Research map / Graph learning / Graph self-supervised learning
gRNAde: Geometric Deep Learning for 3D RNA inverse 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.
Introduces gRNAde, a multi-state GVP-GNN with autoregressive decoding that designs RNA sequences conditioned on one or more 3D backbone conformations, outperforming Rosetta on single-state recovery and achieving 50% wet-lab success on pseudoknots.
Problem
Computational RNA inverse design usually targets a single secondary structure or a single 3D backbone and ignores conformational flexibility, and physics-based 3D methods like Rosetta are slow and hard to use.
Approach
A geometric GNN over a multi-graph of RNA backbone conformations: each state is encoded independently with GVP layers (equivariant to state order), pooled with permutation-invariant Deep Sets over states, and decoded autoregressively to base identities, allowing one model to handle single- and multi-state design.
Key results
- On 14 RNA structures, gRNAde achieves 56% average native sequence recovery versus 45% for Rosetta, 43% for RDesign, 32% for FARNA and 27% for ViennaRNA.
- Multi-state gRNAde improves sequence recovery by roughly 3-5% over the single-state variant on the multi-state test split, with 3 states performing best.
- In Eterna OpenKnot Round 6 wet-lab validation, 52% of gRNAde partial designs and 49% of full designs scored above 80 versus 35% for the best Rosetta variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use gRNAde instead of Rosetta | Single-state fixed-backbone RNA design · 14 PDB RNA structures | native sequence recovery: 0.45 -> 0.5682 | Helped |
| Use multi-state input with 3 states instead of single-state input (ablation) | gRNAde · Multi-state split | native sequence recovery: 0.481 -> 0.531 | Helped |
| Use an equivariant instead of invariant GNN (ablation) | gRNAde · Single-state split | native sequence recovery: 0.531 -> 0.538 | Helped |
| Use autoregressive instead of non-autoregressive decoding (ablation) | gRNAde · Single-state split | 2D self-consistency scMCC: 0.473 -> 0.612 | Helped |
| Use non-autoregressive instead of autoregressive decoding (ablation) | gRNAde · Single-state split | native sequence recovery: 0.538 -> 0.584 | Helped |
| Use random decoding order (ablation) | gRNAde · Single-state split | native sequence recovery: 0.538 -> 0.531 | Hurt |
| Use Deep Symmetric Set pooling instead of Deep Set pooling (ablation) | Multi-state gRNAde · Multi-state split | native sequence recovery: 0.531 -> 0.540 | Helped |
| Limit maximum training RNA length to 500 nucleotides (ablation) | gRNAde · Single-state split | native sequence recovery: 0.438 vs 0.538 | Hurt |
| Replace EternaFold with RNAFold (ablation) | gRNAde evaluation · Single-state split | 2D self-consistency scMCC: 0.612$\pm${ 0.02} -> 0.614$\pm${ 0.03} | No clear effect |
| Use gRNAde instead of the best Rosetta variant | Pseudoknotted RNA design · Eterna OpenKnot Round 6 | design success rate (OpenKnot score >80): 35% -> 52% | Helped |
Surprising findings
- A non-autoregressive decoder gives higher native recovery on the single-state split but substantially worse self-consistency scores than autoregressive decoding.
- The 2D-only ViennaRNA baseline achieves strong 2D self-consistency but near-random 3D self-consistency and recovery.
Limitations
- Does not yet support multiple interacting chains, RNA-partner molecules, or negative design against undesired conformations.
- Self-consistency metrics depend on imperfect RNA structure predictors (EternaFold, RhoFold), and are hard to apply to multi-state design.
- Multi-state models use substantially more GPU memory (12GB for 1 state vs 50GB for 5 states).
Experimental setup
Datasets: RNASolo PDB RNA structures (<=4.0 Å)
Benchmarks: 14 PDB RNA structures single-state benchmark multi-state split test set (100 RNAs) Eterna OpenKnot Round 6 ribozyme fitness landscape (5TU, t1)
Metrics: native sequence recovery perplexity scMCC (EternaFold) scRMSD scTM-score scGDT_TS (RhoFold) OpenKnot score expected maximum fitness improvement
Baselines: Rosetta FARNA ViennaRNA RDesign NAR decoder ablation
Builds on: GVP-GNN ProteinMPNN Deep Sets PyTorch Geometric
Closest papers by meaning
- Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (ICLR 2025 · similarity 0.89)
- TANGO: RNA Topology and Geometry Co-Design (NeurIPS 2026 · similarity 0.89)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.88)
- RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (ICML 2024 · similarity 0.87)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.86)
- Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.