Research map / Probabilistic and scientific modeling / Molecules and proteins
Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective
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.
Reformulates RNA secondary structure prediction as a probabilistic K-Rook matching problem and proposes RFold, which decomposes it into row-wise and column-wise softmax components with a Seq2map attention module, guaranteeing valid structures and running about 8x faster than UFold.
Problem
Deep learning RNA secondary structure predictors generalize poorly, are slow due to iterative constrained optimization, and often output structures violating base-pairing constraints.
Approach
Treat secondary structure prediction as selecting the most probable symmetric K-Rook arrangement, and decompose the matching into row-wise and column-wise softmax components whose product (or average) yields a valid, sparse pairing matrix without iterative constrained optimization.
Key results
- On RNAStralign, RFold reaches F1 0.977 versus 0.915 for UFold, with 100% validity.
- On ArchiveII, RFold achieves F1 0.921, above UFold's 0.905.
- RFold runs in about 0.02 s per sequence, roughly 8x faster than UFold (0.16 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace bi-dimensional optimization with E2Efold optimization (ablation) | RFold · RNAStralign test | F1: 0.977 -> 0.896 | Hurt |
| Replace bi-dimensional optimization with E2Efold optimization (ablation) | RFold · RNAStralign test | Validity: 100.00% -> 50.31% | Hurt |
| Replace bi-dimensional optimization with SPOT-RNA optimization (ablation) | RFold · RNAStralign test | F1: 0.977 -> 0.353 | Hurt |
| Replace Seq2map attention with UFold hand-crafted features (ablation) | RFold · RNAStralign test | F1: 0.977 -> 0.906 | Hurt |
| Replace Seq2map attention with SPOT-RNA outer concatenation (ablation) | RFold · RNAStralign test | F1: 0.977 -> 0.913 | Hurt |
| Remove column-wise component (ablation) | RFold · RNAStralign test | F1: 0.973 -> 0.977 | Helped |
| Remove row-wise component (ablation) | RFold · RNAStralign test | F1: 0.973 -> 0.977 | Helped |
| Remove row-wise and column-wise components (ablation) | RFold · RNAStralign test | F1: 0.995 -> 0.977 | Hurt |
| Remove column-wise component (ablation) | RFold · ArchiveII | F1: 0.914 -> 0.921 | Helped |
| Remove row-wise component (ablation) | RFold · ArchiveII | F1: 0.914 -> 0.921 | Helped |
| Remove row-wise and column-wise components (ablation) | RFold · ArchiveII | F1: 0.025 -> 0.921 | Hurt |
| Remove column-wise component (ablation) | RFold · bpRNA-TS0 | F1: 0.637 -> 0.644 | Helped |
Surprising findings
- Ablating either row-wise or column-wise component alone yields nearly identical performance, suggesting the trained outputs are almost symmetric.
- Enforcing strict validity lowers recall (e.g., 0.635 on bpRNA-TS0), since the model is cautious about predicting pairs.
Limitations
- Strict hard constraints reduce recall relative to precision.
- Validity filtering removes some true non-canonical or abnormal pairs; soft variant only partially addresses this.
- Benchmark datasets contain abnormal samples that violate the constraints, so 100% validity may not reflect real structures.
Experimental setup
Datasets: RNAStralign ArchiveII bpRNA bpRNA-new Rfam 14.2
Benchmarks: RNAStralign ArchiveII bpRNA-TS0 bpRNA-new
Metrics: precision recall F1 validity inference time
Baselines: Mfold RNAfold RNAstructure CONTRAfold LinearFold CDPfold E2Efold UFold SPOT-RNA MXfold2 RTfold Contextfold
Builds on: E2Efold UFold SPOT-RNA RTfold U-Net
Closest papers by meaning
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.90)
- PriFold: Biological Priors Improve RNA Secondary Structure Predictions (AAAI 2025 · similarity 0.90)
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.86)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.85)
- RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching (NeurIPS 2025 · similarity 0.85)
- RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching (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.