Research map / Probabilistic and scientific modeling / Molecules and proteins
RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching
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 RiboFlow, a synergistic flow matching model that co-designs RNA backbone frames, torsion angles, and nucleotide sequences conditioned on target ligands, plus the RiboBind RNA-ligand dataset.
Problem
Designing RNAs that bind specific small-molecule ligands is hard due to RNA conformational flexibility, structural validity, and scarce RNA-ligand structural data.
Approach
A flow matching framework that jointly models RNA backbone frames (SE(3)), torsion angles, and discrete nucleotide types, adds a distance-aware ligand interaction module (IPA refinement with RBF-style ligand-atom distances), uses co-design pre-training on RNASolo, and augments data via dynamic cropping of RiboBind.
Key results
- Co-design pre-training raises AF3-based binding metric (%AF) from 22.8 to 48.4 with fixed ligand structure and true length on the sequence-based split, and validity from 10.8% to 26.8%.
- RiboFlow (+Pre+Crop) reaches 50.1% AF score versus 17.4% for RNA-FrameFlow under the same setting, roughly a 2.9x gain.
- Unconditional generation validity rises to 34.7% for RiboFlow versus 25.0% for RNA-FrameFlow at 50 steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sequence-structure co-design to RNA generation (ablation) | RNA-FrameFlow-R · Unconditional RNA generation, step_50 | validity (%): 26.3 vs 25.0 | Helped |
| Add co-design pretraining (ablation) | RiboFlow · Sequence-based, ligand structure and true length | validity (%): 10.8 -> 26.8 | Helped |
| Add co-design pretraining (ablation) | RiboFlow · Sequence-based, ligand structure and true length | GerNAScore: 0.283 -> 0.437 | Helped |
| Add co-design pretraining (ablation) | RiboFlow · Sequence-based, ligand structure and true length | Vina Top10: -4.25 -> -4.39 | Helped |
| Add co-design pretraining (ablation) | RiboFlow · Sequence-based, ligand structure and true length | sequence recovery (%): 30.9 -> 33.4 | Helped |
| Add dynamic cropping (ablation) | RiboFlow · Sequence-based, ligand structure and true length | AFScore (%): 22.8 -> 34.7 | Helped |
| Add dynamic cropping (ablation) | RiboFlow · Sequence-based, ligand structure and true length | validity (%): 10.8 -> 12.1 | No clear effect |
| Add co-design pretraining and dynamic cropping (ablation) | RiboFlow · Sequence-based, ligand structure and true length | AFScore (%): 22.8 -> 50.1 | Helped |
| Add co-design pretraining and dynamic cropping (ablation) | RiboFlow · Sequence-based, ligand structure and true length | Vina Top10: -4.25 -> -4.30 | No clear effect |
| Add co-design pretraining (ablation) | RiboFlow · Sequence-based, ligand structure and sampling length | validity (%): 10.8 -> 25.9 | Helped |
| Add co-design pretraining and dynamic cropping (ablation) | RiboFlow · Sequence-based, ligand structure and sampling length | AFScore (%): 25.7 -> 54.8 | Helped |
| Re-center coordinates on the ligand centroid (ablation) | RiboFlow + Pre + Crop · Sequence-based, ligand SMILES and sampling length | GerNAScore: 0.460 -> 0.472 | Helped |
Surprising findings
- Sampling with ligand SMILES instead of the true bound ligand structure yields higher predicted binding affinity in several metrics.
- Dynamic cropping did not significantly improve structural validity over pre-training alone, though it helped binding metrics.
Limitations
- Pre-training depends on RNASolo, concentrated in 30-200 nt, so performance on longer RNAs is limited and degrades beyond the training range.
- Binding affinity evaluated only computationally (AF3, Vina, GerNA-Bind) with no wet-lab validation.
- Validity and diversity trade-offs; RiboFlow sequence constraints reduce diversity relative to RNA-FrameFlow-R.
Experimental setup
Datasets: RNASolo RiboBind PDBBind RNAmigos2 HariBoss
Benchmarks: RiboBind sequence-based split RiboBind structure-based split RiboBind few-shot split
Metrics: Validity (scTM) Diversity Novelty Sequence recovery AFScore GerNA-Bind score Vina score RMSD
Baselines: RNA-FrameFlow RNA-FrameFlow-R Random sequences folded with RhoFold
Builds on: RNA-FrameFlow flow matching invariant point attention AlphaFold2 ResNet torsion prediction
Closest papers by meaning
- RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (ICML 2024 · similarity 0.88)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.88)
- TANGO: RNA Topology and Geometry Co-Design (NeurIPS 2026 · similarity 0.86)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.86)
- BAnG: Bidirectional Anchored Generation for Conditional RNA Design (ICML 2025 · similarity 0.86)
- Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective (ICML 2024 · similarity 0.85)
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.