Research map / Probabilistic and scientific modeling / Molecules and proteins
PriFold: Biological Priors Improve RNA Secondary Structure Predictions
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 PriFold, an RNA secondary structure predictor that adds a pairing-frequency positional factor to axial attention and a covariation-based sequence augmentation, pretrained on MARS with masked language modeling.
Problem
Deep-learning RNA secondary structure predictors focus on local structural constraints and ignore global pairing statistics and evolutionary covariation signals.
Approach
Inject pairing-frequency statistics into axial attention scores through a positional factor matrix, and augment training data by substituting paired nucleotides according to covariation-derived transition probabilities among canonical pairs.
Key results
- On RNAStrAlign, PriFold reaches F1 0.988 versus 0.977 for RFold.
- On ArchiveII (cross-dataset generalization from RNAStrAlign), PriFold reaches F1 0.952 versus 0.921 for RFold.
- On bpRNA-TS0, PriFold reaches F1 0.770 versus 0.748 for ERNIE-RNA; ablation shows removing augmentation drops F1 from 0.770 to 0.737.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PriFold instead of RFold | RNA secondary structure prediction · RNAStrAlign | F1: 0.977 -> 0.988 | Helped |
| use PriFold instead of RFold | RNA secondary structure prediction · RNAStrAlign pseudoknot test set | F1: 0.918 -> 0.944 | Helped |
| use PriFold instead of RFold | RNA secondary structure prediction · ArchiveII | F1: 0.921 -> 0.952 | Helped |
| use PriFold instead of ERNIE-RNA | RNA secondary structure prediction · bpRNA-TS0 | F1: 0.748 -> 0.770 | Helped |
| add covariation-based data augmentation (ablation) | PriFold · bpRNA-TS0 | F1: 0.737 -> 0.770 | Helped |
| add the positional factor alongside covariation-based data augmentation (ablation) | PriFold · bpRNA-TS0 | F1: 0.730 -> 0.770 | Helped |
| add the PriFold prediction head (ablation) | RNA-FM · bpRNA-TS0 | F1: 0.680 -> 0.702 | Helped |
| add PriFold prediction head and data augmentation (ablation) | RNA-FM · bpRNA-TS0 | F1: 0.680 -> 0.744 | Helped |
| set positional-factor λ to 0.01 rather than 0 (ablation) | PriFold · bpRNA-TS0 | F1: 0.728±0.18 -> 0.737±0.23 | Helped |
| use λ=0.1 rather than λ=0 (ablation) | PriFold positional factor · bpRNA-TS0 | F1: 0.728±0.18 -> 0.733±0.33 | Helped |
| use λ=0.001 rather than λ=0 (ablation) | PriFold positional factor · bpRNA-TS0 | F1: 0.728±0.18 -> 0.733±0.17 | Helped |
| replace covariation-based augmentation with random replacement (ablation) | PriFold · bpRNA | F1: 0.745 -> 0.770 | Hurt |
Surprising findings
- Random-replacement augmentation (F1 0.745) and C-G exclusive replacement (0.753) also improve over no augmentation, though less than the covariation-based scheme.
- Applying the PriFold head and augmentation to RNA-FM raised its bpRNA F1 from 0.680 to 0.744.
Limitations
- Augmentation is restricted to canonical WC and Wobble pairs and ignores non-canonical pairs.
- Pairing-frequency statistics are taken from PDB and may not reflect the target distribution.
- Gains on bpRNA are modest (F1 0.770 vs 0.748) and depend on a single augmentation ratio tuned on the test set.
Experimental setup
Datasets: MARS RNAStrAlign ArchiveII bpRNA PDB pairing frequencies
Benchmarks: RNAStrAlign ArchiveII bpRNA-TS0 RNAStrAlign pseudoknot subset
Metrics: precision recall F1
Baselines: MFold CONTRAfold RNAstructure RNAfold LinearFold E2Efold UFold RFold SPOT-RNA RNA-FM RiNALMo ERNIE-RNA
Builds on: axial attention (Franke et al. 2023) UFold E2Efold RNA-FM
Closest papers by meaning
- Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective (ICML 2024 · similarity 0.90)
- Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem (ICML 2024 · similarity 0.87)
- RDesign: Hierarchical Data-efficient Representation Learning for Tertiary Structure-based RNA Design (ICLR 2024 · similarity 0.86)
- CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction (AAAI 2025 · similarity 0.86)
- gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.85)
- BAnG: Bidirectional Anchored Generation for Conditional RNA Design (ICML 2025 · similarity 0.84)
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.