Research map / Probabilistic and scientific modeling / Molecules and proteins

TANGO: RNA Topology and Geometry Co-Design

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TANGO, a multi-agent MARL framework that co-designs RNA sequences for both 2D topology and 3D geometry via structural-unit actions, a three-stage curriculum, and frontier-guided multi-objective fine-tuning.

Problem

3D RNA inverse design usually optimizes tertiary geometry alone and neglects secondary base-pairing topology, so designs can fold geometrically similar but topologically wrong.

Approach

Decompose sequence generation into topology-derived structural units (single unpaired positions and jointly designed base pairs) filled by cooperative agents under a shared reward, trained via topology-only pretraining, geometry-conditioned transfer, and hypervolume-based frontier fine-tuning.

Key results

• TANGO raises sc-TM from 0.60 (RIDER) to 0.71 (+18.3%) and sc-BP-F1 from 0.55 to 0.84 (+52.7%).
• Co-design success rate rises from 0.49 (RIDER) to 0.86; solved target rate from 60% to 88%.
• Removing topology-aware initialization degrades all primary metrics; frontier fine-tuning raises hypervolume from 0.031 (TANGO-Base) to 0.170 (TANGO) and 0.374 (TANGO-Div).

Surprising findings

• Geometry-only 3D design baselines can produce geometry-consistent designs with poor recovery of the target base-pairing topology.
• Top-10 sampling from a fixed policy is clearly outperformed by frontier-guided fine-tuning, indicating gains are not merely from selection.

Limitations

• Evaluation relies on structure-prediction oracles (RhoFold, AlphaFold3) rather than experimental validation.
• Benchmark filtered to targets that native sequences refold with TM>=0.45, biasing toward predictable RNAs.
• Non-canonical tertiary interactions are not modeled as discrete pair actions.

Experimental setup

Datasets: PDB single-chain RNA structures (curated, 1,213 targets)

Benchmarks: curated PDB-derived co-design benchmark (50 structurally remote test targets)

Metrics: sc-TM sc-RMSD sc-BP-F1 co-design success rate solved target rate sequence novelty diversity hypervolume

Baselines: gRNAde RDesign R3Design RiboDiffusion RIDER

Builds on: MAPPO GVP-GNN CocoRNA RIDER

Closest papers by meaning

• gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.89)
• RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.88)
• Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (ICLR 2025 · similarity 0.87)
• RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching (NeurIPS 2025 · similarity 0.86)
• RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (ICML 2024 · similarity 0.86)
• Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.84)