Research map / Label-efficient and robust learning / Recognition and long-tail
RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA 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 RDesign, a tertiary-structure-based RNA design model with a new RNA backbone geometry encoding, hierarchical cluster- and sample-level contrastive learning on a hypersphere, and secondary-structure base-pair supervision, evaluated on a new 2218-structure benchmark.
Problem
Designing RNA sequences that fold into a specified three-dimensional tertiary structure is hard due to structural complexity and scarce high-resolution RNA 3D data, and protein design methods transfer poorly to RNA.
Approach
Combines an RNA-specific tertiary structure featurization (local frames built on P atoms, dihedral sin/cos, RBF distances, orientation quaternions) with hierarchical contrastive learning (TM-score cluster-level positives plus confidence-weighted perturbation-based sample-level positives) constrained to a hypersphere, and a base-pair-sharpened supervised loss using secondary structure.
Key results
- RDesign reaches 41.53% overall recovery and 40.89 Macro-F1 on the benchmark test set, versus 28.23% (StructGNN) for the best baseline in recovery.
- On Rfam and RNA-Puzzles, RDesign achieves 56.12% and 50.12% recovery, above the best baseline eM2dRNAs at 33.34% and 37.10%.
- Removing the hyperspherical projection drops benchmark recovery from 41.53% to 30.69%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RDesign rather than the strongest listed baseline, SeqRNN (h=256) | Tertiary structure-based RNA design · Benchmark RNA test set (All) | Recovery (%): 28.24±0.46 -> 41.53±0.38 | Helped |
| use RDesign rather than eM2dRNAs | RNA design · Rfam | Recovery (%): 33.34 ±1.02 -> 56.12±1.03 | Helped |
| use RDesign rather than eM2dRNAs | RNA design · RNA-Puzzles | Macro F1 (×100): 26.91±2.32 -> 49.24±1.07 | Helped |
| add RDesign's RNA-specific tertiary structure modeling rather than classical protein-design modeling (ablation) | RDesign · Benchmark RNA test set (All) | Recovery (%): 36.45 -> 41.53 | Helped |
| add hierarchical representation learning (ablation) | RDesign · Benchmark RNA test set (All) | Recovery (%): 37.12 -> 41.53 | Helped |
| use hyperspherical rather than Euclidean representation space (ablation) | RDesign · Benchmark RNA test set (All) | Recovery (%): 30.69 -> 41.53 | Helped |
| add secondary structure constraints (ablation) | RDesign · Benchmark RNA test set (All) | Recovery (%): 38.55 -> 41.53 | Helped |
Surprising findings
- Protein-design tertiary models (StructMLP, StructGNN, GraphTrans) perform comparably to sequence-only models, suggesting direct transfer from protein design fails for RNA.
- Perplexity results are not aligned with recovery and Macro-F1, with sequence-based models outperforming structure-based ones on perplexity.
Limitations
- Only in silico evaluation; no wet-lab validation.
- Folding of designed sequences assessed only qualitatively due to lack of reliable RNA structure prediction tools.
- Sequences longer than 500 nucleotides were excluded from the dataset.
Experimental setup
Datasets: RNAsolo Protein Data Bank (PDB) RNA structures Rfam RNA-Puzzles
Benchmarks: RDesign benchmark test split Rfam RNA-Puzzles
Metrics: recovery Macro-F1 perplexity
Baselines: SeqRNN SeqLSTM StructMLP StructGNN GraphTrans PiFold MCTS-RNA LEARNA aRNAque eM2dRNAs
Builds on: StructGNN GraphTrans PiFold
Closest papers by meaning
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.88)
- TANGO: RNA Topology and Geometry Co-Design (NeurIPS 2026 · similarity 0.88)
- RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching (NeurIPS 2025 · similarity 0.88)
- Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (ICLR 2025 · similarity 0.87)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.86)
- RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (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.