc\text{-}TM}

and co-design success rate, where success requires both topology and geometry thresholds.

Figure 7: Cross-oracle evaluation of the same TANGO designs using RhoFold and AlphaFold3 . No AlphaFold3 -based refinement is performed.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1 : Sensitivity of TANGO to the joint success criterion. Left: successful design rate when varying the

\mathrm{sc\text{-}TM}

threshold while fixing

\mathrm{sc\text{-}BP\text{-}F1}

>0.50

. Right: successful design rate when varying the

\mathrm{sc\text{-}BP\text{-}F1}

threshold while fixing

\mathrm{sc\text{-}TM}

>0.45

Figure A2 : Comparison between Top-10 sampling and frontier-guided fine-tuning. For each target, we sample sequences from 30 random seeds and select the 10 designs with the highest

\mathrm{sc\text{-}TM}

for evaluation. We compare raw policies, their Top-10 sampling variants, and the full TANGO model. Bars show mean

\pm

standard deviation across targets.

Figure A3 : Comparison between frontier-guided fine-tuning and fixed weighted reward alternatives. Three scalarized reward settings with different

\mathrm{2D}

\mathrm{3D}

weights are compared with TANGO and TANGO-Div on

\mathrm{sc\text{-}TM}

\mathrm{sc\text{-}BP\text{-}F1}

, and sequence diversity. Bars show mean

\pm

standard deviation across targets. Significance markers denote * for

p<0.05

, ** for

p<0.01

, *** for

p<0.001

, and ns for not significant.

Figure A4 : Trade-offs among multiple objectives under fixed weighted rewards. Changing the scalarization weights shifts the balance among

\mathrm{2D}

fidelity,

\mathrm{3D}

fidelity, and diversity, whereas frontier-guided fine-tuning directly optimizes the target-wise trade-off frontier.

Figure A5 : Ablation of the cooperative multi-agent formulation. We compare the default multi-agent TANGO with a single-agent variant under the same training and evaluation protocol. Bars show mean

\pm

standard deviation across held-out targets. Significance markers denote * for

p<0.05

, ** for

p<0.01

, *** for

p<0.001

, and ns for not significant.

Setting	Value
Number of agents	4
Actor learning rate	$1\times 10^{-4}$
Critic learning rate	$1\times 10^{-4}$
Encoder learning-rate scale	0.3
Optimizer	AdamW
Weight decay	$1\times 10^{-4}$

Appendix

Table A1: Stage I rollout and optimization settings.

Setting	Value
Number of agents	4
Actor learning rate	$5\times 10^{-4}$
Critic learning rate	$5\times 10^{-4}$
Learning-rate decay floor	0.3
Buffer size	1024
Epochs per update	10

Appendix

Table A2: Stage II rollout, optimization, and co-design reward settings.

Setting	Value
Warm-start archive size	64
Actor learning rate	$1\times 10^{-4}$
Critic learning rate	$1\times 10^{-4}$
Buffer size	1024
Epochs per update	4
Minibatches per update	2

Appendix

Table A3: Stage III frontier-guided fine-tuning settings.

Explore similar work

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching
Oct 8, 2026Runze Ma, Will Hua, Shuangjia ZhengInverse Design Flow-Based Generative Modeling
Shanghai Jiao Tong University · Independent Researcher
RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning
Sep 29, 2026Zefeng Lin, Xianyong Fang, Tianfan Fu +1Inverse Design Molecular Generation
School of Computer Science and Technology, University of Science and Technology of China, Hefei, Anhui, China · School of Computer Science and Technology, Anhui University, Hefei, Anhui, China · State Key Laboratory for Novel Software Technology at Nanjing University, School of Computer Science, Nanjing University, Nanjing, Jiangsu, China
MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design
Oct 9, 2026Jixin Tang, Ji Guo, Jun ZhaoInverse Design Reward Design for RL
Fudan University

Oct 8, 2026J/K move · Enter open · S save

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching

Runze Ma, Will Hua, Shuangjia Zheng

Shanghai Jiao Tong University · Independent Researcher

Inverse Design Flow-Based Generative Modeling

Open paper AskPDF arXiv

RNA function arises from the coupling of nucleotide sequence and three-dimensional structure, motivating their joint design. Coordinating global folding with nucleotide-level detail remains challenging under limited structural supervision. We introduce La-Ribo, a generative framework for RNA sequence-structure co-design via geometry-latent flow matching. La-Ribo retains a sparse phosphate-sugar--base scaffold and encodes nucleotide identity and local conformation in residue-wise latents. A shared flow network generates both jointly, and an RNA-specific decoder then reconstructs all heavy atoms. To expand supervision, we construct a quality-controlled corpus of 168,561 RNA structures, integrating experimental data with predictions from three folding models, including 10,631 MSA-supported structures generated in this work. La-Ribo improves designability and codesignability over the evaluated baselines across sampling budgets and two refolding models, and the same prior supports scaffold-conditioned inverse folding without additional training.

Abstract

Explore similar work

La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching

RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning

MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design