TANGO: RNA Topology and Geometry Co-Design

Tianmeng Hu, Biao Luo, Ke LiarXiv 2026 · arXiv

TL;DR

TANGO is a novel reinforcement learning framework for RNA inverse design that treats the problem as a joint optimization of 3D geometry and 2D base-pairing topology. By employing a staged curriculum and multi-agent cooperation, TANGO significantly outperforms existing methods, improving tertiary-structure fidelity by 18.3% and secondary-structure fidelity by 52.7% over the strongest baselines.

What Problem Does This Paper Address?

In the world of RNA engineering, "inverse design" is the task of finding a nucleotide sequence that will fold into a specific target structure. While recent deep learning methods have made strides in designing sequences based on 3D coordinates (tertiary geometry), they often overlook the underlying 2D base-pairing patterns (secondary topology).

The authors argue that for RNA, geometry and topology are not redundant; they are complementary. A sequence might fold into a shape that looks roughly correct in 3D space but lacks the specific base pairs (stems, loops, pseudoknots) required for biological function. Conversely, a sequence that satisfies a 2D map might fail to achieve the final 3D arrangement. As shown in Figure 1, existing methods often struggle to satisfy both criteria simultaneously.

Figure 1: RNA 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 sc-TM and secondary-topology fidelity measured by sc-BP-F1. The upper-right region contains designs that satisfy both objectives. Existing 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.

Methods

TANGO (Topology-And-Geometry cO-design) introduces a multi-objective multi-agent reinforcement learning (MARL) approach. Its architecture is built on three key pillars:

  1. Structural Units: Instead of generating sequences one nucleotide at a time, TANGO operates on "structural units." Unpaired residues are handled individually, but paired residues are designed jointly as a single unit. This directly encodes the base-pairing constraints into the action space.
  2. Multi-Agent Cooperation: The sequence is divided among multiple agents that act in parallel. This shortens the "decision horizon"—the number of steps needed to complete a sequence—which helps the model learn from the delayed rewards of structural folding.
  3. Staged Curriculum Training:
    • Stage I (Topology-aware Initialization): The model first learns to satisfy 2D secondary structure scaffolds using a "cheap" 2D folding oracle.
  4. Stage II (Geometry-conditioned Co-design): The policy is transferred to 3D targets. It uses a GVP-GNN (Geometric Vector Perceptron Graph Neural Network) to encode 3D coordinates and zero-initialized adapters to refine the 2D-aware policy with 3D information.
  5. Stage III (Frontier-guided Fine-tuning): The model is fine-tuned for specific targets to expand the Pareto frontier, balancing topology fidelity, geometry fidelity, and sequence diversity.
Figure 2: The TANGO pipeline, showing the structural-unit MARL framework and the three-stage training process.

Datasets

The authors constructed a new benchmark derived from the Protein Data Bank (PDB). They filtered for single-chain RNA structures with high resolution (< 4 Å) and no chain breaks. To ensure the evaluation was meaningful, they only kept targets where current structure predictors (RhoFold and AlphaFold3) could successfully refold the native sequence. The final dataset includes 1,213 targets, with a test set of 50 structurally remote RNAs to test generalization.

Key Insights and Ideas

  • Topology as a Scaffold: The most significant insight is that 2D topology provides a necessary scaffold for 3D assembly. By initializing the model on 2D tasks, the authors provide a "warm start" that makes the much harder 3D optimization tractable.
  • Coordinated Decisions: Treating base pairs as a single unit (a "paired unit") prevents the model from making incompatible choices for residues that must interact, effectively reducing the search space.
  • Multi-Objective Optimization: Rather than optimizing a single scalar score, TANGO's third stage explicitly looks for a diverse set of sequences that represent different trade-offs between shape and pairing accuracy.

Key Results

TANGO was compared against five recent 3D RNA design baselines, including gRNAde and RIDER.

  • Superior Fidelity: TANGO achieved a mean sc-TM (geometry) of 0.71 and sc-BP-F1 (topology) of 0.84. This represents an 18.3% and 52.7% improvement, respectively, over RIDER, the strongest baseline.
  • Higher Success Rate: TANGO solved 88% of test targets (meaning it generated at least one sequence satisfying both geometry and topology thresholds), compared to just 60% for RIDER.
  • Novelty: Despite the high structural accuracy, TANGO's designs remained sequence-diverse and distant from native sequences, suggesting it isn't just memorizing the training data.

Limitations and Future Work

  • Oracle Dependency: The method relies on structure-prediction oracles (like RhoFold) for feedback. Any biases or inaccuracies in these oracles will be inherited by the design process.
  • Scope: The current framework is limited to single-chain RNAs. The authors identify extending TANGO to handle RNA-protein complexes, ligand-bound RNAs, and non-canonical base interactions as vital next steps.
  • Experimental Validation: While the computational results are strong, the authors note that experimental folding and functional assays are needed to fully validate these "in silico" designs.

Summary updated

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.
BibTeX
@misc{hu2026tango,
  author = {Tianmeng Hu and Biao Luo and Ke Li},
  title = {{TANGO: RNA Topology and Geometry Co-Design}},
  year = {2026},
  eprint = {2610.20277},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.20277},
}

Loading related content...

Figure