Research map / Probabilistic and scientific modeling / Molecules and proteins
RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based 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 RNAFlow, a flow-matching model for protein-conditioned RNA sequence-structure co-design that uses a fine-tuned inverse folding network as the score predictor with a frozen RoseTTAFold2NA folding module, plus a trajectory-based inverse folding variant.
Problem
Designing RNA sequences and structures conditioned on a target protein is hard due to RNA conformational flexibility, and fine-tuning large structure prediction networks for generation is computationally expensive.
Approach
Use an RNA inverse folding model (Noise-to-Seq) as the flow matching denoiser: it predicts a sequence from a noised protein-RNA complex, which is folded by a frozen pre-trained RF2NA to produce the structure, so only the inverse folding model is trained. A Traj-to-Seq multi-graph model then re-inverse-folds over the inference trajectory to capture conformational ensembles.
Key results
- On the RF2NA split, RNAFlow-Base + Rescore reaches 10.61 RMSD and 0.53 lDDT versus 14.82 RMSD and 0.34 lDDT for Conditional MMDiff.
- RNAFlow-Traj + Rescoring achieves 0.37 sequence recovery on the RF2NA split versus 0.27 for the LSTM baseline and 0.25 for random.
- On the GRK2 motif-scaffolding task, RNAFlow-Trajectory reaches 7.09 RMSD and 0.54 recovery versus 9.19 RMSD and 0.32 recovery for MMDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove structure MSE loss (ablation) | RNAFlow-Base · Sequence Similarity Split | RMSD: +1.25 RMSD when removed | Helped |
| remove structure MSE loss (ablation) | RNAFlow-Base · Sequence Similarity Split | Recovery: decreases by 1 percentage point when removed | Helped |
| remove protein conditioning (ablation) | RNAFlow-Base · Sequence Similarity Split | RMSD: +2.52 RMSD when removed | Helped |
| remove protein conditioning (ablation) | RNAFlow-Base · Sequence Similarity Split | Recovery: drops by 7 percentage points when removed | Helped |
| use the true RNA centroid and protein orientation instead of the pose guess (ablation) | RNAFlow-Base · Sequence Similarity Split | RMSD: 0.67 better | Helped |
| use the true RNA centroid and protein orientation instead of the pose guess (ablation) | RNAFlow-Base · Sequence Similarity Split | Recovery | No clear effect |
| use ten rather than five flow matching steps (ablation) | RNAFlow-Base · Sequence Similarity Split | RMSD: 0.25 better | Helped |
| use ten rather than five flow matching steps (ablation) | RNAFlow-Base · Sequence Similarity Split | Recovery | No clear effect |
| add Traj-to-Seq trajectory conditioning | RNAFlow-Base · RF2NA Pre-Training Split | Recovery: 0.30 ± 0.02 -> 0.31 ± 0.01 | Helped |
| add Traj-to-Seq trajectory conditioning | RNAFlow-Base · Sequence Similarity Split | Recovery: 0.30 ± 0.01 -> 0.28 ± 0.01 | Hurt |
| add output rescoring | RNAFlow-Base · RF2NA Pre-Training Split | Recovery: 0.30 ± 0.02 -> 0.33 ± 0.02 | Helped |
| add output rescoring | RNAFlow-Base · Sequence Similarity Split | Recovery: 0.30 ± 0.01 -> 0.32 ± 0.03 | Helped |
Surprising findings
- RNAFlow-Trajectory has higher RMSD than RNAFlow-Base on the structure generation tasks, attributed to earlier trajectory structures.
- On short single-stranded RNAs, RNAFlow often does not outperform random sequences, according to the authors.
Limitations
- Large gap in structural accuracy between RNAFlow designs and RF2NA folding upper bound.
- Performance depends on accuracy of protein-RNA folding and docking models, limiting modeling of side-chain-dependent RNAs such as ribozymes.
- Designs have not been validated experimentally.
Experimental setup
Datasets: PDBBind 2020 RNASolo
Benchmarks: PDBBind RF2NA split PDBBind sequence similarity split GRK2 aptamer motif-scaffolding (PDB 3UZS)
Metrics: RMSD lDDT native sequence recovery sequence novelty structure novelty
Baselines: Conditional MMDiff LSTM Random RF2NA (upper bound)
Builds on: flow matching RoseTTAFold2NA GVP-GNN MMDiff RFDiffusion
Closest papers by meaning
- RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching (NeurIPS 2025 · similarity 0.88)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.88)
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.87)
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.86)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.86)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 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.